Abstract
Background/Aims Propensity Score (PS) methods have been used increasingly in observational studies to reduce bias due to confounding. The performance of PS for estimating hazard ratios (HR) has not been systematically studied. Monte Carlo simulations were performed to evaluate four methods of utilizing PS: 1:1 matching, stratification, PS adjustment, and inverse probability weighting (IPW).
Methods Twelve variables (6 binary, 6 continuous), varying in their association with treatment and outcome, were generated. Five of each were related to both treatment and outcome, and the remaining were associated with outcome. Two scenarios were considered. Scenario #1: 20% of the sample was treated and variable distributions differed between groups. Scenario #2: treatment assignment was associated with the variables. One thousand data sets with 5000 samples each were generated. Time-to-event data were generated from an exponential distribution. Each method was evaluated by controlling treatment only, treatment+confounders, treatment+outcome-related variables, and all potential covariates.
Results Controlling treatment only, matching methods yielded the least biased estimate (~40%) in both scenarios if the condition HR >1.0 was true, while PS adjustment performed best if the condition HR <1.0 was true. In both scenarios, an unbiased estimate was attained if, and only if, all potential covariates were controlled, regardless of method.
Conclusions All PS methods resulted in a biased estimation of the true HR if the treatment estimate is not conditioned on all confounders and outcome-related covariates. The least biased estimation occurs when treatment and confounders are controlled in the PS covariate adjustment method and IPW method. Comparing the four methods, stratification on PS performs better than other methods, followed by matching. The bias was reduced by controlling for confounders if treatment and confounders were strongly associated, or by controlling for outcome-related covariates if outcome was strongly associated with covariates. In addition, unbiased estimates are attained without applying any PS methods as long as all potential covariates are controlled in the model.
