Constrained poststratification

Abstract

Adjustment for covariates (or poststratification) is frequently used in the analysis of randomized clinical trials. The purpose of such analysis is mainly to eliminate some residual bias resulting from any imbalance between treatment groups for some important covariates. Usually, covariate effect is modeled with the data at hand. In this paper, we present a new method of poststratification ("constrained poststratification") which consists of estimating the prognostic significance of covariates in a large historical data base, transferring the model's coefficients into the (smaller) randomized trial data set, and estimating treatment effects conditional on this a priori information. In a simulated experiment, constrained poststratification allowed not only reduction of the bias but also enhancement of the efficiency of the estimation of treatment effect.