Bayesian validation framework for dynamic epidemic models

• Published in: Epidemics 2021-12, Vol.37, p.100514-100514, Article 100514
• Main Authors: Dasgupta, Sayan, Moore, Mia R., Dimitrov, Dobromir T., Hughes, James P.
• Format: Article
• Language: English
• Subjects: Bayes Theorem; Bayesian credible interval; Communicable Diseases - epidemiology; Epidemics; Epidemiological model validation; HIV transmission model; Humans; Markov Chain Monte Carlo
• ISSN: 1755-4365; 1878-0067
• DOI: 10.1016/j.epidem.2021.100514

Complex models of infectious diseases are used to understand the transmission dynamics of the disease, project the course of an epidemic, predict the effect of interventions and/or provide information for power calculations of community level intervention studies. However, there have been relatively few opportunities to rigorously evaluate the predictions of such models till now. Indeed, while there is a large literature on calibration (fitting model parameters) and validation (comparing model outputs to data) of complex models based on empirical data, the lack of uniformity in accepted criteria for such procedures for models of infectious diseases has led to simple procedures being prevalent for such steps. However, recently, several community level randomized trials of combination HIV intervention have been planned and/or initiated, and in each case, significant epidemic modeling efforts were conducted during trial planning which were integral to the design of these trials. The existence of these models and the (anticipated) availability of results from the related trials, provide a unique opportunity to evaluate the models and their usefulness in trial design. In this project, we outline a framework for evaluating the predictions of complex epidemiological models and describe experiments that can be used to test their predictions. •Framework for evaluating predictions of epidemiological models (e.g., HIV transmission models).•Evaluating posterior distribution of the model discrepancy using a Bayesian framework.•Allowing for re-calibration of parameters by updating their priors in a MCMC analysis.•Identifying communities where the model fails by providing a goodness of fit evaluation using posterior tail probability.