Propensity Score Modeling in the Presence of Measurement Non-Invariance

• In Event: Data Analysis and Statistics

Abstract

Propensity score analyses assumes no methodological issues with confounding covariates. Yet, assumed observed covariates may be indicators of latent variables which possibly contain measurement error through cross sectional non-invariance between treatment and control groups. This study conducted a Monte Carlo simulation to analyze several latent variable proxies when intercept measurement non-invariance or loading measurement non-invariance was not accounted prior to estimation of the propensity weight. The latent variable proxies included: individual indicators, sum scores, conclusive factor scores, and inclusive factor scores along with the latent variable. Simulation results support the need for strong invariance regardless of proxy, but loading non-invraince was unrelated to bias in the average treatment effect.

Authors

• Joseph E. Miller, University of Texas at Austin
  Presenting Author

• Tiffany A. Whittaker, University of Texas at Austin
  Non-Presenting Author