How to Understand Correlation

The direct answer: correlation is a number, usually called r, that describes the strength and direction of a straight line relationship between two variables. It ranges from minus one to plus one, where plus one means a perfect upward line, minus one means a perfect downward line, and zero means no linear relationship. This guide explains the formula, what the value does and does not tell you, and how to avoid the most common mistake students make with it.

Correlation at a Glance

The Core Idea

Correlation answers one narrow question: if you plot two variables on a scatter plot, do the points roughly form a straight line, and does that line slope up or down? It says nothing about curves, about cause, or about which variable came first. It is a descriptive tool for paired data.

A positive r means that as one variable tends to rise, the other tends to rise. A negative r means that as one rises, the other tends to fall. The closer r is to either extreme, the tighter the points hug a line. The closer r is to zero, the more the cloud of points spreads out. The NIST statistics handbook frames r as one number summarizing a scatter plot, with the same caveat that it only sees straight line patterns NIST.

The Pearson Formula

The sample Pearson correlation coefficient is:

r = sum over i of (x_i minus x bar)(y_i minus y bar) divided by sqrt( sum of (x_i minus x bar)^2 times sum of (y_i minus y bar)^2 )

A cleaner way to read it: r equals the covariance of x and y divided by the product of their standard deviations. The top of the fraction measures how x and y move together. The bottom rescales that movement so r always lands between minus one and plus one.

A small worked example

Work through four points: (1, 2), (2, 3), (3, 5), (4, 4).

The mean of x is (1 + 2 + 3 + 4) divided by 4 = 2.5. The mean of y is (2 + 3 + 5 + 4) divided by 4 = 3.5.

Compute each (x minus x bar)(y minus y bar):

• (1 minus 2.5)(2 minus 3.5) = (minus 1.5)(minus 1.5) = 2.25
• (2 minus 2.5)(3 minus 3.5) = (minus 0.5)(minus 0.5) = 0.25
• (3 minus 2.5)(5 minus 3.5) = (0.5)(1.5) = 0.75
• (4 minus 2.5)(4 minus 3.5) = (1.5)(0.5) = 0.75

The sum is 4.0.

Now the denominator pieces: sum of (x minus x bar)^2 = 2.25 + 0.25 + 0.25 + 2.25 = 5.0. Sum of (y minus y bar)^2 = 2.25 + 0.25 + 2.25 + 0.25 = 5.0. The square root of 5.0 times 5.0 is 5.0.

So r = 4.0 divided by 5.0 = 0.80. The relationship is a fairly strong positive linear one.

Correlation Is Not Causation

The most repeated warning in statistics is that correlation does not imply causation. Two variables can move together for several reasons. One may cause the other. A third hidden variable may drive both. Or the link may be coincidence in a small sample.

The ice cream example

Ice cream sales and drowning deaths rise together in summer. Eating ice cream does not cause drowning. Warm weather drives both. This example, used in many intro textbooks, shows why a high r alone never proves that changing one variable changes the other Khan Academy. The lagging of one series behind another, or a controlled experiment, is what turns a correlation into evidence about cause.

What r Does Not Capture

A correlation of zero means there is no straight line relationship. It does not mean there is no relationship at all. The points could form a clear curve, such as a U shape, with r near zero because the upward and downward pieces cancel. Always look at the scatter plot before trusting r.

Outliers and influential points

Outliers also distort correlation. A single extreme point can pull r strongly toward either end. Because r summarizes the whole data set in one number, it hides that sensitivity unless you plot the data. One point at (4, 40) in a small data set can flip a modest r to a large one, which is why you graph first and compute second.

Reading Strength Without Overclaiming

Students often want a label for each r. A common rough guide, and only a guide:

• 0.0 to 0.3: weak linear association
• 0.3 to 0.7: moderate linear association
• 0.7 to 1.0: strong linear association

These bands are rules of thumb, not laws. A weak correlation in a large medical data set can still be meaningful, while a strong correlation between two meaningless quantities tells you nothing useful. The size of r and the importance of the finding are separate questions, a point worth keeping in mind before you read hypothesis testing, where significance and effect size are likewise different.

A Second Kind of Correlation

Pearson r is not the only measure. When the relationship is monotonic but curved, the Spearman rank correlation often fits better because it looks at the order of the values, not the straight line form. Two variables can have a perfect Spearman correlation while Pearson r is well below 1, because Spearman ignores how far apart the points sit and only cares that one rises as the other rises.

Hand checking versus software

Working one r by hand, as in the four point example above, is the fastest way to see why the formula behaves. Once data sets grow past a dozen points, nobody computes r by hand. Spreadsheets and calculators do it, but the interpretation rules stay the same: graph first, read the sign, then judge the strength, and only then ask whether the link implies cause. Skipping the graph is how a single outlier quietly changes a conclusion.

Correlation Versus Regression

Correlation and regression are cousins, not twins. Correlation reports one number, r, summarizing the linear association. Regression goes further and fits a line, y = a + b x, that you can use to predict y from x. The slope b carries units, so it survives translation into plain language: "each extra hour of study is associated with b more points." r has no units and only describes fit.

When each is the right tool

Use correlation when you only want to know how tightly two things move together, such as screening which variables are worth modeling. Use regression when you need to estimate or predict, or when you want to control for other variables. The two share the same scatter plot, which is why graphing first serves both steps. A strong correlation with no clear reason for the link is a warning to stay away from causal language in either case.

Common Misconceptions

• Assuming a strong correlation means one variable causes the other.
• Treating r = 0 as proof of no relationship at all.
• Trusting r on a data set with a clear curve or a strong outlier.
• Mixing up the sign: a negative r is still a real, strong relationship, just downward.
• Reporting r for variables measured on different scales without checking the scatter plot.
• Calling a correlation "significant" when you mean "large." Significance is a separate test.

Frequently Asked Questions

What does an r of 0.5 mean in plain terms?

It means a moderate positive linear relationship. As one variable rises, the other tends to rise, but the points are fairly spread from a straight line.

Can r be greater than 1?

No. The formula rescales the covariance by the product of the standard deviations, so r is always between minus 1 and plus 1 inclusive.

Why is correlation called linear?

Because r only detects straight line patterns. Curved relationships need other tools such as polynomial regression or a look at the residual plot.

Does a low r mean the variables are unrelated?

Not necessarily. They could follow a strong curved pattern that a straight line summary misses. Always graph the data first.

What is the difference between r and rho?

r is the sample correlation you compute from data. Rho (the Greek letter) is the true population correlation you would get with complete data. Rho is usually unknown and estimated by r.

How do I know if a correlation is statistically real?

You run a hypothesis test on r, often using a t statistic, and report a p value. A significant p value says the linear association in the population is unlikely to be zero, not that the effect is large.

Sources

• NIST/SEMATECH e-Handbook, scatter plot and correlation
• Khan Academy, correlation and causality
• Boston University School of Public Health, correlation