Decomposing wage distributions on a large data set - a quantile regression analysis of the gender wage gap

This paper presents and implements a procedure that makes it possible to decompose wage distributions on large data sets. We replace bootstrap sampling in the standard Machado-Mata procedure with ‘non-replacement subsampling’, which is more suitable for the linked employer-employee data applied in this paper. Decompositions show that most of the glass ceiling is related to segregation in the form of either composition effects or different returns to males and females. A counterfactual wage distribution without differences in the constant terms (or ‘discrimination’) implies substantial changes in gender wage differences in the lower part of the wage distribution.