PURPOSE: Model calibration involves systematically adjusting input parameter sets until model outputs match target endpoints. The conventional method of calibrating longitudinal models to cross-sectional data assumes no changes in epidemiology or practice patterns. The study objective was to use a novel method to calibrate a model of a cervical cancer (CC) vaccine candidate to cross-sectional data, reflecting increased CC screening since the introduction of the Pap test.

METHODS: We developed a Markov cohort model of CC with six-month cycles. We used Nelder-Mead optimization to calibrate the model to age-specific 2000-2003 Surveillance, Epidemiology and End Results (SEER) data on CC incidence and mortality (per 100,000). First, a conventional calibration was conducted that assumed all women had undergone current screening practices (based on health plan data) throughout their lives. Subsequently, we performed a second calibration that stratified the current US female population into two groups: those aged <65 years in 2003 who experienced current cervical screening practices ("<65"); and those aged Ан65 years in 2003 who did not have the benefit of screening at younger ages (">=65"). Therefore, the second calibration required finding a parameter set that simultaneously produced a good fit to the SEER data applicable to patients <65 assuming current screening and Ан65 assuming historical screening. RESULTS: The best input parameter set fit by the conventionally calibrated model (where historical screening was not accurately reflected) produced model-generated CC incidence rates per 100,000 of 9.6 for <65 and 12.1 for >=65 and mortality rates per 100,000 of 2.4 and 8.3, respectively. These rates appear to be consistent with SEER data (incidence: 10.8 and 13.4; mortality: 2.6 and 7.1). However, inserting the parameter set generated by conventional calibration into a model incorporating historical screening patterns resulted in an overestimation of CC mortality in the >=65 group (16.7) compared to SEER (7.1). In the second calibration, which took into account temporal changes in screening practices, the optimal parameter set produced a mortality rate in the >=65 group of 7.3 compared to 7.1 in SEER.

CONCLUSIONS: The parameter set generated by conventional calibration substantially over-estimated CC mortality in older women when used with actual historical screening patterns. These results demonstrate that the effect of secular changes in health practices should be considered when calibrating longitudinal models to cross-sectional data.