(291) Propensity Score-Based Weighting Methods in The Context of Multiple Treatment Options

Publication Number: 965

Background: Several propensity score (PS)-based weighting schemes have been proposed for covariate adjustment over the past decade. These methods could be useful for head-to-head comparison of three or more treatments; however, their relative performance in this context has not been fully elucidated.

Objectives: To compare the performance of matching weights (MW), overlap weights (OW), inverse probability of treatment weights (IPTW), and average treatment effect in the treated weights (ATTW) in the context of seven treatment groups.

Methods: We used US Medicare fee-for-service data (2012-2017) to evaluate the risk of opioid overdose among individuals on long-term opioid therapy initiating one of eight skeletal muscle relaxants. Patients were allocated to seven treatment groups based on an initiated skeletal muscle relaxant and followed from the day after initiation until outcome occurrence, change in therapy, death, or end of data for a maximum of 365 days. PSs were estimated as a function of 97 baseline covariates and used to calculate weights. Weighted Cox proportional hazards regression was used to quantify hazard ratios (HRs) and 95% robust confidence intervals (CIs) for six treatment groups relative to cyclobenzaprine. We also conducted 1:1 PS matching for six pairwise comparisons with cyclobenzaprine as the reference group. Target estimands, statistical efficiency, and HRs with 95% CIs were compared across the four weighting methods and PS matching.

Results: The study included 207,364 eligible patients with treatment group sizes ranging from 3,110 initiating chlorzoxazone/orphenadrine to 77,165 initiating cyclobenzaprine before adjustment. All methods balanced measured covariates with no standardized differences > 0.10. MW and OW standardized all treatment groups to the smallest group (chlorzoxazone/orphenadrine), IPTW standardized all treatment groups to the largest group (cyclobenzaprine), and ATTW and pairwise 1:1 PS-matched cohorts most closely resembled the chosen reference group (cyclobenzaprine) with respect to measured covariates. All approaches suggested an increased risk of opioid overdose for baclofen relative to cyclobenzaprine, with HRs and 95% CIs spanning 2.70 (95% CI 2.06, 3.55) for OW to 3.29 (95% CI 2.44, 4.44) for 1:1 PS matching, and no differences in risk for the other five treatments.

Conclusions: In this head-to-head empirical comparison of seven treatments, the four evaluated PS-based weighting methods yielded similar HRs and 95% CI coverage. Nevertheless, the generalizability of our findings may be limited to situations where the distribution of treatment effect modifiers, if they exist, does not differ across treatment groups.