International Journal of Epidemiology, 2018, 1–13

doi: 10.1093/ije/dyy015
Original article

Original article

Using structural equation modelling to jointly

estimate maternal and fetal effects on
birthweight in the UK Biobank
Nicole M Warrington,1 Rachel M Freathy,2,3 Michael C Neale4 and
David M Evans1,3,5*
1
University of Queensland Diamantina Institute, The University of Queensland, Translational Research
Institute, Brisbane, QLD, Australia, 2Institute of Biomedical and Clinical Science, University of Exeter
Medical School, University of Exeter, Royal Devon and Exeter Hospital, Exeter, UK, 3Medical Research
Council Integrative Epidemiology Unit, University of Bristol, Bristol, UK, 4Virginia Institute for
Psychiatric and Behavioral Genetics, Departments of Psychiatry and Human & Molecular Genetics,
Virginia Commonwealth University, Richmond, VA, USA, 5School of Social and Community Medicine,
University of Bristol, Bristol, UK
*Corresponding author. MRC Integrative Epidemiology Unit, University of Bristol, Oakfield House, Oakfield Grove, Clifton,
BS82BN, UK. E-mail: [email protected]
Editorial decision 16 January 2018; Accepted 25 January 2018

Abstract
Background: To date, 60 genetic variants have been robustly associated with
birthweight. It is unclear whether these associations represent the effect of an
individual’s own genotype on their birthweight, their mother’s genotype, or both.
Methods: We demonstrate how structural equation modelling (SEM) can be used to
estimate both maternal and fetal effects when phenotype information is present for
individuals in two generations and genotype information is available on the older individ-
ual. We conduct an extensive simulation study to assess the bias, power and type 1 error
rates of the SEM and also apply the SEM to birthweight data in the UK Biobank study.
Results: Unlike simple regression models, our approach is unbiased when there is both a
maternal and a fetal effect. The method can be used when either the individual’s own
phenotype or the phenotype of their offspring is not available, and allows the inclusion
of summary statistics from additional cohorts where raw data cannot be shared. We
show that the type 1 error rate of the method is appropriate, and that there is substantial
statistical power to detect a genetic variant that has a moderate effect on the phenotype
and reasonable power to detect whether it is a fetal and/or a maternal effect. We also
identify a subset of birthweight-associated single nucleotide polymorphisms (SNPs) that
have opposing maternal and fetal effects in the UK Biobank.
Conclusions: Our results show that SEM can be used to estimate parameters that would
be difficult to quantify using simple statistical methods alone.

C The Author(s) 2018. Published by Oxford University Press on behalf of the International Epidemiological Association.
V 1
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits
unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 2 International Journal of Epidemiology, 2018, Vol. 0, No. 0

Key words: Structural equation model, maternal effects, fetal effects, birthweight, UK Biobank

Key Messages

• We describe a structural equation model to estimate both maternal and fetal effects when phenotype information is

present for individuals in two generations and genotype information is available on the older individual.
• Using simulation, we show that our approach is unbiased when there is both a maternal and fetal effect, unlike sim-

ple linear regression models. Additionally, we illustrate that the structural equation model is largely robust to random
measurement error and missing data for either the individual’s own phenotype or the phenotype of their offspring.
• We describe how the flexibility of the structural equation modelling framework will allow the inclusion of summary

statistics from studies that are unable to share raw data.

• Using the structural equation model to estimate the maternal and fetal effects of known birthweight-associated loci in

the UK Biobank, we identify three loci that have primary effects through the maternal genome and six loci that have
opposite effects in the maternal and fetal genomes.

Introduction the GCK gene, which cause a defect in the sensing of glu-
Birthweight is a complex trait, and low birthweight is ro- cose by the pancreas, have radically different associations
bustly associated with increased risk of a range of cardio- with birthweight according to their parent of origin.
metabolic diseases in later life.1 It has long been known If inherited paternally, birthweight is lower due to reduced
that birthweight is under the influence of both maternal glucose sensing and consequent reduced insulin secretion,
and fetal genetic sources of variation. Using a large sample which results in reduced growth. But if maternally in-
consisting of the offspring of twins, Magnus illustrated herited (i.e. present in both mother and fetus), birthweight
that more than 50% of the variation in birthweight is is close to the population average because the maternal
caused by fetal genes and less than 20% was caused by ma- hyperglycaemia compensates for the fetal defect in glucose
ternal genes.2 Subsequent studies have reported lower pro- sensing. In the case that the mother has hyperglycaemia
portions of the variance explained by both fetal and due to a GCK mutation, but the fetus does not inherit the
maternal genes, but all have shown that the fetal contribu- mutation, the birthweight is higher due to normal glucose
tion is larger than the maternal contribution.3,4 Using a sensing and thus above-average insulin secretion. This ex-
method that partitions trait variance into components due ample reflects contrasting effects mediated through the
to the maternal and fetal genomes,5 we reported that com- intrauterine environment (i.e. maternal effects) and direct
mon genetic variants in the fetal genome explained effects of the offspring’s genotype.8
approximately 28% of the variation in birthweight, In an attempt to resolve this question, in Horikoshi
whereas common genetic variants in the maternal genome et al.6 we first performed a simple linear regression of an
only explained approximately 8% of the total variance.6 individual’s self-reported birthweight on their own geno-
We and others have begun to investigate the specific type; and then for the UK Biobank women, we performed
regions of the genome that influence fetal growth, using a linear regression of the birthweight of their firstborn
genome-wide association studies (GWAS). In a recent child on their own genotype. We then compared the ma-
GWAS meta-analysis combining data from the Early ternal and fetal effect sizes to get an idea of whether the
Growth Genetics consortium (EGG; http://egg-consortium. locus was operating through the maternal or the individ-
org/) and the UK Biobank,7 we identified 60 single nucleo- ual’s own genotype. However, this approach was subopti-
tide polymorphisms (SNPs) associated with birthweight at mal since it did not consider the correlation between
genome-wide levels of significance.6 One difficulty we maternal and offspring genotypes, and therefore did not
faced in interpreting our results was that it was often not accurately estimate the relative importance of these two
clear whether genetic associations reflected the effect of an potential sources of variation. We also examined the gen-
individual’s own genotype on their birthweight, an effect etic associations with birthweight in cohorts that had
of their mother’s genotype on their birthweight (i.e. mater- genotype information on both mother and offspring.
nal genotype mediated through the intrauterine effect) or Performing an analysis of offspring birthweight on mater-
some combination of both. For example, rare mutations in nal genotype and conditioning on offspring genotype

