Association Is Not a Causal Direction
A scatterplot can show a clear pattern, and a correlation can summarize how strongly two quantitative variables move together in a linear way. Neither one, by itself, tells us why the pattern occurs. If two variables are correlated, it does not follow that changing one will cause the other to change.
This distinction matters whenever a result is used to make a decision. If study time and exam scores are positively correlated, increasing study time might improve scores. But the observed pattern alone does not rule out other explanations: perhaps students who expect to do well study more, or perhaps the paired values happen to line up by chance. The correlation describes the observed values; a causal claim concerns what would happen if one variable were deliberately changed.
Three possibilities are useful to separate. In a direct causal story, changing \(X\) can change \(Y\). In a reverse-causation story, the influence runs from \(Y\) to \(X\), even if someone is tempted to claim that \(X\) causes \(Y\). In a coincidence, the observed association arises by chance in the particular data set, although there is no causal relationship between the variables. Other explanations, such as a shared influence on both variables, can also produce association.
Why a Correlation Cannot Choose the Arrow
A correlation describes paired values without encoding which variable came first in a causal process. In fact, the sample correlation is symmetric: exchanging the labels \(X\) and \(Y\) does not change its value. This follows from the formula
The numerator is unchanged when \(x_i\) and \(y_i\) are swapped, and the two factors in the denominator merely exchange places. Thus \(r(X,Y)=r(Y,X)\). This symmetry is appropriate for a measure of association, but a causal claim is directional: it proposes that changing one variable affects the other.
The same paired values can therefore be consistent with different stories. A scatterplot and \(r\) can tell us about the form and strength of a pattern, but deciding whether it is causal requires information beyond that pattern. For example, we may need to know which event occurred first, how the data were collected, and whether the possible cause was deliberately changed.
Worked Example: A Plausible Direct Effect
Suppose, in a deliberately simplified hypothetical experiment, five similar seedlings are randomly assigned different fertilizer amounts. Other growing conditions are kept alike. The recorded fertilizer amount \(x\), in grams, and plant growth \(y\), in centimeters, are:
\(x: 0, 1, 2, 3, 4\); \(y: 2, 4, 6, 8, 10\).
The means are \(\bar{x}=2\) grams and \(\bar{y}=6\) centimeters. The sums of squared deviations and the sum of cross-products are
Substituting into the sample correlation formula gives
The data show a perfect positive linear association: in these five observations, every additional gram of fertilizer is paired with 2 additional centimeters of growth. Because the fertilizer amounts were assigned and the growing conditions were held alike, the design gives a reason to consider the direct causal explanation that fertilizer affects growth. The correlation itself does not provide that reason; the assignment and controlled conditions do. With only five seedlings, this toy result would not by itself establish how plants generally respond to fertilizer.
Reverse Causation: The Other Direction May Fit
A frequent mistake is to assume that the variable named first, or the variable that seems like a natural predictor, must be the cause. But paired measurements taken at the same time may not establish temporal order. Even when one direction sounds plausible, the reverse direction may also be plausible.
Consider stress scores and hours of sleep measured for five hypothetical people. A positive stress score indicates more reported stress. Suppose the paired observations are stress scores \(x=1,2,3,4,5\) and sleep durations \(y=9,8,7,6,5\) hours. The negative association is clear: higher stress scores are paired with fewer hours of sleep.
Worked Example: Two Stories for the Same Negative Correlation
The means are \(\bar{x}=3\) stress-score points and \(\bar{y}=7\) hours. For these data,
Therefore,
The perfect negative correlation says that the observed values follow a decreasing straight-line pattern. It does not say which variable caused the other. One possible story is that increased stress disrupts sleep. A reverse-causation story is that reduced sleep increases stress. Both stories are compatible with the direction of this association; the paired measurements alone do not choose between them.
To support either causal interpretation, we would need more than the correlation, such as evidence about timing and a study design that addresses other differences between the people. The number \(-1\) describes the pattern in these observations, not the direction of a causal effect.
Coincidence: A Strong Sample Pattern Can Occur by Chance
A correlation need not represent a causal connection at all. When values are paired, a sample can show a noticeable pattern even if the variables were generated independently. This is especially possible in a small data set. Independence in the process that generates the values does not require the sample correlation to equal zero.
Worked Example: A Correlation in Independently Generated Data
Consider five fictional records in a toy exercise. For each record, a number of songs downloaded, \(x\), is generated independently of a quiz score, \(y\). The values happen to be \(x=1,2,3,4,5\) and \(y=2,1,4,3,5\). The exercise stipulates that one set of values was generated without reference to the other; these data are not from a real study.
Both means are 3. The deviations are \((-2,-1,0,1,2)\) for \(x\) and \((-1,-2,1,0,2)\) for \(y\). Their products sum to
Each variable has a sum of squared deviations of 10, so
The sample shows a fairly strong positive linear association, even though the values were generated independently. In a larger collection of repeated independent samples, positive and negative patterns can occur; this one happens to be positive. The result is not evidence that downloading songs causes better quiz scores, or that quiz performance causes downloading. It illustrates why even a large-looking sample correlation is not, on its own, proof of causation.
A correlation does not reveal whether an apparent pattern is a stable feature of a broader process or an accident of the particular observations. More data can help assess how variable a pattern is, but a larger sample by itself does not establish causation. The way data are collected remains important.
What Can Support a Causal Interpretation?
A causal conclusion is more credible when a study deliberately changes a possible cause and compares outcomes under a design that makes the comparison fair. For example, randomly assigning similar experimental units to different treatments can help prevent pre-existing differences from systematically determining who receives which treatment. Researchers must also consider whether the assigned treatment was actually received and whether the study was carried out as intended.
Not every question can be studied by deliberately assigning a cause. It may be unethical or impossible to assign people different levels of a risk or life experience. In those cases, observational data can still show important associations, but a causal interpretation requires additional evidence and careful reasoning about alternative explanations. A correlation alone cannot supply what the data collection process did not establish.
Keep the claim matched to the evidence. If the analysis establishes only that two variables vary together, report an association. If the study design and supporting evidence justify a causal conclusion, state the cause, outcome, population, and circumstances precisely. A strong correlation does not automatically make a causal claim stronger than the design allows.
Common Mistakes
- Treating a high magnitude of \(r\) as proof. A correlation near 1 or \(-1\) describes a strong linear pattern in the observed data. It does not identify a cause.
- Assuming the direction from the order of the variables. Listing one variable first, or calling it a predictor, does not show that it causes the other. The same correlation is obtained when the variables are exchanged.
- Ignoring reverse causation. If the measurements do not establish which event came first, the outcome could plausibly affect the proposed cause instead.
- Calling every unexplained association causal. A sample can show a correlation by coincidence, and other processes can also create association. The observed pattern alone does not distinguish these explanations.
- Assuming more observations automatically prove causation. A larger sample may describe an association more precisely, but it does not by itself change an observational design into an experiment.
Check Your Understanding
For each question, separate what the correlation shows from what a causal claim would require.
- Two variables have a sample correlation of \(r=-0.92\). What does this tell you about their observed linear association, and what does it not establish?
- A study finds that people who sleep less report more stress. Give one direct causal story and one reverse-causation story consistent with this association.
- In the songs-and-quiz example, why does \(r=0.8\) not demonstrate that songs downloaded cause quiz scores to rise?
- Why can a small sample of independently generated paired values have a nonzero correlation?
- What feature of a study could provide stronger support for a causal interpretation than an observational correlation alone?