Literature Review - Systematic Research 4 Pages

M E T A - A N A L Y S I S

Caffeine intake during pregnancy and adverse birth outcomes: a systematic review and dose–response meta-analysis

Darren C. Greenwood • Natalie J. Thatcher •

Jin Ye • Lucy Garrard • Georgina Keogh •

Laura G. King • Janet E. Cade

Received: 7 February 2014 / Accepted: 12 August 2014 / Published online: 2 September 2014

� Springer Science+Business Media Dordrecht 2014

Abstract Caffeine is commonly consumed during preg-

nancy, crosses the placenta, with fetal serum concentra-

tions similar to the mother’s, but studies of birth outcome

show conflicting findings. We systematically searched

Medline and Embase for relevant publications. We con-

ducted meta-analysis of dose–response curves for associ-

ations between caffeine intake and spontaneous abortion,

stillbirth, preterm delivery, low birth weight and small for

gestational age (SGA) infants. Meta-analyses included 60

unique publications from 53 cohort and case–control

studies. An increment of 100 g caffeine was associated

with a 14 % (95 % CI 10–19 %) increase in risk of spon-

taneous abortion, 19 % (5–35 %) stillbirth, 2 % (-2 to

6 %) preterm delivery, 7 % (1–12 %) low birth weight, and

10 % (95 % CI 6–14 %) SGA. There was substantial het-

erogeneity in all models, partly explained by adjustment

for smoking and previous obstetric history, but not by

prospective assessment of caffeine intake. There was evi-

dence of small-study effects such as publication bias.

Greater caffeine intake is associated with an increase in

spontaneous abortion, stillbirth, low birth weight, and

SGA, but not preterm delivery. There is no identifiable

threshold below which the associations are not apparent,

but the size of the associations are generally modest within

the range of usual intake and are potentially explained by

bias in study design or publication. There is therefore

insufficient evidence to support further reductions in the

maximum recommended intake of caffeine, but mainte-

nance of current recommendations is a wise precaution.

Keywords Caffeine � Pregnancy � Miscarriage � Stillbirth � Preterm birth � Birth weight � Small for gestational age infant � Meta-analysis

Introduction

Caffeine is present in many drinks and foods consumed

during pregnancy, most notably in tea, coffee, colas,

energy drinks and chocolate. Although amounts vary due to

brand and brewing methods, and vary from country to

country, on average one 260 ml (10 oz) mug of coffee

contains around 100 mg caffeine. It is also a constituent of

several common medications available over the counter or

on prescription [1]. Given its widespread presence, most

pregnant women consume at least small amounts. How-

ever, its effects on the developing fetus are still not fully

understood [2].

Animal studies suggest possible adverse effects on

reproductive outcomes, including fetal growth, but may not

be relevant for humans because caffeine metabolism varies

greatly across species [2, 3]. In humans, studies have

reached mixed conclusions, in part because of difficulties

measuring caffeine intake, but also because of other clin-

ical influences on fetal growth and birth outcomes [4–6].

Electronic supplementary material The online version of this article (doi:10.1007/s10654-014-9944-x) contains supplementary material, which is available to authorized users.

D. C. Greenwood (&) � J. Ye � L. Garrard � G. Keogh � L. G. King

Division of Epidemiology and Biostatistics, School of Medicine,

University of Leeds, Leeds LS2 9JT, UK

e-mail: [email protected]

N. J. Thatcher

Food Standards Agency, Aviation House, 125 Kingsway,

London WC2B 6NH, UK

J. E. Cade

Nutritional Epidemiology Group, School of Food Science and

Nutrition, University of Leeds, Leeds LS2 9JT, UK

123

Eur J Epidemiol (2014) 29:725–734

DOI 10.1007/s10654-014-9944-x

Only one randomised controlled trial of caffeine reduction

during pregnancy has been conducted to date, concluding

that ‘‘moderate reductions’’ in caffeine intake (of around

200 mg) do not substantially alter birthweight or length of

gestation [7]. Given that the intervention came after the 1st

trimester, did not investigate the more important outcome

of spontaneous abortion, and only intervened on one source

of caffeine (coffee), questions over the safe consumption of

caffeine in pregnancy remain.

Because of the possible associations with restricted fetal

growth, birth defects, miscarriage and stillbirth in humans,

guidance in several countries, including the US and UK,

has continued to consider it a wise precaution to limit

caffeine intake to less than 200 mg immediately before and

during pregnancy [2, 8–10]. In the UK, guidance has

recently changed to reduce the recommended maximum

intakes to this level [11]. However, because of the differing

conclusions reached by several major studies, the strength

of any association, and the possibility of a threshold effect,

has not been fully quantified [2, 8, 12].

To inform guidance, and in the light of recent large

prospective studies, we aimed to pool information from the

different observational studies that present information on

the association between caffeine and the adverse birth

outcomes of spontaneous abortion, stillbirth, preterm birth,

low birth weight and small for gestational age. We aimed

to avoid problems present in many systematic reviews of

observational studies that compare extreme categories by

estimating the dose–response slope over all categories of

intake [13]. In particular, we aimed to quantify the degree

of any association and identify any possible threshold

effects through modelling nonlinear dose–response curves.

Methods

Data sources and searches

We conducted a comprehensive systematic literature

search covering all case–control and cohort studies pro-

viding evidence on dietary caffeine intake and adverse

pregnancy outcomes, including spontaneous abortion or

miscarriage, stillbirth, preterm birth, low birth weight and

small for gestational age (SGA) infants. We searched

MEDLINE and EMBASE online databases for all studies

published in any language up to 15th May 2014 (detailed

search strategy in online table 1). We also hand searched

the reference lists of included studies and relevant review

articles. The guidelines for conducting meta-analysis of

observational studies in epidemiology were used through-

out the design, conduct, analysis, and reporting of this

review [14]. A detailed protocol for this review was

produced, but is not yet available to download. Instead,

methods are provided in detail in this paper.

Study selection

We screened titles and abstracts to remove publications

when it was immediately apparent they were not relevant,

such as editorials and single case-study reports. We

obtained full-text versions of potentially relevant articles.

The process of identifying relevant articles was conducted

independently by five members of the review team (DCG,

JY, LG, GK, LGK). The first author made the final decision

where there were any differences. Only case–control and

cohort studies were eligible for inclusion in the review,

including nested case–control and case-cohort studies.

