Interop: certify a DoWhy / EconML estimate against confounding

DoWhy and EconML answer "what is the effect?" causalrl answers a different question — "does this ship / abstain decision survive hidden confounding?" — and returns a certificate.

Library	Answers	Output
DoWhy / EconML	What is the effect?	point estimate / CATE
causalrl	Does the decision survive hidden confounding?	ship / abstain certificate (tipping-Γ)

The bridge is one typed object, PolicyValueContrast: any estimator that can name, per logged unit, the outcome, the logging propensity, and the two target policies' action probabilities can be certified by certify_estimate.

Install the optional dependencies with pip install causalrl[interop].

DoWhy (propensity weighting)

from causalrl import certify_estimate
from causalrl.interop.dowhy import from_dowhy_estimate

# `estimate` is a fitted DoWhy propensity-based CausalEstimate; y / f / z are the logged
# outcome, binary treatment, and (optionally) a measured confounder.
contrast = from_dowhy_estimate(estimate, outcomes=y, treated=f, confounder_bins=z)
cert = certify_estimate(contrast, labels=("treated", "control"))
print(cert.recommendation)   # "act" or "abstain" — read the verdict, not cert.decision
print(cert)                  # human-readable summary with the tipping-Γ

The fitted propensity scores become the logging propensities, so both the marginal-sensitivity-model (MSM) layer and — with a measured confounder_bins or a structural mi_cap — the pivotality layer apply. Honest scope: the certified sensitivity is on the propensity weights, not on DoWhy's point estimator.

EconML (CATE-induced policy)

from causalrl import certify_estimate
from causalrl.interop.econml import from_econml_cate

# `cate` is a fitted EconML CATE estimator; e0 are the logging/behaviour propensities.
contrast = from_econml_cate(cate, X, outcomes=y, treated=f, logging_propensities=e0)
print(certify_estimate(contrast))

This certifies the value of the induced treatment policy (treat iff τ̂(x) > 0) against its complement under the MSM. MSM layer only — the induced policy varies per unit, so it is not a fixed binary arm and the structural pivotality layer does not apply. The object certified is the induced policy's value, not the CATE point estimate itself.

Runnable examples

• examples/interop_dowhy_certify.py
• examples/interop_econml_certify.py

Each runs the full estimate → certificate flow and prints a DecisionCertificate; both exit cleanly with a skip message if the optional dependency is absent.