Genetic Instrumental Variable Studies of Effects of Prenatal Risk Factors

Abstract

Identifying the effects of maternal risk factors during pregnancy on infant and child health is an area of tremendous research interest. However, of interest to policy makers is unraveling the causal effects of prenatal risk factors, not their associations with child health, which may be confounded by several unobserved factors. In this paper, we evaluate the utility of genetic variants in three genes that have unequivocal evidence of being related to three major risk factors – CHRNA3 for smoking, ADH1B for alcohol use, and FTO for obesity – as instrumental variables for identifying the causal effects of such factors during pregnancy. Using two independent datasets, we find that these variants are overall predictive of the risk factors and are not systematically related to observed confounders, suggesting that they may be useful instruments. We also find some suggestive evidence that genetic effects are stronger during than before pregnancy. We provide an empirical example illustrating the use of these genetic variants as instruments to evaluate the effects of risk factors on birth weight. Finally, we offer suggestions for researchers contemplating the use of these variants as instruments.

1. Introduction

The influence of maternal risk factors during pregnancy on fetal and child health has been widely studied in several disciplines. A large body of literature documents negative effects of maternal smoking during pregnancy on birth weight (Rosenzweig and Schultz 1983; Wehby et al. 2011; Wüst 2010) and child neurological development (Wehby et al. 2011a). Others report a negative effect of excessive alcohol consumption on fetal development (Sullivan 2011; Lemoine et al. 1968; Jones and Smith 1973). In contrast, studies of the effects of maternal body weight on child health have produced mixed results (Bhattacharya et al. 2007; Khashan and Kenny 2009; Torloni et al. 2009).

A major challenge in studying the effects of maternal behaviors and risk factors on child health is the role of unobservable confounders. Maternal prenatal behaviors and risk factors are correlated with several typically unobserved or inadequately measured characteristics that are relevant for fetal/child health. If ignored, these may confound the effects of the behaviors and risk factors of interest. Theory suggests that these unobservable confounders include several personality and preference traits, as well as maternal awareness of any risks to fetal/child health.¹ Not all such factors are adequately observed. For example, maternal psychosocial (stress and depression/anxiety) and nutritional influences may be inadequately captured. Similarly, maternal or family history of health problems, which may reflect certain health risks for future children and modify maternal behaviors, may not be well measured. Therefore, the correlations between the many maternal behaviors and risk factors, several of which are unobservable, make it virtually impossible to separate the effect of one factor from the other in classical regression models that only adjust for observable confounders. Such models will likely result in biased estimation of the effects of maternal behaviors and risk factors of interest on child health outcomes since it is very unlikely – if not impossible – that one can directly adjust for all relevant confounders. This can be shown graphically in a Directed Acyclic Graph (DAG), such as Figure 1. We are interested in the relationship between the risk factor and the outcome, but there may be several unobserved characteristics that affect both, leading to biased estimates.

Figure 1.

Directed Acyclic Graph for IV Approach

An additional complication of these unobservable confounders is that it is impossible in most cases to a-priori determine the direction of the bias based on theory (Wehby et al. 2009). For example, women who care more about their child’s health than others are more likely to avoid risk factors and to engage in healthy behaviors (such as adequate and appropriate nutrition, exercise, stress control, prenatal care, etc.). Such self-selection would result in overestimating (in absolute value) the negative effects of the risk factors of interest on infant health. In contrast, women who perceive greater risks for fetal/child health are more likely to engage in healthy behaviors and to avoid risk factors during pregnancy. For example, women who expect fetal growth complications due to previous personal or family history of such conditions may be more likely to avoid smoking or alcohol use during pregnancy and more likely to adopt healthy behaviors such as prenatal care or appropriate diet. This type of self-selection may result in underestimating the positive effects of healthy behaviors (such as prenatal care) and understating the negative effects of risk factors (such as smoking). Therefore, the net bias is a function of these opposite biases. Given that it is generally impossible to know which bias dominates, it is challenging to a priori determine the direction of the net bias due to unobservable confounders. This emphasizes the need for empirical investigation to evaluate the sign and magnitude of the bias.

Instrumental variable (IV) methods may be employed to account for such bias resulting from unobservable confounders. The premise is to employ variables (called instruments) that are strong predictors of the risk factors of interest (such as maternal smoking during pregnancy), but otherwise are unrelated to the outcome of interest (such as child health) except through their effects on these risk factors. In other words, these instruments should not affect the outcome directly, nor should they relate to the outcome through unobservable confounders. This is presented graphically in Figure 1, where the instrumental variable is associated with the risk factor of interest, but there is no direct link between the instrument and the outcome of interest, nor with the unobserved confounders. Under these conditions, the only variation in the risk factors used to estimate their effects on the outcome is that predicted by the instrument(s). This variation is unrelated to unobservable confounders under the IV assumptions described above. Therefore, an IV model using instruments that satisfy these IV assumptions allows for estimating the causal effects of risk factors on the outcome.

Until now, studies of the causal effects of prenatal health behaviors and risk factors have generally relied on policy instruments, such as changes in cigarette taxes/prices or smoking policies. While useful, these instruments have several limitations, such as ignoring variation in the etiology of behaviors/risk factors for individuals living in the same geographic area at which the instrument is measured. Also, policies may be endogenous to population behavior since they may be motivated by concerns about population health (Wehby, et al. 2011b).

Genes are implicated in many aspects of human behavior. Several recent studies have used genetic variants such as SNPs as instruments for studying causal behavior effects, also known as Mendelian randomization (Wehby et al. 2008); see e.g. (Chen et al. 2008; Timpson et al. 2009; Kivimaki et al. 2008; von Hinke Kessler Scholder et al. 2010; Lawlor et al. 2008; Smith et al. 2007; Smith and Ebrahim 2004; Davey Smith and Ebrahim 2003b; Fletcher 2009; Fletcher and Lehrer 2011; Fletcher 2011; Wehby et al. 2011a; Wehby et al. 2011b; Wehby et al. 2012). The strength originates from the fact that genetic variants are randomly assigned within a family at conception and are not reversely modified by one’s behaviors or environment. However, there are several potential threats to the validity of genetic instruments, with the main ones being pleiotropy (multifunctionality of genes), potential correlations between different genetic variants relevant for behavior/risk factors and outcomes due to linkage disequilibrium (LD) or co-inheritance of variants, and population stratification, leading to biased genetic effects (Davey Smith and Ebrahim 2003a; Lawlor et al. 2008; Didelez, Meng. S., and Sheehan 2010; von Hinke Kessler Scholder et al. 2011; Sheehan et al. 2008; Colhoun, McKeigue, and Davey Smith 2003). Population stratification occurs when genetic variants differ in their allelic distribution (i.e. the frequency of the minor allele) between subpopulations. The most common example of population stratification is that by ancestry, due to its correlation with social constructs such as race and ethnicity. For example, if the allelic distribution and outcome of interest differ systematically between these subgroups, the estimate of interest will be confounded. As such, population stratification may be particularly important to consider when studying diverse ancestral populations like those in the United States. We discuss this issue in further detail below.

In this study, we evaluate the use of genetic variants in three genes, CHRNA3, ADH1B, and FTO as instrumental variables to study the causal effects of maternal smoking and alcohol consumption during pregnancy, and obesity/body weight prior to pregnancy, respectively, on infant and child health. We argue it is important to evaluate the fit of a genetic variant as an instrument for a particular research question of interest on a case-by-case basis, assessing the extent to which the selected variants are:

1. Predictors of these maternal risk factors;

2. Unrelated to other variables that may also affect child health outcomes.

The first evaluation examines the relationship between the genetic variants and the associated risk factor during pregnancy. Most of the evidence on the importance of these genes is based on samples that are very much underrepresented in women of childbearing age and pregnant women. Indeed, less than a handful of studies have specifically evaluated the associations of these genes with maternal prenatal risk factors (Zuccolo et al. 2009; Freathy et al. 2009; Lawlor et al. 2011). However, the effects of genes on behavioral and health phenotypes may differ across the lifecycle. Before and during pregnancy, women go through many biological, hormonal, or behavioral changes, which may modify genetic influences on behavior and risk factors. For instance, many women quit smoking and drinking and modify their diet before and/or during their pregnancy. Given that the effects of certain genetic variants may vary before and during pregnancy compared to a more ordinary time, it is important to assess the effects of these variants specifically during the pregnancy period in order to evaluate their validity as instruments for studying the effects of prenatal risk factors on child health.