Inclusion criteria were studies based on dietary assess-

ment of maternal caffeine intake during pregnancy (not

before), published in any language, with assessment of

caffeine or coffee with more than two categories of expo-

sure, with outcomes including at least one of spontaneous

abortion (fetal loss under 24 weeks gestation or less),

stillbirth (fetal loss after 24 weeks gestation), preterm birth

(before 37 weeks), low birth weight (\2500 g), or small for gestational age infant (defined as a weight below the

10th percentile for a given gestational age), as an outcome.

Studies had to include some estimate of relative risk (RR)

with a measure of uncertainty such as 95 % CIs.

Where results from the same study were presented in

several papers, the results based on the larger sample, with

the most complete assessment of caffeine intake (i.e. total

caffeine rather than just caffeine from coffee), or that with

the most appropriate adjustment for confounding, were

used.

Where caffeine-containing beverages were not pre-

sented as mg of caffeine per day, a serving of coffee was

assumed to contain approximately 100 mg caffeine and

any caffeine-containing beverage not separated into dif-

ferent sources (e.g. coffee, tea or cola) was assumed to

contain 60 mg of caffeine on average [15]. These are

broadly in line with the assumptions made in the studies

included that assessed caffeine intake on the basis of die-

tary intake.

To be included in linear dose–response meta-analyses,

studies needed to present estimates of RR with associated

CIs for at least three categories of intake, alongside a

quantified measure of caffeine or coffee intake, with suf-

ficient detail regarding the numbers of cases and non-cases.

Studies already presenting results as linear dose–response

(e.g. RR per 100 mg/day) were included as well. Where

studies presented results in both ways, the latter was taken

as being the most accurate. To be included in nonlinear

dose–response meta-analyses, more than three categories

726 D. C. Greenwood et al.

123

of intake were required, and results presented only as linear

dose–response curves could not be used.

Data extraction and quality assessment

From the publications identified, we extracted the follow-

ing information: authors, publication year, geographical

region of the study, numbers of cases and non-cases,

whether dietary assessment was prospective (before

delivery) or retrospective, whether assessment of caffeine

was based on multiple dietary sources or just coffee, level

of dietary exposure (either as mean, median, midpoint or

range for each category or unit of increment for continuous

estimates), estimated RRs with confidence intervals, and

characteristics controlled for either by modelling, matching

or stratification. Data extraction was carried out by YJ and

DCG and its accuracy checked by DCG and a sample by

JC.

We assessed the methodological quality of studies using

the Newcastle–Ottawa scale for either case–control or

cohort studies, as appropriate, presented as a risk of bias

table [16]. For cohorts, stars were awarded in the ‘‘selec-

tion’’ category for participants being representative preg-

nancies in terms of caffeine intake, sampling of unexposed

participants from the same community, detailed caffeine

assessment, and demonstration that adverse pregnancy

outcomes were not present at the start of the study. Com-

parability stars were awarded for adjustment for age and

for smoking. These should be interpreted as risk of bias in

the estimate, and not the study itself, as some estimates

were derived from unadjusted descriptive statistics of

secondary exposures. Outcome stars were awarded for

outcomes based on medical records rather than self-report,

for follow-up to end of pregnancy, and for at least 70 %

follow-up. For case–control studies, stars were awarded for

similar criteria, selection stars being awarded for inde-

pendently validated cases, cases representative of all inci-

dent adverse outcomes, appropriately sourced controls, and

controls demonstrably free from the adverse outcome.

Comparability stars were awarded as for cohorts. For case–

control studies, the remaining quality assessment is based

on the exposure definition rather than the outcome defini-

tion, with exposure stars awarded for caffeine assessment

blind to case–control outcome, the same method of caf-

feine assessment for cases and controls, and the same

response rate in both groups (within 10 % points of each

other). All studies were included, regardless of perceived

quality.

Data synthesis and analysis

So that we could legitimately pool results from studies

using different exposure categorisation, we derived a linear

dose–response trend for each study using Greenland and

Longnecker’s method [17]. This method allows estimation

of study-specific linear dose–response slopes and their

associated confidence intervals, based on the results pre-

sented for each category of caffeine intake. These dose–

response slopes, now all representing estimates of the same

quantity, can then be combined into an overall pooled

estimate using standard methods for meta-analysis.

So that we could derive the dose–response trend, we

used the mean or median caffeine intake for each category

if this was presented. We used the category midpoints

when caffeine intake was only reported as a range. When

the highest or lowest categories of intake were unbounded,

we estimated the midpoint by assuming the width of the

category was the same as the next adjacent category. If the

distribution of cases was not presented in the publication,

we estimated these initial numbers from definitions of the

quantiles, assuming equal numbers in each category.

When studies already presented a linear dose–response

trend based on a continuous measure of caffeine intake,

alongside a measure of precision such as a confidence

interval or a standard error, then we used this instead of

deriving the trend indirectly. Where results were presented

for two distinctly separate subgroups of women [18], we

combined the separate subgroup results first using a fixed

effects meta-analysis before they were combined with the

other studies. In this way the between-study heterogeneity

was estimated appropriately. Where results were presented

for two distinct sub-outcomes on the same women [19],

these were first combined using an approach suggested by

Hamling et al. [20]. This combines the fitted cell counts for

the two sub-outcomes, maintaining the same exposure

category totals, combining the two results into one with no

double counting. We then pooled all the estimated dose–

response trends for each study using a random effects

model that takes into account the anticipated between-

study heterogeneity [21].

We also explored potential nonlinear associations

between caffeine and birth outcome. For each study that

presented results for more than three categories of intake

(the minimum required to model a nonlinear curve) we

calculated a restricted cubic spline. This was based on three

knots fixed at 10, 50 and 90 % through the total distribution

of intake [22]. We then combined these using multivariate

meta-analysis [23].

We quantified the percentage of total variation in study

estimates attributable to between-study heterogeneity (I 2 )

and tested between-study heterogeneity using Cochran’s Q

[24]. We assessed the methodological quality of studies

using the Newcastle–Ottowa scale, presented as a risk of

bias table. We also performed a limited number of pre-

defined subgroup analyses to explore aspects of study

quality that may have contributed to the differences in the

Caffeine intake during pregnancy and adverse birth outcomes 727

123

results seen across the studies, such as retrospective or

prospective assessment of intake, source of caffeine, geo-

graphic location, and adjustment for pre-specified potential

confounders. Though caffeine metabolism may be moder-

ated by CYP1A2 activity, smoking status, or nausea [25],

numbers of studies reporting stratified analyses were too

few to enable investigation of effect modification by these

factors. We conducted sensitivity analysis excluding

potentially highly influential studies and studies that might

be considered materially different. Any such analyses are

included in the relevant section of the results. We inves-

tigated potential small study effects, such as publication

bias, using contour-enhanced funnel plots. However, with

small numbers of included studies, exploration of sources

of heterogeneity and of small study effects lacked power.

