The Correlation vs Causation Stress-Tester: Pressure-Test a Data-Driven Claim Before You Act On It
Why this prompt matters
Teams that act on unexamined correlations routinely misallocate budget and headcount for months before the real cause (a channel mix shift, a concurrent feature launch, simple seasonality) surfaces — at which point the reversal costs both the wasted spend and the credibility of whoever championed the original claim.
What we use it for
Your marketing team just told leadership that a new onboarding email sequence increased 30-day retention by 12%, based on a before/after comparison, and the company is about to reallocate a team's entire budget on the strength of that one chart.
Prompt
Role: You are a skeptical senior data scientist whose job is to stress-test causal claims before they influence business decisions. You are not trying to win an argument — you are trying to find out if the claim is actually true. Context: - The claim being made: [DESCRIBE THE CLAIM, e.g. "Our new onboarding email sequence increased 30-day retention by 12%"] - The data behind it: [DESCRIBE THE COMPARISON, e.g. "We compared the 30-day retention of users who received the sequence after its Sept 1 launch against users who signed up before the launch"] - What's at stake: [DESCRIBE THE DECISION RIDING ON THIS, e.g. "Leadership wants to double the lifecycle marketing team's budget and headcount based on this result"] Task: 1. List every plausible alternative explanation for the observed pattern that does not involve the claimed cause (consider confounders, selection effects, seasonality, concurrent changes, regression to the mean, reverse causation, and measurement-window asymmetry). 2. For each alternative explanation, state what evidence would confirm or rule it out. 3. Identify what data the team already has that could test these alternatives, and what data is missing. 4. Propose the minimum additional analysis or experiment needed to meaningfully raise or lower confidence in the causal claim — prefer the cheapest diagnostic that would resolve most of the uncertainty over a full randomized controlled trial. 5. Give a final verdict: "Act on this now," "Investigate further before acting," or "Do not act — too confounded," with two to three sentences of reasoning. Constraints: - Do not accept the causal claim at face value, even if the correlation looks statistically strong. - Flag explicitly if the claim is effectively unfalsifiable given the data described. - Keep the tone rigorous but practical — this is for a business audience making a resourcing decision, not a statistics journal. - Do not pad the analysis with generic caveats that apply to literally any claim; every alternative explanation must be specific to the scenario described. Output Format: - Section 1: Alternative Explanations (numbered list, one line each) - Section 2: Distinguishing Evidence (a table: alternative explanation | what would confirm or rule it out) - Section 3: Data Gap Assessment (what the team already has vs. what's missing) - Section 4: Recommended Next Step (one clear, specific action) - Section 5: Verdict (one of the three options above, with justification)
Result
Section 1: Alternative Explanations
- Seasonality — the post-launch cohort signed up in September, which may systematically differ in intent or tenure from the January–August cohort used for comparison.
- Concurrent product changes — any other onboarding, pricing, or feature changes shipped around September 1 could be driving the retention lift instead of the email sequence.
- Marketing channel mix shift — if the acquisition channel mix changed around the launch date (for example, a shift from paid social toward organic or referral traffic), the new cohort may simply be higher-intent regardless of the emails.
- Regression to the mean — if pre-launch retention was unusually depressed in a specific recent window, the "lift" may partly reflect a natural reversion rather than the sequence's effect.
- Measurement-window asymmetry — the pre-launch cohort has eight months of mature, fully-observed data while the post-launch cohort is recent; if 30-day retention is still being calculated on partially-censored recent signups, the comparison is not apples-to-apples.
- Reverse causation via self-selection — if completion of the email sequence is optional or engagement-gated, only already more-engaged users may be finishing it, which would inflate the retention number for "recipients" without the emails causing anything.
Section 2: Distinguishing Evidence
| Alternative Explanation | What Would Confirm or Rule It Out |
|---|---|
| Seasonality | Compare September signups from the prior year (if the product existed then) against this September's cohort, holding the email sequence constant |
| Concurrent changes | Pull the full changelog for August 25 – September 5 and check for any other shipped change affecting new users |
| Channel mix shift | Break down both cohorts by acquisition channel and compare retention within matched channels |
| Regression to the mean | Plot weekly retention for the 12 months before launch to check whether August was an unusual trough |
| Measurement asymmetry | Confirm every user in the post-launch cohort has a fully elapsed 30-day window as of the analysis date |
| Self-selection | Check whether email sequence completion is optional; if so, compare retention of all assigned users (not just completers) against the baseline |
Section 3: Data Gap Assessment
The team already has: signup dates, retention flags at 30 days, and email send/open logs. Missing: a changelog cross-reference for the launch window, channel attribution broken out by cohort, and completion status for the email sequence (versus mere assignment to receive it).
Section 4: Recommended Next Step
Run a two-week holdback test: randomly assign 15% of new signups to skip the email sequence while keeping everything else identical, then compare 30-day retention between the holdback and recipient groups within the same calendar window. This isolates the sequence's effect from seasonality, channel mix, and concurrent changes in a fraction of the time a full RCT would take, since it only needs to resolve the sign and rough magnitude of the effect, not a precise estimate.
Section 5: Verdict
Investigate further before acting. The 12-point retention gap is large enough to be worth pursuing, but at least three of the six alternative explanations (channel mix shift, concurrent changes, and self-selection via completion) are both plausible and cheap to check. Committing to double a team's headcount on an uncontrolled before/after comparison risks a costly reversal if the holdback test shows a smaller or null effect.
Most data-driven decisions fail not because the data was wrong, but because nobody seriously tried to disprove the first explanation that fit. A retention number goes up after a launch, and the launch gets the credit, even though three or four other things changed in the same window. This prompt is built to interrupt that pattern before a team commits real budget to an unverified causal story.
Why this uses a forced-alternatives structure
The prompt does not ask an AI model to simply “analyze this data” — that framing tends to produce confirmation of whatever claim is already in the prompt, because the model has no built-in incentive to argue against the premise. Instead, the Task section explicitly instructs the model to generate alternative explanations first, before it is allowed to evaluate the original claim. This ordering matters: once a model (or a person) has anchored on a causal story, generating counter-evidence becomes much harder than generating it from a blank slate.
Why the output format uses a distinguishing-evidence table, not a conclusion
A list of alternative explanations is only useful if you know how to tell them apart. That's why Section 2 of the output format pairs every alternative explanation with a specific, checkable piece of evidence — not a vague “investigate further,” but a concrete comparison (channel-level retention, a changelog cross-reference, a prior-year cohort). This turns a skeptical reaction into an actionable diagnostic plan, which is the difference between a prompt that produces anxiety and one that produces a next step.
Why it recommends the cheapest test, not the most rigorous one
A full randomized controlled trial is the gold standard for causal claims, but most teams don't have the time or willingness to run one for every metric that moves. The Constraints section explicitly tells the model to prefer the cheapest diagnostic that would meaningfully move confidence, which is why the example output recommends a short holdback test instead of a lengthy RCT — the goal is resolving the sign and rough size of an effect fast, not publishing a peer-reviewed paper.
How to use this well
Fill in the three bracketed fields with as much specificity as you can — vague inputs produce vague alternative explanations. The prompt works best when you already suspect the causal story might be too convenient; running it on a claim you're fully confident in will mostly just confirm what you already believe, which isn't the point. Use it before a resourcing decision gets made, not after — the whole value is in catching the mistake while it's still cheap to fix.