Differences in target estimands between different propensity score-based weights.

Abstract

Propensity score weighting is a popular approach for estimating treatment effects using observational data. Different sets of propensity score-based weights have been proposed, including inverse probability of treatment weights whose target estimand is the average treatment effect, weights whose target estimand is the average treatment effect in the treated (ATT), and, more recently, matching weights, overlap weights, and entropy weights. These latter three sets of weights focus on estimating the effect of treatment in those subjects for whom there is clinical equipoise. We conducted a series of simulations to explore differences in the value of the target estimands for these five sets of weights when the difference in means is the measure of treatment effect. We considered 648 scenarios defined by different values of the prevalence of treatment, the c-statistic of the propensity score model, the correlation between the linear predictors for treatment selection and the outcome, and by the magnitude of the interaction between treatment status and the linear predictor for the outcome in the absence of treatment. We found that, when the prevalence of treatment was low or high and the c-statistic of the propensity score model was moderate to high, that matching weights, overlap weights, and entropy weights had target estimands that differed meaningfully from the target estimand of the ATE weights. Researchers using matching weights, overlap weights, and entropy weights should not assume that the estimated treatment effect is comparable to the ATE.