"Customers who use our app's wishlist feature spend more. Let's push everyone to use the wishlist!" This kind of conclusion is one of the most common mistakes in business analysis. Two things moving together does not mean one causes the other.
What correlation tells you
Correlation measures how strongly two variables move together in a straight-line way, from โ1 to +1. It says nothing about why.
df[["wishlist_items", "annual_spend"]].corr()Four reasons for correlation without causation
1. A confounding factor
A third factor drives both. Highly engaged customers both use the wishlist and spend more. Engagement is the confounder; forcing casual users to create wishlists may not change their spending at all.
Business example: stores with more staff have higher sales. But bigger, busier stores get more staff and have more sales. Store size drives both.
2. Reverse causation
The arrow points the other way. "Customers who contact support churn more" might mean support causes churn, but more likely customers with problems contact support, and the problems cause churn.
3. Coincidence
With enough variables, some will correlate by pure chance, especially over short periods or small samples.
4. Selection bias
The data only includes a particular group. If you only analyse customers who completed a survey, any pattern may reflect who chose to respond.
How to get closer to causation
- Randomised experiments (A/B tests): the strongest evidence. Randomly encourage some users to use the wishlist and compare. See our A/B testing guide.
- Control for confounders: compare like with like (stores of similar size, customers with similar activity), or include them in a regression.
- Look at timing: a cause must come before the effect.
- Look for a mechanism: is there a plausible reason one would cause the other?
- Natural experiments: a change that affected some regions but not others for reasons unrelated to the outcome.
How to write about it
Use careful language in reports:
- Correlation only: "Customers who use the wishlist tend to spend more."
- With an experiment: "Showing the wishlist prompt increased spending in a randomised test."
That small change in wording protects decision-makers from expensive mistakes, and interviewers notice it.
Always plot it
Correlation coefficients can hide curved relationships, outliers and clusters. Look at a scatter plot before quoting any coefficient. Practise with our relationships and heatmaps sets.
Practise what you just read
Reading helps; writing the code yourself is what sticks. Start here: