“Correlation does not equal causation” is one of the most repeated phrases in science communication. It is useful, but incomplete. Correlations are often how important questions begin. The problem comes when a relationship is treated as a causal explanation before alternative explanations have been considered.
What correlation tells you
If two variables are correlated, they vary together in some systematic way. More of A may be associated with more of B, or more of A may be associated with less of B.
That can be practically useful even when the mechanism is uncertain. Epidemiology often begins with patterns in populations. The question is what those patterns justify us saying next.
What causation adds
A causal claim says that changing A would change B, all else being appropriately accounted for. That is a much stronger statement.
To move from association to causation, researchers try to rule out competing explanations and look for evidence consistent with a causal process.
Confounding: the third-variable problem
Suppose people who perform a certain behavior have lower disease risk. That behavior may be protective. It may also be associated with income, healthcare access, diet, age, smoking, education, or other factors that influence the outcome.
Researchers can adjust statistically for measured confounders, but adjustment is only as good as the variables and models available. Unmeasured confounding can remain.
Reverse causation: the arrow may point the other way
A study might find that people who sleep less have worse health. Poor health could contribute to shorter sleep, shorter sleep could contribute to poor health, or both could occur together.
Timing helps. Prospective studies that measure an exposure before an outcome can strengthen interpretation, but they still do not automatically prove cause.
Randomization is powerful because it changes the comparison
When participants are randomly assigned to an intervention, the groups should differ less systematically at baseline, making it easier to attribute later differences to the intervention. That is part of why randomized trials are so valuable for many treatment and behavior questions.
Randomized trials also have limits: they can be expensive, short, impractical, or unethical for some exposures. We cannot randomly assign people to decades of smoking to test whether smoking causes cancer.
Causal reasoning often uses a mosaic of evidence
Researchers may combine temporal sequence, dose-response patterns, mechanistic evidence, natural experiments, consistency across populations, negative controls, randomized trials where feasible, and other approaches.
This is why “observational” does not mean “worthless,” and “randomized” does not mean “perfect.” The design changes the kinds of bias we worry about.
Language should match design
NIH guidance for communicating health research recommends explicitly stating whether evidence shows association or causation. That simple discipline prevents a large share of headline inflation.
| Study result | Safer language | Language that may overreach |
|---|---|---|
| Observational association | “X was associated with Y.” | “X caused Y.” |
| Randomized intervention effect | “Participants assigned to X had Y outcome over the study period.” | “X works for everyone.” |
| Mechanistic/lab finding | “X affected pathway Y under these conditions.” | “X prevents disease Y in people.” |
What to do when you see a causal headline
- Find the original study.
- Identify the design.
- Ask whether the exposure happened before the outcome.
- Look for major confounders and how they were handled.
- Check whether other evidence points in the same direction.
- See whether the size of the claim matches the size of the result.
The goal is not to refuse causal conclusions. It is to earn them.
Causal confidence often comes from triangulation
Many health questions cannot be answered by one ideal randomized trial. Researchers may instead look for convergence across designs: prospective cohorts, natural experiments, randomized trials where feasible, mechanistic evidence, dose-response patterns, and findings in different populations.
Each design has different weaknesses. When several methods with different sources of bias point in a similar direction, the overall causal case can become more persuasive. When they disagree, the disagreement itself is useful because it shows where the uncertainty still lives.
For a full paper-reading workflow, see how to read a health study. To understand why replication and synthesis matter, see why one study rarely settles a health question.
