Model-Based Analysis in Instrumental Variable Settings: Randomized Experiments with Two-Sided Noncompliance
Lists
Only you see this
Abstract
In this chapter we develop a multiple-imputation, or model-based alternative, to the Neyman-style moment-based analyses for super-population average treatment effects introduced in Chapters 23 and 24. The model-based approach discussed in this chapter has a number of advantages over the moment-based approach, both conceptual and practical. First, it offers a principled way to incorporate the restrictions on the joint distribution of the observed variables that arise from the various exclusion restrictions and the monotonicity assumption. Second, it allows for a straightforward and flexible way to incorporate covariates. In the current chapter we allow for continuous covariates, or at least covariates taking on too many values for analyses to be feasibly conducted separately on subpopulations homogeneous in the covariates’ values. Therefore, we focus on a model-based approach, similar to that in Chapter 8 used in completely randomized experiments. As in Chapter 8, we start by building statistical models for the potential outcomes. A distinct feature of the approach in this chapter is that we also build a statistical model for the compliance behavior. We use these models to simulate the missing potential outcomes and the missing compliance behaviors, and use those in turn to draw inferences regarding causal effects of the primary treatment for the subset of units who would always comply with their assignment.
BibTeX
@inproceedings{imbens2015model,
author = {Guido W. Imbens and Donald B. Rubin},
title = {{Model-Based Analysis in Instrumental Variable Settings: Randomized Experiments with Two-Sided Noncompliance}},
booktitle = {Cambridge},
year = {2015},
doi = {10.1017/cbo9781139025751.026},
}Loading related content...