Understanding the effect of a treatment or action based on data is a fundamental goal in numerous scientific fields, such as economics, biology, and the social sciences. One way to quantify this is the Average Treatment Effect (ATE), defined as the difference between the expected outcome on the treated and untreated population. Unfortunately, ATE is not identifiable from data without making some assumptions on the data-generating process. The most widespread ones require that treatment is assigned randomly within subpopulations (unconfoundedness), and with positive probability on every subpopulation (overlap). However, there are many scenarios where these requirements are not–and cannot–be satisfied, raising the question: Is ATE identification possible beyond unconfoundedness and overlap?
In this talk, I will present a recent work with my collaborators that characterizes ATE identifiability. I will show how this characterization enables ATE identification in important scenarios that prior works could not capture. Finally, I will discuss some sufficient conditions for estimating ATE in these scenarios.
The talk is based on joint work with Yang Cai, Alkis Kalavasis, Anay Mehrotra and Manolis Zampetakis.