We are very pleased to resume the RSSNI Seminar Series. On Wednesday, February 19th, at 1pm GMT, Jack Moore from the University of Limerick will talk about semi-parametric distributional regression survival models. This will be a hybrid event with Jack in-person in room
02/008 in the Peter Froggatt Centre, Queen's University Belfast (Map) and online with MS Teams (Link).
Abstract:
The field of survival analysis is concerned with modelling of time-to-event data, a form of data that arises in various application areas. One of the key applications areas of survival analysis is medical research, where interest lies in survival times of patients, e.g., time to onset or recurrence of a disease, or death due to a disease. Traditional parametric modelling approaches rely on distributions such as the Weibull or Gompertz. These impose strong assumptions on the data at hand, e.g., the shape of the hazard is restricted to be monotonic in the two aforementioned distributions. In contrast, the piecewise exponential model offers a much more general parametric modelling approach: it has the capability of approximating any survival distribution, without prior knowledge of the underlying distribution of the data. This is achieved by partitioning of the time scale into intervals, within which the hazard rate is constant. Despite the versatility of piecewise exponential model, it has historically been under-utilized within the literature. This can perhaps be explained by the popularity of the Cox model, which has the flexibility of a non-parametric baseline hazard but is also restrictive in its proportional hazards assumption. However, in recent years, there has been a resurgence of interest in the piecewise exponential model, with various developments aimed at enhancing its performance and utility.
In this talk, we introduce the piecewise exponential model and present extensions aimed at improving the viability of this model. Specifically, we make use of a distributional regression structure (generalising proportional hazards). Thus, our framework enables flexible modelling of both the underlying baseline hazard and the nature of covariate effects, where the intervals/ pieces and covariates of our model are selected automatically via an adaptive lasso penalisation. We demonstrate the flexibility of our proposal and provide comparisons with existing parametric and semi-parametric approaches.
Read previous write-ups
here.
See previous talks
here.
All welcome - the talk is open to the Public and we particularly wish to encourage an in-person audience
Gilbert MacKenzie (Chair) & Hannah Mitchell (Meetings Host)
Jack Moore from the Dept of Mathematics and Statistics the University of Limerick, Ireland.