Separating family-level and direct exposure effects of smoking during pregnancy on offspring externalizing symptoms: Bridging the behavior genetic and behavior teratologic divide

Abstract

Maternal smoking during pregnancy (MSDP) has been robustly associated with externalizing problems and their developmental precursors in offspring in studies using behavioral teratologic designs (Wakschlag, et al, 2002; Espy, et al, 2011). In contrast, the use of behavior genetic approaches has shown that the effects commonly attributed to MSDP can be explained by family-level variables (D'Onofrio, et al, 2008). Reconciling these conflicting findings requires integration of these study designs. We utilize longitudinal data on a preschool proband and his/her sibling from the Midwest Infant Development Study-Preschool (MIDS-P) to test for teratologic and family level effects of MSDP. We find considerable variation in prenatal smoking patterns both within and across pregnancies within families, indicating that binary smoking measures are not sufficiently capturing exposure. Structural equation models indicate that both conduct disorder and oppositional defiant disorder symptoms showed unique effects of MSDP over and above family level effects. Blending high quality exposure measurement with a within-family design suggests that it is premature to foreclose the possibility of a teratologic effect of MSDP on externalizing problems. Implications and recommendations for future studies are discussed.

INTRODUCTION

Maternal smoking during pregnancy (MSDP) is a modifiable risk factor for externalizing problems in offspring (D’Onofrio et al., 2008; Huizink and Mulder 2006; Wakschlag et al. 2002). Mechanistic studies with high quality exposure measurement have demonstrated links to disruptive behavior, deficits in executive function, and alterations in orbital frontal brain regions that subserve reward processing (Ernst, Moolchan, & Robinson, 2001; Lotfipour, et al, 2009; Wakschlag, et al, 2002). Gene×environment interactions have also been demonstrated, including the interaction of prenatal tobacco exposure with a functional polymorphism in the monoamine oxidase A gene (MAOA) that increases susceptibility to environmental adversity (Kim-Cohen et al. 2006; Wakschlag et al. 2009; McGrath et al. 2012). Adoption studies, which are designed to disentangle heritability from postnatal influences (Leve et al., 2013), have further demonstrated independent effects of MDSP on child externalizing problems in exposed children reared by genetically unrelated adoptive parents (Marceau et al. 2013). These findings are consistent with a wealth of research using animal models, which has consistently shown a teratological link between MSDP and negative outcomes in offspring (Harrod, Lacy, & Morgan, 2012; Lacy, Hord, & Morgan, 2012; LeSage, et al, 2006; Schnieder, et al, 2011; Shy, et al, 2012), further supporting the presence of direct teratologic effects of MSDP.

Whether or not MSDP is a causal risk factor for offspring externalizing problems, however, remains in question. This is due to the challenge of disentangling direct effects of smoke exposure from its correlates, particularly familial predisposition to engage in problem behaviors (Knopik 2009; U.S. DHHS 2014; D’Onofrio et al., 2013). As MSDP cannot be randomly assigned, the difficulties separating direct causal effects from family level variation are substantial. As a result, statistical control alone is inadequate to account for the myriad of potential confounds in the MSDP-child externalizing pathway.

To address this question, a number of recent studies have creatively utilized quasi-experimental genetically-sensitive designs as a means of teasing apart teratologic and familial mechanisms. Most influential have been a series of studies by D’Onofrio and colleagues using a discordant sibling design in several cross-national cohorts. (D’Onofrio et al. 2008; D’Onofrio et al. 2010; Kuja-Halkola et al. 2010; Ellingson et al. 2014; Skoglund et al. 2014). These studies have leveraged within-family differences in sibling exposure to MSDP to test whether individual-level exposure has a unique effect on offspring externalizing behavior when mean exposure level across siblings is accounted for. This design allows such studies to control for familial influences, including family environment and genetic relatedness, when estimating the effects of MSDP on individual outcomes (Dick, et al, 2000). Such studies have consistently failed to find direct effects of MSDP on offspring externalizing behavior, suggesting that elevated symptoms are attributable to the family environment or genetic factors. Taken together, the conflicting findings represented by studies using behavioral teratologic vs. behavior genetic designs have led the United States Surgeon General to conclude that evidence to date for MSDP playing a causal role in offspring externalizing risk is “suggestive but not conclusive” (U.S. DHHS 2014).

Mechanistic studies have typically been designed to test the effects of MSDP using high quality exposure measurement (e.g. interview-based, assessing a range of smoking parameters as they vary across pregnancy). In contrast, the majority of studies with genetically sensitive designs have employed existing and archival data that were not a priori designed to investigate MSDP. Due to the constraints of archival data, such studies have often had to rely on a single yes/no question about smoking (Kuja-Halkola et al. 2010) or a one-time assessment of average cigarettes per day across gestation (D’Onofrio et al. 2008; D’Onofrio et al. 2010; Ellingson et al. 2014; Skoglund et al. 2014). Studies aimed at examining patterns of prenatal smoking suggest that this type of very limited exposure measurement introduces significant bias in several ways.

