Faculty of Actuarial Science & Insurance Seminar with Cecile Proust-Lima , University of Bordeaux
2
Registered
Registration
Registration is now closed (this event already took place).
Details
Abstract: After the diagnosis of a disease, one central objective is to predict cumulative probabilities of events (e.g. clinical relapse) from the individual information collected up to a prediction time, usually including biomarker repeated measurements. Even before a diagnosis, cumulative probability of disease can be computed from the individual screening history or exposure records. Such predictions based on information repeatedly collected over time can be (dynamically) updated as soon as new information becomes available. In this presentation, I will give a short overview about how dynamic predictions can be defined, and what are the difficulties with their computation and the evaluation of their predictive accuracy. For their computation, two main approaches have been proposed: the joint modelling approach which simultaneously models the longitudinal and the time-to-event processes, and the landmarking approach which directly focuses on the time to predict by conditioning on the repeated information collected up to the given landmark time. I will compare the two approaches notably in terms of predictive accuracy, efficiency and robustness to model assumptions using a simulation study. The presentation will be illustrated with the prediction of competing causes of prostate cancer progression from the history of prostate-specific antigen (PSA), the main biomarker in Prostate Cancer.
Where
Bayes Business School, 106 Bunhill Row
Room 2005
106 Bunhill Row, London EC1Y 8TZ, UK