Estimation of global average treatment effect in National Heart, Lung, and Blood Institute (NHLBI) Growth and Health Study.

We propose a procedure to estimate the "time-specific average treatment effect" and "global average treatment effect" for observational studies with outcomes and covariates repeatedly measured over time. This research is motivated by the National Heart, Lung and Blood Institute Growth and Health Study (NGHS), a longitudinal cohort study that aims to evaluate the influences of race and other risk factors on the levels of blood pressure for children and adolescents. As with most longitudinal cohort studies, we do not have a known propensity score model to further discuss the average treatment effects in the NGHS. To solve this problem, a nonparametric machine learning method, the generalized boosted models (GBMs), is used to estimate the propensity score. Based on the estimated propensity score, the "time-specific average treatment effect" can be obtained through the inverse probability weighting methods, then the "global average treatment effect" is also obtained. We apply the proposed GBM-based estimation method to the NGHS blood pressure data and demonstrate through a simulation study that the GBM-based estimation method is superior to the commonly used logistic regression-based method.