Testing Creative Destruction in an Opening Economy: the Caseof the South African Manufacturing Industries�

Philippe Aghion,yJohannes Fedderke,zPeter Howitt,xand Nicola Viegi{

Abstract

This paper analyses the relationship between trade liberalization and economic growth using a Schum-peterian framework of technological innovation and applies it to sector level South African data. Theframework examines direct and indirect e¤ects of trade liberalization on productivity growth. Indirectimpacts operate through a di¤erential impact of trade liberalization on �rms conditional on their dis-tance from the international technological frontier. Estimation results con�rm positive direct impacts oftrade liberalization. Results con�rm also that the greatest positive impact of trade liberalization will beon sectors close to the international technological frontier and that experienced a low level of productmarket competition before liberalization.

Keywords: productivity growth, trade liberalization, South African manufacturingJELCodes: F1, L1, L6, O1, O3.

�The authors acknowledge research support from Economic Research Southern Africa.yHarvard UniversityzEconomic Research Southern Africa and Penn State UniversityxBrown University{Economic Research Southern Africa and University of Pretoria

1

1 Introduction

What policy intervention does a small, less developed economy far from the international technologicalfrontier use in order to promote growth? One intervention that has frequently been posed in the literatureis trade liberalization.Empirical results on the impact of openness and economic growth have steadily accumulated in the

literature. Aghion and Howitt (AH, 2009) report three core results coming from the growth and tradeliterature. First, openness has a direct positive and signi�cant e¤ect on growth. Second, the positive e¤ecton growth of trade liberalization is higher for smaller countries, because of a market size e¤ect or a scalee¤ect whereby the larger the domestic economy relative to the world economy, the less innovation or learning-by-doing domestic producers gain by opening up to trade.1 .Third, growth is less enhanced by openness inmore advanced countries, interpreted as a knowledge spillover e¤ect whereby trade induces knowledge �owsacross countries, such that more advanced countries stand to gain proportionately less from such knowledgespill-overs.2

But there is an additional e¤ect of trade on growth highlighted in AH which is not captured in theliterature: namely that trade liberalization tends to enhance product market competition, by allowing foreignproducers to compete with domestic producers. This in turn should enhance domestic productivity for atleast two reasons: �rst, by forcing the most unproductive �rms out of the domestic market; second, byforcing domestic �rms to innovate in order to escape competition from their new foreign counterparts. Thisframework is complementary to Melitz (2003) which emphasizes the increase in productivity following tradeliberalization due to a better allocation of factors of production. While AH do not identify the speci�cmechanism which induces and aggregate increase in productivity, either by reallocating resources to themost productive �rms as in Melitz, or by technological and skill upgrading, as in Bustos (2011), they suggestthat the positive e¤ect of trade liberalization on productivity and growth is a negative function of thedistance from the technological frontier of national �rms. Moreover AH allows us to think about the e¤ect ofunilateral liberalization, where tari¤s on imports are reduced without a parallel change in export conditions:in this context the increase in productivity will be a response of national �rms to an increase in competitivepressure, but it will only happen if no other instruments of market control is available to the �rm.In this paper we test the theoretical model of Aghion and Howitt in a middle income country context,

using South African manufacturing sector data, for three digit SIC industrial sectors. The case of SouthAfrica is interesting because the country underwent signi�cant trade liberalization; it has a heterogeneousmanufacturing sector and it has signi�cant internal market monopolies. Moreover the South African expe-rience with trade liberalization and economic growth has been extensively analyzed in the literature, whichprovides a useful set of results to use as comparative check of our analysis. Previous studies have examinedthe relationship between pricing power of industry and growth,3 market structure and growth,4 investmentin R&D and human capital and growth,5 and one study has considered the relationship between opennessand growth of total factor productivity in the South African context.6 It found a strong positive correlation,although mitigated by market imperfections, but the speci�cation estimated did not capture the full set oftheoretical considerations detailed below (as is true of most studies examining trade and growth e¤ects).The objective of this paper is to evaluate the composition and the nature of productivity gains (if any)

which result from trade liberalization. Section 2 outlines the open economy Schumpeterian theoretical frame-work employed in the paper. Section 3 provides background on the nature and extent of South African tradeliberalization. In section 4 the empirical strategy of the paper is explained, including the data sets employed,while section 5 reports estimation results which highlight the strict relationship between trade liberalizationand the internal competitive environment. Section 6 concludes discussing how trade liberalization and inter-nal regulatory environment are two necessary and complementary components of outward oriented growthstrategy.

1This result was �rst pointed out by Alesina, Spolaore and Wacziarg (2003).2This knowledge spillover e¤ect has been analyzed at length by Keller (2004) - and see also Sachs and Warner (1995) and

Coe and Helpman (1995).3See Aghion, Braun and Fedderke (2008).4See Fedderke and Szalontai (2009) and Fedderke and Naumann (2011).5See Fedderke (2006).6See Fedderke (2006).

2

2 Theoretical Framework

The theoretical model of this study is provided by the Schumpeterian framework of Aghion and Howitt (1992)extended to an open economy in Aghion and Howitt (2009). The Schumpeterian paradigm has proved to beuseful in considering extensions of models of economic growth beyond the impact of innovation on economicdevelopment. Since the endogeneity of innovation requires an explicit treatment of the source of e¢ ciencygains, the Schumpeterian framework is useful in analyzing the interaction between institutions and economicgrowth. Trade policies are just another set of institutions that de�ne the dimension of the market and thelevel of competition that national �rms face. On the one hand trade liberalization increases the size of thepotential market, thus increasing the expected pro�ts from successful innovation. On the other hand tradeliberalization increases the level of market competition and the ability of the national �rm to successfullyexpand and innovate depends critically on their ability to "compete."

2.1 The Closed Economy

Consider �rst the closed-economy version of the model. A unique �nal good, which also serves as numéraire,is produced competitively using a continuum of intermediate inputs according to:

Yt = L1��

Z 1

0

A1��it x�itdi; 0 < � < 1 (1)

where L is the domestic labor force, assumed constant, Ait is the quality of intermediate good i at time t,and xit is the �ow quantity of intermediate good i being produced and used at time t.Each intermediate sector has a monopolist producer who uses the �nal good as the sole input, with one

unit of �nal good needed to produce each unit of intermediate good. The monopolist�s cost of productionis therefore equal to the quantity produced xit: The price pit at which this quantity of intermediate good issold to the competitive �nal sector is the marginal product of intermediate good i in (1). The monopolistwill choose the pro�t maximizing level of output:

xit = AitL�2=(1��) (2)

with pro�t level:�it = �AitL (3)

where � � (1� �)�1+�1�� .

Equilibrium level of �nal output in the economy can be found by substituting the xit�s into (1), whichyields

Yt = �AtL (4)

where At is the average productivity parameter across all sectors At =R 10Aitdi, and � = �

2�1�� .

