Calibration of Microsimulation Models: A Likelihood-based Approach

• Activity Number: 328
• Type: Topic Contributed
• Date/Time: Wednesday, August 11, 2004 : 10:30 AM to 12:20 PM
• Sponsor: Section on Health Policy Statistics
• Author(s): Amy Knudsen*+ and Milton C. Weinstein and Karen M. Kuntz
• Companies: Harvard School of Public Health and Harvard School of Public Health and Harvard School of Public Health
• Keywords: microsimulation ; model calibration ; likelihood ; colorectal cancer

Abstract

Modeling unobserved processes is challenging since limited data are available to directly inform the transitions between unobserved states. Such models can be estimated by constraining model outcomes to calibrate to observable data. We explored a likelihood-based approach to the calibration of a natural history model of colorectal cancer (CRC). We calibrated the model by simulating life histories under a given set of parameters and comparing model outcomes with data on the prevalence, location, and size of adenomas from autopsy and screening studies, and the incidence, stage, and location of CRC from a cancer registry. We assumed each set of data follows a binomial distribution and calculated two likelihoods for each measure: (1) likelihood of generating the data with a given set of parameters (ie, observed likelihood), and (2) likelihood if the model exactly predicted the data (ie, maximum likelihood). An overall GOF score was calculated as the sum of the differences in -2 log likelihoods. The parameter space was explored using the Nelder-Mead Simplex algorithm. Likelihood-based methods are valuable tools for model calibration and provide insight into limitations in model structure.