Ice cream sales and drowning incidents rise at the same time of year. Nobody seriously believes ice cream causes drowning — summer heat drives both. That example is famous precisely because it is obvious. Most of the correlations you meet at work are not obvious at all, and that is where expensive decisions go wrong.
What correlation actually measures
Correlation describes how two variables move relative to each other. When one goes up and the other tends to go up as well, the correlation is positive. When one rises while the other falls, it is negative. The strength is usually summarised as a number between -1 and 1, where values near zero mean the two barely track each other at all.
Two properties of that number matter more than most people realise. First, it is a description of a pattern, not an explanation of it. Second, the common measure only captures straight-line relationships. A variable that helps up to a point and then hurts — think of study hours versus test performance once exhaustion sets in — can produce a correlation close to zero even though the two are strongly related.
Four reasons two things move together
Whenever you see a correlation, there are at least four explanations competing for your attention, and only one of them is the story you were hoping for.
| Explanation | What is happening | Everyday example |
|---|---|---|
| Direct causation | A really does change B | Raising the price reduces units sold |
| Reverse causation | B changes A, not the other way round | Support tickets and unhappy users |
| Confounding | A third factor drives both | Heat drives ice cream and swimming |
| Coincidence | Random noise in a small sample | Two unrelated trends over five months |
The confounding case is the one that quietly ruins business analysis. A company notices that customers who use its mobile app churn less, and concludes the app causes loyalty. It may be the opposite: already-loyal customers are the ones who bother installing the app. Pouring budget into app installs would then produce far less than the chart promised.
Selection effects: the sample decides the answer
A related trap appears when the data you have is not the data you need. If a training programme is voluntary, the people who enrol are usually more motivated than average. Comparing their results to non-participants measures motivation as much as training. This is called selection bias, and no amount of statistical polish removes it after the fact.
Survivorship bias is its close cousin. Studying only the companies that still exist, the students who finished the course or the products that stayed on the shelf hides every case that failed — which is exactly the group that would tell you what does not work.
How analysts test for real causal links
You cannot prove causation from a chart, but there are practical ways to get closer to it.
- Run an experiment. Randomly assign who receives the change and who does not. Randomisation is what makes the two groups comparable on everything you did not measure. This is the logic behind A/B tests and clinical trials alike.
- Check the direction of time. A cause must precede its effect. If the supposed effect started moving first, the story needs rewriting.
- Look for a dose-response pattern. If more of the cause produces more of the effect in a consistent way, the case strengthens.
- Search for a plausible mechanism. Being able to describe how A would produce B makes accidental correlations easier to dismiss.
- Try to break it. Split the data by region, by season, by customer segment. A real effect usually survives; a coincidence often disappears.
Reading charts with healthy suspicion
Presentation choices can make a weak relationship look overwhelming. A few habits protect you:
- Check the axes. Dual axes with independent scales can make almost any two lines appear to move together.
- Check the time window. Charts that start and end at conveniently chosen dates are a classic way to manufacture a trend.
- Ask about the sample size. With few observations, striking patterns appear by chance all the time.
- Ask what was left out. If a comparison group exists, it should be on the chart.
- Watch for aggregation. A pattern visible in totals can reverse when the data is split into subgroups — a well-documented effect that has embarrassed many analyses.
Why this matters beyond statistics class
Correlations are genuinely useful. Prediction does not require causation: if umbrella sales reliably rise before a certain kind of weather, that signal has value even without a causal story. Trouble starts when you act as if changing the predictor will change the outcome. Forcing umbrella sales up will not make it rain.
So the practical question to ask in any meeting is simple: are we using this relationship to forecast, or to intervene? Forecasting tolerates correlation. Intervening demands evidence of causation. Teams that keep those two purposes separate make far fewer costly mistakes.
Conclusion
Learning to pause at a convincing chart and run through the alternatives — reverse causation, confounding, selection, coincidence — is one of the highest-return habits in data work. It costs a few minutes and saves entire budgets.
If you want to go deeper into how data is collected, analysed and presented, the free Data Science, Business Intelligence and Statistics courses available on Cursa are a natural next step for building that critical eye.