Productivity growth comes from innovations. In each sector, at each date there is a unique entrepreneurwith the possibility of innovating in that sector. He is the incumbent monopolist, and an innovation wouldenable him to produce with a productivity (quality) parameter Ait = Ai;t�1 that is superior to that of theprevious monopolist, by the factor > 1. Otherwise his productivity parameter stays the same: Ait = Ai;t�1:Innovation with any given probability � entails the cost cit(�) = (1� �) � � (�) � Ai;t�1, of the �nal goodin research, where � > 0 is a parameter that represents the extent to which national policies (institutions)encourage innovation, and � is a standard convex cost function. Thus the local entrepreneur�s expected netpro�t is:

Vit = E�it � cit(�) (5)

= ��L Ai;t�1 + (1� �) �LAi;t�1 � (1� �)� (�)Ai;t�1:

Each local entrepreneur will choose a frequency of innovations �� that maximizes Vit: The �rst-order conditionfor an interior maximum @Vit=@� = 0, can be expressed as the research arbitrage equation:

�0 (�) = �L ( � 1) = (1� �) : (6)

3

If the research environment is favorable enough (� is large enough), or the population large enough, so that:

�0 (0) > �L ( � 1) = (1� �) ;

then the unique solution � to (6) is positive, so in each sector the probability of an innovation is that solution(b� = �), otherwise the local entrepreneur chooses never to innovate (b� = 0). Since each Ait grows at the rate � 1 with probability b�; and at the rate 0 with probability 1� b�, the expected growth rate of the economyis:

g = b� ( � 1) :So we see that countries with a larger population and more favorable innovation conditions will be morelikely to grow, and grow faster.

2.2 Opening the Economy

Now open trade in goods (both intermediate and �nal) between the domestic country and the rest of theworld. For simplicity, assume two countries, �home�and �foreign,�with an identical range of intermediategoods and �nal product, and no transportation costs. Within each intermediate sector the world market canthen be monopolized by the lowest cost producer. Asterisks denote foreign-country variables.The immediate e¤ect of this opening up is to allow each country to take advantage of more productive

e¢ ciency. In the home country, �nal good production will equal

Yt =

Z 1

0

Yitdi = L1��

Z 1

0

bA1��it x�itdi; 0 < � < 1 (7)

where bAit is the higher of the two initial productivity parameters bAit = max fAit; A�itg. The e¤ect on theforeign country will be symmetric.Monopolists�pro�t will now be higher than under autarky, because of increased market size. For price

pit, �nal good producers will buy good i up to the point where marginal product equals pit:

xit = bAitL (pit=�) 1��1 and x�it = bAitL� (pit=�) 1

��1 (8)

so that price will depend on global sales relative to global population:

pit = �

Xit

(L+ L�) bAit!��1

: (9)

Accordingly the monopolist�s pro�t �it will equal revenue pitXit minus cost Xit, and pro�t maximizationrequires that:

Xit = bAit (L+ L�)�2=(1��)with price pit = 1=� and pro�t level:

�it = � bAit (L+ L�) : (10)

Substitution of prices pit = 1=� into the demand functions (8) yields

xit = bAitL�2=(1��) and x�it = bAitL��2=(1��)and substituting these into the production functions, �nal good production in the two countries will beproportional to their populations:

Yt = � bAtL and Y �t = � bAtL� (11)

and cross-sectoral average of the bAit�s, bAt = R 10 bAitdi.

4

2.2.1 The Impact of Trade Liberalization on Innovation

The impact of trade liberalization on innovation is analogous to that of competition on innovation.7 Here thestylization is that the competitor comes from the foreign country. Consider the innovation process in a givensector i. In the country where the monopoly currently resides, the country is on the global technology frontierfor sector i, and the local entrepreneur will aim at making a frontier innovation that raises the productivityparameter from bAit to bAit. If so, that country will retain a global monopoly in intermediate product i. Inthe other country, the local entrepreneur will be trying to catch up with the frontier by implementing thecurrent frontier technology. If he succeeds and the frontier entrepreneur fails to advance the frontier thatperiod, then the lagging country will have caught up, both countries will be on the frontier, and we cansuppose that each entrepreneur will monopolize the market for product i in her own country. But if thefrontier entrepreneur does advance the frontier then the entrepreneur in the lagging country will still remainbehind and will earn no pro�t income.The optimization problem of the �rm (5) is than amended considering the new competitive conditions.

The �rst order condition for an interior maximum gives three possible research arbitrage equations, functionof the distance from the technological frontier of the national �rm8 :

A Case A is the case in which the lead in sector i resides in the home country, while the foreign countrylags behind. In this case the open economy research arbitrage equation governing �A:

(1� �)�0 (�A) =� = ( � 1) (L+ L�) + ��AL� (12)

makes clear that for the technology leader innovation will be greater than under the closed economy(compare equation 6). This arises because of:

� Scale e¤ects realized because the successful innovator gets enhanced pro�ts from both markets(L+ L�), not just the domestic market, L, thus giving a stronger incentive to innovate.

� Escape entry e¤ects arising because the unsuccessful innovator in the open economy is at risk oflosing the foreign market to the foreign rival, avoidable by innovation (��AL

�). The unsuccessfulinnovator in the closed economy loses nothing to a foreign rival and thus does not have this extraincentive to innovate.

B Case B is the case in which the domestic and foreign sectors are neck-and-neck. In this case the openeconomy research arbitrage equation governing �B :

(1� �)�0 (�B) =� = ( � 1)L+ ��BL+ (1� ��B) L�

again has scale ((1� ��B) L�) and escape competition (��BL) e¤ects, with symmetrical intuition as for�A above.

C Case C is the case in which the foreign country starts with the lead. Here the open economy researcharbitrage equation governing �C :

(1� �)�0 (�C) =� = (1� ��C)L

shows that sectors behind the world technology frontier will be discouraged from innovating by thethreat of entry because even if it innovates it might lose out to a superior entrant. Provided that theforeign country�s innovation rate is large enough when it has the lead, then the right-hand side of thisresearch arbitrage equation will be strictly less than that of the closed economy (compare equation 6),so we will have �C < �.

It follows that �A > �, �B > �, and ��C > ��, ��B > ��, with �C and ��A indeterminate. It therefore

follows that a su¢ cient (not necessary) condition for the innovation of the domestic economy to be higherunder openness is that j�A + �B j > j�C j, and symmetrically for the foreign country that j��B + ��C j > j��Aj.

7See Aghion et al (2005), Aghion and Gri¢ th (2005), and Aghion et al (2008).8For a detailed derivation of the research arbitrage conditions in open economy see Aghion and Howitt (2009)

5

2.2.2 The Impact of Unilateral Trade Liberalization

The main conclusion of the model, namely that trade openness induces productivity growth by increasingthe return of successful innovation, can be extended to capture the e¤ect of unilateral trade liberalization.With unilateral trade liberalization we identify a policy of reduction of trade barriers aimed at opening thenational economy to competition from foreign imports of intermediate inputs. The competitive productionof the unique �nal good continues to be characterized by (1).Before unilateral liberalization, user of an imported intermediate good faces a tari¤ of � units of the

�nal good per unit of the intermediate input acquired. While demand for intermediate inputs from domesticproducers continues to re�ect the marginal product on intermediate inputs directly:

pit = �L1��

�Aitxit

�1��(13)

that for foreign intermediate inputs re�ects marginal product net of import tari¤s:9

p�it (1 + �) =�Y

�xit= �L1��

�A�itxit

�1��(14)

p�it =1

(1 + �)�L1��

�A�itxit

�1��(15)

Final goods production then employs the intermediate good with the highest productivity net of taxes,so that:

Yt = L1��

Z 1

0

bA1��it x�itdi; 0 < � < 1 (16)

where:

bAit = max(Ait; A�it

(1 + �)1

1��

)(17)

