ARCHIVAL REPORTS

Gene ⴛ Environment Interactions at the Serotonin

Transporter Locus
Marcus R. Munafò, Caroline Durrant, Glyn Lewis, and Jonathan Flint
Background: Although it is universally accepted that human disease and behavior depend upon both environmental and genetic
variation, a view supported by family and twin studies, examples of environmental interactions with genes identified at the molecular level
(G ⫻ E) are not so well established.

Methods: We carried out a systematic review and meta-analysis of the serotonin transporter (5-HTTLPR) polymorphic region ⫻ stressful life
event (SLE) literature and investigated to what extent the main effects reported in this literature are consistent with a number of G ⫻ E
hypotheses. Our aim was to provide a framework in which to assess the robustness of the claim for the presence of an interaction.

Results: The results from our systematic review and meta-analysis indicate that the main effect of 5-HTTLPR genotype and the interaction
effect between 5-HTTLPR and SLE on risk of depression are negligible. We found that only a minority of studies report a replication that is
qualitatively comparable to that in the original report.

Conclusions: Given reasonable assumptions regarding likely genetic and environmental effect sizes, our simulations indicate that pub-
lished studies are underpowered. This, together with other aspects of the literature, leads us to suggest that the positive results for the
5-HTTLPR ⫻ SLE interactions in logistic regression models are compatible with chance findings.

Key Words: Depression, 5-HTTLPR, Gene ⫻ Environment interac- out statistical tests of interaction is, in part, to infer information
tion, meta-analysis, stressful life events about the biological relationships that might underpin the epi-
demiological data.
Epidemiologists observe interaction effects when they ana-

A
lthough it is accepted that human disease and behavior
depend upon both environmental and genetic variation
(1), examples of environmental interactions with genes
identified at the molecular level (G ⫻ E) are not so well
established. Currently the literature looks very similar to the
initial reports of genetic association studies: there have been a
small number of high-profile findings, followed by a mixture of
replications and non-replications (2). This prompted us to con-
duct a review of one of the most prominent G ⫻ E effects
reported to date, that of the moderating effect of the serotonin
transporter (5-HTTLPR) polymorphism on the relationship be-
tween stressful life events (SLE) and depression (3).
Analysis of G ⫻ E has attracted interest for a number of
reasons. First, it might make it easier to detect genetic associa-
tions. Second, advice might become available to at-risk individ-
uals about avoiding (or seeking out) particular environments.
Third, there might be implications for the underlying biological
relationship between genetic variation and the environment (as
in the original 5-HTTLPR ⫻ SLE report) (3). There are three kinds
of relationship to consider (4). Biological synergism occurs when
the presence of a specific environmental exposure and a specific
genetic variant leads to an increased biological response, beyond
that expected if either were present alone. Biological antagonism
occurs when the effect of an environmental exposure is reduced
in the presence of specific genetic variant. Finally, a crossover
effect occurs when the environmental exposure has opposite
effects depending upon the genotype. The purpose of carrying
From the Department of Experimental Psychology (MRM), Academic Unit of
Psychiatry (GL), University of Bristol, Bristol; and the Wellcome Trust
Centre for Human Genetics (CD, JF), University of Oxford, Oxford, United
Kingdom.
Address reprint requests to Marcus R. Munafò, Ph.D., Department of Exper-
imental Psychology, University of Bristol, 12a Priory Road, Bristol BS8
1TU, United Kingdom; E-mail: marcus.munafo@bristol.ac.uk.
Received February 20, 2008; revised June 7, 2008; accepted June 9, 2008.
lyze stratified populations; for example, two groups within the
same population might have different risks of becoming ill, and
the difference in the risk ratios between these groups is an
example of heterogeneity, which can be detected with statistical
tests for interaction. Interaction effects are usually tested with a
regression model and identify when an association between an
environmental variable and the disease differs according to the
presence or absence of a specific polymorphism. If we consider
a regression model with two independent variables (e.g., geno-
type and environment), a standard model for two independent
effects can be illustrated (Figure 1) by two parallel straight lines.
This indicates that the increase in the dependent variable asso-
ciated with one main effect (i.e., genotype) is constant at all
values of the other main effect (i.e., environment) and vice versa.
A statistical interaction occurs when the relationship between
the dependent variable and genotype varies according to the
value of the environmental variable or vice versa. In graphical
terms, this would be represented by lines that are not parallel.
The test of interaction assumes the null hypothesis of two
independent effects, but a statistical interaction, on its own, does
not indicate what kind of departure from a main effects model
has occurred. Different patterns of interaction have different
implications for understanding the underlying biology, as dis-
cussed previously.
In the G ⫻ E literature, continuous (e.g., score on a depres-
sion questionnaire) and categorical (e.g., case or non-case of
depression) outcomes have both been used. Linear regression
models (used for continuous outcomes) are “additive” in that the
absolute difference in depression scores between different ge-
notypes are assumed to be the same, irrespective of the environ-
mental effects. In contrast, logistic regression models (used for
categorical outcomes) assume a “multiplicative” relationship,
such that the ratio of the outcome is constant. This issue is further
complicated because logistic regression models can be con-
ceived as applying to a quantitative trait with a liability threshold

