Doctoral Dissertation Defense: Elande Baro
Advisors: Dr. Yi Huang and Dr. Anindya Roy
Wednesday, July 22, 2015 · 12 - 2 PM
Title: Bayesian Latent Propensity Score Approach for Average Causal Effect Estimation Allowing Covariate Measurement Error
Abstract:
Subclassification on the propensity score is a commonly-used analytic method for estimating the average causal effect of treatment on outcomes in observational studies. This method relies on nonconfoundedness assumption for valid inference on the average causal effect (ACE) estimation, i.e. no unobserved confounder and no covariate measurement error. However, many postmarketing studies rely on observational data where confounders (i.e. covariates) are subject to unobserved measurement error. It has been shown that using the naive propensity score methods ignoring such error will bias the ACE inference. Huang et al. extended the standard propensity score based causal framework to allow covariate measurement error using a latent propensity score, and developed a joint likelihood based approach for consistent ACE estimation using EM algorithm under finite mixture model of continuous outcomes. The numerical performance of the EM algorithm is not always ideal with increasing dimensions of unknown parameters. We extend this work to more flexible outcomes including binary/binomial outcomes using Bayesian MCMC algorithm. Theoretically, we derive the sufficient conditions for the identifiability of the proposed finite mixture models of binary/binomial outcomes. In terms of computation, a Bayesian estimation method is used to fit the latent propensity score model under finite mixture model format. Simulations studies often show superior performance of this newly developed Bayesian approach regarding bias, variance, and MSE criteria, compared to the existing EM algorithm and naıve approach (ignoring the error). We apply this method in post-market evaluation of the health impact of long-term breast pump usage on infant=E2=80=99s cumulative comorbidity index during first year, using the recently collected Infant Feeding Practice Study (II) data.
Abstract:
Subclassification on the propensity score is a commonly-used analytic method for estimating the average causal effect of treatment on outcomes in observational studies. This method relies on nonconfoundedness assumption for valid inference on the average causal effect (ACE) estimation, i.e. no unobserved confounder and no covariate measurement error. However, many postmarketing studies rely on observational data where confounders (i.e. covariates) are subject to unobserved measurement error. It has been shown that using the naive propensity score methods ignoring such error will bias the ACE inference. Huang et al. extended the standard propensity score based causal framework to allow covariate measurement error using a latent propensity score, and developed a joint likelihood based approach for consistent ACE estimation using EM algorithm under finite mixture model of continuous outcomes. The numerical performance of the EM algorithm is not always ideal with increasing dimensions of unknown parameters. We extend this work to more flexible outcomes including binary/binomial outcomes using Bayesian MCMC algorithm. Theoretically, we derive the sufficient conditions for the identifiability of the proposed finite mixture models of binary/binomial outcomes. In terms of computation, a Bayesian estimation method is used to fit the latent propensity score model under finite mixture model format. Simulations studies often show superior performance of this newly developed Bayesian approach regarding bias, variance, and MSE criteria, compared to the existing EM algorithm and naıve approach (ignoring the error). We apply this method in post-market evaluation of the health impact of long-term breast pump usage on infant=E2=80=99s cumulative comorbidity index during first year, using the recently collected Infant Feeding Practice Study (II) data.