Cohorts

We used data from four cohorts: the Avon Longitudinal Study of Parents and Children (ALSPAC; United Kingdom, pregnancies in 1991–1992) [15,16], Born in Bradford (BiB; United Kingdom, pregnancies in 2007–2011) [17,18], Millennium Cohort Study (MCS; United Kingdom, pregnancies in 2000–2001) [19], and the Mother, Father, and Child Cohort Study (MoBa; Norway, pregnancies in 1999–2008) [20]. These cohorts were chosen because they collected data on most of the exposures/outcomes of interest, and had multiple follow-ups during childhood. More information about each cohort is provided in S1 File.

Participants

Within each cohort, we excluded any children from multiple pregnancies (see Fig A in S1 File for sample sizes before and after exclusions). No exclusions were made based on any other variable, but all our analyses were complete case, meaning that parent-child pairs were excluded from analyses if they did not have complete data on the exposure, outcome, and covariates for any specific exposure-outcome analysis. Therefore, the sample size varies between different exposure-outcome associations. For all cohorts, partners were invited by the enrolled study mother. This person cannot always be assumed to be the biological father, and paternity information was not collected by every cohort. Therefore, we use the term ‘partner’ rather than ‘father’.

Parental exposures

To maximise the relevance to public health policy, we focussed on parental health behaviours that are potentially modifiable during pregnancy: tobacco smoking, alcohol consumption, and caffeine consumption. Where data were available (summarised in S2 File), we derived each exposure variable for both mothers and partners. For each health behaviour, we derived several variables based on the nature and timing of exposure, with definitions harmonised across the four cohorts. These were: active smoking, passive smoke exposure, alcohol consumption, episodes of binge drinking, caffeine consumed from tea, coffee and cola (both individually and combined). We considered up to eight timepoints: preconception (in the three to six months before pregnancy), separately in each trimester of pregnancy, in the two years postnatally, at any time during pregnancy, and for active smoking only, early onset (started smoking before age 11) and ever in life. Where possible, we derived ordinal variables with four levels of exposure: heavy, moderate, light and none (for definitions, see S1 File). All other variables were binary (any versus no exposure), or for caffeine consumption only, continuous (mg/day).

In addition to health behaviours, we also explored parental SEP as an exposure class, indicative of the wider social determinants of health. We defined separate binary measures of SEP based on the highest education level achieved and current occupational social class where available. These variables compared the group with the lowest SEP (cohort specific—see S1 File) to all other groups as the reference.

Genetic risk scores for parental exposures

We had access to genetic data for mothers in ALSPAC, BiB and MoBa, and partners (confirmed biological fathers in this instance) in MoBa. We calculated genetic risk scores (GRS) for parental smoking, alcohol, and coffee consumption to incorporate genetically informed analyses within a MR framework [21]. Specifically, GRS were used as proxies for these behaviours in reduced-form (genome-outcome) analyses. GRS were calculated using PLINK V1.9, using genetic variants (SNPs) identified in large scale genome-wide association studies (GWAS) of number of aggregated alcoholic drinks consumed per week [22], number of cups of coffee consumed per day [23], and four smoking-related traits: age of initiation or regular smoking, ever versus never been a regular smoker, cigarettes per day, smoking cessation (current versus former smoker) [22] (Tables A–F in S1 File). Scores were weighted by the effect estimates reported in the corresponding GWAS and converted to standard deviation units (z scores) to facilitate comparing effects across different traits. F-statistics for instrument strength—that is, the strength of the association between each GRS and the exposure it is intended to represent—are shown in Fig B in S1 File.

Child outcomes and age stages

Where data were available (S2 File), we explored a broad range of child outcomes over five health domains: body size and adiposity; psychosocial and cognitive; immunological; serum biomarkers; and blood pressure. Outcomes were measured at up to six different timepoints in childhood (up to age 11) depending on the outcome and data availability in each cohort: at delivery, first year, age 1–2 years, age 3–4 years, age 5–7 years, age 8–11 years. In addition, for some outcomes (e.g., asthma) it was appropriate to develop a variable to describe the outcome occurring or being diagnosed “any time in childhood” (up to age 11). All continuous outcomes were converted to standard deviation units (z scores) so that they could be meta-analysed across different cohorts and visualised on the same scale. To facilitate interpretation of estimates, we also generated binary variables for birthweight, BMI and some psychosocial and cognitive outcomes, based on commonly used clinical thresholds. Outcome variables are described in detail in S1 File.

Covariates

The definition and harmonisation method for all covariates is described in detail in S1 File.

