Research & Analysis / Data Interpretation
Scan findings for causal claims that aren’t supported by the data design — and suggest appropriate phrasing.
Why This Prompt Exists
“Correlation does not imply causation” is the most ignored rule in data analysis.
You get:
- business decisions based on confounded correlations that fail when implemented
- reports that say “X drives Y” when the data only shows they move together
- confidence in findings that reverses when you actually run an experiment
- stakeholders demanding causal actions from correlational evidence
- expensive A/B tests that could have been avoided with honest causal assessment
But causal claims require specific evidence:
- randomization (A/B test, RCT) → strongest
- natural experiment (policy change, discontinuity) → strong with assumptions
- longitudinal with controls (time order, confounders) → suggestive
- cross-sectional correlation (single time point) → not causal
- qualitative mechanism (interviews, process tracing) → supports but doesn’t prove
Without auditing, you’ll accidentally claim causation.
This prompt reviews your findings for causal overreach.
The Prompt
Assume the role of a causal inference auditor who catches overclaims. Your task is to review findings and flag inappropriate causal language. Generate: 1. CAUSAL CLAIMS FOUND - List each claim that implies causation (e.g., "drives," "increases," "reduces," "causes," "leads to") - Quote the exact wording 2. DATA DESIGN ASSESSMENT - Study design (RCT / quasi-experiment / longitudinal / cross-sectional) - Controls included (what confounders are accounted for) - Time order established? (Does X precede Y in the data?) 3. APPROPRIATE PHRASING (per claim) - Current (too causal) → replace with → Suggested (correlational) - Example: "Increasing ad spend drives sales" → "Ad spend is associated with higher sales" 4. REMAINING THREATS TO CAUSALITY - Reverse causation (Y causes X instead) - Omitted variable bias (Z causes both X and Y) - Selection bias (sample not representative) - Measurement error (X or Y measured poorly) 5. RECOMMENDATION - Safe to imply causation (with caveats) - Suggest correlation-only language - Redesign study before making causal claims INPUTS: Finding or report excerpt: [PASTE TEXT WITH CLAIMS] Study design description: [E.G., "Survey of 500 customers at one time point"] Data source: [E.G., "CRM data, observational"] Field/discipline norms: [E.G., "Marketing analytics — causal language is common but often wrong"] RULES: - Be conservative — if unsure, flag as correlation-only - Distinguish between statistical and causal language (many people confuse them) - Suggest alternative interpretations for every causal claim - Note when causal language might be defensible with additional assumptions
How To Use It
- Run this on every report before sharing with stakeholders — especially executives.
- Use it on your own writing — we’re all blind to our own causal overreach.
- When reviewing others’ work, run this to prepare feedback.
- For academic papers, this catches reviewer criticisms before they do.
- Keep a list of safe causal verbs (affects, predicts, is associated with) and dangerous ones (drives, causes, powers).
Example Input
Finding or report excerpt:
“Our analysis shows that companies with higher social media engagement drive 23% more revenue. Therefore, we recommend increasing social media posting to boost revenue.”
Study design description:
“Cross-sectional correlation of 200 companies. Revenue and engagement measured in same quarter. No control variables.”
Data source:
“Public financial reports and social media API”
Why It Works
Most people know correlation isn’t causation — then immediately forget when writing conclusions.
This framework improves outcomes by forcing:
- causal claim identification (catches what you actually wrote)
- design assessment (matches claim strength to evidence strength)
- appropriate phrasing (gives you the right words)
- threat specification (what could still be wrong)
- clear recommendation (safe vs. unsafe)
Great causal auditing doesn’t stop you from making claims — it stops you from making wrong ones.
Build Better AI Systems
Subscribe for advanced prompt engineering, AI coding tools, debugging frameworks, and practical strategies for developers and engineers.