From agent-based models to statistical emulators

In statistical demography information about population processes is inferred from empirical data. In contrast, agent-based approaches focus on aggregate outcomes of individual-level behavioural rules. Given the non-linearities and feedbacks present in agent-based settings, their direct statistical analysis is not always feasible. Hence, in order to bridge the gap between these two perspectives, we propose to utilise Gaussian process emulators, which enable studying the outcomes of rule-based models statistically. The suggested approach includes a sensitivity analysis, assessing the relative importance of different model parameters, and a simple calibration, aimed at selecting plausible parameter values. The discussion is illustrated by presenting a Semi-Artificial Model of Population, which augments an agent-based model of partnership formation with statistical data on natural population change in the United Kingdom. The resulting multi-state model of population dynamics is better aligned with selected aspects of the demographic reality than its underpinning agent-based component alone. The analysis also illuminates important tradeoffs between different parameters and outputs considered.