OBJECTIVES: Simulation models of opioid use disorder (OUD) evaluate the impact of different treatment strategies on population-level outcomes. Researching Effective Strategies to Prevent Opioid Death (RESPOND) is a dynamic population, state-transition model that simulates the Massachusetts OUD population synthesizing data from multiple sources. Structural complexity and scarcity of available data for opioid modeling pose a special challenge to model calibration. We propose a two-step approach for calibrating RESPOND to OUD outcomes: a naïve empirical calibration (EC) to inform priors of a more rigorous Bayesian calibration method.

METHODS: We generate informative priors for the calibrated parameters by first performing a naïve EC using Latin hypercube sampling to identify high-density areas in the multidimensional parameter space comprising arrivals, overdose rates, treatment transition rates, and substance use state transition probabilities. EC accepts proposed parameters when the respective model outputs lie within pre-determined uncertainty ranges of calibration targets: yearly OUD population sizes, admissions to detox facilities, and fatal overdoses. We use results from EC to inform priors of an Approximate Bayesian Computation (ABC) method employing Sequential Monte Carlo (SMC) posterior sampling to search for parameters. SMC modifies ABC rejection sampling by using weights to enhance the performance of the search algorithm by decreasing/increasing the sampling rate from low-/high-probability regions iteratively.

RESULTS: Preliminary runs have shown a good model fit resulting in a maximum absolute relative error of 0.03 between the average model outcomes and the corresponding calibration targets. The calibrated model is also externally validated for active OUD counts, treatment admissions, and overdose rates.

CONCLUSIONS: The combination of an Empirical and Bayesian approach can considerably improve the results both in terms of efficiency and model accuracy when calibrating complex simulation models. The Bayesian approach also allows for assessing the model performance based on the resulting posterior predictive distributions of the calibration targets.