Asian Development Review: Volume 39, Number 1

World Scientific

Vol. 39 • N

o. 1 • M

arch 2022ISSN: 0116-1105

Asian Development ReviewThe Asian Development Review (ADR) is a professional journal that publishes research on development issues relevant to the Asia and Pacific region, specifically members of the Asian Development Bank. The ADR publishes high-quality empirical papers, survey articles, historical analyses, and policy-oriented work. The ADR bridges theoretical work and rigorous empirical studies that advance the understanding of Asia’s development; and it is open to discussions of alternative perspectives on all aspects of development, including globalization, inequality, structural transformation, and poverty. The ADR aims to disseminate the best research, with contributions from scholars in all fields, with the objective of providing the knowledge necessary for designing, implementing, and sustaining effective development policies. Its intended audience comprises a worldwide readership of economists and other social scientists.

Asian Development ReviewVolume 39 · Number 1 · March 2022Asian Development Review

MCI (P) 076/09/2021

Volume 39 2022 Number 1

Mini Symposium on Demographic Change andHuman Capital in Asia 1

Guest Editor: Isaac Ehrlich

A Cross-Country Comparison of Old-Age FinancialReadiness in Asian Countries versus the United States:The Case of Japan and the Republic of Korea 5

Isaac Ehrlich and Yong Yin

Educational Gradients in Disability among Asia’s FutureElderly: Projections for the Republic of Korea andSingapore 51

Cynthia Chen, Jue Tao Lim, Ngee Choon Chia, Daejung Kim,Haemi Park, Lijia Wang, Bryan Tysinger, Michelle Zhao,Alex R. Cook, Ming Zhe Chong, Jian-Min Yuan, Stefan Ma,Kelvin Bryan Tan, Tze Pin Ng, Koh Woon-Puay, Joanne Yoong,Jay Bhattacharya, and Karen Eggleston

Cognitive Functioning among Older Adults in Japan andOther Selected Asian Countries: In Search of a Better Wayto Remeasure Population Aging 91

Naohiro Ogawa, Taiyo Fukai, Norma Mansor, andNurul Diyana Kamarulzaman

Demographic Change, Economic Growth, and Old-AgeEconomic Security: Asia and the World 131

Andrew Mason, Sang-Hyop Lee, and Donghyun Park

Trends in Employment and Wages of Female and MaleWorkers in India: A Task-Content-Of-OccupationsApproach 169

Shruti Sharma

(Continued )

March 23, 2022 1:06:35pm WSPC/331-ADR content ISSN: 0116-1105FA1

EDITORS

ALBERT PARK, Asian Development BankTETSUSHI SONOBE, Asian Development Bank Institute

MANAGING EDITOR

JESUS FELIPE, Asian Development Bank

EDITORIAL TEAM MEMBERS

GEMMA ESTHER B. ESTRADA, Asian Development BankORLEE P. VELARDE, Asian Development Bank

EDITORIAL BOARD

KYM ANDERSON, University of AdelaidePREMA-CHANDRA ATHUKORALA,

Australian National UniversityKLAUS DESMET, Southern Methodist UniversityJESUS FELIPE, Asian Development BankNEIL FOSTER-MCGREGOR, UNU-MERITSHIN-ICHI FUKUDA, The University of TokyoRANA HASAN, Asian Development BankSUNG JIN KANG, Korea UniversityHONGBIN LI, Stanford University

XIN MENG, Australian National UniversityAHMED MUSHFIQ MOBARAK, Yale UniversityNANCY QIAN, Northwestern UniversityKRISLERT SAMPHANTHARAK, University of

California, San DiegoKUNAL SEN, UNU-WIDER and The University of

ManchesterAYA SUZUKI, The University of TokyoMAISY WONG, University of PennsylvaniaJOSEPH E. ZVEGLICH, JR., Asian Development Bank

The Asian Development Review is a professional journal for disseminating the results of economic anddevelopment research relevant to Asia. The journal seeks high-quality papers done in an empirically rigorous way.Articles are intended for readership among economists and social scientists in government, private sector,academia, and international organizations.

The views expressed in this publication are those of the authors and do not necessarily reflect the views andpolicies of the Asian Development Bank (ADB), the Asian Development Bank Institute (ADBI), the ADB Boardof Governors, or the governments they represent.

ADB and ADBI do not guarantee the accuracy of the data included in this publication and accept noresponsibility for any consequence of their use.

By making any designation of or reference to a particular territory or geographic area, or by using the term“country” in this document, ADB and ADBI do not intend to make any judgments as to the legal or other status ofany territory or area.

Please direct all editorial correspondence to the Managing Editor, Asian Development Review, EconomicResearch and Regional Cooperation Department, Asian Development Bank, 6 ADB Avenue, Mandaluyong City,1550 Metro Manila, Philippines. E-mail: [email protected].

Notes: In this publication, \$" refers to United States dollars, unless otherwise stated.ADB recognizes \China" as the People’s Republic of China, \Korea" and \South Korea" as the Republic of

Korea, \United States of America" as the United States, and \Vietnam" as Viet Nam.For more information, please visit the website of the publicationat www.adb.org/publications/series/asian-development-review.

March 23, 2022 3:29:46pm WSPC/ADB ISSN: XXXX-XXXX

CONTENTS — (Continued )

Open Submissions

Disability and Intrahousehold Investment Decisions inEducation: Empirical Evidence from Bangladesh 201

Kamal Lamichhane and Takaki Takeda

The Social Costs of Success: The Impact of World TradeOrganization Rules on Insulin Prices in Bangladesh uponGraduation from Least Developed Country Status 239

Md. Deen Islam, Warren A. Kaplan, Veronika J. Wirtz, andKevin P. Gallagher

Institutions and the Rate of Return on Cattle: Evidencefrom Bangladesh 281

Kazi Iqbal, Kazi Ali Toufique, and Md. Wahid Ferdous Ibon

Impacts of Fuel Subsidy Rationalization on SectoralOutput and Employment in Malaysia 315

Noorasiah Sulaiman, Mukaramah Harun, andArief Anshory Yusuf


Asian Development ReviewVolume 39 · Number 1 · March 2022

World Scientific













Shruti Sharma

(Continued )


CONTENTS — (Continued )

Open Submissions

Disability and Intrahousehold Investment Decisions inEducation: Empirical Evidence from Bangladesh 201

Kamal Lamichhane and Takaki Takeda

The Social Costs of Success: The Impact of World TradeOrganization Rules on Insulin Prices in Bangladesh uponGraduation from Least Developed Country Status 239

Md. Deen Islam, Warren A. Kaplan, Veronika J. Wirtz, andKevin P. Gallagher

Institutions and the Rate of Return on Cattle: Evidencefrom Bangladesh 281

Kazi Iqbal, Kazi Ali Toufique, and Md. Wahid Ferdous Ibon

Impacts of Fuel Subsidy Rationalization on SectoralOutput and Employment in Malaysia 315

Noorasiah Sulaiman, Mukaramah Harun, andArief Anshory Yusuf


Mini Symposium on Demographic Changeand Human Capital in Asia

Older households in aging populations around the world, especially in the Asian

countries, are facing rising challenges in maintaining their overall financial

independence and well-being. The challenges are rising because of interrelated

demographic developments. The basic one is the “demographic transition” consisting

of rising life expectancy and declining fertility, which together contribute to a faster

rise in population aging. Moreover, this challenge is particularly acute in connection to

the rising percentage of older-age groups in the population. Japan stands out relative to

the world with 28% of the population aged 65 and over. The pace of median age

growth is particularly high in East Asia, where it has been rising faster than in any

other region of the world, except Africa. Even more significant is that this pace is

forecasted to remain the highest in the world through 2040.

Is the aging of the population a blessing or a challenge? The blessing is obvious:

living longer may be the most natural and oldest human aspiration and is one of the

most popular blessings in any language. Higher longevity also tends to be the result of

technological breakthroughs that bolster health conditions at all chronological ages.

At the same time, this process creates challenges concerning the impact of the fast pace

of population aging on labor market outcomes, income distribution, financial

preparedness of older-age workers and retirees, the solvency of defined-benefits

public pension systems and public health insurance programs, and their unintended

consequences.

Most of the papers in this issue attempt to explore the broad question in the

preceding paragraph by focusing on one or more of the channels of old-age support

that affect individuals’ ability to maintain their wealth and wellness as they age, in the

aging populations of Asia and the Pacific economies. These channels are equivalent to

four general types of “old-age insurance.” The first is self- or family-insurance.

This channel refers to the ability of old parents to rely on financial support and care

from their extended family—mainly from their adult children. A second “insurance”

February 18, 2022 3:18:50pm WSPC/331-adr 2202001 ISSN: 0116-11052ndReading

Asian Development Review, Vol. 39, No. 1, pp. 1–3DOI: 10.1142/S0116110522020012

© 2022 Asian Development Bank andAsian Development Bank Institute.

https://dx.doi.org/10.1142/S0116110522020012

https://dx.doi.org/10.1142/S0116110522020012

channel is the self-protection and management of privately owned assets that include

guaranteed savings plans and investments in a range of risky assets, such as

commercial property and real estate as well as corporate bonds and stocks that offer

variable returns. A third channel of support involves reliance on employer-based

private pension plans and annuities, as well as private health insurance, which are

forms of “market insurance.” A fourth and major channel of old-age support is “public

insurance” programs in the form of defined-benefits pension systems or

defined-contribution provident funds, and public health insurance programs that

have emerged in more recent decades in the Asian welfare states.

Ehrlich and Yin pursue a cross-country comparison of the financial readiness

of elderly people who are near or at their retirement phase in three developed

economies—Japan, the Republic of Korea (ROK), and the United States. Using the

generally harmonized respective samples from the longitudinal micro-datasets in each

of them, they find that the two Asian economies, especially the ROK, face acute

challenges due to their sharply rising life expectancies along with sharply falling total

fertility rates. By comparing the retirement income systems in the three economies, the

authors find that older Americans benefit from a more developed and better-funded

private pension system than their Japanese and Korean counterparts. They also

compare the degree to which households contribute to their own old-age financial

preparedness through their holding of risky financial assets, which is higher in the

United States.

Chen et al. address the roles of educational attainments and population aging in

determining the health status of future cohorts of older individuals in two Asian

countries experiencing high rates of population aging—the ROK and Singapore.

The idea is to identify the independent role that each of these factors plays in

explaining future disability measures of the elderly population in these countries.

To conduct the study, the authors first forecast the functional status and disability level

among future cohorts of older adults and the disparities in disability prevalence by

educational attainments in 2050. They find a larger increase in the rate of disability

levels and disparities among individuals in the ROK relative to Singapore.

However, when the authors account for the independent role of the higher aging

rate in the ROK relative to Singapore’s, the differences in the trends of disability levels

and disparities disappear. These results indicate that aging has a distinct and

independent positive influence on future disability. The results imply that continuous

aging will necessitate greater medical and caregiving expenditure on older-age groups

in the population.

2 ASIAN DEVELOPMENT REVIEW


Ogawa et al. address an important question in aging research, namely the

measure of population aging itself. The conventional measure is the percentage of the

population aged 65 and over. The authors propose a new index to compute old age,

which they call Cognition-Adjustment Dependency Ratio (CADR). The idea is

interesting since it calls for a shift from a static measure of “aging,” based on a

chronological age, to a more dynamic and variable measure, based on a cognitive

assessment of when it begins. The authors’ main result is that Japan, which has the

largest level of aging by the conventional definition, exhibits a contemporary pattern

of age-related decline in cognitive functioning that is highly comparable to those of

many other developed nations, particularly in the Continental Europe. The authors

interpret the finding to imply that Japan’s “elderly population” is cognitively younger

in age by the CADR index relative to what the conventional definition of

chronological aging would imply.

Mason, Lee, and Park focus on the demographic transition in the Asia and

Pacific economies and the age-related labor income and consumption profiles of

individual households in these economies by studying the National Transfer Accounts,

related administrative data, and system of national accounts. The analysis provides

estimates of the effective labor available to the economy from earnings of the

working-age groups to support income and spending necessary to provide public

funding for the consumption needs of the elderly retirees. The authors’ analysis

indicates that the intensifying increases in life expectancy and imploding fertility rates

generate a potential demographic time bomb in countries like the ROK, where labor

supply and economic growth are raising significantly the cost of the public pension

plans. This implies that old-age support would need to come from some of the other

old-age support channels.

Sharma likewise deals with the role of technological changes that affect the

cognitive content of occupations within and across industries in measuring

employment and wage levels of male and female workers in India over the period

of 1994–2014. The paper uses a task-based approach to analyze labor trends in India

and its possible implications for the measurement of labor productivity, the gender

wage gap, and economic growth.

Isaac Ehrlich

Guest Editor

State University of New York at Buffalo

3


A Cross-Country Comparison of Old-AgeFinancial Readiness in Asian Countriesversus the United States: The Caseof Japan and the Republic of Korea

ISAAC EHRLICH AND YONG YIN¤

We pursue a cross-country comparison of relative financial readiness of olderhouseholds in Japan and the Republic of Korea relative to the United States.Our comparative analysis, using macro-level and harmonized longitudinalhousehold financial data, covers the principal financial channels of old-agesupport: public and private pension plans, family support, andself-management of private financial portfolios. We find that while all threecountries have similar public pension systems, older Americans benefit frommore developed and better-funded public and private pension systems, as wellas individual management of risky financial portfolios. We find thateducational and health attainments of household heads and household wealthlead to a greater tendency to hold and manage risky assets. Our decompositionanalysis also shows that the gap in stock ownership in Asian countries relativeto the United States can be attributed to lower levels of development infinancial and pension markets. However, these gaps have been shrinking morerecently.

⁄Isaac Ehrlich (corresponding author): School of Management, State University of New York at Buffalo.Email: [email protected]; Yong Yin: Department of Economics, State University of New York atBuffalo. Email: [email protected]. This analysis uses data or information from the Harmonized JapaneseStudy of Aging and Retirement (JSTAR) dataset and Codebook, Version B as of August 2014, and fromthe Harmonized Korean Longitudinal Study of Aging (KLoSA) dataset and Codebook, Version C as ofJune 2019, both developed by the Gateway to Global Aging Data. The development of the HarmonizedJSTAR and KLoSA was funded by the National Institute on Ageing (R01 AG030153, RC2 AG036619,R03 AG043052). For more information, please refer to www.g2aging.org. The Asian Development Bankrecognizes “Korea” as the “Republic of Korea.”

This is an Open Access article published by World Scientific Publishing Company. It is distributed underthe terms of the Creative Commons Attribution 3.0 International (CC BY 3.0) License which permits use,distribution and reproduction in any medium, provided the original work is properly cited.

March 23, 2022 10:30:39am WSPC/331-adr 2250004 ISSN: 0116-11052ndReading



https://dx.doi.org/10.1142/S0116110522500044

https://dx.doi.org/10.1142/S0116110522500044

Keywords: education, financial markets development, health and wealth,household portfolio management, old-age financial preparedness and inclusion,public and private pensions

JEL codes: G11, G12, G51

I. Introduction

The financial well-being and inclusion of older age groups and retirees have long

been issues of academic interest as well as public concern in the modern welfare states.

The concern is based on the idea that large segments of the population may not be able

to adequately secure their financial well-being after retirement through their

accumulated savings, and thus need to be supplemented by private (typically

employer- or employee-based plans) and public pension systems. The traditional

private old-age support system has consisted largely of informal “family insurance”

arrangements whereby adult children or extended family members would provide care

and material support for their parents and close relatives. With the development of

modern enterprises and financial markets, old-age support started coming from

employer severance payments to retirees, and employer-based or private pension

funds. To better deal with old-age financial needs, however, welfare states, starting

with Germany in 1889 and the United States (US) in 1935, have established old-age

social insurance programs to supplement the private old-age support mechanisms.

According to the Organisation for Economic Co-operation and Development (OECD

2019), the combined old-age financial support systems in the OECD countries—

private and public—provide a gross pension replacement rate of 49% for men and

48.2% for women earning average incomes. Asian countries have also developed such

systems, but largely only after World War II.

The retirement-income support systems in both Western and Asian countries

have been facing increasing financial vulnerability, however, due to the ongoing

demographic transition process and some slowdown in global economic growth,

especially in developed countries. Total fertility rates, measured as the average number

of children a representative woman would have over her childbearing years, have

fallen significantly from 2.82 during the 5-year period ending in 1955 to 1.57 in the

5-year period ending in 2000 in the more developed region (MDR).1 They have risen

1The more developed region is a term used by the United Nations. The region includes all Europeanand North American countries, plus Australia, New Zealand, and Japan.



only modestly since then and are predicted to reach 1.71 in 2040 (Figure 1). Over the

same period, life expectancy at age 65 in the MDR increased from an average of 13.46

years in 1950–1955 to 16.12 years in 1985–1990 before accelerating to 19.57 years in

2015–2020 (Figure 2). These aging trends have contributed, in turn, to a steady

increase in the old-age to working-age dependency ratio, which is measured as the

ratio of the number of people aged 65 and over for every 100 people aged 20 to 64 in

1950. This number has risen from 13.6 in 1950 to 21 in 1990, and it was predicted to

have risen to 32.7 in 2020 and to increase to an alarming 46.3 by 2040 (Figure 3).

In the US, the inversely related potential old-age support ratio fell from 6 in 1960

to an expected 3.6 in 2020, while in Japan it shrunk from 8.7 to 2.2 in 2015 (Mather

and Kilduff 2020, Knoema 2020). This has put tremendous financial stress on both

public and private defined-benefits pension systems around the world. The challenges

have led to reforms in retirement-income systems in recent years, including increases

in the eligible age for full social security, more use of means tests, and a shift from

defined-benefit to defined-contribution pension schemes in both the public and private

sectors.

One of the potential remedies that could mitigate the increasing prospect of

financial insolvency of private and public pension programs is for households to

Figure 1. Total Fertility Rates: 1955–2040

Note: More developed region includes all European and North American countries plus Australia, New Zealand,and Japan.Source: United Nations. World Population Prospects 2019. https://population.un.org/wpp/ (accessed 3 September2020).

A CROSS-COUNTRY COMPARISON OF OLD-AGE FINANCIAL READINESS 7


improve the management of their individually held retirement assets. Households can

meet this challenge by raising their participation in risky financial markets and

increasing their demand for shares of stocks and corporate bonds they hold in their

individual portfolios. However, the current participation rate of households in risky

financial markets remains low even in the most developed economies. Using the US

Health and Retirement Study (HRS), Ehrlich and Yin (2021) report that just 24% to

34% of American households hold stocks.2 Venti and Wise (2001) used the first wave

of HRS data to illustrate that a considerable amount of the dispersion in wealth among

US households is due to investments in less risky assets such as bonds or bank saving

accounts as opposed to more risky assets such as stocks and corporate bonds.

The recent changing structure of employer-based pension plans in many

countries from defined-benefits to defined-contribution schemes has two opposite

impacts on employees. On the one hand, this change shifts the onus of financial

management of assets—the selection, allocation, and management of the household

Figure 2. Life Expectancy at Age 65: 1955–2040

Note: More developed region includes all European and North American countries plus Australia, New Zealand,and Japan.Source: UnitedNations. “World Population Prospects 2019.” https://population.un.org/wpp/ (accessed 3 September2020).

2Ehrlich, Hamlen, and Yin (2008) report similar percentages using the Survey of ConsumerFinance. It is interesting to note that Ehrlich and Shin (2021) report similar percentages for Europeanhouseholds (24%) using the Survey of Health, Aging and Retirement in Europe (SHARE).



financial portfolios—from employers to employees, thrusting unprepared employees

into a bigger, challenging role. On the other hand, such experiences might exert

offsetting effects going forward by preparing the same employees to better manage

their retirement portfolio during the retirement.

The need for better asset management during retirement has also intensified

because retirees are now facing longer life spans due to continued increases in life

expectancy, especially at older age brackets. This would increase the risk that

aging households would become totally dependent on government old-age pension

plans when they deplete their portfolio of assets. This factor impacts especially

those households that reach retirement age with a relatively low level of accumulated

wealth.

To understand the mechanism of an individual’s asset management choices and

potential portfolio outcomes, Ehrlich, Hamlen, and Yin (2008) developed a theoretical

framework for asset management. This framework was extended in a later study

(Ehrlich and Shin 2010; Ehrlich, Shin, and Yin 2011), focusing on the willingness

to hold both domestic and foreign risky assets and the role of education and

health in asset accumulation by older age groups in the US using HRS data

Figure 3. Demographic Old-Age to Working-Age Dependency Ratio: 1950–2040

Notes: The ratio is defined as the number of individuals aged 65 and over per 100 people between the age of 20and 64. More developed region includes all European and North American countries plus Australia, New Zealand,and Japan.Source: UnitedNations. “World Population Prospects 2019.” https://population.un.org/wpp/ (accessed 3 September2020).



(Ehrlich and Yin 2021). In this paper, we greatly extend the same framework to study

simultaneously the factors affecting the demand for holding risky assets in two Asian

countries—Japan and the Republic of Korea (ROK)—and the US. One reason for

selecting these countries is that they represent three OECD countries in different stages

of development, with the US having the most developed and experienced financial

industry, Japan ranking second behind the US, and the ROK having the least

developed financial industry. Another important reason is that these three countries

have a relatively comparable micro-level longitudinal survey-type data (Health and

Retirement Study [HRS], the Japanese Study of Aging and Retirement [JSTAR], and

the Korean Longitudinal Study of Aging [KLoSA]), although the Japanese survey

differs somewhat by specific characteristics, which we explore in later sections.

The macro-level analysis leads to a better understanding of the role of the individual

determinants of demand for risky asset holding relative to that of the market-level

determinants. More specifically, our analysis has five major objectives:

(i) analyzing and comparing the macro-level accumulated components of wealth by

type of asset across the three countries along with comparing the overall

characteristics of investors in the three countries;

(ii) reviewing the role of old-age pension systems—both private and public—in

supporting the financial well-being of retirees;

(iii) estimating the role of individual investors’ characteristics in explaining household

willingness to hold risky financial assets to achieve better outcomes in

accumulating total and financial assets near and during retirement;

(iv) decomposing the role of individual characteristics, as predicted by our theoretical

model, relative to the role of financial market development in explaining gaps in

households’ willingness to hold risky assets across different countries, by using

an econometric decomposition analysis; and

(v) deriving some general policy implications based on our findings from the

preceding objectives.

The remaining sections of this paper are organized as follows. In Section II, we

briefly describe the retirement-income systems and demographic characteristics of the

three countries we analyze. In Section III, we describe the data surveys used in our

analysis. Section IV contains the basic descriptions of household characteristics,

retirement decisions, income sources, and wealth distributions in these three countries.

In Section V, we develop our theoretical model of asset management and apply it

empirically via a regression analysis using micro data and through a decomposition

analysis. We conclude with the main policy implications of our findings.



II. Comparing the Retirement-Income Systems in Japan,

the Republic of Korea, and the United States

We compare the retirement-income systems in Japan, the ROK, and the US

in terms of their completeness—the combination of both public and private

components—supporting demographic factors, and funding strength.

A. Japan

Japan has a two-tier public pension plan (Rajnes 2007, Social Security

Administration [SSA] 2008): the employees’ pension insurance (EPI), which was

established by law in 1941 and later updated in 1954; and the national pension (NP)

law, which was introduced in 1959 and amended in 1985. The current system can be

described as a social insurance system, which involves a flat-rate benefit for all

residents under the national pension program (first tier) and earnings-related benefits

under the employees’ pension insurance program or other employment-related program.

The full pensionable age for NPwas 60 years old in 2001 but started to increase by 1 year

every 3 years to 65 by 2013. The early pensionable age is currently 60. For EPI, the

eligible age was also 60 in 2001. But starting in 2013 it increased by 1 year every 3 years

for men until it reaches 65 in 2025. For women, the eligible age also started rising in 2018

by 1 year every 3 years until it reaches 65 in 2030. The full benefit requires a minimum of

40 years of contributions, and reduced benefit requires at least 25 years of contributions.

Both NP and EPI also provide disability insurance depending on meeting disability

criteria as well as a contribution requirement (SSA 2008).

The private sector in Japan also offers voluntary employer-sponsored retirement

plans, with retirement typically mandated at age 60. There had been two

defined-benefit (DB) plans offered, both receiving preferential tax treatment (Rajnes

2007). Two laws were passed in 2001 with the aim of reforming these plans. The DB

plans went through a major overhaul to address underfunding problems and allow for

convertibility to a DC plan. In addition, influenced by the expanding US economy and

stock market boom of the 1990s that was enhanced by the growth of DC plans,

primarily 401(k)’s in the US, a separate legislation established both corporate and

individual defined-contribution (DC) plans with the aim of stimulating the flow of

individual retirement-account assets into Japanese financial markets.

In 2019, the gross pension replacement rates for Japanese retirees with mean

earnings were projected by OECD pension models to be 32.5% for public plans and

23.8% for private plans (OECD 2019). The total gross replacement rate, which

includes total mandatory and voluntary plans, is thus 55.8%. By comparison, the



OECD average replacement rate for public plans and private, voluntary plans are

39.6%, and 55.2%, respectively. Compared with other OECD countries, the Japanese

public system is less generous, but it offers better overall income security by

combining private and public sources.

B. Republic of Korea

The ROK has a national pension (NP) system for all citizens except government

employees, private teachers, military personnel, and employees of the special post

office, who are covered in a separate, special occupational pension program. Both

plans are mandatory. These were established by law in 1986 and later updated in 2007.

In 2008, the eligibility age for full pension benefits was set at 60 years old, requiring at

least 20 years of employment coverage. This age is set to gradually increase, however,

to 65 by 2033. The early eligibility age was 55 in 2008, but this is also set to gradually

rise to 60 by 2033 (SSA 2008). NP provides a basic old-age pension for people over

65 with an income below a maximum set by presidential order, and this benefit is not

subject to a minimum of 20 years of coverage. NP provides disability pension as well,

depending on the degree of disability assessed by the National Pension Corporation.

The private sector in the ROK used to have a mandatory severance payment

system (it started out as voluntary in 1953 and became mandatory in 1961) for

businesses with more than five employees, which provided a defined benefit in the

form of a lump-sum payment to employees. It was funded entirely by employers. This

plan was replaced by the Korean retirement benefit system (RBS) in 2005. The RBS

provides a choice between retirement pension plans (RPP) and retirement pay schemes

(RPS), which corporations can form. There are two types of RPP—DC and DB

plans—while RPS is a DB plan by design. There are also voluntary personal pension

plans in the form of individual retirements accounts (IRAs) that have been available in

the ROK since 1994 (OECD 2009).

According to OECD pension models, the gross pension replacement rate in 2019

for retirees with mean earnings in the ROK is 37.7% for mandatory public plans.

There is no estimated figure available, however, for mandatory and voluntary private

plans. Thus, the ROK offers a slightly less generous public pension system compared

to its OECD peers, but the overall gross pension replacement rate is well below the

OECD average of 55.2%.3

3As pointed out by one of our referees, the real-world situation in the ROK is even worse. TheOECD pension models make an unrealistic assumption concerning the number of years in which pensioncontributions have been made in that country. The actual number of contribution years is much lower. Thisis one of the factors contributing to the highest poverty rate among the elderly in the OECD.



C. The United States

In the US, the basic public system is a pay-as-you-go social security system. The

eligibility age for full benefits is 66 years old with at least 40 quarters of coverage, and

the eligible age will be lifted to 67 by 2027 (SSA 2018). Early pensionable age is 62

with the same limitation on coverage. The US system offers a generous disability and

supplemental income security for persons with disabilities and acute needs.

The private sector in the US offers various types of voluntary retirement plans.

Individuals can voluntarily contribute to additional retirement plans typically known

as IRAs. In the US employment-based system, pension plans were originally

dominated by DB plans. Starting around 1980, a quiet revolution took place: DC plans

quickly rose to dominance over DB plans. By 2011, 93% of covered private-sector

workers participated in a DC plan, while only 31% participated in a DB plan (Ehrlich

and Kim 2005). The sum exceeds 100% because the same individual can

simultaneously participate in multiple retirement plans.

The OECD pension models estimate that in 2019 the gross pension replacement

rate for US retirees with mean earnings is 39.4% for mandatory public plans, which is

virtually the same as the OECD average of 39.6%. However, the gross replacement

rate in the US, counting both mandatory and voluntary plans is 70.3%, well above the

OECD average of 55.2%.

D. Comparing the Demographic Changes Affecting the Retirement Systemsin the Three Countries

Figure 1 plots historical and projected total fertility rates for the three countries

under consideration along with the average for the MDR as a benchmark. As the chart

indicates, total fertility rate started to decline significantly in late 1970s for the US, but

it recovered after the 1990s until the early 2000s. It is currently projected to fall below

the replacement rate going forward, but to remain above the average of the MDR. In

Japan, the decline in total fertility rate accelerated in the 1970s, continuing into the

1990s. It is projected to be stable and to even recover somewhat in the near future,

though it will remain significantly below the average in the MDR. The ROK is

undergoing a more dramatic demographic transition. Its total fertility rate fell from

6.33 in 1960 to 2.92 in 1980 (around the time its economy started to take off ) and fell

even further to 1.5 in 2000, a level below the MDR average.4 The decline is projected

4Note that in the recent decade there were years in which the fertility rate fell below one. However,the data exhibited in Figure 1 represent five-year averages, which do not fall below 1.



to continue and then stabilize over the next two decades but will remain well below

that of MDR countries. It is noteworthy that the decline in total fertility rates in the

ROK has been more rapid than that in Japan, and that the ROK’s total fertility rate

would remain significantly lower than that of Japan in the near future.

Figure 4 plots life expectancies at birth for these three countries, along with the

average for MDR. While all countries have experienced steady increase in life

expectancy over the last 5 decades, the rise in the ROK has been more dramatic. It rose

quickly from 52.76 years in 1960 to 70.34 in 1990, and caught up with the US in 2005,

reaching 77.17. It then surpassed both the US and the average of the MDR in 2010 and

has narrowed the gap with Japan in recent years.

Life expectancy trends at age 65 for these countries are plotted in Figure 2. If age

65 is taken to be the unofficial retirement age, the data suggest that in the 1960s people

expected to live another 14 years during retirement. But by 2010 this number

increased to 19 years. For Japan, which exhibits the highest life expectancy at 65,

longevity in retirement is currently expected to be 22.44 years (in 2020) and it is

projected to reach 24.16 years in 2040. Japanese life expectancy at 65 caught up with

that of the US around 1985, much sooner than the ROK, which achieved this feat

around 2010.

Figure 4. Life Expectancy at Birth: 1955–2040

Note: More developed region includes all European and North American countries plus Australia, New Zealand,and Japan.Source: UnitedNations. “World Population Prospects 2019.” https://population.un.org/wpp/ (accessed 3 September2020).



The demographic transition in the three countries we survey—declining fertility

along with rising life expectancy—has generated sharp increases in the shares of

old-age population to total population in these countries. Figure 3 plots the old-age to

working-age dependency ratio, defined as number of individuals aged 65 and over for

every 100 people aged 20 to 64, for these countries. While the US and the ROK

dependency ratios in 2020 of 28.4 and 23.6, respectively, are slightly below the

average of 32.7 for the MDR, Japan’s ratio of 52 is well above the average. The ratio is

projected to go up to 46.3 for the average MDR in 2040. While the US would still see

a figure below this average in 2040 (39.0), both Japan and the ROK (70.7 and 61.6,

respectively) would be well above this average. Thus, based on these ratios, both the

Japanese and Korean retirement-income systems will face a lot more headwinds than

other more developed countries, including the US. This is especially true for the ROK

as it started its mandatory retirement system much later than other countries, and the

rate of increase in this ratio is almost exponential.

As for the extent to which the public pension system is funded, the US is in a

much better place compared to its peers. According to OECD (2020a), in 2019, total

assets in pension funds in the US equaled 87.5% of gross domestic product, while this

figure was only 28.4% for Japan and 11.6% for the ROK. By comparison, the OECD

average was 60.1%.

E. Implications for the Relative Financial Strengths of the Retirement Systemsin the Three Countries

Based on the preceding comparisons of the structure and demographic trends

affecting the private and public retirement systems across the three countries, it seems

apparent that among the three countries, the US has the relatively most complete and

funded public and private retirement system, including the traditional “family-support”

system, due to its relatively higher fertility rate and retirement-income support ratio. Its

demographic pattern going forward is also more favorable compared to the other two

countries. In contrast, the ROK due to its late start and rapid decline in total fertility

rate may face more serious challenges in the near term.

Table 1 reports the net pension average replacement rates of earnings for workers

earning multiples of the mean earnings in the three countries, as estimated by the

OECD pension models (OECD 2019), along with the OECD average replacement

rates serving as reference points. As the table shows, the US provides better

replacement rates of individual earnings relative to both the ROK and Japan, although

all three countries fall below the OECD average rates. Japan, due to its relatively low



reliance on the private sector, provides far less overall coverage relative to both the US

and the ROK. By contrast, the ROK offers comparable replacement rates to those in

the US only for workers earning 50% of mean earnings, while falling slightly short of

Japan in terms of the replacement rate it offers for high earners—retirees who earn

150% of average earnings. This might be a reflection of the ROK’s late start in pension

reforms in the private sector.

It is interesting to note that, even though both Japan and the ROK have higher

life expectancies at birth than the US (84.1 years for Japan and 82.63 years for the

ROK compared to 78.54 years for the US in 2017), both have set a pensionable age

2 years younger than that of the US. The pension eligible age is currently 65 years old

in Japan—the same as the planned pension eligible age in the ROK, compared to 67 in

the US. The lower eligibility ages in Japan and the ROK may accentuate the need for

future potential reforms in the Japanese and Korean systems. In contrast, the OECD

(2019) reported that in 2018, the ROK had the highest average effective age of labor

market exit (72.3 for both men and women), Japan had the second-highest averages

(70.8 for men and 69.1 for women), and the US had the third highest (67.9 for men

and 66.5 for women), while the OECD average was lower (65.4 for men and 63.7 for

women) than in these three countries. These differences translate into a remaining life

expectancy in retirement of 12.9 years for men and 16.3 years for women in the ROK,

15.5 for men and 21.0 for women in Japan, and 16.4 for men and 19.8 for women in

the US. In this regard, all three countries rank below the OECD average, where the

average remaining life expectancy in retirement is 17.8 years for men and 22.5 years

for women. These contrasting statistics suggest that both Japan and the ROK can

mitigate their disadvantages in demographics and workers’ dependencies on pension

Table 1. Net Pension Replacement Rates of Earnings for WorkersEarning Multiples of Mean Earnings

Replacement Rates of Individual Earnings,by Multiples of Mean Earnings

EconomyPensionAge 0.5 1.0 1.5

Japan 65 45.9 36.8 33.3Republic of Korea 65 60.8 43.4 32.6United States 67 61.2 49.4 42.7OECD 66.1 68.3 58.6 54.7

OECD = Organisation for Economic Co-operation and Development.Source: OECD pension models, as reported in OECD. 2019. Pensions at a Glance:OECD and G20 Indicators. Paris: OECD Publishing.



replacement rates because workers in these Asian countries tend to work longer than

those in the US, and much longer than workers in most OECD countries.

III. Data Description and Variable Construction

A. Comparative Samples and Waves

Our study relies on three longitudinal studies sharing a common questionnaire

design: the Health and Retirement Study (HRS) in the US, the Japanese Study of

Aging and Retirement (JSTAR) in Japan, and the Korean Longitudinal Study of Aging

(KLoSA) in the ROK.

The HRS is the seminal longitudinal household survey dataset that has been used

to study retirement and health issues for the elderly in the US (Juster and Suzman

1995). The original survey was conducted in 1992, covering household heads who

were aged 50 to 60 at the time. The survey followed up with the original households

every 2 years since then. A separate survey of the American oldest old population

(AHEAD) was later merged with the original HRS. To maintain a continuing dynamic

survey of the elderly population, a new cohort of households aged 50 and 55 has been

added to the longitudinal survey every 6 years. Such refresher samples were added in

1998, 2004, 2010, and 2016.

The 2015 RAND HRS file version O, which we use in this study, is the result of

several data developments aiming to provide a user-friendly version of HRS (Chien

et al. 2015). It includes final data files from 12 waves (1992 to 2010). This longitudinal

data contains only a subset of variables from the original HRS, but the survey reports

cleaned and processed variables with consistent and intuitive naming conventions and

model-based imputations. Most importantly, it includes a large number of individual

variables, including demographics, job status and history, health, as well as imputed

income and assets. The success of HRS has inspired the development of similar

surveys we use in our three-country study, as detailed in what follows. Notable surveys

also include the English Longitudinal Study of Ageing (ELSA) and the Survey of

Health, Ageing, and Retirement in Europe (SHARE). All these studies share virtually

identical questionnaires. This makes our cross-country comparison possible.

JSTAR is a panel survey of elderly people aged 50 or older conducted by the

Research Institute of Economy, Trade, and Industry (RIETI), Hitotsubashi University,

and the University of Tokyo (Ichimura, Shimizutani, and Hashimoto 2009). The

survey is designed to ensure, to the maximum extent possible, comparability with



HRS, ELSA, and SHARE. To facilitate cross-country comparisons, RIETI also created

the Harmonized JSTAR data to be compatible with the procedures and imputation

methods employed in generating the RAND HRS. The JSTAR version we use is

Version B.

JSTAR surveyed people between the age of 50 and 75 along with their partners.

The first wave was conducted in 2007 with five municipalities. The survey followed

up with these households in wave 2 in 2009. Two more municipalities were added to

the study in wave 2. All households were included in the follow-up wave 3 survey in

2011, along with three new municipalities added to the survey. The unit of analysis in

JSTAR is the household, with survey weights at the household level provided for

analysis. This differs from HRS where both personal-level and household-level survey

weights are provided.

KLoSA is a panel survey in the ROK conducted initially by the Korean Institute

of Labor. KLoSA started the first wave in 2006 for households with at least one person

45 years and older, and the respondents were then surveyed every 2 years. Starting in

wave 3, the data were collected by the Korea Employment Information Service

(KEIS). There were no refresher samples in waves 2 through 4.5 This difference may

cause a downward bias when we aggregate individual data to form the household-level

data. The data we use are the Harmonized KLoSA version C. We have obtained the

original wave data from KEIS and then used a Stata program provided by a team at the

Global Aging Project to generate the Harmonized version C data.

For comparison purposes, we used only waves 8, 9, and 10 from the RAND

HRS, and waves 1, 2, and 3 from the Harmonized JSTAR and KLoSA. We note that

KLoSA data were collected in the same year as the corresponding HRS, while JSTAR

data were collected a year later (see Table 2 for details, which also includes the total

number of units with nonzero survey weights in each wave). In this context, we note

that HRS added a fresh cohort in 2010, leading to a larger number of units compared

with the previous two waves. For KLoSA, the number of units kept falling due to

attrition, and this is also true for wave 9 of HRS compared with wave 8. In contrast,

JSTAR added more municipalities in both the second and third waves, so the number

of units in JSTAR have risen throughout the three waves. To achieve greater

consistency among all three datasets, we also conduct the analysis at the household

level. To capture individual characteristics such as age and education, however, we

also use information on household heads.

5One notable difference about KLoSA is that it does not interview spouses or partners younger than45 years old, a deviation from the treatment in HRS and JSTAR.



We need to point out briefly some differences we observe in these three micro

datasets when using the 2006–2007 wave. First, the household heads in HRS and

JSTAR have similar average ages, with HRS exhibiting a slightly higher average. This

is due to the introduction of a new cohort of individuals aged 50 and 59 into the survey

in 2004, which was the first wave of JSTAR. KLoSA reports a much lower average

age because in the first wave of that survey, the age eligibility for KLoSA was 45,

which was 10 years younger than the early pensionable age of 55 at the time.

Throughout all waves, KLoSA contains much younger households, but the gap with

HRS fell modestly in 2010 when a new cohort was added to HRS. Average age stayed

almost the same between waves 1 and 2 for JSTAR even though there were more

municipalities introduced in the second wave. However, the average age in the JSTAR

sample became older in wave 3, where more municipalities were included.

We should also note that, even though we use the harmonized version of three

micro datasets, these data have their own limitations. For example, JSTAR does not

use a national probabilistic sample. Furthermore, the micro datasets are often

inconsistent with the reported administrative data. Consequently, the data comparisons

we report below should be viewed with caution, subject to this caveat.

B. Comparative Key Variables

To construct comparable variables in nominal local currencies, such as income

and wealth, we first convert them into nominal US dollars using purchasing-power-

parity (PPP) exchange rates (OECD 2020b), and then convert them into 2010 US

Table 2. Wave, Year, and Number of Units

HRS JSTAR KLoSA

Wave 8 1 1Year conducted 2006 2007 2006Number of units with positive weights 12,086 3,521 6,763



HRS ¼ Health and Retirement Study (United States), JSTAR ¼ JapaneseStudy of Aging and Retirement, KLoSA ¼ Korean Longitudinal Study ofAging.Source: Authors’ compilation.



constant dollars by using annual US consumer price indexes (CPI) available from the

Bureau of Labor Statistics (BLS).

Table 3 reports our sample descriptive statistics for the demographic variables

we use in our analysis, along with some key income and wealth variables. A quick

glance would reveal that the three datasets remained quite stable across different

waves.

Household size. The two Asian countries share a similar family size—around

3—despite falling fertility rates, especially in Japan. However, the US has a family

size just slightly above 2. This is simply due to cultural differences since children are

more likely to live with their parents in Asian countries. In all three countries, the

household head is slightly more likely to be male, with the exception of the ROK in

KLoSA’s waves 2 and 3 where only 61% and 69% of household heads are male. We

use “coupled” as a dummy variable indicating whether a couple lives in the household.

Those living separately, divorced, or widowed are not considered coupled. In the US

sample, the coupled rate is slightly above 50% and is stable across waves. In the

Table 3. Sample Descriptive Statistics

HRS JSTAR KLoSA

2006–2007Age 65.8 65.0 59.1Household size 2.12 2.92 3.10Male 52.8% 52.0% 52.0%Coupleda 53.1% 82.4% 47.2%Educational attainmentb 2.57 1.93 1.71Retired 36.3% 14.7% 15.4%Work 44.8% 44.5% 43.5%Receiving private pension 30.3% 16.4% 3.2%Receiving public pension 59.3% 68.9% 10.9%Self-reported good health 71.8% 78.8% 50.9%Income ($’000) 76.31 38.07 23.99Financial wealth ($’000) 140.4 132.0 2.92Homeownership 77.5% 79.7% 53.2%Primary residence ($’000) 217.2 138.0 107.5Net worth ($’000) 531.9 310.0 165.4

2008–2009Age 67.1 65.3 60.6Household size 2.10 2.62 3.09Male 53.4% 58.6% 60.7%Coupled 52.2% 56.8% 47.0%

Continued.



Table 3. Continued.

HRS JSTAR KLoSA

Educational attainment 2.58 2.00 1.68Retired 39.3% 13.3% 19.8%Work 42.9% 46.6% 45.5%Receiving private pension 30.8% 18.3% 3.8%Receiving public pension 62.7% 66.9% 16.2%Self-reported good health 71.7% 78.4% 51.3%Income ($’000) 74.0 31.99 22.27Financial wealth ($’000) 134.5 97.36 7.96Homeownership 77.8% 63.6% 57.3%Primary residence ($’000) 208.4 139.7 127.8Net worth ($’000) 518.0 289.1 204.3

2010–2011Age 64.5 67.3 62.1Household size 2.20 2.70 3.07Male 52.4% 54.5% 68.9%Coupled 54.8% 74.4% 45.9%Educational attainment 2.69 2.15 1.64Retired 33.2% 15.5% 20.0%Work 47.1% 49.3% 44.7%Receiving private pension 22.7% 12.3% 4.5%Receiving public pension 55.3% 78.4% 22.1%Self-reported good health 74.2% 84.8% 48.4%Income ($’000) 71.72 34.54 21.13Financial wealth ($’000) 122.9 98.81 11.12Homeownership 76.4% 45.4% 57.3%Primary residence ($’000) 184.5 97.2 115.8Net worth ($’000) 449.9 238.7 201.5

HRS ¼ Health and Retirement Study (United States),JSTAR ¼ Japanese Study of Aging and Retirement, KLoSA ¼Korean Longitudinal Study of Aging.aCoupled is a dummy variable indicating whether a couple livesin the household. Survey respondents who are separated,divorced, or widowed are not considered coupled.bEducational attainment is classified as 1 for respondents with aneducational attainment less than a high school degree; 2 for highschool graduates; 3 for those with some college; and 4 for thosewith a college degree or higher.Source: Authors’ calculations using data from the Healthand Retirement Study (United States), the Japanese Study ofAging and Retirement, and the Korean Longitudinal Study ofAging.



Korean sample, the coupled rate is surprisingly below 50%. JSTAR reports a

somewhat unstable pattern: this rate is 82.4% in wave 1 and 74.4% in wave 3, but only

56.8% in wave 2.

Educational attainments. Only HRS reports continuous years of education. Both

JSTAR and KLoSA report only educational attainments. For compatibility reasons, we

use the following unified version of educational attainments:6 educational attainment

is classified as 1 for respondents with an educational attainment less than a high school

degree; 2 for high school graduates; 3 for those with some college; and 4 for those with

a college degree or higher. Among the three countries, the US has the highest average

educational attainment in all waves, with the average being above a high school

degree. Japan’s average educational attainment is around a high school degree, while

the ROK’s average is just below a high school degree.

Retirement status. The data concerning individual retirement status are

considerably different in the three countries. In Japan only 13.3% to 15.5% of the

population report being fully retired. In the ROK, the reported percentages are slightly

higher at 15.4% to 20%. But in the US, the reported percentages are much higher—

they range from 33.2% to 39.3%.

Wage earners. The percentage of household heads who report working to earn a

wage is very similar among the three countries, ranging from 42.9% to 49.3% across

different waves.

Public pension receivers. The percentage of households receiving public pension

is highest for Japan, ranging from 66.9% in 2008–2009 to 78.4% in 2010–2011. In the

US, the reported percentages are lower, ranging from 55.3% in 2010–2011 (wave 3) to

62.7% in 2008–2009 (wave 1). The lower number in 2010–2011 (wave 3) is due to the

inclusion of a new younger cohort. The percentages for the ROK rose steadily from

10.9% to 22.1% across the three waves as respondents became older, but they are

significantly lower than those reported in Japan and the US.

Private pension receivers. As for the data on individuals receiving private

pensions, the ROK reported dismal percentages ranging from 3.2% to 4.5%. On

the other hand, JSTAR reported percentages ranging from 12.3% to 18.3%.

The percentages are much higher in the US: around 30% received private pension

in 2006–2007 (wave 1) and 2008–2009 (wave 2) (the 22.7% figure in the third wave in

2010–2011 is due to the influx of a new, younger cohort).

Health status. HRS used a five-point scale for self-reported health status (from

poor to excellent). Both JSTAR and KLoSA included this variable in their

6KLoSA uses a slightly different version, which we convert to our unified version.



questionnaire. Following the literature and our own earlier work, we convert this

health status measure into a dummy variable for people whose self-reported status is

“good” or “better.” Around 80% of Japanese household heads reported their health

being good or better. In the US, this percentage is lower at 72%–74%. However, only

about half of Korean household heads report their health status as good or better.

Whether this difference is due to technical issues such as reporting methods or how

people assess their health status remains a question, but it may affect our regression

results concerning the expected effect of this variable.

Average household income. As expected, the US has a much higher average

income than the two Asian countries. We do see a slightly lower average in the US in

2008–2009 (wave 2) compared with 2006–2007 (wave 1) due to the 2008 financial

crisis. Average Japanese household incomes are about half of those of the US while

the Korean averages are about one-third of those of the US. Note that we use

PPP exchange rates to convert nominal figures. For Japan, this exchange rate is

typically higher than the nominal market exchange rate. This means that if we use the

nominal exchange rate to convert yen to US dollars, the estimated Japanese income

variables would be higher than what is reported in Table 3. For the ROK, the opposite

is true.

Household wealth (net worth). We construct this variable as the sum of

households’ net financial and nonfinancial assets. The former includes stocks, bonds,

and other banking accounts (and IRA accounts in the US), while the latter includes

households’ main residence, other real estate, business assets, vehicles, and other

assets. As expected, household wealth is much higher in the US than in the Asian

countries, with Japanese and Korean household wealth on average about 60% and

40%, respectively, of that of the US. Interestingly, even though the Japanese on

average hold fewer financial assets than their US counterparts, the percentage of these

assets in total wealth is higher. On the other hand, Korean households typically hold

less than 6% of total wealth in financial assets.

Homeownership percentage and value. The first variable is defined by whether

the households’ current residence value is positive or not. Americans have the highest

ownership rate, which is around 80%. It dipped slightly after the financial crisis and

the influx of a new, younger cohort to 76.4%, which is still significantly higher than

the homeownership percentage in the two other Asian countries. The ROK’s

homeownership rates across the three waves remains quite consistent at just under

60%. However, homeownership rates have changed drastically across waves in Japan.

They reached close to 80% in wave 1, fell to 63.6% in wave 2, and dropped to just

45.4% in wave 3. We note, however, that new municipalities were introduced in both



waves 2 and 3, so this drop could be the result of survey design as JSTAR is not a

random probabilistic survey at the national level.

The second variable—the value of homeownership (i.e., primary residence)—

exhibits a more consistent level across waves: the US variable features the highest

average values while the values across the two Asian countries are steady over the

three waves.

IV. Retirement, Pension Dependency, and Wealth Distribution

In this section, we offer additional detailed analysis regarding the variations of

retirement status, working status, income, and wealth among the elderly in the three

countries under study. This analysis reveals some additional insights about the sources

of the differences across the three countries beyond what we learn from the descriptive

statistics.

A. Retirement

We start with the individual retirement status. The survey question asks

respondents to declare whether they consider themselves fully retired. This means

that individuals can stay out of the labor force or run their own business or do some

voluntary work but still claim that they are retired. Table 4 reports the percentage of

household heads who consider themselves retired in the following age brackets: less

than 60; 60–64; 65–69; and 70 and above. The first bracket applies to workers who are

below early pensionable ages, except for Koreans during a transition period discussed

earlier. The second bracket includes workers in early pensionable ages, while the third

bracket applies to people who have reached full retirement age (66 in the US at the

time of the survey), and the fourth bracket can be thought of as late retirement age.

Table 4 indicates that more Americans (36.3%) consider themselves retired

compared to workers in other countries. The two Asian countries’ data reveal that a

significantly lower percentage of workers—around 15%—consider themselves retired

at all three age brackets (the Korean average rose to 20% in wave 3 due to aging of the

only cohort covered in the country’s survey). Due to the unique opportunity to retire

early in the ROK before reaching the age of 60, 8.0% to 9.3% of Koreans chose to do

so, while for Americans this figure is only 3.6% to 5.8% due to social security’s

generous disability programs. Among people who might be eligible for early

retirement, 18.2% to 24.5% of Koreans chose to retire compared to 22.2% to 26.1% of

Americans. Among respondents who are in the full retirement age group, 17.1% to



29.0% Asians consider themselves retired compared to about half of Americans in this

age group (but this figure went down to 47.5% in 2008–2009 because of the financial

crisis). For those aged 70 and older, the retirement percentages are below 30% in Japan

and 37% in the ROK, compared to about 70% in the US.

Because one does not have to claim full retirement to receive pension benefits, it

is interesting to compare the percentage of people who stay in the labor force in the

three countries. The survey asked whether respondents were working for pay, and the

results for the same age brackets in Table 4 are reported in Table 5. Here we see that

the percentage of working household heads are quite similar across the three countries

in all age brackets. However, among Japanese workers younger than 60 this

percentage is above 90%, although it declines in later waves closer to the percentages

seen among workers in the ROK and the US. Surprisingly, the Korean average over all

age brackets is lower than that of the US. In the group eligible for early retirement,

61% of Japanese workers continue to work compared to 55% of Americans in this

Table 4. Percentage Retired by Age

Age Brackets HRS JSTAR KLoSA

2006–2007All ages 36.3% 14.8% 15.4%Age < 60 5.0% 0.3% 8.0%60 � Age < 65 26.1% 5.5% 18.2%65 � Age < 70 50.5% 21.4% 26.8%Age � 70 68.8% 25.8% 28.8%



HRS ¼ Health and Retirement Study (United States),JSTAR ¼ Japanese Study of Aging and Retirement,KLoSA ¼ Korean Longitudinal Study of Aging.Source: Authors’ calculations using data from the Healthand Retirement Study (United States), the JapaneseStudy of Aging and Retirement, and the KoreanLongitudinal Study of Aging.



group. Koreans belonging to this group work the least, as only 36% continue working.

For the full retirement age group, all three countries have numbers around 30%.

However, 13% of Americans work well into late retirement age, while only 9% of

Japanese work into their 70s. In later waves (waves 2 and 3), however, the share of

Japanese older than 70 who reported working (23.8% and 29.1% for waves 2 and 3,

respectively) is higher than in the other two countries. In the ROK, however, only

14.4% to 16.8% of Koreans over 70 years old continue to work. Thus, except for wave

1, the two Asian countries’ workers retire later in life and continue working well into

full retirement age.

B. Public and Private Pensions

Next, we compare pension coverages for these three countries. We first consider

public pensions. Table 6 reports percentages of people receiving public pension by age

groups. It is quite clear that Japan has the widest public pension coverage among the

Table 5. Percentage Working by Age





HRS ¼ Health and Retirement Study (United States),JSTAR ¼ Japanese Study of Aging and Retirement,KLoSA ¼ Korean Longitudinal Study of Aging.Source: Authors’ calculations using data from theHealth and Retirement Study (United States), theJapanese Study of Aging and Retirement, andthe Korean Longitudinal Study of Aging.



three countries, while the public pension coverage for the ROK clearly lags behind the

other two countries due to the fact that its pension system was developed quite late.

For people who are eligible for full public pension, 92.3% of Americans and 95.8% of

Japanese are covered, while only 30.1% of Koreans are covered (wave 1). For people

aged 70 and older, coverage is almost universal in Japan and the US, while only 10.8%

of Koreans in wave 1 and 27% of Koreans in wave 3 received coverage. Even people

in the 60–69 age group in the ROK who have a better chance of meeting the eligibility

criteria, only around 30% of them in wave 1 and 50% in wave 3 reported receiving

coverage.

Table 7 reports coverage rates by age bracket for people receiving private

pensions. In this pension category, Americans clearly have a significant advantage

over their Asian counterparts. Among Americans in their early retirement age, 29%

Table 6. Percentage Receiving PublicPension by Age





HRS ¼ Health and Retirement Study (UnitedStates), JSTAR ¼ Japanese Study of Aging andRetirement, KLoSA ¼ Korean LongitudinalStudy of Aging.Source: Authors’ calculations using data from theHealth and Retirement Study (United States), theJapanese Study of Aging and Retirement, andthe Korean Longitudinal Study of Aging.



receive private pensions. The coverage rates reach 41% when they reach full

retirement age and it goes to almost 50% when they are in their 70s. The coverage

rates dipped after the financial crisis, as reported in the 2010–2011 wave, but they are

still much higher than the coverage rates of their Asian counterparts. For the Japanese,

the coverage rates hover around 20%, and only 12.6% of Japanese in their 70s claimed

pension from private sources. For Koreans, the coverage rates are never higher than

9% across all waves in all the age brackets reported in Table 7. One caveat is that the

old income-retirement systems in Japan (before 2001) and the ROK (before 2005)

offered lump sum severance payments. It is not clear how these payments have

impacted the coverage rates reported in these two countries.

To assess the degree of dependency on pensions of old-age groups in the three

countries in our sample, we first consider the income distribution of the whole sample

of old-age groups in each country and those who are fully retired. Table 8 reports this

Table 7. Percentage Receiving Private Pensionby Age





HRS ¼ Health and Retirement Study (United States),JSTAR ¼ Japanese Study of Aging and Retirement,KLoSA ¼ Korean Longitudinal Study of Aging.Source: Authors’ calculations using data fromthe Health and Retirement Study (United States), theJapanese Study of Aging and Retirement, and theKorean Longitudinal Study of Aging.



distribution of income at the 10th, 25th, 50th (median), 75th, and 90th percentiles of

income. The two tail percentiles (10th and 90th) are included in order to see the impact

of pensions on people at the tails of the distribution, which typically have skewed tails.

This analysis has potentially important policy implications as well, because those who

are most affected by any policy reforms are those in the extreme ends of the spectrum.

The median incomes reported in Table 8 for the three countries are somewhat

different from the average incomes we report in the summary statistics presented in

Table 3. More specifically, the median income for Japan is closer to that of the US. For

the whole sample, the median income in Japan is 80% of that of the US. For people

who are fully retired, the gap is even smaller: the median income for that group in

Japan is almost 90% of that of the US. The lower tails are similar for these two

countries although the US maintains its advantages over Japan in the upper tails,

especially at the 90th percentile, with the Japanese 90th percentile income equal to just

50% of US income in the whole sample, and two-thirds in the fully retired subsample.

This clearly is the driving force behind the average income gap between these two

countries.

For the ROK, both the median income and the average income are about one-

third of US median and average income, respectively. However, the lower tail looks

very different from that of the US. The whole lower tail of the Korean income

Table 8. Income Distribution for Whole Sample and Retired ($’000)

Whole Sample Retired

10% 25% Median 75% 90% 10% 25% Median 75% 90%

2006–2007HRS 10.6 19.5 41.3 81.1 146.1 10.9 17.1 30.9 54.3 91.8JSTAR 8.7 18.9 32.9 52.4 73.4 8.7 15.7 27.1 41.9 61.2KLoSA 0.0 0.2 11.8 33.6 65.0 0.0 0.0 1.7 13.7 38.6

2008–2009HRS 10.1 18.8 40.5 79.6 142.4 10.3 16.4 30.3 56.0 98.1JSTAR 6.2 13.2 24.3 39.7 61.8 7.1 15.9 26.5 39.7 59.5KLoSA 0.0 1.6 13.9 32.5 57.4 0.0 0.2 2.9 14.9 33.9

2010–2011HRS 10.5 19.8 42.5 85.1 156.2 10.8 17.1 31.6 56.2 94.0JSTAR 9.0 17.1 28.9 45.1 69.0 9.2 18.1 27.1 39.3 58.7KLoSA 0.0 2.0 14.1 29.5 57.1 0.0 0.4 4.0 15.7 33.3

HRS ¼ Health and Retirement Study (United States), JSTAR ¼ Japanese Study of Aging andRetirement, KLoSA ¼ Korean Longitudinal Study of Aging.Source: Authors’ calculations using data from the Health and Retirement Study (United States), theJapanese Study of Aging and Retirement, and the Korean Longitudinal Study of Aging.



distribution is close to 0, which is also the case for the fully retired subset. At the same

time, Table 8 indicates that the ROK has managed to narrow the gap in the upper tail

of the income distribution compared to the US relative to the larger gap in the lower

tail. The 90th percentile income is 44% of that of the US for the whole sample and

42% in the fully retired subsample. The fat right tail of the income distribution in the

ROK is quite striking: the median income is only one-third of the upper quartile for

the whole sample while the median income for the fully retired is only one-tenth of the

upper quartile. It seems that many Korean retirees have to resort to their savings to

support their consumption.

The median percentage of total income received from total pension income can

be used as a rough measure of dependency on pension income during retirement.

Because relatively few Asians would claim they are fully retired, we calculate the

median percentage for people over the age of 65. However, while public pension

coverage in both Japan and the US is almost universal for this age group, there is still

limited pension coverage (well below 50%) for Koreans in the same group. For this

reason, Table 9 compares the median percentage of income received from pensions

just for Japan and the US. We compute it for all the age groups above 65 in our sample

and break it down by income quartiles.

Table 9. Median Percentage of Total Income Received fromTotal Pension Income for 65 and Older by Income Quartile

Income Quartiles

Overall 1st 2nd 3rd 4th

2006–2007HRS 85.7 100 95.4 76.3 36.9JSTAR 77.3 120 97.1 76.0 36.3

2008–2009HRS 84.1 100 93.3 72.2 31.7JSTAR 80.0 103 96.4 80.0 40.0

2010–2011HRS 85.1 100 95.6 76.2 32.4JSTAR 76.7 103 94.3 70.0 41.0

HRS ¼ Health and Retirement Study (United States), JSTAR ¼Japanese Study of Aging and Retirement.Note: We chose not to report the results for the Korean LongitudinalStudy of Aging (KLoSA) because median percentages are mostly zero,the highest being 10% for the fourth quartile in 2010.Source: Authors’ calculations using data from the Health and RetirementStudy (United States) and the Japanese Study of Aging and Retirement.



Table 9 shows that for Americans over age 65, 85.7% of their retirement income

comes from public and private pensions. For the Japanese in the same age group,

however, only 77.3% of retirement income comes from public and private pensions.

This evidence implies that the US has a more rewarding retirement system than Japan.

At the same time, to the extent that most of the pension coverage comes from public

pensions, our finding implies that Americans rely on the public pension system more

than their Japanese counterparts. Indeed, Table 9 shows that people at the lowest

income quartiles, on average, rely entirely on pension income, and those in the second

lowest quartiles rely heavily on pension income as well. For those whose incomes are

in the third quartile, however, more than one-quarter of their incomes come from

sources other than pensions. For people in the highest income bracket, the median

percentage coming from non-pension sources is only 63%. In sum, we can say that

three-quarters of people in both the US and Japan rely heavily on the retirement

pension system for their old-age income. Even wealthier people (in the highest

quartile) derive a substantial portion of their incomes from the pension system.

C. Wealth, Financial Wealth, and Stock Ownership

The previous section, which focused on the extent to which old-age income in

the three countries is dependent on the existence of private and public pensions, raises

a question about the role of accumulation and management of private savings as an

additional channel which affects the financial well-being of individuals and

households. The question is important because if the majority of people derive their

income just from the pension system, especially if the latter is in the form of defined

benefits rather than defined contributions, such dependency limits the opportunities for

older and retired people to cope with unforeseen financial and health predicaments

affecting themselves or their family members. Since high degrees of dependency on

pension income are the result of insufficient private wealth accumulation, we explore

in this section the distribution of personal wealth in the three countries under study.

Table 10 reports the distribution of household net worth at the 10th, 25th, 50th

(median), 75th, and 90th percentiles in 2010 US constant dollars (see Section III.B for

a definition of household net worth) for the whole sample of older age groups as well

as the fully retired in the three countries. A quick overview indicates that Japan and the

US share similar wealth distributions, especially for the whole sample. The Korean

wealth distribution is quite different, however, and resembles to a large extent the

country’s income distribution, which immediately indicates the large influence of

private and public pensions in the country.



While the Japanese wealth distribution is thicker in the lower half of the

distribution and even at its median in wave 1, the American wealth distribution is

thicker in the upper half, which clearly dominates the distribution. Looking only at

those who are fully retired, the American distribution dominates the Japanese over the

whole distribution and across all waves. The same applies to the comparison between

the American and the Korean distributions. Just like the income distribution, the

wealth distribution of the ROK is heavily skewed and wealth accumulation seriously

lagging that of other developed countries. This is largely the result of higher human

and physical capital accumulation in the US relative to the two Asian countries in

recent centuries. The obvious implication is that Americans enjoy a higher level of

retirement income relative to their Asian counterparts. An interesting question to ask is

to what extent is this the result of private accumulation of assets as opposed to the

private and public pension systems in the three countries.

We are tempted to think that the larger accumulation of wealth in the US might

be due to their relatively large holdings of stocks and financial assets. However, as

reported in Table 11, this conjecture is not entirely correct. Looking at the whole

sample, the Japanese almost dominate the Americans over the entire financial wealth

distribution. Moreover, it is surprising to see that the median value of financial assets

in Japan is almost five times higher than that of the US. Even for people who are fully

Table 10. Wealth Distribution for Whole Sample and Retired ($’000)


10% 25% Median 75% 90% 10% 25% Median 75% 90%

2006–2007HRS 0.7 44.0 202.3 557.2 1,219.0 4.1 71.8 269.4 643.7 1,308.0JSTAR 0.0 61.2 227.1 436.8 742.5 0.0 52.4 209.7 436.8 698.9KLoSA 0.0 5.6 70.0 191.5 401.6 0.0 2.8 65.8 202.2 432.8

2008–2009HRS 0.6 40.7 189.2 532.9 1,174.0 3.0 64.8 249.5 625.7 1,290.0JSTAR 0.0 26.5 164.2 361.4 713.6 8.7 84.4 216.3 379.5 697.3KLoSA 0.0 12.9 91.9 253.0 515.7 1.0 2.58 77.4 257.9 555.2

2010–2011HRS 0.0 25.7 154.0 464.0 1,063.0 4.1 69.5 241.8 593.8 1,204.0JSTAR 0.0 9.9 135.3 328.7 615.5 0.0 45.1 180.5 357.3 604.5KLoSA 0.0 11.9 95.5 249.7 500.0 0.0 3.57 71.4 237.8 444.9




retired, the median value of financial assets in the US is almost half of that in Japan.

Americans do hold more financial assets, however, in the upper tail—the 90th

percentile of the whole sample—and even in the two upper quartiles of the financial

wealth distribution of the fully retired in wave 1. Koreans hold far fewer financial

assets compared with these two countries. Even the value of financial assets at the 90th

percentile in the ROK is just 60% of the median value in Japan. This should not be a

total surprise because economic development in the ROK started much later than in

Japan. But it is still odd to see that more than half of Korean households do not hold

any financial wealth.

We turn next to stock ownership in all three countries as reported in Table 12.

Following the conventional definition, this classification includes both direct stock

ownership and or indirect ownership through other financial instruments such as

mutual funds. This category excludes pension plans, even those based on DC plans,

according to the way the survey questionnaire is structured. The overall stock

ownership in the US is around 30%, which means 30% of American households hold

some form of stocks, while in Japan stock ownership experienced a sharp increase

from 13% in 2006–2007 (wave 1) to over 20% in 2010–2011 (wave 3). Stock

ownership in the ROK, however, is almost negligible—in all survey years, less than

5% of Korean households reported owning stocks. Table 12 shows, however, that

Table 11. Financial Wealth Distribution for Whole Sample and Retired ($’000)


10% 25% Median 75% 90% 10% 25% Median 75% 90%

2006–2007HRS �3.7 0.5 10.8 92.0 326.7 0.0 1.3 27.0 153.6 447.9JSTAR 0.0 13.8 52.4 131.0 262.1 0.0 8.7 52.4 124.5 262.1KLoSA �25.2 0.0 0.0 7.0 30.8 �9.6 0.0 0.0 5.6 33.5

2008–2009HRS �4.6 0.0 10.1 100.3 330.3 0.0 1.0 27.4 152.0 461.0JSTAR 0.0 2.65 30.9 123.6 220.7 0.0 8.8 44.6 176.5 285.1KLoSA �6.5 0.0 0.0 7.7 38.7 0.0 0.0 0.0 6.45 38.7

2010–2011HRS �8.0 0.0 7.0 75.5 300.0 �0.2 1.0 25.0 142.0 432.1JSTAR 0.0 9.0 54.1 143.8 261.7 0.0 18.0 74.0 162.0 316.0KLoSA 0.0 0.0 0.0 9.5 39.2 0.0 0.0 0.0 6.0 35.7




stock ownership rises when we restrict the sample to include only households that

report positive overall wealth and positive financial wealth. In the latter group, stock

ownership reaches almost 40% of all US households. It also increases to around 15%

in waves 1 and 2 for Japanese households with positive financial wealth, and it shoots

up to 24% in wave 3.

The data on stock holding can provide an explanation as to why Americans hold

more financial assets in the upper tails of the distribution of financial assets. The

driving force behind the thick upper tails of the distribution is stock ownership, due to

the higher expected values of stocks relative to fixed-income assets. The median value

of personal stocks in the US is $70,300 in 2006–2007 (wave 1), 2.7 folds higher than

the median of $26,200 in Japan. This gap between the US and Japan stays the same

even in 2010–2011 (wave 3) when stock ownership rate in Japan increases

dramatically to 24%. The median value of personal stocks in the ROK is only

$14,000 in 2006–2007 (wave 1). Comparing the fully retired households with

households in the whole sample, we see an increase in stock ownership by retirees in

the US and Japan (except in wave 1), while stock ownership becomes smaller for

Table 12. Stock Ownership

HRS JSTAR KLoSA

WholeSample Retired

WholeSample Retired

WholeSample Retired

2006–2007Overall 27.9% 32.6% 13.1% 11.6% 2.4% 2.1%Given positive wealth 30.2% 34.3% 14.4% 13.4% 2.8% 2.7%Given positive financial 36.1% 38.7% 15.1% 12.7% 4.5% 4.2%Median value ($’000) 70.3 108.2 26.2 34.9 14.0 19.5



HRS ¼ Health and Retirement Study (United States), JSTAR ¼ Japanese Study of Aging andRetirement, KLoSA ¼ Korean Longitudinal Study of Aging.Source: Authors’ calculations using data from the Health and Retirement Study (United States),the Japanese Study of Aging and Retirement, and the Korean Longitudinal Study of Aging.



retirees in the ROK. However, in all three countries, the median retiree holds a higher

value of stocks than the median household in the whole sample.

The policy ramifications from our analysis of stock ownership are quite clear. If

more American households were willing to hold stocks, then they would be likely to

hold more financial wealth and total wealth over the entire wealth distribution relative

to Japanese households. The lesson from this analysis especially for Korean

households and policy makers appears to be that households may gain significantly

from policies promoting stock ownership in the ROK. The lesson may have already

started to sink in: in recent years, some financial reports have indicated that the overall

stock ownership by Koreans has started to crack the 10% barrier. But this pace of

increase may not be fast enough to improve the country’s wealth distribution,

especially among current or soon-to-be retirees, to a level comparable to that of Japan

and the US.

V. The Asset Management Model and Participation in Risky Asset Holding

Given the apparent importance of the contribution of stock ownership to the

accumulation of financial wealth and overall wealth, and thus the promotion of

financial well-being for retirees, this section focuses on the determinants of

participation in stock ownership. This is a well-researched topic in the literature.

Traditionally, however, the approach used in the household finance literature to explain

this issue (see, for example, Campbell 2006) has tended to focus on preferences and

bequest motives as well as underlying income and savings constraints, or the role of

financial literacy. Ehrlich, Hamlen, and Yin (EHY) (2008) and Ehrlich, Shin, and Yin

(ESY) (2011) develop a labor-theoretic model of asset management based on noisy

rational expectations (Grossman and Stiglitz 1976, Verrecchia 1982) and the

determinants of productive savings through asset management (Ehrlich and

Ben-Zion 1976). Our point of reference is what we call “the asset management

hypothesis” (AMH). We present the outline of this approach and implement it

empirically by following the econometric model that Ehrlich and Yin (2021)

developed to analyze the HRS data from the US. We then modify this model to apply

to all three datasets used in our current study. The value added of this application of the

AMH lies not only in a comparative analysis across countries but also in a

decomposition analysis we pursue empirically to determine the separate contribution

of individual asset management relative to the distinct contribution of financial

markets within which individuals manage their assets. This decomposition analysis is

similar to the method used in labor economics to address racial or gender gap.



A. The Asset Management Hypothesis

The standard assumption in the finance literature that equilibrium prices of

financial assets are “fully revealing” of all relevant information on the performance of

these assets raises a puzzle at the micro level: it leaves no incentive for individuals to

collect any private information (see, for example, Grossman and Stiglitz 1976). But

how, then, do prices become fully revealing? To resolve this puzzle, it seems necessary

to augment the equilibrium theory of efficient markets with complementary micro

foundations. In the context of centrally traded homogeneous assets, the idea is that the

public information revealed by market prices is incomplete or imperfect even in

equilibrium, leaving room for at least some asset management activity by

heterogeneous investors endowed with different abilities, opportunities, and prior

knowledge.

The asset management hypothesis, as formulated by EHY (2008), for example, is

a natural extension of the “costly information” or “noisy prices” literature. In

extending the basic framework of the “noisy” rational expectations model (see,

for example, Verrecchia 1982), EHY specify a precision production function of a

Cobb–Douglas type, which is a function of one’s human capital and time spent on

information acquisition. By developing market equilibrium properties associated with

imperfectly informed markets, the EHY and extended models yield solutions for the

optimal demand for investment in information precision, based on both market and

private signals, and the associated expected individual demand for alternative risky

assets, as well as the resulting overall portfolio returns.

In particular, the theoretical models lead to an explicit reduced form specification

of the determinants of the expected demand for a risky asset (RA), which stems from

greater demand for asset management activity, that is, gaining private information about

the productivity of traded assets. Such activity can be shown to enhance the expected

demand for risky asset holding due to a more precise assessment of the riskiness of these

assets. Modified to account for asset management by those approaching retirement age

or retirees, the reduced-form regression can be outlined as follows:

E(RA) ¼ �(HC,H,w,FA): ð1ÞIn equation (1), E(RA) is the expected demand for risky assets. HC is human capital

(or knowledge), either general or specific, which in turn can be modeled as a function

of schooling (S) and experience (Exp), or HC ¼ f (S,Exp); H stands for individual

health, both physical and mental, which can be taken as a constraint on productive

time that can be devoted to asset management; w denotes the opportunity cost of asset

management time, usually proxied by the wage rate; and FA denotes investor’s



accumulated portfolio size—a proxy for experience in managing assets and lower

fixed costs of entry to risky markets due to scale economies.

We do not explore the role of the wage rate in our current regression analysis

because this information is not reported in the sample of older workers who are still

active and is naturally absent in the sample of fully retired workers who earn no wages

during their retirement phase. Partly because of this fact, we expect that general

education, which raises efficiency in managing assets as well as one’s wage earnings

power, will tend to have an unambiguous positive effect on asset management and the

expected demand for risky assets, E(RA), especially in the case of retired persons.

More generally, the regression results reported in EHY (2008) and ESY (2011)

indicate that the impact of education on asset management efficiency exceeds that of

the opportunity cost of asset management time, which should be the case especially for

retirees who earn no wages.

From this expected demand for RA, we also derive an expected portfolio return

equation:

E(Portfolio Return) ¼ Ψ (S,Exp,H,w,FA): ð2ÞThe asset management hypothesis thus implies that investors in general, including

older workers and retirees who have higher educational attainments and are in better

health, will be more efficient in managing risky portfolios and thus more willing to

participate in (enter into) the market for risky assets. By participation we mean

willingness to hold at least some risky assets in their financial portfolios, which can be

best proxied by stocks (RA > 0). The hypothesis also implies that people with higher

education or better health would tend to have a larger expected demand (i.e., average

demand over the long term) for risky assets, and thus allocate a larger percentage of

their portfolios to risky assets. Such tendency to manage, rebalance, and hold more

risky securities ultimately yields a higher corresponding market return for individuals

with more education and better health, independent of their attitudes toward

financial risks.

Note, however, that the hypothesis does not imply that asset management directly

affects the returns on the observed risky market portfolio, comprised of just risky

securities. The return on any risky asset held is determined strictly by the market in

which it is traded. However, higher general and specific knowledge and better health

enable investors to assess risk more accurately and select and rebalance efficient

portfolio compositions, for example, when portfolios include securities that are traded

in different international markets. Therefore, these investors have, on average, a

greater demand for or a larger portion of risky assets in their financial portfolios (ESY

2011). A more informed selection of risky assets can thus yield higher returns on



individuals’ overall financial portfolios by increasing the share of the risky assets in a

portfolio comprised of both stocks (including risky corporate bonds) and fixed-income

assets.

B. Empirical Implementation

We implement the basic ideas of the asset management hypothesis by using the

three longitudinal datasets described in the preceding sections, with particular

emphasis on the role of educational attainments and health status in enhancing the

expected demand for stocks due to better information and physical and mental health,

which enhances effective asset management. We concentrate on the participation

decision due to data limitation. In addition, some of the demographic variables used in

Ehrlich and Yin (2021) are available only in HRS, such as race and cohort. The

regression specification we use to implement the basic decision variables in our model

is thus given in equation (3) as follows:

Pr(Rah ¼ 1) ¼ a0 þ a1Ageþ a2Age2 þ a3Educationþ a4Health

þ a5log(income)þ a6log(wealth)þ a7Gender

þ a8Coupled þ a9Retired þ a10PrivatePension: (3)

The dummy variable, Rah, is an indicator of whether the household has any stocks, as

reported in the three surveys. We choose educational attainment as an education

proxy because continuous years of education are not available in JSTAR and KLoSA.

Health is a health-status index, based on the individual’s self-reporting, indicating

whether the self-reported health is good or better. Income is household income while

wealth is household net worth. Gender is a dummy variable which is 1 for male and 0

for female, while coupled and retired are dummy variables to distinguish couples that

are living in a household and heads of household who consider themselves retired,

respectively. Finally, PrivatePension is a dummy variable indicating whether

households receive any private pension. A complete definition of these variables is

given in Section III.B.

We run a probit regression analysis of equation (3) using the HRS, JSTAR, and

KLoSA data on households with positive survey weights, positive wealth (due to the

need to perform a logarithmic transformation on this variable), and household heads

older than 50. The results are reported in Tables 13–15 for all three waves.

The regression results are quite robust across the three waves. In what follows,

therefore, we focus largely on the results obtained across the three samples based on

the 2006–2007 wave.



C. Regression Results

A quick inspection of the results across the three samples in Tables 13–15

indicates that the probit regression specification yields the most robust results using the

HRS sample, where all coefficients are statistically significant at 5%, the only

exception being the coefficient of age, which is significant at the 10% level. In the

probit analysis using JSTAR and KLoSA, some of the demographic variables become

insignificant, in particular gender, coupled, and retired.

Table 13. Participation Decision for Stocks,2006–2007 Wave

Characteristics HRS JSTAR KLoSA

Intercept �6.438 �2.281 0.265(�9.42) (�2.70) (2.47)

Age �0.0348 0.0579 �0.00112(�1.83) (2.16) (�3.45)

Age2 0.00032 �0.00045 0.0000787(2.40) (�2.17) (3.19)

Education 0.205 0.067 0.0216(12.4) (7.79) (7.95)

Health 0.128 0.0097 �0.00293(3.38) (0.49) (�0.63)

log(income) 0.105 0.0154 0.0005(5.94) (3.28) (0.82)

log(net worth) 0.401 0.0238 0.0102(29.9) (3.66) (6.63)

Gender �0.0639 �0.0138 �0.0064(�1.98) (�0.73) (�1.34)

Coupled �0.0974 �0.00102 �0.0017(�2.77) (�0.04) (�0.34)

Retired 0.103 0.0138 0.0036(2.70) (0.57) (0.63)

PrivatePension 0.0823 0.0772 �0.0146(2.38) (3.91) (�1.45)

HRS ¼ United States Health and Retirement Study, JSTAR ¼Japanese Study of Aging and Retirement, KLoSA ¼ KoreanLongitudinal Study of Aging.Notes: The table shows the results of a probit analysis on stockownership for households with positive net worth and householdheads older than 50. t-statistics are in parentheses.Source: Authors’ calculations using data from the Health andRetirement Study (United States), the Japanese Study of Agingand Retirement, and the Korean Longitudinal Study of Aging.



The most important result of our probit analysis is that it verifies the key

proposition of the asset management hypothesis: the educational attainment of the

household head has a statistically positive and significant impact on the probability

that households will be holding risky assets, which in this application are represented

by stocks. We observe this significant effect of education across all countries as well as

across all waves within the three countries, with no exception.

Due to the nonlinearity in the probit model, it is difficult to assess the quantitative

impacts of educational attainments on participation solely based on regression results.



Intercept �6.123 �1.329 0.529(�8.25) (�2.17) (4.25)

Age �0.0444 0.0235 �0.0194(�2.16) (1.21) (�5.13)

Age2 0.000374 �0.000182 0.000135(2.62) (�1.21) (4.74)

Education 0.192 0.0727 0.0181(11.20) (9.48) (5.59)

Health 0.0733 0.00737 0.0049(1.79) (0.44) (0.87)

log(income) 0.103 0.0372 0.00197(5.97) (1.05) (2.27)

log(net worth) 0.414 0.0426 0.0118(29.40) (9.51) (6.70)

Gender �0.0946 0.00244 �0.00827(�2.83) (0.15) (�1.49)

Coupled �0.0775 0.00654 �0.0188(�2.13) (0.42) (�3.31)

Retired 0.0959 �0.00638 0.0118(2.47) (�0.29) (6.70)

PrivatePension 0.0576 0.0972 0.00836(2.62) (5.20) (0.48)




To generate a quantitative measure of the difference in participation, we compare the

participation probabilities across households attaining high school education versus

some college education. We make sure that the two groups are identical, except for

their different educational attainments, by assuming that all their other characteristics

follow the sample-weighted averages. For the 2006–2007 wave, the differences in the

probability of holding risky assets between these two groups are 7.96%, 7%, and

5.42% in the HRS, JSTAR, and KLoSA, respectively. Turning into percentage



Intercept �4.899 �0.571 �0.0637(�8.05) (�0.58) (�0.65)

Age �0.0716 �0.0058 �0.00056(�4.19) (�0.19) (�0.19)

Age2 0.000552 0.00004 0.00000057(2.62) (0.17) (0.03)

Education 0.163 0.0509 0.0144(9.92) (3.82) (5.39)

Health 0.161 0.00054 0.000825(4.10) (0.02) (0.18)

log(income) 0.0932 0.00103 0.000725(6.00) (1.20) (1.02)

log(net worth) 0.402 0.0627 0.0085(31.00) (7.76) (5.82)

Gender 0.00462 �0.0457 �0.0191(0.15) (�1.69) (�3.94)

Coupled �0.0588 0.0101 �0.00976(�1.73) (2.89) (�2.07)

Retired 0.0327 0.030 �0.00666(0.086) (0.81) (�1.23)

PrivatePension 0.141 0.114 0.0126(3.93) (1.63) (1.32)




differences, a higher educational attainment would boost stock ownership by 21.53%,

27.3%, and 34.56%, respectively.7

Our secondary conjecture, that better health would more likely lead people to

own stock, is verified only by our regression analysis using HRS data. This variable

becomes insignificant in both JSTAR and KLoSA. A possible reason is that the health

variable is a subjective measure of how well people perceive their health to be, and

these perceptions may differ across countries. In Japan, as we see in the descriptive

statistics, most people rate their health status to be quite high. The opposite seems to

be the case in the ROK (see descriptive statistics in Section III.B). The sharply

divergent concentrations of the individual data may bias the regression coefficient

associated with the health dummy downwards. The lower participation rate in stock

holding might bias the impact of health as well since, as a proxy for productive asset

management, it becomes less relevant.

The total effect of the age variable is found to be U-shaped in the HRS

regressions. But the age effect is not symmetrical across waves in the Asian countries:

it is hump-shaped in Japan (in wave 1), U-shaped in the ROK (in waves 1 and 2), and flat

in other waves (in Japan in waves 1 and 2 and in the ROK in wave 3). The U-shaped age

effect in the HRS regressions means that older people tend to hold stocks. We compute

the age at which older US investors start raising their stock holdings to be around age 60,

59.4, and 64.9 in the 2006–2007, 2008–2009, and 2010–2011waves, respectively, which

is reasonable since at this age wealth accumulation becomes more substantial for most

households, although this may be the case mainly in the US.

Income is included merely as a control variable. Without its presence, the

education effect might mistakenly pick up its explanatory power. Its positive

coefficient is statistically significant in regressions using HRS and JSTAR data. In

contrast, wealth is included as an efficiency measure, since it may represent experience

at managing assets. Its positive impact is statistically significant in all regressions.

In regressions based on the HRS sample, both retirement status and having

private pension as income have a statistically positive impact as well. This might be an

indication of the presence of spillover effects from having pension funds, especially

those based on defined contributions or IRAs, since individual account holders have a

direct responsibility for making decisions about portfolio composition in such funds.

However, the market for such funds appears to be weaker in Japan and especially in

the ROK.

7We did this comparison for other waves as well, using the corresponding data sets, and foundcomparable results.



D. Decomposition

How much of the difference in observed stock ownership is due to the difference

in the structure and efficiency of the financial markets that is separate from the effect of

individual characteristics? Christelis, Georgarakos, and Haliassos (2013) used a

method originally developed in labor economics to assess the distinct effect of the

labor market environment in explaining the impact of gender on the entire distribution

of wage earnings. The authors used this method to assess the effect of what they call

the “environment” on wealth distribution as well as on all asset ownership.

Basically, the difference in participation rates (pr) between the base country and

country i can be decomposed as follows:

prbase � pr i ¼ fprbase � bpr i, baseg þ f bpr i, base � pr ig, ð4Þwhere bpr i, base is the predicted stock ownership for country i if people in that country

move to live in the base country. In other words, it is the counterfactual stock

ownership of households in country i. The first difference term on the right side of

equation (4) can thus be thought of as the covariates effects, because it illustrates the

difference in the observed stock ownership when people in both countries are facing

the same coefficients, or the same environments, in the base country. The second

difference term on the right side of equation (4) signals the coefficients effects, because

it is the difference between the counterfactual ownership rates for country i and the

actual ownership rates in that country. This difference can be interpreted as the effect

of the efficiency of the financial (including pension) market on the difference in the

observed household participation rates in holding stocks.

We apply this decomposition method to all waves of our three micro datasets by

using the US as the base. The results are reported in Table 16.

Our first observation is that all decomposition effects are statistically significant

at the 5% level. It is interesting to see most of the differences due to both the covariates

and institutional effects have positive signs, the only exception being the difference

due to the coefficients effect in JSTAR in 2011. In other words, the observed difference

in the participation rates, which indicates that the US has a higher participation rate

than the Asian countries, is due to two factors: (i) the higher level of key explanatory

variables in the US relative to the two Asian countries, such as education and wealth;

and (ii) a financial market environment in the US that is more accommodating to

participation than in the other Asian countries.

Another major observation is that the magnitudes of the differences are generally

larger in the ROK than in Japan. Furthermore, the relative weight of the covariates

effects versus the financial markets efficiency effects varies between Japan and the



ROK. In Japan, the weight of the independent effects is more equal than in the ROK.

This implies that the financial market effect has more weight in the ROK relative to

Japan in explaining the participation gap in the ROK relative to the US.

The differences due to covariates are very robust for both countries across waves.

They vary from 7.72% for Japan in wave 1 to almost 10% in wave 2 and 9.16% in

wave 3. For the ROK, this difference steadily decreases from 20.78% in wave 1, down

to 17.15% in wave 2, and further down to 15.85% in wave 3. The explanation for the

decline in the ROK is quite straightforward. The most powerful driving force in stock

ownership is the household’s net worth, and we see a gradual, albeit slow rise in net

worth in ROK over this time period relative to the US.

As for the difference in the effects of the financial market environment on

participation, we observe a widening gap in the ROK: the difference due to coefficients

grows from 7.43% in wave 1 to 8.90% in wave 2 and further expands to 11% in wave

3. This might be a result of a combination of improvement in institutional arrangement

in the US (more DC plans encourage more stock ownership) and a lagging

development in institutional arrangement in the ROK (slow development of the stock

market and modernization of private pension plans). This pattern needs to be further

analyzed in future research.

