Logistic Regression Classifiers with Longitudinal Data

• Activity Number: 428
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Daniel Jeske*+ and Jun Li and Xin Zhang and Vance Wong and Brian Noland
• Companies: University of California and Unviversity of California and University of California and Alere Corporation and Alere Corporation
• Keywords: Mixed Models ; Functional Data Analysis

Abstract

We propose a robust two-group classification methodology for longitudinal data applications. A nonparametric mixed effects (NME) model is fit to training data. Nonparametric basis functions are used to capture the underlying group-dependent time trends and subject-specific random intercepts and slopes are used to tune the model for subject heterogeneity. The NME is used to obtain empirical best linear unbiased predictions (eBLUPs) for the slopes and intercept of each subject. A logistic regression model whose explanatory variables are eBLUPs is fit and used to evaluate the posterior membership probability for, say Group 1. A classification threshold for the posterior membership probability is determined from a receiving operating curve (ROC) analysis. New subjects are classified by evaluating their eBLUPs, inserting them into the logistic regression model, and comparing the posterior membership probability for Group 1 to the selected threshold. We illustrate the use of this methodology and then compare and contrast it's effectiveness to an alternative approach where the measured responses at each time point are used directly in a logistic regression classifier.