The second evaluation aims at indirectly testing whether there are any patterns suggesting that the instruments affect child outcomes other than through affecting the maternal risk factors of interest. Such evidence would violate the IV condition requiring the instruments to only affect the outcome through these factors.

Several genes have been implicated in smoking, alcohol use, and obesity. However, replicability of the reported associations is crucial, as we now know that many initial genetic associations fail to replicate (Colhoun et al. 2003). Indeed, there are only a few genes that have unequivocal evidence of being related to these risk factors in the literature. These include CHRNA3 for smoking, ADH1B for alcohol use, and FTO for obesity and body weight.

CHRNA3 is thought to influence smoking through coding for a receptor of nicotinic acetylchoine, a metabolic form of nicotine (OMIM 2011). CHRNA3 variants have been found to be significantly associated with nicotine dependence measures, including cigarettes per day and smoking intensity (Saccone et al. 2007; Saccone et al. 2009; Berrettini et al. 2008; Thorgeirsson et al. 2008; Weiss et al. 2008; Stevens et al. 2008; Sherva et al. 2008; Baker et al. 2009; Keskitalo et al. 2009; Liu et al. 2010; The Tobacco and Genetics Consortium 2010).

ADH1B is involved in alcohol metabolism by coding for a subunit of alcohol dehydrogenase (ADH) which catalyzes the rate-limiting step for ethanol metabolism (OMIM 2011). The risk allele in the functional variant rs1229984 increases the production of acetaldehyde and consequently alcohol toxicity symptoms believed to lead to lower alcohol consumption. ADH1B has been linked with a variety of alcohol consumption and dependence measures in multiple populations (Reich et al. 1998; Whitfield et al. 1998; Saccone et al. 2000; Loew et al. 2003; Wall et al. 2005; Saccone et al. 2005; Duranceaux et al. 2006; Zintzaras et al. 2006; Luo et al. 2006; Zhang et al. 2007; Tolstrup et al. 2008; Zuccolo et al. 2009; Birley et al. 2009; Macgregor et al. 2009; Sherva et al. 2009; Liu et al. 2011; Ghosh et al. 2008).

Although its function is not fully understood, FTO has been associated with obesity related traits both in adults and children (Scuteri et al. 2007; Scherag et al. 2010; Frayling et al. 2007; Melen et al. 2010; Meyre et al. 2009; Thorleifsson et al. 2009; Cheung et al. 2011; Willer et al. 2009; Dina et al. 2007; Tung and Yeo 2011). The increase in obesity is thought to occur through an increase in food and energy intake (Cecil et al. 2008; Haupt et al. 2009; Jonsson and Franks 2009; Speakman, Rance, and Johnstone 2008; Timpson et al. 2008) and a decrease in satiety (Wardle et al. 2008; den Hoed et al. 2009; Rutters et al. 2010).²

Using two independent datasets, we find that the genetic variants in these three genes are likely to be appropriate and useful instruments for these prenatal risk factors as well as some suggestive evidence of intensified genetic effects during pregnancy. We also provide an empirical example of using these genetic variants to evaluate the effects of maternal smoking, alcohol consumption, and body weight on child birth weight.

2. Methods and Data

2.1 Evaluating if the instruments fit the two IV conditions

We assess the fit of the selected genetic variants as instruments by evaluating the extent to which they satisfy the two IV conditions:

1. They are a predictor of the maternal risk factor of interest;

2. They are unrelated to other variables that may also affect child health.

We investigate the first condition by evaluating the strength of the association between each genetic variant and its related risk factor. This is done by regressing the risk factor on the genetic instrument using an ordinary least squares (OLS) regression and calculating the F statistic for the instrument effect on the risk factor. A common rule of thumb for a non-weak instrument is an F-statistic of at least 10 (Staiger and Stock 1997).³ Lower F-statistics indicate weaker instruments, which are more likely to lead to bias in the IV estimates and may require additional inference approaches. We use a linear probability model for binary risk factors to follow a standard interpretation of the instrument effects across all models and to evaluate the instruments’ strength using a common metric based on the F-statistic threshold. We repeat the estimation for binary risk factors using logit and find similar results to OLS (results available from the authors upon request). Linear probability models provide a good approximation of the average effect in large samples.⁴

Each of the evaluated SNPs has two alleles. SNPs may be coded in different ways under different assumptions. Under a dominant model, an indicator for the presence of the risk allele in either homozygote or heterozygote forms would be used. Under a recessive model, an indicator for the homozygote genotype of the risk allele would be used. Under an additive model, a variable representing the number of risk allele copies (0, 1, 2) would be appropriate. We specify the FTO and CHRNA3 to have three values assuming an additive model, which is consistent with the literature and is supported in our data. Given that the majority of individuals of European ancestry are homozygous for the common allele of rs1229984 in ADH1B, we use a binary indicator for the risk allele assuming a dominant model, which is also the model supported in our data.

Evaluating the second IV assumption is less straightforward. In fact, it is impossible to fully test this assumption due to the role of unobservables. Therefore, the foremost validation should be based on the current knowledge about the functions of the gene (and genetic variant) and whether they may relate to child outcomes other than through the maternal risk factor of interest. We evaluate this by reviewing the literature on the functions and expression of these genes. In addition, an ad-hoc check of the second IV assumption is to evaluate the association between the genetic variants and observed characteristics that may confound the relationship between the risk factors of interest and child health, but which are not affected by the risk factors (i.e. not on their causal pathway to child health). The premise is that if a genetic variant randomizes unobserved confounders, it would also be expected to randomize observed ones. While this does not provide a complete validation, it would reveal if there are systematic associations between the genetic variants and other characteristics, which would reduce the evidence for the fit of these variants as instruments. We evaluate the association between the genetic variants and such observable characteristics including maternal demographic, socioeconomic, and health indicators, pregnancy wantedness, and others by regressing the instrument on a vector of these characteristics. The two datasets that we employ vary in the available background measures, but they both capture the main maternal demographic, socioeconomic and health characteristics.

2.2 Datasets

2.2a. Danish National Birth Cohort (DNBC) dataset

The first dataset we employ in this study is based on a subsample from the Danish National Birth Cohort (DNBC). The DNBC is a cohort study of more than 100,000 pregnant women and their babies in Denmark in 1996–2003 (Olsen et al. 2001). Pregnant women were identified and recruited during their first prenatal visit, which generally occurred in the first trimester. The women were interviewed for their health and socioeconomic backgrounds, and provided DNA samples. The women were interviewed again around the beginning of the third trimester and after delivery.