In Japan, we see a quick reversal of the difference in the impact of the financial

market efficiency on participation—while this difference accounts for almost half of

the total difference in wave 1 at 8.24%, it narrows to 5.24% in wave 2, and then turns

negative at �4.65% in wave 3. The negative difference in 2010–2011 means that

Table 16. Decomposition of Differences in Participation Decision for Stocks UsingUnited States as the Base

Japan Republic of Korea

WaveTotal

Difference

DifferenceDue to

Covariates

DifferenceDue to

CoefficientsTotal

Difference

DifferenceDue to

Covariates

DifferenceDue to

Coefficients

(1) 2006–2007 15.96% 7.72% 8.24% 28.21% 20.78% 7.43%(11.2) (11.9) (42.2) (16.9)

(2) 2008–2009 15.06% 9.82% 5.24% 26.05% 17.15% 8.90%(17.5) (9.4) (32.4) (16.8)

(3) 2010–2011 4.51% 9.16% �4.65% 26.85% 15.85% 11.00%(15.8) (�8.0) (29.9) (20.8)

Notes: See Section V.D for a description of the decomposition analysis. t-statistics are in parentheses.Source: Authors’ calculations using data from the Health and Retirement Study (United States), theJapanese Study of Aging and Retirement, and the Korean Longitudinal Study of Aging.



Japan now holds an advantage in pension market efficiency relative to the US, which

encourages stock ownership. Whether this interpretation is valid, however, requires

further research. This is because there are several factors that may lead to this

interpretation. One factor is that JSTAR is not a national probabilistic survey, and it

added new municipalities in both waves 2 and 3. So the reversal in apparent market

productivity may merely reflect the difference between the newly added municipalities

and the old ones. Another possible explanation is that by 2011 private pension plan

reforms that were introduced by 2001 in Japan may have begun bearing fruit. We leave

this issue to further research.

VI. Conclusion

In this paper, we attempt to estimate the degree of financial readiness of older

cohorts in two relatively developed Asian countries by identifying the main channels

affecting their financial readiness or government dependency and assessing their

relative importance through a comparative analysis within and across the two

countries and in comparison to a reference country. We selected Japan and the ROK

as representatives of the Asian countries and added the US to serve as a reference

point for both for two reasons: all three countries have conducted similar micro-level

longitudinal surveys of the financial wellness and health of elderly individuals and

households using the same survey methodologies. These surveys report information

on the three main channels of old-age support for older households and retirees: the

income benefits they derive from the public and private pension systems in their

corresponding countries and the accumulated values of their households’ total and

financial assets. The latter channel allows us to study the extent to which households’

wealth and asset management contribute to their own financial readiness at old age.

All countries have experienced a demographic transition that has tended to accentuate

the degree of dependency of older and retired cohorts on old-age financial support

systems in their respective countries. The rapid aging of the populations, especially in

Japan and the ROK, makes the analysis of the determinants of the households’

financial readiness during their retirement a topic of academic importance and policy

relevance.

To compare the strength of old-age support systems in the three countries, we

first looked at the macro environments in these countries to determine the extent to

which demographic transitions have affected these countries’ old-age dependency

ratios. We find that the two Asian countries, especially the ROK, are facing an



increasing overall level of dependency on both their public and private pension

systems due to rising life expectancies along with sharply declining total fertility

rates relative to the US. These demographic trends raise serious potential challenges

to the financial viability of old-age retirement systems in the Asian countries relative

to the US.

We then briefly assessed the retirement-income systems in these countries,

including both the public and private pension systems. Japan and the US have a clear

advantage over the ROK in terms of the financial strengths of their pension systems,

not only because they are more developed but also because they were established

much earlier than the Korean systems. Although the public pension systems in Japan

and the US are comparable in terms of their current financial viability, we find that US

households obtain higher benefits from their private pension system compared to their

Japanese and Korean counterparts, and their private pension system is better funded as

well. In contrast, the Korean private pension system is much weaker.

We examined the three countries further by studying the extensive personal and

financial household data provided in their micro-level longitudinal surveys—the HRS,

JSTAR, and KLoSA surveys—to further analyze the differences among the three

countries in terms of their old-age financial readiness. We find that Americans

typically retire earlier than their Asian counterparts—both Japanese and Koreans

workers tend to work into their 70s, whereas more Americans retire at early and full

pension-eligible ages. We also find that the replacement rate of income in retirement is

comparable in Japan and the US for retirees earning average incomes, while Koreans

have much lower replacement rates.

Using our micro-level data, we were also able to construct aggregate

distributions of income, net worth, and financial assets for the three countries.

While the US dominates both countries in all distributions, it only dominates Japan at

the upper part of their respective distributions of net worth and financial assets. The

corresponding net worth and financial wealth in the ROK lag significantly behind

those of the other two countries.

Turning finally to the extent to which individuals and households contribute to

their own financial readiness through their accumulated savings and asset allocation

decisions, we examine the role of a key determinant of wealth accumulation and old-

age financial readiness in developed countries: stock ownership. To determine the

extent to which this private source of old-age financial well-being is functioning in the

three countries, we implement the asset management model of the demand for risky

assets developed in EHY (2008). Specifically, we conduct a probit analysis of stock



ownership using the data on individual-level characteristics of household heads that

are expected to contribute to the willingness to hold stocks.

Our empirical analysis strongly confirms the key proposition of our model. The

regression results we obtain, using samples of older-age cohorts and retired cohorts in

each of the three countries and all wave samples within each country, show that

educational attainments of household heads and household wealth accumulation exert

a significant positive effect on the willingness of the households to hold stocks in their

financial portfolios. In addition, we utilize an econometric model to decompose the

total gap in stock ownership across countries into the portion that is attributable to

differences in the individual characteristics of investors and the portion that is due to

the differences in the level of development of the financial markets and pension

systems that exist in the respective countries. We find that much of the total difference

in stock holding between Korean, and to some extent Japanese, households and their

US counterparts is attributable to differences in the household’s individual

characteristics. However, our decomposition analysis also shows that a large part of

the gap in stock ownership is attributable to the institutional characteristics of the

different financial systems, largely because of the level of development of the private

pension sector in the Asian countries, especially in the ROK relative to the US. A more

encouraging finding from this analysis is that the difference in the institutional

characteristics of the pension systems between the US and Japan, and to some extent

even the ROK, have started shrinking in later waves of the households’ individual

longitudinal surveys we explore. Whether this trend will continue going forward may

be a subject of future research.

Our model suggests that the significant contributions of household asset

management and the level of development of the pension systems in the three

countries to the willingness of investors to hold stocks and other risky assets are

expected to enhance portfolio returns as well (for concrete evidence see Ehrlich and

Yin (2021) and Ehrlich and Shin (2021)). Our findings thus offer important policy

implications on the way the financial readiness of older and retired cohorts can be

improved in all surveyed countries, especially in the ROK. This is achievable through

a faster development of the markets for private pensions as well as via a concerted

effort to use special education programs to raise the financial knowledge of older age

groups. In future research, we plan to further explore more specific policy reforms

based on our findings, which might help improve the financial readiness of older age

groups and lower their dependency on the public support systems in their respective

countries.



References

Campbell, John. 2006. “Household Finance.” Journal of Finance 61 (4): 1553–1604.Chien, Sandy, Nancy Campbell, Chris Chan, Orla Hayden, Michael Hurd, Regan Main, Joshua

Mallett, Craig Martin, Colleen McCullough, Erik Meijer, Michael Moldoff, PhilipPantoja, Susann Rohwedder, and Patricia St. Clair. 2015. “RAND HRS DataDocumentation, Version O. Labor & Population Program, RAND Center for the Studyof Aging.” http://hrsonline.isr.umich.edu/modules/meta/rand/randhrso/randhrs_O.pdf.

Christelis, Dimitris, Dimitris Georgarakos, and Michael Haliassos. 2013. “Differences inPortfolios across Countries: Economic Environment versus Household Characteristics.”The Review of Economics and Statistics 95 (1): 220–36.

Ehrlich, Isaac, and U. Ben-Zion. 1976. “Asset Management, Allocation of Time and Returns toSavings.” Economic Inquiry 14 (4): 558–86.

Ehrlich, Isaac, and Jinyoung Kim. 2005. “Endogenous Fertility, Longevity, and EconomicDynamics: Using a Malthusian Framework to Account for the Historical Evidence onPopulation and Economic Growth.” Journal of Asian Economics 15 (6): 789–806.

Ehrlich, Isaac, and Jong Kook Shin. 2010. “Human Capital and Imperfectly InformedInternational Financial Markets.” American Economic Review 100 (2): 244–49.

_____. 2021. “Can Human Capital and Asset Management Improve the Financial Performanceof Older Age Groups? Evidence from Europe.” Unpublished.

Ehrlich, Isaac, and Yong Yin. 2021. “The Role of Asset Management, Education, and Health inFinancial Decisions of Older Adults.” Unpublished.

Ehrlich, Isaac, William A. Hamlen, and Yong Yin. 2008. “Asset Management, Human Capital,and the Market for Risky Assets.” Journal of Human Capital 2 (3): 217–62.

Ehrlich, Isaac, Jong Kook Shin, and Yong Yin. 2011. “Private Information, Human Capital, andOptimal ‘Home Bias’ in Financial Markets.” Journal of Human Capital 5 (3): 255–301.

Grossman, Sanford, and Joseph Stiglitz. 1976. “Information and Competitive Price Systems.”The American Economic Review 66 (2): 246–53.

Ichimura, Hidehiko, Satoshi Shimizutani, and Hideki Hashimoto. 2009. JSTAR First Results2009 Report. Research Institute of Economy, Trade and Industry (RIETI).

Juster, F. Thomas, and Richard Suzman. 1995. “An Overview of the Health and RetirementStudy.” Journal of Human Resources 30: S7–S56.

Knoema. 2020. “Japan — Ratio of Population Aged 20–64 per Population 65+ Years.” https://knoema.com/atlas/Japan/topics/Demographics/Dependency-Ratios/Potential-support-ratio-20-64-per-65.

Mather, Mark, and Lillian Kilduff. 2020. “The U.S. Population Is Growing Older, and theGender Gap in Life Expectancy Is Narrowing.” Population Reference Bureau. https://www.prb.org/resources/u-s-population-is-growing-older/.

Organisation for Economic Co-operation and Development (OECD). 2009. OECD PrivatePensions Outlook 2008. Paris: OECD Publishing.

_____. 2019. Pensions at a Glance: OECD and G20 Indicators. Paris: OECD Publishing._____. 2020a. Pension Funds in Figures. Paris: OECD Publishing._____. 2020b. “Purchasing Power Parities (PPP).” https://doi.org/10.1787/1290ee5a-en

(accessed 29 October 2020).


March 23, 2022 5:33:57pm WSPC/331-adr 2250004 ISSN: 0116-11052ndReading

Rajnes, David. 2007. “The Evolution of Japanese Employer-Sponsored Retirement Plans.”Social Security Bulletin 67 (3): 89–104.

Social Security Administration (SSA). 2008, 2018. Social Security Programs Throughout theWorld: Asia and the Pacific. Woodlawn, Maryland.

United Nations. 2019. “World Population Prospects 2019.” Population Division, Department ofEconomic and Social Affairs. https://population.un.org/wpp/ (accessed 3 September2020).

Venti, Steven F., and David A. Wise. 2001. “Choice, Chance, and Wealth Dispersion atRetirement.” In Aging Issues in the United States and Japan, edited by Seiritsu Ogura,Toshiaki Tachibanaki, and David Wise, 25–64. Chicago: University of Chicago Press.

Verrecchia, Robert. 1982. “Information Acquisition in a Noisy Rational Expectation Economy.”Econometrica 50 (6): 1415–30.



Educational Gradients in Disabilityamong Asia’s Future Elderly: Projectionsfor the Republic of Korea and Singapore

CYNTHIA CHEN, JUE TAO LIM, NGEE CHOON CHIA,DAEJUNG KIM, HAEMI PARK, LIJIA WANG, BRYAN TYSINGER,

MICHELLE ZHAO, ALEX R. COOK, MING ZHE CHONG,JIAN-MIN YUAN, STEFAN MA, KELVIN BRYAN TAN, TZE PIN NG,

KOH WOON-PUAY, JOANNE YOONG, JAY BHATTACHARYA,AND KAREN EGGLESTON¤

Asia is home to the most rapidly aging populations in the world. This studyfocuses on two countries in Asia that are advanced in terms of theirdemographic transition: the Republic of Korea and Singapore. We developed ademographic and economic state-transition microsimulation model based onthe Korean Longitudinal Study of Aging and the Singapore Chinese HealthStudy. The model was employed to compare projections of functional statusand disability among future cohorts of older adults, including disparities in

⁄Cynthia Chen: Saw Swee Hock School of Public Health, National University of Singapore and NationalUniversity Health System, Singapore; and Schaeffer Center for Health Policy and Economics, Universityof Southern California, United States (US). E-mail: [email protected]; Jue Tao Lim: Saw Swee HockSchool of Public Health, National University of Singapore and National University Health System,Singapore. E-mail: [email protected]; Ngee Choon Chia: Department of Economics, National Universityof Singapore, Singapore. E-mail: [email protected]; Daejung Kim (corresponding author): Departmentof Health Care Policy Research, Korea Institute for Health and Social Affairs, Republic of Korea. E-mail:[email protected]; Haemi Park: Department of Social Welfare, Daejeon University, Republic ofKorea; Lijia Wang: Department of Statistics and Actuarial Science, University of Waterloo, Canada; BryanTysinger: Schaeffer Center for Health Policy and Economics, University of Southern California, US;Michelle Zhao: Stanford University, US; Alex R. Cook: Saw Swee Hock School of Public Health, NationalUniversity of Singapore and National University Health System, Singapore; Ming Zhe Chong: Saw SweeHock School of Public Health, National University of Singapore and National University HealthSystem, Singapore; Jian-Min Yuan: Department of Epidemiology, Graduate School of Public Health,University of Pittsburgh, US; and Division of Cancer Control and Population Sciences, UPMC HillmanCancer Center, US; Stefan Ma: Epidemiology and Disease Control Division, Ministry of Health, Singapore;





https://dx.doi.org/10.1142/S0116110522500056

https://dx.doi.org/10.1142/S0116110522500056

disability prevalence by educational attainment. The model also projectsincreasing disparities in the prevalence of activities-of-daily-living disabilityand other chronic diseases between those with low and high educationalattainment. Despite overall increases in educational attainment, all elderly,including those with a college degree, experience an increased burden offunctional disability and chronic diseases because of survival to older ages.These increases have significant economic and social implications, includingincreased medical and long-term care expenditures, and an increased caregiverburden.

Keywords: ADL disability, microsimulation model, Republic of Korea,Singapore

JEL code: I14

I. Introduction

The Republic of Korea and Singapore are home to the most rapidly aging

populations in the world and are projected to have continually increasing proportions

of older adults in the coming decades. These countries are projected to be “top-heavy”

societies because of continued increase in life expectancy coupled with low,

below-replacement fertility rates. Although the changes in population size and

demographics in these countries have been studied, the complex evolution of the

health and functional disparities of the future elderly has not been fully explored.

These changes will have important implications for social protection systems,

including the financing and delivery of long-term care and health care.

Both Singapore and the Republic of Korea experienced rapid industrialization,

with exceptionally high average annual real gross domestic product growth rates of

Kelvin Bryan Tan: Policy Research and Economics Office, Ministry of Health, Singapore; Tze Pin Ng:Yong Loo Lin School of Medicine, National University of Singapore, Singapore; Koh Woon-Puay:Healthy Longevity Translational Research Programme, Yong Loo Lin School of Medicine, NationalUniversity of Singapore, Singapore; and Singapore Institute for Clinical Sciences, Agency for ScienceTechnology and Research, Singapore; Joanne Yoong: Center for Economic and Social Research,University of Southern California, US; Jay Bhattacharya: School of Medicine, Stanford University, US;Karen Eggleston: Shorenstein Asia Pacific Research Center, Stanford University, US. We thank KenwinMaung for the helpful discussions and statistical support that have contributed to this work. The researchwas supported by the National Medical Research Council (MOH-HSRGMH18may-0001) and the NationalInstitute on Aging of the National Institutes of Health (P30AG024968). The Singapore Chinese HealthStudy was supported by US National Institutes of Health (R01CA144034 and UM1CA182876). KohWoon-Puay was supported by the National Medical Research Council, Singapore (NMRC/CSA/0055/2013). The funding sources had no role in the design and conduct of the study; collection, management,analysis, or interpretation of the data; preparation, review, or approval of the paper; and decision to submitthe paper for publication. The Asian Development Bank recognizes “China” as the People’s Republic ofChina and “Korea” and “South Korea” as the Republic of Korea.



7.38% and 7.42%, respectively, from 1960 to 2017 (World Bank 2017). They transited

from low-income to high-income economies within 50 years. Out of 188 economies

ranked in the Human Development Index, Singapore and the Republic of Korea are

considered to be highly developed economies with Human Development Index values

of 0.925 (ranked 4th) and 0.901 (ranked 18th), respectively (United Nations

Development Programme 2018). Their high scores are in part due to good universal

health care accessibility and high levels of educational attainment.

Our conceptual framework for studying the Republic of Korea and Singapore is

based on hypotheses and evidence from a large literature showing that human capital

accumulation is associated with broad socioeconomic development (Lee and Mason

2010) and can have a large effect on individuals and the economy. Good health allows

for more schooling and better absorption of knowledge, and education provides people

with the skills that translate into higher labor productivity and wages (Barro 2013).

Thus, higher educational attainment can affect health directly (e.g., through better

knowledge of how lifestyle choices impact health) as well as through its association

with higher socioeconomic status, including (depending on the health system)

improved financial access to quality health care and health insurance (Cutler and

Lleras-Muney 2010; Leopold and Engelhartdt 2012; Basu, Jones, and Dias 2018).

While many studies have examined the return-on-investment from education, few

have investigated how health disparities across education groups vary across time in

Asia. This endeavor may be challenging given the complex economic and social

interdependence between demographic change and population health. The Republic of

Korea and Singapore are both fast-growing industrialized countries, with sparse

mineral resources and human capital as their primary resource. While both the

Republic of Korea and Singapore are aging rapidly, they have differing trends in

educational attainment across genders. As the formation of human capital developed

differently over time in each country, this might have had different impacts on health.

Therefore, a study involving both countries allows us to examine human capital’s

comparative impacts on population health and the consequent disparities over time.

This may provide supporting evidence to aid policy makers in formulating effective

responses, especially for disadvantaged groups.

Education is linked to prevention of chronic disease (Harris 2007), disability

(Lutz et al. 2007, K. C. and Lentzner 2010) and mortality (Baker et al. 2011).

Higher-educated and wealthier individuals are less prone to physical disability later in

life (Gjonça, Tabassum, and Breeze 2009), and they possess longer disability-free life

expectancy (Minicuci and Noale 2005, Nogueira and Reis 2014). Thus, health

EDUCATIONAL GRADIENTS IN DISABILITY AMONG ASIA’S FUTURE ELDERLY 53


differences by educational attainment are expected, although it is not clear if these

educational gradients in health are steepening as the average education or human

capital of a society increases. Moreover, other factors might crowd out education’s

beneficial impact on health outcomes (Behrman and Wolfe 1987; Fuchs 2004; Conti,

Heckman, and Urzua 2010). A recent study in the People’s Republic of China found

steepening educational gradients in health, although the gradient flattens with age and

is mediated by economic and social resource variables (Chen et al. 2017). Lastly, age

and income also have a complex interplay with morbidity and mortality (McDonough

et al. 1997, Schnittker 2004, Kennedy et al. 2014). As such, a projection of the health

and functional disability of the middle-aged and elderly population needs to account

for socioeconomic and demographic trends, as well as competing risk factors to

provide an accurate picture for policy making.

Improvements in the educational composition of the population can also have a

considerable impact on decreasing the future prevalence of disability at any given age

(Lutz et al. 2007). There are many mechanisms. For example, Caldwell (1996) showed

that women’s education has a direct effect on the health status of the population

through improved maternal–child health. Gender disparities in education also

narrowed as educational attainment increased in the Republic of Korea and Singapore.

In the past, women were unlikely to be formally employed and had fewer educational

opportunities. However, much has changed and both countries have made great strides

in education; for example, in Singapore, female literacy rates rose from 34% in 1957 to

94% in 2010 (Moi 2010). This resulted in increased female wages (median income

rose from $2,863 in 2010 to $3,518 in 2014) and labor force participation (21.6% in

1957 to 58.6% in 2014), and induced better health outcomes, as an increase in literacy

rates is associated with increases in health-care access and utilization (Sudore et al.

2006, Bustamante et al. 2012). In addition, higher education is associated with lower

mortality, an improvement in psychosocial health, a reduction in bad lifestyle habits,

and lower occupational and health risks from jobs that demand manual labor such as

agriculture or construction (Currie and Hyson 1999, Deaton and Paxson 1999,

National Research Council 2012). Lower mortality among the better-educated, in turn,

implies survival to older ages. As a result, our conceptual framework suggests that the

increasing educational attainment of the Korean and Singaporean populations will lead

to lower age-standardized disability rates but higher overall disability rates because of

a higher proportion of the oldest-old and their associated disabilities. The educational

gradient in disability may decline or increase, depending on the relative health benefit

of better education among the low- versus high-educated groups.



To take into consideration the dynamic interplay between a rapidly aging

population, increases in educational attainment, age-dependent susceptibility to health

conditions, and functional disability in Asia, we develop a Future Elderly Model

(FEM) for the Republic of Korea and Singapore. The FEM is a sociodemographic and

economic state-transition microsimulation model with flexibility as a key strength: it

can take into account the evolving educational attainment of future elderly cohorts and

competing risks when projecting functional disability outcomes for the future elderly

population. We employ these models to project and compare functional disability

across the two countries, including disparities in functional disability prevalence by

educational attainment and gender, up to 2050. The models in this paper are similar in

spirit to the American FEM (Goldman et al. 2005), but with country-specific

prevalence rates of diseases, demographic trends, and transition probabilities for

different health states. The Republic of Korea’s estimates are based on the harmonized

version of the Korean Longitudinal Study of Aging (KLoSA), while the Singapore

model is based on the Singapore Chinese Health Study (SCHS).

We find that from 2015 to 2050, the educational gradient in health and functional

status will steepen, with a wider gap in the prevalence of functional disability between

those with low and high educational attainment. For example, in Singapore, the

disparity in functional disability prevalence, as measured by limitation in at least one

activity of daily living (ADL), starts at a difference of 9.5 percentage points in 2015

and climbs by 2050 to a difference of over 23 percentage points (31.9% among those

with low education, compared to 8.7% among the college-educated). Functional

disability as measured by instrumental activities of daily living (IADL) shows a

similar pattern. The Republic of Korea’s model projects even larger disparities by

2050, as exemplified by a difference of 41.7 percentage points in the prevalence of any

functional disability between low- and high-educational attainment groups (i.e., 51.7%

among those with less than a high school education, compared to 10.0% among the

college-educated). These disparities are driven almost entirely by population aging and

increasing education so that the population in the lower educational attainment group

is a small, selected disadvantaged group. Improved survival of those with low

education and their declining share of the population leads to their concentration

among the oldest-old (therefore with the highest disability burden), especially in the

Republic of Korea.

The remainder of this paper is organized as follows. Section II describes the data

and method for operationalizing the FEM for the Republic of Korea and Singapore.

Section III presents the results on functional disability and the associated educational

gradient in the future elderly by country. The last section identifies economic and



policy implications, as well as areas for mutual learning and improvement. Further

details on the models for individual countries are reported in the appendixes.1

II. Method

A. Data

Data were drawn from comprehensive longitudinal surveys of older adults from

two rapidly aging economies, the Republic of Korea and Singapore. The KLoSA is a

longitudinal study of individuals over age 45 years old in the Republic of Korea. It was

designed to help researchers and policy makers better understand the socioeconomic

determinants and consequences of aging. The survey includes a rich set of questions

regarding the economic standing, physical and psychological health, demographics,

and social networks of aged persons. We used the harmonized KLoSA dataset, which

contains Waves 1, 2, 3, and 4 as of October 2015. The first wave of the KLoSA survey

was conducted in 2006 and included 10,254 respondents aged 45 years and over. The

second wave was done in 2008 with 8,688 respondents. The third wave was done in

2010 with 7,920 respondents. The fourth wave was done in 2012 with 7,486

respondents.

The SCHS is a prospective cohort study of ethnic Chinese men and women aged

45–74 years old in the baseline who were followed up for a mean duration of 12 years.

Inclusion criteria were similar to an earlier study (Chen et al. 2019b), with respondents

being either citizens or permanent residents, residing in government-built housing, and

belonging to either of the two major dialect groups of Chinese in Singapore (Hokkien

and Cantonese). The baseline study (n ¼ 63,257) was collected between 1993 and

1999, follow-up 1 (n ¼ 52,325) was collected between 1999 and 2004, and follow-up

2 (n ¼ 39,528) was collected between 2006 and 2010. At baseline, each participant

completed an in-person interview at their home using a structured questionnaire that

requested information about demographic characteristics, self-reported height and

weight, smoking status, current physical activity, occupational exposure, medical

history, and family history of cancer. A follow-up telephone interview asked

participants for an update on their tobacco and alcohol use as well as medical history.

We used the Singapore Longitudinal Aging Study (SLAS)—a smaller cohort

that consists of 2,804 subjects aged 55 years or above interviewed in 2004–2005,

1The online Appendixes can be accessed at: https://sph.nus.edu.sg/wp-content/uploads/2022/03/Appendix_FEM_SGSK_v2.pdf.



https://sph.nus.edu.sg/wp-content/uploads/2022/03/Appendix_FEM_SGSK_v2.pdf

https://sph.nus.edu.sg/wp-content/uploads/2022/03/Appendix_FEM_SGSK_v2.pdf

2007–2008, and 2010–2011—to model functional disability. Older adults who were

citizens or permanent residents aged 55 years or above were identified by door-to-door

census and invited to participate voluntarily in the study.

B. Microsimulation Model

The FEM microsimulation model includes disease transitions and accounts for

population aging within the middle-aged and elderly population. Using the FEM, we

are able to model individual-level longitudinal dynamics, allowing for greater

heterogeneity than cell-based approaches. Moreover, populations entering the

microsimulation reflect current trends in sociodemographic characteristics and health

status based on data from population surveys. The FEM was successfully implemented

first to support decision-making related to Medicare and Medicaid, the public health

insurance and welfare programs for the elderly and needy, respectively, in the United

States (Goldman et al. 2005). Since then, the FEM has been used for dozens of

important papers for informing health and social issues in aging societies such as the

United States, European Union, and Japan (Michaud et al. 2011, Chen et al. 2016,

Chen et al. 2019b). Its ongoing development is supported by the National Institute on

Aging through the University of Southern California’s Roybal Center for Health

Policy Simulation (University of Southern California 2019). In using the same FEM

model to project functional disability for Singapore and the Republic of Korea, we aim

to: (i) highlight potential differences in the aging experience by gender and education

in each country to inform social and health-care policy, and (ii) provide a common

platform for international comparison to identify and compare challenges across

countries.

Individuals in the FEM are characterized by their socioeconomic status and

health states. From period to period, they transit from one health state to another with

probabilities that are estimated from the data (see Section II.C). The Singapore FEM is

modeled with a 6-year time step, as the SCHS was surveyed with a mean of 6 years

between each study wave. The Republic of Korea’s FEM model follows a 2-year time

step mirroring the mean time between waves of the KLoSA study. We then interpolate

linearly the outcomes to yearly rates. As there was no difference in results when

presenting with 6-year compared to 5-year intervals, we reported the 5-year charts for

ease of interpretation. Also, a replenishing cohort is introduced every period at the

default starting age, set at 55–60 years old for Singapore and 51–52 years old for the

Republic of Korea. Individuals within the simulation exit from the model at the time of

death. Hence, although some individuals are lost to mortality, the population expands



with renewal. KLoSA and SCHS are post-stratified by age group, gender, and

education. They are then rebalanced with the corresponding population weights using

population projections obtained from the International Institute for Applied Systems

Analysis and the Vienna Institute of Demography projections for Singapore. In the

Republic of Korea, the education profile was rebalanced using the Barro–Lee

projections. This adjustment is made for both the initial cohort as well as the

replenishing cohorts over the simulation cycles. The distribution of socioeconomic

status, particularly education, and health status vary across the new cohorts depending

on the trends derived from the data and on the time of their introduction. In each

simulation cycle, the population moves forward in time and individuals transit to new

states probabilistically; this results in an observable distribution of health in the

population at the end of each cycle. The replenishing cohort model, which determines

the distribution of states among individuals in the new cohorts, is estimated in Stata,

while the overarching simulation is implemented in C++ for computational efficiency.

To calculate the Monte Carlo confidence intervals, we repeated the sample process

described above 1,000 times and used the 2.5th and 97.5th percentiles as the prediction

interval, similar to other studies (Pericchi and Walley 1991, O’Brien et al. 2009, Chen

et al. 2019a).

C. Health Transition Model

The FEM uses a discrete piecewise linear hazard model on longitudinal data to

estimate transitional probabilities for selected measures of health status. The hazard of

getting a disease, dying, or becoming disabled depends on risk factors such as gender,

education, lifestyle, comorbidities, functional status, and age (Appendix 4).

The resulting hazards will then be used as parameters in the microsimulation model

to determine how individuals transit between health states from one period to the next.

The health transition model is estimated in Stata. We used probit regressions to

estimate the probability of transition to each health condition, controlling for

education, demographic variables, and comorbidities at the previous period.

We treated all diseases as an absorbing state—as there is currently no cure for

chronic diseases such as diabetes, hypertension, heart disease, or stroke—in reflection

of the question asked: “Have you ever been told . . . by a doctor?” In the Republic of

Korea, variables were measured with a 2-year lag in the KLoSA (Jang 2016).

In Singapore, all independent variables were measured with a 6-year lag in the SCHS

and represent the respondent’s characteristics from 1993 to 2010 (National University

of Singapore 2018). In both countries, as diseases were treated as an absorbing state,

we assumed that those with a chronic condition in the previous wave would continue



to have that disease in the current wave. As such, transition probabilities were only

estimated on those who did not suffer from a specific condition at the time of the

previous survey wave.

The transition model is the core of the FEM. The transition model describes the

health status of individuals and how it evolves between periods over the course of the

simulation. The model’s main variables consist of sociodemographic and behavioral

variables denoted as AGEit, GENDERi, EDUCATIONi, MARITAL_STATUSi,

SMOKE_CURRENTit, SMOKE_EVERit, and BMIit, as well as a set of indicator

variables for diseases: DISEASEit ¼ fDISEASE1it, . . . , DISEASENitg with reference

to the N chronic conditions, where i refers to the individual and t refers to the time.

In addition, a mortality indicator variable denoted DIEDit was created and is equal to 1

if the simulated individual i dies at time t during the simulation or 0 otherwise.

The baseline cohort is defined at the initial time period (t ¼ 1). The variables of AGEit,

GENDER i, EDUCATION i, MARITAL_STATUS i, SMOKE_CURRENT it,

SMOKE_EVERit, BMIit, and DISEASEit in the baseline cohort are retrievable from

the KloSA and SCHS datasets. All of these samples are allocated with a specific

weight such that the number of individuals in each age, sex, and education level

category is consistent with that of the population distribution in each country.

Figure 1 provides a schematic overview of the model. The FEM simulation starts

from 2008 with initial populations aged 51 years and above in the Republic of Korea

and Singapore. We then predict outcomes using our estimation for the initial cohort

(Appendix 4) to begin the simulation. Individuals who survive are defined as those for

whom the transition model predicts DIEDit not equal to 1 (nondecedents) at the end of

the corresponding year. Projections on policy outcomes are then made for that year.

Thereafter, the FEM moves to the following time period within the simulation cycle

when a new cohort enters. The current cohort and the replenishing cohort form the

new population of those aged 51 years and above (Republic of Korea) or 55 years and

above (Singapore), which then proceeds through the transition module as before. The

replenishing cohort enters the simulation every 2 years for the Republic of Korea and

every 6 years for Singapore. This process of replenishing new cohorts and transiting

across various health states is repeated until we reach the final year of the simulation.

Each individual module is explained in more detail in Appendix 4.

D. Functional Disability Model

We defined functional disability as having any ADL disability such as

limitation in washing, dressing, feeding, toileting, mobility, and transferring



(Ministry of Health 2018), or having any IADL disability such as limitation in taking

transportation, shopping, managing money, making phone calls, doing household

chores, and meal preparation. The former measures an individual’s physical ability to

perform basic tasks, whereas the latter includes higher-order tasks that measure an

individual’s engagement and management of resources. We projected future functional

disability using probit regressions for any ADL and IADL disability, and ordered

probit regressions for disability measured on an ordinal scale. In the ordered probit

regression for each country, we summed the number of ADL functional disabilities in

washing, dressing, feeding, toileting, mobility, and transferring. We then modeled

adlstat as three categories: (i) having no disability, (ii) having one or two disabilities,

or (iii) having three or more disabilities. We modeled adlstat as an ordinal outcome

using ordered probit regression. Similar analysis was performed for IADL.

The covariates included were age, gender, educational attainment, body mass index,

marital status, and chronic diseases.

Figure 1. Schematic Overview of Future Elderly Model Simulation in theRepublic of Korea and Singapore

Population

T0

Replenishingcohort

T1

Health andspendingoutcomes

T2

Replenishingcohort

T1


T1


T0

Health Transitions Module Module

Policy OutcomesReplenishing Cohort Module

Population

T1

Population

T2

Source: Authors’ illustration.



E. Educational Attainment

Only educational attainment was modeled according to countries’ specific

categorization. In the Republic of Korea, low attainment is defined as less than high

school, middle attainment is defined as high school, and high attainment is defined as

having a college education. In Singapore, low attainment is defined as having a

primary education or less (6 years of education or less); middle attainment as having a

secondary school education, technical education, or diploma (7–15 years of

education); and high attainment as having a college education or above (at least 15

years of education). We wanted to capture each country’s specific distribution of

relative educational attainment, as this distribution better captures each country’s

dispersion in human capital and helps us therefore to understand the relationship

between educational differences and health disparities. Given this categorization, those

with low educational attainment in the Republic of Korea are on average more

educated than those categorized as low educational attainment in Singapore, whereas

high educational attainment is comparable across the two countries. Our appendixes

also include figures depicting the model projections when using a consistent definition

of low educational attainment for both countries, defined as having less than a college

education. Education was used as a measure of socioeconomic status, and we used

these two terms interchangeably.

III. Results

A. Country Characteristics

The United Nations (UN) has projected both the Republic of Korea and

Singapore to age rapidly in the future (Figure 2). This aging within the middle-aged

and older population reflects the large post-war baby boomer cohorts, subsequent

low fertility, and improved survival (lower age-specific mortality) across the

population. Using UN data, the proportion aged 65 years and above in 2015 was

13.0% in the Republic of Korea and 11.7% in Singapore; in 2050, these shares will

have increased to 35.3% and 33.6%, respectively. The old-age support ratio, defined

as the population aged 20–64 years (working age) divided by the population aged 65

years and above, declines at similar rates in both countries from about 6 working

adults per elderly person in 2015 to about 1.5 working adults per elderly person in

2050. Using the UN projections, Figure 3 shows that both countries will have an

increasing percentage of the elderly with a college education among both men and



women. The projection also shows that the number of the elderly with low or no

education will decrease continuously. Using UN data, in 2015 the percentage of

older adults with low or no formal education was higher in females compared to

males: for the Republic of Korea, 64% (females) and 30% (males); for Singapore,

Figure 2. Share of Population Aged 65 Years and Above

0

5

10

15

20

25

30

35

40

2015 2020 2025 2030 2035 2040 2045 2050

Perc

enta

ge a

ged

65+

Republic of Korea Singapore

Source: United Nations Department of Economic and Social Affairs. 2017. World Population Prospects: The2017 Revision. https://www.un.org/development/desa/publications/world-population-prospects-the-2017-revi-sion.html.

Figure 3. Percentage of Low and High Educational Attainment among Those Aged65 Years and Above

Low edu female Low edu male

High edu female High edu male

–64%

1970 1980 1990 2000 2010 2020 2030 2040 2050

Republic of Korea

–30%

1%%–4%

39% F

1970 1980 1990 2000 2010 2020 2030 2040 2050

Singapore

–18%

50% M

0%

20%

40%

60%

80%

100%

–67%

–6%

36% F

–13%

42% M

0%

20%

40%

60%

80%

100%

3%%% FF M

18% F 13% M

–46%

edu ¼ education, M ¼ male, F ¼ female.Source: International Institute for Applied Systems Analysis and the Vienna Institute of Demography projectionextracted from United Nations Department of Economic and Social Affairs. 2017. World Population Prospects:The 2017 Revision. https://www.un.org/development/desa/publications/world-population-prospects-the-2017-revision.html.



67% (females) and 46% (males). By 2050, the percentage is projected to have

decreased substantially for both countries: for the Republic of Korea, 3% (females)

and 1% (males); for Singapore, 18% (females) and 13% (males). Over the same

period, there will be a sharper increase in college education among women than

men: for females, from 4.3% in 2015 to 38.9% in 2050 in the Republic of Korea

and from 6.0% in 2015 to 36.1% in 2050 in Singapore; compared with males, from

18.1% in 2015 to 50.2% in 2050 in the Republic of Korea and from 13.1% in 2015

to 41.8% in 2050 in Singapore.

Thus, educational attainment among our replenishing cohorts was projected to

increase (Figure 4). In the Republic of Korea, the proportion of primary school

graduates entering middle school was 99.9% in 2000, and the proportion of middle

school graduates entering high school was 99.7% (Ministry of Education 2000).

Our model captures how this change in educational attainment of future older adults

will shape trends in health, incorporating both differences in mortality and morbidity.

Based on our FEM simulations, Figure 5 shows the aging pyramid in both countries

with growing numbers of elderly with college education.

In both countries, chronic conditions and functional disability were projected to

rise during the review period. IADL disability is projected to be much higher in

Singapore than the Republic of Korea in 2050 (Figure 6). The prevalence of chronic

disease burden for diabetes, stroke, and heart disease will be higher in the Republic of

Korea.

By 2050, the total number of people with any ADL disability was projected to

increase to 275,000 (18.9%) in Singapore and 2.48 million (15.9%) in the Republic of

Figure 4. Educational Attainment among Future Middle-Aged and Elderly in Singapore andthe Republic of Korea: Trends in the Replenishing Cohorts in the Future Elderly Model

Highsch ¼ high school, PSLE ¼ primary school leaving examination.Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.



Korea, with an increasing rate of disability among those aged 65 years and above

(Figure 7). To adjust for the differences in the age structure and account for the strong

influence of age on the risk of ADL disability, the prevalence rates were

age-standardized to the combined populations of the Republic of Korea and Singapore

in 2015 and expressed as 100,000 person-years for the elderly aged 65 years and

above. Data were analyzed using 5-year age groups (65–69, 70–74, 75–79, 80–84, and

85+ years) for both countries. After changes in the age structure were removed, both

countries exhibit very similar trends. The age-standardized rates (ASRs) for both

countries remained stable across time, with slight declines in age-standardized rates in

later years that partly reflect the increase in the share with a college education, who are

in better health.

Figure 5. Population Pyramids by Education for the Middle-Aged and Elderly, 2020–2040

(b) Singapore

(a) Republic of Korea

edu ¼ education.Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.



Figure 6. Proportion of Those Aged 65 Years and Above with Disability andChronic Conditions, 2050

0.0

0.1

0.2

0.3

0.4

Prop

ortio

n w

ith c

hron

ic c

ondi

tions

0.0

0.1

0.2

0.3

0.4

Legend

IADL Overall Diabetes Heart Disease Stroke ADL Overall

Republic of KoreaSingapore

2015 2020 2025 2030 2035 2040 2045 20502015 2020 2025 2030 2035 2040 2045 2050

ADL ¼ activities of daily living, IADL ¼ instrumental activities of daily living.Note: Confidence bounds represent the 95% prediction interval from 1,000 Monte Carlo simulation.Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.

Figure 7. Disability Rates (Actual and Age-Standardized) per 100,000 People Aged 65 Yearsand Above

Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.



B. Educational Gradients in Disability

The model projects a continuing—and in some cases steepening—educational

gradient in disability and chronic conditions, with those of lower education

experiencing a higher prevalence of disability and chronic diseases. This result was

consistent across genders, although varying in magnitude across countries over time.

For men in Singapore, in 2015, the difference in the prevalence of poor health

between elderly with low and high educational attainment was largest for IADL

disability (21.3 percentage points), followed by diabetes (6.4 percentage points) and

ADL (5.8 percentage points). For women in Singapore, the difference was largest for

IADL disability (16.7 percentage points), followed by diabetes (15.5 percentage

points) and hypertension (15.2 percentage points). For men in the Republic of Korea,

in 2015 the absolute difference in the prevalence between elderly with low and high

educational attainment was largest for IADL disability (7.0 percentage points),

followed by heart disease (5.2 percentage points), and ADL disability (4.1 percentage

points). For women, the difference was largest for hypertension (20.0 percentage

points), followed by IADL disability (12.2 percentage points) and diabetes (10.7

percentage points).

Over time, ADL disability is projected to rise sharply for the elderly with low

educational attainment in the Republic of Korea from 10.3% in 2015 to 51.7% in

2050, while increasing at a slower pace in Singapore from 13.1% to 31.9% over the

same period (Figure 8a). In part, this reflects the fact that low-educated Koreans will

be concentrated exclusively among the oldest-old by 2050 (see Figures 2 and 3),

rather than also including those in their 60s and 70s as in Singapore. For female

elderly with high educational attainment, the prevalence of ADL disability was

projected to be higher in Singapore compared to the Republic of Korea. The disparities

in the burden of ADL disability appeared larger for the Republic of Korea and are

projected to increase in the future as the low-education group becomes a tiny fraction

of the population, whereas the difference is projected to stabilize for Singapore in the

future. Similar trends were observed in both males (Figure 8b) and females

(Figure 8c). Also, the difference in ADL disability among low- and high-education

groups is projected to be larger for females than males in both countries.

IADL disability is projected to rise sharply for elderly with low education in the

Republic of Korea from 19.9% in 2015 to 71.8% in 2050, while it increases less

quickly for Singapore from 27.2% to 57.0% during the same period (Figure 9a).

Although the overall projected prevalence of IADL disability is lower in Singaporean

elderly with high education, this is driven mainly by males (Figures 9b and 9c).



Figure 8. Disparities in Activities of Daily Living Disability byCountry, Education, and Gender

0.0

0.2

0.4

0.6

Prev

alen

ce o

f A

DL

dis

abili

ty

(a) Overall

i) P(ADL|low ed)

0.0

0.2

0.4

0.6

Prev

alen

ce o

f A

DL

dis

abili

ty

ii) P(ADL|high ed)

0.0

0.2

0.4

0.6

Dif

fere

nce

in p

reva

lenc

e

2015

iii) P(ADL|low ed) − P(ADL|high ed)

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

SINKOR SINKOR KOR SIN

Dif

fere

nce

in p

reva

lenc

e

ii) P(ADL|high ed)

(b) Males

i) P(ADL|low ed) iii) P(ADL|low ed) − P(ADL|high ed)

0.0

0.2

0.4

0.6

0.8

Prev

alen

ce o

f A

DL

dis

abili

ty

0.0

0.2

0.4

0.6

0.8

Prev

alen

ce o

f A

DL

dis

abili

ty

0.0

0.2

0.4

0.6

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050


Dif

fere

nce

in p

reva

lenc

e

iii) P(ADL|low ed) − P(ADL|high ed)

0.0

0.2

0.4

0.6

Prev

alen

ce o

f A

DL

dis

abili

ty

(c) Females

i) P(ADL|low ed)

0.0

0.2

0.4

0.6

Prev

alen

ce o

f A

DL

dis

abili

ty

ii) P(ADL|high ed)

0.0

0.2

0.4

0.6

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050


ADL ¼ activities of daily living, high ed ¼ high education, KOR ¼ Republic of Korea, low ed ¼ low education,SIN ¼ Singapore.Notes: From left to right, by country:1. Proportion of ADL disability among elderly with low education from 2015 to 2050.2. Proportion of ADL disability among elderly with high education from 2015 to 2050.3. Absolute difference in proportion of ADL disability between elderly in high- and low-education groups from2015 to 2050. Lines represent the mean difference in ADL disability prevalence between high- and low-educationfor the Republic of Korea and Singapore. Confidence bounds represent the 95% prediction interval from MonteCarlo uncertainty of 1,000 simulation.Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.



Figure 9. Disparities in Instrumental Activities of Daily Living Disability byCountry, Education, and Gender

0.0

0.2

0.4

0.6

0.8

Prev

alen

ce o

f IA

DL

dis

abili

ty

(a) Overalli) P(IADL|low ed)

0.0

0.2

0.4

0.6

0.8

Prev

alen

ce o

f IA

DL

dis

abili

ty

ii) P(IADL|high ed)

0.0

0.2

0.4

iii) P(IADL|low ed) − P(IADL|high ed)

0.6

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

Dif

fere

nce

in p

reva

lenc

e


0.00

0.25

0.50

0.75

1.00

Prev

alen

ce o

f IA

DL

dis

abili

ty

(b) Males)de wol|LDAI(P )i

0.00

0.25

0.50

0.75

1.00

Prev

alen

ce o

f IA

DL

dis

abili

ty

ii) P(IADL|high ed)

0.0

0.2

0.4

0.6

2015

2020


2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

Dif

fere

nce

in p

reva

lenc

e


0.0

0.2

0.4

0.6

0.8

Prev

alen

ce o

f IA

DL

dis

abili

ty

(c) Femalesi) P(IADL|low ed)

0.0

0.2

0.4

0.6

0.8

Prev

alen

ce o

f IA

DL

dis

abili

ty

ii) P(IADL|high ed)

0.0

0.2

0.4

0.6

2015

2020


2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

Dif

fere

nce

in p

reva

lenc

e


IADL ¼ Instrumental activities of daily living, high ed ¼ high education, KOR ¼ Republic of Korea, low ed ¼low education, SIN ¼ Singapore.Notes: From left to right, by country:1. Proportion of IADL disability among elderly with low education from 2015 to 2050.2. Proportion of IADL disability among elderly with high education from 2015 to 2050.3. Absolute difference in proportion of IADL disability between elderly in high- and low-education groups from2015 to 2050. Lines represent the mean difference in IADL disability prevalence between high- and low-education for the Republic of Korea and Singapore. Confidence bounds represent the 95% prediction intervalfrom Monte Carlo uncertainty of 1,000 simulation.Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.



The educational disparities in the burden of IADL disability is projected to increase in

the Republic of Korea in the future, whereas the difference is projected to stabilize for

Singapore. Also, the educational disparities appeared larger for females in both

countries.

For the most part, diabetes prevalence is projected to increase among elderly in

both the low-education group (Republic of Korea: 23.3% in 2015 to 31.0% in 2050;

Singapore: 25.1% in 2015 to 37.1% in 2050) and the high-education group (Republic

of Korea: 20.7% in 2015 to 29.9% in 2050; Singapore: 16.6% in 2015 to 20.4% in

2050) (Figure 10a). However, in the Republic of Korea, there is a decline in the

growth rate of diabetes prevalence among elderly with low education in both males

(Figure 10b) and females (Figure 10c). As such, there is a decreasing trend in the

difference in diabetes prevalence among the low- and high-education groups for the

Republic of Korea, but a greater difference in diabetes prevalence in Singapore.

Low-educated males in Singapore have a higher prevalence of diabetes compared to

the Republic of Korea.

While stroke prevalence is projected to increase rapidly in the Republic of Korea

among the low-educated (9.3% in 2015 to 21.8% in 2050) and high-educated (9.6% in

2015 to 15.2% in 2050) elderly, Singapore’s prevalence is projected to increase at a

much slower rate for both the low educated (7.1% in 2015 to 13.1% in 2050) and high

educated (4.7% in 2015 to 6.6% in 2050) (Figure 11a). A similar pattern is seen for

both males (Figure 11b) and females (Figure 11c). The educational difference in

stroke prevalence increased in the Republic of Korea and remained stable in Singapore

during the review period.

Heart disease prevalence is projected to increase in both the Republic of Korea

and Singapore among the low-education group (Republic of Korea: from 12.9% in

2015 to 22.3% in 2050; Singapore: from 13.7% in 2015 to 24.7% in 2050) and in the

high-education group in the Republic of Korea from 14.8% in 2015 to 19.9% in 2050.

In Singapore, there is a slight decline from 12.3% in 2015 to 12.1% in 2050

(Figure 12a). In the Republic of Korea, males were less likely to experience heart

disease than their female compatriots, but the converse was found in Singapore.

As such, while the overall growth in heart disease prevalence in the Republic of Korea

tapers off during the review period, the decline was driven by the drop in prevalence

among males (Figure 12b), as heart disease in females continued to increase

(Figure 12c). By contrast, in Singapore the overall disparities in the prevalence of

heart disease were projected to increase. The difference in heart disease prevalence

among the low and high educated is projected to decrease in the later years in the

Republic of Korea and remain stable in Singapore.



Figure 10. Disparities in Diabetes Prevalence by Country, Education, and Gender

(c) Females

0.0

0.1

0.2

0.3

0.4

Prev

alen

ce o

f di

abet

es

(a) Overall

i) P(Diabetes|low ed)

0.0

0.1

0.2

0.3

0.4

Prev

alen

ce o

f di

abet

es

ii) P(Diabetes|high ed)

0.00

0.05

0.10

0.15

iii) P(Diabetes|low ed)−P(Diabetes|high ed)

0.20

Dif

fere

nce

in p

reva

lenc

e

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050


Dif

fere

nce

in p

reva

lenc

e

ii) P(Diabetes|high ed)

0.0

0.1

0.2

0.3

0.4

0.5

Prev

alen

ce o

f di

abet

es

(b) Males

i) P(Diabetes|low ed)

0.0

0.1

0.2

0.3

0.4

0.5

Prev

alen

ce o

f di

abet

es

0.0

0.1

0.2

2015

2020


0.3

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050


Dif

fere

nce

in p

reva

lenc

e

ii) P(Diabetes|high ed)P(Diabetes|low ed)

0.0

0.1

0.2

0.3

0.4

Pre

vale

nce

of d

iabe

tes

0.0

0.1

0.2

0.3

0.4

Pre

vale

nce

of d

iabe

tes

0.0

0.1

0.2


0.3

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050


high ed ¼ high education, KOR ¼ Republic of Korea, low ed ¼ low education, SIN ¼ Singapore.Notes: From left to right, by country:1. Proportion of diabetes among elderly with low education from 2015 to 2050.2. Proportion of diabetes among elderly with high education from 2015 to 2050.3. Absolute difference in proportion of diabetes between elderly in high- and low-education groups from 2015 to2050. Lines represent the mean difference in diabetes prevalence between high- and low-education for theRepublic of Korea and Singapore. Confidence bounds represent the 95% prediction interval from Monte Carlouncertainty of 1,000 simulation.Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.



Figure 11. Disparities in Stroke Prevalence by Country, Education, and Gender

(b) Males

(c) Females

(a) Overall

high ed ¼ high education, KOR ¼ Republic of Korea, low ed ¼ low education, SIN ¼ Singapore.Notes: From left to right, by country:1. Proportion of stroke among elderly with low education from 2015 to 2050.2. Proportion of stroke among elderly with high education from 2015 to 2050.3. Absolute difference in proportion of stroke between elderly in high- and low-education groups from 2015 to2050. Lines represent the mean difference in stroke prevalence between high- and low-education for the Republicof Korea and Singapore. Confidence bounds represent the 95% prediction interval from Monte Carlo uncertaintyof 1,000 simulation.Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.



Figure 12. Disparities in Heart Disease Prevalence by Country, Education, and Gender

0.0

0.1

0.2

0.3

Prev

alen

ce o

f he

art d

isea

se

0.0

0.1

0.2

0.3

Prev

alen

ce o

f he

art d

isea

se

ii) P(Heart Disease|high ed)

0.00

0.05

0.10

0.15

0.20

Dif

fere

nce

in

pre

vale

nce

2015

2020

iii)P(Heart Disease|low ed)−P(Heart Disease|high ed)

2025

2030

(a) Overalli) P(Heart Disease|low ed)

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050



0.0

0.1

0.2

0.3

0.4

Prev

alen

ce o

f he

art d

isea

se

(b) Malesi) P(Heart Disease|low ed)

0.0

0.1

0.2

0.3

0.4

Prev

alen

ce o

f he

art d

isea

se


0.00

0.05

0.10

0.15

0.20

Dif

fere

nce

in p

reva

lenc

e

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050



Dif

fere

nce

in p

reva

lenc

e


(c) Femalesi) P(Heart Disease|low ed)

0.0

0.1

0.2

0.3

0.4

Pre

vale

nce

of h

eart

dis

ease

0.0

0.1

0.2

0.3

0.4

Pre

vale

nce

of h

eart

dis

ease

0.00

0.05

0.10

0.15

0.20

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050


high ed ¼ high education, KOR ¼ Republic of Korea, low ed ¼ low education, SIN ¼ Singapore.Notes: From left to right, by country:1. Proportion of heart disease among elderly with low education from 2015 to 2050.2. Proportion of heart disease among elderly with high education from 2015 to 2050.3. Absolute difference in proportion of heart disease between elderly in high- and low-education groups from2015 to 2050. Lines represent the mean difference in heart disease prevalence between high- and low-educationfor the Republic of Korea and Singapore. Confidence bounds represent the 95% prediction interval from MonteCarlo uncertainty of 1,000 simulation.Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.



Figure 13. Disparities in Hypertension Prevalence by Country, Education, and Gender

(b) Males

(c) Females

(a) Overall

0.00

0.25

0.50

0.75

1.00

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

SINKOR SINKOR

i ) P(Hypertension|low ed)

0.00

0.25

0.50

0.75

1.00

ii) P(Hypertension |high ed)

0.0

0.1

0.2

0.3

Dif

fere

nce

in p

reva

lenc

e

KOR SIN

Prev

alen

ce o

f h

yper

tens

ion

Prev

alen

ce o

f h

yper

tens

ion

iii) P(Hypertension|low ed)−P(Hypertension|high ed)

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

SINKOR SINKOR

Dif

fere

nce

in p

reva

lenc

e

KOR SIN

i ) P(Hypertension|low ed) ii) P(Hypertension |high ed) iii) P(Hypertension|low ed)−P(Hypertension|high ed)

0.00

0.25

0.50

0.75

1.00

0.00

0.25

0.50

0.75

1.00

Prev

alen

ce o

f h

yper

tens

ion

Prev

alen

ce o

f h

yper

tens

ion

0.0

0.1

0.2

0.3

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

2015

2020

2025

2030

2035

2040

2045

2050

SINKOR SINKOR

Dif

fere

nce

in p

reva

lenc

e

KOR SIN

i ) P(Hypertension|low ed) iii) P(Hypertension|low ed)−P(Hypertension|high ed)

Pre

vale

nce

of h

yper

tens

ion

ii) P(Hypertension |high ed)

Pre

vale

nce

of h

yper

tens

ion

0.00

0.25

0.50

0.75

1.00

0.00

0.25

0.50

0.75

1.00

0.0

0.1

0.2

0.3

high ed ¼ high education, KOR ¼ Republic of Korea, low ed ¼ low education, SIN ¼ Singapore.Notes: From left to right, by country:1. Proportion of hypertension among elderly with low education from 2015 to 2050.2. Proportion of hypertension among elderly with high education from 2015 to 2050.3. Absolute difference in proportion of hypertension between elderly in high- and low-education groups from2015 to 2050. Lines represent the mean difference in hypertension prevalence between high- and low-educationfor the Republic of Korea and Singapore. Confidence bounds represent the 95% prediction interval from MonteCarlo uncertainty of 1,000 simulation.Source: Authors’ calculations using the Future Elderly Model based on data from the Korean Longitudinal Studyof Aging and the Singapore Chinese Health Study.



Table

1.Preva

lence

ofCon

ditionsin

2050

forthoseAge

65yearsan

dAbov

ebyCou

ntry,

Education

,an

dGender

Singa

pore

Republic

ofKorea

Difference

inPreva

lence

(Low

versus

HighEducation

)

Con

dition

Low

Education

HighEducation

Low

Education

HighEducation

Singa

pore

Republic

ofKorea

Overall

ADLdisability

0.31

90.08

70.51

70.10

00.23

30.41

7IA

DLdisability

0.57

00.20

10.71

80.211

0.36

90.50

7Diabetes

0.37

10.20

40.31

00.29

90.16

70.02

2Strok

e0.13

10.06

60.21

80.15

20.06

50.06

6Heartdisease

0.24

70.12

10.22

30.19

90.12

60.02

7Hyp

ertension

0.84

00.77

00.80

90.59

50.07

00.21

4

Fem

ale

ADLdisability

0.37

30.09

40.52

10.07

50.27

80.44

6IA

DLdisability

0.55

80.15

80.71

20.14

20.40

00.56

9Diabetes

0.34

00.17

90.31

70.29

70.16

20.02

9Strok

e0.10

80.05

80.21

00.12

20.05

00.08

8Heartdisease

0.19

90.06

50.22

90.20

90.13

40.02

7Hyp

ertension

0.83

20.70

90.82

80.60

30.12

30.22

5

Male

ADLdisability

0.22

30.07

90.50

30.12

20.14

40.38

0IA

DLdisability

0.59

30.24

30.74

00.27

00.35

00.47

0Diabetes

0.42

70.22

90.28

50.30

20.19

8�0

.044

Strok

e0.17

20.07

40.25

00.17

90.09

80.07

3Heartdisease

0.33

40.17

60.20

00.19

10.15

90.03

8Hyp

ertension

0.85

50.82

90.73

60.58

80.02

60.14

8

ADL¼

activ

ities

ofdaily

living,

IADL¼

instrumentalactiv

ities

ofdaily

living.

Sou

rce:

Autho

rs’calculations

usingtheFutureElderly

Mod

elbasedon

data

from

theKoreanLon

gitudinalStudy

ofAging

andtheSingapo

reChinese

Health

Study.



Interestingly, while hypertension prevalence is similar in the Republic of Korea

and Singapore among elderly with low education, elderly with high education in

Singapore have a higher prevalence of hypertension compared to the Republic of

Korea (Figure 13a). The increase in the educational gradient for hypertension tapers

off in both countries between 2015 and 2050, driven in the Republic of Korea more by

decreasing growth in male hypertension prevalence, while in Singapore it is driven by

both genders (Figures 13b and 13c).

Table 1 summarizes the projected prevalence rate of all conditions in 2050

(Figures 7–12). The respective results were described in the above paragraphs.

Diabetes is the only chronic condition for which there is a projected decline in the

educational gradient such that among men in the Republic of Korea in 2050 the

prevalence of diabetes is higher among those with a college education than among

men with less than a high school education.

Sensitivity analysis was performed comparing the elderly aged 55–64, 65–74,

and 75–84 years, where low education was recorded as having no college education

(high school or less) and high education was having at least a college education.

Overall, we found decreasing trends in disparities for both the Republic of Korea and

Singapore (Appendix 7). In general, the elderly with a college education had a lower

prevalence of diseases compared to the elderly without a college education, except for

diabetes (aged 65–74 and 75–84 years) and stroke in the Republic of Korea. While

disparities in diabetes increase in Singapore, this was driven by the sharper rise in

prevalence among elderly without a college education.

IV. Discussion

These findings show associations between education and health for two rapidly

aging countries in East Asia. Overall, despite increases in educational attainment, we

project an increasing prevalence of functional disability as well as chronic conditions

from 2015 to 2050, due entirely to survival to older ages and thus population aging

within the over-50 population in each country. The elderly with high educational

attainment are projected to have a lower prevalence of functional disability and chronic

diseases compared to the elderly with low educational attainment across time.

This was found to be consistent across gender in both the Republic of Korea and

Singapore. After stratification by age group (55–64, 65–74, and 75–84 years old), we

found disparities in functional disability between low- and high-education groups in

each age group. However, these disparities decrease across time in both countries.



Individuals with high educational attainment also experience an increasing burden of

functional disability and chronic diseases. This is consistent with studies that found

diminishing marginal health returns to education (Harris 2007). While those with high

educational attainment are more efficient producers of health and may have greater

ability to pay for life-protection activities, their lower age-specific mortality and hence

survival to older ages leads to greater functional disability and chronic disease burden.

Those with relatively lower education also benefited from individual and social

investments that lowered age-specific morbidity less than age-specific mortality,

relative to their better-educated counterparts. If we remove these survival

improvements by holding age structure constant through age standardization, we

find that functional disability remains stable across time.

By comparing disease projections by education and across population groups, we

can better understand how health disparities change dynamically. In what follows, we

discuss in more detail the underlying factors that may be responsible for the dynamic

pattern of changes in survival and disability projected by our simulation, including

differential self-protective behaviors and early-life exposures shaping life expectancy

and age-specific health across different groups, as well as the economic and social

implications of our results.

Our simulation model builds on a rich literature that shows that mortality and

morbidity are partially endogenous choices of human capital rather than biological

endowments. Public health measures, prevention, and primary care have supported a

rapid demographic transition in both Singapore and the Republic of Korea,

representing what a complementary literature characterize as investments in life

protection at the individual, community, and national levels (Ehrlich 2000, Ehrlich and

Yin 2005). These activities include control of infectious diseases, better nutrition, and

enhanced access to the technological frontier in treating acute conditions, as well as

screening for and treating chronic disease, through organized financing and universal

health coverage in both countries. Thus, the primary finding from our projections is

that in both countries, across men and women of different educational attainment,

improved survival will lead to population aging within the over-65 population, a

pattern intensified within successive cohorts and compressed over a shorter period than

in many other parts of the world.

The significant increase in the proportion of younger cohorts with a college

education itself represents substantial individual, household, and social investments in

self-protective behaviors in both countries (through public schools, private tutoring,

and subsidized access to other educational programs). These populations clearly

invested in human capital at earlier stages of the life course, which contributed to the



observed lower age-specific mortality rates, as demonstrated by successive cohorts

rapidly increasing the proportion with higher educational attainment (see Figures 3–5).

This significant increase in educational attainment in both countries, especially among

women, is important to take into account when considering the health needs of the

future elderly. As shown in Figure 3, the percentage of women with a college

education increased steadily across birth cohorts; in the Republic of Korea, women

with less than a high school education will represent only 3% of women aged 65 years

and above by 2050. In Singapore, gender parity in education predates the compulsory

education policy in 2000 (Attorney-General’s Chambers Compulsory Education Act

2000). There has been a huge transition into higher education among females.

By 2007, parity was achieved at all levels of education, and women in college now

outnumber men by a 10% margin (Pan 2013, Thong 2017). This trend is not isolated to

Asia but also observed in other high-income economies (Fiske 2012). While women

have seen a faster increase in college education, they have yet to catch up with men on

some other aspects of socioeconomic status (SES) such as earnings, which might

contribute to a delay in seeking medical or long-term care. Women tend to live longer,

incur higher health-care expenditures, and utilize more health-care resources.

Empowering females through education leads to better outcomes in health and

income. As a larger proportion of women reach old age with greater education, they

benefit from longer survival. Moreover, we find that health disparities between the

shrinking group of low-educated women and the growing group of high-educated

women for certain chronic diseases (e.g., hypertension) are projected to decline.

The projected male–female differences in functional disability shown in our

simulations also reflect changing social norms and opportunities for wage-income

streams over the life course—both because of greater formal sector female labor force

participation conditional on educational attainment and because of greater increases in

human capital for females (starting from a lower base of schooling among the oldest

female cohorts). Increased educational attainment and the associated growth in the

age-earning profiles of women relative to those of men may allow future elderly

women to afford the increased amount of health care necessary for them in old age,

perhaps contributing to a further narrowing of disparities (Owens 2008; Vaidya,

Partha, and Karmakar 2012; Doumas et al. 2013).

Our results do not permit an interpretation of causality as our FEM model uses a

reduced form equation to estimate the association between educational attainment and

mortality. However, studies have shown that education may enhance knowledge on

living a healthy life, leading to improved choices that affect health and shape mortality

risks over the life span through self-protective behavior and life-protection investments



over the life cycle (Kenkel 1991, Ehrlich 2000, Ehrlich and Yin 2005). Age-specific

mortality risk is lowered through self-protective inputs such as medical care services,

diet, and exercise, collectively termed as life-protection (Ehrlich and Yin 2005).

Higher-educated adults also possess higher wealth and have a higher incentive to

protect this wealth by insuring more against mortality risk through life insurance and

annuities (Ehrlich and Yin 2005). Thus, they may benefit more from these cumulative

or persistent effects of life-protection outlay earlier in their lives, where the existence

of insurance increases the likelihood of larger health endowments (Ehrlich 2000,

Ehrlich and Yin 2005). In addition, when focusing on the middle-aged and elderly as

we do, one could argue that it is more likely that education influences choices of use of

time and goods that affect health (Kenkel 1991). Our simulations suggest a generally

increasing educational disparity, with ADL and IADL disability showing stark

differences in prevalence between the high- and low-education groups in both

Singapore and the Republic of Korea—larger than disparities in chronic diseases such

as hypertension and diabetes.

These results are consistent with the current body of the literature on educational

disparities on health outcomes in both Eastern and Western countries. Higher

educational attainment of both males and females in the Republic of Korea led to

comparatively lower odds of having ADLs or IADLs (Kye et al. 2014), and a

Singapore-based projection showed that individuals with higher education have lower

risk of functional disability status over their life course (Ansah et al. 2015). There are

numerous reasons why those with higher education may enjoy better health status

(Cubbin and Winkleby 2005, Chandola et al. 2008, Berkman 2009). The elderly with

higher socioeconomic status is more efficient producers of health due to more

informed health choices and better network effects, which lead to better health

outcomes (Cutler and Lleras-Muney 2010). Higher educational attainment early in life

is usually correlated with healthier lifestyle choices in adulthood such as less smoking,

more physical activity, better weight management, and greater accumulation of healthy

years of life over the life span (Clarke et al. 2009; Basu, Jones, and Dias 2018).

The elderly with low educational attainment may not benefit as much from advances in

medicine due to poorer adherence to treatment (Goldman and Lakdawalla 2001), in

addition to barriers from the financial and time investments required (Goldman and

Lakdawalla 2001). The positive association between SES inequality and poorer health

outcomes supports the general notion that inequality affects health negatively

(Kawachi, Kennedy, and Wilkinson 1999). Disparities in life expectancy between the

well-educated and those with less education are also well documented (Olshansky

et al. 2012; Goldring, Lange, and Richards-Shubik 2016; Sasson 2016).



Subgroup analysis done on the 65–74 and 74–85 year-old age groups showed that

disparities in chronic conditions, ADL disability, and IADL disability by education

exist but decrease over time (Appendix 7). Our result is consistent with the existing

literature showing that compositional changes in education can have considerable

impact on decreasing future disability prevalence (Lutz et al. 2007). We find that

increasing educational attainment among the populations of the Republic of Korea and

Singapore will lead to decreasing age-standardized disability rates. However, due to

the rapidly aging population in both countries, the overall crude disability rates will

increase because of a higher proportion of the oldest-old and their associated disability.

While elderly with a college education had fewer diseases compared to those without a

college education, the disparities are projected to decline over time as survival

continues to improve among the least advantaged and disparities converge among the

oldest-old. Nevertheless, disparities persist, in part because several beneficial assets

accumulate with education: more effective coping skills, better access to preventative

services, better use of resources, and the network of family and friends where

adaptation of positive health behavior is being reinforced (Winkleby, Fortmann, and

Barrett 1990). Thus, improving human capital through the improvement of the

educational composition may offer a way to avoid some of the possible negative

consequences associated with rapid aging (Lutz et al. 2007).

In addition to the evolving educational gradient in disability, our simulations

capture the likely future differential health impacts of social developments over the

past half century in both countries. Younger cohorts have also been exposed over

a larger share of their life span to social safety nets and expanding social welfare

systems—such as the Republic of Korea’s National Health Insurance since 1989 and

long-term care insurance since 2008—that provide greater insurance value to those of

low SES, helping to smooth consumption and lower the cost of self-protective

behaviors such as preventive care. This growth of the social protection system may

help to explain the dynamic pattern of improved survival to older ages for both low

and high SES individuals, as well as the decreasing educational gradient in some

chronic diseases.

Since a substantial share of life-protection investments occur at earlier ages than

we model, our estimates must take as given the young-adult choices and investments

that determined heterogeneity of age-specific mortality and morbidity by sex and

educational attainment of our modeled populations. These earlier investments could

explain a large share of the disparities we document and project. For example,

modeling the United States, Ehrlich and Yin (2005) find that the impact of life

protection on life expectancy at age 30 accounted for a higher percentage of remaining



life expectancy, and a greater share of the educational disparities in life expectancy,

than at age 65.

Our results also indicate an increase in functional disability among the future

elderly in both countries, across genders and education groups, primarily because of an

increase in the number of oldest-old. Depending on whether current cohorts’

self-protection activities decrease morbidity risks at the same rate as mortality risk, the

future elderly may experience compression of morbidity at older ages (i.e., reductions

in age-specific morbidity) that may reduce projected disability at older ages.

Such activities should be encouraged since increased prevalence of functional

disability has significant economic implications. First, functional disability has direct

medical costs, impedes extension of labor force participation and work productivity,

and may negatively impact standards of living (Fried et al. 2001, Loyalka et al. 2014).

Indirect costs from caregiver burden can be significant. Managing disability could

mean substantial increases in primary and secondary medical care and long-term care

expenditure for both individuals and health-care systems. In the low-education group

for both Singapore and the Republic of Korea, we projected a drastic increase in

functional disability compared to the high-education group. Compared to their better

educated peers, individuals in the low-education group are less likely to benefit from

human capital accumulation through employment (Redding 1996). They are also more

likely to experience functional disability that induces early retirement or suspension of

work with limited likelihood of reentering the workforce permanently (Dahl, Nilsen

IV, and Vaage 2000).

With rapidly aging populations, both the Republic of Korea and Singapore face a

higher proportion of their respective populations at a period in life when life-protection

efforts are reaching diminishing returns (given higher biological risk of mortality at

older ages), so spending in self-protection may decrease. Yet, as noted, the future

elderly of both countries will also be much wealthier with more human capital than

earlier cohorts, especially among women; this latter pattern pushes in the opposite

direction toward higher investments in health and life protection. Similarly, although

higher health endowments of the current young cohorts might seem to suggest that we

would expect lower life-protection efforts in the future, models such as Ehrlich and

Yin (2005) point out that longer life expectancies also generate a wealth effect that

increases the value of life-protection investments. Thus, our estimates might be

conservative to the extent that such investments generate a further compression of

morbidity not included in our projections. Policies that encourage prevention and

healthy lifestyles, especially supporting the vulnerable, could help both countries

narrow the gap by education and sex in healthy life expectancy.



A. Limitations

Modeling in this complex environment necessarily reflects only the best

available information and is subject to certain limitations. Our study uses data from

surveys that only included community-dwelling individuals and did not include those

in nursing homes or other institutionalized patients, thus underestimating the disability

and chronic disease burden in both the Republic of Korea and Singapore.

Nevertheless, our results are still relevant for policy planning to cater to the growing

needs of the elderly with disabilities living in the community. The number of nursing

home beds per 1,000 population aged 65 years and over is 24 for the Republic of

Korea and 26 for Singapore, or 2.4% and 2.6% of the over-65 population, respectively.

Although our forecast of dependency is an underestimate, the risk of institutionalization

is low in the Republic of Korea and Singapore. Due to small numbers, we were also

unable to model the severity of disability after adjusting for individual heterogeneity in

Singapore. While similar ADL tasks were assessed in both the Republic of Korea and

Singapore, Singapore had an additional question on IADL tasks. Specifically, the

question was whether the respondent needs help using transportation. As such,

Singapore has one more IADL task compared to the Republic of Korea and would have a

slightly overestimated prevalence for IADL disability. Nevertheless, among older adults

who needed help with any IADL tasks, only 2% of older adults needed help with

transportation only. The rest of the older adults required help in IADL tasks as captured in

surveys in both the Republic of Korea and Singapore.

From 2010 to 2015, Singapore had a racial composition of around 74% Chinese,

14% Malay, 9% Indian, and 3% other races (Department of Statistics 2012). As a

result, our chronic disease projections may be an underestimate as minority groups

have a greater chronic disease burden (Venketasubramaniam et al. 2005, Williams and

Mohammed 2009, Lee et al. 2009, Sharma et al. 2012). Transition models were

estimated using panel data, which suffer from issues of nonresponse and attrition.

Recall bias may be possible as health status is reported in 2-year and 6-year surveys

for the Republic of Korea and Singapore, respectively. We were also unable to model

shorter disease dynamics in Singapore as our transitional probabilities were estimated

based on the survey with a 6-year median follow-up. However, as we are comparing

the disparity in socioeconomic standing between the low- and high-educational

attainment groups, the difference in years between survey waves may not bias our

results. Singapore has a quota for college education (Toh 2017), which makes the

proportion of individuals in that group perhaps more selective compared to the

Republic of Korea, although a nontrivial fraction of individuals in both countries can



also obtain education abroad. In addition, the Republic of Korea’s improvement in

education has been so dramatic that the country will no longer have elderly with less

than a high school education in the replenishing cohort by 2030 (see Figure 4);

therefore, the higher disparities we found in the Republic of Korea are largely driven

by the fact that those with low education in the Republic of Korea are older on average

than those with low education in Singapore. We also assumed that the protective effect

of college education on functional disability is maintained over time. However, as it

was harder for post-war generation elderly to attain a college degree compared to the

more recent replenishing cohorts, we might have overestimated the disparities since

we assumed the same protective effect of college education across time.

In addition, older adults born after industrialization spend their childhood in a

more affluent society than previous generations. As such, an increase in health

problems in the future could be overestimated. Nevertheless, these birth cohorts are

also less involved in physically laborious work and tend to have diets with a higher

composition of fat and sodium. The changes in work, lifestyle, and diet may lead to

higher incidence of health problems in the future, as shown for example by the

increased prevalence and declining educational gradient in diabetes, especially evident

for the Republic of Korea. Lastly, Singapore’s incoming cohort is slightly older,

starting at aged 55 years and above, whereas the Republic of Korea’s cohort is aged 51

years and above. Nevertheless, as most diseases tend to occur later in life, we do not

expect our disparity projections to be significantly biased by that difference. We are

also unable to tease apart the difference between educational attainment versus the

selection of high-ability people due to entrance exams to enter high school, thus the

effect of improved educational attainment on old-age health could be overestimated.

It would be valuable to extend these microsimulation results with more detailed data to

estimate the dynamic evolution of age-specific hazards of mortality, morbidity, and

functional disability by sex and educational attainment.

B. Conclusion

Our work allows comparisons of the challenges rapidly aging countries in Asia

face by providing a common platform (FEM) to derive projections. Although similar

microsimulations have been conducted in projecting elderly health and functional

disability for Japan (Fukawa 2007), the Republic of Korea (Kye et al. 2014), and

Singapore (Ansah et al. 2015), it is unclear whether valid cross-country comparisons

can be made given the underlying differences in model estimation, assumptions, and

projection timelines. Previous research comparing elderly health in Asia tended to



report trends without modeling the micro-dynamics (Lee and Shinkai 2003, Ofstedal

et al. 2007). Thus, our comparison of future developments in elderly functional

disability and diseases, along with the associated educational gradient, contributes to

the literature on each country and to the literature comparing aging internationally,

especially because of the marked differences in health-care and social support systems

across countries.2

Our microsimulation analyses indicate that from 2015 to 2050, the educational

gradient in health and functional status will steepen, with a wider gap in the prevalence

of disability between older adults with low and high educational attainment.

For example, in Singapore, the disparity in functional disability prevalence—as

measured by limitation in at least one ADL—starts at 9.0 percentage points in 2015

and is projected to increase to 23.3 percentage points by 2050. The Republic of

Korea’s model projects even larger disparities by 2050, despite having more years of

schooling among the low-education group (i.e., high school versus primary school),

with a 41.7-percentage-point difference in the prevalence of ADL disability between

low- and high-educational attainment groups (51.7% compared to 10%).

These disparities also hold across most other chronic diseases such as heart disease.

For the Republic of Korea, the low-educational attainment group by 2050 represents

only 1% of men and 3% of women, and these most-disadvantaged elderly are entirely

concentrated in the oldest-old group (Figure 3). Their rates of disability are higher than

in a population aged 65 years and above that includes “young-old” in their 60s and 70s

as well as those in their 80s and 90s such as Singapore. Age-standardized disability

rates are constant or declining in both countries (Figure 7).

Overall, the model projects an increasing prevalence of functional disability as

well as chronic conditions from 2015 to 2050. Elders with high educational attainment

are projected to have a lower prevalence of functional disability and chronic conditions

compared to elderly with low educational attainment. Even with increases in

educational attainment, reduced mortality and improved survival lead to greater

disease prevalence among the elderly across all educational groups due to a higher

proportion of over-80 within the over-65 population. While both the Republic of

Korea and Singapore have universal health coverage, and the Republic of Korea has

long-term care insurance, there might be differences in the out-of-pocket burden for

medical care and gaps in the provision of and access to long-term care services that

exacerbate disparities. Health insurance enhances access to care and reduces the risk of

2For example, Singapore employs a compulsory savings plan linked to an individual retirementaccount without social risk pooling, while schemes in East Asia are characterized by employer-based pay-as-you-go plans.



catastrophic medical expenditures, lowering the cost of health investments through

self-protection. Nevertheless, medical care in adulthood may be imperfectly able to

compensate for the exposures to less healthy environments in childhood that cumulate

through differential education and other human capital. Moreover, the benefits of

consumption smoothing from universal health-care coverage may be offset somewhat

by the moral hazard of low out-of-pocket costs for care, leading to fewer preventive

activities (ex ante moral hazard) and overuse of curative services (ex post moral

hazard), as evident in our projections by the increase in some chronic diseases among

the high-SES group.

To the best of our knowledge, our study is the first to compare the progression of

educational disparities in disability across two rapidly aging Asian societies,

accounting for their complex interrelationship with sociodemographic and health

behaviors as covariates that evolve across time. By studying these evolving patterns of

morbidity and disability, our study complements the literature on life-protection

activities that endogenizes mortality while abstracting from effects on morbidity

(Ehrlich and Yin 2005), further underscoring the importance of investment in healthy

aging and control of chronic disease so that added years of life can be relatively

healthy ones. We projected a widening disparity in health outcomes across education

and between genders for both Singapore and the Republic of Korea. Lastly, we

delineated possible mechanisms explaining the education–gender disparity in health

outcomes and suggested possible policy actions to narrow those disparities in

super-aging societies.

References

Ansah, John P., Rahul Malhotra, Nicola Lew, Chi-Tsun Chiu, Angelique Chan, Steffen Bayer,and David B. Matchar. 2015. “Projection of Young-Old and Old-Old with FunctionalDisability: Does Accounting for the Changing Educational Composition of the ElderlyPopulation Make a Difference?” PloS One 10 (5): e0126471.

Attorney-General’s Chambers, Government of Singapore. 2003. Compulsory Education Act2000. Singapore Statutes Online. https://sso.agc.gov.sg/Act/CEA2000.

Baker, David P., Juan Leon, Emily G. Smith Greenaway, John Collins, and Marcela Movit.2011. “The Education Effect on Population Health: A Reassessment.” Population andDevelopment Review 37 (2): 307–32.

Barro, Robert. 2013. “Education and Economic Growth.” Annals of Economics and Finance14 (2A): 277–304.



Basu, Anirban, Andrew M. Jones, and Pedro Rosa Dias. 2018. “Heterogeneity in the Impact ofType of Schooling on Adult Health and Lifestyle.” Journal of Health Economics57 (2018): 1–14.

Behrman, Jere R., and Barbara L. Wolfe. 1987. “How Does Mother’s Schooling Affect FamilyHealth, Nutrition, Medical Care Usage, and Household Sanitation?” Journal ofEconometrics 36 (1–2): 185–204.

Berkman, Lisa F. 2009. “Social Epidemiology: Social Determinants of Health in the UnitedStates: Are We Losing Ground?” Annual Review of Public Health 30: 27.

Bustamante, Arturo Vargas, Hai Fang, Jeremiah Garza, Olivia Carter-Pokras, Steven P. Wallace,John A. Rizzo, and Alexander N. Ortega. 2012. “Variations in Healthcare Access andUtilization among Mexican Immigrants: The Role of Documentation Status.” Journal ofImmigrant and Minority Health 14 (1): 146–55.

Caldwell, Brian J. 1996. “School Reform for the Knowledge Society: An EconomicPerspective.” Australian Economic Review 29 (4): 416–22.

Chandola, Tarani, Annie Britton, Eric Brunner, Harry Hemingway, Marek Malik, MeenaKumari, Ellena Badrick, Mika Kivimaki, and Michael Marmot. 2008. “Work Stress andCoronary Heart Disease: What Are the Mechanisms?” European Heart Journal 29 (5):640–648.

Chen, Brian K., Sze-Chuan Suen, Lena Schoemaker, Hawre Jalal, Michael Hurley, HidekiHashimoto, Karen Eggleston, and Jay Bhattacharya. 2016. “Forecasting Trends inDisability in a Super-Aging Society: Adapting the Future Elderly Model to Japan.”National Bureau of Economic Research Working Paper Series No. 21870.

Chen, Cynthia, Francisco Cervero Liceras, Stefan Flasche, Sucitro Sidharta, Joanne Yoong,Neisha Sundaram, and Mark Jit. 2019a. “Effect and Cost-Effectiveness of PneumococcalConjugate Vaccination: A Global Modelling Analysis.” The Lancet Global Health 7 (1):e58–e67.

Chen, Cynthia, Jue Tao Lim, Ngee Choon Chia, Lijia Wang, Bryan Tysinger, JulieZissimopoulos, and Ming Zhe Chong. 2019b. “The Long-Term Impact of FunctionalDisability on Hospitalization Spending in Singapore.” The Journal of the Economics ofAgeing 14: 100193.

Chen, Qiulin, Karen Eggleston, Wei Zhang, Jiaying Zhao, and Sen Zhou. 2017. “TheEducational Gradient in Health in China.” The China Quarterly (London) 230: 289–322.

Clarke, Philippa, Patrick M. O’Malley, Lloyd D. Johnston, and John E. Schulenberg. 2009.“Social Disparities in BMI Trajectories Across Adulthood by Gender, Race/Ethnicity andLifetime Socio-Economic Position: 1986–2004.” International Journal of Epidemiology38 (2): 499–509.

Conti, Gabriella, James J. Heckman, and Sergio Urzua. 2010. “The Education-HealthGradient.” The American Economic Review 100 (2): 234–38.

Cubbin, Catherine, and Marilyn A. Winkleby. 2005. “Protective and Harmful Effects ofNeighborhood-Level Deprivation on Individual-Level Health Knowledge, BehaviorChanges, and Risk of Coronary Heart Disease.” American Journal of Epidemiology162 (6): 559–68.

Currie, Janet, and Rosemary Hyson. 1999. “Is the Impact of Health Shocks Cushioned bySocioeconomic Status? The Case of Low Birthweight.” The American Economic Review89 (2): 245–50.



Cutler, David M., and Adriana Lleras-Muney. 2010. “Understanding Differences in HealthBehaviors by Education.” Journal of Health Economics 29 (1): 1–28.

Dahl, Svenn-Age, Oivind Anti Nilsen, and Kjell Vaage. 2000. “Work Or Retirement? ExitRoutes for Norwegian Elderly.” Applied Economics 32 (14): 1865–76.

Deaton, Angus, and Christina Paxson. 1999. “Mortality, Education, Income, and Inequalityamong American Cohorts.” National Bureau of Economic Research Working PaperSeries No. 7140.

Department of Statistics, Government of Singapore. 2012. Population Trends 2012. http://costofhealthyliving.com/wp-content/uploads/2013/11/population2012.pdf.

Doumas, Michael, Vasilios Papademetriou, Charles Faselis, and Peter Kokkinos. 2013. “GenderDifferences in Hypertension: Myths and Reality.” Current Hypertension Reports 15 (4):321.

Ehrlich, Isaac. 2000. “Uncertain Lifetime, Life Protection, and the Value of Life Saving.”Journal of Health Economics 19 (3): 341–67.

Ehrlich, Isaac, and Yong Yin. 2005. “Explaining Diversities in Age-Specific Life Expectanciesand Values of Life Saving: A Numerical Analysis.” Journal of Risk and Uncertainty31 (2): 129–162.

Fiske, E. B. 2012. World Atlas of Gender Equality in Education. Paris: UNESCO.Fried, T. R., E. H. Bradley, C. S. Williams, and M. E. Tinetti. 2001. “Functional Disability and

Health Care Expenditures for Older Persons.” Archives of Internal Medicine (1960)161 (21): 2602.

Fuchs, Victor R. 2004. “Reflections on the Socio-Economic Correlates of Health.” Journal ofHealth Economics 23 (4): 653–61.

Fukawa, Tetsuo. 2007. “Health and Long-term Care Expenditures of the Elderly in JapanUsing a Micro-simulation Model.” The Japanese Journal of Social Security Policy 6 (2):199–206.

Gjonça, E., F. Tabassum, and E. Breeze. 2009. “Socioeconomic Differences in PhysicalDisability at Older Age.” Journal of Epidemiology and Community Health (1979)63 (11): 928–74.

Goldman, Dana P., Baoping Shang, Jayanta Bhattacharya, Alan M. Garber, Michael Hurd,Geoffrey F. Joyce, Darius N. Lakdawalla, Constantijn Panis, and Paul G. Shekelle. 2005.“Consequences of Health Trends and Medical Innovation for the Future Elderly.” HealthAffairs (Project Hope) 24 (Suppl. 2): W515–17.

Goldman, Dana, and Darius N. Lakdawalla. 2001. “Understanding Health Disparities AcrossEducation Groups.” National Bureau of Economic Research Working Paper SeriesNo. 8328.

Goldring, Thomas, Fabian Lange, and Seth Richards-Shubik. 2016. “Testing for Changes in theSES-Mortality Gradient when the Distribution of Education Changes Too.” Journal ofHealth Economics 46 (2016): 120–30.

Harris, Douglas N. 2007. “Diminishing Marginal Returns and the Production of Education: AnInternational Analysis.” Education Economics 15 (1): 31–53.

Jang, Soong-Nang. 2016. “Korean Longitudinal Study of Ageing: Overview of ResearchDesign and Contents.” In Encyclopedia of Geropsychology, edited by Nancy Pachana.Singapore: Springer.



