Three mistakes in current stratified analyses and confident patient subgroup identification for randomized controlled trials, using a Subgroup Mixable Estimation app

• Activity Number: 195 - Topics in Personalized/Precision Medicine - II
• Type: Contributed
• Date/Time: Tuesday, August 4, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Siyoen Kil* and Jason Hsu
• Companies: LSK Global PS and Ohio State University
• Keywords: Subgroup identification; Biomarkers; Causal Inference; Logic-respecting and collapsible efficacy measures

Abstract

This presentation demonstrates three mistakes in current stratified analyses and a new principle called Subgroup Mixable Estimation (SME) for confident subgroup identification, with the theory given in another related presentation. Numerical examples will be demonstrated to show how unconscious adaptation of mixing by prevalence strategy (current computer packages’ stratified analyses) for continuous efficacy measure to ratio efficacy does NOT produce logic-respecting mixture efficacy. While hazard ratio and odds ratio are non-collapsible so not mixable, response ratio and time ratio are logic-respecting and can be mixed. An interactive app visually displays confidence intervals for efficacy in subgroups and their mixtures as the user varies the value of a biomarker cut-point via a slider. Realistic data examples are given that show logic-respecting (and therefore collapsible) efficacy measures such as ratio of median survival times and relative response lead to confidence identification of patient subgroups, while non-collapsible efficacy measures such as hazard ratio and odds ratio will incorrectly target patients, confusing prognostic biomarkers for predictive biomarkers.