A Comparison of Predictive Marginals Estimated from Logistic Regression Models and Log-Linear Regression Models with Categorical Data

• Activity Number: 314
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Chaoyang Li*+ and Earl S. Ford and Catherine A. Okoro and Tara W. Strine and Jin-Mann S. Lin and Lina S. Balluz
• Companies: Centers for Disease Control and Prevention and Centers for Disease Control and Prevention and Centers for Disease Control and Prevention and Centers for Disease Control and Prevention and Centers for Disease Control and Prevention and Centers for Disease Control and Prevention
• Keywords: predictive marginals ; prevalence ; complex survey data ; standardization ; diabetes ; obesity

Abstract

Predictive marginals have been used as a useful tool to compare adjusted prevalence estimates between subgroups controlling for differences in the distribution of covariates in survey data analyses. Little is known about whether different regression models may yield similar predictive marginal estimates. We analyzed data from the 2008 Behavioral Risk Factor Surveillance System to empirically compare the predictive marginal estimates of obesity, diabetes, and myocardial infarction among non-Hispanic whites, non-Hispanic blacks, Hispanics, and adults with other race/ethnicity using logistic regression and log-linear regression analyses. Results showed that with adjustment for age and sex, log-linear regression models yielded predictive marginal estimates similar to the age- and sex-adjusted prevalence estimates obtained from the direct standardization, whereas logistic regression models yielded inflated predictive marginal estimates (relative difference ranged from 29.7% to 75.2%). Log-linear regression models appear to perform better than logistic regression models for estimating predictive marginals, particularly when the population prevalence is low and/or sample size is small.