Abstract:
|
Propensity score (PS) methods are popular strategies for controlling for observed confounders in causal inference. However, for PS methods to work well, there should be a sufficient overlap in the covariate distribution between treatment groups. This suggests (a) defining an alternative estimands, and (b) restricting analysis to a subset of cases with common support. Conditioning on covariates that are strong predictors of treatment assignment but not of outcome may severely limit the number of cases with common support, resulting in inefficient inference. In this paper, we address: (1) defining appropriate restrictions of the causal estimand; (2) strategies for identifying a common support region; (3) the impact of covariate inclusion on a PS based strategy in identifying common support; (4) accounting for model uncertainty in estimation; and (5) comparing alternative estimation methods, including inverse propensity weighting and a robust multiple imputation based approach that we proposed recently (Penalized Spline of Propensity Methods for Treatment Comparison or PENCOMP). We apply these methods to estimate the effects of antiretroviral treatment on CD4 counts in HIV+ patients.
|