Where data were available, we derived harmonised variables for potential confounders of the relationship between parental health behaviours and child health (based on the established definition that confounders must be known or plausible causes of both the exposure and outcome): parents’ ethnicity (Asian, Black, white, mixed or other); parents’ age at conception or delivery (years); parents’ highest education (ordered categories, as defined in each cohort); parents’ occupational social class (ordered categories, as defined in each cohort); mother’s parity (nulliparous versus parity ≥1).

We also generated harmonised variables for potential mediators (i.e., variables that potentially sit on the causal pathway between an exposure and an outcome) to include in a sensitivity analysis for some exposure-outcome combinations: gestational age at delivery (weeks); birthweight (kg); child’s caffeine consumption before age 2 (where available); child’s exposure to passive smoke before age 2 (where available).

Finally, all models were adjusted for sex (male, female) and age of the child at outcome measurement (years) to increase the precision of the estimate.

Statistical analysis in each cohort

We conducted linear (for continuous outcomes) and logistic (for binary outcomes) multivariable regression models for each possible exposure/outcome combination, running up to eight models per exposure/outcome combination. These models and their rationales are outlined in Table 1; further details and the directed acyclic graphs (DAGs) are provided in S3 File. Where a cohort did not collect data on a particular covariate, it was omitted from the adjustment set for that cohort.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Regression models that were run for each cohort.

https://doi.org/10.1371/journal.pmed.1005153.t001

The minimally-adjusted models (models 1a/1b) included parent-of-interest’s ethnicity because this was a particularly important source of variation in health behaviours, especially for Born in Bradford, which had high ethnic heterogeneity. We considered the standard set of confounder-adjusted models (models 2a/2b and 3a/3b) to be our main models.

The mediator-adjusted models 4a and 4b were run as a sensitivity analysis and results should be interpreted with caution because adjusting for mediators on the causal pathway can introduce collider bias [24]. Mediators were included only where the outcome was measured after the potential mediator (e.g., child’s exposure to passive smoke postnatally was not included in models where the outcome was birthweight).

To allow exploration of sex-specific effects (which are supported by some DOHaD studies [25]), all models were additionally run stratified by sex of the child. We did not compare results between the sexes and the results of these models are not discussed further in this paper, but full results are available for others to explore and download via EPoCH Explorer.

Processing of cohort results

To uphold the privacy of participants, we removed results from analyses where the total sample size was <20, or any binary exposure or outcome group had n < 5. We also removed results with an implausible effect estimate, which likely arose due to the model failing to converge (see S3 File).

Meta-analysis

Where more than one cohort contributed results to an exposure/outcome combination, we estimated the weighted average effect across cohorts by inverse-variance weighted meta-analysis using the metafor package in R. Fixed-effects meta-analysis was used, because this is an appropriate method when there are a small number of studies [26]. Heterogeneity was assessed using I² and the heterogeneity P-value, and visually via forest plots.

Processing of final results

We generated results from 594,326 analyses, which included a mix of regression results from single cohorts (where only one cohort had the required data for that analysis) and meta-analysis results (where two or more cohorts had the required data). P-values for each model were then corrected for multiple testing using a false discovery rate (FDR) of 5%. To aid visual comparisons between binary and continuous outcome results, we converted the log odds ratios for binary outcomes to Cohen’s D [27] (standardised mean difference) which is a unitless measure of effect (equation provided in S3 File). Effect estimates are presented in their raw format (i.e., not as Cohen’s D) in the full set of results. To compare models in this paper succinctly, we categorised results by applying commonly used thresholds: FDR-adjusted P < 0.05 to define “statistical significance”, and D > 0.2 to define anything other than a “small” effect.

Causal inference approaches and triangulation

Our analyses allowed us to apply multiple causal inference approaches to strengthen causal interpretation of associations. In the analyses presented in this paper, the causal hypothesis was that there is a direct effect of the parental exposure during pregnancy. The approaches we applied were: (1) multivariable regression adjusting for potential confounders; (2) applying the principles of MR by instrumenting parental health behaviours with GRS [11]; (3) employing co-parent’s exposure as a negative control to account for shared familial and environmental confounding [13]; (4) using postnatal exposure as a negative control for pregnancy exposure [13]; (5) exploring dose-response relationships where possible [12]. Detailed methodology, including the assumptions and source of bias for each of these approaches, is outlined in S3 File. Causal evidence was evaluated based on predefined criteria (Table 2). Triangulation [7] was conducted for each exposure-outcome combination by summarising the number of analyses that provided causal evidence. When the mother was the parent-of-interest, strongest causal evidence was defined as support from five agreeing lines of evidence. When the partner was the parent-of-interest, we excluded the parental comparison negative control from this count—since the underlying assumptions are less likely to hold in this context (see discussion in S3 File)—and therefore defined strongest causal evidence as support from four agreeing lines of evidence.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Causal inference approaches and pragmatic criteria for evidence suggestive of a causal effect during pregnancy.

