Gestational Weight Loss and Restrictive Weight Gain in Maternal Obesity: A Target Trial Emulation

Detailed Description Background. The 2009 IOM guidelines recommend a uniform gestational weight gain (GWG) target of 5.0-9.0 kg for all pregnant people with obesity, without distinguishing between Class I (BMI 30.0-34.9 kg/m²), Class II (BMI 35.0-39.9 kg/m²), and Class III (BMI ≥ 40.0 kg/m²) obesity. This single-range approach is increasingly questioned because the cardiometabolic consequences of obesity are not uniform across severity strata. For individuals with Class II-III obesity, restrictive gain or modest gestational weight loss may reduce maternal hypertensive disorders of pregnancy (HDP) and normalize fetal overgrowth (large for gestational age, LGA), while maternal adipose reserves may protect against small-for-gestational-age (SGA) birth. A pragmatic randomized trial assigning pregnant people to gestational weight loss raises ethical and clinical equipoise concerns, and conventional observational analyses of GWG are vulnerable to gestational-age-related survival bias and baseline confounding. Objectives. This study applies a formal target trial emulation framework to estimate the causal effect of three GWG strategies on a co-primary benefit-risk outcome pair: maternal HDP (efficacy) and fetal SGA (safety). A key secondary outcome is LGA. The central hypothesis is that the comparative benefit-risk profile of restrictive gain or weight loss versus standard IOM gain differs meaningfully across obesity classes, particularly in Class III obesity. Design and data source. This is a retrospective, nationwide target trial emulation using the U.S. National Vital Statistics System (NVSS) Natality Public Use Files. The primary emulation cohort is drawn from the 2024 singleton live-birth cohort. A 10-year physiological reference cohort (2015-2024), restricted to uncomplicated pregnancies, is used to standardize GWG into race- and gestational-age-specific Z-scores. Emulated strategies. To eliminate gestational-age-dependent survival bias, dynamic, gestational-age-specific IOM bounds are calculated for each delivery week from 24 to 41 weeks: a first-trimester gain of 1.1 kg plus a second/third-trimester velocity of 0.17 kg/week (lower bound) to 0.27 kg/week (upper bound). Participants are assigned to one of three strategies: Gestational Weight Loss (net GWG \< 0.0 kg); Restrictive Gain (GWG ≥ 0.0 kg but below the dynamic IOM lower bound); or Standard IOM Gain (GWG within the dynamic IOM range; reference group). Individuals exceeding the dynamic IOM upper bound are not included in the three-arm comparison, because the study is designed to ask whether guidelines should be lowered, not whether they should be exceeded. Causal estimation. Within each obesity class, generalized propensity scores for the three strategies are estimated via multinomial logistic regression, conditioned on continuous pre-pregnancy BMI and prespecified sociodemographic and clinical confounders. Propensity score overlap weighting (targeting the average treatment effect in the overlap population, ATO) is used to achieve covariate balance while avoiding extreme weight variance. ATO-weighted modified Poisson regression with robust sandwich variance estimators is used to estimate relative risks and absolute risk differences for HDP, SGA, and LGA. Risk differences are translated into numbers needed to treat for HDP/LGA prevention and numbers needed to harm for SGA induction. Secondary and sensitivity analyses. Prespecified effect modification is evaluated within the Class III stratum across parity, race/ethnicity, and maternal age, with ATO weights re-estimated de novo in each subgroup. A biological negative control outcome (infant male sex) is analyzed to detect residual structural confounding. Sensitivity analyses include alternative first-trimester gain assumptions (0.0 kg and 2.0 kg), E-value quantitative bias analysis for statistically significant causal estimates, and multiple imputation by chained equations to address missing data, alongside the primary complete-case analysis. Reproducibility. Data extraction, physiological Z-score modeling, and cohort construction are performed in Python 3.9.6. Overlap weighting, modified Poisson regression, marginal effect extraction, and E-value calculation are performed in R (v4.5.1).