New techniques for household microsimulation, and their application to Australia

Abstract

Household microsimulation models are sometimes used by national governments to make long-term projections of proposed policy changes. They are costly to develop and maintain, and sometimes have short lifetimes. Most present national models have limited interactions between agents, few regions and long simulation cycles. Some models are very slow to run. Overcoming these limitations may open up a much wider range of government, business and individual uses. This thesis suggests techniques to help make multi-purpose dynamic microsimulations of households, with fine spatial resolutions, high sampling densities and short simulation cycles. Techniques suggested are: * simulation by sampling with loaded probabilities * proportional event alignment * event alignment using random sampling * immediate matching by probability-weighting * immediate 'best of n' matching. All of these techniques are tested in artificial situations. Three of them - sampling with loaded probabilities, alignment using random sampling and best of n matching - are successfully tested in the Cumpston model, a household microsimulation model developed for this thesis. Sampling with loaded probabilities is found to give almost identical results to the traditional all-case sampling, but be quicker. The suggested alignment and matching techniques are shown to give less distortion and generally lower runtimes than some techniques currently in use. The Cumpston model is based on a 1% sample from the 2001 Australian census. Individuals, families, households and dwellings are included. Immigration and emigration are separately simulated, together with internal migration between 57 statistical divisions. Transitions between 8 person types are simulated, and between 9 occupations. The model projects education, employment, earnings and retirement savings for each individual, and dwelling values, rents and housing loans for each household. The onset and development of diseases for each individual are simulated. Validation of the model was based on methods used by the Orcutt, CORSIM, DYNACAN and APPSIM models. Iterative methods for model calibration are described, together with a statistical test for creep in multiple runs. The model takes about 85 seconds to make projections for 50 years with yearly simulation cycles. After standardizing for sample size and projection years, this is a little slower than the fastest national models currently operating. A planned extension of the model is to 2.2 million persons over 2,214 areas, synthesized from 2011 census tabulations. Using multithreading where feasible, a 50-year projection may take about 10 minutes.