Alcohol as a Risk Factor for Type 2DiabetesA systematic review and meta-analysis

DOLLY O. BALIUNAS, MSC1,2

BENJAMIN J. TAYLOR, MSC1,2

HYACINTH IRVING, MA1

MICHAEL ROERECKE, MSC1,2

JAYADEEP PATRA, PHD1

SATYA MOHAPATRA, PHD1

JURGEN REHM, PHD1,2,3

OBJECTIVE — To clarify the dose-response relationship between alcohol consumption andtype 2 diabetes.

RESEARCH DESIGN AND METHODS — A systematic computer-assisted and handsearch was conducted to identify relevant articles with longitudinal design and quantitativemeasurement of alcohol consumption. Adjustment was made for the sick-quitter effect. We usedfractional polynomials in a meta-regression to determine the dose-response relationships by sexand end point using lifetime abstainers as the reference group.

RESULTS — The search revealed 20 cohort studies that met our inclusion criteria. A U-shaped relationship was found for both sexes. Compared with lifetime abstainers, the relativerisk (RR) for type 2 diabetes among men was most protective when consuming 22 g/day alcohol(RR 0.87 [95% CI 0.76–1.00]) and became deleterious at just over 60 g/day alcohol (1.01[0.71–1.44]). Among women, consumption of 24 g/day alcohol was most protective (0.60[0.52–0.69]) and became deleterious at about 50 g/day alcohol (1.02 [0.83–1.26]).

CONCLUSIONS — Our analysis confirms previous research findings that moderate alcoholconsumption is protective for type 2 diabetes in men and women.

Diabetes Care 32:2123–2132, 2009

D iabetes is a major public healthproblem with long-term conse-quences including loss of vision;

kidney failure; amputations; gastrointes-tinal, genitourinary, and cardiovascularsymptoms; and sexual dysfunction (1).Several factors increase the risk of diabe-tes, including being overweight, lack ofphysical activity, and family history of di-abetes (2). There is growing consensusthat alcohol consumption is an influenc-ing factor. The biological mechanism isuncertain, but there are several factorsthat may explain the relationship, includ-ing increases in insulin sensitivity aftermoderate alcohol consumption (3),changes in levels of alcohol metabolites(4), increases in HDL cholesterol concen-

trations (5), or via the anti-inflammatoryeffect of alcohol (6).

The exact nature of the dose-responserelationship remains unclear (7). Severalreviews have suggested a U-shaped rela-tionship or a protective effect of moderateconsumption with some question aboutthe effect of higher levels of alcohol con-sumption (7–10). However, these reviewsare narrative. Two quantitative reviewshave been conducted. Carlsson et al. (11)categorized consumption into predeter-mined moderate- and high-consumptiongroups and used current abstainers or lowconsumers as the reference group. In theiranalysis, moderate consumption was as-sociated with a 30% reduced risk of dia-betes among men (relative risk [RR] 0.72

[95% CI 0.67–0.77]) and women (0.68[0.61–0.75]). The risk associated withhigh consumption was described as beingunclear. In the other meta-analysis, inwhich alcohol consumption was treatedcontinuously, a U-shaped relationshipwas found for both men and women, witha more protective effect of moderate con-sumption observed for women (12).However, in both of these reviews, thereference group was composed of formerdrinkers and lifetime abstainers. Becauseformer drinkers may be inspired to ab-stain due to health concerns, they mayactually be at increased risk of developingdiabetes, known as the sick-quitter effect(13). Our goal, therefore, was to examinethe relationship between alcohol con-sumption and the risk of type 2 diabetesby conducting a meta-analysis that uses aflexible modeling approach and that, forthe first time, uses lifetime abstention asthe reference category.

RESEARCH DESIGN ANDMETHODS — Non–insulin-depen-dent diabetes (type 2) was the outcome.Although this outcome can be measuredin various ways, the current World HealthOrganization (WHO) clinical diagnosticcriteria were considered the gold standardfor this meta-analysis. These criteria de-fine diabetes by a fasting plasma glucose(FPG) level �7.0 mmol/l or a venousplasma concentration �11.1 mmol/l 2 hafter a 75-g oral glucose challenge (14).The American Diabetes Association(ADA) also includes as sufficient criteriasymptoms of hyperglycemia and a ran-dom plasma glucose concentration�11.1 mmol/l (1). The criteria, however,changed in 1996 (ADA) and 1999(WHO) from an FPG �7.8 mmol/l.

Articles were found via a search of thefollowing sources: Medline (via OVIDand PubMed), the Cumulative Index toNursing and Allied Health Literature(CINAHL), the Excerpta Medica Database(EMBASE), CAB Abstracts, World HealthOrganization Library Information System(WHOLIS), the System for Informationon Grey Literature in Europe (SIGLE), theAlcohol and Alcohol Problems Science

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

From the 1Centre for Addiction and Mental Health, Toronto, Ontario, Canada; the 2Dalla Lana School ofPublic Health, University of Toronto, Toronto, Ontario, Canada; and the 3Technische UniversitatDresden, Dresden, Germany.

