Abstract

In this talk, I will discuss the concepts of near-redundancy, practical non-identifiability, and weak-identifiability through two examples in survival analysis. First, I will present a general hazard structure that encompasses the classical proportional hazards and accelerated failure time models. This model structure helps identify the reasons behind the aforementioned inferential issues in some data sets. In the second example, I will present a recent methodology for analysing survival data based on modelling the hazard function using systems of ODEs. Here, parameter interpretation sheds light on why weak identifiability may occur in certain data sets. I will also discuss other models beyond survival analysis where finite-sample inferential issues may arise. Overall, I will aim to convey that when inferential issues arise, statistical models and data must be evaluated or judged together. Furthermore, practical identifiability problems in a specific data set should not be viewed as a fatal flaw of the model.