No. Correlation does not by itself prove causation. A correlation shows that two variables are statistically associated according to a particular measure, whereas causation requires evidence that changes in one variable produce changes in another through a defensible causal mechanism under specified conditions. An association can arise because one variable causes the other, because the second causes the first, because both are influenced by another variable, because of selection effects, because of measurement problems or because of random variation. Correlation is therefore evidence of a pattern, not automatically evidence of a causal relationship. This distinction is fundamental to scientific reasoning because many observational datasets contain numerous correlations, including associations that disappear when additional variables or alternative explanations are considered.
Confounding is one of the main reasons a correlation can fail to demonstrate causation. A confounder is a variable that is associated with the exposure and the outcome and can distort their observed relationship. A classic example is the relationship between ice cream sales and drowning incidents, both of which can increase during warm weather. Temperature and seasonal conditions influence both variables, creating an association without ice cream purchases causing drowning.
In more technical causal analysis, confounding is defined in relation to the causal structure being investigated, and simply observing a third variable is not enough to determine whether it is a confounder. Researchers can address confounding through randomisation, restriction, matching, stratification, regression adjustment or other methods, but each method has assumptions and potential limitations.
Reverse causation is another reason correlation cannot establish a causal direction. If two variables are associated, the observed relationship alone may not reveal which variable influences which. For example, an observational association between physical activity and health could be compatible with multiple causal structures because health status may affect activity levels as well as activity affecting health. Establishing temporal order is therefore important in causal reasoning.
Longitudinal studies can help by measuring variables over time, although temporal precedence alone does not prove causality. Researchers also need to consider confounding, selection, measurement error, plausible mechanisms and alternative explanations. The direction of association in a statistical dataset is therefore not necessarily equivalent to the direction of a causal effect.
Selection effects can create associations even when the variables are not causally related in the way an analyst assumes. If inclusion in a dataset depends on variables related to both the exposure and outcome, conditioning on that selection process can produce a misleading relationship. Collider bias is a formal example in causal graphical reasoning. Measurement error can also generate or distort correlations.
If an instrument systematically changes its readings according to another variable, an apparent association may reflect the measurement process rather than the underlying phenomena. Data processing choices, missing data mechanisms and selective reporting can similarly alter observed relationships. Consequently, causal inference requires careful consideration of how observations entered the dataset and how variables were measured, not merely calculation of an association coefficient.
Randomised controlled experiments provide particularly strong evidence for causal effects because random assignment can distribute known and unknown confounding factors probabilistically between treatment groups. If the intervention is the principal systematic difference between groups and outcomes are measured appropriately, differences in outcomes can be interpreted under the assumptions of the experimental design as evidence concerning the causal effect of the intervention. Randomisation does not eliminate every problem. Noncompliance, loss to follow up, measurement error, inadequate sample size and interference between participants can complicate interpretation. Nevertheless, randomised experiments provide a stronger basis for causal inference than a simple observational correlation because the allocation mechanism creates a clearer basis for comparing counterfactual outcomes.
Observational studies can nevertheless provide substantial evidence about causation when randomised experiments are impossible, unethical or impractical. Epidemiology, astronomy, climate science and many areas of social and environmental research necessarily rely heavily on observational data. Researchers can strengthen causal inference using longitudinal designs, natural experiments, instrumental variables, regression discontinuity, difference in differences, propensity methods, causal graphical models and carefully specified adjustment strategies. These methods rely on assumptions that must be evaluated rather than treated as automatic substitutes for randomisation. Converging evidence from multiple study designs, consistent temporal patterns, plausible mechanisms, dose response relationships and successful predictions under interventions can collectively strengthen a causal explanation.
Causal inference is also distinct from finding a statistically significant association. A very small correlation can become statistically significant in a very large dataset, while a genuine causal effect can fail to reach conventional significance in a small or noisy study. Conversely, a strong correlation can be generated by confounding or selection. Statistical models can estimate conditional associations without necessarily identifying causal effects unless their assumptions correspond to an appropriate causal design.
Terms such as causal effect, counterfactual, intervention, potential outcome and directed acyclic graph are used in modern causal inference to formalise questions about what would happen under specified interventions. The central issue is not simply whether variables are related but what would change if one variable were actively altered while relevant conditions were held appropriately constant.
The correct scientific conclusion from a correlation is therefore conditional and limited. Correlation can identify a relationship that deserves investigation and can be compatible with causation, but it does not logically establish causation on its own. To support a causal claim, researchers seek evidence concerning temporal order, plausible mechanism, confounding, selection, measurement quality, experimental or quasi experimental design and robustness across analyses and datasets.
The strength of causal evidence depends on the question and the design rather than on one universal statistical threshold. Scientific reasoning therefore treats correlation as one component of evidence and distinguishes clearly between observing that variables move together and demonstrating that changing one produces a change in the other.