Corresponding author: Dolly Baliunas, dolly.baliunas@utoronto.ca.Received 6 February 2009 and accepted 22 July 2009.DOI: 10.2337/dc09-0227© 2009 by the American Diabetes Association. Readers may use this article as long as the work is properly

cited, the use is educational and not for profit, and the work is not altered. See http://creativecommons.org/licenses/by-nc-nd/3.0/ for details.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be herebymarked “advertisement” in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

R e v i e w s / C o m m e n t a r i e s / A D A S t a t e m e n t sM E T A - A N A L Y S I S

care.diabetesjournals.org DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 2123

(ETOH), Web of Science, and the AlcoholIn Moderation (AIM; an alcohol industrydatabase) databases. The databases weresearched for reports published from 1January 1980 to 31 January 2008, withthe following keywords: alcohol or etha-nol, diabetes, case-control or cohort orprospective, and risk. Animal studies,commentaries, editorials, letters, and re-view articles were excluded. No languagerestriction was applied. A simplifiedsearch using the terms alcohol or ethanolas well as diabetes was used for WHOLISand SIGLE, which could not support thecomplex strategy. AIM is not a searchableindex, but a selective list of articles wasreviewed.

The results of the search are shown inFig. 1. The strategy resulted in 1,615 hitsafter the removal of duplicates. The ab-stracts or complete publications were re-viewed and excluded if they contained noindication of a measure of association be-tween alcohol and either morbidity ormortality due to diabetes (n � 1,561), ifthe measure was cross-sectional (n � 8),

or if fewer than three levels of alcohol con-sumption (i.e., no dose-response data)were reported (n � 5). For all non-English articles, the authors were able toascertain eligibility. Of the 41 publica-tions remaining, 3 (student theses) couldnot be obtained and 6 did not containenough information to quantify, for eachalcohol group, consumption in grams perday and/or the measure of association.Two publications reporting consumptionusing continuous measures could not berevised to provide a measure of associa-tion. In cases where more than one pub-lication was generated from a given study,the most comprehensive analysis wasused. After exclusion of 6 such duplicates,24 articles remained. The references ofthese and relevant review articles werechecked for additional publications, and25 articles were identified. One remainedafter applying the same criteria describedabove. Combining the database and handsearches, 25 articles were identified fordata abstraction. However, four were ex-cluded because the number of cases of

diabetes per alcohol exposure group,which was required for the analysis, wasnot reported. One was excluded becausethe measure was for a recent drinking oc-casion and not a typical day. Thus, 20articles were included in the analysis(15–34).

Data abstracted included descriptorsof study design in addition to measures ofassociation. Alcohol consumption wasconverted to grams per day (if not origi-nally reported as such). For studies thatreported ranges of alcohol consumptionfor the categories, the midpoint was used.When the highest category was openended, three-quarters the width of theprevious interval was added to the lowerlimit. Where consumption was reportedin drinks and not grams, the grams ofpure alcohol equivalent described in thearticle, if stated, was used as a conversionfactor; if not stated, conversion was basedon typical drink sizes of the country (35).In one case, due to ambiguity over a sus-pected misprint, the authors were con-tacted via e-mail for clarification (16). 1.

Figure 1—Flow diagram of literature search for the relationship between alcohol consumption and risk of type 2 diabetes.

Alcohol and type 2 diabetes

2124 DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 care.diabetesjournals.org

The measures of effect abstractedwere hazard ratios, odds ratios, and RRsbut are referred to hereafter using thegeneral term RR. Where RRs were notspecifically presented but sufficient infor-mation was available, they were calcu-lated. Where some consumption groupwas used as the reference, the RRs werereformulated to make abstainers the ref-erence group. For studies for which vari-ous estimates including more or lesscovariates were reported, and a choice ex-isted as to which to include, those thatcontrolled for the most potential con-founders not on the causal pathway werechosen. In five cases, crude measureswere used when no other measure wasavailable. In one study, where males andfemales were analyzed together, the RRswere applied to both men and women.

The studies contained two types ofreference groups: lifetime and current al-cohol abstainers. In order to use lifetimeabstention as the reference, RRs for thosestudies that had current abstention as thereference were adjusted; for each sex,studies that had both current and lifetimeabstainers were used to determine theoverall prevalence of former drinkersamong current abstainers and the RR offormer drinkers relative to lifetime ab-stainers (weighted by precision). The RRfor current abstention was then re-

weighted by the overall prevalence andthe RR calculated above and then multi-plied by the dose-specific RRs.

The literature search, review, and ab-straction were carried out by D.B. To en-sure accuracy in abstraction, a limiteddouble entry was performed by B.T. andthe results compared. Both authorsagreed on 5/5 articles reviewed for inclu-sion/exclusion and 605/664 data pointsabstracted over 10 articles. Where dis-agreements existed, both authors re-viewed the materials together until aconsensus was reached.

To assess potential publication bias,separate funnel plots were drawn for con-sumption �20 g/day and �20 g/day. Inthe absence of a known cutoff for decreas-ing and increasing diabetes risk, thesecategories were chosen because they corre-spond to the WHO’s low-risk drinkinglevel guidelines (36). The estimates wereprepooled using the inverse variance–weighted method because funnel plotmethodology assumes one overall RR perarticle. We assessed publication bias us-ing the tests of Egger and Begg (37,38).The Q test was used to assess the presenceof heterogeneity (39). Additionally, the I2

statistic was used to measure inconsis-tency across studies and represents theproportion of variability in the estimates

that is due to between-study variation(40).

