Dealing with Censored Earnings in Register Data

Abstract Earnings are often top-coded (right-censored) in administrative registers. The censoring threshold in the case of Germany is the limit value for social security contributions, leading to a substantial fraction of censoring: For example, about 12 % of male workers in West Germany are affected, rising to above 30 % for highly educated prime-aged workers. This missing right tail of the earnings distribution constitutes a major problem for researchers studying earnings inequality and top incomes. We overcome this challenge by taking a distributional approach and semi-parametrically modelling the right tail as being Pareto-like. Non-censored earnings survey data matched to administrative records, derived from the SOEP-RV project, let us operate in a laboratory-like setting in which the targets are known. Our approach outperforms alternative imputation methods based on Tobit regressions.