0006-3223/09/$36.00 BIOL PSYCHIATRY 2009;65:211–219

doi:10.1016/j.biopsych.2008.06.009 © 2009 Society of Biological Psychiatry
212 BIOL PSYCHIATRY 2009;65:211–219
Figure 1. Presence or absence of interaction between two independent effects. In the absence of an interaction (left panel), the value of the main effect of Life
Events (E) remains constant (b1) at both levels of the main effect of Genotype (G), where this also remains constant (b2). In the presence of an interaction (right
panel), the value of the main effect of Life Events (E) differs as a function of Genotype (G). However, statistical significance of the interaction does not provide
information about the nature of the interaction.
superposed upon it (5). However, it is not entirely clear how
statistical interaction in linear regression models relates to statis-
tical interaction in logistic regression models. Although there is
some consensus that departure from an additive model is most
likely to accurately reflect a biological interaction (6), the critical
point here is that the presence of a statistical interaction per se is
not necessarily informative regarding the exact nature of any
underlying biological interaction.
Therefore, the way in which observed data depart from the
independent main effects model is an important determinant of
the statistical significance of the main genetic effect, but it is also
critical in understanding any underlying biological interaction.
The nature of the interaction also has implications for whether
the interaction is easier to detect than a main effect. For example,
in the original 5-HTTLPR ⫻ SLE report (3) participants with one
or two copies of the short allele (S) of the promoter polymor-
phism had an increased risk of depression in the presence of one
or more life events, compared with individuals homozygous for
the long allele (L). The main effect of genotype was not
statistically significant in this study. However, the interaction
reported would lead to a significant main effect of genotype
given sufficient statistical power because, on average, those with
the S allele would have an increased risk of depression.
In this report, we had two aims: first, to carry out a systematic
review and meta-analysis of the 5-HTTLPR ⫻ SLE literature; and,
second, to ask to what extent the main effects reported in the
5-HTTLPR ⫻ SLE literature are consistent with a number of G ⫻
E hypotheses. We set out to explore, by simulation, the set of
possible circumstances in which the type of synergistic interac-
tion originally reported (3) could occur without a corresponding
main effect of genotype. Our aim was to provide a framework in
which to assess the robustness of the claim for the presence of an
interaction.
Methods and Materials
Systematic Review
The PubMed, EMBASE, and PsycINFO databases were
searched to the end of December 2007 with combinations of the
terms “serotonin transporter”, “5-HTTLPR”, “life event$”, “life
stress$”, “depress$”, “interaction” and “moderation”. Bibliogra-
phies of studies identified by our search were also hand-
searched, and authors were contacted to identify work under
review or in press, with the latter resulting in the identification of
www.sobp.org/journal
M.R. Munafò et al.
two studies (7,8). Studies were included if they met three criteria:
1) assessment of 5-HTTLPR genotype, 2) assessment of SLE
comparable to that used in the original report, and 3) assessment
of major depression or depressed mood. The authors MM and GL
determined the inclusion of studies for review.
Meta-Analysis
Studies identified by our systematic review were included in
a meta-analysis, where data were available that enabled this (i.e.,
frequency of depression grouped by 5-HTTLPR genotype and
SLE exposure). Genotype frequencies for each study were used
to calculate whether or not they deviated significantly from
Hardy-Weinberg equilibrium among controls. Discrepancies in
data extraction were resolved by mutual consent.
Data were initially analyzed within a fixed effects framework,
and odds ratios (OR) were pooled with inverse variance methods
to generate a summary OR and 95% confidence interval (CI). A
fixed effects framework assumes that the effect of interest is
constant across studies, and between-study variation is consid-
ered to be due to chance or random variation. The assumption
was checked with a ␹2 test of goodness of fit for homogeneity,
and the I2 statistic was used to estimate the degree of heteroge-
neity. When significant association was observed in the presence
of significant between-study heterogeneity, the analysis was
repeated within a random effects framework and ORs pooled
with Der Simonian and Laird methods. A random effects frame-
work assumes that between-study variation is due to both chance
or random variation and an individual study effect. In all cases,
the significance of the pooled OR was determined with a Z test.
Genotypes were grouped according to the presence (SS or SL)
or absence (LL) of the S allele (9), and life events were grouped
according to the presence (1⫹) or absence (0) of an SLE, to allow
the calculation of ORs. In addition, interaction ORs were calcu-
lated for each study. Main effects were tested without the
interaction term.
Simulation
The basic parameters of our simulation were chosen to
maximize comparability with the initial report of a 5-HTTLPR ⫻
SLE interaction effect. Our model included a genetic effect (G),
an environmental effect (E), and their interaction (I). The genetic
locus was specified to have two variants (S and L) and three
possible genotypes (SS, SL, and LL), with the S allele increasing
susceptibility to the outcome of interest (y). No assumptions of
M.R. Munafò et al.
Figure 2. Example interaction model. An example of the type of interaction
assumed, with the S allele: 1) dominant (squares), 2) additive (circles), and 3)
recessive (triangles). The main effect of environment (E) is assumed to ac-
count for 2% of the total phenotypic variance, and the main effect of geno-
type (G) is assumed to account for .3% of the total phenotypic variance. LL, S
allele absent; SL, one S allele present; SS, two S allele present.
additivity or dominance at this locus were made. The environ-
mental effect was specified as having two levels, unexposed (E1)
and exposed (E2), corresponding to 0 and 1⫹ SLE. Main effects
were tested without the interaction term.
A full description of the equations used in our simulations is
provided as Supplement 1. A model of interaction was assumed
where there is no effect of genotype on the phenotype in the
unexposed group E1, comparable to that described in the original
5-HTTLPR ⫻ SLE report. The effect of environmental exposure
on the phenotype is affected by the genotype present at the locus
of interest (Figure 2).
Results
Systematic Review
A total of 33 studies were identified by the search strategy, of
which 15 (1,3,10 –23) met all three criteria for inclusion (Table 1).
One of these (11) reported data that was also reported in
another, larger study (15).
Of the original 33 studies identified, 7 (7,8,24 –28) were
identified that met one or two but not all three criteria for
inclusion and were therefore excluded (Table 2).
The original report of a moderating effect of the 5-HTTLPR
polymorphism on the association between SLE and risk of major
depression (3) indicated that, with no exposure to SLE, there was
no difference between 5-HTTLPR genotype groups in risk of
major depression. However, as the number of SLE increased, the
risk of major depression increased among those individuals
carrying one or more copies of the S allele (in a dose-dependent
manner), with risk greater among the SS group than the SL group.
Among LL homozygotes, increasing SLE did not result in a
significantly increased risk of major depression.
Failure to distinguish between the different types of statistical
interaction and the statistical modelling used has, in our view, led
to unwarranted claims for replication. Of the 14 independent
studies identified by our search strategy (Table 1), including the
original report (3), only 2 further studies (21,22) reported a
statistically significant interaction apparently similar to that ob-
served in the original report with a categorical outcome within a
BIOL PSYCHIATRY 2009;65:211–219 213
multiplicative model. In one of these the SS and SL groups did
not differ at high stress levels (21), whereas in the other the SL
and LL groups did not differ, so that higher risk of depression at
high stress levels was only observed in the SS compared with the
L⫹ group (22), both in contrast with the original report. A further
three studies (11,14,15,18) claimed replication with a continuous
outcome within an additive model. Even within these studies,
however, there were specific differences in the nature of the
interaction reported compared with the original report, com-
monly in whether the SS and SL groups differed in risk of
depression at high stress levels. Three studies reported no
evidence of a statistically significant interaction (12,17,23), al-
though one interpreted their results as offering “modest support”
on the basis of sub-group analyses (23). A further five studies
reported a significant interaction that was apparently different
from that observed in the original report, three with multiplica-
tive models (10,19,20) and two with additive models (13) or
nonparametric methods (16). Reports using multiplicative mod-
els included higher risk of depression in the S⫹ compared with
the LL group at high stress levels but higher risk in the LL
compared with the S⫹ group at the lowest levels (in one case
among females only) (10,16). One study reported higher risk in
the S⫹ group at moderate but not at high stress levels (20).
Because these studies have not distinguished between the dif-
ferent types of statistical interaction, it cannot be said whether
they are in fact qualitatively similar or different, although any
deviation from the main effects model will contribute to the
statistical significance of the interaction term. Although any
judgement regarding the comparability of the qualitative nature
of an interaction effect is somewhat subjective, in our view some
claims of replication seem not to be justified.
Of the 14 studies identified by our systematic review that we
consider to represent attempts at replication sufficiently similar to
the original report, 10 reported categorical outcomes, of which
only 1 (17) provided summary data that would enable a meta-
analysis. We contacted the authors of the remaining nine eligible
studies to request a release of summary data, to enable inclusion.
Authors were contacted on multiple occasions, up to a maximum of
three times. This request was successful in the case of the original
study and one replication study (3,10). In one study we were able
to extract the relevant data from a figure (21), and in another we
used the “abuse” SLE sub-construct (23). We were therefore able to
conduct a meta-analysis with k ⫽ 5 studies (Table 3).
Meta-Analysis
Meta-analysis indicated no evidence of a significant main
effect of genotype (OR 1.07, 95% CI .90 –1.27, Z ⫽ .80, p ⫽ .42),
with evidence of moderate between-study heterogeneity (␹2 ⫽
9.34, p ⫽ .053, I2 ⫽ 57.17) and evidence of a significant main
effect of environment (OR 2.08, 95% CI 1.77–2.44, Z ⫽ 8.88, p ⬍
.001) with evidence of substantial between-study heterogeneity
(␹2 ⫽ 17.52, p ⫽ .002, I2 ⫽ 77.17). Random-effects analyses did
not alter these results substantially.
There was also no evidence of a significant G ⫻ E interaction
effect (OR 1.16, 95% CI .89 –1.49, Z ⫽ 1.11, p ⫽ .27), with no
evidence of significant between-study heterogeneity (␹2 ⫽ 5.25,
p ⫽ .26, I2 ⫽ 23.80). Stratifying by presence (1⫹) or absence (0)
of an SLE indicated no evidence of a significant main effect of
genotype in either the presence (OR 1.16, 95% CI .93–1.45, Z ⫽
1.34, p ⫽ .18) or absence (OR .94, 95% CI .71–1.25, Z ⫽ .40, p ⫽
.69) of an SLE, and these estimates did not differ significantly
from each other (p ⫽ .26).
These results were not altered substantially when SLE were
www.sobp.org/journal
214 BIOL PSYCHIATRY 2009;65:211–219 M.R. Munafò et al.

