Dynamic prediction in event-history analysis
Research Objectives:
In many clinical studies, survival models are needed, that consider multiple events, which are either mutually exclusive (competing risks) or occur sequentially (multistate models). One important aim of such models is estimation of transition probabilities, probabilities of future evolutions of the multistate model, given the current state. The dynamic aspect is that these predictions need to be updated as more information in the event-history of the patient is revealed over time. The classical approach to this problem is based on first estimating the transition intensities. These estimated transition intensities then form the building blocks for the calculation of transition probabilities through formulas like the well-known Aalen-Johansen formula. Within the last decade, a number of alternative methods have been proposed for prediction, based on pseudo-values, direct binomial regression, vertical modelling and landmarking. The former two methods directly model the transition probability of interest, depending on covariates, where the time of prediction is held fixed.
Description of work:
Issues that will be studied in this project are:
1) How can these models be updated?
2) How optimal are these updated models?
3) How do time-dependent covariates and/or time-dependent covariate effects influence the models and their performance?
4) How can these models be used to address complex problems in prognostic and therapeutic studies?
Host Institution: Leiden University Medical Center