Abstract:
|
Disproportionate stratified sampling is common in research settings where one stratum or (sub)population is of greater interest compared to another, and has progressively become prevalent in clinical trials with advances in identification of predictive biomarkers. Biomarker-enriched trials are highly desirable for patients, sponsors, regulators and payers as they enhance the benefit-risk ratio for trial subjects enabling more efficient trials. Such a sampling scheme however, introduces bias in estimating effects in the overall population unless information is properly synthesized. Statistical inference for the treatment effect in the overall population is of interest for various reasons including lack of specificity or sensitivity of biomarker assays and uncertainty of association between the biomarker and clinical outcome. Significant complexities in inference arise in this context of censored time-to-event endpoints since standard univariate summaries, such as the ubiquitous hazard-ratio, are nonlinear in nature and standard log-rank tests may not be represented as a simple Wald-type test statistic. This paper provides a solution to this problem.
|