Table 1. Studies Eligible for Inclusion

Age of Number
Depression Depression Available for
Study Yr Design Sampling % Male Ancestry Measure Measure Stress / Life Event Measure Response Rate Analysis

Caspi 2003 Birth Cohort All births in Dunedin, 42% European DIS 26 (mean) Life events occuring between 82% 847
New Zealand 16 and 21 years or 0,1,2, or
more episodes of
maltreatment between 3
and 11 yrs
Eley 2004 Cross- Adults screened for 42% Not Stated SMFQ (⬎12 10-20 Combined stress index (split Not Stated 377
Sectional another study asked or ⬍3) at median) from SPQ, (sample not
Survey to recruit their parental educational level representative)
children & LTE in previous 6 months

Kaufman 2004 Cross- Children in care and 46% Mixed MFQ 5-15 Maltreatment (children who Not Stated 101
a Sectional community controls had been in care) (sample not
Survey selected from representative)
adverts

Gillespie 2005 Cross- Volunteer twin sample 42% European SSAGA (DSM- 15-25 LTE 29% 1091
Sectional (Australian National IV lifetime
Survey Twin Register) depression)

Grabe 2005 Case-Control Random household 31% Not Stated BL-38 20-79 Unemployment or number of 69% 976
sample chronic diseases

Kendler 2005 Longitudinal Population based twin 50% European Interview 35 (mean) List of 11 personal life events 96% 549
Twin Study sample (DSM-III-R and 4 “network” life events
major with rating of long-term
depression) threat

Jacobs 2006 Longitudinal Not stated 0% European SCL-90 27 (mean) IRLE Not Stated 384
Twin Study (depressive
dimension)
Kaufman 2006 Cross Children in care and 49% Mixed MFQ and 5-15 Maltreatment (children who Not Stated 196
b Sectional community controls KSADS PL had been in care) (sample not
Survey selected from representative)
adverts

