A correlation coefficient is a single number, r, between −1 and +1 that tells you whether two variables move together and how tightly. What it does not tell you is what you’re allowed to conclude from that number — an r of 0.60 can mean something entirely different depending on how the data were collected, whether the relationship is actually a straight line, and whether one stray data point or a hidden subgroup is doing all the work. That gap between “here is r” and “here is what r means” is where most correlation coefficients get misread, and it’s the gap this page is built to close.
You’ll get the formula and a full worked calculation below, but the core of this page is judgment: how to read a given r value correctly, and the specific ways r quietly misleads people who take it at face value — nonlinearity, restricted range, outliers, and mixing subgroups that individually tell a different story than the pooled data does. For the broader concept of correlation — what it means and why it matters before you ever compute a number — see what correlation means in statistics.
What a Correlation Coefficient Actually Measures
A correlation coefficient summarizes how consistently two variables move together across a set of paired measurements. When x rising tends to accompany y rising, the coefficient is positive. When x rising tends to accompany y falling, it’s negative. When there’s no consistent straight-line pattern either way, r sits near zero — though, as the misleading section below covers, “near zero” is not the same as “unrelated.”
The version used almost everywhere is the Pearson correlation coefficient, named for Karl Pearson, who formalized it in the 1890s. It’s symmetric (correlating x with y gives the same answer as y with x), dimensionless (changing units doesn’t change r), and it only ever measures the linear part of a relationship. Two other coefficients — Spearman’s ρ and Kendall’s τ — extend the idea to ranked or curved-but-monotonic data; the Pearson correlation coefficient article covers how those relate to Pearson’s r and to covariance in more depth than belongs here.
A positive coefficient shows up wherever two quantities share a driver in the same direction — study hours and exam scores, altitude and air pressure (inverted), advertising spend and short-run revenue. A negative one shows up just as often — exercise frequency and resting heart rate, elevation and temperature. Neither sign says more about the strength of the relationship than the other; only the absolute value does, which is the subject of the interpretation section further down.
The Correlation Coefficient Formula
The coefficient of correlation formula expresses r as the covariance of x and y scaled by the product of their standard deviations:
r = Σ[(xᵢ − x̄)(yᵢ − ȳ)] / √[ Σ(xᵢ − x̄)² · Σ(yᵢ − ȳ)² ]
where xᵢ and yᵢ are the paired values, x̄ and ȳ are the sample means, and Σ sums over all n pairs. An algebraically identical raw-score form skips computing the means first:
r = [ nΣxᵢyᵢ − (Σxᵢ)(Σyᵢ) ] / √{ [ nΣxᵢ² − (Σxᵢ)² ] · [ nΣyᵢ² − (Σyᵢ)² ] }
Both forms give the same result. The deviation form is the one worth understanding: the numerator is the sum of cross-products of deviations from each mean, and dividing by the product of the two standard deviations is exactly what forces the result to stay inside [−1, +1], as the NIST/SEMATECH e-Handbook of Statistical Methods, §1.3.5.11 — Bivariate Data and Scatter Plots sets out. When x and y deviate from their means in the same direction more often than not, the cross-products skew positive and r climbs toward +1; when they deviate in opposite directions, r moves toward −1; when the two are unrelated, positive and negative cross-products cancel out and r drifts toward 0.
How to Find the Correlation Coefficient: A Worked Example
A bottling plant’s quality team wants to know whether a fill-line sensor drifts as it goes longer without recalibration. Over five consecutive shifts, they log the hours since the sensor was last calibrated (x) against the number of bottles flagged out-of-tolerance that shift (y):
| Shift i | Hours since calibration (xᵢ) | Bottles flagged (yᵢ) |
|---|---|---|
| 1 | 1 | 2 |
| 2 | 2 | 4 |
| 3 | 3 | 5 |
| 4 | 4 | 4 |
| 5 | 5 | 5 |
Step 1: Compute the sample means
Sum of x: 1 + 2 + 3 + 4 + 5 = 15, so x̄ = 15 / 5 = 3.
Sum of y: 2 + 4 + 5 + 4 + 5 = 20, so ȳ = 20 / 5 = 4.
