How to tell if something is correlation or causation?
Correlation means two variables move together; causation means one directly changes the other. To tell them apart, check the study design, ask whether chance, reverse causality or a confounder could explain the link, and remember that only controlled or well-designed studies can establish cause and effect.
Two things rising and falling together is one of the most seductive patterns in statistics. The Australian Bureau of Statistics gives the classic warning: a correlation between variables does not automatically mean that a change in one causes the change in the other. The correlation coefficient, usually written as r, only describes the size and direction of a relationship, on a scale from +1.0 to –1.0. It says nothing about cause and effect.
[1]What a correlation actually tells you
A positive r means the variables move in tandem. A negative r means they move in opposite directions. A value of zero indicates no relationship. But the ABS notes a limitation worth remembering: the coefficient usually measures linear relationships. A tradesperson charging a flat hourly rate shows a straight-line link between hours and income; one who charges an initial call-out fee plus an hourly fee that progressively decreases over a long job may show a coefficient close to zero even though a real relationship exists.
[1]Three reasons a correlation misleads
- Chance. The relationship is coincidental. A Mayo Clinic-authored orthopaedic research paper notes that rising incidence of breast cancer can correlate strongly with rising rates of joint replacement surgery simply because both climb over time.
- Reverse causality. The direction of the arrow is backwards — the supposed outcome may be driving the supposed cause.
- Confounding. A third factor influences both variables. The ABS gives the sunscreen-and-ice-cream example: both rise in hot weather, so the season, not any link between them, explains the pattern.
A worked comparison
Imagine a headline saying a hospital's newer hip implants are associated with fewer dislocations. Before accepting a causal story, ask the questions the orthopaedic paper raises. Were the groups comparable? The paper describes a registry scenario where an unmeasured confounder — prior spine disease — influenced both which implant was chosen and the dislocation risk, masking or distorting the true relationship. Compare that with a controlled study, which the ABS describes as the most effective way of establishing causality: split a population into two comparable groups, give each a different treatment, and compare outcomes. If the groups differ noticeably, the different treatments may have caused the different outcomes. The catch is that controlled studies are not always ethical or practical, which is why observational studies exist — and why their correlations need careful interpretation rather than automatic causal reading.
[1][2]Questions to ask of any causal claim
- What was the study design? The orthopaedic paper notes that causal direction is typically demonstrated in experimental and longitudinal studies, not one-off observational snapshots.
- Could a third factor explain both variables, the way season drives sunscreen and ice-cream sales?
- Could the arrow point the other way?
- Could the pattern be coincidence, especially when both variables trend over time?
- If a randomised trial was not possible, did the researchers use methods to address confounding, such as propensity score calibration or instrumental variable approaches?
- Could the relationship be non-linear, hiding from a simple correlation coefficient?
One nuance cuts the other way too. The orthopaedic paper points out that correlation is necessary for a causal relationship, even though it is never sufficient. So a correlation is not evidence of causation, but its absence is evidence against one.
[2]A practical next step: next time you see a causal claim built on survey or registry data, write down the three rival explanations — chance, reverse causality and confounding — and check whether the study addressed any of them. If it did not, treat the claim as a hypothesis worth further research, not a settled cause.
[1][2]Sources and dates
- https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/correlation-and-causation · accessed 2026-10-18
- https://pmc.ncbi.nlm.nih.gov/articles/PMC10010939/ · accessed 2026-10-18
- https://thelongerlook.com/ · accessed 2026-10-18