Sjoberg 2006 Cross- All students in a 2 year 40% Not Stated DSRS 16 and 19 Composite psychosocial 4% 180
Sectional period index (presence or absence
Survey of one or more stressors)

Surtees 2006 Cross- Patients on General 53% Not Stated HLEQ 40-74 Modified LTE of 8 childhood 61% 4175
Sectional Practice registers and 16 adult adverse
Survey events

Taylor 2006 Cross- Adverts on university 43% Mixed BDI 18-29 Early family environment Not Stated 118
Sectional campus using RFQ and 10 adult (sample not
Survey major life events representative)

Wilhelm 2006 Cohort Sample from 33% Not Stated CIDI and DIS 47 (mean) Life events recorded in 77% 127
postgraduate (lifetime relation to episode of
teachers course depression) depression

Cervilla 2007 Cross- Consecutive attenders 28% Not Stated CIDI and 18-75 LTE (previous 6 months) 80% 737
Sectional of ad hoc group of ICD-10
Survey clinics (depressive
episode)

Kim 2007 Cross- Population register Not Stated East Asian GMS 65⫹ LTE (previous 12 months) Not Stated 732
Sectional
Survey

Scheid 2007 Cross- Women attending 0% European CES-D Majority Life events from childhood to Not Stated 568
Sectional prenatal screening 20-34 adulthood
Survey clinics

www.sobp.org/journal
M.R. Munafò et al. BIOL PSYCHIATRY 2009;65:211–219 215

Hardy Result: Adjustment / Results of

Weinberg Triallelic Grouping: Result: Grouping: Stress / Stress / Control for Adjusment /
Equilibrium Genotyping Genotype Genotype Life Event Life Event Result: Interaction Description of Interaction Statistical Model Confounding Control

Yes No SS vs. SL p ⫽ .47 0,1,2,3,4⫹ p ⬍ .001 p ⫽ .05 SS and SL higher risk of Logistic Regression Gender Only adjusted
vs. LL depression than LL at high (additive and results given
stress levels (SL lower risk multiplicative)
than SS)

Yes No SS vs. SL p ⫽ .07 Split at median p ⫽ .75 p ⫽ .09 SS and SL higher risk of Logistic Regression None n/a
vs. LL depression than LL at high (multiplicative)
stress levels, but LL higher
risk than SL and SS at low
stress levels (females only)
Not Stated No SS vs. SL p ⫽ .05 Comparison of p ⫽ .05 p ⫽ .04 SS higher depression scores Linear Regression Age, Gender, Not possible to
vs. LL maltreated and than SL and LL in (additive) Ancestry, compare pre-
control group maltreated group Income and post-
adjustment
results
Yes No SS vs. SL p ⫽ .51 0, 1, 2, 3, 4 p ⫽ .01 p ⫽ .43 LL higher risk of depression Logistic Regression Age, Gender Only adjusted
vs. LL than SL and SS at high (multiplicative) results given
stress levels
(nonsignificant)
Yes No SS/SL vs. ns Presence or Sig. Women: p ⫽ .05 S⫹ higher risk of depression ANOVA (additive) Age, Gender, Only adjusted
LL absence of (unemployment) with unemployment than Education, results given
unemployment and p ⫽ .001 LL (risk of depression Marital
or chronic (chronic disease). declines among LL); S⫹ Status,
disease Men: ns higher risk of depression Social
than LL with intermediate Support
levels of chronic disease
Not Stated No SS vs. SL/ p ⫽ .28 Analyses for both p ⫽ .02 p ⫽ .04 SS higher risk of depression Cox Regression Gender Only adjusted
LL one SLE vs. none than L⫹ at high stress (multiplicative) results given
and using the levels
threat rating in 4
categories
Yes No SS vs. SL Not Stated 0,1,2,3,4,5⫹ p ⫽ .001 p ⫽ .04 SS higher risk of depression Multi-Level Model None n/a
vs. LL than SL and LL at high stress (additive)
levels
Not stated No SS vs. SL ns Comparison of ns p ⫽ .03 Detailed results not provided Linear Regression Age, Gender, Only adjusted
vs. LL maltreated and but text suggests SS higher (additive) Ancestry, results given
control group risk of depression than SL Income
and LL with maltreatment
(50% overlap with previous
study sample)
Yes No SS vs. SL Not Stated Risk vs. no risk on Not Stated p ⫽ .004 SS higher risk of depression Non-Parametric Not Stated n/a
vs. LL psychosocial at than SL and LL at high (additive?)
index stress levels (risk of
depression declines
among LL)
Yes No SS vs. SL ns Number of life Sig. For adult life LL higher risk of depression Logistic Regression Gender Only adjusted
vs. LL events events: OR .96 than SL and SS at high (multiplicative) results given
(.84-1.09) stress levels
(nonsignificant)
Yes No SS vs. SL ns Divided both Sig. p ⫽ .008 (early SS higher risk of depression ANOVA (additive) None n/a
vs. LL early family and family life than SL and LL at high
adult SLEs into events) and stress levels (SL higher risk
two categories p ⫽ .02 (adult life of depression than LL and
events) SS at low stress levels)
Yes No SS vs. SL p ⫽ .60 0, 1, 2, 3⫹ p ⫽ .006 p ⫽ .036 (5-yr SS and SL higher risk of Logistic Regression Gender Only adjusted
vs. LL adverse event, depression than LL at high (multiplicative) results given
but not previous stress levels (LL higher risk
yr) of depression than SL and
SS at low stress levels)
Yes No SS vs. SL/ ns 0, 1, 2⫹ 38% of Sig. Sig. SS higher risk of depression Logistic Regression Gender, Only adjusted
LL sample are in than SL and LL at (multiplicative) Family results given
highest moderate stress levels (no History
category difference at high stress
levels)
Yes No SS vs. SL p ⫽ .38 0, 1, 2, 3⫹ p ⫽ .001 p ⫽ .01 SS and SL higher risk of Logistic Regression Age, Gender, Only adjusted
vs. LL depression than LL at high (multiplicative) Education, results given
stress levels (SL same risk Cognitive
as SS) Function,
Disability
Yes No SS vs. SL p ⫽ .54 Presence or p ⬍ .001 p ⫽ .19 (p ⫽ .05 for SS higher risk of depression Logistic Regression None n/a
vs. LL absence abuse sub- than SL and LL at high (multiplicative)
construct only stress levels (non-
when participants significant)
on medication
excluded)

www.sobp.org/journal
216 BIOL PSYCHIATRY 2009;65:211–219 M.R. Munafò et al.

