|
Abstract:
|
Causal inference in empirical studies needs assumptions about certain "stability" of the underlying data generating processes. For instance, a fundamental assumption underlying the Neyman-Rubin potential outcomes, the Stable Unit Treatment Value Assumption, is explicit about the stability of the intervention of interest, and another crucial assumption for observational studies, unconfoundedness or ignorability, is implicit about the stability of the background features of the units across treatment groups. In fact, many causal inference problems invoke various forms of "stability", although they appear with different names such as "homogeneity," "no interaction," "invariance," "transportability" and so on. I will review these concepts in causal inference and try to unify them under the general "stability principle."
|