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 International Journal of Epidemiology, 2018, Vol. 0, No. 0 3

Figure 1. Diagram of the structural equation model (SEM) used for the simulation study and the UK Biobank analysis of birthweight. The three observed
variables (in squares) are the birthweight of the individual (BW), the birthweight of their offspring (BWO) and the genotype of the individual (SNP). The
latent variables (in circles) are the genotypes for the individual’s mother (GG) and the genotype of the individual’s first offspring (GO). The total variance
of the latent genotypes for the individual’s mother (GG) and offspring (GO) and for the observed SNP variable is set to U [i.e. variance(GG) ¼ U, variance
(SNP) ¼ 0.75U þ 0.25U, variance (GO) ¼ 0.75U þ 0.25U]. The m and f path coefficients refer to maternal and fetal effects, respectively. The residual error
terms for the birthweight of the individual and their offspring are represented by E and EO, respectively, and we estimate the variance of both of these
terms in the SEM. The covariance between residual genetic and environmental sources of variation is given by q.

should yield an unbiased estimate of the mother’s genetic association when maternal and fetal effects operate in op-
influence on her child’s birthweight, and likewise a regres- posite directions. We also show how our framework can
sion of offspring birthweight on offspring genotype condi- easily combine summary results data from additional co-
tioning on maternal genotype should produce an unbiased horts, including previous large scale GWAS meta-analyses,
estimate of the fetal contribution on birthweight. The dif- involving either maternal or offspring phenotypes. Using
ficulty however is that there is a paucity of cohorts in the the UK Biobank data,7 we provide strong evidence to sug-
world that have birthweight data as well as genotype data gest that several of the known birthweight-associated SNPs
on both mothers and children, meaning that such an ana- exert effects acting in opposite directions on birthweight
lysis is likely to have low power to resolve maternal and through the maternal and fetal genotypes.
fetal effects.
To better estimate the maternal and fetal genetic contri-
butions to birthweight for each of the 60 genome-wide sig- Methods
nificant variants reported in Horikoshi et al.6 we used a
structural equation modelling (SEM) approach with birth- Simulations
weight data from the UK Biobank. Our method enables us We performed simulations to investigate the bias, power and
to model both grand-maternal and offspring genotypes type one error rate of the SEM for modelling both the indi-
(which were absent in the UK Biobank) as latent factors, vidual’s own genetic effect (referred to as the ‘fetal effect’)
and to estimate maternal and fetal effects on birthweight in and maternal genetic effects on birthweight. The model we
the same statistical model. To investigate the properties of used for generating the data is illustrated in Figure 1, and the
our approach, we first performed a series of simulations to: R code used for performing these simulations is provided in
(i) quantify any bias in the effect estimates for the maternal the Supplementary material (available as Supplementary
and fetal effects; and (ii) estimate power to detect maternal data at IJE online).
and fetal effects and type 1 error. We also assessed the ef- For each scenario, we generated 10 000 replicates of
fect of allele frequency and measurement error in birth- 30 000 maternal-offspring pairs. For each replicate we gen-
weight (which can often be an issue with self-report) on erated grandparental (on the maternal side) and paternal
our estimates. We show that our method provides accurate genotypes at a single locus. Assuming autosomal Mendelian
estimates of maternal and fetal effects under a range of dif- inheritance, additivity and unit variance, latent variables for
ferent scenarios, and increased power to detect genetic the genotype of the individual’s mother (i.e. grand-maternal

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 4 International Journal of Epidemiology, 2018, Vol. 0, No. 0

