Bellot's paper exposes two layers of instability in causal estimation

Causal inference from observational data has long been known to be fragile. Even when a causal effect is theoretically identifiable — meaning it can be uniquely expressed in terms of the observed data distribution — the effect can still be discontinuous: two arbitrarily close data distributions may yield very different causal effects. This was formally established by Robins and Ritov (1997) and is independent of the chosen estimator. Alexis Bellot's new paper, 'Two Layers of Instability in Causal Estimation,' builds on this by exposing a second, estimator-dependent layer of instability.

Bellot shows that many standard point estimates — such as inverse propensity weighted (IPW) estimators and regression-based estimators — can be interpreted as point summaries of a multimodal distribution over the space of structural causal models (SCMs). Because these summaries are not the full posterior, they can jump discontinuously as the data changes. In contrast, explicit posterior means and medians (e.g., from Bayesian causal models) are shown to be continuous in the data distribution. The paper offers a decision-theoretic taxonomy tied to the loss function implicitly optimized: estimators that minimize a loss misaligned with the causal effect itself are more prone to instability. This work has direct implications for scientists and engineers relying on causal methods for robust decision-making.