Kawachi, Ichiro, Kennedy, Bruce P., and Wilkinson, Richard G. 1999. Income Inequality andHealth (Vol. 1). New York: New Press.

Kenkel, Don. 1991. “Health Behavior, Health Knowledge, and Education.” Journal of PoliticalEconomy 99 (2): 287–305.

Kennedy, Brian K., Shelley L. Berger, Anne Brunet, Judith Campisi, Ana Maria Cuervo, ElissaS. Epel, and Claudio Franceschi. 2014. “Geroscience: Linking Aging to ChronicDisease.” Cell 159 (4): 709–13.

Kye, Bongoh, Erika Arenas, Graciela Teruel, and Luis Rubalcava. 2014. “Education, ElderlyHealth, and Differential Population Aging in South Korea: A Demographic Approach.”Demographic Research 30: 753–94.

Lee, Raymond, Siew-Pang Chan, Yiong-Huak Chan, Jennifer Wong, Diana Lau, and KennethNg. 2009. “Impact of Race on Morbidity and Mortality in Patients with Congestive HeartFailure: A Study of the Multiracial Population in Singapore.” International Journal ofCardiology 134 (3): 422–25.

Lee, Ronald, and Andrew Mason. 2010. “Fertility, Human Capital, and Economic Growth Overthe Demographic Transition.” European Journal of Population 26 (2): 159–82.

Lee, Yunhwan, and Shoji Shinkai. 2003. “A Comparison of Correlates of Self-Rated Health andFunctional Disability of Older Persons in the Far East: Japan and Korea.” Archives ofGerontology and Geriatrics 37 (1): 63–76.

Leopold, Liliya, and Henriette Engelhartdt. 2012. “Education and Physical Health Trajectoriesin Old Age. Evidence from the Survey of Health, Ageing and Retirement in Europe(SHARE).” International Journal of Public Health 58 (1): 23–31.

Loyalka, Prashant, Lan Liu, Gong Chen, and Xiaoying Zheng. 2014. “The Cost of Disability inChina.” Demography 51 (1): 97–118.

Lutz, W., K. C. Samir, H. T. A. Khan, S. Scherbov, and G. W. Leeson. 2007. Future Ageing inSoutheast Asia: Demographic Trends, Human Capital and Health Status IR-07-026.https://go.exlibris.link/W6l6gJKw.

McDonough, P., G. J. Duncan, D. Williams, and J. House. 1997. “Income Dynamics and AdultMortality in the United States, 1972 through 1989.” American Journal of Public Health(1971) 87 (9): 1476–83.

Michaud, Pierre-Carl, Dana Goldman, Darius Lakdawalla, Adam Gailey, and Yuhui Zheng.2011. “Differences in Health between Americans and Western Europeans: Effects onLongevity and Public Finance.” Social Science & Medicine (1982) 73 (2): 254–63.

Minicuci, N., and M. Noale. 2005. “Influence of Level of Education on Disability Free LifeExpectancy by Sex: The ILSA Study.” Experimental Gerontology 40 (12): 997–1003.

Ministry of Education, Government of the Republic of Korea. 2000. Korea Statistical YearbookEducation. http://kess.kedi.re.kr/eng.

Ministry of Health, Government of Singapore. 2018. About Eldershield. Singapore.Moi, Kho Ee. 2010. Economic Pragmatism and the “Schooling” of Girls in Singapore.

Singapore.National Research Council. 2012. Aging and the Macroeconomy: Long-Term Implications of an

Older Population. Washington, DC: The National Academies Press. https://doi.org/10.17226/13465.

National University of Singapore. 2018. Introduction to the Singapore Chinese Health Study.Singapore.



Nogueira, Dália, and Elizabeth Reis. 2014. “Longevity and Health: Trends in FunctionalLimitations and Disability among Older Adults in Portugal, 1985–2005.” AgeingInternational 39 (3): 195–209.

O’Brien, Katherine L., Lara J. Wolfson, James P. Watt, Emily Henkle, Maria Deloria-Knoll,Natalie McCall, and Ellen Lee. 2009. “Burden of Disease Caused by StreptococcusPneumoniae in Children Younger than 5 Years: Global Estimates.” The Lancet (BritishEdition) 374 (9693): 893–902.

Ofstedal, Mary Beth, Zachary Zimmer, Albert I. Hermalin, Angelique Chan, Yi-Li Chuang,Josefina Natividad, and Zhe Tang. 2007. “Short-Term Trends in Functional Limitationand Disability among Older Asians: A Comparison of Five Asian Settings.” Journal ofCross-Cultural Gerontology 22 (3): 243–61.

Olshansky, S. Jay, Toni Antonucci, Lisa Berkman, Robert H. Binstock, Axel Boersch-Supan,John T. Cacioppo, and Bruce A. Carnes. 2012. “Differences in Life Expectancy Due toRace and Educational Differences are Widening, and Many may Not Catch Up.” HealthAffairs Web Exclusive 31 (8): 1803–13.

Owens, Gary M. 2008. “Gender Differences in Health Care Expenditures, Resource Utilization,and Quality of Care.” Journal of Managed Care Pharmacy 14 (3 Suppl.): 2.

Pan, Jessica. 2013. “The Rise and Rise of Women in Higher Education.” The Straits Times,March 7. https://fass.nus.edu.sg/srn/2013/03/07/the-rise-and-rise-of-women-in-higher-education/.

Pericchi, Luis Raúl, and Peter Walley. 1991. “Robust Bayesian Credible Intervals and PriorIgnorance.” International Statistical Review 59 (1): 1–23.

Redding, Stephen. 1996. “The Low-Skill, Low-Quality Trap: Strategic Complementaritiesbetween Human Capital and R & D.” The Economic Journal (London) 106 (435):458–70.

Samir, KC, and Harold Lentzner. 2010. “The Effect of Education on Adult Mortality andDisability: A Global Perspective.” Vienna Yearbook of Population Research 8: 201–35.

Sasson, Isaac. 2016. “Trends in Life Expectancy and Lifespan Variation by EducationalAttainment: United States, 1990–2010.” Demography 53 (2): 269–93.

Schnittker, Jason. 2004. “Education and the Changing Shape of the Income Gradient inHealth.” Journal of Health and Social Behavior 45 (3): 286–305.

Sharma, Vijay K., Georgios Tsivgoulis, Hock Luen Teoh, Benjamin K. C. Ong, and Bernard P.L. Chan. 2012. “Stroke Risk Factors and Outcomes among various Asian Ethnic Groupsin Singapore.” Journal of Stroke and Cerebrovascular Diseases 21 (4): 299–304.

Sudore, Rebecca L., Kala M. Mehta, Eleanor M. Simonsick, Tamara B. Harris, Anne B.Newman, Suzanne Satterfield, Caterina Rosano, Ronica N. Rooks, Susan M. Rubin, HilsaN. Ayonayon, and Kristine Yaffe. 2006. “Limited Literacy in Older People andDisparities in Health and Healthcare Access.” Journal of the American Geriatrics Society54 (5): 770.

Thong, Yong Jun. 2017. “Women’s Quiet Achievement on the Education Front.” Hidden DataStories. https://blog.data.gov.sg/hidden-data-stories-womens-quiet-achievement-on-the-education-front-2a541c442430?gi=2a60d364b2d8.

Toh, Ee Ming. 2017. “Skills, Not Degrees, at a Premium Now.” Today. 8 May. https://www.todayonline.com/singapore/what-matters-ability-not-university-degree-ong-ke-yung.



United Nations Department of Economic and Social Affairs. 2017. World PopulationProspects: The 2017 Revision. https://www.un.org/development/desa/publications/world-population-prospects-the-2017-revision.html.

United Nations Development Programme. 2018. Human Development Indices and Indicators2018 Statistical Update. New York.

University of Southern California. 2019. The Future Elderly Model. https://healthpolicy.usc.edu/future-elderly-model.

Vaidya, Varun, Gautam Partha, and Monita Karmakar. 2012. “Gender Differences in Utilizationof Preventive Care Services in the United States.” Journal of Women’s Health(Larchmont, N.Y. 2002) 21 (2): 140.

Venketasubramaniam, Narayanaswamy, Louis C. S. Tan, Suresh Sahadevan, Jing J. Chin,Ennapadam S. Krishnamoorthy, Ching Y. Hong, and Seang M. Saw. 2005. “Prevalence ofStroke among Chinese, Malay, and Indian Singaporeans: A Community-Based Tri-RacialCross-Sectional Survey.” Stroke (1970) 36 (3): 551–56.

Williams, David R., and Selina A. Mohammed. 2009. “Discrimination and Racial Disparities inHealth: Evidence and Needed Research.” Journal of Behavioral Medicine 32 (1): 20.

Winkleby, Marilyn A., Stephen P. Fortmann, and Donald C. Barrett. 1990. “Social ClassDisparities in Risk Factors for Disease: Eight-Year Prevalence Patterns by Level ofEducation.” Preventive Medicine 19 (1): 1.

World Bank. 2017. Open Data. https://data.worldbank.org/ (accessed 2 August 2019).



Cognitive Functioning among OlderAdults in Japan and Other Selected AsianCountries: In Search of a Better Way

to Remeasure Population Aging

NAOHIRO OGAWA, TAIYO FUKAI, NORMA MANSOR,AND NURUL DIYANA KAMARULZAMAN

¤

⁄Naohiro Ogawa (corresponding author): Asian Development Bank Institute (ADBI), Japan; TheUniversity of Tokyo, Japan; and University of Malaya, Malaysia. E-mail: [email protected],[email protected]; Taiyo Fukai: Graduate School of Economics, The University of Tokyo, Japan. E-mail:[email protected], [email protected]; Norma Mansor: Social Wellbeing Research Centre,Faculty of Economics & Administration, University of Malaya, Malaysia. E-mail: [email protected];Nurul Diyana Kamarulzaman: Social Wellbeing Research Centre, Faculty of Economics & Administration,University of Malaya, Malaysia. E-mail: [email protected]. We thank the Managing Editor and theanonymous referees for helpful comments and suggestions. We are also grateful to the Research Institute ofEconomy, Trade and Industry (RIETI), Hitotsubashi University, Japan, and The University of Tokyo,Japan, for allowing us to use the Japanese Study of Aging and Retirement (JSTAR) dataset. Our thanks arealso to the Center for Aging Society Research/Research Center of the National Institute of DevelopmentAdministration, Thailand, for permitting us to access the Health, Aging, and Retirement in Thailand(HART) dataset. Moreover, our analysis for the People’s Republic of China has used data and informationfrom the Harmonized CHARLS dataset and Codebook, Version C as of April 2018, developed by theGateway to Global Aging Data, USA, and funded by the National Institute on Aging, USA (R01AG030153, RC2 AG036619, and R03 AG043052). For more information, readers can refer to https://g2aging.org. Furthermore, our analysis for India has used data and information from the LongitudinalAgeing Study in India (LASI) Pilot microdata and documentation, which was funded by the NationalInstitute on Aging/National Institutes of Health, USA (R21AG032572, R03AG043052, and R01AG030153). The LASI Project is funded by the Ministry of Health and Family Welfare, Government ofIndia, the National Institute on Aging (R01 AG042778 and R01 AG030153), and the United NationsPopulation Fund, India. We have also used the LASI data that were produced by the International Institutefor Population Sciences, Mumbai, Harvard T.H. Chan School of Public Health, Boston, and University ofSouthern California, Los Angeles, and distributed by the University of Southern California with fundingfrom the Ministry of Health and Family Welfare, Government of India, the National Institute on Aging(R01 AG042778 and R01 AG030153), and the United Nations Population Fund, India. The AsianDevelopment Bank recognizes “China” as the People’s Republic of China, “South Korea” as the Republicof Korea, and “United States of America” as the United States.





https://dx.doi.org/10.1142/S0116110522500068

https://dx.doi.org/10.1142/S0116110522500068

Japan is the oldest society in the world. It has the highest proportion of thepopulation aged 65 and over, a demographic indicator that has been used bydemographers for more than a century. One of the main objectives of this study is toapply a new indicator—the cognition-adjusted dependency ratio (CADR)—toremeasure the level of population aging from an innovative point of view. To computethis new index, we apply the mean age-group-specific immediate recall scores forJapan and four other Asian countries, and we compare the results with those derivedfrom the United States and various developed nations in Europe. Our analysis showsthat Japan’s pattern and level of age-related decline in cognitive functioning are highlycomparable to those of many other developed nations, particularly in ContinentalEurope.Among the otherAsian countries,Malaysia shows a pattern of change similarto countries in Southern Europe, although Malaysia has slightly lower scores thanSouthern Europe in all age groups. More importantly, these comparative resultsbased on CADR are astonishingly different from the corresponding resultsobtained from conventional old-age dependency ratios. The Japanese case is themost salient example.

Keywords: cognition-adjusted dependency ratio, cognitive functioning,immediate word recall, population aging

JEL codes: J11, J14

I. Introduction

Since the second half of the 1960s, the tempo of world population growth has

been gradually slowing down due to substantial fertility declines in various countries,

both developed and developing. Population aging has become a worldwide

phenomenon, attracting growing attention from researchers and policy makers

particularly for its escalating economic and social costs (Sanderson and Scherbov

2010). The field of demography has increasingly recognized that while the 20th

century was the century of “population explosion,” the 21st century is becoming the

century of “population aging” (Hermalin 2003; Lutz, Sanderson, and Scherbov 2004;

United Nations 2007; Clark, Ogawa, and Mason 2007; Fu and Hughes 2009;

Uhlenberg 2009; Arifin and Ananta 2009; Tuljapurkar, Ogawa, and Gauthier 2010;

Eggleston and Tuljapurkar 2010; Lee and Mason 2011; Park, Lee, and Mason 2012;

Kendig, McDonald, and Piggott 2016).

At present, almost 60% of the world population inhabits Asia, making it the most

populous region in the world. Also, the proportion of Asia’s population aged 65 and

over in the world’s elderly population has been continuously rising since the end of

World War II. In 1950 it was 44%, but it reached 57% by 2020 and is now projected to

grow to 62% in 2050 (United Nations 2019).



In parallel with such rapid growth of its older population, Asia has also witnessed

dramatic changes in its demographic landscape, particularly its population’s age

composition. Asia’s total dependency ratio, which is expressed as the ratio of the

number of dependents to the working-age population {[(0–14 years old) þ (65 years

old and over)]/(15–64 years old)}, reached its peak value (0.81) in 1966, after which

its projected long-term trend showed a U-shaped pattern reaching its trough value

(0.47) in 2015. In addition, there have been substantial inter-country differences in the

trends and levels of population aging within Asia in the past several decades (Mason

2001, Lee and Mason 2011, Ogawa et al. 2021).1

To compare the burden of population aging across countries, we frequently use

conventional demographic indicators such as the age dependency ratio, which is

defined as the ratio of the number of elderly people to the working-age population [(65

years old and over)/(15–64 years old)], and the index of aging, expressed as the ratio

of the number of elderly persons to the young population [(65 years old and over)/(0–14

years old)]. Based on these commonly used demographic indicators, we characterize and

rank how old countries are. Although these demographic indicators are readily available

to researchers, one of their most serious limitations is that they are exclusively based on

chronological age distributions. Because of this, they fail to provide a powerful base for

deriving persuasive conclusions on the consequences of and possible responses to

population aging. To cope with this major drawback, Skirbekk, Loichinger, and Weber

(2012) recommend a new approach in which age variation in cognitive abilities among

older persons is incorporated into a revised version of the conventional total dependency

ratio, with a view to comparing the extent of aging across countries from an innovative

standpoint. It is important to note that this new approach has become feasible primarily

thanks to an increasing number of surveys collecting individual data on cognitive

abilities among older persons in numerous countries, both developed and developing,

particularly since the 1990s.

Among them, the Health and Retirement Study (HRS), a longitudinal survey of a

representative sample of United States (US) citizens over the age of 50, is the most

well-known and has served as a public resource for data on aging since 1990. The

HRS has a number of sister studies in many countries all over the world. In recent

1Apparently, such inter-country differences in the age-composition transformation have contributedto generating marked differences in the magnitude and timing of the “first demographic dividends,” whichhave, in turn, facilitated a remarkable economic growth in various Asian countries. The economic“miracle” of East Asian economies between 1960 and 1997 is a salient example (Bloom and Williamson1998).

COGNITIVE FUNCTIONING AMONG OLDER ADULTS 93


years, such studies have been implemented in five Asian countries: Japan (the

Japanese Study of Aging and Retirement or JSTAR), the People’s Republic of China

(PRC; the China Health and Retirement Longitudinal Study or CHARLS), India (the

Longitudinal Ageing Study in India or LASI), Thailand (the Health, Aging, and

Retirement in Thailand or HART), and Malaysia (the Malaysia Ageing and Retirement

Survey or MARS).2

By drawing heavily on microlevel datasets gathered from these surveys, we

compute age-specific cognitive abilities among adults aged 50 and over in each of

these Asian countries, and then compare the differences in their cognitive

performance. We also compare them with their counterparts in selected Western

countries. In the second half of the paper, we examine, by applying microlevel data

from the Asian surveys to the regression model, how and to what extent the cognitive

abilities of older adults in each country are related to a host of demographic,

socioeconomic, and biopsychological factors. Based on the computed results, we

discuss both similarities and dissimilarities of the relationships between cognitive

functioning and its covariates such as demographic, socioeconomic, and medical

factors in the five Asian countries. Subsequently, we briefly discuss the likely future

trends in older adults’ cognitive abilities in these countries.

The paper is organized as follows. Section II discusses cognition measures and

matters related to them to facilitate later on in the paper a variety of analyses on

inter-country variations in cognitive functioning of older workers in the five selected

Asian countries, the US, and a number of industrialized nations in Europe. To provide

a solid base for conducting such analyses, Section III reviews several important earlier

studies, which have examined numerous key factors linking the relationships between

cognitive functioning and a host of demographic and socioeconomic factors.

Section IV describes the data from the five Asian surveys mentioned earlier, which

will be used in Section V to compute the mean age-group-specific immediate recall

scores in the Asian countries. These scores will be compared to those for Europe and

the US, as derived from an earlier study by Skirbekk, Loichinger, and Weber (2012).

In Section VI, we relate the computed mean age-group-specific immediate recall score

to population aging using the cognition-adjusted dependency ratio (CADR).

In Section VII, we attempt to identify the factors associated with immediate recall

2Besides these five countries, the HRS was extended to the Republic of Korea (the KoreanLongitudinal Study of Aging or KLoSA). Unfortunately, the immediate recall data gathered in KLoSAwere not measured in a way comparable to those employed in this study. For this reason, we have excludedKLoSA from our study.



scores among the adults aged 50–79. Section VIII summarizes the main findings with

a few policy implications.

II. Measuring Cognitive Functioning

Over the past few decades, rapid population aging worldwide has compelled

several countries in Europe, Asia, and elsewhere to gradually postpone the mandatory

retirement age to maintain financial solvency and sustainability of their public pension

schemes (Clark and Ogawa 1992, Clark et al. 2008). At the same time, because

cognition affects the capacity to acquire and use information, improving cognitive

functioning at older ages has been adopted as a top public health priority in many

countries to enable individuals to make good decisions and, ultimately, to remain

independent and care for themselves longer (Maurer 2010). However, cognitive

functioning of older workers can vary widely across countries, which can create large

differences between them in the severity of various problems arising from aging

(Skirbekk, Loichinger, and Weber 2012).

Due to the importance of cognition and cognitive variation among older adults,

the HRS and its sister studies have made cognitive measurement a priority (Ofstedal,

Fisher, and Herzog 2005; Weir, Lay, and Langa 2014). In general, the following

activities are regarded as cognitive processes: thinking, knowing, remembering,

judging, and problem-solving. Both fluid intelligence and crystallized intelligence are

used in these cognitive activities (van Aken et al. 2016).

Fluid intelligence is the ability to use logic and solve problems in new or novel

situations without resorting to pre-existing knowledge. Fluid intelligence plays a role

in the creative process, and we often use it to handle nonverbal tasks such as

mathematical problems and puzzles. On the other hand, crystallized intelligence is the

ability to make use of information or knowledge previously acquired through

education and experience. We usually employ crystallized intelligence when we

encounter verbal tasks, such as reading comprehension or grammar. Crystallized

intelligence is generally retained or even improved over time. By contrast, because

fluid intelligence is rooted in physiological functioning, it typically peaks in young

adulthood (approximately at an age of 25) and then steadily declines. Although fluid

and crystallized intelligence represent two distinctly different sets of abilities, they

often work jointly. For instance, when taking a test in mathematics, we use

mathematical formulas and notations such as (þ) and (�), which come from



crystallized intelligence as pieces of pre-existing knowledge, but also utilize fluid

intelligence to develop strategies and derive solutions to accomplish the task.

In the HRS family of studies, different cognitive tests have been used to measure

specific mental capacities. For instance, the capacities frequently tapped in previous

research investigations are episodic memory, numeracy, orientation, attention, working

memory, and verbal fluency.

One of the key domains of measuring cognition in an aging population is

short-term memory, which is frequently evaluated using word recall tasks (Weir, Lay,

and Langa 2014).3 The ability to recall words read from a randomly selected list of a

certain number of given words generally declines with age.4 This ability is usually

measured in two ways: (i) immediate word recall and (ii) delayed word recall.

In immediate word recall, the respondent reads a list of 10 words and, after a very brief

interval, recalls as many words as possible, not necessarily in order, within one minute.

In delayed word recall, the respondent is asked to recall as many words as possible

approximately five minutes after the immediate word recall task out of the same list the

respondent had read for the immediate recall task.

The number of words and the length of time allowed for recalling the words can

vary from survey to survey. For example, in the HRS, for both immediate and delayed

recall tasks, a respondent reads one out of four possible lists of 10 words and then has

two minutes to recall the words. Despite being given two minutes, a majority of HRS

respondents do not make use of the second minute: 90% of the respondents used less

than 49.2 seconds in the immediate recall task, and less than 50.4 seconds in the

delayed recall task. Based on these results, Skirbekk, Loichinger, and Weber (2012)

assert that recalling the words within one or two minutes does not affect the validity of

inter-country comparative results in cognitive abilities.

In addition to short-term memory (e.g., immediate word recall and delayed word

recall), a person’s working memory is also used in the literature to measure variation in

cognitive functioning. A common procedure for assessing working memory is a

task called serial-7s, which is the repeated subtraction of sevens starting from 100.

This activity involves numeric ability as well as the ability to attend to a task, thus

falling into the category of fluid intelligence.

3Word recall measures are designed to capture the ability to remember and use relevant informationwhile in the middle of an activity such as information processing. For this reason, word recall measures fallinto the category of fluid intelligence.

4The ability to recall words is sensitive to brain changes that often occur at an early stage ofAlzheimer’s disease.



Apart from these measures of fluid intelligence, indicators of crystallized

intelligence such as verbal fluency are also used to measure cognitive functioning.

A well-known task in verbal fluency is naming animals—respondents list as many

names of animals as they can in one minute.5

In view of the availability and comparability of HRS-type survey data among the

five Asian countries, we confine our analysis of the cognitive functioning of older

adults to immediate word recall scores.6 Another reason for this restriction is that an

overwhelming majority of empirical studies conducted outside Asia thus far have been

based on such scores, which means that we can compare our results with these studies

(e.g., Weber et al. 2014; Bonsang, Skirbekk, and Staudinger 2017).

III. Earlier Studies Pertaining to Cognitive Functioning among Older Adults

The 20th century saw considerable growth in cognitive functioning in many

countries. The factors that induced such cognitive improvements include greater

exposure to cognitive stimulation through better education, improved living

conditions, steady improvements in health, and declining average family size

triggered by lower fertility and changing marriage values (Sundet, Borren, and

Tambs 2008; Lynn 2009).

The study conducted by Skirbekk, Loichinger, and Weber (2012) examined

inter-country age variation in cognitive functioning by measuring the immediate recall

scores. The authors computed the mean age-group-specific immediate recall score

using data from the HRS, the World Health Organization Study on global AGEing and

adult health (SAGE), and the Survey of Health, Ageing and Retirement in Europe

(SHARE). Caution should be exercised, however, in interpreting their results. For each

5-year age group, the mean value of the immediate recall score for older persons in a

certain age group was computed from each relevant survey, but there are some

5Orientation is another measure of crystallized intelligence. Orientation is measured using a set oftests involving simple questions about the date and day of the week. The HRS contains additional itemsconcerning the names of US presidents and vice-presidents.

6Immediate word recall has been shown to be important for a variety of outcomes, ranging fromfinancial decision-making to the risk of developing dementia (Fein, McGillivray, and Finn 2007; Skirbekk,Loichinger, and Weber 2012). Moreover, technological advances and changes in working proceduresimply that the importance of the ability to learn and remember is increasing (Machin and Van Reenen1998). Employers are particularly interested in whether their employees are able to learn new workprocedures and process new information (Munnell, Sass, and Soto 2006), which also suggests thatemployers view the ability to immediately recall information as advantageous to labor market performance.



differences in the way the respondents were tested by the interviewers. That is,

respondents in as many as 18 countries were given one minute for recalling, but

respondents in the US were allowed two minutes.7 Furthermore, the interviewers read

out the 10 words to be recalled only once in all surveys except for SAGE, where

interviewers read out the words three times before the respondents recalled the words.

Despite these differences in the way the data on immediate word recall were collected,

the computed results showed a statistically significant age-related decline in all the

countries within the 50–84 age group.

In the face of rapid societal improvements over time, particularly during the 20th

century, cognitive gender differences continue to be a source of scientific and political

debate, and the magnitude, pattern, and explanation of these differences remain

important research topics. By using data from SHARE, Weber et al. (2014)

investigated gender differences in cognitive performance in the middle-aged and

older populations across 13 European countries. They found that the magnitude of the

differences varied systematically across cognitive tasks, birth cohorts, and

geographical regions. In addition, both the living conditions and educational

opportunities the individuals were exposed to during their formative years were

related to increased gender differences favoring women in episodic memory

(immediate word recall scores), decreased gender differences in the case of numeracy

(arithmetic computation), and the elimination of differences in verbal fluency (animal

naming). It is also interesting to note that their analysis of immediate word recall

scores shows that although women in Northern Europe perform at a higher level than

men across all birth cohorts, the pattern is different in Central and Southern Europe. In

Central Europe, the female advantage is found only for birth cohorts born in 1932 or

later, but not in earlier cohorts. In Southern Europe, there is even less of a female

advantage, which gradually switches to a male advantage in earlier cohorts.

Weir, Lay, and Langa (2014) also examined gender inequality in cognition, by

analyzing data from the PRC’s CHARLS and India’s LASI pilot survey, as well as

from SAGE, using individual data derived from cognitive tests such as immediate

word recall, orientation, serial-7s, and listing the names of animals. In both countries

and in virtually all the cognitive tasks, men performed considerably better than

women. In addition, the study found that despite some notable differences in survey

samples and measures, a strong general association of cognition in older ages with

education emerges as a potential explanation for the gender gaps and cohort

7In the HRS 92–94 version, 20 words were given to the respondents.



differences. They also found that the female disadvantage in cognition is greater in

India than in the PRC, both before and after controlling for education.

It is generally considered that being married is associated with a healthier

lifestyle and greater daily social interaction (Fuller 2010). These behaviors may

improve cognitive reserve and reduce dementia (Kuiper et al. 2015). In this context, as

briefly mentioned in footnote 4, the incidence of Alzheimer’s disease (one of the

subtypes of dementia) is closely connected with the ability to recall words.8 For this

reason, it is highly conceivable that being married is positively related to the ability to

perform short-term memory tasks such as immediate word recall. More importantly, a

recent study carried out by Sommerlad et al. (2018), which is based on a systematic

review and meta-analysis of 15 studies on the association between marital status and

the risk of developing dementia, shows that being married is associated with a

significantly smaller risk of dementia compared to lifelong single people.9 Hence,

changing one’s marital status may affect cognitive abilities throughout one’s life.10

It is well known that old age tends to be related to a host of health risks such as

cardiac infarction and cerebral hemorrhage (Slomski 2014). Furthermore, it is

increasingly recognized that cognitive functioning tends to be a good predictor of

future morbidity and mortality (Negash et al. 2011). Therefore, individuals with higher

cognitive abilities are more likely to be healthier and live longer than those with low

cognitive abilities. Cognitive abilities predict individual productivity better than any

other observable individual characteristic, and they are increasingly relevant for labor

market performance (Skirbekk, Loichinger, and Weber 2012). Moreover, this finding

is applicable to many countries, both developed and developing, and in different

settings, both urban and rural (Behrman, Ross, and Sabot 2008).

Over the past few decades, the number of seniors have been increasing in labor

markets at an accelerating pace. Because certain cognitive abilities decline

substantially at late adult ages, most studies previously conducted on older workers

have focused on those aged 50 and over (Anderson and Craik 2000). A substantial

fraction of these seniors can remain in the labor market for a long time, but how long

they stay depends on how long they can retain high cognitive performance.

8Among dementia subtypes, Alzheimer’s disease occupies the largest share in most countries in theworld. In Japan, for example, approximately 70% of persons with dementia fall under the category ofAlzheimer’s disease.

9The following three Asian economies are included in this meta-analysis: Japan; Taipei,China; andthe Republic of Korea.

10There is no significant difference in the risk of dementia among those currently married, divorced,or widowed.



Staying in the labor market could even improve cognitive performance.

Using six waves of the HRS (1998–2008), Bonsang, Adam, and Perelman (2012)

demonstrated that retirement negatively influences cognitive functioning for older

Americans. This finding suggests that reforms aimed at promoting labor force

participation at an older age may not only ensure the sustainability of social security

systems but also create positive health externalities for older individuals. However, a

study by de Grip et al. (2015) based on a Dutch survey dataset have found the opposite

result. Using data from the Maastricht Aging Study (MAAS), they examined the

relation between retirement and cognitive development in the Netherlands and showed

that retirees experienced lower decline in cognitive flexibility than those who remained

employed.11

Primarily due to the growing availability of representative surveys on the

cognitive functioning of elderly persons in different countries and regions, an

increasing number of empirical studies on the determinants of cognitive performance

among the elderly have been carried out in recent years. In addition, almost all of these

surveys have used highly comparable questionnaires, thus making inter-country

comparisons feasible. One salient example is the study carried out by Maharani and

Tampubolon (2016). Using data from the 2006 round of the English Longitudinal

Study of Ageing (ELSA) and the 2007 round of the Indonesian Family Life Survey

Wave 4, the authors examined the associations between central obesity, as measured

by waist circumference, and the cognition level in adults aged 50 and over in England

and Indonesia. Conducting regression analysis, after controlling for some selected

demographic, socioeconomic, and biomedical variables, they found that centrally

obese respondents had lower cognition levels than non-centrally obese respondents in

England, while the opposite was true for Indonesia.

Similarly, using data gathered in rural Central Java, LaFave and Thomas (2017)

examined the relationship between the respondents’ height and cognitive ability.

By and large, taller workers earned more. In lower income settings, an adult’s height is

normally a marker of strength, which is rewarded in the labor market. Adult height is

also a proxy for cognitive performance or other dimensions of human capital such as

school quality, a proxy for health status, and a proxy for family background and

genetic characteristics. Taking these observations into account, the authors conducted

a regression analysis and showed that the respondents’ cognitive abilities were

significantly related to their height.

11The authors also found that the decline in information processing speed after retirementparticularly holds for those who are less educated.



By drawing on data derived from SHARE, Doblhammer, van den Berg, and

Fritze (2013) examined cognitive functioning at the age of 60 and over. In their study,

a total of 17,070 persons in 10 SHARE member countries were included in the

analysis of several domains of cognitive functioning, which were linked to

macroeconomic conditions during their birth year.12 One of the main findings of

this study was that economic conditions at birth significantly influenced cognitive

functioning in late life in various domains. Another finding was that economic

recessions adversely affected numeracy, verbal fluency, and recall abilities, as well as

scores on omnibus cognitive indicators.

Furthermore, Bordone, Scherbov, and Steiber (2015) investigated if and why

individuals aged 50 and over who were born into more recent cohorts performed better

in terms of cognition than their counterparts of the same age born into earlier cohorts, a

phenomenon called the “Flynn effect.” They used data from two waves of ELSA and

the German Socio-Economic Panel (SOEP) surveys and showed that cognitive test

scores of participants aged 50 and over in the later wave were higher than those of

participants aged 50 and over in the earlier wave. In addition to identifying the Flynn

effect based on the two cross-sectional waves, they pointed out that the reason why

they used two waves was because a repeat cross-sectional design overcomes potential

bias of retest effects. They also showed that although compositional changes of the

older population in terms of education partly explain the Flynn effect, the increasing

use of modern technology (i.e., computers and mobile phones) in the first decade of

the 2000s also accounts for it.

IV. Description of Data Sources Used

In the rest of the paper, we aim to shed light on the age-specific pattern of

cognitive abilities among older adults in Japan and four selected Asian countries, and

then offer a statistical analysis of the demographic, biomedical, and socioeconomic

factors associated with cognitive functioning in each country. To facilitate these

quantitative analyses, we employ the following survey datasets: JSTAR for Japan,

CHARLS for the PRC, LASI pilot survey for India, HART for Thailand, and MARS

for Malaysia.

12The 10 countries included in the study are Austria, Belgium, Denmark, France, Germany, Italy,the Netherlands, Spain, Sweden, and Switzerland.



A. Japanese Study of Aging and Retirement

JSTAR is a longitudinal, interdisciplinary survey that collects internationally

comparable data on middle-aged and older adults. The JSTAR project commenced in

2007 and the survey has been implemented in 2-year intervals. Because JSTAR is a

sister survey compatible with the HRS, a considerable proportion of the content

included in the JSTAR questionnaire is comparable to the content in the other four

Asian surveys, which were also modeled after the HRS. JSTAR’s design and sample

methodology are described elsewhere (Ichimura, Hashimoto, and Shimizutani 2009).

The baseline sample consists of male and female respondents aged 50–75 from

10 Japanese municipalities.13 The respondents were randomly chosen from household

registries in each of the 10 cities, towns, or villages. The sample size and the average

response rate at the baseline were approximately 8,000 and 60%, respectively. JSTAR

collects a wide range of variables, including economic, social, family, and health

conditions of the sampled respondents. As for cognition-related variables, JSTAR

gathers data on cognitive tasks such as short-term memory (both immediate and

delayed word recall) and serial-7s.

Caution should be exercised in interpreting our results because we use data only

from the first round of JSTAR from the following three groups: the five municipalities

surveyed in 2007 (Takikawa, Sendai, Adachi, Kanazawa, and Shirakawa), the two

municipalities added in 2009 (Naha and Tosu), and the three that joined the survey in

2011 (Chofu, Tondabayashi, and Hiroshima). This data treatment is chosen for the

purpose of avoiding problems that arise from nonrandom dropout and retest-practice

effects associated with cognitive tests in longitudinal surveys, including JSTAR

(Thorvaldsson et al. 2006; Skirbekk, Bordone, and Weber 2014). As is the case with

most internationally comparable surveys such as SHARE, the JSTAR respondents

listened to 10 words read out by the interviewers and were given one minute each to

recall them, both in the immediate and delayed word recall tasks.

B. China Health and Retirement Longitudinal Study

CHARLS is a nationally representative longitudinal survey of persons 45 years

of age or older and their spouses, and includes assessments of the social, economic,

and health circumstances of community residents in the PRC. CHARLS examines

13These 10 municipalities joined the JSTAR project at different points in time: Takikawa, Sendai,Adachi, Kanazawa, and Shirakawa joined in 2007; Naha and Tosu in 2009; and Tondabayashi, Hiroshima,and Chofu in 2011.



health and economic adjustments to the rapidly aging population of the PRC.

The national baseline sample size is 10,287 households and 17,708 individuals,

covering 150 counties in 28 provinces. The first national baseline wave was fielded

from June 2011 to March 2012, followed by wave 2 in 2013 and wave 3 in 2015.

Core CHARLS questionnaires include numerous sections dedicated to demographics,

family structure and changes, health status and functioning, general health now and

before the age of 16, physician-diagnosed chronic illnesses, lifestyle and health-related

behaviors (smoking, drinking, and physical activities), subjective expectation of

mortality, activities of daily living (ADLs), instrumental activities of daily living

(IADLs), helpers, cognition testing (short-term memory task: two minutes to recall 10

words), depression (Center for Epidemiological Studies Depression Scale or CES-D),

health care and insurance, work, retirement and pension, income and consumption,

and assets (individual and household).

The interviewers conduct and carry equipment for measurements of health

functioning and performance in respondents’ households. These include the

anthropometric measurements of height, weight, waist circumference, lower right

leg length and arm length, lung capacity, grip strength, speed in repeated chair stand

test, blood pressure, walking speed, and balance tests.

C. Longitudinal Ageing Study in India

In 2010, a LASI pilot survey was undertaken in four Indian states (Karnataka,

Kerala, Punjab, and Rajasthan) on a targeted sample of 1,600 individuals aged 45 and

older and their spouses. To capture regional variation, two northern states (Punjab and

Rajasthan) and two southern states (Karnataka and Kerala) were included in the

survey. Punjab is an example of an economically developed state, while Rajasthan is

relatively poor, with very low female literacy, high fertility, and persisting gender

disparities. Kerala, which is known for its relatively efficient health-care system, has

undergone rapid social development and is included as a potential harbinger of how

other Indian states might evolve.

The survey questionnaire consisted of sections such as the household roster,

housing environment, household consumption, individual income of all household

members, household real estate, household financial and non-financial assets, and

household debts. In addition, the survey gathered various information concerning

family and social network, social activities, psychosocial measures, life satisfaction,

health conditions, and health-care utilization. In the section on mental health, the

following cognitive task scores were collected: time orientation, short-term memory



(two minutes to recall 10 words), verbal fluency (animal naming), numeric ability

(counting backwards from 20), and computation (serial-7s).

D. Health, Aging, and Retirement in Thailand

The primary objective of the HART project is to create a national longitudinal

and household panel dataset on aging in Thailand.14 HART is a biannual household

panel survey designed to provide panel data on the multidisciplinary dimensions of

aging in older Thai adults, including (i) demographic characteristics, (ii) family and

transfers, (iii) health and cognition, (iv) employment and retirement, (v) income, (vi)

assets and debts, and (vii) life expectations and life satisfaction. Five thousand and six

hundred households from five regions and Bangkok and its vicinity were sampled to

represent national households. More concretely, 13 provinces were selected for

forming a household panel in the baseline survey. In each household, one member

aged 45 and over was selected as the respondent.15

The data collected from the national longitudinal survey in 2015 (wave 1) and

2016 (wave 2) are maintained in the data archive at the Intelligence and Information

Center of the National Institute of Development Administration, Bangkok. The

cognitive test consisted of three tasks: (i) word recall (immediate and delayed word

recall tasks: two minutes to recall 10 words), (ii) numeracy (serial-7s), and (iii) data

memory. Because cognitive test scores are available only in wave 2, we draw upon the

individual data gleaned in wave 2 for our statistical analysis on the cognitive

performance of older adults in Thailand.

E. Malaysia Ageing and Retirement Survey

MARS is a longitudinal study launched in 2018 which aims to produce

nationally representative data on topics related to aging. MARS was motivated by the

country’s aging population and the importance of having such data to formulate and

implement relevant policies. The baseline sample consists of households from all

states in Malaysia, which were randomly selected based on Malaysia’s 2010

Population and Housing Census. The Department of Statistics Malaysia (DOSM)

selected the sample using a multistage sampling procedure. For each selected

household, any member aged 40 and above who lived in the household most of the

14HART is harmonized with the HRS and its sister studies, including JSTAR, CHARLS, LASI, andMARS.

15For a more detailed description of HART, see: https://g2aging.org/?section=study&studyid=44.



time would be eligible to be selected as a respondent. Should there be more than one

eligible member, a maximum of three oldest eligible members would be selected.

The sample size of the first wave was 5,613 respondents with a response rate of 84%.

These respondents will be interviewed every 2 years to measure changes in their health

and economic and social circumstances.

MARS collects comprehensive information on various aspects of life and

personal experiences covering six sections: (i) respondent background; (ii) family

information and support; (iii) health and health-care utilization; (iv) work and

employment; (v) income and consumption; and (vi) savings and assets. The cognitive

abilities of respondents are measured in the health section where they are required to

perform several tasks such as word recall (both immediate and delayed: two minutes to

recall 10 words), serial-7s, time orientation, and semantic fluency (animal naming).

The word recall task was included in the questionnaires of all five Asian

countries. We have carefully examined the inter-country comparability of the words

asked in the immediate word recall tasks. Because all the countries developed their

survey questionnaires through a close contact with the US HRS team and its

international network, the words chosen for the immediate word recall test are not only

very basic but also similar. Respondents were assigned a list of words from multiple

word lists. In India’s LASI pilot survey, for example, the interviewer randomly

assigned one of the three lists each consisting of 10 words to a respondent.

Interviewers for the PRC’s CHARLS, Thailand’s HART, and Malaysia’s MARS

randomly assigned one out of four lists each consisting of 10 words to a respondent.16

To overcome language barriers in India, the questionnaire was translated into four

regional languages: Hindi, Kannada, Malayalam, and Punjabi. In Malaysia, the

questionnaire was prepared in the following four languages: English, Malay,

Chinese/Mandarin, and Tamil. It is conceivable that these additional adjustments

incorporated in the questionnaires to overcome language barriers reduce some of the

potential biases likely to emerge in inter-country comparative analyses of the immediate

word recall task.

V. Inter-Country Comparison of Immediate Word Recall Scores

By closely following the computational steps taken by Skirbekk, Loichinger, and

Weber (2012), we compute the mean age-group-specific immediate recall scores for

16Unlike the other Asian countries, Japan’s JSTAR had only one list of 10 words for measuringimmediate word recall. However, the 10 Japanese words selected for this task are very basic andcommonly used not only in Japan but also elsewhere in Asia.



the five Asian countries by drawing on the microlevel data derived from JSTAR,

CHARLS, LASI pilot survey, HART, and MARS. In addition, we quantitatively show

where the cognitive abilities of the five Asian countries stand relative to developed

Western countries, which have been computed by Skirbekk, Loichinger, and Weber

(2012) using survey data from the HRS and SHARE.17 Caution should be exercised in

interpreting the computed results because the time periods covered for the analyzed

countries vary to a certain extent. For the Asian countries, the first wave of JSTAR was

undertaken from 2007 to 2011, while CHARLS data were obtained in 2011–2012.

Furthermore, HART was conducted in 2017, MARS in 2018, and the LASI pilot

survey in 2010.18 In contrast, the HRS data were gathered in 2006–2007, while the

SHARE datasets for various European countries were generated from the 2006–2007

round. Except for HART in 2017 and MARS in 2018, the datasets for JSTAR,

CHARLS, LASI pilot survey, HRS, and SHARE employed in this study are relatively

comparable in terms of the time period covered, circa 2010.

The study by Skirbekk, Loichinger, and Weber (2012) computed the immediate

word recall scores for all the countries reported in this study, drawing on the

microlevel survey data together with their sampling weights. To keep the Asian results

compatible with those derived from the study by Skirbekk, Loichinger, and Weber

(2012), we have attempted to use the RAND harmonized version of each country

survey in Asia. One of the great advantages of using the RAND harmonized versions

is that the sampling weights are computed for each data file.19 At the time of writing

this paper, however, the RAND harmonized versions with sampling weights were

available only for JSTAR, CHARLS, and the LASI pilot survey.

We have encountered another limitation with the harmonized JSTAR. As pointed

out in footnote 13, the 10 survey sites did not join the harmonized JSTAR in the same

year but in three different years. Moreover, to avoid problems arising from the

nonrandom dropout and retest-practice effects associated with cognitive tests in

longitudinal surveys, we have used only the immediate recall scores from the first

round for each JSTAR survey site (cohort 1 residing in the five survey sites in 2007,

cohort 2 living in the two sites in 2009, and cohort 3 in the three sites in 2011). For this

17We are grateful to Skirbekk and his associates for providing us with the data on immediate wordrecall scores used in their study published in the Proceedings of the National Academy of Sciences of theUnited States of America (PNAS) in 2012.

18At the time of writing this paper, we did not have access to the individual data gathered in mainwave 1 and wave 2 of LASI conducted during 2016–2020. We plan to update our findings for India when anew dataset becomes available.

19Since 1989, the RAND Center for the Study of Aging has been producing harmonized versions ofvarious national aging survey data files to facilitate international comparative research on aging.



reason, three different sets of computed sampling weights covering different numbers

of survey sites exist. Technically, no single set of sampling weights can be computed

for the entire sample combining the three different cohorts. Hence, in this study, we do

not use the sampling weights available in the harmonized JSTAR.20 Despite this

limitation, we still use the harmonized JSTAR data file (without sampling weights) for

computation since the JSTAR dataset was carefully cleaned by the country team who,

in collaboration with RAND, conducted a consistency check. For this reason, we

assume that various types of data entry errors and unreasonable outliers have been

expunged before the datasets became available for public use.

With MARS and HART, we have not been able to obtain information on

sampling weights to retain the comparability of the computed results. In the case of

MARS, the Social Wellbeing Research Centre of the University of Malaya is currently

collaborating with the RAND Center to rearrange the survey data file to be in line with

RAND’s harmonized version, and their work is expected to be completed by the end of

2021. Furthermore, in their preliminary computations, they have found that there is

only a very small difference between weighted and unweighted results, which seems to

indicate that the use of sampling weights may not be critically important. Thailand’s

HART has not started to compute its sampling weights.

Thus, in the rest of the paper, we will employ as the base for our computation the

harmonized versions of the LASI pilot survey, CHARLS, and JSTAR. In addition, we

will use the original MARS and HART to analyze Malaysia and Thailand, respectively.

Figure 1 compares the mean age-group-specific immediate recall scores21 for the

five Asian countries, the US, and three European regions.22 Clearly, the mean

age-group-specific immediate recall scores continuously decline with age in virtually

all the countries and regions.23 For the sake of clear exposition, we have plotted the

results in Figure 1 using a solid line for the countries computed from the harmonized

data files, and a dotted line for the remaining countries computed without sampling

weights.

20We have compared the results for the immediate recall scores, calculated with and withoutsampling weights for each of the three cohorts. The computed results show virtually the same for eachcohort, which seems to indicate that our analytical results are fairly comparable whether or not we usesampling weights. The plotted results for the three cohorts are available from the authors upon request.

21The scores are expressed in terms of the number of words recalled, ranging from 0 to 10.22The SHARE data were used for the following three European regions: Northern Europe

(Denmark, Ireland, and Sweden), Continental Europe (Austria, Belgium, Czech Republic, France,Germany, the Netherlands, Poland, and Switzerland), and Southern Europe (Greece, Italy, and Spain). Thethree European groups of economies were set up by Skirbekk, Loichinger, and Weber (2012), who alsoadded England to the Northern Europe group, using data collected by ELSA.

23In the case of the Northern European countries, the immediate recall score increased slightly fromthe 50–54 age group to the 55–59 age group.



Figure 1 reveals a few interesting results. First, the US has the highest score for

the 50–54 age group (6.1 words recalled out of 10), followed by the Northern

European group (six words). Both the US and the Northern European group show

a similar declining pattern with age in almost all age groups except for 50–54 and

80–84. Second, immediate recall age trajectories for Continental European countries

and Japan are fairly comparable.24 Furthermore, though not displayed in Figure 1,

Japan’s trajectory of change in the age-group-specific immediate recall scores is

Figure 1. Mean Age-Group-Specific Immediate Recall Scores in Selected Economies,circa 2010

PRC ¼ People’s Republic of China, USA ¼ United States.Note: “Northern Europe” includes Denmark, England, Ireland, and Sweden; “Continental Europe” comprisesAustria, Belgium, Czech Republic, France, Germany, the Netherlands, Poland, and Switzerland; and “SouthernEurope” covers Greece, Italy, and Spain.Sources: Skirbekk, Vegard, Elke Loichinger, and Daniela Weber. 2012. “Variation in Cognitive Functioning as aRefined Approach to Comparing Aging across Countries.” Proceedings of the National Academy of Sciences ofthe United States of America 109 (3): 770–74; and authors’ calculations based on data from the following:(i) Japanese Study of Aging and Retirement of the Research Institute of Economy, Trade and Industry (RIETI),Hitotsubashi University, Japan, and The University of Tokyo, Japan; (ii) China Health and RetirementLongitudinal Study of the National School of Development of Peking University, Beijing; (iii) the pilot portion ofthe Longitudinal Ageing Study in India of the Harvard T.H. Chan School of Public Health, Boston, theInternational Institute for Population Sciences, Mumbai, and the University of Southern California, Los Angeles;(iv) Malaysia Ageing and Retirement Survey of the Social Wellbeing Research Centre of the University ofMalaya, Kuala Lumpur; and (v) Health, Aging, and Retirement in Thailand of the Center for Aging SocietyResearch/Research Center of the National Institute of Development Administration, Bangkok.

24JSTAR covers subjects aged 50–75. Because the age group 75–79 includes only those JSTARrespondents who are 75 years old, we have excluded this group from the analysis due to its skeweddistribution.



between the Netherlands and France—Japan’s scores are slightly lower than those for

the Netherlands but are consistently higher than those for France by a considerable

margin.

Among the five Asian countries, Japan’s age-group-specific immediate recall

scores are the highest in all groups until the 70–74 age group. Note that India’s score

(4.5 words) for the 75–79 age group is higher than the corresponding value for the

Continental European countries. This result needs to be interpreted with great caution.

The total number of observations in India’s LASI pilot survey used for calculating the

age-group-specific immediate recall scores is 1,007, but the number of observations

for the 75–79 age group is only 65.25 For this reason, the reliability of India’s score for

those aged 75–79 is open to question.26

The age-group-specific immediate recall scores for the Southern European group

show a pattern of change similar to that for Malaysia, although Malaysia has slightly

lower scores than Southern Europe in all age groups. Furthermore, Thailand exhibits a

declining pattern in age-group-specific immediate recall scores and has the lowest

scores among the five Asian countries in the 60–64 age group and older.

Attention should be drawn to the pattern of change in the PRC’s

age-group-specific immediate recall scores. In the age groups 50–54 and 55–59, the

PRC’s scores are marginally lower than those for Thailand, but in the remaining age

groups, the PRC has substantially higher scores than Thailand. Moreover, the PRC

overtakes Malaysia at ages 75–79.

In Figure 1, we have also drawn a horizontal dotted line at score 4 to facilitate an

interesting discussion. Let us briefly turn our attention to Thailand and Continental

Europe. In the case of Thailand, the average score for the 60–64 age group plunges

below four words, which is a result obtained by those aged 80–84 in Continental

Europe. Although the age difference amounts to approximately 20 years, the cognitive

performances of these two groups are at the same level (four words). This suggests a

huge difference in cognitive functioning between Thailand and the countries in

Continental Europe. Such inter-country differences in cognitive abilities are likely to

constitute a crucial and decisive drawback in the future to the transfer of new

digitalized technologies and innovative production methods from advanced countries

with higher cognition levels to the countries with lower cognition. More importantly,

25The number of observations for India’s age group 80–84 is only 32.26The cohorts that are presently 50 years and older in India have grown up during a period of

widespread poverty and high mortality and, as a result, the population has been positively selected in termsof cognitive performance at a more advanced age (Skirbekk, Loichinger, and Weber 2012). We plan tosubstantiate the validity of this view once we gain access to data from waves 1 and 2 of LASI(2016–2020).



in view of the slow process of cohort replacement, those countries whose seniors

already have higher cognitive levels today are very likely to continue to be at an

advantage for many decades to come. Thus, the legacy of low cognition among the

older populations of today’s developing countries will put them at a disadvantage for a

very long time (Skirbekk, Loichinger, and Weber 2012; Weir, Lay, and Langa 2014).

VI. Introducing Cognitive Performance into the Measurement of

Population Aging

In this section, we relate the computed mean age-group-specific immediate recall

score to the context of population aging. For this purpose, we draw upon a new

indicator that focuses on cognition and demographic change: the cognition-adjusted

dependency ratio, which was proposed by Skirbekk, Loichinger, and Weber (2012).

The formula for CADR is expressed as follows:

CADR ¼ jfx 2 P j (mx < 5) ^ (agex � 50)gjjfx 2 P j (15 � agex < 50)g [ f(mx � 5) ^ (agex � 50)gj ,

where mx represents the memory score of person x, agex represents the age of person x,

while P stands for the population. To compute CADR, we have applied the mean

age-group-specific immediate recall scores for Japan and other countries in Asia, as

well as in the US and Europe, to the relevant age-composition data derived from the

United Nations (UN) population projection prepared in 2019.27 This formula implies

that if a country has a low value of CADR, then it is effectively “younger,” since it has

a lower share of seniors with poor cognitive performance.

The calculated results are displayed in Table 1.28 Although Japan’s CADR value

(0.22) is higher than the corresponding values for the US and Northern Europe

(Denmark, England, Ireland, and Sweden), Japan’s dependency ratio adjusted by

age-specific cognitive scores is fairly comparable to that (0.18) of Continental Europe

(Austria, Belgium, Czech Republic, France, Germany, the Netherlands, Poland, and

27Because CADRs have already been computed based on the data derived from the 2009 UNpopulation projection for the year 2005 for many European countries by Skirbekk, Loichinger, and Weber(2012), we have applied the 2005 age-composition data gleaned from the 2019 UN population projectionto all the Asian countries except Japan to facilitate inter-country comparisons. In the case of Japan, becauseof the unique survey setup of JSTAR (2007–2011) described in Section V, we have applied the 2010 agecomposition.

28As mentioned earlier, data for age groups 75–79 and 80–84 are not available in JSTAR. Tocalculate CADR, however, cognitive scores for these two old-age groups are required. For this purpose, wehave conducted a linear extrapolation based on the data for those aged 50–74, and the linearity has beenconfirmed by comparing the extrapolated values with the observed values for other countries.



Switzerland), and is considerably lower than that (0.32) of Southern Europe (Greece,

Italy, and Spain).

More importantly, these comparative results based on the CADRs are

astonishingly different from those shown in Table 2, which reports the conventional

age-composition indicators such as old-age dependency ratios and age dependency

ratios for various countries, both developed and developing. Among the countries

listed in Table 2, Japan’s population is by far the oldest, but based on the CADRs

listed in Table 1, Japan’s is fairly close to the medium level observed among the

European countries. This finding seems to justify the UN’s recent efforts to raise

awareness regarding the urgent need for remeasuring population aging in both

developed and developing nations with a view to formulating effective policies for

coping with aging.29

Table 1. International Comparison of Cognition-Adjusted Dependency Ratio Scores,circa 2010

Economy CADR Score

United States 0.10Northern Europe (Denmark, England, Ireland, and Sweden) 0.12Continental Europe (Austria, Belgium, Czech Republic, France, Germany,

the Netherlands, Poland, and Switzerland)0.18

Southern Europe (Greece, Italy, and Spain) 0.32

AsiaJapan 0.22India 0.10People’s Republic of China 0.20Thailand 0.21Malaysia 0.12

CADR ¼ Cognition-adjusted dependency ratio.Sources: Skirbekk, Vegard, Elke Loichinger, and Daniela Weber. 2012. “Variation in CognitiveFunctioning as a Refined Approach to Comparing Aging across Countries.” Proceedings of theNational Academy of Sciences of the United States of America 109 (3): 770–74; and authors’calculations based on data from the following: (i) Japanese Study of Aging and Retirement of theResearch Institute of Economy, Trade and Industry, Hitotsubashi University, Japan, and The Universityof Tokyo, Japan; (ii) China Health and Retirement Longitudinal Study of the National School ofDevelopment of Peking University, Beijing; (iii) the pilot portion of the Longitudinal Ageing Study inIndia of the Harvard T.H. Chan School of Public Health, Boston, the International Institute forPopulation Sciences, Mumbai, and the University of Southern California, Los Angeles; (iv) MalaysiaAgeing and Retirement Survey of the Social Wellbeing Research Centre of the University of Malaya,Kuala Lumpur; and (v) Health, Aging, and Retirement in Thailand of the Center for Aging SocietyResearch/Research Center of the National Institute of Development Administration, Bangkok.

29For instance, the United Nations Population Division organized, in collaboration with the Institutefor Applied System Analysis, Laxenburg, an expert group meeting called “Measuring Population Aging:Bridging Research and Policy” in Bangkok on February 25–26, 2019.



Furthermore, in Table 1, Japan’s CADR (0.22) is the highest among the five

Asian countries, followed by Thailand (0.21) and the PRC (0.20). Because the PRC

and Thailand have recently passed the first demographic dividend stage,30 as

illustrated in Figure 2, their aging process will be accelerating in the future, so their

CADR values will also be swiftly rising in the years ahead.

In contrast, both Malaysia and India have a considerably younger age

composition than the other three Asian countries, as reported in Table 2. Moreover,

Table 2. Inter-Country Comparison of Selected Age-Composition Indices

Region or Country Year Total Dependency Ratio Age Dependency Ratio

Northern AmericaUnited States 2010 49.7 19.4

Northern EuropeDenmark 2010 52.9 25.5United Kingdom 2010 51.7 25.1Ireland 2010 46.6 16.1Sweden 2010 53.2 27.9

Continental EuropeAustria 2010 48.2 26.4Belgium 2010 52.0 26.4Czech Republic 2010 42.2 22.0France 2010 54.6 26.1Germany 2010 51.8 31.2The Netherlands 2010 49.2 23.0Poland 2010 40.2 18.9Switzerland 2010 47.0 24.8

Southern EuropeGreece 2010 49.9 27.5Italy 2010 52.7 31.3Spain 2010 46.7 25.2

AsiaIndia 2010 56.3 8.0People’s Republic of China 2010 35.6 11.4Japan 2010 55.9 35.1Malaysia 2010 49.0 7.4Thailand 2010 39.0 12.4

Source: Authors’ calculations based on the data from the United Nations. 2019. World PopulationProspects: The 2019 Revision. New York: United Nations.

30For a more detailed description of the first demographic dividend, see Ogawa et al. (2021, 44–52).



as depicted in Figure 2, Malaysia and India are still enjoying the benefits of the first

demographic dividend. Depending on their future fertility trends, their CADR values

may vary greatly in the future. For instance, assuming the UN’s low-fertility-variant

population projection, Malaysia’s CADR will be higher than Japan’s current level by

the mid-2060s.

Figure 2. Comparison of the Temporal Change in the Magnitude of theFirst Demographic Dividend

PRC = People’s Republic of China.Note: Each year enclosed in parentheses represents the survey year in which the data of per capita labor incomeand consumption for each country were gathered.Source: Adapted from Figure 9 in Ogawa, Naohiro, Norma Mansor, Sang-Hyop Lee, Michael Abrigo, and TahirAris. 2021. “Population Aging and the Three Demographic Dividends in Asia.” Asian Development Review38 (1): 32–67.



VII. Factors Associated with Cognitive Functioning among Older Adults

in the Five Asian Countries

In this section, we attempt to identify the factors associated with immediate recall

scores among the adults aged 50–79 who are included in recent aging surveys in the

five Asian countries by running a linear regression.

Before going any further, caution should be exercised with regard to our

empirical analysis. In our regressions, the dependent variable representing immediate

recall scores and a few explanatory variables such as education and work status have a

problem of causal ordering. However, we are not able to resolve this endogeneity issue

due to the absence of powerful instrumental variables in the datasets available to us.31

Thus, the regression results presented in this section primarily indicate associations

between the dependent variable and the explanatory variables that cannot be

interpreted in terms of causal effects. Nevertheless, our statistical analysis can be used

to see whether the relationships between the immediate recall scores among the

respondents and their individual attributes that have been discovered in various

Western countries can also be confirmed in the context of the five selected Asian

countries.

Let us first look at the computational results derived from the harmonized JSTAR

(without sampling weights) in relative detail. As shown at the bottom of Table 3,

the total number of observations is 4,873. The dependent variable is the number of

words recalled by the respondent immediately after 10 words were read out by the

interviewer. Except for the respondents’ height, all other explanatory variables are

dummy variables, with the dagger notation (†) representing the reference group. In this

31In our regression model, the issue of causal ordering between the dependent variable, cognitiveperformance, and some explanatory variables, such as education and work status, needs to be properlyaddressed. In the past, numerous studies have been undertaken which shed light on the relationshipsbetween cognitive functioning (measured in terms of immediately recalled words) and a host of othervariables (demographic, socioeconomic, cultural, psychosocial, biomedical, etc.). However, most of thesestudies have not addressed the issue of potential endogeneity bias in their estimations, primarily because ofthe unavailability of appropriate instrumental variables. The issue of causal ordering has been solvedsuccessfully only in a very limited number of studies, including a study by Atalay, Barret, and Staneva(2019) and another by Schneeweis, Skirbekk, and Winter-Ebmer (2014). These studies successfullyaddressed the issue of causal ordering by drawing heavily on powerful instrumental variables created basedon the variation caused by major policy reforms. Although we have, in the hope of addressing the issue ofendogeneity, attempted to identify appropriate instrumental variables by going through various datasetsavailable in the five Asian countries, our attempts have met no success at the time of revising our paper.Thus, following many earlier studies on this research topic, we confine ourselves in this study to examiningthe association between individuals’ cognitive performance and their demographic and socioeconomicbackgrounds. The issue of endogeneity remains to be addressed in our future work.



regression, we introduced the following 10 explanatory variables: age groups

(50–54, 55–59, 60–64†, 65–69, and 70–74), sex (man, woman†), marital status

(currently married†, widowed, divorced/separated, and single), work status (working,

not working†), education (junior high school†, senior high school, junior college, and

university or higher), self-rated health status (excellent, very good, good, fair†, and

poor), CES-D (�16, <16†), IADLs, height (centimeters), and survey cohorts

(cohort 1† consisting of those residing in Takikawa, Sendai, Adachi, Kanazawa, and

Shirakawa in 2007, cohort 2 comprising those living in Naha and Tosu in 2009, and

cohort 3 consisting of those residing in Tondabayashi, Hiroshima†, and Chofu in

2011).

Table 3. Regression Analysis of Immediate Recall Scores among Those Aged 50 and Overin Japan

(Dependent Variable = Immediate Recall Score)

Explanatory Variable Coefficient t-value Explanatory Variable Coefficient t-value

Age Self-rated health status50–54 0.262 3.44*** Excellent 0.179 2.29**55–59 0.060 0.89 Very good 0.122 1.6160–64† — — Good 0.105 1.4665–69 �0:236 �3:46*** Fair† — —

70–74 �0:499 �6:91*** Poor �0:255 �1:58Sex CES-D (20 items)Man �0:569 �7:99*** �16 �0:215 �1:56Woman† — — <16† — —

Marital status IADLs (sum) 0.129 2.26**Currently married† — — Height 0.555 1.33Widowed 0.013 0.16 Survey cohortDivorced/separated �0:169 �1:48 cohort 1 (5 cities)† — —

Single �0:354 �3:02*** cohort 2 (2 cities) �0:098 �1:68*Work status cohort 3 (3 cities) 0.137 2.29**Working �0:017 �0:33 Intercept 3.626 5.20***Not working† — —

EducationJunior high school† — —

Senior high school 0.416 7.37***Junior college 0.575 7.34***University or higher 0.748 9.59***

CES-D ¼ Center for Epidemiological Studies Depression Scale, IADLs ¼ instrumental activities of dailyliving.Notes: † denotes the reference group. Adjusted R-squared ¼ 0:102. Number of observations ¼ 4,873.***, **, and * indicate 1%, 5%, and 10% levels of statistical significance, respectively.Source: Authors’ estimates based on data from the Japanese Study of Aging and Retirement of theResearch Institute of Economy, Trade and Industry, Hitotsubashi University, Japan, and The University ofTokyo, Japan.



The respondents’ age and education have been incorporated in this regression to

capture the effect of two types of intelligence on cognitive functioning. Fluid

intelligence refers to the ability to reason and think flexibly, while crystallized

intelligence refers to the accumulation of knowledge, facts, and skills throughout life

(Cattell 1978). The explanatory variable, age, is expected to capture the change in fluid

intelligence, which peaks approximately at an age of 25. Because the respondents

included in the regression are 50 years or older, the estimated coefficients are expected

to have negative signs. The other explanatory variable (educational attainment) is

intended to capture the effect of education on crystallized intelligence, which is based

on facts and rooted in experiences. As we age and accumulate new knowledge and

understanding, crystallized intelligence becomes stronger. More importantly, because

education can also improve learning techniques such as memorization skills, education

helps improve performance in fluid intelligence even in the case of immediate word

recall. Therefore, since fluid abilities are improved by crystallized intelligence to a

substantial degree, we expect the estimated coefficient for education to have a positive

sign. In addition, we can anticipate that the higher the level of education the larger the

estimated coefficient will be.

As discussed in Section III, the magnitude, pattern, and explanation of cognitive

gender differences remain important research topics. As demonstrated in a SHARE-

based study undertaken by Weber et al. (2014), the magnitude of the gender

differences in cognitive performance in middle-aged and older populations across 13

European countries varies systematically across cognitive tasks, birth cohorts, and

geographical regions. Bonsang, Skirbekk, and Staudinger (2017) have also found that

both living conditions and educational opportunities to which individuals are exposed

during their formative years are related to increased gender differences, favoring

women in immediate word recall scores. Whether these findings based on the

European data are applicable to Japan and other Asian countries will be examined later

in this study.

As for the other explanatory variables, the health-related variables such as

self-rated health status, CES-D,32 and IADLs33 are expected to be associated with

cognitive performance. Moreover, a respondent without a spouse is likely to be left

alone without anybody to communicate with, which may weaken his or her cognitive

32Scores on the CES-D range from 0 to 60, where higher scores suggest a greater presence ofdepression symptoms. A score of 16 or higher is interpreted as indicating a risk for depression.

33In JSTAR, the respondents were asked 15 questions pertaining to IADLs, and the variable’s score,which ranges from 0 to 15 (IADLs sum), represents the number of activities that the respondent has nodifficulty performing, such as shopping, preparing meals, housekeeping, managing finances, takingresponsibility for having medication in correct dosages at the right time, etc.



functioning. Similarly, whether the respondent holds a job is likely to affect his or her

level of life satisfaction and career development, both of which may affect cognitive

performance.

The respondent’s height has been incorporated in the regression because adult

height is closely related to childhood nutritional condition which, in turn, affects

cognitive functioning and other dimensions of human capital, such as school ability

(Weir, Lay, and Langa 2014; LaFave and Thomas 2017).

We have also included in the regression a set of explanatory variables representing

survey cohorts, which differ significantly in terms of the level of urbanization of the areas

where the respondents live and their lifestyles. It is quite conceivable that, because a

considerable proportion of the respondents included in cohort 3 live in wealthy urban

areas such as Chofu in Tokyo, this cohort is more likely to be exposed to modern

technologies, such as the Internet and computers, than their counterparts in cohort 1.34

It is plausible that those who often use such modern technologies, by doing so, stimulate

their crystallized intelligence (Bordone, Scherbov, and Steiber 2015). For these reasons,

we expect that modern technologies will be more significantly associated with the

cognitive score in cohort 3 than in cohort 1.

Table 3 shows the estimated results derived from the JSTAR dataset. Except for

work status and height, all explanatory variables introduced in the regression are

statistically significant, with the coefficients having expected signs.

As expected, the cognitive abilities of Japanese older adults are negatively

associated with age. It is important to observe that education is positively related to

immediate recall scores—the higher the educational level, the better the cognitive

performance. Another important finding is that the respondent’s own health evaluation

(self-related health status) and physical limitations (IADLs) are also positively

associated with the immediate recall score. Moreover, women show a considerably

higher cognitive score than men, which is comparable to the pattern widely seen in the

Northern and Central European regions. Those who are currently married have higher

cognitive abilities than those who have never been married. In view of the rising

prevalence of lifetime singlehood in Japan over the past few decades, this variable may

play an increasingly important role in the future. Where respondents live also plays a

role in cognitive performance—the coefficient for cohort 3, which includes a

considerable number of respondents who live in relatively wealthy residential areas, is

not only statistically significant but also positive, which agrees with our a-priori

expectation.

34For example, Shirakawa Town, which is included in cohort 1, is predominantly rural.



Let us now compare these JSTAR-based regression results with those estimated

based on CHARLS, the LASI pilot survey, HART, and MARS. The results based on

the PRC’s CHARLS in Table 4 show that the coefficients of all the explanatory

variables, except for marital status and work status, are statistically significant with

theoretically expected signs. Compared to JSTAR, the coefficients for age, sex,

education, and self-rated health status are statistically significant for both datasets with

the theoretically expected signs, while marital status is statistically significant only for

Japan. However, unlike in Japan, both height (childhood nutritional conditions) and

CES-D (representing the level of depression) yielded statistically significant results in

the case of the People’s Republic of China.

Table 5 displays the regression results based on India’s LASI pilot survey data.

As mentioned, the number of observations in this dataset is relatively small—only 832

observations, which casts some doubt on the reliability of some of the estimated

results. For instance, age, sex, work status, and CES-D are not statistically significant.

However, education is a statistically significant predictor at the 1% significance level.

It is also worth noting that Punjab, which is the most developed state among the four

Indian states included in the pilot, exhibits a considerably higher cognitive

performance.

Table 6 presents the regression results estimated from Thailand’s HART dataset.

Age, sex, marital status, education, and CES-D are significant predictors with

theoretically expected signs. Due to the paucity of data, however, we could not

incorporate the explanatory variables representing self-rated health status and IADLs.

All provinces except for Surin have higher cognitive abilities than Bangkok (reference

group). This result is rather unexpected, and we do not have a reasonably good

explanation at hand.

Table 7 shows the regression results based on Malaysia’s MARS data.

All categories of the explanatory variables are statistically significant. As expected,

cognitive functioning decreases as age increases. In addition, the estimated coefficients

for sex, marital status, education, depression signs,35 IADLs, and work status are

statistically significant. Work status, unlike in other Asian countries, has a positive

coefficient.36

35The depression symptom score is constructed using 17 negative and positive statements related toa respondent’s experienced psychological well-being. The response scale for positive statements isinversely converted. The total score was calculated as the aggregate for all 17 statements. The total scoresthus range from 17 to 85, with the scores in the top 15th percentile (44 or higher) interpreted as indicating ahigher risk of depression (�44, <44†).

36Although the Malaysian survey data indicate that many of those still working are in theagriculture sector, it is not clear why this contributes to increasing their cognitive functioning.



Another significant finding is the link between the respondent’s health and

cognitive ability, whereby poor self-rated health and higher signs of depression relate

negatively to immediate recall scores. Nutrition, as represented by the respondent’s

height, is also seen to play an important role in cognitive functioning. We also observe

that the coefficient for the more urbanized states, such as Kuala Lumpur, Pulau Pinang,

and Perak, are positive and statistically significant. Better cognitive ability in these

states may be attributed to greater exposure and utilization of technology in the

subjects’ daily lives.

Table 4. Regression Analysis of Immediate Recall Scores among Those Aged 50 and Overin the People’s Republic of China



Age Self-rated health status50–54 0.059 0.79 Very good 0.250 2.47***55–59 0.070 1.18 Good 0.136 1.79*60–64† — — Fair 0.116 2.15***65–69 �0:049 �0:43 Poor† — —

70–74 �0:598 �7:73*** Very poor �0:209 �2:08**75–79 �0:923 �9:78*** CES-D (10 items)

Sex �10 �0:402 �7:98***Man �0:137 �2:07** <10† — —

Woman† — — Height 2.568 7.24***Marital status RegionCurrently married† — — Urban community† — —

Partnered 0.328 1.02 Rural �0:244 �3:70***Widowed 0.024 0.33 Intercept 0.182 0.33Divorced/separated 0.194 1.39Single �0:320 �1:25

Work statusWorking �0:105 �1:52Not working† — —

EducationLess than lowersecondary†

— —

Upper secondaryand vocational

0.819 10.79***

Tertiary 1.545 7.21***

CES-D ¼ Center for Epidemiological Studies Depression Scale.Notes: † denotes the reference group. Adjusted R-squared ¼ 0:136. Number of observations ¼ 8,738.***, **, and * indicate 1%, 5%, and 10% levels of statistical significance, respectively.Source: Authors’ estimates based on data from the China Health and Retirement Longitudinal Study ofthe National School of Development of Peking University, Beijing; see http://charls.pku.edu.cn/index/en.html (accessed 5 August 2020).



Several points of interest emerge from the foregoing discussions on the

regression results for the five Asian countries. First, in all five Asian countries, the

cognitive abilities of older adults decline with age. Second, education is highly and

positively associated with immediate recall scores. Third, health condition is positively

related with cognitive performance, and height is positively linked to better cognitive

abilities, which implies that nutritional condition in childhood plays an important role

in developing cognitive functioning at a later stage.

Fourth, those who are currently married have higher cognitive abilities than those

who have never been married. In view of the recent gradual shift from universal

Table 5. Regression Analysis of Immediate Recall Scores among Those Aged 50 and Overin India



Age Self-rated health status50–54 0.114 0.63 Excellent 0.285 0.6155–59 0.161 0.84 Very good 0.424 1.70*60–64† — — Good 0.171 0.7965–69 �0:230 �1:01 Fair† — —

70–74 �0:183 �0:66 Poor �0:604 �1:5875–79 �0:328 �1 CES-D (8 items)

Sex �8 �0:257 �1:84*Man �0:012 �0:06 <8† — —

Woman† — — Height 1.712 1.78*Marital status State

Currently married† — — Punjab† — —

Widowed �0:326 �1:69* Rajasthan �0:563 �2:98***Divorced/separated 0.082 0.20 Kerala �0:718 �3:93***Single �0:036 �0:07 Karnataka �0:879 �4:26***

Work status Intercept 2.593 1.72*Working �0:015 �0:10Not working† — —

Education

Junior high school† — —

Senior high school 0.682 4.12***Junior college 1.247 4.22***University or higher 1.333 4.12***

CES-D ¼ Center for Epidemiological Studies Depression Scale.Notes: † denotes the reference group. Adjusted R-squared ¼ 0:126. Number of observations ¼ 832.***, **, and * indicate 1%, 5%, and 10% levels of statistical significance, respectively.Source: Authors’ estimates based on data from the Longitudinal Ageing Study in India of the HarvardT.H. Chan School of Public Health, Boston, the International Institute for Population Sciences,Mumbai, and the University of Southern California, Los Angeles; see https://g2aging.org/downloads(accessed 5 August 2020).



marriage to lifetime singlehood in Japan and other Asian countries, policy makers and

researchers should pay more attention to Asia’s changing marriage patterns in the

years to come, particularly from a standpoint of cognitive performance among older

adults.

Fifth, women show considerably higher cognitive scores than men. Nevertheless,

to gain further insights into Asia’s gender differences in cognition, we plotted male and

female age-specific immediate recall scores for the five Asian countries in Figure 3.

Although our regression results have uniformly indicated that women have higher

Table 6. Regression Analysis of Immediate Recall Scores among Those Aged 50 and Overin Thailand

(Dependent Variable ¼ Immediate Recall Score)


Age CES-D (20 items)50–54 0.485 3.58*** �16 �0:760 �2:94***55–59 0.393 3.13*** <16† — —

60–64† — — Height 0.001 0.2365–69 �0:281 �2:25** State70–74 �0:652 �4:88*** Bangkok† — —

75–79 �0:931 �6:90*** Samut Prakan 0.901 3.12***Sex Nonthaburi 0.062 0.10

Man �0:231 �2:31** Pathum Thani 0.558 1.61Woman† — — Sing Buri 0.845 3.49***

Marital status Chanthaburi 0.973 4.05***Currently married† — — Surin �0:161 �0:62Widowed �0:025 �0:25 Khon Kaen 0.948 4.05***Divorced/separated 0.235 1.28 Chiang Mai 1.250 5.29***Single �0:397 �2:35** Uttaradit 0.737 2.64***

Work status Phetchabun 0.576 2.45**Working 0.054 0.060 Krabi 0.477 1.72*Not working† — — Songkhla 1.119 4.83***

Education Intercept 1.663 1.91*No formal education† — —

P1–P6 1.041 5.76***M1–M6 1.661 8.14***First year collegeor higher

2.295 7.21***

CES-D ¼ Center for Epidemiological Studies Depression Scale, M ¼ secondary education, and P ¼primary education.Notes: † denotes the reference group. Adjusted R-squared ¼ 0:170. Number of observations ¼ 2,264.***, **, and * indicate 1%, 5%, and 10% levels of statistical significance, respectively.Source: Authors’ estimates based on data from the Health, Aging, and Retirement in Thailand of theCenter for Aging Society Research/Research Center of the National Institute of DevelopmentAdministration, Bangkok.



scores than men in these Asian countries,37 Figure 3 reveals considerable differences

across the five Asian countries—in the PRC and India, for example, women have

distinctively lower cognitive scores than men. To account for the gap between our

Table 7. Regression Analysis of Immediate Recall Scores among Those Aged 50 and Over inMalaysia



Age Depression symptom score50–54 0.307 3.51*** �44 �0:229 �2:83***55–59 0.272 3.21*** <44† — —

60–64† — — IADLs (sum) 0.113 7.46***65–69 �0:178 �1:91* Height 0.017 3.87***70–74 �0:478 �4:39*** State75–79 �0:788 �5:66*** Johor† — —

Sex Kedah �0:195 �1:52Man �0:602 �7:41*** Kelantan �0:053 �0:39Woman† — — Kuala Lumpur 1.028 5.42***

Marital status Melaka �0:436 �1:88*Currently married† — — Negeri Sembilan �0:334 �1:93*Widowed �0:050 �0:60 Pahang �0:309 �2:36**Divorced/separated �0:176 �1:06 Perak 0.513 4.17***Single �0:270 �1:71* Perlis �0:154 �0:44

Work status Pulau Pinang 0.358 2.22**Working 0.140 2.01* Sabah �0:008 �0:07Not working† — — Sarawak �0:55 �4:53***

Education Selangor 0.165 1.45No schooling† — — Terengganu �0:299 �1:99**Primary 0.426 4.64*** Intercept 0.532 0.78Secondary 1.010 10.54***Tertiary 1.537 11.49***

Self-rated health statusVery good �0:123 �0:88Good �0:039 �0:63Moderate† — —

Poor �0:125 �1:34Very poor �0:444 �1:65*

IADLs ¼ Instrumental activities of daily living.Notes: † denotes the reference group. Adjusted R-squared ¼ 0:219. Number of observations ¼ 3,680.***, **, and * indicate 1%, 5%, and 10% levels of statistical significance, respectively.Source: Authors’ estimates based on data from the Malaysia Ageing and Retirement Survey of the SocialWellbeing Research Centre of the University of Malaya, Kuala Lumpur.

37In the case of India, the estimated coefficient for sex is statistically insignificant, presumably dueto the small sample size.



Figure 3. Immediate Recall Scores in Five Asian Countries by Sex and Age Group

PRC = People’s Republic of China.Sources: Authors’ calculations based on data from the following: (i) Japanese Study of Aging and Retirement ofthe Research Institute of Economy, Trade and Industry, Hitotsubashi University, Japan, and The Universityof Tokyo, Japan; (ii) China Health and Retirement Longitudinal Study of the National School of Development ofPeking University, Beijing; (iii) the pilot portion of the Longitudinal Ageing Study in India of the Harvard T.H.Chan School of Public Health, Boston, the International Institute for Population Sciences, Mumbai, and theUniversity of Southern California, Los Angeles; (iv) Malaysia Ageing and Retirement Survey of the SocialWellbeing Research Centre of the University of Malaya, Kuala Lumpur; and (v) Health, Aging, and Retirement inThailand of the Center for Aging Society Research/Research Center of the National Institute of DevelopmentAdministration, Bangkok.



regression results and the patterns in Figure 3, we need to pay attention to gender

differences in educational attainment in these Asian countries. Once we control for

education as we did in our regressions, women have an advantage over men in

cognition, suggesting that Asia’s gender difference in cognition performance is

primarily caused by gender gaps in education.

Although it falls outside the scope of this study, we are planning to carry out a

series of simulation exercises in a future study by using the regression results for the

five Asian countries generated here. In the case of Japan, for instance, the statistical

results indicate that the cognitive ability of Japanese elderly persons is likely to

improve due to the following potential factors: (i) the level of education among those

50 years and over is expected to rise at a phenomenal rate, as shown in Figure 4;

(ii) future generations of the elderly are likely to have an advantage over past

generations because children’s nutrition started to improve considerably in Japan in the

late 1950s when the school lunch program was introduced nationwide; and (iii) the use

of modern communication technologies among the elderly is likely to increase at a

remarkable rate because the overwhelming majority of young cohorts have already

been exposed to extensive use of computers and mobile phones, as illustrated in

Figure 5. By conducting various simulation exercises of this nature, we will be able to

project to what extent cognitive functioning among older adults in Japan will improve,

and how high Japan’s CADR will be in the years ahead.

Figure 4. Changes in Educational Composition in Japan by Sex, 1920–1980

CG ¼ University or higher, HSG ¼ senior high school, JCG ¼ junior college, and JSG ¼ junior high school.Source: Statistics Bureau of Japan. 2013. Population Census 2010. Tokyo: Japan Statistical Association.



VIII. Concluding Remarks

In recent years, the five Asian countries intensively analyzed in this paper have

been facing increasingly difficult policy challenges induced by rapid population aging.

Among these five countries, Japan’s level of population aging has been the most

pronounced over the past few decades. Japan has the highest proportion of those aged

65 and over, an indicator which has been used by demographers for more than a

century. One of the main objectives of this study was to introduce from an innovative

angle a new index to measure the level of population aging to shed a different light on

policy-oriented research on this phenomenon. To compute this new index—the

cognition-adjusted dependency ratio—we applied the mean age-group-specific

immediate recall scores for Japan and four other Asian countries and compared the

computed results with those derived from the US and various developed nations in

Europe.

Our computed results have shown that Japan’s pattern and level of age-related

decline in cognitive functioning are highly comparable to those of many other

developed nations, particularly those in the group designated as Continental Europe in

previous research. This finding seems to have a few important policy implications for

the aging Japanese economy, particularly its labor market. The population census data

show that the size of Japan’s total labor force, after reaching a peak in 1995, has been

shrinking continuously, while the overall labor force participation rate has been on a

Figure 5. Age-Specific Pattern of Using the Internet in Japan, 2016

Source: Ministry of Internal Affairs and Communications. 2017. Communications Usage Trend Survey. https://www.soumu.go.jp/johotsusintokei/tsusin_riyou/data/eng_tsusin_riyou02_2017.pdf.



downward trend since 1970. Despite these substantial changes, Japan’s age-based

employment practices, which comprise lifetime employment, seniority wage system,

and mandatory retirement age, remained virtually intact (Kato 2016). Particularly,

Japan’s policies related to the mandatory retirement age—requiring workers to leave

the company at a relatively young age, typically at an age of 60—are considered

extreme compared to the practices in other industrialized nations.38 Due to the

existence of such age-based employment practices, many Japanese businesses that

face fierce competition from overseas rivals have been confronted in recent years with

a shortage of highly qualified workers with specialized skills acquired from career

experiences. Our research finding is not yet widely known in Japan’s labor market, but

once the market recognizes that older Japanese have reasonably good cognitive

performance, the finding could provide a strong incentive for many employers to

modify or even abandon the long-running age-based employment practices. This

would allow a sizable number of older Japanese with a reasonably good level of

cognitive functioning to be recruited, which would likely generate a considerable

amount of economic dynamism.

It is also worth noting that among the selected Asian countries, Malaysia shows a

pattern of change in age-specific cognitive functioning that is similar to the Southern

European group, although Malaysia has somewhat lower scores than Southern Europe

in all age groups.

More importantly, these inter-country comparative results based on cognition-

adjusted dependency ratios are astonishingly different from the results emerging from

the conventional old-age dependency ratios. This conclusion seems to justify the UN’s

recent efforts to raise awareness regarding the urgent need for remeasuring population

aging with a view to formulating more efficient and effective policies to cope with

rapid population aging in both developed and developing nations.

References

Anderson, Nicole, and Fergus Craik. 2000. “Memory in the Aging Brain.” In The OxfordHandbook of Memory, edited by Endel Tulving and Fergus Craik, 411–25. Oxford:Oxford University Press.

Arifin, Evi Nurvidya, and Aris Ananta, eds. 2009. Older Persons in Southeast Asia: AnEmerging Asset. Singapore: Institute of Southeast Asian Studies.

38In Western industrialized nations, cognitive performance is becoming increasingly important forlabor market outcomes (Gordo and Skirbekk 2013).



Atalay, Kadir, Garry Barret, and Anita Staneva. 2019. “The Effect of Retirement on ElderlyCognitive Functioning.” Journal of Health Economics 66 (C): 37–53.

Behrman, Jere, David Ross, and Richard Sabot. 2008. “Improving Quality versus Increasing theQuantity of Schooling: Estimates of Rates of Return from Rural Pakistan.” Journal ofDevelopment Economics 85 (1–2): 94–104.

Bloom, David, and Jeffrey Williamson. 1998. “Demographic Transitions and EconomicMiracles in Emerging Asia.” The World Bank Economic Review 12 (3): 419–55.

Bonsang, Eric, Stephane Adam, and Sergio Perelman. 2012. “Does Retirement AffectCognitive Functioning?” Journal of Health Economics 31: 490–501.

Bonsang, Eric, Vegard Skirbekk, and Ursula Staudinger. 2017. “As You Sow, So shall YouReap: Gender-Role Attitudes and Late-Life Cognition.” Psychological Science 28 (9):1201–13.

Bordone, Valeria, Sergei Scherbov, and Nadia Steiber. 2015. “Smarter Every Day: TheDeceleration of Population Ageing in Terms of Cognition.” Intelligence 52: 90–96.

Cattell, Raymond. 1978. The Scientific Use of Factor Analysis. New York: Plenum Press.Clark, Robert, and Naohiro Ogawa. 1992. “Employment Tenure and Earnings Profiles in Japan

and the United States: Comments.” The American Economic Review 82 (1): 336–45.Clark, Robert, Naohiro Ogawa, and Andrew Mason, eds. 2007. Population Aging,

Intergenerational Transfers and the Macroeconomy. Cheltenham, UK and Northampton,MA: Edward Elgar.

Clark, Robert, Naohiro Ogawa, Sang-Hyop Lee, and Rikiya Matsukura. 2008. “Older Workersand National Productivity in Japan.” In Population Aging, Human Capital Accumulation,and Productivity Growth: A Supplement to Vol. 34, 2008: Population and DevelopmentReview, edited by Alexia Prskawetz, David Bloom, and Wolfgang Lutz, 257–74.Hoboken, NJ: Blackwell Publishing (On Behalf of the Population Council).

de Grip, Andries, Arnaud Dupuy, Jelle Jolles, and Martin van Boxtel. 2015. “Retirement andCognitive Development in the Netherlands: Are the Retired Really Inactive?” Economics& Human Biology 19: 157–69.