genotype; GG), the individual’s own genotype (SNP) and iv. SEM estimating only the maternal effect, to conduct a
offspring’s genotype (GO) were generated. The individual’s likelihood ratio test for the fetal effect; this likelihood
own birthweight variable (BW), for each family i, was ratio test P-value was compared with the Wald test
generated using the following equation: P-value for the maternal effect;
v. SEM with neither fetal nor maternal paths (i.e. both
pffiffiffiffiffiffiffi pffiffiffiffiffiffiffi fixed to zero); this model was fit to conduct a
BWi ¼ VM  GGi þ VO  SNPi þ bU  Ui þ ei
likelihood ratio test of the overall SNP effect, and the
where VM denotes the variance in birthweight explained by P-value from this test is referred to as the two degrees
the maternal genotype (‘maternal effect’), GG is a latent vari- of freedom (2DF) test P-value.
able indexing the genotype of the individual’s mother, VO is Bias was defined as the mean difference between the
the variance in birthweight explained by the individual’s estimated SNP effect and the true parameter across the
own genotype (‘fetal effect’), SNP is the genotype of the indi- 10 000 simulations, and was calculated for both the ma-
vidual, U is a standard normal random variable representing ternal and fetal effects. A 95% confidence interval was
all residual genetic and environmental sources of similarity calculated around the bias to give an indication of the un-
between mother and offspring, bU is the total effect of U on certainty in the estimate. Power was defined as the propor-
the individual’s own birthweight and E is a random normal tion of tests that reached P < 0.05 under the alternative
variable with mean zero and variance needed to ensure that hypothesis, and type 1 error rate the proportion of tests
BW has unit variance asymptotically. that reached P < 0.05 under the null hypothesis.
Similarly, offspring birthweight (BWO), for each family
i, was generated using the following equation:
Additional simulations investigating
pffiffiffiffiffiffiffi pffiffiffiffiffiffiffi
BWOi ¼ VM  SNPi þ VO  GOi þ bUO  Ui þ eOi measurement error
In the UK Biobank, female participants were asked to re-
where GO is a latent variable indexing the offspring geno-
port the birthweight of their first offspring to the nearest
type, bUO is the total effect of U on offspring birthweight and
pound. After appropriate data cleaning, this left six
EO is a random normal variable with mean zero and variance
discrete birthweight values for the offspring (see below).
needed to ensure that BWO has unit variance asymptotically.
We therefore conducted a second set of simulations to in-
In all simulations, the regression of phenotype on residual
vestigate the effect of this type of measurement error, using
shared genetic and environmental factors was set to 0.5
the same method as described above but rounding the
(i.e. bU ¼ bUO ¼ 0.5). We considered the effects of: allele fre-
birthweight of the offspring to the nearest unit.
quency (p ¼ 0.99, p ¼ 0.90 or p ¼ 0.5); the strength of the
Given that birthweights of both individuals and their
fetal genetic effect on birthweight (VO ¼ 0%, VO ¼ 0.02% or
offspring are self-reported in the UK Biobank, we also as-
VO ¼ 0.04%); and the strength of the maternal genetic effect
sessed the potential effect of measurement error on both
on birthweight (VM ¼ 0%, VM ¼ 0.01% or VM ¼ 0.02%).
variables. To do this, we added a normally distributed error
We simulated the fetal and maternal genetic effects to have
component to both simulated birthweight measurements,
both increasing and decreasing effects on birthweight.
which is referred to as discrimination or classical measure-
For each simulated dataset we fit a series of models:
ment error.9 For example, an individual’s own birthweight
i. linear models regressing either the individual’s own with measurement error was generated as follows:
birthweight or, for the women, the birthweight of their
offspring on the SNP (individual’s own genotype), BWi ¼ BWi þ si ; si  Nð0; d Þ
which respectively estimate the fetal and maternal gen-
where d* was chosen to produce a specific R2 value for the
etic effects on birthweight; this is equivalent to the
regression of BW* on BW, using the following equation:
model typically used in genetic studies of birthweight6
and was used for comparison purposes;
VarðBWÞ
ii. SEM estimating both maternal and fetal effects as illus- R2 ¼
VarðBW Þ þ VarðsÞ
trated in Figure 1; P-values were calculated using
Wald tests; BWoi was generated in the same way. We varied the value
iii. SEM estimating only the fetal effect; this model was fit of R2 (1.00, 0.75, 0.50, 0.25), where lower values of R2 repre-
to conduct a likelihood ratio test for the maternal ef- sent increasing measurement error. We used a subset of mater-
fect and the likelihood ratio test P-value was compared nal and fetal effect sizes to get an idea of whether the impact
with the Wald test P-value for the fetal effect; of measurement error is influenced by effect size; neither a

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 International Journal of Epidemiology, 2018, Vol. 0, No. 0 5

maternal or fetal effect (VO ¼ 0% ¼ VM), large fetal effect individuals who were part of multiple births and were excluded
and no maternal effect (VO ¼ 0.04%, VM ¼ 0%), no fetal ef- from the analyses. Of the 9034 individuals who reported their
fect and a large maternal effect (VO ¼ 0%, VM ¼ 0.02%), own birthweight at both baseline and follow-up, 401 (4% of
large fetal and maternal effect (VO ¼ 0.04%, VM ¼ 0.02%) individuals with repeat birthweight reports) were excluded be-
and large fetal and maternal effect in opposite directions. cause the two values differed by more than 0.5 kg. For those
All simulations were conducted with an allele frequency of individuals who reported different values between baseline and
p ¼ 0.5. follow-up (<0.5 kg) we took the baseline measure for the ana-
lyses. Finally, we excluded individuals who reported their
own birthweight to be <2.5 kg or >4.5 kg [24 138 (9%)
Additional simulations investigating missing data individuals with birthweight <2.5 kg and 14 065 (5%)
We also assessed the impact of when individuals did not have individuals with birthweight >4.5 kg], as these are implausible
both their own and their offspring birthweight available. for live term births before 1970. In total, 233 662 individuals
Supplementary Figure 1 (available as Supplementary data at had data on their own birthweight matching our inclusion
IJE online) illustrates the three components of the SEM used criteria.
to incorporate individuals with missing data; the first compo- Women in the UK Biobank were also asked to report the
nent models individuals with complete data, the second birthweight of their first child to the nearest pound. We used
component models genotyped individuals who report their the same inclusion criteria as for their own birthweight, leav-
own phenotype but not their offspring’s, and the third com- ing 210 405 individuals with birthweight of their first child
ponent models genotyped mothers who report their offspring [51 (0.6%) excluded because the multiple reports of off-
phenotype data but not their own. These three components spring birthweight differed by >0.5 kg; 5838 (3%) excluded
are fit to the three subsets of data and then the likelihoods with offspring birthweight <2.5 kg; and 473 (0.2%)
from each model are combined. Modelling the data in this excluded with offspring birthweight >4.5 kg], 109 205 of
way avoids list-wise deletion of cases due to missing pheno- whom had also reported their own birthweight.
type information and makes maximum use of the observed Genotype data from the May 2015 release were available
data. If data are missing at random then our full information on a subset of 152 248 individuals. In addition to the quality
maximum likelihood approach returns asymptotically un- control metrics performed centrally by the UK Biobank, we
biased parameter estimates, and the most precise estimates excluded individuals who were related. We defined a subset of
that have this property.10 We simulated four additional scen- ‘White European’ ancestry samples using a K-means (K ¼ 4)
arios, all with minor allele frequency of p ¼ 0.5 and a total clustering approach based on the first four genetically deter-
sample size of 30 000 individuals: (i) 15 000 individuals mined principal components. A subset of 89 296 unrelated
with both their own and their offspring’s birthweight and individuals with genotype data, a valid birthweight for them-
15 000 individuals with their own birthweight only; (ii) selves or their first child and genetically of ‘White European’
15 000 individuals with both their own and their offspring’s ancestry were included in the analysis. Of these, 24 962 were
birthweight and 15 000 individuals with their offspring’s men who only reported their own birthweight. Among the
birthweight only; (iii) 15 000 individuals with both their own women, 8723 reported only their own birthweight, 24 645 re-
and their offspring’s birthweight, 7500 individuals with their ported only that of their first child and 30 966 reported both.
own birthweight only and 7500 individuals with their off- We adjusted both the individual’s own birthweight and the
spring’s birthweight only; and (iv) 15 000 individuals with birthweight of their first child for the principal components
their own birthweight only and 15 000 individuals with their that were associated with birthweight, adjusted the individ-
offspring’s birthweight only (i.e. no individuals with both ual’s own birthweight for sex (sex was not reported for the
birthweight measures, and therefore only the second and offspring) and then created z-scores. A subset of 58 autosomal
third components of Supplementary Figure 1 are fit and the SNPs out of the 60 birthweight-associated SNPs6 were ex-
term q in Figure 1 can not be estimated). Given that we tracted from the imputed files provided by UK Biobank and
observed very close correspondence between the likelihood aligned to the birthweight-increasing allele (rs62240962 was
ratio and Wald tests, we only conducted Wald tests because not available and rs11096402 is on the X chromosome).
these were computationally easier to perform. We fit the SEM to the data from each of these 58 autosomal
SNPs to estimate the maternal and fetal genetic effects on
birthweight. To confirm our results, we compared them with
UK Biobank those from a conditional linear regression model in a subset
UK Biobank phenotype data were available on 502 643 indi- of 12 909 individuals with both maternal and offspring
viduals, of whom 279 959 reported their own birthweight at genotype data from the EGG consortium, as presented in
either the baseline or follow-up visits. There were 7693 Horikoshi et al.6 Specifically, Horikoshi et al.6 reported:

