Heterogeneous Effects of Cesarean Delivery Across Maternal and Infant Outcomes Using Population-Based Causal Machine Learning

Cesarean delivery is associated with maternal and infant outcomes, but population-average estimates may obscure variation across individuals. This study examined associations between delivery mode and four maternal and infant outcomes and explored heterogeneity in estimated treatment effects. We analyzed Pregnancy Risk Assessment Monitoring System (PRAMS) data from 2016–2020, including 197,942 women with classifiable delivery modes. Least absolute shrinkage and selection operator regression was used for covariate screening, followed by extreme gradient boosting for propensity score estimation. Stabilized inverse-probability-of-treatment weights were combined with PRAMS analytic weights for population-average logistic regression. An X-learner framework was applied to explore heterogeneity in estimated treatment effects across baseline maternal and pregnancy characteristics. Compared with vaginal delivery, cesarean delivery was associated with higher odds of infant transfer (OR, 2.27; 95% CI, 1.99–2.60; p < 0.001) and postpartum depressive symptoms (OR, 1.11; 95% CI, 1.06–1.16; p < 0.001), and lower odds of breastfeeding initiation (OR, 0.85; 95% CI, 0.81–0.88; p < 0.001). The estimated odds of infant death were lower but not statistically significant (OR, 0.77; 95% CI, 0.57–1.03; p = 0.082). Exploratory X-learner analyses indicated greater heterogeneity for infant transfer and breastfeeding initiation, whereas postpartum depressive symptoms and infant death exhibited comparatively limited variability in estimated conditional effects. Cesarean delivery was associated with differential maternal and infant outcomes, with estimated heterogeneity varying across outcomes. Combining population-average regression with exploratory machine learning approaches may provide additional insight into these associations. However, residual confounding, particularly by clinical indication, and the lack of external validation limit causal interpretation and immediate clinical application.