Improving the synthetic data generation process in spatial microsimulation models

Simulation models are increasingly used in applied research to create synthetic micropopulations and predict possible individual-level outcomes of policy intervention. Previous research highlights the relevance of simulation techniques in estimating the potential outcomes of changes in areas such as taxation and child benefit policy, crime, education, or health inequalities. To date, however, there is very little published research on the creation, calibration, and testing of such micropopulations and models, and little on the issue of how well synthetic data can fit locally as opposed to globally in such models. This paper discusses the process of improving the process of synthetic micropopulation generation with the aim of improving and extending existing spatial microsimulation models. Experiments using different variable configurations to constrain the models are undertaken with the emphasis on producing a suite of models to match the different sociodemographic conditions found within a typical city. The results show that creating processes to generate area-specific synthetic populations, which reflect the diverse populations within the study area, provides more accurate population estimates for future policy work than the traditional global model configurations.