Matthew Goldstein

A stepwise discrete variable selection procedure

A stepwise variable selection procedure for multinomial discrimination is presented and discussed. Based upon the work of Kullback and Hills, stopping rules are proposed and illustrated for a set of data on communication buyer behavior.

On a test of independence for contingency tables

Using the concept of distributional distance, a test statistic is proposed FOR the hypothesis of independence in multidimensional contingency tables. A Monte Carlo Study is done to empirically compare the power of the proposed test to the Pearson x2 and the likelihood ratio test- Further, the nonnull distribution under various spike alternatives is tabulated

Selection of Variates for the Two-Group Multinomial Classification Problem

Abstract A sample-based procedure for selecting an optimal subset of variables for the two-group multinomial classification problem is presented. Based on the work of Glick, we examine two variations of a procedure which looks at differences in discriminant scores as a method for assessing worthiness of various subsets of variables. Two examples illustrate the methods.

Multivariate Analysis Methods and Applications.

Selected Aspects of Multivariate Analysis. Principal Components Analysis. Factor Analysis. Multidimensional Scaling. Cluster Analysis. Multiple Regression. Some Practical Considerations: Data Analysis Problems. Cross-Classified Frequency Data. Canonical Correlation Analysis. Discriminant Analysis: The Two-Group Problem. Multiple Discriminant Analysis and Related Topics. Linear Structural Relations (LISREL). Latent Structure Analysis. Appendixes. References. Index.

A Measure Of Separability And Random Zeros In Statistical Classification

The setting for this study is the two-group multinomial classification problem. Based on a measure of the log odds in favor of one particular group, a large sample confidence interval for a measure of separability is derived. The asymptotic result employed assumes that all states have positive observed frequencies. Realizing that this assumption is often violated, we consider a method based upon log-linear representation of state frequencies to first remove any random zeros before attempting to effect a classification. The method is illustrated in a data set dealing with the behavioral consequences following hypoxic trauma.