Correlation vs causation in health research

01

Association describes a pattern; causation asks about a change

An association means that an exposure and an outcome vary together. The exposure might be a medicine, a behavior or an environmental condition. A causal question asks whether changing that exposure would change the outcome, compared with a meaningful alternative, in a defined population and period.

Start by rewriting a headline as a question: “Compared with what, for whom, and over how long?” Replacing one food with another is a different question from simply adding it. A study that observes people who already chose a behavior is also different from one that assigns the behavior. Precise questions make misleading comparisons easier to spot.

Sources for this section: [1]

02

A worked example of confounding

Imagine a study in which people who buy a fitness app have fewer hospital visits. This is an illustration, not a reported study. Buyers might already have more time to exercise, higher incomes or better access to preventive care. A background factor could influence both buying the app and later health, creating a common-cause explanation.

The right conclusion is not that the app cannot help. It is that the observed comparison does not isolate its contribution. To investigate, ask what differed before purchase and how the researchers measured and handled those differences. A large sample can estimate an association very precisely while leaving this problem unresolved.

Sources for this section: [1]

03

Reverse causation and the limits of adjustment

Now imagine that people who walk less are more likely to receive a diagnosis later. Early, undiagnosed illness might already be limiting their walking. The time of diagnosis is not necessarily the time the disease began. That is why “the habit was measured first” may be insufficient to rule out reverse causation.

An “adjusted” estimate compares groups after statistically accounting for selected measured factors. Its credibility depends on choosing the right factors, measuring them well and using appropriate assumptions. Adjustment cannot automatically recover missing information, and adjusting for a consequence of the exposure may answer a different question. Look for the variable list and sensitivity analyses, not just the word “adjusted.”

Read uncertainty in an effect estimate

Sources for this section: [1] [3]

04

What randomization helps with

Proper random assignment makes treatment groups comparable in expectation for both measured and unmeasured baseline factors. Concealing the upcoming assignment prevents recruitment decisions from undermining that process. Blinding is different: it reduces the influence of knowing the assigned treatment on behavior or assessment.

A trial can still be affected by missing outcomes, departures from the intended treatment or selective reporting. Conversely, an observational study can contribute valuable evidence when a trial is impractical or unethical. Evaluate how a study addresses the causal question rather than treating its design label as a guarantee or a reason for automatic dismissal.

Sources for this section: [2]

05

Five questions to ask of the next health headline

• What exposure, comparator, population and outcome were actually studied?
• Could a common cause explain the association?
• Could early illness have changed the behavior before diagnosis?
• Which factors were adjusted for, and were important data missing?
• Do the effect size, uncertainty and other studies support the wording of the headline?

A small p-value does not decide whether confounding exists. A confidence interval describes uncertainty under the analysis assumptions, not protection against every bias. You do not need to solve the study’s entire statistical model to notice when “linked with” has become “prevents” without a convincing causal argument.

See what a p-value can and cannot establish

Sources for this section: [1] [2] [3]