Abstract

Distributional generalized additive models (GAMs) extend classical GAMs by allowing multiple parameters of the response distribution to depend on covariates through additive predictors. While this increased flexibility facilitates richer probabilistic modelling, it also introduces substantial challenges for model interpretation, checking, selection and communication, particularly when multiple linear predictors are linked to distributional parameters via nonlinear multivariate parametrisations.

This talk focuses on the development of visualisation and diagnostic tools tailored to this broader class of models. In particular, we consider how effect visualisation, uncertainty quantification and diagnostic methods can be adapted to settings in which covariates simultaneously influence several distributional parameters, often through nonlinear parameterisations. We discuss approaches for disentangling and summarising covariate effects across parameters, highlighting how changes in smooth effects propagate to interpretable features of the response distribution. Overall, the proposed extensions aim to make complex distributional GAMs more interpretable and accessible, enabling practitioners to better understand, validate and communicate the insights they provide.