Correlation and causation

The core idea

In observational studies the researcher merely observes the exposure and the health condition, so if there is a clear difference in disease rate between those exposed and others, the exposure is said to be "associated" with the disease; this is an association, not proof of cause.

Even with statistical adjustments, associations do not prove causation, and the distance between association and causation may be large.

Why it matters to you

Observational studies are useful: they help discover patterns and signals and raise new hypotheses worth testing, but their results are read as an association that needs confirmation.

Why a link can mislead

Confounders: in the NCCIH example, women who use chamomile may differ from others in many ways, such as level of physical activity, and these differences may explain the result.

Statistical adjustment is limited: it addresses only the differences that were measured, and any factor the researchers did not measure remains a possible source of confounding.

Chance: studies with small samples are more prone to results produced by chance, and their results may be inconclusive.

Timing of measurement: a cross-sectional study measures the exposure and the condition at the same time, and so is considered weaker for causal analysis, though suitable for describing prevalence.

How causes are tested

In a trial, the researcher determines each participant's exposure by a controlled process, participants are allocated by lot, then disease rates are compared between the groups; allocation by chance helps avoid bias.

When trials are not available, confidence in non-randomised studies increases when there is a strong association, or a relationship between the amount of exposure and the size of the effect, or when the effect of possible residual confounders would work against the observed result.

Repeating the same result in multiple studies makes it more reliable.

Common pitfalls in the news

Reading "was associated with" as if it meant "caused" or "protects against": association alone does not prove causation.

Changing your health habit based on a single observational study: a single study rarely proves anything.

Limits of this page

This page explains the general concept, and does not judge any particular association published in the news or in studies.

Even randomised trials can be infiltrated by bias, which is why their design and conduct are examined, and confidence in their evidence may be downgraded.

Common questions