The Impact of Covariate Measurement Error on Risk Prediction

• Activity Number: 42
• Type: Contributed
• Date/Time: Sunday, July 29, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Polyna Khudyakov*+ and Malka Gorfine and David M. Zucker and Donna Spiegelman
• Companies: Harvard School of Public Health and Technion and Hebrew University of Jerusalem and Harvard School of Public Health
• Keywords: Risk Prediction ; Measurement Error ; AUC ; Brier Score ; Binary Outcome

Abstract

In risk prediction models, key factors are often measured with error, and this may affect the quality of prediction. We studied the impact of covariate measurement error on risk prediction based on logistic regression models. We compared the performance of predictions from the model based on the costly true covariate (true model) with those based on an inexpensive surrogate covariate (surrogate model). The comparison was based on the area under the receiver operating characteristic curve (ROC-AUC), Brier score, and the ratio of the observed and expected number of events. We showed that the performance of the surrogate model depends in an important way not only on the amount of error in the mismeasured covariate, but also on the relationship between the error-prone and the error-free covariates. Specifically, we showed that as the correlation between the true and error-free covariates increases, the difference in the ROC-AUC between the true and the surrogate models decreases. This result suggests that in the presence of error-free covariates that are highly correlated with the error-prone covariate, the costly true covariate has a negligible advantage over the surrogate covariate.