Propensity Score Weighting Using Overlap Weights: A New Method Applied to Regorafenib Clinical Data and a Cost-Effectiveness Analysis.

Abstract

BACKGROUND\nIn situations of markedly different population characteristics and weak population overlap, inverse propensity score (PS) weights suffer from extreme values. The new propensity score weighting method using overlap weights (PSOW) overcomes this limitation by estimating the overlap population at the point of highest mutual overlap, thus may be preferred to other balancing methods (trimming, target, or inverse weights) in some situations.\n\n\nOBJECTIVES\nTo evaluate the performance of PSOW with regorafenib effectiveness data from previously treated patients with metastatic colorectal cancer based on the Czech national registry data (regorafenib) and a global phase 3 randomized clinical trial (RCT) (placebo). The second goal was to assess the cost-effectiveness of regorafenib versus placebo.\n\n\nMETHODS\nIndividual data on progression-free survival (PFS)/overall survival (OS) were balanced via PSOW for age, sex, Eastern Cooperative Oncology Group performance status, number of treatment lines, metastatic colorectal cancer location, KRAS mutation, and time from metastases estimated using logistic regression. The weighted Kaplan-Meier PFS/OS curves were used in a 3-state partitioned survival model. The R code is provided.\n\n\nRESULTS\nIn comparison with target or inverse PS weights, PSOW showed remarkable performance measured by effective sample size and PS weight distribution or extreme weights despite the weak overlap between the registry and RCT. In the registry or RCT cohort, regorafenib provided better survival compared with the RCT. The new PSOW hazard ratio for OS was 0.53 (RCT: 0.79), which is conservative compared with inverse or target weights with a hazard ratio of 0.44 and 0.27, respectively.\n\n\nCONCLUSION\nThis is the first use of PSOW for clinical data and cost-effectiveness analysis. It is promising in cases of weak or small population overlap and makes pharmacoeconomic modeling, in such cases, feasible.