All analyses were conducted using Stata version 13.1 [26].

Results

Literature search

In total, 91 relevant publications were identified, of which

14 contained replicated information, and 17 did not contain

sufficient information for inclusion in the meta-analyses

(Fig. 1). This left 60 unique publications from 53 studies

(26 cohort and 27 case–control) with sufficient information

for inclusion in dose–response meta-analyses. Of these 53

studies, 29 were from the US, 21 from Europe, 2 from

South America, and 1 from Cuba. Characteristics of the

included studies are shown in online tables 2a to 2e

alongside an assessment of risk of bias for each study.

One publication only presented miscarriage combined

with stillbirth [27]. These two outcomes could have very

different risk factors, but because of the low incidence of

stillbirth relative to miscarriage, this publication was

included in the meta-analysis of miscarriage, with sensi-

tivity analysis to confirm this did not materially influence

the pooled estimate. One study of low birth weight also

included some small for gestational age infants [28]. This

was included in the low birth weight analysis as these

formed the majority of events, but with sensitivity analysis

to confirm that this did not substantially change the results.

One study provided additional information on miscarriage

and stillbirth separately that was incorporated into the

meta-analyses [29]. One publication provided results for

the association between preterm delivery and caffeine

separately for small for gestational age infants and normal

for gestational age infants [18]. These two subgroups were

first combined using a fixed effects meta-analysis before

14 2

1

14

242

129

=343

101)

113

(n=13 n=180)

Embase (n=150)

22

60

91

Fig. 1 Article retrieval and screening process flow chart

728 D. C. Greenwood et al.

123

they were pooled with the other studies. One study pre-

sented the association between caffeine and preterm

delivery with premature rupture of the membranes sepa-

rately from those without premature rupture of the mem-

branes [19]. These were first combined using an approach

suggested by Hamling et al. [20] before pooling with the

other studies. One publication was based on women with

type 1 diabetes [30]. Because of the non-general popula-

tion, a sensitivity analysis was conducted to confirm that

this did not materially influence the pooled estimate.

Where a publication presented results for different defini-

tions of small for gestational age, the method most con-

sistent with the other studies was used [31]. However, one

publication with a combined outcome of ultrasound-based

diagnosis of intrauterine growth restriction and low birth

weight, could not be included in either category [32].

Miscarriage

Our literature search identified 35 unique publications from

32 studies investigating the association between caffeine

and miscarriage. Data were extracted from 26 of these

studies (14 cohort, 12 case–control), 13 from Europe and

13 from the US. The pooled estimate of RR from linear

dose–response meta-analysis was 1.14 (95 % CI

1.10–1.19) per 100 mg/day of caffeine (P \ .001) (Fig. 2a). Five of the six studies that could not be included

in meta-analyses supported this direction of association,

with all but one statistically significant [33–37], and the

remaining study finding no evidence of any association

[38].

There was substantial heterogeneity between the studies

(I 2

= 89 %; 95 % CI 85–92 %; Q = 230; df = 25;

P \ .001). The estimated between-study variance was unchanged by excluding the study that had combined

miscarriage and stillbirth into a single outcome [27], and

this resulted in a negligible change to the pooled estimate

(RR = 1.14; 95 % CI 1.09–1.18; P \ .001). Excluding the study with greatest influence [39] decreased the estimated

between-study variance by more than half, but resulted in

only a marginally reduced estimate (RR = 1.11; 95 % CI

1.07–1.14; P \ .001), so was subsequently excluded from the subgroup analyses.

There was no evidence that any study characteristics

investigated, such as retrospective caffeine assessment or

adjustment for specific potential confounders, were

Pooled estimate

Tolstrup et al, 2003 Rasch et al, 2003

Dominguez-Rojas, 1994

Weng et al, 2008

Wisborg et al, 2003

Parazzini et al, 1998

Mills et al, 1993

George et al, 2006

Srisuphan et al, 1986

Fenster et al, 1997

Bech et al, 2005

Savitz et al, 2008

Agnesi et al, 1997

Maconochie et al, 2007

Armstrong et al, 1992

Fenster et al, 1991

Infante-Rivard et al, 1993

Wilcox et al, 1990

Khoury et al, 2004

Wen et al, 2001

Greenwood et al, 2010

Kline et al, 1991

Giannelli et al, 2003

Cnattingius et al, 2000

Hansteen et al, 1990

Study

Dlugosz et al, 1996

1.14 (1.10, 1.19)

1.04 (1.00, 1.07) 1.25 (1.13, 1.38)

2.04 (1.80, 2.30)

1.29 (1.10, 1.51)

1.11 (1.04, 1.18)

1.37 (1.27, 1.49)

1.15 (0.88, 1.51)

1.16 (0.93, 1.44)

1.32 (1.01, 1.72)

1.07 (0.97, 1.19)

1.04 (1.01, 1.08)

0.92 (0.79, 1.07)

1.36 (1.06, 1.74)

1.02 (0.98, 1.07)

1.02 (1.00, 1.03)

1.06 (0.97, 1.17)

1.32 (1.15, 1.52)

Estimated

1.18 (0.74, 1.89)

1.86 (0.98, 3.54)

1.26 (1.02, 1.55)

1.08 (0.86, 1.37)

1.07 (1.03, 1.11)

1.18 (1.04, 1.34)

1.07 (1.00, 1.15)

1.04 (1.01, 1.07)

RR (95% CI)

1.11 (0.94, 1.31)

.5 1 2 3

RR per 100 mg/day of caffeine

a

Pooled estimate

Matijasevich et al, 2006

Study

Greenwood et al, 2011

Bech et al, 2005

Linn et al, 1982

Wisborg et al, 2003

1.19 (1.05, 1.35)

1.32 (1.17, 1.49)

RR (95% CI)

1.53 (1.29, 1.83)

1.03 (0.94, 1.14)

1.08 (0.88, 1.32)

1.10 (1.02, 1.18)

