|
Abstract:
|
Randomized controlled trials (RCTs) typically aim to demonstrate a causal effect of a treatment on an outcome. Due to voluntary participation and inclusion/exclusion criteria, inferences made from RCTs only naturally apply to populations similar to trial participants; as a result, research has increased on extending RCT results to a broader target population. One method for estimating the target population average treatment effect (TATE) is to use inverse probability of selection weighted estimators, where the probability of selection, or propensity score, is often estimated parametrically. Positivity assumptions and similar considerations for these estimators may practically dictate that only a subsample of the RCT data or target population should be used in extending inferences. In these cases, one should consider how model misspecification in the multi-step analysis may impact estimation. Our theoretical results and simulated data examples suggest that re-estimating the propensity score using this subsample is beneficial to estimating a target subpopulation average treatment effect.
|
