The basic idea
Two things are correlated when they tend to change together. When one is higher, the other is usually higher too, or perhaps usually lower. That is a pattern in the data. Causation is a stronger claim: that changing one thing actually produces a change in the other.
The distinction matters because the first is easy to observe and the second is hard to establish. Headlines often slide from “people who do X tend to have Y” to “X leads to Y,” and that step needs more support than the pattern alone provides.
An everyday example
Imagine, in a purely hypothetical town, that ice cream sales and sunburn cases both rise and fall together through the year. Someone might suggest that ice cream causes sunburn, or that sunburn makes people crave ice cream. Neither is likely. A more plausible explanation is a third factor: sunny, hot weather. Hot days send people outdoors, which means more ice cream and more sun exposure.
This third factor is called a confounder, or a lurking variable. It is connected to both things and produces a pattern between them even though neither directly affects the other.
Four reasons a pattern can mislead
A hidden third factor
As above, something else may drive both. In real research, age, income, general health and location are common confounders. For example, if people who drink a certain beverage seem healthier, it could be that people who already take good care of their health are more likely to choose it.
Reverse direction
Sometimes the causation runs the opposite way from what we assume. Suppose a study found that people who use a particular app report lower stress. It might be that the app helps, but it might also be that less-stressed people have the energy to use it. Without more information, both stories fit.
Coincidence
If you look at enough pairs of measurements, some will line up by chance. Over a short stretch of time, two unrelated trends can move together simply because both happen to be rising or falling. The more comparisons that are made, the more likely it is that an impressive-looking match turns up for no reason at all. This is one reason a surprising pattern deserves a second look before being believed.
Selection effects
How data were collected can create a pattern. If a survey only reaches people who respond online, or a study only includes people who stayed in a program, the group may not resemble everyone, and the pattern may be an artifact of who was included.
What it takes to support a causal claim
No single test settles it, but several kinds of evidence make causation more plausible:
- Controlled experiments. When people or items are assigned to a group by chance, such as through randomization, the groups are usually similar apart from the thing being tested. Differences between them are then more likely to be due to that thing. This is the logic behind the scientific method applied in research.
- Timing. The supposed cause should come before the effect.
- A plausible mechanism. It helps if there is a reasonable explanation for how one thing could affect the other, though a missing explanation does not prove no effect exists.
- Consistency. Similar results appearing across different groups, places and methods are more convincing than a single result.
- Dose and response. In some cases, a larger amount of the exposure going with a larger effect adds support, though this is not always expected.
- Ruling out alternatives. Researchers try to account for confounders, either by design or with statistical adjustments, though adjustments can never be sure to capture everything.
Sometimes experiments are impractical or unethical, so researchers rely on careful observational studies. Those can be valuable, and many important findings in public health came from them, but they usually give a more cautious level of confidence.
Questions to ask when you see a headline
- What exactly was measured, and was it measured the same way in everyone?
- Is this an experiment or an observation of what people already do?
- Could a third factor explain both?
- Could it run in the other direction?
- How big is the effect, and how many people were studied?
- Is the claim stated cautiously (“is associated with”) or boldly (“causes”)?
Words matter here. Phrases such as “linked to,” “associated with” and “tends to go along with” describe correlation. They are not mistakes, but they do not mean “causes,” and a careful reader keeps the two apart.
A balanced view
It would be wrong to conclude that correlations are useless. They are often the first clue that something deserves investigation, and in some cases they point toward real causes. They can also be useful for prediction even when the cause is unknown: if a pattern reliably holds, it can help forecast, even if it cannot explain.
The mistake is treating the pattern as an explanation. “Correlation is not causation” is not a way to dismiss every finding; it is a reminder to ask what else might be going on before deciding what to believe or do.