Table 2. Studies Not Eligible for Inclusion with the corresponding interaction effect size specified according
to this model.
Study Year Reason for Exclusion
These results suggest that for a range of assumptions regard-
Fox (24) 2005 The outcome was “behavioral inhibition” ing the main effects of genotype and environment and the model
rather than depression. of genetic action, the statistical power available for sample sizes
Gibb (25) 2006 Patients were of varying diagnoses with comparable to those in the majority of G ⫻ E studies published
no comparison between depressed vs. to date is low. However, it should be noted that these results
non-depressed participants. depend on the proportion of environmentally exposed individ-
Zalsman (26) 2006 Patients had bipolar disorder, currently uals in the sample. The rarer the exposed individuals, the larger
depressed, in 20% of cases. the interaction effect will be in comparison with the genetic main
Gunthert (27) 2007 The outcome was anxiety rather than effect (Table 5). In the studies reviewed in the meta-analysis, the
depression. frequency of the SLE 1⫹ group was relatively high in absolute
Stein (28) 2007 The outcome was anxiety rather than terms, and so one would therefore expect to find a genetic main
depression. effect if an interaction were present (i.e., the interaction effect
Araya (7) under review The outcome was parent-rated size is roughly equal to the genetic main effect size).
“emotional symptoms” rather than
depression.
Middeldorp (8) in press The outcome was self-reported “anxious Discussion
depression” rather than depression.
Our results indicate, first, that only a minority of studies in our
view report a replication that is qualitatively comparable to that
grouped with a higher threshold for presence (2⫹) versus in the original report; second, that the main effect of 5-HTTLPR
absence (0 –1). These results are non-conservative, because they genotype and the interaction effect between 5-HTTLPR and SLE
do not take into account the nature of the interaction term and on risk of depression are negligible; and third, that given
rely solely on its magnitude. A recent criticism (2) of the largest reasonable assumptions regarding likely genetic and environ-
attempt at replication (17), which failed to find evidence of mental effect sizes, published studies are underpowered. At
interaction, was that it relied on self-report measures of SLE present, the positive results for the 5-HTTLPR ⫻ SLE interactions
weakly related to risk of depression. This is not borne out by our in logistic regression models are still compatible with chance
analysis: there is a significant association between measures of findings.
life events and depression, and under a more conservative Our literature review revealed considerable diversity of opin-
random effects model this effect remains statistically significant. ion on the requirements for replication of G ⫻ E reports (Table
1). Failure to distinguish between different types of statistical
Simulation interaction is partly to blame. In an additive statistical model, a
We used the estimates of the main effects of G and E from our synergistic interaction effect might reflect an underlying biolog-
meta-analysis to investigate by simulation the likelihood of ically synergistic relationship, whereas in a multiplicative model
detecting G ⫻ E, under different models. The effect of the even without a statistical interaction it can be argued that
environmental variable accounts for about 2% of the phenotypic biological synergism is occurring between two independent
variance, and the main effect of genotype is either nonexistent or factors (29). In a multiplicative model, if both genetic and
negligible. It follows that the genetic main effect and the effect of environmental main effects are present this implies that the
the interaction are not independent for this type of interaction absolute increase in risk associated with the environment is
(Supplement 1). If one is specified, the other is automatically larger in the presence of the risk allele, compared with its
determined—that is, if the interaction is such that the environ- absence. It has also been argued (4, 30) that, for complex disease
mental effect is increased in the presence of a gene, then there phenotypes, it is difficult to make any specific inference about
will also be a main effect of the gene when averaged across all underlying biological mechanisms from the rather indirect evi-
levels of the environmental variable. Therefore, for the type of dence from observational studies in humans. It is therefore
interaction and sample frequencies of environmentally exposed important to identify the genetic (and environmental) main
and unexposed individuals used in the simulations, the interac- effects.
tion effect is also either nonexistent or negligible, consistent with A large body of data, outside of the studies reviewed here,
the estimate derived from our meta-analysis. indicates that the 5-HTTLPR effect on depression is very small or
Table 4 indicates the power available for a range of sample negligible (31,32). Indeed, the possibility that the polymorphism
sizes, given a range of genetic and environmental effect sizes, has no association with the disorder at all has not been excluded.

