A Simulation Based Evaluation of Sample Size Methods for Biomarker Studies

Abstract

Cancer researchers are often interested in identifying biomarkers that are indicative of poor outcomes (prognostic biomarkers) or response to specific therapies (predictive biomarkers). In designing a biomarker study, the first statistical issue encountered is the sample size requirement for adequate detection of a biomarker effect. In biomarker studies, the desired effect size is typically larger than those targeted in therapeutic trials and the biomarker prevalence is rarely near the optimal 50% . In this article, we review sample size formulas that are routinely used in designing therapeutic trials. We then conduct simulation studies to evaluate the performances of these methods when applied to biomarker studies. In particular, we examine the impact that deviations from certain statistical assumptions (i.e., biomarker positive prevalence and effect size) have on statistical power and type I error. Our simulation results indicate that when the true biomarker prevalence is close to 50% , all methods perform well in terms of power regardless of the magnitude of the targeted biomarker effect. However, when the biomarker positive prevalence rate deviates from 50% , the empirical power based on some existing methods may be substantially different from the nominal power, and this discrepancy becomes more profound for large biomarker effects. The type I error is maintained close to the 5% nominal level in all scenarios we investigate, although there is a slight inflation as the targeted effect size increases. Based on these results, we delineate the range of parameters within which the use of some sample size methods may be sufficiently robust.