We conducted the meta-regressionusing linear, first-order, and second-order fractional polynomial regression ofthe inverse variance–weighted data to es-timate a best-fitting curve (41). Best-fitcurves were assessed using decreased de-viance compared with the referencemodel. Comparisons of curves to deter-mine the best fit were made using a �2

distribution (41).Two sensitivity analyses were con-

ducted.Because themeta-analysisbyKop-pes et al. (12) suggested that studies thatused self-report of diabetes status re-ported a more protective effect than thosethat tested for diabetes, we assessedwhether the model varied by self-reported outcome. Also, we abstracted es-timates that did not adjust for potentialintermediates on the causal pathway toavoid over adjustment. However, thismay have resulted in using estimates thatwere not sufficiently adjusted. Therefore,we repeated the analysis using the most-adjusted estimates available, selectingmodels for men and women based on fitand comparability with the main analysis.All analyses were conducted using Statasoftware (version 10.1; StatCorp, CollegeStation, TX).

Figure 2—Scatter plot of the RR estimates of type 2 diabetes reported in the 20 studies included in the analyses. Each study provides more than oneRR estimate. The area of each circle is proportional to the precision of the RR estimate.

Baliunas and Associates

care.diabetesjournals.org DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 2125

Tab

le1—

Cha

ract

eris

tics

of20

coho

rtst

udie

sin

clud

edin

the

anal

yses

Firs

tau

thor

(yea

rof

publ

icat

ion)

Loca

tion

Sex

Age

atba

selin

e(y

ears

)n

Follo

w-u

pA

lcoh

olco

nsum

ptio

n

Dia

bete

sde

finit

ion

Adj

ustm

ents

Dur

atio

n(y

ears

)E

vent

s(n

o.)

No.

cate

gori

esD

escr

ipti

on

Aja

ni(2

000)

U.S

.M

en40

–85

20,9

5112

.776

66

No.

ofdr

inks

:rar

ely/

neve

r,1–

3/m

onth

,1/w

eek,

2–4/

wee

k,5–

6/w

eek,

or�

1/da

y

Self-

repo

rt(p

hysi

cian

part

icip

ants

)A

ge,B

MI,

trea

tmen

tas

sign

men

t,sm

okin

gst

atus

,and

phys

ical

acti

vity

Beul

ens

(200

5)T

heN

ethe

rlan

dsW

omen

49–7

016

,330

6.2

760

7g/

wee

k:te

etot

aler

,0–4

.9,

5–29

.9,3

0–69

.9,7

0–13

9.9,

140–

209.

9,or

�21

0

Self-

repo

rtan

d/or

urin

ary

gluc

ose

stri

pan

d/or

Dut

chre

gist

erof

hosp

ital

disc

harg

edi

agno

sis

Age

and

BMI

Car

lsso

n(2

003)

Finl

and

Men

Not

repo

rted

10,9

7019

.427

74

g/da

y:ab

stai

ners

,�5.

0,5.

0–29

.9,o

r�

30.0

Link

age

toFi

nnis

hna

tion

alho

spit

alA

gean

dBM

I

Wom

enN

otre

port

ed11

,808

19.8

297

4g/

day:

abst

aine

rs,�

5.0,

5.0–

19.9

,or

�20

.0di

scha

rge

regi

ster

and

nati

onal

drug

regi

ster

;pa

tien

tsw

hore

ceiv

edon

lyor

alm

edic

atio

n,di

etar

yth

erap

y,or

insu

linas

anad

juva

nttr

eatm

ent

tohy

pogl

ycem

icdr

ugs

wer

eca

tego

rize

das

type

2di

abet

icC

onig

rave

(200

1)U

.S.

Men

40–7

546

,892

10.9

1,57

17

g/da

y:0,

0.1–

4.9,

5.0–

9.9,

10.0

–14.

9,15

.0–2

9.9,

30.0

–49.

9,or

�50

.0

Self-

repo

rt;c

rite

ria:

any

of1)

one

orm

ore

clas

sic

sym

ptom

sw

ith

FPG

�7.

8m

mol

/l,no

nfas

ting

gluc

ose

�11

.1m

mol

/l,or

OG

TT

�11

.1m

mol

/l,2)

elev

ated

plas

ma

gluc

ose

leve

lson

two

diff

eren

toc

casi

ons,

3)hy

pogl

ycem

iatr

eatm

ent

(but

not

type

1di

abet

ic)

Age

and

BMI

deV

egt

(200

2)T

heN

ethe

rlan

dsM

enan

dw

omen

com

bine

d50

–75

1,32

2N

otre

port

ed24

13

g/da

y:0,

�10

,or

�10

FPG

�7.

0m

mol

/lor

OG

TT

�11

.1m

mol

/l;or

subj

ects

alre

ady

bein

gtr

eate

dfo

rdi

abet

es

Age

and

sex

Djo

ussé

(200

7)U

.S.

Men

64–9

51,

909

6.3

109

4D

rink

s/w

eek:

neve

r,fo

rmer

,�1,

1–7,

or�

7U

seof

insu

linor

oral

hypo

glyc

emic

agen

ts;

Age

,BM

I,ed

ucat

ion,

and

smok

ing

stat

usW

omen

63–9

52,

746

6.3

125

4D

rink

s/w

eek:

neve

r,fo

rmer

,�1,

1–7,

or�

7or

FPG

�7.

0m

mol

Alcohol and type 2 diabetes

2126 DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 care.diabetesjournals.org

Tab

le1—

Con

tinu

ed

Firs

tau

thor

(yea

rof

publ

icat

ion)

Loca

tion

Sex

Age

atba

selin

e(y

ears

)n

Follo

w-u

pA

lcoh

olco

nsum

ptio

n

Dia

bete

sde

finit

ion

Adj

ustm

ents

Dur

atio

n(y

ears

)E

vent

s(n

o.)

