|
Abstract:
|
Common remedies to minimize confounding bias are either to stratify on unbalanced factors or to adjust for these factors by including them as covariates in analytical models. Though traditional stratified analysis provides a robust test of comparison, increasingly studies are using regression models as the main analytical tools to adjust for potential confounding factors as well as to identify prognostic predictors. While these straightforward covariate adjusted multiple regression models can also provide similar treatment effects, there lacks detailed breakdown of subgroup analyses results and strata information. Here, we consider an approach that combines both stratification and covariate adjustment methods, i.e. adjust unbalanced factors using propensity score methods and stratify the factor by subgroup information when interaction between the stratified factor and the treatment is absent. The detailed method is discussed for various different scenarios. Multiple statistical models are used to illustrate the proposed method using a large registry study and a randomized clinical trial. Meanwhile, the potential for decrease in power and inflation of type I error are evaluated.
|