Doblhammer, Gabriele, Gerald van den Berg, and Thomas Fritze. 2013. “Economic Conditionsat the Time of Birth and Cognitive Abilities Late in Life: Evidence from Ten EuropeanCountries.” PLoS ONE 8 (9): e74915.

Eggleston, Karen, and Shripad Tuljapurkar, eds. 2010. Aging Asia: The Economic and SocialImplications of Rapid Demographic Change in China, Japan, and South Korea. Stanford,CA: Walter H. Shorenstein Asia-Pacific Research Center and Baltimore, MD: TheBrookings Institution.

Fein, George, Shannon McGillivray, and Peter Finn. 2007. “Older Adults make LessAdvantageous Decisions than Younger Adults: Cognitive and Psychological Correlates.”Journal of the International Neuropsychological Society 13 (3): 480–89.

Fu, Tsung-hsi, and Rhidian Hughes. 2009. Ageing in East Asia: Challenges and Policies for theTwenty-First Century. Volume 4. Comparative Development and Policy in Asia Series.London and New York: Routledge.

Fuller, Theodore. 2010. “Relationship Status, Health, and Health Behavior: An Examination ofCohabiters and Commuters.” Sociological Perspectives 53 (2): 221–45.



Gordo, Laura, and Vegard Skirbekk. 2013. “Skill Demand and the Comparative Advantage ofAge: Job Tasks and Earnings from the 1980s to the 2000s in Germany.” LabourEconomics 22 (C): 61–69.

Hermalin, Albert, ed. 2003. The Well-Being of the Elderly in Asia: A Four-CountryComparative Study. Ann Arbor, MI: The University of Michigan Press.

Ichimura, Hidehiko, Hideki Hashimoto, and Satoshi Shimizutani. 2009. “Japanese Study ofAging and Retirement: First Results.” RIETI Discussion Paper Series 09-E-047, TheResearch Institute of Economy, Trade and Industry, Tokyo.

Kato, Takao. 2016. “Productivity, Wages and Unions in Japan.” Conditions of Work andEmployment Series No. 73, International Labour Office, Geneva.

Kendig, Hal, Peter McDonald, and John Piggott, eds. 2016. Population Ageing and Australia’sFuture. Acton, ACT: Australian National University Press.

Kuiper, Jisca, Marij Zuidersma, Richard Voshaar, Sytse Zuidema, Edwin van den Heuvel,Ronald Stolk, and Nynke Smidt. 2015. “Social Relationships and Risk of Dementia: ASystematic Review and Meta-analysis of Longitudinal Cohort Studies.” Ageing ResearchReviews 22: 39–57.

LaFave, Daniel, and Duncan Thomas. 2017. “Height and Cognition at Work: Labor MarketProductivity in a Low-Income Setting.” Economics & Human Biology 25: 52–64.

Lee, Ronald, and Andrew Mason, eds. 2011. Population Aging and the Generational Economy:A Global Perspective. Cheltenham, UK: Edward Elgar Publishing.

Lutz, Wolfgang, Warren Sanderson, and Sergei Scherbov. 2004. The End of World PopulationGrowth in the 21st Century: New Challenges for Human Capital Formation andSustainable Development. London: Earthscan.

Lynn, Richard. 2009. “Fluid Intelligence But not Vocabulary has Increased in Britain, 1979–2008.” Intelligence 37 (3): 249–55.

Machin, Stephen, and John Van Reenen. 1998. “Technology and Changes in Skill Structure:Evidence from Seven OECD Countries.” The Quarterly Journal of Economics 113 (4):1215–44.

Maharani, Asri, and Gindo Tampubolon. 2016. “National Economic Development Status mayAffect the Association between Central Adiposity and Cognition in Older Adults.” PLoSONE 11 (2): e0148406.

Mason, Andrew, ed. 2001. Population Change and Economic Development in East Asia:Challenges Met, and Opportunities Seized. Stanford, CA: Stanford University Press.

Maurer, Jurgen. 2010. “Height, Education and Later-Life Cognition in Latin America and theCaribbean.” Economics & Human Biology 8 (2): 168–76.

Ministry of Internal Affairs and Communications. 2017. Communications Usage Trend Survey.https://www.soumu.go.jp/johotsusintokei/tsusin_riyou/data/eng_tsusin_riyou02_2017.pdf.

Munnell, Alicia, Steven Sass, and Mauricio Soto. 2006. “Employer Attitudes towards OlderWorkers: Survey Results.” Work Opportunities for Older Americans, Series 3, June,Center for Retirement Research at Boston College.

Negash, Selam, Glenn Smith, Shane Pankratz, Jeremiah Aakre, Yonas Geda, Rosebud Roberts,David Knopman, Bradley Boeve, Robert Ivnik, and Ronald Petersen. 2011. “SuccessfulAging: Definitions, and Prediction of Longevity and Conversion to Mild CognitiveImpairment.” The American Journal of Geriatric Psychiatry 19 (6): 581–88.



Ofstedal, Mary Beth, Gwenith Fisher, and Regula Herzog. 2005. Documentation of CognitiveFunctioning Measures in the Health and Retirement Study. Ann Arbor, MI: Institute ofSocial Research, University of Michigan.

Ogawa, Naohiro, Norma Mansor, Sang-Hyop Lee, Michael Abrigo, and Tahir Aris. 2021.“Population Aging and the Three Demographic Dividends in Asia.” Asian DevelopmentReview 38 (1): 32–67.

Park, Donghyun, Sang-Hyop Lee, and Andrew Mason, eds. 2012. Aging, Economic Growth,and Old-Age Security in Asia. Cheltenham, UK and Northampton, MA: Edward Elgar.

Sanderson, Warren, and Sergei Scherbov. 2010. “Remeasuring Aging.” Science 329 (5997):1287–88.

Schneeweis, Nicole, Vegard Skirbekk, and Rudolf Winter-Ebmer. 2014. “Does EducationImprove Cognitive Performance Four Decades after School Completion?” Demography51 (2): 619–43.

Skirbekk, Vegard, Valeria Bordone, and Daniela Weber. 2014. “A Cross-Country Comparisonof Math Achievement at Teen Age and Cognitive Performance 40 Years Later.”Demographic Research 31: 105–18.

Skirbekk, Vegard, Elke Loichinger, and Daniela Weber. 2012. “Variation in CognitiveFunctioning as a Refined Approach to Comparing Aging across Countries.” Proceedingsof the National Academy of Sciences of the United States of America 109 (3): 770–74.

Slomski, Anita. 2014. “Midlife Diabetes, Hypertension, May Affect Cognition Later in Life.”JAMA 311 (20): 2056.

Sommerlad, Andrew, Joshua Ruegger, Archana Singh-Manoux, Glyn Lewis, and GillLivingston. 2018. “Marriage and Risk of Dementia: Systematic Review andMeta-analysis of Observational Studies.” Journal of Neurology, Neurosurgery andPsychiatry 89 (3): 231–38.

Statistics Bureau of Japan. 2013. Population Census 2010. Tokyo: Japan StatisticalAssociation.

Sundet, Jon, Ingrid Borren, and Kristian Tambs. 2008. “The Flynn Effect Is Partly Caused byChanged Fertility Patterns.” Intelligence 36 (3): 183–91.

Thorvaldsson, Valgeir, Scott Hofer, Stig Berg, and Boo Johansson. 2006. “Effects ofRepeated Testing in a Longitudinal Age-Homogeneous Study of Cognitive Aging.”The Journals of Gerontology, Series B: Psychological Sciences and Social Sciences61 (6): 348–54.

Tuljapurkar, Shripad, Naohiro Ogawa, and Anne Gauthier, eds. 2010. Ageing in AdvancedIndustrial States: Riding the Age Waves—Volume 3. International Studies in Population.London and New York: Springer.

Uhlenberg, Peter, ed. 2009. International Handbook of Population Aging. London and NewYork: Springer.

United Nations. 2007. World Economic and Social Survey 2007: Development in an AgeingWorld. New York: United Nations.

_____. 2019. World Population Prospects: The 2019 Revision. New York: United Nations.van Aken, Loes, Roy Kessels, Ellen Wingbermuhle, William van der Veld, and Jos Egger.

2016. “Fluid Intelligence and Executive Functioning More Alike than Different?” ActaNeuropsychiatrica 28 (1): 31–37.



Weber, Daniela, Vegard Skirbekk, Inga Freund, and Agneta Herlitz. 2014. “The Changing Faceof Cognitive Gender Differences in Europe.” Proceedings of the National Academy ofSciences of the United States of America 111 (32): 11673–78.

Weir, David, Margaret Lay, and Kenneth Langa. 2014. “Economic Development and GenderInequality in Cognition: A Comparison of China and India, and of SAGE and the HRSSister Studies.” The Journal of the Economics of Ageing 4: 114–25.



Demographic Change, Economic Growth,and Old-Age Economic Security:

Asia and the World

ANDREW MASON, SANG-HYOP LEE,AND DONGHYUN PARK¤

Asia is aging, although there is significant heterogeneity across subregions andeconomies. Population aging poses two strategic challenges for the region:sustaining economic growth and delivering old-age economic security. In thispaper, we leverage the lifecycle perspective—that individuals’ consumptionand labor income differ at each age—and the National Transfer Accountsdatabase to construct and analyze key economic indicators. Our analysisconfirms that demographic change will challenge the region’s future growthand increase the cost of funding the consumption of the elderly. We also findthat it will have a substantial impact on the public finances of some Asianeconomies.

Keywords: population aging, demographic change, Asia, economic growth,old-age economic security

JEL codes: J11, J14

⁄Andrew Mason: East-West Center, Honolulu, HI, United States. E-mail: [email protected];Sang-Hyop Lee: East-West Center, Honolulu, HI, United States and University of Hawai‘i at Mānoa,Honolulu, HI, United States. E-mail: [email protected]; Donghyun Park (corresponding author):Asian Development Bank, Metro Manila, Philippines. E-mail: [email protected]. We thank the ManagingEditor and the anonymous referees for helpful comments and suggestions.





https://dx.doi.org/10.1142/S0116110522500019

https://dx.doi.org/10.1142/S0116110522500019

I. Introduction

Asia and the Pacific, and indeed the entire world, is in the midst of a

demographic transition that has fundamental implications for economic growth,

generational equity, public finances, and other important features of national and

regional economies. It is widely recognized that population aging may adversely affect

economic dynamism, exacerbate inequality between generations, and harm public

finances. More broadly, older populations pose two strategic challenges for Asia:

(i) sustaining economic growth in the face of less favorable demographics; and

(ii) delivering old-age economic security for a large and growing elderly population.

In this paper, we undertake a granular analysis of demographic and economic data to

improve our understanding of the economic impact of the demographic transition in

Asia and the world.

Demographic data point to two interrelated and stylized global trends: population

aging and slowing population growth. Significantly, Asian economies, which enjoyed

a substantial dividend from large workforces in the past, will see effective labor growth

drop sharply between 2020 and 2060. While the populations of Asia and the world are

aging and growing more slowly, there is substantial heterogeneity within the world

and Asia. A key objective of this paper is to explore this heterogeneity within Asia

and the Pacific.

A lifecycle perspective allows us to better understand exactly how demographic

change affects the economy at any point in time. The key principle here is that the

profiles of average labor income and consumption depend heavily on age. Individuals

consume more than they earn when they are young or old, and they earn more than

they consume during productive middle-ages. While these broad stylized facts hold for

all economies, there are significant differences across economies. The lifecycle

perspective dates back to some of the earliest literature on population and economic

growth. What is new here are the comprehensive efforts to quantify how economic

data vary over the lifecycle and how patterns vary across economies due to

development level and other factors.

Combining the lifecycle profiles of consumption and labor income with

population estimates and projections from the United Nations (United Nations

Population Division 2019), we are able to project three key indicators that capture the

impact of demographic transition on the economy. These are (i) effective number of

workers; (ii) effective number of consumers; and (iii) support ratio, which is the ratio

of (i) to (ii). Asia will see a sharp decline in the effective number of workers in the

coming decades. As a result, the first demographic dividend, or the boost to economic



growth due to rising support ratios, will dwindle across the region. The decrease will

be most pronounced in East Asia, which has experienced a rapid decline in fertility.

Population aging gives rise to two strategic challenges: sustaining economic

growth and securing adequate resources for old-age consumption. The resources for

old-age consumption can come from two sources, either lifecycle savings or

contemporaneous private or public transfers. We estimate the old-age GAP ratio

(OAGAP), or the amount of resources required to meet the old-age deficit. Funding the

old-age gap through 2060 would require funding in excess of 40% of contemporaneous

labor income in East Asia and more than 20% of labor income in the rest of developing

Asia.1 We estimate the retirement wealth that would be required to fund old-age needs.

The estimates vary widely across developing Asia’s subregions and economies. We also

show that, compared with relying on funded approaches to meeting old-age needs,

relying on transfers is relatively cheap in the beginning but very expensive when aging

sets in.

Finally, population aging will have sizable implications for the public finances of

developing Asian economies. Public transfers are one of the main avenues for funding

the gap between labor income and consumption at old age. The size of public

transfer inflows and outflows will differ at each age. For example, during productive

middle-ages we can expect outflows to the government (i.e., tax payments) to exceed

inflows from the government (i.e., benefits). We use such age profiles to project the

impact of population on public finances. The old-age gap, or the difference between

labor income and consumption for those aged 65 years and above, varies significantly

among economies and is rising significantly in the Republic of Korea, Thailand, and

Taipei,China. As a result, net public transfers to those aged 65 and older are projected

to rise sharply by 2060 to 15% and 25% of total labor income in the Republic of Korea

and Taipei,China, respectively, although they will remain a much more modest 4% in

Thailand.

The contribution of this paper to the literature is to provide new National

Transfer Accounts (NTA) data for Asia and the Pacific and to assess the potential

impact of population change on key dimensions of the economy. These data are

sufficient to highlight very important differences within and between the subregions of

Asia and the Pacific. We emphasize some of the simple, first-order effects of

population rather than the complexities that have been explored elsewhere in our own

work and the works of others. For example, we do not consider the impact of

fertility on human capital (Becker, Murphy, and Tamura 1990; Lee and Mason 2010;

1Developing Asia refers to the 46 developing member economies of the Asian Development Bank.

DEMOGRAPHIC CHANGE, ECONOMIC GROWTH, AND OLD-AGE ECONOMIC SECURITY 133


Lee et al. 2014); relationship between population and capital deepening (Tobin 1967);

endogeneity of health-care spending, fertility, and economic growth (Ehrlich and Yin

2013, de la Croix 2017); or the possibility that population change will lead to secular

stagnation and low interest rates in advanced economies (Eggertsson, Lancaster, and

Summers 2019). These issues are more appropriately addressed in simulation or

econometric models that are usually constructed for a single economy.

Another aspect of this study is that the availability of NTA data is limited. The

growth of NTA has been rapid and there is widespread interest. The methods have

been codified through close cooperation with the United Nations (UN) Department of

Economic and Social Affairs. Resources for training and implementation have been

provided by many agencies and foundations. Estimates are now available for 19

economies in Asia and the Pacific, 14 of which have a World Bank income status of

upper–middle income or below. However, NTAs for some very important economies

have not been constructed in recent years. Institutional support is essential to maintain

up-to-date estimates, construct time series, and extend coverage to economies that

have not yet participated.

II. Setting the Stage: Demographic Transition

The global demographic transition began with a decline in death rates,

particularly among infants and children. This led to rapid population growth and a

high concentration of the global population at young ages. During the 1970s, many

economies entered a new phase of their demographic transition when declining birth

rates led to slower population growth and a major shift in age distributions, a reduced

concentration at child ages, and an increased concentration at working ages. As the

demographic transition proceeds, low fertility rates lead to even slower population

growth, population decline in some economies, and further changes in age structure

(i.e., a lower concentration at working ages and a higher concentration at older ages).

Important differences in age structure across regions of the world are shown in

Figure 1. Age structure as measured here incorporates differences across economies in

the extent to which members of each age group rely on their own labor income to fund

their own consumption. The “GAP ratio” is an estimate of the gap between total

consumption and total labor income at each age as a percentage of total labor income.

The significance of this distinction and the differences across economies are discussed

in more detail below.

The GAP ratio varies considerably across the world and within Asia and the

Pacific because some economies are very far along in their demographic transition and



Figure 1. GAP Ratios by Region for 2020, 2040, and 2060 (%)

Notes: GAP ratios are the gap between consumption and labor income as a percentage of total labor income forchildren and young adults aged 0−24 years, seniors aged 65 years and older, and both age groups combined. Tofacilitate comparison with the Asian Development Bank classifications, the developed member economies in Asia(Australia, Japan, and New Zealand) are not included. Values for developing member economies are reported inthe figure. Values are the simple averages of values for each economy belonging to the regional group in question.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



others are at a relatively early stage. Currently, countries in Africa are relatively early

in their demographic transitions. They have a very high GAP ratio because the material

needs of children greatly exceed what they produce through their labor. In 2020, almost

63% of the total labor income of African economies was needed to fund the gap between

consumption and labor income for those under the age of 25. The burden of supporting

children and youth in Africa is expected to decline very sharply over the next 40 years.

The projected changes in the child GAP ratio are also projected to be substantial in Asia

and in Latin America and the Caribbean. Based on the situations in Europe and North

America, the child GAP ratio might stabilize at about one-quarter of total labor income

for each region in the future, but that will depend on trends in consumption among

children and young adults and features of labor markets that are uncertain.

Aging, as measured by the old-age GAP ratio, is a universal phenomenon

occurring in all regions of the world. The countries of Europe and North America have

higher old-age GAP ratios, and that will continue to be the case in the future. The

situations are similar in Asia and in Latin America and the Caribbean, although Asia is

not quite as “old.” Both regions will experience sharp increases in the old-age share of

their respective populations. African and Pacific Island economies are, on average,

aging more slowly than economies in other parts of the world.

The second global trend is slower growth. Throughout the world, the number of

children is growing more slowly than in the past. In many economies, the number of

children has declined substantially. This has been matched, with a delayed effect, by

slower growth or a decline in the number of workers. This slowdown in population

growth has fundamental implications for economic growth. Given the productivity of

effective labor, slower growth in the number of workers means slower growth in gross

domestic product (GDP). This trend for each major region is captured in Figure 2

using estimates and projections of the annual growth rate of effective labor. Effective

labor is the population weighted to capture age variation in labor force participation,

unemployment, hours worked, and wages.

Starting in the 1970s, rapid growth in the number of effective workers has led to

more rapid growth in GDP throughout the world. But the era of population-driven

economic growth is coming to an end everywhere except in Africa. There, growth in the

effective number of workers was estimated to add 2.9 percentage points per year to

economic growth in 2020. In Asia and the Pacific economies, growth in the number of

effective workers was estimated to add only 1.5 percentage points to GDP growth per

year. The effect is somewhat smaller in Latin America and the Caribbean. Among

countries in North America, growth in the number of effective workers was estimated to

add only 0.4 percentage points per year to GDP growth in 2020. In Europe, growth in

the effective labor force is already estimated to have a negative impact on GDP growth.



Over the next few decades, population numbers per se will no longer be an

important driver of GDP growth. Africa will increasingly become the only region of

the world where growth in the effective number of workers will play an important role.

Within the broad context outlined above, demographic conditions are very

diverse in Asia and the Pacific, with some economies very far along in their

demographic transition and others at a relatively early stage. The diversity is quantified

using two measures—the old-age GAP ratio and the growth of the effective labor

force—that are expected to play a crucial role in macroeconomic performance.

The left-hand side panel in Figure 3 provides current values (projected values for

2020) distinguishing members of the Asian Development Bank by region.2 Economies

2In the figures, Central Asia consists of Armenia, Azerbaijan, Georgia, Kazakhstan, the KyrgyzRepublic, Tajikistan, Turkmenistan, and Uzbekistan. East Asia includes Hong Kong, China; Mongolia; thePeople’s Republic of China; the Republic of Korea; and Taipei,China. South Asia covers Afghanistan,Bangladesh, Bhutan, India, Maldives, Nepal, Pakistan, and Sri Lanka. Southeast Asia includes BruneiDarussalam, Cambodia, Indonesia, the Lao People’s Democratic Republic, Malaysia, Myanmar, thePhilippines, Singapore, Thailand, Timor-Leste, and Viet Nam. The Pacific consists of Fiji, Papua NewGuinea, Samoa, Solomon Islands, Tonga, and Vanuatu. Other economies are Australia, Japan, and NewZealand (also referred to here as developed member economies). ADB placed on hold its assistance inAfghanistan effective 15 August 2021 (https://www.adb.org/news/adb-statement-afghanistan).

Figure 2. Annual Growth Rates of Effective Labor for Regions of the World in2020, 2040, and 2060 (%)

Notes: To facilitate comparison with the Asian Development Bank classifications, the developed membereconomies (Australia, Japan, and New Zealand) are not included. Values for developing member economies onlyare reported in the figure. Values are the simple averages of values for each economy belonging to the subregionalgroup in question.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



that are earliest in their demographic transition are frequently found in South Asia. In

Nepal, for example, effective labor force growth is very rapid at about 4% per annum

and the old-age GAP ratio is 8.4%. In many economies, the old-age GAP ratio is less

than in Nepal. Thirteen economies have a GAP ratio of less than 5%. Among the

developing Asian economies, those that are furthest along in their demographic

transition are a varied group. Growth in the effective labor force is negative in four

East Asian economies—the People’s Republic of China (PRC); Hong Kong, China;

the Republic of Korea; and Taipei,China—but also in Armenia, Georgia, and

Thailand. The extent of aging in these economies varies greatly, however. The PRC is

not yet very old, for example, with an old-age GAP ratio of only about 6%.

As the demographic transition proceeds in Asia and the Pacific, further aging and

slower growth will become pervasive. By 2060, the old-age GAP ratio will exceed

20% in 14 developing member economies, compared with only two—Georgia and

Hong Kong, China—in 2020. Growth of the effective labor force will be negative or

essentially zero in 21 developing member economies in 2060, compared with six

today. Aging and slower growth are pervasive in the region, however. Every

Figure 3. Old-Age GAP Ratios and Growth Rates of the Effective Labor Force in Asia andthe Pacific, 2020 and 2060

Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



developing member economy in Asia and the Pacific will have a population that is

older and growing more slowly in 2060 than in 2020.

Both demographic and economic factors explain the distinctive features of regions

and economies, which are highlighted further below. Although the broad outlines of the

age transition appear to be common to all economies in Asia, there is enormous

diversity across subregions and economies. Most East Asian economies have

experienced a rapid fertility decline, leading to rapid population aging and slowing

population growth. The same is true of some Southeast Asian economies like Thailand

and Singapore. However, in Singapore, the effects of low fertility have had less effect

because of its liberal immigration policies. Elsewhere in Asia, especially in South Asia,

the demographic transition is an evolving event where birth and death rates continue to

decline more slowly and steadily in many cases than in other rapidly aging economies.

Economic behavior matters as well. In Japan, the effects of aging are reinforced

by high levels of health-care spending. In the Republic of Korea, the economic effects

of aging are more moderate because consumption by seniors is low relative to

consumption by children and younger adults. In India, the impact of aging is affected

by the sharp drop in labor income at age 60 due to labor policies.

III. The Generational Economy and the Lifecycle: An Overview

Understanding how a change in population will influence the economy builds on

an understanding of the lifecycle. What does it mean to be young or to be old from an

economic perspective? Chronological age is widely used with those under 15 years old

often classified as children, those 15–64 years old as working-age adults, and those

65 years and older as seniors. However, the meaning of childhood or youth varies from

economy to economy and is changing over time depending on many factors. How long

do children remain in school and how quickly can they secure quality jobs when they

leave school? Likewise, the meaning of old age depends on the health of older

individuals, attitudes and policies toward working at older ages, public policies toward

pensions and health-care spending, and other features of later years of life. Several

other scholars have also recently explored different concepts of age that are not based

on years since birth but other markers of the aging process (Balachandran et al. 2019,

Sanderson and Scherbov 2010).

The results presented make use of detailed data from the NTA project and

population projections by age prepared by the UN to explore how changing age

structure is likely to interact with age-specific economic behavior to influence

the macroeconomy (United Nations Population Division 2013, Lee and Mason 2011).



The NTA framework provides a comprehensive accounting of any economy that

incorporates age details while maintaining broad consistency with the UN System of

National Accounts. The NTA flow accounts can be expressed in simplest terms by the

following identity:

C(x, t)� Yl(x, t) ¼ T(x, t)þ Ya(x, t)� S(x, t), ð1Þwhere C, Yl, T , Ya, and S are the aggregate consumption, labor income, net transfers,

asset income, and saving, respectively, of persons aged x in year t. The left-hand side

of equation (1) captures the basic lifecycle problem: the considerable differences

between consumption and labor income depending on the age of individuals. The

right-hand side, referred to as age reallocations, quantifies the economic mechanisms

available to deal with the lifecycle. Economies can rely on net transfers from parents to

children, for example, or from taxpayers to retirees, as another example, to fund

lifecycle needs. The alternative economic mechanism is asset-based reallocations.

People can rely on asset income or dis-saving to fund lifecycle needs.

The terms in equation (1) are broad macroeconomic measures, and NTA

accounts are constructed to conform to each economy’s total income accounts data.

The classification of these data by age provides a simple mechanism for assessing how

demographic change will influence the economy. The lifecycle patterns of behavior are

quantified using per capita age profiles. To illustrate, aggregate consumption and labor

income are equal to

C(x, t) ¼ �c(x, t)P(x, t),

Yl(x, t) ¼ �yl(x, t)P(x, t),ð2Þ

where �c(x, t) is the per capita consumption of persons aged x in year t, �yl(x, t) is the per

capita labor income of persons aged x in year t, and P(x, t) is the population aged x in

year t.

Considering economies at different income levels illustrates some of the key

features of the lifecycle of consumption and labor income. In Figure 4, labor income

and consumption are compared across economies by calculating the per capita values

relative to the average value for the 30–49-year-old age group. This normalization

allows us to compare age shapes for economies at very different levels of

development. The estimates are available for 70 economies, which have been grouped

by income class. The median value for each income group is charted.

At young ages, children consume less than adults in all economies irrespective of

their income level, but consumption increases as children mature and, in particular, as

they enter school. The increase is particularly sharp in high-income, low-fertility

economies due to high levels of spending on education. The age pattern of consumption

at adult ages also varies with the level of income. In low-income economies,



consumption per adult is lower at older ages. Adults in their mid-20s are consuming

about 10% more while adults at age 70 are consuming about 10–15% less than adults in

the 30–49-year-old age range. Consumption is flatter for adults living in middle-income

economies. And for high-income economies, on average, consumption is much higher at

older ages. By age 60, adults in high-income economies are consuming about 10% more

than those in the 30–49-year-old age range. Under the influence of spending on health

care, adults in their late 80s are consuming about 20%more than adults aged 30–49 years.

Labor income is low for the young, rises with age during the 20s and 30s,

reaches a plateau during the 40s, and declines beginning in the 50s and early 60s.

Labor income is concentrated in a narrower age range in high-income economies, with

labor income rising at later youth ages and declining to low levels at an earlier point in

old age. Even though older adults are healthier in high-income economies, they have

lower labor income relative to earnings at prime ages than do older adults in middle- or

low-income economies.

Figure 4. Consumption and Labor Income by Age and Income Group

Notes: Values for each economy are expressed relative to the mean value of those in the 30–49-year-old agegroup. Average values for each income group are calculated as the median value for the economies belonging tothat group. See Appendix for additional information.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



The remaining results presented in this paper emphasize Asia and the Pacific. NTA

estimates are now available for the 19 Asia and the Pacific economies shown in

Figure 5. (These estimates have been constructed by research teams in each of the

19 economies with support provided by the NTA project and the UN Population Fund.)

The broadest patterns reported above generally hold for individual economies, but there

are important differences in the details. In high-income economies, for example,

consumption rises with age in Japan and Australia, but not in the Republic of Korea,

Singapore, or Taipei,China. Labor income reaches a peak at very young ages in the PRC

and Viet Nam. In India, labor income declines very sharply at age 60 due to the effects of

mandatory retirement on high-paid workers in the formal sector. Of all NTA economies,

including those outside Asia and the Pacific, the PRC has very low consumption at

every age. Some like the Philippines and Timor-Leste have high consumption relative to

their labor income. This can occur due to high net transfers from the rest of the world or

due to high total income from natural resources, although this is not a factor for the

economies shown in Figure 5. These differences are important for assessing how

changes in population age structure will influence economic growth.

An important general feature of the labor income profiles is that they do not rely on

an arbitrary definition of the working-age population. The labor income profile for each

economy reflects actual labor force participation plus three other important dimensions

of labor: unemployment, hours worked, and earnings per hour. These elements are

critical to assessing the contribution of adults in the working ages and older.

Reallocation systems receive less attention in this paper, but in every economy the

lifecycle deficit of children is funded almost entirely by transfers. For young children,

private (familial) transfers are particularly important. For older children, public transfers

approach or even exceed private transfers due to the high level of public spending on

education in many economies. The reallocation system for funding old-age needs is

much more varied. In continental Europe and Latin America, public transfer systems

play a very prominent role. In Australia, Canada, Japan, the United Kingdom, and the

United States, seniors depend on a more balanced combination of public transfers and

asset-based reallocations. In several East Asian economies, net private transfers from

children still remain important to seniors. In many Asian developing economies for

which data are available, seniors rely heavily on asset-based reallocations.

A. Transforming Population and Lifecycle Profiles into NTA Indicators

Population data and age profiles of consumption and labor income are combined

to produce indicators that capture the effects of population trends on the aggregate



Figure 5. Per Capita Consumption and Labor Income in International Dollars by IncomeGroup in Asia and the Pacific (Purchasing Power Parity-Adjusted)

C ppp = per capita consumption, Yl ppp = labor income.Source: National Transfer Accounts, www.ntaccounts.org (accessed 15 March 2020).



economy (Mason et al. 2017). The approach taken here is to answer a simple question.

Suppose the age patterns of consumption and labor income do not change in the future,

how would changes in population age structure lead to strains or imbalances in the

macroeconomy? The results are not a forecast of changes in the macroeconomy.

Without question, patterns of consumption and labor income must change to restore

balance in the macroeconomy. The goal is to assess the nature, timing, and magnitude

of those imbalances.

Three measures are emphasized in this subsection—effective labor, effective

consumers, and the support ratio—that capture the impact of age structure on

economic growth. Additional indicators are considered below. In all cases, we project

values by combining population projections by age with per-age profiles held at the

baseline value estimated from NTA.

For any economy, the effective number of workers measures the impact of

population change on effective labor by incorporating age variation in labor force

participation, unemployment, hours worked, and wages (or estimated productivity for

those who are self-employed or unpaid family workers). Effective labor in year t is as

follows:

L(t) ¼X!x¼0

yl(x, b)P49x¼30 yl(x, b)

P(x, t), ð3Þ

where L(t) is the total number of effective workers in year t, yl(x, b) is per capita labor

income of persons aged x in the base year, and P(x, t) is the population of age x in the

year t. This formulation counts adults in the 30–49-year-old age range as one effective

worker, on average, while those at each single year of age are counted as more or less

than one effective worker depending on how their per capita labor income compares

with that of the 30–49-year-old age group. The age profiles of per capita labor income

shown in Figure 5 are used for this purpose.

The effective number of consumers captures the reality that needs, as well as

productivity, vary over the lifecycle. The effective number of consumers is given as

follows:

N(t) ¼X!x¼0

c(x, b)P49x¼30 c(x, b)

P(x, t), ð4Þ

where N(t) is the total number of effective consumers in year t, c(x, b) is per capita

consumption of persons aged x in the base year, and P(x, t) is the population of age x in

the year t. This formulation counts adults in the 30–49-year-old age range as one

effective consumer, on average, while those at each single year of age are counted as



more or less than one effective consumer depending on how their per capita

consumption compares with that of the 30–49-year-old age group. The age profiles of

per capita consumption shown in Figure 5 are used for this purpose.

The support ratio is the ratio of the number of effective workers to the number of

effective consumers in year t:

SR(t) ¼ L(t)N(t)

: ð5Þ

IV. The Great Slowdown of Labor Force Growth

At the global level, the effective labor force was growing rapidly during the last

half of the 20th century, providing a tailwind to economic growth. The winds are

shifting, however, and for many economies labor force growth is slowing or even

turning negative. The first-order effects of population on GDP are captured by a simple

identity

Y(t) ¼ Y(t)L(t)

L(t): ð6Þ

Output or GDP in year t, Y(t), is equal to output per effective worker times the

number of effective workers. The relationship in growth terms is as follows:

gr(Y(t)) ¼ grY(t)L(t)

� �þ gr(L(t)), ð7Þ

where gr() represents the growth rate of the argument. Given the growth in output per

effective worker, a 1 percentage point increase (decrease) in the effective number of

workers leads to a 1 percentage point increase (decrease) in the growth of GDP.

The impact of demography on GDP growth is remarkably diverse in Asia and the

Pacific. In 2020, the effective labor force was expected to have grown most rapidly

(between 3.5% and 3.6%) in Afghanistan, Cambodia, and Solomon Islands.

In contrast, the most rapid declines (between �0.5% and �0.6%) were expected in

Georgia, the PRC, and Japan. In most economies, demographic change is currently

providing a tailwind for GDP growth. The effective labor force was expected to grow

in 2020 in all Central Asian economies except Armenia and Georgia, all South Asian

economies, all Southeast Asian economies except Thailand, and all Pacific economies.

East Asia is an exception to this regional pattern. Total effective labor was projected to

decline in 2020 in all East Asian economies except Mongolia. Among the developed

member economies, Australia and New Zealand were also projected to experience

moderate growth in their respective effective labor force (Figure 6).



With rare exception, however, growth in the effective number of workers is

expected to decline over the coming decades, serving as a significant drag on GDP

growth. Some economies that are experiencing a moderate decline now are projected

to experience a much more rapid decline in the future. The rate of growth of the

effective number of workers is projected to decline by more than 1 percentage point

per annum by 2060 in Japan, the Republic of Korea, and Taipei,China. In many

economies, the effective number of workers will continue to grow over the next two

decades but much more slowly than today or in the recent past. The growth rate of the

effective number of workers in Bangladesh, for example, is projected to decline from

1.7% per year in 2020 to 0.3% per year in 2040. The growth rate for the effective

number of workers in India is projected to decline by a full percentage point between

2020 and 2040.

Projections of effective labor for 2040 are somewhat more reliable than other

indicators because everyone 20 years and older has already been born. Projections

Figure 6. Effective Labor Force Annual Rates of Growth in 2000, 2020, 2040, and 2060 (%)

AFG = Afghanistan; ARM = Armenia; AUS = Australia; AZE = Azerbaijan; BAN = Bangladesh; BHU = Bhutan;BRU = Brunei Darussalam; CAM = Cambodia; FIJ = Fiji; GEO = Georgia; HKG = Hong Kong, China; IND =India; INO = Indonesia; JPN = Japan; KAZ = Kazakhstan; KOR = Republic of Korea; KGZ = Kyrgyz Republic;LAO = Lao People’s Democratic Republic; MAL = Malaysia; MLD = Maldives; MON = Mongolia; MYA =Myanmar; NEP = Nepal; NZL = New Zealand; PAK = Pakistan; PNG = Papua New Guinea; PHI = Philippines;PRC = People's Republic of China; SAM = Samoa; SIN = Singapore; SOL = Solomon Islands; SRI = Sri Lanka;TAP = Taipei,China; TAJ = Tajikistan; THA = Thailand; TIM = Timor-Leste; TON = Tonga; TKM =Turkmenistan; UZB = Uzbekistan; VAN = Vanuatu; VIE = Viet Nam.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



beyond 2040 are less reliable due to uncertainty about fertility trends. Accepting UN

projections at face value indicates that the slowdown in the effective number of workers

will continue. The decline is projected to moderate in a few economies—such as

Georgia, Japan, and Taipei,China—that have already experienced a very rapid decline.

Projections of effective number of workers for smaller economies are also subject

to uncertainty to the extent that their future will depend to a large degree on

immigration. Singapore and Hong Kong, China are two obvious cases in point.

V. The First Demographic Dividend: The Adverse Impact of

Population Aging on Growth in Asia

The first demographic dividend refers to the positive effect of demographic

change on standards of living due to a rise in the support ratio; that is, when the

effective number of workers is growing more rapidly than the effective number of

consumers (Bloom and Canning 2001, Bloom and Williamson 1998, Mason 2001,

Mason and Lee 2007). Extending the identities presented in equations (2) and (3),

output per effective consumer is equal to the following:

Y(t)N(t)

¼ Y(t)L(t)

SR(t): ð8Þ

Given the productivity Y(t)=L(t), income per effective consumer varies directly with

the support ratio. In growth terms,

grY(t)N(t)

� �¼ gr

Y(t)L(t)

� �þ gr(SR(t)), ð9Þ

where gr() represents the growth rate of the argument. The growth rate of income per

effective consumer is equal to the growth rate of productivity plus the growth rate of

the support ratio.

The effect of population on the support ratio is referred to as the first

demographic dividend while the effect of population on productivity is referred to as

the second demographic dividend. The rise in the support ratio over the demographic

transition is referred to as the dividend phase.

The first dividend was especially pronounced in East Asia. Using the values for the

median economy as a marker, the support ratio began to increase in the early 1970s and

continued to be favorable for about four decades (Figure 7). The first dividend has now

turned negative as population aging is leading to a decline in the support ratio.

This condition is expected to persist to 2060 and beyond. During its peak years, mostly

in the 1980s and 1990s, the dividend added more than 1% per year to GDP per effective



consumer. At the trough, the decline in the support ratio will be about 1% or less per

year and will persist for two decades before tapering off. Comparing the least and the

most favorable periods, the downturn in economic growth as measured in GDP per

effective consumer due to population change amounts to a swing of 2 percentage points.

The actual and projected first dividends in other regions of Asia and the Pacific

are similar to the East Asia pattern, although the swings are somewhat more moderate.

The median support ratio began to increase in the 1970s everywhere in the region

except in South Asia, where the dividend phase was delayed until the early 1990s.

The rise and fall of the dividend is less pronounced outside of East Asia because

fertility declines were not as rapid in other subregions. The historical and projected

first dividends have been or will be most moderate in the Pacific and developed

member economies. Swings in the support ratio of developed member economies were

generally moderate, although in Japan the impact of aging has already begun and is

substantial.

Figure 7. The First Demographic Dividend in Subregions of Asia and the Pacific, 1970–2060(% Per Year)

Notes: The median economy value is shown as the black dotted line for each subregional grouping. The firstdividend is equal to the rate of growth of the support ratio. See Appendix for more information.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



Some highlights of our findings on the first demographic dividend in Asia and

the Pacific include the following:

. 1950–1999: the effective labor force was growing in every member economy of the

Asian Development Bank except Georgia and Armenia;

. 2000: the growth of the effective labor force ended in Japan; and

. 2016–2017: the growth of the effective labor force ended in the PRC; Hong Kong,

China; the Republic of Korea; Taipei,China; and Thailand.

Based on median values, the dividend phase began and ended in subregions of Asia

and the Pacific at the following times:

. 1968–1999: dividend phase in developed member economies;

. 1970–1975: dividend phase began in Central Asia, East Asia, Southeast Asia, and

the Pacific;

. 1992: dividend phase began in South Asia;

. 2014 and 2018: dividend phase ended in East Asia and Central Asia, respectively;

. 2028 and 2038: dividend phase will end in Southeast Asia and South Asia,

respectively; and

. 2076: dividend phase will end in the Pacific.

VI. Funding the Consumption of the Elderly

Aging is beginning to have important economic implications for many

economies in Asia and the Pacific. The pace of aging is accelerating so the

interactions between aging and the economy will become increasingly important in the

coming decades (Mason and Lee 2018).

The economic impact of aging depends on the institutions and economic

mechanisms on which the elderly rely to fund old-age needs. One possibility is

lifecycle saving. Assets can be accumulated during the surplus ages to fund deficits, by

relying on asset income and dis-saving, during old age. This mechanism is explored in

more detail below. The second possibility is that societies may rely on inter-age

transfers with resources flowing from surplus ages to old-age deficit ages. These

transfers may be private, funded by family members, often between co-resident family

members. Or they may be public, funded through public transfer programs such as

publicly funded health-care programs that support the elderly or through public

pension programs based on pay-as-you-go principles.



Transfers are governed by an iron rule: transfer inflows must be matched by

transfer outflows. Changes in age structure produce an imbalance between inflows and

outflows that requires rebalancing. How large is that imbalance? That depends on the

extent of aging and the importance of transfers in meeting lifecycle needs as

determined by policy and practice in any particular economy. Given the per capita age

profile of consumption and labor income, the maximum imbalance from any change in

age structure would arise if the lifecycle deficit were funded entirely by transfers.

The OAGAP measures contemporaneous resources that would be required to

meet the old-age deficit given the projected population at each age and per capita

consumption and labor income. The old-age deficit is measured as a percentage of total

labor income:

OAGAP(t) ¼ 100�P!

x¼65 P(x, t)(c(x, b)� yl(x, b))Yl(t)

, ð10Þ

where P(x, t) is the population aged x in year t, c(x, b) is per capita consumption, and

yl(x, b) is per capita labor income of persons aged x in the base year. The OAGAP has

a simple interpretation. It is the share of total labor income in year t that would be

required to fund the lifecycle deficit of the elderly in that year. It is also the maximum

tax on labor income that would be required to fund old-age needs entirely. Changes in

the OAGAP over time for any economy quantify the imbalances created by changes in

the population age structure.

The OAGAP values in 2020 for 41 economies in Asia and the Pacific are

presented by subregion in Figure 8. The values are calculated using projected

populations by age in 2020 and consumption and labor income profiles in the base

year (see Appendix for the base year for each economy). The 2020 ratios fall below

3% in a handful of economies: Tajikistan, Afghanistan, Bangladesh, and the Lao

People’s Democratic Republic. The highest values for each subregion are found in

Georgia (Central Asia; 20.9%); Hong Kong, China (East Asia; 21.9%); Sri Lanka

(South Asia; 15.2%); Thailand (Southeast Asia; 13.6%); and Fiji (Pacific; 6.8%).

Among the developed member economies, the highest 2020 OAGAP value belongs to

Japan (43.6%).

Subregional averages for 2020 are calculated as the simple average for

economies in each subregion. Excluding developed member economies (30.4%), the

highest subregional average is in East Asia (12.4%), while in other developing

subregions the values vary from 8.5% to 4.4% (Figure 9).

Aging will produce substantial lifecycle imbalances in developing member

economies over the next four decades. Aging is most advanced in East Asia, with the

average value expected to rise to 41.5% in 2060, an increase of almost 30 percentage



points from 2020. The projected OAGAP values in 2060 are very similar for Central

Asia, South Asia, and Southeast Asia at 21% to 22%, with an increase in the OAGAP

of 13–16 percentage points for each subregion during the review period. The Pacific

economies are aging more slowly, with a projected OAGAP value of only 8.4% in

2060, which is a 4-percentage-point increase compared with 2020. For developed

member economies, the average OAGAP is projected to reach 53.7% by 2060,

reflecting a 24-percentage-point increase compared with 2020.

The bottom line is that maintaining the existing lifecycle profiles in the presence

of population aging would require funding in excess of 40% of labor income in East

Asia and more than 20% in the rest of developing Asia by 2060. This funding could

rely to some extent on transfers but also on asset-based reallocations: asset income,

and dis-saving. To explore this possibility, we consider the implications of aging from

a longitudinal perspective.

Figure 8. OAGAP Values for Economies in Asia and the Pacific in 2020 (%)

Notes: The OAGAP measures contemporaneous resources that would be required to meet the old-age deficitgiven the projected population at each age and per capita consumption and labor income. The old-age deficit ismeasured as a percentage of total labor income.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



VII. Lifecycle Asset Accumulation for Old Age in Asia

Old-age needs can be met relying exclusively on current income through a

combination of public and private transfers. The old-age gap measures presented

above are particularly relevant for assessing transfer programs, but many economies,

particularly in Asia, do not rely heavily on transfer systems to meet old-age needs.

Instead, they rely on assets to fund old-age needs. In this section, we consider the

implications of population aging for asset-based systems relying again on a polar case

where economies rely exclusively on assets and not at all on transfers to fund old-age

needs (i.e., the gap between consumption and labor income at old ages).

Asset-based systems are more complex than transfer systems. During their working

years, each cohort must accumulate the lifecycle wealth on which it will rely during their

retirement phase.We call this the preretirement or accumulation phase. To some extent, the

accumulation of retirement wealth occurs at all working ages. For example, workers may

participate in employment-based funded pension plans. We abstract from this, however,

and assume that the lifecycle surplus for young adults is devoted to supporting children,

while the lifecycle surplus for older adults is devoted exclusively to accumulating

retirementwealth. The accumulation phase is definedby the age range overwhich lifecycle

wealth is positive and rising. See Mason et al. (2017) for more details.

Figure 9. OAGAP Values for the Subregions of Asia and the Pacific in 2020, 2040, and2060 (%)

Notes: The OAGAP measures contemporaneous resources that would be required to meet the old-age deficitgiven the projected population at each age and per capita consumption and labor income. The old-age deficit ismeasured as a percentage of total labor income. Regional values are simple averages of the values for economiesin each region.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



The accumulation phase is followed by the retirement phase. During this phase,

lifecycle wealth is declining but sufficient to meet anticipated retirement needs in the

future. The cohort relies on asset income and dis-saving to fund the old-age lifecycle

deficit. (Note that the peak of lifecycle wealth may occur later than the age at which

the lifecycle deficit first occurs because seniors may rely on only some part of their

asset income to fund the lifecycle deficit. For simplicity, we refer to the retirement

phase with reference to the peak of lifecycle wealth, but this has no bearing on the

analysis.)

Estimates of the retirement wealth that would be required to fund old-age needs

are presented for economies in Asia and the Pacific in 2020 in Figure 10 (see

Appendix for details). There is a wide range in the values. In Japan, lifecycle wealth is

12 times the total labor income, with similarly high values found in Hong Kong,

China; Australia; and New Zealand. In nine additional economies, retirement wealth

ranges from about 5 to 7.5 times the total labor income. Much lower values of

Figure 10. Lifecycle Retirement Wealth as a Percentage of Total Labor Income forAsia and the Pacific in 2020

Lao PDR = Lao People’s Democratic Republic, WR = aggregate lifecycle retirement wealth, YI = labor income.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



retirement wealth are found in seven economies, with the lowest in Tonga (49%) and

the Lao People’s Democratic Republic (44%).

These values might strike people as extraordinarily high when compared, for

example, with capital–output ratios that are often in the 4–5 range in rich countries.

It is important to keep two considerations in mind, however. The values presented here

take labor income as the denominator, which would be about two-thirds of net national

income and even smaller when compared with GDP. In Japan, for example, the ratio of

lifecycle wealth to GDP would be less than 8. Second, economies do not rely

exclusively on assets for their retirement. In the analysis presented elsewhere, we find

that Japan relies on transfer for about half of old-age needs. Applying this value would

bring the demand for retirement assets to fund old-age needs to around four times the

GDP in Japan.

The result for the PRC is also somewhat surprising. Retirement wealth is only

243% of labor income even though the PRC will be aging very rapidly over the

coming decades. Moreover, labor income among the elderly in the PRC is relatively

low. These two factors are outweighed, however, by the PRC’s very low level of

consumption among those in the retirement and preretirement phases.

There are pronounced regional differences in lifecycle wealth across Asia and the

Pacific (Figure 11). Currently, East Asia has the highest subregional average value at

Figure 11. Lifecycle Retirement Wealth as a Percentage of Total Labor Income in2020, 2040, and 2060

WR = Aggregate lifecycle retirement wealth, YI = labor income.Notes: Subregional values are simple averages of values for the member economies.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



561% of total labor income. In Central Asia and Southeast Asia, lifecycle wealth falls

between 300% and 400%, while lower values are found in South Asia (259%) and the

Pacific (149%). Between 2020 and 2060, the retirement wealth relative to labor

income is projected to almost double in Central Asia, East Asia, and the Pacific, and to

more than double in South Asia and Southeast Asia.

In every economy in Asia and the Pacific, retirement assets are projected to rise

much more rapidly than labor income. This would be possible only if an economy’s

total saving rates were sufficient, but how high would saving rates have to be to

achieve the projected wealth paths that underlie the results presented in Figure 11?

And how would those saving rates compare with transfers if an economy were to rely

on transfers rather than assets to fund old-age needs?

The connection between saving and wealth operates under the influence of two

opposing effects. In many economies in Asia and the Pacific, slower growth in the

effective labor force is leading to downward pressure on the rate of growth of labor

income. This will reduce the need for capital widening. A lower saving rate will

maintain the existing ratio of assets to labor income if labor income is growing

more slowly. This effect of slower economic growth is countered by the demand for

capital deepening, an increase in the ratio of asset to labor income, due to population

aging.

The existence of these two effects is easily formalized. We use the following

terminology. Aggregate lifecycle retirement wealth is represented by WR(t), labor

income by Yl(t), saving by S(t), and the rate of growth of total labor income by

gr(Yl(t)):

WR(t þ 1) ¼ WR(t)þ S(t) ð11Þ

and

Yl(t þ 1) ¼ (1þ gr(Yl(t)))Yl(t): ð12Þ

Dividing equation (11) by equation (12) gives us the following:

WR(t þ 1)Yl(t þ 1)

¼ WR(t)þ S(t)(1þ gr(Yl(t)))Yl(t)

: ð13Þ

Letting w ¼ WR=Yl, the ratio of wealth to labor income, and s ¼ S=Yl, the ratio of

saving to labor income, and rearranging terms gives us the following:

(1þ gr(Yl(t)))w(t þ 1) ¼ w(t)þ s(t): ð14Þ



The saving rate required to produce the wealth path from the beginning of period t to

the beginning of period t þ 1 is as follows:

s(t) ¼ gr(Yl(t))w(t)þ Δw(t), where Δw(t) ¼ w(t þ 1)� w(t): ð15Þ

Rearranging the terms, we have

s(t) ¼ gr(Yl(t))þΔw(t)w(t)

� �w(t): ð16Þ

Or letting gr(w(t)) represent the growth rate of the ratio of wealth to labor income, we

have

s(t) ¼ (gr(Yl(t))þ gr(w(t)))w(t): ð17Þ

In general, the demographic transition will lead to a temporarily elevated saving

rate. Both growth terms in equation (17) will be high. Growth in total labor income

will be high because of growth in the effective labor force. Growth in wealth relative to

labor income will be high due to the effects of aging. The growth effects will be

reinforced by the upward trends in wealth. These effects are likely to dissipate over

time, however. Growth in total labor income will decline with growth in the effective

labor force. Growth in the wealth-to-labor-income ratio will decline as the age

distributions of economies stabilize. The outcome in the distant future will depend on

many details that are unknowable. Simulations can be used, however, to consider

possible effects. The results presented in Figure 6 rely on the assumptions and results

presented above. The growth in total labor income is equal to growth in productivity

(1.5% per annum) and growth in the effective labor force (see Figure 6 for growth in

the effective labor force at 20-year intervals). The simulated ratios of wealth to labor

income are based on the same assumptions and methods used in Figures 10 and 11.

The economies of Asia and the Pacific follow a broadly similar pattern of rising

over the demographic transition and then falling as the effects of the demographic

dividend dissipate and the forces of aging set in (Figure 12). Simulated saving for East

Asian economies reached a peak of about 25% of labor income in 2013 and is

projected to decline through the mid-2050s, reaching a low of only about 9% of total

labor income. The simulated saving rate in Southeast Asian economies is similar in

that the peak value also occurs in 2013, but the peak is lower and saving rates remain

at a high plateau for longer. The patterns for Central Asian and South Asian economies

are similar but the simulated saving rate peaks are not reached until the mid-2030s.

The simulated saving rates for the Pacific economies are generally lower, rising slowly,

and do not reach a peak until around 2070.



Simulated savings and the old-age deficits as a share of total labor income are

compared in Figure 12. An important and general lesson is illustrated. Funding

old-age needs by accumulating assets requires a commitment of resources long before

aging sets in. Taking East Asia in 2020 as an example, the gap between consumption

and labor income could be funded with only 12% of total labor income; but in the

same year, accumulating the assets needed to fund old-age needs in the future would

require 23% of total labor income.

There is a natural temptation then to rely on transfer programs rather than asset

accumulation to deal with old-age needs. One impact of this approach, however, is

very evident in Figure 12. By 2033, the resources needed to fund a transfer system in

East Asia would exceed the resources needed to accumulate pension assets. Thereafter,

the resource requirements diverge sharply. By 2050, funding pensions through

transfers would require 38% of total labor income, while the required funding for

pension assets would drop to 10% of total labor income.

Figure 12. Simulated Saving and Old-Age Deficit as a Share of Total Labor Income in theSubregions of Asia and the Pacific, 1975–2060




A second implication of the asset-based approach has been discussed more

extensively. Higher saving rates and greater capital accumulation will lead to higher

productivity and higher standards of living, which is often referred to as the second

demographic dividend. As a practical matter, however, governments are relying on a

combination of public transfers and capital accumulation to fund old-age needs. In the

following section, we consider the impact of aging on the public sector in economies

for which detailed data are available.

VIII. Aging and Its Implications for Asia’s Public Finances

The connection between aging and the public sector is very important in Asia

and the Pacific as it is everywhere. In some economies, the impact of aging on public

finances is of immediate concern. As discussed above, aging is expected to lead to an

increase in the old-age resource gap, and in some economies, governments are heavily

involved in filling that resource gap for their older populations. In economies with

rapid aging and expansive public support for the elderly, the pressure to increase

public spending may outstrip the availability of public resources. In many economies,

the public sector is playing a more limited role. This is due in part to the fact that aging

is delayed and in part to the fact that the public sector is less involved in filling the

resource gap for their senior citizens.

Figure 13. Public Transfer Inflows and Outflows by Age in Indonesia (2005) and theRepublic of Korea (2012)

Notes: Values are expressed relative to the per capita labor income of persons aged 30–49 years.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



The differences in old-age support systems are illustrated using estimates of

public transfer inflows and outflows by age in Indonesia and the Republic of Korea

(Figure 13). The Republic of Korea’s public sector is much larger than Indonesia’s.

At peak ages in the Republic of Korea (early 40s), public transfer outflows are about

40% of the per capita labor income of an adult aged 30–49 years, compared with about

15% for those at their peak age in Indonesia (mid-40s). If we consider the importance

of the public sector based on public transfer inflows, the Republic of Korea’s public

sector is also much larger than Indonesia’s.

The age patterns of inflows also differ markedly between the Republic of Korea

and Indonesia. Inflows to children are higher than inflows to prime age adults in both

countries, although the inflows are especially high in the Republic of Korea compared

with Indonesia and almost every other economy in the world. Of particular importance

to the impact of aging is the very large difference in public transfer inflows to seniors.

In Indonesia, we see virtually no tendency for inflows to be higher at older ages;

neither major spending on pensions nor that on health lead to higher support for the

elderly. The Republic of Korea, on the other hand, has a pattern that is similar to many

high-income economies: substantial public transfer inflows to the elderly.

Age profiles like those presented in Figure 13 are used below to project the

impact of population aging on public finances in the absence of public sector reform

for eight developing member economies—Cambodia, the PRC; India; Indonesia; the

Philippines; the Republic of Korea; Taipei,China; and Thailand—and two developed

Figure 14. Old-Age GAP Ratio as a Percentage of Total Labor Income for 10 Economiesin Asia and the Pacific in 2020, 2040, and 2060

Notes: The old-age GAP ratio is the difference between consumption and labor income for those 65 years andolder. Values are projected using per capita consumption and labor by age and population projections.Source: Authors’ calculations based on the National Transfer Accounts, www.ntaccounts.org (accessed 15 March2020).



member economies—Australia and Japan. Before those results are presented,

however, we present our earlier estimates of the impact of aging on the old-age gap

as a ratio of total labor income in those economies for 2020, 2040, and 2060. These

results show the maximum impact of aging on public sector spending irrespective of

the public policy pursued (Figure 14).

The OAGAP is rising sharply in three developing member economies: the

Republic of Korea, Thailand, and Taipei,China. By 2060, resources needed by the

elderly in those economies will range from 40% of total labor income in Thailand to

almost 70% in Taipei,China. The percentage of resources needed in Australia by 2060

will be less at about 38%, while in Japan the OAGAP is projected to reach 76% of

total labor income.

In the other five developing economies in Asia and the Pacific—Cambodia, the

PRC, India, Indonesia, and the Philippines—the maximum impact of aging on public

transfer systems is much smaller. But the increases are by no means inconsequential.

The old-age GAP ratios will be three–four times as large in 2040 as they were in 2020.

Still, the results suggest that very generous old-age support systems in these

economies are an attractive option. Expansive programs may be feasible in the shorter

term, but they could well prove to be unsustainable in the more distant future

depending on how they are designed.

In Figure 15, we consider projections of net public transfers if the normalized

age profiles of inflows and outflows are held constant relative to labor income of

those aged 30–49 years. Of the five economies experiencing medium-level gains in the

Figure 15. Projected Net Public Transfers to Those Aged 65 Years and Older as aPercentage of Total Labor Income in 2020, 2040, and 2060




old-age GAP ratio, none but the PRC will experience significant upward pressure on

net public transfers. By 2060, the projected value of net public transfers to the elderly

will increase to 1% of total labor income in Cambodia and by less in the other four

economies. In the Philippines, net transfers to seniors are projected to be negative,

with seniors receiving less in benefits than they pay in taxes. The PRC is an exception

among these economies with net public transfers projected to reach between 10% and

15% of total labor income by 2060.

The importance of policy in conjunction with aging is also apparent in the

projections for the Republic of Korea; Taipei,China; and Thailand. Net public transfers

to the elderly are relatively modest in Thailand and, hence, the projected values in

2060 reach only 4% of the total labor income. The increases are much more substantial

in the Republic of Korea and Taipei,China, where the net public transfers are projected

to reach 15% and 25%, respectively, in 2060. In Japan, the projected increases in net

public transfers to the elderly are very substantial, reaching 45% of total labor income

by 2060.

IX. Conclusions

For many decades to come, the landscape in Asia and the Pacific will be

dominated by two demographic trends: slower population growth and population

aging. The region is diverse and, hence, the timing and severity of these trends will

vary considerably from economy to economy. However, these trends will intensify, not

dissipate, with significant implications for economies over the foreseeable future.

Demographic transition will pose two strategic challenges for the region: maintaining

economic growth in the face of less favorable demographics and securing adequate

resources to meet the consumption needs of the elderly.

By and large, our analysis confirms the conventional wisdom that population

aging will adversely affect Asia’s economic growth. GDP and other aggregate

economic indicators can be expected to grow more slowly in the future. In the past,

recent growth in the effective labor force led to more rapid GDP growth, but this will

not be the case in the future. In some instances, where fertility rates are very low, the

effective labor force may shrink substantially and GDP may actually decline.

Per capita income could also grow more slowly because of a decline in the support

ratio. The effective number of workers is expected to grow more slowly than the

effective number of consumers.



Our analysis also underlines the challenges of securing adequate resources for

the region’s large and growing elderly population. Generational features of economies

in Asia and the Pacific are in the midst of change. Other things being equal, aging will

lead to a very substantial increase in the resources required to support seniors. The

implications of this change are very complex and difficult to anticipate. Some of the

most important effects may be felt at the family level rather than the aggregate level.

Providing support to the elderly could prove to be a significant financial burden to

prime-age adults, particularly in societies where spending on children is reaching high

levels. In many families, the financial costs may prove to be less important than the

burdens from caregiving that often fall on women.

In some societies, providing for seniors is less a family responsibility and more a

social responsibility. In most of the region’s economies, public transfers to the elderly

are less important than in many European and Latin American countries. In those

places where public transfer systems are important, tax revenues are expected to rise

much more slowly than promised benefits. Either tax rates will have to rise or benefits

will have to be curtailed, or both. Other economies may hope to implement more

extensive old-age support systems in response to the greater needs of seniors in aging

societies. A cautious approach is surely warranted.

Many working-age adults may accumulate assets in anticipation of their

retirement. They may participate in funded pension programs, buy a home, pay down

debt, or accumulate assets in many other ways. For now, elevated saving rates will be

essential to fund future retirement needs. Eventually, however, accumulating assets

will ease the financial burden imposed by an aging society.

“Demographics is destiny” refers to the widely held pessimistic view that

economies are powerless in the face of population aging; that is, the demographic

transition to older populations is often blamed for the loss of economic dynamism.

However, there are many things that Asian policy makers can do to mitigate the impact

of population aging on economic growth and old-age economic security. For example,

governments can invest more in education and human capital so that higher labor

productivity can mitigate the reduction of the first demographic dividend. Another

example is creating a working environment that enables seniors to remain productive

for a longer time. This will contribute to both old-age economic security and economic

growth. In short, public policy is hardly impotent in the quest to achieve a more benign

demographic destiny in Asia and elsewhere.



Appendix. Methodology

A1. Consumption and Labor Income Profiles

Consumption and labor income profiles have been estimated by members of the

NTA network using methods described in the United Nations’ National Transfer

Accounts Manual: Measuring and Analyzing the Generational Economy (United

Nations Population Division 2013). These members belong to the Asia Regional

Group led by Sang-Hyop Lee, and their work has relied heavily on financial support

from the United Nations Population Fund and their home institutions.

Labor income consists of wages and earnings of employees and estimates of the

value of labor of the self-employed and unpaid family workers. The age profile is

affected by age variation in labor force participation, unemployment, hours worked,

and productivity. Consumption consists of both public and private consumption, with

separate estimates for health, education, and all other consumption expenditure

combined.

All estimates are based on surveys and administrative data. They are adjusted to

match aggregate data from the System of National Accounts.

Data are available for the following economies included in Table A1, as well as

the year for which data have been estimated.

Age profiles have been rescaled to match the most recently available System of

National Accounts data, but important changes may have occurred in the age patterns

of estimates in some economies. More recent estimates for Japan (not yet released by

the Statistics Bureau) show considerable stability in the age profiles. Efforts are

underway to update estimates for India, but they are not yet completed.

Table A1. Data Availability by Income Group

High Income Upper-Middle Income Lower-Middle Income Low Income

Australia (2010) People’s Republicof China (2014)

Bangladesh (2010) Nepal (2011)

Japan (2004) Malaysia (2009) Cambodia (2009)

Republic of Korea (2015) Maldives (2016) India (2004)

Singapore (2013) Thailand (2013) Indonesia (2012)

Taipei,China (2015) Lao People’s DemocraticRepublic (2012)

Mongolia (2014)

Philippines (2015)

Timor-Leste (2011)

Viet Nam (2012)

Source: Authors’ compilation.



A2. Lifecycle Retirement Wealth

An economy relies on lifecycle wealth, broadly defined, to fund old-age needs.

For a cohort of age x in year t, lifecycle wealth is equal to the present value of

prospective consumption less the present value of prospective labor income of

members of that cohort over the remainder of their lifetimes. The accumulation

of lifecycle wealth can be divided into two phases: the retirement phase and the

preretirement (or accumulation) phase. During the retirement phase, lifecycle wealth

will be declining as members of the cohort rely on their wealth to fund consumption in

excess of labor income. During the preretirement phase, lifecycle wealth is rising,

being accumulated as members of the cohort approach retirement. The cohort’s

lifecycle wealth will depend on the per capita consumption and labor income at each

age and the number of cohort members who are still alive.

For members of the cohort, wealth consists in part of assets accumulated through

saving or from bequests received in previous years. But wealth also consists of the

value of prospective net transfers. Prospective public transfers have a value determined

by the prospective benefits received from public programs less the prospective taxes

paid to support public programs. Likewise, the prospective private transfers have a

value that depends on the prospective private transfers received less the prospective

private transfers given. We refer to these two forms of lifecycle wealth as assets and

transfer wealth.

Assets and transfer wealth are equivalent as a means of meeting lifecycle needs.

In other respects, they are quite different, however. Assets are invested and lead to an

increase in capital, a rise in labor productivity, and possibly a decline in interest rates.

Transfer wealth is not invested and consists of nothing more than an obligation of

future generations to transfer resources to the cohort in question. Indeed, transfer

wealth must be balanced by transfer debt, including the net obligations of future

generations.

Formally, per capita lifecycle wealth for the cohort aged x in year t, wr(x, t), for x

equal to or greater than the beginning of preretirement (xp), is equal to the present

value of consumption less the present value of labor income,

wr(x, t) ¼ PVc(x, t)� PVyl(x, t) for x � xp,

PVcr(x, t) ¼X!�xp

z¼0

(1þ r)�zc(xþ z, t þ z) for x � xp,

PVylr (x, t) ¼X!�xp

z¼0

(1þ r)�zyl(xþ z, t þ z) for x � xp,

ðA1Þ



where c and yl are prospective per capita consumption and labor income, respectively,

over the remaining lifetime of the cohort aged x in year t. Total lifecycle wealth is

equal to the following:

WR(t) ¼ PVCR(t)� PVYlR (t),

PVCR(t) ¼X!x¼xp

P(x, t)PVcr (x, t),

PVYlR (t) ¼X!x¼xp

P(x, t)PVylr (x, t),

ðA2Þ

where WR(t) is the total lifecycle wealth in year t; PVCR(t) is the total prospective

consumption and PVYlR (t) is the total prospective labor income, both in present value

terms in the year t; and P(x, t) is the population of age x in year t. All values are

calculated for preretirement and retirement age cohorts.

The calculations presented here are based on the consumption and labor income

profiles for each economy, an exogenous rate of growth of the profiles by 1.5% per

annum, and a discount rate of 3% per annum. More detailed information about the

calculation of the lifecycle phases and the values of xp calculated for each economy is

available in Mason et al. (2017).

A3. Accumulating Wealth and Saving

In this subsection, we consider the connection between savings in wealth under

the influence of two opposing effects. Aging is leading to an increase in wealth relative

to total labor income, but at the same time, slower growth in labor income has the

opposite effect. Here is the derivation using the following terminology: lifecycle

wealth isW(t) and labor income is Yl(t), the rate of growth of total labor income is g(t),

and S(t) is the net saving of lifecycle wealth during the year t,

W(t þ 1) ¼ W(t)þ S(t): ðA3ÞLet g(t) be the growth rate of total labor income:

Yl(t þ 1) ¼ (1þ g(t))Yl(t): ðA4Þ

Dividing equation (A3) by equation (A4) gives the following:

W(t þ 1)Yl(t þ 1)

¼ W(t)þ S(t)(1þ g(t))Yl(t)

: ðA5Þ



Letting w ¼ W=Yl, the ratio of wealth to labor income, and s ¼ S=Yl, the ratio of

saving to labor income, and rearranging the terms, we have

(1þ g(t))w(t þ 1) ¼ w(t)þ s(t): ðA6ÞThe saving rate required to produce the wealth path from the beginning of period t to

the beginning of period t þ 1 is as follows:

s(t) ¼ g(t)w(t)þ Δw(t), where Δw(t) ¼ w(t þ 1)� w(t): ðA7ÞRearranging the terms, we have

s(t) ¼ g(t)þ Δw(t)w(t)

� �w(t): ðA8Þ

Or letting gr(w(t)) represent the growth rate of the ratio of wealth to labor income, we

have the following:

s(t) ¼ (g(t)þ gr(w(t)))w(t): ðA9Þ

References

Balachandran, Arun, Joop de Beer, K. S. James, Leo van Wissen, and Fanny Janssen. 2019.“Comparison of Population Aging in Europe and Asia Using a Time-consistent andComparative Aging Measure.” Journal of Aging and Health 32 (5–6): 340–51.

Becker, Gary S., Kevin M. Murphy, and Robert S. Tamura. 1990. “Human Capital, Fertility,and Economic Growth.” Journal of Political Economy 98 (5): S12–37.

Bloom, David E., and David Canning. 2001. “Cumulative Causality, Economic Growth, andthe Demographic Transition.” In Population Matters: Demographic Change, EconomicGrowth, and Poverty in the Developing World, edited by Nancy Birdsall, Allen C. Kelley,and Steven W. Sinding, 165–200. Oxford: Oxford University Press.

Bloom, David E., and Jeffrey G. Williamson. 1998. “Demographic Transitions and EconomicMiracles in Emerging Asia.” The World Bank Economic Review 12 (3): 419–56.

de la Croix, David. 2017. “Did Longer Lives Buy Economic Growth? From Malthus to Lucasand Ben-Porath.” In Demographic Change and Long-Run Development, edited byM. Cervellati and U. Sunde, 69–89. Cambridge, MA: The MIT Press.

Eggertsson, Gauti B., Manuel Lancaster, and Lawrence H. Summers. 2019. “Aging,Output Per Capita, and Secular Stagnation.” American Economic Review: Insights1 (3): 325–42.

Ehrlich, Isaac, and Yong Yin. 2013. “Equilibrium Health Spending and Population Aging in aModel of Endogenous Growth: Will the GDP Share of Health Spending Keep Rising?”Journal of Human Capital 7 (4): 441–47.



Lee, Ronald, and Andrew Mason. 2010. “Fertility, Human Capital, and Economic Growth overthe Demographic Transition.” European Journal of Population 26 (2): 159–82.

Lee, Ronald, and Andrew Mason, eds. 2011. Population Aging and the Generational Economy:A Global Perspective. Cheltenham: Edward Elgar.

Lee, Ronald, Andrew Mason, and Members of the NTA Network. 2014. “Is Low FertilityReally a Problem? Population Aging, Dependency, and Consumption.” Science 346(6206): 229–34.

Mason, Andrew, ed. 2001. Population Change and Economic Development in East Asia:Challenges Met, Opportunities Seized. Stanford: Stanford University Press.

Mason, Andrew, and Ronald Lee. 2007. “Transfers, Capital, and Consumption over theDemographic Transition.” In Population Aging, Intergenerational Transfers and theMacroeconomy, edited by Robert Clark, Andrew Mason, and Naohiro Ogawa, 128–62.Cheltenham, United Kingdom and Northampton, United States: Elgar Press.

_____. 2018. “Intergenerational Transfers and the Older Population.” In IntergenerationalTransfers and the Older Population, edited by The National Academies Press, 187–214.Washington, DC: National Academies of Sciences, Engineering, and Medicine.

Mason, Andrew, Ronald Lee, Michael Abrigo, and Sang-Hyop Lee. 2017. “Support Ratios andDemographic Dividends: Estimates for the World.” United Nations Population DivisionTechnical Paper No. 2017/1.

Sanderson, Warren, and Sergei Scherbov. 2010. “Remeasuring Aging.” Science 329 (10):1287–88.

Tobin, James. 1967. “Life Cycle Saving and Balanced Economic Growth.” In Ten EconomicStudies in the Tradition of Irving Fisher, edited by William Fellner, 231–56. New York:Wiley.

United Nations Population Division. 2013. National Transfer Accounts Manual: Measuringand Analyzing the Generational Economy. New York: United Nations.

_____. 2019. World Population Prospects: The 2019 Revision. New York: United Nations.



Trends in Employment and Wagesof Female and Male Workers in India:

A Task-Content-of-Occupations Approach

SHRUTI SHARMA

This paper uses the task-content-of-occupations framework to analyze trends inemployment and wages of female and male workers in the Indian labor marketfrom 1994 to 2017. Workers are classified into four main occupationalcategories: nonroutine cognitive, routine cognitive, nonroutine manual, androutine manual. Decomposing the changes in employment shares intobetween-industry changes and within-industry changes across occupationalcategories reveals that within-industry employment changes have increasinglyplayed an important role, suggesting the growing importance of using thetask-content framework to analyze labor market trends. The biggest increase inemployment shares is for nonroutine cognitive occupations for both female andmale workers. The wage analysis reveals that, on average, the gender wage gaphas been lowest in routine cognitive occupations for most of the period ofanalysis. However, the analysis finds no consistent, significant changes inwages based on occupational specialization during the period of analysis.

Keywords: employment, gender, occupations, wages, tasks

JEL codes: J20, J24, J30, J31

Shruti Sharma: Department of Social Sciences, Human Services and Criminal Justice, Borough ofManhattan Community College, City University of New York. Email: [email protected]. Supportfor this project was provided by a PSC-CUNYAward, jointly funded by The Professional Staff Congressand The City University of New York. I am grateful to the participants of the Faculty FellowshipPublication Program 2020 (City University of New York), the managing editor, and two anonymousreferees for very helpful suggestions on this paper.





https://dx.doi.org/10.1142/S0116110522500020

https://dx.doi.org/10.1142/S0116110522500020

I. Introduction

Labor markets in both developed and developing countries have been

significantly impacted by technological innovations and advancements (James 1999;

Autor, Levy, and Murnane 2003; Bruckner, LaFleur, and Pitterle 2017). India, since its

economic reforms in the 1990s, has also experienced an increase in technological

adoption brought about by substantial growth of its technology sector (Kapur 2002,

Heeks 2015) as well as trade-induced technological transfers (Goldberg et al. 2010,

Topalova and Khandelwal 2011, Bas and Berthou 2012, Sharma 2018). It is therefore

important to incorporate the likely impacts of such technological developments in

analyzing India’s labor market.

Traditional literature that examines the impact of technology on workers focuses

on the skills of workers, hypothesizing that the impact of technological advancement

on the labor force should be skill biased (Tinbergen 1974, 1975, Bound and Johnson

1992, Katz and Murphy 1992). However, Acemoglu and Autor (2011) develop a

framework that focuses on tasks involved in various occupations. These are

distinguished along two dimensions—whether they are routine or not (which captures

the differential effects of technology adoption), and whether they are cognitive or not

(which captures the effects of skills or education). Accordingly, they divide

occupations into four categories—routine manual, routine cognitive, nonroutine

manual, and nonroutine cognitive. The hypothesis is that technology is likely to

displace routine tasks, whether these are cognitive or not. In their paper, they find

evidence of job polarization for the United States (US)—a decline in the share of

“middle-skilled” occupations, which are mainly routine occupations, both cognitive

and manual, while the share of workers in nonroutine cognitive and nonroutine manual

occupations increased. This framework performs better than previous models in

explaining the trends in employment and wages of US workers in recent decades

(Acemoglu and Autor 2011, Autor and Dorn 2013).

While the Indian context is very different compared to a developed country such

as the US when it comes to technological adoption, it is nonetheless useful to

understand labor market trends from the perspective of the task content of various

occupations given the increase in technological adoption and advancement in India in

the period after liberalization (Kapur 2002, Heeks 2015). This paper applies the

task-content-of-occupations framework to analyze the Indian labor market separately

for female and male workers and to examine the differences between them. One can

expect technology adoption (captured by whether occupations are routine or

nonroutine) and education (captured by whether occupations are cognitive or



noncognitive) to impact gender inequality in labor markets. There is empirical

evidence that adoption of technology can reduce gender wage gaps. Juhn, Ujhelyi, and

Villegas-Sanchez (2014) show that technological adoption through trade liberalization

in Mexico reduced the demand for “brawn-intensive skills” and favored employment

of blue-collar female workers. There is also evidence that female workers with

cognitive skills earn higher wages and suffer from lower gender wage gaps than those

with noncognitive skills, although education does not entirely reduce the gender wage

gap (Gunewardena, King, and Valerio 2018). This paper studies the trends in

employment and wages of female and male workers separately to capture the

differences in average trends that emerge from using a task-content-of-occupations

framework.

To examine these trends, the paper uses data on India’s labor force for the years

1994, 2000, 2005, 2008, and 2012 from the National Sample Survey of India, and for

the year 2017 from the Periodic Labor Force Survey. The data provide detailed

information on the principal activity, occupational code, wages, gender, age,

education, region, and industry of workers. To create occupational categories, as in

Acemoglu and Autor (2011), information on the task content of each occupation was

obtained from the Occupational Network (O�NET) database and merged onto the

main worker-level dataset.

Employment analysis reveals that, on average, for both female and male

workers, the share of nonroutine cognitive occupations increased the most during

the period of analysis, mainly at the cost of the share of occupations in nonroutine

manual occupations. This was followed by an employment decomposition exercise,

which further revealed interesting findings. Examining whether changes in

employment shares are mainly driven by changes in industrial structure (referred

to as “between-industry” changes hereinafter) or by changes in the occupational

shares within industries (referred to as “within-industry” changes hereinafter), I find

that in the more recent decade (2005–2017), within-industry changes in occupational

shares primarily influenced total changes across all occupational categories. In the

previous decade (1994–2005), however, changes in industrial structure played a more

important role in determining total change in employment shares. This shows that

within-industry changes in occupational categories are becoming increasingly

important in determining changes in employment shares of occupations, which

suggests that the demand and supply of workers performing certain tasks within

industries have been increasingly responsible for shifts in employment trends. In

particular, within-industry changes in nonroutine cognitive shares increased

significantly between the two periods and contributed to an overall increase in the

TRENDS IN EMPLOYMENT AND WAGES OF FEMALE AND MALE WORKERS IN INDIA 171


employment share for both female and male workers. However, there are some

differences across female and male workers across occupational categories. During

2005–2017, while within-industry declines in employment shares for male routine

manual workers was the primary factor behind the changes in total employment shares,

the same was not true for female workers in routine manual occupations. In fact, the

employment share of routine manual female workers increased whereas male workers

experienced a decline in this category. This could be indicative of the fact that

routinization, especially inmanual occupations, affectsmaleworkersmore severely and

could serve to reduce the perceived differences between the abilities of female and male

workers. Overall, these findings emphasize the importance of within-industry changes

in occupational structures in determining trends in employment and occupational

structures in India.

Examining trends in wages across occupational categories also yield interesting

findings. Average real wages and relative wages (measured as the ratio of average

wages earned by female workers to male workers) increased during the period of

analysis. Workers in nonroutine cognitive occupations earned the highest wages

throughout the period, followed by those employed in routine cognitive occupations,

for both male and female workers. The analysis also reveals that average wages earned

by routine manual workers are higher than those of nonroutine manual workers. When

considering relative wages, the gender wage gap on average is lowest for almost the

entire period for routine cognitive occupations, followed by nonroutine cognitive

occupations. However, examining how wages have changed over time based on the

occupational specialization of the workers after controlling education, age, and region

fixed effects reveals that there have been no significant changes in trends across all

occupational categories. This could be possible because within-industry occupational

shares in employment have only recently started playing an important role. Thus, these

changes might become more prominent in the future.

The paper is divided into six sections. A description of the datasets used in the

analysis and the task-content model is described in Section II. Section III presents

summary statistics and trends in employment and wages of workers after classifying

them into four categories based on the task-content model. A decomposition of the

percentage change in employment across these categories into between-industry and

within-industry changes can be found in Section IV. An analysis of changes in wages

of male and female workers based on their occupational category is presented in

Section V. Section VI concludes.



II. Data

This paper uses data from the employment-unemployment rounds of the National

Sample Survey (NSS) of India and the Periodic Labor Force Survey (PLFS). These are

household-level datasets that provide information on the activities of all household

members. This includes whether they are employed, their occupation, industry of

employment, wages, age, educational qualification, gender, and the region they are

employed in. Both datasets use a sampling methodology that is representative of the

Indian labor force, and they are available as repeated cross sections. Data for years

1994, 2000, 2005, 2008, and 2012 are from the NSS, and data for the year 2017 are

from the PLFS. The last NSS employment-unemployment round available is from the

year 2012, and labor force data are only available in the form of the PLFS since then.

All the rounds used in this study are thick rounds.1 Occupation data are reported using

a three-digit National Classification of Occupations (NCO) code, and industry data are

reported using a four-digit National Industrial Classification (NIC) code. For the

purpose of the study, any member who reported an occupation code as a current

employment activity is considered employed. Workers who did not report an industry

of employment have been dropped from the analysis.2 The data include both full-time

and part-time workers as well as workers from both the formal and informal sectors.

Salaried as well as self-employed workers and wage laborers are included.

The industry classifications change over time—NIC-1987 for the year 1994,

NIC-1998 for the years 2000 and 2005, NIC-2004 for the year 2008, and NIC-2008

for the years 2012 and 2017, and concordances used from the Ministry of Statistics

and Programme Implementation were used to map all classifications to the NIC-2004

classification. The occupation classifications also change over time—NCO-1968 for

the years 1994, 2000, and 2005, and NCO-2004 for the years 2008, 2012, and 2017,

and a concordance from the Ministry of Labor and Employment was used to map

NCO-1968 to NCO-2004 classification.

Data on wages are reported as weekly wages in Indian rupees. Real wages are

obtained by using a deflator from the consumer price index data of the Organisation

1Thick rounds consider a large sample of households (which ranges from roughly 100,000 to125,000 in this dataset) and are conducted approximately every five years. Thin rounds, on the other hand,consider 35% to 40% of thick round samples and are conducted in the intervening years.

2As a result, a total of 1.2% workers were dropped over six years. Of the workers excluded from theanalysis, 43% were in routine manual occupations whereas 50% were in nonroutine manual occupations.Of the workers dropped in routine manual occupations, 73% were male workers while 27% were female.For nonroutine manual workers, this breakdown was 87% male workers and 13% female workers.



for Economic Co-operation and Development (OECD 2021). I consider the wages

reported for the current activity of the workers for each year of the survey.3

To obtain the task content of occupations, this paper follows Acemoglu and

Autor (2011). The O�NET database version 20.3 from April 2016 is used to obtain

this information. This dataset provides the task content of occupations using various

descriptors—variables that describe various tasks involved in each occupation—with

values assigned along different scales for each occupation. I consider a subset of

descriptors to classify O�NET occupations into nonroutine cognitive, routine

cognitive, nonroutine manual, and routine manual occupations. The descriptors are

based on the abilities, skills, work context, and work activities used for each

occupational category.

The descriptors used in this study are provided in Table A1 of Appendix.4 The

value for the “importance” of each descriptor for all occupations is normalized, and the

values across all descriptors for each occupational category are added up to obtain a

score. This score is then normalized, and each occupation is classified as belonging to

an occupational category based on which occupational category has the highest

normalized score for that occupation code. For instance, a high normalized value for

the descriptor “the amount of time spent making repetitive motions” would suggest

that an occupation is intensive in routine manual tasks, whereas a high normalized

value for the descriptor “thinking creatively” would mean that an occupation is

intensive in nonroutine cognitive tasks. Once we obtain this classification, data from

O�NET are then merged onto the NSS data. This is accomplished by first using a

concordance from the O�NET occupation classification to the ISCO-1988 occupation

codes (which is the same classification used by NCO-2004). For years in which NSS

data report occupations using NCO-1968 codes, a concordance between NCO-1968

and NCO-2004 is used. Table A2 of Appendix shows a mapping between the

one-digit NCO division and the percentage of three-digit NCO occupations in that

division classified as either routine manual, nonroutine manual, routine cognitive, or

3The wages are reported as weekly earnings. I use the “principal activity” occupation code specifiedfor individual persons in the dataset to classify the workers in their occupational categories. The wages arereported for the “current activity” of a person, which could have multiple entries. In 97% of the cases, thereis only one current activity for a person, and in all these cases, it matches the principal activity. For theremaining 3% of the cases where there are multiple entries for the current activity, I sum up the wages forall the current activities. These are mainly workers classified as “Agriculture and Fishery workers” or as“Elementary occupations.” However, all current activities for these workers fall within nonroutine manualoccupations.

4Unlike Sharma (2016), which uses an interaction of level and importance values reported to obtainthe score for each task category, I consider only the “importance” of each occupational category based onAcemoglu and Autor (2011) to match their analysis. This explains why Sharma (2016) finds only threemain occupational categories, whereas this study finds four main occupational categories.



nonroutine cognitive. This can provide an understanding of how one-digit NCO codes

roughly match occupational categories.

The following section describes the trends in employment and wages for female

and male workers based on this classification of occupations.

III. Summary Statistics

In analyzing the trends in female and male workers in India, it is important to

first understand the labor force participation rates. Figure 1 shows the labor force

participation rates from 1994 to 2017. Labor force participation rate for male workers

was 84.7% in 1994 and declined to 83% in 2017. For female workers, the labor force

participation rate was 30.5% in 1994 and declined to 22.2% in 2017.

In terms of gender composition of the workforce, Figure 2 shows that female

workers comprised 19% of the workforce in 2017, down from 23% in 1994, whereas

the share of male workers increased to 81% in 2017 from 77% in 1994.

The next set of statistics is based on examining the task content of occupations,

which would help identify occupations that are experiencing the biggest changes in

employment shares.

Table 1 reports summary statistics for employment of all workers and male and

female across all occupational categories. I compute the shares of workers in

nonroutine manual, nonroutine cognitive, routine manual, and routine cognitive

occupations for both male and female workers (Table A3). Figure 3 presents the

trends in employment obtained from this classification. I find that for both male and

female workers, the share of workers employed in nonroutine manual occupations5 is

the highest, although it has been decreasing over time. For male workers, this share

declined from 81% in 1994 to 72% in 2017, whereas for female workers, the share

declined from 80% in 1994 to 65% in 2017. For both female and male workers, the

biggest increase in employment shares is for nonroutine cognitive workers. For routine

cognitive occupations, on the other hand, the share of male workers employed in this

category decreased from 5% in 1994 to 4% in 2017, while for female workers, this

share increased from 2% in 1994 to 3% in 2017. Finally, for routine manual

occupations, employment shares in this category increased from 18% in 1994 to 21%

5Unlike Acemoglu and Autor (2011), I do not drop occupations in the agriculture sector, all ofwhich classified as nonroutine occupations, to be able to provide an analysis of all possible sectors.Mechanization in the agriculture sector is not properly captured in the classification of occupations asroutine and nonroutine, which might explain why this analysis presents a decline in the share of nonroutinemanual occupations.



in 2017 for female workers; however, for male workers, this share declined from 12%

in 1994 to 10% in 2017. This reflects the share in the economy and growth of India’s

manufacturing sector (Sharma and Singh 2013) where most routine manual

occupations are concentrated.

Figure 1. Labor Force Participation Rate

Source: Author’s calculations using data from the National Sample Survey of India (1994, 2000, 2005, 2008,2012) and the Periodic Labor Force Survey (2017).

Figure 2. Share of Employment by Gender




The next set of summary statistics considers trends in average and relative wages.

Wages are weekly and in Indian rupees, and have been deflated using the consumer

price index (OECD 2021). Figure 4 shows that average real wages for both male and

female workers increased from 1994 to 2017, and Figure 5 shows that relative wages

Table 1. Summary Statistics: Employment (million)

1994 2000 2005 2008 2012 2017

AllMale 151.0 186.0 215.0 214.0 227.0 250.0Female 45.1 54.2 70.0 57.1 56.6 59.5

Routine manualMale 14.0 21.3 24.6 25.6 23.5 25.7Female 6.9 10.5 13.1 11.9 11.5 12.5

Routine cognitiveMale 7.4 8.8 8.4 9.3 9.0 10.2Female 0.84 1.1 1.3 1.1 1.2 1.6

Nonroutine manualMale 122.0 144.0 168.0 154.0 166.0 179.0Female 36.0 40.5 52.9 40.0 39.0 38.8

Nonroutine cognitiveMale 7.4 11.8 14.0 25.4 28.8 35.7Female 1.4 2.1 2.7 4.1 4.9 6.7

Source: Author’s calculations using data from the National Sample Survey of India (1994,2000, 2005, 2008, 2012) and the Periodic Labor Force Survey (2017).

