Evaluating Data Fusion Methods to Improve Income Modeling

Income is an important economic indicator to measure living standards and individual well-being. In Germany, different data sources yield ambiguous evidence for analyzing the income distribution. The Tax Statistics (TS)—an income register recording the total population of more than 40 million taxpayers in Germany for the year 2014—contains the most reliable income information covering the full income distribution. However, it offers only a limited range of socio-demographic variables essential for income analysis. We tackle this challenge by enriching the tax data with information on education and working time from the Microcensus, a representative 1 percent sample of the German population. We examine two types of data fusion methods well suited to the specific data fusion scenario of the TS and the Microcensus: missing-data methods and performant prediction models. We conduct a simulation study and provide an empirical application comparing the proposed data fusion methods, and our results indicate that Multinomial Regression and Random Forest are the most suitable methods for our data fusion scenario.