Distributional Treatment Effect with Latent Rank Invariance
Treatment effect heterogeneity is of great concern in policy evaluation. However, existing analyses have mostly been limited to summary measures such as an average treatment effect, due to the fundamental limitation that we cannot simultaneously observe both potential outcomes for a given unit. In this paper, I identify and estimate moment-identified distributional treatment effect (DTE) parameters, such as the marginal distribution of treatment effect. The key identifying assumption is that the two potential outcomes are conditionally independent given a latent variable. Interpreting this latent variable as underlying individual-level heterogeneity, I motivate the identifying assumption as `latent rank invariance.' In the conditional independence framework, the DTE parameters are identified given two proxy variables. In implementation, I assume a finite support on the latent variable, imposing a finite mixture structure on the identifying assumption. Using Neyman orthogonality, I establish asymptotic normality of the estimator, enabling inference for DTE parameters.