Modelling clinical trials in heterogeneous samples.

Individual variation in genetic, phenotypic and environmental factors could lead to significant differences in rates of drug metabolism, in clinical responses to drugs, and in drug side effects. The impact of population heterogeneity on treatment effect estimation and on assessment and application of clinical trial findings has been less than fully studied. In this paper, we studied the properties of models that reflect population heterogeneity of clinical trial samples by unmeasured covariates such as genetic susceptibility. The impact of heterogeneity on the estimation of treatment effect in a two-armed placebo or active-controlled clinical trial was quantified using logistic regression models. We also proposed a two-stage clinical trial, where the effects of individual drug-response-related covariates were estimated in a 'training data set' in the first stage of the trial. In the second stage of the trial, a subgroup of individuals with enhanced (or reduced) sensitivity to drug treatments in a heterogeneous risk cohort was identified based on information about their drug-related characteristics. Only these 'responders' were included in the subsequent efficacy test. Our simulation results showed that population heterogeneity could lead to biased estimation of treatment effect not only in the contaminated groups but also in the uncontaminated groups. A two-stage trial could greatly increase the power of an efficacy test over a 'full trial' even with fewer individuals enrolled in the trial; results would apply to a more highly specified population. The bigger the covariate effects, the more efficient the two-stage trial is compared to a 'full trial'. Copyright 2005 John Wiley & Sons, Ltd.