Additive smooth models in general

Simon Wood, University of Bath

Regression models for independent exponential family responses built using random effects and penalized reduced rank spline smoothers are popular, with the link between smoothers and random effects providing a reliable computational and inferential framework for their practical use. This talk will discuss a framework for more general models built in terms of smooth functions. Examples includes Tweedie (with unknown power parameter), negative binomial (with unknown 'theta'), beta, scaled t and ordered categorical additive smooth regression, as well as additive Cox proportional hazard models, GAMLSS models such as zero inflated Poisson and Gaussian location scale models, and multivariate Gaussian additive models. Methods for smoothing parameter selection, and model selection will be covered along with example applications.