Evaluating Data Fusion Methods to Improve Income Modeling.

This article focuses on integrating Germany's Tax Statistics (TS), a comprehensive income register covering over 40 million taxpayers with high-quality income data but limited socio-demographic variables, and the Microcensus (MC), a representative 1% population survey rich in socio-demographic information such as education and working time. To enrich the TS with these socio-demographic variables, the study evaluates four data fusion methods—Multinomial Logistic Regression, Predictive Mean Matching, Recursive Binary Partitioning (Rpart), and Random Forest—using simulation studies and empirical applications. Results indicate that incorporating income as a common variable improves fusion accuracy, with Multinomial Regression and Random Forest performing best, though each exhibits some bias in estimating parameters. The study demonstrates that these methods can effectively create an integrated database combining reliable income data with socio-demographic indicators to support detailed income distribution analyses in Germany.