Gestational weight gain and long-term maternal obesity risk: A multiple bias analysis

Abstract

BACKGROUND

Lifecourse research provides an important framework for chronic disease epidemiology. However, data collection to observe health characteristics over long periods is vulnerable to systematic error and statistical bias. We present a multiple-bias analysis using real-world data to estimate associations between excessive gestational weight gain and midlife obesity, accounting for confounding, selection, and misclassification biases.

METHODS

Participants were from the multi-ethnic Study of Women’s Health Across the Nation. Obesity was defined by waist circumference measured in 1996 to 1997 when women were age 42 to 53. Gestational weight gain was measured retrospectively by self-recall and was missing for over 40% of participants. We estimated relative risk (RR) and 95% confidence intervals (CI) of obesity at midlife for presence versus absence of excessive gestational weight gain in any pregnancy. We imputed missing data via multiple imputation, and used weighted regression to account for misclassification.

RESULTS

Among the 2,339 women in this analysis, 937 (40%) experienced obesity in midlife. In complete case analysis, women with excessive gestational weight gain had an estimated 39% greater risk of obesity (RR=1.4, CI=1.1, 1.7), covariate-adjusted. Imputing data, then weighting estimates at the guidepost values of sensitivity=80% and specificity=75%, increased the RR (95% CI) for obesity to 2.3 (2.0, 2.6). Only models assuming a 20-point difference in specificity between those with and without obesity decreased the RR.

CONCLUSIONS

The inference of a positive association between excessive gestational weight gain and midlife obesity is robust to methods accounting for selection and misclassification bias.

INTRODUCTION

Lifecourse research provides an important framework for understanding chronic disease. In epidemiology, the lifecourse approach acknowledges the contributions of physical, social, and behavioral factors in long-term disease risk (1). Although many types of study designs can address lifecourse questions, observing health characteristics over multiple decades remains challenging (2). A major obstacle to lifecourse studies is managing systematic error in data collection. Systematic error may or may not bias estimates of association away from the true effect, and intuition around these biases is generally poor (3, 4). Quantitative bias analysis is an important but under-utilized tool in lifecourse epidemiology, representing a range of accessible analytic options to quantify the potential bias around an estimate (5, 6).

We present an illustrative case study of a multiple bias analysis using real-world data. We describe the analytic approach motivated by a practical example that can be generalized to other lifecourse studies. Our long-term goal is to evaluate whether excessive gestational weight gain contributes to maternal cardiovascular risk across midlife through a pathway that includes obesity. We previously demonstrated an association between the number of pregnancies with excessive gestational weight gain and midlife maternal BMI in the Study of Women’s Health Across the Nation (SWAN). However, as we evaluated the hypothesized causal pathway, we became concerned that a portion of the association that we observed may have been due to bias from common sources of systematic error in studies of maternal health.

We identified two potential sources of systematic error in the SWAN reproductive history data: participant attrition and use of self-recall to measure pregnancy characteristics. Over 40 percent of cohort participants who would otherwise be eligible for this analysis had missing data, primarily due to study dropout. Because participant attrition is often related to demographic and clinical characteristics that are associated with health outcomes (7–9), substantial loss to follow-up may induce selection bias. Misclassification bias was also a concern in these data. Our primary exposure, gestational weight gain, was measured by retrospective self-report at an average of 30 years after the participants’ last birth. Self-recall is a common measure for pregnancy weight characteristics (10). Moderate validity and reliability have been documented for self-recalled pregnancy weight characteristics, and its use generally has not been found to bias estimates of association (10–12). However, bias analyses from a large birth cohort concluded that measurement error in self-reported prepregnancy weight biased association estimates between weight and birth outcomes (13).

Having identified these challenges, we sought to address both selection bias and misclassification bias in our data. The aim of this manuscript is to describe the approach that we used to quantify potential bias around the estimate of association between excessive gestational weight gain and maternal obesity at midlife in the SWAN cohort.

