|
Abstract:
|
For cost-effectiveness and efficiency, large-scale epidemiologic cohorts are being assembled within large health-care providers who use electronic health records. For estimating absolute risk from such cohorts, there are challenges that incident disease is interval-censored between irregular visits and prevalent disease is left-censored. Furthermore, because prevalent disease is not always immediately diagnosed, disease diagnosed at future visits is a mixture of truly incidence disease and undiagnosed prevalent disease. The second challenge we address is to estimate absolute risk from phase 2 case-control studies nested within such cohorts. We propose inverse-probability weighted (IPW) prevalence-incidence models, which are mixture models for interval-censored incident disease and point-mass left-censored prevalent disease. We employ iterative algorithms to estimate parameters for semiparametric and weakly-parametric prevalence-incidence models. We apply the models to subcohort sample cervical cancer screening data at Kaiser Permanente Northern California and calculate absolute risks of cervical precancer/cancer for different human papillomavirus types.
|
