Correlation Coefficient Calculator

Four datasets, one correlation, nothing in common

Francis Anscombe built four datasets in 1973 to make a point that no amount of arguing had made. Load them in the tool above and switch between them.

All four agree to three decimal places on every classical summary. r = 0.816. r² = 0.667. Mean x = 9 exactly. Mean y = 7.50. Variance of y = 4.12. Fitted line y = 3.00 + 0.500x. Everything a statistical report normally contains is the same.

The scatter plots have nothing in common. The first is a genuine noisy line. The second is a clean parabola — a perfect relationship that a linear measure cannot see. The third is a straight line to seven figures with one outlier, and that single point drags r down from 0.99999655 to 0.816. The fourth has no relationship at all: ten points share the same x, and one point at x = 19 manufactures the entire correlation. Delete it and r is undefined, because x never varies.

Here is the part usually left out, and it is the useful part. Spearman’s ρdoes separate them: 0.818, 0.691, 0.991, 0.500. Where Pearson reports four identical numbers, Spearman reports four distinctly different ones.

That gives you an automatic check that costs one extra calculation. When Pearson and Spearman differ substantially, the relationship is not linear, or a small number of points are doing most of the work. Neither coefficient tells you which — but the disagreement tells you to look, and the tool flags it.

It is still not a substitute for the plot. Anscombe’s first and fourth datasets have Pearson and Spearman in reasonable agreement, and only one of them contains a relationship. The disagreement check catches many failures cheaply; the scatter plot catches all of them, which is why the tool draws it before it reports anything.

Which coefficient answers your question

The three measure genuinely different things, and picking by habit rather than by question is how a real relationship gets reported as absent.

Pearson’s r asks: how close are these points to a straight line? It is the default and it is the right choice for roughly linear relationships between roughly symmetric variables. It is sensitive to outliers, because it works with the actual distances.

Spearman’s ρ asks: does y rise when x rises? It works on ranks, so it detects any monotone relationship regardless of shape — and a perfect curve gets ρ = 1 where Pearson reports something considerably less. Replacing values with ranks also caps the influence of an outlier at one rank position, which makes it far more robust.

Kendall’s τ asks: what share of pairs agree on direction? It counts, for every pair of observations, whether both variables move the same way. That interpretation is unusually concrete — a τ of 0.6 means 80% of pairs are concordant — and it behaves better than Spearman on small samples and heavily tied data. It is systematically smaller than ρ on the same data, which is not a disagreement but a difference of scale.

Use Spearman or Kendall when the data is ordinal — survey responses, rankings, grades. Pearson on a five-point Likert scale is treating “agree” minus “neutral” as a real distance equal to “neutral” minus “disagree”, which the scale does not promise.

One implementation warning, because it is invisible. Spearman is often computed with the shortcut 1 − 6Σd²/(n(n² − 1)). That formula is only valid with no tied values. With ties it is silently wrong — on one small example here it returns 0.2321 where the correct answer is 0.1989 — 16.7% too high, with nothing to indicate a problem. This page computes Spearman as Pearson on midranks, which is correct either way.

Correlation and causation, stated usefully

“Correlation does not imply causation” is true, universally repeated, and too vague to act on. The useful version is the list of specific things that produce a correlation without a causal link.

A common cause. Ice cream sales and drowning deaths correlate because both rise with temperature. Neither causes the other; a third variable drives both. This is the most common case by far, and the fix is to measure the third variable and control for it.

Reverse causation. The relationship is real and runs the other way. Hospitals correlate with illness. The correlation is symmetric; causation is not, and a coefficient cannot distinguish direction.

Selection effects. Correlations can be created by how the sample was chosen. Among admitted students, test scores and interview scores often correlate negatively even when they are independent in the applicant pool — because admission required doing well on at least one, so anyone weak on the first must have been strong on the second. This one is genuinely counter-intuitive and appears constantly in observational data.

Coincidence. Test enough variable pairs and strong correlations appear by chance. With 100 unrelated variables there are 4,950 pairs, and at a 5% threshold roughly 250 will look significant. This is why a correlation found by searching needs a much stronger threshold than one predicted in advance.

