Correlation: What It Means and How to Read the Coefficient

Correlation measures how closely two variables move together. The usual way to quantify it is the correlation coefficient, written r for a sample: it was devised by Karl Pearson in the early twentieth century, according to OpenStax.

What the Correlation Coefficient Measures

Pearson’s correlation coefficient runs from -1 to +1, according to the NIST/SEMATECH e-Handbook of Statistical Methods. A value of +1 means a perfect increasing relationship between the two variables, and a value of -1 a perfect decreasing one: when one variable rises, the other falls at a constant rate. According to OpenStax, the closer the value sits to either extreme, the tighter the linear relationship between the two variables. Anyone who wants to review how data is organized and described before reaching this point can start with the page on descriptive statistics; the bigger picture of how probability and statistics work together is on the page about probability and statistics.

The first move in any correlation analysis, before r is even calculated, is the scatter plot: the points drawn on a chart show at a glance whether they follow an upward trend, a downward one, or no recognizable pattern at all. This visual check is why the scatter plot is called the standard tool for spotting whether two variables are associated, according to OpenStax.

Limits of the Coefficient: Outliers and Non-Linear Relationships

Pearson’s coefficient measures only linear relationships, and that produces two practical limits. A value of r = 0 signals that no straight-line pattern shows up between the two variables; according to OpenStax, a curved pattern can still sit underneath it and go unregistered by the coefficient. A coefficient can even turn out statistically significant while the scatter plot shows a curve rather than a straight line: a limit built into the tool itself.

The second limit concerns isolated data points.

Worked example: in an OpenStax exercise on a real dataset, the correlation coefficient was 0.6631 with one unusual point included; removing that single point raised the coefficient to 0.9121. A practical rule of thumb for flagging an outlier, from the same source, is a point sitting more than two standard deviations above or below the best-fit line. Not every outlier should be dropped, though: some carry real information about the phenomenon under study and belong in the analysis.

Correlation and Causation

Even when it’s real and well measured, a strong correlation between two variables doesn’t show that one causes the other: it’s the principle known as correlation does not imply causation, according to OpenStax. The coefficient describes how closely two variables move together, but it stays silent on why they do. Deciding whether an observed relationship is genuine or a coincidence calls for the tools of probability, not the correlation coefficient alone.

Correlation vs. Regression

Correlation and linear regression start from the same data but answer different questions. Correlation measures the strength and direction of a linear association between an independent variable x and a dependent variable y. Linear regression, instead, finds the line that best fits that data: according to OpenStax, the method — called least squares — makes the sum of the squared errors between the observed points and the line as small as possible. Outliers, in this context, are precisely the data points sitting farthest from that line.

Correlation answers “how linked are these two variables.” Regression answers “what value would y take, given a certain x.”