regression-and-forecasting-reviewlisted

# Skill: regression-and-forecasting-review > **Invoked by:** `applied-statistician` (primary). Escalate model-as-product / heavy ML feature-engineering to `ravenclaude-core/data-engineer`; financial-model structure to `finance` (when installed). > > **When to invoke:** "what drives <outcome>?"; "build/forecast <metric> for next quarter"; "review this regression — is it sound?" > > **Output:** the model-family choice + rationale, the assumption checks for that family, the intervals to report, and a pitfall screen (leakage / overfitting / causal overreach). ## Regression 1. **Pick the family by outcome type** ([`../../knowledge/test-selection-decision-tree.md`](../../knowledge/test-selection-decision-tree.md)): continuous → **OLS**; binary → **logistic GLM**; count → **Poisson / negative-binomial GLM**. 2. **Check the assumptions that matter:** linearity (residual-vs-fitted), independence, homoscedasticity (else robust SEs), no severe multicollinearity (VIF), influential points (Cook's distance). Normality of *residuals* matters for small-sample inference, not for the point estimates. 3. **Report coefficients with CIs**, and translate to plain language ("each +1 in X is associated with +β in Y, 95% CI […]"). 4. **Causal overreach guard:** a regression coefficient is **association, not cause**, unless the data came from a randomized/causal design. If the user says "drives/causes," route to [`../../knowledge/causal-inference-primer.md`](../../knowledge/causal-inference-primer.