This paper introduces a new method for asking causal questions about entire distributions, not just averages. The authors build “counterfactual densities,” which ask things like: what would the wage distribution of East Germans look like if they had the same relationship between wages and characteristics as West Germans, or if they had West German characteristics? This lets them break the observed difference into two parts: a distribution effect (changes in the wage pattern conditional on characteristics) and a covariate effect (changes from a different mix of characteristics like age or education).

To make these ideas practical the authors work with conditional densities — functions that show the full range of possible wages for people with given characteristics. They analyze those densities inside a mathematical space called a Bayes Hilbert space. That name sounds technical, but the key point is that this choice keeps estimated densities non-negative and ensures they add up to one, which raw regression fits can violate. For estimation they use a flexible additive model built from basis functions (for example, splines) and a Poisson-based approximation to the likelihood, which the authors argue avoids both the strong assumptions of simple parametric models and the “curse of dimensionality” of fully nonparametric methods.

The method is closely tied to a form of Oaxaca–Blinder decomposition, a common tool for breaking group differences into parts. Here the decomposition is multiplicative, focusing on ratios between densities rather than absolute differences. The authors say that ratios work better in regions where densities are low, and the density view can show features such as bimodal shapes or a spike at zero income. Those features matter for wage data, which can be skewed, have two modes, and include many zeroes (for example, people without wage income).