Modeling Log-linear and Logit Models in Categorical Data Analysis
The association between categorical variables is analyzed using the mutual information approach complied with the multivariate multinomial distributions. Schematic decompositions of mutual information are employed for characterizing log-linear and logit models. A geometric analysis of the conditional mutual information is proposed for selecting indispensable predictors and their interaction effects for constructing log-linear and logit models. The new approach to selecting the most concise logit model also facilitates search for the minimum AIC model with a finite set of predictors. The proposed constructive schemes are illustrated in analyzing a contingency table of data collected in a study on the risk factors of ischemic cerebral stroke.
READ FULL TEXT