No.

cate

gori

esD

escr

ipti

on

Hod

ge(2

006)

Aus

tral

iaM

en40

–69

12,2

144

179

5g/

day:

neve

rdr

inke

r,fo

rmer

drin

ker,

�10

,10

–20,

20–3

0,or

�30

Part

icip

ants

mai

led

ques

tion

nair

eco

veri

ngdi

agno

sis

ofdi

abet

es,

Age

,BM

I,co

untr

yof

birt

h,di

etar

ygl

ucos

ein

take

,die

tary

ener

gyW

omen

40–6

919

,208

418

34

g/da

y:ne

ver

drin

ker,

form

erdr

inke

r,�

10,

10–2

0,or

�20

wit

hve

rific

atio

nfr

ompr

imar

yca

reph

ysic

ian

inta

ke,a

ndw

aist

-to-

hip

rati

o

Hol

broo

k(1

990)

U.S

.M

en40

–79

221

1431

4g/

wee

k:no

ndri

nker

,0.1

–84

.3,8

4.4–

176.

0,or

�17

6.1

FPG

�14

0m

g/dl

,OG

TT

�20

0m

g/dl

(11.

1m

mol

/l),o

rse

lf-re

port

Age

Wom

en40

–79

303

1444

4g/

wee

k:no

ndri

nker

,0.1

–41

.3,4

1.4–

117.

4,or

�11

7.5

ofdi

abet

esdi

agno

sis

bya

phys

icia

n

Hu

(200

6)Fi

nlan

dM

en35

–74

10,1

8813

.451

73

g/w

eek:

none

,1–1

00,o

r�

100

Link

age

ofin

foin

Finn

ish

nati

onal

hosp

ital

Age

;BM

I;st

udy

year

;ed

ucat

ion;

syst

olic

Wom

en35

–74

11,1

9713

.444

73

g/w

eek:

none

,1–1

00,o

r�

100

disc

harg

ere

gist

eran

dna

tion

aldr

ugre

gist

erbl

ood

pres

sure

;bre

ad,

vege

tabl

e,fr

uit,

saus

age,

coff

ee,a

ndte

aco

nsum

ptio

n;sm

okin

gst

atus

;and

phys

ical

acti

vity

Kao

(200

1)U

.S.

Men

45–6

45,

423

5.3

547

6D

rink

s/w

eek:

lifet

ime

abst

aine

r,fo

rmer

drin

ker,

�1,

1.1–

7,7.

1–14

,14.

1–21

,or

�21

Any

of1)

FPG

�7.

0m

mol

,2)

nonf

asti

nggl

ucos

e�

11.1

mm

ol,

3)cu

rren

tus

eof

diab

etic

med

icat

ions

,

Non

e

Wom

en45

–64

6,83

85.

456

96

Dri

nks/

wee

k:lif

etim

eab

stai

ner,

form

erdr

inke

r,�

1,1.

1–7,

7.1–

14,1

4.1–

21,o

r�

21

or4)

posi

tive

resp

onse

to“H

asa

doct

orev

erto

ldyo

uth

atyo

uha

ddi

abet

es?”

Lee

(200

3)K

orea

Men

25–5

54,

055

483

5g/

wee

k:ab

stai

ner,

1–90

,91

–180

,181

–360

g,or

�36

0g

Seru

mfa

stin

ggl

ucos

eco

ncen

trat

ion

�12

6m

g/dl

orta

king

diab

etes

med

icat

ion

Non

e

Lee

(200

4)U

.S.

Wom

en55

–69

35,6

9811

1,92

13

g/da

y:0,

1–14

,or

�15

Self-

repo

rtN

one

Mei

sing

er(2

002)

Ger

man

yM

en35

–74

3,05

27.

512

83

g/da

y:0,

0.1–

39.9

,or

�40

.0Se

lf-re

port

diag

nosi

sor

taki

ngan

tidi

abet

esA

ge,B

MI,

and

surv

ey

Wom

en35

–74

3,11

47.

685

3g/

day:

0,0.

1–19

.9,o

r�

20.0

med

icat

ions

Baliunas and Associates

care.diabetesjournals.org DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 2127

Tab

le1—

Con

tinu

ed

Firs

tau

thor

(yea

rof

publ

icat

ion)

Loca

tion

Sex

Age

atba

selin

e(y

ears

)n

Follo

w-u

pA

lcoh

olco

nsum

ptio

n

Dia

bete

sde

finit

ion

Adj

ustm

ents

Dur

atio

n(y

ears

)E

vent

s(n

o.)

No.

cate

gori

esD

escr

ipti

on

Stam

pfer

(198

8)U

.S.

Wom

en34

–59

85,0

514

524

5g/

day:

0,�

1.5,

1.5–

4.9,

5.0–

14.9

,or

�15

.0Se

lf-re

port

byqu

esti

onna

ire,

then

supp

lem

enta

ryqu

esti

onna

ire

rega

rdin

gcl

assi

csy

mpt

oms

wit

hFP

G�

140

mg/

dlor

rand

ompl

asm

agl

ucos

e�

200

mg/

dl(o

rat

leas

ttw

oel

evat

edpl

asm

agl

ucos

ele

vels

ifno

sym

ptom

s)

Age

,BM

I,an

dca

lori

cin

take

Stro

dl(2

006)

Aus

tral

iaW

omen

70–7

48,

896

323

13

Dri

nks/

day:

none

orra

rely

,1–2

,or

�3

Ask

edw

heth

era

doct

orha

dto

ldth

emth

eyha

da

diag

nosi

sof

diab

etes

Non

e

Tsu

mur

a(1

999)

Japa

nM

en35

–61

6,36

29.