Table 3. Summary Data for Studies Included in the Meta-Analysis

LL SL⫹SS
SLE 0 SLE 1⫹ SLE 0 SLE 1⫹
Study Year MDD⫹ MDD⫺ MDD⫹ MDD⫺ MDD⫹ MDD⫺ MDD⫹ MDD⫺

Caspi (3) 2003 8 (1%) 71 (8%) 30 (4%) 155 (18%) 18 (2%) 166 (20%) 77 (9%) 320 (38%)
Eley (10) 2004 25 (7%) 17 (5%) 12 (3%) 9 (2%) 92 (25%) 90 (24%) 77 (21%) 47 (13%)
Surtees (17) 2006 41 (1%) 648 (15%) 62 (2%) 598 (15%) 75 (2%) 1292 (32%) 120 (3%) 1224 (30%)
Kim (21) 2007 4 (1%) 31 (4%) 10 (1%) 51 (7%) 22 (3%) 147 (20%) 165 (23%) 300 (41%)
Scheid (23) 2007 10 (2%) 77 (14%) 30 (5%) 54 (10%) 34 (6%) 191 (34%) 63 (11%) 103 (18%)
SLE, stressful life event; LL, S allele absent; SL⫹SS, S allele present; MDD, major depressive disorder.

www.sobp.org/journal
M.R. Munafò et al. BIOL PSYCHIATRY 2009;65:211–219 217

Table 4. Power Analysis

Genotype Interaction
Effect Size (% variance) Power (␣ ⫽ .05) Power (␣ ⫽ .05)
Model Genotype Environment Interaction n ⫽ 200 n ⫽ 500 n ⫽ 1000 n ⫽ 200 n ⫽ 500 n ⫽ 1000

Dominant .1 1.5 .08 6% 9% 13% 7% 10% 15%

Dominant .3 1.5 .25 11% 18% 33% 8% 18% 27%
Dominant .5 1.5 .41 14% 29% 52% 12% 23% 44%
Dominant .1 2.0 .08 6% 9% 13% 7% 9% 10%
Dominant .3 2.0 .25 10% 19% 33% 9% 15% 26%
Dominant .5 2.0 .41 13% 30% 50% 11% 23% 45%
Dominant .1 2.5 .08 6% 10% 11% 6% 8% 11%
Dominant .3 2.5 .25 10% 19% 35% 9% 17% 27%
Dominant .5 2.5 .41 15% 27% 52% 12% 22% 42%
Recessive .1 1.5 .08 8% 11% 15% 6% 9% 13%
Recessive .3 1.5 .25 11% 23% 41% 11% 18% 33%
Recessive .5 1.5 .41 16% 35% 61% 14% 29% 54%
Recessive .1 2.0 .08 7% 9% 15% 6% 10% 15%
Recessive .3 2.0 .25 10% 23% 39% 10% 18% 33%
Recessive .5 2.0 .41 18% 34% 63% 15% 30% 53%
Recessive .1 2.5 .08 8% 11% 15% 5% 10% 15%
Recessive .3 2.5 .25 13% 24% 42% 9% 19% 32%
Recessive .5 2.5 .41 15% 35% 64% 15% 28% 55%
Additive .1 1.5 .08 9% 10% 17% 6% 11% 17%
Additive .3 1.5 .25 13% 24% 47% 11% 19% 39%
Additive .5 1.5 .41 17% 36% 64% 16% 32% 57%
Additive .1 2.0 .08 7% 10% 16% 7% 11% 15%
Additive .3 2.0 .25 13% 24% 45% 11% 22% 36%
Additive .5 2.0 .41 17% 39% 68% 13% 34% 56%
Additive .1 2.5 .08 7% 10% 18% 5% 9% 14%
Additive .3 2.5 .25 10% 22% 45% 10% 21% 36%
Additive .5 2.5 .41 14% 38% 66% 16% 34% 60%
Meta-Analysis .1 1.5 .08 7% 12% 18% 6% 10% 15%
Meta-Analysis .3 1.5 .25 12% 25% 45% 9% 22% 37%
Meta-Analysis .5 1.5 .41 17% 37% 68% 16% 29% 59%
Meta-Analysis .1 2.0 .08 7% 10% 16% 6% 12% 17%
Meta-Analysis .3 2.0 .25 11% 23% 45% 10% 22% 35%
Meta-Analysis .5 2.0 .41 19% 37% 68% 15% 32% 57%
Meta-Analysis .1 2.5 .08 8% 10% 18% 7% 11% 15%
Meta-Analysis .3 2.5 .25 12% 24% 45% 9% 21% 39%
Meta-Analysis .5 2.5 .41 17% 40% 66% 15% 33% 57%

The results of our meta-analysis are consistent with this and this broader literature on the effect of 5-HTTLPR on depression.
further suggest that the 5-HTTLPR and SLE interaction effect is Under these assumptions, including the type of interaction, the
negligible. Genetic association studies, although not unique in size of the interaction effect will be similar to that of the genetic
this respect, are notorious for non-replication and inconsisten- main effect. With a main effect of genotype of up to .5% of
cies (33). Failure to replicate does not mean that an original
finding is necessarily incorrect—inconsistent findings in different Table 5. Impact of Variation in Proportion of Environmentally Exposed
populations might be due to genetic heterogeneity, where Individuals
different loci or even different alleles at the same locus contribute
to the effect. Our results cannot demonstrate that the original Effect Size (% variance)
report of the interaction is a false positive. However, our results SLE 0 SLE 1⫹ Genotype Environment Interaction
do indicate that, under the assumption that there is a common
variant of small effect acting at the 5-HTTLPR locus, this finding .45 .55 .30 2.00 .24
could have arisen by chance alone. The large number of .20 .80 .30 2.00 .07
subsequent reports of statistically significant interaction effects, .30 .70 .30 2.00 .13
albeit apparently different in nature to the original report in many .40 .60 .30 2.00 .20
cases, suggests either that minor differences between studies .50 .50 .30 2.00 .30
generate different patterns of interaction or that publication bias .60 .40 .30 2.00 .45
or an undisclosed degree of multiple testing within studies might .70 .30 .30 2.00 .70
influence the reporting of results. .80 .20 .30 2.00 1.20
.90 .10 .30 2.00 2.70
The assumptions used in our simulations are based on
estimates derived from our meta-analysis but are compatible with SLE, stressful life event.

