Inference Bounds on the Joint Distribution of Treatment Outcomes
30 Pages Posted: 18 Jan 2014
Date Written: January 16, 2014
The paper generalizes ideas presented on bounds of treatment effects (Manski (1990)) to bounding the joint distribution of treatment outcomes. Frechet-Hoeffding bounds are sharp for the distribution of treatment outcomes from unconfounded data and Manski bounds are sharp for the distribution of treatment outcomes from confounded data. In general, unconfounded data does not provide sharper bounds than confounded data. Frechet-Hoeffding bounds can be made sharper with the addition of observational characteristics of the agent that are associated with differential treatment outcomes (signals). Manski bounds can be sharpened with the addition of observational information affecting the agent's selection into treatment (instruments). Additionally, Manski bounds can be made sharper with the use of behavioral assumptions.
Keywords: bounds, joint distributions, treatment effects, heterogenous treatment effects
JEL Classification: C10, C14
Suggested Citation: Suggested Citation