Academic

The topic of analyzing time-to-event data where individual units are subject to multiple causes for the event occurrence has been well-studied for decades. A case that has received lion’s share of attention in this context is when the event results from the earliest onset of a cause, a framework traditionally dubbed as that of competing risks. Examples are replete in manufacturing applications, software development, clinical trials, social sciences and risk analysis. Research objectives in such instances range from understanding a component’s contribution towards the overall system failure to assessing the role of any prognostic factor on the component failure history in the presence of the competing causes. Earlier work in competing risks analysis utilized a series system (observing the minimum of several lifetimes) formulation in terms of latent event times. It is well known that such a formulation is fraught with the issue of identifiability, unless one can assume the different causes to act independently. Alternative formulation through cause-specific quantities attempt to formulate a model that has direct links to the observables and avoids imposing a dependence structure on the causes. This is an expository talk on the nuances of analyzing competing risks data. We shall cover both single and recurrent event data in this context. In keeping with the wide applicability of the framework, the methodology will be illustrated using a warranty claim database for a fleet of automobiles as well as registry data on cause of death of patients diagnosed with breast cancer. In some situations, the cause of system failure is not known exactly, but can be narrowed down to a subset of potential causes. This phenomenon, referred to as masking, is often the result of incomplete or partial information on the failures arising in destructive experiments. In my talk, I shall integrate the framework of masking and indicate how it can be handled easily using a Bayesian approach. The final topic of my talk is semi-competing risks which investigates recurrent events under a dependent termination framework.