O CEAUL - Centro de Estatística e Aplicações da Universidade de Lisboa promove o seminário "Prediction of Long-Term Effects in Susceptible Population", por Eduardo Janotti Cavalcante (São Paulo Research Foundation (FAPESP), Brazil / CEAUL, Faculdade de Ciências, Universidade de Lisboa).
Abstract: Survival modeling often requires extrapolation beyond observed follow-up periods, making long-term dynamics difficult to characterize reliably. Existing approaches typically rely on restrictive assumptions, such as cure-rate formulations or constrained tail specifications, which can limit extrapolation performance. We propose a Bayesian semiparametric survival model that combines Extreme Value Theory (EVT) with latent Gaussian processes, enabling the joint modeling of central and extreme hazard behavior without threshold selection. The framework accommodates a flexible, covariate-dependent extreme value index while avoiding strong parametric assumptions. Using survival data from InCor, we show that the proposed approach improves the representation of long-term hazard dynamics while maintaining coherence across the hazard function. The resulting methodology provides a flexible and practical tool for settings where accurate tail characterization is essential.