Abstract:
|
In an observational study, the estimation of causal treatment effect is prone to selection bias. Such bias may result from confounders that lead to systematic difference in covariates between patients exposed to one treatment and those exposed to other treatment(s). Another source of selection bias is that patients tend to drop out of the study and therefore the outcomes of interest of these censored patients may not be observed, which may yield invalid causal inference when the censoring is not random. To correct the two types of selection bias, we propose a weighted regression framework, which estimates causal effects by weighting the score function of the outcome model by the product of inverse probability of treatment and inverse probability of remaining uncensored. We focus on the comparison for a binary outcome among multiple treatment groups. We apply the proposed method to estimate the effect of four common treatments for metastatic castration-resistant prostate cancer, using claims data from a large national private health insurance network with the outcome being admission to the emergency room within a short time window of treatment initiation.
|