Verification Bias in Assessment of Mammography Accuracy

• Activity Number: 22
• Type: Contributed
• Date/Time: Sunday, August 11, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section*
• Author(s): Yingye Zheng and William Barlow*+
• Affiliation(s): University of Washington and Group Health Cooperative
• Keywords: Bias correction ; Cancer screening ; ROC Analysis ; Weighted estimating equations

Abstract

Accuracy of mammography is typically assessed by ROC analysis. An ordinal rating of likelihood of cancer is assigned by the radiologist that is compared to cancer status within one year of the mammogram. Verification bias can occur if a gold standard determination of cancer status is unavailable and the screening test result influences ascertainment of the outcome. Gray, Begg, and Greene (1984) and Rodenberg and Zhou (2000) provide approaches that allow for the correction due to biased ascertainment. We consider an alternative approach using weighted estimating equations (WEE) (Lipsitz, Ibrahim, Zhao, 1999). An ordinal regression model is applied to obtain parametric ROC curves, but the outcome is weighted to reflect uncertainty in verification. The analysis incorporates two additional regression models, one for being verified (given covariates and the mammographic outcome) and one for disease (given covariates). Only one of these two regression models needs to be correctly specified for the ROC analysis to give consistent estimates. The technique is compared to other techniques including the EM approach and an inverse probability weighting process.