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 6 International Journal of Epidemiology, 2018, Vol. 0, No. 0

Figure 2. Bias in effect estimates with an allele frequency of 0.5 and varying maternal and fetal effect sizes using two linear models that assess the
maternal and fetal effects independently (left panel), or the structural equation model (SEM, right panel) assessing both the maternal and fetal effects
simultaneously.

(i) the association between maternal genotype and offspring maternal effect, then the fetal effect estimated from the lin-
birthweight after conditioning on offspring genotype (i.e. ear model is unbiased. The same pattern of bias occurs for
their estimate of the maternal effect); and (ii) the association the maternal effect estimates. Conversely, the SEM is un-
between offspring genotype and offspring birthweight after biased for both the maternal and fetal effects as it simultan-
conditioning on maternal genotype (i.e. their estimate of the eously models both effects. The bias and 95% confidence
fetal effect). intervals for all simulation results are presented in
Supplementary Table 1 (available as Supplementary data at
Results IJE online).
When there was measurement error in either the
Bias individual’s own birthweight or the birthweight of their
Figure 2 shows the bias calculated from the simulations for offspring, the estimates of maternal and fetal effects were
the linear model and the SEM in the simulations with allele unbiased (Table 1 for abridged results, and full results in
frequency of 0.5 and all 30 000 individuals whom had com- Supplementary Table 2, available as Supplementary data
plete data for both their own birthweight and the birth- at IJE online). However, there was a decrease in the preci-
weight of their offspring. The fetal effect estimates from the sion of the estimate (i.e. increase in the standard error) as
standard linear model are biased wherever there is a mater- the measurement error increased (Table 1 for abridged re-
nal effect that is not being modelled. For example, in the sults, and full results in Supplementary Table 2). For a
scenarios where there is both a fetal and maternal effect, the small number of scenarios where the birthweight of the
estimated fetal effect approximately equals the true fetal ef- offspring was distributed as it is in the UK Biobank, a
fect plus half the true maternal effect. In other words, the small bias was introduced from the SEM (Supplementary
bias of the estimated fetal effect is approximately half the Table 3, available as Supplementary data at IJE online);
true maternal effect. In the scenarios where there is no this bias differed across allele frequency and true effect size

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
by guest
on 19 February 2018
Table 1. The effect of measurement error in the individual’s own birthweight and the birthweight of their offspring on bias, precision and power in the structural equation model
(SEM)
pffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffi
Measurement error V O ¼ 0; V M 5 0 V O ¼ 0; V M 5 0.014 V O ¼  0:020; V M 5 0 V O ¼  0:020; V M 5 0.014 V O ¼ 0:020; V M 5 0.014

R2 BW R2 BWO Mean Mean Power Mean Mean Power Mean Mean Power Mean Mean Power Mean Mean Power
estimate standard estimate standard estimate standard estimate standard estimate standard
error error error error error

Estimated fetal effect

Measurement error for offspring birthweight only
1.00 1.00 0.000 0.008 0.048 0.000 0.008 0.048 0.020 0.008 0.738 0.020 0.008 0.738 0.020 0.008 0.741
1.00 0.75 0.000 0.008 0.050 0.000 0.008 0.049 0.020 0.008 0.711 0.020 0.008 0.712 0.020 0.008 0.702
1.00 0.50 0.000 0.009 0.048 0.000 0.009 0.051 0.020 0.009 0.655 0.020 0.009 0.656 0.020 0.009 0.634
1.00 0.25 0.000 0.010 0.048 0.000 0.010 0.048 0.020 0.010 0.515 0.020 0.010 0.516 0.020 0.010 0.490
1.00 UKBBa 0.000 0.008 0.047 0.000 0.008 0.047 0.020 0.008 0.733 0.020 0.008 0.736 0.020 0.008 0.732
Measurement error for individual’s own birthweight only
1.00 1.00 0.000 0.008 0.048 0.000 0.008 0.048 0.020 0.008 0.738 0.020 0.008 0.738 0.020 0.008 0.741
International Journal of Epidemiology, 2018, Vol. 0, No. 0

0.75 1.00 0.000 0.009 0.048 0.000 0.009 0.049 0.020 0.009 0.616 0.020 0.009 0.618 0.020 0.009 0.612
0.50 1.00 0.000 0.011 0.047 0.000 0.011 0.047 0.020 0.011 0.448 0.020 0.011 0.449 0.020 0.011 0.453
0.25 1.00 0.000 0.015 0.046 0.000 0.015 0.046 0.020 0.015 0.248 0.020 0.015 0.248 0.020 0.015 0.253
Estimated maternal effect
Measurement error for offspring birthweight only
1.00 1.00 0.000 0.008 0.052 0.014 0.008 0.456 0.000 0.008 0.052 0.014 0.008 0.456 0.014 0.008 0.457
1.00 0.75 0.000 0.009 0.047 0.014 0.009 0.350 0.000 0.009 0.047 0.014 0.009 0.351 0.014 0.009 0.351

