IMPROVING METHODS FOR MISSING DATA, CAUSAL INFERENCE & LATENT-TRANSITION ANALYSIS

点击翻译按钮获取中文摘要

Nonresponse, causal inference and latent variables are closely related; from a statistical viewpoint, all three can be regarded as missing-data problems. This project is a continuation of the efforts by Dr. Schafer and his colleagues since 1996 to enhance the science of drug abuse prevention and treatment through improved methods for handling missing data. It also continues the work previously led by Dr. Collins to model stage-sequential development of substance use and related phenomena through improved methods of latent-transition analysis (LTA). To these we have added another area: developing new tools for inference about causal effects from observational studies and broken randomized experiments. This project has six Specific Aims. First, we will work to improve practices used to analyze incomplete data in prevention and treatment through application, education and software development. Second, we will investigate the properties of the new "doubly robust" regression methods for missing values in longitudinal research. Third, we will work to develop robust strategies for imputing missing values thought to depart from the usual assumption of missing at random (MAR). Fourth, we will develop techniques for analyzing incomplete data when some of the missing values are thought to be MAR but others are not. Fifth, we will continue to develop, implement and apply Bayesian methods for statistical inference in LTA. Sixth, we will develop procedures for imputing counterfactual outcomes for causal inference in the presence of confounding.