From Illogical Efficacy Measures to Mistakes in Current Computer Packages – All That Is Wrong in Biomarker Subgroup Identification with Real Example and Proposed Solution

• 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): Jason Hsu* and Yi Liu and Bushi Wang
• Companies: Ohio State University and Nektar Therapeutics and Boehringer Ingelheim Pharmaceuticals, Inc.
• Keywords: Subgroup identification; Estimand; Biomarkers; Causal Inference; Logic-respecting and collapsible efficacy measures; Subgroup Mixable Estimation

Abstract

Subgroup analysis in current computer packages are misleading for binary and time-to-event outcomes, because they over-extend stratified Least Squares means (LSmeans) analysis for continuous outcomes. Hsu (1992) pointed out the initial mistakes in the Means statement of SAS Proc GLM, which was corrected in the LSmeans statement in Proc GLM and Proc Mixed. Stratified LSmeans analyses are correct for continuous outcomes. Causes of the over extension are theoretically quite deep, best explained by a new concept called "logic-respecting" efficacy measures (Ding et al 2017 Statistics in Medicine, Lin et al 2019 Biometrical Journal). This concept is stronger than "collapsibility" in Causal Inference. Unfortunately, misleading stratified analysis for binary and time-to-event outcomes in current practice has led to the incorrect impression that some biomarkers are predictive (for immunotherapy) when they are purely prognostic. Fortunately, in the Randomized Controlled Trial (RCT) setting, a new principle called Subgroup Mixable Estimation (SME), in conjunction with LSmeans, provides correct treatment comparisons stratified on subgroups for binary and time-to-event outcomes.