Combining ROC Curve Analysis and Generalized Logistic Regression to Estimate Continuous Cutpoints Discriminating Three Categories

• Activity Number: 183
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Rey DeCastro*+
• Companies: CDC/NCEH
• Keywords: NHANES ; cotinine ; tobacco exposure ; ROC curve ; generalized logistic regression ; sensitivity and specificity

Abstract

Standard ROC curve analysis enables estimation of a cutpoint or threshold for a continuous measurement in order to discriminate two discrete states. Recent years have seen progress in extending this technique to cases of more than two states. In this paper, a two-stage approach is explored where generalized logistic regression is first used to estimate predicted probabilities for each state associated with the continuous measurement. The second stage processes the predicted probabilities using ROC curve analysis to estimate two cutpoints discriminating the first state from the second and the second state from the third. Measurements of serum cotinine -- a biomarker of tobacco smoke exposure -- from the National Health and Nutrition Examination Survey (NHANES) 1999-2012 are used to discriminate three ordinal categories of tobacco smoke exposure among survey participants: unexposed, environmentally exposed, directly exposed (via first-hand smoking).