Multinomial logistic regression and prediction accuracy for interval-censored competing risks data

Interval-censored competing risks data are ubiquitous in biomedical research fields. The direct parametric modeling of the cumulative incidence functional (CIF) is appealing due to its intuitive probability interpretation and easy implementation. This dissertation is to study and extend the multinomial logistic regression (MLR) model to interval-censored competing risks data. The MLR model naturally guarantees the additivity property of the event-specific probabilities under competing risks. A cubic B-Spline-based sieve method is then adopted to add flexibility into the proposed MLR model. The second study objective is to develop the prediction error (PE) as a model-free metric to evaluate and validate the prediction accuracy for interval-censored competing risks data. Adopting the method of the pseudo-value estimator, this dissertation work proposes a novel approach to estimate the PE under the interval-censored competing risks setting. Simulation studies are presented to assess performance of the MLR model and the PE in different scenarios. The proposed methods were then applied to a community-based study of cognitive impairment in aging population. Public Health Significance: Interval-censored competing risks data could be often encountered in biomedical research that is essential for public health, such as rehabilitation and pain medicine. The proposed methods provide precise yet flexible modeling of such data with straightforward interpretation on how predictors affect the CIF, as well as useful tools to evaluate and validate the prediction accuracy of the developed models