Prediction and inference under competing risks in high dimension

Venue: Room 200-9, Run Run Shaw Business Administrative Building, Yuquan Campus

Speaker:

Ronghui Xu, Dr./ Professor

Professor of Statistics and Biostatistics

Department of Family Medicine and Public Health and Department of Mathematics

University of California, San Diego

Abstract:

Prediction and Inference under Competing Risks in High Dimension - An EHR Demonstration Project for Prostate Cancer

The work was motivated by the analysis project using the linked SEER-Medicare database to predict cancer versus non-cancer mortality in men of age 65 years or older who were diagnosed with prostate cancer. We consider existing R package implementations that are computationally feasible for such data sets with up to 100,000 human subjects and over 20,000 claim codes. We carried out simulation studies to compare lasso for cause-specific hazards regression and boosting for Fine-Gray sub-distribution regression models with different approaches to choosing the penalty parameters, with the goal of prediction accuracy of the cumulative incidence rates at 2 and 5 years from baseline. Separately, we also developed methods for constructing confidence intervals of the regression effects of predictors (clinical variables and claim codes) under the Fine-Gray model in high dimensions, using the one-step estimator along the line of van de Geer et al. (2014) and Zhang and Zhang (2014). The extension though, is non-trivial under competing risks and Cox type model formulation. The results of the SEER-Medicare database analysis will be presented.