Purpose: For Markov models of disease processes with substantial prognostic uncertainty, the assessment of internal and external consistency is important. Assessment of consistency is not standardized.

Methods: A framework for assessing external consistency was described in the context of a cervical cancer model for CEA of screening options in Ontario.

1) Structure: A 19-state Markov model was used to simulate lifetime events along the causal pathway from transient and persistent human papillomavirus (HPV) infection, cervical disease (CD) and invasive cervical cancer (ICC) of a birth cohort aged 13. The model also incorporated the effectiveness of Pap smear every 3 years for women aged >18.

2) Populate: Six-month transition probabilities (⁶TP) of incidence and clearance of transient infection (TI) were estimated using data from a low-risk group of Canadian women aged 15-49 (n=259) and the McGill-Concordia student cohort aged 17-42 (n=621). Progression and regression from persistent infection, CD and ICC Stage I were stochastically represented by 16 uniform distributions of ⁶TP estimates derived from published HPV models. Published data for other ICC stages were used.

3) Target: The five calibration variables (target distributions) were lifetime risk of ICC (0.7 ± 0.18%), risk of ICC-related death (0.3 ± 0.19%), age-standardized-hysterectomy-corrected ICC rate (17 ± 3.2/100,000), and the median retention time for High Risk (HR: 14.0 ± 1.25 months) and Low Risk (LR: 12.4 ± 1.12 months) infection.

4) Calibrate: Monte-Carlo simulations (1st-2nd order) were conducted to derive modeled outcomes, sojourn times and calibration estimates . Using the target distributions, calibration estimates were converted into Z-scores. The deviation from each target distribution (i.e., [Z-score]²) was C₁² distributed and the overall deviation C₅² distributed. The simulated configuration that minimized both the overall and individual deviations was deemed optimal.

Results: The model was initially calibrated to 0.7% lifetime risk of ICC, 0.01% ICC-related death, 14.5/100,000 for the adjusted ICC rate, 18.2 and 17.5 months for HR and LR infection, respectively (C₅² = 13.2; p=0.02). This was not an optimal solution. A heuristic search through the 16-dimension ⁶TP space is ongoing to improve predicted ICC-related death and retention time.

Conclusions: For Markov models of complex disease processes, calibration is essential prior to use in CEA studies. The proposed framework outlines one systematic approach to improve the conduct and reporting of such calibration.