Title: Mixed-effects additive transformation models
Authors: Balint Tamasi - University of Zurich (Switzerland) [presenting]
Torsten Hothorn - University of Zurich (Switzerland)
Abstract: Statistical models that accommodate non-normal, potentially correlated data and allow for nonlinear predictor-outcome relationships are crucial in applied regression settings. Traditional approaches typically rely on the idea that conditional response distribution can be fully captured with a few parameters of a predefined distribution type. Picking the correct distribution is often difficult in practice, and misspecifications can lead to incorrect inference. Transformation models provide a general and flexible approach to modeling the whole conditional distribution that forgoes the a priori specification of the response distribution by estimating its shape from the data. Mixed-effects additive transformation models extend the transformation model framework with random effects and penalized additive terms. The resulting model class can be readily used for modeling complex, dependent data structures (e.g., grouped data, temporal or spatial heterogeneity) in the presence of non-linear effects. Fully parametric likelihood-based estimation and inference of the model are discussed, and a fast and efficient R implementation is presented. The motivating example is an ecological experiment on carrion decomposition times under various environmental settings, where the non-normal, interval-censored time-to-event response, potential nonlinear covariate effects and grouped data structure render traditional regression tools inapplicable.