The correlation coefficient is a number between -1 and +1 that tells you how strongly two things are related and in what direction. A value of +1 means they move perfectly together, -1 means they move in perfectly opposite directions, and 0 means there’s no linear relationship at all. It’s one of the most common statistics you’ll encounter in research, news headlines, and data analysis, and understanding what it actually tells you (and what it doesn’t) can change how you interpret information.
What the Number Tells You
The correlation coefficient, usually written as “r,” captures two things at once: the direction of a relationship and its strength. If r is positive, both variables tend to increase together. Height and weight, for example, generally have a positive correlation. If r is negative, one variable tends to go up while the other goes down. Hours of TV watched and physical fitness scores often show a negative correlation.
The strength of the relationship increases as r moves away from zero in either direction. An r of 0.85 represents a strong positive relationship. An r of -0.85 represents an equally strong relationship, just in the opposite direction. Zero sits in the middle, meaning the two variables have no predictable linear pattern between them.
How to Judge Weak, Medium, and Strong
A widely used set of benchmarks comes from the statistician Jacob Cohen. He proposed that an r of 0.10 represents a small correlation, 0.30 is medium, and 0.50 is large. These thresholds might seem surprisingly low if you’re expecting values close to 1, but in real-world data involving human behavior, health, or social patterns, correlations above 0.50 are genuinely strong findings. A correlation of 0.30 between exercise frequency and mood scores, for instance, would be considered a meaningful medium-sized relationship.
These benchmarks aren’t universal rules. In physics or engineering, where measurements are extremely precise, researchers might consider anything below 0.90 weak. In psychology or public health, where dozens of factors influence any outcome, a correlation of 0.25 can be practically important. Context matters more than rigid cutoffs.
Turning r Into Something More Intuitive
One of the most useful tricks for interpreting a correlation is squaring it. If you take r and multiply it by itself, you get what’s called the coefficient of determination, or r². This number tells you the percentage of variation in one variable that’s explained by the other.
Say the correlation between study time and exam scores is r = 0.70. That sounds pretty strong. But square it and you get 0.49, meaning study time accounts for about 49% of the differences in exam scores. The other 51% comes from things like prior knowledge, sleep quality, test anxiety, or natural aptitude. Even a correlation that feels large leaves a lot unexplained. A more modest r of 0.30 gives you an r² of just 0.09, meaning only 9% of the variation is accounted for. This is why researchers pay close attention to r² rather than just the raw correlation.
Positive, Negative, and No Correlation
If you plot your data on a scatterplot, the pattern of dots reveals the correlation visually. A positive correlation looks like dots rising from the lower left to the upper right. A negative correlation shows dots falling from the upper left to the lower right. No correlation looks like dots scattered randomly with no clear trend.
Here’s an important subtlety: r only measures linear relationships, meaning straight-line patterns. Two variables can have a very real, very strong relationship that’s curved, and r might come out close to zero. Imagine plotting the relationship between effort and performance. At first, more effort means better performance. But past a certain point, overworking leads to burnout and performance drops. That U-shaped or hill-shaped pattern is a genuine relationship, but a standard correlation coefficient would miss it entirely because the ups and downs cancel each other out.
Why You Should Always Look at the Data
A famous demonstration called Anscombe’s quartet shows exactly why a single number can be misleading. A statistician named Francis Anscombe created four completely different datasets that all produce the same correlation coefficient of 0.82 and the same regression line. In one dataset, the points follow a nice linear trend. In another, the relationship is clearly curved. In a third, all points fall on a perfect line except for one extreme outlier dragging the result off course. In the fourth, all the points cluster at one x-value except for a single distant point that creates the illusion of a trend.
Same r value, four wildly different stories. The lesson is that a correlation coefficient summarizes data, but it can’t replace actually looking at your data. A scatterplot takes seconds to make and can reveal patterns, outliers, or curved relationships that r alone would hide.
Sample Size Changes What Counts as Significant
A correlation of r = 0.40 means something very different if it comes from 15 data points versus 15,000. With a small sample, random chance can easily produce correlations that look impressive but are just noise. With a very large sample, even tiny correlations like r = 0.05 can be “statistically significant,” meaning they’re unlikely to be due to chance, while still being practically meaningless.
Larger samples reduce the impact of random error and increase precision, making it easier to detect real but small relationships. This is why researchers report both the correlation value and a p-value alongside it. The p-value tells you how likely it is that you’d see a correlation this strong by pure chance if no real relationship existed. But statistical significance doesn’t mean the relationship is important or useful. A correlation of 0.04 between shoe size and income might be statistically significant with a million data points, but nobody would use it to predict anything.
Correlation Does Not Mean Causation
This phrase gets repeated so often it almost loses its punch, but it captures a genuinely critical point. A high correlation between two variables doesn’t tell you that one causes the other. There are three common reasons two things can be correlated without a direct causal link.
- Reverse causation: You assume A causes B, but B actually causes A. Sleep quality and mood are correlated, but poor mood can disrupt sleep just as easily as poor sleep can lower mood.
- Confounding variables: A third factor drives both. Ice cream sales and drowning deaths are correlated because hot weather increases both, not because ice cream causes drowning.
- Coincidence at scale: With enough variables, some will correlate by pure chance. Per capita cheese consumption in the U.S. correlates with the number of people who die by becoming tangled in bedsheets. That tells you nothing meaningful.
Researchers have published studies linking childhood obesity to criminal behavior based on correlations, only to have others point out that the apparent link was driven by confounding factors like poverty, neighborhood environment, and access to resources. A strong r value can tempt people into drawing causal conclusions that the data simply can’t support.
Different Types for Different Data
The most common version, the Pearson correlation, works best when both variables are measured on a continuous scale (like weight, temperature, or test scores), the relationship between them is roughly linear, and there aren’t extreme outliers pulling the result in one direction.
When those conditions aren’t met, other options exist. The Spearman correlation handles data that’s ranked or ordered rather than precisely measured, like satisfaction ratings on a 1-to-5 scale. It also works better when the relationship is consistently increasing or decreasing but follows a curve rather than a straight line, and it’s more resistant to outliers. Kendall’s tau is another option for ranked data, often preferred when sample sizes are small. All three types produce values on the same -1 to +1 scale, so the interpretation stays the same even if the math behind them differs.
For most everyday encounters with correlations in news articles, health studies, or business reports, you’re looking at Pearson’s r. Knowing what it can and can’t tell you, checking whether the sample size is reasonable, and remembering to think about causation before jumping to conclusions will get you further than most people who just glance at the number.