The 6 Hidden Assumptions of Pearson’s Correlation (And How to Check Them)

Karl Pearson’s Coefficient of Correlation (r) is one of the first statistics we learn because the output is so simple: a single number between -1 and 1.

But that simplicity is a trap. If you feed Pearson’s formula the wrong kind of data, it will still calculate a number—it will just be mathematically meaningless. To ensure your correlation is actually telling the truth, your data must pass these six assumptions.

1. Continuous Variables

Pearson’s math requires both variables to be continuous (measured on an interval or ratio scale). They must be able to take any value within a range—like height, temperature, or test scores.

If even one of your variables is categorical (like “Yes/No”) or ordinal (like a 1st to 5th place ranking), the distances between your data points aren’t mathematically equal.

• The Fix: If you have ranked data, use Spearman’s Rank Correlation instead.

2. Paired Observations

This is a structural requirement of the formula: every data point must have a partner. For every observation of your independent variable ($X$), there must be a corresponding observation of your dependent variable ($Y$) from the exact same subject.

• The Rule: You cannot compute a correlation if one dataset has 12 observations and the other has 10. They must match perfectly.

3. Linearity

Pearson only measures straight lines. It assumes that as $X$ increases, $Y$ increases or decreases at a relatively constant rate.

If your data forms a curve—for example, the relationship between stress and performance, which forms a U-shape where moderate stress is good but extreme stress is bad—Pearson’s formula will get confused. It might spit out an r of 0, falsely tricking you into thinking there is no relationship at all, when in fact there is a very strong non-linear relationship.

4. No Extreme Outliers

While statistically there is no “rule” that data cannot contain outliers, Pearson’s r is incredibly sensitive to them.

Think of the line of best fit as a seesaw. A single extreme outlier (usually defined as a point lying more than 3 standard deviations away from the mean) has massive “leverage.” It can drag a completely flat line upward, artificially creating a strong correlation out of thin air.

Use this sandbox to see exactly how a single extreme outlier, or a change in the data’s shape, can instantly break Pearson’s math:

5. Homoscedasticity (Equal Variances)

The word comes from the Greek homo (same) and skedastikos (able to disperse). It simply means “equal scatter.”

For Pearson to work, the spread of your data points should be roughly the same all the way across the line of best fit. If you draw a boundary around your scatterplot, it should look like a uniform tube or a cigar.

If the variance changes—for example, if the points are tightly packed at the low end of the $X$-axis but fan out wildly at the high end—you have a violation called heteroscedasticity. Your scatterplot will look like a cone or a megaphone. This means your error rate is inconsistent, making the correlation coefficient unreliable for predictions.

6. Normality

Finally, both variables should roughly approximate a normal distribution (the classic bell curve). In normally distributed data, most data points hover close to the mean, with fewer points trailing off into the extremes. While Pearson is somewhat robust to minor violations of normality, heavily skewed data will distort your results.

The Golden Rule: You can verify almost all of these assumptions with a single step. Always generate a scatterplot before computing the correlation coefficient. It will instantly reveal non-linear curves, outliers, and megaphone-shaped variances that the raw math hides.