The specific dataset that we employ in our study comes from the Prematurity Genome Wide Association Study (GWAS), which included a subsample from the DNBC of 1,000 preterm birth infants and 1,000 randomly selected term infants and their mothers. The vast majority of the sample mothers are thought to be of European ancestry. The GWAS genotyped the maternal DNA samples for 550K SNPs using a high density panel (Illumina Human660W-Quadv1_A platform) at the Center for Inherited Disease Research (CIDR). The covered genes included CHRNA3, FTO, and ADH1B. The SNPs in these genes reported in the literature to have the largest effects on the risk factors of interest were not covered in the GWAS panel. However, there are other covered SNPs in strong or perfect LD with these SNPs in CHRNA3 and FTO. LD, the correlation between alleles of two genetic variants (SNPs), is evaluated by measures of association such as a Pearson correlation coefficient or its squared term, the r-squared (r²).⁵ We report the results for one SNP per gene that is in high or perfect LD with the SNP in the literature with the largest effect on the risk factor of interest and that has the most explanatory power for the risk factor in this sample.⁶ There are no genotyped SNPs in the DNBC GWAS panel in strong LD with rs1229984 in ADH1B. Therefore, we only evaluate this SNP in the second dataset we use as described below.

We evaluate rs12914385 in CHRNA3 for its validity as an instrument for smoking measures in the DNBC sample. This SNP is in strong LD (r² of 0.836) with rs1051730 which has been consistently found to have one of the strongest effects on smoking behavior. We also evaluate rs8050136 in FTO as an instrument for body mass index (BMI) and obesity. This SNP is in perfect LD (r² of 1.0) with rs9939609, which is the SNP with the most convergent evidence in the literature for having the strongest effects on BMI/obesity (Frayling et al. 2007; Cauchi et al. 2009; Cornes et al. 2009; Hubacek et al. 2008).

The measures on smoking and obesity are obtained from the interviews conducted with the mothers during pregnancy. Our analytical sample ranges from 1811 to 1898 observations depending on the number of women with complete data for each analysis that we employ. CHRNA3 has been mainly reported in previous studies to be related to the number of cigarettes. Therefore, we evaluate the effects of the CHRNA3 variant on the average number of cigarettes per day during pregnancy (including 0 for non-smokers) and on number of cigarettes among smokers. Also, we evaluate its effect on any smoking during pregnancy. When possible, evaluating the genetic effects separately for before and during pregnancy is important to identify if these are modified by pregnancy. However, smoking before pregnancy is not measured in the DNBC dataset but is measured in the second dataset described below.⁷ We evaluate the effects of the FTO variant on obesity (BMI>30) and body mass index (BMI) before pregnancy. Table 1 describes the measures of risk factors and genetic instruments used in the DNBC analysis.⁸

Table 1.

Descriptive Statistics for Risk Factors, Birth Weight, and Genetic Instruments in DNBC

Variable	Description	Mean	Std. Dev.	N
Smoking	0/1 indicator for any smoking during pregnancy	0.280	0.449	1900
Cigarettes	Average number of cigarettes smoked per day	1.541	3.757	1813
BMI	Maternal body mass index	23.527	4.308	1811
Obese	0/1 indicator for obese mothers (BMI >30)	0.087	0.282	1811
Birth weight	Child birth weight in grams	3095.077	871.277	1914
rs12914385	Number of T alleles of maternal rs12914385 in CHRNA3	0.749	0.685	1934
rs8050136	Number of A alleles of maternal rs8050136 in FTO	0.799	0.692	1937

2.2b. Avon Longitudinal Study of Parents and Children (ALSPAC)

The second dataset is a cohort of children born in one geographic area (Avon) in England. Women eligible for enrolment in the population-based ALSPAC study had an expected delivery date between 1 April 1991 and 31 December 1992. Approximately 85% of eligible mothers enrolled, leading to about 14,000 pregnancies. The Avon area has approximately 1 million inhabitants and is broadly representative of the UK as a whole, although slightly more affluent than the general population (Golding, Pembrey, and Jones 2001); see www.bristol.ac.uk/alspac for more a detailed description of the data).

Although the cohort focused on the children and their development, ALSPAC also includes regular data collection on the mothers. For all analyses, we exclude mothers of non-white ethnic origin to reduce the risk of population stratification. All genotyping was performed by KBioscience (http://www.kbioscience.co.uk). SNPs were genotyped using the KASPar chemistry, which is a competitive allele-specific PCR SNP genotyping system using FRET quencher cassette oligos (http://www.kbioscience.co.uk/genotyping/genotyping-chemistry.htm).

We use the SNPs in the (maternal) genes CHRNA3, ADH1B and FTO that have been reported in the literature to have the strongest effects on the risk factors of interest (rs1051730, rs1229984 and rs9939609, respectively).

We observe maternal smoking and alcohol consumption at three points in time during pregnancy: at 8 weeks, 18 weeks and 32 weeks gestation. We refer to these as the first, second, and third trimester respectively.⁹ We use dummy variables indicating any smoking and alcohol consumption at these three time points, as well as counts indicating the number of cigarettes and units of alcohol.¹⁰ We also observe whether the mother binged, measured only in the second trimester and defined as drinking the equivalent of two pints of beer, four glasses of wine, or four pub measures of spirit. Similar to the DNBC analysis, we estimate the gene effects on maternal risk factors before pregnancy including for smoking (any smoking and number of cigarettes per day) and any alcohol consumption. Finally, we evaluate the FTO effects on BMI and obesity, which are obtained based on self-reported measures of pre-pregnancy height and weight. Table 2 describes the variables for the ALSPAC sample.¹¹

Table 2.

Descriptive Statistics for Risk Factors, Birth Weight, and Genetic Instruments in ALSPAC

Variable	Description	Mean	Std. Dev.	N
Mother smokes pre-pregnancy	0/1 indicator for any smoking during pregnancy	0.318	0.466	10521
Mother smokes in Trimester 1	0/1 indicator for any smoking in first trimester	0.194	0.396	9864
Mother smokes in Trimester 2	0/1 indicator for any smoking in second trimester	0.193	0.395	10521
Mother smokes in Trimester 3	0/1 indicator for any smoking in third trimester	0.199	0.399	9177
No. of cigarettes in Trimester 1	Number of cigarettes a day in first trimester	2.090	5.114	9864
No. of cigarettes in Trimester 3	Number of cigarettes a day in third trimester	2.148	5.249	9177
Mother drinks pre- pregnancy	0/1 indicator for any alcohol during pregnancy	0.928	0.258	10449
Mother drinks in Trimester 1	0/1 indicator for any alcohol in first trimester	0.310	0.463	9500
Mother drinks in Trimester 2	0/1 indicator for any alcohol in second trimester	0.291	0.454	10414
Mother drinks in Trimester 3	0/1 indicator for any alcohol in third trimester	0.331	0.471	5849
No. of alcoholic units pre-pregnancy	Categorical variable indicating the number of units of alcohol per week pre-pregnancy	1.594	0.785	10449
No. of alcoholic units in T1	No. of units of alcohol a week in first trimester	1.699	4.285	9500
No. of alcoholic units in T2	No. of units of alcohol a week in second trimester	1.521	3.854	10414
No. of alcoholic units in T3	No. of units of alcohol a week in third trimester	1.730	3.702	5849
Bingeing	Mother binged in second trimester	0.307	0.890	10400
BMI	Maternal Body Mass Index	22.95	3.822	9403
Obese	0/1 indicator for mother being obese (BMI>30)	0.055	0.229	9403
Birth weight	Child birth weight in grams	3407.10	559.437	10869
rs1051730	Number of T alleles of maternal rs1051730 (CHRNA3)	0.667	0.671	7830
rs1229984	0/1 indicator for one or two copies of the A allele of maternal rs1229984 (ADH1B)	0.050	0.222	7818
rs9939609	Number of T alleles of maternal rs9939609 (FTO)	0.799	0.690	7832

3. Results

We first show the results from evaluating the extent to which the instruments satisfy the first IV condition. We describe the effects of each instrument on the related risk factor in separate sections for each gene. Next, we show the results of the indirect tests of the second IV condition that the instruments are unrelated to confounders. In that section, we also describe the correlations between the risk factors and observable confounders to further make the case for the need of IV analysis. Finally, we provide an applied example for using these instruments to evaluate the effects of selective prenatal risk factors on birth weight.