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

1.00 0.50 0.000 0.011 0.046 0.014 0.011 0.245 0.000 0.011 0.046 0.014 0.011 0.246 0.014 0.011 0.245
1.00 0.25 0.000 0.015 0.047 0.014 0.015 0.142 0.000 0.015 0.047 0.014 0.015 0.142 0.014 0.015 0.142
1.00 UKBBa 0.000 0.008 0.052 0.014 0.008 0.426 0.000 0.008 0.052 0.014 0.008 0.424 0.014 0.008 0.433
Measurement error for individual’s own birthweight only
1.00 1.00 0.000 0.008 0.052 0.014 0.008 0.456 0.000 0.008 0.052 0.014 0.008 0.456 0.014 0.008 0.457
0.75 1.00 0.000 0.008 0.050 0.014 0.008 0.418 0.000 0.008 0.051 0.014 0.008 0.418 0.014 0.008 0.418
0.50 1.00 0.000 0.009 0.051 0.014 0.009 0.375 0.000 0.009 0.051 0.014 0.009 0.374 0.014 0.009 0.375
0.25 1.00 0.000 0.010 0.050 0.014 0.010 0.283 0.000 0.010 0.051 0.014 0.010 0.283 0.014 0.010 0.283

a
UKBB indicates offspring birthweight distributed as in the UK Biobank, where only six discrete values are available.
7
 8 International Journal of Epidemiology, 2018, Vol. 0, No. 0

Figure 3. Power of the two degrees of freedom test using the structural equation model (SEM) assessing both the maternal and fetal effects simultan-
eously, and power of the two linear models (LM) that assess the maternal and fetal effects independently. Note, power from the linear models is artifi-
cially inflated due to the bias in the effect estimate, but they are presented here as a comparison with what would be provided from a standard
genetic analysis of birthweight. Power is presented for simulations with a minor allele frequency of 0.5.

for both the maternal and fetal effects, and no clear pattern Supplementary Table 1). For example, the power is greater
was observed. The bias was less than 4% of the true value for the fetal effect estimated using the linear model over
in all scenarios and substantially lower than the bias intro- the Wald test from the SEM when the maternal effect is
duced in the linear models. zero. Nevertheless, there is still substantial power to detect
In simulations where either the individual’s own birth- an effect using the Wald test in the SEM with a ¼ 0.05,
weight or that of their offspring was missing, the SEM con- with 74% power to detect a variant that explains 0.04%
tinued to produce unbiased estimates (Supplementary Table of the variance, 45% power for one explaining 0.02% of
1). However, in the simulations where all individuals only the variance and 25% power for one explaining 0.01%
had their own birthweight or the birthweight of their off- of the variance, in a sample of 30 000 individuals with
spring (i.e. no individuals had both birthweight measures), both their own and their offspring’s birthweight. However,
we detected a small bias in the maternal effect estimate (bias the two degrees of freedom test has very similar power to
approximately 0.0003, or less than 3% of the true value). the linear model when the SNP had either a fetal or mater-
nal effect only, and greater power in most scenarios than
testing either maternal or fetal effects individually using
Power/type 1 error the Wald test (Figure 3 and Supplementary Table 1). It is
The power and type 1 error results for all simulations from worth nothing that the power estimates from the linear
the SEM and the linear models for the fetal and maternal models are artificially inflated due to the bias introduced in
effects are presented in Supplementary Table 1. the linear models; however, we have included them in the
The linear model has greater power than the Wald test figure as they give an indication of what the power of a
in the SEM when the SNP has either a fetal or maternal standard genetic analysis would be. This indicates that the
effect only (i.e. when the effect estimate is unbiased; SEM can detect when a SNP affects birthweight, but it has

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 International Journal of Epidemiology, 2018, Vol. 0, No. 0 9

Figure 4. Power from the structural equation model (SEM) with different combinations of individuals reporting their own birthweight (BW) or their off-
spring’s birthweight (BWO). Power for the fetal effect is presented from the simulations where there is no maternal effect; however, similar estimates
were obtained when there was a maternal effect (see Supplementary Table 1 for full results). Similarly, for the maternal effect, results are presented
from simulations where there is no fetal effect. Power is presented for simulations with a minor allele frequency of 0.5.

lower power to detect whether the effect is driven by the there is only a small decrease in power to detect the fetal effect
mother or the offspring. For example, when the variant ex- when information on the offspring birthweight is not available
plains 0.04% of the variance using the individual’s own in 50% of the individuals, and for power to detect the mater-
genotype and 0.02% of the variance using the mother’s nal effect when the individual’s own birthweight is not avail-
genotype, with a ¼ 0.05, the SEM has 74% power to de- able in 50% of the individuals. Interestingly, the SEM can still
tect the fetal effect, 45% power to detect the maternal ef- be used to estimate maternal and fetal effects when the sample
fect and 100% power to detect any effect of the variant consists of some individuals only measured on their own
using the two degrees of freedom test in a sample of 30 000 birthweight, and others who have only reported the birth-
individuals with both their own and their offspring’s birth- weight of their offspring. However, the power to detect either
weight. Similarly, with a ¼ 0.05 and 30 000 individuals a fetal or a maternal effect is approximately half of that when
with complete data, when the variant explains 0.02% of birthweight data are available on both individuals in the pair.
the variance using the individual’s own genotype and As expected, the type 1 error from the linear model is
0.01% of the variance using the mother’s genotype, the inflated in situations where the estimated effect is biased
SEM has 45% power to detect the fetal effect, 25% power (Supplementary Table 1). However, the type 1 error is well
to detect the maternal effect and 95% power to detect any controlled when using the SEM. It remains controlled
effect of the variant using the two degrees of freedom test. when birthweight of the offspring is distributed as in the
Power for both fetal and maternal effects is reduced when UK Biobank (Supplementary Table 3), when there is meas-
there is measurement error in either the individual’s own birth- urement error in either the individual’s own birthweight or
weight or the birthweight of their offspring, due to the de- the birthweight of their offspring (Table 1, Supplementary
crease in precision of the estimate (Table 1 and Supplementary Table 2) or when data are not available on both the indi-
Table 2). This decrease in power was the same across the dif- viduals own birthweight and their offspring’s birthweight
ferent true effect sizes for the fetal and maternal effects. (Supplementary Table 1).
Figure 4 shows the power when not all individuals have The difference between P-values estimated using the
complete data for both maternal and offspring birthweight. Wald test and the likelihood ratio test in the SEM was neg-
Power is greatest when information is available on both the ligible (Supplementary Table 1 for mean difference), indi-
individual’s own and their offspring’s birthweight; however, cating that the Wald test was adequate.

