Correlation describes a pattern between measurements. Causation means that changing one thing produces a change in another. Two things can move together because one influences the other, because something else influences both, or because the apparent connection is misleading.
The difference matters whenever someone turns “these things are linked” into “do this and you will get that result.” A pattern can be useful evidence without being a complete explanation.
The question that separates them
Imagine a shop introduces a new window display and sales rise the following week. The owner says the display caused the increase.
That may be true. But perhaps the same week included a holiday, a local event or a discount. A before-and-after comparison records the change; it does not automatically isolate what produced it.
The causal question is: “What would sales have been during that same period without the new display?” You cannot directly observe both versions of the same week, so a good investigation needs a credible comparison.
Three explanations to consider before celebrating
| Possible explanation | Question to ask |
|---|---|
| One thing influences the other. | What evidence shows the proposed cause changed the outcome? |
| A shared factor influences both. | What else changed at the same time? |
| The pattern reflects chance or selection. | Does it remain when we examine more suitable data? |
The shop example is fictional. Its purpose is to keep several explanations open until the evidence distinguishes them.
A shared cause can make unrelated actions look connected
Suppose a café records more cold-drink sales and more outdoor-seat use on warm days. Buying a cold drink is not necessarily what makes people sit outside. Temperature can influence both choices.
A factor affecting the relationship you are studying may be called a confounder. Simply listing one is not proof that it explains the whole pattern, but it tells you what needs investigation.
The Australian Bureau of Statistics explains the distinction between correlation and cause and effect. It cautions that a correlation alone cannot establish that one measurement caused another to change.
The direction can run the other way
Imagine a company notices that its busiest teams send the most messages. It decides that sending more messages will make every team busier and more productive.
Perhaps the workload creates the messages: more tasks require more coordination. Increasing messages without increasing useful work may simply create interruptions.
Ask what happened first and what mechanism could connect the two. Timing is necessary for a proposed cause to produce a later outcome, but timing alone is not sufficient evidence. “After” does not automatically mean “because of.”
The dataset can make a relationship look stronger than it is
Look at how the observations were selected. A report that includes only successful customers may miss everyone who tried the same approach and saw no benefit. A short period can also hide a longer trend or a seasonal cycle.
For example, a shop’s weekly sales and advertising spending might both rise during a festive season. A graph of those weeks can look persuasive while leaving the season itself out of the story.
Ask whether the pattern survives different reasonable time periods, whether unusual observations dominate it, and whether the measurements mean the same thing across the comparison.
Why an impressive graph cannot finish the argument
A scatter plot can reveal patterns, unusual observations and possible relationships. NIST’s statistical handbook makes clear that a scatter plot cannot by itself prove cause and effect.
A correlation coefficient is also a summary, not a mechanism. Common measures such as Pearson’s correlation describe linear association. A value close to zero does not rule out every possible curved relationship.
Read the axes and examine the underlying observations where possible. A chart can help you ask better questions; it cannot supply missing study design.
What makes causal evidence stronger?
A well-designed experiment deliberately changes an input and compares outcomes. Random assignment can help make comparison groups similar on average, reducing systematic differences in who receives the change.
JMP’s statistics guide discusses controlled experiments and randomisation as ways to investigate causal relationships. The design, measurement and execution still matter; the label “experiment” is not a guarantee.
Sometimes experiments are impractical or unethical. Researchers can use observational designs and other evidence to investigate causal explanations, but their assumptions and limitations need scrutiny. Correlation alone being insufficient does not mean causation is unknowable.
A practical business example
Return to the shop display. A useful small investigation could compare a planned change across comparable locations or periods while recording prices, promotions and customer traffic. If feasible, randomise which eligible locations receive the display first.
Decide in advance what outcome matters: total sales, profit, conversion rate or something else. A display that increases revenue while sharply reducing margin may not deliver the benefit the owner wanted.
A small trial gives evidence for the conditions tested. Record uncertainty, check for other changes and avoid generalising one encouraging week to every future location.
Five questions for the next “X causes Y” headline
- Was this an observation, an experiment or a review of several studies?
- Did the proposed cause come before the outcome?
- Could another factor influence both?
- What comparison supports the claim?
- How far does the evidence justify applying the result?
If the evidence only establishes an association, use wording such as “was associated with.” If a study provides causal evidence, describe its actual scope rather than stretching its conclusion.
Common questions
Does correlation mean the relationship is useless?
No. It can help identify patterns, support prediction and guide further investigation. Prediction and explaining what will happen after an intervention are different tasks.
Does a strong correlation prove causation?
No. A strong pattern can still reflect a shared influence or another explanation.
Can a real cause have a weak overall correlation?
Yes. Other influences, measurement problems or the shape of the relationship can obscure a simple overall pattern.
Related: Use our root-cause investigation guide to test explanations, and percentage-points explainer to keep numerical comparisons clear.
