Pearson or Spearman: which correlation should you use?

They answer different questions, and the choice is usually settled by looking at a scatter plot.

Pearson

Measures linear association between two continuous variables. Use it when the relationship looks like a straight line, both variables are roughly interval-scaled, and there are no extreme outliers.

Pearson is highly sensitive to outliers. A single stray point can move it substantially, in either direction.

Spearman

Measures monotonic association: whether one variable consistently increases as the other does, without requiring the relationship to be a straight line. It works on ranks.

Use it for ordinal data such as Likert scales, for curved but consistently increasing relationships, or when outliers are present and legitimate. Because it uses ranks, an extreme value is just the highest rank rather than a large number.

Kendall

Also rank-based. Preferred over Spearman for small samples and where there are many tied ranks. Its values are typically lower than Spearman's for the same data, which is a property of the statistic, not a weaker relationship.

Always plot first

A scatter plot answers this question faster than any rule. It also catches the cases where no correlation coefficient is appropriate, such as a U-shaped relationship, which is strong but not monotonic, and will return a coefficient near zero from both Pearson and Spearman.

Partial correlation

When a third variable may be driving both, partial correlation gives the association between two variables with the third held constant.

In Plotara

Pearson, Spearman, Kendall and partial correlation, with a scatter or joint plot alongside so you can see the shape before choosing.