Scalable Inference for Individual-Based Epidemic Models

Lorenzo Rimella

Università degli studi di Bergamo

Individual-based models allow epidemiologists to capture risk at the individual level. However, the computational cost of exact likelihood evaluation for partially observed individual-based models grows exponentially with population size, necessitating approximate inference. In this contribution, we explore scalable approximations to the likelihood for individual-based models that avoid the exponential-in-population computational cost. In particular, we analyse two approaches: Simulation-Based Composite Likelihood [1], which uses a composite likelihood approximation, and Categorical Approximate Likelihood [2], which constructs an approximation by substituting expectations in the same vein as assumed density filters. This work analyses the advantages and disadvantages of both methods and further compares them with block particle filters, while also outlining directions for future research.

Keywords: Epidemiology, Approximate Inference, Hidden Markov Models