Essays on Instrumental Variables and Identification in Causal Inference

This dissertation studies instrumental variable (IV) models when standard assumptions underlying causal inference may be violated. While IV methods are widely used in empirical economics, their causal interpretation typically relies on strong conditions such as exclusion restrictions, independence, and the absence of spillovers. The three chapters in this dissertation develop econometric frameworks that relax these assumptions while maintaining informative identification of causal parameters. The first chapter develops a general framework for identifying causal effects in settings with spillovers, where individuals’ outcomes and treatment decisions may depend on the behavior of peers within a known group. It introduces generalized local average controlled spillover and direct effects (LACSEs and LACDEs), extending the local average treatment effect (LATE) framework to environments with spillovers. The chapter also defines marginal controlled spillover and direct effects (MCSEs and MCDEs), which extend the marginal treatment effect (MTE) framework to settings with spillovers and are nonparametrically identified using continuous instruments. These marginal effects serve as building blocks for a wide class of policy-relevant treatment effects (PRTEs). Semiparametric and parametric estimators are proposed, and an empirical application using Add Health data documents heterogeneous education spillovers within best-friend networks. The second chapter studies identification in binary-treatment, binary-instrument models when the exclusion restriction may fail. It introduces the local average treatment-controlled direct effect (LATCDE) and the local average instrument-controlled direct effect (LAICDE) and derives sharp identified bounds for the full vector of causal parameters under random assignment and monotonicity. The framework is extended to one-way spillover settings, where treatment assigned to one individual may directly affect the outcomes of others. The third chapter proposes a procedure for testing identification assumptions in parametric models using testable implications based on intersection bounds. When identifying assumptions are rejected, the chapter develops an alternative framework that relaxes parametric structure and IV assumptions and derives identified bounds for potential outcome distributions, allowing partial identification of causal parameters such as marginal treatment effects.