Moving together versus causing
Two measurements are correlated when they tend to change together. When one is higher, the other is usually higher too (a positive correlation), or usually lower (a negative one). Causation is a stronger claim: that changing one thing actually produces a change in the other.
Headlines often blur the two. “People who drink coffee live longer” describes a pattern; it does not by itself say that coffee is the reason. The pattern is worth noticing, but several different explanations can produce it.
A classic example: heat
Ice cream sales and sunburn cases both tend to be higher in summer. Nobody seriously thinks ice cream causes sunburn. Hot, sunny weather leads more people to buy ice cream and also to spend time in the sun. The weather is a third factor, often called a confounder, that drives both.
Confounders are the most common reason a correlation misleads. They are not always as obvious as the weather.
Hypothetical example: study apps
Imagine a survey in which students who use a particular kind of study app score higher on tests. This is an invented illustration. Possible explanations include:
- The app helps, so using it causes better scores.
- Students who are already motivated are more likely to download and use apps and also more likely to score well. Motivation is the confounder.
- Students who are doing well feel confident and so use the app more. Here the direction is reversed: good scores lead to app use.
- The pattern is a coincidence of a small or unusual sample.
The data alone cannot tell you which of these is true. They only show that scores and app use go together in this group.
Reverse causation
Sometimes the arrow points the other way. If people who exercise more report better mood, it might be that exercise improves mood, or that people in a better mood are more likely to exercise, or both. Untangling this usually requires information about what came first, ideally measured over time.
Coincidence
With enough data, unrelated things will sometimes line up by chance. Over a short period, two numbers that both happen to trend upward can look strongly related even when there is no link at all. A pattern that appears in a small sample and fails to appear in the next is a warning sign. The longer a pattern holds up in different groups and settings, the more seriously it deserves to be taken.
How researchers get closer to cause
No method gives certainty, but some designs are much better at ruling out alternatives.
- Randomized experiments assign people to groups by chance, so that other factors are, on average, spread evenly between them. If the groups then differ, the intervention is the most likely reason. The thinking behind this is related to the ideas in controlled experiments.
- Comparing before and after, and comparing with a group that did not receive the change, helps when randomization is not possible.
- Statistical adjustment can account for known confounders, but it cannot account for those nobody measured.
- Biological or logical plausibility, a dose-response pattern (more exposure, more effect), and repeated results from independent teams all add weight.
Observational studies, where researchers watch what happens without assigning anything, are valuable and often the only ethical option, particularly in health. They are good at finding patterns worth investigating. Their findings usually need to be described with care: “associated with” rather than “causes.”
Reading claims in the news
Phrases like “linked to,” “associated with,” “tied to” and “more likely among” describe correlation. Words like “causes,” “leads to” and “prevents” claim more. When you see a causal claim, a few questions help:
- Was this an experiment with a comparison group, or an observation?
- Could another factor explain both things?
- Could the cause run in the other direction?
- How large and how varied was the sample?
- Has anyone else found the same result?
It also helps to look at the size of an effect. A relationship can be statistically reliable and still small. Understanding the difference between relative and absolute change, which draws on percentage calculations, can stop a modest effect from sounding dramatic. For example, a risk rising from 2 in 1,000 to 3 in 1,000 is a 50 percent relative increase (an extra 1 on top of 2), but the absolute change is one additional case in every thousand people.
A caution in the other direction
Saying “correlation is not causation” should not become a way to dismiss every finding. Some correlations do reflect real causes; the link between smoking and lung disease was established through a combination of strong observational evidence and other kinds of support. The phrase is a reminder to ask for more, not a verdict that no connection exists.
The same reasoning applies to everyday decisions about your own study habits. If a change in routine is followed by better results, it might have helped, but exam difficulty, sleep and many other factors were probably changing too. Treating your own experience as a small, imperfect experiment, and staying a little uncertain about what caused what, is usually the safest approach.