Causal Claims without Overreach

Mechanism (causal) — when evidence justifies it

Use only with randomised evidence, strong quasi-experimental identification, or triangulation that reasonably closes major bias pathways.

• X caused an increase in Y by … (use when the design rules out confounding and temporal precedence is clear)
• The intervention produced a … change in Y
• Mediated pathway: X → M → Y; the indirect effect accounted for …%
• Removing X decreased Y by …, consistent with a causal effect
• We identify a causal effect of X on Y using … (instrument, RD, diff-in-diff, IV strength checks)
• Sensitivity analyses suggest the effect is robust to unmeasured confounding up to Γ = …

Association (non-causal) — observational stance

When design or data do not justify causal verbs, keep claims associative.

• X was associated with higher odds of Y
• We observed a relationship between X and Y after adjusting for Z
• X correlated with Y (r = …); causality cannot be inferred
• Exposure to X coincided with changes in Y over time; residual confounding may remain
• The association attenuated after controlling for …
• Findings are hypothesis-generating and require prospective confirmation

Prediction/Prognosis — forward-looking language

Predictive models optimise accuracy, not identification. Avoid causal verbs entirely.

• Model M predicts Y with AUROC = … / RMSE = … on held-out data
• Predictors A, B, C contributed most to performance (SHAP …)
• The model estimates risk; it does not imply that modifying A reduces Y
• Calibration across deciles indicates over/under-prediction in … strata
• We report drift monitoring and plan external validation

Hedging strength — calibrated modality

• may, might, could (exploratory/weak)
• appears to, is consistent with, suggests (moderate)
• indicates, supports (stronger, still non-causal)
• causes, leads to, effects of (causal; use only when justified)

Edit pattern: replace over-assertive verbs with calibrated alternatives while preserving the core claim.

Phrases to avoid or rewrite

• X proves Y → X supports Y
• X leads to Y (observational) → X is associated with Y
• Because X occurred, Y happened → Y occurred alongside X; alternative explanations remain
• Intervention X will reduce Y (predictive model) → Model identifies individuals at higher risk of Y
• No effect (under-powered null) → No statistically significant difference was detected (power …)

Response-to-reviewers snippets

• We revised causal verbs to associative language where design limitations preclude identification (lines …)
• We added a statement distinguishing prediction from causation and clarified that changing predictor A may not alter outcome Y (lines …)
• We expanded the methods to justify causal terminology: assignment mechanism, temporal ordering, robustness checks (lines …)
• We replaced “proves/causes” with “supports/associated with” and added sensitivity analyses (lines …)

Pre-submission checklist