Step 2: Deviations, squared deviations, and cross-products
| xᵢ | yᵢ | xᵢ − x̄ | yᵢ − ȳ | (xᵢ − x̄)² | (yᵢ − ȳ)² | (xᵢ − x̄)(yᵢ − ȳ) |
|---|---|---|---|---|---|---|
| 1 | 2 | −2 | −2 | 4 | 4 | 4 |
| 2 | 4 | −1 | 0 | 1 | 0 | 0 |
| 3 | 5 | 0 | +1 | 0 | 1 | 0 |
| 4 | 4 | +1 | 0 | 1 | 0 | 0 |
| 5 | 5 | +2 | +1 | 4 | 1 | 2 |
| Σ | 10 | 6 | 6 |
Step 3: Apply the formula
r = 6 / √(10 × 6)
= 6 / √60
= 6 / 7.746
≈ 0.775
An r of roughly 0.775 tells the quality team that flagged bottles do climb fairly reliably as time-since-calibration increases, but the fourth shift (4 hours, only 4 flags instead of the 5 the trend would predict) is a reminder that “fairly reliably” is not “always” — real production data rarely sits exactly on a line. Squaring r gives the coefficient of determination, r² = 0.775² ≈ 0.600: about 60% of the shift-to-shift variation in flagged bottles lines up with hours since calibration, leaving 40% down to other factors (bottle batch, line speed, operator, ambient temperature).
Try the Calculator
Enter the same five pairs above — or substitute your own x and y values — to confirm the result instantly. The calculator computes r, r², and the number of paired values.
To work with larger datasets or explore the output labels in more detail, open the dedicated correlation coefficient calculator page. You can also browse all available tools on the calculators hub.
What an r Value Actually Licenses You to Say
Once you have r, two independent questions matter: which direction the relationship runs, and how much strength the number represents. Getting the sign right is easy; getting the strength claim right is where people overreach.
Direction is read straight off the sign: positive means the variables rise together, negative means one rises as the other falls, and r = 0 means no linear pattern — not necessarily no pattern at all (more on that below).
Strength is read off the absolute value, using rough conventions like the ones in Penn State STAT 501 — Regression Methods, Lesson 1:
| |r| value | Commonly described as | |------------|----------------------| | 0.9 – 1.0 | Very strong | | 0.7 – 0.9 | Strong | | 0.5 – 0.7 | Moderate | | 0.3 – 0.5 | Weak to moderate | | 0.0 – 0.3 | Weak or negligible |
Treat that table as a starting point, not a verdict — an r of 0.40 is a headline result in psychology and an unusable sensor in precision manufacturing. What the table does settle is comparisons: when you’re asked which value of r indicates a stronger correlation, or which of a set of candidates represents the strongest or weakest correlation, the sign never enters into it — only the absolute value does. r = −0.90 beats r = +0.75 as the stronger relationship, because |−0.90| = 0.90 exceeds |+0.75| = 0.75. Given a set like r₁ = −0.90, r₂ = +0.75, r₃ = −0.30, r₄ = +0.10, the strongest correlation is r₁ (largest |r|) and the weakest is r₄ (smallest |r|), regardless of how the question happens to be worded or hyphenated. The most common exam mistake is picking the largest positive number instead of the largest magnitude — a near-perfect negative correlation is still near-perfect.
One more caveat worth stating plainly: strength conventions assume a reasonable sample size. r = 0.50 from five data points and r = 0.50 from five hundred carry very different amounts of evidence, even though the coefficient itself is identical.
Four Ways a Correlation Coefficient Can Mislead You
A single r value is a compression of a lot of structure into one scalar, and compression loses information. These are the four failure modes that trip up people who report r without a second look.
It only sees straight lines
Pearson’s r measures linear association exclusively. If x and y trace out a clean U-shape — imagine drug dosage on the x-axis and a measured effect that improves, peaks, then declines on the y-axis — the rising half and falling half of the curve can cancel almost perfectly, handing you an r close to 0 even though the two variables are tightly, predictably related. The fix is procedural, not statistical: plot the data before you trust the number.
It shrinks under a restricted range
Correlating within a narrow slice of one variable understates the relationship that exists across the full range. If the bottling-plant example above only logged shifts where the sensor had been calibrated one or two hours earlier — a restricted slice of the x-axis — the visible trend would look far weaker than the true relationship across the full 1-to-5-hour range, simply because most of the variation in x has been sliced away before the correlation was ever computed. Comparing correlations computed on different-width samples of the same relationship is comparing different questions, not different answers to the same one.
A single outlier can move it a long way
Because the formula squares deviations from the mean, one extreme point can swing r substantially in either direction — sometimes conjuring a correlation that isn’t really there, sometimes erasing one that is. If your data includes points far outside the typical spread, check whether r survives their removal before reporting it. Pearson vs Spearman covers when a rank-based coefficient is the more honest choice for exactly this situation.