First, prenatal smoking fluctuates substantially. Whereas non-pregnant adults tend to maintain fairly stable levels of smoking patterns over time (Chassin et al., 2000), pregnant women exhibit complex patterns, with repeated attempts to quit, and fluctuations in number of cigarettes smoked per day (Pickett et al. 2003, Pickett et al. 2005) due to unique motivations and exceptionally high social pressures to quit during pregnancy (Pickett, et al, 2003; Eiden, et al, 2013). A second issue concerns recall bias and non-disclosure that contributes to misclassification of smokers versus non-smokers or exposed versus non-exposed. This threatens the detection of subtle or long-term effects of MSDP (England et al. 2007, Pickett et al. 2009). For example, a woman who stopped smoking as soon as she learned she was pregnant may retrospectively report being a nonsmoker during pregnancy, truthfully, from her own long-term perspective. She would be classified in most studies as a nonsmoker or quitter, yet her fetus would be at risk for any adverse outcomes associated with first-trimester exposure. In this way, misclassification can lead to an underestimate of the true risk of prenatal exposure to smoking. Previously reported simulations of the effects of misclassifying smoking during pregnancy in epidemiological studies have found that the underestimation of the relative risk for smoking on hypothesized adverse outcomes ranged from < 10% to 55% under varying misclassification scenarios (Pickett et al., 2003).

Second, the majority of studies on MSDP to date have compared women who smoke during pregnancy with women who do not, thereby lumping never smokers with smokers who quit early in pregnancy (Massey & Compton, 2013). The few existing studies that have specifically compared smokers who quit early in pregnancy (pregnancy quitters) with smokers who continued smoking during pregnancy (persistent pregnancy smokers) have shown significant differences in history of antisocial behavior between these groups (Wakschlag, Pickett, Middlekamp et al., 2003; Kodl & Wakschlag, 2004; Pickett, Wilkinson, & Wakschlag, 2009; McGrath et al., 2012). Importantly, while never smokers may differ from ever smokers, pregnancy quitters are also significantly different from persistent pregnancy smokers (Massey, Bublitz, Magee, et al., 2015, Massey, Estabrook, O’Brien et al., 2015). Broadly speaking, the continuum of never smokers, pregnancy quitters, and persistent pregnancy smokers may represent a continuum of risk.

To address these substantial potential errors, interview-based methods that anchor retrospective recall in the context of calendric events salient to each participant significantly enhance the accuracy and reliability of report, as this method mimics the manner in which individuals naturally structure memories (Brigham et al. 2008; Lewis-Esquerre et al. 2005). While repeated prospective, multi-level MSDP assessment is optimal, interview-based methods that assess patterns of MSDP across gestation have demonstrated good sensitivity and specificity (Pickett et al., 2009). Furthermore, some women who are not identified prospectively during pregnancy due to stigma associated with smoking disclosure are actually better identified retrospectively during the postpartum period (Pickett et al., 2009). What these findings highlight is the extent to which a single gross measure of “typical day smoking” is likely to misrepresent level of MSDP. This is a particularly salient source of error in regard to the question of interest, i.e. how does differential exposure impact risk of externalizing behavior. Binary measures of exposure (smoked or did not smoke during pregnancy) reduce the many ways women can differ and change into two groups: those whose differences cross the threshold from “smoker” to “non-smoker” by that particular study’s definition, and those whose differences do not cross that threshold. The existence of any further variation in magnitude and timing of exposure will be falsely constrained to those two categories, adding error and reducing statistical power to detect the effects of MSDP.

In summary, genetically sensitive designs have provided robust controls for familial and genetic confounds of MSDP, but have had significant limitations in exposure measurement. Behavior teratologic (mechanistic) studies have utilized more comprehensive measures of exposure, including repeated prospectively-assessed smoking patterns with biological data, but have not utilized within-family designs. The goal of the current study is to substantively advance the debate about whether or not MSDP has a teratologic effect on offspring externalizing behavior by blending paradigms from both of these approaches. We combine a within-family design (assessing differential exposure effects on siblings) with comprehensive interview-based measurement of smoking patterns across each pregnancy.

METHODS

Participants

Mothers were initially recruited during pregnancy for the Midwest Infant Development Study (MIDS), which followed women from the prenatal period until their infant was 4 weeks of age (N = 369) (for detailed description see, Espy et al. 2011; Wiebe et al., 2015). Women were recruited over a 4.5 year period using study fliers distributed to all obstetric clinics in a rural area in Southern Illinois and a mid-sized city in Nebraska. As MIDS was designed to compare tobacco-exposed and non-exposed infants on developmental outcomes while minimizing the influence of other prenatal exposures, smokers were oversampled while women who reported illegal drug use or consumed four or more alcoholic drinks in a single sitting were excluded. Forty-seven percent of women reported active smoking at enrollment. Non-smokers were broadly matched to smokers by demographic factors known to be associated with cigarette smoking (i.e., lower educational attainment (< 14 years), majority race/ethnicity (Caucasian), and Medicaid status). Sixty-five percent of women were enrolled prior to 16 weeks of gestation and all women were enrolled prior to 28 weeks. The present paper is derived from a preschool follow-up of the MIDS (MIDS-P) when the probands were approximately 5 years of age. The analytic sample for the current study consisted of MIDS families who participated in MIDS-P (N = 299), regardless of whether or not those families had participating siblings (n = 217) (see Figure 1) or not. All subsequent sample descriptions and percentages reflect a sample size of 299.

Figure 1. Flow chart depicting derivation of participants in analytic sample.

Flow chart of MIDS sample selection decisions. Full sample at proband intake (n=369) is reduced to analytic sample (n=299) through attrition and exclusion based on comorbid disorders.