The tari¤ operates as a wedge between national and international productivity. Hence, as before, liber-alization will increase the productivity of the �nal good sector and corresponding labour income. However,the pro�t of the intermediate input producing monopolist, will inevitably decrease under liberalization, sincethe reduction in tari¤s, being unilateral, will reduce aggregate pro�ts due to the substitution of domesticintermediate goods with foreign, without compensation through an increase in pro�ts due to an increase inthe market size for the surviving domestic �rms. Thus the impact is certainly negative on monopoly pro�ts.On the other hand, the impact of liberalization on innovation and research is more ambiguous. In an

unilateral trade liberalization, the incentive to innovate comes only from the need of incumbent �rms tomaintain market share in the local market. As before we have three di¤erent possibilities:

A For a �rm that is a technological leader unilateral liberalization of imports does not change the optimallevel of research spending. Because the dominance in the local market is guaranteed by technologicalsuperiority and not by trade barriers, the impact of liberalization generates neither a scale e¤ect ofcapturing foreign markets (foreigner has not liberalized by assumption), nor an escape entry e¤ect.10

B For �rms that are close to the technological frontier the result is signi�cantly di¤erent. When the countryimposes trade barriers to import, the national �rm incentive to innovate comes only from the possibilityof gaining control of the foreign market. Thus with unilateral trade barrier, the expected pro�t of theentrepreneur in the home country is:

EUB = f�B [L+ (1� ��B)L�] + (1� �B)g� � (1� �)� (�B)9We assume that domestic is small, so that domestic demand does not a¤ect world price.10Notice that this conclusion is derived under the assumption of Step-by-Step innovation process ( see Aghion and Howitt

2009). Because there cannot be leapfrogging in the innovation process, the national monopoly of the leading �rm is guaranteedby technological superiority, which might be contested only in the future after a series of unsuccessful innovation attempts.

6

and the corresponding research arbitrage equation would be:

(1� �)�0 (�)�

= ( � 1)L+ (1� ��)L� (18)

where the second term gives the incentive to innovation produced by the access to export market. Thereduction on import tari¤s introduces the possibility that a foreign successful innovator might gaincontrol of national market, thus increasing the incentive to innovate for the national �rms. In fact theresearch arbitrage equation after liberalization is:

(1� �)�0 (�)�

= ( � 1)L+ (1� ��)L� + ��L (19)

Thus the rate of innovation of �rms close to the technological frontier (net of any remaining tari¤) willincrease after the unilateral opening of the economy.

C For �rms lagging behind the technological frontier, the e¤ect of unilateral opening will be unchanged fromthe case of bilateral liberalization, and the rate of innovation after liberalization will be lower thanunder the closed economy.

The theory gives a rich set of empirical predictions After a reduction in trade barriers, any sector in theeconomy should experience an increase in productivity, either because the �nal sector adopts more advancedforeign technology or because only the more productive intermediate good �rms survive international com-petition (selection e¤ects). At the same time the positive e¤ect of trade liberalization on productivity wouldbe limited by the extent that national �rms can maintain anti-competitive behaviour. This limits the extentof penetration of more productive foreign �rms and technology. The furthest away from the technologicalfrontier a country is, the greatest the productivity gain that can be achieved via trade liberalization, be-cause of the jump in technological adoption that trade openness allows. On the other hand, the adoption ofmore advanced foreign technology in the �nal good sector will reduce the variety of national production ofintermediate goods, with the exit of the least productive �rms.A second set of results refer to the dynamic e¤ect of an increase in competition on the incentives to

invest and innovate. For this the most important variable is the distance from the technological frontierof national �rms and the size of the potential market for successful innovators. Thus �rms closer to thetechnological frontier will have an incentive to increase investment in technological innovation as a defensivemeasure against foreign competitors and as an instrument to gain access to a much larger potential market.This should be particularly true for �rms located in small countries. Instead, if a �rm is far away from thetechnological frontier, trade liberalization will discourage further investment in innovation. This, coupledwith the selection e¤ect, should induce signi�cant exit from the market of ine¢ cient �rms.The rest of the paper will take these theoretical observations to the data, looking at the experience of

trade liberalization in South Africa from the end of the 1980s to the present.

3 The South African Experience of Trade Liberalization and Growth

South Africa represents an interesting case where trade liberalization can be well located in time and at thesectoral level. After an early import substitution strategy, policies of trade liberalization were followed fromthe early 70�s to the early 80�s. This timid process of liberalization was interrupted and partly reverted afterthe debt crisis of 1985 and the following economic downturn. It is only at the beginning of the 90�s and inparticular with the new democratic government in 1994 that the policy takes a decisive turn towards a moreliberal trade regime. Although in the literature there is some disagreement about the extent of e¤ective tradeliberalization,11 Edwards (2005) provides the most recent re-evaluation of the extent to which South Africahas liberalized its trade since the late 1980s. He �nds that signi�cant progress has been made in terms ofreducing tari¤ protection. In particular, between 1994 and 2004, the e¤ective protection in manufacturingfell from 48% to 12:7%.11See for example Fedderke and Vaze (2001) for a sceptical view on the real extent of trade liberalization in South Africa,

and Rangasamy and Harmse (2003) for a more positive assessment

7

Table 1 shows the change in e¤ective rate of protection and nominal tari¤s (both including surcharges)for the sectors used in this study between 1994 and 2003 and the e¤ective rate of protection and the nominaltari¤s at the end of the sample. It is evident that there is a generalized reduction in trade protection, butthe extent of liberalization is di¤erentiated across sectors.The 5th column of table 1 shows the change in employment in each industry from 1994 to 2003 relative to

the average change in employment in the manufacturing sectors used in the study, which saw a reduction ofemployment of 11%. It is clear that some sectors have seen a signi�cant reduction in employment, expeciallyfootwear and textile, which are still protected by relative high e¤ective rate of protection, and glass & glassproducts, basic iron & steel and basic non-ferrous metals, which have seen instead a signi�cant reductionin tari¤ protection. Other sectors instead have seen a signi�cant expansion in employment following thechange in tari¤ protection. Overall the correlation between change in tari¤s and change in employment isnot signi�cant while change in employment is correlated (0.4) with changes in pricing power, shown in thelast column of table 1. The objective of the empirical analysis is to make sense of these di¤erent patternsrelating the di¤erent responses of manufacturing sectors to trade liberalization to their distance from thetechnological frontier, as suggested by the theory.

4 Empirical Strategy

4.1 The Data Sets

For this study we employ industry level data from a number of distinct sources:

1. Industry-level panel data for South Africa from the Trade and Industry Policy Strategies (TIPS)database. The data employed for this study focuses on the three digit manufacturing industries, overthe 1988-2003 period. Variables for the manufacturing sector include the output, capital stock, andlabour force variables their associated growth rates, the distribution of value added between factor

8

inputs, and the skills composition of the South African manufacturing labour force by manufacturingsector.

2. Industry-level panel data for South Africa and the USA and 28 manufacturing industries from 1963-2003, obtained from UNIDO�s International Industry Statistics 2004. These data are used to computeSouth Africa and US industry total factor productivity and distance from the technological frontierof South African 3-digit manufacturing industry .This data set contains yearly information on output,value added, total wages, and employment, gross capital formation and the distribution of value addedbetween factor inputs. From the gross capital formation data we compute capital stock data on thebasis of the perpetual inventory methodology.12

3. For measures of industry pricing power, we employ the estimated values of the mark-up of price overmarginal cost of production of Aghion et al (2008) obtained from the Roeger (1995) methodology.

