Title: Joint modelling approaches to survival analysis via likelihood-based boosting techniques
Authors: Colin Griesbach - FAU Erlangen-Nuernberg (Germany) [presenting]
Elisabeth Waldmann - Friedrich-Alexander-Universitaet Erlangen-Nuernberg (Germany)
Andreas Groll - Technical University Dortmund (Germany)
Abstract: When analyzing data where event-times are recorded alongside a longitudinal outcome, one commonly used approach in practice is separate modeling of the two outcomes without considering any interaction effects. Especially in survival analysis one main interest is incorporating time-varying covariates into the model. This however is quite a challenge, since popular methods like the extended cox regression produce biased results. Joint modeling on the other hand combines a longitudinal and a survival submodel in one single joint likelihood and thus accounts for interactions like time-varying covariates measured with error, which can be often found in follow-up studies. Previous works proposed algorithms to fit joint models via component-wise gradient boosting techniques which focus on minimizing the predictive risk, offer advantages like variable selection and also work with high dimensional data. However, gradient boosting leads to problems in the survival part of the model, since time-varying effects can not be estimated so easily. Likelihood-based boosting approaches on the other hand are, as verified in various literature, capable of handling time-dependent covariates in survival analysis, since likelihood-based boosting directly optimizes the likelihood by using newton algorithms with a component-wise updating procedure.