Transition Model for Incomplete Longitudinal Binary Data

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Compared with longitudinal designs in other fields, at least three distinct features are observed with the designs in substance abuse treatment studies: (1) behavioral correlates of drug dependence result in missing values in the data matrix due to either nonresponse or dropout, (2) the maximum number of repeated measures is large, and (3) binary repeated measures, as opposed to continuous measures, are most often seen. This B/START proposal aims to identify optimal methods for studying the probability of developing strategies to conduct incomplete binary longitudinal data analysis. Determined by the above three features, transition models, based on Markov stochastic process, provide a more appropriate modeling strategy than other longitudinal modeling choices such as marginal models using quasi-likelihood functions and generalized linear mixed models. Computationally, transition models for binary repeated measures are easier to be fitted and applied after the data matrix has been reformed, since they are just logistic or Iogit regression models. Making use of the past responses in predicting the future ones usually produces analytical inferences that are more meaningful and interpretable. Large number of repeated measures on each experiment subject makes Markov process modeling more appealing. Using transitional models, we also have more choices to handle missing data. The proposed project will develop, compare, and evaluate two missing data strategies: multiple partial imputation (MPI), and multicategory-logit model (MLM). In MPI approach, intermittent missing data are imputed several times with missing data due to dropout left as they are, and then transition models will be fitted for each of these partially imputed data sets, and finally the multiple results are combined to make one final inference. In MLM approach, status of missingness is treated as a third category to extend the repeated measures into three-category ones.