Multinomial Logistic Regression and Prediction Accuracy for Interval-Censored Competing Risks Data

• Activity Number: 320 - Innovative Approaches for Modeling Time-to-Event Data in the Presence of Competing Risks and/or Time-Varying Covariates
• Type: Topic Contributed
• Date/Time: Tuesday, August 9, 2022 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Yu Cheng* and Yongli Shuai and Jong H Jeong
• Companies: University of Pittsburgh and University of Pittsburgh and University of Pittsburgh
• Keywords: Brier score; B-Spline; Cumulative incidence function; Parametric model; Prediction error

Abstract

Interval-censored competing risks data are ubiquitous in biomedical research. The direct parametric modeling of the cumulative incidence function (CIF) is appealing for its intuitive probability interpretation and easy implementation. Thus, this work first extends the multinomial logistic regression (MLR) model to interval-censored competing risks data. A B-Splines-based sieve method is utilized to estimate the baseline function of the log-odds of cause-specific events, which substantially improves the MLR model's flexibility. More importantly, we adopt a modified Brier score for competing risks outcomes to assess the proposed model’s prediction error (PE) and estimate the modified Brier score by replacing the unidentifiable event status arising from interval censoring with a Jackknife pseudo-value estimator. Simulation results are presented to confirm the desirable performance of the MLR model and its PE under different scenarios. We illustrate the proposed methods by analyzing the interval-censored competing risks data from a community-based study and show that the PE captures the improvement in modeling cognitive impairment in this aging population.