Recruitment and Procedures

Families who participated in the MIDS were recruited for the MIDS-P as probands approached their fifth birthdays. After obtaining informed consent, a one hour home visit was conducted, followed by a 3 hour lab visit about one week later in which performance-based assessments were conducted. In addition to these intensive data, MIDS-P also included maternal reports of siblings’ prenatal exposure to cigarettes and their externalizing behavior. In order to obtain parallel measures on proband and sibling externalizing behavior, we included siblings who were within the age-band covered by the Stonybrook Symptom Checklists (3 – 18 years: Gadow and Sprafkin 1998, Gadow and Sprafkin 2000, Gadow and Sprafkin 2004). Mothers’ reports of sibling externalizing behavior were collected during the home visit, while questionnaires about probands were administered during the lab visit to minimize the possibility of bias introduced by “spillover” from one child to another. During the home visit, we employed a time-line follow-back developed for this study designed to ground maternal recall of smoking behavior during each child’s pregnancy as accurately as possible (Clark et al., 2009). Procedures in the MIDS and MIDS-P were approved by the Institutional Review Boards at the University of Nebraska at Lincoln, the University of Illinois at Carbondale (MIDS) and Northwestern University (MIDS-P), respectively. Mean ages were 5.5 years for probands (range = 3.5 – 7.2 years) and 9.09 for siblings (range = 3 – 18 years). Fifty-one percent (50.9 %) of probands and 54.4% of siblings were boys.

Measures

Maternal smoking during pregnancy (MSDP)

Mothers reported on their smoking patterns during a timeline follow-back interview developed for this study, administered at the 5-year home visit (Clark et al., 2009). In this “bounded recall” method (Eisenhower et al. 1991), participants are cued to mentally represent the time frame of interest by picturing landmark events at the time (Brigham 2008; Glasner & van der Vaart, 2009; Sobell and Sobell, 1996;). For each child by order of birth, mothers confirmed the written birth date of the child and the age at which they became pregnant. They were reminded of the approximate time of year that they would have become pregnant and were asked to build up a vivid mental representation of their living circumstances, important occasions or events, work or study environment, daily routines, and friend and family relationships at the time. Mothers were then cued to recall specific events for each trimester of their first pregnancy, e.g. ‘where were you when you found out you were pregnant? Did you have morning sickness during that trimester?’ They were encouraged to use these representations as a guide for thinking about their smoking behavior and to provide their estimate of the typical amount they smoked, the least amount they smoked and the most they smoked before and after pregnancy, and during each trimester, on an 8-point scale from < 1 cigarette per day to 2 or more packs per day. This process was repeated using different cues for each successive pregnancy. Timeline follow-back methods have shown good reliability and validity in previous studies of smoking (Brown, Burgess, Sales, Whiteley, Evans, & Miller, 1998; Lewis-Esquerrea et al., 2005).

Over three quarters of the mothers (78.0%) in the analytic sample reported any lifetime smoking (Table 1). Based on the timeline follow-back data, 36.9% of probands and 40.9% of siblings were prenatally exposed at one or more trimesters. Of those who smoked during pregnancy, mean typical use per trimester on the 8-point scale was 2.93, 2.56, and 2.53 (SD =2.06, 2.08, and 2.09, respectively), corresponding to a position between the categories: “1 – 3 cigarettes per day” (2nd) and the “4–7 cigarettes per day” (3rd). Twenty-seven probands (9%) and 35 siblings (12%) were exposed to heavy “typical” daily maternal cigarette use, defined as a half pack per day or more for one or more trimesters.

Table 1.

Detailed smoking histories of mothers

	n	%
Not smoking at baseline (early pregnancy)
Never smoked in lifetime	81	22
Smoker who quit > 1 year before last menstrual period	12	3
Smoker who quit < 1 year before last menstrual period	67	18
Smoking at baseline	209	57

Externalizing symptoms

The DSM-IV-based Stony Brook Symptom Checklists were used to quantify externalizing symptoms in probands and their siblings (Gadow and Sprafkin 1998, Gadow and Sprafkin 2000, Gadow and Sprafkin 2004). The Stony Brook includes the Early Childhood Symptom Inventory (ECI) for children ages 3 – 6 years; the Childhood Symptom Inventory (CSI) for children ages 6 – 12 years; and the Adolescent Symptom Inventory (ASI) for children aged 13 – 18 years. For each age-appropriate inventory, the 18-item Attention Deficit Hyperactivity Disorder (ADHD) scale (with two subscales, as described below), the 8-item Oppositional Defiant Disorder (ODD) scale, and the Conduct Disorder (CD) scale (10 items in the ECI, 15 items in the CSI and ASI) were utilized to capture externalizing symptoms across childhood from ages 3 to 18 years. In these scales, mothers were asked to rate how often their child engaged in specific behaviors, i.e. “runs about or climbs on things when asked not to do so” on a 5-point scale (0 = Never to 4 = Very often). ADHD was scored as two subscales: inattentive (ADHD-IN) and hyperactivity/impulsivity (ADHD-HY). The five CD items added to the ECI to form the CSI and ASI were omitted from this analysis to prevent the creation of differences between siblings due to the use of differing versions of the scale. These items were developmentally inappropriate for 3 to 6-year-old children (“has stolen while confronting a victim”, “broken into someone else's house, building, or car”, “often stays out at night despite parental prohibitions”, “has run away from home overnight”, “often truant from school”). Each of the four externalizing subscales showed strong evidence of unidimensionality by parallel analysis (Horn, 1965), with ratios of first to second eigenvalues between 3.66 (CD) and 6.38 (ADHD-IN).

