4G-03 - Comparison of Propensity Score Matching vs. Inverse Probability Treatment Weighting: A Simulation Study in the CorEvitas Psoriasis Registry

Background: In observational studies, covariate imbalance, if not addressed, could lead to confounding and reduced study validity. Various propensity score (PS) based methods are available to address covariate imbalance such as matching and inverse probability of treatment weighting (IPTW). These methods may have power, bias, and precision tradeoffs and thus be more advantageous in different settings.

Objectives: To compare PS matching and IPTW using simulated time-to-event data with varying covariate imbalance by evaluating 1) effective sample size (ESS), 2) standardized mean difference (SMD), 3) bias, and 4) precision.

Methods: Performance of PS 1:1 matching with a caliper of 0.2 and IPTW with stabilized weights was evaluated using simulated data informed by real-world data from the prospective, non-interventional multi-center CorEvitas Psoriasis Registry. Three population data sets were created to represent low, moderate, and high overall baseline covariate imbalance (mean SMDs of 0.18, 0.28, and 0.42, respectively) across 13 potential confounders (12 binary, 1 continuous) using a 1:1 treatment-to-control ratio. N=1000 samples were drawn from each population data set to calculate the ESS, mean SMD, bias, and precision of the estimated hazard ratio (HR) from an unadjusted Cox model.

Results: The ESS was highest using IPTW. The ESS preserved in the treated group by IPTW vs. PS matching was 85% vs. 70%, 74% vs. 60%, and 50% vs 41% for low, moderate, and high covariate imbalance, respectively. IPTW and PS matching yielded similar mean SMD when covariate imbalance was low to moderate, while PS matching outperformed IPTW, yielding a lower mean SMD, when covariate imbalance was high. Similarly, with low covariate imbalance, IPTW had the lowest HR bias, but as covariate imbalance increased, PS matching yielded a lower HR bias than IPTW. Lastly, precision of the HR estimate was consistently higher using IPTW, due to greater ESS, compared to PS matching.

Conclusions: Based on these simulations, with low covariate imbalance, IPTW is the better method to address covariate imbalance. However, as covariate imbalance increased, PS matching more completely adjusted for observed confounding. These simulations also showed that while the IPTW models utilize all observations unlike PS matching methods, due to weighting, the ESS for IPTW analyses is lower than the raw SS, especially when covariate imbalance is high. Additional simulations with unequal treatment to control population ratios are forthcoming to investigate if this pattern holds in other settings.