Estimated

.5 1 2 3

RR per 100 mg/day of caffeine

c

Pooled estimate

Bakker et al, 2010 Boylan et al, 2008

McDonald, 1992

Pastore, 1995 Peacock, 1995

Fortier et al, 1993

Bracken et al, 2003

de Souza et al, 2005 Khoury et al, 2004

Study

Olsen, 1991

Bicalho et al, 2002

Williams, 1992

Berkowitz, 1982

Sengpiel et al, 2013

Chiaffarino et al, 2006

1.02 (0.98, 1.06)

1.01 (0.93, 1.11) 1.12 (1.01, 1.24)

1.02 (0.99, 1.04)

0.92 (0.82, 1.04) 0.96 (0.85, 1.09)

0.98 (0.86, 1.11)

1.19 (0.99, 1.44)

1.75 (0.74, 4.14) 1.53 (0.97, 2.40)

RR (95% CI)

1.02 (0.98, 1.07)

0.79 (0.67, 0.93)

1.13 (1.06, 1.19)

0.98 (0.83, 1.14)

Estimated

0.98 (0.92, 1.05)

0.95 (0.78, 1.15)

.5 1 2 3

RR per 100 mg/day of caffeine

e .5

1 1 .5

2 3

4 5

E st

im a

te d

R R

0 200 400 600 800 1000

Caffeine intake (mg/day)

Best fitting cubic spline

95% confidence interval

b

.5 1

1 .5

2 3

4 5

E st

im a

te d

R R

0 200 400 600 800 1000

Caffeine intake (mg/day)

Best fitting cubic spline

95% confidence interval

d

.5 1

1 .5

2 3

4 5

E st

im a

te d

R R

0 200 400 600 800 1000

Caffeine intake (mg/day)

Best fitting cubic spline

95% confidence interval

f

Fig. 2 Dietary caffeine intake and estimated relative risk of miscar- riage, stillbirth, and preterm birth. a, c, e Forest plots of linear dose– response trends with pooled estimates from random-effects meta-

analysis per 100 mg/day caffeine intake for miscarriage (a), stillbirth (c), and preterm birth (e). b, d, f Summary nonlinear dose–response

curves from multivariate random effects meta-analysis of restricted

cubic spline curves for miscarriage (b), stillbirth (d), and preterm birth (f), using zero intake as the reference intake. Tickmarks on the horizontal axis indicate the location of category medians, means, or

midpoints for included studies

Caffeine intake during pregnancy and adverse birth outcomes 729

123

associated with higher or lower estimates in subgroup

analyses (online table 3a). However, there was evidence of

considerable asymmetry in the funnel plot (online fig-

ure 1a), to the extent that small-study effects such as

publication bias cannot be ruled out for this outcome.

The nonlinear dose–response meta-analysis showed a

small but consistently increasing incidence of miscarriage

associated with increased daily caffeine intake (Fig. 2b).

There was little evidence of any nonlinear association, such

as a threshold effect, in the plot.

Stillbirth

Our literature search identified eight unique publications

from five studies investigating the association between

caffeine and stillbirth. Data were extracted from all five of

these studies (three cohort, two case–control), three from

Europe, one from the US and one from Uruguay. The

pooled estimate of RR from linear dose–response meta-

analysis was 1.19 (95 % CI 1.05–1.35) per 100 mg/day of

caffeine (P = .007) (Fig. 2c). All the identified studies

were included in meta-analyses.

There was substantial heterogeneity between the studies

(I 2

= 82 %; 95 % CI 59–92 %; Q = 22; df = 4;

P \ .001). Studies that only considered coffee had sub- stantially lower estimates than those that considered caf-

feine from multiple sources (P = .02). Stratifying on this

study characteristic reduced the percentage of total varia-

tion in estimates attributable to between-study heteroge-

neity to \50 % in both groups (online table 3b). Again there was evidence of asymmetry in the funnel plot (online

figure 1b), but the number of studies was small.

The nonlinear dose–response meta-analysis showed a

small but consistently increasing incidence of stillbirth

associated with increased daily caffeine intake (Fig. 2d).

There was little evidence of any nonlinear association.

Preterm delivery

The literature search identified 21 unique publications from

20 studies investigating the association between caffeine

and preterm delivery. Data were extracted from 15 of these

studies (8 cohort, 7 case–control), 6 from Europe and 9

from the US. The pooled estimate of RR from linear dose–

response meta-analysis was 1.02 (95 % CI .98–1.06) per

100 mg/day of caffeine (P = .42) (Fig. 2e). Of the five

studies that could not be included in meta-analyses, four

suggested some positive association [40–43], and one did

not [44].

There was substantial heterogeneity between the studies

(I 2

= 63 %; 95 % CI 34–79 %; Q = 38; df = 14;

P = .001). Studies with prospective dietary assessment of

caffeine intake tended to have more positive associations

with preterm delivery than those with retrospective

assessment (P = .04) (online table 3c). There was no

evidence of any asymmetry in the funnel plot (online

figure 1c).

The nonlinear dose–response meta-analysis showed a

generally straight flat line with no evidence of any non-

linear association (Fig. 2f).

Low birth weight

We identified 14 unique publications from 14 studies

investigating the association between caffeine and low

birth weight. Data were extracted from 11 of these studies

(6 cohort, 5 case–control), 3 from Europe, 7 from the US

and 1 from Brazil. The pooled estimate of RR from linear

dose–response meta-analysis was 1.07 (95 % CI

1.01–1.12) per 100 mg/day of caffeine (P = .01) (Fig. 3a).

Of the three studies not included in meta-analysis, one

found a strong association between caffeine and low birth

weight [45], whilst the other two found no evidence of any

association [40, 46].

There was substantial heterogeneity between the studies

(I 2

= 75 %; 95 % CI 55–86 %; Q = 40; df = 10;

P \ .001). Excluding the study with additional small for gestational age infants [28], had a negligible impact on the

pooled estimate or the estimated heterogeneity (RR = 1.06;

95 % CI 1.01–1.12; P = .02). Studies that adjusted for

maternal education or socio-economic factors had substan-

tially lower estimates than those that did not (P = .02)

(online table 3d). There was some evidence of asymmetry in

the funnel plot (online figure 1d), leaving open the possi-

bility of small-study effects such as publication bias.

The nonlinear dose–response meta-analysis showed a

small but consistently increasing incidence of low birth

weight associated with increased daily caffeine intake

(Fig. 3b). There was little evidence of any nonlinear

association.

