Application of information theory in generalized additive models
by
Hong Gu
Dept. of Math. & Stat., Dalhousie Univ., Halifax, NS, Canada
Coauthors: Mu Zhu, Dept. of Statistics and Actuarial Science, University of Waterloo

The concept of mutual information (MI) provides a good measure for the

strength of dependence between variables. MI for two variables can be

deemed as a generalized nonlinear version of the widely used correlation

coefficient in linear space. We first utilize the concept of mutual information (MI)

to recast the smoothing procedures of generalized additive models (GAM) into

a procedure of maximizing MI criterion. We further develop a new procedure called

Partial Generalized Additive Models (PGAM) which fits GAM on a set of conceptually

independent nonlinearly transformed predictors. PGAM can avoid the concurvity issues

of GAM and improve the interpretations of the model.

PDF

Date received: March 11, 2007