| What the data show | What it supports | Planning use |
|---|---|---|
| Correlation coefficient (r) near 0 | No meaningful correlation between the variables | Treat the measures as unrelated in this dataset and do not search for a cause. |
| Strong correlation, such as r = -0.85 between sleep hours and stress | The variables move together; causation cannot be determined from this data alone | Read it as a lead worth testing, not as a conclusion. |
| A third variable drives both measures, as hot weather does for ice cream sales and drowning incidents | An apparent direct relationship that is not causal | Look for confounders before accepting the link. |
| Exposure comes before the outcome (temporality) | One of Hill’s considerations for a causal reading | Check the timing and the direction of the arrow. |
| Hill’s considerations met in observational data | A more credible causal case, though causation typically still requires experimental data | Act on it cautiously and wait for experimental confirmation where stakes are high. |
The figures and definitions in this table come from an MCAT data-interpretation guide.
Correlation vs causation in medical studies comes down to this: a strong association shows that two variables move together, not that one causes the other. Causal claims need confounding checked, direction established and, typically, experimental data.
Readers following research coverage can find further explainers in our medical news today section.
How to treat a strong correlation in a headline
A strong correlation tells a reader that two measures rise and fall together. It does not establish that one causes the other, and treating it as proof is one of the most common errors in reading data, according to the same guide.
The correlation coefficient, r, ranges from -1 to +1. A value near 0 means no meaningful correlation, while values toward either end mean the variables track each other closely. The sign only gives the direction of the pattern. A negative r, such as -0.85 between sleep hours and stress, means more sleep goes with less stress. The guide stresses that causation cannot be determined from that data alone.
That gap between pattern and cause shapes how a reader should respond to a headline:
- Treat a strong r as a prompt for questions. It marks a relationship worth investigating, which is different from a finding that justifies changing behaviour.
- Ask what else could explain the pattern. The guide’s data-interpretation checklist asks whether alternative explanations exist for the observed relationship, and that question should come before any claim about cause.
- Keep three explanations in view. A correlation can arise because one variable causes the other, because a third variable drives both, or through coincidence.
- Separate chance from cause. Coincidence is only one of those explanations, and a guide on what is a p-value in medical research explains how that part of the question is assessed.
The second of those explanations, a hidden third variable, is the one that most often misleads, so it deserves its own section.
Confounding variables and how they mislead
A confounding variable is a variable that influences two others and so creates the appearance of a direct relationship between them. The guide’s standard example is a strong correlation between ice cream sales and drowning incidents. Neither causes the other. Hot weather drives both.
The same logic applies in medicine. Whenever a study links an exposure to a health outcome, a reader should ask what else travels with the exposure and might produce the outcome independently. The guide is explicit that correlational data does not rule out confounding variables. A tidy statistical pattern therefore cannot, by itself, clear a confounder from the picture.
Reverse causation
Direction is a separate trap. If one variable causes the other, the arrow could point either way, and a correlation is silent about which. Sleep and stress illustrate this. A reader cannot tell from the coefficient alone whether poor sleep raises stress, whether stress shortens sleep, or whether both are true. Reverse causation means the supposed outcome is actually producing the supposed exposure. Timing evidence is the main defence, which is why temporality features in the checklist below.
Observational studies and randomised trials
An observational study records exposures and outcomes as they occur, whereas an experiment assigns the exposure. That difference matters because observational data cannot, on its own, exclude confounders.
So when can an observational study support a causal claim? It can make the case more credible when the evidence holds up against a structured set of considerations and when alternative explanations have been eliminated. Even then, the guide notes that establishing causation typically still requires experimental data and careful analysis. A randomised trial is the standard experimental design, which is why a trial result usually carries more weight than an observational association on the same question.
For readers, the practical split is straightforward:
- Observational finding, no experiment yet. Treat the result as a credible hypothesis at best, and weigh it against the checklist below.
- Observational finding backed by experimental data. The causal reading becomes much stronger, provided the analysis was careful.
- Experimental data alone. It still deserves the usual scrutiny of design and analysis, since the guide calls for careful analysis in every route to causation.
Effect size also shapes interpretation. A guide on relative risk vs absolute risk shows how the strength of an association is reported in health news, which feeds directly into the first Hill consideration.
The Bradford Hill criteria as a reader’s checklist
The Bradford Hill criteria are a framework first proposed by Sir Austin Bradford Hill in 1965 to judge whether an association is likely to be causal. They work best as a set of questions rather than a scorecard, and the guide lists several that a reader can apply directly.
- Temporality. The cause must precede the effect. This directly answers the reverse-causation problem, since an outcome that appeared first cannot have been caused by a later exposure.
- Strength of association. A stronger association is harder to dismiss, though strength alone says nothing about whether a confounder is behind it.
- Consistency across studies. A finding that appears repeatedly in different studies is more credible than one that appears in a single dataset.
- Elimination of alternative explanations. This is the hardest item. It asks whether confounders, reverse direction and coincidence have each been addressed.
Running the earlier examples through the list shows how it works. The ice cream and drowning pattern fails at the last item, because hot weather is an obvious alternative explanation. The sleep and stress correlation of -0.85 is strong, yet nothing in the data settles timing or direction, so it stalls at the first item. Neither case reaches a causal conclusion, and that is the correct outcome.
One limit applies to every use of the checklist. The guide states that the criteria on their own do not settle the matter, because causation typically still requires experimental data and careful analysis. A reader should treat a claim that passes the checklist as well supported but open to confirmation, and a claim that fails it as not yet shown.
Frequently asked questions
What does a correlation coefficient near zero tell me?
A value of r near 0 indicates no meaningful correlation between the variables, according to this data-interpretation guide. The scale itself runs from -1 to +1.
Can the Bradford Hill criteria prove that something causes a disease?
No. They make a causal reading more or less credible, but causation typically still requires experimental data and careful analysis.
Is a hidden third variable the only reason a correlation can be misleading?
No. A correlation can also arise because one variable causes the other, with the direction unclear, or through coincidence. A third variable driving both is only one of the explanations to rule out.
Who proposed the Bradford Hill criteria?
Sir Austin Bradford Hill proposed the framework in 1965. Temporality, strength of association, consistency across studies and elimination of alternative explanations are among its considerations.
