Empirical calibration of a simulation model of opioid use disorder

Simulation models of opioid use disorder (OUD) aim at evaluating 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 an empirical calibration approach applicable to complex simulation models in general.

We implement an empirical approach to calibrate RESPOND to multiple targets: annual fatal opioid-related overdoses, detox admissions, and OUD population sizes. The empirical calibration involves Latin hypercube sampling for searching a multidimensional parameter space comprising arrivals, overdose rates, treatment transition rates, and substance use state transition probabilities. The algorithm accepts proposed parameters when the respective model outputs lie within pre-determined target uncertainty ranges. This is an iterative process resulting in a set of parameter values for which the model closely fits all the calibration targets. We validated the model assessing its accuracy to projections important for shared decision-making of OUD outside the training data.

The empirical calibration resulted in a model that fits well both calibration and validation targets. The flexibility of the algorithm allowed us to explore structural and parameter uncertainty, reveal underlying relationships between model parameters and identify areas of model improvement for a more accurate representation of the OUD dynamics.

The proposed empirical calibration approach is an efficient tool for approximating parameter distributions of complex models, especially under complete uncertainty. Empirically calibrated parameters can be used as a starting point for a more comprehensive calibration exercise, e.g., to inform priors of a Bayesian calibration. The calibrated RESPOND model can be used to improve shared decision-making for OUD.