A Monte Carlo Simulation on the Propensity Score Weighting with Principal Component Scores
Propensity score weighting is increasingly used to help causation in non-experimental, observational studies. Nevertheless, the main challenge is to estimate the propensity score with an optimal set of covariates because oftentimes there are too many available variables. Traditionally, it is recommended to use all available variables to estimate the propensity score. In practice, the set of possible covariates is often too large and sometimes correlated with each other, resulting in separation or misspecification issues in estimating PS. This study proposed to integrate the propensity score weighting with the principal component analysis (PCA) to transfer the large number of covariables into a manageable number of principal components so as to improve the performance of the propensity score weighting in terms of a) separation rate and b) proportion reduction in selection bias. Through Monte Carlo simulation, this study evaluated the performance of two primary propensity score weighting methods in combination with the traditional all-inclusive strategy and three principal component strategies under various conditions. The results evidenced that the PCA strategies have a great potential for reducing selection bias due to non-randomization in observational research and eliminating the risk of separation issues.
Embargo status: Restricted to TTU community only. To view, login with your eRaider (top right). Others may request the author grant access exception by clicking on the PDF link to the left.