https://doi.org/10.1371/journal.pmed.1005153.t002

Comparison to low SEP

For each outcome, we compared the estimated effect of each health behaviour in pregnancy (ever versus never) to the estimated effect of either measure of low SEP (i.e., based on education or occupation; lowest group versus higher groups). For these analyses, we used model 2a (i.e., the standard adjustment set) to ensure that the health behaviour model was adjusted for SEP while maximising power and minimising selection by not adjusting for the co-parent’s health behaviour. If the association between low SEP and the outcome is similar to, or greater than, the association between the health behaviour and the outcome, then the effect of the health behaviour might reflect residual confounding by SEP. In this case, the health behaviour may not be a specific or strong causal factor but rather part of a broader pattern of disadvantage associated with low SEP.

EPoCH Explorer

To share, visualise, and draw inferences from our results, we developed EPoCH Explorer, an RShiny web app, which can be accessed at https://gcsharp.shinyapps.io/EPoCH/. Via EPoCH Explorer, users can view cohort data summaries and generate data visualisations of regression and meta-analysis summary statistics. Verbal interpretations of results are provided, as well as a causal inference report that presents all information needed to triangulate and contextualise the causal evidence for any exposure specified by the user. EPoCH Explorer also allows users to download the full set of summary statistics for EPoCH.

Open science and reproducibility

The general analysis plan was outlined prospectively in the funding proposal, which is included, along with a list of changes and their rationales, in S4 File. It has been developed iteratively and checked (see S4 File) to improve reproducibility. This paper is written in accordance with the Strengthening Reporting of Observational Studies in Epidemiology (STROBE) guidelines for reporting of observational [28] (checklist in S5 File) and MR studies [29] (checklist in S6 File).

Ethics

Each cohort obtained ethical approval and written participant consent for primary data collection. No separate ethical approval was sought for this secondary data study, but it was approved by the executive boards for ALSPAC (ref B3123), BIB, and MoBa (MCS data were accessed via the UK Data Service directly).

Sample sizes

Without stratification by sex, sample sizes ranged from 35 to 232,139 (mean = 18,554) and varied according to the model, the availability of data across cohorts, and missing values for participants within cohorts. For our standard models 2a and 2b (unstratified by sex, adjusted for the standard set of covariates), the number of analyses involving more than one cohort was 8955/29146 (31%) for model 2a and 7760/28171 (28%) for model 2b. See Table A in S7 File for a detailed summary of the number of cohorts and sample sizes for each model. The maximum sample size decreased by up to 22% after adjustment for a standard set of confounders (i.e., from model 1–2), but only up to 3.1% after further adjustment for potential mediators (i.e., from model 2 to model 4). It is not possible to make a direct comparison with model 3 because this model was only run for health behaviours in the second trimester, third trimester, or postnatally, and total sample sizes were lower for these exposure variables.

Sample characteristics

Fig 1 summarises the percentage of the sample (pooled across cohorts where multiple cohorts contributed) that were in the exposed group for each binary exposure variable. Mothers self-reported lower levels of smoking and alcohol consumption than partners did, particularly during pregnancy. For caffeine consumption, differences between parents and pregnancy status were less apparent. Less than 23% of parents were categorised in the lowest SEP groups. In all cohorts, low SEP was correlated with high smoking (and caffeine consumption except in MoBa). In all cohorts except BiB, low SEP was weakly correlated with low alcohol consumption (Fig A in S7 File). In general, cohorts had prioritised collecting data on mothers, so more variables were available (S2 File), and the data were more complete compared to data on partners. S8 File contains a detailed breakdown of sample characteristics by cohort, including the distribution and level of missingness for each variable.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Summary of parental exposures in EPoCH.

Percentages are the exposed sample pooled across all contributing cohorts, divided by the total non-missing sample pooled across those cohorts. SEP, socioeconomic position.

https://doi.org/10.1371/journal.pmed.1005153.g001

General trends

Across all models (i.e., models 1a to 4b across all health behaviours, GRS, SEP, and outcomes), effect estimates were generally small (absolute Cohen’s D < 0.2) and P-values large (FDR-adjusted P > 0.05). In our standard models, the percentage of results with D > 0.2 and FDR-P < 0.05 (i.e., strong statistical evidence of a larger-than-small associations) was 6% before adjustment for the co-parent’s exposure (model 2a), and 3% after (model 2b).