www.sobp.org/journal
218 BIOL PSYCHIATRY 2009;65:211–219
phenotypic variance, our simulations (Table 4) indicate that
none of the published studies has reasonable power (⬎ 80%)
to detect an interaction under the assumptions in our simula-
tion (which entail an interaction effect corresponding to up to
.8% of phenotypic variance).
The relationship between sample size and power to detect an
interaction is complex. When the proportion of environmentally
exposed individuals is quite low, the size of the interaction effect
will be larger than the size of the genetic main effect, because
data from the unexposed individuals dominate the genetic main
effect and overwhelm the signal from the exposed individuals
(Table 5). In this case it would be possible, with a small sample,
to detect an interaction effect and not a genetic main effect.
However, if the genetic main effect size is exactly zero, then the
interaction effect size must also be zero for the type of interaction
we have considered. Power will also depend on the number of
degrees of freedom in the test: the more degrees of freedom, the
lower the power to detect an effect of a given size. The test for
an interaction will, in general, have at least the same number of
degrees of freedom as the test for a main effect and often more.
Therefore, the power to detect an interaction of the same size as
a main effect will be no greater than the power to detect that
main effect.
In practice, without knowledge of the type of interaction
present (i.e., synergism, antagonism, crossover) one lacks infor-
mation regarding how the interaction effect relates to the main
effects. Statistical tests do not discriminate between different
types of interaction but simply consider whether there is depar-
ture from the main effects model. Different types of interaction
will imply different relationships between the main effects and
the interaction effect. Nevertheless, this does not alter the fact
that the size of the interaction effect is determined by the main
effects and the type of interaction present. When the environ-
mental and genetic main effect sizes are known, as well as the
type of interaction operating, the size of the interaction effect is
fixed, given certain assumptions (i.e., genotype frequencies,
proportion of environmentally exposed individuals, model of
genetic action). The results presented here are true for the type
of interaction we have assumed. Assuming a different type of
interaction would have varying effects but would not alter the
interdependency of the effect sizes. Because, in practice, the
nature of the interaction effect might not be known, uncertainty
will remain regarding these assumptions, but it is nevertheless
possible to consider a limited range of alternatives. As data
accumulate, the range of possibilities narrows.
It is possible to obtain robust evidence for the involvement of
a small effect locus by establishing stringent significance thresh-
olds on the basis of prior probabilities of association (34),
although in practice this is not done consistently. Most impor-
tantly, testing for statistical interactions multiplies the potential
problem of multiple testing, particularly in the context of rela-
tively non-conservative significance thresholds. Without rigorous
attention to the qualitative nature of replication, G ⫻ E studies
provide even further opportunities for confusion in the genetic
association literature. The positive findings for the 5-HTTLPR ⫻
SLE interactions are compatible with chance findings, particularly
given the absence of any main effect for the 5-HTTLPR polymor-
phism with depression. The results from our meta-analysis,
although limited to dichotomous comparisons of the presence or
absence of the S allele and one or more SLE, indicate that the
main effect of 5-HTTLPR genotype and the interaction effect
between 5-HTTLPR and SLE on risk of depression are negligible.
Similarly, our simulations indicate that given any reasonable
www.sobp.org/journal
M.R. Munafò et al.
assumption regarding likely genetic effect size, and across a
range of environmental effect sizes and models of genetic action,
statistical power in samples of less than several thousand will be
low except for the most optimistic of scenarios.
The author MM is supported in part by a National Alliance for
Research on Schizophrenia and Depression Young Investigator
Award; CD is supported by a Medical Research Council fellow-
ship; JF is supported by the Wellcome Trust. We are grateful to
George Davey Smith and Stan Zammit for discussion of the issues
included here and comments on several sections of this manu-
script and to the several authors who released data in a format
that enabled their inclusion in the meta-analysis. The author
MM had full access to all of the data in the study and takes
responsibility for the integrity of the data and the accuracy of the
data analysis.
All authors reported no biomedical financial interests or
potential conflicts of interest.
Supplementary material cited in this article is available
online.
1. Kendler KS, Prescott CA (2006) Genes, Environment and Psychopathol-
ogy. New York: Guildford Press.
2. Uher R, McGuffin P (2008): The moderation by the serotonin transporter
gene of environmental adversity in the aetiology of mental illness:
Review and methodological analysis. Mol Psychiatry 13:131–146.
3. Caspi A, Sugden K, Moffitt TE, Taylor A, Craig IW, Harrington H, et al.
(2003): Influence of life stress on depression: Moderation by a polymor-
phism in the 5-HTT gene. Science 301:386 –389.
4. Thompson WD (1991): Effect modification and the limits of biological
inference from epidemiologic data. J Clin Epidemiol 44:221–232.
5. Risch NJ (2000): Searching for genetic determinants in the new millen-
nium. Nature 405:847– 856.
6. Rothman KJ, Greenland S (1998): Modern Epidemiology, 2nd ed. Philadel-
phia: Lippincott Williams & Wilkins.
