Seminar by Xianwei Li

Variable Selection and Prediction for Multistate Models under Intermittent Observation

Psoriatic Arthritis (PsA) is a chronic condition that can cause joint damage. It is known that HLA antigens play an important role in joint destruction and disability in PsA. The University of Toronto Psoriatic Arthritis Clinic maintains a registry of around 1600 PsA patients who are scheduled to make clinic visits at 6-month intervals, but the actual visit times vary considerably over time and across individuals. We aim to use the resulting data on joint damage, along with baseline serum samples, to study a large group of HLA markers and predict disease progression at a particular time horizon. To do so we formulate a multistate model of joint damage and formulate a penalized likelihood for variable selection in the setting where the processes are under intermittent observation. An innovative expectation-maximization (EM) algorithm is then developed which can flexibly incorporate different penalty functions for variable selection and allows one to exploit existing packages for penalized regression. An application of the proposed algorithm to the motivating registry involves selecting HLA markers in sacroiliac joint damage in PsA. We discuss how to extend this model to incorporate a death state, where deaths are subject to right-censoring. We also use the sacroiliac joint damage data to illustrate how to assess predictive performance with validation data.