METHODS

Participants

The Study of Women’s Health Across the Nation is a prospective, multiethnic, multi-center study designed to follow women through the menopause transition. The study was conducted at seven sites: Boston, MA, Chicago, IL, Detroit, MI, Los Angeles, CA, Oakland, CA, Newark, NJ, and Pittsburgh, PA. Each site recruited women who identified as Non-Hispanic White as well as women from one additional race or ethnic group: Non-Hispanic Black, Hispanic, Japanese, or Chinese. Enrollment began in 1996 with the following primary eligibility criteria: age 42 to 52, having at least one menstrual period in the previous 3 months, no exogenous hormone use in the previous 3 months, intact uterus, at least one ovary, and self-identification with a designated racial or ethnic group recruited by site. More information on the sampling strategy for the study has been published previously (14). Institutional Review Board approval was obtained with each site institution and written consent given by all participants.

Figure 1 is a participant and data collection flow-chart. Participants were eligible for this analysis if they reported a history of live birth at the baseline interview (n=2733, conducted 1996 to 1997). A full reproductive history questionnaire was administered at the 13th follow-up visit (conducted in 2011 to 2013). This questionnaire collected data on prepregnancy weight and gestational weight gain for each birth. A subset of the women identified as eligible at baseline were retained through the 13^th follow-up visit and completed the questionnaire.

Figure 1.

Participant and Data Collection Flow-Chart

We excluded women from this analysis if they reported at baseline a history of stillbirth or multifetal birth (n=154). We hypothesized that weight gain during these births may have a different impact on long-term maternal health than that of live or singleton births. Women with a history of underactive thyroid (n=195) were also excluded due to known associations between hypothyroidism and pregnancy complications (15, 16), and because we did not know whether the reported thyroid condition was diagnosed before, during, or after pregnancies. We excluded women for missing the outcome (midlife waist circumference), pregnancy outcome data (i.e. live birth, stillbirth, miscarriage, or abortion), later reporting a birth that occurred after the outcome assessment, or later reporting conflicting information on pregnancy outcomes (n=45). The analytic sample is comprised of 1340 women with complete data and 999 with some imputed data, totaling 2339 participants.

Measures

Outcome: Midlife Abdominal Obesity.

The primary outcome was midlife abdominal obesity based on waist circumference measured at the SWAN baseline visit when women ranged in age from 42 to 53 years. Waist circumference was measured by trained staff according to a standard protocol. Abdominal obesity was defined as a waist circumference > 80 cm for Japanese and Chinese women and ≥ 88 for White, Hispanic, and Black women. All participants included in the analytic sample had a measure for waist circumference, therefore no outcome values were imputed.

Exposure: History of Excessive Gestational Weight Gain.

We defined the primary exposure as ever having a pregnancy with excessive gestational weight gain. Total gestational weight gain for each live birth was collected by retrospective self-report at visit 13, when women ranged in age from 56 to 68. Participants were asked to recall prepregnancy weight, gestational weight gain amount, and gestational age for each birth. We calculated prepregnancy body mass index (BMI) with retrospective prepregnancy weight collected at visit 13 and height measured in-clinic at the baseline visit. Each pregnancy was categorized as having inadequate, adequate, or excessive gestational weight gain per the Institute of Medicine’s 2009 guidelines (17). These guidelines represent the impact of gestational weight gain on maternal health and birth outcomes based on the current epidemiologic literature. We categorized pregnancies reported as term births by adequacy range for total gestational weight gain. We categorized births reported as preterm by adequacy of weight gain rate in the second and third trimester (see Table 1). We assumed 4.4 pounds of gain in the first trimester for preterm births, consistent with the guidelines’ range of 0.5 to 2 kg (1.1–4.4 pounds). In sensitivity analysis, shifting the definition to use the lower limit (1.1 pounds) had a modest effect on estimates indicating that the overall results in this paper are robust to this preterm birth threshold. Questionnaire items from the reproductive history form are included as eAppendix C. Alternative BMI cutoffs were used in creating the gestational weight gain adequacy variables for Japanese and Chinese participants, with overweight defined as ≥23 kg/m² and obese as ≥25 kg/m². This is consistent with recommendations from the Western Pacific Region WHO (18) and prior research in Japanese and Chinese populations living in North America (19, 20).

