Abstract:
|
In screening data, disease onset occurs between the last time observed to be disease free and the time of diagnosis (interval censoring), but also has a point mass at time zero because of prevalent disease at the initial screen. Furthermore, prevalent disease is not always immediately diagnosed, which is common for people with negative screening results. Consequently, some disease diagnosed at future screens is actually undiagnosed prevalent disease. We demonstrate that standard analysis methods, which ignore those features, generally underestimate early risks and overestimate later risks. We propose parametric logistic-Weibull and semiparametric logistic-Cox mixture models for survival analyses of screening data. Parameters are estimated with direct likelihood maximization for the logistic-Weibull, but the logistic-Cox requires a variant of the EM algorithm. Using these models, we re-analyze data from the cervical cancer screening program at Kaiser Permanente Northern California that underlie current U.S. screening and management guidelines for cervical cancer. We show that Kaplan-Meier estimates can be grossly misleading, and our models provide superior risk estimates.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.