Essays on the Econometrics of Causal Inference

Abstract

Causal inference is a constant theme in econometrics. In this dissertation, I explore three issues related to policy analysis within the context of complex data structures. In the first chapter, I perform a sensitivity analysis on various treatment effect parameters when duration outcomes are dependently censored. Endogeneity in the censoring process often arises from competing events/risks or when research subjects are lost to follow-up. The dependent censoring mechanism is characterized by Archimedean copulas in this chapter, which allows for the derivation of bounds on policy effects in a closed form. The strength of these identification results reflects the extent of prior information about the censoring process, which is no longer required to be uninformative. The second chapter examines the marginal effect of counterfactual changes in a covariate on the unconditional quantiles of an outcome, when the two variables are not part of the same dataset. We leverage variables that are common to both datasets to recover information missing from the main dataset. Under assumptions of rank similarity and conditional independence, we establish a set of nonparametric identification results on the unconditional quantile effects. In the final chapter, a new difference-in-differences estimator is introduced. This estimator is not only doubly robust against the misspecification of nuisance function models, but it also accommodates compositional changes in the covariates--a common feature of repeated cross-sectional data. This chapter further details a trade-off related to such changes: the asymptotic bias of doubly robust estimators that erroneously exclude compositional changes, versus the efficiency loss resulting from incorrect disregard of them.