Table 1.

Gestational Weight Gain Adequacy Definitions

Prepregnancy BMI Category	BMI Category Definition (kg/m2)		Range for Adequate Total GWG (pounds)a	Range for Adequate GWG Rate, 2nd and 3rd Trimester (pounds/week)a
	NH White, NH Black, and Hispanic	Japanese and Chinese ethnicity
Underweight	< 18.5	< 18.5	28–40	1.0–1.3
Normal weight	18.5–24.9	18.5–22.9	25–35	0.8–1.0
Overweight	25.0–29.9	23.0–24.9	15–25	0.5–0.7
Obese (all classes)	≥ 30.0	≥ 25.0	11–20	0.4–0.6

Abbreviations: BMI, body mass index; GWG, gestational weight gain; NH: Non-Hispanic.

^aAdequacy ranges: Institute of Medicine (US) and National Research Council (US) Committee to Reexamine IOM Pregnancy Weight Guidelines; Rasmussen KM, Yaktine AL, editors. Weight Gain During Pregnancy: Reexamining the Guidelines. Washington (DC): National Academies Press, 2009.

Covariates.

Covariates collected at the baseline visit – concurrent with the outcome assessment – were age (years), race/ethnicity (Non-Hispanic Black, Chinese, Japanese, Hispanic, Non-Hispanic White), parity, education level (categorized as high school or less, some college/college degree, or post-college study), age at first pregnancy (years), time since last pregnancy (years), smoking status (current, previous, or never smoker), difficulty in paying for basics (somewhat hard/very hard, or not very hard), menopausal status (premenopause or early perimenopause), daily caloric intake (kcal), physical activity (score), and stress level (score). We also included study site as a covariate.

Covariates measured by retrospective self-report at follow-up 13 were number of pregnancies with a gestational hypertensive disorder and number of pregnancies with gestational diabetes.

Auxiliary Variables.

Multiple imputation models included all analysis variables and characteristics we hypothesized to be associated with loss-to-follow up, based on previous literature (7–9, 21–27). These were: current health insurance (yes/no), current employment status (yes/no worked for pay in the last 2 weeks), language acculturation (high versus low or medium), domestic violence (yes/no report of being “Slapped, kicked, or otherwise hurt by husband/partner or someone else important to you” in the past year), very upsetting or stressful life event in the past year (yes/no), marital status (yes/no currently married), comorbidities (yes/no self-report of ever had: heart attack/myocardial infarction, angina, diabetes, arthritis or osteoarthritis, high blood pressure or hypertension, high cholesterol, overactive thyroid, osteoporosis, or stroke), social support (score 0–16), depression (CES-D scale score 0–60), hostility/cynicism (score 0–13), and four quality of life scores (0–100) calculated from the 36-Item Short Form Health Survey (SF-36): physical functioning, pain, vitality, and social functioning. All auxiliary variables were measured at the baseline visit.

Statistical Analysis

Participant Characteristics.

We present participant characteristics overall and stratified by missing data status. We also summarize participant characteristics among women with complete data, stratified by whether women had reported any pregnancies with excessive gestational weight gain or none. Categorical variables are shown as number (%), continuous variables with approximately normal distribution are shown as mean and standard deviation, and those with a skewed distribution are shown by median with first and third quartile values.

Analysis Models.