745

65

ml/d

ay:a

bsta

iner

,0.1

–19

.0,1

9.1–

29.0

,29.

1–50

.0,o

r�

50.1

FPG

�7.

8m

mol

/l,O

GT

T�

11.1

mm

ol/l,

orFP

G�

7.0

mm

ol/l

Age

,BM

I,sm

okin

gst

atus

,lei

sure

tim

eph

ysic

alac

tivi

ty,

pare

ntal

diab

etes

,and

FPG

leve

lW

aki(

2004

)Ja

pan

Men

40–5

912

,913

Not

repo

rted

703

4g/

day:

abst

aine

ran

din

freq

uent

drin

ker,

�23

.0,2

3.1–

46.0

,or

�46

.1

Self-

repo

rt�H

asa

doct

orev

erto

ldyo

uth

atyo

uha

vedi

abet

es?�

Age

,BM

I,sm

okin

gst

atus

,fam

ilyhi

stor

yof

diab

etes

,lei

sure

tim

eph

ysic

alac

tivi

ty,

Wom

en40

–59

15,9

80N

otre

port

ed48

04

g/da

y:ab

stai

ner

and

infr

eque

ntdr

inke

r,�

4.9,

5.0–

11.5

,or

�11

.6

and

hype

rten

sion

Wan

nam

ethe

e(2

002)

U.K

.M

en40

–59

5,22

116

.819

85

Uni

ts/w

eek:

none

,�1,

1–15

,15–

42,o

r�

42D

iagn

osis

ofdi

abet

esno

tac

cept

edon

basi

sof

self-

com

plet

edqu

esti

onna

ire

unle

ssco

nfirm

edin

prim

ary

care

reco

rds

Age

and

BMI

Alcohol and type 2 diabetes

2128 DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 care.diabetesjournals.org

RESULTS — Characteristics of the 20studies included are presented in Table Intotal, data from 477,200 individuals, in-cluding 12,556 incident cases of diabetes,were included. Six studies included onlymen, five only women, eight both menand women (separately), and one menand women combined together. All wereprospective cohorts. Adjustment for con-founders varied. All but four adjusted forage at minimum; for these, only crudemeasures were available once those thatadjusted for factors on the causal pathwaywere excluded. Diabetes ascertainmentvaried from self-report and data linkage tonational registers to an oral glucose toler-ance test (OGTT), the current clinicalgold standard.

Figure 2 shows the RRs for the indi-vidual studies in the meta-analyses forwomen and men, respectively. Markedheterogeneity was found for both women(Q � 69.97, P � 0.004, I2 � 40%, 95%CI 13–58%) and men (Q � 79.58, P �0.008, I2 � 35%, 95% CI 8–53%). Ran-dom-effects models were used for all sub-sequent analyses . No significantpublication bias was detected.

Among women, the best-fittingmodel was the second-degree model withpowers 0.5 and 3 (and function �1x0.5 ��2x3) (P � 0.001). Among men, the best-fitting model was the second-degree frac-tional polynomial with powers 1 and 1(and function �1x � �2xlnx) (P � 0.007).

Figure 3 displays the relationship be-tween the risk of alcohol consumptionand type 2 diabetes among women andmen. For both sexes, the relationship wasU shaped. For women, the protective ef-fect was greatest at the 24 g/day level, witha risk reduction of 40% compared withlifetime abstainers (95% CI 0.52–0.69).Alcohol consumption remained protec-tive until just under 50 g/day. For men,the protective effect of alcohol consump-tion was greatest at 22 g/day, with the riskof diabetes being �0.87 times that of life-time abstainers (95% CI 0.76–1.00), andremained protective until consumption of�60 g/day. Thus, for both women andmen, the protective effect of alcohol con-sumption on incident type 2 diabetes wasgreatest with the consumption of abouttwo drinks per day. Similarly, for bothmen and women, higher levels of con-sumption (above �50 g/day for womenand 60 g/day for men) were no longerprotective but actually increased the riskfor diabetes.

The mode of ascertainment of diabe-tes, self-report versus objective measure-T

able

1—C

onti

nued

Firs

tau

thor

(yea

rof

publ

icat

ion)

Loca

tion

Sex

Age

atba

selin

e(y

ears

)n

Follo

w-u

pA

lcoh

olco

nsum

ptio

n

Dia

bete

sde

finit

ion

Adj

ustm

ents

Dur

atio

n(y

ears

)E

vent

s(n

o.)

No.

cate

gori

esD

escr

ipti

on

Wan

nam

ethe

e(2

003)

U.S

.W

omen

25–4

210

9,70

58.

193

55

g/da

y:lif

elon

gab

stai

ner,

form

erdr

inke

r,�

5.0,

5.0–

14.9

,15.

0–29

.9,o

r�

30.0

Rep

ort

onbi

enni

alqu

esti

onna

ire,

then

supp

lem

enta

ryqu

esti

onna

ire

rega

rdin

gsy

mpt

oms:

befo

re19

971)

�1

clas

sic

sym

ptom

san

dFP

G�

7.8

mm

ol/l

orra

ndom

plas

ma

gluc

ose

�11

.1m

mol

/l,2)

�2

elev

ated

plas

ma

gluc

ose

resu

lts

ondi

ffer

ent

occa

sion

s(F

PG�

7.8

mm

ol/l,

rand

ompl

asm

agl

ucos

e�

11.1

mm

ol/l,

and/

orO

GT

T�

11.1

mm

ol/l

inab

senc

eof

sym

ptom

s),o

r3)

trea

tmen

tw

ith

hypo

glyc

emia

med

icat

ions

Age

Wei

(200

0)U

.S.

