Correcting for confounding in analyzing receiver operating characteristic curves

A method is described for modeling a receiver operating curve as a function of confounding covariates when the outcome of the screening test is a continuous variate. A parametric survival model is proposed for modeling the distribution of the screening test outcome as a function of true disease status and other confounding covariates. The sensitivity and specificity of the screening test at any ''cut-point'' along the range of the screening test outcome may be estimated easily from the estimated survival distribution. Confidence intervals and an estimate of the area under the curve are derived.