Prediction of Disease Status: Transition Model Approach for Repeated Measures

This paper develops models for prediction of disease status from longitudinal data. The estimation of area under curve (AUC) is illustrated on the basis of estimates of sensitivity and specificity for repeated binary outcomes of disease status. There are several research papers in this field on cross-sectional data but only a few dealt with the repeated observations. This paper shows the procedures to deal with repeated observations employing Markov models. These procedures employ covariate dependent Markov models for estimating sensitivity and specificity, which in turn, produce the estimates for area under curve. The tests for equality of areas under curve for two models are also suggested. An application is illustrated for depression data from the Health and Retirement Survey, USA. The results indicate that the transition model approach can reveal the matching of disease status very efficiently; an estimate of more than 0.96 was obtained for the AUC for a transition model based prediction of disease from the depression data.