Men

30–7

98,

633

614

95

g/w

eek:

abst

aine

r,1–

61.8

,61

.9–1

22.7

,122

.8–

276.

6,or

�27

6.6

FPG

�7.

0m

mol

/lor

hist

ory

ofdi

abet

espl

uscu

rren

tin

sulin

ther

apy

Age

,par

enta

ldia

bete

s,an

dye

ars

offo

llow

-up

Baliunas and Associates

care.diabetesjournals.org DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 2129

ment, did impact the risk relation withvolume of alcohol exposure but only formen; there was no effect for women. Ac-cordingly, we repeated the analyses sepa-rately for self-report versus no self-reportin men and found a linearly decreasingdose-response relationship in the stud-ies with self-report (15,18,27,31),aswell as a model similar to the main anal-ysis in the rest of the studies. The resultin the group based on self report was

mainly influenced by two studies(15,18) that accounted for 81% of theobservations.

The second sensitivity analysis, per-formed using the most adjusted estimatesavailable, resulted in models remarkablysimilar to those in the main analysis forboth sexes. For men, the U-shaped rela-tionship was most protective at a con-sumption of 22 g/day and crossed backover the RR 1 at 62 g/day. For women, the

U-shaped relationship, as with the mainanalysis, was most protective at a con-sumption of 25 g/day and crossed backover to deleterious effect at 51 g/day.

CONCLUSIONS — Our meta-analy-sis confirms the U-shaped relationshipsbetween average amount of alcohol con-sumed per day and risk of incident type 2diabetes among men and women, al-though a more protective effect of moder-

Figure 3—Pooled and fitted RR estimates and 95% CI band. A: The highest single alcohol consumption measure for women was 52.35 g/day, thusx-axis is scaled to 60 g/day. B: Among men, the single highest alcohol consumption measure was 80.04 g/day.

Alcohol and type 2 diabetes

2130 DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 care.diabetesjournals.org

ate consumption was found for women.For women, the protective effect at mod-erate consumption and hazardous effectat higher consumption were both statisti-cally significant. For men, the protectiveeffect was statistically significant, but forhigher consumption the CI did not ex-clude the RR 1.

Previous reviews found a protectiveeffect of moderate alcohol consumptionbut limited evidence for a deleterious ef-fect of heavy consumption. In compari-son to the 2 previous meta-analyses, ouranalysis included 20 cohort studies in to-tal: an additional 6 studies not includedby Koppes et al. and an additional 10 notincluded by Carlsson et al. The meta-analyses by Koppes et al. used a total of 15studies but included 2 that did not meetour inclusion criteria. That by Carlsson etal. included 13 studies, including 4 thatdid not meet our criteria. Thus, althoughthere is substantial overlap in the data in-cluded in the respective analyses, our ar-ticle is the most comprehensive in termsof the amount of data contributing to thesummary estimates.

Meta-analyses are vulnerable to sev-eral biases. Although the gold standard indiabetes ascertainment is the OGTT, theindividual studies ascertained diabetesstatus in various ways including self-report, linkage with national registries,and clinical tests. Additionally, becausethe diagnostic criteria cutoffs changedduring the 1990s, even use of clinical testswas not consistent from study to study. Asa result, some misclassification was likely.It is difficult to gauge the extent or type ofany misclassification, although it is likelyto be nondifferential and hence would at-tenuate any true causal association. In ad-dition to misclassification due to testingdifferences, there is also potential misclas-sification due to the use of self-reportingof diabetes. However, we investigated thepotential role of self-reported outcomesversus clinical tests and found no differ-ence in results for women. Lastly, somemisclassification of alcohol consumptionmay have occurred but would likelypresent as underreporting and not over-reporting of consumption, resulting in ashift of the relationship curve to the left.That is, associations that appear to exist ata given consumption level in our analysiswould, in fact, exist at some higher level.

Alcohol consumption, and the result-ing health effects, is more complex thanmere volume of consumption measuredat one point in time. Though several stud-ies measured alcohol more than once

(18,19,28,30,33), only one study usedmore than one alcohol measurement in itsmain analysis (17). Additional alcoholmeasurements would add weight to thevalidity and relevance to the alcohol mea-sure because it is long-term consumptionthat tends to be of medical and publichealth concern. Additionally, the way inwhich alcohol is consumed (i.e., withmeals or bingeing on weekends) affectsvarious health outcomes (42). Thus, itmay be the case that the risk of diabetesassociated with heavy alcohol consump-tion is due to consumption mainly on theweekend as opposed to the same amountspread over a week. A few studies mea-sured drinking pattern but did notpresent any analysis or did not present itin an analysis that combined it with vol-ume (17,18,21).

Our meta-analysis addresses the sickquitter effect by making lifetime absten-tion the reference category. For diabetes,ignoring the sick quitter effect will tend tooverestimate the benefit of moderate con-sumption and underestimate the risk ofheavy consumption. However, adjustingfor this required some estimate of the pro-portion of abstainers who are lifetime ab-stainers. For men, this was based on threestudies (19,21,24) in which 37.6% of cur-rent abstainers were former drinkers and,for women, four studies (19,21,24,33) inwhich 49.6% of current abstainers wereformer drinkers. These estimates are a po-tential source of bias because the true un-derlying proportion of former drinkersmay be higher or lower. Given the lack ofa valid external estimate, we felt our data-driven approach was reasonable.

