Model-based classification of clustered binary data with non-ignorable missing values

Abstract

A hierarchical logistic regression model with nested, discrete random effects is proposed for the unsupervised classification of clustered binary data with non-ignorable missing values. An E-M algorithm is proposed that essentially reduces to the iterative estimation of a set of weighted logistic regressions from two augmented datasets, alternated with weights updating. The proposed approach is exploited on a sample of Chinese older adults, to cluster subjects according to their cognitive impairment and ability to cope with a Mini-Mental State Examination questionnaire.