Pearson correlation does not strictly require normality to calculate the correlation coefficient itself, but normality becomes important when you want to test whether that coefficient is statistically significant. This distinction trips up a lot of people, and the short answer is: it depends on what you’re doing with the result.
What Normality Actually Affects
The Pearson correlation coefficient (r) is a mathematical formula. You can compute it on any dataset with two continuous variables, regardless of how those variables are distributed. The number you get back is still a valid description of the linear relationship in your sample. No normality required for that step alone.
The normality assumption kicks in when you want to generalize beyond your sample. If you’re running a hypothesis test to get a p-value, or constructing a confidence interval around r, the underlying math assumes the data come from a normally distributed population. Specifically, it assumes bivariate normality: the two variables together follow a bell-shaped distribution in two dimensions. When bivariate normality holds, the conditional mean of one variable is a linear function of the other, which is exactly the kind of relationship Pearson r is designed to detect.
There’s a stronger claim, too. The Pearson correlation coefficient fully characterizes the relationship between two variables only when the data are drawn from a multivariate normal distribution. In plain terms, if your data are bivariate normal, r tells you everything you need to know about how the two variables relate. If they’re not, r might miss important parts of the picture, like curved relationships or clusters.
The Assumptions That Matter Most
Normality gets the most attention, but it’s not the only prerequisite for a meaningful Pearson correlation. Here are the core assumptions:
- Continuous data. Both variables need to be measured on an interval or ratio scale. Pearson correlation is not designed for ordinal rankings or categorical labels.
- Linearity. Pearson r measures linear relationships only. If two variables have a strong curved relationship, r can come back close to zero even though the variables are clearly related. This is arguably the most important assumption. If it’s violated, the coefficient is misleading.
- Homoscedasticity. The spread of one variable should be roughly consistent across all levels of the other variable. If the scatter in your data fans out like a cone, the correlation may not accurately reflect the relationship’s strength at different points.
- No extreme outliers. Pearson r is calculated from actual data values, not ranks, which makes it sensitive to extreme observations. A single outlier can inflate or deflate r dramatically.
Of these, linearity is the biggest deal. A violation of linearity fundamentally undermines what the coefficient is trying to measure. Violations of normality, by contrast, mainly affect the reliability of p-values and confidence intervals rather than the coefficient itself.
How to Check Before You Calculate
Before running a Pearson correlation, it’s worth spending a few minutes on diagnostics. A scatterplot of the two variables is the single most useful check. It will reveal nonlinear patterns, outliers, and unequal spread faster than any formal test.
For normality specifically, you can look at each variable individually using a histogram or a Q-Q plot (a graph that compares your data’s distribution to a theoretical normal distribution). If the points on a Q-Q plot fall roughly along a straight diagonal line, your data are approximately normal. Formal tests like the Shapiro-Wilk test give you a p-value for the null hypothesis that your data are normally distributed, which can be helpful when visual inspection is ambiguous.
Keep in mind that with large samples (roughly 30 or more observations), the central limit theorem means the sampling distribution of r approaches normality even when the underlying data are somewhat skewed. So mild departures from normality are less of a concern with bigger datasets. With small samples, normality matters more because the p-values become unreliable faster.
When to Use Spearman or Kendall Instead
If your data clearly violate normality, contain extreme outliers, or show a monotonic but nonlinear relationship (one variable consistently increases as the other increases, just not in a straight line), Spearman’s rank correlation is the standard alternative. It works by converting your data to ranks and then calculating the correlation on those ranks, which makes it resistant to outliers and skewed distributions.
Spearman’s coefficient detects any monotonic relationship, not just linear ones. This makes it more flexible. In medical research, for example, a comparison showed that when data were not normally distributed, the Spearman coefficient detected monotonic trends more accurately than Pearson. If you find a weak Pearson correlation but a strong Spearman correlation, the relationship likely exists but isn’t linear.
Kendall’s tau is another rank-based option, generally preferred when you have a small sample size or many tied ranks. In practice, Spearman is more commonly reported, but both serve the same basic purpose: measuring association without requiring normality.
Practical Decision Framework
Here’s how to think about this in practice. If you just need a quick descriptive summary of how two continuous variables move together in your dataset, Pearson r works fine regardless of distribution. If you need to report a p-value or confidence interval, check your data first. Look at the scatterplot for linearity and outliers, check histograms or Q-Q plots for each variable’s distribution, and consider your sample size.
If both variables are roughly normal, the relationship looks linear, and you don’t have extreme outliers, Pearson correlation with its associated p-value is appropriate. If any of those conditions are clearly violated, switch to Spearman. Many researchers report both coefficients side by side, which gives readers a fuller picture: if the two values are similar, the relationship is likely linear and normality isn’t distorting the result. If they diverge, something interesting is happening in the data that deserves a closer look.