The impact of covariate imbalance on the size of the logrank test in randomized clinical trials.

This paper examines the baseline imbalance incidental to stratified random sampling in clinical trials. We assume that patients are randomly sampled from a stratum-specific population and assigned to either treatment or control groups. Prognostic factors not used for the stratification cause within-stratum patient heterogeneity. The patient heterogeneity is investigated with an analysis of the prognostic index, or the log-relative-hazard. An index V, used to measure the degree of baseline imbalance between the groups, is defined using the prognostic index. The degree of baseline imbalance resulting from the stratified random sampling of a population of heterogeneous patients is examined. A method to estimate the exact size of the stratified logrank test is developed and applied to data obtained from a cancer clinical trial. Copyright 2000 John Wiley & Sons, Ltd.