🤖 AI Summary
This study addresses a critical limitation in conventional evaluation of combination therapies—the frequent neglect of heterogeneity in individual patient responses to single agents, which induces bias in treatment effect estimation. The authors propose a model-free statistical inference framework that, without imposing assumptions on functional forms, systematically tackles the non-identifiability and nonlinearity arising from such heterogeneity. They establish partial identifiability conditions for cross-world target parameters and introduce an outcome-based optimal matching strategy that achieves √N-consistent estimation and asymptotically valid confidence intervals. Reanalysis of the ACTG 175 trial demonstrates that this approach substantially enhances the accuracy, reliability, and transparency of efficacy assessment for combination therapies.
📝 Abstract
Combination therapy, a treatment approach that involves two or more monotherapies, is widely considered to enhance therapeutic efficacy across different medical conditions. It was often believed that combination therapy is beneficial because of pharmacological interactions between its component monotherapies. However, through laboratory experiments, pharmacologists have recently noted that the benefits of some combination therapies might be largely driven by varying patient-level responses to their component monotherapies. Without accounting for such patient-level heterogeneity, classical statistical inference frameworks for combination therapy might be inadequate and overly optimistic. In this paper, we introduce a novel and model-free statistical inference framework to complement the classical one and evaluate combination therapy after adjusting for patient-level heterogeneity in responses to monotherapies. We address the non-identifiability and nonlinearity issue inherent in adjustment of patient-level heterogeneity and establish conditions for the (partial) identifiability of the cross-world target parameter. We develop an outcome-based optimal matching scheme to achieve asymptotic normality and construct $\sqrt{N}$-rate confidence intervals for the target parameter, thereby enabling reliable, efficient and transparent evaluation of combination therapy amid patient-level heterogeneity. The benefits of the proposed framework are demonstrated through a reanalysis of the ACTG 175 trial.