The Impact of Different Screening Model Structures on Cervical Cancer Incidence and Mortality Predictions: The Maximum Clinical Incidence Reduction (MCLIR) Methodology

Background. To interpret cervical cancer screening model results, we need to understand the influence of model structure and assumptions on cancer incidence and mortality predictions. Cervical cancer cases and deaths following screening can be attributed to 1) (precancerous or cancerous) disease that occurred after screening, 2) disease that was present but not screen detected, or 3) disease that was screen detected but not successfully treated. We examined the relative contributions of each of these using 4 Cancer Intervention and Surveillance Modeling Network (CISNET) models. Methods. The maximum clinical incidence reduction (MCLIR) method compares changes in the number of clinically detected cervical cancers and mortality among 4 scenarios: 1) no screening, 2) one-time perfect screening at age 45 that detects all existing disease and delivers perfect (i.e., 100% effective) treatment of all screen-detected disease, 3) one-time realistic-sensitivity cytological screening and perfect treatment of all screen-detected disease, and 4) one-time realistic-sensitivity cytological screening and realistic-effectiveness treatment of all screen-detected disease. Results. Predicted incidence reductions ranged from 55% to 74%, and mortality reduction ranged from 56% to 62% within 15 years of follow-up for scenario 4 across models. The proportion of deaths due to disease not detected by screening differed across the models (21%–35%), as did the failure of treatment (8%–16%) and disease occurring after screening (from 1%–6%). Conclusions. The MCLIR approach aids in the interpretation of variability across model results. We showed that the reasons why screening failed to prevent cancers and deaths differed between the models. This likely reflects uncertainty about unobservable model inputs and structures; the impact of this uncertainty on policy conclusions should be examined via comparing findings from different well-calibrated and validated model platforms.