When a Small Correlation Misleads
Pearson’s correlation is useful for describing the direction and strength of a linear pattern in paired quantitative data. But a small correlation does not guarantee that a scatterplot has no structure. If the points follow a curve, an upward trend in one part of the plot can be balanced by a downward trend in another. The overall linear summary can then be small even when the pattern is clear.
This distinction matters whenever a numerical summary is reported without the scatterplot it summarizes. A correlation close to zero may describe a cloud with no visible structure—or a precise U-shaped curve. Those situations call for different conclusions, even though their correlations may be alike.
The numerator adds the products of paired deviations from the two sample means. If observations with above-average \(x\) tend also to have above-average \(y\), and below-average \(x\) tend to pair with below-average \(y\), these products tend to be positive. If one variable tends to be above its mean when the other is below, the products tend to be negative. In a curved pattern, however, positive and negative products can cancel. This cancellation is the key to understanding why a strong nonlinear pattern can have \(r=0\), or a value close to it.
An Exact U-Shaped Pattern
Consider a hypothetical set of measurements where \(x\) is a position in centimeters relative to a reference point and \(y\) is a device’s response in arbitrary response units. The response follows a U-shaped rule: positions farther from the reference point have larger responses.
Worked Example: A Perfect U With Zero Correlation
Suppose the five paired measurements are shown below. These hypothetical values follow the exact rule \(y=x^2\).
| \(x\) (cm) | \(y\) (response units) |
|---|---|
| -2 | 4 |
| -1 | 1 |
| 0 | 0 |
| 1 | 1 |
| 2 | 4 |
First calculate the sample means: $$ \bar{x}=\frac{-2-1+0+1+2}{5}=0,\qquad \bar{y}=\frac{4+1+0+1+4}{5}=2. $$ The products of paired deviations are $$ (-2-0)(4-2)=-4,\quad (-1-0)(1-2)=1,\quad (0-0)(0-2)=0, $$ $$ (1-0)(1-2)=-1,\quad (2-0)(4-2)=4. $$ Their sum is \(-4+1+0-1+4=0\). The sums of squared deviations in the denominator are $$ \sum(x_i-\bar{x})^2=4+1+0+1+4=10, $$ $$ \sum(y_i-\bar{y})^2=4+1+4+1+4=14. $$ Therefore, $$ r=\frac{0}{\sqrt{10\cdot14}}=0. $$
The zero correlation does not mean the measurements have no relationship: every response is exactly \(x^2\). It means the positive and negative contributions to the linear summary cancel. On the left side of the U, increasing \(x\) is associated with a decreasing \(y\); on the right side, increasing \(x\) is associated with an increasing \(y\). A single straight-line summary does not capture both parts of this curve.
Why Symmetry Can Make the Correlation Zero
The calculation above reflects a general symmetry argument. Consider paired \(x\)-values that occur as opposites, such as \(-2\) and \(2\), with matching \(y\)-values. For an even curve such as \(y=x^2\), each pair has the form \((-a,b)\) and \((a,b)\). The average \(x\)-value is zero when all such pairs are included. Their contributions to the correlation numerator are $$ (-a)(b-\bar{y})+(a)(b-\bar{y})=0. $$ The same cancellation applies to a pair at any distance \(a\). If there is also a central observation at \(x=0\), its contribution to the numerator is zero. Thus the full numerator is zero. If both variables vary, the denominator is positive, so \(r=0\).
This is a short justification for a broader lesson: a zero covariance, and therefore a zero Pearson correlation when both variables vary, rules out a net linear pattern in the data. It does not rule out a curved pattern. Symmetry is one common way for positive and negative contributions to cancel, but cancellation can also be approximate rather than exact.
A Circular Pattern With Zero Correlation
A nonlinear relationship need not be a U-shaped function. A circular pattern provides a different example: the coordinates can be tightly constrained by a circle, even though knowing \(x\) does not specify a single \(y\)-value. In the following hypothetical dataset, each point lies exactly on a circle centered at the origin.
Worked Example: Points Exactly on a Circle
Suppose a sensor records the following twelve pairs of coordinates, in arbitrary units: \((5,0),(-5,0),(0,5),(0,-5)\); \((3,4),(3,-4),(-3,4),(-3,-4)\); and \((4,3),(4,-3),(-4,3),(-4,-3)\). Every point satisfies \(x^2+y^2=25\), so the points lie exactly on a circle of radius 5.