Figure 3. Share of Employment across Occupational Categories




Figure 4. Average Wage by Gender


Figure 5. Relative Wage

Note: Relative wage is measured as the ratio of the average wage of female workers to the average wage of maleworkers.Source: Author’s calculations using data from the National Sample Survey of India (1994, 2000, 2005, 2008,2012) and the Periodic Labor Force Survey (2017).



(measured as the ratio of average wages of female workers to average wages of male

workers) have also increased for the period of analysis.

Table 2 presents the summary statistics for the average weekly real wages earned

by all workers, male workers, and female workers. This is provided for the entire labor

Table 2. Average Weekly Real Wages across Occupational Categories (�)

All Male (M) Female (F) Relative Wage (F/M)

1994

All 858.96 948.92*** 565.13*** 0.60(2.93) (3.41) (5.16)

Routine manual 956.09 1,031.77*** 362.99*** 0.35(5.79) (6.14) (8.55)

Routine cognitive 1,314.12 1,321.55** 1,264.80** 0.96(7.65) (8.13) (22.44)

Nonroutine manual 427.42 488.19*** 285.08*** 0.58(1.56) (1.99) (1.82)

Nonroutine cognitive 2,011.37 2,113.09*** 1,698.37*** 0.80(10.07) (11.62) (18.93)

2000

All 1,096.38 1,203.20*** 748.24*** 0.62(4.03) (4.68) (7.36)

Routine manual 1,180.30 1,241.38*** 530.80*** 0.43(7.59) (8.00) (14.51)

Routine cognitive 1,730.15 1,748.76*** 1,606.92*** 0.92(10.85) (11.49) (32.28)



2005

All 1,135.25 1,246.88*** 771.20*** 0.62(4.27) (4.96) (7.77)

Routine manual 1,075.50 1,138.43*** 459.45*** 0.40(7.04) (7.50) (11.27)




Continued.



Table 2. Continued.

All Male (M) Female (F) Relative Wage (F/M)

2008

All 1,135.48 1,228.93*** 819.10*** 0.67(4.24) (4.91) (7.95)

Routine manual 1,163.70 1,210.86*** 574.10*** 0.47(7.60) (8.00) (15.01)




2012

All 1,599.33 1,701.22*** 1,195.20*** 0.70(6.48) (7.30) (13.38)

Routine manual 1,286.59 1,335.65*** 688.38*** 0.52(8.43) (8.90) (16.57)




2017

All 1,859.17 1,940.36*** 1,431.70*** 0.74(4.70) (5.05) (12.10)

Routine manual 1,606.90 1,708.99*** 730.83*** 0.43(7.52) (7.86) (15.12)


Nonroutine manual 1,283.83 1,364.15*** 781.53*** 0.57(5.04) (5.48) (9.71)


�¼ Indian rupee.Notes: Standard errors in parentheses. ** and *** indicate that the means of weekly real wagesfor male and female workers are significantly different at p < 0:05 and p < 0:01, respectively.Source: Author’s calculations using data from the National Sample Survey of India (1994,2000, 2005, 2008, 2012) and the Periodic Labor Force Survey (2017).



force as well as by occupation type. The data show that there exists a gender wage gap

on average and across all occupation types, with wages earned by male workers

significantly higher than that of female workers across all years.

Figure 6 presents the average log real wages for female and male workers across

occupational categories. For both female and male workers, average trends reveal that

wages are highest for workers in nonroutine cognitive occupations, followed by

routine cognitive occupations. The lowest average wages are earned by workers in

nonroutine manual occupations.

The relative wages shown in Figure 7 are a ratio of average weekly real wages of

female workers to the average weekly real wages of male workers. The figure shows

that relative wages are highest for female workers in nonroutine cognitive and routine

cognitive occupations. However, these wages have been declining over time. On the

other hand, gender wage inequality is higher in manual occupations but has been

exhibiting a decreasing trend over time. It is also interesting to note that gender wage

inequality for the most part has been lowest in routine cognitive occupations. The fact

that the gender wage gap is lower for cognitive occupations is not surprising. Evidence

from Gunewardena, King, and Valerio (2018) shows that obtaining cognitive skills

reduces the gender wage gap for female workers. Sections IV and V delve into these

trends in employment and wages in more detail.

Figure 6. Average Wage by Gender and across Occupation Types




IV. Employment Decomposition

It is important to examine whether changes in employment shares in the four

categories of occupations obtained in this paper are mainly driven by a change in

industrial structure or by changes in employment shares within occupations. This

enables us to understand whether, for example, an increase in employment shares in

nonroutine cognitive occupations is mainly driven by an expansion of industries that

predominantly employ nonroutine cognitive workers or an increase in the share of

nonroutine cognitive workers within industries. Following Acemoglu and Autor

(2011), I use the given shift-share instrument to determine the total change in

employment shares:

ΔEjt ¼ ΔEBt þ ΔEW

t : ð1ÞChanges in employment shares of occupations between industries, ΔEB

t , or changes in

occupation shares within-industry, ΔEWt , can explain the total change in the share of

employment ΔEjt, where j is the occupation and t is the time. I can further express this

as follows:

ΔEjt ¼X

k

ΔEkt�jk þX

j

Δ�jkt Ek: ð2Þ

Figure 7. Relative Wage by Occupation Type

Note: Relative wage is measured as a ratio of average weekly real wage of female workers to the average weeklyreal wage of male workers.Source: Author’s calculations using data from the National Sample Survey of India (1994, 2000, 2005, 2008,2012) and the Periodic Labor Force Survey (2017).



ΔEkt represents the change in industry k’s share over the period under consideration,

whereas Ek represents the average employment share. Similarly, Δ�jkt gives the

change in occupation j’s share of industry k’s employment over the period under

consideration, whereas �jk gives the average share. The industries in this analysis are

at the NIC four-digit level.

Table 3 reports the decomposition of changes in employment shares, in

percentage points, of the four main occupation categories for all workers, male

workers, and female workers. I divided the data into roughly 2 decades to examine

Table 3. Employment Decomposition (Percentage Points)

1994–2005 2005–2017 1994–2017

All

Routine manualTotal Δ 2.55 �0.94 1.61Between industry Δ 2.22 �0.06 1.88Within industry Δ 0.33 �0.88 �0.27

Routine cognitiveTotal Δ �0.80 0.44 �0.36Between industry Δ �1.00 0.72 �0.50Within industry Δ 0.20 �0.28 0.14

Nonroutine manualTotal Δ �3.10 �7.32 �10.42Between industry Δ �4.58 �1.66 �7.68Within industry Δ 1.48 �5.66 �2.74

Nonroutine cognitiveTotal Δ 1.35 7.82 9.17Between industry Δ 0.36 2.86 3.39Within industry Δ 0.99 4.96 5.78

Male

Routine manualTotal Δ 2.13 �1.20 0.93Between industry Δ 1.93 �0.23 1.49Within industry Δ 0.20 �0.97 �0.56



Continued.



how this decomposition has been changing over time, in addition to reporting the

decomposition for the entire period under consideration. One can expect differences

over time because technological adoption and advancements have increased

significantly in India since 1990. For instance, the output of the IT sector alone

increased from $2.21 billion in 1994–1995 to $123.22 billion in 2014–15. While the

IT sector’s output increased by about $26 billion in the first decade of the analysis, it

increased by almost quadruple that amount in the following decade (Heeks 2015).

I find that within-industry changes in employment shares of occupations became

more important in the second period (2005–2017) when there was likely greater

technological adoption compared to the first period (1994–2005). For instance,

within-industry changes accounted for only 12.9% of the total change in routine

manual occupations in 1994–2005, but increased to 93.6% in 2005–2017. In fact,

within-industry changes contributed to a slight increase in the employment share of

Table 3. Continued.

1994–2005 2005–2017 1994–2017


Female

Routine manualTotal Δ 3.47 2.21 5.68Between industry Δ 2.65 2.60 5.25Within industry Δ 0.82 �0.39 0.43

Routine cognitiveTotal Δ �0.05 0.84 0.79Between industry Δ �0.26 1.12 0.60Within industry Δ 0.21 �0.28 0.19



Δ ¼ change in employment shares in percentage points.Source: Author’s calculations using data from the National SampleSurvey of India (1994, 2000, 2005, 2008, 2012) and the PeriodicLabor Force Survey (2017).



these occupations in the first period but caused the employment share to shrink in the

second period. The growing importance of within-industry shifts over time is

consistent across almost all occupational categories and for both female and male

workers. This highlights the need for following a task-content-of-occupations

approach to understand changes in employment structures, especially with greater

technological advances and adoption over time.

The results from the analysis suggest that routinization could be causing a

decline in within-industry employment shares in routine occupational categories. For

instance, for both male and female workers, the contribution of within-industry

employment share to the total change in employment shares for routine manual

occupations went from positive in 1994–2005 to negative in 2005–2017. It is

interesting to note, however, that in the period 2005–2017 for male workers in routine

manual occupations, within-industry declines accounted for 80% of the total decline in

employment shares, whereas for female workers this decline was much smaller than

the positive increase in employment share between industries. Similarly, for routine

cognitive occupations for male workers, within-industry changes during 2005–2017

accounted for the larger percentage of the total change in employment shares, whereas

this change was smaller for female workers. This might suggest that routinization has

differential impacts on female and male workers, affecting the employment of male

workers more adversely than that of female workers. Mechanization and technological

adoption, while automating jobs, also reduces the perceived differences in female and

male workers (Juhn, Ujhelyi, and Villegas-Sanchez 2014).

It is also interesting to note that while within-industry employment shares of

nonroutine cognitive occupations have been increasing as expected, the

within-industry employment shares for nonroutine manual occupations have been

declining. The latter runs counter to the hypothesis that automation should be

increasing the relative demand for workers in nonroutine occupations. This could be

explained by the fact the task-content framework does not capture mechanization that

can occur within the agriculture sector, which is primarily categorized as nonroutine

manual in the occupational categorization. Also, there may be supply-side effects

with workers moving to cognitive occupations as they gain greater access to

education. In fact, Acemoglu and Autor (2011) drop the agriculture sector from their

analysis. Table A4 of Appendix shows the employment decomposition after leaving

out the agriculture sector from the analysis. Compared to Table 3, Table A4 shows a

smaller decline in the employment shares of nonroutine manual occupations, and a

bigger decline in the employment of routine occupations, which is more in line with

expectations. Table A5 provides an employment decomposition at the one-digit



NCO for female workers to provide a description of each occupational type.

For instance, technicians and associate professionals experienced the biggest increase

in the share of employment overall, and also the largest within-industry increase in

employment shares.

An important concern is whether the between-industry changes in employment

shares could be a result of the changing classification of NIC codes over the years.

Table A6 of Appendix reports the share of employment in the data that is affected by

these changes in classification. Table A7 then reports the results of the employment

decomposition exercise without the employment data that have changed because of

changes in industrial classification. This change does not affect the analysis

significantly, and the main inferences stay the same.

V. Wage Analysis

In this section, I analyze whether routinization of tasks has impacted the returns

to workers based on their occupational specialization. To do so, I follow Acemoglu

and Autor (2011) in examining how wages of workers have changed over time

depending on their occupational specialization. They divide workers into three main

categories: workers performing routine tasks, those in nonroutine cognitive tasks, and

those in occupations intensive in nonroutine manual tasks. The main hypothesis is that

increased routinization leads to a decline in wages of workers in occupations intensive

in routine tasks. This should be reflected in a relative increase in wages of workers in

both nonroutine cognitive and nonroutine manual occupations.

To estimate this, Acemoglu and Autor first create demographic groups based on

gender, education, age, and region and then construct employment shares of workers in

routine occupations, nonroutine cognitive occupations, and nonroutine manual

occupations at the beginning of the period of their analysis, with the assumption that

workers self-select into each of these categories based on their comparative advantage

or task specialization. I similarly create these demographic groups and then construct

γRsejk, γNRMsejk , and γNRCsejk , which are the shares of workers employed in routine occupations

(including both routine manual and routine cognitive), nonroutine manual occupations,

and nonroutine cognitive occupations in each demographic group, respectively, for the

year 1994, which is the first year of the analysis. The age and education buckets used for

creating these cohorts are presented in Appendix in Tables A8 and A9, respectively.

NSS state-region codes (125 codes) were used for the region category. The variables

gender, education, age, and region are denoted by s, e, j, and k, respectively. Then based



on Acemoglu and Autor’s analysis, I estimate the following:

Δwsejkt ¼ β0 þ β1 � γNRMsejk � ti þ β2 � γNRCsejk � ti þ ti þ θe þ θj þ θk þ �sejkt: ð3ÞγRsejk was dropped from the regression because by construction γRsejk þ γNRMsejk þ γNRCsejk ¼ 1.

The education, region, and age fixed effects are denoted by θe, θj, and θk, respectively,

and ti is the time (year) dummy where t stands for time. Δwsejkt stands for the change

in mean log wages for each demographic group in the analysis. The difference in this

analysis compared to Acemoglu and Autor (2011) is that the change in wages is not

within a decade but across the various time periods available for this analysis.

The estimation results are presented in Table 4. Model 1 reports the estimates for

male workers without fixed effects, which are included in Model 2. Similarly, Model 3

reports the results for female workers without fixed effects, which are then included in

Model 4. There are more male workers than female workers in the survey, which is

reflected in the total number of observations for each.

Table 4. Wage Analysis

Male withoutFixed Effects

Male withFixed Effects

Female withoutFixed Effects

Female withFixed Effects

(1) (2) (3) (4)

Nonroutine manual2000 share� 2000 time dummy �0.210** �0.288*** �0.156 �0.0863

(0.0902) (0.0985) (0.131) (0.135)2005 share� 2005 time dummy 0.212** 0.134 �0.108 �0.0380

(0.103) (0.108) (0.137) (0.137)2008 share� 2008 time dummy 0.0343 �0.0434 0.118 0.188*

(0.0797) (0.0864) (0.0989) (0.107)2012 share� 2012 time dummy �0.0191 �0.0968 �0.191* �0.121

(0.0830) (0.0898) (0.110) (0.113)2017 share� 2017 time dummy 0.314*** 0.237*** �0.106 �0.0364

(0.0795) (0.0849) (0.119) (0.123)

Nonroutine cognitive2000 share� 2000 time dummy �0.181 �0.158 0.106 0.198

(0.114) (0.119) (0.177) (0.186)2005 share� 2005 time dummy 0.159 0.182 �0.244 �0.152

(0.112) (0.116) (0.166) (0.173)2008 share� 2008 time dummy 0.0388 0.0620 �0.00166 0.0903

(0.0993) (0.109) (0.128) (0.146)2012 share� 2012 time dummy �0.0639 �0.0408 �0.235* �0.143

(0.0865) (0.0931) (0.140) (0.155)2017 share� 2017 time dummy �0.240** �0.217** �0.172 �0.0797

(0.0970) (0.101) (0.152) (0.163)

Continued.



The results obtained are mixed. Relative to routine occupations, male workers in

nonroutine manual occupations experienced a significant decline in their wages in

2000, which is robust to the inclusion of fixed effects. These relative returns however

increased significantly in 2017 for both Models 1 and 2. This provides some support

for the fact that returns to workers specialized in nonroutine manual tasks have been

increasing recently relative to routine occupations. On the other hand, while wages of

male workers in nonroutine cognitive occupations did not change significantly relative

to workers in routine occupations in most years, a significant decline is observed for

2017, which runs counter to expectations. Female workers in nonroutine manual and

nonroutine cognitive occupations, on the other hand, for most years, did not

experience any significant changes in wages relative to those in routine occupations

Table 4. Continued.

Male withoutFixed Effects

Male withFixed Effects

Female withoutFixed Effects

Female withFixed Effects

(1) (2) (3) (4)

2000 year dummy 0.424*** 0.550*** 0.347*** 0.286**(0.0659) (0.0873) (0.0888) (0.140)

2005 year dummy �0.387*** �0.261*** �0.235** �0.296**(0.0700) (0.0884) (0.0988) (0.145)

2008 year dummy 0.191*** 0.317*** 0.181*** 0.119(0.0567) (0.0796) (0.0671) (0.129)

2012 year dummy 0.124** 0.250*** 0.227*** 0.165

(0.0594) (0.0840) (0.0748) (0.130)2017 year dummy �0.0552 0.0711 0.0608 �0.000817

(0.0580) (0.0812) (0.0866) (0.144)

Region fixed effect No Yes No YesEducation fixed effect No Yes No YesAge fixed effect No Yes No Yes

Observations 3,165 3,165 2,005 2,005R2 0.239 0.259 0.124 0.152

Notes: Standard errors in parentheses. *p < 0:10, **p < 0:05, ***p < 0:01. Models 1�4 in eachcolumn present a separate ordinary least squares regression of stacked changes in mean log weekly realwages by cohort and year, where cohorts are created using sex, age group (Table A8), education group(Table A9), and state (125 State-region codes are reported in NSS) of the workers in the NationalSample Survey (1994, 2000, 2005, 2008, 2012) and Periodic Labor Force Survey (2017) data.Occupation shares are calculated for each demographic group in 1994 and interacted with decadedummies. Occupations are grouped into three categories: (1) nonroutine cognitive; (2) nonroutinemanual; and (3) routine—both cognitive and manual. The routine group is the omitted category in theregression models. For the year dummy variables, the reference year is 1994.Source: Author’s calculations using data from the National Sample Survey of India (1994, 2000, 2005,2008, 2012) and the Periodic Labor Force Survey (2017).



during the period of analysis. One exception is an increase in the wages of female

workers in nonroutine manual occupations to those in routine occupations for 2008.

The coefficients of the time dummies in this analysis provide an estimate of the

wage trends of workers specializing in routine occupations at the beginning of the

analysis. For female workers in routine occupations, wages increased significantly in

2000 and then declined significantly in 2005, which is consistent with the hypothesis

that the returns to these occupations should decline over time. For male workers in

routine occupations, on the other hand, the results are not consistent—wages initially

increase significantly in 2000, then decline significantly in 2005, and then increase

thereafter.

This analysis is a preliminary exercise in determining how returns to workers in

various occupational specializations are evolving over time. While there is some

evidence of a decline in returns to workers in routine occupations over time, especially

for female workers, this is not strongly corroborated. Given that a developing country

such as India is still lagging in terms of technological adoption compared to the US,

one might observe these effects on wages becoming more prominent as technological

adoption increases in the future. The fact that changes in within-industry employment

shares have only recently become prominent in impacting overall changes in

employment across occupational categories also suggests that we can probably expect

these results to be stronger in the future.

VI. Conclusion

This paper uses a task-content-of-occupations approach to analyze trends in

employment and wages of male and female workers in India from 1994 to 2017. This

approach provides a better framework to take into account the impact of technological

advancements and automation while analyzing these labor market trends. Accordingly,

it classifies male and female workers into four occupational categories: routine

cognitive, nonroutine cognitive, routine manual, and nonroutine manual. The paper

considers the trends in male and female workers separately to understand how

technological advancements on average might be impacting these two groups of

workers differently.

Analyzing the employment trends I find that, on average, the share of workers in

nonroutine manual work is highest for both male and female workers but declining

over time, whereas the share of nonroutine cognitive workers is increasing.

An employment decomposition exercise, which uses a shift-share instrument to

determine the extent to which changes in employment shares across occupations are



due to changes in industrial structure compared to changes in within-industry

occupational shares, reveals interesting results. I find that in the more recent decade,

changes in within-industry employment shares are more important in determining the

overall change in occupational shares. This suggests that using the task-content

framework to understand changes in employment trends is becoming more important

in the Indian context. I find that, as expected, there is a decline in within-industry

shares of employment for routine cognitive and routine manual occupations. This

decline in occupational shares in routine occupations is the main factor driving the

overall changes in employment shares of male workers, while the same is not true for

female workers. This suggests that routinization might not be impacting female

workers as adversely as male workers. Within-industry employment shares of

nonroutine cognitive occupations for both female and male workers have been

increasing, in line with expectations.

I also examine average trends in wages across all occupational categories.

For both male and female workers, average wages are highest for nonroutine cognitive

occupations, followed by routine cognitive occupations and then routine manual

occupations, with the lowest wages for nonroutine manual occupations. The gender

wage gap is lowest in cognitive occupations—with routine cognitive occupations

recording the lowest gap for most of the period of analysis. A wage analysis that

considers changes in wages for workers specializing in a certain occupational category

at the beginning of the period, however, finds that there are no consistent, significant

changes in the earnings of workers of nonroutine occupations compared to those in

routine occupations.

The paper highlights the importance of using a task-content framework in

analyzing trends in employment and wages of workers by showing that changes in

employment shares across occupational categories are increasingly driven by

within-industry changes in occupational shares. Nonroutine cognitive occupations

experienced the biggest increases in employment shares—both between and within

industries—and we can expect the employment shares of workers in routine

occupations to decline over time. Labor market policies need to take these findings

into account as policy makers prepare for the future of the Indian workforce.

References

Acemoglu, Daron, and David Autor. 2011. “Skills, Tasks and Technologies: Implications forEmployment and Earnings.” In Handbook of Labor Economics, Volume 4, edited byOrley Ashenfelter and David Card, 1043–171. Amsterdam: Elsevier-North Holland.



Autor, David, Frank Levy, and Richard J. Murnane. 2003. “The Skill Content of RecentTechnological Change: An Empirical Exploration.” The Quarterly Journal of Economics118 (4): 1279–333.

Autor, David H., and David Dorn. 2013. “The Growth of Low-Skill Service Jobs andthe Polarization of the US Labor Market.” American Economic Review 103 (5): 1553–97.

Bas, Maria, and Antoine Berthou. 2012. “The Decision to Import Capital Goods inIndia: Firms’ Financial Factors Matter.” The World Bank Economic Review 26 (3):486–513.

Bound, John, and George Johnson. 1992. “Changes in the Structure of Wages in the1980’s: An Evaluation of Alternative Explanations.” American Economic Review 82 (3):371–92.

Bruckner, Matthias, Marcelo LaFleur, and Ingo Pitterle. 2017. “Frontier Issues: The Impact ofthe Technological Revolution on Labor Markets and Income Distribution.” UnitedNations Department of Economic and Social Affairs.

Goldberg, Pinelopi Koujianou, Amit Kumar Khandelwal, Nina Pavcnik, and Petia Topalova.2010. “Imported Intermediate Inputs and Domestic Product Growth: Evidence fromIndia.” The Quarterly Journal of Economics 125 (4): 1727–67.

Gunewardena, Dileni N., Elizabeth M. King, and Alexandria Valerio. 2018. “More ThanSchooling: Understanding Gender Differences in the Labor Market When Measures ofSkill Are Available.” The World Bank Policy Research Working Paper Series 8588.

Heeks, Richard. 2015. “Indian IT/Software Sector Statistics: 1980–2015 Time Series Data.”Retrieved from Reserve Bank of India. https://ict4dblog.wordpress.com/2015/04/28/indian-itsoftware-sector-statistics-1980-2015-time-series-data/ (accessed 31 May 2021).

James, Jeffrey. 1999. “New Technologies, Employment and Labor Markets in DevelopingCountries.” In Globalization, Information Technology and Development, 116–54.London: Palgrave Macmillan UK.

Juhn, Chinhui, Gergely Ujhelyi, and Carolina Villegas-Sanchez. 2014. “Men, Women, andMachines: How Trade Impacts Gender Inequality.” Journal of Development Economics106 (C): 179–93.

Kapur, Devesh. 2002. “The Causes and Consequences of India’s IT Boom.” India Review 1:91–110.

Katz, Lawrence F., and Kevin M. Murphy. 1992. “Changes in Relative Wages, 1963–1987:Supply and Demand Factors.” The Quarterly Journal of Economics 107 (1): 35–78.

Organisation for Economic Co-operation and Development (OECD). 2021. Consumer PriceIndex: All Items for India [INDCPIALLMINMEI]. Retrieved from FRED, FederalReserve Bank of St. Louis. https://fred.stlouisfed.org/series/INDCPIALLMINMEI(accessed 14 May 2021).

Sharma, Shruti. 2016. “Employment, Wages and Inequality in India: An Occupations and TasksBased Approach.” The Indian Journal of Labor Economics 59 (4): 471–87.

_____. 2018. “Heterogeneity of Imported Intermediate Inputs and Labor: Evidence from India’sInput Tariff Liberalization.” Applied Economics 50 (11): 1171–87.

Sharma, Shruti, and Nirvikar Singh. 2013. “Information Technology and Productivity in IndianManufacturing India Policy.” India Policy Forum 9 (1): 189–238.

Tinbergen, Jan. 1974. “Substitution of Graduate by Other Labor.” Kyklos 27 (2): 217–26.



_____. 1975. Income Differences: Recent Research. Amsterdam: North-Holland PublishingCompany.

Topalova, Petia, and Amit Khandelwal. 2011. “Trade Liberalization and Firm Productivity: TheCase of India.” The Review of Economics and Statistics 93 (3): 995–1009.

Appendix

Table A1. Classification of Occupations Based on Tasks

Occupation Type Tasks

Routine manual 4.C.3.d.3 Pace determined by speed of equipment4.A.3.a.3 Controlling machines and processes

4.C.2.d.1.i Spend time making repetitive motions

Routine cognitive 4.C.3.b.7 Importance of repeating the same tasks4.C.3.b.4 Importance of being exact or accurate

4.C.3.b.8 Structured vs. unstructured work (reverse)

Nonroutine manual 4.A.3.a.4 Operating vehicles, mechanized devices, or equipment4.C.2.d.1.g Spend time using hands to handle, control, or feel objects, tools, or

controls1.A.2.a.2 Manual dexterity

1.A.1.f.1 Spatial orientation

2.B.1.a Social perceptiveness

Nonroutine cognitive 4.A.2.a.4 Analyzing data/information4.A.2.b.2 Thinking creatively

4.A.4.a.1 Interpreting information for others

4.A.4.a.4 Establishing and maintaining personal relationships

4.A.4.b.4 Guiding, directing, and motivating subordinates

4.A.4.b.5 Coaching/developing others

Source: Acemoglu and Autor (2011)



Table

A2.

Classification

ofOccupations

Division

Dom

inan

tCategory

Total

3-Digit

Occupations

%Rou

tine

Man

ual

%Rou

tine

Cog

nitive

%Non

routine

Man

ual

%Non

routine

Cog

nitive

Legislators,senior

officials,andmanagers

Non

routinecogn

itive

70

00

100

Professionals

Non

routinecogn

itive

180

00

100

Techn

icians

andassociateprofession

als

Non

routinecogn

itive

200

405

55Clerks

Rou

tinecogn

itive

70

100

00

Service

workers

Rou

tinecogn

itive

837

5013

0Skilledagricultu

reandfisheryworkers

Non

routinemanual

60

010

00

Craftsandtrades-related

workers

Rou

tinemanual

1656

044

0Plant

andmachine

operatorsandassemblers

Rou

tinemanual

2080

020

0Elementary

occupatio

nsNon

routinemanual

933

1145

11

Sou

rce:

Autho

r’scalculations

usingdata

from

theNationalSam

pleSurveyof

India(199

4,20

00,20

05,20

08,20

12)andthePeriodicLabor

Force

Survey

(201

7).



Table A3. Share of Workers across Occupation Types (%)

1994 2000 2005 2008 2012 2017

All

Routine manual 11 13 13 14 12 12Routine cognitive 4 4 3 3 4 4Nonroutine manual 81 77 78 72 72 70Nonroutine cognitive 4 6 6 11 12 14

Male


Female


Source: Author’s calculations using data from the National Sample Survey of India(1994, 2000, 2005, 2008, 2012) and the Periodic Labor Force Survey (2017).

Table A4. Employment Decomposition without Agriculture

1994–2005 2005–2017 1994–2017

All

Routine manualTotal Δ 1.39 �5.08 �3.69Between industry Δ 0.87 �3.39 �2.76Within industry Δ 0.52 �1.69 �0.93

Routine cognitiveTotal Δ �2.66 �0.25 �2.91Between industry Δ �2.96 0.30 �2.94Within industry Δ 0.30 �0.55 0.03

Nonroutine manualTotal Δ 0.08 �3.95 �3.87Between industry Δ �2.28 3.47 0.88Within industry Δ 2.36 �7.42 �4.88

Nonroutine cognitiveTotal Δ 1.18 9.30 10.48Between industry Δ �0.68 2.65 1.73Within industry Δ 1.86 6.65 8.75

Continued.



Table A4. Continued.

1994–2005 2005–2017 1994–2017

Male


Routine cognitiveTotal Δ �3.04 �0.40 �3.44Between industry Δ �3.31 0.15 �3.53Within industry Δ 0.27 �0.55 0.09

Nonroutine manualTotal Δ 0.88 �4.97 �4.09Between industry Δ �1.42 2.52 1.24Within industry Δ 2.30 �7.49 �5.33

Nonroutine cognitiveTotal Δ 1.31 9.23 10.54Between industry Δ �0.75 2.59 1.51Within industry Δ 2.06 6.64 9.03

Female



Nonroutine manualTotal Δ �2.39 �1.07 �3.46Between industry Δ �5.11 6.22 �0.88Within industry Δ 2.72 �7.29 �2.58


Δ ¼ change in employment shares in percentage points.Source: Author’s calculations using data from the National SampleSurvey of India (1994, 2000, 2005, 2008, 2012) and the Periodic LaborForce Survey (2017).



Table A5. Employment Decomposition for Female Workers atOne-digit NCO Divisions

1994–2005 2005–2017 1994–2017

Legislators, senior officials, and managers

Total Δ 0.64 3.33 3.97Between industry Δ 0.22 0.70 1.00Within industry Δ 0.42 2.63 2.97

Professionals

Total Δ 0.51 4.11 4.62Between industry Δ 0.32 3.43 2.91Within industry Δ 0.19 0.68 1.71

Technicians and associate professionals

Total Δ 0.48 4.54 5.02Between industry Δ �0.08 4.16 1.23Within industry Δ 0.56 0.38 3.79

Clerks

Total Δ �0.11 0.77 0.66Between industry Δ �0.15 1.00 0.55Within industry Δ 0.04 �0.23 0.11

Service workers

Total Δ 0.92 3.02 3.94Between industry Δ 0.61 3.44 3.41Within industry Δ 0.31 �0.42 0.53

Skilled agriculture and fishery workers

Total Δ �5.79 �16.33 �22.12Between industry Δ �5.85 �14.53 �21.50Within industry Δ 0.06 �1.80 �0.62

Crafts and trades related workers

Total Δ 2.46 �0.45 2.01Between industry Δ 2.03 1.09 2.85Within industry Δ 0.43 �1.54 �0.84

Plant and machine operators and assemblers

Total Δ 0.64 �0.84 �0.20Between industry Δ 0.10 �0.20 0.20Within industry Δ 0.54 �0.64 �0.40

Continued.




1994–2005 2005–2017 1994–2017

Elementary occupations

Total Δ 0.25 1.83 2.08Between industry Δ 0.27 2.47 2.68Within industry Δ �0.02 �0.64 �0.60

Δ ¼ change in employment shares in percentage points, NCO ¼National Classification of Occupations.Source: Author’s calculations using data from the National SampleSurvey of India (1994, 2000, 2005, 2008, 2012) and the Periodic LaborForce Survey (2017).

Table A6. Industries and their Share of Employment

YearTotal Numberof Industries

Number ofNew

Industries

EmploymentShare of

New Industries (%)

Number ofIndustriesLeaving

EmploymentShare of Industries

Leaving (%)

1994 238 — — — —

2000 302 65 4.35 1 0.012005 305 5 0.09 2 0.012008 300 2 0.002 7 0.012012 303 7 0.01 4 0.112017 307 4 0.02 0 0.00

— means data not available.Source: Author’s calculations using data from the National Sample Survey of India (1994, 2000, 2005,2008, 2012) and the Periodic Labor Force Survey (2017).

Table A7. Employment Decomposition without the Effects ofChanges in Industrial Classification

1994–2005 2005–2017 1994–2017

All

Routine manualTotal Δ 2.65 �0.55 2.10Between industry Δ 2.67 0.36 2.71Within industry Δ �0.02 �0.91 �0.61

Routine cognitiveTotal Δ �1.35 0.24 �1.11Between industry Δ �1.15 0.42 �0.72Within industry Δ �0.20 �0.18 �0.39

Continued.




1994–2005 2005–2017 1994–2017

Nonroutine manualTotal Δ �1.68 �6.56 �8.24Between industry Δ �1.52 �1.12 �2.91Within industry Δ �0.16 �5.44 �5.33


Male

Routine manualTotal Δ 2.15 �0.92 1.23Between industry Δ 2.30 0.08 2.09Within industry Δ �0.15 �1.00 �0.86

Routine cognitiveTotal Δ �1.54 �0.04 �1.58Between industry Δ �1.32 0.19 �1.22Within industry Δ �0.22 �0.23 �0.36

Nonroutine manualTotal Δ �1.30 �5.97 �7.27Between industry Δ �1.07 �0.38 �1.64Within industry Δ �0.23 �5.59 �5.63


Female

Routine manualTotal Δ 3.75 3.60 7.35Between industry Δ 3.22 4.33 7.35Within industry Δ 0.53 �0.73 0.00

Routine cognitiveTotal Δ �0.58 0.82 0.24Between industry Δ �0.48 0.80 0.44Within industry Δ �0.10 0.02 �0.20


Continued.



Table A8. Age Buckets

Age Bucket Age Range (years)

1 < 182 18–243 25–344 35–445 45–546 55–647 65þSource: Author’s categories.


1994–2005 2005–2017 1994–2017

Nonroutine cognitiveTotal Δ �0.42 5.64 5.22Between industry Δ 0.00 1.63 2.32Within industry Δ �0.42 4.01 2.90

Δ ¼ change in employment shares in percentage points.Source: Author’s calculations using data from the National SampleSurvey of India (1994, 2000, 2005, 2008, 2012) and the PeriodicLabor Force Survey (2017).

Table A9. Education Buckets

Education Bucket Label Years of Education

1 Below middle school Up to grade 82 Secondary/higher secondary Grade 9 to grade 123 Higher education College or more

Source: Author’s categories.



Disability and Intrahousehold InvestmentDecisions in Education: Empirical

Evidence from Bangladesh

KAMAL LAMICHHANE AND TAKAKI TAKEDA¤

Investment disparity in the education of personswith disabilitiesmay be larger on thepart of parents, in part resulting from predicted lower returns to the investment due tomistaken beliefs about their capabilities, or actual lower returns due to barriers in thelabor market. Using a nationally representative dataset fromBangladesh and utilizingthe framework of the Engel curve, we investigate intrahousehold investmentdecisions in education between children with and without disabilities. The resultsof the hurdle model show the existence of disability bias in enrollment decisions,whereas individual-level analysis suggests that bias exists on educationalexpenditure after children with disabilities enroll in school. Additionally, thoughwe observe a lower level of bargaining power among household heads oneducational investments for their children with disabilities, interaction effectssuggest the importance of greater income stability and maternal education statusbeing instrumental to improving the education of persons with disabilities.

Keywords: Bangladesh, disability bias, Engel curve, hurdle model, investmentin education

JEL codes: I2, E2, O1

⁄Kamal Lamichhane (corresponding author): Faculty of Human Sciences, University of Tsukuba,Tsukuba, Japan. E-mail: [email protected]; Takaki Takeda: Independent Researcher.E-mail: [email protected]. We thank the Bangladesh Bureau of Statistics for the inclusion of globallyaccepted questions on disability in their survey, which has enabled us to research the topic. In addition, wealso wish to thank Aya Suzuki at the University of Tokyo, Toby Long at Georgetown University, DanielMont at the Center for Inclusive Policy Research, Guenwoo Lee at Hitotsubashi University, and Juan N.Martínez and Takahiro Tsujimoto for their helpful comments while drafting the paper. Last but not theleast, we wish to thank two anonymous reviewers for their comments that have helped improve the paper.The Asian Development Bank recognizes “Vietnam” as Viet Nam.





https://dx.doi.org/10.1142/S0116110522500032

https://dx.doi.org/10.1142/S0116110522500032

I. Introduction

Studies have shown the effect of household and individual factors such as

parental education, family size, and the gender of the child on educational investment

for children (Kingdon 2002, Sawada and Lokshin 2009). Similarly, there is ample

evidence regarding gender disparities in intrahousehold resource allocation

(Subramanian and Deaton 1991; Subramanian 1995; Deaton 1997; Burgess and

Zhuang 2000; Lancaster, Maitra, and Ray 2003; Kingdon 2005; Aslam and Kingdon

2008; Lancaster, Maitra, and Ray 2008; Himaz 2010; Masterson 2012; Azam and

Kingdon 2013).

On the other hand, why school enrollment and educational attainment of persons

with disabilities (PWDs) are lower than those of non-PWDs is still a matter of intense

debate. Several studies have discussed discriminatory attitudes, parents’ financial

constraints, and institutional barriers as some of the plausible reasons for the lower

level of schooling of PWDs (Lamichhane 2013, 2015; Lamichhane and Kawakatsu

2015; Takeda and Lamichhane 2018). Despite growing attention regarding the social

inclusion and economic empowerment of PWDs in recent years, to the best of our

knowledge, except for a study by Rosales-Rueda (2014) on children with mental

health conditions in the United States, there are no studies that examine the

relationship between children’s disability and household investment decisions for

education. The lack of rigorous studies on this topic and the importance of providing

evidence-based policy implications on the education of PWDs is the main motivation

for this study. Given the fact that there is dearth of specific data on disability and

intrahousehold resource allocation in education in Bangladesh, our study has the

potential to be a powerful planning tool for policy makers.

Investment disparity in the education of PWDs may be largely due to parents, as

in many countries discrimination toward PWDs is widespread. Consequently,

household financial constraints, combined with discriminatory attitudes on the part

of parents, may negatively affect their decision to invest in education of PWDs, in part

resulting from the predicted lower returns to education of PWDs due to mistaken

beliefs about their capabilities, or actual lower returns due to barriers in the labor

market. Thus, the education of PWDs may be partly driven by shifts in parental

investment strategies that may give priority to their nondisabled children. Therefore,

utilizing a large and nationally representative Household Income Expenditure Survey

(HIES) dataset from 2010 of a developing economy, Bangladesh, we aim to detect the

biases against children’s education caused by disability status in the process of

intrahousehold investment decisions.



According to Lamichhane and Kawakatsu (2015), the education system in

Bangladesh comprises 3 years of preprimary, 5 years of compulsory primary, 3 years

of junior secondary, 2 years of secondary, and 2 years of higher secondary education.

Education is only compulsory at the primary level and is free up to that level, with girls

continuing to receive free education up to the secondary level (Lamichhane 2015).

Bangladesh has initiated policies such as the National Education Policy, 2010 and a

series of Primary Education Development Programs (e.g., PEDP II and III) to meet its

constitutional obligations to provide a uniform, mass-oriented, and universal system of

education as well as its international commitments to educating all school-aged

children within the mainstream education system. A study by Ahmmed, Sharma, and

Deppeler (2012) discussed teachers’ perceptions of school support for implementing

inclusive education. While inclusive education is still gaining momentum in

Bangladesh, disability issues have been recognized in the country’s midterm

development plans, gradually placing policy attention on critical issues of education.

As one of the world’s least developed countries per United Nations classification

(World Bank 2021), Bangladesh is characterized by little or no access to social benefits

and little implementation on the ground to ensure that vulnerable people, including

individuals with disabilities, get what they are promised. The study of Chowdhury and

Foley (2006) shows how persons with disabilities in rural Bangladesh can slide into

economic impoverishment once they are labeled as such due to the various

deprivations that their impairments expose them to. Bangladesh developed a national

policy in 1995 emphasizing the provision of services for PWDs. It also enacted the

comprehensive disability law known as the Disability Welfare Act in April 2001,

which aimed to protect the rights of PWDs. Additionally, Bangladesh ratified the

Convention on the Rights of Persons with Disabilities in November 2007 and ratified

the optional protocol in May 2008. In 2013, a new Disability Act in line with the

convention was enacted. Despite such legal accomplishments, the increased

participation of PWDs in educational, social, and economic sectors is still critical

(Lamichhane 2015).

Additionally, Lamichhane and Kawakatsu (2015) examined the determinants of

school participation between children with and without disabilities in Bangladesh and

found that those with disabilities are less likely to participate in school. However, once

the sample is restricted to those with disabilities only, their study shows that household

monthly expenditure and working-age members are positively correlated with the

probability of school participation. If PWDs are deprived of education, this will hurt

their quality of life. For example, Lamichhane and Sawada (2013) found that in Nepal

wage returns to the investment in education of PWDs were between 19.2% and 25.6%,

DISABILITY AND INTRAHOUSEHOLD INVESTMENT DECISIONS IN EDUCATION 203


which is two or three times higher than estimates for the general population

(Psacharopoulos and Patrinos 2004). While studies on disability and education in

Bangladesh are rare, with this paper we aim to at least partially fill the gap in existing

knowledge by examining the existence of disability bias and related factors associated

with parental decisions on educational investments. By doing so, we intend to help

identify constraints preventing PWDs from enjoying the multifaceted benefits of

education. The research questions posed herein are as follows: Does disability bias

exist in intrahousehold investment in education? If so, where does such bias exist, in

the stage of enrollment or thereafter? What are the key factors affecting parental

investment decisions with regard to the education of children with disabilities

(CWDs)? Although there is a serious lack of scientific and evidence-based information

on disabilities, Bangladesh provides a good setting for studying parental decisions on

intrahousehold resource allocation in education due to the availability of nationally

representative data.

II. Dataset from Bangladesh

We use the large-scale and nationally representative 2010 HIES dataset published

by the Bangladesh Bureau of Statistics. This dataset includes a wide variety of

information on the country’s socioeconomic situation—including demographic

characteristics, educational attainment, employment status, and access to facilities,

among others—and consists of more than 12,000 households: 20 each from 612

primary sampling units. Out of the total sample of 55,580 household members, there

are 16,696 school-aged children (6–18 years old) who are supposed to attend school in

Bangladesh.

Additionally, the survey identifies people’s disability status based on a short set

of questions recommended by the Washington Group on Disability Statistics (2020).

The questions focus on the difficulties people face in seeing, hearing, walking,

cognition, self-care, and communication according to a four-point scale: (1) no

difficulty, (2) some difficulty, (3) a lot of difficulty, and (4) cannot do. From these

questions, we obtain information on the type and severity of activity limitations.

Having some difficulty is considered to be a moderate limitation, whereas a lot of

difficulty and cannot see, hear, speak, walk, and so on are considered to be severe

limitations. The inclusion of the Washington Group on Disability Statistics’ questions

is also helpful for the international comparability of the situation regarding disability.

In the 2010 HIES, disability modules were asked under section three (health),



subsection (d). Together with 18 questions, including the question on onset of

impairment, two additional questions on any difficulty caused by disability at home, in

the workplace, or at school were included in the 2010 HIES. Sadly, the 2016 HIES

included only six questions about participants’ impairment as part of a household

information roster, excluding all other questions such as onset of impairment and any

difficulty caused by disability at home, work, and school. The exclusion of disability

modules in the 2016 HIES dataset led us to believe that the 2010 HIES was more

comprehensive in relation to the information on disability. For this reason, 2010 HIES

data have been used in this study.

III. Empirical Strategy and Related Literature

We investigate the existence of disability bias by examining the discriminatory

allocation of educational expenditure within households. We use the Engel curve

framework, which has been used in numerous previous studies to detect gender bias in

intrahousehold resource allocation. It is equally important to detect disability bias

through the Engel curve framework, which seeks differential treatment within

households indirectly. Such bias can be identified by examining how the household

composition of people with and without disabilities affects household expenditure on

education. To detect these biases, we use both an indirect and direct method. The

indirect method, known as household-level analysis, is based on conventional Engel

curve methodology, while the direct method refers to individual-level analysis that

uses individual-child-level data.

A. Household-Level Analysis

We follow Subramanian and Deaton (1991) and Deaton (1997) and employ the

Working and Leser specification, extended by adding household demographic

composition to Working’s Engel curve (Working 1943). To estimate the Engel

curve, the equation is also relaxed for nonlinearity of the log of per capita expenditure

along with the shape of the Engel curve

si ¼ αþ β lnxini

� �þ γ ln (ni)þ

XK�1

k¼1

θknkini

þ �zi þ "i, ð1Þ

where si is the budget share of total educational expenditure of ith household; xi is the

total expenditure of that household; ni is the number of members (i.e., household size)



of the ith household; nki is the number of household members in the kth disability-age

cohort where k ¼ 1, 2, . . . ,K; zi is a vector of other household characteristics (i.e.,

information on household head, religion, and dependency ratio) that are incorporated

as controls; and "i is the error term. Accordingly, ln (xi=ni) is the natural log of total

per capita expenditure of the ith household. When detecting bias, we elaborate

disability-age category variables, which are divided into 14 categorical groups for

PWDs and non-PWDs following Kingdon (2005): 0–4 years, 5–9 years, 10–14 years,

15–19 years, 20–24 years, 25–60 years, and over 60 years. The variable for over 60

years old is omitted as this is the base observation against the results of variables for

other groups. Using these disability-age categories, nki=ni is the share of the kth

disability-age fraction and the value of coefficient θk is the effect of household

composition by disability-age category on the budget allocation with the difference in

the household size of each household being considered. As our main objective is to

identify disability bias, we test the difference of the marginal effect (DME) of the

disability-age category variable (difference among the same age category) only for

variables along with school age, and the following null hypothesis is tested for each

school age (i.e., children aged 5–9, 10–14, and 15–19 years old):

θkCWD ¼ θknCWD, ð2Þwhere CWD is a child with disabilities and nCWD is a child without disabilities. k

refers to a given age category within school age roughly in line with the Bangladeshi

education system from ages 5 to 19 years old (5–9 ¼ primary education,

10–14 ¼ junior secondary and secondary education, 15–19 ¼ higher secondary and

higher education).

Additionally, as a strategy to identify disability bias, we employ the hurdle model

(Cragg 1971, Cameron and Trivedi 2005, Wooldridge 2010), which has also been used

to examine gender bias (Kingdon 2005, Aslam and Kingdon 2008, Himaz 2010, Azam

and Kingdon 2013).1 While the Engel curve model (equation [1]) has been estimated

using ordinary least squares (OLS) and including all households regardless of

children’s enrollment status, Kingdon (2005) highlighted the failure of such

conventional research and proposed that there should be two possible channels

through which pro-male bias is observed in expenditure: (i) via zero purchases for

daughters and positive purchases for sons, and (ii) conditional on positive purchases

for both daughters and sons via lower expenditure on daughters than on sons. More

succinctly, bias against girls must exist in two types of decision-making patterns by

1The hurdle model is also referred to as a two-part model and is primarily applied by Duan et al.(1983) for forecasting medical expenses (Cameron and Trivedi 2005, Wooldridge 2010).



parents—that is, whether to enroll their children (zero-versus-positive spending

decision) and how much to pay for their education on the condition that a child is

already in school (spending conditioned on enrollment decision). The Engel curve

method with conventional OLS neglects the first part of decision-making (neglect zero

expenditure) and this leads to the downward bias of estimation results.2

The hurdle model detects bias efficiently under censoring and two types of

decision-making. The simple hurdle model is as follows:

P(s ¼ 0jx) ¼ 1� �(xγ), ð3Þ

ln (s)j(x, s > 0) � Normal(xβ,�2), ð4Þ

where s is the budget share of educational expenditure in total household expenditure,

x is a vector of other explanatory variables, and γ and β are parameters to be estimated.

Once we obtain the results of the maximum likelihood estimation of the probit model

(binary choice of whether s > 0) and OLS conditional on nonzero expenditure (s > 0)

using the hurdle model results, we calculate the combined marginal effect (CME),

which shows the effect of x on outcomes of both models (i.e., probability of s > 0 and

the amount of sjs > 0). This CME is calculated as follows (Kingdon 2005)3:

@E(sjx)@x

¼ fγ’(xy)þ �(xγ)βg exp xβ þ �2

2

� �: ð5Þ

Estimators are derived separately in each stage of the hurdle model in equations (3)

and (4). �( � ) is the cumulative normal density function and ’(xy) is the standard

normal density function. We apply the model to investigate the existence of disability

bias between CWDs and non-CWDs within households.

Moreover, this model can solve the averaging problems of the conventional

Engel model (Kingdon 2005). Additionally, educational expenditure is used in

household analysis as if this were expenditure for all household members, but this is

actually a more personalized cost. Specifically, while enrollment itself is lower, the

educational cost for CWDs is considered to be higher than that for non-CWDs as

educational materials such as textbooks in braille or providing sign language,

equipment, and inaccessible infrastructure may require additional costs. Addressing

these issues is crucial for CWDs.

2Deaton (1997) states that a large proportion of households pay nothing for education.3For more specific calculation for derivation of an equation, see Kingdon (2005).



B. Individual-Level Analysis

We also employ the direct method with individual-child-level data to check the

robustness of our analysis to confirm disability-based bias, as well as to consider

disability-based bias and parents’ investment behaviors in detail. Here, equation (1) is

slightly changed as follows:

pij ¼ αþ β lnxini

� �þ γ ln (ni)þ θ impairmentij þ �zij þ "ij, ð6Þ

where pij is personal educational expenditure on the jth child in the ith household, xi is

total household expenditure, ni is household size, impairment is a dummy variable

(1 ¼ disabled, 0 ¼ otherwise), and zij is a vector of other child-specific (includes

timing of getting disability) and household characteristics. In this individual

regression, the impairment dummy captures the effect of disability bias. We use

interaction terms between children’s disability status and other characteristics to test

whether there are heterogeneous effects of disability depending on certain

characteristics. Here, disability status is interacted with a female dummy, mother’s

years of schooling, father’s years of schooling, and employment status of the

household head. These terms can reveal whether female CWDs suffer more bias

relative to male CWDs, and whether parents’ education and income stability alter the

levels of disability bias.

Table 1 reports the results for three equations: the unconditional OLS model (D),

the probit model for binary choice of school enrollment (A), and the conditional OLS

model conditioned on enrollment (B). Additionally, we add a column for CME in (C).

We restricted the household-level analysis to households that have children aged 5–19

years old, while individual-level analysis is conducted for each schooling level group

of children (i.e., 5–9, 10–14, and 15–19 years old). Estimations for the hurdle model

are calculated using the “twopm” command in Stata (Belotti et al. 2015).

While the estimation framework does not enable us to determine a clear direction

of causality, we think that this allows us to pursue our objectives to determine the

existence of bias itself. Additionally, the estimation can be complicated by a problem

of endogeneity. Among independent variables, monthly expenditure and dependency

ratio can be the index of poverty level and this may cause child impairment

(impairment dummy) through malnutrition and serious illness. This may underestimate

the impact of the impairment dummy variable, making it hard to identify a causal

relationship between disability status and individual educational expenditure, which

leads to failure to disclose disability bias. To handle this, we include household-level

(family) fixed effects to consider all time-invariant heterogeneity of household



characteristics and minimize the effect of the impairment dummy. Filmer (2008)

argues the effectiveness of incorporating household fixed effects to deal with

endogeneity of disability status that may arise from poverty in the developing world,

which is presented in the final section.4 Additionally, a Durbin–Wu–Hausman test is

unable to reject the null hypothesis that household size is exogenous. Therefore, we do

not present the result of the instrumental variable regression model. To control district-

level characteristics, we also incorporate district-level fixed effects for all analysis.

IV. Results and Findings

A. Descriptive Results

We observe significant differences in enrollment between CWDs and non-CWDs

for both rural and urban areas (Table 1). In particular, the primary enrollment rates of

non-CWDs are 23 percentage points and 17 percentage points higher than that of their

counterparts in urban and rural areas, respectively. We observe a similar trend for the

secondary level enrollment rate. Enrollment rates for non-CWDs are higher by more

than 10 percentage points in both areas. Similarly, at the tertiary level, we also find that

CWDs are in a disadvantageous position. Additionally, Table 1 shows mean

differences of literacy status (reading and writing skills) where differences are

observed in reading skills for both groups.

Table 2 shows the investment breakdown of educational expenditure, which is

divided into six categories. Admission includes admission, seasonal, and registration

fees. Tuition includes annual tuition and examination fees. Books includes textbooks

and exercise books. Other costs include hostel costs, conveyance, internet and e-mail

fees, meals, and other costs. Generally, tuition fees are the single-largest category of all

expenditures. Other costs are higher in rural areas than in urban areas, which may

indicate why transportation fees account for a relatively large proportion.

Table 3 shows the educational expenditure differential by schooling level and

disability status for both urban and rural areas, and for both the entire child sample and

the sample restricted to enrolled children only. The sample that includes out-of-school

children suggests that there is a significant difference in educational expenditure at the

primary level in rural areas. Surprisingly, there is little difference in educational

investment by disability status for enrolled children. This may be because school

4The possibility of the problem of endogeneity of the impairment dummy is also rejected by theDurbin–Wu–Hausman test.



Table 1. Differences in Educational Indices by Area and Disability Status

Urban Rural

Non-CWDs CWDs

MeanDifference(Non-CWDs

VersusCWDs) t-Value

Non-CWDs CWDs

MeanDifference(Non-CWDs

VersusCWDs) t-Value

Enrollment (school-aged children)Primary level 0.69 0.47 0.23*** 2.69 0.67 0.51 0.17*** 3.36Secondary level 0.45 0.34 0.11* 1.89 0.39 0.27 0.13*** 2.75Higher level 0.39 0.25 0.14* 1.86 0.28 0.18 0.10** 2.31

Literacy (school-aged children)Reading 0.76 0.59 0.17*** 4.76 0.67 0.58 0.10*** 3.58Writing 0.97 0.97 0.00 0.09 0.96 0.98 �0:02 �1:41

CWDs ¼ children with disabilities.Notes: *, **, and *** represent significance in gap by disability at the 10%, 5%, and 1% levels,respectively. Enrollment rate is calculated using the proportion of enrolled children within school-agedchildren divided by the population of school-aged children. School age is defined following Kingdon(2005): primary level (aged 5–9 years), secondary (aged 10–14 years), and higher (aged 15–19 years). Thedefinition of CWDs is children with disabilities below the age of 20 years old who acquire impairmentssuch as visual, hearing, physical, cognitive, and communication.Source: Authors’ calculations using data from the 2010 Bangladesh Household Income ExpenditureSurvey.

Table 2. Breakdowns of Educational Expenditure by Area (%)

Urban Rural

Public Private Madrasa Public Private Madrasa

Admission fee 13.17 14.87 5.90 6.27 8.22 5.82Tuition fee 48.36 52.96 43.54 35.81 40.71 33.61Books 14.49 13.81 17.96 19.13 20.71 25.26Uniform 5.65 5.14 8.93 10.45 8.04 11.45Contribution (donation) 0.07 0.02 0.00 0.48 0.15 0.00Other costs 18.26 13.20 23.67 27.86 22.17 23.86

Notes: Educational expenditure is divided into six categories. Admission includes admission fees,seasonal fees, and registration fees. Tuition includes annual tuition and examination fees. Booksincludes textbooks and excise books. Other costs are the sum of hostel costs, conveyance, internetand e-mail fees, tiffin (lunch), and other costs.Source: Authors’ calculations using data from the 2010 Bangladesh Household IncomeExpenditure Survey.



Table3.

Difference

ofEducation

alExp

enditure

byAreaan

dDisab

ility

Status(Tk)

Urban

Rural

Non

-CW

Ds

CW

Ds

MeanDifference

(Non

-CW

Ds

VersusCW

Ds)

t-Value

Non

-CW

Ds

CW

Ds

MeanDifference

(Non

-CW

Ds

VersusCW

Ds)

t-Value

Education

alexpenditure

forbothenrolledan

dnon

enrolledchild

ren

Primarylevel

3,08

2.09

1,94

3.33

1,83

8.75

1.59

1,14

6.76

709.07

437.69

*1.90

Secon

dary

level

6,76

2.93

4,34

5.00

2,41

7.93

*1.92

2,85

8.27

1,98

2.99

875.28

**2.51

Higherlevel

7,86

2.18

7,44

2.39

419.79

0.11

3,87

5.85

4,15

7.38

�281:53

�0:33

Education

alexpenditure

forenrolledchild

ren

Primarylevel

4,59

8.08

3,64

3.75

954.33

0.57

1,53

1.15

1,13

2.02

399.13

1.25

Secon

dary

level

8,01

2.52

7,011.25

1,00

1.27

0.60

3,38

1.98

2,98

7.71

394.27

0.90

Higherlevel

15,916

.88

16,373

.25

�456:37

�0:06

9,38

5.83

11,997

�2,611:17

�1:41

CWDs¼

child

renwith

disabilities,Tk¼

Bangladeshtaka.

Notes:*and**

representsign

ificanceat

the10

%and5%

levels,respectiv

ely.Schoo

lingageisdefinedfollo

wingKingd

on(200

5):prim

arylevel

(aged5–

9years),secon

dary

(aged10

–14

years),and

high

er(aged15

–19

years).T

hedefinitio

nof

CWDsischild

renwith

disabilitiesbelowtheage

of20

who

acqu

ireim

pairmentssuch

asvisual,hearing,

physical,cogn

itive,andcommun

ication.

Sou

rce:

Autho

rs’calculations

usingdata

from

the20

10BangladeshHou

seho

ldIncomeExp

enditure

Survey.



enrollment of CWDs can require more costs than that of non-CWDs when the

infrastructure is not disability friendly, as well as when the proper support systems are

not in place. However, once enrolled, the results reveal only small differences in

educational expenditure.

Table 4 reports the descriptive statistics of the variables used for the regression

analysis. This table also shows differences in child enrollment and educational

spending between urban and rural areas. In the results of the t-test among variables

between these areas, we can observe significant differences at the 1% level for the

budget share of educational expenditure, total household expenditure, household size,

and dependency ratio, clearly showing the advantages regarding education for

households in urban areas in terms of income level, number of household members,

and family structure. These results indicate that disparities in income level and

educational achievement exist between urban and rural areas in Bangladesh.

B. Regression Results of Household-Level Analysis

Table 5 presents the results of household-level analysis for factors affecting the

budget share of educational expenditure with disability-age category variables. We

perform regression analysis separately for the entire sample, urban areas, and rural

areas. Column (1) is the marginal effect in the probit model of whether parents enroll

their child in school. Column (2) is the result of conditional OLS analysis when a child

is enrolled. Column (3) shows the CME derived from the results of the probit and

conditional OLS models. Finally, column (4) is the result of unconditional OLS

analysis.

In the results of the entire sample and those broken down by rural or urban area,

the log of monthly expenditure per capita has a positive and significant effect on the

budget share for educational expenditure. These results reveal that income level is a

strong predictor of educational investment, as found in Glewwe and Patrinos (1999)

and Glewwe and Jacoby (2003). Household size has a positive correlation with

educational expenditure, and this variable works appropriately as a control variable

since larger households generally have more educational expenditure. The positive and

significant effect of a household head’s years of schooling on educational expenditure

for their children is consistent as the educational level of household head or parents is

considered a strong predictor of investment in their children’s education (Behrman and

Deolalikar 1995, Haveman and Wolfe 1995, Cameron and Heckman 1998). Similarly,

the dummy variable of female household head shows that female bargaining power has

a positive effect on both school enrollment and educational investment for their



Table4.

Descriptive

Statisticsof

Cov

ariatesat

theHou

seholdan

dIndividual

Levels

Urban

Rural

Number

ofObs.

Mean

Std.Dev.

Number

ofObs.

Mean

Std.Dev.

MeanDifference

(Urban

–Rural)

t-Value

Variablesforhou

sehold-level

analysis

Share

ofeducationalexpend

iture

4,39

40.05

0.07

7,81

30.03

0.05

0.01

7***

15.62

Total

mon

thly

householdexpend

iture

(Tk)

4,39

413

,852

.34

11,803

.18

7,81

39,65

5.47

7,42

2.88

4,19

6.87

2***

24.08

Hou

seho

ldsize

(num

berof

person

s)4,39

44.48

1.81

7,81

34.58

1.93

�0:107**

*�3

:01

Share

ofno

n-CWDsaged

5–9years

4,39

90.10

0.14

7,84

00.12

0.15

�0:014**

*�5

:11

Share

ofCWDsaged

5–9years

4,39

90.00

0.02

7,84

00.00

0.02

�0:001*

�1:94

Share

ofno

n-CWDsaged

10–14

years

4,39

90.10

0.14

7,84

00.11

0.14

�0:004

�1:36

Share

ofCWDsaged

10–14

years

4,39

90.00

0.03

7,84

00.00

0.02

0.00

040.73

Share

ofno

n-CWDsaged

15–19

years

4,39

90.09

0.14

7,84

00.08

0.13

0.00

8***

3.06

Share

ofCWDsaged

15–19

years

4,39

90.00

0.02

7,84

00.00

0.02

�0:001

�1:46

Years

ofeducationof

householdhead

4,38

75.95

5.72

7,82

33.08

4.29

2.87

5***

31.43

Hou

seho

ldhead

isfemale(¼

1)4,38

70.13

0.33

7,82

30.15

0.36

�0:026**

*�3

:99

Hou

seho

ldhead

with

impairment(¼

1)4,38

70.14

0.35

7,82

30.18

0.38

�0:037**

*�5

:29

Hou

seho

ldhead

isfulltim

ewageworker(¼

1)4,38

70.41

0.49

7,82

30.28

0.45

0.12

9***

14.75

Muslim

(¼1)

4,39

90.89

0.32

7,84

00.87

0.33

0.01

2**

1.99

Hindu

(¼1)

4,39

90.11

0.31

7,84

00.11

0.31

0.00

10.15

Dependencyratio

4,39

90.35

0.21

7,84

00.39

0.23

�0:047**

*�1

1:20

Con

tinued.



Table

4.Con

tinued.

Urban

Rural

Number

ofObs.

Mean

Std.Dev.

Number

ofObs.

Mean

Std.Dev.

MeanDifference

(Urban

–Rural)

t-Value

Variablesforindividual-level

analysis

Fem

ale(¼

1)19

,636

0.50

0.50

35,796

0.51

0.50

�0:006

�1:42

Age

(years)

19,636

27.02

18.93

35,796

26.45

19.95

0.56

6***

3.25

Impairment(¼

1)19

,636

0.08

0.27

35,796

0.09

0.29

�0:012**

*�4

:75

Visualim

pairment(¼

1)19

,636

0.06

0.24

35,796

0.06

0.24

�0:002

�0:91

Hearing

impairment(¼

1)19

,636

0.02

0.14

35,796

0.03

0.16

�0:007**

*�5

:16

Phy

sicalim

pairment(¼

1)19

,636

0.02

0.13

35,796

0.03

0.16

�0:011**

*�7

:92

Cog

nitiv

eim

pairment(¼

1)19

,636

0.01

0.10

35,796

0.01

0.12

�0:004**

*�4

:24

Com

mun

icationim

pairment(¼

1)19

,636

0.01

0.08

35,796

0.01

0.10

�0:002**

*�3

:05

Mother’syearsof

scho

oling

9,76

04.61

4.76

18,622

2.51

3.56

2.09

7***

41.80

Father’syearsof

scho

oling

8,22

35.86

5.67

15,512

3.16

4.34

2.69

4***

40.77

CWDs¼

child

renwith

disabilities,Obs:¼

observations,Tk¼

Bangladeshtaka.

Notes:*,

**,and**

*representsign

ificanceat

the10

%,5%

,and1%

levels,respectiv

ely.Amon

gvariablesforho

usehold-levelanalysis,femalehead

ofho

usehold,

householdhead

with

impairment,ho

useholdhead

isfulltim

ewageworker,andMuslim

andHindu

aredu

mmyvariablesthat

take

avalueof

either

0or

1.Amon

gvariablesforindividu

al-level

analysis,female,im

pairment,andtypesof

impairment(i.e.,visual,hearing,

physical,cogn

itive,and

commun

ication)

aredu

mmyvariablesthat

take

avalueof

either

0or

1.Sou

rce:

Autho

rs’calculations

usingdata

from

the20

10BangladeshHou

seho

ldIncomeExp

enditure

Survey.



Table5.

Resultsof

Hou

sehold-Level

Analysis

All

Urban

Rural

Probit

Con

di-

tion

alOLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Uncondi-

tion

alOLS

Probit

Con

di-

tion

alOLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Uncondi-

tion

alOLS

Probit

Con

di-

tion

alOLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Uncondi-

tion

alOLS

IndependentVariables

(1)

(2)

(3)

(4)

(1)

(2)

(3)

(4)

(1)

(2)

(3)

(4)

Log

ofmonthly

expenditu

repercapita

0.480***

3.982***

0.082***

0.087***

0.442***

4.737***

0.083

0.096*

0.587***

4.289***

0.126***

0.132***

(0.099)

(0.473)

(0.026)

(0.029)

(0.154)

(0.860)

(0.053)

(0.057)

(0.146)

(0.571)

(0.020)

(0.018)

Squareof

logof

monthly

expenditu

repercapita

�0:023***

�0:236***

�0:004**

�0:004**

�0:020**

�0:278***

�0:004

�0:005

�0:030***

�0:261***

�0:007***

�0:008***

(0.006)

(0.030)

(0.002)

(0.002)

(0.009)

(0.053)

(0.003)

(0.004)

(0.009)

(0.037)

(0.001)

(0.001)

Log

ofhouseholdsize

0.296***

0.207***

0.021***

0.023***

0.296***

0.185***

0.027***

0.029***

0.295***

0.236***

0.020***

0.021***

(0.009)

(0.045)

(0.002)

(0.001)

(0.014)

(0.071)

(0.003)

(0.002)

(0.012)

(0.059)

(0.002)

(0.002)

Disab

ility-age

category

variab

les

Share

ofCWDsaged

5–9years

0.733***

�0:132

0.037**

0.022

0.582**

0.404

0.036

0.036

0.853***

�0:218

0.036**

0.024

(0.128)

(0.559)

(0.015)

(0.014)

(0.243)

(0.805)

(0.033)

(0.028)

(0.145)

(0.690)

(0.017)

(0.016)

Share

ofnon-CWDsaged

5–9years

1.149***

�0:169

0.065***

0.043***

1.218***

0.000

0.098***

0.075***

1.116***

�0:230

0.049***

0.030***

(0.035)

(0.219)

(0.008)

(0.004)

(0.066)

(0.355)

(0.015)

(0.009)

(0.041)

(0.279)

(0.008)

(0.005)

Share

ofCWDsaged

10–14

years

0.811***

1.423***

0.081***

0.077***

1.050***

0.402

0.099***

0.079**

0.721***

1.932***

0.070***

0.072***

(0.127)

(0.420)

(0.017)

(0.019)

(0.200)

(0.743)

(0.033)

(0.039)

(0.156)

(0.484)

(0.018)

(0.020)

Share

ofnon-CWDsaged

10–14

years

1.147***

1.758***

0.107***

0.105***

1.083***

1.534***

0.130***

0.128***

1.179***

1.927***

0.093***

0.095***

(0.038)

(0.215)

(0.008)

(0.005)

(0.066)

(0.340)

(0.014)

(0.010)

(0.045)

(0.279)

(0.008)

(0.005)

Share

ofCWDsaged

15–19

years

0.651***

1.555***

0.080***

0.083***

0.513**

3.296***

0.133***

0.117***

0.675***

0.932

0.061**

0.075***

(0.127)

(0.590)

(0.023)

(0.022)

(0.216)

(0.860)

(0.044)

(0.043)

(0.149)

(0.719)

(0.025)

(0.026)

Share

ofnon-CWD

child

ren15–19

years

0.525***

1.920***

0.096***

0.098***

0.588***

1.942***

0.135***

0.124***

0.493***

1.926***

0.076***

0.086***

(0.048)

(0.282)

(0.012)

(0.007)

(0.078)

(0.473)

(0.020)

(0.014)

(0.058)

(0.345)

(0.013)

(0.008)

Con

tinued.



Table

5.Con

tinued.

All

Urban

Rural

Probit

Con

di-

tion

alOLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Uncondi-

tion

alOLS

Probit

Con

di-

tion

alOLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Uncondi-

tion

alOLS

Probit

Con

di-

tion

alOLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Uncondi-

tion

alOLS


(1)

(2)

(3)

(4)

(1)

(2)

(3)

(4)

(1)

(2)

(3)

(4)

Hou

seho

ld-level

characteristics

Years

ofschoolingfor

householdhead

0.006***

0.060***

0.002***

0.002***

0.007***

0.057***

0.002***

0.003***

0.005***

0.054***

0.002***

0.002***

(0.001)

(0.003)

(0.000)

(0.000)

(0.001)

(0.004)

(0.000)

(0.000)

(0.001)

(0.004)

(0.000)

(0.000)

Household

head

isfemale

0.029***

0.172***

0.007***

0.009***

0.013

0.199***

0.008**

0.012***

0.035***

0.156***

0.008***

0.008***

(0.011)

(0.043)

(0.002)

(0.002)

(0.018)

(0.072)

(0.003)

(0.003)

(0.013)

(0.054)

(0.002)

(0.002)

Household

head

with

disabilities

0.010

�0:090*

�0:002

�0:001

0.037*

�0:180**

�0:002

0.000

�0:003

�0:053

�0:002

�0:002

(0.011)

(0.049)

(0.002)

(0.002)

(0.020)

(0.088)

(0.004)

(0.004)

(0.013)

(0.058)

(0.002)

(0.002)

Household

head

isfulltim

ewageworker

0.018***

0.121***

0.007***

0.006***

0.018

0.048

0.007***

0.007***

0.021**

0.164***

0.006***

0.006***

(0.007)

(0.028)

(0.001)

(0.001)

(0.011)

(0.044)

(0.002)

(0.002)

(0.009)

(0.038)

(0.001)

(0.001)

Muslim

�0:017*

�0:136***

�0:006***

�0:006***

�0:01

�0:148***

�0:007**

�0:006*

�0:016

�0:110**

�0:004**

�0:005***

(0.009)

(0.038)

(0.002)

(0.002)

(0.015)

(0.057)

(0.003)

(0.003)

(0.012)

(0.051)

(0.002)

(0.002)

Dependencyratio

0.069

�0:114

�0:008

0.001

0.104

0.354

0.017

0.009

0.047

�0:398

�0:019

�0:004

(0.049)

(0.288)

(0.012)

(0.005)

(0.085)

(0.496)

(0.021)

(0.011)

(0.058)

(0.348)

(0.013)

(0.005)

Fixed

effects

District-levelfixedeffect

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Constant

�20:686***

�0:417***

�24:316***

�0:503**

�21:305***

�0:560***

(1.896)

(0.112)

(3.508)

(0.222)

(2.252)

(0.070)

Num

berof

Observatio

ns12,178

7,957

12,178

12,178

4,382

2,947

4,382

4,382

7,796

5,010

7,796

7,796

CWDs¼

child

renwith

disabilities,OLS¼

ordinary

leastsquares.

Notes:Stand

arderrors

arein

parentheses.*,

**,and**

*representsign

ificanceat

the10

%,5%

,and1%

levels,respectiv

ely.Stand

arderrors

ofcolumns

(1)

(binaryprob

it),(2)

(con

ditio

nalOLS),and(4)(uncon

ditio

nalO

LS)arerobu

standclusteredby

householdto

consider

thepo

ssibility

thatresidu

alswith

inthe

sameho

useholdarelik

elyto

becorrelated.For

column(3)(resultof

combinedmarginaleffect:prob

itþcond

ition

alOLS),thestandard

errorisob

tained

bybo

otstrapp

ingwith

400replications.Dependent

variable

inprob

itmod

el(colum

n(1))analysisisabinary

variable

ofwhether

aho

useholdhadeducational

expend

iture

inthepast12

mon

ths.The

naturallogof

shareof

educationalexpend

iture

andabsolutevalueof

shareof

educationalexpend

iture

aredepend

ent

variablesof

columns

(2)and(4),respectiv

ely.

Allequatio

nsinclud

edistrict-level

fixedeffects.

Sou

rce:

Autho

rs’calculations

usingdata

from

the20

10BangladeshHou

seho

ldIncomeExp

enditure

Survey.



children. This finding partly suggests that if a mother receives more education, she is

likely to allow her children to receive more and better education. Takeda and

Lamichhane (2018) found that mother’s education can be an important predictor for

CWDs to receive better education, claiming that strong female bargaining power can

increase their understanding of CWDs in the household. Regardless of disability, some

studies have also discussed the positive effect of the strong bargaining power of

women (Thomas 1994, Duflo and Udry 2004).

The dummy variable of household head with disabilities also provides interesting

results. While it shows an insignificant effect on the probability of children’s

enrollment in the analysis using the entire sample (i.e., all areas in Bangladesh)

(column [1]), this variable is found to have a negative effect on educational

expenditure at the 10% significance level (column [2]). Similarly, the dummy variable

of disabled household head is positively correlated (column [1]) with child enrollment

at the 10% significance level in the probit model, while it negatively affects the

conditional share of expenditure at the 5% level (column [2]) in urban areas. These

interesting findings suggest that if the household head has some form of disability,

they are more likely to understand the value of access to education and thus enroll their

children in schools. At the same time, as PWDs are more likely to experience poverty,

they may not be able to afford educational fees once their child is enrolled. Finally, the

dummy variable for household head as full-time wage earner is a strong predictor of

educational expenditure, again suggesting that the income stability of parents and

household heads is important in terms of investing in their children’s education.

We test the DME for disability-age variables for each school-aged group to

confirm whether a biased allocation of parental investment toward CWDs exists

(Table 6). The DME shows a difference in the value of non-CWDs minus the

marginal effect of CWDs. We test whether this difference is statistically significant

(equation [2]) and observe a clear bias against CWDs in the results for the entire

sample. In column (1), which examines disability bias in enrollment using the probit

model analysis, the DME of children aged 5–9 years and 10–14 years are positive and

statistically significant at the 1% level. This finding suggests the likelihood of bias in

age for primary and secondary education. In column (2), which is the conditional OLS

regression model, no significant results for the DME are observed.

We attempt to identify biases in educational expenditure by restricting to

households with only enrolled children but found insignificant results. This finding

suggests that the bias or discrimination for CWDs exists at the stage of deciding

whether to enroll them in school, but once they are enrolled, they may not experience

significant bias. In column (3), which shows the CME of the probit and conditional



Table

6.Summaryof

Difference

inMarginal

Effect�10

0of

Hou

sehold-Level

Analysis

AllAreas

Urban

Rural

Probit

Con

di-

tion

alOLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Uncondi-

tion

alOLS

Probit

Con

di-

tion

alOLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Uncondi-

tion

alOLS

Probit

Con

di-

tion

alOLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Uncondi-

tion

alOLS


(1)

(2)

(3)

(4)

(1)

(2)

(3)

(4)

(1)

(2)

(3)

(4)

Childrenaged

5–9years

41.59***

�3:75

2.84**

2.12

63.62***

�40:40

6.17**

3.93

26.27*

�1:19

1.23

0.61

(0.001)

(0.943)

(0.034)

(0.126)

(0.008)

(0.584)

(0.041)

(0.151)

(0.064)

(0.985)

(0.440)

(0.706)

Childrenaged

10–14

years

33.58***

33.51

2.68*

2.83

3.26

113.27*

3.11

4.89

45.79***

�0:43

2.29

2.29

(0.008)

(0.371)

(0.086)

(0.140)

(0.869)

(0.098)

(0.317)

(0.211)

(0.003)

(0.992)

(0.183)

(0.257)

Childrenaged

15–19

years

�12:61

36.51

1.55

1.51

7.52

�135

:47

0.12

0.62

�18:18

99.4

1.48

1.10

(0.295)

(0.494)

(0.432)

(0.501)

(0.714)

(0.066)

(0.978)

(0.883)

(0.199)

(0.129)

(0.507)

(0.681)

OLS¼

ordinary

leastsquares.

Notes:Eachvalueshow

sthedifference

inthemarginaleffect

(DME)of

disability-agecatego

ryvariables,which

ismultip

liedby

100.

The

DMEisthe

difference

inthecoefficientor

marginaleffect

ofachild

with

disabilitiesandachild

with

outdisabilities(¼

non-CWDs–CWDs)

foreach

agecatego

ry;

positiv

evalueshow

stheexistenceof

disabilitybias.The

figu

resin

parenthesesareP-valuesof

theF-testthat

DMEis

equalto

zero.*,

**,and**

*representsign

ificancein

gapby

disabilityat

the10

%,5%

,and1%

levels,respectiv

ely.

Sou

rce:

Autho

rs’calculations

usingdata

from

the20

10BangladeshHou

seho

ldIncomeExp

enditure

Survey.



OLS models, the DMEs in cohorts of children aged 5–9 years and 10–14 years are

positive and significant at the 5% and 10% levels, respectively, while in column (4) for

the unconditional model, no cohorts show significant effects. Again, these results

suggest that there is a bias mainly in the enrollment decision; the significant difference

is reflected in the probit model and CME. Moreover, disability bias is found at the

primary and secondary levels for all areas, indicating that such bias exists at the stage

of providing basic education.

Parents’ investment motives regarding children’s education could reflect unequal

allocations to the differential returns of CWDs and non-CWDs. We assume that

parents’ expectation of returns to education of CWDs is much lower than that of non-

CWDs (i.e., the expected contribution to household income is lower). This gap in

parents’ expectation of returns is larger than the gap between boys and girls, even in

developing countries, as parents expect to face more physical and institutional barriers

when they raise CWDs compared to non-CWDs (including girls). Findings for both

rural and urban areas consistently suggest the likelihood of biases, mainly in the

decision whether to enroll their child. Unlike the results in rural areas, we did not find

significant differences for secondary level enrollment for CWDs in urban areas.

Additionally, although it is important to identify which impairment groups are

more vulnerable to household investment disparities for their schooling and which

impairments drive the disability bias, due to the smaller sample size, we are not able to

perform statistical analysis. Nonetheless, we want to see descriptively if there exist any

gaps among different impairment groups. We have presented this descriptive table in

the appendix. Except for the primary level, individuals with visual impairments have

higher enrollment rates among different disability types: 51% and 31% for secondary

and higher-level enrollment, respectively. For enrollment at the primary level,

participants who are deaf and hard of hearing have higher enrollment rates (57%).

Likewise, the enrollment rate for participants with physical impairments (29%, 14%,

and 0%), cognitive difficulties (17%, 7%, and 7%), and communication difficulties

(12%, 6%, and 0%) is observed for primary, secondary, and higher education,

respectively. Furthermore, for literacy in reading and writing, which are considered

important components for academic achievement, participants with visual

impairments, hearing difficulties, and physical impairments have higher literacy

rates than those with cognitive and communication difficulties. Literacy rates of 71%,

49%, 39%, 22%, and 22% are observed, respectively, for participants with visual

impairments, hearing impairments, physical impairments, cognitive difficulties, and

communication difficulties. As Lamichhane (2013) and Takeda and Lamichhane

(2018) acknowledge, our findings indicate that if the different needs of each disability



group are not addressed, then children with severe difficulties may face significant

barriers in education.

For writing literacy, we find a similar trend: groups of children with visual or

hearing impairments (98%), physical impairments (95%), cognitive difficulties (95%),

and communication difficulties (96%). Though we find gaps for enrollment depending

on the disability type, we cannot conclude if they are consistent as the already small

sample size further decreases at the secondary and higher education levels, thus not

allowing us to explore it further. Similarly, the question on reading and writing literacy

asked only whether the participant can read or write a letter or not. These are plausible

reasons for the higher performance of individuals with different impairment groups in

reading and writing. Furthermore, though we find an enrollment disparity for girls

versus boys with disabilities at the primary level, based on our econometric analysis,

we see an improvement in gender parity in the enrollment of girls at the primary

education level, which can be attributed to various demand-side interventions taking

place in Bangladesh.

C. Regression Results of Individual-Level Analysis

1. Reaffirming Disability Bias

We run individual-level regressions and compare the results of household-level

analysis with individual child data. Table 7 reports the results of individual-level

analysis for children aged 5–9 years (entire sample). In addition to the impairment

dummy, we prepared two variables that explain the timing of acquiring disability:

impairment at birth and impairment acquired during enrollment. These are

incorporated into our estimation with the aim of identifying differences in disability-

based bias toward CWDs depending on when the impairment first occurred. As these

dummy variables can be used as interaction terms between impairment dummy and

impairment timing, the results (coefficients and marginal effects) need to be interpreted

with a combination of these variables. We also present the results of estimation with

and without interaction term variables between other individual and household

characteristics (i.e., interactions with female, parents’ education, and household head

is full-time wage worker).

In Table 7 (primary level children), the impairment dummy has a negative effect

on the probability of enrolling in elementary school at the 1% significance level

(column [1]). This finding suggests that CWDs are less likely to enroll in primary

school due to bias on the part of their parents, who may not consider their CWDs as a

subject for investment in education. In addition, the dummy variable of impairment



Table

7.Resultsof

Individual-Level

Analysis(A

llAreas:Aged5–9Years)

Aged5–

9years

Probit

Probitwith

Interaction

Con

ditional

OLS

Con

ditional

OLSwith

Interaction

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Com

bined

Margina

l

Effect

(Probitþ

Con

ditional

OLS)with

Interaction

Uncond

itional

OLS

Uncond

itional

OLSwith

Interaction


(1)

(1)

(2)

(2)

(3)

(3)

(4)

(4)

Log

ofmon

thly

expenditu

repercapita

0.09

7***

0.097*

**1.022*

**1.02

3***

1.391*

**1.39

3***

2,94

0.078*

**2,94

3.689*

**

(0.015

)(0.015

)(0.049

)(0.049

)(0.094

)(0.095

)(245

.002)

(244

.848)

Log

ofho

useholdsize

0.00

70.007

0.002

0.00

00.047

0.04

513

8.16

413

1.26

2

(0.019

)(0.019

)(0.070

)(0.070

)(0.120

)(0.121

)(242

.910)

(242

.565)

Individual

characteristics

Fem

ale

0.02

1**

0.021*

*�0

:083**

�0:078

**0.074

0.07

5�1

34:956

�131:214

(0.010

)(0.010

)(0.033

)(0.033

)(0.073

)(0.073

)(96.69

6)(98.36

2)

Age

0.13

7***

0.137*

**0.213*

**0.211*

**1.047*

**1.04

5***

460.52

4***

456.61

4***

(0.003

)(0.003

)(0.014

)(0.014

)(0.024

)(0.024

)(33.41

5)(33.41

2)

Impairment

�0:155

***

�0:177**

0.025

0.78

5**

�0:982**

�0:557

�497:632

830.84

3**

(0.056

)(0.077

)(0.288

)(0.323

)(0.455

)(0.666

)(347

.111)

(373

.028)

Impairmentby

birth

�0:155

*�0

:160*

�0:326

� 0:591

*�1

:246*

�1:479**

�346:848

�776:135**

(0.085

)(0.087

)(0.417

)(0.338

)(0.662

)(0.664

)(449

.503)

(377

.773)

Impairmentacquired

during

enrollm

ent

0.181*

0.199*

�0:25

�0:552

*0.984

0.87

4�3

35:077

�742:783**

(0.104

)(0.109

)(0.351

)(0.310

)(0.699

)(0.793

)(500

.550)

(376

.489)

Con

tinued.



Table

7.Con

tinued.

Aged5–

9years

Probit

Probitwith

Interaction

Con

ditional

OLS

Con

ditional

OLSwith

Interaction

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Com

bined

Margina

l

Effect

(Probitþ

Con

ditional

OLS)with

Interaction

Uncond

itional

OLS

Uncond

itional

OLSwith

Interaction


(1)

(1)

(2)

(2)

(3)

(3)

(4)

(4)

Hou

seholdcharacteristics

Dependencyratio

0.023

0.024

�0:436**

*�0

:432

***

�0:174

�0:166

607.67

962

9.91

3

(0.044

)(0.044

)(0.153

)(0.153

)(0.292

)(0.294

)(452

.140)

(452

.520)

Mother’syearsof

schooling

0.010***

0.009***

0.041***

0.041***

0.095***

0.091***

134.451***

133.586***

(0.002

)(0.002

)(0.006

)(0.006

)(0.013

)(0.012

)(20.36

2)(20.66

3)

Father’syearsof

schooling

0.006***

0.006***

0.036***

0.037***

0.063***

0.066***

102.469***

106.438***

(0.002

)(0.002

)(0.006

)(0.006

)(0.012

)(0.012

)(14.71

8)(14.91

3)

Hou

seho

ldhead

isfemale

�0:013

�0:013

0.093

0.09

3�0

:015

�0:012

�195:44

�197:052

(0.038

)(0.038

)(0.119

)(0.119

)(0.264

)(0.264

)(310

.046)

(310

.650)

Hou

seho

ldhead

with

disabilities

0.00

10.002

�0:086*

�0:083

�0:055

�0:051

172.89

917

9.37

7

(0.017

)(0.017

)(0.052

)(0.052

)(0.111)

(0.113

)(260

.389)

(260

.511)

Household

head

isfulltim

ewageworker

0.013

0.011

0.115**

0.115**

0.168*

0.160*

202.648

199.823

(0.013

)(0.013

)(0.046

)(0.046

)(0.091

)(0.092

)(156

.478)

(158

.880)

Muslim

0.02

60.026

�0:216**

*�0

:214

***

0.008

0.00

6�4

22:647**

�421:178**

(0.017

)(0.017

)(0.064

)(0.063

)(0.111)

(0.111)

(192

.789)

(192

.856)

Interactionterm

s

Impairment�

Fem

ale

0.004

�0:213

�0:133

108.49

(0.080

)(0.288

)(0.677

)(331

.699)

Con

tinued.



Table

7.Con

tinued.

Aged5–

9years

Probit

Probitwith

Interaction

Con

ditional

OLS

Con

ditional

OLSwith

Interaction

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Com

bined

Margina

l

Effect

(Probitþ

Con

ditional

OLS)with

Interaction

Uncond

itional

OLS

Uncond

itional

OLSwith

Interaction


(1)

(1)

(2)

(2)

(3)

(3)

(4)

(4)

Impairment�

Mother’syearsof

schooling

0.016

�0:005

0.10

1�7

5:80

6

(0.013

)(0.039

)(0.118

)(60.56

6)

Impairment�

Father’syearsof

schooling

�0:013

�0:139

***

�0:188

�276:359**

*

(0.013

)(0.048

)(0.120

)(53.77

2)

Impairment�

Hou

seho

ldhead

is

fulltim

ewageworker

0.062

�0:21

0.24

4�1

82:221

(0.108

)(0.361

)(0.950

)(444

.352)

Fixed

effect


Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Constant

�2:178**

*�2

:166

***

�23,92

4:92

3***

�23,91

6:15

1***

(0.433

)(0.433

)(2,259

.818

)(2,256

.811)

Num

berof

Observatio

ns5,694

5,694

4,259

4,259

5,694

5,694

5,694

5,694

OLS¼

ordinary

leastsquares.