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 10 International Journal of Epidemiology, 2018, Vol. 0, No. 0

evident that most of the 58 SNPs only have evidence for a

fetal effect, which is unsurprising given how the SNPs were
selected. Three SNPs primarily have a maternal effect
(EBF1, ACTL9 and MTNR1B) and eight SNPs have
evidence for both. Perhaps the most interesting SNPs are
those where the birthweight-increasing allele identified in
Horikoshi et al.6 has opposite effects on birthweight
through the fetal and maternal genotype, half of which are
known type 2 diabetes loci (HHEX-IDE, CDKAL1,
ADCY5 and ANK1-NXK6-3).
Supplementary Table 4 and Supplementary Figure 3 (avail-
able as Supplementary data at IJE online) present the full re-
sults from the two linear models (fetal and maternal effects)
and the SEM. These results show that for SNPs where the ma-
ternal and fetal effects go in opposite directions (for example,
the HHEX-IDE, CDKAL1, ADCY5 and ANK1-NKX6-3
Figure 5. Fetal and maternal effect size estimated using the structural
equation model (SEM) for the 58 birthweight-associated SNPs in the UK
loci), the fetal effect estimated in the GWAS6 would have
Biobank. All SNPs are aligned to the birthweight-increasing allele re- been reported to be smaller than its true effect.
ported in Horikoshi et al.6 The colour of each dot indicates the maternal To confirm our results, we compared estimates of mater-
genetic association P-value for birthweight generated using the Wald
nal and fetal effects obtained from the SEM with those from
test: orange, P < 0.001; yellow, 0.001  P < 0.05; white P  0.05. Gene
names are provided for those loci with large effects. the conditional linear regression model implemented in
Horikoshi et al.6 in the subset of EGG cohorts with both ma-
Timing ternal and fetal genotype data (N ¼ 12 909 individuals).
The SEM can be fitted with either the raw data or observed Supplementary Figure 4 (available as Supplementary data at
covariance matrices. As seen in Supplementary Figure 2 IJE online) displays forest plots for the maternal and fetal ef-
(available as Supplementary data at IJE online), the compu- fects for each of the 58 birthweight-associated SNPs, using
tational time is approximately 100 times faster using the co- both the SEM in the UK Biobank and conditional linear re-
variance matrices than the raw data. There is not a gression in the EGG cohorts. The confidence intervals sur-
substantial difference in computational time between rounding the estimates from the conditional linear regression
datasets with different amounts of missing data for the analyses are larger than those from the SEM, due to the
phenotype of the individual or their offspring when fitting smaller sample size in the former study (12 909 individuals
the model using the raw data, but there is a difference for in the conditional regression analysis versus 89 296 individ-
variants with lower minor allele frequencies which take lon- uals in the SEM). Estimates obtained using both procedures
ger to run than common variants. When using covariance were similar for most SNPs. Formally, after Bonferroni cor-
matrices, however, it takes slightly longer to fit the model rection for the 58 tests, no significant heterogeneity was de-
when data are missing for either the individual or their off- tected between estimates from the SEM and estimates from
spring, because the model fits two or three sub-models sim- the conditional linear regression for either the maternal or
ultaneously (i.e. one for each of the complete data subsets: the fetal effects (heterogeneity P > 0.05/58 ¼ 9 x 1 04). The
one for individuals with both phenotypes, one for indi- largest heterogeneity between the SEM and the conditional
viduals with their own phenotype only and one for individ- regression for the maternal effect was at the ACTL9 locus
uals with their offspring’s phenotype only). The estimates (I2 ¼ 90.2%, P ¼ 0.001), where the conditional linear re-
from the model fit with the raw data are the same as those gression resulted in a (non-significant) negative estimate of
using with covariance matrices. In comparison with a linear the effect of the maternal T allele on offspring birthweight,
model, the SEM using covariance matrices takes about three whereas the SEM resulted in a positive estimate of the effect
times as long to compute with a sample size of 30 000 indi- of the same allele on offspring birthweight. However, the
viduals (Supplementary Figure 2). result from the maternal GWAS analysis presented in
Horikoshi et al.6 showed a similar direction of effect as the
SEM. It may be that there are differences between EGG and
the UK Biobank in how birthweight is measured or analysed,
UK Biobank which may be responsible for this discrepancy (for
Figure 5 presents the results from the SEM for each of the example, many studies in EGG correct birthweight for gesta-
58 birthweight-associated SNPs in the UK Biobank. It is tional age whereas this is not done in UK Biobank). Further

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 International Journal of Epidemiology, 2018, Vol. 0, No. 0 11

