|
Abstract:
|
Propensity score methods (PSMs) have gained immense popularity in the last three decades. They are widely used to control for confounding, reduce bias, and estimates causal treatment effects. Of the four major PSMs used, the propensity score stratification (PSS) method often yields results that are unreliable and for which there are limited empirical and theoretical justifications.
For this presentation, we take a closer look at the weights used to aggregate stratum-specific treatment effects into the overall average treatment effect (ATE). The goal is to understand the current limitations and provide reliable solutions. To this end, we demonstrate that current PSS weights rely on underlying assumptions that are not always true and rarely verified in practice. This explains, in part, some the discrepancies with other PSMs. As an alternative, we introduce efficient weight estimates for the ATE, which have currently used PSS weights as a special case. We assess the performances of the proposed weight estimates through simulation studies, considering several data-generating scenarios. Finally, we illustrate the proposed method using the publicly-available Lalonde and Linder dataset
|
