MODELING HISTORICAL TRENDS IN CERVICAL CANCER USING A MULTI-COHORT MODEL

Purpose: When using models to assist in decision making, single birth cohorts are often simulated assuming population homogeneity over time. For diseases that have changed over time, whether attributable to changes in risk factors or improvements in interventions, the ability of a model to incorporate temporal trends may be important. Using a previously-developed model of cervical carcinogenesis, we assessed model fit to historical cervical cancer data after introducing model features that allow inputs to vary over multiple birth cohorts of women over time.

Method: We adjusted a previously-calibrated microsimulation model of cervical carcinogenesis to allow for major temporal changes with respect to cervical cancer prevention and control in the U.S., including screening uptake and cancer treatment effectiveness. We used publicly-available data to inform changes in the corresponding model inputs of screening coverage and cancer survival, as well as background mortality, in consecutive cohorts of U.S. women born each year between 1900-2000. Combining outcomes from multiple birth cohorts at appropriate ages, we produced model estimates of cervical cancer incidence and mortality over a 50-year period to compare to data from U.S. cancer registries from 1975-2000.

Result: Consistent with the empirical data, model results showed a decrease in cancer incidence and mortality over time as screening increased and treatment improved. The model predicted age-adjusted cervical cancer incidence rates (per 100,000 women) of 10.6 in 1975, 9.7 in 1985, 8.8 in 1995, and 8.2 in 2000; corresponding data from the U.S. Surveillance Epidemiology and End-Result (SEER) cancer registry were 14.8 in 1975, 10.2 in 1985, 8.9 in 1995, and 7.7 in 2000. The reduction in cervical cancer incidence observed empirically between 1975 and 1995 is greater than that predicted from the model (48% versus 27%, respectively); however information on screening patterns at the individual level and changing test performance of screening over time are limited. When varying assumptions of historical screening frequency in sensitivity analysis, modeled results more closely matched observed cancer rates.

Conclusion: When modeling a disease that has changed over time, simulating multiple birth cohorts with time-varying attributes may be an effective means of assessing model performance and reflecting important heterogeneities in the population; however, these benefits should be considered against the purpose of the model and the computing resources required for such an exercise.