investigation of this locus needs to be undertaken before any To illustrate the method, we used birthweight because
strong conclusions can be drawn. there is clear evidence that both maternal and fetal effects
exist.2–4,11 However, the method could be useful for many
other phenotypes, especially pregnancy outcomes and early
Discussion developmental traits. As long as phenotype information is
In this article, we have presented a method for estimating present for individuals in two generations and genotype in-
and testing maternal and fetal effects. The approach uses formation is available on the older individual, then it is
data from mother-offspring pairs for whom genotype data possible to use this method to estimate both maternal and
are available for the mothers only and phenotype data are fetal effects. This could include phenotypes where genome-
available on both individuals. Our method is (asymptotically) wide association meta-analyses already exist, such as
unbiased when both maternal and fetal effects exist, which measures of size at birth including length12 and head cir-
improves on the traditional linear model which estimates cumference,13 maternal phenotypes during pregnancy such
each effect separately while assuming the other to be absent. as gestational weight gain,14 or developmental phenotypes
The approach is flexible and can be used when either the in- during childhood such as language development.15
dividual’s own phenotype or the phenotype of their offspring The most common study design used when trying to esti-
is not available. The ability to incorporate studies with only mate fetal and maternal effects is to have maternal/offspring
an individual’s own or their offspring’s phenotype, in com- pairs, with phenotype information on the offspring and
bination with mother/offspring pairs with complete data, will genotype information on both the mother and the offspring.
transform many aspects of perinatal research, as it provides a These studies are then analysed using a standard linear re-
large increase in statistical power to disentangle maternal and gression model adjusting for both the maternal and the off-
fetal effects which have been difficult to resolve until now. spring genotype, which is often referred to as ‘conditional
Data from males are included in the SEM in two ways. analysis’. One of the benefits of the SEM we describe here is
First, genotyped males who report their own phenotype that the coefficients for the maternal and fetal effects are on
(birthweight), but not their offspring’s phenotype, are the same scale as the coefficients from a conditional analysis,
modelled in the top half of Figure 1. For example, these and therefore a meta-analysis could be conducted across
males contributed directly to estimation of the fetal effect multiple cohorts with different study designs. Alternatively,
of genotype on birthweight (see the coefficient labelled ‘f’ because the model can be fit with observed covariance matri-
that is on the path from their own SNP to their own BW as ces, if the phenotypes of the mother and offspring are both
illustrated in the top half of Figure 1) and also indirectly to standardized and the effect allele frequency is known, then
estimation of the maternal effect on birthweight, since their the summary statistics (allele frequency, beta coefficient from
observed genotype (SNP) is correlated with their mother’s the regression model and variance of the phenotype) from an
unmeasured latent genotype at the same locus (GG) unconditional analysis for either the fetal effect or the mater-
(see the coefficient ‘m’ that is on the path from SNP to GG nal effect can be incorporated into this SEM. This makes it a
to BW in the top half of Figure 1). Second, male data con- potentially very powerful approach, as cohorts with pheno-
tribute to estimation in this SEM when the offspring of a type data and genotypes from mother, child or both, can all
UK Biobank female with genotype data is male and she re- be incorporated. It also avoids the need to share raw data,
ports his phenotype (i.e. birthweight of offspring BWO). which can be problematic for some cohorts, but still allows
For example, these offspring males contribute directly to for all cohorts to be included in the analysis and therefore
the estimation of the maternal effect on birthweight the sample size maximized.
(see the coefficient ‘m’ that is on the path from SNP to One of the biggest advantages of this SEM is that it is
BWO in the lower half of Figure 1) and indirectly to the robust to missing data, either for the individual’s own
estimate of the fetal effect on birthweight, since the male’s phenotype or the phenotype of their offspring. This is an
latent genotype (GO) is correlated with his mother’s advantage over conditional analysis, which uses only those
observed genotype at the same locus (SNP) (see the coeffi- mother/offspring pairs that have genotype data from both
cient ‘f’ that is on the path from SNP to offspring genotype persons. It can even be used when no individuals have
GO to offspring birthweight BWO in the lower half of phenotypes measured on both themselves and their off-
Figure 1). It is important to note that the inclusion of males spring; however, the power to detect a maternal or fetal ef-
in this way is not equivalent to estimating a paternal effect. fect is reduced and a small bias is introduced to the
To estimate paternal effects, one would need information maternal effect estimate. There are unlikely to be many
on males’ own genotype, their own phenotype and their studies with this study design, as the majority would have
offspring’s phenotype. Our SEM purely involves resolving a combination of individuals with complete data with
maternal and fetal effects. individuals missing data for their own phenotype or the

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 12 International Journal of Epidemiology, 2018, Vol. 0, No. 0

