A Causal Inference Pipeline for Injection Open-Source Methodology and Implementation

Growing seismicity in the Permian Basin has renewed industry and regulatory interest in understanding how and where salt-water-disposal (SWD) operations trigger earthquakes. We present the first fully open-source, end-to-end causal-inference pipeline that formalises physical mechanisms in a directed acyclic graph (DAG); links > 660 000 SWD injection records to > 6 000 TexNet earthquake reports; quantifies total, direct, and pressure-mediated effects across 20 spatial radii (1–20 km) at both well- and event-level resolutions; validates every estimate with 50-iteration bootstrap confidence intervals plus three refutation tests; and performs formal sensitivity analyses to unobserved confounding using partial-R2 (Cinelli–Hazlett) and robust stress-tests. Near-field wells (≤ 5 km) amplify earthquake magnitude roughly twenty-fold relative to 20 km wells, exhibiting a mixed mechanical/pressure signature; beyond ∼10 km the effect is almost entirely pressure-mediated. A 5–7 km radius maximises predictive accuracy (event-level R2 ≈ 0.55) while retaining material causal signal. Formal sensitivity results show that at mid- to far-field offsets (≥ 4 km) an unmeasured confounder would need to explain more than half of the residual variance of both treatment and outcome to nullify the effect, stability scores are high at mid- to far-field radii, with pockets of fragility at 3–5 km where causal mechanisms overlap, reinforcing robustness alongside bootstrap and refutation checks. We perform two complementary sensitivity analyses: (i) partial R2 grids à la Cinelli–Hazlett, and (ii) refuter-based robustness checks using DoWhy (DoWhy Library 2024) (unobserved-confounder, bootstrap, and data-subset refuters), summarised with a simple stability score derived from the dispersion of refuted effects. For operators and midstream facilities, induced events create unplanned downtime, regulatory exposure, and public-safety risk. This is especially true for high-rate SWD near mapped faults. Our results therefore address a practical question: where and how to prioritise controls that reduce hazard per barrel curtailed while preserving operations. All code, data, and notebooks are released under an MIT licence at github.com/Project-Geminae/induced-seismicity, enabling immediate replication and extension by operators, service companies, and researchers.