We did not find widespread evidence of much stronger effects in mothers than partners: after mutual adjustment (model 2b), 51% of analyses had a larger effect estimate for mothers than partners, and 52% had a smaller p-value for mothers than partners. The percentage of model 2b results with D > 0.2 and FDR-P < 0.05 was 7% for mothers, compared to 5% for partners.

We observed some variation between exposure and outcome classes (Fig 2 and Fig B in S7 File). Across all outcomes, the strongest associations (D > 0.2 and FDR-P < 0.05) were seen with low SEP. In model 2b (standard adjustment with adjustment for the exposure in the co-parent), 15% of the results for low SEP had D > 0.2 and FDR-P < 0.05, compared to 6% for smoking, 3% for alcohol, and 0.4% for caffeine. Across all exposures, strongest associations were seen with psychosocial and cognitive outcomes (especially when the exposure was low SEP, smoking, or partner’s alcohol consumption) and body size and composition (especially when the exposure was smoking). For smoking, caffeine, and SEP, we saw a consistent pattern: a slightly higher proportion of strong associations for mothers compared to partners. In contrast, for alcohol, the proportion of strong associations was higher for partners. Notably, the direction of the partner association was often opposite to that seen for mothers (particularly for psychosocial outcomes), suggesting a potential protective effect of partner’s alcohol consumption during pregnancy. However, associations with the partner alcohol GRS (Fig F in S7 File; F-statistic for instrument strength = 10 (Fig B in S1 File)) were mostly null, indicating that this apparent protective effect could be explained by residual or unmeasured confounding. We found no clear patterns according to the timing of exposure or age at which outcomes were assessed (Figs C to E in S7 File).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Summary of results by exposure and outcome class.

Presented for the standard regression model with adjustment for co-parent’s exposure (model 2b). Dark green indicates the strongest evidence (FDR < 0.05) of a larger-than-small effect (D > 0.2). SEP, Socioeconomic Position; BMI, Body Mass Index; SDQ, Strengths and Difficulties Questionnaire; CBCL, Child Behavioural Checklist; M-CHAT, Modified Checklist for Autism in Toddlers; SCQ, Social Communication Questionnaire; SCDC, Skuse Social Communication Scores; ASQ, Ages and Stages Questionnaire; LDL, Low Density Lipoprotein; HDL, High Density Lipoprotein; IL6, Interleukin-6; IL7, Interleukin-7; Apo-A1, Apolipoprotein A1; Apo-B, Apolipoprotein B.

https://doi.org/10.1371/journal.pmed.1005153.g002

Triangulation of evidence for causal effects in pregnancy

In model 2b (standard adjustment with adjustment for the exposure in the co-parent), the number of analyses where the exposure was a health behaviour that occurred during pregnancy was 10,564 for mothers and 6,049 for partners. Of these, 1,099 (10%) and 585 (9%), respectively, had FDR-P < 0.05. Fig 3 illustrates triangulation of causal evidence for each of these associations. Note that a lack of evidence does not indicate evidence against a causal effect during pregnancy; it may reflect limited data availability or differences in assumptions across contexts. For most exposure-outcome combinations, the direction of association is consistent between the equivalent observational (MVR) and GRS (MR-informed) analyses. However, for caffeine consumption, this is not always the case, which could reflect residual confounding in the observational (MVR) estimates, or bias in the MR-informed results if the GRS is a “weak instrument”—that is, only weakly related to the reported exposure (see S1 File for F-statistics for instrument strength).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Triangulation of evidence for effects of parental health behaviours during pregnancy.

MVR, multivariable regression (model 2b); PC, parental comparison negative control (model 2b); MR, mendelian randomisation-informed genetic risk score (model 2a); D, dose–response (model 2b); PN, postnatal comparison negative control (model 3b); T, triangulation summary indicating the number of lines of evidence across different analytical approaches that agree on the direction of association and have regression P < 0.05. BMI, Body Mass Index; SDQ, Strengths and Difficulties Questionnaire; CBCL, Child Behavioural Checklist; M-CHAT, Modified Checklist for Autism in Toddlers; SCQ, Social Communication Questionnaire; SCDC, Skuse Social Communication Scores; ASQ, Ages and Stages Questionnaire; LDL, Low Density Lipoprotein; HDL, High Density Lipoprotein; IL6, Interleukin-6; IL7, Interleukin-7; Apo-A1, Apolipoprotein A1; Apo-B, Apolipoprotein B.

