Extreme proxies for causal identification: deconfounding with outliers

Abstract

Causal identification under unobserved confounding typically relies on additional structure, such as a valid instrument or proxy variables. We show that heavy tailed variables offer an alternative route to identification by revealing treatment variation for which confounding becomes asymptotically negligible. Our key assumption is extreme propagation: when an unobserved confounder takes an extreme value, some observed variable in the system (called a safeguard) will become extreme with large probability. This way, selecting extreme treatment observations while safeguards remain non-extreme makes confounding asymptotically negligible relative to the treatment. We call this certification. We derive graphical conditions specifying which latent confounders must be monitored by safeguards and distributional conditions ensuring effective certification. The framework includes valid-instrument and double-proxy structures as special cases, as well as confounded instruments that are invalid in most approaches that do not exploit heavy tails. Geometrically, certification isolates the extremal ray whose slope equals the causal effect. This connects causal identification with multivariate extreme-value geometry and motivates a family of novel simple estimators. Simulations demonstrate performance gains under nonlinear confounding and heavy tails, while revealing limitations in light-tailed regimes (where cancellation can conceal extreme confounding from the safeguards). We validate our methodology in simulation studies and by recovering the ground-truth flow in a river discharge application.