Small for gestational age

A total of 18 unique publications were identified investi-

gating the association between caffeine and small for

gestational age infants, based on 18 studies. Data were

extracted from 15 studies (10 cohort, 5 case–control) that

could be included in meta-analysis, 6 from Europe and 9

from the US. The pooled estimate of RR from linear dose–

response meta-analysis was 1.10 (95 % CI 1.06–1.14) per

100 mg/day of caffeine (P \ .001) (Fig. 3c). The three studies that could not be included in meta-analysis pro-

vided mixed results, with one suggesting an association

[47], one not [40], and a third suggesting no overall asso-

ciation, but some evidence of effect modification by

smoking [48].

730 D. C. Greenwood et al.

123

There was substantial heterogeneity between the studies

(I 2

= 64 %; 95 % CI 38–80 %; Q = 39; df = 14;

P \ .001). Studies that adjusted for smoking (P = .04) and studies that adjusted for previous adverse pregnancy out-

comes (P = .05) tended to have lower estimates than those

that did not adjust for these potential confounders (online

table 3e). There was some evidence of asymmetry in the

funnel plot (online figure 1e), leaving open the possibility

of small-study effects such as publication bias.

The nonlinear dose–response meta-analysis showed a

small but consistently increasing incidence of small for

gestational age infants associated with increased daily

caffeine intake (Fig. 3d). There was little evidence of any

nonlinear association.

Discussion

We have, for the first time, quantified with precision the

association between caffeine and adverse birth outcomes,

based on 60 publications from 53 separate cohort and case–

control studies of adverse pregnancy outcomes. Meta-

analysis of the associations between caffeine intake contain

nearly 15,000 cases of miscarriage from 180,000 women,

700 still births from 120,000 women, 8,000 preterm

deliveries from nearly 110,000 women, 5,000 low birth

weight infants from nearly 78,000 women, and nearly

12,000 small for gestational age infants from 160,000

women. The evidence covers a variety of countries with

different levels of intake, including non-consumers and

categories consuming over 1,000 mg/day. This pooled

evidence allows the associations between caffeine intake

during pregnancy and these adverse outcomes to be

described in greater detail than previously possible, and in

a manner that allows the shape of the dose–response curve

to be described.

A small but quantifiable association was observed

between caffeine intake during pregnancy and incidence of

miscarriage, stillbirth and low birth weight. There is also a

similar sized association between caffeine intake during

pregnancy and small for gestational age. There was no

evidence of an association between caffeine intake and

preterm delivery. For all outcomes the dose–response

curves are fairly linear, with no evidence of any ‘‘threshold

effect’’ or ‘‘plateau’’ in the dose–response curves. Hetero-

geneity is generally high, with little between-study heter-

ogeneity being explained by aspects of study design or

analysis investigated.

The size of the associations are relatively modest within

the range of intakes consumed by the majority of women in

Pooled estimate

Caan et al, 1989

Santos et al, 1998

Bracken et al, 2003

Olsen, 1991

Study

Bicalho et al, 2002

Martin et al, 1987

McDonald, 1992

Boylan et al, 2009 Bakker et al, 2010

Linn et al, 1982

Fenster et al, 1991

1.07 (1.01, 1.12)

1.26 (0.97, 1.64)

0.92 (0.81, 1.04)

1.12 (0.88, 1.43)

1.05 (1.01, 1.09)

RR (95% CI)

0.86 (0.75, 0.98)

1.48 (1.24, 1.77)

Estimated

1.02 (1.00, 1.05)

1.31 (1.01, 1.69) 1.10 (1.01, 1.20)

1.08 (1.03, 1.13)

1.19 (0.97, 1.47)

.5 1 2 3

RR per 100 mg/day of caffeine

a

Pooled estimate

Bakker et al, 2010

Vik et al, 2003

Boylan et al, 2009

McDonald, 1992

Mills et al, 1993

Sengpiel et al, 2013

Study

Parazzini et al, 2005

Bracken et al, 2003

Xue et al, 2007

van den Berg et al, 2013

Fenster et al, 1991

Bicalho et al, 2002

Fortier et al, 1993

Grosso et al, 2001 Rondo et al, 1996

1.10 (1.06, 1.14)

1.14 (1.05, 1.25)

1.31 (1.09, 1.56)

1.14 (1.02, 1.27)

1.04 (1.02, 1.06)

1.73 (1.25, 2.40)

1.07 (1.04, 1.11)

RR (95% CI)

1.02 (0.94, 1.10)

1.12 (0.93, 1.34)

1.09 (1.05, 1.13)

1.15 (1.03, 1.29)

1.27 (1.03, 1.55)

0.87 (0.69, 1.09)

1.13 (1.03, 1.23)

0.98 (0.83, 1.15) 1.19 (1.02, 1.38)

Estimated

.5 1 2 3 RR per 100 mg/day of caffeine

c

.5 1

1 .5

2 3

4 5

E st

im a te

d R

R

0 200 400 600 800 1000

Caffeine intake (mg/day)

Best fitting cubic spline

95% confidence interval

b

.5 1

1 .5

2 3

4 5

E st

im a te

d R

R 0 200 400 600 800 1000

Caffeine intake (mg/day)

Best fitting cubic spline

95% confidence interval

d

Fig. 3 Dietary caffeine intake and estimated relative risk of

low birth weight and small for

gestational age. a, c Forest plots of linear dose–response trends

with pooled estimates from

random-effects meta-analysis

per 100 mg/day caffeine intake

for low birth weight (a) and small for gestational age (c). b, d Summary nonlinear dose– response curves from

multivariate random effects

meta-analysis of restricted cubic

spline curves for low birth

weight (b) and small for gestational age (d), using zero intake as the reference intake.

Tickmarks on the horizontal

axis indicate the location of

category medians, means, or

midpoints for included studies

Caffeine intake during pregnancy and adverse birth outcomes 731

123

the included studies, and within the range of intake cur-

rently recommended in most countries during pregnancy.

In addition, the size of the associations are small relative to

some established risk factors such as maternal smoking, but

similar to others such secondhand smoke [49]. It is there-

fore important to interpret any public health implications

regarding caffeine intake in the context of known lifestyle

risk factors.

It is also important to interpret these results alongside

the clinical implications of the outcomes. Whilst the con-

sequences of small for gestational age infants are less

severe than miscarriage or stillbirth, small for gestational

age has been associated with an increased risk of perinatal

mortality and morbidity, including perinatal asphyxia.