3.1 Evaluating if instruments are significantly related to risk factors of interest

3.1a CHRNA3

Table 3 lists the coefficients of the regressions of smoking measures on CHRNA3 variants and the F statistics for variant effects (panel A for the Danish sample and panel B for the ALSPAC sample). In the DNBC sample, the CHRNA3 variant has significant effects on the total number of cigarettes smoked, and on the number of cigarettes among smokers, with F statistics of 9 and 10.6, respectively, suggesting overall a reasonably non-weak instrument. Each copy of the risk allele increases the total number of cigarettes by about 0.39 cigarettes per day and the number of cigarettes among smokers by about 1.2 cigarettes per day. In contrast, CHRNA3 has an insignificant effect on the probability of smoking during pregnancy.

Table 3.

Effects of Genetic Instruments in CHRNA3 on Several Maternal Smoking Measures

Panel A. DNBC
	Number of cigarettes per day during pregnancy	Number of cigarettes per day during pregnancy among smokers	Any smoking during pregnancy
rs12914385	0.389***	1.197***	0.012
	(0.142)	(0.366)	(0.015)
F statistic	9.05	10.662	0.621
R2	0.005	0.024	0.0003
N	1811	445	1898

Panel B. ALSPAC
	Any smoking prior to pregnancy	Any smoking, Trimester 1	Any smoking, Trimester 2	Any smoking, Trimester 3
rs1051730	0.005	0.021***	0.017**	0.015**
	(0.008)	(0.007)	(0.007)	(0.007)
F statistic	0.322	8.296	6.291	4.339
R2	0.000	0.001	0.001	0.001
N	7405	7023	7405	6485

	Number of cigarettes a day, trimester 1	Number of cigarettes a day among smokers, trimester 1	Number of cigarettes a day, trimester 3	Number of cigarettes a day among smokers, trimester 3
rs1051730	0.407***	0.923***	0.264***	0.476*
	(0.094)	(0.259)	(0.101)	(0.288)
F statistic	18.534	12.664	6.875	2.735
R2	0.003	0.010	0.001	0.002
N	7023	1356	6485	1288

Notes: The Table reports the instrument effects (OLS regression coefficients and their standard errors in parentheses) on several smoking measures separately for the DNBC and ASLPAC datasets.

^*p < 0.10,

^**p < 0.05,

^***p < 0.01.

In the ALSPAC data, the CHRNA3 variant has significant effects on the binary indicators of smoking during pregnancy with F-statistics between 4 and 8 but has an insignificant effect on smoking prior to pregnancy. The CHRNA3 variant has also significant effects on the number of cigarettes smoked per day during the first and third pregnancy trimesters, with an F statistic of 18 and 7, respectively, and on the number of cigarettes smoked among smokers, with an F statistic of 13 and 3 for the first and third trimesters, respectively. Recall that we do not observe the count of the number of cigarettes during the second trimester. The analyses show that each copy of the risk allele increases the total number of cigarettes per day by 0.26–0.41 cigarettes and the number of cigarettes among smokers by 0.5–1 cigarette per day.

3.1b ADH1B

Table 4 lists the effects of the variants in ADH1B on the various alcohol measures and the F statistics in the ALSPAC sample.¹² The ADH1B rare allele variant decreases the probability of consuming alcohol mainly during pregnancy. Women who carry at least one copy of the rare allele are between 7 and 10 percentage points less likely to drink during pregnancy. The F-statistics for the ADH1B effects during pregnancy are between 9 (for third trimester) and 20 (for first trimester), suggesting a relatively strong instrument. There is a smaller and marginally significant effect on alcohol consumption prior to pregnancy; women with the rare variant are about 2.9 percentage points less likely to drink (F-statistic of 3.4). When we examine the number of alcoholic units consumed per week during pregnancy, we also find strong genetic effects, with those carrying the rare allele of the ADH1B variant drinking between 0.6 and 0.9 units per week less than those homozygous for the common allele; F-statistics range from 9.2 to 55, with the largest effects on the first trimester. The ADH1B variant also shows significant effects on binge drinking during pregnancy, with those carrying at least one rare allele being on average 13 percentage points less likely to binge.

Table 4.

Effects of the Genetic Instrument in ADH1B on Several Maternal Alcohol Measures in ALSPAC

	Any alcohol Trimester 1	Any alcohol Trimester 2	Any alcohol Trimester 3
rs1229984	−.100***	−0.072***	−0.086***
	(0.022)	(0.022)	(0.029)
F statistic	20.529	10.429	9.011
R2	0.002	0.001	0.002
N	6754	7340	4553

	Number of alcohol units, Trimester 1	Number of alcohol units, Trimester 2	Number of alcohol units, Trimester 3	Binge drinking, Trimester 2
rs1229984	−0.888***	−0.610***	−0.580***	−0.129***
	(0.120)	(0.132)	(0.191)	(0.034)
F statistic	55.242	21.379	9.193	14.355
R2	0.002	0.001	0.001	0.001
N	6754	7340	4553	7316

Notes: The Table reports the instrument effects (OLS regression coefficients and their standard errors in parentheses) on several alcohol consumption measures in the ALSPAC dataset.

^*p < 0.10,

^***p < 0.01

3.1c FTO

Table 5 shows the FTO variant effects on obesity and BMI before pregnancy (panel A for DNBC; panel B for ALSPAC). In both datasets, the FTO variant has large and significant effects on obesity status and BMI, with F-statistics of 20.9 and 16, respectively, in the DNBC dataset, and 16.5 and 35.5, respectively, in ALSPAC. In the DNBC sample, each copy of the risk allele of rs8050136 increases the probability of obesity by 4 percentage points and increases mean BMI by 0.6 points, respectively. In ALSPAC, each copy of the risk allele of rs9939609 increases the probability of obesity by 2 percentage points and BMI by 0.4 points. .

Table 5.

Effects of Genetic Instruments in FTO on Maternal Obesity and BMI before Pregnancy

	Obesity	BMI
Panel A. DNBC
rs8050136	0.044***	0.587***
	(0.01)	(0.146)
F statistic	20.87	16.119
R2	0.011	0.009
N	1811	1811

Panel B. ALSPAC
rs9939609	0.018***	0.427***
	(0.004)	(0.072)
F statistic	16.524	35.486
R2	0.003	0.006
N	6634	6634

Notes: The Table reports the instrument effects (OLS regression coefficients and their standard errors in parentheses) on obesity and BMI for the DNBC and ALSPAC datasets.

^***p < 0.01.

3.2 Evaluating if instruments are unrelated to confounders

We first check the associations between the study factors and several background demographics, socioeconomic status, and maternal health characteristics. The goal is to evaluate self-selection into each of these risk factors and the extent to which they may relate to child health outcomes through unobserved confounders. Consistent patterns of association between the risk factors and observable background characteristics suggest there may also be unobservable confounders. The associations are obtained from an OLS regression of the risk factor of interest on these background characteristics. After that, we show the associations between these same background characteristics and the SNPs to see if these characteristics are randomized by the genetic variants. Consistent patterns of correlations between the genetic variants and these background characteristics would suggest that the variants may be related to other unobservable confounders.

Here we summarize the main associations between the risk factors and the background characteristics (detailed results are in Tables A3 and A4 in the Appendix for the DNBC and ALSPAC datasets, respectively). In the DNBC sample, individuals who smoke more are less likely to be married, have a stable relationship, or plan their pregnancy. Furthermore, they have more previous live births but are younger at menarche (marginally significant). Also, individuals who smoke more are more likely to have mental health problems and to be employed. Furthermore, any smoking is associated with worse general health. Similar to the DNBC sample, smoking is generally characterized by an inverse social gradient in ALSPAC: smokers are younger, less educated, of lower social class, and more likely to be a lone parent than non-smokers. The findings are similar when examining the number of cigarettes.

