2 p.m. - 3 p.m. Location: JO 4.614
Philip E. Cheng
Institute of Statistical Science
A Geometric Approach to Modeling Categorical Variables With Applications to Medical Research
Information identities defined with multivariate multinomial distributions are used to examine the association between a finite set of discrete variables, in particular, categorical variables in a contingency table. For a response variable and its covariates, mutual information is employed to identify indispensable predictors as main effects of a logit model, and information identities are decomposed to identify the interaction effects of the predictors. The geometric information analysis is proposed to yield the most parsimonious logit models and log-linear models. It also serves as a convenient vehicle to finding the AIC models. The proposed method is examined with a contingency data table from a medical study of the ischemic cerebral stroke with risk factors.
Sponsored by the Department of Mathematical Sciences