4. Given the focus of the present study on the impact of trade liberalization on productivity growth,accurate measures of openness or protection are crucial. For our openness indicators, we employ dataon e¤ective rates of protection and scheduled nominal tari¤ rates obtained from Edwards (2005).13

While most indicators employed for this study are available over the 1970-2004 or 1970-2002 period, thetrade measures are restricted to the 1988-2004 period. In addition, data comparability issues between the USand SA reduced the total number of comparable sectors from 28 to 23 sectors. The list of sectors includedin the panel is that speci�ed in Table 1. This generated a panel of dimension 23� 17 = 391 observations.14

4.2 The Distance From Frontier Measures

Following Aghion et al (2005) and Aghion and Gri¢ th (2005), we generate an industry and time speci�cmeasure of distance from the technological frontier, under the assumption that the USA constitutes thetechnological leader for South African industry.15 The measure we employ is given by

Mi;t = tfpSA;i;t=tfpUS;i;t (20)

where the measure of distance from the frontier, M , for industry i in year t, in country X = [SA;US], isthe di¤erence between total factor productivity (TFP) in the US from that in SA for that industry andyear. TFP is computed by means of the primal decomposition, with factor shares given by the share oflabour remuneration in value added. We compute the distance measure both by comparing US TFP withRand-denominated and Dollar-denominated South African TFP to ascertain how much real exchange ratedepreciation have helped the competitiveness of national �rms. Figure 1 shows the data patterns

12Since the comparison of distance from the frontier is conducted over the 1970-2002 period, and data for the US is availablefrom 1963, implementation employed a seven year lead, under an assumption of 15% depreciation rates.13Note that Edwards (2005) also contain measures of export taxes and anti-export bias. We also used both these measures

in estimation, with symmetrical results. In the case of the anti export bias measure, however, strong sectoral outliers renderthe measure less reliable in the sense of raising standard errors. Full results available from the authors on request.14On the nature and quality of this data see the discussion in Aghion et al (2008).15South Africa relative to its trading partners has developing country characteristics. Due to her international isolation,

there has been a heavy reliance on trading with developed countries, making the US (rather than African countries) a relevantcomparator. Edwards and Schöer (2002) report that 85% (57%) of South African imports (exports) in 1990 were sourced fromthe 25 OECD countries, declining to 72% (53%) in 1999. South African trade is thus heavily biased toward developed countriesover the sample period of this study, though middle income and developing countries have begun to feature more prominentlyin South African trade during the 1990s.

9

0.5

1

0.0

2.0

4.0

6.0

8

.05

.1.1

5.2

.1.2

.3.4

01

23

0.5

11.

5

0.5

11.

5

0.1

.2.3

.4

02

0.5

1

0.2

.4.6

01

23

02

46

8

0.5

11.

5

0.5

11.

5

01

23

0.5

11.

5

0.2

.4.6

0.5

11.

52

0.5

11.

5

1990 1995 2000 2005 1990 1995 2000 2005 1990 1995 2000 2005 1990 1995 2000 2005 1990 1995 2000 2005

Basic iron & steel Basic non-ferrous metals Beverages Food Footwear)

Furniture Glass & glass products Leather & leather products Machinery & equipment Metal products excl. machinery

Non-metall ic minerals Paper & paper products Plastic products Professional & scientific equip. Rubber products

TV, radio & communication eq. Textiles Transport equipment Wearing apparel Wood & wood products

Distance Rands Distance Dollars

Year

Graphs by Sector

Distance from Technological Frontier (USA) by Sector - TFP Leve

We �nd three broad patterns in the data.One grouping of thirteen sectors sees a steady widening of the technological gap between South African

and US TFP. While for six sectors the widening gap occurs from a base that is already very low (de�ned asless than 10% of US TFP productivity levels),16 for two sectors there is a dramatic increase of the distance ofTFP productivity from relatively close levels (de�ned as greater than 50% of US TFP productivity levels),17

and for four sectors the growing productivity gap occurs for mid-range productivity sectors (de�ned asbetween 10% and 50% of US TFP productivity levels).18

A second grouping of 5 sectors sees a narrowing of the TFP productivity gap between South Africa andthe US - though for a number of these sectors the �nal few years sees a reversal in the trend. Again, thereis a distinction between one sector for which the productivity gain has been substantial (to the point ofrising to TFP productivity levels that exceed that of the US),19 and four sectors for which the gain has beenmoderate.20

The third grouping of �ve sectors sees a catch-up of South African TFP productivity levels with theUS from 1988 through the mid-1990s, but with a subsequent reversal in the catch-up. In the case of onesector this decline is both dramatic and o¤ a relatively high base,21 for two sectors the decline occurs o¤ amid-level plateau,22 one sector experiences both substantial catch-up but equivalent decline toward the endof our sample period,23 and for one sector the movements are small leaving the sector at moderate US TFPproductivity levels throughout.24 All these patterns are shown in Figure 1.

16Beverages, Tobacco, Leather & leather products, Industrial chemicals, Basic non-ferrous metals, Other manufacturingequipment.17Footwear and Paper & paper products.18Food, Non-metallic mineral products, Basic iron & steel, Metal products.19Plastics & plastic products.20Wearing apparel, Wood & wood products, Furniture, Rubber & rubber products.21Television, radio & communication equipment.22Textiles and Professional & scienti�c equipment.23Glass & glass products.24Machinery & equipment.

10

What is particularly noteworthy is that productivity catch-up for South African manufacturing sectorsdoes not in general occur in sectors that are obviously natural resource extractive, where South Africa shouldhave most of its comparative advantage Non-metallic minerals, Basic iron & steel, Basic non-ferrous metals,Metal products, and Paper & paper products all consistently lose ground relative to US productivity levels,and in the case of virtually all of these sectors South African TFP productivity is never close to US levels -the only possible exception is Paper & paper products.However, given the �ndings of Aghion et al (2008) on the impact of market structure on productivity

growth, and of Fedderke (2006) on the impact of poor human capital endowments and low R&D investmentby South African manufacturing on productivity growth, these �ndings are not surprising.

