Latent Propensity Score for Average Causal Effect Estimation Allowing Covariate Measurement Error

• Activity Number: 595
• Type: Contributed
• Date/Time: Wednesday, August 7, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Yi Huang*+ and Karen Bandeen-Roche and Xiaoyu Dong and Andrew Raim and Cunlin Wang
• Companies: University of Maryland Baltimore County and Johns Hopkins Bloomberg School of Public Health and FDA and University of Maryland, Baltimore County and OTS/CDER/FDA
• Keywords: Causal Inference ; propensity score ; covariate measurement error ; finite mixture model ; nondifferential error ; Average causal effect

Abstract

The covariates are often measured with error in biomedical and policy studies, which is a violation of the strong ignorability assumption. The naive approach is to ignore the error and use the observed covariates in current propensity score framework for average causal effect (ACE) estimation. However, after extending the standard causal assumptions behind propensity score approaches allowing covariate measurement errors, we showed that the naive approach typically produces biased ACE inference. In this talk, we developed a finite mixture model framework under continuous outcomes for ACE estimation, where the joint likelihood captures the uncertainty in propensity score subclassification arising from the unobserved measurement error. Simulations studies are presented to show the performance of this newly developed approach vs. naïve approach. This approach is used to evaluate the health impact of regular breast pump usage on infant's health using Infant Feeding Practice Study II data. This is a joint work with Xiaoyu Dong, Andrew Raim, Karen Bandeen-Roche, and Cunlin Wang.