An Evaluation of Two Synthetic Small-Area Microdata Simulation Methodologies: Synthetic Reconstruction and Combinatorial Optimisation

Abstract

An essential requirement for all microsimulation models is an initial set of population microdata. If processes are to be modelled at the subregional level, these microdata must be spatially detailed. Unfortunately, confidentiality and sample size restrictions preclude the provision of such microdata via either census or social survey. As a result, a number of techniques have been developed to synthetically generate the requisite microdata.

This chapter reports on two techniques developed in the UK: ‘synthetic reconstruction’ and ‘combinatorial optimisation’. The basic methodology of each approach is outlined, following which recent innovations in their detailed implementation are introduced. The performance of both approaches is necessarily affected by the nature of between-place differences, which are shown to be surprisingly resistant to conventional data reduction techniques. Having considered this spatial variability, a new framework is introduced for the evaluation and validation of synthetic microdata, which includes multiple realisations and the use of measures of fit based around the Z-score and two derivations: Σ Z ² and RSSZ. Finally, an evaluation of the output from each approach is presented. Combinatorial optimisation is shown to produce estimates with less bias, and significantly less variance, than those produced via synthetic reconstruction. The resulting synthetic microdata are shown to have a very high degree of fit to estimation constraints and to produce good estimates for margin-constrained distributions.