R-package agentBayes: Likelihood-based statistical methods for agent-based models

Statistically analysing interacting particle systems remains challenging because the governing equations are analytically intractable. Existing solutions include moment closure methods with pseudolikelihood-based frameworks, and likelihood-free frameworks based on extensive simulations, both relying on heuristic choices whose validity is difficult to predict. As a resolution, we rigorously derive an asymptotically exact expression for the likelihood of agent-based models (ABMs) operating in continuous space and time that can be formulated as reactant–catalyst–product (RCP) models. We derive an expression for the conditional density of agents given information about the current and earlier distributions of neighbouring agents. We utilize this expression to construct an asymptotically exact likelihood that applies to both spatial snapshot and time-series data. We implement the likelihood expression and a Bayesian parameter estimation framework in the R-package agentBayes and demonstrate its utility in biological research and beyond with simulated case studies and empirical data on the evolution of cancer cell populations.Author summary: Many study systems across biology, finance, physics and the social sciences consist of discrete interacting agents, yet fitting agent-based models to data remains challenging as most models are too complicated to be treated analytically. We resolve this fundamental challenge and derive tractable likelihoods for a broad class of agent-based models and data, such as spatial snapshot and time-series data. Accompanied by our statistical software, we show how to apply our framework to address biological research questions at the example of the coevolution of cancer cell populations.