Abstract

Recent work by Reiss and Ogden (JRSSB, 2009) gives some grounds for preferring REML to GCV as the criteria for smoothness estimation in GAMMs. This talk briefly discusses the relative merits of prediction error (e.g. GCV) and likelihood (e.g. REML) based smoothness selection. It is then shown that REML is more computationally challenging than GCV in a GAMM setting, but an approach to REML optimization is presented which allows REML based estimation of GAMMs to be as robust and efficient as GCV based estimation. Practical illustrations are provided.