causal-vs-predictive

causal-vs-predictive

Overview

A model can be an excellent predictor of an outcome while being silent on what causes it — and the two get conflated constantly, because the same word ("driver," "important feature") is used for both. This skill is the check that catches "X reduces/causes/drives Y, confirmed" language when only an associational comparison actually ran.

When to Use

• Writing up feature importance, SHAP findings, or a bivariate/EDA relationship in /ds-explain or /ds-explore.
• Writing a recommendation in /ds-report that implies intervening on a feature (targeting a segment for a changed offer, pushing customers toward an option, recommending a policy change) — as opposed to ranking or scoring individuals using that feature, which doesn't need this check.
• NOT for: the model's predictive validity itself (that's /ds-evaluate) — a model can be a perfectly valid predictor and still say nothing about causal effects. Using a score to prioritize outreach or flag transactions needs the score to be predictively valid, not causally identified.

Core Process

1. For every feature described as a "driver" of the target, ask which of two claims is actually being made:
   • Predictive/ranking claim: "this feature helps distinguish who is more likely to have outcome Y" — supported directly by permutation importance/SHAP/correlation. No further check needed.
   • Causal/interventional claim: "changing this feature (for a given individual) would change their outcome" — this is what "reduces," "causes," "drives" (in the active sense), or a recommendation to intervene actually asserts.
2. If it's a causal claim, ask what would have to be true for the correlation to reflect a real causal effect rather than a confound. The single most common confound to check explicitly: self-selection — did the subject choose this feature's value themselves (a contract length, a plan tier, a loyalty program, an opt-in)? If so, the feature may be a symptom of the outcome's underlying propensity, not a cause of it.
3. If no causal identification strategy is available or stated (a randomized experiment, a natural experiment, an instrument, or at minimum an explicit, defensible argument for why the obvious confound doesn't apply), reword the claim as an association: "X is associated with Y," not "X reduces/causes/ drives Y."
4. Effect size is not evidence either way — a large, clean, monotonic correlation is exactly as consistent with a large confound as with a large causal effect. Don't let a dramatic effect size substitute for the identification argument.