Comparison of Logistic Regression Model and CART Model: An Application in Studying Polytobacco Users in a Tri-Ethnic Sample

• Activity Number: 295
• Type: Contributed
• Date/Time: Tuesday, August 5, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Health Policy Statistics Section
• Author(s): Yang Lei*+ and Nikki Nollen and Qing Yu and Jianghua He and Matthew S. Mayo
• Companies: KU Medical Center and Kansas University Medical Center and Kansas University Medical Center and Kansas University Medical Center and Kansas University Medical Center
• Keywords: stratified samples ; model selection ; CART ; partial interaction

Abstract

Other forms of tobacco use are gaining in popularity because most tobacco control policies are aimed at cigarettes only. For policy development purposes, it is of the utmost importance to identify who are using multiple tobacco products. Multiple logistic regression models were mostly used in previous studies. We are proposing to conduct the classification and regression trees model (CART) in contrast with a predictive logistic regression model. Our data includes 2,376 participants from an online cross-sectional survey administered through Survey Sampling International. The sample is stratified to obtain equal samples of each of the three ethnicities: Whites, African Americans, and Latinos. A hold-out validation strategy is used to obtain independent training and validation samples for both the logistic regression model and the CART model. The logistic regression model based on Frank Harrell's model selection strategy and AIC criterion can achieve better model fit with the cost of increased complexity. The CART model is able to capture the most predictive factors and reveals partial interaction effects that identify target populations of policy significance.