Short answer
Correlation means variables are associated; causation means one affects an outcome. Moving from association to a causal explanation requires evidence that the proposed cause precedes the effect and that plausible alternative explanations have been addressed—not merely that the variables change together. 1
On this page
At a glance
Scroll to compare all columns →
| Question | Correlation | Causation |
|---|---|---|
| What does it describe? | An observed association | A cause-and-effect relationship |
| What must be assessed? | Whether variables are related | Timing, association, and plausible alternatives |
| Does statistical adjustment settle it? | It can refine an association | Measured adjustments alone do not prove causality |
These distinctions follow the research-design criteria and the cautions about adjusted observational findings. 1 2
What each thing is
Correlation describes a pattern in the data. Causation is an explanation of how an exposure affects an outcome. In health statistics, that effect can be probabilistic: the causal question concerns an outcome’s likelihood, rather than whether every exposed person experiences it. 1
Key differences
The key difference is the explanatory burden. Establishing association addresses covariation. A causal assessment also asks whether the proposed cause came first and whether other factors could account for the outcome. NLM presents these as research-design criteria, not as a guarantee that any study meeting a checklist has settled the question. 1
How to tell them apart
Ask what evidence supports the claim beyond the association: Is the sequence established? What competing explanations were examined? Which differences were measured? If the evidence is only an association, describe it that way. This rule has a limit: adjustment may address measured differences while leaving unmeasured ones unresolved. 1 2
Where they overlap
Correlation and causation are not mutually exclusive labels. An association may be part of the evidence for a causal relationship; covariation is one of NLM’s assessment criteria. Association studies also help researchers find patterns and formulate causal hypotheses, even when they cannot establish those hypotheses themselves. 1 2
Edge cases
A statistically significant, adjusted association remains an important boundary case. NCCIH describes a chamomile–mortality association that remained significant for women after adjustment for several known factors. Its caution is that chamomile users could still differ in unmeasured ways, such as broader lifestyle patterns. Adjustment strengthens the analysis without closing every explanatory gap. 2
Why the distinction exists
Researchers often want to know whether an exposure changes health outcomes, not simply whether exposed people have different outcomes. Keeping the terms separate prevents an observed pattern from being mistaken for its explanation and keeps attention on internal validity—the problem of plausible competing explanations. 1
Common misconceptions
Statistical significance does not turn correlation into causation, and controlling for known factors does not control for every possible difference. Conversely, an association that does not establish causation is not useless: it can generate research hypotheses. Nor are randomized controlled trials the only methods for assessing causal relationships. 1 2
Examples
First, the reported chamomile association supports investigating a possible benefit, but it does not establish that chamomile prolonged life. 2 Second, hypothetically, people taking a medication might have lower blood pressure. That pattern alone does not establish a medication effect; researchers would also need to assess timing and plausible alternative explanations. 1