Compositional difference-in-differences for categorical outcomes

Abstract: In difference-in-differences (DiD) settings with categorical outcomes, treatment effects often operate on both total quantities (e.g., voter turnout) and category shares (e.g., vote distribution across parties). In this context, linear DiD models can be problematic: they suffer from scale dependence, may produce negative counterfactual quantities, and are inconsistent with discrete choice theory. We propose compositional DiD (CoDiD), a new method that identifies counterfactual categorical quantities, and thus total levels and shares, under a parallel growths assumption. The assumption states that, absent treatment, each category's size grows or shrinks at the same proportional rate in treated and control groups. In a random utility framework, we show that this implies parallel evolution of relative preferences between any pair of categories. Analytically, we show that it also means the shares are reallocated in the same way in both groups in the absence of treatment. Finally, geometrically, it corresponds to parallel trajectories (or movements) of probability mass functions of the two groups in the probability simplex under Aitchison geometry. We extend CoDiD to i) derive bounds under relaxed assumptions, ii) handle staggered adoption, and iii) propose a synthetic DiD analog. We illustrate the method's empirical relevance through two applications: first, we examine how early voting reforms affect voter choice in U.S. presidential elections; second, we analyze how the Regional Greenhouse Gas Initiative (RGGI) affected the composition of electricity generation across sources such as coal, natural gas, nuclear, and renewables.