Overall gestational weight gain mediates the relationship between maternal and child obesity

When studying the relationship between maternal and child obesity, previous literature has considered GWG to be a confounder to be adjusted for [27] or interaction term [19, 21, 25, 26, 28]. These studies showed evidence of a synergistic effect of adverse maternal and child outcomes when the mother has both a high BMI and gains excessive weight during pregnancy. However, there is little research on quantifying the direct and indirect mechanism of maternal obesity on child obesity, mediated through GWG. Current literature consistently shows a relationship between maternal pre-pregnancy obesity, GWG, and later child obesity [21, 29], as depicted in Fig. 1(SAS Code included in Supplemental Material). In this figure, the thick arrow illustrates the direct effect of maternal pre-pregnancy obesity on child obesity, and the arrows from maternal pre-pregnancy obesity to GWG to child obesity illustrate the indirect, mediated effect. This mediated path is of interest because it has not been previously quantified and GWG is a modifiable risk factor, which presents an opportunity for public health intervention.

The purpose of this study was to estimate the casual effect of maternal obesity on child obesity, mediated through gestational weight gain. We hypothesize that the indirect effect of maternal obesity on child obesity through GWG is positive.

The study participants were the offspring of 2595 women who received prenatal care in the Duke or Durham Regional health care system in Durham, NC between 2005 and 2009. Only women 18 years of age and older with the ability to speak English were eligible for this study. Women were excluded if they did not intend to have custody of the child, had HIV, or planned to receive obstetric care outside of Duke or Durham Regional Hospitals, or planned to move out of the area. Only children of enrollees with available clinical and survey data were considered for this study (n = 2,267). The offspring of the women formed an ongoing, longitudinal birth cohort. Figure 2 shows the progression of the full cohort to the sample selected for this study. Anthropomorphic data from the cohort was collected via electronic medical records (EMRs) during each voluntary visit to a health care facility. Because the children belong to an open cohort that uses EMRs, the age of the last recorded measurement varies. Some children in the cohort had not reached age 4, and others did not visit a study health care facility between the ages of 48 and 59 months (n = 1,417). These children were very similar to those included in the study (p > 0.05). Less than 10% of women (n = 86) that were considered underweight were also excluded from the analysis, as their offspring experience a separate set of adverse outcomes [30, 31].

Pre-pregnancy obesity, measured as continuous pre-pregnancy body mass index, was calculated using self-reported height and weight in kilograms per square meter of height (kg/m²). Weight was measured three months before gestation and height measured at the first office visit. In n = 269 women who visited Duke University Medical Center six month prior to pregnancy, the correlation between self-reported and clinic measured weight was high, (rho = 0.99, p <  0.0001).

The mediator of interest was GWG. It was computed during each clinic and study visit as the difference between the self-reported weight at last menstrual period and last weight measured up to 7 days before delivery, including the delivery visit. The clinic used the Tanita scale and stadiometer to measure weight and height—these are accurate within 0.1 pounds (lbs) and 1/8 of 1 in., respectively. Categories for GWG were inadequate, adequate, and excessive. Adequate GWG was based on the Institute of Medicine (IOM) guidelines by BMI status of 28–40 lbs. for underweight women, 25–35 lbs. for ideal weight women, 15–25 lbs. for overweight women, and 11–20 lbs. for obese women. Studies have shown that the amount of weight gained during pregnancy is linearly associated with child BMI z-score, (i.e. higher GWG means higher child BMI) [22, 32]. Therefore, the final GWG variable was dichotomized into excessive and not excessive weight gain.

Valeri and VanderWeele (2013) introduced a mediation analysis method using the counterfactual approach to account for confounders in order to estimate the controlled direct, natural direct, and indirect effects [33]. In context, the natural indirect effect (NIE) is the average difference in child BMI had all the women been of normal weight and then gained excessive gestational weight versus all women being normal weight and gained non-excessive gestational weight. The natural direct effect (NDE) is the average difference in child BMI had all the women been obese and gained non-excessive weight versus all women of normal weight and gained non-excessive gestational weight. Non-excessive gestational weight gain is the category of GWG that one would expect a normal-weight mother to gain. The controlled direct effect (CDE) is similar to the NDE when mothers gain a specific amount of weight during pregnancy. This method is also able to capture the potential interaction between the exposure and mediator, which describes the relationship between maternal obesity and GWG. Assumptions for a causal mediation analysis are: no unmeasured confounding between the exposure and outcome, exposure and mediator, mediator and outcome, and none of the mediator-outcome confounders are induced (or caused) by the exposure. The assumptions were satisfactory for our research aim and data.