4.3 The Empirical Speci�cation

Our model hypothesizes two fundamental e¤ects of trade liberalization. First, trade liberalization increasesaggregate productivity (and wages) through selection e¤ects and increases in competitive pressure. Second,innovation in sectors in which �rms are su¢ ciently close to the technological frontier reacts positively to anincrease the level of product market competition due to trade liberalization. Where they lag considerablybehind the frontier, the impact of the liberalization reverses. Thus increasing distance from the technologicalfrontier should raise the deleterious e¤ect of trade liberalization on productivity growth.25

To test these predictions of the theoretical framework, we examine productivity dynamics in South Africanmanufacturing sectors for the period 1988-2002 for the three digit SIC level data. The baseline speci�cationtests for a direct linear impact of the trade protection measure on productivity growth:

�Ait = a1Pi;t + �i + �t + uit (21)

where �Ait denotes productivity growth in sector i in year t, measured by TFP growth or growth in outputper worker. Pit is a measure of e¤ective trade barriers as detailed in the data section above. The P measureis given by measures of nominal tari¤s and of e¤ective protection rates. The measure is thus an inverseof openness. Finally, �i represent �xed e¤ects, and �t time e¤ects. The theoretical prior is that a1 < 0,since protection (as an inverse of openness) induces a decrease in the level of productivity growth eitherby preventing an increase of technological innovation by more advanced �rms or by preventing the importof new technology. In order to capture some of the entry-exit �ows that are central to the theoreticalanalysis, we also use the sectoral change in the output to labour ratio �(Yit=Lit) as an alternative measureof productivity growth for all speci�cations we estimate.Since the theoretical framework outlined above also identi�es an indirect impact of trade liberalization,

that di¤erentiates the impact of liberalization conditional on the distance of the industrial sector from thetechnological frontier, we supplement the baseline speci�cation by incorporating the impact of distance fromthe technological leader into the productivity growth dynamics. Formally, for the three digit SIC level datawe specify:

�Ait = a0 + a1Pi;t + a2Mi;tPi;t + �i + uit (22)

where terms are de�ned as before, Mi;t denotes the distance from the technological frontier de�ned inequation (20), and the term MitPi;t represents an interaction term that captures the relationship betweenopenness and technological innovation. Additional priors are that a2 < 0, such that the maximum e¤ect oftrade liberalization occurs in sectors that are closer to the technological frontier.Speci�cation (22) ignores one important additional factor known to be relevant to productivity growth

in South African manufacturing. Aghion et al (2008) demonstrated both that product market competitionhas strong predictive power for productivity growth in South African manufacturing, and that pricing powerof domestic producers in manufacturing appears to be substantial. Rodrik (2008) by contrast has arguedthat the relative price of manufacturing in the South African economy has declined, due in considerable

25A third impact of trade liberalization that emerges from our model is that the magnitude of the positive e¤ect of increasingopenness on productivity depends of the increase in the potential market size for winning �rms. This requires a comparison ofthe relative market size winning �rms have access to, ex ante and ex post trade liberalization. An immediate proxy that o¤ersitself is a comparison of a countries market size relative to world markets. However, this would be true only in the absence of anydi¤erentials in the rates at which countries lowered their trade barriers. Since we have data only on the levels of tari¤ barriersfor South Africa (and not its trading partners), we are unable to employ this proxy of the scale e¤ect of trade liberalization.We therefore do not model this e¤ect empirically.

11

measure to the rising import penetration associated with the liberalization of the economy, placing domesticproducers under a pro�t squeeze. For this reason we also test for the robustness of our �ndings on a1 anda2, by controlling for the impact of a Lerner index of pricing power.26 We also allow for the possibilityof interaction between product market competition and trade protection and distance from technologicalfrontier, providing us with:

�Ait = a0 + a1Pi;t + a2Mi;tPi;t + a3Lit + a4LitPi;t + a5LitMi;t + �i + uit (23)

with all terms de�ned as before, and with L denoting the measure of pricing power.Additional priors are that a3 < 0, given that lower product market competition hampers the escape

competition e¤ect. a4 < 0, since a decrease in trade protection should have greater productivity growthe¤ects for sectors with low rather than high product market competition. We do not have explicit theoreticalpriors for a5. Nonetheless, we posit a5 < 0 on the grounds that sectors with strong competitive pressure farfrom the technological frontier would be vulnerable to entry without the possibility of recourse to an escapecompetition response by incumbent producers.

5 Results

Our estimation strategy proceeds as follows: �rst we consider the direct e¤ects of trade protection on industryproductivity growth, as speci�ed in equation (21), and the non-linear speci�cation given by equation (22)under two alternative productivity measures, and for two measures of trade protection, e¤ective protectionrates and nominal tari¤ rates respectively.We follow the baseline estimation with a series of robustness checks. It is possible that the protection

measures we use are not exogenous to productivity growth. Sectors that are subject to low productivitygrowth have an incentive to lobby for protection. We allow for this possibility by estimating (21) and (22)under the systems GMM methodology. A further concern using panel estimation methodology might be thatthe underlying assumption of homogeneity across the manufacturing sectors required for the pooling of thedata is not appropriate, resulting in bias and inconsistency of estimation results. For this reason we employthe pooled mean group (PMG) estimation methodology and control for manufacturing group heterogeneity.Finally we check the impact of product market competition using speci�cations (23).

5.1 The Direct Productivity E¤ect of Trade Protection

We start by estimating the speci�cation (21). For the panel of South African manufacturing sectors listedin the �rst column of Table 1, we control for industry �xed e¤ects. Results are reported in Table 2A for theTFP productivity growth measure, and Table 2B for growth in the output labour ratio. In columns 1 to 5we present the results using the e¤ective rate of protection as trade liberalization measure while in columns6 to 9 we show the results using nominal tari¤s.

[INSERT TABLE 2A AND 2B ABOUT HERE]

Estimation results consistently con�rm the negative direct impact of trade protection on productivitygrowth (a1 < 0) both for the TFP growth measure and the growth in output per worker. We note that forboth e¤ective rate of protection and nominal tari¤ rate measures the negative impact is robust to industry�xed e¤ects (column 2 and 7). Symmetrically, results are robust to the inclusion of time e¤ects in the case of

26Mark-ups are obtained following the contributions by Hall (1990) and Roeger (1995) by means of:

NSR = �(p+ q)� � ��(w + l)� (1� �) ��(r + k)= (�� 1) � � � [� (w + l)��(r + k)]

where � = P=MC, with P denoting price, and MC denoting marginal cost. Under perfect competition � = 1, while imperfectlycompetitive markets allow � > 1. � denotes the di¤erence operator, lower case denotes the natural log transform, q, l, and kdenote real value-added, labour, and capital inputs, and � is the labour share in value-added. See the additional discussion inFedderke et al (2007).

12

13

the TFP productivity growth measure - though not for the labor productivity growth measure (see columns3 of Tables 2A and 2B).27

Both sets of results imply an economically signi�cant marginal impact. A 10 percentage point decreasein e¤ective protection increases TFP growth by 0:2� 0:4 percentage points and growth of output per workerby 0:1� 0:3 percentage points. A reduction in nominal tari¤s by 10 percentage points increases TFP growthby 1� 2 percentage points and growth of output per worker between 0:5 and 1 percentage points.We also allow for the possibility of endogeneity of trade protection, by employing the dynamic GMM es-

timator,28 as well as the possibility of possible heterogeneity across industry group beyond the time invariantindustry heterogeneity controlled for by �xed e¤ects, by employing a the pooled mean group estimator.29

We note that the data employed for estimation under both pooled OLS, the Within estimator, GMM andPMGE is uniformly stationary (see the Appendix reporting panel unit root tests), rendering the distributionalassumptions underlying the results from the regressors consistent.30

Our results prove to be robust to controlling for the possibility of endogeneity of the trade protectionmeasure under systems GMM, as reported table in columns 4 for e¤ective protection rates, and 8 for nominaltari¤s, of Tables 2A (for TFP growth) and 2B (for �(Yit=Lit)) respectively. The negative impact and botheconomic and statistical signi�cance of trade protection on productivity growth is maintained under the GMMestimator. We employ the two-step GMM estimator,31 under higher order lags as instruments in estimationas reported in Table 2A and 2B.32 The Sargan test of the overidentifying moment restrictions, distributed �2

under the null that the restrictions are valid, in all instances provides support for the instruments employed(by not rejecting the null). The AR tests, con�rm the presence of AR(1), and absence of AR(2), as requiredto be consistent with the AR structure in the GMM speci�cations.33

