Estimating the Accuracy of Mammography from a Repeatedly Screened Population

• Activity Number: 103
• Type: Invited
• Date/Time: Monday, August 4, 2003 : 10:30 AM to 12:20 PM
• Sponsor: WNAR
• Author(s): Diana L. Miglioretti*+ and Patrick J. Heagerty
• Companies: GHC Center for Health Studies and University of Washington
• Keywords: marginal model ; multilevel model ; GEE ; Bayesian ; longitudinal data ; clustered data

Abstract

Estimating mammography accuracy in a repeatedly screened population poses multiple challenges. One challenge is accounting for correlation within both women and radiologists. Typically, relatively few observations are collected on many women and most outcomes are true negatives; thus, hierarchical models using the standard assumption of normally distributed random effects are difficult to fit and may not be reasonable. In addition, interest is often in the population-average effects, which are not directly modeled by hierarchical models. Analyses are further complicated by the fact that women are not necessarily nested within radiologists. To overcome these challenges, we propose two approaches. The first is a three-step generalized estimating equations approach that utilizes standard software. The second is a marginalized multilevel model that combines a logistic regression model to estimate the influence of covariates on the marginal mean and a separate conditional logistic regression model that captures both the serial dependence within women and the correlation within radiologists. The methods will be illustrated with data from the Breast Cancer Surveillance Consortium.