The Standard of Evidence
Regulatory proceedings in financial services, healthcare, environmental liability, and employment law increasingly distinguish between statistical association and causal attribution. A correlation between an action and an adverse outcome is not sufficient to establish liability — the question is whether the action caused the outcome, and whether it would have occurred regardless.
This is Rung 3 in legal language. “But for” causation — would the harm have occurred but for the defendant’s action? — is a counterfactual question. Answering it requires a structural causal model, not a regression coefficient. The regression coefficient tells you the average association. The counterfactual question is about this specific harm, in this specific case, under the specific conditions that obtained.
Where Causal Models Apply
| Proceeding type | Causal question |
|---|---|
| Financial mis-selling | Would this customer have purchased this product had the disclosure been compliant? |
| Environmental liability | Did this discharge cause this ecological harm, or was the baseline already compromised? |
| Employment discrimination | Did protected characteristic status causally affect the promotion decision, after adjusting for performance? |
| Insurance coverage disputes | Was the covered peril the proximate cause, or was it an excluded condition that was the actual cause? |
What the Model Produces
A Structural Causal Model produces a probability distribution over the counterfactual outcome — P(harm would have occurred | action not taken, everything else held constant). This is a quantified causal claim with explicit assumptions, testable against the observed data, and defensible in cross-examination because every assumption is explicit in the graph.
The model also produces the attribution fraction — what proportion of the observed harm is attributable to the defendant’s action versus background causes. This is the formal basis for proportional liability calculations. See also the Insurance Attribution and Criminal Causation cases.
If your regulatory exposure includes a causal question — and most do — the model is the answer the correlation cannot provide.