Parental History of Antisocial Behavior

In addition to the within-family design, parental history of antisocial behavior was also controlled in all analyses using the Antisocial Behavior Questionnaire, which has been normed in an epidemiologic sample for men and women separately, and includes items about childhood and adult antisocial behaviors (Zoccolillo, 2000; Huijbregts, Seguin, Zoccolillo et al., 2007).

Statistical Approach

To facilitate a genetically sensitive design, a proband child and the sibling closest in age to the proband were included in analyses. A single sibling was included to most closely mirror existing genetically sensitive models and simplify analyses, while the selection of the closest in age minimizes differences in family environment that would attenuate or understate family level effects, impairing its use as a control on direct MSDP effects. Families were included in the sample regardless of smoking status to best represent the population, increase power to detect MSDP to externalizing effects by including the full range of MSDP, and avoid biasing correlations between MSDP measures within a family caused by excluding extreme MSDP scores. As noted previously, the majority (78%) of women in the sample had a smoking history.

Validation of MSDP measure

There are three primary parts to this analysis. First, the presence of adequate variability in smoking exposure between proband and sibling pregnancies was verified. Next, we addressed the relationship between prospective and retrospective recall. Since both retrospective and prospective exposure data were available for probands, we compared the probands’ retrospective exposure measures (from MIDS-P) to prospective measures of exposure collected at the 16-, 28- and 40-weeks from MIDS prenatal assessments as a validity check. These comparisons are affected by several methodological differences between the measures. Whereas the retrospective followback data were by trimester, the prospective reporting were month by month since the last prenatal visit. Similarly, retrospective reporting assessed “Most”, “Typical”, and “Least” cigarettes per day, while prospective assessments queried “average cigarettes” per day. As prenatal visits varied slightly in their timing and occurred exclusively in the second and third trimesters, we used canonical correlation analyses to compare prospective and retrospective measures as sets, and to supplement direct bivariate comparisons.

Observed Pearson and polyserial correlations between retrospective and prospective self-reported exposure are shown in Table 2. Polyserial correlations, which appropriately treat the retrospective smoking categories as ordered categories, show much stronger associations between prospective and retrospective exposure, with most correlations greater than 0.8 and many greater than 0.9. Pearson correlations, which assume the retrospective smoking categories are numeric, show correlations between 0.5 – 0.8, with both the 16- and 28-week prospective reports most closely tracking second trimester retrospective reporting. Regardless, both measures of agreement between prospectively- and retrospectively-reported smoking were comparable or higher to correlations previously measured in another large prenatal smoking cohort (Pickett et al., 2005, Pickett et al., 2009).

Table 2.

Observed Correlations between Prospective and Retrospective SDP

		Prospective (in weeks)
		Pearson			Polyserial
		16	28	40	16	28	40
Retrospective (by Trimester)	1st - Most	.595	.638	.483	.702	.772	.576
	1st - Typical	.640	.698	.533	.765	.847	.642
	1st - Least	.717	.709	.614	.877	.864	.752
	2nd - Most	.658	.747	.568	.801	.908	.696
	2nd - Typical	.678	.775	.603	.833	.947	.744
	2nd - Least	.755	.779	.671	.941	.954	.838
	3rd - Most	.665	.765	.590	.809	.924	.721
	3rd - Typical	.693	.804	.634	.854	.973	.783
	3rd - Least	.767	.795	.692	.946	.955	.857

The canonical correlation analysis showed significant multiple correlations between two pairs of canonical dimensions, with each pair of dimensions representing different linear combinations of prospective and retrospective variables. The first dimensions represent positive associations between all variables (R = .829, F (27, 299) = 7.258, p < .001) and account for a large amount of the shared variation between the two sets of variables (the first pair of dimensions account for between 58.5% and 96.3% of variance in the nine retrospective items, and between 70.3% and 96.9% of variance in the three prospective items). While the second pair of dimensions also showed a significant association (R = .479, F(16, 206) = 2.159, p = .007), the orthogonalization inherent to the canonical correlation procedure means that this set accounts for a very small amount of variance in the prospective and retrospective variables (0.06% to 5.5% and 2.2% to 13.8%, respectively) and must be interpreted conditional on the first set. Complete canonical correlation results are presented in Appendix 1. Put together, the bivariate and canonical correlations showed strong agreement between prospective and retrospective exposure reporting and support their use in subsequent analyses.

Testing for direct effects of MSDP independent of family level effects

Finally, the predictive relationship between prenatal tobacco exposure (shared and differential exposure across siblings) and proband and sibling externalizing symptoms were assessed using structural equation modeling, testing both the presence of any family-level MSDP effect and a direct effect of MSDP controlling for family-level MSDP, as shown in Figure 2. This model included latent exposure measures for each child predicting first, second, and third trimester exposure measurements. These latent exposure measures covaried within families and predicted child externalizing symptoms both directly and through a latent variable representing the family effect. Twelve versions of the model were run, each including one of the four externalizing subscales and either “Most”, “Typical”, or “Least” daily exposure. In addition to the path diagram shown in Figure 2, child gender, age, and parental antisocial measures were entered as covariates in all models. All modeling was completed using OpenMx 2.0 (Neale, et al, 2015), using full information maximum likelihood.

