Statistical Methods for the Evaluation of a Cancer Screening Program

Abstract

Estimation of the sensitivity and specificity of cancer screening tests using data from population-level databases is complicated by the lack of independent confirmation of test results using a "gold standard''. The true sensitivity and specificity are unknown and errors in measurement can occur due to subjective clinical judgment, technical imperfections or interpretational differences. A further complication is clustered data (such as patients nested within examiners within screening centre), which are common in population-based screening. We propose a cancer screening model that accommodates the partially unobserved disease status, clustered data structures, general covariate effects, and the dependence between exams.

The model is applied to the estimation of the diagnostic accuracy of mammography and clinical breast examination using a cohort consisting of women 50 to 69 years of age screened at the OBSP between January 1, 2002 and December 31, 2003. When offered in addition to mammography, we found CBE may benefit women using hormone therapy but not likely benefit women with dense breast tissues.

The thesis also discusses two measures of interest, the length of the pre-clinical state and the false negative rate. Two estimation procedures are proposed to model the pre-clinical state duration, the false negative rate of screening exam, and the underlying incidence rate in the screened population. Both methods assume the sojourn time follows a negative exponential distribution, but we consider two different forms for the false negative rate: 1) constant with time and 2) an exponential function to compensate for the fact that lesions should become easier to detect the closer they are to become clinically evident. The proposed methods are illustrated with another cohort of women who were first screened through the OBSP between January 1, 2003, and December 31, 2004 and were followed up until December 31, 2009.