7. Araya R, et al. (under review): Effects of stressful life events, maternal
depression and 5-HTTLPR genotype on emotion symptoms in pre-ado-
lescent children.
8. Middeldorp CM, Boomsma DI (in press): Genetics and psychopathology.
In: Berntson GG, Cacioppo JT, editors. Handbook of Neuroscience for
Behavioral Sciences. New York: Wiley.
9. Heils A, Teufel A, Petri S, Stöber G, Riederer P, Bengel D, Lesch KP (19961):
Allelic variation of human serotonin transporter gene expression. J Neu-
rochem 66:2621–2624.
10. Eley TC, Sugden K, Corsico A, Gregory AM, Sham P, McGuffin P, et al.
(2004): Gene-environment interaction analysis of serotonin system
markers with adolescent depression. Mol Psychiatry 9:908 –915.
11. Kaufman J, Yang BZ, Douglas-Palumberi H, Houshyar S, Lipschitz D,
Krystal JH, Gelernter J (2004): Social supports and serotonin trans-
porter gene moderate depression in maltreated children. Proc Natl
Acad Sci U S A 101:17316 –17321.
12. Gillespie NA, Whitfield JB, Williams B, Heath AC, Martin NG (2005): The
relationship between stressful life events, the serotonin transporter
(5-HTTLPR) genotype and major depression. Psychol Med 35:101–111.
13. Grabe HJ, Lange M, Wolff B, Völzke H, Lucht M, Freyberger HJ, et al.
(2005): Mental and physical distress is modulated by a polymorphism in
the 5-HT transporter gene interacting with social stressors and chronic
disease burden. Mol Psychiatry 10:220 –224.
14. Jacobs N, Kenis G, Peeters F, Derom C, Vlietinck R, van Os J (2006):
Stress-related negative affectivity and genetically altered serotonin
transporter function: Evidence of synergism in shaping risk of depres-
sion. Arch Gen Psychiatry 63:989 –996.
15. Kaufman J, Yang BZ, Douglas-Palumberi H, Grasso D, Lipschitz D,
Houshyar S, et al. (2006): Brain-derived neurotrophic factor-5-HTTLPR
gene interactions and environmental modifiers of depression in chil-
dren. Biol Psychiatry 59:673– 680.
16. Sjoberg RL, Nilsson KW, Nordquist N, Ohrvik J, Leppert J, Lindström L,
Oreland L (2006): Development of depression: Sex and the interaction
M.R. Munafò et al.
between environment and a promoter polymorphism of the serotonin
transporter gene. Int J Neuropsychopharmacol 9:443– 449.
17. Surtees PG, Wainwright NW, Willis-Owen SA, Luben R, Day NE, Flint J
(2006): Social adversity, the serotonin transporter (5-HTTLPR) polymor-
phism and major depressive disorder. Biol Psychiatry 59:224 –229.
18. Taylor SE, Way BM, Welch WT, Hilmert CJ, Lehman BJ, Eisenberger NI
(2006): Early family environment, current adversity, the serotonin trans-
porter promoter polymorphism, and depressive symptomatology. Biol
Psychiatry 60:671– 676.
19. Wilhelm K, Mitchell PB, Niven H, Finch A, Wedgwood L, Scimone A, et al.
(2006): Life events, first depression onset and the serotonin transporter
gene. Br J Psychiatry 188:210 –215.
20. Cervilla JA, Molina E, Rivera M, Torres-González F, Bellón JA, Moreno B, et
al. (2007): The risk for depression conferred by stressful life events is
modified by variation at the serotonin transporter 5HTTLPR genotype:
Evidence from the Spanish PREDICT-Gene cohort. Mol Psychiatry 12:
748 –755.
21. Kim JM, Stewart R, Kim SW, Yang SJ, Shin IS, Kim YH, Yoon JS (2007):
Interactions between life stressors and susceptibility genes (5-HTTLPR
and BDNF) on depression in Korean elders. Biol Psychiatry 62:423– 428.
22. Kendler KS, Kuhn JW, Vittum J, Prescott CA, Riley B (2005): The interac-
tion of stressful life events and a serotonin transporter polymorphism in
the prediction of episodes of major depression: A replication. Arch Gen
Psychiatry 62:529 –535.
23. Scheid JM, Holzman CB, Jones N, Friderici KH, Nummy KA, Symonds LL,
et al. (2007): Depressive symptoms in mid-pregnancy, lifetime stressors
and the 5-HTTLPR genotype. Genes Brain Behav 6:453– 464.
24. Fox NA, Nichols KE, Henderson HA, Rubin K, Schmidt L, Hamer D, et al.
(2005): Evidence for a gene-environment interaction in predicting be-
havioral inhibition in middle childhood. Psychol Sci 16:921–926.
BIOL PSYCHIATRY 2009;65:211–219 219
25. Gibb BE, McGeary JE, Beevers CG, Miller IW (2006): Serotonin transporter
(5-HTTLPR) genotype, childhood abuse, and suicide attempts in adult
psychiatric inpatients. Suicide Life Threat Behav 36:687– 693.
26. Zalsman G, Huang YY, Oquendo MA, Burke AK, Hu XZ, Brent DA, et al.
(2006): Association of a triallelic serotonin transporter gene promoter
region (5-HTTLPR) polymorphism with stressful life events and severity
of depression. Am J Psychiatry 163:1588 –1593.
27. Gunthert KC, Conner TS, Armeli S, Tennen H, Covault J, Kranzler HR
(2007): Serotonin transporter gene polymorphism (5-HTTLPR) and anx-
iety reactivity in daily life: A daily process approach to gene-environ-
ment interaction. Psychosom Med 69:762–768.
28. Stein MB, Schork NJ, Gelernter J (2008): Gene-by-environment (seroto-
nin transporter and childhood maltreatment) interaction for anxiety
sensitivity, an intermediate phenotype for anxiety disorders. Neuropsy-
chopharmacology 33:312–319.
29. Clayton D, McKeigue PM (2001): Epidemiological methods for studying
genes and environmental factors in complex diseases. Lancet 358:
1356 –1360.
30. Eaves LJ (2006): Genotype ⫻ Environment interaction in psychopathol-
ogy: Fact or artifact? Twin Res Hum Genet 9:1– 8.
31. Lotrich FE, Pollock BG (2004): Meta-analysis of serotonin transporter
polymorphisms and affective disorders. Psychiatr Genet 14:121–129.
32. Willis-Owen SA, Turri MG, Munafò MR, Surtees PG, Wainwright NW,
Brixey RD, Flint J (2005): The serotonin transporter length polymor-
phism, neuroticism, and depression: A comprehensive assessment of
association. Biol Psychiatry 58:451– 456.
33. Munafo MR, Flint J (2004): Meta-analysis of genetic association studies.
Trends Genet 20:439 – 444.
34. Colhoun HM, McKeigue PM, Davey Smith G (2003): Problems of
reporting genetic associations with complex outcomes. Lancet 361:
865– 872.
www.sobp.org/journal