Causal Identification Under Deep Uncertainty
Classical causal identification assumes the analyst can enumerate the relevant variables. We relax this assumption: when the variable set itself is uncertain, what can be identified? We give necessary and sufficient conditions, and demonstrate them on a climate-policy case study where the standard approach fails.
Problem
The do-calculus tells us what is identifiable given a known causal graph. It says nothing about whether the graph itself is the right one.
Contribution
We introduce *graph-robust identification*: a causal claim is graph-robust if it holds for every member of a hypothesis class of graphs consistent with the data. We give necessary and sufficient conditions, an algorithm, and a worst-case sample complexity bound.
Case study
Re-analysis of a published climate-policy attribution finding: the original conclusion does not survive graph-robust identification. We identify the specific additional observation that would settle the question.