When investigating predictive biomarkers, beware of the qualitative interaction.

Abstract

2575 Background: The investigation of predictive biomarkers in clinical trials often involves estimation of the biomarker by treatment interaction effect (Royston and Sauerbrei, Journal of Clinical Oncology 2008). A so-called qualitative interaction (Gail and Simon, Biometrics 1985) occurs when a subgroup of patients appear to benefit from treatment while another subgroup is harmed. Spurious, statistically significant interactions are more likely to be observed as qualitative interactions than so called quantitative interactions, i.e. interactions where no subgroup is harmed. This issue is exemplified by recent study results (e.g., http://www.cancernetwork.com/gastrointestinal-cancer/amgen-halts-rilotumumab-development-due-increased-death-signal), highlighting the challenges to developing personalized therapies. Methods: We graphically and statistically illustrate how, under the null hypothesis of no treatment or biomarker effects, spurious qualitative interactions occur much more frequently than quantitative interactions, potentially leading to erroneous conclusions regarding a biomarker’s clinical utility. Following the approach of Mackey and Bengtsson (Contemporary Clinical Trials 2013), we propose a strategy for investigating potential predictive biomarkers that focuses on detecting clinically meaningful treatment effects, as opposed to focusing on interactions which have greater variability. We provide sample size formulae and an inference approach for use in proof-of-concept oncology clinical trials with time-to-event data. Results: By targeting inference on the maximal and minimal biomarker effects (i.e. the largest- and smallest levels of clinical benefit), the proposed method yields lower false positive rates and higher power for detecting the presence of biomarker subgroups than standard approaches. For fixed false positive rates and power, sample size savings were found to be as large as 40%. Conclusions: This new inferential method improves on existing approaches of identifying predictive biomarkers by minimizing false positive risks associated with qualitative, biomarker by treatment interactions.