Econometric microsimulation models that simulate the effects of taxation and social benefit legislation on the disposable incomes of individuals and households are widely used by social scientists and policymakers worldwide. The results produced by these models have a degree of uncertainty arising from multiple sources. One of these is sampling error that is caused by the fact that the simulation is performed on a sample of the total population of interest. However, assessment of the accuracy of results through the estimation of sampling variability caused by this error is still largely absent in the microsimulation literature.
The users of econometric microsimulation models are often interested in the values of certain inequality and poverty indicators. This thesis presents variance estimation methods that can be employed to produce variance estimates for these indicators. The main focus is on bootstrap and linearization methods for variance estimation and the indicators considered are the at-risk-of-poverty threshold (ARPT), the at-risk-of-poverty rate (ARPR) and the Gini coefficient. The efficiency of variance estimation methods is tested in a simulative study performed on a data set produced by the SISU microsimulation model developed by Statistics Finland. The methods are also employed in a hands-on case study to help assess the effects of an actual legislative reform simulated by the SISU model.
It is found that both bootstrap and linearization methods for variance estimation produce relatively good variance estimates for the indicators considered, with linearization being the more effective of the two. However, high outlier incomes are shown to cause difficulties in the variance estimation of the Gini coefficient with both methods.