A pattern-mixture model for censored binary longitudinal data

A pattern-mixture model is proposed to deal with binary longitudinal data subject to right censoring. We extend the mixture model of Hogan and Laird, which jointly models interval-scaled longitudinal data subject to censoring, to the censored repeated binary responses. With an assumption that the repeated measures are correlated only through the previous measurement, we adapted a transitional logistic regression model fitted to the distribution of repeated binary outcomes combined with a nonparametric form for the nonresponse process. The estimates are obtained by the method of the Expectation-Maximization (EM) algorithm with weights combined with the Quasi-Newton method to maximize the complete data likelihood in the maximization steps. The appropriate test statistics are described. An example is given to illustrate the proposed model. Sensitivity analysis is performed to address the adequacy and validity of this approach.