Exposure, outcome, and covariate characteristics were determined with the guidance of a directed acyclic graph (Figure 2). Covariates include characteristics hypothesized to be confounders, potential mediators, and those associated only with the outcome (used to improve efficiency). The base model estimated relative risk and 95% confidence intervals of abdominal obesity for ever- versus never-having excessive gestational weight gain using generalized linear regression with a log link. We modeled regression under the Poisson distribution due to convergence issues with the binomial distribution. Unadjusted, minimally adjusted, and fully covariate-adjusted models are presented. We conducted two sensitivity analyses on the base model. First, we estimated the relative risk of midlife obesity for the number of excessive gestational weight gain pregnancies. Second, we stratified analysis based on obesity status prior to pregnancy.

Figure 2.

Directed acyclic graph of effects hypothesized in this analysis

Accounting for Missing Data.

Our imputation method was multiple imputation by chained equations, a widely used, flexible method to address missing data. We included all analysis and auxiliary variables in the imputation model. We imputed missing reproductive exposures as continuous values representing the number of pregnancies with excessive gestational weight gain, inadequate gestational weight gain, hypertensive disorder of pregnancy, and gestational diabetes. Imputation models used continuous values to maintain the level of detail available in our data. This approach can result in illogical imputed values (for example having more hypertensive pregnancies than births). However, simulation studies have demonstrated that using raw values to impute characteristics with logical restrictions produces less bias in analysis models compared to limiting or rounding the range of values (28–31). We also imputed missing data for some covariate variables.

We hypothesized that preterm birth status modified the effect of inadequate gestational weight gain on maternal midlife waist circumference. Experiencing a preterm birth is associated with long-term maternal cardiovascular risk (32–34), and inadequate gestational weight gain is associated with risk of preterm birth (35, 36). Preliminary data visualizations in complete case data supported the hypothesis. Therefore, we looked for an imputation method that could accommodate non-linear relationships. Classification and regression tree algorithms have been put forward in the literature as a promising method to create imputed datasets that maintain interactions (37–39). We also included a more traditional approach using predictive mean matching and logistic regression within the multiply imputed chain equation model for comparison. eFigure 1 illustrates the imputation process using classification and regression tree and traditional multiply imputed chain equations.

We created ten imputed datasets for each imputation method. Diagnostics included visual inspection of trace plots to assess convergence of models. In addition, we compared the distributions of the primary imputed variables between observed and imputed values. While under the missing at random assumption the distribution of imputed versus observed values may differ, distributions should be consistent conditional on the probability of being observed. Therefore, we also estimated predicted probabilities of being observed within each imputed dataset, averaged the probabilities across imputed datasets per participant, and plotted them against each imputed variable (40).

Accounting for Misclassification.

To adjust estimates for misclassification of the exposure, we first calculated misclassification weights based on validation studies of gestational weight gain recall. Misclassification weight calculations are based on the sensitivity and specificity of the measure of interest, using the method described by Johnson et al (41). We estimated relative risk of midlife obesity in misclassification-weighted, covariate-adjusted regression models with pooled estimates using classification and regression tree-imputed datasets. See eAppendix B for a sample of the code used to build these models.

Validation studies of pregnancy weight characteristics measured by maternal recall are well summarized by Headen et al (10). Women often underestimate pre-pregnancy weight (13) and overestimate gestational weight gain, resulting in a trend of over-reported excessive gestational weight gain prevalence in the literature (10). Among studies measuring pregnancy weight characteristics multiple years after birth, the mean deviation from the true value was less than 1 kg. However, the magnitude of error varied widely among women and was greater among those with higher BMIs and those of minority race or ethnicity (10). These trends have been supported in more recent validation studies (11, 42–44).

Although pregnancy weight recall is well studied, there is little published that can be translated directly into sensitivity and specificity values of the measure. Studies define misclassification inconsistently, with most presenting only a measure of correlation. Therefore, we relied on studies with published data tables to directly calculate observed sensitivity and specificity as guideposts in creating weights. We assumed that women’s self-recall of pregnancy weight was better than chance, i.e. sensitivity + specificity > 1.