There is also a body of literature suggesting that it is

associated with adverse effects in adult life [50, 51], such

as increased incidence of obesity, hypertension, hyper-

cholesterolemia, cardiovascular disease, and type 2 diabe-

tes [52–54], If shown to be causal, a small association

could therefore still be of importance from a public health

perspective.

Given the observational nature of the evidence, we

cannot draw inferences on the causal nature of the asso-

ciation identified in this review. Meta-analyses of obser-

vational studies are prone to the same biases as the studies

they pool evidence across. Therefore the evidence from

case–control studies are particularly susceptible to selec-

tion bias and recall bias, and all the studies are susceptible

to uncontrolled confounding. In addition, all these obser-

vational studies are liable to bias from measurement error

in using self-report measures to estimate the dietary intake

of caffeine. One particular issue common to the majority of

studies was the lack of an objective measure of exposure to

tobacco smoke. Smoking is a potentially very strong con-

founder: smokers both consume more caffeine than non-

smokers (because smokers’ altered CYP1A2 activity leads

to faster caffeine clearance) and have much higher rates of

adverse birth outcomes [55]. It is therefore important to

measure smoking objectively, using a repeated biomarker

such as cotinine, to avoid measurement error bias, which in

this case could lead to exaggerated associations from only

partially controlling for its confounding effects [56].

The large heterogeneity observed in the meta-analyses

also requires caution to be exercised in the interpretation of

the results. Whilst earlier meta-analyses have also observed

substantial heterogeneity [4, 5, 57], these may be explained

in part by their pooling studies using different categories of

intake. We tried to avoid this by placing each study on the

same scale, pooling dose–response trends instead [13].

Beyond this, we investigated other potential sources of

heterogeneity through a small number of a subgroup

analyses specified in advance. We used this to explore

whether different study characteristics were associated

with the observed differences in the results. These included

study design, the method of caffeine intake assessment and

adjustment for pre-specified potential confounders such as

smoking. Whilst each of these were associated with some

of the heterogeneity, it was not consistent across the out-

come groups. Though heterogeneity was generally high, it

mostly reflected variation in the size of the association,

rather than whether there was an association. Heterogeneity

associated with small-study effects such as publication bias

was also observed for the meta-analyses of miscarriage,

stillbirth, low birth weight and small for gestational age.

It is possible that reduction in caffeine may be a marker

for a healthier pregnancy and that caffeine is not the cause

of the adverse outcomes [58, 59]. None of the studies

reviewed in this paper have adequately addressed this

issue; simply adjusting for nausea does not correct for this

potential bias and subgroup analysis suggested it made

little difference to the estimates. This potential bias there-

fore remains the most prominent argument against a causal

role of caffeine. Neither are pragmatic trials immune to this

potential bias, where greater compliance with the inter-

vention may be associated with healthier pregnancy.

Only one large double-blind randomised controlled trial

of caffeine reduction during pregnancy and subsequent

birthweight has been conducted to date [7]. Over 1,000

Danish women were recruited, each consuming over three

cups of coffee a day, and randomised to either caffeinated

or decaffeinated coffee. However, the trial did not assess

the important outcomes of miscarriage or stillbirth, ignored

caffeine intake during the first trimester when caffeine

consumption changes markedly and the majority of fetal

deaths occur [7, 29], and did not measure compliance

through objective biomarkers of caffeine intake. In addi-

tion, the intervention focussed on coffee intake rather than

caffeine as a whole, whilst there is evidence from other

countries that cola drinks, tea and chocolate may all con-

tribute at least as much caffeine to the diet during preg-

nancy [6, 29]. These features of the trial limit the extent to

which its results can contribute towards discussion of

caffeine intake as a whole, or the association with mis-

carriage and stillbirth. In the absence of any other sub-

stantive trial data, our meta-analysis of observational

studies provides a valuable resource.

If the observed association is causal, it is possible that it

may be due to caffeine itself, to one of its metabolites, or a

combination of them. Of the four primary routes of caf-

feine metabolism in humans, 3-demethylation is quantita-

tively the most important, the caffeine being converted to

paraxanthine by CYP1A2. Studies have shown there to be

varying levels of CYP1A2 activity in humans and there is

considerable inter-individual variation in caffeine metab-

olism [60]. Measures of caffeine consumption therefore do

not necessarily indicate the levels of caffeine and caffeine

732 D. C. Greenwood et al.

123

metabolites in the maternal or fetal circulation. A small

number of studies have measured levels of caffeine and its

metabolites in maternal or umbilical cord blood rather than

assessing caffeine consumption [55, 61, 62], though given

the range of possible exposures, this was not the focus of

this review. Linking phenotype with genotype is, however,

an area for possible future research.

Given the observational nature of the studies, the het-

erogeneity and small-study effects, it is not possible to

conclude that these associations are causal. The modest

sizes of the associations are such that it is possible they

could be explained by any or all of these potential biases.

However, the plausible biological mechanisms, the evi-

dence from animal studies, the mounting evidence from

different observational human studies, and the dose–

response slopes, provide some evidence to support the

current recommendations limiting caffeine intake during

pregnancy, such as restricting to less than 200 mg/day, as a

precaution in case the associations really are causal. Whilst

the associations are modest in size, they are potentially

important at a public health level, and for infants already at

elevated risk of adverse outcomes.

In summary, combining results from a large number of

studies has allowed associations between caffeine intake

and adverse pregnancy outcomes to be quantified with

precision and discern a modest but significant association

with caffeine intake that could only be adequately quanti-

fied by pooling results. A number of questions still remain

to be answered. These include confirming causality, such

as identifying whether caffeine is the causal agent, one of

its metabolites, or whether the associations are completely

explained by publication bias or caffeine being a marker of

healthy pregnancy. Whilst these issues are unresolved, our

results confirm the precautionary guidance adopted by

countries recommending limiting caffeine consumption

during pregnancy.

Acknowledgments This review was funded by the Food Standards Agency (Contract T01033). We would like to acknowledge the con-

tribution of Alastair Hay, Kay White and Nigel Simpson from the

University of Leeds for comments on preliminary analyses and Gary

Welsh from the Food Standards Agency Information Services and the

University of Leeds Health Sciences Library for assistance with the

literature searches.

Conflict of interests The authors have no competing interests.