Allowing for the possibility of heterogeneity across manufacturing sectors by means of the PMG estimator,in columns 5 for e¤ective protection rates and 9 for nominal tari¤s, of Tables 2A (TFP growth) and 2B(�(Yit=Lit)) respectively, further con�rms the robustness of the �nding of the negative impact of tradeprotection on productivity growth.34 In application, the PMG estimator con�rms the homogeneity of theimpact of the trade protection measure in the long-run speci�cation (see the Hausman h-test statistic), aswell as adjustment to long-run equilibrium by virtue of the negative PHI coe¢ cient. Indeed, under the ARDLspeci�cation of the PMGE estimator, the strength of the long run impact of trade protection if anythingincreases from that obtained under the static �xed e¤ects estimation.Results for the TFP and output per worker productivity growth results are symmetrical throughout.The empirical results are thus encouraging for our posited theoretical framework. That the impact of

trade protection should be pervasively negative, even without controlling for the indirect e¤ects identi�ed bythe theory, and under alternative estimation methodologies that control both for potential endogeneity andpanel group heterogeneity, suggests some robustness of the link between trade dispensations and growth.

27Note: for the sake of parsimony, we report only the inclusion of time e¤ects in the case of the ERP protection measure.Results are symmetrical.28See Arellano and Bond (1991) and Blundell and Bond (1998).29See Pesaran, Shin and Smith (1999) A fuller discussion of the estimator is included in an appendix.30While the PMGE can deal with non-stationary data, it is also consistent with stationary data. The case of stationary

regressors for the PMGE is explicitly derived in Pesaran et al (1999: section 4.1), while the case of non-stationary data is dealtwith in section 4.2 of the paper. Thus the PMGE can be used either for stationary, or non-stationary data.31Since under potential endogeneity of the non-autoregressive regressors, the weight matrix employed to construct the single

step GMM-estimator is no longer known, the gain in e¢ ciency in the two-step estimator is held to be substantial in the literature.See the discussion in Bond (2002).32Reason for the higher order lags is that in the presence of MA(1)-processes in the error structure, the �rst di¤erenced error

term under GMM becomes MA(2). Hence lower order lags of the instruments are no longer valid. While results under the useof lower order lags are consistent with those reported, we report those that preclude the possiblity of the impact of the MAprocesses.33The AR-structure implied by the GMM structure appears well speci�ed, as con�rmed by the fact that the AR-coe¢ cient

under OLS estimation lies above, and that obtained under Within Group estimation lies below that obtained under GMMestimation. Since we know analytically that the OLS estimator is biased upward, the Within is biased downward, while theGMM estimator of the AR coe¢ cient is consistent, it is reassuring that empirically the GMM estimates of the AR coe¢ cientproves to be located between the OLS and Within limiting cases, providing further con�dence in the GMM speci�cation. Seealso the discussion in Bond (2002).34An Appendix provides an more detailed exposition of the PMG estimator. Here we note that it exploits the statistical

power of the panel data structure by imposing homogeneity on the long run association between variables, while allowing forindustry heterogeneity in the short run dynamics, and hence (through the solution of the implied di¤erence equations) in thesteady states to which di¤erent industries converge.

14

It remains to be seen whether controlling for the interaction e¤ects implied by our theory adds additionalinsight.

5.2 Testing the Robustness of Productivity E¤ect of Trade Protection

Under the speci�cation (22) we explicitly control for the di¤erential impact of trade liberalization on produc-tivity growth for sectors that di¤er in terms of their distance from the technological frontier. To determinethe robustness of our results we again estimate (22) under both productivity growth measures, for bothmeasures of trade protection, and under �xed e¤ects, GMM (to control for endogeneity), and PMG (tocontrol for sector heterogeneity) estimation methodologies.Results are reported in Tables 3A and 3B for the TFP and output-labour ratio productivity growth mea-

sures respectively. Columns (1) and (5) report results under �xed e¤ects estimation for e¤ective protectionrates and nominal tari¤ rates respectively. Columns (3) and (6) report GMM, and (4) and (7) report thePMGE estimations.

Note immediately that the negative impact of trade protection on productivity growth is con�rmed forall speci�cations that control for the impact of distance from the technological frontier (with the singleexception of the speci�cation in labor productivity growth and nominal tari¤s). More importantly in thepresent context, our results also con�rm our prior that sectors that are closer to the technological frontierbene�t more from trade liberalization in productivity growth terms (a2 < 0), than do sectors far from thetechnological frontier. The result is again robust across the alternative productivity growth measures, thealternative trade protection measures, as well as the estimation methodologies that control for sectorialheterogeneity by means of �xed e¤ects or the more elaborate dynamic PMGE methodology,35 as well as theGMM approach that allows for endogeneity of the protection measure.

35Only for the PMGE methodology under the nominal tari¤ measure does the interaction term report coe¢ cients of thewrong sign (for both productivity growth measures) - but under this speci�cation the coe¢ cients is statistically insigni�cant.

15

Note that the indirect impact of trade liberalization conditional on distance from technological frontieris of signi�cant magnitude in economic terms. For a 10 percentage point decrease in trade protection, eachpercentage point increase in distance from the technological frontier lowers productivity growth by between0:1 and 0:2 percentage points under the TFP productivity growth measure, and 0:13 and 0:44 percentagepoints under the output-labour ratio measure.Trade liberalization thus stands to bene�t sectors closer to the technological frontier considerably more,

than those that lag the frontier.

5.3 Controlling for the Impact of Product Market Competition

In Table 4 we report the results of estimations under the speci�cation given by (23). We allow for the impactof our measure of pricing power evident in South African manufacturing sectors, and for di¤erent levelsof pricing power for a di¤erential impact of both trade liberalization and distance from the technologicalfrontier. Columns 1 to 3 report results for TFP growth using the e¤ective rate of protection and columns 4to 6 report results using nominal tari¤ as protection measure.

Our �ndings of both the direct negative impact of trade protection and the greater positive impact oftrade liberalization for sectors closer to the technological frontier are robust to the additional controls forproduct market competition.The impact of pricing power continues to have the statistically signi�cant negative impact on productivity

growth (a3 < 0) noted in Aghion et al (2008). Indeed, the economic size of the impact is approximatelydouble that found for the closed economy model of the earlier study. Aghion et al (2008) found that a 0:1unit increase in the Lerner index resulted in the loss of approximately one percentage point in productivitygrowth as measured by TFP. Under the present open economy speci�cations the implication is that a 0:1unit increase in the Lerner index now results in a 1 to 2 percentage point loss in productivity growth under

16

the TFP productivity growth measure (see columns 1 and 4), and a loss of between 2 and 3 percentagepoints under output-labour ratio productivity growth (see columns 7 and 10).36

For the pricing power interaction terms, our results suggest that trade liberalization has greater produc-tivity growth e¤ects for sectors with low rather than high product market competition (a4 < 0).37

The results also support the inference that sectors with strong competitive pressure far from the techno-logical frontier experience lower productivity growth (a5 < 0).Our �ndings on the impact of trade liberalization (direct and indirect) are therefore robust to the inclusion

of the measure for pricing power in the form of the proxy of the Lerner index.