Several significant associations are also found between alcohol consumption and background characteristics in the ALSPAC sample.¹³ Mothers who consume alcohol either before or during pregnancy are older, higher educated, and more likely to be employed. Similar results are obtained when examining the number of alcoholic units (available upon request).

Obesity and BMI are also associated with several background characteristics. In DNBC, obesity is associated with an increase in the number of previous miscarriages, a decrease in general health, younger age at menarche, and increase in employment. Similar associations are found for BMI (except for employment which is negative but insignificantly related). In ALSPAC, heavier mothers are less educated, and of lower social class. BMI is positively related to being employed.

These results show strong patterns of associations between maternal risk factors and a large set of background characteristics. As datasets do not provide measures of all relevant determinants of child health outcomes that are correlated with the maternal risk factors of interest, it is impossible to directly control for all such confounders.

Next, we summarize the associations between the SNPs and background characteristics obtained from OLS regressions of the SNPs on these characteristics (detailed results are in Tables A5 and A6 in the Appendix). In stark contrast to the study risk factors, we find no systematic differences in these characteristics by the variants of each SNP. In the DNBC sample, there are no significant associations between the SNPs and any of these characteristics, except for a marginally significant negative association between the CHRNA3 risk variant and being married or in a stable relationship and a marginally significant positive association between the FTO risk variant and previous number of miscarriages, though this effect may go via the effect of FTO on BMI, as higher BMI is associated with increased miscarriage risk (Metwally et al. 2008). Similar to the DNBC, the ALSPAC sample only shows few marginally significant associations. Carrying the risk allele of FTO is positively related to the probability that the mother is a lone parent. However, we do not find any systematic associations with socioeconomic position, social class, or other evaluated characteristics, as none of these other covariates are associated with the SNPs. Hence, these findings indicate that the various genotypes of these SNPs effectively randomize observable background characteristics that are relevant to the study risk factors, suggesting that they are expected to also randomize unobservable confounders.

We further evaluate the second IV assumption by reviewing the literature and current knowledge about gene expression and functions. CHRNA3 is expressed in the central and peripheral nervous systems where it is involved in detecting chemical signals through synapses as well as in other body parts including tongue, stomach, and others (OMIM, 2011). Other than nicotine dependence, CHRNA3 has been linked to lung cancer (Tournier and Birembaut 2011; Timofeeva et al. 2011) and bladder cancer (Kaur-Knudsen et al. 2011). Nicotinic acetylcholine receptors may also play a role in other health related issues such as chronic obstructive pulmonary disease (COPD) (Saccone et al. 2010; Nakamura 2011; Kaur-Knudsen et al. 2011; Kim et al. 2011). Moreover, a meta-analysis found that CHRNA3 may be related to peripheral arterial disease (PAD) (Zintzaras and Zdoukopoulos 2009). All of these conditions and diseases however, are highly correlated with smoking, where smoking is believed to be the causal factor. Hence, these associations are not likely to invalidate the use of the CHRNA3 variant as an instrument, but are simply another causal effect of smoking.

ADH1B is mainly expressed in the liver, as well as in the central and peripheral nervous system, the stomach, nose, renal system, and reproductive system. ADH1B has been associated with several cancers including upper aerodigestive tract (McKay et al. 2011), esophageal (Zhang et al. 2010a; Zhang et al. 2010b; Yang et al. 2010), oral and laryngeal cancers (Marichalar-Mendia et al. 2011), with potential variation between populations. As ADH1B is expected to affect these conditions through its effects on alcohol consumption, these effects do not compromise using the ADH1B variant as an instrument.

FTO is expressed in the central nervous system, placenta, hypothalamus, and pancreatic islets. While its function is not yet fully understood, some studies have found associations between FTO and diabetes independent of the effect on body weight (Hertel et al. 2011; Rees et al. 2011), but many other large GWA studies find that the association pertains to obesity and not diabetes (Frayling et al. 2007). Variants in FTO have also been related to age at menarche (Elks et al. 2010; Mumby et al. 2011), although there is no significant association between the evaluated FTO variant and age at menarche in our DNBC sample. However, age at menarche itself is known to be strongly associated with BMI (e.g. Wattigney et al. 1999). To the extent that FTO affects age at menarche and diabetes through its effect on maternal body weight, this does not violate the IV assumptions. If maternal FTO variants directly affect infant and child health however, this would invalidate their use instruments, potentially biasing the IV estimates. However, the overall lack of significant associations between the observed characteristics and FTO variants provides some support for using these variants as instruments. Nonetheless, as with all instruments, caution should be used when using genetic variants as instrumental variables for studying causal effects.

3.3 Empirical Example

As an empirical illustration of using the variants in these genes as instruments for studying the effects of prenatal risk factors on infant health, we apply these variants as instruments to study the effects of maternal smoking, alcohol use, and BMI on birth weight, which is measured in the DNBC and ALSPAC datasets (descriptive statistics for birth weight are in Tables 1 and 2).¹⁴ We evaluate the effect of each risk factor using 2SLS estimation for the IV model and OLS as a reference model.¹⁵

We first discuss the results from the DNBC sample shown in Table 6. Cigarettes are negatively and significantly associated with birth weight under OLS. Each cigarette is associated with about 30 gram decrease in birth weight on average under OLS. Under the IV model, cigarettes have larger (in absolute value) negative effects on birth weight by at least two times than OLS but the effects are overall statistically insignificant and not statistically different from OLS estimates based on a Hausman-type test (Hausman 1978). The estimated IV effect of about 55 gram decrease per cigarette in the total sample combining smokers and non-smokers is very close to a previous estimate using other genetic estimates in a Norwegian sample (Wehby et al, 2011a). In the IV estimation limited to smokers, birth weight decreases by about 90 grams with each cigarette (marginally significant); however, the difference between the OLS and 2SLS estimates is still not significant (p=0.16). These results suggest that the adverse effects of prenatal smoking on birth weight may be underestimated in a classical OLS model, consistent with several previous studies that have suggested that maternal smoking effects on birth weight and other child health outcomes may be underestimated when unobservable confounders are ignored (e.g. Permutt and Hebel 1989; Lien 2005; Wehby et al. 2011a; Wehby et al. 2011b; Wehby et al. 2011c).

Table 6.

Effects of Risk Factors on Birth Weight under OLS and IV (2SLS) Models with Genetic Instruments in DNBC

Prenatal risk factor	OLS	IV (2SLS)
	Birth Weight (grams)	Birth Weight (grams)
Cigarettes	−28.411***	−55.434
	(5.239) [1792]	(73.937) [1792] {0.712}
Cigarettes among smokers	−29.551***	−93.318**
	(6.963) [441]	(50.100) [441] {0.161}
BMI	9.657**	−111.637*
	(4.773) [1791]	(58.860) [1791] {0.016}

Note: The table reports the coefficients from the regressions of birth weight on prenatal risk factors; standard errors are in parentheses; sample sizes are in brackets; p values for significance of the difference between OLS and 2SLS estimates are in curly brackets. A separate regression is estimated for prenatal factor.

^*p < 0.1,

^**p < 0.05,

^***p < 0.01

The lack of significance of the smoking effects under the IV model is in part due to limited power given that only part of the variation in cigarettes that is predicted by the instrument is used in the estimation, which is expressed in very large standard errors. This highlights the particular need for large samples in IV estimations. Similarly, the lack of significance of differences between the OLS and 2SLS estimates is in part driven by this limited variation and large variances of IV estimates.

Unlike cigarettes, BMI is positively associated with birth weight under OLS. However, the BMI effect on birth weight switches to negative and is large and marginally significant under the IV model. Each unit increase in BMI leads to a 111 gram decrease in birth weight; this is significantly different from the OLS estimate (p=0.016).