References

1. Ministry of Agriculture Fisheries and Food. Survey of caffeine

and other methylxanthines in energy drinks and other caffeine-

containing products (updated). London: Ministry of Agriculture,

Fisheries and Food. 1998. Report No.: 144.

2. Committee on Toxicity of Chemicals in Food Consumer Products

and the Environment. Statement on the reproductive effects of

caffeine. Food Standards Agency. 2001.

3. Wilcox A, Weinberg C, Baird D. Caffeinated beverages and

decreased fertility. The Lancet. 1988;332:1453–6.

4. Santos IS, Victora CG, Huttly S, Morris S. Caffeine intake and

pregnancy outcomes: a meta-analytic review. Cad Saude Publica.

1998;14:523–30.

5. Fernandes O, Sabharwal M, Smiley T, Pastuszak A, Koren G,

Einarson T. Moderate to heavy caffeine consumption during

pregnancy and relationship to spontaneous abortion and abnormal

fetal growth: a meta-analysis. Reprod Toxicol. 1998;12:435–44.

6. Boylan SM, Cade JE, Kirk SFL, Greenwood DC, White KLM,

Shires S, et al. Assessing caffeine exposure in pregnant women.

Br J Nutr. 2008;100:875–82.

7. Bech BH, Obel C, Henriksen TB, Olsen J. Effect of reducing

caffeine intake on birth weight and length of gestation: ran-

domised controlled trial. BMJ. 2007;334:409.

8. Committee on Toxicity of Chemicals in Food Consumer Products

and the Environment. Statement on the review of caffeine.

Department of Health. 1984.

9. Committee on Toxicity of Chemicals in Food Consumer Products

and the Environment. Statement on the reproductive effects of

caffeine. Food Standards Agency. 2008.

10. Organisation of Teratology Information Specialists. Caffeine and

pregnancy. 2006. http://www.otispregnancy.org/pdf/caffeine.

Accessed 13 June 2014.

11. NHS Choices. Caffeine and pregnancy. http://www.nhs.uk/news/

2008/11November/Pages/Newcaffeineadviceinpregnancy.aspx.

Accessed 13 June 2014.

12. American College of Obstetricians and Gynecologists. Education

pamphlet AP001—Nutrition during pregnancy. Washington DC:

American College of Obstetricians and Gynecologists; 2008.

13. Yu WW, Schmid CH, Lichtenstein AH, Lau J, Trikalinos TA.

Empirical evaluation of meta-analytic approaches for nutrient and

health outcome dose–response data. Res Syn Meth. 2013;4:

256–68.

14. Stroup DF, Berlin JA, Morton SC, Olkin I, Williamson GD,

Rennie D, et al. Meta-analysis of observational studies in epi-

demiology: a proposal for reporting. J Am Med Assoc. 2000;

283:2008–12.

15. NHS Choices. Should I limit caffeine during pregnancy? http://

www.nhs.uk/chq/Pages/limit-caffeine-during-pregnancy.aspx.

Accessed 13 June 2014.

16. Wells GA, Shea B, O’Connell D, Peterson J, Losos M, Tugwell

P. The Newcastle–Ottawa Scale (NOS) for assessing the quality

of nonrandomised studies in meta-analyses. http://www.ohri.ca/

programs/clinical_epidemiology/oxford.asp. Accessed 13 June

2014.

17. Greenland S, Longnecker MP. Methods for trend estimation from

summarized dose–response data, with applications to meta-ana-

lysis. Am J Epidemiol. 1992;135:1301–9.

18. Chiaffarino F, Parazzini F, Chatenoud L, Ricci E, Tozzi L,

Chiantera V, et al. Coffee drinking and risk of preterm birth. Eur

J Clin Nutr. 2006;60:610–3.

19. Williams MA, Mittendorf R, Stubblefield PG, Lieberman E,

Schoenbaum SC, Monson RR. Cigarettes, coffee, and preterm

premature rupture of the membranes. Am J Epidemiol. 1992;

135:895–903.

20. Hamling J, Lee P, Weitkunat R, Ambuhl M. Facilitating meta-

analyses by deriving relative effect and precision estimates for

alternative comparisons from a set of estimates presented by

exposure level or disease category. Stat Med. 2008;27:954–70.

21. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control

Clin Trials. 1986;7:177–88.

Caffeine intake during pregnancy and adverse birth outcomes 733

123

22. Harrell FE Jr, Lee KL, Pollock BG. Regression models in clinical

studies: determining relationships between predictors and

response. J Natl Cancer Inst. 1988;80:1198–202.

23. White IR. Multivariate random-effects meta-analysis. Stata J.

2009;9:40–56.

24. Higgins JP, Thompson SG. Quantifying heterogeneity in a meta-

analysis. Stat Med. 2002;21:1539–58.

25. Boylan SM, Greenwood DC, Alwan N, Cooke MS, Dolby VA,

Hay AW, et al. Does nausea and vomiting of pregnancy play a

role in the association found between maternal caffeine intake

and fetal growth restriction? Matern Child Health J. 2013;17:

601–8.

26. StataCorp. Stata statistical software: Release 13.1. College Sta-

tion, TX: Stata Corporation; 2013.

27. Infante-Rivard C, Fernandez A, Gauthier R, David M, Rivard GE.

Fetal loss associated with caffeine intake before and during

pregnancy. JAMA. 1993;270:2940–3.

28. Caan BJ, Goldhaber MK. Caffeinated beverages and low birth-

weight: a case–control study. Am J Public Health. 1989;79:

1299–300.

29. Boylan S, Cade JE, Dolby VA, Greenwood DC, Hay AWM, Kirk

SFL, et al. Maternal caffeine intake during pregnancy and risk of

fetal growth restriction: a large prospective observational study.

BMJ. 2008;337:a2332.

30. Khoury JC, Miodovnik M, Buncher CR, Kalkwarf H, McElvy S,

Khoury PR, et al. Consequences of smoking and caffeine con-

sumption during pregnancy in women with type 1 diabetes.

J Matern Fetal Neonatal Med. 2004;15:44–50.

31. Sengpiel V, Elind E, Bacelis J, Nilsson S, Grove J, Myhre R,

et al. Maternal caffeine intake during pregnancy is associated

with birth weight but not with gestational length: results from a

large prospective observational cohort study. BMC Med. 2013;

11:42.

32. Spinillo A, Capuzzo E, Nicola SE, Colonna L, Egbe TO, Zara C.

Factors potentiating the smoking-related risk of fetal growth

retardation. Br J Obstet Gynaecol. 1994;101:954–8.

