|
Abstract:
|
When can propensity score methods solve problems that standard covariate adjustment cannot? While some view propensity score methods primarily as a tool for dimension reduction of a complex covariate space (with all important covariates measured), the problem of biased estimates of treatment effect can arise with even a single covariate. We examine a simple model, conceptually and via simulations, to show that standard regression can appropriately handle differing covariate distributions between treatment groups, but when model misspecification is also present, estimates from standard regression are biased. We compare analyses using crude estimates (no adjustment), regression, and propensity score methods. When covariate distributions differ by treatment group and models are mis-specified, only the propensity score method provides an unbiased estimate of treatment effect.
|
