Faculty Scholarly Dissemination Grants

Title

Logistic Regression Versus Discriminant Analysis: Empirical Simulations Investigating Prediction Performance in Biometrical Data

Department

Statistics Department

College

College of Liberal Arts and Sciences

Disciplines

Physical Sciences and Mathematics

Abstract

Historically, logistic regression has been the standard for binary response classification in biometrical settings. Increased emphasis on decision making has given rise to competing techniques and discriminant analysis is a logical alternative in such contexts. Recommendations differentiating the two techniques have focused on the discriminant analysis assumptions but findings regarding prediction performance and these rules-of-thumb have been inconsistent. Real data of various characteristics were investigated for empirical simulations. The data sets presented differ in size, the nature of predictor variables and the underlying assumptions. Repeated subsampling utilized two training-to-test set ratios and the impact of outliers was also investigated. Logistic regression performed comparably to discriminant analysis in situations where inferior performance was expected. There was a performance drop by discriminant analysis in the presence of categorical predictors. If one's prediction application involves continuous predictors but goals also include odds ratios and inference, one may want to reconsider logistic regression as it may still give strong prediction performance

Conference Name

Joint Statistical Meetings

Conference Location

San Diego, California

This document is currently not available here.

Share

COinS