Dissertation Defense: Concordance Measures for Variable Screening and Model Evaluation with Competing Risks Data

We focus on analysis of time-to-event data with competing risks. In the first project, we make additional assumption of natural ordered event status, and propose a time-dependent model-free variable screening method for high-dimensional data that evaluate the discrimination ability of a biomarker to distinguish multiple event status simultaneously. The proposed method utilizes the Volume under the ROC surface (VUS), which measures the concordance between values of biomarkers and event status at certain time points. We show that the VUS possesses the sure screening property, i.e., true important covariates can be retained with probability tending to one. Simulations and data analysis show that VUS appears to be a viable screening metric and is robust to data contamination.

In the second project, we provide a systematic examination of model evaluation metrics that evaluate the discrimination ability of prognostic models. Most of the existing metrics focus on how a particular cause of event can be discriminated from the healthy control by the prognostic models when competing events exist, and one metric, the polytomous discrimination index (PDI), additionally provides an overall evaluation of diagnostic accuracy of a group of models for predicting all competing events. A systematic comparison of PDI with other existing methods is missing. We thus fill this gap and illustrate the performance of different model evaluation metrics under various scenarios via simulation studies and data analysis. Two natural extensions of concordance index are also considered, and we discuss their ability of model evaluation.

Advisor: Dr. Yu Cheng