|
Abstract:
|
For prognostic models in medicine with multimodal or bathtub-shaped hazard function, 2 questions need to be answered: 1) do(es) the change-point(s) exist, 2) if so, a) does it reflect the existence of a mixture population, with different hazard functions for subgroups of subjects, or b) is it part of the natural course of the disease. For simple forms of the hazard plot, testing for a constant/linear functional form vs. an alternative of piecewise linear hazard can answer the first question. Next, survival regression models can test for patient subgroups with hazard functions of different forms, although these will not necessarily align directly with the change-points identified. We will illustrate a combination of statistical methods that test both hypotheses on a prognostic model of risk of suicide attempt for subjects in a major depressive episode.
|
