On November 18th, Philip Cooney presented recent work ‘Hazard Changepoint Analysis with Collapsing Changepoint Models’. An abstract for the talk is below:
There are a wide variety of applications for statistical models which assess how the parameters underlying a data generating process may change over time. One function which is subject to change is the hazard rate in survival analysis. The hazard rate quantifies the instantaneous failure rate of a subject which has not failed at a given time point. Because the survival probabilities are directly related to sum of the hazard function, changes in this function over time are of interest in a variety of situations. Matthews & Farewell (1982) suggest a real world application whereby physicians are interested in determining whether the hazard of relapse in leukemia is constant or time varying. Both Frequentist and Bayesian methods exist for changepoint analysis of hazard functions. Bayesian methods have the advantage in that they can account for uncertainty in the location of the changepoint(s), however, most previous methods focus is on estimating the location of a known number of changepoints. In this presentation we discuss a technique previously used in other types of changepoint analysis Wyse & Friel (2010) to the estimate an unknown number of changepoints in a hazard function. We present results from a simulation study and an application to real world data. Our approach is attractive because we can estimate both uncertainty in the changepoint locations for a given changepoint model and obtain a probability interpretation for the number of changepoints.
Matthews, D. E. & Farewell, V. T. (1982), ‘On Testing for a Constant Hazard against a ChangePoint Alternative’, Biometrics 38(2), 463–468. Wyse, J. & Friel, N. (2010), ‘Simulation-based Bayesian analysis for multiple changepoints’.
Slides for Philip’s presentation are available here.