Table 7 reports the results from a similar application using the ALSPAC dataset in addition to evaluating the effects of alcohol consumption on birth weight. The results for cigarette number and BMI are overall consistent with those using the DNBC data. OLS shows a negative association between maternal smoking during pregnancy and the child’s birth weight, with each cigarette on average being associated with a reduction in birth weight of 14 grams. Using the variant in CHRNA3 to instrument for the number of cigarettes doubles this estimate. We find similar negative results when conditioning the analysis on those mothers who smoke at least one cigarette, and the difference between the IV and OLS estimates is marginally significant. Although the actual point estimates differ between the DNBC and ALSPAC sample, both show that smoking reduces birth weight and that the IV estimates are larger than OLS. The large standard errors, however, preclude us overall from rejecting the OLS estimate.

Table 7.

Effects of Risk Factors on Birth Weight under OLS and IV (2SLS) Models with Genetic Instruments in ALSPAC

Prenatal risk factor	OLS	IV (2SLS)
Number of cigarettes (first trimester)	−14.035*** (1.280) [6947]	−28.270 (23.554) [6947] { 0.542}
Number of cigarettes among (first trimester) smokers	−3.422 (2.201) [1341]	−42.571 (25.323) [1341] {0.087}
Any alcohol use (firsttrimester)	32.877* (14.769) [6683]	355.315 (332.896) [6683] {0.327}
Number of units of alcohol (first trimester)	1.285 (1.572) [6697]	−187.651 (301.782) [6697] {0.272}
BMI	19.991*** (1.969) [6566]	−112.017 (196.182) [6566] {0.363}

Note: The table reports the coefficients from the regressions of birth weight on prenatal risk factors; standard errors are in parentheses; sample sizes are in brackets; p values for significance of the difference between OLS and 2SLS estimates are in curly brackets. A separate regression is estimated for prenatal factor.

^*p < 0.1,

^***p < 0.01

Similar to the DNBC results, we find that maternal BMI is positively related to child birth weight in the OLS analysis, which changes sign in the IV. In addition, the ALSPAC and DNBC analysis report very similar findings: a one unit increase in maternal BMI reduces the child’s birth weight by about 112 grams.

Finally, the association between alcohol consumption and birth weight in ALSPAC is positive in both the OLS and IV analysis, with the IV point estimate being considerably larger than that in the OLS, although the difference is not statistically significant. However, once we examine the number of drinks rather than the binary indicator, we find a clear negative effect of alcohol exposure on child birth weight.

Despite the IV estimates being relatively imprecise, with large standard errors, meaning we cannot always reject the OLS based on the IV estimates, the analyses show that a simple OLS regression may lead to different conclusions from the IV model. Assuming the IV conditions hold, these differences are likely to be due to unobservable factors that affect both the outcome and risk factors of interest. As we show in this illustrative example, not accounting for such unobservables may lead to bias in the estimates.

4. Discussion

This paper evaluates the utility of variants in three genes, CHRNA3, ADH1B, and FTO to be employed as instruments for studying the causal effects of maternal smoking and alcohol intake during pregnancy and obesity/BMI before pregnancy on child health outcomes. Using two independent datasets, we find these genetic variants to have overall relatively strong effects on these measures from a statistical perspective, with F-statistics generally close to or above 10. Furthermore, we find no patterns of significant correlations between these instruments and several maternal demographic, socioeconomic, and health characteristics that are correlated with the study risk factors, supporting the hypothesis that these genetic variants randomize confounders. These findings are consistent with the lack of evidence in the literature to date for strong effects of these three genes on other important confounders.

The study provides support for using these variants as instruments for maternal risk factors during pregnancy. However, as is the case for every (genetic as well as non-genetic) instrument, investigators should always validate the fit of their instruments in each dataset and for the specific outcomes they are studying. We note however, that the lack of evidence to date about other functions of the evaluated variants of interest that could confound their use as instruments does not mean that they do not have such functions. As shown above though, unlike the genetic instruments, the risk factors are highly correlated with several observable characteristics. This suggests that these factors are also likely to be correlated with unobservable confounders. Therefore, employing genetic instrumental variables to study the effects of these risk factors on child health could significantly reduce the bias in classical estimates of these effects.

The motivation for studying the fit of these variants specifically for risk factors during and around pregnancy is the premise that genetic effects may be modified by pregnancy. Effects from previous studies which focused on general and older samples may not be generalizable to risk factors during pregnancy. In this study, similar to Zuccolo et al. (2009) and Freathy et al. (2008) we find some suggestive evidence that genetic effects are intensified during compared to before pregnancy. The stronger effects during pregnancy could suggest an interaction between the variant and the environment; the latter being pregnancy. The alcohol-related symptoms among those with the ADH1B rare allele for example, may be perceived as more unfavorable during than before pregnancy, or these women may find it easier to quit as a result of these symptoms.

Although we aim to evaluate the fit of these variants as instruments for maternal prenatal risk factors, one may argue that the genes are more general, as their effects are not necessarily specific to this period. In other words, mothers who carry the rare allele variant of CHRNA3 and/or ADH1B are likely to smoke more and/or drink less anytime throughout life including after pregnancy, and those with the risk allele of FTO are likely to be slightly heavier throughout life. Hence, any estimated effects of these risk factors on child health outcomes may not necessarily be solely due to prenatal effects, but may partly reflect risk factors after pregnancy. The extent of this analytical challenge will depend on the period at which child outcomes are measured. While this is unlikely to be of relevance to neonatal outcomes, it may be an issue for longer-term child outcomes that may be affected by maternal risk factors after pregnancy. In these cases, one can instrument for both prenatal and postnatal behaviors risk factors if there are enough instruments. If not, one may compare the prenatal behaviors effects between models that alternatively control for or exclude postnatal risk factors (if these cannot be instrumented for due to the lack of instruments).

We were able to evaluate the genetic instruments in CHRNA3 and FTO in two independent datasets and find consistent results, which supports the validity of the findings. We could not evaluate the ADH1B instrument (rs1229984) in the DNBC sample since it was not genotyped and there were no other genotyped SNPs in strong LD with this variant. Therefore, it is important to evaluate this instrument in other datasets in the future.

One study caveat is that both of our samples are very homogenous on race/ethnicity. Although this reduces the bias in instrument effects due to population stratification, it may limit the generalizability of the study results to other populations. This highlights the need to study the utility of these genetic variants as instruments in other populations and samples. However, it is important to recognize the potential bias due to population stratification when studying diverse populations such the United States. If the allelic distributions of the genetic instruments vary by ancestry, the IV estimates may be biased. Furthermore, direct adjustment for self-reported measures of race and ethnicity may be inadequate to remove this bias, as these may not effectively account for differences in allelic frequency by ancestry. Although this is not likely to affect our estimation, we report in Table A7 in the Appendix the minor allele frequencies for the three main evaluated SNPs in CHRNA3, ADH1B, and FTO. There is a substantial variation in the allelic frequencies of these SNPs by ancestry. For example, the frequency of the minor allele for rs1051730 in CHRNA3 is about 39%, 10% and 3% in individuals of European, Sub-Saharan African, and Asian (Han Chinese) ancestry, respectively. Therefore, when using genetic instruments that vary in allelic distribution by ancestry and data from populations with diverse ancestries, it is important to appropriately account for population stratification using effective methods such as genome-wide based indicators of ancestry based on principal component analysis when possible (Paschou et al. 2007).

IV models provide an estimate of the local average treatment effect (also known as LATE), which is the effect that applies to individuals whose behavior/risk factor was affected by the instrument, and whose behavior/risk factor would have changed had they been in a different instrument group (Angrist et al. 1996). These individuals are commonly described as the group on the margin (or “marginal” group). However, it is impossible to identify those individuals since none is observed under different instrument groups. Therefore, an important question is the generalizability of the LATE estimate to the total population. Again, this question is not specific to genetic instruments per se, but to all instruments. Two complications for the generalizability of the LATE estimate might arise in the context of genetic instruments: 1- gene-environment interactions and 2- limited explanatory power of instruments. We discuss each of these below in detail.

