Title: First-hitting-time models for high-dimensional data: A statistical boosting approach
Authors: Riccardo De Bin - University of Oslo (Norway) [presenting]
Vegard Stikbakke - University of Oslo (Norway)
Abstract: In the recent years, increasing attention has been given to first-hitting-time models, at least in the context of survival analysis. In biomedical applications, the idea is to model the health status as a stochastic process, for example a Brownian motion or a Gamma process, that degrades until it reaches a critical level (threshold), which may represent the death of a patient or the recurrence of a disease. The parameters of these processes (e.g., location and scale parameters in a Brownian motion) can depend on covariates, as well as the threshold. We develop a boosting algorithm to extend the use of first-hitting-time models to high-dimensional contexts. In particular, we focus on the situation in which low-dimensional clinical data must be combined with high-dimensional genetic data to build a prediction model. We show that the integration of these two sources of data in a first-hitting-time model is intuitive and avoids complicated weighting procedures. Finally, the novel approach is applied to a real data example.