We identified two studies with sufficient published data to calculate sensitivity and specificity. McClure et al (45) assessed the validity of maternal recall of gestational weight gain adequacy among 503 women at an average follow-up of 8 years postpartum. Based on their published data, we calculated that the overall sensitivity and specificity of recalling excessive gestational weight gain in a single pregnancy was 80% and 73%, respectively. Bodnar et al (46) compared gestational weight gain adequacy based on birth certificate data, which relies on self-report of prepregnancy weight collected at delivery, with medical records in 1204 women. This validation sample was selected using a balanced design stratified by race, weight, and gestational age categories from a large birth registry sample (n=853,559). From supplementary materials we calculated sensitivity of 85% and specificity of 86% within the validation sample for reporting high gestational weight gain (defined as reporting total gestational weight gain > 80^th percentile). We also applied these rates to the reported agreement in the larger birth registry sample, resulting in sensitivity of 79% and specificity of 92%. We then tested a range of sensitivity and specificity values around these benchmarks.

We further hypothesized that women with the outcome of abdominal obesity at midlife may be more likely to over-report excessive gestational weight gain than those without obesity. This misclassification scenario is of particular concern as it could induce an artifactual positive association between excessive gestational weight gain and midlife obesity. We tested 20 combinations of sensitivity and specificity stratified by outcome status. Our references for differential misclassification were again from supplementary data published by Bodnar (46). Within the validation sample, women with an obese BMI (n=575) had sensitivity of 76% and specificity of 85% compared to those with underweight, normal, or overweight prepregnancy BMI (n=618), who had sensitivity of 94% and specificity 87%. When agreement was applied to the registry sample, women reported with lower sensitivity and higher specificity compared to the validation sample.

Software.

Imputation models were run using the R mice package (47) in R version 3.6.1 (48). All other analyses were run in SAS v. 9.4 (SAS Institute, Cary, NC, USA).

RESULTS

Participant Characteristics

The analytic sample included 2339 women representing 5605 births. Reproductive history or covariate data were missing for 999 (43%) participants (Figure 1). Of the 999 women with missing data, 590 (59%) were inactive in the study by visit 13, including 71 deaths. Women with missing data were more likely to be Black or Hispanic than White, had higher mean parity (2.5 births versus 2.3 births), and were more likely to have lower educational attainment compared to those with complete data (Table 2). Overall, 937 (40%) experienced the outcome of midlife abdominal obesity. Women with missing data were more likely to experience midlife obesity (45%) compared to those with complete data (36%). Among women with full data (n=1340 participants, 3097 births), 544 (41%) reported at least one pregnancy with excessive gestational weight gain. Supplementary eTable 1 shows participant characteristics stratified by excessive gestational weight gain among complete cases.

Table 2.