https://doi.org/10.1371/journal.pmed.1005153.g003

For mothers, we found five agreeing lines of evidence across analytical approaches for an effect of smoking during pregnancy on: small for gestational age (SGA) at delivery; waist circumference at age 5–7, BMI and overweight/obese at age 8–11; conduct problems, externalising traits, and hyperactivity/inattention at ages 3–4 and 5–7; aggression at age 5–7; total behavioural issues at age 3–4; and depressive symptoms at age 8–11 (full results and heterogeneity statistics presented in Table B in S7 File). It should be noted that, when adjusting for child’s and co-parent’s GRS (i.e., Model 4b) in the MR analysis (which can be an appropriate strategy to reduce genetic confounding and collider bias [30]—see S3 File for a discussion), P-values were >0.05 for the body size and composition outcomes, possibly due to low statistical power. The body size and composition outcomes were all more strongly associated with maternal smoking than with low SEP (i.e., model 2a MVR effect estimates were further from the null). However, the psychosocial outcomes were more strongly associated with low SEP than maternal smoking.

For partners, we found four agreeing lines of evidence for an effect of smoking during pregnancy on BMI at age 5–7, BMI, overweight/obese, and weight at age 8–11; internalising traits at age 1–2 and 5–7; total behavioural issues at age 1–2, and social communication difficulties at age 1–2. We also found four agreeing lines of evidence for an effect of partner’s alcohol consumption during pregnancy on internalising traits (full results and heterogeneity statistics presented in Table C in S7 File). Additional adjustment for child’s and co-parent’s GRS (Model 4b) resulted in P-values >0.05 for the body size and composition outcomes and the associations with alcohol consumption. All identified outcomes were more strongly associated with low SEP than with partner health behaviours. It should also be noted that, although the parental comparison negative control was not included in the triangulated evidence for partner effects (because the assumptions are less likely to hold), the maternal effect was generally further from the null than the partner effect for these exposure-outcome combinations. This pattern may indicate residual confounding by maternal exposure, or the presence of independent maternal and partner effects. The exceptions were smoking-social communication difficulties and alcohol-internalising traits, where findings instead suggest specific partner effects.

Example of using EPoCH Explorer to triangulate causal evidence

Fig 4 shows screenshots from EPoCH Explorer to illustrate how a researcher might use the app to refine and assess evidence for (as an example) a hypothesis about prenatal influences on child psychosocial outcomes. To select an outcome of interest, the researcher could start by generating a Manhattan plot of the minimally adjusted (model 1a) results for psychosocial traits, which would reveal that the strongest associations are between maternal smoking during pregnancy and child externalising traits at age 5–7. To explore this hypothesis further, they could generate a causal inference report (example in S9 File), which would show that: (1) the MVR association persists after adjusting for potential confounders; (2) the direction of the estimate for the smoking initiation GRS aligns with the MVR estimate; (3) the maternal estimate is larger than the partner’s; (4) the estimate is larger for smoking during pregnancy than postnatally, (5) a dose-response relationship exists. These five lines of evidence support a causal effect of maternal smoking in pregnancy on externalising traits at age 5–7 years. The effect estimate for maternal smoking ever in pregnancy (model 2a β = 0.21 (95% confidence interval (CI) [0.18,0.25], P = 1.4 × 10⁻³¹) is smaller than the effect estimates for low SEP (education β = 0.30,95% CI [0.26, 0.34], P = 2 × 1 0⁻⁴⁶, occupation β = 0.26, 95% CI [0.21, 0.31], P = 1.5 × 10⁻²³), which suggests the effect may not be specific to smoking, but part of a broader pattern of disadvantage. The researcher could then perform a ‘Deep Dive’ to explore (for model 2b) cohort sample sizes, summary statistics, a forest plot, and a verbal interpretation of the findings. This shows that three cohorts (ALSPAC, MCS, MoBa) contributed to the meta-analysis, with a sample size of 40,014 (n smokers = 2,971; 7%), and low heterogeneity (I² = 23.4; heterogeneity P-value = 0.27). The researcher could also use EPoCH Explorer to investigate the impact of stratifying by sex, or adjusting for mediators (i.e., comparing to model 4). Finally, the researcher could download all results to analyse offline and/or to compare to their own data.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Screenshots from EPoCH Explorer.

(A) The navigation bar and top of the welcome page. (B) An interactive volcano plot comparing maternal to partner estimates for smoking and psychosocial outcomes. (C) An interactive Manhattan plot showing the distribution of P-values by exposure class for all psychosocial outcomes. (D) A ‘Deep Dive’ summary of the results from model 2B, including a verbal interpretation of results and an interactive forest plot. EPoCH Explorer can be accessed at: https://gcsharp.shinyapps.io/EPoCH/.

https://doi.org/10.1371/journal.pmed.1005153.g004