Classification Accuracy of Logistic Regression Models Under Misspecifications

• Activity Number: 26
• Type: Contributed
• Date/Time: Sunday, August 3, 2014 : 2:00 PM to 3:50 PM
• Sponsor: IMS
• Author(s): Yongzhao Shao*+ and Enhan Zhang
• Companies: New York University-School of Medicine and New York University School of Medicine
• Keywords: area under the curve ; ROC ; classifier ; logistic regression ; model misspecification ; risk prediction models

Abstract

Numerous biomarkers have emerged due to recent advancement in genomics and high-throughput biotechnology. A problem of interest is to decide whether a set of newly-emerged biomarkers have added values in predicting disease risk or treatment responses. One approach is to compare two nested logistic regression models (with and without the new biomarkers). The commonly used comparison of the nested prediction models include testing significance of the new biomarkers (e.g. via Wald test) or testing significance of the improved classification performance (e.g. via DeLong's test for the incremental AUC under the ROC). The test statistics of those two approaches (i.e. Wald's versus DeLong's tests) have equivalent null hypotheses when regression models are specified correctly. However, in practice, model misspecifications can adversely affect the performance of resulting classifiers. This paper investigates the impact on the discriminatory/predictive accuracy of logistic predictive models from two common types of misspecifications: (1) the log odds ratios of the predictors are estimated with biases and (2) the binary outcomes are misclassified in absence of an affordable gold standard.