Participant Characteristics at Time of Midlife Waist Circumference Assessment

		Total	Missing Data Status
		(n=2339)	Complete Case Analysis Sample (n=1340)	Missing Reproductive History or Covariate (n=999)
Age, mean (SD)		46.4 (2.7)	46.6 (2.6)	46.2 (2.7)
Race/Ethnicity, n (%)
Black		705 (30)	364 (27)	341 (34)
White		999 (43)	628 (47)	371 (37)
Chinese		188 (8)	123 (9)	65 (7)
Hispanic		235 (10)	89 (7)	146 (15)
Japanese		212 (9)	136 (10)	76 (8)
Education, n (%)
High school or less		645 (28)	298 (22)	347 (36)
Some college or degree		1239 (54)	745 (56)	494 (51)
Post-college study		431 (19)	297 (22)	134 (14)
Smoking Status, n (%)
Never smoker		1358 (58)	823 (61)	535 (54)
Past smoker		544 (23)	335 (25)	209 (21)
Current smoker		435 (19)	182 (14)	253 (25)
Difficulty Paying for Basics, n (%)
Not very hard		1332 (57)	846 (63)	486 (49)
Somewhat or very hard		991 (43)	494 (37)	497 (51)
Menopausal Status, n (%)
Pre menopause		1261 (54)	745 (56)	516 (53)
Early perimenopause		1061 (46)	595 (4)	466 (48)
Perceived Stress Score, mean (SD)		8.6 (3.0)	8.5 (2.9)	8.8 (3.0)
Daily Caloric Intake (kcal), mean (SD)		1870 (779)	1834 (740)	1920 (826)
Physical Activity Score, mean (SD)		7.7 (1.8)	7.8 (1.8)	7.5 (1.8)
Parity, mean (SD)		2.4 (1.1)	2.3 (1.1)	2.5 (1.2)
Years Since Last Birth, mean (SD)		15.1 (6.7)	15.0 (6.8)	15.2 (6.7)
Age Pregnant First Time, mean (SD)		23.4 (5.5)	24.0 (5.5)	22.6 (5.4)
Abdominal obesity, n (%)		937 (40)	486 (36)	451 (45)
Waist circumference (cm), median (Q1, Q3)		83 (74, 95)	81 (73, 93)	85 (75, 97)

Abbreviations: SD, standard deviation; Q1, first quartile; Q3, third quartile.

Complete Case Analysis

Among women with complete data, ever-having excessive gestational weight gain was associated with a relative risk for abdominal obesity of 1.8 (1.5, 2.2) in the unadjusted model (Table 3). This was attenuated to 1.4 (1.1, 1.7) in the minimally adjusted model. The addition of potential mediators—pregnancy complications and preterm births— did not substantively change estimates for excessive gestational weight gain. Sensitivity analyses estimating relative risk for each excessive gestational weight gain pregnancy were consistent with primary results (eTable 2). Models stratified by obesity status were consistent among those without prepregnancy obesity (eTable 3).

Table 3.

Complete Case and Pooled Regression Estimates: Relative Risk of Midlife Obesity

Predictor	Complete Case	Pooled Mean Matching	Pooled CART
	n=1340	n=2339	n=2339
	RR (95% CI)	RR (95% CI)	RR (95% CI)
Model 1: Unadjusted
Ever had excessive GWG pregnancy(ies)	1.8 (1.5, 2.2)	1.7 (1.5, 1.9)	1.7 (1.5, 1.9)
Model 2: Minimally Adjusteda
Ever had excessive GWG pregnancy(ies)	1.4 (1.1, 1.7)	1.2 (1.1, 1.5)	1.4 (1.2, 1.6)
Number inadequate GWG pregnancies	0.96 (0.86, 1.1)	0.97 (0.88, 1.1)	0.98 (0.91, 1.1)
Parity	1.1 (0.98, 1.2)	1.1 (1.1, 1.2)	1.1 (1.0, 1.2)
Model 3: Fully Adjusteda
Ever had excessive GWG pregnancy(ies)	1.4 (1.1, 1.7)	1.3 (1.1, 1.5)	1.4 (1.2, 1.6)
Number inadequate GWG pregnancies	0.96 (0.86, 1.1)	0.97 (0.89, 1.1)	0.98 (0.91, 1.1)
Parity	1.1 (0.97, 1.2)	1.1 (1.0, 1.2)	1.1 (1.0, 1.2)
Number hypertensive pregnancies	1.1 (0.93, 1.4)	1.1 (0.92, 1.2)	1.1 (0.95, 1.3)
Number gestational diabetes pregnancies	1.2 (0.89, 1.7)	1.1 (0.91, 1.4)	1.0 (0.82, 1.3)
Number of preterm births	0.97 (0.79, 1.2)	0.99 (0.86, 1.1)	1.0 (0.89, 1.1)

Abbreviations: CART, classification and regression tree; CI, confidence interval; GWG, gestational weight gain; RR, relative risk.