The point set is balanced around both axes, so \(\bar{x}=\bar{y}=0\). For each group with coordinates \(3\) and \(4\), changing one coordinate’s sign produces matching positive and negative \(xy\) products. The axis points also have \(xy=0\). Therefore, $$ \sum (x_i-\bar{x})(y_i-\bar{y})=\sum x_i y_i=0. $$ For the \(x\)-coordinates, the two axis points contribute \(2(25)=50\) to the sum of squares. The four points with \(|x|=3\) contribute \(4(9)=36\), and the four with \(|x|=4\) contribute \(4(16)=64\). Thus \(\sum x_i^2=50+36+64=150\). By the same symmetry, \(\sum y_i^2=150\). Hence $$ r=\frac{0}{\sqrt{150\cdot150}}=0. $$
Here \(r=0\) despite an exact geometric constraint: every pair obeys \(x^2+y^2=25\). The circular pattern is not a single-valued curve of \(y\) against \(x\), since most \(x\)-values correspond to both a positive and a negative \(y\)-value. But the plotted points still have unmistakable structure. The example shows why correlation alone cannot describe every kind of pattern in paired data.
A Nearly Zero Correlation on a U-Shaped Pattern
Exact symmetry is not necessary for a small correlation. A slight departure from symmetry can leave the curved pattern clear while making the correlation close to zero rather than exactly zero. The next example shows the calculation with such data.
Worked Example: A Small Nonzero Correlation
In another hypothetical set of device measurements, let \(x\) be position in centimeters and \(y\) be response in arbitrary units. Suppose the pairs are \((-3,9),(-2,4),(-1,1),(0,0),(1,1),(2,4),(3,10)\). The values closely follow the U-shaped rule \(y=x^2\), except that the final response is 10 rather than 9.
The \(x\)-values sum to zero, so \(\bar{x}=0\). The \(y\)-values sum to \(29\), so \(\bar{y}=29/7\). The numerator can be simplified because \(\sum x_i=0\): $$ \sum (x_i-\bar{x})(y_i-\bar{y}) =\sum x_i y_i-\bar{y}\sum x_i =\sum x_i y_i. $$ Substituting the data gives $$ \sum x_i y_i=(-3)(9)+(-2)(4)+(-1)(1)+0(0)+(1)(1)+(2)(4)+(3)(10)=3. $$ For \(x\), the sum of squared deviations is \(9+4+1+0+1+4+9=28\). For \(y\), first sum the squared values: \(9^2+4^2+1^2+0^2+1^2+4^2+10^2=215\). Then subtract \(n\bar{y}^{\,2}\) to obtain the sum of squared deviations: $$ \sum(y_i-\bar{y})^2=215-7\left(\frac{29}{7}\right)^2 =215-\frac{841}{7} =\frac{664}{7}. $$ Consequently, $$ r=\frac{3}{\sqrt{28(664/7)}}=\frac{3}{\sqrt{2656}}\approx 0.0582. $$
The correlation is close to zero, with a slight positive sign. That sign describes the small net linear tendency in these observations; it does not summarize the prominent U-shaped pattern. A scatterplot would show the response falling toward the middle and rising again on either side. Reporting only \(r\approx0.0582\) would leave out that important feature.
How to Read a Small Correlation
A small value of \(r\) should be interpreted narrowly: the data show little overall linear association as summarized by Pearson’s correlation. It is not, by itself, evidence that the variables are unrelated. A curve can be strong and systematic while having little net linear association, as the U-shaped and circular examples demonstrate.
For paired quantitative data, look at the scatterplot as well as the correlation. The plot can reveal whether the points follow a curve, form a circular shape, or lack a visible pattern. Correlation compresses the data into a single number, while the plot preserves the arrangement that determines what kind of pattern is present.
A small correlation also does not establish that the variables are independent. Independence is a stronger statement about the joint behavior of variables; a zero linear summary alone does not establish it. In particular, the exact relation \(y=x^2\) in the first example makes clear that one variable can be determined by the other even when their correlation is zero.
Check whether the pattern is roughly straight, curved, circular, or visually unstructured before relying on a single numerical summary.
Describe its size and sign as a summary of the overall linear association, not as a complete description of every pattern in the data.
If the scatterplot shows a U-shape or another curve, describe that pattern directly rather than calling the variables unrelated because \(r\) is small.
Common Mistake
Wrong reasoning: “The correlation is almost zero, so there is no relationship between these variables.” This treats Pearson’s correlation as a measure of every possible kind of relationship.
Correction: A small Pearson correlation indicates little overall linear association. The data may still follow a strong nonlinear pattern. Check the scatterplot, describe its shape, and interpret the correlation specifically as a linear summary. For example, the exact U-shaped measurements above have \(r=0\), but their values satisfy \(y=x^2\).
Check Your Understanding
Use the ideas in this tutorial to explain what a small correlation does and does not tell you.
- For the pairs \((-2,4),(-1,1),(0,0),(1,1),(2,4)\), explain why the correlation is zero even though the points follow a clear rule.
- What feature of a symmetric U-shaped pattern causes the terms in the correlation numerator to cancel?
- Points lie exactly on the circle \(x^2+y^2=16\) and are balanced around both axes. Can their Pearson correlation be zero? Explain why.
- A scatterplot has a clear curve, but its Pearson correlation is \(0.04\). Write one sentence that interprets the correlation without claiming that the variables have no relationship.
- Why is it misleading to report a small correlation without looking at the scatterplot?