Figure 2. Path Diagram of MSDP to Externalizing Model.

Path diagram of fitted MSDP to externalizing effect. With the exception of the variances and covariances of the raw MSDP variables, all paths are labeled, such that any paths with the same label are constrained to equality. To fit the model, a set of MSDP variables ("Most", "Typical", or "Least" per day) as well as relevant covariates (age, sex, and anti-social behavior variables for both parents, not presented) are entered as predictors of an externalizing domain ("ADHD-IN", "ADHD-HY", "ODD", or "CD", labeled “Ext’ in the figure). While the total effect of any MSDP variable on externalizing consists of the sum of multiple paths, the difference between the implied covariance between the within-child and cross-child MSDP to externalizing effects is equal to (1 − sigma_SDP) * Beta_Individual. If there is a stronger effect of MSDP within-child than cross-child, Beta-Individual will be non-zero and positive.

RESULTS

Variability in MSDP

The first step in this analysis was to broadly support the supposition that smoking behavior varies within a pregnancy and shows variability across pregnancies for future analyses. Figure 3 shows the relationship between probands and siblings on “Typical” daily smoking by trimester. Each point on the plot represents reported “Typical” cigarettes per day for a particular trimester, corresponding to both the earlier (x-axis) and later (y-axis) pregnancies. Each woman is represented by a red dot (first trimester), a green dot (second trimester) and a blue dot (third trimester), connected by a dotted line. While approximately two-thirds of the data points correspond to no smoking during that trimester in either pregnancy, the remaining third of the sample showed considerable variation in their typical smoking behavior. Approximately 8% of the sample (one-quarter of smokers) were consistent in level smoked across the two presented pregnancies. Another 12% (one-third of smokers) smoked during the first pregnancy only, while just over 4% (one-eighth of smokers) restarted for the later pregnancy after not smoking during the earlier pregnancy. The remaining 7% of the sample smoked during both pregnancies, but at different levels. Beyond these broad categorizations, 18.5% of women fall into more than one of the groups just listed, indicating between-pregnancy variation in within-pregnancy changes in smoking behavior.

Figure 3. Comparing “Typical” daily use across pregnancies and trimesters.

Comparison of "Typical" Daily use across trimesters and pregnancies. Each mother is represented by three points, with one point for each trimester

While there was noticeable variation in number of cigarettes per day that should be modeled, smoking behavior was quite stable. The second and third trimesters showed particularly high correlations between “Typical” (r₂₃=.960, CI=[.950, .967]), “Most” (r₂₃=.953, CI=[.943, .961]), and “Least” cigarette use (r₂₃=.966, CI=[.958, .972]) within the proband pregnancy. First trimester use, while still highly correlated with the other trimesters, was notably weaker for “Typical” (r₁₂=.849, CI=[.819, .875]; r₁₃=.818, CI=[.783, .839]), “Most” (r₁₂=.831, CI=[.798, .860]; r₁₃=.796, CI=[.756, .829]), and “Least” use (r₁₂=.865, CI=[.838, .888]; r₁₃=.831, CI=[.758, .859]), perhaps because many women attempted to quit or reduce when they learned they were pregnant in the middle of the trimester.

The statements that smoking during pregnancy is both variable and stable are apparently conflicting, but they peacefully co-exist. A large portion of the consistency in smoking behavior is due to women who have never smoked, whose consistent scores of zero on all measures inflate correlations. The magnitude of the correlations across method and trimester point to a single exposure factor per pregnancy, but the variation shows that each additional trimester provides some additional information for informing subsequent analyses.

Relationship between MSDP and Externalizing Symptoms

The within-pregnancy variability found in the first part of the Results must be modeled to present a more parsimonious representation of exposure for the purposes of predicting externalizing symptoms. Nine measures of smoking were present for both probands and siblings: most, least, and typical cigarettes per day for the first, second, and third trimesters. As discussed in the Methods section, the three measures for each trimester have an inherent structure that complicates inclusion in a structural equation model; any measure of “typical” daily smoking must be no greater than “most” per day and no less than “least” per day. Models that included all nine variables simultaneously showed poor fit due to these relationships, so we modeled exposure separately for the “most”, “typical”, and “least” variables. Each model created a single latent exposure factor for each pregnancy, predicting three trimester-specific exposure variables, which will in turn predict a particular externalizing scale. With only three time points and strong covariances among the three measures, only a single factor per pregnancy could be supported. Extensions of this model into growth curve modeling proved unstable, and growth factors accounted for little variance in MSDP in these models. All models include the MSDP externalizing effects shown in Figure 2, and also include age, sex, and parent anti-social variables as predictors of externalizing.

The results for each of the twelve models (four externalizing outcomes by three versions of the exposure variable) are given in Table 3. Four of the twelve tests show total or overall effects of MSDP on externalizing symptoms, three of which show significant effects of a direct pathway over and above family level effects. The most consistent independent effect of MSDP was for CD symptoms. CD symptoms showed a relationship with MSDP for all three methods of assessing exposure and showed some effect of MSDP on CD symptoms (X²₂ = 11.639, 10.349, and 8.687; p = .003, .006, and .012 for Least, Typical, and Most, respectively). Once family level variability in MSDP was controlled for, both the “Typical” and “Least” methods of exposure measurement maintained a direct effect (X²₁ = 5.589 and 7.872; p = .018 and .005, respectively), while the “Most” measure did not (X²₁ = 3.738, p = .053). ODD symptoms showed effects of “Least” MSDP (X²₂ = 7.767, p = .021) and maintained a direct effect above and beyond family level effects (X²₁ = 4.817, p = .028). However, there was no evidence for any pathways to ODD symptoms from “Typical” or “Most” exposure (X²₂ = 4.175, 2.928, p =.124, .231). In contrast, ADHD-IN and ADHD-HY showed no effects of MSDP once family level effects were controlled for (p = .107 – 545).