6 Conclusion and Evaluation

This paper has provided a new approach for the examination of the linkage between trade liberalization andproductivity growth.The theoretical framework employed in the paper, while acknowledging a direct impact of openness on

growth, also serves to highlight that the impact of trade liberalization on growth may also operate throughindirect channels. Speci�cally, the prediction is that sectors that are closer to the world technological frontiershould bene�t more from trade liberalization.We report empirical results from panel estimations at industry level for the South African manufacturing

sector.Our results con�rm that the productivity growth impact of trade liberalization is positive, both in terms of

its direct and in terms of its indirect impacts. Trade protection has a direct negative impact on productivitygrowth. Sectors which are closer to the international technological frontier bene�t most signi�cantly fromtrade liberalization. These �ndings are robust to controlling for potential endogeneity of the trade protection

36While Rodrik (2008) was therefore correct to caution that the trade context is important to the quanti�cation of the impactof pricing power on productivity growth, the impact of trade liberalization is not such as to eliminate the impact of pricingpower - instead it enhances its importance. Not controlling for the reduction in trade protection biases the impact of pricingpower downward. Further, and crucially for the policy context we also note that liberalization of the South African economy isnot only incomplete at present, but that the regulatory framework on output markets in South Africa is more stringent thaneven for the OECD (see OECD, 2008: 65), thus creating non-tari¤ barriers to entry that continue to protect incumbent �rmseven under trade liberalization.37Note that there is no clear theory that speci�es the length of time it takes for dynamics to take e¤ect. Given only adjustment

costs, there is no need for the impact to be instantaneous. In the absence of close theoretical guidance therefore, our reasoningis straightforwardly one of sequencing: change in protection (-2) impacts pricing power (-1) and the resultant combined e¤ectimpacts on productivity growth. This is applied consistently in all instances of this interaction term - see also Table 7.

17

measure by means of GMM estimation, and manufacturing sector heterogeneity by means of �xed e¤ectsand pooled mean group estimation, as well as to controlling for the impact of product market competition.Additional �ndings indicate that sectors with low product market competition bene�t most signi�cantly

from trade liberalization. Further, we report results that indicate that sectors close to the technologicalfrontier, that have greater product market competition experience higher productivity growth.Our results also strengthen �ndings of a positive impact of product market competition on productivity

growth. Under our open economy speci�cations, controlling for trade liberalization serves to verify thepositive impact of competitive pressure on productivity growth, and suggests that the impact of competitionis in fact greater than suggested by speci�cations that do not control for the extent of trade protection. Our�ndings indicate a productivity growth impact of double the magnitude obtained from estimations that donot control for trade protection measures.These results highlight the strict relationship between trade liberalization and the internal competitive

environment. While trade liberalization have a positive e¤ect on productivity growth (but not necessarilyon sector size), this e¤ect is possible only if internal market conditions are competitive enough. The resultsseems to suggest that the reason we do not see a technological catching up in some sectors after tradeliberalization is that the market power of the national �rms is actually increased, thus limiting in a di¤erentway the e¤ect of international competition on productivity. The main policy implication of the study istherefore that trade liberalization and addressing the restrictiveness of the regulatory environment are twonecessary and complementary components of outward oriented growth strategy.What remains to analyse is the extent of "destruction" caused by trade liberalization, for which ideally

we would need information about entry and exit of �rms in di¤erent sectors. Nevertheless we can noticethat, according to the theory, the process of destruction and the process of technological catching up arepositive correlated, because only �rms able to compete internationally would be able to survive in the newcompetitive environment. It is therefore arguable that the limited technological catching up that we see inthe data is probably a re�ection of the limited amount of overall liberalization.

7 Appendix: TFP Measure for South Africa and US

To calculate the Total Factor Productivity at the industry level we have used Industry-level panel data forSouth Africa from the Trade and Industry Policy Strategies (TIPS) database. The variables in the datasetfor the manufacturing sector include output, capital stock, and labour force variables and their associatedgrowth rates, the distribution of value added between factor inputs, and the skills composition of the SouthAfrican manufacturing labour force by manufacturing sector. The TFP is than calculate directly usingproduction function methods, assuming a Cobb-Douglas production function with capital and labour sharesof value added calculated from the data.For the United States instead we use Industry-level data for 28 manufacturing industries from 1963-2003,

obtained from UNIDO�s International Industry Statistics 2004. This data set contains yearly informationon output, value added, total wages, and employment, gross capital formation and the distribution of valueadded between factor inputs From the gross capital formation data we compute capital stock data on thebasis of the perpetual inventory method. The remaining of the computation is in line with the computationfor South Africa. The main objective was to make the two series of TFP comparable, thus using the sameproduction function method.The measure of distance is than calculated by the ratio of TFP in SA sectors relative to corresponding

US sectors.

8 Appendix: The Pooled Mean Group Estimator

The relationships proposed by our theory tend to hold in the long-run and deviate from its equilibrium pathin the short-run. The underlying economic theory does not explore these issues, but there is still a potentialneed to control for dynamics. To this end we use the Pooled Mean Group (PMG) dynamic heterogeneouspanel estimator of Pesaran, Shin and Smith (1999).

18

Consider the unrestricted error correction ARDL(p; q) representation:

�yit = �iyi;t�1 + �0ixi;t�1 +

p�1Xj=1

�ij�yi;t�j +

q�1Xj=0

�0ij�xi;t�j + �i + "it; (24)

where i = 1; 2; :::; N; t = 1; 2; :::; T , denote the cross section units and time periods respectively. Here yit is ascalar dependent variable, xit (k�1) a vector of (weakly exogenous) regressors for group i, and �i represents�xed e¤ects. We allow the disturbances "it�s to be independently distributed across i and t, with zero meansand variances �2i > 0, and assume that �i < 0 for all i. Then, there exists a long-run relationship betweenyit and xit:

yit = �0ixit + �it; i = 1; 2; :::; N; t = 1; 2; :::; T; (25)

where �i = ��0i=�i is the k � 1 vector of the long-run coe¢ cients, and �it�s are stationary with possiblynon-zero means. This allows (24) to be written as the error correction model:

�yit = �i�i;t�1 +

p�1Xj=1

�ij�yi;t�j +

q�1Xj=0

�0ij�xi;t�j + �i + "it; (26)

where �i is then the error correction coe¢ cient measuring the speed of adjustment towards the long-runequilibrium.This general framework gives the formulation of the PMG estimation, which allows the short-run co-

e¢ cients and error variances to di¤er freely across groups, but the long-run coe¢ cients to be homoge-nous; i.e. �i = � 8 i. Group-speci�c short-run coe¢ cients and the common long-run coe¢ cients arecomputed by pooled maximum likelihood estimation. Denoting these PMG estimators by ~�i, ~�i, ~�ij , ~�ij

and ~�, we obtain them by �̂PMG =PN

i=1~�i

N , �̂PMG =PN

i=1~�i

N , �̂jPMG =PN

i=1~�ij

N , j = 1; :::; p � 1, and�̂jPMG =

PNi=1

~�ijN ; j = 0; :::; q � 1, �̂PMG = ~�. PMGE provides an intermediate case between the dynamic

�xed e¤ects (DFE) estimator which imposes the homogeneity assumption for all parameters except for the�xed e¤ects, and the mean group estimator (MGE) proposed by Pesaran and Smith (1995), which allows forheterogeneity of all parameters. It exploits the statistical power o¤ered by the panel through long-run ho-mogeneity, while admitting short-run heterogeneity. To test the validity of long-run homogeneity we employa Hausman (1978) test (denoted h test in the paper) on the di¤erence between MG and PMG estimates oflong-run coe¢ cients.Note that the PMG estimator, while able to deal with non-stationary data (Pesaran, Shin and Smith,

1999: section 4.2), does not preclude stationary data in estimation (Pesaran, Shin and Smith, 1999: section4.1).