Notes:Stand

arderrors

arein

parentheses.*,

**,and**

*representsign

ificancein

themarginaleffectsof

independ

entvariablesat

the10

%,5%

,and1%

levels,respectiv

ely.

Stand

arderrors

ofcolumn(1)(binaryprob

it),column(2)(con

ditio

nalOLS),andcolumn(4)(uncon

ditio

nalOLS)arerobu

stand

clusteredby

householdto

consider

thepo

ssibility

that

residu

alswith

inthesameho

useholdarelik

elyto

becorrelated.For

column(3)(resultof

combined

marginaleffect:prob

itþcond

ition

alOLS),thestandard

errorisob

tained

bybo

otstrapp

ingwith

400replications.T

hedepend

entv

ariablein

theprob

itmod

el(colum

n(1))analysisisabinary

variableof

whether

aho

useholdhashadanyeducationalexp

enditure

inthepast12

mon

thsforeach

child

.The

naturallog

ofeducationalexp

enditure

andtheabsolutevalueof

educationalexp

enditure

foreach

child

arethedepend

entv

ariables

ofcolumns

(2)and(4),respectiv

ely.All

equatio

nsinclud

edistrict-level

fixedeffects.

Sou

rce:

Autho

rs’calculations

usingdata

from

the20

10BangladeshHou

seho

ldIncomeExp

enditure

Survey.



from birth has a negative effect and the dummy variable of impairment while in

education has a positive effect at the 10% significance level. This finding suggests that

fewer children with congenital impairments who are enrolled in education are more

likely to experience disability-based bias compared to children acquiring impairment

during primary school age. For conditional OLS, significant results are shown only for

estimations with interaction terms. The impairment dummy has a positive effect on

educational expenditure for enrolled CWDs at the 5% significance level. Both the

dummy variables of impairment at birth and acquired during enrollment have negative

effects at the 10% significance level. These results suggest that even if children have

impairments, enrolled CWDs can receive more educational investment. We can

assume that such investment includes extra expenditure particular to their needs.

However, the degree of these extra investments decreases for children who are

impaired from birth or acquire it during primary school age. Importantly, CME shows

different results from unconditional OLS and can appropriately extract the negative

disability-based bias. The disability dummy in the result of CME without interaction

terms is negatively significant at the 5% level, while that of unconditional OLS

without interaction terms is not significant. Combined with the results of the dummy

variable of impairment at birth, results are consistent with the fact that children with

congenital impairments face disability bias more than those who acquire it later in life.

On the other hand, considering both the enrollment decision and how much to spend

for enrolled children, we find that the major difficulty for CWDs is having the

opportunity to enroll in schools. Once enrolled, parents are likely to invest in them as

results show that the impairment dummy is negative regarding enrollment but positive

or not significant for conditional educational expenditure.

Tables 8 and 9 report the results for children aged 10–14 years and 15–19 years,

respectively. The disability dummy in Table 8 is negatively correlated with individual

educational expenditure for all estimations except for unconditional OLS without

interaction terms. Compared with the results in Table 7 (primary school age), the

results in Table 8 suggest that CWDs are disadvantaged, having a significantly lower

probability of enrollment given the lower level of investment in education by their

parents. Based on this finding, we cannot reject the possibility of disability bias even if

CWDs are enrolled in secondary school. Due to disability-related biases, parents are

likely to consider that education is less important for CWDs. Therefore, in terms of

investment, since secondary education is not compulsory in Bangladesh, it is likely

that parents do not feel an obligation to educate CWDs beyond primary level. Finally,

in Table 9, we obtain no significant results regarding these disability-related variables,

indicating that hardly any disability bias exists in higher education.



Table8.

Resultsof

Individual-Level

Analysis(A

llAreas:Aged10–1

4Years)

Aged10–14years

Probit

Probit

Con

ditional

OLS

Con

ditional

OLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Com

bined

Margina

lEffect

(Probitþ

Con

ditional

OLS)

Unconditional

OLS

Uncond

itional

OLS


(1)

(1)

(2)

(2)

(3)

(3)

(4)

(4)

Log

ofmon

thly

expenditu

repercapita

0.143*

**0.14

6***

0.971*

**0.97

1***

1.904*

**1.920*

**2,940.07

8***

5,26

8.570*

**(0.014

)(0.014

)(0.038

)(0.038

)(0.108

)(0.107

)(245.002

)(378

.583)

Log

ofho

useholdsize

0.017

0.01

80.117*

*0.118*

*0.231*

0.239*

138.16

478

1.60

6**

(0.017

)(0.017

)(0.051

)(0.051

)(0.124

)(0.124

)(242.910

)(351

.163)

Individual

characteristics

Fem

ale

0.097*

**0.09

9***

�0:035

�0:039

0.702*

**0.716*

**�1

34:956

230.39

9(0.009

)(0.009

)(0.026

)(0.026

)(0.072

)(0.073

)(96.69

6)(147

.094)

Age

�0:047**

*�0

:048

***

0.223*

**0.22

3***

�0:171**

*�0

:173**

*46

0.52

4***

552.72

5***

(0.003

)(0.003

)(0.009

)(0.009

)(0.026

)(0.027

)(33.41

5)(47.34

0)Im

pairment

�0:227**

*�0

:156

***

�0:393**

�0:498

***

�2:049**

*�1

:597**

*�4

97:632

�1,636:995**

(0.052

)(0.050

)(0.158

)(0.173

)(0.407

)(0.408

)(347.111)

(735

.784)

Impairmentat

birth

�0:019

�0:017

�0:205

�0:111

�0:318

�0:219

�346:848

�996:727

(0.068

)(0.062

)(0.250

)(0.236

)(0.545

)(0.509

)(449.503

)(837

.842)

Impairmentacquired

during

enrollm

ent

0.217***

0.210***

0.339*

0.351**

1.928***

1.889***

�335:077

1,99

9.258*

**(0.063

)(0.056

)(0.189

)(0.169

)(0.517

)(0.482

)(500.550

)(753

.266)

Con

tinued.



Table

8.Con

tinued.

Aged10–14years

Probit

Probit

Con

ditional

OLS

Con

ditional

OLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Com

bined

Margina

lEffect

(Probitþ

Con

ditional

OLS)

Unconditional

OLS

Uncond

itional

OLS


(1)

(1)

(2)

(2)

(3)

(3)

(4)

(4)

Hou


Dependencyratio

�0:087**

�0:085

**0.226*

*0.22

2**

�0:466*

�0:460

607.67

91,41

0.162*

*(0.036

)(0.036

)(0.105

)(0.105

)(0.281

)(0.283

)(452.140

)(672

.861)

Mother’syearsof

schooling

0.015***

0.016***

0.031***

0.030***

0.141***

0.146***

134.451***

270.593***

(0.002

)(0.002

)(0.005

)(0.005

)(0.015

)(0.016

)(20.36

2)(34.30

7)Father’syearsof

schooling

0.010*

**0.011*

**0.030*

**0.03

1***

0.103*

**0.107*

**10

2.46

9***

182.87

1***

(0.002

)(0.002

)(0.004

)(0.004

)(0.012

)(0.013

)(14.71

8)(24.97

3)Hou

seho

ldhead

isfemale

0.006

0.00

2�0

:125

�0:123

�0:06

�0:087

�195:440

698.66

4(0.040

)(0.040

)(0.133

)(0.133

)(0.349

)(0.350

)(310.046

)(1,315

.242

)Hou

seho

ldhead

with

disabilities

�0:026**

�0:025

*�0

:036

�0:036

�0:225**

�0:217**

172.89

958

.622

(0.013

)(0.013

)(0.038

)(0.038

)(0.107

)(0.107

)(260.389

)(238

.451)

Household

head

isfulltim

ewageworker

�0:005

0.00

00.087*

*0.07

8**

0.038

0.062

202.64

872

1.66

3***

(0.013

)(0.013

)(0.035

)(0.036

)(0.106

)(0.108

)(156.478

)(226

.229)

Muslim

�0:003

�0:003

�0:135**

*�0

:136

***

�0:137

�0:136

�422:647

**�3

59:698

(0.015

)(0.015

)(0.043

)(0.043

)(0.123

)(0.123

)(192.789

)(222

.773)

Interactionterm

sIm

pairment�

Fem

ale

�0:048

0.18

4�0

:209

194.38

8(0.048

)(0.162

)(0.405

)(760

.037)

Con

tinued.



Table

8.Con

tinued.

Aged10–14years

Probit

Probit

Con

ditional

OLS

Con

ditional

OLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Com

bined

Margina

lEffect

(Probitþ

Con

ditional

OLS)

Unconditional

OLS

Uncond

itional

OLS


(1)

(1)

(2)

(2)

(3)

(3)

(4)

(4)

Impairment�

Mother’syearsof

schooling

�0:009

0.01

2�0

:059

51.821

(0.009

)(0.024

)(0.076

)(167

.526)

Impairment�

Father’syearsof

schooling

�0:006

�0:029

*�0

:069

�209:280**

(0.007

)(0.015

)(0.068

)(99.17

1)Im

pairment�

Household

head

isfulltim

ewageworker

�0:094

0.43

3**

�0:344

852.63

2(0.071

)(0.221

)(0.594

)(1,247

.332

)

Fixed

effect


Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Constant

�2:287**

*�2

:271

***

�45,15

4:36

7***

�45,13

3:66

9***

(0.374

)(0.374

)(3,356

.073

)(3,371

.935

)Num

berof

Observatio

ns5,44

15,441

4,58

64,586

5,441

5,441

5,44

15,441

OLS¼

ordinary

leastsquares.

Notes:S

tand

arderrorsarein

parentheses.*,

**,and

***representsignificancein


independ

entv

ariables

atthe10

%,5

%,and

1%levels,

respectiv

ely.

Stand

arderrors


it),column(2)(con

ditio


ditio

nalOLS)arerobu

standclusteredby

householdto

consider

thepo

ssibility

thatresidu

alswith

inthesameho

useholdarelik

elyto

becorrelated.F

orcolumn(3)(resultof

combinedmarginaleffect:

prob

itþcond

ition

alOLS),thestandard

erroris

obtained

bybo

otstrapp

ingwith

400replications.The

depend

entvariable

intheprob

itmod

el(colum

n(1))

analysis

isabinary

variable

ofwhether

aho

useholdhadanyeducationalexpend

iture

inthepast

12mon

thsforeach

child

.The

naturallogof

educational

expend

iture

andtheabsolute

valueof

educationalexpend

iture

foreach

child

arethedepend

entvariablesof

columns

(2)and(4),respectiv

ely.

Allequatio

nsinclud

edistrict-level

fixedeffects.

Sou

rce:

Autho

rs’calculations

usingdata

from

the20

10BangladeshHou

seho

ldIncomeExp

enditure

Survey.



Table9.

Resultsof

Individual-Level

Analysis(A

llAreas:Aged15–1

9Years)

Aged15–19years

Probit

Probit

Con

ditional

OLS

Con

ditional

OLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Com

bined

Margina

lEffect

(Probitþ

Con

ditional

OLS)

Unconditional

OLS

Uncond

itional

OLS


(1)

(1)

(2)

(2)

(3)

(3)

(4)

(4)

Log

ofmon

thly

expenditu

repercapita

0.235*

**0.23

5***

0.854*

**0.85

3***

2.493*

**2.494*

**10

,358

.898

***

10,387

.140

***

(0.021

)(0.021

)(0.049

)(0.049

)(0.172

)(0.172

)(2,058

.823

)(2,064

.466

)Log

ofho

useholdsize

0.053*

0.05

6**

0.269*

**0.26

5***

0.603*

**0.624*

**1,782.711*

*1,75

9.864*

*(0.028

)(0.028

)(0.067

)(0.068

)(0.223

)(0.225

)(741.187

)(738

.537)

Individual

characteristics

Fem

ale

0.131*

**0.13

5***

�0:119**

*�0

:124

***

1.082*

**1.112*

**�1

76:560

�289:628

(0.014

)(0.014

)(0.033

)(0.033

)(0.126

)(0.128

)(672.459

)(699

.117

)Age

�0:076**

*�0

:076

***

0.074*

**0.07

4***

�0:621**

*�0

:621**

*�8

7:22

6�8

5:71

8(0.005

)(0.005

)(0.013

)(0.013

)(0.044

)(0.044

)(186.921

)(188

.112

)Im

pairment

�0:081

�0:034

0.121

0.01

7�0

:646

�0:291

1688.746

1391.512

(0.084

)(0.088

)(0.280

)(0.320

)(0.810

)(0.934

)(1,287

.444

)(1,615

.110

)Im

pairmentat

birth

�0:109

�0:026

�0:346

�0:270

�1:129

�0:363

�676:478

�989:758

(0.113

)(0.115

)(0.401

)(0.405

)(1.060

)(1.182

)(3,049

.862

)(2,725

.554

)Im

pairmentacquired

during

enrollm

ent

0.104

0.183*

0.051

0.044

0.930

1.615

�830:625

�151:496

(0.106

)(0.107

)(0.307

)(0.303

)(1.045

)(1.140

)(2,032

.072

)(1,914

.476

)

Con

tinued.



Table

9.Con

tinued.

Aged15–19years

Probit

Probit

Con

ditional

OLS

Con

ditional

OLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Com

bined

Margina

lEffect

(Probitþ

Con

ditional

OLS)

Unconditional

OLS

Uncond

itional

OLS


(1)

(1)

(2)

(2)

(3)

(3)

(4)

(4)

Hou


Dependencyratio

�0:046

�0:050

�0:008

�0:009

�0:406

�0:440

1,103.66

31,19

0.330

(0.051

)(0.051

)(0.126

)(0.126

)(0.423

)(0.424

)(1,319

.547

)(1,316

.366

)Mother’syearsof

schooling

0.022***

0.022***

0.021***

0.020***

0.205***

0.205***

748.261***

741.794***

(0.003

)(0.003

)(0.006

)(0.006

)(0.027

)(0.027

)(205.037

)(209

.653)

Father’syearsof

schooling

0.014*

**0.01

4***

0.015*

**0.01

5***

0.126*

**0.127*

**22

1.17

2***

231.85

1***

(0.002

)(0.002

)(0.005

)(0.005

)(0.020

)(0.020

)(70.52

8)(69.16

7)Hou

seho

ldhead

isfemale

�0:045

�0:046

�0:093

�0:090

�0:443

�0:448

�2,279:240

�244

5:86

(0.097

)(0.097

)(0.189

)(0.190

)(0.954

)(0.958

)(2,079

.269

)(2,008

.789

)Hou

seho

ldhead

with

disabilities

�0:024

�0:022

0.051

0.04

9�0

:183

�0:168

146.67

814

4.70

3(0.020

)(0.020

)(0.044

)(0.044

)(0.163

)(0.163

)(522.590

)(513

.782)

Household

head

isfulltim

ewageworker

0.007

0.011

0.086**

0.090**

0.102

0.144

�176:290

�76:71

1(0.019

)(0.019

)(0.040

)(0.040

)(0.163

)(0.167

)(1,404

.275

)(1,441

.509

)Muslim

�0:002

�0:002

�0:129**

*�0

:130

***

�0:082

�0:087

�814:591

�786:218

(0.022

)(0.023

)(0.050

)(0.050

)(0.198

)(0.200

)(630.650

)(633

.883)

Interactionterm

sIm

pairment�

Fem

ale

�0:133

0.31

9�0

:997

4,20

5.383*

(0.087

)(0.240

)(0.930

)(2,315

.953

)

Con

tinued.



Table

9.Con

tinued.

Aged15–19years

Probit

Probit

Con

ditional

OLS

Con

ditional

OLS

Com

bined

Marginal

Effect

(Probitþ

Con

ditional

OLS)

Com

bined

Margina

lEffect

(Probitþ

Con

ditional

OLS)

Unconditional

OLS

Uncond

itional

OLS


(1)

(1)

(2)

(2)

(3)

(3)

(4)

(4)

Impairment�

Mother’syearsof

schooling

0.00

70.04

70.082

251.39

1(0.020

)(0.032

)(0.219

)(598

.876)

Impairment�

Father’syearsof

schooling

�0:01

�0:024

�0:099

�361:507

(0.012

)(0.026

)(0.131

)(243

.893)

Impairment�

Household

head

isfulltim

ewageworker

�0:137

�0:313

�1:354

�5,084:805

(0.110

)(0.264

)(1.266

)(3,437

.347

)

Fixed

effect


Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Constant

0.365

0.39

2�7

7,36

6:44

1***

�77,53

7:62

6***

(0.478

)(0.477

)(18,01

6.46

6)(18,08

4.26

3)Num

berof

Observatio

ns3,59

83,598

1,88

31,883

3,598

3,598

3,59

83,598

OLS¼

ordinary

leastsquares.

Notes:S

tand

arderrorsarein

parentheses.*,

**,and

***representsignificancein


independ

entv

ariables

atthe10

%,5

%,and

1%levels,

respectiv

ely.

Stand

arderrors


it),column(2)(con

ditio


ditio

nalOLS)arerobu

standclusteredby

householdto

consider

thepo

ssibility

thatresidu

alswith

inthesameho

useholdarelik

elyto

becorrelated.F

orcolumn(3)(resultof

combinedmarginaleffect:

prob

itþcond

ition

alOLS),thestandard

erroris

obtained

bybo

otstrapp

ingwith

400replications.The

depend

entvariable

intheprob

itmod

el(colum

n(1))

analysis

isabinary

variable

ofwhether

aho

useholdhadanyeducationalexpend

iture

inthepast

12mon

thsforeach

child

.The

naturallogof

educational

expend

iture

andtheabsolute

valueof

educationalexpend

iture

foreach

child

arethedepend

entvariablesof

columns

(2)and(4),respectiv

ely.

Allequatio

nsinclud

edistrict-level

fixedeffects.

Sou

rce:

Autho

rs’calculations

usingdata

from

the20

10BangladeshHou

seho

ldIncomeExp

enditure

Survey.



2. Disability Bias for Girls

While analyzing girls and disability, we generally find pro-female bias in

enrollment at all levels of education and pro-male bias in investment on education for

enrolled children in primary and higher education. This might reflect the evidence that

the returns to education for girls are the same as or even higher than for boys

(Behrman and Deolalikar 1995), and therefore, parents might allow girls to enroll in

school. Additionally, recent and ongoing global initiatives for girls’ education in line

with the United Nations Sustainable Development Goals and other related plans could

have contributed to increasing access to education for girls. These findings are

consistent with previous studies on intrahousehold gender bias such as Azam and

Kingdon (2013) and Kingdon (2005). Though Bangladesh is expanding access to

basic education for girls, when compared to their male counterparts, girls may still face

challenges to continuing their education as results show a negative effect on

educational expenditure for enrolled children. The difference between girls and CWDs

is that girls are found to not face gender bias for enrollment, while disability bias exists

for enrollment and even beyond the primary level of education. Moreover, girls with

disabilities are expected to face severe discrimination and are regarded as being in a

disadvantageous position compared to boys with disabilities, as the former may

experience both gender and disability bias. Although we tried to examine it by

incorporating an interaction term between the female dummy and impairment dummy,

we generally could not obtain significant findings to support this argument.

3. The Effects of Parental Education on Disability

Unlike other studies that have shown that parents’ years of schooling generally

have a positive effect on a child’s education (Behrman and Deolalikar 1995, Cameron

and Heckman 1998), we find a negative correlation between a CWD’s father’s

education and educational expenditure on enrolled children, whereas no correlation is

observed between a CWD’s mother’s education and child enrollment. Takeda and

Lamichhane (2018) found that in India the mother’s, but not necessarily the father’s,

education level can be an important predictor of the school enrollment of CWDs. In

this sense, investment in girls’ education is crucial and should be increased as these

girls will become mothers who affect enrollment decisions in the future.

Finally, we also incorporate a dummy variable of whether the household head is

a fulltime wage earner, as we do in the household-level analysis. The interaction terms

between the impairment dummy and this variable show a positive effect on educational

expenditure for CWDs aged 10–14 years in the entire sample. This finding indicates



that the income stability of the household head is connected to a greater probability of

their CWDs receiving more investment in secondary education.

4. Bargaining Power of Persons with Disabilities

Finally, we test whether PWDs having decision-making power in the household

helps improve the educational situation of their children. This is with the casual

observation that individuals with disabilities who are in a decision-making position

may have a deep and fair understanding of the importance of education for all children

regardless of disability status. Following Masterson (2012), we incorporate a variable

to test female bargaining power or objective bias. We incorporate a dummy variable of

female household head as a proxy for female bargaining power.

Although previous studies demonstrate that strong female bargaining power is

crucial to improving their children’s education (Thomas 1994, Duflo and Udry 2004),

we mostly find no significant results in our individual-level analysis, whereas the

results are positive and significant in the household-level analysis. We obtain

significant results for the variable for household head with disabilities. In Table 7, the

dummy variable for household head with disabilities is negative in conditional OLS

(without interaction), suggesting the difficulty for such household heads regarding

educational spending even if their children are already enrolled.

In Table 8, this variable is negatively correlated with the enrollment decision, as

shown in the probit model. Again, this finding suggests that household heads with

disabilities face difficulty sending their children to secondary school itself. This finding

is consistent with the educational situation in Bangladesh in which primary education

up to grade 5 is free and compulsory, enabling parents to send their children to school

at low cost. Once their children are in secondary level education, they have to pay

costs such as tuition. Financing for education may, therefore, be a challenge for rural

household heads with a disability who may be struggling with poverty. These results

lead to our interpretation that the bargaining power of household heads with a

disability is conditional on income stability. Regardless of how deep their

understanding of the importance of education for their children is, they cannot have

bargaining power without being financially stable themselves. Some of the findings in

our individual-level analysis that show a negative effect of household heads with

disabilities on enrollment support this argument.

5. Household Fixed Effect

By controlling all household-level characteristics with a household fixed-effects

model, we rigorously check whether disability bias exists. As shown in Table 10, we



Table

10.

Summaryof

Marginal

Effectof

Impairm

entDummywithHou

seholdFixed

Effects

All

Urban

Rural

Probit

Con

ditional

OLS

Unconditional

OLS

Probit

Con

ditional

OLS

Unconditional

OLS

Probit

Con

ditional

OLS

Unconditional

OLS

Age

Category

(1)

(2)

(3)

(1)

(2)

(3)

(1)

(2)

(3)

Childrenaged

5–9years

�0:478**

*�0

:012

�325

:664

�0:427**

*�0

:274

�869:829

�0:433

***

�0:437

�419:577**

(0.110

)(0.437

)(231

.414

)(0.084

)(0.255

)(865

.247

)(0.123

)(0.293

)(204

.000

)Childrenaged

10–14

years

�0:332**

*�0

:350

�1,522

:422

**�0

:428**

*�0

:467

�2,158:441*

�0:185

***

�0:437

�1,113:266*

(0.074

)(0.232

)(628

.317

)(0.080

)(0.353

)(119

0.12

2)(0.050

)(0.293

)(629

.094

)Childrenaged

15–19

years

0.06

2�0

:247

1,86

8.02

10.58

9***

�0:341

3,91

7.54

8�0

:099

�0:194

774.44

7

(0.112

)(0.850

)(2,384

.077

)(0.106

)(1.180

)(3,210

.105

)(0.132

)(0.674

)(3,186

.215

)

OLS¼

ordinary

leastsquares.

Notes:Rob

uststandard

errors

arein

parentheses.*,

**,and**

*representsign

ificanceat

the10

%,5%

,and1%

levels,respectiv

ely.

For

each

column,

the

marginaleffectsof

theim

pairmentd

ummyarepresented.The

equatio

nsarefittedon

lyto

thesubsam

pleof

households

who

have

atleasttwochild

ren,on

ebeing

disabled.Individu

alcharacteristics(i.e.,femaledu

mmy,

age)

areinclud

edas

controls.

Sou

rce:

Autho

rs’calculations

usingdata

from

the20

10BangladeshHou

seho

ldIncomeExp

enditure

Survey.



obtain mostly similar results with previous tables showing individual-level analysis,

confirming that our findings are robust.

V. Conclusion

Utilizing a nationally representative dataset from Bangladesh, we examine

disability bias in household investment decisions regarding education. We apply the

hurdle model, which enables us to consider investment decisions more systematically,

to household analysis based on the Engel curve. Consequently, we find the existence

of disability bias on the part of parents, especially with regard to the enrollment

decision for their children. Results from the direct method using the individual child

dataset suggest that there is also a possibility of disability bias in investment decisions

even for children who are already enrolled.

Additionally, individual-level analysis provides ample evidence of disability

bias. Variables on the bargaining power of PWDs suggest that they have low

bargaining power in terms of educational investment for their children. Similarly,

interaction effects suggest the importance of income stability and mother’s education

as instrumental in improving the education of disabled children.

By investing heavily in non-CWDs, parents attempt to provide them with a

competitive advantage in the acquisition of both resources and mates. However,

wealthier and economically stable parents opt for a more opportunistic strategy of

educational investment that does not discriminate between offspring regardless of

disability status. Therefore, at the household level, strategies aiming to increase the

financial stability of parents who have CWDs are important. It is also equally

necessary to design programs to increase parents’ awareness of the fact that investment

in education for CWDs produces two- or three-times higher wage returns (Lamichhane

and Sawada 2013). Investment discrimination regarding disability has a detrimental

effect on the accumulation of human capital, thereby depriving both individuals and

societies of the benefits of private and social returns. This is, therefore, a sufficient

reason for adopting affirmative action plans and anti-discrimination educational

policies. Besides programs that raise awareness, government support programs such as

conditional cash transfers or other encouraging alternatives may be effective in

reducing parental investment disparity and increasing access to quality education.

The main purpose of this study is to identify whether disability-based bias exists

and to clarify the direct causality between a child’s disability and parents’ investment

decisions. However, our research is a preliminary attempt at examining this issue, and



further study with a more organized dataset is required to examine the causality.

Similarly, the presence of a disabled child may affect the family to a large extent. In

particular, the disability of a child may affect the mother’s fertility behavior; that is,

caring for a disabled child may require more time and financial resources, preventing

the family from having another child. As this important information is not captured by

the 2010 Bangladesh HIES, we could not examine the effect of birth order and whether

there is a difference in the number of children if one of the older children is disabled

compared to one of the younger children. While we cannot reject the possible bias

related to these issues, future research can examine the effect of disability on fertility

decisions, which in turn may affect educational investment decisions.

Additionally, although we were able to examine the disability biases for

intrahousehold investment decisions in education generally, we could not identify

which types of impairment actually drive the disability bias. Depending on the type of

impairment and its severity, a CWD’s needs can differ as can the required extra

educational cost. If individual needs arising from the type and severity of impairment

are not addressed, then children with severe difficulties may face significant barriers in

education. Though we find gaps for enrollment depending on disability type in our

descriptive analysis, we cannot conclude if this is actually the case. Due to the smaller

sample size for each disability type, we were not able to perform statistical analysis.

The availability of a robust dataset with the inclusion of a short set of disability-related

questions recommended by the Washington Group on Disability Statistics (2020)

would help conduct future research on the topic by focusing on which impairments

group are more vulnerable to household investment disparities as they relate to

schooling.

References

Ahmmed, Masud, Umesh Sharma, and Joanne Deppeler. 2012. “Variables Affecting Teachers’Attitudes Towards Inclusive Education in Bangladesh.” Journal of Research in SpecialEducational Needs 12 (3): 132–40.

Aslam, Monazza, and Geeta G. Kingdon. 2008. “Gender and Household EducationExpenditure in Pakistan.” Applied Economics 40 (20): 2573–91.

Azam, Mehtabul, and Geeta G. Kingdon. 2013. “Are Girls the Fairer Sex in India? RevisitingIntra-Household Allocation of Education Expenditure.” World Development 42 (C):143–64.

Behrman, Jere R., and Anil B. Deolalikar. 1995. “Are There Differential Returns to Schoolingby Gender? The Case of Indonesian Labor Markets.” Oxford Bulletin of Economics andStatistics 57 (1): 97–117.



Belotti, Federico, Partha Deb, Willard G. Manning, and Edward C. Norton. 2015. “twopm:Two-Part Models.” The Stata Journal 15 (1): 3–20.

Burgess, Robin, and Juzhong Zhuang. 2000. “Modernisation and Son Preference.” Suntory andToyota International Centres for Economics and Related Disciplines, London School ofEconomics Development Economics Discussion Paper No. 29.

Cameron, Colin A., and Pravin Trivedi. 2005. Microeconometrics: Methods and Applications.New York: Cambridge University Press.

Cameron, Stephen V., and James Heckman. 1998. “Life Cycle Schooling and DynamicSelection Bias: Models and Evidence for Five Cohorts of American Males.” Journal ofPolitical Economy 106 (1998): 262–333.

Chowdhury, Jahan, and Dermot Foley. 2006. “Economics of Disability: An Empirical Study ofDisability and Employment in the Bangladesh District of Chuadanga.” Disability StudiesQuarterly 26 (4).

Cragg, John G. 1971. “Some Statistical Models for Limited Dependent Variables withApplication to the Demand for Durable Goods.” Econometrica 39 (5): 829–44.

Deaton, Angus. 1997. The Analysis of Household Surveys: A Microeconometric Approach toDevelopment Policy. Baltimore: Johns Hopkins University Press for the World Bank.

Duan, Naihua, Willard G. Manning, Carl N. Morris, and Joseph P. Newhouse. 1983. “AComparison of Alternative Models for the Demand for Medical Care.” Journal ofBusiness & Economic Statistics 1 (2): 115–26.

Duflo, Esther, and Christopher Udry. 2004. “Intrahousehold Resource Allocation in Coted’Ivoire: Social Norms, Separate Accounts and Consumption Choices.” National Bureauof Economic Research Working Paper 10498.

Filmer, Deon. 2008. “Disability, Poverty, and Schooling in Developing Countries: Results from14 Household Surveys.” The World Bank Economic Review 22 (1): 141–63.

Glewwe, Paul, and Hanan G. Jacoby. 2003. “Economic Growth and the Demand for Education:Is There a Wealth Effect?” Journal of Development Economics 74 (1): 33–51.

Glewwe, Paul, and Harry A. Patrinos. 1999. “The Role of the Private Sector in Education inVietnam: Evidence from the Vietnam Living Standards Survey.” World Development27 (5): 887–902.

Haveman, Robert, and Barbara Wolfe. 1995. “The Determinants of Children Attainments: AReview of Methods and Findings.” Journal of Economic Literature 33 (4): 1829–78.

Himaz, Rozana. 2010. “Intrahousehold Allocation of Education Expenditure: The Case ofSri Lanka.” Economic Development and Cultural Change 58 (2): 231–58.

Kingdon, Geeta G. 2002. “The Gender Gap in Educational Attainment in India: How MuchCan Be Explained?” The Journal of Development Studies 39 (2): 25–53.

_____. 2005. “Where Has All the Bias Gone? Detecting Gender Bias in the IntrahouseholdAllocation of Educational Expenditure.” Economic Development and Cultural Change53 (2): 409–51.

Lamichhane, Kamal. 2013. “Disability and Barriers to Education: Evidence from Nepal.”Scandinavian Journal of Disability Research 15 (4): 311–24.

_____. 2015. Disability, Education and Employment in Developing Countries: From Charity toInvestment. New Delhi: Cambridge University Press.



Lamichhane, Kamal, and Yoshito Kawakatsu. 2015. “Disability and Determinants ofSchooling: A Case from Bangladesh.” International Journal of Educational Development40 (C): 98–105.

Lamichhane, Kamal, and Yasuyuki Sawada. 2013. “Disability and Returns to Education in aDeveloping Country.” Economics of Education Review 37 (C): 85–94.

Lancaster, Geoffrey, Pushkar Maitra, and Ranjan Ray. 2003. “Endogenous Power, HouseholdExpenditure Patterns, and New Tests of Gender Bias: Evidence from India.” Departmentof Economics, Monash University Technical Report Discussion Paper No. 20/03.

_____. 2008. “Household Expenditure Patterns and Gender Bias: Evidence from SelectedIndian States.” Oxford Development Studies 36 (2): 133–57.

Masterson, Thomas. 2012. “An Empirical Analysis of Gender Bias in Education Spending inParaguay.” World Development 40 (3): 583–93.

Psacharopoulos, George, and Harry A. Patrinos. 2004. “Returns to Investment in Education:A Further Update.” Education Economics 12 (2): 111–34.

Rosales-Rueda, Maria F. 2014. “Family Investment Responses to Childhood HealthConditions: Intrafamily Allocation of Resources.” Journal of Health Economics 37(2014): 41–57.

Sawada, Yasuyuki, and Michael Lokshin. 2009. “Obstacles to School Progression in RuralPakistan: An Analysis of Gender and Sibling Rivalry Using Field Survey Data.” Journalof Development Economics 88 (2): 335–47.

Subramanian, Shankar. 1995. “Gender Discrimination in Intra-Household Allocation in India.”Unpublished.

Subramanian, Shankar, and Angus Deaton. 1991. “Gender Effects in Indian ConsumptionPatterns.” Sarvekhshana 14 (4): 1–12.

Takeda, Takaki, and Kamal Lamichhane. 2018. “Determinants of Schooling and AcademicAchievements: Comparison between Children with and without Disabilities in India.”International Journal of Educational Development 61 (C): 184–95.

Thomas, Duncan. 1994. “Like Father, like Son; Like Mother, like Daughter: Parental Resourcesand Child Height.” Journal of Human Resources 29 (4): 950–88.

Washington Group on Disability Statistics. 2020. The Washington Group Short Set onFunctioning. https://www.washingtongroup-disability.com/fileadmin/uploads/wg/Docu-ments/Questions/Washington_Group_Questionnaire__1_-_WG_Short_Set_on_Function-ing.pdf.

Wooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Section and Panel Data, SecondEdition. Cambridge: MIT Press.

Working, Holbrook. 1943. “Statistical Laws of Family Expenditures.” Journal of the AmericanStatistical Association 38 (221): 43–56.

World Bank. 2021. “Least Developed Countries: UN Classification.” https://data.worldbank.org/region/least-developed-countries:-un-classification (accessed 16 October 2021).



Appendix

Table A1. Impairment-Wise Comparison of Children with and without Disabilities(Aged 5–19 Years)

Type of Impairment

Visual Hearing Physical Cognitive Communication Non-CWDs

Enrollment (among school-aged children)Primary level 0.42 0.57 0.29 0.17 0.12 0.68Secondary level 0.51 0.19 0.14 0.07 0.06 0.41Higher level 0.31 0.20 0.00 0.07 0.00 0.32

Literacy (among school-aged children)Reading 0.71 0.49 0.39 0.22 0.22 0.66Writing 0.98 0.98 0.95 0.95 0.96 0.96

CWDs ¼ children with disabilities.Notes: Enrollment rate is calculated using the proportion of enrolled children within school-aged childrendivided by the population of school-aged children. School age is defined following Kingdon (2005):primary level (aged 5–9 years), secondary level (aged 10–14 years), and higher level (aged 15–19 years).Source: Authors’ calculations using data from the 2010 Bangladesh Household Income ExpenditureSurvey.



The Social Costs of Success: The Impact ofWorldTradeOrganizationRules on InsulinPrices in Bangladesh upon Graduationfrom Least Developed Country Status

MD. DEEN ISLAM, WARREN A. KAPLAN, VERONIKA J. WIRTZ,AND KEVIN P. GALLAGHER

¤

In 2021, the United Nations Committee on Development Policy adopted aresolution that Bangladesh would graduate from least developed country(LDC) status after a period of 5 years. This means that in 2026 Bangladeshwould have to forego its exemption to intellectual property (IP) provisions ofthe World Trade Organization (WTO). Bangladesh has taken advantage of thepolicy space it was granted under the LDC exemption to build a genericmedicines industry that not only serves Bangladesh but also other LDCs. Weexamine how IP provisions in the WTO will impact the price of insulin inBangladesh and the subsequent impacts on welfare and poverty. We find thatLDC graduation will trigger a significant jump in insulin prices that couldcause about a 15% decline in the welfare of households in Bangladesh with oneor more members living with diabetes, increasing the poverty rate of suchhouseholds unless policy adjustments are carried out.

Keywords: affordability, cost of illness, insulin, intellectual property,low-income country

JEL codes: I10, I18, I32, I38

⁄Md. Deen Islam (corresponding author): Department of Economics and Global Policy DevelopmentCenter (GDPC), Boston University, USA. E-mail: [email protected]; Warren A. Kaplan, Department ofGlobal Health and GDPC, Boston University, USA. E-mail: [email protected]; Veronika J. Wirtz, Departmentof Global Health and GDPC, Boston University, USA. E-mail: [email protected]; Kevin P. Gallagher,Frederick S. Pardee School of Global Studies and GDPC, USA. E-mail: [email protected].





https://dx.doi.org/10.1142/S0116110522500093

https://dx.doi.org/10.1142/S0116110522500093

I. Introduction

Least developed countries (LDCs) are exempt from granting pharmaceutical

patents until 1 January 2033 (World Trade Organization [WTO] 2015). In addition,

LDC members of the WTO have the option of not filing patent mailbox applications

and obtaining exclusive marketing rights until January 2033 (WTO 2015). This

implies that LDC members have the freedom to reject a pharmaceutical patent

application if the exemption is active. This temporary exemption is important to ensure

access to essential medicines in LDCs. The temporary exemption may facilitate local

production of generic versions of many essential medicines among those LDC

members who are capable, while allowing others to import generic medicines.

However, once this temporary exemption is over, LDC members must ensure

patent protection and provide exclusive marketing rights for any patented medicines.

This change may greatly restrict access to essential medicines in low-income

countries. We use the case of Bangladesh’s LDC graduation to carry out an ex ante

analysis of the impact of such graduation on access to insulin, a lifesaving medicine

for individuals with diabetes.

As an LDC, Bangladesh does not presently need to comply with global

commitments under the WTO’s Trade Related Intellectual Property Rights (TRIPS)

provisions, commonly referred to as the TRIPS Agreement. Currently, Bangladesh can

produce the generic version of any medicine, and patent protection for

pharmaceuticals is not allowed. In 2021, the United Nations recommended

Bangladesh for graduation from the LDC category in 2026. Consequently, firms

will no longer be able to produce copies of medicines that are on patent in Bangladesh

after the country’s graduation from LDC status. Household out-of-pocket expenditure

as a percentage of total health expenditure in Bangladesh was more than 67% in 2015,

of which more than 75% was on pharmaceuticals (Government of Bangladesh,

Ministry of Health and Family Welfare 2016). This implies that prices of some

medicines may increase significantly after 2026, which will place an even larger

burden of health expenditure on households.

Higher prices can affect access to medicines in several ways. First, higher prices

of medicines may force some households to stop taking medicines or take less than the

recommended dose. Second, households may also reduce other forms of consumption,

such as food or spending on children’s education, to cope with the additional

expenditure on medicines. Thus, higher prices of medicines not only affect their usage

but may also reduce consumption of foods, education, and other essential amenities

that are necessary to lead a healthy life. This paper estimates the impact of these



different types of expenditure substitution. We estimate the changes in household

welfare following the implementation of pharmaceutical patenting and stricter

intellectual property rights (IPRs) that would potentially increase the prices of some

medicines. For this purpose, we choose the market for insulin to estimate these effects.

Insulin is a good tracer medicine to measure the effects of stronger IPR on access

to medicines for several reasons. First, some types of insulin would still be under

patent (in other countries) after Bangladesh’s LDC graduation, which implies that IPR

provisions will be a binding constraint on the insulin market. Second, the burden of

diabetes is increasing in Bangladesh. More than 10% of adults have diabetes (mostly

type 2), and more than 70,000 deaths per year are attributable to diabetes or high blood

glucose (World Health Organization [WHO] 2016). This means that insulin is widely

required to satisfy the health needs of the population. Finally, expenditure on insulin is

mostly out of pocket (WHO 2016). Thus, after Bangladesh’s LDC graduation, the

price of insulin may significantly increase as patented versions are imported.

In this paper, we use 2016 Household Income and Expenditure Survey (HIES)

data (Bangladesh Bureau of Statistics [BBS] 2019) and the quadratic almost ideal

demand system (QUAIDS) to estimate household substitution patterns between food,

medicines, and education for households with potential expenditure on insulin. In

addition, we estimate the loss in household welfare and increase in household poverty

resulting from the higher prices of insulins. Unlike other ex ante studies that

investigate a similar question for different medicines in other LDCs or developing

countries, we use household-level data to estimate elasticities of medicine demand and

perform welfare analysis.

There are several advantages of using household data rather than the market

share data of different brands and generic medicines, or aggregate sales and average

prices data. First, household data allow us to control many characteristics of a

household and individuals living in the household, which are important determinants

of demand for medicines along with the price of medicine. Thus, controlling for those

characteristics will enable us to estimate the demand parameters consistently and

efficiently. Second, household data enable us to estimate the different types of

substitution between medicines and other important expenditure items, such as food

and education. Third, sales data for different brands or generic medicines are often

proprietary, and it can be very hard and expensive to get access to that data. Moreover,

sales data may not be very representative, especially for LDCs. On the other hand,

HIES data are available for most LDCs, which is the best representative sample of the

population. In addition, HIES data are often publicly available. Thus, our paper

provides an effective way to estimate the demand parameters of insulin and perform

THE SOCIAL COSTS OF SUCCESS 241


household welfare analysis with household data for Bangladesh, which could also be

applied for any other medicine and HIES data of any other LDC to carry out a similar

analysis.

The paper finds that household demand for insulin is highly price inelastic, even

more inelastic than household demand for food. The price elasticity of insulin is less

than 1 in absolute value, and the price of insulin could increase more than 11 times its

current price if a stronger IPR regime facilitates an unregulated monopoly for insulin;

this would have a significant welfare effect for households with members who need

insulin. We find that the aggregate annual expenditure of those households goes up by

$336 million, which can be as low as $148 million and as high as $656 million. The

welfare cost of the unregulated monopoly of insulin would vary from $71 million to

$408 million under various estimation methods and measures of welfare. Moreover,

the increase in the price of insulin would have a serious impact on household poverty:

poverty rates for households needing insulin could increase between 3 and 40

percentage points.

The rest of the paper is organized as follows. Section II provides some

background on Bangladesh’s LDC graduation and the current status of IP regulation

and the pharmaceutical industry in Bangladesh. Section III is a discussion of relevant

studies. Section IV details the methodology and estimation techniques with a

description of the data and sources. Section V shows the estimation results along with

the household welfare and poverty analysis. Section VI discusses some policy

implications, the limitations of our analysis, and our conclusions.

II. Background

Bangladesh is in the process of making its transition out from the group of LDCs

(United Nations [UN] 2020). This involves a country meeting a graduation threshold

under at least two of the following three predefined criteria: per capita income, human

assets, and economic vulnerability.1 Decisions on inclusion into, and graduation from,

1Income criterion is based on a 3-year average estimate of gross national income per capita for2011–2013, based on the World Bank Atlas method (under $1,025 for inclusion and above $1,230 forgraduation, as applied in the 2018 triennial review).

The Human Assets Index is based on indicators of (i) nutrition: percentage of populationundernourished; (ii) health: mortality rate for children aged 5 years or under; (iii) education: the grosssecondary school enrolment ratio; and (iv) adult literacy rate.

The Economic Vulnerability Index is based on indicators of (i) population size; (ii) remoteness; (iii)merchandise export concentration; (iv) share of agriculture, forestry, and fisheries; (v) share of populationin low elevated coastal zones; (vi) instability of exports of goods and services; (vii) victims of disastertriggered by natural hazard; and (viii) instability of agricultural production.



the list of LDCs is made by the UN General Assembly based on recommendations

from the Committee for Development Policy (CDP), a subsidiary body of the UN

Economic and Social Council. The CDP is, among other things, mandated to review

the category of LDCs every 3 years and to monitor their progress after graduation

from the category (Bhattacharya 2009). In March 2018, the CDP found that

Bangladesh met the criteria for graduation for the first time by satisfying all three

criteria. Bangladesh met the graduation criteria in the triennial review in 2021, and

therefore the CDP recommended the country for graduation from the LDC category in

2026 (United Nations 2021).

LDC classification accords a country duty-free access to the richer economies of

the world, exemption from IPR enforcement, and other economic benefits (UN 2020).

The loss of LDC privileges for Bangladesh would carry with it a 3-year grace period,

during which time Bangladesh must prepare itself for graduation. The most visible

trade-related implication of LDC graduation is the loss of preferential market access,

such as the loss of concessions granted to LDCs under the global system of trade

preferences among developing countries (UN 2019). Since LDCs are also exempt

from the trade-related aspects of the TRIPS Agreement, graduation from LDC status

may have significant implications for IPR enforcement in Bangladesh, which will

have to be addressed by the pharmaceutical and software industries, among others

(UN 2019).

Bangladesh has a burgeoning manufacturing capability and a relatively

self-sufficient pharmaceutical sector. Companies generally manufacture finished

medicine formulations by assembling known generic and, in some cases, patented

components. Since pharmaceutical patents in Bangladesh were suspended in 2008, this

created opportunities for local generic production of medicines patented outside

Bangladesh, with several generic companies supplying the same medicine. For

example, local firms manufacturing medicines patented abroad include Incepta,

Beximco, Beacon, Renata, Square, and Eskayef. Domestically produced medicines

patented abroad include sofosbuvir, sitagliptin, linagliptin, vildagliptin, rivaroxaban,

and empagliflozin (Islam et al. 2017). Some firms have been engaged in producing

active pharmaceuticals ingredients, excipients, and solvents that are used as raw

material in producing the final medicine formulations. Innovative R&D activity is,

however, virtually nonexistent in the Bangladesh pharmaceutical industry as it is a

generics market and generic formulations represent the main business of the

Bangladesh pharmaceutical industry. Presently, the market consists of approximately

8,000 generic products and 258 firms with manufacturing capability, in addition to

imported already-patented products (Islam, Rahman, and Al-Mahmood 2018). This



local production supplies over 95% of Bangladesh’s pharmaceutical needs, and about

80% of these medicines are generics. The top 30–40 companies by value dominate

almost the entire market in which the top 10 hold a 70% domestic market share, and

the top two—Beximco and Square Pharma—capture over 25% of the market (Islam,

Rahman, and Al-Mahmood 2018). In brief, the Bangladesh pharmaceutical market can

be divided as follows:

1. High-end products (anti-cancer, insulin, and vaccines) produced by

multinationals—if on patent, they are not patented yet in Bangladesh;

2. Branded generics (antibiotics, GI medicines);

3. Low-end generics; and

4. Contract manufacturing (domestic and export).

The dynamic nature of the Bangladesh pharmaceutical industry contrasts with its

long-standing IP system. Patent rules and procedures are governed by the original

Patents and Designs Acts of 1911. Bangladesh has not replaced or amended the 1911

Act. It only issued a Notification in 2008 that applications for pharmaceutical and

agrochemical product patents were to be suspended since LDC members of the WTO

could exempt pharmaceutical products from patent protection. This waiver has been

extended until 2033 by the TRIPS Council. Bangladesh can benefit from these

transition periods but only if it retains LDC status (Chowdhury 2018). Some

companies in Bangladesh can make high-end products like insulin to compete with

multinationals (Mohiuddin 2018). This is important as Bangladesh ranks as one of the

10 countries with the highest number of people with diabetes globally (IHME 2019).

A recent scoping review for Bangladesh (Biswas et al. 2016) found that a final

estimate of diabetes prevalence, obtained after pooling data from individual studies

among 51,252 participants, was 7.4%, somewhat less than the estimated overall global

prevalence of 9.3% (Saeedi et al. 2019). For Bangladesh, with 165 million inhabitants

in 2020 (World Bank 2020), this means there are 11.6 million people with diabetes,

about half of them undiagnosed. Undiagnosed diabetes is more likely among people of

lower socioeconomic status (Hasan et al. 2019). The prevalence of diabetes is higher

in males compared to females in urban areas and vice versa in rural areas. Analyses

revealed an increasing trend of diabetes prevalence among both the urban and rural

populations.

Type 2 diabetes is the most common form of diabetes worldwide, comprising

over 90% of all cases (WHO 2019). Management of type 2 diabetes includes diet,

physical exercise, and weight management (National Institute of Diabetes and

Digestive and Kidney Disease [NIDDK] 2020). Some patients with type 2 diabetes



require medication such as oral anti-diabetes medicines and, in some cases, insulin

(NIDDK 2020). Patients with type 1 diabetes require insulin. Since patients with

diabetes have a higher risk of developing cardiovascular diseases, they may also

require additional medicines (NIDDK 2020). Generally, insulin is more expensive than

several commonly used oral anti-diabetes medicines that have been marketed for many

decades and are available at a low price; these generics are recommended as a first-line

pharmacological treatment for diabetes (WHO 2015).

Diabetes has emerged as a major public health problem worldwide, especially in

low- and middle-income countries, where more than 80% of all people with diabetes

are living. The International Diabetes Federation estimated that the global prevalence

of diabetes among adults in 2013 was 8.3%, or roughly 382 million people, and this

was projected to increase more than 592 million in less than 25 years, which might be

a conservative estimate. Southeast Asia accounts for close to one-fifth of all diabetes

cases worldwide and the prevalence of diabetes is projected to increase by 71% in this

region by 2035. The International Diabetes Federation Diabetes Atlas: Fourth Edition

projected in 2009 that diabetes prevalence in Bangladesh would increase more than

50% by 2017, ranking Bangladesh 8th in the number of people with diabetes globally.

The economic and human costs provoked by diabetes in a large population such as

in Bangladesh will continue to be substantial. This study estimates the effect of

graduation out of LDC status and the attendant changes in IP protection for

pharmaceuticals on the price of insulin and the subsequent impacts on welfare and

poverty in Bangladesh.

III. Literature Review

This paper builds on an emerging body of literature on the impacts of trade and

investment treaties on access to medicines. A full assessment of this literature can

be found in Islam et al. (2019). This literature is commonly grouped into two

categories—ex ante analyses that examine the extent to which proposed policies might

impact access to medicines, and ex post analyses that examine the impact of trade and

investment treaties that have already occurred. This paper falls in the ex ante category,

attempting to estimate the extent to which access to insulin will be jeopardized in

Bangladesh under a scenario where it loses its exemption from the TRIPs Agreement

under the WTO if it graduates from LDC status in the coming years.

Most ex post studies find that trade and investment treaties adversely impact

access to medicines in developing countries but to a lesser degree than do ex ante



studies. With respect to ex post studies, some analyses look at the impacts of

WTO-related provisions and others look at free trade agreements (FTAs). Of the WTO

studies, Kyle and Qian (2014) examined the impact of IPR in the TRIPS Agreement

on the launch of new medicines, prices, and sales using data from 59 countries at

varying levels of development. They found that patented medicines have higher prices

and quantities sold, and that new medicine launches were unlikely without patent

protection. Other studies examine impacts from FTAs that have more stringent

provisions than the TRIPS Agreement, particularly those of the United States (US).

Examples of this literature are studies that examine the US–Jordan FTA and find that

the FTA increased prices of essential medicines and delayed market entry of generics

(Abbott et al. 2012). Shaffer and Brenner (2009) examined the Central American Free

Trade Agreement and found that it reduced access to generics already on the market

and delayed entry of other generics. Most recently, Trachtenberg et al. (2020) found

that the US–Chile trade agreement increased both the price and sales volume of

biologics.

This study builds on a set of ex ante studies that predictably estimate adverse

impacts given the underlying assumptions they deploy from economic theory. The

outcomes that ex ante studies predict reflect the models’ underlying assumptions,

which are rooted in economic theory. When a firm is granted a patent, economic theory

predicts the firm will supply a restricted quantity at a higher price because the patent

grants the producing firm a temporary monopoly over the product (Baker 2016).

Akaleephan et al. (2009) used a trade liberalization framework and attempted to

find effects on prices and quantities following a reduction in tariffs or other trade

barriers to estimate the potential cost savings in Thailand resulting from an absence of

TRIPS-plus provisions and increased price competition between innovative and

generic producers of 74 international nonproprietary-name imported medicines. These

authors found that a proposed US–Thailand treaty would increase medical expenses

and reduce the entry of generic medicines.

Chaves et al. (2017) used the IPR impact aggregate model to project the impact

of TRIPS-plus provisions of the Mercosur–European Union FTA on the public

expenditures and domestic sales of antiretroviral medicines and hepatitis C medicines

in Brazil. They reckoned that the treaty would increase medicine expenditures and

decrease sales by domestic producers.

This paper is like the work of Chaudhuri, Goldberg, and Jia (2006) and Dutta

(2011) in terms of the nature of the research question being investigated. Chaudhuri,

Goldberg, and Jia (2006) used a two-stage budgeting framework (using data from

1999 to 2000) to investigate the effects on prices and welfare when one or more



domestic generics are withdrawn from the quinolone market in India due to the TRIPS

Agreement of the WTO.2 That study found considerable consumer welfare losses from

a reduction in the variety of products available on the market after TRIPS. We used

household survey data to estimate the effects of stronger IP laws in the market for

insulin in Bangladesh and obtained similar results of welfare loss as in Chaudhuri,

Goldberg, and Jia (2006) and Dutta (2011).

IV. Methodology, Estimation Framework, and Data

To estimate the effect of graduating from LDC status on the prices of essential

medicines such as insulin, we analyze the effect of introducing patent protection for

pharmaceuticals in Bangladesh. This introduction will potentially reduce competition

in the pharmaceutical market, and even the market of innovative medicines might be

monopolized by the patent holder if there is no further regulation of medicine prices.

Hence, analyzing the effects of Bangladesh’s LDC graduation on medicine prices is

akin to estimating the price effect due to the pharmaceutical market becoming more

monopolized through new patent protection and the withdrawal of generic versions of

innovative medicines from the local market.

In this paper, we estimate the demand for insulin in Bangladesh as the burden of

Type 2 diabetes is increasing in Bangladesh and the price of insulin affects many

persons with Type 2 diabetes. We combine a variety of data sources for this purpose.

To estimate the demand elasticities for pharmaceutical products and/or medicines,

previous studies used market share data. For example, Chaudhuri, Goldberg, and Jia

(2006) and Dutta (2011) used IQVIA market share data of different brands or generics

of quinolones in the Indian market to examine the impact of the WTO agreement.

While IQVIA market sales data of quinolones are representative of the Indian market,

IQVIA market share data only cover 2% of total sales of medicines in Bangladesh,

which is not representative enough to carry out a rigorous demand parameter

estimation. Hence, we use the household-level expenditure data on medicine and other

items instead of market share data. The household-level data have the advantage of

reporting the cost of medicines faced by households rather than the price reported by

manufacturers, but the drawback of using household-level data is that it does not

provide the quantity or price of medicines but rather the total cost of medicines per

person monthly or annually.

2Quinolones are a subsegment of systematic antibacterials.



Accordingly, for our estimation purpose we use Bangladesh’s 2016 HIES data

for information on different categories of expenditures (e.g., food, medicines, and

education); household characteristics (e.g., income, number of members, and

geography of residence); and household head’s characteristics (e.g., age, gender,

religion, employment status, and employment sector). The summary statistics of these

variables are provided in Tables A1 and A2 in Appendix 1. From the HIES data, we

select the households with at least one member with diabetes. The 2016 HIES was

conducted by BBS from April 2016 to March 2017 (BBS 2019). This most recent

HIES is the most extensive household survey in Bangladesh.

The HIES data provide the most granular information on a wide range of

individual and household characteristics. The survey was conducted at three levels

(urban and rural breakdown, district level, and household level) and was designed to

represent different socioeconomic groups in every part of the country. A sample design

was adopted for the 2016 HIES with 2,304 primary sampling units in eight

administrative and geographical divisions (Barisal, Chittagong, Dhaka, Khulna,

Mymensingh, Rajshahi, Rangpur, and Sylhet) and 64 districts selected from the last

Housing and Population Census in 2011. Within each primary sampling unit, 20

households were selected for interviews. The final sample size was 46,080 households.

The sample was stratified at the district level and included a total of 132 substrata: 64

urban, 64 rural, and 4 main city corporations (BBS 2019). Details of the survey design

of the 2016 HIES can be found in International Household Survey Network (2020).

From the 2016 HIES, we construct our sample consisting of all households with

at least one member suffering from diabetes. We excluded individuals who are

suffering from multiple chronic diseases as there is no breakdown of medicine

expenditure in the HIES. Finally, we have a sample of 1,125 households with at least

one member suffering from only one chronic disease (diabetes). We complement the

HIES data with insulin prices from the Directorate General of Drug Administration

(DGDA) of Bangladesh, where prices of all approved insulins and their respective

strengths are reported.

To measure the effects of stronger IPR on the use of insulin and consumption of

other essential items, we model a household’s decision problem of allocating income

in broad expenditure categories such as food, medicine, and education. We estimate

the parameters at this stage using a version of QUAIDS.

Traditionally, elasticities of demand are estimated using a nested logit model of

demand or a full random coefficient logit model of demand (Dutta 2011; Chatterjee,

Kubo, and Pingali 2015). One potential issue of these demand models is that the

demand for any medicine such as insulin generally depends on the physicians’



prescription, especially if patients are not very well informed. So, taste for a particular

brand of insulin is unlikely to be independent across consumers, which violates the

key assumption in those demand modeling strategies. Moreover, to estimate a nested

logit model of demand or a full random coefficient logit model of demand, we need

data on sales of different brands and generic types of insulin, which are not available in

the case of Bangladesh. IQVIA does have some sales data for Bangladesh, but the

coverage is very limited and not representative. Hence, we choose the QUAIDS

framework, which allows us to estimate the price elasticity of insulin using the

household’s expenditure on insulin. One advantage of using the household data to

estimate the elasticities is that we can control many household characteristics, which is

important in estimating the elasticities more consistently. The QUAIDS framework

requires expenditure shares on these expenditure categories, price or price index, total

household income, and other household-level controls, all of which are available in the

2016 HIES. Here, we use the Poi (2012) specification of QUAIDS, which incorporates

the demographic variables.

A. Demand

The QUAIDS model in our estimation framework is based on the following

indirect utility function used in Banks, Blundell, and Lewbel (1997):

lnV(p,m) ¼ lnm� ln a(p)b(p)

� ��1

þ �(p)

� �, ð1Þ

where ln a(p) is the transcendental logarithm function of prices or costs of individual

expenditure items, pi:

ln a(p) ¼ α0 þX3i¼1

αi ln pi þ12

X3i¼1

X3j¼1

γij ln pi ln pj ð2Þ

and b(p) is the Cobb–Douglas price aggregator, defined as follows:

b(p) ¼Y3i¼1

pβii

and �(p) is defined as follows:

�(p) ¼X3i¼1

�i ln pi:



Here, we need to estimate parameters fαi, βi, γi,�ig except α0, which is generally set tosome value lower than the lowest value of lnm (income) in the sample (Deaton and

Muellbauer 1980; Banks, Blundell, and Lewbel 1997). The set of parameters satisfy

some conditions:

adding up:P3

i¼1 αi ¼ 1, homogeneity:P3

i¼1 βi ¼ 0, Slutsky symmetry:P3j¼1 γij ¼ 0,

P3i¼1 �i ¼ 0, and γij ¼ γji.

Now, we specify the expenditure share equation of expenditure item i by

applying the Roy’s identity to equation (1)

!i ¼ αi þX3j¼1

γij ln pj þ βi lnm

a(p)

� �þ �i

b(p)ln

m

a(p)

� �� 2, i�f1, 2, 3g, ð3Þ

where !i is the household’s budget share for expenditure category i; and here we only

consider expenditure on three items: food (1), medicine (2), and education (3), !i is

defined as follows:

!i �piqiPj pjqj

¼ piqim

, j�f1, 2, 3g,

where qi is the quantity of item i and pi is the price or cost of expenditure category j, m

is the household income spent on food, medicine, and education.

B. Demographics

Household and household head characteristics can be incorporated into

the QUAIDS framework using the scaling techniques first used by Ray (1983).

Poi (2002), using this scaling technique, introduces the demographic variables into the

QUAIDS model. Suppose Z is the vector of demographic variables and e(p, u) is

the expenditure function. Ray’s scaling method decomposes the expenditure function

into a scaling function, which depends on prices, level of utility, and demographics,

and an expenditure function, which depends on prices and level of utility only.

Specifically,

e( p, u,Z) ¼ m0(p, u,Z)� e(p, u):

Here, the scaling function m0(p, u,Z) takes the following form:

m0(p, u,Z) ¼ �m0(Z)� �(p, u,Z),

where �m0(Z) is the part of the scaling function that depends on demographics only;

that is, a larger family will have a larger expenditure on food compared to a smaller



family, and a family with more school-aged children is likely to have higher

educational expenditure than a family with no school-aged children. The second part

�(p, u,Z) accounts for the interaction between the consumption pattern and

demographics; that is, a family with a member with diabetes may consume a different

type of food compared to a family with no such member. Ray (1983) parametrizes

�m0(Z) and �(p, u,Z) as follows:

�m0(Z) ¼ 1þ � 0Z

�(p, u,Z) ¼uQ3

j¼1 pβjj

Q3j¼1 p

η0jZj � 1

� 1u �

P3j¼1 �j ln pj

,

where � and η are vectors of parameters to be estimated. The expenditure share

equations specified in (3) become

!i ¼ αi þX3j¼1

γij ln pj þ (βi þ η0jZ) lnm

�m0(Z)a(p)

� �

þ �i

b(p)c(p,Z)ln

m

�m0(Z)a(p)

� �� 2, ð4Þ

where c(p,Z) ¼Q3i¼1 p

η0iZi and the additional adding-up condition:

P3i¼1 ηi ¼ 0.

C. Elasticities

The uncompensated price elasticity of demand for good i with respect to the price

of good j (�ij) is derived in Poi (2012) and given as follows:

�hij ¼d ln qid ln pj

¼ �δij þ1!i

γij � βi þ η0iZþ 2�i

b(p)c(p,Z)ln

m

�m0(Z)a(p)

� ��

� αj þXk

γik ln pk

!� (βj þ η0jZ)�i

b(p)c(p,Z)ln

m

�m0(Z)a(p)

� �� 2!,

where δij ¼ 1 if i ¼ j and 0 otherwise, and h is the index for households. The

expenditure or income elasticity for good i (�i) is derived as follows:

�hi ¼

d ln qid lnm

¼ 1þ 1!i

βi þ η0iZþ 2�i

b(p)c(p,Z)ln

m

�m0(Z)a(p)

� �� :

The formula for price elasticities here is at the household level. The price elasticities at

the market level are then the average of the household-level price elasticities.



D. Econometric Issues

The 2016 HIES does not provide any information on whether a household with a

person who is living with diabetes needs to purchase insulin for that member, so to

estimate the demand parameters and elasticities for insulin demand, we construct a

sample that has the highest probability of including the households that purchase

insulin. For this purpose, we use the maximum retail price of each registered insulin to

estimate the cost per daily dose as defined by WHO, and then calculate the monthly

cost of insulin for an individual. First, we estimate the bounds on insulin cost per

month for an individual, and our calculation shows that the monthly cost of using only

insulin ranges from 436 taka (Tk) to Tk1,925. Second, for the purpose of this study we

assume that the individuals who use only noninsulin diabetes medicines are in the

lower bound of the abovementioned price range of insulins. That is, individuals whose

monthly cost of diabetes medicines is below Tk436 are assumed to use only

noninsulin diabetes medicines. The distribution of the costs of diabetes medicines is

shown in Figure 1. From the distribution of costs of diabetes medicines, we obtain that

around 47% of observations (534 out of 1,125) are below the lower bound of Tk436.

Thus, the proportion of households using only insulin, insulin plus noninsulin, or

expensive noninsulin medicines is about 53%. Hence, our sample for the analysis is

Figure 1. Distribution of Monthly per Person Costs of Diabetic Medicines

Source: Authors’ calculations based on Bangladesh Bureau of Statistics. 2019. Report on the Household Incomeand Expenditure Survey 2016. Dhaka: Statistics and Information Division, Ministry of Planning, Government ofBangladesh.



the 38% of households with at least one member with diabetes in which per person

costs of medicines range from Tk436 to Tk1,925 (424 out of 1,125).

Here, we do not include households with members suffering from diabetes along

with other chronic illnesses. Our assumptions seem plausible given that Mohiuddin

(2019) found that in Bangladesh about 15% of patients with diabetes use only insulin,

whereas Islam et al. (2017) found that about 41% of patients with diabetes use insulin

in Bangladesh.

Since our sample includes only those households that have at least one member

with diabetes and the per person costs of medicines range from Tk436 to Tk1,925, the

bounds on the cost of medicines ensure that our sample includes almost all households

spending on insulin; however, this does not ensure the exclusion of households whose

expenditures on medicine fall within the bounds, but these expenditures are not on

insulin. This may introduce a sample selection bias into our estimation. To minimize

this bias, we perform a Heckman type correction for selection bias. This correction is

performed in two stages. In the first stage, we estimate the following Probit model:

Prob(D ¼ 1jX) ¼ �(Xθ), ð5Þwhere X is the vector of explanatory variables that includes different individual

characteristics such as age, gender, education, and ethnicity, as well as individual

household characteristics such as household income, location, religion, household

head’s education, age, and gender. θ is a vector of unknown parameters, and � is the

cumulative distribution function of the standard normal distribution. Here, vector X

could be the same as Z or different than Z; that is, we can use the variable vector Z in

place X, or we could use a subset of Z with some other control variables to construct X.

The indicator variable D is defined as follows:

D ¼ 1, if monthly cost of diabetic medicine is less than BDT436

0, otherwise

�

Estimation of this Probit model yields results that can be used to predict the probability

for everyone with diabetes that uses only noninsulin diabetes medicines given the

various individual and household characteristics. We use this estimated Probit model

to predict the probability that an individual uses only noninsulin diabetes medicines

for our sample, which comprises individuals with diabetes with monthly costs of more

than Tk436. These predictions will be unbiased and consistent if the error terms in the

Probit model are uncorrelated with control variables and are normally distributed.

After estimating this Probit model, we obtain the correlation between the predicted

values and the residuals of the model and this correlation is almost 0 (�0.0004).

So, we can maintain the assumption that the error terms of the Probit model and the set



of control variables are uncorrelated. Assuming that these assumptions are satisfied,

we estimate the probabilities that the individuals using only insulin or insulin with

other noninsulin medicines in our sample using the estimated Probit model:

Prob(D ¼ 0jX) ¼ 1� �(Xθ̂): ð6ÞUsing these estimated probabilities, we estimate the inverse Mills ratio as follows:

�(Xθ̂) ¼ �(Xθ̂)

1� �(Xθ̂), ð7Þ

where � is the probability density function. After estimating the inverse Mills ratio, we

estimate the QUAIDS model, where now Z includes �(Xθ̂) as an additional control

along with the other control variables described above. Assuming that the error terms

are jointly normal, we estimate the QUAIDS model including the Mills ratio as an

additional demographic variable.

A second issue in the estimation of the QUAIDS model is that the costs of

diabetes medicines might be correlated with other unobserved individual or household

characteristics (Islam et al. 2019). To overcome this problem, we construct an

instrumental variable (IV) for the cost of diabetes medicines. To construct this IV, we

argue that the cost of diabetes medicines of an individual might be correlated with

unobserved individual and household characteristics, but these unobserved

characteristics are orthogonal to the cost of medicines of individuals residing in the

same geographic area. Thus, we use the average cost of medicines in the smallest

geographic unit of the HIES as the IV for cost of diabetes medicines, as the price or

cost of diabetes medicines is correlated within the same geographic region, but

orthogonal to a specific individual’s or household’s characteristics, where the average

is calculated by a leave-one-out method. That is, the IV for the cost of medicines for

individuals in household h residing in region r is the average cost of medicines for all

individuals residing in the same region r except members of household h. Let us refer

to this IV as IV1, so IV1 is given as follows:

IV1hr ¼P

h�r n h pdhrNr

, ð8Þ

where IV1hr is the IV for the cost of medicines of individuals in household h in region

r, pdhr is the cost or price of diabetes medicines of an individual in household h in

region r,P

h�r n hpdhr is the sum of the cost of diabetes medicines of all individuals in

region r except for individuals living in household h, and Nr is the total number of

individuals with diabetes living in region r and incurring a cost of medicines ranging



from Tk436 to Tk1,930. Similarly, we construct an IV for the prices of food and

education.

Another issue in estimating the demand parameters is that error terms u may be

spatially correlated as costs of diabetes medicines are generally correlated with the

types of health care provider such as public hospitals, private hospitals, and

pharmacies, and we have certain types of health care providers in each region (Islam

et al. 2019). This may introduce heteroscedasticity in the QUAIDS model and hence

reduce the efficiency of the estimators. To eliminate the heteroscedasticity due to

spatial correlation in error terms u, we cluster the standard errors at the union or ward

level, which is the lowest administrative unit in Bangladesh.

E. Computing Counterfactual Price Changes

To determine the range of potential increases in the prices of insulin following

Bangladesh’s graduation from LDC, we use estimated demand elasticities to compute

the ranges of markups and marginal costs based on the current prices of insulin and

insulin market structure. Since the expenditure items in our QUAIDS model are

defined broadly (i.e., food, medicine, and education), it is expected that the price

elasticities of demand would be very low. Hence, it would be impossible to determine

the insulin prices under the monopoly market structure ensured by stronger IP laws as

a monopoly’s equilibrium output is always at the elastic part of the market demand

curve. To compute the counterfactual prices of insulin under monopoly market structure,

we need to estimate the slope of the demand function of insulin so that we can use this

slope to estimate the price elasticities of demand at different points on the demand curve.

This estimated elasticity is then used to derive the optimal monopoly markup. Here, we

assume that the market demand for insulin is linear in insulin prices and estimate this

linear demand function by estimating the following regression equation:

!2 ¼ ’0 þ ’1p2 þ ’2 �! þ Z 0�þ u, ð9Þ

where !2 is the household expenditure on insulin, p2 is the price of insulin faced by the

household, �! is the minimum level of income necessary to ensure a subsistence level of

food consumption for the household. �! is calculated by multiplying the household size

and the national lower poverty level income as reported in the final report of the 2016

HIES (BBS 2019); Z 0 is the vector of household and household head’s characteristics; uis the error term; and ’0, ’1, ’2, and � are parameters to be estimated. Here, the main

parameter of interest is’1, which then is used to calculate the slope of the insulin demand



curve with respect to insulin price as follows:

d!2

dp2¼ d(p2q2)

dp2¼ q2 þ p2

dq2dp2

¼ ’̂1,

�b ¼ dq2dp2

¼ ’̂1

�p2� �q2

�p2,

ð10Þ

where �b is the slope of the demand curve evaluated at the average price and quantity of

the insulin.We also verify the estimated slope of the inverse demand curve using the own

price elasticity of insulin demand obtained from our QUAIDS model as follows:

0 ¼ p2 þ q2dp2dq2

¼ 1þ 1E22

, ð11Þ

dp2dq2

¼ 1þ 1E22

� �p2�q2

¼ 1�b, ð12Þ

where we use the fact that at the midpoint of the demand curve, marginal revenue is 0.

Once we have the estimated slope of the insulin demand curve, we can estimate the price

elasticities of the insulin demand curve

E22 ¼ �bp2q2

: ð13Þ

Now, we can find the elasticities at different points of the demand curve. With these

estimated elasticities, we can find the optimal markup for the monopoly. In addition to

simulating the counterfactual markup and price under monopoly market structure, we

also use the average insulin price in Pakistan, where the pharmaceutical market is less

regulated and strong IP laws govern the market (Basant 2007). Nevertheless, most types

of insulin are very affordable in Pakistan compared to other South Asian countries. The

main reason that a stronger IPR regime did not lead to exorbitant price increases for

insulin in Pakistan is the provision of the insulin supply by the public sector (Ewen et al.

2019). The reasons that we choose current insulin prices in Pakistan as another

counterfactual price are as follows: (i) this provides an interesting scenario where strong

IP laws coexist with public sector participation, which enables greater access to insulin;

and (ii) the size and characteristics of the economy of Pakistan are comparable to those of

Bangladesh.

F. Welfare Analysis

To have insights into the welfare effects of a stronger IPR regime in post-LDC

Bangladesh under two counterfactual prices—simulated prices under monopoly



market structure and prices in a less regulated neighboring country (Pakistan)—we use

several measures of welfare as elaborated by Araar and Verme (2016). Our first

measure is the consumer surplus (CS), defined as the difference between willingness to

pay and the market price of insulin. The measure of CS is given as follows:

CS ¼Z p2 0

p2D(p2)dp2, ð14Þ

where p2 and p2 0 are the current and counterfactual prices of insulin, D(p2) is the

demand function of insulin. Here, to estimate the CS we need to know the Marshallian

demand function D(p2). For a linear demand system and moderate change in prices,

CS can be estimated using the following equation:

CS ¼ �x2 Δp2(1þ 0:5E22Δp2): ð15ÞFor the problem concerned in this paper, the price changes could be significantly

higher and so the above formula will provide a highly overstated estimate for CS. Araar

and Verme (2016) derived an approximation CS formula for a large price change:

CS ¼ �x2 Δp2 1� 0:5Δp221þ Δp2

� �: ð16Þ

CS as a measure of welfare is somewhat restrictive as it assumes that the marginal utility

of real income is constant and there is no distributional effect of price changes. It also

captures only the partial equilibrium effect and does not perfectly measure the change in

welfare if the changes in prices are large. However, CS is a straightforward and easy-to-

estimate welfare measure, which would be a good standard to compare with other

measures of welfare. The next two welfare measures that we estimate are compensating

variation (CV) and equivalent variation (EV). These measures are defined as follows:

CV ¼ e(p2, v0)� e(p2 0 , v0) ¼

Z p2 0

p2h(p2, v

0)dp2, ð17Þ

EV ¼ e(p2, v1)� e(p2 0 , v1) ¼

Z p2 0

p2h(p2, v

1)dp2, ð18Þ

where v0 and v1 are levels of generic indirect utility before and after the implementation

of a stronger IPR regime, respectively, e(.) is the generic expenditure function, and h(.)

is the Hicksian demand function. Here, CV is the monetary compensation required to

bring the consumer back to the original utility level after the price change, and EV is the

monetary change required to obtain the same level of utility after the price change (Araar

and Verme 2016). One computational problem in calculating CVand EV is that we need

to know the indirect utility level before or after the changes in prices. One solution to



this computational problem is to derive CV and EV from CS as given in Chipman and

Moore (1980)

CV ¼ (1� e�CS=m)m, ð19ÞEV ¼ (eCS=m � 1)m, ð20Þ

where m is the income level. In addition to these measures of welfare, there are two

simple straightforward measures of welfare: Laspeyres Variation (LV), which is defined

as the exact change in income necessary to purchase the initial bundle of goods at prices

after and before the change in the IPR regime. LV is defined as follows:

LV ¼ e(p 02,X

0)� e(p2,X0), ð21Þ

where X 0 is the initial bundle of goods purchased before the change in prices. The

second measure is the Paasche Variation (PV), which is defined as the exact change in

income required to purchase the final bundle of goods at prices after and before the

change in the IPR regime. PV is given as follows:

PV ¼ e(p 02,X

1)� e(p2,X1), ð22Þ

where X 1 is the final bundle of goods purchased after the change in prices due to a

change in the IPR regime. To estimate LVor PV, we just need the information of quantity

purchased before or after the change in the policy regimes and the associated changes in

prices, whereas to estimate the other measures of welfare requires some knowledge or

assumptions on the demand function or the utility function.

V. Results

A. Price and Expenditure Elasticities

Table B1 in Appendix 2 reports the parameter estimates of our QUAIDS model.

The estimated uncompensated price elasticities and expenditure elasticities are

reported in Table 1. Here, all elasticities are the average elasticities across all

households in the sample. The price elasticities are denoted as Eij, where subscript i

denotes the expenditure on item i, and j denotes the price of item j. The estimate of

price elasticity of food, E11, is consistently estimated across different models; E11

ranges from 93.7% to 99.0% under different specifications. The price elasticities of

insulin have expected negative signs only under IV specification, and these vary from



92.7% to 94.3%, whereas the price elasticities of education vary from 14.3% to 25.5%

under various specifications but are not statistically significant.

The cross-price elasticities show interesting demand patterns as well. The cross-

price elasticities between food and insulin (E12) or education (E13) are always negative

under all specifications and statistically different from zero. This indicates that

expenditure on food falls in response to an increase in the price of insulin or education.

However, the cross-price elasticities between insulin and food (E21) or education (E23)

are positive under IV specifications, which indicates that an increase in the price of

food or education may not lead to a decrease in demand for insulin.

B. Marginal Costs and Markups

Currently, the market for insulin in Bangladesh is oligopolistic. To find the

markups in this market, we assume that the marginal cost (MC) of producing insulin is

constant and the same for all producers. If there are n firms in the market with the same

Table 1. Uncompensated Price and Expenditure Elasticitiesof Major Expenditure Items in Bangladesh

Not Corrected Corrected

OLS IV OLS IV

Price elasticitiesE11 �0.988*** �0.945*** �0.990*** �0.937***E12 �0.103*** �0.004*** �0.106*** �0.004***E13 �0.071*** �0.054*** �0.072*** �0.060***E21 �0.043*** 0.621*** �0.042*** 0.563***E22 0.377*** �0.927*** 0.413*** �0.943***E23 0.120*** 0.107*** 0.125*** 0.090***E31 0.062*** �2.011*** 0.082*** �2.124***E32 �0.010 0.013 �0.010 0.003E33 �0.180 �0.255 �0.180 �0.143

Expenditure elasticitiesE1 1.162*** 1.003*** 1.168*** 1.001***E2 �0.454*** 0.203*** �0.495*** 0.289***E3 0.133 2.251*** 0.111 2.258***

IV ¼ instrumental variable, OLS ¼ ordinary least squares.Notes: Subscript 1 refers to food, subscript 2 is insulin, and subscript 3 iseducation. ***p < 0:01, **p < 0:05, *p < 0:1.Source: QUAIDS model estimates based on data from the BangladeshBureau of Statistics. 2019. Report on the Household Income andExpenditure Survey 2016. Dhaka: Statistics and Information Division,Ministry of Planning, Government of Bangladesh.



MC, c, the markup is defined as follows:

P� c

p¼ � 1

n

Q

P

dP

dQ¼ � 1

nED:

The current insulin market in Bangladesh is to some extent competitive. There are

seven domestic producers of insulin supplying 50 different insulin products in

Bangladesh (DGDA 2019). The differences in these products are in terms of dosages

size and the producers. In addition, there are six foreign producers, who have

registered a combined 65 insulin products in Bangladesh (DGDA 2019). The licenses

of products of two foreign producers expired in 2015 and early 2016.3 Hence, there are

now 11 suppliers of insulin in Bangladesh. Thus, the markup is given by the following

formula: 11þ 1

11*jE22 j

� �. The MCs are calculated using the bounds of insulin prices, which

is the amount paid by a household for 1 month of insulin supply. We use the maximum

retail prices reported by DGDA to estimate this monthly expenditure on insulin, which

is found to range from Tk436 to Tk1,925. The estimated markups and bounds of MC

of a 1-month insulin supply are reported in Table 2.

From Table 2, we can see that the current markups range from 1.107 to 1.109,

reflecting the fact that the current market is characterized by some competitive forces

as the markups are around 10% over the MC. The lower and upper bounds of MC

range from Tk393.24 to Tk1,740.73 and from Tk393.97 to Tk1,743.94, respectively.

These bounds reflect end user MCs rather than MCs at the production level.

3Eli Lilly & Company’s license in the US expired in May 2016. Lilly’s license in France expired inMay 2015. https://www.dgda.gov.bd/index.php/2013-03-31-05-16-29/registered-imported-drugs.

Table 2. Implied Markups and Marginal Costs of Insulin in Bangladesh underCurrent Market Structure (Tk)


Lower Bound Upper Bound Lower Bound Upper Bound

Price elasticities �0.93 �0.93 �0.94 �0.94Markups 1.109 1.109 1.107 1.107Marginal costs 393.24 1,740.73 393.97 1,743.94

Tk ¼ Bangladesh taka.Source: Authors’ calculations based on data from the Bangladesh Bureau of Statistics. 2019.Report on the Household Income and Expenditure Survey 2016. Dhaka: Statistics andInformation Division, Ministry of Planning, Government of Bangladesh; and DirectorateGeneral of Drug Administration. “Registered Products.” Government of Bangladesh.http://www.dgda.gov.bd/index.php/manufacturers/allopathic (accessed 14 October 2019).



Thus, these are the MCs of all value added of insulin production: from production to

final purchase by households.

C. Demand Function of Insulin and Counterfactual Prices

We estimate the insulin demand function as specified in regression equation (9)

using the IV for insulin prices. The result of the regression equation (9) is reported in

Table B3 in Appendix 2. The estimates of the coefficient of insulin price (p2IV) are

negative under the estimation strategies of not correcting and correcting for sample

selection bias. We use the estimate of the coefficient of p2IV in a regression corrected

for sample selection bias, and this estimate is ’̂1 ¼ �0:11. The estimated slope

coefficient, �b ¼ �0:00137, is given as follows:

�b ¼ dq2dp2

¼ ’̂1

�p2� �q2

�p2¼ ’̂1

�p2� �p2�q2

�p 22

¼ �0:11884:16

� 973:33884:162

¼ �0:00137,

where Tk973.33 is the average monthly household expenditure on insulin and

Tk884.16 is the average of monthly price of insulin. Now, the elasticity of insulin

demand at the average price and quantity of insulin is given as follows:

E22 ¼ �b�p

�q¼ �b

�p 2

�p�q¼ �0:00137� 884:162

973:33¼ �1:10:

Using this elasticity of insulin demand measured approximately at the midpoint of the

insulin demand curve, we can find the maximum markups:

1

1þ 1jE22j

!¼ jE22j

1þ jE22j¼ 11:01:

This markup shows that under an unregulated monopoly, the insulin price could be

more than 11 times higher than current insulin prices, where the current markup of

insulin in Bangladesh is about 1.1. Using the estimated markup under an unregulated

monopoly and the upper and lower bounds of MC as reported in Table 2, we estimate

maximum possible counterfactual prices of insulin, which are reported in Table 3.

These counterfactual prices show the most extreme situations of an increase in insulin

prices in Bangladesh following its graduation from LDC status and the enforcement

of strong IP laws. Thus, these provide some bounds on the prices of insulin in a

worst-case scenario.

For the Pakistan price counterfactual, we use the average insulin prices reported

in Ewen et al. (2019), where the insulin prices for several low- and middle-income

countries including Pakistan were surveyed in 2016. Since our sample is from the



2016 HIES, we use the insulin prices in Pakistan as reported in Ewen et al. (2019).

These prices are shown in Table 4, where we only show the average insulin prices in

the private sector (private pharmacies, hospitals, and clinics), as reported in Ewen et al.

(2019), since the public sector insulin price is very similar to the private sector price

for any specific type of insulin. However, the glargine analogue insulins are only

available in the private sector, particularly at private retail pharmacies.

The average insulin price per 1,000 international unit (IU) in Pakistan in 2016

ranged from about $4.50 to $7.89, except for the glargine analogue. Using the Tk–$

exchange rate in June 2016 from Bangladesh Bank, the central bank of Bangladesh,

Table 3. Counterfactual Markups and Prices of Insulin in Bangladesh under anUnregulated Monopoly (Tk)


Lower Bound Upper Bound Lower Bound Upper Bound

Marginal costs 393.24 1,740.73 393.97 1,743.94Counterfactual markups 11.01 11.01 11.01 11.01Counterfactual prices 4,329.59 19,165.43 4,337.59 19,200.78Change in prices 3,893.59 17,235.43 3,901.59 17,270.78

Tk ¼ Bangladesh taka.Source: Authors’ calculations based on data from the Bangladesh Bureau of Statistics. 2019. Reporton the Household Income and Expenditure Survey 2016. Dhaka: Statistics and Information Division,Ministry of Planning, Government of Bangladesh; and Directorate General of Drug Administration.“Registered Products.” Government of Bangladesh. http://www.dgda.gov.bd/index.php/manufac-turers/allopathic (accessed 14 October 2019).

Table 4. Insulin Prices in Pakistan in 2016 ($ per 1,000 IU)

Cartridge Vial

Private RetailPharmacies

Private Hospitalsand Clinics

Private RetailPharmacies

OriginalBrand Bio-similar



Short-acting human 5.81 4.50 5.81 4.72Intermediate-acting human 5.81 4.6730/70 human 5.15 4.48 5.82 7.89Glargine analogue 28.60 20.65

IU ¼ international unit.Source: Authors’ calculations using data from Table 2 in Ewen, Margaret, Huibert-Jan Joosse, DavidBeran, and Richard Laing. 2019. “Insulin Prices, Availability and Affordability in 13 Low-Income andMiddle-Income Countries.” BMJ Global Health 4 (3): e001410.



these average prices correspond to Tk352.80–Tk618.58 per 1,000 IU, whereas the

average monthly cost of insulin per person in Bangladesh is about Tk884.16.4 Since

1,000 IU of insulin is approximately the monthly supply of insulin for an individual,

the average monthly insulin cost for most types of insulin is significantly higher in

Bangladesh than in Pakistan. However, average prices for long-acting insulins

such as glargine analogues range from $20.65 to $28.60, which corresponds to

Tk1,618.18–Tk2,246.16, higher than the average monthly insulin costs in Bangladesh.

Here, we use the price of the original brand of glargine analogues in Pakistan as the

counterfactual price of insulin in Bangladesh under stricter IP laws. To estimate the

upper bound of price gain and loss in welfare, we take the difference between this

price, Tk2,246.16, and the current monthly average cost of insulin per person in

Bangladesh, Tk8,84.16, which implies a potential 154% increase in the average

monthly cost of insulin in Bangladesh.

D. Welfare Analysis

The welfare estimates are reported in Table 5. The welfare loss estimates in this

table are aggregate national-level estimates. The welfare losses in the “Upper bound”

column correspond to upper bound price changes in columns 2 and 4 of Table 4.

Similarly, the welfare losses in columns 3 and 4 in Table 5 correspond to lower bound

price changes in columns 1 and 3 of Table 4. The welfare estimates in column 5 of

Table 5 are calculated for the counterfactual price increase from the average price of

Tk884.16. The welfare loss estimates in column 6 of Table 5 are calculated by using

the originator’s price of long-acting insulin glargine analogues in Pakistan. All these

estimates of welfare loss show the worst-case scenario, which entails maximum

welfare losses under an unregulated monopoly because of stronger IP laws after

Bangladesh’s graduation from LDC status.

The first row of Table 5 is the measure of aggregate increases in household

expenditures due to an increase in insulin prices following Bangladesh’s graduation

from LDC status. Here LV and PV measures are the same, as we use the same

composition of goods before and after changes in insulin prices. The upper bound of

the aggregate increase in household expenditure under an unregulated monopoly is

about $656 million per year, whereas the lower bound is about $148 million per year.