^aModels 2 and 3 adjusted for variables shown as well as study site, age at outcome measure, race/ethnicity, education, smoking, difficulty paying for basics, menopausal status, stress score, caloric intake, physical activity score, years since last birth, age first pregnant, and BMI category prior to the first pregnancy.

Accounting for Missing Data

Compared to the observed data, imputed datasets showed higher proportions of women with excessive gestational weight gain pregnancies and preterm births (eTable 4). Diagnostic scatter plots of the number of excessive gestational weight gain pregnancies against the predicted probability of being observed per woman appeared consistent between the observed and imputed data (eFigure 2).

Table 3 presents estimates of the risk of midlife obesity for the main exposures based on the observed and the imputed data. Both imputation methods resulted in consistent but diminished associations between ever had excessive gestational weight gain and midlife abdominal obesity. Pooled estimates in fully adjusted models attenuated the observed RR (95% CI) of 1.4 (1.1, 1.7) to 1.3 (1.1, 1.5) using traditional multiply imputed chain equation datasets and 1.4 (1.2, 1.6) using classification and regression tree-imputed datasets. As expected, pooled estimates had improved precision over the complete case analysis.

Accounting for Misclassification

Of the 1340 women observed with a full reproductive history questionnaire, 269 of the 486 (55%) with midlife obesity reported ever having excessive gestational weight gain, compared to 275 of the 854 (32%) without midlife obesity. All models weighted for non-differential misclassification resulted in increased estimates compared to the observed (eFigure 3, eTable 5). Weighting the covariate-adjusted pooled estimates from classification and regression tree-imputed data at the guidepost values of sensitivity=80% and specificity=75% increased the RR (95% CI) for obesity from 1.4 (1.2, 1.6) to 2.3 (2.0, 2.6).

Figure 3 and eTable 5 present estimates of RR of obesity for those with excessive gestational weight gain compared to those without, adjusted for confounding and selection bias, and weighted for misclassification assuming that misreporting differed by outcome status. Our guidepost assumption of sensitivity=95%, specificity=85% (without obesity), sensitivity=75%, specificity=85% (with obesity), resulted in a RR of 2.5 (2.2, 3.0). In all combinations tested, only models assuming a 20-point higher specificity among those without midlife obesity, compared to those with obesity, decreased the magnitude of the estimate. Models weighted for sensitivity and specificity values based directly from published validation studies increased the estimates. Figure 4 summarizes the results across analyses.

Figure 3.

Regression Estimates Adjusted for Confounding, Missing Data, and Misclassification Differential by Outcome Status: Relative Risk of Midlife Obesity for Ever- versus Never-Had Excessive Gestational Weight Gain

Abbreviations: GWG, gestational weight gain; SE0, sensitivity among those without the outcome (midlife obesity); SP0, specificity among those without the outcome; SE1, sensitivity among those with the outcome; SP1, specificity among those with the outcome; RR, relative risk; CI, confidence interval.

All models are adjusted for study site, age at outcome measure, race/ethnicity, education, smoking, difficulty paying for basics, menopausal status, stress score, caloric intake, physical activity score, years since last birth, age first pregnant, BMI category prior to the first pregnancy, inadequate gestational weight gain, hypertensive pregnancies, pregnancies with gestational diabetes, and preterm birth. Estimates are pooled across regressions on 10 CART-imputed datasets.

Figure 4.

Summary Figure of Regression Estimates: Relative Risk of Midlife Obesity for Ever- versus Never-Had Excessive Gestational Weight Gain

Abbreviations: RR, relative risk; CI, confidence interval.

Figure presents relative risk of midlife obesity for ever- versus never-had excessive gestational weight gain. Relative risk and 95% confidence intervals are displayed on the primary y-axis. Assumed sensitivity and specificity (%) of the self-recall gestational weight gain measure are displayed on the secondary y-axis. X-axis shows individual model numbers. Model 1 is unadjusted complete case analysis. Model 2 is covariate-adjusted complete case analysis. Model 3 is the covariate-adjusted pooled estimate from CART-imputed data. Models 4–17 are covariate-adjusted pooled estimates from CART imputation assuming shown sensitivity and specificity values.