Pooling groups can flip the sign entirely
The most dramatic failure isn’t a weak correlation — it’s a correlation that reverses once subgroups get combined. Picture two production lines at the bottling plant: on Line A, longer time-since-calibration correlates with fewer flagged bottles (its newer sensor drifts slowly and gets recalibrated fast); on Line B, the same relationship holds separately but at consistently higher flag counts and longer calibration intervals. Pool both lines into one dataset without accounting for which line each shift came from, and the line identity — a confounding variable — can make the combined correlation run positive even though it’s negative within each line. This is Simpson’s paradox, and it is exactly why a correlation coefficient computed on aggregated data deserves a check for hidden subgroups before you trust its sign, let alone its magnitude.
Correlation Coefficients Don’t Prove Causation
A high r tempts people to conclude that one variable causes the other, and it can’t support that conclusion on its own — a correlation is equally consistent with direct causation, a shared third cause, or plain coincidence in a small or heavily-tested sample. The bottling-plant r of 0.775 is a case in point: it’s consistent with calibration drift causing more flagged bottles, but it can’t rule out a third factor, like ambient humidity, that independently affects both the sensor and the bottling process. Distinguishing those explanations needs controlled experiments or formal causal-inference methods, not a bigger r. The full reasoning, including the classic ice-cream-and-drownings example, lives in correlation vs causation.
Frequently Asked Questions
What is a correlation coefficient?
A correlation coefficient is a standardized number, always between −1 and +1, that captures the strength and direction of the linear relationship between two numeric variables. Pearson’s r is the standard version. Values near +1 or −1 indicate a strong linear relationship (positive or negative); values near 0 indicate a weak or absent linear relationship.
How do I find the correlation coefficient by hand?
Compute the sample means x̄ and ȳ, subtract each mean from its column of values to get deviations, multiply each x-deviation by the matching y-deviation to get cross-products, sum those cross-products for the numerator, then sum the squared x-deviations and squared y-deviations separately, multiply those two sums, take the square root, and divide. The worked example above walks through every step of how to find the correlation coefficient with a full table.
Which of the following r values represents the strongest correlation — and which the weakest?
Whichever value has the largest absolute value is the strongest, and whichever has the smallest absolute value is the weakest, no matter how the r-values are hyphenated or phrased in the question. For candidates −0.90, +0.75, −0.30, and +0.10, the strongest is −0.90 (|−0.90| = 0.90) and the weakest is +0.10 (|+0.10| = 0.10). Sign tells you direction; only magnitude tells you strength.
What’s the difference between a correlation coefficient and covariance?
Covariance measures the same co-movement between two variables but stays in their original units (kg·cm, dollars·days), which makes it impossible to compare across datasets. Dividing covariance by the product of the two standard deviations produces r — dimensionless and always within [−1, +1]. The Pearson correlation coefficient article works through that derivation in full.
Can two variables have r = 0 and still be related?
Yes. Pearson’s r captures only linear association. A perfect U-shaped or quadratic relationship — y = x², for instance — can produce r = 0 exactly, because the rising and falling halves of the curve cancel out. It’s one of the four ways r misleads, covered above, and the reason a scatter plot should always accompany a reported correlation coefficient.
What counts as a good correlation coefficient?
There’s no universal cutoff. r > 0.50 reads as a strong effect in psychology and social science; r = 0.30 can be meaningful at a population level in epidemiology; precision engineering may treat anything below 0.95 as unacceptably noisy. The right threshold depends on your field, your sample size, and what’s actually riding on the answer.
Summary
A correlation coefficient compresses the strength and direction of a linear relationship into one number between −1 and +1, computed by standardizing the covariance of two variables against their standard deviations. Sign tells you direction only; absolute value tells you strength, which is why the largest-magnitude r in a set is always the strongest correlation and the smallest-magnitude r is always the weakest, whatever the sign.
What the number can’t do on its own is rule out the four failure modes covered here: a curved (non-linear) relationship that r reads as weak or zero, a restricted range that understates the true association, a single outlier doing most of the work, or a Simpson’s-paradox-style reversal hiding inside pooled subgroups. Compute r, but also plot the data, check for outliers, and ask whether the data were pooled across groups that might behave differently — the calculator above will do the arithmetic for any dataset you have, but the judgment calls are still yours to make.