The aggregate increase in household expenditure would be significantly lower, about

4The Tk–$ exchange rate in June 2016 was 78.4. https://www.bb.org.bd/econdata/exchangerate.php.



Table

5.Annual

Agg

rega

teWelfare

Lossesin

Ban

glad

eshunder

anUnregu

latedMon

opolyan

dPak

istanPrice

Cou

nterfactual

Upper

Bou

nd

Low

erBou

nd

Not

Corrected

Corrected

Not

Corrected

Corrected

Average

Price

Cou

nterfactual

Pak

istanPrice

Cou

nterfactual

LV¼

PV

Tkmillion

51,347

.49

51,450

.48

11,599

.73

11,623

.54

26,367

.08

4,05

7.65

$million

654.94

656.27

147.96

148.26

336.31

51.76

CS

Tkmillion

25,675

.23

25,726

.73

5,80

1.35

5,81

3.26

13,185

.03

2,03

0.31

$million

327.49

328.15

74.00

74.15

168.18

25.90

CV

Tkmillion

31,883

.49

31,962

.69

6,07

1.70

6,08

4.74

14,662

.37

2,06

2.52

$million

406.68

407.69

77.45

77.61

187.02

26.31

EV

Tkmillion

21,385

.61

21,421

.22

5,55

2.47

5,56

3.38

11,961

.42

1,99

9.02

$million

272.78

273.23

70.82

70.96

152.57

25.50

CS¼

consum

ersurplus,

CV¼

compensating

variation,

EV¼

equivalent

variation,

LV¼

Laspeyres

Variatio

n,PV

¼Paasche

Variatio

n,Tk¼

Bangladeshtaka.

Sou

rce:

Autho

rs’calculations

basedon

data

from

theBangladeshBureauof

Statistics.20

19.Reporton

theHou

seho

ldIncomean

dExpenditure

Survey

2016

.Dhaka:StatisticsandInform

ationDivision,

Ministryof

Plann

ing,

Gov

ernm

entof

Bangladesh.



$52 million per year if the insulin prices in Bangladesh stayed at a similar level to

insulin prices in Pakistan.

For an increase in the price of insulin, the relationship among losses in CS, CV,

and EV are as follows: CV > CS > EV. From Table 5, we can see that these

relationships are satisfied. From the figures in third (CV) and fourth (EV) rows in

Table 5, the annual aggregate loss in welfare under an unregulated monopoly will

range from $71 million to $407 million. However, under the Pakistan price

counterfactual, the annual loss in welfare would be around $26 million.

The welfare effect of an increase in insulin prices at the household level is in

Table 6, which reports the increase in household expenditure and the increase in

expenditure as a percentage of household average income per year for three

counterfactual scenarios: largest upper bound estimate (upper bound IV), smallest

lower bound estimate (lower bound OLS), and the Pakistan price counterfactual.

The annual welfare impacts of stronger IP laws could be from $51.8 million

across impacted households (Pakistan price counterfactual) to an upper bound of

$656.3 million under an unregulated monopoly (Table 6). According to a review of the

literature (Biswas et al. 2016), the incidence of people with diabetes in Bangladesh is

estimated to be between 4.5% and 35.0%, with the “pooled preference” being 7.4%.

The average number of people in a household in Bangladesh is 4.06 (BBS 2019). The

cost per impacted household per year would therefore range from $17.6 to $223.1,

which implies a 0.7% to about a 9.1% decline in affected household incomes.

E. Poverty Impact

An increase in the price of insulin because of stricter IP laws would also have a

significant impact on the poverty incidence for households that require access to

lifesaving insulin for the members of those households with diabetes. To show the

effect of a price rise in insulin on household poverty, we estimate the rate of poverty

for the households with members with diabetes, especially with members needing

insulin. Table 7 shows the absolute number of people and households and rates of

poverty under the upper and lower poverty lines at the national level, households with

persons having diabetes, and households with members requiring insulin.

Table 7 shows there are about 39.33 million households in Bangladesh, and out

of them, 12.89% fall below the lower poverty line and 24.28% fall below upper

poverty line.5 The corresponding poverty rates for households with at least one

5Lower and upper poverty line incomes are defined in BBS (2019).



member living with diabetes are 25.36% and 33.54%, respectively, and for households

needing insulin they are 20.99% and 27.44%, respectively. The absolute number of

households that would fall below the lower and upper poverty lines as a result of an

increased price in insulin along with the percentage increase from the initial level are

reported in Table 8.

Table 6. Household-Level Welfare Analysis of Insulin Price Increases in Bangladesh

Incidence of Diabetes (Range of Estimates)

Total 4.5% 7.4% 35%

Population (2016) 161,356,000 7,261,020 11,940,344 56,474,600Households 1,788,429 2,940,971 13,910,000

Increase in Expenditure and Welfare Loss per Household

AggregateWelfare Loss ($) Dollars per Household per Year

Pakistan price counterfactual 51,760,000 28.9 17.6 3.7Lower bound OLS 147,960,000 82.7 50.3 10.6Upper bound IV 656,270,000 337.0 223.1 47.2

Impact per Affected Household(Average annual income per household: $2,447a)

Welfare as % of HouseholdAverage Income

Pakistan price counterfactual 1.18% 0.72% 0.15%Lower bound OLS 3.38% 2.06% 0.43%Upper bound IV 15.00% 9.12% 1.93%

Welfare as % of GDP (2016 Bangladesh GDP: $221 billion)

Pakistan price counterfactual 0.02%Lower bound OLS 0.06%Upper bound IV 0.27%

GDP ¼ gross domestic product, IV ¼ instrumental variable, OLS ¼ ordinary least squares.aAverage annual income per household is calculated by multiplying average monthly family income ofTk15,988 (2016 HIES) by 12 and then converting into the United States dollars by dividing by 78.4(Tk–$ exchange rate in June 2016) with the average household size, i.e., 4.06 members per household(2016 HIES).Note: Ranges of estimates according to Mohiuddin, Abu Kholdun. 2018. “An A–Z of Pharma IndustryReview: Bangladesh Perspectives.” PahrmaTutal 6 (12): 64–78.Source: Authors’ calculations based on data from the Bangladesh Bureau of Statistics. 2019. Report onthe Household Income and Expenditure Survey 2016. Dhaka: Statistics and Information Division,Ministry of Planning, Government of Bangladesh; and World Bank. “World Development Indicators.”https://databank.worldbank.org/source/world-development-indicators (accessed 1 June 2020).



Poverty estimates in Table 8 are reported only for the upper bound and lower

bound of price change under an unregulated monopoly scenario, estimated using IVs

for prices with correction for sample selection bias and the price change under the

Pakistan counterfactual policy regime. The numbers of households that fall below the

lower poverty line are 5.2 million and 5.1 million for the upper and lower bounds of an

unregulated monopoly counterfactual scenario, respectively, which are about 3.0% and

0.8% higher than the initial level. For the Pakistan price counterfactual, the increase is

much smaller, with only about a 0.24% increase from the initial level of poverty.

Among all households with at least one member with diabetes, 0.27 million

households are estimated to be below the lower poverty line, which increases to 0.42

million, 0.30 million, and 0.28 million under each of the three counterfactual

scenarios, respectively. These increases represent a rise in poverty rates ranging from

4.61% to 58.27% from the current level of poverty for these households. Out of all

households that require insulin for one or more members, 0.08 million fall below the

lower poverty line, which increases by 194.51% to 0.23 million under the upper bound

and 49.06% to 0.12 million under the lower bound of an unregulated monopoly

counterfactual scenario. Under the Pakistan price counterfactual scenario, the number

of households that fall below the lower poverty line is 0.09 million, which is 15.40%

higher than the initial level. The pattern of increases in the poverty rates are similar

under the upper poverty line.

Table 9 reports the numbers of households that are below the lower and upper

poverty lines, and the percentage increase in poverty from the initial level of poverty

under the various counterfactual price increase scenarios. We also estimated the

poverty rates as a fraction of total households for three different aggregate levels.

After an increase in the insulin price under a stricter IPR regime, the fraction of

total households that fall below the lower poverty line ranges from 12.92% to 13.28%

Table 7. Initial Level of Poverty in Bangladesh

Lower Poverty Line Upper Poverty Line

TotalHouseholds(million)

Households inPoverty(million)

PovertyRate(%)

Householdsin Poverty(million)

PovertyRate(%)

1. National 39.33 5.07 12.89 9.55 24.282. Households with diabetes 1.05 0.27 25.36 0.35 33.543. Households needing insulin 0.38 0.08 20.99 0.10 27.44

Source: Authors’ calculations based on data from the Bangladesh Bureau of Statistics. 2019. Report onthe Household Income and Expenditure Survey 2016. Dhaka: Statistics and Information Division,Ministry of Planning, Government of Bangladesh.



Table8.

Pov

erty

Rates

inBan

glad

eshafteran

Increase

inInsulin

Prices

Hou

seholdsin

Hou

seholdsin

Pov

erty

after

aPrice

Increase

(million):

Increase

inHou

sehold

Pov

erty

Rates

(%):

Pov

erty

before

aPrice

Increase

Upper

Bou

nd

Low

erBou

nd

Pak

istan

Price

Upper

Bou

nd

Low

erBou

nd

Pak

istan

Price

(million)

Low

erPov

erty

Line

1.National

5.07

5.22

5.11

5.08

3.05

0.77

0.24

2.Hou

seho

ldswith

diabetes

0.27

0.42

0.30

0.28

58.27

14.69

4.61

3.Hou

seho

ldsneedinginsulin

0.08

0.23

0.12

0.09

194.51

49.06

15.40

Upper

Pov

erty

Line

1.National

9.55

9.69

9.58

9.57

1.45

0.36

0.18

2.Hou

seho

ldswith

diabetes

0.35

0.49

0.39

0.37

39.45

9.80

4.92

3.Hou

seho

ldsneedinginsulin

0.10

0.24

0.14

0.12

133.23

33.11

16.62

Sou

rce:Autho

rs’calculations

basedon

datafrom


Statistics.20

19.Reporton

theHou

seho

ldIncome

andExpenditure

Survey

2016


ationDivision,

Ministryof

Plann

ing,

Gov

ernm

entof

Bangladesh.



Table

9.Pov

erty

Rates

inBan

glad

eshbeforean

dafteran

Increase

inInsulin

Pricesan

dPercentage

Chan

ge

Hou

seholdPov

erty

Rates

after

Price

Increase

Percentage

PointIncrease

inHou

seholdPov

erty

Rates

Hou

seholdPov

erty

Ratebefore

Upper

Bou

nd

Low

erBou

nd

Pak

istan

Price

Upper

Bou

nd

Low

erBou

nd

Pak

istan

Price

Price

Increase

Low

erPov

erty

Line

1.National

12.89

13.28

12.99

12.92

0.39

0.10

0.03

2.Hou

seho

ldswith

diabetes

25.36

40.14

29.09

26.53

14.78

3.73

1.17

3.Hou

seho

ldsneedinginsulin

20.99

61.81

31.28

24.22

40.82

10.30

3.23

Upper

Pov

erty

Line

1.National

24.28

24.63

24.37

24.33

0.35

0.09

0.04

2.Hou

seho

ldswith

diabetes

33.54

46.78

36.83

35.20

13.23

3.29

1.65

3.Hou

seho

ldsneedinginsulin

27.44

64.01

36.53

32.01

36.56

9.09

4.56

Sou

rce:Autho

rs’calculations

basedon

datafrom


Statistics.20

19.R

eporto

ntheHou

seho

ldIncomean

dExpenditure

Survey

2016


ationDivision,

Ministryof

Plann

ing,

Gov

ernm

entof

Bangladesh.



under these three counterfactual scenarios, with an increase in poverty rates ranging

from 0.03 percentage points to 0.39 percentage points. Similarly, among all the

households with at least one person with diabetes, the share of households that will be

under the lower poverty line increases from an initial 25.36% to 26.53%–40.14%, with

the increase in poverty rates ranging from 1.17 percentage points to 14.78 percentage

points. We can see a very substantial increase in household poverty among the

households needing insulin. Here, under the three different counterfactual scenarios,

the share of households that fall below the lower poverty line among all households

needing insulin range from 24.22% to 61.81% from an initial poverty rate of 20.99%,

which indicates an increase in poverty rates ranging from 3.23 percentage points to

40.82 percentage points for those households. We see a very similar pattern in

increased poverty rates when we use the upper poverty line instead of the lower

poverty line.

VI. Discussion and Conclusion

This paper is built on the previous theoretical and empirical insights to estimate

the potential impact of Bangladesh’s LDC graduation on its population living with

diabetes in general and insulin users in particular. To date, few if any studies deploy an

ex ante partial equilibrium framework that estimates price changes due to trade policy

change and then links those results to household behavior models and data. We model

and then estimate the potential impact of LDC graduation on the price of insulin in

Bangladesh and then link those price changes. Following those estimates, we calculate

demand elasticities and relate them to Bangladeshi household data to determine the

impacts of those potential price changes in household wealth.

Our findings have significant policy ramifications as well. Bangladesh has a

high incidence of diabetes and insulin users, as well as a fairly thriving domestic

industry that supplies those treatments to patients in need. We find that prices of

insulin would increase significantly in Bangladesh due to LDC graduation and the

subsequent requirement to comply with the IPR provisions of the WTO. What is

more, such price changes would also have significant welfare impacts for the

population. LDC graduation would trigger a significant jump in insulin prices that

could cause a 1%–15% decline in the welfare (i.e., an increase in expenditure) of

households with diabetes, increasing the poverty rate of households with diabetes by

54%–58% and of those needing insulin by 15%–195% unless policy adjustments

were carried out.



Our estimates of the impact of an increase in insulin price under a stronger IPR

regime on household welfare and poverty has some important data limitations. These

limitations emanated from the lack of detailed expenditure information on medicines

by individuals with diabetes. The 2016 HIES of Bangladesh does not provide

disaggregated data on types of diabetic medicines, i.e., whether an individual with

diabetes needs insulin or noninsulin medicines, and it contains no information on the

quantity of medicines needed per day or per month. To construct the sample for our

analysis, we needed to infer the households needing insulin from the expenditure on

medicines for chronic disease reported in the 2016 HIES and compare this expenditure

to an interval constructed using administrative data on monthly insulin costs. It was

likely that there would be some households needing insulin but not included in our

sample if the household’s monthly expenditure on medicines fell below the lower

bound of the cost of insulin constructed using administrative data. Similarly, there

would be some households that do not need insulin but expenditure on medicines by

those households fell within the interval. In the prior scenario, our household welfare

and poverty estimates would be underestimated, and in the latter scenario, these would

be overestimated. Hence, without additional information on medicine expenditure by

the households with members living with diabetes, we could not determine the

direction of bias that our constructed sample may induce.

Another data limitation in the 2016 HIES is that it seems to underrepresent the

fraction of the population suffering from diabetes. In the final report, 186,078

individuals were included in the survey, but only 2,238 individuals were reported to be

living with diabetes, which is about 1.2% of the sample. However, it has been

estimated that about 10% of the population of Bangladesh are suffering from diabetes,

with half of them going undiagnosed (WHO 2016). Hence, we would expect about 5%

of the individuals in our sample to report a diagnosis of diabetes. The

underrepresentation of individuals with diabetes in the 2016 HIES would also cause

a downward bias in estimation. Thus, in this case, our estimated effects of an increase

in insulin price on households’ welfare and poverty are conservative estimates, which

signifies that the true welfare cost of a stricter IPR regime in Bangladesh after its

graduation from LDC status would be significantly higher.

That said, this paper should not be the last word on this subject for Bangladesh,

but rather it should start a discussion. As noted earlier, this analysis suffers from a lack

of data availability in a transparent manner. Better data collection and dissemination

will be paramount in achieving a better understanding of these issues in economics in

general and in Bangladesh in particular.



References

Abbott, Ryan B., Rania Bader, Lina Bajjali, Taher Abu ElSamen, Thamer Obeidat, HananSboul, Mustafa Shwayat, and Ibrahim Alabbadi. 2012. “The Price of Medicines inJordan: The Cost of Trade-Based Intellectual Property.” Journal of Generic Medicines9 (2): 75–85.

Akaleephan, Chutima, Suwit Wibulpolprasert, Rungpetch Sakulbumrungsil, Paithip Luan-gruangrong, Anchalee Jitraknathee, Achara Aeksaengsri, Siripa Udomaksorn, VirojTangcharoensathien, and Sripen Tantivess. 2009. “Extension of Market Exclusivity andIts Impact on the Accessibility to Essential Medicines, and Drug Expense in Thailand:Analysis of the Effect of TRIPs-Plus Proposal.” Health Policy 91 (2): 174–82.

Araar, Abdelkrim, and Paolo Verme. 2016. Prices and Welfare. Washington, DC: Poverty andEquity Global Practice Group, World Bank Group.

Baker, Brook K. 2016. “Trans-Pacific Partnership Provisions in Intellectual Property,Transparency, and Investment Chapters Threaten Access to Medicines in the US andElsewhere.” PLoS Medicine 13 (3): e1001970.

Bangladesh Bureau of Statistics. 2019. Report on the Household Income and ExpenditureSurvey 2016. Dhaka: Statistics and Information Division, Ministry of Planning,Government of Bangladesh.

Banks, James, Richard Blundell, and Arthur Lewbel. 1997. “Quadratic Engel Curves andConsumer Demand.” Review of Economics and Statistics 79 (1997): 527–39.

Basant, Rakesh. 2007. “Intellectual Property Rights Regimes: Comparison of Pharma Prices inIndia and Pakistan.” Economic and Political Weekly 42 (39): 3969–77.

Bhattacharya, Debapriya. 2009. “SABER.” http://saber.eaber.org/node/23153 (accessed 22January 2020).

Biswas, T., A. Islam, L. B. Rawal, and S. M. S. Islam. 2016. “Increasing Prevalence of Diabetesin Bangladesh: A Scoping Review.” Public Health 138 (2016): 4–11.

Chatterjee, Chirantan, Kensuke Kubo, and Viswanath Pingali. 2015. “The Consumer WelfareImplications of Governmental Policies and Firm Strategy in Markets for Medicines.”Journal of Health Economics 44 (2015): 255–73.

Chaudhuri, Shubham, Pinelopi K. Goldberg, and Panle Jia. 2006. “Estimating the Effects ofGlobal Patent Protection in Pharmaceuticals: A Case Study of Quinolones in India.”American Economic Review 96 (5): 1477–514.

Chaves, Gabriela, Walter Britto Gasper, and Marcela Fogaça Viera. 2017. “Mercosur–EU FreeTrade Agreement: Impact Analysis of TRIPS-Plus Measures Proposed by the EU onPublic Purchases and Domestic Production of HIV and Hepatitis C Medicines in Brazil.”Minist�erio da Saúde, Brazil.

Chipman, John S., and James C. Moore. 1980. “Compensating Variation, Consumer’s Surplus,and Welfare.” The American Economic Review 70 (5): 933–49.

Chowdhury, Alam Mohammad Aktarul. 2018. “Enforcement of Intellectual Property Rights inBangladesh: To What Extent Is It TRIPS-Responsive?” Beijing Law Review 9 (2019):425–38.

Deaton, Angus, and John Muellbauer. 1980. “An Almost Ideal Demand System.” TheAmerican Economic Review 70 (3): 312–26.



Directorate General of Drug Administration. “Registered Products.” Government ofBangladesh. http://www.dgda.gotav.bd/index.php/manufacturers/allopathic (accessed 14October 2019).

Dutta, Antara. 2011. “From Free Entry to Patent Protection: Welfare Implications for the IndianPharmaceutical Industry.” The Review of Economics and Statistics 93 (1): 160–78.

Ewen, Margaret, Huibert-Jan Joosse, David Beran, and Richard Laing. 2019. “Insulin Prices,Availability and Affordability in 13 Low-Income and Middle-Income Countries.” BMJGlobal Health 4 (3): e001410.

Government of Bangladesh, Ministry of Health and Family Welfare. 2016. BangladeshNational Health Accounts, 1997–2015. Dhaka: Health Economics Unit.

Hasan, Md. Mehedi, Fariha Tasnim, Md. Tariqujjaman, and Sayem Ahmed. 2019.“Socioeconomic Inequalities of Undiagnosed Diabetes in a Resource-Poor Setting:Insights from the Cross-Sectional Bangladesh Demographic and Health Survey 2011.”International Journal of Environmental Research and Public Health 16 (1): 115.

IHME. 2019. “Bangladesh.” http://www.healthdata.org/bangladesh (accessed 5 March 2020).International Diabetes Federation. 2009. IDFDiabetes Atlas: Fourth Edition. https://diabetesatlas.

org/idfawp/resource-files/2006/07/IDF_diabetes_atlas_ fourth_edition.pdfInternational Household Survey Network. 2020. “Household Income and Expenditure Survey,

2016–2017.” Bangladesh Bureau of Statistics. https://catalog.ihsn.org/index.php/catalog/7399 (accessed 28 February 2021).

Islam, Md. Deen, Warren A. Kaplan, Danielle Trachtenberg, Rachel Thrasher, Kevin P.Gallagher, and Veronika J. Wirtz. 2019. “Impacts of Intellectual Property Provisions inTrade Treaties on Access to Medicine in Low and Middle Income Countries: ASystematic Review.” Globalization and Health 15 (1): 88.

Islam, Shariful Sheikh Mohammed, Andreas Lechner, Uta Ferrari, Michael Laxy, JochenSeissler, Jonathan Brown, Louis W. Niessen, and Rolf Holle. 2017. “Healthcare Use andExpenditure for Diabetes in Bangladesh.” BMJ Global Health 2 (1): e000023.

Islam, Sharmin, Aminur Rahman, and Abu Kholdun Al-Mahmood. 2018. “BangladeshPharmaceutical Industry: Perspective and the Prospects.” Bangladesh Journal of MedicalScience 17 (4): 519–25.

Islam, Sheikh Mohammed Shariful, Andreas Lechner, Uta Ferrari, Michael Laxy, JochenSeissler, Jonathan Brown, Louise W Niessen, and Rolf Holle. 2017. “Healthcare Use andExpenditure for Diabetes in Bangladesh.” BMJ Global Health 2 (1): e000033.

Kyle, Margaret, and Yi Qian. 2014. “Intellectual Property Rights and Access to Innovation:Evidence from TRIPS.” National Bureau of Economic Research Working Paper 20799.

Mohiuddin, Abu Kholdun. 2018. “An A–Z of Pharma Industry Review: BangladeshPerspectives.” PahrmaTutal 6 (12): 64–78.

. 2019. “Diabetes Fact: Bangladesh Perspective.” International Journal of DiabetesResearch 2 (1): 14–20.

National Institute of Diabetes and Digestive and Kidney Disease. Diabetes. https://www.niddk.nih.gov/health-information/diabetes (accessed 1 June 2020).

Poi, Brian P. 2002. “Dairy Policy and Consumer Welfare.” In Three Essays in AppliedEconometrics. Chapter II. Ann Arbor: Department of Economics, University ofMichigan.



. 2012. “Easy Demand-System Estimation with QUAIDS.” The Stata Journal 12 (3): 433–46.

Ray, Ranjan. 1983. “Measuring the Costs of Children: An Alternative Approach.” Journal ofPublic Economics 22 (1): 89–102.

Saeedi, Pouya, Inga Petersohn, Paraskevi Salpea, Belma Malanda, Suvi Karuranga, NigelUnwin, Stephen Colagiuri, Leonor Guariguata, Ayesha A. Motala, Katherine Ogurtsova,Jonathan E. Shaw, Dominic Bright, and Rhys Williams. 2019. “Global and RegionalDiabetes Prevalence Estimates for 2019 and Projections for 2030 and 2045: Results fromthe International Diabetes Federation Diabetes Atlas, 9th edition.” Diabetes Research andClinical Practice 157 (2019): 107843.

Shaffer, Ellen R., and Joseph E. Brenner. 2009. “A Trade Agreement’s Impact on Access toGeneric Drugs.” Health Affairs 28 (5): w957–w968.

Trachtenberg, Danielle, Warren A. Kaplan, Veronika J. Wirtz, and Kevin P. Gallagher. 2020.“The Effects of Trade Agreements on Imports of Biologics: Evidence from Chile.”Globalization and Development. doi: https://doi.org/10.1515/jgd-2019-0030.

United Nations. 2019. “Ex Ante Assessment of the Possible Impacts of the Graduation ofBangladesh from the Category of Least Developed Countries (LDCs).” https://www.un.org/development/desa/dpad/wp-content/uploads/sites/45/IA-Bangladesh-2019.pdf.

. 2020. “UN Office of the High Representative for the Least Developed Countries,Landlocked Developing Countries and Small Island Developing States.” http://unohrlls.org/about-ldcs/criteria-for-ldcs/.

. 2021. “Graduation of Bangladesh, Lao People’s Democratic Republic and Nepal from theLDC Category.” CDP News, November 24. https://www.un.org/development/desa/dpad/2021/graduation-of-bangladesh-lao-peoples-democratic-republic-and-nepal-from-the-ldc-category/.

World Bank. “World Development Indicators.” https://databank.worldbank.org/source/world-development-indicators (accessed 1 June 2020).

World Health Organization. 2015. Bangladesh Health System Review. Manila: Regional Officefor the Western Pacific.

. 2016. Diabetes Country Profiles: Bangladesh. https://www.who.int/diabetes/country-profiles/bgd en.pdf?ua=1.

. 2019. “Diabetes.” Factsheets, November 10. https://www.who.int/news-room/fact-sheets/detail/diabetes.

World Trade Organization. 2015. “General Council–Least Developed Country Members–Obligations under Article 70.8 and Article 70.9 of the TRIPS Agreement with Respect toPharmaceutical Product Patents.” 30 November. https://docs.wto.org/dol2fe/Pages/FE Search/FE S S006.aspx?Query=(%20@Symbol=%20(wt/l/971%20))&Langua-ge=ENGLISH&Context=FomerScriptedSearch&languageUIChanged=true#.



Appendix 1. Summary Statistics

Table A1. Summary Statistics: Mean and Standard Deviation of Household andHousehold Head’s Characteristics

All Households with at Least OneMember with Diabetes Households Needing Insulin

Obs. Mean SD Obs. Mean SD

Household size 1,124 6.38 3.34 424 6.44 3.35Average age 1,124 35.03 12.24 424 35.15 11.59Head age 1,124 51.48 13.17 424 51.49 12.80Monthly income (Tk) 1,125 28,716.56 27,074.13 424 30,936.63 30,831.34Monthly food

expenditure (Tk)1,124 9,210.79 66,10.99 424 9,368.45 4,855.21

Monthly medicineexpenditure (Tk)

1,125 1,128.66 26,80.49 424 973.33 545.03

Monthly educationexpenditure (Tk)

1,015 1,562.88 2,673.55 383 1,786.54 3,034.72

Obs: ¼ observations, SD ¼ standard deviation, Tk = Bangladesh taka.Source: Authors’ calculations based on data from the Bangladesh Bureau of Statistics. 2019. Reporton the Household Income and Expenditure Survey 2016. Dhaka: Statistics and Information Division,Ministry of Planning, Government of Bangladesh.

Table A2. Summary Statistics: Proportions of Household andHousehold Head’s Characteristics

All with at Least OneMember with Diabetes

HouseholdsNeeding Insulin

Obs.Proportion

(%) Obs.Proportion

(%)

Location Rural 618 51.07 202 47.64Urban 592 48.93 222 52.36

House ownership Does not own 104 8.60 32 7.55Owns a house 1,106 91.40 392 92.45

Religion Non-Muslim 143 11.82 48 11.32Muslim 1,067 88.18 376 88.68

Members attending school 0 392 32.37 134 31.601 389 32.12 139 32.782 297 24.53 106 25.003 98 8.09 35 8.25More than 3 35 2.90 10 2.36

Continued.



Appendix 2. Additional Tables


All with at Least OneMember with Diabetes

HouseholdsNeeding Insulin

Obs.Proportion

(%) Obs.Proportion

(%)

Members older than 60 years 0 661 54.63 231 54.481 418 34.55 146 34.432 129 10.66 46 10.853 2 0.17 1 0.24

Members with noncommunicablediseases

1 643 53.41 217 51.302 449 37.29 168 39.723 83 6.89 32 7.57More than 3 29 2.41 6 1.42

Household head’s employment status Unemployed 269 22.25 96 22.64Employed 940 77.75 328 77.36

Household head’s employment sector Agriculture 252 20.83 87 20.52Industry 129 10.66 51 12.03Service 829 69.00 286 67.00

Obs: ¼ observations.Source: Authors’ calculations based on data from the Bangladesh Bureau of Statistics. 2019. Report onthe Household Income and Expenditure Survey 2016. Dhaka: Statistics and Information Division,Ministry of Planning, Government of Bangladesh.

Table B1. Coefficients of QUAIDS Model


OLS IV OLS IV

αα1 0.736*** 0.884*** 0.704*** 0.917***α2 0.124*** 0.259*** 0.147*** 0.240***α3 0.141*** �0.143 0.149*** �0.157*

ββ1 0.164*** 0.224*** 0.111* 0.174β2 �0.097*** 0.184*** �0.098*** �0.245**β3 �0.067*** �0.040 �0.013 0.072

Continued.



Table B1. Continued.


OLS IV OLS IV

γγ11 0.064*** 0.034 0.062*** 0.040γ21 �0.041*** 0.012 �0.040*** 0.010γ31 �0.023*** �0.046* �0.022*** �0.050**γ22 0.055*** �0.023* 0.055*** �0.022*γ32 �0.014*** 0.012* �0.015*** 0.011*γ33 0.036*** 0.034 0.036*** 0.039

�

�1 0.005*** �0.030*** 0.006*** �0.030***�2 �0.003*** 0.029*** �0.003*** 0.030***�3 �0.003* 0.002 �0.003** 0.001

ηη Hsize 1 0.001 0.018*** 0.003 0.021**η Hsize 2 0.001 �0.007** 0.000 �0.006*η Hsize 3 �0.003 �0.011** �0.003* �0.015**η AvgAge 1 �0.001** 0.001 �0.001* 0.002η AvgAge 2 0.000 �0.001 0.000 �0.001η AvgAge 3 0.001* �0.000 0.000** �0.001η NumSch 1 �0.011*** �0.050*** �0.012*** �0.047***η NumSch 2 �0.003** 0.011** �0.002 0.011***η NumSch 3 0.014*** 0.039*** 0.013*** 0.036***η old60 1 �0.002 �0.20** �0.004 �0.023**η old60 2 �0.001 0.007* 0.000 0.006η old60 3 0.003 0.012** 0.003 0.017**η NumNCD 1 0.018*** �0.002 0.017*** �0.003η NumNCD 2 �0.009*** 0.005 �0.008*** 0.006η NumNCD 3 �0.009*** �0.003 �0.008** �0.003η HeadAge 1 �0.000 �0.001 0.000 �0.001η HeadAge 2 �0.000 0.000 �0.000 0.000η HeadAge 3 0.000 0.001 0.000 0.001

η HeadEmpl 1 �0.007 �0.050*** �0.004 �0.055***η HeadEmpl 2 0.004 0.026*** 0.002 0.029***η HeadEmpl 3 0.004 0.024* 0.003 0.027**η HeadSector 1 �0.008 �0.017* �0.009 �0.019η HeadSector 2 0.005 0.010* 0.005 0.013**η HeadSector 3 0.003 0.007 0.004 0.006η HeadMuslim 1 �0.014 0.003 �0.012 0.010η HeadMuslim 2 0.012** �0.008 0.011** �0.008η HeadMuslim 3 0.002 0.005 0.001 �0.002η House 1 �0.035*** �0.019 �0.032*** �0.016

Continued.



Table B2. Suppliers of Insulin in Bangladesh

Suppliers of Insulin in Bangladesh

Domestic Producers (50 products) Import (65 products)

1. Advanced Chemical Industries Limited 1. Eli Lilly & Company, USA (License expired as of 2016)2. Aristopharma Limited 2. Lilly France S.A.S3. Beximco Pharmaceuticals Ltd. 3. Novo Nordisk A/S4. Drug International Ltd. 4. Novo Noris Producao Pharmaceutica do Brasil Ltd.

Brazil.5. Incepta Pharmaceuticals Ltd. 5. Novo Nordisk Production SAS (License expired

as of 2018)6. Popular Pharmaceuticals Ltd. 6. Sanofi Aventis Deutschland7. Square Pharmaceuticals Ltd.

Source: Government of Bangladesh, Directorate General of Drug Administration.

Table B1. Continued.


OLS IV OLS IV

η House 2 0.017** �0.000 0.017** 0.005η House 3 0.018*** 0.019 0.015*** 0.011η Urban 1 �0.016*** �0.035** �0.015*** �0.031η Urban 2 0.002 �0.002 0.003 0.000η Urban 3 0.015*** 0.037*** 0.012*** 0.031*η IMR 1 0.027 0.019η IMR 2 0.008 0.028η IMR 3 �0.036 �0.047

�

� Hsize �0.019 1.230 0.001 0.290� AvgAge �0.006 0.298 �0.003 0.177� NumSch 0.035 �1.266* 0.011 �1.752� old60 0.018 �1.491 �0.006 �0.271� NumNCD 0.190** �0.531 0.126** �0.613� HeadAge �0.001 �0.056 0.000 �0.099� HeadEmpl �0.052 �0.393 �0.016 �1.958� HeadSector �0.069 �0.090 �0.047 �2.001� HeadMulim �0.163 7.168* �0.107 7.748*� house �0.321** 0.220 �0.233** �1.473� Urban �0.037 �1.653 �0.052 �1.832� IMR �0.310* 16.22*

IV ¼ instrumental variable, OLS ¼ ordinary least squares.Source: Authors’ calculations based on data from the Bangladesh Bureauof Statistics. 2019. Report on the Household Income and ExpenditureSurvey 2016. Dhaka: Statistics and Information Division, Ministry ofPlanning, Government of Bangladesh.



Table B3. Estimates of Insulin Demand Equation

Dependent Variable is Total Expenditure on Insulin


Coefficient SE Coefficient SE

p2IV �0.11 0.22 �0.112 0.22�! 0.03 0.02 0.027 0.02AvgAge 7.85* 4.29 7.803* 4.38NumSch �52.94* 32.05 �53.19 32.45old60 �48.42 36.22 �47.94 37.36NumNCD 74.04* 38.54 74.45* 39.34HeadAge 0.65 3.01 0.490 4.32HeadGender �172.700 109.90 �169.0 130.00HeadEduc 12.35 33.64 14.95 59.41HeadEmpl 186.90** 80.60 184.6** 91.51HeadSector 36.74 46.47 33.27 80.14Urban 106.30 78.48 106.5 78.65HeadMuslim 92.49 86.32 95.46 102.90House �28.37 102.40 �23.91 132.50IMR 76.30 1,436.70N 421 421

Adjusted R20.066 0.066

N ¼ number of observations, SE ¼ standard error.Source: Authors’ calculations based on data from the BangladeshBureau of Statistics. 2019. Report on the Household Income andExpenditure Survey 2016. Dhaka: Statistics and InformationDivision, Ministry of Planning, Government of Bangladesh.



Institutions and the Rate of Returnon Cattle: Evidence from Bangladesh

KAZI IQBAL, KAZI ALI TOUFIQUE,AND MD. WAHID FERDOUS IBON

¤

This study extends the recent debate on the rate of return on cattle rearing inIndia, triggered by Anagol, Etang, and Karlan (2017) and followed by others,to the Bangladeshi context and finds that the apparent paradox of widespreadcattle rearing despite negative returns in India is absent in Bangladesh. We usea nationally representative two-year panel data for rural Bangladesh and findthat the average and marginal returns on raising cows and bullocks are positiveand high in both 2011 and 2015. We show that appreciation of the value ofcattle is the major contributing factor to positive returns. The existence of cattlemarkets where cattle can be freely traded for slaughter, milk production, or forany other purpose—which is constrained to various degrees in India—is thekey to high and positive returns in Bangladesh.

Keywords: Bangladesh, livestock, poverty, rural development

JEL codes: C23, L25, O12, Q12

I. Introduction

Anagol, Etang, and Karlan (2017) unleashed a debate over raising livestock in

India. They used survey data collected from Uttar Pradesh to estimate returns from

raising livestock and found that the median return to cows was �7%, and 51% and

⁄Kazi Iqbal (corresponding author): Bangladesh Institute of Development Studies (BIDS), Dhaka,Bangladesh. E-mail: [email protected]; Kazi Ali Toufique (deceased): Bangladesh Institute ofDevelopment Studies (BIDS), Dhaka, Bangladesh; Md. Wahid Ferdous Ibon: Department of Economics,University of Dhaka, Bangladesh. E-mail: [email protected].



Asian Development Review, Vol. 39, No. 1, pp. 281–313DOI: 10.1142/S011611052250007X


https://dx.doi.org/10.1142/S011611052250007X

https://dx.doi.org/10.1142/S011611052250007X

45% of households earned negative returns on cows and buffaloes, respectively.

When labor costs were included, they found that more than half of milk cows had

negative returns. This, they claimed, contradicts the fundamental tenets of capitalism

where activities generating negative returns would have been given up. They discussed

a comprehensive list of factors that could explain the puzzle of negative returns:

measurement errors, preference for illiquid savings, insurance, variation of returns

over the years, labor market failures, milk market failures, and social, cultural, and

religious values.

Two subsequent papers joined the debate and added some interesting

dimensions. Attanasio and Augsburg (2018) argued that the data used by Anagol,

Etang, and Karlan (2017) came from a drought year characterized by scarcity of fodder

and lower milk production, resulting in low returns. Attanasio and Augsburg used a

three-year panel data to show that returns were positive in normal years (good rain,

low fodder costs, and higher milk production because of better nutrition) and negative

in drought years (bad rain and high fodder costs). Subsequently, Gehrke and Grimm

(2018) joined the debate to check the generalizability of the results and introduced an

analysis of marginal returns and economies of scale. They also found that most

households operated at unprofitable levels and returns to livestock varied by quality of

cattle, size of stock, and annual rainfall. Those with cattle of better quality had higher

returns and those with a larger herd enjoyed economies of scale due to decreasing

labor costs. To summarize, the debate has generated four main results: returns to

raising livestock are predominantly negative, they vary from one year to another, there

are scale economies with larger farms showing diminishing costs, and cattle of better

quality generate higher returns.

This study contributes to the debate by extending the geographical horizon and

testing the results mentioned above in the context of livestock rearing in Bangladesh.

This shift in geographical focus from India to Bangladesh involves a shift in religious

and cultural beliefs that have relevance for the debate. There are no restrictions on

buying and selling of cows, slaughtering, movement across the country, or on the

consumption of beef in Bangladesh. Unrestricted and widespread markets for cattle

imply that they can be sold at any time in their life cycle, and their value as an asset is

not constrained by trading restrictions. Thus, Bangladesh provides a counterfactual to

India where most states have restrictive cattle trading and slaughtering policies.

We use household-level panel data from the Bangladesh Integrated Household

Survey (BIHS) of the International Food Policy Research Institute (IFPRI), which

is representative of the rural areas of all eight administrative divisions of the country.

For the sake of brevity and clarity, we combine survey years 2011 and 2015 for our



main analysis. However, we also present results for 2011 and 2015 separately in the

online appendix to highlight year-specific characteristics.1

The key finding of this study is the predominance of positive and high rates of

return to raising cattle (i.e., large ruminants such as bullocks and cows) in

Bangladesh.2 We also find that the average annual return from raising livestock is

31.2% for the combined sample, with a higher return in 2015. These results are also

robust for small ruminants. Our data reveal that appreciation of the value of cattle is

the major contributing factor to large positive returns. Over a period of 12 months, the

value of cattle appreciated by 39%, and this figure is as high as 53% for households

that own only bullocks. In contrast, Anagol, Etang, and Karlan (2017) estimated that

the median value of cows depreciated by 3.1%, while Gehrke and Grimm (2018)

estimated a 40% depreciation in the value of cattle.3

We find that average returns tend to decrease with herd size, reaching a

maximum when herd size is equal to one. Thus, unlike India, there are diseconomies

of scale from livestock rearing in Bangladesh. BIHS data do not provide information

on the breed of the livestock but there is secondary evidence showing that, similar to

India, livestock of better variety generate higher returns in Bangladesh (Jabbar et al.

2005, Kabir and Talukder 1999, Gisby 2010).

Higher returns from raising livestock in Bangladesh are largely due to

appreciation of the market value of livestock. This is particularly true for bullocks

that are raised for their meat value. There is no reason to believe that Bangladeshi

farmers are more “rational” than Indian farmers or face more competitive markets for

milk, fodder, and other inputs and outputs. The difference lies in the cultural and

religious contexts that maximize the market value of cattle in Bangladesh but not in

India. Article 48 of the Constitution of India mandates the state to prohibit the

slaughter of cows and calves and other milk and draft cattle (Ministry of Law and

Justice 2020). Out of 29 states in India, 24 currently have various regulations

prohibiting the slaughter of cows. States such as Kerala, Meghalaya, Mizoram,

and Nagaland have no restrictions on cattle slaughtering. Other factors including

state-level politics may have also constrained the cattle market. For example, in both

1The online appendix can be accessed at: https://www.researchgate.net/publication/354776441_ONLINE_APPENDIX.

2Livestock includes cows (female) and bullocks (male) only. Buffaloes are not included becausethey represent only 6% of the cattle population of Bangladesh. We use the term cattle as consisting of cowsand bullocks only and use it interchangeably with the word livestock.

3The difference in depreciation rates of these two studies is largely driven by the nature of thesample, as the sample of Anagol, Etang, and Karlan (2017) includes a large number of heifers (Gehrke andGrimm 2018).

INSTITUTIONS AND THE RATE OF RETURN ON CATTLE 283


Uttar Pradesh (surveyed by Anagol, Etang, and Karlan 2017) and Andhra Pradesh

(surveyed by Gehrke and Grimm 2018), slaughtering cows is banned, but bullocks and

buffaloes can be slaughtered upon obtaining a “fit-for-slaughter” certificate that the

animal is not economical or is not likely to become economical for the purpose of

breeding or draft/agricultural operations (Citizens for Justice and Peace 2018).

However, transportation of cattle within Uttar Pradesh and between this state and other

states (without permit) is forbidden, but not in Andhra Pradesh. This indicates a wide

variation in institutional setups across states, which may have adversely affected the

performance of the market for buying and selling livestock.

Our study has a bearing on the literature on anti-poverty programs based on asset

(livestock) transfers and the return on capital of microenterprises in developing

countries. There is a growing evidence that suggests that if livestock, generally cows

or goats, are given away to extremely poor households along with a set of

complementary inputs such as skills training, health support, consumption support,

etc., the beneficiaries tend to increase their labor supply and, as a result, their income

and assets likewise increase (Bandiera et al. 2017, Banerjee et al. 2015). In a recent

paper, Banerjee, Duflo, and Sharma (2020) found that such programs also have

substantial long-term impacts. Taking better advantage of opportunities to diversify

into more productive wage employment and migration has been found to contribute to

such successes. However, these studies could not separate the impact of livestock from

other inputs as the interventions were delivered as a bundle. Therefore, the rate of

return on livestock can be very high due to the presence of other complementary

inputs.

The rest of the paper is organized as follows. The next section describes the data

and presents relevant descriptive statistics. Revenue, costs, profit, and average and

marginal returns to raising livestock are estimated in Section III, while heterogeneity

in returns is analyzed in Section IV. We perform robustness checks in Section V.

We discuss overall findings in Section VI, and Section VII draws the conclusions.

II. Context, Data, and Descriptive Statistics

A. Livestock Sector in Bangladesh

The livestock sector plays an integral part of the rural economy of Bangladesh.

About 37.6% of rural households in Bangladesh had at least one livestock in 2015

(IFPRI 2015). It is estimated that about 20% of employment in the rural economy is



directly associated with the livestock subsector, and this figure is about 50% when we

consider indirect employment (GOB 2018).

Though the livestock sector grew by 3.2% in 2016, its contribution to gross

domestic product (GDP) fell from 2.2% in 2008 to 1.7% in 2016 (GOB 2016).

The contribution of the livestock sector to the overall agriculture sector was almost

static at around 13% during the same period. The number of bovine populations

(cattle, buffalo, goat, and sheep) has increased since 1960, but their growth could not

match the growth of the human population (Huque and Huda 2016). As a result, the

per capita bovine population in Bangladesh declined from 0.4 in 1960 to 0.2 in 2018.

Meat consumption in Bangladesh steadily increased from 11.6 grams per person

per day (g/p/d) in 1995 to 18.6 g/p/d in 2010 (BBS 1995, BBS 2010a). Beef consumption

increased by about 11% between 2010 and 2016 (BBS 2016). Per capita milk

consumption in Bangladesh is about 126 milliliter per person per year which is

significantly lower than other countries in the region (Kabir, Islam, and Reza 2018).

According to the household income and expenditure survey (BBS 2010a), 79% of

Bangladeshi households eat meat no more than two days in a fortnight. Only about

63% of households drink milk and 60% eat eggs in a fortnight (Toufique and Belton

2014).

Existing studies estimating returns from livestock rearing in Bangladesh have

found them to be generally profitable. All the studies estimated returns without taking

into consideration appreciation or depreciation of livestock assets except for the study

undertaken by Gisby (2010).4

B. Data

There are several advantages of using the Bangladesh Integrated Household

Survey (BIHS of the International Food Policy Research Institute [IFPRI] dataset

[IFPRI 2011, IFPRI 2015]). First, the sample is nationally representative of rural

Bangladesh and representative of rural areas of each of the seven administrative

divisions of the country. Second, the BIHS conducted two rounds of surveys in 2011

and 2015 on the same households. We combine them to create household-level panel

data, which allow us to control household-level time-invariant heterogeneity in

estimating marginal returns. Third, since the objective of the survey is to study

agriculture, food security, nutrition, and poverty, detailed data are collected on

4Table A14 of the online appendix lists the names, data, methodology, and main findings of therelevant literature in the context of Bangladesh. Note that the appendix is available as an online supplementdue to space limitations of the journal.



livestock raised by households. The survey contains information on livestock

ownership (large/small ruminants and poultry), including the livestock’s current and

last year’s values, revenues, and costs.

It is imperative to discuss whether the 2 sample years are normal years, as

year-specific shocks can influence the rate of return on cattle substantially. Data from the

World Bank’s Climate Change Knowledge Portal (2021) show that the average amounts

of rainfall in 2011 and 2015 were 233.2 and 212 millimeters, respectively. Note that

average rainfall for the period 1990–2015 was 194.8 millimeters, indicating that our

sample years experienced flood, particularly in 2011. However, our data show that only

48 households (4.5% of the working sample) in 2011 and 34 households (3.9% of the

working sample) in 2015 experienced crop loss due to floods. We found that farm

income as a share of total household incomewas 60% in both 2011 and 2015. Hence, the

loss of income due to flood was not significant in our sample. However, the floodwaters

may wash away haystacks and damage grass fields. The market demand for cattle and

milk may also decrease due to floods. But our data show that the rate of appreciation of

cattle value was higher in 2011 than in 2015 (44% versus 33%). However, other factors

such as higher fodder costs and wages for hired labor may lead to an underestimation of

costs to some extent in 2011. If this is the case, we may be estimating some lower

bounds of the rate of return on livestock rearing. This, in fact, corroborates our overall

results that the rate of return on livestock rearing is positive and high in Bangladesh.

The number of households included in the first (2011) and second rounds (2015)

of the BIHS was 4,423 and 4,419, respectively. Out of these households, 41.1%

(1,817) and 37.2% (1,643) owned cattle in the first and second rounds of the surveys,

respectively (Table 1). Farm animals in Bangladesh consist of cattle, buffalo, goat, and

sheep, but in this study, we only considered cattle, which consist of cows and

bullocks. As buffalo is not a common type of livestock in Bangladesh, we dropped

buffalo-owning households from the analysis.5

The BIHS data provide information at the household level only. The data record

the total number of cattle by type (cows, bullocks) for each household, and associated

revenues and costs are also presented in the aggregate. Information on cattle was

collected for the last 12 months in both survey rounds. The BIHS data recorded all

transactions in cattle that resulted in either depletion or accumulation of stock at the

end of the survey year.6

5In the agricultural census of 2008, only 0.7% of rural households reported owning buffaloes (BBS2010b). In the BIHS, only 20 households in 2011 and 10 in 2015 reported owning a buffalo.

6BIHS data provide two data points (beginning and end of the survey year) on size and value ofcattle stock in both rounds of surveys (2011 and 2015).



The size of the herd can change from the initial period for the following reasons:

(i) sale of cattle, (ii) purchase of a cattle, (iii) exchange of cattle as gift, (iv) lease of

cattle from other households, (v) slaughter of cattle for consumption of meat, and (vi)

loss of cattle due to theft or death. We could only account for cattle sales (i) and loss of

cattle (vi) because the price of cattle sold and the value of cattle lost are reported in the

survey. For cases (ii) to (v), the information to calculate the rate of appreciation and

hence the rate of return on cattle is incomplete. Therefore, we drop cases (ii) to (v) in

our analysis, which leaves us with 1,065 and 884 households in 2011 and 2015,

respectively (Table 1).

To capture the heterogeneity of returns by gender of cattle, we divide all

households with livestock into three categories: (i) households with only bullocks, (ii)

households with only milk cows, and (iii) households with both bullocks and milk

cows (i.e., households with cattle of both gender). Note that Anagol, Etang, and Karlan

(2017) and Gehrke and Grimm (2018) considered milk cows and milk buffaloes only.

C. Descriptive Statistics

We first check whether the households omitted from our sample are different

from the sample households in a systematic way since we had to drop 41% and 46% of

cattle-owning households from our 2011 and 2015 sample, respectively, as shown in

Table 1. The table also shows the number and percentage of households with

livestock, households dropped from the sample, and households in the working sample

for each type of cattle. If the households were dropped in a systematic way, it would

create a selection bias in our sample. To this end, we report in Table 2 the distribution

of the four categories of households that were dropped from the sample due to

incomplete information on purchases, gifts, lease, and home consumption of cattle for

2011 and 2015. The table shows that purchases comprise about 60% of the dropped

sample in both years. The second-largest category at about 21%–23% is leasing. In the

sample of cattle-owning households, only 17% of the households that purchased in

2011 likewise purchased cattle in 2015. Hence, these are not the same households

purchasing cattle in both years. If they were the same households, one could argue that

these households always expand herd size and thus share some common

characteristics that have bearing on returns. This potential “systematic pattern”

could create a bias in our sample. Our data also show that the households that

purchased cattle come from different income groups. About 43% of the dropped

samples are from the first two quintiles of per capita expenditure. As we will see later,

average returns tend to be lower for these lower quintiles. Since the share of the



Table

1.Number

ofHou

seholdswithLivestock,Hou

seholdsDropp

ed,an

dHou

seholdsin

WorkingSam

ple

byTyp

eof

Cattle

2011

2015

Hou

seholdswith

No.

ofHHswith

Livestock

No.

ofHHs

Dropped

No.

ofHHsin

WorkingSam

ple

No.

ofHHs

withLivestock

No.

ofHHs

Dropped

No.

ofHHsin

WorkingSam

ple

Onlybu

llocks

641

319

322

593

357

236

(35.27

)(42.42

)(30.23

)(36.09

)(47.0)

(26.70

)

Onlymilk

cows

569

188

381

495

185

310

(31.31

)(25.00

)(35.77

)(30.12

)(24.4)

(35.07

)

Bothbu

llocksandmilk

cows

607

245

362

555

217

338

(33.40

)(32.57

)(33.99

)(33.77

)(28.6)

(38.24

)

All

1,81

775

21,06

51,64

375

988

4(100

.00)

(100

.00)

(100

.00)

(100

.00)

(100

.00)

(100

.00)

HH¼

household.

Note:

The

numbers

inparenthesesarecolumnpercentagesandmay

notsum

upto

100du

eto

roun

ding

.Sou

rce:

Autho

rs’calculations

usingdata

from

theInternationalFoo

dPolicyResearchInstitu

te,“B

angladeshIntegrated

Hou

seho

ldSurvey(BIH

S)

2011–12

,”https://d

ataverse.harvard.edu

/dataset.xhtml?persistentId=hd

l:190

2.1/21

266and“B

angladeshIntegrated

Hou

seho

ldSurvey(BIH

S)20

15,”

https://d


/dataset.xhtml?persistentId=do

i:10.79

10/DVN/BXSYEL(bothaccessed

1June

2018

).



Table2.

Distribution

ofHou

seholdsDropped

byReason

HHsDropped

Dueto

Incomplete

Inform

ationon

No.

ofHHs

Dropped

in20

11

Percentage

ofHHs

Dropp

edin

2011

(N¼

752)

No.

ofHHs

Dropped

in20

15

Percentage

ofHHs

Dropped

in20

15(N

¼75

9)

1Purchaseof

cattle

441

58.64

468

61.66

2Cattle

received

asgift

618.11

587.64

Cattle

givenas

gift

344.52

212.77

3Lease

ofcattlefrom

otherHHs

174

23.14

159

20.95

4Slaug

hter

ofcattleforow

nconsum

ption

425.59

536.98

Total

752

100.00

759

100.00

HH

¼ho

usehold,

N¼

numberof

households.

Sou

rce:Autho

rs’calculations

usingdatafrom

theInternationalF

oodPolicyResearchInstitu

te,“BangladeshIntegrated

Hou

seho

ldSurvey(BIH

S)

2011–12

,”https://d



l:190

2.1/21

266

and

“Bangladesh

Integrated

Hou

seho

ldSurvey

(BIH

S)

2015

,”https://d



i:10.79


1June

2018

).



dropped sample is not disproportionately higher in these lower quintiles, it lowers the

concern for overestimating the rate of return on cattle. The shares of gifts and home

consumption of cattle are both very low. Those who leased cattle are not owners, so we

drop this group from the sample.

We compare the characteristics of cattle-owning households that are included in

the sample with those that are dropped from the sample in Table 3. There is no

statistically significant difference between the working sample and the dropped

sample in per capita monthly expenditure. This suggests that the two groups are

similar in terms of income status. However, per capita food expenditure per month

was slightly lower in the working sample than the dropped sample in 2011.

The difference was 76 taka (Tk), which could buy about 2 kilograms of rice in 2011.

On the contrary, in 2015, per capita food expenditure was slightly higher for the

working sample at a difference of Tk59. Thus, this difference is not systematic over

the years. These random and meager differences are very unlikely to produce any

selection problems.

Table 3 also shows that the working sample is richer in landownerhip than the

dropped sample. This difference is driven by quintile 1 (lowest) and quintile 5

(highest) of the land distribution. Note that it is the homestead land that determines the

ownership and herd size of the cattle (Table A12 in online appendix). Since there is no

difference in homestead land between our working sample and the dropped sample, the

generalization of our results is less likely to suffer from a sample selection problem.

The fact that the household head is 2 years older than the dropped sample (48 versus

46) in 2015 is less likely to affect the generalization of results.

To further highlight that the working sample compares well with the dropped

sample in farming characteristics, we report the p-value of the mean differences in

Table 4. In particular, we focus on those variables that enter into the calculation of

returns. We observe that there is hardly any difference in the components of revenue

and costs in livestock rearing between these two samples in both years. There is no

statistically significant difference between these two samples in the case of revenues

from manure and milk, and costs of labor and fodder. The cost of fodder is slightly

higher for the working sample than the dropped sample, and the p-value of the mean

difference is 0.097 in 2015.

In short, the socioeconomic and the farming characteristics are very similar for

the working and dropped samples, lending support for the generalization of results of

the working sample. Note that we have a limited set of variables and there might still

be some systematic differences, but we cannot check for those. However, given the set

of variables we have, we are confident that there is no systematic selection bias.



Table3.

Hou

seholdCharacteristicsof

CattleOwners:

WorkingSam

ple

vs.Dropped

Sam

ple

2011

2015

CattleOwners

(WorkingSam

ple)

CattleOwners

(Dropped

Sam

ple)Difference

inMean

CattleOwners

(WorkingSam

ple)

CattleOwners

(Dropped

Sam

ple)Difference

inMean

Mean

SD

Mean

SD

p-value

Mean

SD

Mean

SD

p-value

Maleho

useholdhead

0.93

0.26

0.92

0.26

0.67

80.92

0.27

0.91

0.27

0.96

9Age

ofho

useholdhead

45.97

13.17

45.03

13.19

0.13

548

.212

.88

45.89

12.8

0.00

0Hou

seho

ldsize

4.6

1.71

4.56

1.83

0.63

24.74

1.79

4.6

1.8

0.10

2Hou

seho

ldhead

isliterate

0.44

0.49

0.41

0.49

0.21

50.45

0.49

0.44

0.49

0.46

2Male–femaleratio

1.25

0.89

71.21

0.88

0.26

81.23

0.87

1.27

0.89

0.32

8Per

capita

food

expend

iture

(Tk,

mon

thly)

1,02

961

31,10

570

70.01

31,05

567

399

662

30.07

0Per

capita

totalexpend

iture

(Tk,

mon

thly)

2,42

72,35

42,53

92,25

20.31

02,51

42,37

52,37

53,24

60.31

8Hom

estead

land

owned(decim

al)

11.01

12.91

10.5

13.45

0.41

610

.41

12.07

9.8

11.54

0.29

7Total

land

owned(decim

al)

97.00

170.00

77.00

129.00

0.00

710

4.00

179.00

85.00

159.00

0.01

9Observatio

ns1,06

575

288

475

9

SD¼

standard

deviation,

Tk¼

Bangladeshtaka.

Notes:Allfigu

resarein

constant

2011

taka.$1

¼Tk7

4.2.

Sou

rce:

Autho

rs’calculations

usingdata

from

theInternationalFoo


te,“B

angladeshIntegrated

Hou

seho

ldSurvey(BIH

S)20

11–12

,”https://d



l:190

2.1/21

266

and“B

angladesh

Integrated

Hou

seho

ldSurvey(BIH

S)20

15,”

https://d

ataverse.

harvard.edu/dataset.x

html?persistentId=do

i:10.79


1June

2018

).



Table

4.Characteristicsof

CattleFarming:

Com

parison

betweenDropped

andWorkingSam

ples

WorkingSam

ple

Dropped

Sam

ple

MeanDifference

CattleFarmingVariables

NMean(Tk,yearly)

SD

NMean(Tk,yearly)

SD

p-value

2011

Milk

revenu

e48

210

,164

23,354

.60

286

8,83

6.00

4,27

5.63

0.38

5Manurerevenu

e1,05

31,97

42,19

0.70

670

1,83

0.00

2,44

5.60

0.10

7Fod

dercost

883

4,87

010

,880

.21

619

5,14

3.00

9,32

0.26

0.61

2Fam

ilylabo

rcost

1,05

56,97

45,37

1.44

714

6,76

8.21

4,95

4.18

0.41

5Fam

ilylabo

rcost:male

895

4,34

85,08

7.90

628

4,13

8.33

3,91

0.29

0.38

5Fam

ilylabo

rcost:female

994

3,48

72,25

4.42

679

3,28

9.59

2,82

0.85

0.113

Wagelabo

r23

8,75

115

,038

.26

712

,736

.36

16,739

.37

0.49

1Wagelabo

r:male

218,93

715

,293

.28

712

,736

.36

16,739

.37

0.52

3

2015

App

reciation

884

6,35

612

,276

.13

260

11,342

18,937

.60

0.64

4Manurerevenu

e76

51,54

01,69

2.97

400

1,59

82,93

4.97

0.64

3Fod

dercost

725

4,69

47,56

8.19

600

4,113

8,49

8.93

0.09

7Fam

ilylabo

rcost

882

3,86

42,02

6.66

732

3,79

62,119.95

0.25

8Fam

ilylabo

rcost:male

800

2,55

21,62

2.45

655

2,50

01,69

6.87

0.27

8Fam

ilylabo

rcost:female

846

1,61

499

4.75

710

1,54

595

9.92

0.08

5Wagelabo

r7

3,58

83,05

3.30

34,54

13,87

3.91

0.66

4Wagelabo

r:male

73,58

83,05

3.30

34,54

13,87

3.91

0.66

4

SD¼

standard

deviation,

Tk¼

Bangladeshtaka.

Notes:Allmon

etaryfigu

resarein

constant

2011

taka.$1

¼Tk7

4.2.

Nis

thenu

mberof

households

with

positiv

evalues

forvariou

scattle

farm

ingvariables.The

samplesize

offemalewagelabo

ristoosm

allto

estim

atethep-valueof

themeandifference.

Sou

rce:

Autho

rs’calculations

usingdata

from

theInternationalFoo


te,“B

angladeshIntegrated

Hou

seho

ldSurvey

(BIH

S)20

11–12

,”https://d



l:190

2.1/21

266and“B

angladeshIntegrated

Hou

seho

ldSurvey

(BIH

S)20

15,”

https://d



i:10.79


1June

2018

).



Next, we present the main characteristics of cattle farming of the sample

households in Table 5.7 The BIHS collected information on the current value of

livestock as well as their value a year before as reported by respondents. The values of

the stock per household at the end of the reference periods were Tk33,810 in 2011 and

Tk37,089 in 2015, respectively, in real terms, and the average values of the stock at the

end of the reference period were Tk15,581 and Tk16,411, respectively. The increase in

average value of the stock may indicate that the quality of cattle increased between the

survey years. The herd size is highest for those with both milk cows and bullocks.

On the other hand, herd size is slightly higher for those with only bullocks than those

with only milk cows. This indicates that many households raise livestock to slaughter

for meat production. In both cases, the herd size increased during the survey periods.

The herd size of mixed farms slightly declined.

Table 5 also presents the components of revenues and costs. Note that average

appreciations (i.e., meat value), defined as the average of the change in the value of the

stock between the initial and terminal periods are reported in detail in Table 6. Average

revenue from selling milk increased between the survey years, by 6.4% in real terms

for households with positive milk revenue. However, the contribution of manure as a

source of revenue decreased over time. Fodder contributed the most in the cost of

rearing livestock in Bangladesh. Average fodder cost was about Tk4,870 in 2011 and

Tk4,694 in 2015 for households with positive fodder costs. The decrease in fodder

costs is not statistically significant. Wage labor was hired by only a few households.

Family labor was mostly employed in these farms. We observe that family labor costs

have declined in real terms and no clear explanation can be given. Female family

members worked more hours on cattle rearing than male members. In 2015, male

members spent about 368 hours compared to 417 hours by female members.

At the outset, it is important to note that we do not make the returns based on

imputed values of family labor salient in our discussion of results. There are two major

reasons, which are explained in the following. The explanation also sheds light on why

returns are consistently low and negative in most cases in 2011 while high and positive

in 2015 when we value family labor at market price.

First, the households in our sample are smallholders—the average cattle size is

about 2.2 in 2011 and 2.3 in 2015. Thus, the time that households spend per day to

take care of the cattle is also very low. Our data show that time spent by male members

of a household on livestock in a year is 368 hours, that is, about one hour per day.

7Note that we also present household characteristics of the cattle owners and nonowners in theonline appendix in Table A1.



Table

5.Characteristicsof

CattleFarming,

2011

and20

15

2011

2015

MeanDifference


NMean

SD

NMean

SD

p-value

Value

ofstockof

cattleperHH

(Tk)

1,06

533

,810

.00

29,547

.00

884

37,089

.00

30,294

.17

0.02

6Average

cattlevalue(Tk)*

1,06

515

,581

.00

8,44

4.29

884

16,411.00

8,94

1.44

0.03

5Herdsize:on

lybu

llocks

322

1.91

1.19

236

2.03

1.48

0.28

3Herdsize:on

lymilk

cows

381

1.72

1.00

310

1.91

1.39

0.03

7Herdsize:bo

thmilk

cowsandbu

llocks

362

2.87

1.66

338

2.73

1.52

0.25

0Herdsize:all

1,06

52.17

1.41

884

2.26

1.51

0.18

1Milk

revenu

e(Tk,

yearly)

482

10,164

.00

23,354

.60

486

10,811.00

18,462

.54

0.63

2Manurerevenu

e(Tk,

yearly)

1,05

31,97

4.00

2,19

0.70

765

1,54

0.00

1,69

2.97

0.00

0Revenue

from

calves

(Tk,

yearly)

391

8,46

7.00

3,08

8.34

327

8,28

8.00

3,06

0.95

0.43

6Fod

dercost(Tk,

yearly)

883

4,87

0.00

10,880

.21

725

4,69

4.00

6,56

8.19

0.70

3Value

ofcattlelost(Tk,

yearly)

1529

,833

.00

36,941

.88

616

,912

.00

15,183

.65

0.42

2Fam

ilylabo

rcost(Tk,

yearly)*

1,05

56,97

4.00

5,37

1.44

882

3,86

4.00

2,02

6.66

0.00

0Fam

ilylabo

rcost:male(Tk,

yearly)*

895

4,34

8.00

5,08

7.90

800

2,55

2.00

1,62

2.45

0.00

0Fam

ilylabo

rcost:female(Tk,

yearly)*

994

3,48

7.00

2,25

4.42

846

1,61

4.00

994.75

0.00

0Total

timespenton

livestock

inayear

(hou

rs)

1,06

066

7.00

488.86

883

794.00

397.72

0.00

0Malefamily

timespenton

livestock

inayear

(hou

rs)

895

367.00

425.53

884

368.00

273.54

0.93

8Fem

alefamily

timespenton

livestock

inayear

(hou

rs)

994

364.00

232.93

884

417.00

277.41

0.00

0Total

family

timespenton

livestock

inayear

(hou

rs)

1,05

565

5.00

465.82

882

788.00

395.40

0.00

0Hired

timespenton

livestock

inayear

(hou

rs)

2272

9.00

599.37

785

1.00

659.68

0.65

0

Con

tinued.



Table

5.Con

tinued.

2011

2015

MeanDifference


NMean

SD

NMean

SD

p-value

Num

berof

calves

391

1.18

0.43

327

1.15

0.43

0.43

6Total

valueof

cattlesold

(Tk,

yearly)

141

27,732

.00

19,846

.02

153

26,389

.00

25,062

.70

0.61

2Wagelabo

r(Tk,

yearly)

238,93

7.00

15,038

.26

73,58

8.00

3,05

3.30

0.38

0Wagelabo

r:male(Tk,

yearly)

218,75

1.00

15,293

.28

73,58

8.00

3,05

3.30

0.37

2Wagelabo

r:female(Tk,

yearly)

22,42

5.00

813.17

0...

...

...

*¼

calculated

values

(not

repo

rted

values),...¼

data

notavailable,

HH¼

household,

SD¼

standard

deviation,

Tk¼

Bangladeshtaka.

Notes:A

llmon

etaryfigu

resarein

constant

2011

taka.$

1¼

Tk7

4.2.

The

valueof

thestockof

cattleperho

useholdisdefinedas

thetotalv

alue

ofcattle

ownedby

allsampleho

useholds

dividedby

thenu

mberof

sampleho

useholds.Average

cattlevalueisdefinedas

thetotalvalueof

thestockof

cattle

ownedby

thesampleho

useholds

dividedby

thenu

mberof

cattle.

Nis

thenu

mberof

households

with

positiv

evalues

forvariou

scattlefarm

ing

variables.

Sou

rce:

Autho

rs’calculations

usingdata

from

theInternationalFoo


te,“B

angladeshIntegrated

Hou

seho

ldSurvey(BIH

S)

2011–12

,”https://d



l:190

2.1/21

266and“B

angladeshIntegrated

Hou

seho

ldSurvey(BIH

S)20

15,”

https://d



i:10.79


1June

2018

).



Literature suggests that time spent on livestock does not hamper regular work in farm

or nonfarm employment. Typically, the time is spent on taking cattle to the grass field,

bathing, feeding, etc., which the male members do in between their regular work or

after work (Gisby 2010). Hence, the opportunity costs of time spent on livestock is

negligible. Gisby (2010, 22) noted “The majority of those looking after cattle stated

that they would be ‘idle’ in the absence of having to care for cattle.” The household

members do not have to give up any remunerated work time for taking care of cattle.

Hence, monetized family labor is negligible in our analysis.

Second, hourly wage of hired labor is what makes the returns in 2011

significantly lower than the returns in 2015. Note that male wage rates in 2011 and

2015 were Tk11.9 and Tk4.2 per hour, respectively. Female wage rate in 2011 was

Tk9.6 per hour. These lead to yearly labor costs of Tk8,751 in 2011 and only Tk3,588

in 2015 (Table 5). Thus, when we use these figures to impute the costs of family labor,

the figures for 2011 inflate. Our inspection of the data suggests that there are several

“large values” that are driving the very high wage rates in 2011. The standard

deviation of total wages in 2011 was Tk15,038 whereas it was only Tk3,053 in 2015.

We checked wage-related information from the BIHS employment module: the

average hourly wage in employment in livestock was about Tk4 in 2011 and Tk4.8 in

2015. This indicates that the hourly wage in 2015 is a more plausible figure than that in

2011. The survey of literature on Bangladesh (Table A14 in the online appendix) for

the period 2010–2014 suggests that hired labor costs were much lower than in the

BIHS data.8

III. Returns to Cattle Holding

The return to livestock comprises two parts: (i) the flow of profits from the sale of

livestock products such as milk or manure, and (ii) the appreciation or depreciation

of the value of livestock. We follow the empirical specification given by Gehrke and

Grimm (2018). Let the production function of a household from raising livestock be:

Q ¼ Af (K, L,X,F): ð1Þ

8Gisby (2010) found that monthly family labor cost (if paid for the opportunity cost of raising asingle livestock) varies from Tk108 to Tk295 based on cattle type and variety. Halim et al. (2010) recordeda family labor cost of Tk4,877 per year per cattle. Mondal, Sen, and Rayhan (2010) estimated an averagetotal labor (both family and hired) cost of Tk11.4 and Tk20.2 per cow per day for local-variety andcrossbred cows, respectively.



In this production function,Q ¼ milk, calves, andmanure;K ¼ current/end period value

of cattle; L ¼ labor; X ¼ land; F ¼ fodder; and A ¼ household- and region-specific

characteristics that influence the total factor productivity (TFP) of the inputs. Gehrke

and Grimm (2018) lumped X and F together and noted that land entered into the

production function through F. Since we have data on homestead land, we treat X and

F separately. Note that in the case of bullocks, the production function becomes trivial

as Q includes only revenue from manure. The profit function is given by:

� ¼ P � Q� cK � wL� gF � rX � Sþ δK, ð2Þwhere P ¼ price of outputs; w ¼ wage rate (both market and imputed); g ¼ price of

fodder; r ¼ rent of land; c ¼ other costs associated with K (medical, purchase, or sale

of related costs, etc.); S is the value of cattle lost; and δ ¼ rate of appreciation/

depreciation.9 Following Gehrke and Grimm (2018), we assume the price of capital to

be zero. We also assume that the rental value of land, r, is equal to 0. We first calculate

the household-level annual profit and then estimate average and marginal returns of

raising livestock. The following equation shows the average return on raising livestock:

�

K¼ P � Q

K� c� wL

K� gF

K� S

Kþ δ: ð3Þ

Following Gehrke and Grimm (2018), we estimate marginal returns using both linear

production and Cobb–Douglas production technology, using pooled OLS and panel

fixed effects (FE). First, with the linear production function, we estimate profit (�it) as a

function of the value of herd size (value of the livestock, K) and several control variables

at the household level:

�it ¼ α0 þ α1 log (Kit)þ α2(Xit)þ eit, ð4Þwhere i and t stand for households and time, respectively. Xi includes labor cost,

homestead land owned, fodder cost, herd size, and a year dummy. α1 is the marginal

return of holding livestock of value K for a one-year period. Second, we use a

Cobb–Douglas production technology to estimate marginal return. The marginal return

of this type of production function is the first derivative of equation (2) with respect toK:

� 0(Kit) ¼ P � Q 0(Kit)� cþ δ: ð5Þ

9δ ¼ (K � K0)=K0, where K0 is the value of the cattle stock at the initial period and K is the endperiod value of the same stock. For households with constant herd sizes between the two survey periods,calculation of the rate of appreciation/depreciation is straightforward (δ ¼ (K � K0)=K0). In cases withnonconstant herd sizes and households with calves, calculation of δ involves several intermediate stepsdescribed in the online appendix.



If Q is a Cobb–Douglas production function such that

Qit ¼ A � K α1it L

α2it F

α3it , ð6Þ

where all parameters lie strictly between 0 and 1. We get α1 from the following

logarithmic transformation of the production function in equation (6):

log (PQit) ¼ α0 þ α1 logKit þ α2 log Lit þ α3 logFit þ α4Xit þ 2it, ð7Þwhere Xit includes log of homestead land, herd size, and a year dummy. We then plug in

the estimated value of α1 into equation (5) and get the marginal return from raising

livestock in the following form:

� 0(Kit) ¼ P � α1 �Qit

Kit� cþ δ: ð8Þ

A. Appreciation and Depreciation

In Anagol, Etang, and Karlan (2017), the price of a cattle, P(t), is reported by the

farmer. They used self-reported values of dairy animals to establish a relationship

between cattle values and age. This helped them to estimate appreciation as the

difference in cattle values over a period of one year, P(t)� P(t � 1). On the other

hand, Attanasio and Augsburg (2018) could not estimate depreciation because their

data did not contain information on the age of livestock, although they had information

on the value of livestock. Gehrke and Grimm (2018) did not follow Anagol, Etang,

and Karlan (2017) on the estimation of appreciation or depreciation of cattle because

they also did not have information on age. Instead, Gehrke and Grimm (2018) used

information from secondary sources and assumed that a cow depreciates by 1,240

Indian rupees (�) (US$27) every year, and that the end-of-fertility value of a cow

is �1,400 (US$30), which is based on an annual depreciation of 20%.

The approach taken by these papers for estimating appreciation reflects the

institutional setup that exists in the respective Indian states: “Since cattle cannot be

sold for slaughter, this implies that the value of a cow will be zero once it is no longer

of reproductive age” (Gehrke and Grimm 2018, 682). This rationalized the use of age

as an appropriate indicator for estimating appreciation (Anagol, Etang, and Karlan

2017), the use of secondary sources of information (Gehrke and Grimm 2018), or

ignoring measurement of depreciation altogether (Attanasio and Augsburg 2018).

Our method of estimating appreciation is similar to Anagol, Etang, and Karlan

(2017) in the sense that we take the difference between initial and terminal values of

cattle as appreciation, and then divide the difference by the initial value to get the rate

of appreciation. The BIHS data include, for each year, the current value of cattle and



their value in the previous 12 months, as reported by the respondents. This is

appropriate in the Bangladeshi context where there is no restriction on selling cattle for

slaughter. The value of livestock depends not only on age but also on other factors

such as weight of cattle, location, and time. Market conditions may also change

between the initial and terminal period, and the appreciation rate may reflect these

market conditions. These factors are incorporated by the respondents while reporting

the value of their cattle. The BIHS data do not present information on age, weight, or

breed of the cattle and hence we could not estimate the relationship between these

factors and the value of cattle. We would expect these to be incorporated by the

respondent while valuing the livestock. Dependence on self-reported values of cattle

likely involves overreporting by the sellers as they have private information on the

quality of the cattle that is not available to the buyers. This problem is less acute when

cattle are sold for slaughter, as weight is more visible than the potential milk

production of a cow. We also think that any bias that is generated from reported values

of cattle by the respondents is carried over to the next period and thus can cancel out

when estimating how much the cattle stock has appreciated. This asymmetric

information is not handled adequately in existing studies and this study is no

exception. As mentioned before, we have information on the value of cattle sold

during the reference period. We have considered that value as the end period value of

the cattle.

B. Revenue

The BIHS data provide information on three items of revenue: milk, manure, and

calves.

1. Milk and Manure

Households report value and quantity of milk and manure sold in the last 12

months. We consider net milk production by accounting for the amount of spoil.

We used two sets of prices to determine the value of milk. For households who sold

milk in the market, we estimated the price by dividing the value of the milk sold by the

quantity of milk sold. About 53% and 55% of households sold milk in the market in

2011 and 2015, respectively. For households that consumed all the milk they

produced, they provided information on the value of that milk consumed and this

allowed us to determine the market price. Though Anagol, Etang, and Karlan (2017)

mentioned that households may value their milk higher than the price it can fetch in

the market, we find that there is hardly any difference between the self-reported price



and market price. This is true for the full sample as well as for all eight administrative

divisions in the country. The BIHS data also provide information on manure.

The BIHS data lumped together revenues from manure and milk by aggregating

small ruminants such as goat and sheep and large ruminants. It is therefore not possible

to separate milk and manure revenues for households that raise both cattle and small

ruminants. About a quarter of households who own cattle also have small ruminants.

However, milk revenues from small ruminants are very small: they comprise 6% and

9% of average milk revenues for 2011 and 2015, respectively. Small ruminants, on the

other hand, hardly have any relevance to manure revenue.

2. Calves

The BIHS data provide information on the number of calves born in the last

12-month period, but not calves’ market price. We take the calf price to be Tk7,151 in

both 2011 and 2015 from the data collected for the Final Impact Evaluation Survey of

the Second Participatory Livestock Development Project (BIDS 2010). We consider

the calves as a separate component of revenue in the return calculation; they are not

part of herd size.

C. Costs

Unlike revenue, cost data are presented by type of livestock. Three components

of costs are recorded: feed or fodder, medicine or treatment, and labor costs.

1. Fodder

Fodder is a major component of the costs of raising livestock in Bangladesh.

With the gradual diminishing of grazing grounds and other common property

resources, most of the cattle in Bangladesh are stall-fed. Common items of fodder are

straw, green grass, and concentrate. The BIHS data do not provide any information on

collected or home-produced fodder and report the value of purchased fodder only.

Halim et al. (2010) have found that home-produced and collected fodder comprise

about 20% of total purchased fodder when valued at market price. We used this

information to inflate fodder costs by 20% and estimate average and marginal returns.

2. Labor

The BIHS data provide labor use and labor cost information by gender as well by

source (family and hired). Labor use is presented as the number of hours spent on



raising livestock. Family labor is unpaid but hired labor is paid, and the BIHS records

information on hired labor cost. About 98%–99% of labor time used in livestock

farming is provided by family labor. Thus, we present labor cost in two ways: (i) assume

it is 0 for family labor and (ii) monetize family labor. The process of monetizing family

labor is described in the online appendix. Table 6 presents the breakdown of costs and

revenue for three different types of households for the 2 years combined.

As is evident from Table 6, the highest stock appreciation is observed for the

households who only raise bullocks; within a year, the average value of their stock

increases by more than half, and this is true for both years (Tables A2 and A3 in the

online appendix). Relative appreciation is similar (higher for bullocks) to those found

by Gisby (2010) who found that bulls appreciated more than cows, both the local

variety (13% compared to 6%) as well as the crossbred variety (21% compared to 12%).

For the full sample, the average annual appreciation at the household level was 39%.

As expected, the share of milk revenue is highest for households with only milk

cows. Fodder and family labor are the major cost components for all household

categories. Cost and revenue components are similar in 2011 and 2015.

D. Profits and Average and Marginal Returns

Profit from raising livestock is positive (excluding family labor cost) in all

categories in the combined sample (Table 7) as well as for both 2011 and 2015

(Tables A4 and A5 in the online appendix). Once family labor is valued at market

price, profit decreases considerably and becomes negative for households with

bullocks only. The value of calves and the revenue from selling cattle (Table 5)

increase the profits for households with only milk cows and those selling cattle,

respectively.

Both average and marginal returns are positive for the combined sample in both

years except for the average return (with family labor) in 2011. For all household

categories together, the average return (without family labor) from holding cattle is

31.2%. Marginal return using the Cobb–Douglas production function is positive and

high at about 48% (Table 7).10 Intuitively, the annual return from investing an

additional one dollar to the existing stock is about 48 cents.11

10We are aware that the value of herd size (K) can be endogenous due to simultaneity bias. Higherprofit from livestock rearing can motivate households to raise certain breeds and thus affect the value of K.Due to lack of valid instruments, we did not consider instrumental variables estimation.

11See Table A10 of the online appendix for the regressions to estimate the marginal returns ofcapital stock.



Table

6.Com

pon

ents

ofRevenuean

dCost(C

ombined

Tab

lefor20

11an

d20

15)

Revenue

Costs

Appreciation

Initial

Period

CattleValue

(K0)

EndPeriod

Cattle

Value

(K1)

(K1�K

0)*1

00=(K

0)Milk

(Tk)

Man

ure

(Tk)

Calf

(Tk)

Wag

eLab

or(Tk)

Fam

ilyLab

or(Tk)

Purchased

and

Non

purchased

Fod

der

Cost(Tk)

Medicine

andOther

Cost(Tk)

Valueof

Cattle

Lost(Tk)

Hou

seho

ldswith

only

bullo

cks(N

¼55

8)19

,184

29,351

53NA

1,39

6NA

545,13

03,88

429

254

9

Hou

seho

ldswith

only

milk

cows(N

¼69

1)21

,039

30,086

436,36

91,56

03,67

327

5,41

94,26

738

135

0

Hou

seho

ldswith

both

bullo

cksandmilk

cows(N

¼70

0)

34,022

42,187

248,21

82,00

04,97

524

05,94

15,89

751

30

Fullsample(N

¼1,94

9)24

,620

34,222

395,20

91,67

13,08

9111

5,52

44,74

340

328

1

N¼

numberof

households,NA

=no

tapplicable,Tk¼

Bangladeshtaka.

Notes:$1

¼Tk7

4.2.

App

reciation(depreciation)

istherealrateof

increase

(decrease)

ofthecattlestockin

12mon

ths.Allrevenu

eandcostcompo

nentsare

theaveragenu

mbers

forthesampleho

useholds

(N¼

1,94

9 ).Non

purchasedfodd

ercosthasbeen

assumed

as20

%of

thepu

rchasedfodd

ercost.

Sou

rce:Autho

rs’calculations

usingdatafrom

theInternationalFoo


te,“BangladeshIntegrated

Hou

seho

ldSurvey(BIH

S)20

11–12

,”https://d



l:190

2.1/21

266and“B

angladeshIntegrated

Hou

seho

ldSurvey(BIH

S)20

15,”

https://d

ataverse.

harvard.edu/dataset.x

html?persistentId=do

i:10.79


1June

2018

).



Table

7.Average

andMarginal

Returnsfrom

RaisingLivestock

(Com

bined

Tab

lefor20

11an

d20

15)

Total

Annual

Profit(Tk)

Average

Return

(%)

With

Fam

ilyLab

orWithou

tFam

ilyLab

orWithFam

ilyLab

orWithou

tFam

ilyLab

orMarginal

Return

Hou

seho

ldswith

only

bullo

cks

(N¼

558)

�3,101

2,02

8�1

8.22

8.28

0.54

Hou

seho

ldswith

only

milk

cows

(N¼

691)

5,67

311,092

13.79

40.09

0.54

Hou

seho

ldswith

both

bullo

cksand

milk

cows(N

¼70

0)8,21

814

,159

20.75

40.67

0.37

Fullsample(N

¼1,94

9)4,07

59,59

97.12

31.19

0.48

N¼

numberof

households,Tk¼

Bangladeshtaka.

Notes:$

1¼

Tk7

4.2.

“With

family

labo

r”im

pliesmon

etized

valueof

family

labo

r,and“w

ithou

tfam

ilylabo

r”im

plieson

lyhired

labo

rcosts.

Sou

rce:

Autho

rs’calculations

using

data

from

theInternationalFoo

dPolicy

Research

Institu

te,“B

angladesh

Integrated

Hou

seho

ldSurvey(BIH

S)20

11–12

,”https://d



l:190

2.1/21

266and“B

angla-

desh

Integrated

Hou

seho

ldSurvey(BIH

S)20

15,”

https://d



i:10.79

10/DVN/

BXSYEL(bothaccessed

1June

2018

).



IV. Heterogeneity in Average and Marginal Returns

In this section, we discuss the heterogeneity of returns from two aspects: (i) herd

size and (ii) household food expenditure. For the first aspect, we explore if there are

any economies of scale in herd size; and for the second, we check whether rates of

return change for different poverty groups as indicated by household food expenditure

quintiles (including home-produced food expenditure).

Table 8 shows that for both survey years combined, average return is maximized

at herd size one and minimized at herd size four when family labor cost is not

monetized. However, when we value family labor at market price, no systematic

patterns exist in the relationship between average returns and herd size. This is true for

2011 (Table A6 in the online appendix). But we find that even if we consider family

labor, we observe the highest average returns at herd size one in 2015 (Table A7 in the

online appendix). Thus, it appears that there are no economies of scale in raising

livestock in Bangladesh. To shed more light on this issue, we examine the sources of

the costs of rearing that can potentially lead to economies of scale and whether such

costs were incurred in large amounts in the context of rural Bangladesh. Cattle rearing

involves the following cost components: fodder, medicine, space, and hired labor to

take care of the cattle. Consider fodder and medicine first. The per unit cost (the cost

for one cow or bullock) of fodder and medicine is unlikely to vary much with herd size

since these costs are very specific to each animal. In the case of space in the homestead

where cattle are housed, there is scope for economies of scale. However, for a small

herd size, which is about 2.2 animals, additional cattle may cost little in terms of

Table 8. Herd Size and Returns (Combined Table for 2011 and 2015)

Herd Size

Average Valueof Total

Stock (Tk)

Average Returnwith FamilyLabor (%)

Average Returnwithout FamilyLabor (%)

MarginalReturn Observations

1 17,130 7.4 41.04 0.58 7422 15,804 7.97 29.62 0.48 6083 14,492 6.32 23.11 0.34 3174 14,839 2.4 17.62 0.47 152> 4 14,856 9.03 17.87 0.25 130

Tk ¼ Bangladesh taka.Note: $1 ¼ Tk74.2.Source: Authors’ calculations using data from the International Food Policy Research Institute,“Bangladesh Integrated Household Survey (BIHS) 2011–12,” https://dataverse.harvard.edu/dataset.xhtml?persistentId=hdl:1902.1/21266 and “Bangladesh Integrated Household Survey (BIHS) 2015,”https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/BXSYEL (both accessed 1June 2018).



extending the space or house. Moreover, this cost is a fixed cost, which is not

considered in the costing exercise. Hired labor is the potential source of economies of

scale—the average cost of taking care of the cattle may drop significantly with herd

size, particularly for smaller sizes. However, only 22 households in 2011 and 7

households in 2015 used hired labor for cattle rearing. Hence our finding of no

economies of scale in cattle rearing by smallholders is not implausible.

Poorer households raise more livestock. Table 9 shows that in the combined

sample, the incidence of livestock rearing by households in the bottom quintile (Q1) is

more than twice of those households in the top quintile (Q5). We find that although

poorer households raise more livestock, they earn a lower annual return than their

richer counterparts (Q4 and Q5). To examine further, we regress the incidence of cattle

rearing and herd size on the quintiles of food consumption, controlling for other

covariates. The results show that the lower quintiles are more likely to raise cattle and

have lower herd size. The results for 2011 and 2015 are reported in Tables A8 and A9

of the online appendix.