The effects of genetic variants on behaviors and risk factors may be modified by environmental factors. Although such gene-environment interactions should not bias the IV estimate, ignoring them may mask important differences between subgroups defined by these environmental factors. As such, this may limit the generalizability of the results. While it is impossible to identify the “marginal” group to whom the LATE applies, some subgroups of certain environmental characteristics may include a larger representation of the “marginal” individuals than others. In general, gene-environment interactions are not expected to affect the internal validity of the IV estimate, but rather the external validity and interpretation of the estimates. Therefore, it is important to evaluate any suspected gene-environment interactions in the first stage. If significant interactions are found, the IV model (both first and second stages) should be stratified by the environmental variables with significant interactions, provided that the instrument has sufficiently strong effects on the risk factor in the stratified subgroups (Wehby et al. 2008).

It is important to evaluate the extent of variation in the risk factor that the instrument explains, beyond whether the instrument passes the minimum threshold to be considered non-weak from a statistical perspective (e.g. an F-statistic above 10).¹⁶ An instrument may be considered statistically strong, but it may still explain a small percentage of the variation in the risk factor. This does not bias the analysis, but it may lead to an imprecise estimate with potentially limited generalizability as it is based on only little variation in the risk factor. Again, the issue here is whether the LATE estimate that is applicable to the “marginal” group may generalize to the majority of the population for whom the risk factor was not explained by the instrument. This will depend on whether the effects of the risk factors being instrumented for on outcomes may vary between subgroups that have different underlying etiologies for these risk factors. Therefore, it is important that researchers consider the “clinical relevance” of the variation explained by the genetic instruments and implications for generalizability.

In our case, the instruments explain a small percent of the variation in the risk factors (see R² in Tables 3 through 5). The CHRNA3 instruments explain about 0.2–2.4% of the variation in number of cigarettes among smokers. Similarly, the ADH1B variants explain about 0.1–0.2% of the variation in alcohol consumption during pregnancy. The FTO variants explain about 0.3–1% of the variation in obesity and BMI. The small explanatory power is not surprising given the complex etiology of these risk factors. In fact, most models of socioeconomic and environmental characteristics explain less than 10% of the variation in such risk factors.¹⁷ However, the small variation explained by the genetic instruments raises a question about the applicability of the IV estimates to the larger population. In this particular case, there is no consistent knowledge that the effects of these risk factors on infant health outcomes vary significantly between maternal groups with different etiologies for these risk factors. For example, there is no expectation a priori that the effects of smoking on birth weight vary between those who smoke more cigarettes because of lower prices or because of having the risk CHRNA3 variant or both. As the genetic etiologies of these risk factors are further unraveled allowing for the use of additional instruments, it is possible to evaluate the IV estimates of the risk factor effects under different instruments. If the IV estimates are overall insensitive to different instrument (sets), then this would support the generalizability of the results to other groups, and the validity of the instruments and IV approach (Wehby et al. 2008; Davey Smith 2011; Palmer et al. 2011; von Hinke Kessler Scholder et al. 2010).

Appendix

Table A1.

Appendix Descriptive Statistics for Background Characteristics in DNBC

Variable	Description	Mean	Std. Dev.	N
Maternal age	Maternal age in years	29.565	4.244	1930
Child: boy	0/1 indicator for a male birth	0.525	0.500	1937
Married/stable relationshipa	0/1 indicator for mother being married or a stable relationship	0.989	0.106	1839
Previous live births	Number of previous live births at time of 1st DNBC interview	0.727	0.833	1841
Previous miscarriages	Number of previous miscarriages at time of 1st DNBC interview	0.258	0.608	1839
Unplanned pregnancy	0/1 indicator for not planning the pregnancy	0.099	0.298	1842
Average healthb	0/1 indicator for mother reporting her health as “average”	0.429	0.495	1840
Not so good healthb	0/1 indicator for mother reporting her health as “not so good”	0.047	0.211	1840
Mental problems	0/1 indicator for any previous or current mental health problems	0.126	0.332	1840
Menarche	Mother’s age at menarche in years	13.330	1.421	1678
Employed	0/1 indicator for being employed at the time of the 1st DNBC interview	0.213	0.409	1842

^aNotes: reference category is single mothers

^breference category is good health

Table A2.

Descriptive Statistics for Background Characteristics in ALSPAC

Variable	Description	Mean	Std. Dev.	N
Mother’s age	Maternal age in years	28.27	4.850	10999
Child: girl	0/1 indicator for female birth	0.485	0.500	11428
Lone Parent	0/1 indicator for mother being a lone parent	0.070	0.255	8576
# siblings	Number of siblings of cohort child	0.871	0.958	9009
Locus of control	Mother’s locus of control	−0.033	0.991	8605
CCEI Trimester 2	Crown-Crisp Experimental Index, Trimester 2	13.29	7.583	9279
EPDS Trimester 2	Edinburgh Postnatal Depression Score, Trimester 2	6.759	4.757	9717
Mother is happy	Variable ranging from 1–4 with 4 being most happy	3.260	0.805	6733
Mother is content	Variable ranging from 1–4 with 4 being most content	3.100	0.908	6692
Maternal education	Variable ranging from 1–4; 4 is highest education	2.297	0.926	10372
Social class	Variable ranging from 1–6; 6 is lowest social class	3.021	1.320	9008
Mother is employed	0/1 indicator for mother being employed	0.480	0.500	8531

Notes: Maternal educational includes four indicators: less than ordinary (O) level, O-level only, advanced (A) level that permits higher educational study, and having a university degree. We use the standard (reversed, so that higher values correspond to higher social classes) UK classification of social class based on occupation (professional (I), managerial/technical (II), non-manual skilled (IIInm), manual skilled (IIIm), semi-skilled (IV) and unskilled (V)). EPDS (Edinburgh Postnatal Depression Score) and CCEI (Crown-Crisp Experimental Index) aim to capture aspects of maternal mental health. EPDS indicates to what extent the mother is at risk of perinatal depression; CCEI captures a broader definition of mental health, measuring general anxiety, depression and somaticism. Higher scores mean the mother is more affected.

Table A3.

Coefficients from OLS Regressions of Study Risk Factors on Background Characteristics in DNBC

	Number of cigarettes	Any smoking during pregnancy	Obese	Pre- pregnancy BMI
Maternal age	−0.024	−0.011***	−0.000	0.029
	(0.026)	(0.003)	(0.002)	(0.029)
Child: boy	0.163	0.027	0.015	0.177
	(0.187)	(0.022)	(0.014)	(0.210)
Married/stable relationship	−4.748***	−0.292***	−0.008	0.619
	(0.858)	(0.099)	(0.064)	(0.980)
Previous live births	0.295**	0.022	−0.002	0.106
	(0.128)	(0.015)	(0.009)	(0.143)
Previous miscarriages	0.049	0.006	0.035***	0.351**
	(0.160)	(0.018)	(0.011)	(0.175)
Unplanned pregnancy	0.986***	0.107***	−0.019	−0.476
	(0.323)	(0.037)	(0.024)	(0.364)
Average health	0.199	0.036	0.072***	1.050***
	(0.193)	(0.022)	(0.014)	(0.218)
Not so good health	0.593	0.134**	0.069**	1.487***
	(0.475)	(0.053)	(0.034)	(0.515)
Mental problems	1.377***	0.120***	−0.038	−0.560
	(0.359)	(0.040)	(0.026)	(0.392)
Menarche	−0.117*	−0.013*	−0.016***	−0.457***
	(0.067)	(0.008)	(0.005)	(0.075)
Employed	0.497**	0.054**	0.040**	−0.135
	(0.237)	(0.027)	(0.017)	(0.266)
Constant	7.834***	0.977***	0.260*	27.508***
	(1.399)	(0.161)	(0.103)	(1.580)
N	1607	1668	1643	1643
R2	0.057	0.048	0.035	0.047

Notes: The Table reports the coefficients (with standard errors in parentheses) from OLS regressions of the study risk factors on several background characteristics in the DNBC dataset. All characteristics are included simultaneously in the regression.

