Choosing a correlation coefficient
Use Pearson for linear association, Spearman for association between ranks, and Kendall for concordance between pairs of rankings.
Fit
When it matters, and when it does not
It matters when
- You want to describe how two variables move together.
- You can inspect paired values and choose the relationship of interest before testing.
It matters less when
- You want to know whether two instruments give interchangeable readings. Use an agreement analysis. Read: Bland-Altman method comparison
- You need an adjusted relationship with an outcome. Consider a regression model. Read: Linear regression
Before you start
What to check in your study
- The scale of each variableDistinguish measured quantities from ordered categories.
- The shape of the relationshipPlot the values. A monotonic relationship keeps moving in one direction, even if it curves.
- Tied values and repeated observationsPlan how to handle ties and dependence within people.
- Each pair of values comes from an independent unit.How to check: Repeated values per person need a method that models that dependence.
- The interval method fits the coefficient and data.How to check: Bivariate normal assumptions concern usual Pearson inference. Choose a suitable alternative when those assumptions fail.
Plain language
How it works
Pearson summarizes a straight-line relationship using the measured values. It is sensitive to extreme observations and to the range of values sampled. It can be computed without normal data; its usual tests add distribution assumptions.
Spearman is Pearson correlation applied to ranks, using average ranks for ties. It describes a monotonic relationship rather than just a straight line. It is less sensitive to extreme magnitudes, but is not immune to problems.
Kendall compares concordant with discordant pairs: whether two observations have the same order on both variables. Tau-b adjusts for ties. None of these coefficients captures every curved pattern or establishes a causal effect.
Worked example
A worked example
- Setting
- Two instruments measure the same quantity.
- Question
- Does perfect association show that they agree?
- Approach
- Instrument A reads 1, 2 and 3 units. Instrument B reads 6, 7 and 8 units for the same samples.
- Result
- Their mean readings are 2 and 7 units: a difference of 5 units. Pearson and Spearman are both 1.
- What it means
- The instruments rank samples identically and follow a straight line, yet disagree by 5 units on every sample.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Correlation measures association, not agreement or cause. A coefficient near zero can hide a strong curved relationship.
- Choosing Spearman solely because a normality test fails.Instead: Decide whether the target is linear association or rank association. Inspect shape, tails and outliers.
- Comparing Kendall and Pearson by the same strength cut-offs.Instead: The coefficients use different scales of association. State the coefficient, its interval and the scatter plot.
Inside Mallard
How Mallard uses this
Mallard lists Choosing a correlation coefficient under Core research concepts.
concept.choosing_correlation, the id a plan cue records.Work it out
Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Comparing the Pearson and Spearman correlation coefficients across distributions and sample sizes: A tutorial using simulations and empirical data (opens in a new tab) · Psychol Methods · doi 10.1037/met0000079
- Asymptotic Properties of Spearman's Rank Correlation for Variables with Finite Support (opens in a new tab) · PLoS One · doi 10.1371/journal.pone.0145595
- [Mathematical treatment of rank correlation--a comparison of Spearman's and Kendall's coefficients] (opens in a new tab) · Z Psychol Z Angew Psychol