Causal pitfalls in the decomposition of wage gaps

The decomposition of gender or ethnic wage gaps into explained and unexplained components (often with the aim to assess labor market discrimination) has been a major research agenda in empirical labor economics. This paper demonstrates that conventional decompositions, no matter whether linear or non-parametric, are equivalent to assuming a (probably too) simplistic model of mediation (aimed at assessing causal mechanisms) and may therefore lack causal interpretability. The reason is that decompositions typically control for post-birth variables that lie on the causal pathway from gender/ethnicity (which are determined at or even before birth) to wage but neglect potential endogeneity that may arise from this approach. Based on the newer literature on mediation analysis, we therefore provide more attractive identifying assumptions and discuss non-parametric identification based on reweighting.