Identifying treatment effects on categorical outcomes in IV models

Abstract: This paper provides a nonparametric framework for causal inference with categorical outcomes under binary treatment and binary instrument settings. We decompose the observed joint probability of outcomes and treatment into marginal probabilities of potential outcomes and treatment, and association parameters that capture selection bias due to unobserved heterogeneity. Under a novel identifying assumption, association similarity, which requires the dependence between unobserved factors and potential outcomes to be invariant across treatment states, we achieve point identification of the full distribution of potential outcomes. Recognizing that this assumption may be strong in some contexts, we propose two weaker alternatives: monotonic association, which restricts the direction of selection heterogeneity, and bounded association, which constrains its magnitude. These relaxed assumptions deliver sharp partial identification bounds that nest point identification as a special case and facilitate transparent sensitivity analysis. We illustrate the framework in an empirical application, estimating the causal effect of private health insurance on health outcomes.