33. Al-Ansary LA, Babay ZA. Risk factors for spontaneous abortion:

a preliminary study on Saudi women. J R Soc Health.

1994;114:188–93.

34. Axelsson G, Rylander R, Molin I. Outcome of pregnancy in

relation to irregular and inconvenient work schedules. Br J Ind

Med. 1989;46:393–8.

35. Furuhashi N, Sata S, Suzuki M, Hiruta M, Tanaka M, Takahashi

T. Effects of caffeine ingestion during pregnancy. Gynecol Obstet

Invest. 1985;19:187–91.

36. Parazzini F, Bocciolone L, Fedele L, Negri E, La VC, Acaia B.

Risk factors for spontaneous abortion. Int J Epidemiol. 1991;20:

157–61.

37. Windham GC, Von BJ, Waller K, Fenster L. Exposure to envi-

ronmental and mainstream tobacco smoke and risk of spontane-

ous abortion. Am J Epidemiol. 1999;149:243–7.

38. Pollack AZ, Buck Louis GM, Sundaram R, Lum KJ. Caffeine

consumption and miscarriage: a prospective cohort study. Fertil

Steril. 2010;93:304–6.

39. Dominguez-Rojas V, De Juanes-Pardo JR, Astasio-Arbiza P,

Ortega-Molina P, Gordillo-Florencio E. Spontaneous abortion in

a hospital population: are tobacco and coffee intake risk factors?

Eur J Epidemiol. 1994;10:665–8.

40. Eskenazi B, Stapleton AL, Kharrazi M, Chee WY. Associations

between maternal decaffeinated and caffeinated coffee con-

sumption and fetal growth and gestational duration. Epidemiol-

ogy. 1999;10:242–9.

41. Tough SC, Newburn-Cook CV, White DE, Fraser-Lee NJ, Faber

AJ, Frick C, et al. Do maternal characteristics and past pregnancy

experiences predict preterm delivery among women aged 20 to

34? J Obstet Gynaecol Can. 2003;25:656–66.

42. Kesmodel U, Olsen SF, Secher NJ. Does alcohol increase the risk

of preterm delivery? Epidemiology. 2000;11:512–8.

43. Mikkelsen TB, Osterdal ML, Knudsen VK, Haugen M, Meltzer

HM, Bakketeig L, et al. Association between a Mediterranean-

type diet and risk of preterm birth among Danish women: a

prospective cohort study. Acta Obstet Gynecol Scand. 2008;

87:325–30.

44. Jarosz M, Wierzejska R, Siuba M. Maternal caffeine intake and

its effect on pregnancy outcomes. Eur J Obstet Gynecol Reprod

Biol. 2012;160:156–60.

45. Stusser R, Paz G, Ortega M, Pineda S, Infante O, Martin P, et al.

Risk of low birth weight in the Plaza de la Habana region. Bol

Oficina Sanit Panam. 1993;114:229–41.

46. Jadsri S, Jadsri C. Parental smoking habits and infant birth

weight. Asia Pac J Public Health. 1995;8:134–7.

47. Alonso J, Sosa C, Verde ME, Balsamo A, Moraes M, Zolessi M,

et al. Risk factors for term small for gestational age. A case-

control study in an uruguayan population. Int J Gynecol Obstet.

2012;119:S276.

48. Infante-Rivard C. Caffeine intake and small-for-gestational-age

birth: modifying effects of xenobiotic-metabolising genes and

smoking. Paediatr Perinat Epidemiol. 2007;21:300–9.

49. Leonardi-Bee J, Britton J, Venn A. Secondhand smoke and

adverse fetal outcomes in nonsmoking pregnant women: a meta-

analysis. Pediatrics. 2011;127:734–41.

50. Barker DJ. Childhood causes of adult diseases. Arch Dis Child.

1988;63:867–9.

51. Nathanielsz PW. Fetal programming: how the quality of fetal life

alters biology for a lifetime. NeoReviews. 2000;1:e126–31.

52. Eriksson M, Wallander MA, Krakau I, Wedel H, Svardsudd K.

The impact of birth weight on coronary heart disease morbidity

and mortality in a birth cohort followed up for 85 years: a pop-

ulation-based study of men born in 1913. J Intern Med.

2004;256:472–81.

53. Ong KK, Dunger DB. Birth weight, infant growth and insulin

resistance. Eur J Endocrinol. 2004;151(Suppl 3):U131–9.

54. Levy-Marchal C, Jaquet D. Long-term metabolic consequences

of being born small for gestational age. Pediatr Diabetes.

2004;5:147–53.

55. Grosso LM, Triche EW, Belanger K, Benowitz NL, Holford TR,

Bracken MB. Caffeine metabolites in umbilical cord blood,

cytochrome P-450 1A2 activity, and intrauterine growth restric-

tion. Am J Epidemiol. 2006;163:1035–41.

56. Cook DG, Peacock JL, Feyerabend C, Carey IM, Jarvis MJ,

Anderson HR, et al. Relation of caffeine intake and blood caf-

feine concentrations during pregnancy to fetal growth: prospec-

tive population based study. BMJ. 1996;313:1358–62.

57. Maslova E, Bhattacharya S, Lin SW, Michels KB. Caffeine

consumption during pregnancy and risk of preterm birth: a meta- analysis. Am J Clin Nutr. 2010;92:1120–32.

58. Leviton A, Cowan L. A review of the literature relating caffeine

consumption by women to their risk of reproductive hazards.

Food Chem Toxicol. 2002;40:1271–310.

59. Peck JD, Leviton A, Cowan LD. A review of the epidemiologic

evidence concerning the reproductive health effects of caffeine

consumption: a 2000–2009 update. Food Chem Toxicol. 2010;

48:2549–76.

60. Landi MT, Sinha R, Lang NP, Kadlubar FF. Human cytochrome

P4501A2. IARC scientific publications 1999. Issue 148.

61. Klebanoff MA, Levine RJ, DerSimonian R, Clemens JD, Wilkins

DG. Maternal serum paraxanthine, a caffeine metabolite, and the

risk of spontaneous abortion. N Engl J Med. 1999;341:1639–44.

62. Klebanoff MA, Levine RJ, Clemens JD, Wilkins DG. Maternal

serum caffeine metabolites and small-for-gestational age birth.

Am J Epidemiol. 2002;155:32–7.

734 D. C. Greenwood et al.

123