5. Conclusion
A mixture of log-normal distributions was found to be a convenient and powerful explanatory model of the UK income distribution. We have demonstrated how a Gibbs sampler can be used to estimate this type of mixture when we elicit more precise prior information which helps to reduce the usual label switching problem. Using the UK FES data, we have managed to identify and characterize income groups.
We were able, in this context, to provide a Bayesian inference for commonly used inequality indices that are decomposable. Using a Rao-Blackwellization, we could provide a plausibly more numerically accurate evaluation of posterior standard deviations. We have extended the method to indices that are not decomposable at the price of a one dimensional numerical integration, showing how it works for the Gini and the Atkinson indices.
As a final remark, we can note by inspecting the graph of the posterior predictive density (see Fig. 5) that the last member of the log-normal mixture does not seem satisfactory for modelling high incomes. In order to have a large right tail for the third group, we must have a large value for View the MathML source. A Pareto tail would have been intellectually more satisfactory, corresponding to a hybrid mixture of two log-normals and a Pareto. Hybrid mixtures are not common in the literature. This topic is left for future research.