The included articles varied in adjust-ment for potential confounders. It was notpossible to include only articles thath ad-justed for the same factors. The sensitivityanalysis in which the most fully adjustedestimates were used resulted in a relation-ship similar to that found in the mainanalysis. This consistency of results addsweight to the validity of our findings.

We found significant heterogeneityamong men and women, which was ex-pected because of the different methodsused in the individual studies. Most stud-ies were conducted in Western countrieswith the exception of three studies of Jap-anese and Korean populations; however,this was not a significant source of heter-ogeneity for men or women (analyses notshown). Study type cannot be a source ofheterogeneity because only cohort studieswere included, though the variable fol-low-up time may have contributed. The

finding of heterogeneity indicates thatthere are additional factors such as pat-tern of consumption that should be con-sidered in future research.

Our findings confirm previous re-search, both individual studies andsummary estimates, of the U-shaped rela-tionship between average alcohol con-sumption and risk of diabetes in bothmen and women. Although the biologicalmechanism responsible for this relation-ship is still a matter of research, severalpossibilities exist including increased in-sulin sensitivity with low levels of alcoholconsumption. These factors, together,add weight to the argument for a causalrole of alcohol consumption in diabetes.Alcohol consumption in men and womenshould thus be limited to moderateamounts, and heavy consumption shouldbe discouraged. Moreover, the balance ofrisk of alcohol consumption on otherdiseases and health outcomes, even atmoderate levels of consumption, mayoutweigh the positive benefits with re-gard to diabetes.

Acknowledgments— This work was finan-cially supported by a small contribution of theGlobal Burden of Disease (GBD) Study to J.R.

No potential conflicts of interest relevant tothis article were reported.

We thank the core group of the Compara-tive Risk Assessment within the GBD 2005Study for Alcohol for their support and com-ments on the general methodology and an ear-lier version of this article: Robin Room, TheoVos, and Rosana Norman.

References1. American Diabetes Association. Diagnosis

and classification of diabetes mellitus. Di-abetes Care 2008;31(Suppl. 1):S55–S60

2. Mayor S. International Diabetes Federationconsensus on prevention of type 2 diabetes.Int J Clin Pract 2007;61:1773–1775

3. Hendriks HF. Moderate alcohol con-sumption and insulin sensitivity: observa-tions and possible mechanisms. AnnEpidemiol 2007;17(Suppl. 5):S40–S42

4. Sarkola T, Iles MR, Kohlenberg-MuellerK, Eriksson CJ. Ethanol, acetaldehyde, ac-etate, and lactate levels after alcohol in-take in white men and women: effect of4-methylpyrazole. Alcohol Clin Exp Res2002;26:239–245

5. Rimm EB, Williams P, Fosher K, CriquiM, Stampfer MJ. Moderate alcohol intakeand lower risk of coronary heart disease:meta-analysis of effects on lipids and hae-mostatic factors. Br Med J 1999;319:1523–1528

6. Imhof A, Froehlich M, Brenner H, Boeing

Baliunas and Associates

care.diabetesjournals.org DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 2131

H, Pepys MB, Koenig W. Effect of alcoholconsumption on systemic markers of in-flammation. Lancet 2001;357:763–767

7. Klatsky AL. Alcohol, cardiovascular dis-eases and diabetes mellitus. PharmacolRes 2007;55:237–247

8. Howard AA, Arnsten JH, Gourevitch MN.Effect of alcohol consumption on diabetesmellitus: a systematic review. Ann InternMed 2004;140:211–219

9. Zilkens RR, Puddey IB. Alcohol and car-diovascular disease: more than one para-dox to consider. Alcohol and type 2diabetes: another paradox? J CardiovascRisk 2003;10:25–30

10. Conigrave KM, Rimm EB. Alchol for theprevention of type 2 diabetes mellitus?Treat Endocrinol 2003;2:145–152

11. Carlsson S, Hammar N, Grill V. Alcoholconsumption and type 2 diabetes: meta-analysis of epidemiological studies indicatesa U-shaped relationship. Diabetologia2005;48:1051–1054

12. Koppes LL, Dekker JM, Hendriks HF,Bouter LM, Heine RJ. Moderate alcoholconsumption lowers the risk of type 2 di-abetes: a meta-analysis of prospective ob-servational studies. Diabetes Care 2005;28:719–725

13. Shaper A, Wannamethee G, Walker M.Alcohol and mortality in British men: ex-plaining the U-shaped curve. Lancet1988;332:1267–1273

14. World health organization. Definition andDiagnosis of Diabetes Mellitus and Inter-mediate Hyperglycemia. Geneva, WorldHealth Org., 2006

15. Ajani UA, Hennekens CH, Spelsberg A,Manson JE. Alcohol consumption andrisk of type 2 diabetes mellitus among USmale physicians. Arch Intern Med 2000;160:1025–1030

16. Beulens JW, Stolk RP, van der SchouwYT, Grobbee DE, Hendriks HF, Bots ML.Alcohol consumption and risk of type 2diabetes among older women. DiabetesCare 2005;28:2933–2938

17. Carlsson S, Hammar N, Grill V, Kaprio J.Alcohol consumption and the incidenceof type 2 diabetes: a 20-year follow-up ofthe Finnish twin cohort study. DiabetesCare 2003;26:2785–2790

