How do scientists tell correlation from causation?

Correlation means two variables are associated: when one changes, the other tends to change too. Causation means a change in one directly produces a change in the other. Scientists separate the two mainly by running randomized controlled trials, which are described as the gold standard for direct evidence. When trials are not possible, they gather several kinds of indirect evidence and check whether other explanations can be ruled out. [1] [2]

An orthopaedic research review names three common reasons a correlation is mistaken for causation. The first is coincidence: rising rates of breast cancer or liver disease may track rising joint replacements purely by chance. The second is reverse causality, where the supposed effect actually comes first. For example, a very low body mass index is linked to higher short-term mortality after joint replacement, partly because low weight reflects poor underlying health. The third, and the most common, is confounding, where a third factor distorts the apparent relationship between two variables. [2]

An ecology paper states the same idea as the Common Cause Principle. If two variables are statistically dependent, then either X causes Y, Y causes X, or a third variable causes both. [3]

In a randomized controlled trial, participants are randomly assigned to receive an intervention or to serve as controls. Because of this, a difference between the groups can be attributed to the intervention, since randomization minimizes the influence of measured and unmeasured confounders. A data science textbook likewise calls randomized experiments the most reliable way to infer causation from data. [1] [4] [2]

Trials are not always possible for ethical or practical reasons. There is no trial showing that smoking causes lung cancer, because the indirect evidence is so strong that running one would be unethical. Most analyses in science, medicine and the social sciences therefore use observational data, where causal conclusions depend on assumptions that can be hard to verify. [1] [4]

Without a trial, scientists reason inductively, and each piece of evidence strengthens the case even if none is persuasive alone. Temporality is the only absolute requirement: the cause must come before the effect, though long lags, such as 20 to 60 years between asbestos exposure and mesothelioma, can obscure this. Other supports include a strong association, consistent results across studies using different methods, and a plausible mechanism, such as lab work showing how a substance damages cells. A dose-response relationship, where larger exposures produce larger effects, is also highly supportive, although some causes act only above a threshold, as with the minimum infectious dose for many infections. [1]

Other sources give similar checklists. The data science textbook adds the elimination of alternative explanations such as confounders, and the National Library of Medicine page lists biological plausibility, coherence and specificity of association among its criteria. [4] [5]

A Nature Methods article states that correlation implies association but not causation, and that causation implies association but not necessarily correlation, since correlation measures specifically increasing or decreasing trends. [6]