Empirically Estimating an Agent-Based Model of School Choice on Household-Level Register Data

In agent-based models (ABMs), the estimation of model parameters from data is much less straightforward than in traditional but more limited techniques of modelling data, such as regression. For most ABMs, the likelihood of a parameter vector given the data cannot be written down explicitly nor sampled from, ruling out commonly used techniques such as maximum likelihood estimation and Markov chain Monte Carlo sampling. This study proposes a methodology to estimate ABMs on household-level data, for an ABM tailored to primary school choice and segregation in the Netherlands. It explores the interplay between the micro-, meso-, and macro-levels in school choice dynamics, highlighting the limitations of conventional methodologies in capturing such interactions. By estimating an ABM directly on household-level data using neural ratio estimation, the study enhances the realism of the ABM, shedding light on choice processes and mechanisms driving segregation. It unveils that heuristic-based models better capture household behaviours than traditional models of rational action, challenging existing assumptions. This study not only advances understanding of school choice dynamics, but also provides a estimation framework applicable to ABMs of other social systems, paving the way for more realistic and validated ABMs.