Essays in econometrics
- In this dissertation, I propose novel approaches to causal inference in the settings characterized by an explicit clustering structure. I study different aspects of this problem, considering settings with few large clusters as well as with many small clusters. The dissertation consists of two essays. The first essay proposes a new model for causal inference in the settings with few large clusters and cluster-level treatment assignment. The second essay studies causal inference questions in the settings with many clusters of moderate size and individual-level treatment assignment. In the first essay, I construct a nonlinear model for causal inference in the empirical settings where researchers observe individual-level data for few large clusters over at least two time periods. It allows for identification (sometimes partial) of the counterfactual distribution, in particular, identifying average treatment effects and quantile treatment effects. The model is flexible enough to handle multiple outcome variables, multidimensional heterogeneity, and multiple clusters. It applies to the settings where the new policy is introduced in some of the clusters, and a researcher additionally has information about the pretreatment periods. I argue that in such environments we need to deal with two different sources of bias: selection and technological. In my model, I employ standard methods of causal inference to address the selection problem and use pretreatment information to eliminate the technological bias. In case of one-dimensional heterogeneity, identification is achieved under natural monotonicity assumptions. The situation is considerably more complicated in case of multidimensional heterogeneity where I propose three different approaches to identification using results from transportation theory. The second essay is co-authored with Guido Imbens. We develop a new estimator for the average treatment effect in the observational studies with unobserved cluster-level heterogeneity. We show that under particular assumptions on the sampling scheme the unobserved confounders can be integrated out conditioning on the empirical distribution of covariates and policy variable within the cluster. To make this result practical we impose a particular exponential family structure that implies that a low-dimensional sufficient statistic can summarize the empirical distribution. Then we use modern causal inference methods to construct a novel doubly robust estimator. The proposed estimator uses the estimated propensity score to adjust the familiar fixed effect estimator.
|Type of resource
|electronic resource; remote; computer; online resource
|1 online resource.
|Benkard, C. Lanier
|Benkard, C. Lanier
|Degree committee member
|Stanford University, Graduate School of Business.
|Statement of responsibility
|Submitted to the Graduate School of Business.
|Thesis Ph.D. Stanford University 2018.
- © 2018 by Dmitry Arkhangelskiy
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...