Aggregation. Correlations between group averages are usually far stronger than the same correlation between individuals, and applying one to the other is the ecological fallacy. Country-level data is especially prone to this — a strong relationship across 40 national averages can be near zero within every one of those countries.

What a correlation genuinely provides is a screening tool and a prediction. If x and y correlate, x helps predict y whether or not it causes it. That is worth something on its own, and it is a claim you can defend.

Reading r without overstating it

Four habits that keep an r honest.

Square it before interpreting the strength. r = 0.5 sounds like half a relationship; it accounts for 25% of the variance. r = 0.7 accounts for 49%. The coefficient is on a scale that flatters weak relationships, and r² is the number that answers “how much does knowing x help?”

Do not use fixed strength labels. “0.7 is strong” depends entirely on the field. An r of 0.3 is a substantial effect in psychology and unacceptably poor for a calibration instrument. Compare against what is typical in your domain, not against a table.

Statistical significance is not strength. At n = 1000, an r of 0.07 is significant at the 1% level — and it accounts for half a percent of the variance. With enough data every nonzero correlation becomes significant. Report the coefficient and the sample size together; a p-value alone hides which one you have.

Restricting the range shrinks r. Compute a correlation only among high scorers and it drops sharply, because you have removed most of the variation it needs. This is why validity coefficients look weak when computed on people who were selected using the very test being validated — a structural artefact, not evidence that the test fails.

For the fitted line itself, its uncertainty and its residual behaviour, the linear regression calculator goes further than the summary shown here.

Controlling for a third variable

The most common cause of a correlation without a causal link is a shared third variable. Partial correlation is the arithmetic that removes it, and it is worth knowing because it turns a caveat into a calculation.

A partial correlation is the correlation between x and y after removing what z explains of both. Regress x on z, regress y on z, then correlate the two sets of residuals. What is left is the association that z does not account for.

The classic result is that it can collapse to nothing. Ice cream sales and drowning deaths correlate strongly; control for temperature and the partial correlation falls near zero. That is a direct demonstration that the original association carried no information about a link between them.

It can also go the other way, which is the surprising case. Two variables can be uncorrelated overall and strongly correlated once a third is held fixed — a suppressor variable was hiding the relationship. So a zero correlation is not evidence of no relationship any more than a strong one is evidence of a causal link.

The limits are real and worth stating. Controlling for z only works if you measured z, and measured it well — a noisy measurement removes only part of its influence and leaves residual confounding that looks like a genuine effect. And controlling for a variable that sits between x and y on the causal path removes the very effect you were trying to measure.

Which is the honest summary: partial correlation controls for the confounders you thought of and measured. That is a real advance on doing nothing, and it is not the same as controlling for confounding.

Sources and methodology

References for the quartet and the rank statistics.

• Anscombe, “Graphs in Statistical Analysis”, The American Statistician 27(1), 1973 — the original four datasets
• NIST/SEMATECH e-Handbook — scatter plots and the linear correlation coefficient
• R documentation for cor — Pearson, Kendall and Spearman, and its handling of ties
• Linear Regression Calculator — for the fitted line, its intervals and its residuals
• P-Value Calculator — for the significance test behind the reported p
• How we calculate — methodology

Method. Pearson is computed from centred sums rather than the algebraically identical (nΣxy − ΣxΣy) form, which subtracts two large nearly-equal numbers and can lose most of its significant digits — or return an |r| above 1 — on data far from the origin. Ranks use midranks, so Spearman is correct under ties where the familiar 6Σd² shortcut is not; Kendall is tau-b for the same reason. A constant variable returns undefined rather than zero, because “no information” and “no association” are different findings. The suite asserts that all four Anscombe datasets agree on r, r², both means and the fitted line to three decimals while their Spearman values differ by up to 0.49, that the shortcut disagrees with the correct Spearman under ties, and that the centred and computational forms diverge on offset data. That engine is verified on every change against 79 assertions. The count and the per-case breakdown are published on the formula verification page.

Related calculators

Where this goes next:

More in Math, or browse all calculators.