18. Conigrave KM, Hu BF, Camargo CA, Jr,Stampfer MJ, Willett WC, Rimm EB. Aprospective study of drinking patterns inrelation to risk of type 2 diabetes amongmen. Diabetes 2001;50:2390–2395

19. Djousse L, Biggs ML, Mukamal KJ, Sisco-

vick DS. Alcohol consumption and type 2diabetes among older adults: the Cardio-vascular Health Study. Obesity 2007;15:1758–1765

20. de Vegt F, Dekker JM, Groeneveld WJA,Nijpels G, Stehouwer CDA, Bouter LM,Heine RJ. Moderate alcohol consumptionis associated with lower risk for incidentdiabetes and mortality: the Hoorn Study.Diabetes Res and Clin Pract 2002;57:53–60

21. Hodge AM, English DR, O’Dea K, GilesGG. Alcohol intake, consumption pat-tern and beverage type, and the risk oftype 2 diabetes. Diabet Med 2006;23:690 – 697

22. Holbrook TL, Barrett-Connor E, WingardDL. A prospective population-based studyof alcohol use and non-insulin-dependentdiabetes mellitus. Am J Epidemiol 1990;132:902–909

23. Hu G, Jousilahti P, Peltonen M, Bidel S,Tuomilehto J. Joint association of coffeeconsumption and other factors to the riskof type 2 diabetes: a prospective study inFinland. Int J Obes 2006;30:1742–1749

24. Kao WH, Puddey IB, Boland LL, WatsonRL, Brancati FL. Alcohol consumptionand the risk of type 2 diabetes mellitus:atherosclerosis risk in communitiesstudy. Am J Epidemiol 2001;154:748–757

25. Lee DH, Ha MH, Kim JH, Christiani DC,Gross MD, Steffes M, Blomhoff R, JacobsDR Jr. Gamma-glutamyltransferase anddiabetes: a 4 year follow-up study. Diabe-tologia 2003;46:359–364

26. Lee DH, Folsom AR, Jacobs DR Jr. Dietaryiron intake and type 2 diabetes incidence inpostmenopausal women: the Iowa Wom-en’s Health Study. Diabetologia 2004;47:185–194

27. Meisinger C, Thorand B, Schneider A,Stieber J, Doring A, Lowel H. Sex differ-ences in risk factors for incident type 2diabetes mellitus: the MONICA Augsburgcohort study. Arch Intern Med 2002;162:82–89

28. Stampfer MJ, Colditz GA, Willett WC,Manson JE, Arky RA, Hennekens CH,Speizer FE. A prospective study of mod-erate alcohol drinking and risk of diabetesin women. Am J Epidemiol 1988;128:549–558

29. Strodl E, Kenardy J. Psychosocial andnon-psychosocial risk factors for the newdiagnosis of diabetes in elderly women.Diabetes Res Clin Pract 2006;74:57–65

30. Tsumura K, Hayashi T, Suematsu C, EndoG, Fujii S, Okada K. Daily alcohol con-sumption and the risk of type 2 diabetesin Japanese men: the Osaka Health Sur-vey. Diabetes Care 1999;22:1432–1437

31. Waki K, Noda M, Sasaki S, Matsumura Y,Takahashi Y, Isogawa A, Ohashi Y, Kad-owaki T, Tsugane S, JPHC Study Group.Alcohol consumption and other risk fac-tors for self-reported diabetes amongmiddle-aged Japanese: a population-based prospective study in the JPHCstudy cohort I. Diabet Med 2005;22:323–331

32. Wannamethee SG, Shaper AG, Perry IJ,Alberti KG. Alcohol consumption and theincidence of type II diabetes. J EpidemiolCommunity Health 2002;56:542–548

33. Wannamethee SG, Camargo CA, Jr, Man-son JE, Willett WC, Rimm EB. Alcoholdrinking patterns and risk of type 2 dia-betes mellitus among younger women.Arch Intern Med 2003;163:1329–1336

34. Wei M, Gibbons LW, Mitchell TL, Ka-mpert JB, Blair SN. Alcohol intake andincidence of type 2 diabetes in men. Dia-betes Care 2000;23:18–22

35. World Health Organization. InternationalGuide for Monitoring Alcohol Consumptionand Related Harm. Geneva, World HealthOrg., 2000

36. Babor TF, Higgin-Biddle JC, Saunders JB,Monteiro MG; World Health Organiza-tion. The Alcohol Use Disorders Identifica-tion Test: Guidelines for Use in PrimaryCare. Geneva, World Health Org., 2001

37. Egger M, Davey Smith G, Schneider M,Minder C. Bias in meta-analysis detectedby a simple, graphical test. Br Med J 1997;315:629–634

38. Begg CB, Mazumdar M. Operating char-acteristics of a rank correlation test forpublication bias. Biometrics 1994;50:1088–101

39. Cochran WG. The combination of esti-mates from different experiments. Bio-metrics 1954;10:101–129

40. Higgins JP, Thompson SG. Quantifyingheterogeneity in a meta-analysis. Stat Med2002;21:1539–1558

41. Royston P. A strategy for modelling theeffect of a continuous covariate in medi-cine and epidemiology. Stat Med 2000;19:1831–1847

42. Rehm J, Gmel G, Sempos CT, Trevisan M.Alcohol-related morbidity and mortality.Alcohol Res Health 2003;27:39–51

Alcohol and type 2 diabetes

2132 DIABETES CARE, VOLUME 32, NUMBER 11, NOVEMBER 2009 care.diabetesjournals.org