Auditing Causal Teleconnections: An Identification-First Framework for Regional SST and Extreme Precipitation in ERA5

Extreme precipitation in the western United States is shaped by coupled ocean--atmosphere variability, but observational teleconnection analyses do not by themselves define causal estimands. We present an identification-grounded causal workflow for ERA5 monthly reanalysis panels over two western-US regions from 1979-2023. The workflow separates discovery from identification, documents a projection from time-series discovery output to a lag-0 acyclic panel graph, and validates a target estimand with IPW, doubly robust estimation, falsification checks, temporal holdout, and targeted sensitivity analyses. The primary estimand is the risk-difference effect of warm regional sea-surface temperature anomalies on 90th-percentile precipitation extremes during ONDJFM in the Pacific Northwest. In the full sample, IPW estimates 0.146 (95% CI 0.026, 0.237), and DR estimates 0.160 (95% CI 0.040, 0.245). In a temporal holdout with nuisance models fit on 1979-1999 and evaluated on 2000-2023, IPW estimates 0.180 (95% CI 0.016, 0.361), and DR estimates 0.066 (95% CI 0.045, 0.085). The effect is not western-US-wide and remains sensitive to broad observed-covariate adjustment. The contribution is a transparent climate-ML workflow for making causal claims, regional boundaries, and sensitivity explicit.