Do we need frailty modeling in the development of prognostic models?

Abstract

Prognostic models are developed to assist in clinical decision making particularly in regards to patient treatment and follow-up time. The Cox proportional hazards model is the most extensively used in analysis. As such almost all prognostic models in survival data are developed based on Cox models. However, overreliance on the the Cox model alone is common even in the presence of clustering ignoring the intrinsic heterogeneity and dependency present in such data. This project highlight the need for prognostic model building to account for clustering with a case study in Traumatic Brain Injury(TBI). Two approaches are used to account for clustering, the fixed effect and the frailty models. By comparing model parameter estimates and standard errors for these methods against the naive Cox, we demonstrate that lack of accounting for heterogeneity may lead to invalid inference as model parameter estimates have underestimated standard errors leading to inflated type I error. Comparison of