Table 3.

Tests of Individual (Direct) Effects of MSDP on Externalizing Symptoms

					MSDP-Direct		MSDP-Total
Outcome	Cigs/Day	CFI	TLI	RMSEA	χ21	p	χ22	p
ADHD-IN	Most	0.937	0.933	0.059	0.424	0.515	2.301	0.316
	Typical	0.934	0.930	0.061	0.475	0.491	2.871	0.238
	Least	0.886	0.879	0.083	0.366	0.545	1.911	0.385
ADHD-HY	Most	0.942	0.939	0.057	1.022	0.312	3.589	0.166
	Typical	0.937	0.933	0.060	1.319	0.251	4.470	0.107
	Least	0.895	0.889	0.080	1.138	0.286	3.591	0.166
ODD	Most	0.951	0.948	0.052	0.522	0.470	2.928	0.231
	Typical	0.948	0.945	0.054	1.428	0.232	4.175	0.124
	Least	0.899	0.893	0.078	4.817	0.028	7.767	0.021
CD	Most	0.950	0.947	0.053	3.738	0.053	8.687	0.013
	Typical	0.946	0.943	0.055	5.589	0.018	10.349	0.006
	Least	0.900	0.894	0.078	7.872	0.005	11.639	0.003

Note. Model fit statistics for full Figure 2 model including both direct and family level MSDP effects. MSDP-Direct is a likelihood ratio test comparing model from Figure 2 with and without freely estimated direct path from MSDP for a given pregnancy to that child’s externalizing symptoms. MSDP-Total is a likelihood ratio test comparing model from Figure 2 with and without both direct and family-level effects of MSDP.

Effect sizes are reported as partial R² for the direct MSDP effect, family MSDP effect, and total MSDP effects together in Table 4. Methods such as Cohen’s d are not appropriate for continuous predictors, and the high correlation between individual and family smoking complicates interpretation of standardized regression parameters. Regardless of method, the effects of MSDP were relatively small. All significant effects showed partial R² between 1.1 and 1.5%, corresponding to partial correlations between MSDP and externalizing symptoms of .105 to .126 once covariates and family MSDP are controlled for. The total effects (direct and family) of MSDP are slightly larger (up to 1.8% for CD - “Least”), and the family level effects range from 0.0% to 1.1% in models where MSDP has any effect at all. Because MSDP for any given pregnancy and family level effects are correlated, part of the total MSDP effect cannot be attributed specifically to either a direct or a family-level effect, instead remaining related to the variance shared between them.

Table 4.

Effect Sizes Tests of Individual (Direct) Effects of MSDP on Externalizing Symptoms

Outcome	Cigs/Day	pR2MSDP	pR2Direct|Famil	pR2Family|Direct
ADHD-IN	Most	0.003	0.001	0.000
	Typical	0.004	0.000	0.000
	Least	0.001	0.000	0.000
ADHD-HY	Most	0.006	0.002	0.000
	Typical	0.007	0.002	0.000
	Least	0.004	0.001	0.000
ODD	Most	0.006	0.001	0.002
	Typical	0.006	0.004	0.002
	Least	0.014	0.012	0.002
CD	Most	0.013	0.008	0.002
	Typical	0.016	0.011	0.002
	Least	0.018	0.015	0.004

Note. Effect size, as measured by variance explained partialling out covariates, for total MSDP effect (Direct & Family together) and the direct and family effects, each partialling out the other. All models include child age and sex as covariates.

DISCUSSION

The use of varied scientific designs in an attempt to elucidate mechanisms that underlie the robust association between MSDP and offspring externalizing problems has led to controversy about the nature of these associations. Behavior genetic approaches have suggested that the effects commonly attributed to MSDP can be explained by family-level variables (D’Onofrio et al. 2008, D’Onofrio et al. 2010), while behavior teratologic approaches have supported a direct effect of MSDP. In this study, we capitalized on strengths of both approaches to address this important question. The primary finding of this study is that when behavioral teratologic and within family designs are blended, MSDP has direct effects on conduct and oppositional defiant syndromes, with family exposure accounted for and parental history of antisocial behavior controlled, but no direct effects on ADHD.

Consistent with a host of mechanistic studies, these patterns suggest links specific to disruptive behaviors rather than the broader externalizing spectrum. Direct effects were generally small (partial R² between 1.1% and 1.5 %), in part due to controlling for familial effects. Notably, however, the direct effects accounted for approximately half of the total MSDP variation, indicating that teratologic effects are of similar magnitude as effects of genetics and family environment in this sample.