The outcome of interest is child obesity at age 4. Body size was measured using the continuous BMI age- and sex-specific, standardized z-scores. The z-scores were calculated by first computing the traditional BMI of weight (kg)/height (m)², then transformed according to the 2000 CDC growth curve for children aged 2 to 19. The age, height, and weight of child were recorded during each visit, resulting in multiple follow-up measurements per year. Only children that reached age 4 were eligible for inclusion, and the last recorded height-weight pair was selected to calculate the BMI when the child was between 48 and 59 months.

Maternal race, socioeconomic status (SES), gestational diabetes, and smoking status was measured via self-report at baseline before delivery. These variables made up the minimally sufficient set needed for adjustment obtained from a causal directed acylic graph (Fig. 1), a planning tool used to reduce bias in epidemiologic studies [34]. Highest educational attainment served as a proxy for SES. Smoking status was dichotomized into no smoking during pregnancy and smoking for any period during pregnancy. In addition, a cohort variable was included to account for any differences in the enrollment periods, if present. Most children were enrolled between 2006 and 2008 or between 2009 and 2011. Additional parturition data, including maternal age, delivery mode, gestational age, infant sex and birth weight, were extracted via EMRs, and were included in a separate sensitivity analysis. Pregnancy complications, like gestational diabetes, were assessed during prenatal visits and EMRs.

The mediation analysis utilized the modified product method approach that allows for additional covariates and interaction described by Valeri and VanderWeele [33] in order to estimate the direct effect of the exposure on the outcome, as well as the effect the exposure on the outcome mediated through GWG. The model included the confounders from the minimal sufficient set. Logistic regression was used to estimate the effect of the exposure on the mediator, and linear regression to estimate the effect of the exposure and mediator on the outcome. Pre-pregnancy BMI was considered exposed at BMI of 30 and unexposed at BMI of 22. The parameter estimates from the two models were used to estimate the controlled direct effect (CDE), natural direct effect (NDE), and the natural indirect effect (NIE).

A sensitivity analysis was conducted to assess the validity of the chosen model using a fully adjusted model with all confounders including parturition and maternal clinical data. Standard errors were calculated using bootstrapping with 1000 samples within the mediation macro to estimate the direct and indirect effects of the exposure on the outcome (see Supplement). Child height and weight data were cleaned using the MonoInc package in the R programming language prior to calculating the BMI [35]. The MonoInc package identified data that departed from the monotonic pattern, then imputed using two methods, Last & Next (average of previous and next value) and fractional linear regression. The final value was a weighted average (0.3 and 0.7) of the imputed values. Multiple imputation using FCS discriminant method for arbitrary missing patterns was employed to produce more reliable estimates by imputing missing covariate data [36]. Bivariate comparisons of children by weight status were made using a t-test or Chi-square test for continuous and categorical variables, respectively. These remaining analyses were carried out in SAS 9.4.

There are limitations that should be considered. Variables associated with obesity, such as diet and physical activity, were not available in this dataset. The DAG constructed for this study does not indicate that maternal or child diet and exercise are confounders that need to be adjusted for in the models. In fact, several studies have found that physical activity and diet are strongly associated with age and SES, both of which were included in the model [44‐46]. In addition, our sample size was relatively small compared to other studies. This was partly due to the nature of how the data was collected (i.e. EMRs). However, the follow-up time did not differ by maternal or child weight status at age 4, and the prevalence of child obesity at age 4 and adult obesity are consistent with the national averages [1]. In addition, a study in the Ohio WIC population with a large sample size found a significant association between maternal and child obesity, but did not find a significant effect of GWG on child age 4 obesity status [27]. Despite a smaller sample size, this study was able to find a significant association between maternal BMI, GWG, and child obesity. Also, a common concern is that self-reported pre-pregnancy weight from which BMI is computed may be under-reported due to recall or social desirability bias. We assessed this possibility and showed, in 260 pregnant women, that self-reported weight and height were highly (99%) correlated with clinic measures done within 3 months of self-reports, and 92% in those reporting within 6 months (reference Bernard Fuemmeler, 2016).

The strengths of this study include a racially-diverse population, a longitudinal birth cohort with almost complete maternal data that establishes a temporal relationship between the outcome and exposure. This study also utilized innovative methods to quantify a causal association between maternal and child obesity. Gestational weight gain mediated 8.1% of the effect of maternal BMI on child BMI. To our knowledge, the mediated relationship between maternal and child BMI has not been previously quantified. Even with a significant direct effect of the exposure on the outcome, there was still a significant positive effect mediated through GWG.