DoWhy and refutation tests

DoWhy structures causal analysis in steps: model (assumptions), identify (estimand), estimate (number), and refute (stress-test). That workflow prevents jumping from data to a coefficient without naming what must be true.

Refutation tests include placebo treatment (shuffle treatment labels — effect should vanish), random common cause (add noise covariate — effect should be stable), and subset validation.

Why refutation matters

A result that fails refutation is a warning, not a formality. A result that passes is not guaranteed correct — but it survived a simple falsification attempt. Combine refutation with domain review and alternative estimators.