The strongest effects of MSDP were for CD symptoms, which showed a total MSDP effect for all measurements of exposure and direct effects for “Least” and “Typical” daily use. The effect sizes for these models were largest of all models, with total MSDP accounting for approximately 2% of variance in conduct disorder symptoms, 1.1 to 1.6% attributable to the direct effect of exposure and the remainder to familial effects. The effects on ODD symptoms were also significant, showing total effects on “Least” and “Typical” daily exposure and a direct effect on “Least” exposure. This is also aligned with prior work, where direct relations of MSDP to CD have been most consistently demonstrated (Wakschlag & Hans, 2002). Given that associations were strongest for CD symptoms, the young age of the present sample may be one reason for the relatively small amount of variance explained, as DSM-IV CD symptoms are largely developmentally improbable or impossible in younger children (Wakschlag, Tolan, & Leventhal, 2010). We partially mitigated this only including CD symptoms present in the ECI version of the CD subscale. The use of developmentally-sensitive measures to capture latent antisocial patterns at younger ages will be important for shedding light on this question, particularly as capturing patterns at younger ages in close proximity to exposure is most feasible for explicating this question efficiently in future prospective behavioral teratologic designs (Wakschlag, et al, 2014). In addition, we used continuous symptom measures, which may also attenuate effect size relative to those derived from extreme group designs. In contrast to CD, MSDP was unrelated to ADHD after accounting for familial effects. This is consistent with the studies noted above that have directly tested unique associations to specific externalizing disorders (Wakschlag & Hans, 2002; Nigg & Breslau, 2007).

The variation in prenatal smoking merits further discussion. Despite the high correlations between trimester-to-trimester smoking behavior, there exists considerable variability as well. In addition to the between-pregnancy differences shown in Figure 3, 18.5% of the sample cannot be described by any one category in the Figure. There is further variability in the timing of within- and between-pregnancy changes. This variability is of high importance in considering the optimal study design suggesting that high quality exposure measurement is critical for pinpointing whether and how MSDP has teratologic effects on children’s behavior, and should be explored in future studies with a blended teratologic-behavior genetic design for parsing neurodevelopmental effects.

The methods utilized in this paper are not the only ways that this effect can be tested. Fundamentally, any test that attempts to disentangle individual- and family-level effects is a test of whether the relationship between independent and dependent variables is stronger within-child than across children within the same family. If the MSDP to externalizing effect were entirely teratologic, sibling exposure would have no effect on a given child’s outcomes once that child’s exposure was accounted for. If the effect were entirely due to family-level variables, then the effect of a given child’s exposure on his or her outcomes would be no stronger than the effect of his or her sibling’s exposure. Our findings suggest that both are at play. The effect of MSDP on externalizing symptoms and other outcomes can and should be studied using a variety of statistical methods, provided those methods fundamentally address the strength of within child MSDP-outcomes relationships relative to his or her siblings’ MSDP status.

Limitations and Directions for Future Research

The primary limitation in this analysis is the sample size. Many of the prior studies have utilized very large population based cohorts to enable them to find sufficient numbers of discordant siblings. The quality of exposure measurement and the oversampling for MSDP in the MIDS-P offset our relatively small sample to some extent. Still, a non-clinical sample of 299 families (217 of whom include siblings) provides limited power to find small effects or model the complex longitudinal associations among multivariate measures of smoking behavior. In this analysis, we lack power to model complex dependencies between “Least”, “Typical” and “Most” items in the same model, or determine whether the slight increase in effect sizes when “Least” items predict externalizing symptoms. The presence of repeated measurements and more fine grained assessments (i.e., approximate number of cigarettes per day rather than binary smoking variables) also helps maintain added power above comparable studies with fewer exposure variable and categorical measurement. However, larger samples are needed to fit more complex models and provide greater precision in parameter estimates.

A further issue in our analysis was our inability to include all nine measurements of exposure (Most, Typical, and Least exposure at each trimester) in one multivariate model. The three measurements at each trimester have non-linear dependencies that complicate both their own inter-correlations as well as their correlations with other exposure and outcome variables. Put differently, if “Most” use at a given trimester is zero, then “Typical” and “Least” use must also be zero, and thus those variables have zero variance and undefined covariance with all other variables in this instance. As “Typical” use must be no greater than “Most” use and no less than “Least” use, these create complex inequalities that influence all parts of the model. We are not aware of any methods specifically defined for this type of data. The closest analogue is ipsative data analysis (Baron, 1996), a set of methods for when a set of test items are ranked in order, such that one item has a score of 1, another item a score of 2, and so on. These data share the feature that the value of one variable has direct effects on the available values for all other variables. The most common recommendations for ipsative data are that “the assumptions of the factor model, as usually stated, are inappropriate for data with ipsative properties (Jackson, 1980; p. 219). Given our reliance on multivariate and latent measures of exposure, we avoided this problem by repeating the analyses on each variation of exposure measurement. This problem could be avoided with more proximal exposure measurement that does not aggregate the variability in daily exposure during the measurement process.

Relatedly, there is a general trend in our results that the strongest effects come from “Least” daily exposure, while the weakest are found for “Most” daily exposure. It is possible that “Least” daily use is a better indicator of total exposure, better reflects attempts to quit smoking than the other exposure measures, or is less susceptible to self-presentation bias than the other exposure measures. We do not have an explicit test of this trend, and are not making specific conclusions about this effect. However, this apparent difference in effects should be explored in future studies, and researchers interested in self-reported MSDP should note that “typical” or “average” use is neither the only way nor necessarily the most powerful way to measure effects. This recommendation is apparently contradictory to our discussion of the problems in analyzing all exposure data simultaneously. While the methodological problems caused by interdependent measures of exposure are real, problems associated with having too much data are generally preferable to those associated with having too little data.