9 Appendix: Panel Unit Root Tests

We performed the Im-Pesaran-Shin (2003) and Levin�Lin�Chu (2002) Panel Unit Root Tests, on the fol-lowing variables: TFP Productivity Growth (TFP); growth in labor productivity (Y/L); the e¤ective rateof protection (ERP), the nominal tari¤ rate (Nom Tar), and distance from technological frontier (M). Laglength is selected on the basis of information criteria. For all variables test statistics reject the null ofnon-stationarity.38

Results are reported in Table A-1.

References

[1] Aghion, P., Bloom, N., Blundell, R., Gri¢ th, R., and Howitt, P. (2005). �Competition and Innovation:An Inverted-U Relationship�, Quarterly Journal of Economics, 120(2), pp. 701-728.

38The sole exception is the Levin � Lin �Chu test statistic for M, which only rejects the null at the 14% level. However,Im-Pesaran-Shin con�rms the stationarity property of the variable.

19

[2] Aghion, P., Braun, M., and Fedderke, J.W. (2008). �Competition and Productivity Growth in SouthAfrica�, Economics of Transition, 16(4), pp. 741-68.

[3] Aghion, P. and Gri¢ th, R. (2005). �Competition and Growth: Reconciling Theory and Evidence�,Cambridge MA., MIT Press.

[4] Aghion, P. and Howitt, P. (1992). �A Model of Growth through Creative Destruction�, Econometrica,60(2), pp. 323-351.

[5] Aghion, P. and Howitt, P. (2009).Understanding Economic Growth, Cambridge MA., MIT Press.

[6] Alesina, A, Spolaore, E. and Wacziarg, R. (2005) �Trade, Growth and the Size of Countries�in Aghion,P and .Durlauf, S.N (eds.) Handbook of Economic Growth 1B, Elsevier, Chapter 23, pp. 1499-1542.

[7] Arellano, M., and Bond, S.R. (1991). �Some Tests of Speci�cation for Panel Data: Monte Carlo Evidenceand an Application to Employment Equations�, Review of Economic Studies, 58, pp. 277-97.

[8] Blundell, R., and Bond, S.R. (1998). �Initial conditions and moment restrictions in dynamic panel datamodels�Journal of Econometrics, 87, pp. 115-43.Bond, S.R., (2002) Dynamic Panel Data Models: AGuide to Micro Data Methods and Practice, cemmap working paper CWP09/02.

[9] Bond, S.R. (2002). �Dynamic Panel Data Models: A Guide to Micro Data Methods and Practice�,Cemmap Working Paper 09/02, http://www.cemmap.ac.uk/wps/cwp0209.pdf.

[10] Bustos, P. (2011). �Trade Liberalization, Exports and Technology Upgrading: Evidence on the impactof MERCOSUR on Argentinian Firms�American Economic Review, 101 (1), pp. 304-340.

[11] Coe, D.T., and Helpman, E. (1995). �International R&D Spillovers�, European Economic Review, 39(5),859-87.

[12] Edwards, L. (2005). �Has South Africa Liberalised its Trade?�, South African Journal of Economics,73(4), pp. 754-75.

[13] Edwards, L. and Schöer, V. (2002). �Measures of competitiveness: A dynamic approach to South Africa�strade performance in the 1990�s�, South African Journal of Economics, 70, pp. 1008-1046.

[14] Fedderke, J.W. (2006) �Technology, Human Capital and Growth: evidence from a middle income countrycase study applying dynamic heterogeneous panel analysis�in South African Reserve Bank, Banco deMexico and The People�s Bank of China (eds.) Economic Growth, Proceedings of a G20 seminar heldin Pretoria, South Africa, on 4-5 August 2005.

20

[15] Fedderke J.W., Kularatne, C., and Mariotti, M. (2007) �Mark-up Pricing in South African Industry�Journal of African Economies, 16(1), pp. 28-69.

[16] Fedderke, J.W., and Naumann, D.(2011) �An Analysis of Industry Concentration in South AfricanManufacturing, 1972-2001�,Applied Economics, Volume 43 (22), pp. 2919-2939.

[17] Fedderke, J.W., and Szalontai, G. (2009) �Industry Concentration in South African Manufacturing:Trends and Consequences�Economic Modelling, 26(1), pp. 241-50.

[18] Fedderke, J.W., and Vaze, P. (2001) �The Nature of South Africa�s Trade Patterns�, South AfricanJournal of Economics, 69(3), pp. 436-73.

[19] Hall, R.E. (1990) �The Invariance Properties of Solow�s Productivity Residual�, in P. Diamond (Ed..)Growth, Productivity, Unemployment, Cambridge MA: MIT Press.

[20] Keller, W. (2002). �Geographic Localization of International Technological Di¤usion�, American Eco-nomic Review, 92(1), pp. 170-92.

[21] Im, K.S., Pesaran, M.H. and Shin, Y. (2003). �Testing for Unit Roots in Heterogeneous Panels�Journalof Econometrics, 115, pp. 53-74.

[22] Levin, A., Lin, C-F. and Chia-Shang, J.C. (2002). �Unit Root Tests in Panel Data: Asymptotic andFinite Sample Properties�, Journal of Econometrics, 108, pp. 1-24.

[23] Melitz, M.J. (2003). �The Impact of Trade on Intra-Industry Reallocations and Aggregate IndustryProductivity�, Econometrica, 71, pp. 1695-1725.

[24] Pesaran, M.H. and Smith, R.P. (1995). �Estimating Long-Run Relationships from Dynamic Heteroge-neous Panels�, Journal of Econometrics, Vol. 68, pp. 79-113.

[25] Pesaran, M.H., Shin, Y. and Smith, R. (1999). �Pooled Mean Group Estimation of Dynamic Heteroge-neous Panels�, Journal of the American Statistical Association, Vol. 94, pp. 621-634.

[26] Rangasamy, L. and Harmse, C.(2003). �The Extent of Trade Liberalisation in the 1990s: Revisited�,South African Journal of Economics, 71(4): pp. 705-728.

[27] Rodrik, D. (2008). �Understanding South Africa�s Economic Puzzles�, Economics of Transition, 16(4),pp. 769-97.

[28] Roeger, W. (1995) �Can Imperfect Competition Explain the Di¤erence Between Primal and Dual Pro-ductivity Measures? Estimates for US Manufacturing�, Journal of Political Economy, 103, pp.316-30.

[29] Sachs, J.D., and Warner, A. (1995). �Economic Reform and the Process of Global Integration�, BrookingsPapers on Economic Activity, Economic Studies Program, The Brookings Institution, 26(1, 25th A),pages 1-118.

[30] Tre�er, D. (2004) �The Long and Short of the Canada-US Free Trade Agreement�,American EconomicReview, 94(4), pp. 870-95.

21