DISCUSSION

We applied quantitative bias analysis to address potential systematic error in this case study using real-world data from the SWAN cohort. We found that the association between a history of excessive gestational weight gain and risk of midlife obesity among parous women persisted after accounting for selection and misclassification biases. Imputation of missing exposure data in over 40% of participants attenuated the point estimate but did not change the interpretation. Misclassification due to self-recall of the exposure was unlikely to account for the association. Most plausible scenarios to adjust for misclassification increased the magnitude of the estimates.

Self-recall is a practical approach to collect reproductive history information. Validation studies have shown a small mean difference but high variability when comparing maternal self-recall to in-person measures of pregnancy weight characteristics (10). In our data, all non-differential misclassification scenarios tested moved estimates away from the null. This is an often assumed, but not guaranteed, phenomenon depending on patterns of confounding (49). Misclassification scenarios reflecting higher rates of over-reporting among women with the outcome of midlife obesity (e.g. lower specificity) also increased the magnitude of estimates in most tested combinations, including the two scenarios based on published validation data. Models that decreased the estimates required that women with midlife obesity have moderate recall if they had experienced excessive gestational weight gain (sensitivity of 85%), but much poorer recall if they had not experienced excessive gestational weight gain compared to their counterparts without midlife obesity (specificity of 75% versus 95%). In contrast, as the sensitivity of recall among those without midlife obesity increased, estimates increased.

Our results did not show evidence that preterm birth or pregnancy complications accounted for the association between excessive gestational weight gain and midlife obesity risk. We include estimates for these characteristics, parity, and inadequate gestational weight gain to demonstrate the contribution of each in the models. However, because this analysis was designed specifically to estimate the effect of excessive gestational weight gain, estimates for covariates should be interpreted with caution (50).

Midlife is a critical period for cardiovascular risk factors in women. Changes in level of reproductive hormones as women move through the menopause transition, cease taking hormonal birth control, or initiate hormone replacement therapy, can contribute to risk factors including waist circumference (51–55). We note that, although the SWAN study was designed to be generalizable to a large proportion of US women (14), it excluded those with hormone use in the 3 months prior to enrollment. Given the relationship between reproductive hormones and weight, our results may not be generalizable to women with exogenous hormone use in early midlife.

A strength of this analysis is the comprehensive nature of the SWAN study. We were able to draw from a wide range of descriptive variables in analytic and imputation models. A further strength was the study’s rigorous data collection methods and validity of measures, including important confounders such as physical activity and diet. Our analytic outcome, waist circumference, was collected in-clinic by trained staff. Finally, the study sample represented women from five racial or ethnic groups including those of Japanese and Chinese descent, who are underrepresented in reproductive history studies within the US.

A lack of validity research directly relevant to our data is a limitation to this analysis. We identified two published studies that reported sufficient detail to calculated sensitivity and specificity of self-recall. One of these had very short recall time (recall at birth), and likely serves more as a best-case scenario than a proxy for our data. Sensitivity analyses are highly dependent on the bias parameters chosen by the investigator. Results can vary widely depending on sensitivity and specificity values or distributions tested (56, 57). Because the accuracy of a measure in a specific cohort is unknown without internal validation data, it is important to obtain documented rates from the literature when possible.

In summary, we sought to provide an applied example of quantitative bias analysis in lifecourse epidemiology while investigating a clinically relevant question in reproductive health. We estimated the risk of midlife obesity associated with a history of excessive gestational weight gain, and explored the susceptibility of this estimate to common sources of statistical bias. We found that systematic error was unlikely to account for the observed association. Useful information on risk factors can be gained from observational data even in the presence of likely systematic error, as evidenced by this analysis.