phenotype of their offspring. Additionally, it is robust to ACTL9 and EBF1). This result is consistent with the condi-
measurement error involving either the individual’s own tional analysis of 12 909 mother-offspring pairs in Horikoshi
phenotype or the phenotype of their offspring. However, et al.6 for MTNR1B and EBF1. The initial finding of a fetal
increasing measurement error in both phenotypes will in- effect appears to be due to the bias in the linear model, and
crease the standard errors and therefore decrease statistical therefore the fetal effect size was approximately half of the
power, similar to the effect of measurement error on ordin- maternal effect size estimated using the SEM. We also identi-
ary least squares regression. fied six SNPs where the maternal and fetal effects were in op-
There are four potential limitations to this SEM. First, al- posite directions (variants in ADCY5, CDKAL1 and ABCC9
though we have shown that the model fits well under a range and near HHEX-IDE, ANK1-NKX6-3 and DTL).
of minor allele frequencies, in some situations it has difficulty Interestingly, four of these SNPs that exhibited maternal and
optimizing with low frequency variants (generally with minor fetal effects in opposing directions (three of which were con-
allele frequency < 5%). This can often be resolved using the firmed using the conditional analysis in Horikoshi et al.6) are
mxTryHard function in OpenMx, which makes several at- known type 2 diabetes loci (ADCY5, HHEX/IDE, CDKAL1
tempts to optimize the model and returns results from the and ANK1), consistent with what is known regarding the
most optimal fit. Second, the SEM assumes multivariate nor- underlying biology at these loci.19 Importantly, the existence
mality between the observed variables and linearity between of an opposing maternal effect at these loci would not have
the genotypes and phenotypes. We have simulated our pheno- been detected had only unconditional linear regressions of
types to be normally distributed and ensured that birthweight offspring birthweight on maternal genotype been performed,6
data in the UK Biobank were approximately normally distrib- further highlighting the importance of our method in disen-
uted. In the case of non-normality, an appropriate data trans- tangling maternal and fetal effects on perinatal phenotypes.
formation can help ensure that the assumption of In summary, we describe a new method for estimating
multivariate normality is satisfied. Third, we assumed addi- unbiased maternal and fetal effects using studies where
tive genetic effects for both the fetal and maternal contribu- genotype data are available for only the individual and not
tions. An additive model is used in the vast majority of their offspring. We have shown that the type 1 error rate of
genome-wide association studies in the literature, and theory the method is appropriate, there is substantial statistical
and data show that the overwhelming majority of genetic loci power to detect a genetic variant that has a moderate effect
act in an additive fashion.16 If these assumptions do not hold, on the phenotype and reasonable power to detect whether it
then we expect reduced power to detect a fetal or maternal ef- is a fetal and/or maternal effect. We have also illustrated
fect. Finally, we assume only main effects and no interaction that this method could be useful for accurate estimation of
between the maternal and fetal genotypes. fetal and maternal effects in large genetic studies, such as
The SEM using observed covariance matrices takes ap- genome-wide association studies, as the computational time
proximately three times longer to compute than an uncondi- is not substantially larger than the standard linear model.
tional linear model. Therefore, there is potential for this
method to be used in large genomic studies, such as
genome-wide studies. A new method for fitting SEMs in Supplementary Data
genome-wide association studies in a computationally Supplementary data are available at IJE online.
feasible fashion has recently been developed,17 which may
facilitate analyses involving more complicated models like
ours. We note that tests of genetic association have trad- Funding
itionally been performed in the fixed effects part of SEMs N.M.W. is supported by a National Health and Medical Research
(i.e. the ‘model for the means’). In contrast, we have mod- Council Early Career Fellowship (grant number APP1104818).
D.M.E. is funded by an Australian Research Council Future
elled SNP effects in the covariance part of the model, which
Fellowship (grant number FT130101709) and a Medical Research
has allowed us to model latent genotypes. We have shown Council programme grant (grant number MC_UU_12013/4). Access
that within the confines of our study, accuracy of estimates to the UKBB study data was funded by University of Queensland
of maternal and fetal effects appear to be robust to the in- Early Career Researcher Grant (2014002959).
herent non-normality of individual-level SNP data, which is
to be expected in the case of exogenous variables.18
A recent study by Horikoshi et al.6 found three SNPs that Acknowledgement
were significantly associated with birthweight using the indi- This research has been conducted using the UK Biobank Resource
vidual’s own genotype (i.e. have a ‘fetal effect’); our analyses (application 12703). We thank George Davey Smith for comments
on an earlier draft of this manuscript.
indicate that the effects at these loci are driven by a maternal
rather than a fetal effect (variants in MTNR1B, and near Conflict of interest: None declared.

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018
 International Journal of Epidemiology, 2018, Vol. 0, No. 0 13

References 10. Little RJA, Rubin DB. Statistical Analysis With Missing Data.
Hoboken, NJ: Wiley, 1987.
1. Barker DJ, Hales CN, Fall CH, Osmond C, Phipps K, Clark PM. 11. Ounsted M, Scott A, Ounsted C. Transmission through the fe-
Type 2 (non-insulin-dependent) diabetes mellitus, hypertension male line of a mechanism constraining human fetal growth. Ann
and hyperlipidaemia (syndrome X): relation to reduced fetal Hum Biol 1986;13:143–51.
growth. Diabetologia 1993;36:62–67. 12. van der Valk RJ, Kreiner-Moller E, Kooijman MN et al. A novel
2. Magnus P. Causes of variation in birth weight: a study of common variant in DCST2 is associated with length in early life
offspring of twins. Clin Genet 1984;25:15–24. and height in adulthood. Hum Mol Genet 2015;24:1155–68.
3. Magnus P. Further evidence for a significant effect of fetal genes 13. Taal HR, St Pourcain B, Thiering E et al. Common variants at
on variation in birth weight. Clin Genet 1984;26:289–96. 12q15 and 12q24 are associated with infant head circumference.
4. Lunde A, Melve KK, Gjessing HK, Skjaerven R, Irgens LM. Nat Genet 2012;44:532–38.
Genetic and environmental influences on birth weight, birth 14. Warrington N, Richmond R, Fenstra B et al. Maternal and fetal
length, head circumference, and gestational age by use of genetic contribution to gestational weight gain. Int J Obes
population-based parent-offspring data. Am J Epidemiol 2007; (Lond) 2017, Oct 9. doi: 10.1038/ijo.2017.248. [Epub ahead of
165:734–41. print.]
5. Eaves LJ, Pourcain BS, Davey Smith G, York TP, Evans DM. 15. St Pourcain B, Cents RA, Whitehouse AJ et al. Common vari-
Resolving the effects of maternal and offspring genotype ation near ROBO2 is associated with expressive vocabulary in
on dyadic outcomes in genome wide complex trait analysis infancy. Nat Commun 2014;5:4831.
(‘M-GCTA’). Behav Genet 2014;44:445–55. 16. Hill WG, Goddard ME, Visscher PM. Data and theory point to
6. Horikoshi M, Beaumont RN, Day FR et al. Genome-wide associ- mainly additive genetic variance for complex traits. PLoS Genet
ations for birth weight and correlations with adult disease. 2008;4:e1000008.
Nature 2016;538:248–52. 17. Verhulst B, Maes HH, Neale MC. GW-SEM: A statistical pack-
7. Allen NE, Sudlow C, Peakman T, Collins R. UK biobank data: age to conduct Genome-Wide Structural Equation Modeling.
come and get it. Sci Transl Med 2014;6:224ed4. Behav Genet 2017;47:345–59.
8. Hattersley AT, Beards F, Ballantyne E, Appleton M, Harvey R, 18. Bollen KA. Structural Equations With Latent Variables.
Ellard S. Mutations in the glucokinase gene of the fetus result in Hoboken, NJ: Wiley, 1989.
reduced birth weight. Nat Genet 1998;19:268–70. 19. Beaumont RN, Horikoshi M, McCarthy MI, Freathy RM.
9. Pierce BL, VanderWeele TJ. The effect of non-differential meas- How can genetic studies help us to understand links be-
urement error on bias, precision and power in Mendelian ran- tween birth weight and type 2 diabetes? Curr Diabetes Rep
domization studies. Int J Epidemiol 2012;41:1383–93. 2017;17:22.

Downloaded from https://academic.oup.com/ije/advance-article-abstract/doi/10.1093/ije/dyy015/4855879

by guest
on 19 February 2018