^*p < 0.10,

^**p < 0.05,

^***p < 0.01.

Table A4.

Coefficients from OLS Regressions of Study Risk Factors on Background Characteristics in ALSPAC

	Any smoking before pregnancy	Any smoking during pregnancy	Number of cigarettes (first trimester)	Any alcohol before pregnancy	Any Alcohol during 1st pregnancy trimester	Obese	Pre- pregnancy BMI
Mother’s age	−0.008***	−0.003*	−0.024	0.004***	0.011***	−0.000	0.028
	(0.002)	(0.001)	(0.018)	(0.001)	(0.002)	(0.001)	(0.016)
Child: girl	−0.011	−0.013	−0.060	0.005	−0.012	−0.009	−0.184
	(0.013)	(0.010)	(0.123)	(0.007)	(0.014)	(0.007)	(0.113)
Lone parent	0.128**	0.102**	1.202*	−0.025	0.032	−0.032	−0.832**
	(0.042)	(0.038)	(0.544)	(0.024)	(0.041)	(0.017)	(0.313)
# siblings	−0.010	0.008	0.262**	−0.024***	0.005	0.009*	0.144
	(0.007)	(0.006)	(0.084)	(0.005)	(0.009)	(0.004)	(0.074)
Locus of control	0.062***	0.046***	0.451***	−0.005	0.026**	0.009*	0.202**
	(0.008)	(0.006)	(0.085)	(0.004)	(0.009)	(0.004)	(0.073)
CCEI Trimester 2	0.003	0.002	0.031*	0.000	0.001	0.001	0.050***
	(0.001)	(0.001)	(0.013)	(0.001)	(0.002)	(0.001)	(0.012)
EPDS Trimester 2	0.004	0.002	0.019	0.001	0.001	−0.002	−0.068***
	(0.002)	(0.002)	(0.022)	(0.001)	(0.002)	(0.001)	(0.020)
Mother is happy	−0.023	−0.011	−0.129	−0.009	−0.029	−0.005	−0.008
	(0.013)	(0.011)	(0.127)	(0.007)	(0.015)	(0.007)	(0.128)
Mother is content	0.004	−0.003	−0.062	0.007	0.028*	0.007	0.050
	(0.012)	(0.010)	(0.108)	(0.006)	(0.013)	(0.007)	(0.112)
Mother’s education	−0.035***	−0.031***	−0.433***	0.012**	0.035***	−0.013**	−0.368***
	(0.008)	(0.007)	(0.081)	(0.005)	(0.010)	(0.004)	(0.072)
Father’s social class	−0.025***	−0.019***	−0.212***	−0.000	0.015*	−0.014***	−0.293***
	(0.006)	(0.004)	(0.056)	(0.003)	(0.006)	(0.003)	(0.053)
Mother is employed	−0.000	−0.007	−0.141	0.014*	0.036*	0.007	0.393***
	(0.013)	(0.010)	(0.123)	(0.007)	(0.015)	(0.007)	(0.115)
Constant	0.686***	0.402***	4.021***	0.812***	−0.185**	0.134***	23.740***
	(0.064)	(0.053)	(0.659)	(0.036)	(0.069)	(0.038)	(0.633)
N	4234	4220	4220	4223	4065	4015	4015
R2	0.087	0.076	0.071	0.023	0.026	0.022	0.044

Notes: The Table reports the coefficients (with standard errors in parentheses) from OLS regressions of the study risk factors on several background characteristics in the ALSPAC dataset. All characteristics are included simultaneously in the regression.

^*p < 0.10,

^**p < 0.05,

^***p < 0.01.

Table A5.

Coefficients from OLS Regressions of Genetic Instruments on Background Characteristics in DNBC

	rs12914385 (CHRNA3)	rs8050136 (FTO)
Maternal age	−0.003	−0.005
	(0.005)	(0.005)
Child: boy	−0.021	−0.019
	(0.034)	(0.034)
Married/stable relationship	−0.282*	−0.065
	(0.154)	(0.154)
Previous live births	0.002	0.009
	(0.023)	(0.023)
Previous miscarriages	0.014	0.047*
	(0.028)	(0.028)
Unplanned pregnancy	0.058	−0.009
	(0.058)	(0.058)
Average health	−0.020	0.018
	(0.035)	(0.035)
Not so good health	−0.025	0.135
	(0.082)	(0.082)
Mental problems	0.010	0.056
	(0.063)	(0.063)
Menarche	0.003	0.007
	(0.012)	(0.012)
Employed	−0.011	0.010
	(0.043)	(0.043)
Constant	1.077***	0.872***
	(0.253)	(0.253)
N	1667	1668
R2	0.004	0.005

Notes: The Table reports the coefficients (with standard errors in parentheses) from OLS regressions of the study SNPs on several background characteristics in the DNBC dataset. All characteristics are included simultaneously in the regression.

^*p < 0.10,

^***p < 0.01.

Table A6.

Coefficients from OLS Regressions of Genetic Instruments on Background Characteristics in ALSPAC

	rs1051730 (CHRNA3)	rs1229984 (ADH1B)	rs9939609 (FTO)
Mother’s age	0.000	−0.001	0.002
	(0.003)	(0.001)	(0.003)
Child: girl	−0.007	−0.001	−0.044
	(0.025)	(0.008)	(0.025)
Lone parent	0.063	0.047	0.183*
	(0.071)	(0.031)	(0.072)
# siblings	0.012	0.005	−0.015
	(0.015)	(0.005)	(0.016)
Locus of control	0.006	−0.008	0.004
	(0.015)	(0.005)	(0.015)
CCEI Trimester 2	0.000	0.000	0.001
	(0.003)	(0.001)	(0.003)
EPDS Trimester 2	0.001	0.000	−0.001
	(0.004)	(0.002)	(0.004)
Mother is happy	−0.002	0.010	−0.000
	(0.026)	(0.010)	(0.026)
Mother is content	0.009	−0.009	0.004
	(0.023)	(0.009)	(0.023)
Mother’s education	0.008	0.001	0.006
	(0.017)	(0.006)	(0.016)
Father’s social class	−0.010	−0.002	−0.006
	(0.011)	(0.004)	(0.011)
Mother is employed	−0.027	−0.013	0.044
	(0.025)	(0.009)	(0.026)
Constant	0.662***	0.075*	0.747***
	(0.117)	(0.036)	(0.122)
N	2943	2932	2958
R2	0.002	0.004	0.005

Notes: The Table reports the coefficients (with standard errors in parentheses) from OLS regressions of the study SNPs on several background characteristics in the ALSPAC dataset. All characteristics are included simultaneously in the regression.

^*p < 0.10,

^***p < 0.01.

Table A7.

Minor Allele Frequencies for the Genetic Instruments by Ancestry

Gene/SNP (minor allele)	Minor Allele Frequency (%)
	European	Sub-Saharan African	Asian (Han Chinese)
CHRNA3/rs1051730 (C)	38.5	9.7	3.5
ADH1B/rs1229984 (A)	<1%	∼0%	75.6
FTO/rs9939609 (A)	46.0	11.6	50.9

Note: The table reports the minor allele frequencies for three SNP instruments in three ancestral populations based on HapMap allele frequencies (obtained from NCBI dbSNP).