V. Robustness Check

Since we have dropped a large sample for which we could not calculate the

appreciation rate, we use the working sample’s average appreciation rates of cattle to

Table 9. Livestock Variables and Per Capita Food Expenditure Quintiles (Combined Tablefor 2011 and 2015)

Per CapitaFoodExpenditureQuintiles (Q)

% of HHswith

LivestockHerdSize

AverageValue ofTotal

Stock (Tk)

AverageReturnwithoutFamily

Labor (%)

AverageReturnwith

FamilyLabor (%)

MarginalReturn

Q1 56.70 2.31 14,298 28.10 4.14 0.46Q2 43.27 2.17 15,075 30.76 6.36 0.49Q3 36.01 2.20 15,220 27.92 3.45 0.45Q4 32.35 2.16 16,777 34.51 9.35 0.59Q5 27.32 2.21 18,421 34.67 12.32 0.39

Tk ¼ Bangladesh taka.Note: $1 ¼ Tk74.2.Source: Authors’ calculations using data from the International Food Policy Research Institute,“Bangladesh Integrated Household Survey (BIHS) 2011–12,” https://dataverse.harvard.edu/dataset.xhtml?persistentId=hdl:1902.1/21266 and “Bangladesh Integrated Household Survey (BIHS) 2015,”https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/BXSYEL (both accessed 1June 2018).



impute the rates for the dropped sample as a robustness check. These appreciation

rates are 0.4 for 2011 and 0.3 for 2015. With these imputed rates, the average returns

without monetizing family labor are 25.6% in 2011 and 36.8% in 2015 for the dropped

sample (Table A13 in the online appendix). These two rates compare very well with

the appreciation rates of the working sample for both years. The marginal returns of

livestock rearing for both working and dropped samples are also very similar.

In the context of rural Bangladesh, households depend less on the market

for fodder. We find that our results of high average returns are robust to accounting for

imputed fodder costs. All the above rates of returns we have discussed so far are for

large ruminants only. Since a large number of households are also raising small

ruminants (goats and sheep) along with large ones, the rate of return must also be high

for small ruminants. We find that the average returns from raising only small ruminants

are 26% and 35% in 2011 and 2015, respectively (Table A11 in the online appendix).

The marginal returns are also high at 42% and 53% in these 2 years.

The above exercises suggest that our estimates of returns from rearing livestock

are very robust.

VI. Discussion on the Debate

This paper is motivated by the debate triggered by Anagol, Etang, and Karlan

(2017) and the availability of the nationally representative BIHS data that contain

information on livestock holdings. This allowed us to estimate rates of return from

rearing livestock in rural Bangladesh and check whether these are similar to those

found in India. The estimation of positive and high rates of return kept us pondering

why, unlike India, this is the case in Bangladesh. We strongly believe that this is due to

two factors. First, a cow in Bangladesh has value not only for the capacity to produce

milk but also for producing meat for consumption. Second, markets for buying and

selling cattle for meat consumption freely exist that help farmers dispose of their cattle

whenever needed and for whatever purpose. As an asset, a cow is therefore more

liquid in Bangladesh than in India.

A milk cow has two main attributes: it provides milk, and it also provides meat.

A bull, on the other hand, has only one major attribute, that of providing meat.

The relevance of livestock as draft power has severely diminished over time in

Bangladesh due to the mechanization of agriculture, and other benefits such as manure

from livestock are minor. In India, livestock are raised mainly to produce milk, and

meat can be viewed as a by-product. This is not the case in Bangladesh where a cow



can be sold not only for its milk but also for its meat. This is the basis of our

classification of households into three categories, those who own only bullocks and

only milk cows, and those who have both. About 35% of the households raise only

milk cows. About 27%–30% of the households rear only bullocks. Those who rear

only cows do so mainly because of milk, either for home consumption, to sell at the

market, or both. We do not have enough information about the extent of selling cows

for slaughter. A survey carried out by Toufique, Iqbal, and Ibon (2018) on the cattle

markets established for the sacrificial Eid-ul-Adha found that about 8% of the cattle

sold for slaughter are cows. A cow is least preferred because of the common

perception that a pregnant cow cannot be sacrificed, and buyers do not want to take the

risk of buying a cow. The BIHS data show that about 30% of cattle sold in a year

comprise cows. This figure includes cows bought for milk, but a part of it represents

those sold for meat. Thus, a cow can be sold at any time in its life and has value for both

milk and meat.

In both Uttar Pradesh and Andhra Pradesh, cows can never be slaughtered

irrespective of their age and therefore cows are traded only for their milk.12

The existence of a market for cattle, where they can be sold anytime either for milk or

meat or both, implies that the value of cattle is determined not only by age alone but

also by their breed (recognized by Anagol, Etang, and Karlan 2017), overall health,

buyers’ preference, condition of the market, etc. When a bull or cow can be bought for

slaughter, the value of the cattle includes the value of the meat, a component that is

largely missing in most Indian contexts. Since a cattle market for meat consumption

does not exist in many states in India or exist only for a limited type of cattle, such as

old, worn out, or unproductive cattle, the value of cattle is not maximized because the

meat value is almost zero.13 India’s beef industry is predominantly based on the

slaughtering of water buffaloes. According to the existing meat export policies in

India, the export of beef (meat of cows, oxen, and calves) is prohibited.

A reexamination of depreciation of the value of livestock by Anagol, Etang, and

Karlan (2017) justifiably used age as a determinant of the value of a cow in the context

of restrictions on slaughtering or meat consumption in most states in India. In a linear

depreciation method, as used by Gehrke and Grimm (2018), the value of livestock

always declines by a fixed amount each year as there is a finite period during which

holding livestock is useful. This observation leads Gehrke and Grimm (2018, 682) to

12In Andhra Pradesh, bullocks can be slaughtered when it is certified unproductive, normally at theage of 15.

13However, there is evidence that beef/buffalo meat consumption has increased among the Hindupopulation in India (See Bansal 2016, Sathyamala 2018).



comment: “we assume that the market value of a cow approaches 0 with the end of

its fertility.”

When cattle can be slaughtered for meat consumption and there are competitive

markets for buying and selling cattle, the price of cattle can be considered as a more

reliable indicator of the current value of cattle. The cattle owner will likely sell,

ignoring distress sales, when the value of the cattle is highest, and this may depend on

many factors, although age and number of lactations continue to play a major role.

For example, in Bangladesh, cattle can fetch the highest value before the Eid-ul-Adha

when cattle are sacrificed for religious purposes. Many households buy and fatten

cattle to get the most of their investment from this occasion. Appreciation of buffaloes

is also estimated by using the same method. But since there are no restrictions on

transactions of buffaloes, including slaughtering them, age may not be the only

determinant of their value. In the case of buffaloes, weight is likely to be a better

determinant of asset value. We also observe a wide variation of values of both cows

and buffaloes for any given age.

The BIHS data include the value of cattle over a period of 1 year as reported by

the respondents. This helped us to estimate the appreciation or depreciation of cattle by

taking the price difference during this period. These values not only incorporate the

age of the cow but also the value of meat or other factors that are relevant to prices.

We find appreciation to be very high and mostly positive. For example, the price of

bullocks increased by more than half within a year. Such high appreciation is also

reported by Gisby (2010). Gisby (2010) estimated asset gain as measured by the

difference between buying and expected selling prices of cattle. For bullocks, they find

a monthly asset gain of approximately 13% for local-variety bulls and 22% for

crossbred bulls. The corresponding figures for cows are 6% and 12%, respectively.

We also notice that the rate of appreciation is very different in the two methods

used by Anagol, Etang, and Karlan (2017) and Gehrke and Grimm (2018). In Anagol,

Etang, and Karlan (2017), depreciation for the full sample is around 3.1% of the

median value of a cow. In the method used by Gehrke and Grimm (2018), cows

depreciate by 20% each year. In the full sample, depreciation is 40% of the value of the

cattle. Attanasio and Augsburg (2018, 318) consider appreciation and depreciation to

be a “minor source of costs, so the neglect is unlikely to introduce significant biases.”

Thus, there is a wide gap in the estimates of appreciation between Anagol, Etang, and

Karlan (2017) and Gehrke and Grimm (2018). The high incidence of negative returns

from rearing livestock was found both with relatively low depreciation (Anagol,

Etang, and Karlan 2017) as well as with relatively high depreciation (Gehrke and

Grimm 2018). On the other hand, looking across states in India, a lower incidence of



negative returns can be expected in Andhra Pradesh than in Uttar Pradesh, because

rules on slaughtering cattle are less restrictive in Andhra Pradesh. We have already

mentioned that unproductive bullocks can be slaughtered in Andhra Pradesh, and there

are various formal and informal restrictions on the movement of cattle in Uttar

Pradesh. Besides, possession of beef is also illegal in Uttar Pradesh. The transport of

cattle is important because there are incentives to move cattle to states where cattle

trading and slaughtering are less restrictive. To cite Gehrke and Grimm (2018, 682),

“Of course, reports exist throughout the country of unproductive animals being sold

off to other states in which cattle slaughter is not prohibited.”

In the Indian context, various restrictions on slaughtering have strong

implications on returns to livestock. Negative returns imply that returns from the

sale of livestock products generally fail to account for the depreciation of livestock as

an asset. If various restrictions on livestock trading were absent or less restrained, we

could have observed relatively higher returns. Since cows cannot be slaughtered, a

farmer must retain relatively older cattle for some time as it is difficult to dispose of

them. For example, for a farm with 10 crossbred cows, the National Dairy

Development Board of India (NDDB 2019) has recommended that farmers dispose of

cows that have already had three lactation periods to maintain the productivity of the

herd.14 This also indicates that cows, even when they are allowed to be slaughtered

before they reach the age of 15 or beyond, are not that productive. Allowing cows to

be slaughtered even at the age of 15 helps farmers to get some revenue that can be used

to replenish the stock. The overall public opinion about slaughtering cows also affects

the price of cows, as traders find it difficult or risky to move cows across and within

states when there are strong restrictions against transporting them, such as in Uttar

Pradesh. The restrictions on slaughtering cows have already given rise to the problem

of stray cows. These are mostly cows abandoned by farmers as they become infertile

and hardly get any buyers. These problems are more acute in states with more

restrictive slaughtering rules. In Uttar Pradesh (more restrictive), 5.1% of cattle are

stray cattle compared to 0.4% in Andhra Pradesh (less restrictive).15 These restrictions

on cattle trade or movement surely create a black market for trading cows. This is

recognized by Gehrke and Grimm (2018, 682) but ignored. This suggests that farmers

can find an informal avenue to dispose of their cows, which may have somewhat

increased the rate of return.

14See http://www.dairyknowledge.in/content/10-crossbred-cow-farm (accessed 3 October 2019).15Indian Livestock Census 2012 figures cited in https://thewire.in/politics/modi-government-cow-

slaughter-stray-cattle (accessed 3 October 2019).



VII. Conclusion

By using a nationally representative panel dataset for rural Bangladesh, this

paper finds that, unlike in India, the rates of return from raising cattle in Bangladesh

are high and positive. Positive rates of return in India are either explained by a good

year when fodder costs are low (Attanasio and Augsburg 2018) or by the existence of

economies of scale where households with larger herd sizes only get positive returns

(Gehrke and Grimm 2018). We have argued that positive and high rates of return in

Bangladesh are explained by the existence of a market for cattle in an institutional

setup where there is no moral or religious stigma attached to meat consumption or

trading. Existence of this market adds a new dimension to the relationship between age

and market value of cattle because cattle have value beyond milk and draft power. This

increases the extent of appreciation of cattle of Bangladesh. A market that is missing in

most states in India is present in Bangladesh, and this market increases the value of

livestock held by smallholders.

We find that the average rate of return on cattle rearing is about 31% and the

marginal return is about 48%. First, these findings are not at odds with the literature on

the rate of return on microenterprises (De Mel, McKenzie, and Woodruff 2008).16

We have also documented the rate of return on cattle found in other studies in

Bangladesh (Table A14 in the online appendix). Though the return figures in these

studies are accounting returns, they indicate a very high rate of return for cattle.

For example, Sarma, Raha, and Jørgensen (2014) found that the rate of return was

52% in the Pabna and Sirajganj districts. Earlier Jabbar et al. (2005) also estimated

44% and 55% returns for crossbred and local cows, respectively. Second, the high

interest rate on borrowing in the rural financial market also lends support to high

returns from cattle rearing. Rural finance in Bangladesh is dominated by microfinance

institutions and informal moneylenders. The effective rate of return on microcredit can

be as high as 43% (Faruqee and Khalily 2011). Borrowers also resort to moneylenders

for additional funds and the average annual interest rate of moneylenders was found to

be about 103% (Mallick 2012). In our sample, we find that 65% of households in 2011

and 70% of households in 2015 had outstanding loans with microfinance institutions

or other sources. Third, note that when we value family labor at market prices, the

average rate of return drops significantly. For example, in 2015, the average return is

23% when incorporating family labor and 39% without family labor. It is the

unaccounted family labor that makes the average return very high. This point is also

16De Mel, McKenzie, and Woodruff (2008) found a 6% rate of return per month in Sri Lanka.



highlighted in the microfinance literature to explain higher repayment rates (Emran,

Morshed, and Stiglitz 2021).

Our findings have strong implications for livestock development and poverty

reduction in Bangladesh. Higher rates of return for livestock rearing indicate that there

is scope for further development of the livestock sector. Since rates of return are higher

for poorer households, the possibility to reduce poverty through livestock transfers

remains. There are, however, some worrying signs that justify caution. Though the

returns from livestock are high, they are declining. The poorest households have

reduced livestock rearing more than others, but the extent of livestock rearing for all

households has been falling. This may happen for factors that could not be analyzed

with available data. We think that inadequate livestock services, high costs of fodder,

and other factors could have set this trend. It should be emphasized that a successful

asset-transfer-based anti-poverty program in Bangladesh must be bundled with a

provision of livestock services and transfer of cash.

References

Anagol, Santosh, Alvin Etang, and Dean Karlan. 2017. “Continued Existence of CowsDisproves Central Tenets of Capitalism?” Economic Development and Cultural Change65 (4): 583–618.

Attanasio, Orazio, and Britta Augsburg. 2018. “Holy Cows or Cash Cows? The EconomicReturn to Livestock in Rural India.” Economic Development and Cultural Change 66 (2):307–30.

Banerjee, Abhijit, Esther Duflo, Nathanael Goldberg, Dean Karlan, Robert Osei, WilliamParienté, Jeremy Shapiro, Bram Thuysbaert, and Christopher Udry. 2015. “AMultifaceted Program Causes Lasting Progress for the Very Poor: Evidence from SixCountries.” Science 348 (6236): 1260799.

Bandiera, Oriana, Robin Burgess, Narayan Das, Selim Gulesci, Imran Rasul, and MunshiSulaiman. 2017. “Labor Markets and Poverty in Village Economies.” The QuarterlyJournal of Economics 132 (2): 811–70.

Banerjee, Abhijit, Esther Duflo, and Garima Sharma. 2020. “Long-term Effects of the Targetingthe Ultra Poor Program.” NBER Working Paper 28074, National Bureau of EconomicResearch, Cambridge, MA.

Bangladesh Bureau of Statistics (BBS). 1995. Household Income and Expenditure Survey1995, Ministry of Planning, Government of the People’s Republic of Bangladesh.

_____. 2010a. Household Income and Expenditure Survey 2010, Ministry of Planning,Government of the People’s Republic of Bangladesh.

_____. 2010b. Census of Agriculture 2008: Structure of Agricultural Holdings & LivestockPopulation, Volume 1. Ministry of Planning, Government of the People’s Republic ofBangladesh.



_____. 2016. Household Income and Expenditure Survey 2016, Ministry of Planning,Government of the People’s Republic of Bangladesh.

Bangladesh Institute of Development Studies (BIDS). 2010. Final Impact Evaluation Survey ofSecond Participatory Livestock Development Project (PLDP-II). BIDS, Dhaka.

Bansal, Samrath. 2016. “More Indians Eating Beef, Buffalo Meat.” The Hindu. October 29.http://www.thehindu.com/news/national/%E2%80%98More-Indians-eating-beef-buffalo-meat%E2%80%99/article16085248.ece.

Citizens for Justice and Peace. 2018. “Cow Slaughter Prevention Laws in India.” https://cjp.org.in/cow-slaughter-prevention-laws-in-india/.

De Mel, Suresh, David McKenzie, and Christopher Woodruff. 2008. “Returns to Capital inMicroenterprises: Evidence from a Field Experiment.” The Quarterly Journal ofEconomics 123 (4): 1329–72.

Emran, M. Shahe, A. K. M. Mahbub Morshed, and Joseph E. Stiglitz. 2021. “Microfinance andMissing Markets.” Canadian Journal of Economics 54 (1): 34–67.

Faruqee, Rashid, and M. A. Baqui Khalily. 2011. “Interest Rates in Bangladesh MicrocreditMarket.” Institute of Microfinance Policy Paper, Dhaka, Bangladesh.

Gehrke, Esther, and Michael Grimm. 2018. “Do Cows Have Negative Returns? The EvidenceRevisited.” Economic Development and Cultural Change 66 (4): 673–707.

Gisby, Laura. 2010. “Relative Profitability of Crossbred versus Local Cattle Rearing UnderATP (DRAFT) Chars Livelihoods Programme.” Innovation, Monitoring, and LearningDivision, Government of the People’s Republic of Bangladesh.

Government of Bangladesh (GOB). 2016. “Bangladesh Economic Review.” EconomicAdvisor’s Wing, Finance Division, Ministry of Finance, Government of the People’sRepublic of Bangladesh.

_____. 2018. “Summary of Livestock Economy.” Department of Livestock Services (DLS),Government of the People’s Republic of Bangladesh.

Halim, M. A., M. A. Kashem, J. U. Ahmed, and M. Hossain. 2010. “Economic Analysis of RedChittagong Cattle Farming System in Some Selected Areas of Chittagong District.”Journal of the Bangladesh Agricultural University 8 (2): 271–76.

Huque, K. S., and N. Huda. 2016. “Agriculture Census-Based Dynamics of Changes inLivestock Farming Bangladesh—A Review.” Bangladesh Journal of Animal Science45 (1): 1–6.

International Food Policy Research Institute (IFPRI). 2011. “Bangladesh Integrated HouseholdSurvey (BIHS) 2011–12.” https://dataverse.harvard.edu/dataset.xhtml?persisten-tId=hdl:1902.1/21266 (accessed 1 June 2018).

_____. 2015. “Bangladesh Integrated Household Survey (BIHS) 2015.” https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/BXSYEL (accessed 1 June2018).

Jabbar, Mohammad A., S. M. Fakhrul Islam, Christopher Delgado, Simeon Ehui, M. AminulIslam Akanada, M. I. Khan, and M. Kamruzzaman. 2005. “Policy and Scale FactorsInfluencing Efficiency in Dairy and Poultry Production in Bangladesh.” InternationalLivestock Research Institute (ILRI), Nairobi, Kenya; Systemwide Livestock Programme(SLP), Addis Ababa, Ethiopia; and Bangabandhu Sheikh Mujibur Rahman AgriculturalUniversity (BSMRAU), Salana, Gazipur, Bangladesh.



Kabir, Abid Ul, Sirajul Islam, and Hasib Reza. 2018. “A Study on Milk Value Chains for PoorPeople in Bangladesh.” IDS Bulletin 49 (1).

Kabir, M. H., and R. K. Talukder. 1999. “Economics of Small-Scale Dairy Farmingin Bangladesh under the Government Support Programme.” Animal Bioscience 12 (3):429–34.

Mallick, Debdulal. 2012. “Microfinance and Moneylender Interest Rate: Evidence fromBangladesh.” World Development 40 (6): 1181–89.

Ministry of Law and Justice, Government of India. 2020. “The Constitution of India.” https://legislative.gov.in/constitution-of-india.

Mondal, R. K., S. Sen, and Shah Johir Rayhan. 2010. “A Comparative Economic Analysis ofLocal Breed and Cross Breed Milk Cow in a Selected Area of Bangladesh.” Journal ofScience Foundation 8 (1&2): 23–29.

National Dairy Development Board of India (NDDB). 2019. Dairy Knowledge Portal. http://www.dairyknowledge.in/content/10-crossbred-cow-farm (accessed 3 October 2019).

Sarma, P. K., S. K. Raha, and H. Jørgensen. 2014. “An Economic Analysis of Beef CattleFattening in Selected Areas of Pabna and Sirajganj Districts.” Journal of the BangladeshAgricultural University 12 (1): 127–34.

Sathyamala, C. 2018. “Meat-Eating in India: Whose Food, Whose Politics, and WhoseRights?” Policy Futures in Education 17 (7): 878–91.

Toufique, Kazi Ali, and Ben Belton. 2014. “Is Aquaculture Pro-Poor? Empirical Evidence ofImpacts on Fish Consumption in Bangladesh.” World Development 64: 609–20.

Toufique, Kazi Ali, Kazi Iqbal, and Wahid Ferdous Ibon. 2018. “Survey of Qurbani Cattle Hut-2018.” Bangladesh Institute of Development Studies, Dhaka.

World Bank. Climate Change Knowledge Portal. 2021. https://climateknowledgeportal.worldbank.org/download-data (accessed 1 January 2021).



Impacts of Fuel Subsidy Rationalizationon Sectoral Output and Employment

in Malaysia

NOORASIAH SULAIMAN, MUKARAMAH HARUN,AND ARIEF ANSHORY YUSUF

¤

Large allocations for fuel subsidies have long put the Government ofMalaysia’s budget under great strain. Using a computable general equilibrium(CGE) model, this paper evaluates the impact of fuel subsidy rationalizationon sectoral output and employment. Employment is classified intooccupational categories and skill levels. Fuel subsidies were measuredusing the disaggregation of prices for petrol, diesel, and other fuel products.Findings show that removing fuel subsidies would hit economic performancethrough high input costs, specifically for industries closely attached to thepetroleum refinery sector. The manufacturing sector has the largest reductionin output and employment. Nevertheless, high- and medium-skilled laborforces experience increased demand. To increase economic efficiency, thesavings from the removal of fuel subsidies should be put toward policies suchas sales tax reduction. This study provides useful information for policymakers in evaluating or updating current subsidy policies to reduce economiclosses.

⁄Noorasiah Sulaiman (corresponding author): Center for Sustainable and Inclusive Development Studies,The National University of Malaysia. E-mail: [email protected]; Mukaramah Harun: Universiti UtaraMalaysia College of Business. E-mail: [email protected]; Arief Anshory Yusuf: Center forSustainable Development Goals Studies, Padjadjaran University. E-mail: [email protected] for this paper was supported by the Project of Economy and Environment Program for SoutheastAsia (EEPSEA-2011). We thank the managing editor and the anonymous referees for helpful commentsand suggestions. The Asian Development Bank recognizes “China” as the People’s Republic of China.





https://dx.doi.org/10.1142/S0116110522500081

https://dx.doi.org/10.1142/S0116110522500081

Keywords: computable general equilibrium model, employment, fuel subsidy,sectoral output, subsidy removal

JEL codes: H29, E23, E24, D58

I. Introduction

The high level of uncertainty over future global oil prices, which are more

influenced by international market conditions than domestic factors, places a domestic

economy in a very precarious position. An increase in fuel prices affects government

spending on fuel subsidies in many countries, including Malaysia. The latest data

reveal that government spending on fossil fuel subsidies has cost Southeast Asia

$17 billion (International Energy Agency 2017). The fuel subsidy has been identified

as the primary cause of Malaysia’s ballooning fiscal deficit, which threatens to make

the country’s economic position unsustainable (Economic Planning Unit [EPU] 2010).

Besides straining the budget, as the fuel subsidy continually raises the issue of

fiscal balance (Anand et al. 2013), the subsidy boosts the demand for fossil fuels and

discourages energy efficiency (Liu and Li 2011), leading to negative environmental

impacts (Li, Shi, and Bin 2017) and fuel smuggling (Asian Development Bank 2016).

Also, the fuel subsidy, which was primarily created to help the poor, has benefited the

wealthy population more. The poorest 20% of the population get only 7% of

the subsidy’s benefit, while the wealthiest 20% receive a disproportionate 43%

(del Granado, Coady, and Gillingham 2012).

Rapid industrialization has led to the domination of fuel usage in the industrial

sector, which has made Malaysia the third-largest energy consumer in Southeast Asia

(International Energy Agency 2015). Therefore, when the managed-float mechanism

for fuel prices comes into effect, fuel usage for all economic activities would be

based on market rates, which means those economic activities would be exposed to

high fluctuating costs. High-cost burdens those domestic sectors that are dependent

upon fuel and other energy products in their production processes. Thus, sectors

that are characterized by large shares of fuel-based inputs would be significantly

affected. Subsequently, decision-making concerning production activities will be

influenced too.

The impact of oil price hikes on commercial and industrial users includes

increased production costs (Middle East Economic Survey 2016). For example, in

Saudi Arabia, the Saudi Cement Company expected annual production costs to rise by

$18 million due to the removal of fuel subsidies (Trade Arabia 2015). Energy price

increases caused by subsidy reforms highlight that cost increases occur both directly



and indirectly (Rentschler, Kornejew, and Bazilian 2017). Notably, energy-intensive

manufacturing firms experience substantial changes to their cost structures, with

adverse implications for profitability (Bazilian and Onyeji 2012).

Such implications can have knock-on effects on economic activity, employment,

and eventually on households (Kilian 2008). Although producers can pass on the

cost to their consumers, the output is reduced due to increased production costs.

Furthermore, employment decisions are impacted as high production costs lead to

reduced desirability for producing more output, which lowers the demand for

employment. Rentschler, Kornejew, and Bazilian (2017) highlight that cost increases

(both direct and indirect) do not necessarily reflect competitiveness losses since firms

have various ways to mitigate and pass on price shocks. Initially, high fuel prices are

often associated with low output and, in turn, low employment. Moreover, the high

costs of goods and services discourage spending by households and the government,

leading to lower economic growth.

Malaysia’s fiscal capacity runs at an unsustainable level as subsidies to maintain

low fuel prices constitute a huge portion of the government’s annual budget. The

country must run a fiscal deficit when excessive spending on subsidies has to support

rising fuel prices. When the crude oil price hovers between $65 and $85 per barrel

under normal circumstances, the estimated fuel subsidy is between 9 billion ringgit

(RM) and RM11 billion per annum (EPU 2008). When the crude oil price peaked in

2008 at more than $100 per barrel, the Malaysian government’s total fuel subsidy was

RM15 billion (EPU 2008).

Figure 1 shows that the petroleum subsidy comprised a large percentage of

public spending in Malaysia from 2004 to 2010, ranging from a low of 10.1% to a

high of 26.4% in 2008. It is more than the combined total spending on agriculture and

rural development, health, and housing, which are all critical to the country’s

development. The fiscal deficit, which was about 2.7% in 2007, climbed rapidly to

7.0% in 2009 as a result of the global oil price spike and its impact on the cost of fuel

subsidies (EPU 2010). The large substitution effect of the petroleum subsidy can be

seen by comparing it to other sectors. In other words, it indicates that a potentially high

amount of savings from a cut in the petroleum subsidy can be utilized for other policy

priorities that can have potentially more benefit for the Malaysian people.

Subsidy reforms and their impact have been thoroughly examined in developed

and developing countries (Clements et al. 2014). These studies have emphasized

welfare effects by looking at the distributional impact of fuel subsidy reform on

households. These investigations have been conducted in developing countries

(del Granado, Coady, and Gillingham 2012) such as Indonesia (Yusuf and

IMPACTS OF FUEL SUBSIDY RATIONALIZATION 317


Resosudarmo 2008) and Malaysia (Solaymani and Kari 2014; Li, Shi, and Bin 2017).

Even though prior studies have investigated the potential adverse effects of fuel

subsidy reform on households, research on the energy industry’s impact is still

required. Hence, this study aims to analyze the impact of fuel subsidy removal on the

energy industry, specifically on sectoral output and employment (according to skill and

occupation), by considering macroeconomic performance.

This study contributes to the current literature based on three perspectives. First,

this study contributes via a comprehensive examination of the fuel subsidy removal’s

impact on the sectoral output and employment by classifying fuel into petrol

(gasoline), diesel, and other fuels. Next, by extending the impact on labor into

occupational categories and skill levels, this study also observes the impact of fuel

subsidy removal on aggregate employment according to occupations and skills.

The final aspect pertains to the model’s contribution via incorporating a detailed

description of the labor market’s structure in the economy through nine occupational

categories and three skill levels.

The remainder of this paper is structured as follows. Section II explains the

database and labor classifications applied in the model. Section III describes the

methodology, model validation, benchmark scenario, and simulations. Section IV

presents the results and discussion, and Section V concludes.

Figure 1. Petroleum Subsidy Expenditure as a Share of Total Government Spendingin Malaysia

Source: Ministry of Finance, Government of Malaysia. 2011. Malaysia Economic Reports 2004/05–2010/11.Putrajaya.



II. Data

This study uses the 2005 Malaysia Input–Output (IO) Table consisting of 120

industries and commodities (Department of Statistics Malaysia 2010). We disaggregated

the subsector of petroleum refinery into three types of fuel commodities: petrol, diesel,

and other fuel products (liquified petroleum gas, coke, and gas), thus bringing the total to

122 industries (see Table A1 of the Appendix). The disaggregation is based on the

Malaysian Standard Industrial Classification (Department of Statistics Malaysia 2000),

while the fuel share is based on the National Energy Balance (Energy Commission

2005). The disaggregation is in line with the subsidy provided according to fuel products,

Table 1. Database of the Computable General Equilibrium Model

Sector

Producer Investor Household Export Government

Extension Matrix 1–122 1 1 1 1

Basic flows ofintermediateinputs,domestic

122� 122 V1dom V2dom V3dom V4dom V5dom

Basic flows ofintermediateinputs, import

122� 122 V1imp V2imp V3imp V4imp V5imp

Taxes 1� 122 V1tax V2tax V3tax V4tax V5tax

Labor Occupationalcategory

9� 122 V1lab

Skill levels 3� 122

Capital 1 V1cap

Land 1 V1lnd

Other costs 1 V1oct

V1dom ¼ domestic intermediate goods, V2dom ¼ domestic investment, V3dom ¼ household domesticconsumption, V4dom ¼ domestic production on export, V5dom ¼ domestic government expenditure,V1imp ¼ imported intermediate good, V2imp ¼ investment on imported capital, V3imp ¼ householdconsumption on import, V4imp ¼ imported goods, V5imp ¼ government expenditure on import,V1tax ¼ taxes on producer, V2tax ¼ taxes on investor, V3tax ¼ taxes on household, V4tax ¼ taxes onexport, V5tax ¼ taxes on government expenditure, V1lab ¼ labor, V1cap ¼ capital, V1lnd ¼ land, andV1oct ¼ other costs.Source: Noorasiah Sulaiman and Mukaramah Harun. 2015. “Valuing the Impact of RationalizingMalaysia’s Fuel Subsidies on its Macroeconomic Performance.” In Economy-Wide Analysis of ClimateChange in Southeast Asia: Impact, Mitigation and Trade-Off, edited by A.A. Yusuf, Arvin Hermanto, A.R.Irlan, K. Ahmad, N. Sulaiman, and M. Harun, pp. 146–201. Los Banos: WorldFish and Economy andEnvironment Program for Southeast Asia.



i.e., petrol for passenger vehicles, and diesel and other fuel products for non-passenger

vehicles.

Table 1 presents the database that comprises production, primary factors, and

final demand components, as well as detailed labor variables. The rows in the matrix

represent the supply side. In particular, the table presents linkages among economic

activities such as the relationships among production value, production cost, selling

price, market clearing conditions for the commodity, primary inputs, other macro

indicators, and the price index.

This study develops and includes an extension of the employment matrix,

according to occupations and skills in the Malaysia IO Table, to analyze the impact of

fuel subsidy removal on employment across subsectors. Employment data are obtained

from the Labour Force Survey 2005 (Department of Statistics Malaysia 2006).

As shown in Table 2, labor is classified into nine occupational categories and skill

levels based on the Malaysian Standard Classifications of Occupation (Ministry of

Human Resources 2008) and Sulaiman and Ismail (2019).

III. Methodology

The computable general equilibrium (CGE) model used in this study is based on

the generic (ORANI-G) CGE model developed by Horridge (2006). The programming

utilizes the General Equilibrium Modelling Package to analyze the impact of subsidy

removal on sectoral output and employment according to occupation and skill. In this

study, a static CGE model is developed, which includes the assumption that the

Table 2. Classification of Labor

Skill Level Occupational Category Job Description

High 1 Senior officials and managers2 Professionals

3 Technicians and associate professionals

Medium 4 Clerical workers5 Service workers

6 Skilled agriculture workers

7 Craft workers

8 Plant and machine operators

Low 9 Elementary workers

Source: Ministry of Human Resources, Government of Malaysia. 2008. MalaysiaStandard Classification of Occupations, 2nd Edition. Putrajaya.



subsidy will be returned to the economy, so that aggregate employment remains

constant. Thus, the analysis focuses on the structural implications of subsidy removal

on sectoral output and employment growth.

The model is calibrated according to subsidy removal by fuel types, whereas

employment is based on occupation and skill. The inclusion of occupation and skill

into the Malaysia IO Table, and thus into the CGE model, allows the in-depth

realization of the impact of fuel subsidy removal on output and employment growth.

In analyzing the impact of the labor market in the CGE model, further assumptions are

made (Meagher, Adams, and Horridge 2000). The demand side of the labor market

assumes that labor by occupational type is demanded by industry according to constant

elasticity of substitution (CES) functions. Meanwhile, the supply side assumes that

labor is supplied according to the constant elasticity of transformation functions.

Thus, both labor demand and supply are supposed to be in equilibrium. Similar labor

skills are assumed substitutable among industries, and relative wage rates are assumed

to adjust to clear labor markets by occupation.

Since a fuel price increase is an endogenous variable, it would not directly affect

output and employment among industries. The subsidy’s removal would indirectly

affect production. As this study analyzes subsidy removal on fuel commodities, removal

of the subsidy would increase cost in terms of transportation due to diesel and other fuel

products used for non-passenger vehicles. Output and employment would remain high

in moderately competitive industries, with both falling in the least competitive ones.

The cost saving from subsidy removal is assumed as aggregate tax revenue from all

indirect taxes. The cost-saving return to households via cash transfer alleviates the

increased cost of living, specifically among low-income households. Household spending

would increase as a result of the cash transfer, resulting in higher aggregate demand.

Therefore, growth in aggregate demand would impact output and employment

among subsectors of the economy. The price elasticities of demand are expected not to

respond to fuel demand in the short run because the endogenous shock of the fuel price

increase has a smaller impact on demand for fuel products from economic subsectors

and households.

Nevertheless, demand for fuel can be elastic, as all inputs can be changed in the

long run. In developing countries, cash transfer programs are now more prevalent,

both as long-term poverty alleviation measures and to lessen the adverse effects of

certain types of reforms that may impact the poor. It has been successful to use these

transfers to reform fuel subsidies. Yusuf (2018) found that cash transfers funded by

cost savings in fuel subsidies minimize disparity in Indonesia more than other

approaches.



Figure 2 presents the production function of the model, based on a nested

production system. With the assumption of the CES functions, industries are supposed

to choose a mix of inputs that minimize production costs for their output level when

different CES production technologies are assumed. The industry’s output is classified

into domestic and export. Both require intermediate goods, primary factors, and other

costs in their production process.

Fuel subsidies are implemented into the model through net indirect taxes of fuel

products (petrol, diesel, and other fuels). Removal of the subsidy would result in

higher fuel prices that impact other sectors of the economy, especially those using fuel

more extensively. For household consumption, the model adopts the linear expenditure

system demand function derived from the maximization of a Klein–Rubin (1947)

utility function, which distinguishes between necessary and luxury demand goods

(Pollack and Wales 1992).

The next level of the primary-factor branch of the production nest consists of a

CES combination of labor, capital, and land. At the lowest level, the industry- and

occupation-specific labor demand are combined using a CES production technology to

Figure 2. Nested Production Structure

CES = constant elasticity of substitution.Source: Authors’ illustration.



obtain the occupation-aggregated labor input. Labor by occupational categories is

represented by occupation 1 to occupation 9 and then classified into three skill levels.

The labor market extension in the CGE modeling techniques relies on the

strength of its capacity to consider available information on the structural linkages

between industries, occupations, and skills. Therefore, the model has a significant

feature of disaggregated employment into occupational categories and skill levels to

examine the impact of distributed labor demand and the output production across

industries by three types of fuel products.

A. Impacts of Fuel Subsidy Removal

The demand function is contingent on the impacts of energy disaggregated into

diesel, petrol, and other fuel. Specified as follows, the elasticity of demand in a

constant estimation with a range of parameters is

c ¼ b": ð1ÞTherefore, we can express that as

Δc ¼ "(B1 � B0): ð2ÞMeanwhile, the impact of household utilization is determined by the equation

Δc ¼ C1 � C0, ð3Þwhere Δc is the change in energy consumption when the fuel price increases; " is the

long-run price elasticity of energy demand; B0 and C0 indicate the price of energy and

its consumption before the subsidy removal policy, respectively. The B1 and C1 refer

to the price of energy and its consumption after the subsidy removal policy.

B. Employment Impacts

It is assumed that all production factors are variable. Producers can rent capital

and land in the agriculture sector. Intermediate inputs of capital and land are assumed

fixed between industries. Production specifications for the model are nested. Demand

for inputs for each industry ( j ) is determined by the cost-minimizing function subject

to Leontief ’s production function in equation (4). Inputs in the production structure

are composite commodities (i) (hþ 1, s), intermediate inputs, and other costs (hþ 2).



Therefore, the production function is

LeontiefI 1iyT 1iy

( )

¼ T 1j Aj, ð4Þ

where

I 1iy is an effective input for good i for current production in industry j,

Aj is the level of activity for industry j, and

T 1y and T 1

j are the coefficients for technological change.

Based on equation (5), composite commodity I 1iy is used in every industry with a

combination of export and import goods based on the CES technology. Primary input

I 1(hþ1, x)j also includes a combination of labor, capital, and land integrated based on the

CES technology. The CES technology refers to the combination of exported and

imported commodities, which explains that these two sources are imperfect substitutes

for input demand that vary according to relative price changes:

I 1iy ¼ CESx¼1, 2, 3I 1(hþ1, x)j

T 1(hþ1, x)j

( )

i ¼ 1, . . . , h (h is differential in production),

j ¼ 1, . . . , h (h is differential in industries):ð5Þ

The CES specification allows for an inter-labor replacement for the primary factors’

composite and intermediate inputs, depending on price changes relative to skill level.

In the input demand function, the production of each industry’s output level and the

input price (excluding the composite labor demand) are exogenous factors.

Consequently, minimizing costs can be solved with the input functions in the form

of percentage change by choosing the following equation:

I 1ij , I1(ix)j, I

1(hþ1, x)j, I

1(hþ1, 1, n)j: ð6Þ

To minimize,

Xh

i¼1

X2

x¼1

P1(ix)jI

1(ix)j þ

Xn

n¼1

P1(hþ1, 1, n)jI

1(hþ1, 1, n)j

þX3

x¼2

P1(hþ1, 1, x)jI

1(hþ1, 1, x)j þ P1

hþ2I1hþ2, j: (7)

I 1ij is the demand for effective intermediaries and primary inputs for industry j;

I 1(ix)j is the demand for intermediate inputs for import and export in industry j;

I 1(hþ1, x)j is the demand for primary factor input for industry j, including the capital,

labor, and L;

I 1(hþ1, 1, n)j is the demand for labor for types of skill levels for industry j;



P1(ix)j is the price for the intermediate input of import and export;

P1(hþ1, 1, n)j is the price for the primary factor s for industry j; and

P1hþ2 is the price for other costs.

The function of demand for primary factors in equation (8) exhibits an increase

in industry j, following the average cost of labor, capital, and land, and causing

replacement factors:

I 1(hþ1, s)j ¼ yj � �1(hþ1, s)j p1

(hþ1, s) �X

x

F 1(hþ1, s)jp

1(hþ1, s)j

!

þ c1j þ c1hþ1, j

þ c1(hþ1, s)j �X

x

F 1(hþ1, s)c

1(hþ1, s)j): (8)

F 1(hþ1, s) is the share of labor, capital, and land for the payment of primary factor inputs

in industry j.

Equation (9) is a demand function of labor according to skill level. It also

includes changes in technical variables. In the absence of technological changes, an

increase in labor price for specific skills relative to other skilled labor costs will

increase the consumption of such labor more slowly than other labor:

I 1(hþ1, 1, r)j ¼ x(hþ1, 1)j � δ1(hþ1, 1, r)j p1(hþ1, 1, r)j �

X

r

F 1(hþ1, 1, r)jp

1(hþ1, 1, r)j

!

þ c1(hþ1, 1, r)j � �1(hþ1, 1)j c1(hþ1, 1, r)j �

X

r

F 1(hþ1, 1, r)jp

1(hþ1, 1, r)j

!

,

j ¼ 1, . . . , h: (9)

F 1(hþ1, r) is the share of costs for labor (at a skill or occupational level).

This can be explained as

F 1(hþ1, 1, r)j ¼

P1(hþ1, 1, r)jI

1(hþ1, 1, r)j

Pnr¼1 P

1(hþ1, 1, r)jI

1(hþ1, 1, r)j

: ð10Þ

δij denotes the elasticity of substitution between imported goods. Meanwhile, domestic

goods are the input for the production in industry j. Therefore, this specifies that if the

cost of any one source increases, it will cause a relative decrease in the input demand

of that particular source.

C. The Closure

The conditions give the closure of the model on the (i) government balance and

(ii) saving–investment balance. The government balance follows the condition that



saving is endogenously determined as the difference between the government’s

disposable income and total expenditure. The saving–investment balance condition

requires that investment is saving-driven, with gross fixed investment that derives

from the sum of aggregate saving. The real exchange rate can be flexible, while gross

saving and interest rates are fixed in nominal terms.

Government consumption and public transfers to households are fixed.

Furthermore, investment, the nominal exchange rate, government saving, and real

investment expenditure are also considered endogenous. Labor at three skill levels

and nine occupational types are mobile across subsectors, and capital supply is

exogenously fixed. Endogenous factor prices clear the corresponding labor and capital

markets, so there is no unemployment model. The model results must be considered

short term since the model is static with a fixed total factor supply.

D. Parameter and Model Validation

This study sets out the parameters of the model. The parameters are Armington

elasticities between domestic and imported commodities for different CES functions

for intermediate-use investment demand and household demand obtained from the

Global Trade Analysis Project (2008). The parameters also include commodity-

specific export elasticities, constant elasticity of transformation between domestic and

export supply commodities from the ORANI-G model (Horridge 2006), and the

elasticities of substitution between labor types by skill level utilized from the literature

(Meagher, Adams, and Horridge 2000). Sensitivity analysis is done for each elasticity

parameter, and the results show that the models are not sensitive to the value for

different parameters.

The model’s calibration is accomplished by a validation test to verify database

construction, specifying the equations and closure of the model. The model establishes

a validation test for the implementation of the CGE model (Horridge 2006). First, this

study has performed the nominal and real homogeneity tests. The test considers the

system of equations, and economic agents respond to the changes according to the

relative prices, not the absolute price level. The results imply that if all exogenous

nominal variables of the model change, then all endogenous nominal variables will

also change, while real variables remain unchanged. Likewise, if all real exogenous

variables of the model change, then all real endogenous variables will change, while

nominal variables remain unchanged.

Second, the model should pass a conditional balance equal to gross domestic

product (GDP) from the income and expenditure sides. Thus, the percentage changes



in GDP are uniform. Third, the zero-profit condition test is performed with the total

cost of output equal to the total sale of commodities.

E. Benchmark Scenario and Simulations

The model simulation is outfitted with CES production and utility functions, with

indirect taxation affecting inputs, and consumption as the subsidy is removed through

the increase in prices. As an endogenous variable, the percentage change in prices can

be used to simulate the impacts of different policy scenarios in the CGE model

(Horridge 2006). Moreover, the endogenous variable of the price is used to determine

a subsidy removal on the fuel commodities.

For efficiency in fuel consumption and real cost for the industry, this study

developed a benchmark scenario used as a baseline for the model. The benchmark of

the scenario is based on the price increase for petrol and diesel from the Means of

Platts Singapore, which is a calculation of petrol prices done by a company based in

Singapore called Platts. The retail prices of petrol and diesel in Malaysia are

determined through the automatic price mechanism, ensuring that the difference

between retail and actual prices is borne by subsidies and sales tax exemptions

(EPU 2005).

In this study, we calculate the price changes in petrol and diesel as practiced in

Malaysia. As shown in Table A2 of the Appendix, the price increases in petrol and

diesel are considered as the base case and alternative scenarios. For petrol, the

subsidized rate is RM1.62 per liter. Without the subsidy, the actual cost would be

RM2.45 per liter. Hence, the government is bearing 83 sen: 59 sen in foregone taxes

and 24 sen in subsidies.1 On the other hand, the retail price is RM1.28 per liter for

diesel, but the actual cost is RM2.07 per liter. Thus, the subsidy is 59 sen per liter

with a foregone tax of 20 sen per liter. Besides, the unsubsidized price for other fuels

(e.g., liquified petroleum gas) is higher (i.e., RM2.39 per kilogram), but the retail price

is only RM1.45 per kilogram, and the subsidy is 94 sen (EPU 2005).

The level of price increase in fuel products affects government spending on

subsidies. Based on this situation, the fuel subsidy will be markedly higher, especially

when the fuel price per barrel has increased remarkably. The alternative scenario

considers the largest price increase to reflect the highest percentage when the subsidy

is removed. Table A2 of the Appendix shows that when the increase in fuel is larger

than $100 per barrel, the subsidized petrol and diesel are RM2.70 and RM2.58 per

1The Malaysian ringgit is divided into 100 sen.



liter, respectively, representing 30% and 40% of the subsidy removal. Table 3 presents

two scenarios that are examined in this study. In each scenario, three different

simulations are designed (i.e., for petrol, diesel, and other fuels). In the base case

scenario, the subsidy removal starts with a 10% increase in prices; the simulation of

fuel products can be seen in SIM1 (petrol), SIM2 (diesel), and SIM3 (other fuels). The

minimum 10% removal in subsidy per liter is chosen as the base model benchmark to

observe the general impact. It is represented by a minimum subsidy removal of 25 sen

per liter of fuel consumed for petrol, diesel, and other fuels.

Furthermore, it is crucial to examine the extent to which the policy instrument of

subsidy is fully removed to ensure a competitive market. Without the subsidies, the

prices for petrol and diesel are higher by 30% and 40%, respectively. Prices with those

subsidies correspond to the largest amount that the government commits for subsidies.

IV. Results and Discussion

The results on the impact of demand for labor are analyzed in two stages. First,

the macroeconomics scenario is discussed for the selection of macro variables. At the

macro level, by conducting each simulation, employment growth is determined

according to occupational categories and skill levels. At the sectoral level, labor is

Table 3. Subsidy Removal Simulations for Petrol, Diesel, and Other Fuels

Simulation Base Case Scenario Description

SIM1 f0tax_s(“C44aPetrol”) 25 sen per liter subsidy removal foris increased by 10% all users of petrol

SIM2 f0tax_s(“C44bDiesel”) 25 sen per liter subsidy removal foris increased by 10% all users of diesel

SIM3 f0tax_s(“C44cOthFuel”) 25 sen per liter subsidy removal foris increased by 10% all users of other fuel products

Simulation Alternative Scenario Description

SIM1 f0tax_s(“C44aPetrol”) RM1.32 per liter subsidyis increased by 30% removal for all users of petrol

SIM2 f0tax_s(“C44bDiesel”) RM1.70 per liter subsidy removalis increased by 40% for all users of diesel

SIM3 f0tax_s(“C44cOthFuel”) RM1.32 per liter subsidy removalis increased by 30% for all users of other fuel products

RM ¼ Malaysian ringgit, SIM ¼ simulation.Source: Authors’ calculations based on data from the Department of StatisticsMalaysia, Government of Malaysia. 2010. Malaysia Input–Output Tables 2005.Putrajaya.



aggregated into the employment rate to determine the impact of subsidy removal on

employment growth. Similarly, at the subsector or industrial level, subsidy removal is

analyzed based on the type offuel product to examine the impact on employment and output

growth. Price elasticities of demand for fuels at the sectoral level are also estimated.

A. Key Results

Table 4 presents the base case scenario and the alternative scenario of subsidy

removal for selected macro variables and employment by occupation type and skill

level. The impacts of each scenario can be seen in the aggregate employment level by

occupational category and skill level. The base case scenario assumes a price increase

of 10% for each fuel commodity. In contrast, the alternative scenario is based on fuel

price increases of 30% (petrol) and 40% (diesel and other fuels). In both scenarios, the

results for all simulations (SIM1: petrol, SIM2: diesel, SIM3: other fuel products)

indicate that the price has increased for all users.

As the table shows, the fuel subsidy’s reduction by 10% of the fuel price

increase positively impacts macroeconomic variables for GDP, exports, and imports.

GDP (in real terms) increases by 0.05% for SIM1 and SIM2, whereas for SIM3, it

increases by 0.04%. Similar results are obtained for the alternative scenario, where the

subsidy reductions are 30% (petrol) and 40% (diesel and other fuels) of the fuel price

increase. GDP increases by 0.139% (SIM1), 0.185% (SIM2), and 0.119% (SIM3).

Likewise, the increase in exports rises from 0.244% with a 10% fuel subsidy reduction

to 0.721% with the fuel subsidy’s full removal (SIM1), from 0.258% to 1.007%

(SIM2), and from 0.205% to 0.608% (SIM3). The gain in imports also increases from

0.243% with a 10% fuel subsidy reduction to 0.73% with the fuel subsidy’s full

removal (SIM1), from 0.256% to 1.030% (SIM2), and from 0.204% to 0.613%

(SIM3). A similar study by Solaymani and Kari (2014) also found that removing

energy subsidies increases real GDP. However, total exports and imports decline,

which is the opposite of the results obtained from this study.

On the other hand, household consumption and government expenditure are

impacted negatively under the base case and the alternative scenarios. Household

consumption under a 10% fuel subsidy reduction falls by 0.035%, 0.036%, and

0.043% for SIM1, SIM2, and SIM3, respectively. Similarly, government expenditure

drops by 0.132%, 0.139%, and 0.111%, respectively. Nevertheless, both variables

show a larger contraction for the alternative scenario (full subsidy removal). The larger

decline in household consumption indicates that the cash transfer program does not

fully compensate for a higher cost of living due to the rising prices of goods and

services brought about by the increase in fuel prices.



Table

4.Key

ResultsforSelectedMacro

Variables(%

)

BaseCaseScenario

AlternativeScenario

Macroecon

omic

Variable

SIM

1SIM

2SIM

3SIM

1SIM

2SIM

3

GDP(real)

0.05

10.05

30.04

30.13

90.18

50.119

Hou

seho

ldexpend

iture

�0.035

�0.036

�0.029

�0.103

�0.143

�0.088

Gov

ernm

entexpend

iture

�0.132

�0.139

�0.111

�0.387

�0.537

�0.327

Exp

orts

0.24

40.25

80.20

50.72

11.00

70.60

8Im

ports

0.24

30.25

60.20

40.73

01.03

00.61

3

Skill

Level

Employm

entbyOccupation

High

1.Seniorofficialsandmanagers

0.17

00.18

00.14

30.50

60.71

00.42

72.

Professionals

0.16

40.17

30.13

80.48

70.68

20.41

0

3.Techn

icians

andassociate

profession

als

0.12

00.12

70.10

10.35

60.49

70.30

0

Total

0.45

40.48

00.38

21.34

91.88

91.13

4

Medium

4.Clericalworkers

0.15

80.16

70.13

30.46

90.65

70.39

65.

Service

workers

0.18

10.19

10.15

20.53

70.75

20.45

2

6.Skilledagricultu

reworkers

0.15

20.16

10.12

80.45

30.63

40.38

1

7.Craftworkers

0.12

60.13

30.10

60.37

30.52

10.31

4

8.Plant

andmachine

operators

0.119

0.12

60.10

00.35

20.49

20.29

7Total

0.73

60.77

80.61

92.18

43.05

61.52

6

Low

9.Elementary

workers

0.15

00.15

80.12

60.44

40.62

10.37

4

GDP¼

grossdo

mestic

prod

uct,SIM

¼simulation.

Sou

rce:Autho

rs’calculations

basedon

datafrom

theDepartm

entof

StatisticsMalaysia,Gov

ernm

entof

Malaysia.20

10.M

alaysia

Inpu

t–Outpu

tTables

2005

.Putrajaya.



All results rely upon the assumption of providing extra revenue to households

through cash transfers to lower the cost of living. The results highlight that Malaysia’s

budget is distributing the cash transfers otherwise to other production factors.

A positive sign for real GDP implies that the fuel subsidy’s removal would reduce the

government’s budget and positively affect productive sectors. Specifically, the removal

of the diesel subsidy has a larger impact, as shown by SIM2, compared to the removal

of the subsidy for petrol (SIM1) and other fuels (SIM3).

Table 4 also shows that the aggregate demand for labor (the employment aspect)

has recorded positive signs by occupational categories and skill levels. In the base case

scenario, employment expansion for all types ranges from 0.100% to 0.191%; under

the alternative scenario, it ranges from 0.297% to 0.752%. Specifically, diesel subsidy

removal (SIM2) would significantly impact aggregate employment expansion for all

skill levels and occupational categories. Furthermore, the service worker category

experiences significant employment expansion.

B. Output and Employment Effects

Rationalizing fuel subsidies would have a significant impact on industries. In the

second stage of focusing on output and employment growth by type of industry, capital

growth and technical change are constant when production costs increase. With a

larger subsidy removal for fuel commodities, the additional cost to industry will

directly affect output and labor usage. On the other hand, an industry with a relatively

high rate of return will attract investment and enjoy a relatively high capital growth

rate. As a result, a relatively low percentage of employment growth is achieved for a

given rate of output growth. Also, for labor-intensive industries, employment

expansion will increase labor productivity.

Since this study concentrates on the mitigation scenario to see the maximum

impact of subsidy removal, the alternative scenario is given priority in the discussions.

Hence, the relative growth (both positive and negative) in the base case scenario

would always be a benchmark to examine the mitigation’s impact in general. Thus, the

relative growth rates in the alternative scenario are quite similar to those in the base

case scenario. For all simulations, subsidy removal shows a contraction in output and

employment for most of Malaysia’s economic subsectors. The output and employment

contractions range from 0.005% to 0.582% and from 0.011% to 3.951% for the base

case scenario. Moreover, the alternative scenario registers larger contractions from

0.025% to 2.425% for output and from 0.051% to 14.484% for employment.

Figure 3 presents the output and employment effects for each scenario according

to subsectors based on the findings for SIM1, SIM2, and SIM3. Within the agriculture



Figure 3. Agriculture Sector: Output and Employment

Source: Authors’ calculations based on data from the Department of Statistics Malaysia, Government of Malaysia.2010. Malaysia Input–Output Tables 2005. Putrajaya.



sector, the oil palm subsector shows the largest contraction in output and a significant

reduction in employment for all the simulations in the base case scenario. Similar

results are obtained for the alternative scenario. All simulations have a larger impact

on output and employment, with the largest contraction belonging to SIM2 (petrol).

It shows a contraction of 1.141% for output and 2.505% for employment. However, the

forestry and logging subsector have the largest contractions in employment at 0.678%

(SIM1), 0.717% (SIM2), and 0.569% (SIM3). Thus, despite the adverse impact of

subsidy removal, there are also some positive impacts on output and employment.

A positive impact on output and employment can be seen in subsectors such as

vegetables, food crops, other agriculture, other livestock, paddy, fruits, poultry, and

rubber. For all simulations in the alternative scenario, the growth in output ranges from

0.029% to 1.721%, while employment growth ranges from 0.430% to 4.062%.

The findings indicate that employment growth is relatively larger than the output

growth for all simulations in the alternative scenario.

In Figure 4, all subsectors of the mining sector suffer from the fuel subsidy’s

removal. It highlights the marginal impact on output and employment ranging from

�0.635% to 0.617% and from �1.070% to 1.056%, respectively. The stone, clay, and

Figure 4. Mining Sector: Output and Employment




sand quarrying subsector is the most affected, followed by the metal ore and other

mining and quarrying subsectors. The alternative scenario shows a reduction in output

and employment from 0.012% to 1.282% and from 0.555% to 3.620%, respectively.

The construction sector’s residential and special trade work subsectors

experienced a smaller expansion in output and employment (Figure 5). In contrast,

the nonresidential and civil engineering subsector experienced a contraction in output

and employment, even though it was relatively small.

Figure 6 exhibits the impacts on output and employment among subsectors of

the manufacturing sector. For the alternative scenario, SIM1 shows output and

employment contractions of 1.732% and 10.721%, respectively. For SIM2, the

contractions are 2.425% and 14.484%, respectively; while for SIM3, they are 1.459%

and 9.501%, respectively. Under SIM1, the basic chemicals products; cement, lime,

and plaster; and sheet glass and glass products subsectors each have a reduction in

output ranging from 1.070% to 1.730%. SIM2 generates relatively larger output

contractions in these subsectors ranging from 1.144% to 2.425%, while SIM3’s

Figure 5. Construction Sector: Output and Employment




Figure 6. Manufacturing Sector: Output and Employment




contractions range from 1.040% to 1.459%. In terms of employment, the contractions

range from 2.596% to 10.721% (SIM1), from 3.647% to 14.484% (SIM2), and from

2.184% to 9.501% (SIM3).

Expansions in output for all simulations of the alternative scenario were

registered for the following subsectors: semiconductor devices, tubes, and circuit

boards; TVs, radio receivers, transmitters, and associated goods; domestic appliances;

industrial machinery, measuring, checking, and industrial process equipment; soap,

perfumes, cleaning, and toilet preparations; office, accounting, and computing

machinery; electric lamps and lighting equipment; electrical machinery and apparatus;

and ship, boat building, and bicycles. These industries are all associated with

multinational companies and/or foreign direct investment. The output expansion was

largest for SIM2, ranging from 1.918% to 5.973%, followed by SIM1 (from 1.363% to

4.231%) and SIM3 (from 1.146% to 3.551%). The employment growth was largest for

SIM1, ranging from 5.024% to 9.827%, followed by SIM2 (from 7.034% to 13.906%)

and SIM3 (from 4.232% to 8.242%).

Fuel products show a significant contraction, both in output and employment,

particularly in the full subsidy removal (i.e., alternative) scenario. For both the base

case and full subsidy removal scenarios, the subsectors of petrol, diesel, and other fuel

products suffer large contractions in output and employment in each simulation.

Observing the full subsidy removal scenario, for SIM1, these industries experience

output reductions ranging from 0.856% to 1.177%, while for employment, the decline

ranges from 7.857% to 10.721%. SIM2 shows the largest effects, with contractions

ranging from 1.208% to 1.604% for output and from 11.002% to 14.484% for

employment. A similar result is obtained for SIM3, with output and employment

reductions ranging from 0.719% to 1.040% and from 6.618% to 9.501%, respectively.

Overall, these results imply that subsidy removal increases the price of fuels so that the

cost of production increases, indicating a decline in output and employment in fuel

product industries and among manufacturing subsectors.

Finally, for the service sector, subsectors closely related to petroleum refining

exhibit a larger impact on output and employment for all simulations in the alternative

scenario (Figure 7). The subsectors include other private services, restaurants, air

transport, land transport, port and airport operation, water transport, electricity and gas,

and waterworks. For those subsectors, output and employment decline under SIM1

from 0.323% to 1.585%, from 0.457% to 2.228% (SIM2), and from 0.271% to

1.333% (SIM3). Large declines are recorded in the output of other private services,

with contractions of 1.585% (SIM1), 2.228% (SIM2), and 1.333% (SIM3). Similarly,

for restaurants, the reductions in output are 1.277% (SIM1), 1.799% (SIM2), and



Figure 7. Service Sector: Output and Employment (%)




1.333% (SIM3). Among the transportation sector, air transport and land transport are

the most affected subsectors in output reduction for all alternative scenarios in

this study.

Furthermore, all simulation results register even larger contractions in

employment. The subsectors of electricity and gas and waterworks register the largest

reductions in employment. For electricity and gas, the declines are 3.910% (SIM1),

5.472% (SIM2), and 3.297% (SIM3); while for waterworks, the reductions are 3.065%

(SIM1), 4.323% (SIM2), and 2.576% (SIM3) (Figure 7). The results are similar for the

subsectors of water transport, other public services, restaurant, air transport, real estate,

port and airport operation services, land transport and highways, and bridge and tunnel

operations, with each subsector experiencing a significant contraction in employment

for all simulations.

From the results obtained, a few inferences are formed. Based on the scenario

analysis, first, the findings reveal that SIM2 (diesel) has a larger impact on the output

and employment contraction than SIM1 and SIM3 for both the base case and

alternative scenarios. It is supported by the fact that the government-borne subsidy for

diesel is substantially larger than for petrol and other fuels. Specific subsectors,

especially those under the manufacturing sector, are affected the most when the

subsidy is removed. Notwithstanding, fuel subsidy reforms significantly influenced

sectoral output through increased production costs due to an increase in the prices of

intermediate inputs (Rentschler, Kornejew, and Bazilian 2017). Furthermore, the larger

contribution in both domestic and imported inputs shows that intermediate input is the

major component of total factor productivity growth for the manufacturing sector

(Sulaiman 2012). Also, the manufacturing sector is supported by upstreaming (as

consumers) and down-streaming industry (as suppliers) linkages (Sulaiman and Fauzi

2017), implying that those manufacturing subsectors deal with the transportation of

intermediate inputs and finished products from the supplier to the consumers.

Second, in general, both scenarios (base and alternative) indicated that all sectors

experience either a contraction or expansion in output and employment, particularly

both are larger for the manufacturing sector. A contraction in output usually reduces

the need for employment, and vice versa in the case of output and employment

expansion. The subsectors most influenced by multinational firms’ production

experience an increase in output and employment. As mentioned, the semiconductor

devices; tubes, circuit boards, TVs, radio receivers, and transmitters associated goods;

and measuring, checking, and industrial process equipment subsectors all have

substantial ties to foreign producers (i.e., multinational corporations). On the other

hand, output and employment contractions are experienced by local producers. Oil and



fats; clay and ceramics; sheet glass and glass products; basic chemicals; and cement,

lime, and plaster are examples of locally produced goods. These findings are

corroborated by Sulaiman, Rashid, and Hamid (2012), who revealed that multinational

corporations are more efficient in utilizing both domestic and imported inputs than

local manufacturers.

Third, this study found a larger contraction in output for subsectors closely

related to the fuel sector, both in the manufacturing and service sectors. The findings

are comparable to that of an Indonesian study that reached the same conclusion (Yusuf

and Resosudarmo 2008). Furthermore, the study observed that fuel subsidy removal

tends to increase the price of industrial outputs that are highly dependent on fuel such

as in the transportation, energy, fishery, and industrial sectors. Increased production

costs result from rising oil prices for commercial and industrial customers (Middle

East Economic Survey 2016). Increases in energy prices from subsidy reforms result

in direct and indirect cost increases (Rentschler et al. 2017). Thus, subsectors that used

more energy, in particular, would have a significant impact on their cost structures

(Bazilian and Onyeji 2012).

Fourth, the alternative scenario shows the contraction in the employment rate is

greater than that in output. This finding is rational because firms will react by not

hiring new workers rather than reducing their production units. Thus, to cover the cost

of a price increase due to a subsidy removal, firms would prefer to minimize labor

compared to reducing the output produced, implying that the decline in the labor used

would adversely impact the output. Similarly, expansion in employment growth is

larger than the output growth even though it is not as big as the contraction, resulting

in a moderate impact on the distribution of employment across industries.

Such consequences may have repercussions for economic activity, jobs, and,

ultimately, households (Kilian 2008). Even though producers can pass on the cost

to their customers, the industry’s output is being reduced as production costs rise

(Harun et al. 2018). Therefore, employment decisions must be carefully considered

because high production costs reduce the desire to produce more output, depressing

the employment rate. According to Rentschler, Kornejew, and Bazilian (2017), cost

increases (both direct and indirect) may not indicate competitiveness losses because

firms have a variety of techniques to manage and pass on price shocks. Initially, high

fuel prices are often associated with lower output and, as a result, lower employment.

Furthermore, high prices for goods and services deter household and government

spending, resulting in slower economic growth.

Finally, based on the analysis presented, it is notable that when the revenues from

subsidy removal policies are channeled back into the economy via payments to



households, the resulting increase in demand will stimulate the economy. In Malaysia,

studies such as Solaymani and Kari (2014) and Loo and Harun (2020) identified the

harmful effects that would come from the implementation of the fuel subsidy reform

and urged for mitigating measures. The study found that the integration of a transfer of

government income to rural households would increase pro-poor growth and reduce

the negative impacts on all households’ real incomes with a slight improvement.

Specifically, Loo and Harun (2020) emphasized that (i) cash transfers are needed to

cope with the underlying high price resulting from the high fuel consumption price,

and (ii) the vulnerable—primarily low-income households and the poor—are the ones

hit hardest. Cash transfers are particularly preferable in the short term, where extended

time is needed for behavioral change, while developmental investments have

long-term benefits.

C. Sensitivity Analysis

Table 5 shows the price elasticity effects on fuel consumption estimates for the

whole economy, in which the elasticities range from �0.14 to �0.94. The ranges in

demand elasticity in key sectors are as follows: agriculture sector (�0.14 to �0.49),

mining sector (�0.43 to �0.44), manufacturing sector (�0.14 to �0.49), construction

sector (�0.46 to �0.63), and service sector (�0.35 to �0.94). It is not surprising that

demand in the agriculture and manufacturing sectors is relatively inelastic because

price elasticities could have been lower even for the long-run period, as revealed by

Bohi and Zimmerman (1994) and Graham and Glaister (2002).

Table 5. Elasticities of Demand for Fuel and Household Consumption

Price Elasticities ofDemand for Fuels

Expenditure Elasticitiesof Household Consumption

Sector Range Mean Value Range Mean Value

Agriculture �0.14 to �0.49 �0.27 0.32 to 1.15 0.63Mining �0.43 to �0.44 �0.44 1.01 to 1.04 1.03Construction �0.46 to �0.63 �0.58 0.28 to 0.63 1.34Manufacturing �0.14 to �0.49 �0.32 0.28 to 1.01 0.74Services �0.35 to �0.94 �0.56 0.47 to 0.87 1.28Total mean �0.40 0.89

Source: Authors’ calculations (using the Klein–Rubin utility function) based on datafrom the Department of Statistics Malaysia, Government of Malaysia. 2010. MalaysiaInput–Output Tables 2005. Putrajaya.



However, prior studies found that long-run price elasticities tend to be much

higher than in the short run due to the three crucial conclusions on the sensitivity of

price changes on fuel demand in the long run (Graham and Glaister 2002, Plante

2014). First, behavioral responses to cost changes occur over time, implying demand

has a larger impact than short-run elasticity. Second, the range of responses included

changes by vehicle type and location decisions. Third, policy options are more

comprehensive in the long run.

Under the linear expenditure (Klein–Rubin) system, the elasticity of marginal

utility of income is estimated, which shows that the long-run household expenditure

elasticities for the whole economy are relatively larger, ranging from 0.28 to 1.15. In

contrast, the mean value ranges from 0.63 to 1.34. This result is in line with prior

studies that have reported income elasticity of fuel consumption in the range of 0.6–1.6

(Graham and Glaister 2002).

The subsidy removal minimizes deadweight losses in the economy by

incorporating an efficient fiscal policy to give a complete result (Plante 2014).

Remarkably, the Government of Malaysia has distributed the savings from

rationalizing fuel subsidies in the form of a direct cash assistance program to

low-income groups to mitigate fuel price increase (e-BR1M 2018). The program

provides RM500 ($159) in cash aid to households with a monthly income of RM3,000

($953) or below. Even though the cash transfer program is not a popular practice, it

can save about 70% of the government’s current expenses on fuel subsidies and benefit

low-income groups, while the benefits of fuel subsidies are biased toward high-income

groups.

V. Conclusions

The simulations investigate the impact of fuel price increases on macroeconomic

variables, sectoral outputs, and employment. In addition, the disaggregation of fuel

commodities into petrol, diesel, and other fuel products has enabled examining the

impact of price increases for these fuel commodities separately according to industry.

The subsidy removal and subsequent increase in fuel prices will reduce selected

sectors’ activities in the Malaysian economy. Still, a reduction in the general sales tax

has an expansionary impact on the broader segments of the economy. This effect will

more than compensate for the contractionary impact, resulting in a positive net gain for

the overall economy. On the other hand, removing fuel subsidies can have immediate

negative effects on macroeconomic indicators like GDP through an increase in the cost



of production, a rise in the consumer price index, and a reduction in employment.

Nevertheless, the net macroeconomic impact is positive when revenue from the

subsidy removal is given back to the economy through cuts in the general sales tax.

It indicates that the pre-reform fuel pricing policy has distorted resource allocation.

Therefore, a departure from such a policy reform would be the right move toward

having a more efficient economy.

The contraction in output is larger in the manufacturing sector due to increased

diesel prices versus petrol and other fuels. It is not surprising because a larger

proportion of the government’s fuel subsidy expenditure goes to diesel rather than

petrol and other fuels. However, negative impacts, such as increased production costs

due to higher fuel prices, would be more than offset by the positive impacts of

government revenue reallocation. Most subsectors exhibit a drop in output due to an

increase in the cost borne by all users. Some producers have no choice concerning the

consumption of petroleum products; they will not reduce their use and thus, their cost

of production will increase, resulting in contractions in industry output. Nonetheless,

some industries experience a positive impact on output and employment.

Furthermore, the distributional impacts across different types of labor vary and

are much larger than impacts on output. Hence, the results show that the effect of

subsidy removal would be felt much more by unskilled labor, such as service workers

and elementary workers, compared to skilled labor. On the other hand, the most

uniform impact across workers implies that the distributional impact of the reform

would be neutral. Under the subsidy rationalization program, the Government of

Malaysia has planned for a gradual subsidy removal for other subsidized items, primarily

food (e.g., wheat flour, cooking oil, sugar, and rice). This is because it accounts for a

sizable portion of the operating expenditure in Malaysia’s national budget.

An estimated RM25 billion worth of subsidies is allocated in the budget

annually, depending on price changes. The sugar subsidy was eliminated on

26 October 2013, and the rice subsidy was completely removed on 1 November

2015. On 1 November 2016, the government announced that the cooking oil subsidy

would be phased out (Ministry of Finance Malaysia 2018). All of these actions are part

of the subsidy rationalization program, necessitating further analysis and future

research.

This study, therefore, suggests that the design of subsidy removal has to include

mitigating measures that address the well-being of the Malaysian people, especially

from the perspective of employment. Awell-designed subsidy rationalization program

would not only increase the acceptance level of the reform but underpin sustainable

economic development. Regarding the fuel subsidy rationalization that has been



focused on in this study, the integration of a cash transfer program would strengthen

economic performance by increasing real GDP growth, aggregate output, and

employment, and by improving the trade balance.

References

Anand, Rahul, David Coady, Mohammad Adil, Thakoor Vimal, and James P. Walsh. 2013.“The Fiscal and Welfare Impacts of Reforming Fuel Subsidies in India.” InternationalMonetary Fund Working Papers 13/128.

Asian Development Bank. 2016. “Fuel Smuggling: 12 Things to Know.” https://www.adb.org/news/features/fuel-smuggling-12-things-know (accessed 15 October 2019).

Bazilian, Morgan, and Ijeoma Onyeji. 2012. “Fossil Fuel Subsidy Removal and InadequatePublic Power Supply: Implications for Businesses.” Energy Policy 45 (C): 1–5.

Bohi, Douglas R., and Mary B. Zimmerman. 1994. “An Update on Econometrics Studies ofEnergy Demand.” Annual Review of Energy 9 (1): 105–54.

Clements, Benedict, David Coady, Stefania Fabrizio, Sanjeev Gupta, and Baoping Shang. 2014.“Energy Subsidies: How Large Are They and How Can they Be Reformed?” Economicsof Energy and Environmental Policy 3 (1): 1–17.

del Granado, Francisco J. A., David Coady, and Robert Gillingham. 2012. “The UnequalBenefits of Fuel Subsidies: A Review of Evidence for Developing Countries.” WorldDevelopment 40 (11): 2234–48.

Department of Statistics Malaysia, Government of Malaysia. 2000. Malaysian StandardIndustrial Classification. Putrajaya.

_____. 2006. Labour Force Survey 2005. Putrajaya._____. 2010. Malaysia Input–Output Tables 2005. Putrajaya.e-BR1M. 2018. Initial Portal. https://ebr1m.hasil.gov.my/ (accessed 23 January 2018).Economic Planning Unit, Prime Minister’s Office, Government of Malaysia. 2005.

Socioeconomic Indicators 2005. Putrajaya._____. 2008. Socioeconomic Indicators 2008. Putrajaya._____. 2010. Tenth Malaysia Plan 2011–2015. Putrajaya.Energy Commission, Ministry of Energy, Green Technology and Water, Government of

Malaysia. 2005. National Energy Balance. Putrajaya.Global Trade Analysis Project. 2008. “GTAP 7 Data Base.” Department of Agricultural

Economics, Purdue University. https://www.gtap.agecon.purdue.edu/databases/v7/(accessed 11 March 2016).

Graham, Danie J., and Stephen Glaister. 2002. “The Demand for Automobile Fuel: A Survey ofElasticities.” Journal of Transport Economics and Policy 36 (1): 1–25.

Harun, Mukaramah, Siti Hadijah Che Mat, Wan Roshidah Fadzim, Shazida Jan Mohd Khan,and Mohd Saifoul Zamzuri Noor. 2018. “The Effects of Fuel Subsidy Removal on InputCosts of Productions: Leontief Input–Output Price Model.” International Journal ofSupply Chain Management 7 (5): 529–34.



Horridge, Mark J. 2006. ORANI-G: A Generic Single-Country Computable GeneralEquilibrium Model. Centre for Policy Studies, Monash University.

International Energy Agency. 2015. Southeast Asia Energy Outlook 2015: World EnergyOutlook Special Report. Paris.

_____. 2017. Southeast Asia Energy Outlook 2017: World Energy Outlook 2017 SpecialReport. Paris.

Kilian, Lutz. 2008. “The Economic Effects of Energy Price Shocks.” Journal of EconomicLiterature 46 (4): 871–909.

Klein, Lawrence R., and H. Rubin. 1947. “A Constant Utility Index of the Cost of Living.”Review of Economic Studies 15 (2): 84–87.

Li, Yingzhu, Xunpeng Shi, and Su Bin. 2017. “Economic, Social, and Environmental Impactsof Fuel Subsidies: A Revisit of Malaysia.” Energy Policy 110 (C): 56–61.

Liu, Wei, and Hong Li. 2011. “Improving Energy Consumption Structure: A ComprehensiveAssessment of Fossil Energy Subsidies Reform in China.” Energy Policy 39 (7): 4134–43.

Loo, Sze Ying, and Mukaramah Harun. 2020. “The Assessment of Direct AgriculturalInvestment and Cash Transfer on Households in Malaysia: An Evidence of CompensationMechanism for Fuel Subsidy Removal.” International Journal of Business and Society21 (1): 300–12.

Meagher, Gerald A., Philip D. Adams, and Mark J. Horridge. 2000. “Applied GeneralEquilibrium Modeling and Labour Market Forecasting.” CoPS Working Paper No. IP-76.

Middle East Economic Survey. 2016. “Riyadh Cuts Fuel Subsidies, Petchem Producers Countthe Cost.” http://archives.mees.com/issues/1618/articles/53472.

Ministry of Finance, Government of Malaysia. 2011. Malaysia Economic Reports 2004/05–2010/11. Putrajaya.

_____. 2018. Estimated Federal Expenditure. http://www.treasury.gov.my/index.php/bajet/anggaran-perbelanjaan-persekutuan.html (accessed 22 January 2018).

Ministry of Human Resources, Government of Malaysia. 2008. Malaysia StandardClassification of Occupations, 2nd Edition. Putrajaya.

Plante, Michael. 2014. “The Long-Run Macroeconomic Impacts of Fuel Subsidies.” Journal ofDevelopment Economics 107 (C): 129–43.

Pollack, Robert A., and Terrence J. Wales. 1992. Demand System Specification and Estimation.New York: Oxford University Press.

Rentschler, Jun, Martin Kornejew, and Morgan Bazilian. 2017. “Fossil Fuel Subsidy Reformsand Their Impacts on Firms.” Energy Policy 108 (C): 617–23.

Solaymani, Saeed, and Fatimah Kari. 2014. “Impacts of Energy Subsidy Reform on theMalaysian Economy and Transportation Sector.” Energy Policy 70 (C): 115–25.

Sulaiman, Noorasiah. 2012. “An Input–Output Analysis of the Total Factor ProductivityGrowth of the Malaysian Manufacturing Sector, 1983–2005.” Jurnal Ekonomi Malaysia46 (1): 147–55.

Sulaiman, Noorasiah, and M. Ahmad Fauzi. 2017. “Identifying Key Sector of EmploymentPotential in Malaysia.” International Journal of Applied Business and EconomicResearch 15 (8): 577–93.



Sulaiman, Noorasiah, and Rahmah Ismail. 2019. “Re-Aligning the Needs of Manpower for theVision 2020: The Case of the Manufacturing Sector in Malaysia.” International Journalof Business and Management Science 9 (3): 407–23.

Sulaiman, Noorasiah, Zakariah Rashid, and Khalid Hamid. 2012. “Productivity Improvement inthe Utilization of Domestic and Imported Inputs in Resource and Non-Resource-BasedIndustries: 1983–2005.” International Journal of Management Studies 19 (1): 87–114.

Trade Arabia. 2015. “Energy Price Hike to Cost Saudi Cement $18M.” http://www.tradearabia.com/news/CONS_297717.html.

Yusuf, Arief. 2018. “The Direct and Indirect Effect of Cash Transfers: The Case of Indonesia.”International Journal of Social Economics 45 (5): 793–807.

Yusuf, Arief, and Budy Resosudarmo. 2008. “Mitigating Distributional Impact of Fuel PricingReform: The Indonesian Experience.” ASEAN Economic Bulletin 25 (1): 32–47.



Appendix

TableA1.

Subsectorsof

theMalay

sian

Econom

y

No.

Sectoran

dSubsector

Agriculture

1Paddy

29Other

food

processing

60Rub

berprod

ucts

2Foo

dcrop

s30

Animal

feed

61Plastic

prod

ucts

3Vegetables

31Wineandspirits

62Sheet

glassandglassprod

ucts

4Fruits

32Softdrinks

63Clayandceramics

5Rub

ber

33Tob

acco

prod

ucts

64Cem

ent,lim

e,andplaster

6Oilpalm

34Yarnandcloth

65Con

creteandotherno

n-metallic

minerals

7Flower

plants

35Finishing

oftextiles

66Iron

andsteelprod

ucts

8Other

agricultu

re36

Other

textiles

67Basic

precious

andno

n-ferrou

smetals

9Pou

ltryfarm

ing

37Wearing

apparel

68Castin

gof

metals

10Other

livestock

38Leather

indu

stries

69Structuralmetal

prod

ucts

11Forestryandlogg

ing

39Foo

twear

70Other

fabricated

metal

prod

ucts

12Fishing

40Saw

millingandplaningof

woo

d71

Indu

strial

machinery

Mining

41Veneersheets,plyw

ood,

etc.

72General-purpo

semachinery

13Crude

oilandnaturalgas

42Builders’

carpentryandjoinery

73Special

purposemachinery

14Metal

oremining

43Woo

denandcane

containers

74Dom

estic

appliances

15Stone,clay,andsand

quarrying

44Other

woo

dprod

ucts

75Office,accoun

ting,

compu

tingmachinery

16Other

miningandqu

arrying

45Paper,paperprod

ucts,andfurnitu

re76

Electricalmachinery

andapparatus

Con

struction

46Pub

lishing

77Other

electrical

machinery

17Residential

47Printing

78Insulatedwires

andcables

18Non

-residential

48Petrolprod

ucts

79Electriclamps

andlig

htingequipm

ent

19Civilengineering

49Dieselprod

ucts

80Sem

i-cond

uctordevices,circuitbo

ards,etc.

20Special

tradeworks

50Other

fuel

prod

ucts

81TV,radioreceivers,transm

itters,etc.

Man

ufacturing

51Basic

chem

icals

82Medical,surgical

&orthop

edic

appliances

21Meatandmeatprod

uctio

n52

Fertilizers

83Measuring

,checking

,etc.

22Preservationof

seafoo

d53

Paintsandvarnishes

84Optical

instruments,etc.

23Preservationof

fruitsandvegetables

54Pharm

aceutical,chem

ical,etc.

85Watches

andclocks

Con

tinued.



Table

A1.

Con

tinued.

No.

Sectoran

dSubsector

24Dairy

prod

uctio

n55

Soap,

perfum

es,cleaning

,etc.

86Motor

vehicles

25Oils

andfats

56Other

chem

ical

prod

ucts

87Motorcycles

26Grain

mills

57Tires

88Shipandbo

atbu

ilding,

bicycles,etc.

27Bakeryprod

ucts

58Rub

berprocessing

89Other

transportequipm

ent

28Con

fectionery

59Rub

berglov

es90

Other

manufacturing

91Recyclin

g10

1Portandairportop

erationservices

112

Researchanddevelopm

ent

Service

102

Highw

ay,bridge,andtunn

elop

eration

113

Professional

92Electricity

andgas

103

Com

mun

ication

114

Businessservices

93Waterworks

104

Banks

115

Pub

licadministration

94Who

lesale

andretailtrade

105

Financial

institu

tions

116

Edu

catio

n95

Accom

mod

ation

106

Insurance

117

Health

96Restaurants

107

Other

financialinstitu

tions

118

Defense

andpu

blic

order

97Landtransport

108

Realestate

119

Other

public

administration

98Water

transport

109

Ownershipof

dwellin

gs12

0Private

non-profi

tinstitu

tion

99Airtransport

110

Rentalandleasing

121

Amusem

entandrecreatio

nal

100

Rentalandleasing

111

Com

puterservices

122

Other

privateservices

Sou

rce:

Departm

entof

StatisticsMalaysia,

Gov

ernm

entof

Malaysia.

2010

.MalaysiaInpu

t–Outpu

tTables

2005

.Putrajaya.



Table A2. The Benchmark Scenario

Price Per Liter (RM)

July 2005 June 2008

No. Item Petrol Diesel Petrol Diesel

1 Product cost 2.13 2.10 4.28 4.252 Other costs 0.32 0.23 0.32 0.23

(i)þ Alpha 0.05 0.04 0.05 0.04(ii)þ Operational cost 0.10 0.10 0.10 0.10(iii)þ Oil company’s margin 0.05 0.02 0.05 0.02(iv)þ Fuel retailer’s margin 0.12 0.07 0.12 0.07

3 Actual price 2.45 2.33 4.60 4.484 Retail price 1.62 1.28 2.70 2.585 Taxþ subsidy (0.83) (1.05) (1.90) (1.90)

Sales tax foregone 0.58 0.20 0.59 0.20Subsidy 0.25 0.85 1.32 1.70

6 Price increase (%) 10.20 36.50 28.70 38.00Base case scenario 0.25� 10% 0.25� 10%

Alternative scenario 1.32� 30% 1.70� 40%

RM ¼ Malaysian ringgit.Notes:a. Product cost is based on average current Means of Platts Singaporeþ alpha (a constant: 5 sen for

petrol and 4 sen for diesel).b. Other costs are the summation of i, ii, iii, and iv, which are fixed elements.c. Actual price is obtained from product cost and other costs (aþ b).d. Retail price is obtained from actual price� (taxþ subsidy).e. Sales tax is obtained from actual price� subsidy.f. Subsidy per liter of petrol and diesel obtained in July 2005 and June 2008, respectively.g. Price increase is obtained from subsidy divided by actual price (a proportion of subsidy removed from

the actual price).Source: Authors’ calculations based on data from the Department of Statistics Malaysia, Government ofMalaysia. 2010. Malaysia Input–Output Tables 2005. Putrajaya.



Vol. 39 No. 1 °c 2022, Asian Development Bank and Asian Development Bank Institute

Publication Stock No. JRN220114-2

Asian Development Review (ADR)

The Asian Development Review (ISSN 0116-1105, E-ISSN 1996-7241) is published twice a year (March,September) by World Scientific Publishing Co. Pte. Ltd., 5 Toh Tuck Link, Singapore 596224. A free electronicfull-text version of the Asian Development Review is available from World Scientific Publishing Co. Pte. Ltd.

The Asian Development Review is available under the Creative Commons Attribution 3.0 (CC BY 3.0) license:https://creativecommons.org/licenses/by/3.0. By using the content of this publication, you agree to be bound bythe terms of said license. This CC license does not apply to non-ADB and ADBI copyright materials in thispublication. If the material is attributed to another source, please contact the copyright owner or publisher of thatsource for permission to reproduce it. ADB, ADBI, and World Scientific Publishing Co. Pte. Ltd. cannot be heldliable for any claims that arise as a result of your use of the material. Please address inquiries to the PermissionsDepartment of World Scientific Publishing Co. Pte. Ltd at [email protected].

Subscriptions, changes of address, and single-copy orders should be addressed to Journal Department, WorldScientific Publishing Co. Pte. Ltd., 5 Toh Tuck Link, Singapore 596224, or 27 Warren Street, Suite 401-402,Hackensack, NJ 07601, USA, or 57 Shelton Street, Covent Garden, London WC2H 9HE, UK.

Typeset by Stallion Press

Email: [email protected]

March 23, 2022 7:06:08pm WSPC/ADB ISSN: XXXX-XXXX

World Scientific

Vol. 39 • N

o. 1 • M

arch 2022ISSN: 0116-1105

Asian Development ReviewThe Asian Development Review (ADR) is a professional journal that publishes research on development issues relevant to the Asia and Pacific region, specifically members of the Asian Development Bank. The ADR publishes high-quality empirical papers, survey articles, historical analyses, and policy-oriented work. The ADR bridges theoretical work and rigorous empirical studies that advance the understanding of Asia’s development; and it is open to discussions of alternative perspectives on all aspects of development, including globalization, inequality, structural transformation, and poverty. The ADR aims to disseminate the best research, with contributions from scholars in all fields, with the objective of providing the knowledge necessary for designing, implementing, and sustaining effective development policies. Its intended audience comprises a worldwide readership of economists and other social scientists.

Asian Development ReviewVolume 39 · Number 1 · March 2022Asian Development Review

MCI (P) 076/09/2021













Shruti Sharma

(Continued )


Asian Development Review: Volume 39, Number 1

Documents

Transcript of Asian Development Review: Volume 39, Number 1