A note on prognostic accuracy evaluation of regression models applied to longitudinal autocorrelated binary data

Background: Focus of this work was on evaluating the prognostic accuracy of two approaches for modelling binary longitudinal outcomes, a Generalized Estimating Equation (GEE) and a likelihood based method, Marginalized Transition Model (MTM), in which a transition model is combined with a marginal generalized linear model describing the average response as a function of measured predictors. Methods: A retrospective study on cardiovascular patients and a prospective study on sciatic pain were used to evaluate discrimination by computing the Area Under the Receiver-Operating- Characteristics curve, (AUC ), the Integrated Discrimination Improvement (IDI) and the Net Reclassification Improvement (NRI) at different time occasions. Calibration was also evaluated. A simulation study was run in order to compare model’s performance in a context of a perfect knowledge of the data generating mechanism. Results: Similar regression coefficients estimates and comparable calibration were obtained; an higher discrimination level for MTM was observed. No significant differences in calibration and MSE (Mean Square Error) emerged in the simulation study; MTM higher discrimination level was confirmed. ConclusionS: The choice of the regression approach should depend on the scientific question being addressed: whether the overall population-average and calibration are the objectives of interest, or the subject-specific patterns and discrimination. Moreover, some recently proposed discrimination indices are useful in evaluating predictive accuracy also in a context of longitudinal studies.