This interdependence between MSDP variables leads to an apparent multiple testing issue. Our results include twelve very similar versions of the same analysis, differing only on which MSDP and externalizing variables are included. While we freely admit that we are running multiple tests of the same effect, the high correlations and interdependence between these sets of variables means that these analyses cannot be viewed as twelve independent tests. Absent the non-linear dependence between the “Most”, “Typical”, and “Least” variables, these models could have been collapsed into a more parsimonious set of models. Rather than overcorrect with a Bonferroni correction or one of the competing methods, we present raw p-values and effect sizes and attempt to assess the total effect across versions of the test.

The makeup of our sample merits further discussion. The majority of women (78%) in MIDS-P had a smoking history, while 22% consisted of women who never smoked. We chose to include these women for several reasons. First, we were interested in maintaining as much of the sample as possible to maximize statistical power, retain the representativeness of the original sample, and to make our analyses as generalizable as possible. As the smoking rate in the general public continues to decrease, analyses that exclude non-smokers pertain to an increasingly small and extreme subsample of the general population, and results from those studies do not generalize either to the general population or to smokers-only samples from different eras. Further, existing work shows that women who quit smoking during pregnancy differ from those who do not in important ways (Massey & Compton, 2013), in particular, with respect to problem behavior history (Wakschlag, Pickett & Middlekamp et al., 2003; Pickett, Wilkinson & Wakschlag, 2009; Kodl & Wakschlag, 2004). We feel that the difference between smokers who quit and those who do not are substantively and quantitatively comparable to the difference between ever smokers and never smokers, and look forward to continued debate on methods to sample this increasingly complex smoking risk construct.

Finally, our interview-based measures of exposure were detailed but retrospective. Though our results show substantial agreement between prospective and retrospective reporting, the optimal design would be prospectively characterized exposure that combines the assessment of patterns via self-report with bioassays (Dukic, et al., 2007) for multiple pregnancies, with follow-up on sibling outcomes. In addition to supporting a direct effect of exposure once family level variation is accounted for, our approach points to the importance of more nuanced exposure measurement. Discovering associations between exposure and child outcomes depend on reliable measurement for both variables. If our goal is to find that differences in one variable (MSDP) predict differences in another variable (child externalizing outcomes), our analyses will be inherently limited by whichever variable is less reliably measured. Whether our analyses are completed with regression, ANOVA, SEM, or a more complex method, the association between MSDP and outcomes will be attenuated by unreliability in either measure. Future directions in the study of this effect must include the most reliable and accurate measures of prenatal exposure possible to separate teratologic and family level effects.

Conclusions

It has recently been suggested that ‘familial factors [and not MSDP] are causing the intergenerational transmission of adverse outcomes in families in which offspring are exposed to MSDP’ (Ellingson et al., 2014). Our findings suggest that this conclusion is premature due to the important shortcomings of both behavior genetic and behavior teratologic approaches, when employed separately. The literature cited in the introduction of this paper showed a stark methodological divide between mechanistic studies with high quality exposure measurement without genetically sensitive designs, and genetically sensitive designs with lower quality exposure measurement. This is a false dichotomy, as answering this question clearly requires reliable and repeated measurement of exposure within the framework of a genetically sensitive design. It is likely that collecting this type of rich data would require adding a supplemental data collection to an ongoing panel study or other longitudinal design featuring women likely to become pregnant one or more times during the study interval with developmental follow-up across early childhood. Such a design would not only provide much more powerful tests of the effects of MSDP, but also allow for more nuanced tests of the timing of exposure and assessment of externalizing patterns within a developmental context.

Appendix 1

Correlations between Prospective and Retrospective measures

	Measure	Dimension 1	Dimension 2	Dimension 3
Retrospective	1st – Most	.765	.025	.415
	2nd – Most	.890	.182	.300
	3rd - Most	.991	.211	.211
	1st – Typical	.837	.067	.385
	2nd - Typical	.925	.195	.200
	3rd - Typical	.960	.234	.102
	1st - Least	.881	−.227	.178
	2nd - Least	.961	−.122	.079
	3rd - Least	.982	−.117	.014
Prospective	16 weeks	0.758	.176	.038
	28 weeks	0.816	.071	.020
	40 weeks	0.695	.095	−.112

Note. This table shows the correlations between the three canonical dimensions and the set of prospective and retrospective variables.

Appendix 2

Power Simulation

To test our assertion that our longitudinal assessment increases power over a single assessment of SDP, we created a small Monte Carlo simulation. Data were generated using either a single continuous SDP indicator or three indicators correlated 0.8. We varied the sample size (values of 250, 500, 750 and 1000) and the size of the individual level effect from our model in Figure 2 (B_direct of 0.1, 0.2, 0.3 and 0.4). The direct effect was held constant at 0.1, while the within-family correlations of both SDP and externalizing were held constant at 0.5. These values are reasonably close to those found in our empirical data, but with a much wider range of sample sizes to inform future studies. We created 250 datasets for each condition, generated an empirical power estimate as the proportion of studies within a cell that find the effect, and presented all results in Figure A2.

We found three occasion exposure models with power of .8 have their power reduced to .602 when reduced to a single measure. Three occasions with power of .9 have power of .735 with a single measure.

Figure A2: Power of one- and three-occasion designs.