Modification of Synthetic Reconstruction Method with Positional Information for All Households in Japan

Real-scale social simulation using agent-based models is increasingly important for evaluating public policies such as disaster mitigation and economic measures at an individual-citizen level. However, administrative microdata (e.g., registries and tax records) are typically unavailable due to privacy constraints, motivating the synthesis of synthetic microdata from publicly released tabulations. This work refines a sample-free simulated-annealing (SA) framework to generate nationwide synthetic microdata for Japan (≈1,900 municipalities), including rich attributes such as household structure, age/sex, employment-related variables, and geolocation. A key limitation of municipality-by-municipality synthesis is that tabulations for municipalities with fewer than 200,000 residents are published in coarser bins (e.g., 5-year age groups), requiring estimation or downscaling and potentially causing inconsistencies with prefecture-level statistics after aggregation. The proposed approach synthesizes all municipalities within each prefecture in a single run, explicitly aligning the survey scope of tabulations with municipality identifiers to avoid estimation and excessive downscaling. To reduce within-bin bias introduced by coarse constraints, additional 1-year overall age distributions are incorporated, and the SA swap/transition rule is redesigned based on municipality size. Nationwide experiments for 2010 and 2015 demonstrate a substantial reduction in discrepancies against prefecture-level tabulations, providing a more consistent foundation for distributing municipality-attributed, geolocated synthetic microdata for large-scale social simulations.