Correlation coefficients
| Coefficient | Range | Assumptions | Use |
|---|---|---|---|
| Pearson r | −1 to +1 | Linear relationship, roughly normal | Linear correlation on continuous vars |
| Spearman ρ | −1 to +1 | Monotonic relationship | Rank correlation; robust to outliers |
| Kendall τ | −1 to +1 | Concordant / discordant pairs | Small-sample rank correlation |
| Point-biserial | −1 to +1 | One binary + one continuous | Binary vs continuous correlation |
| Phi φ | −1 to +1 | Two binary variables | 2×2 table correlation |
Interpretation of |r|
| |r| | Strength |
|---|---|
| 0.00 – 0.10 | Negligible |
| 0.10 – 0.30 | Weak |
| 0.30 – 0.50 | Moderate |
| 0.50 – 0.70 | Strong |
| 0.70 – 1.00 | Very strong |
Caveats
- Correlation ≠ causation. Confounders, selection bias, and reverse causality can all create spurious correlations.
- Pearson r only measures linear association — a perfect parabola can have r = 0.
- Outliers can drive r artificially high or low — always plot the scatter.
- Statistical significance (p-value) of r depends on sample size — big n makes tiny correlations "significant" without practical meaning.
- R² (coefficient of determination) = r² — proportion of variance explained.
Worked example: Pearson and Spearman by hand
Take x = 1, 2, 3, 4, 5 and y = 2, 4, 5, 4, 5. The means are 3 and 4. Deviations from the mean: x gives −2, −1, 0, 1, 2 and y gives −2, 0, 1, 0, 1. The sum of products is 4 + 0 + 0 + 0 + 2 = 6, the sum of squared x deviations is 10, and of y is 4 + 0 + 1 + 0 + 1 = 6. So r = 6 / √(10 × 6) = 6 / 7.746 = 0.775 and r² = 0.60: about 60 percent of the variance in y is linearly explained by x. On the table above that is “very strong”.
For Spearman, rank the data first. The y values 2, 4, 5, 4, 5 rank as 1, 2.5, 4.5, 2.5, 4.5 (ties share the average rank). Running the same calculation on ranks gives a sum of products of 7 and sum of squared rank deviations of 10 for x and 9 for y, so ρ = 7 / √(10 × 9) = 0.738.
Sample size matters. With only 5 points, r = 0.775 is not statistically significant at the 5 percent level: the critical value for 3 degrees of freedom is 0.878 (two-tailed). The same r from 30 points would be very significant. A parabola shows the other limit: x = −2 to 2 with y = x² has r = 0 even though y is perfectly determined by x.
Choosing a coefficient
- Use Pearson when the plot looks like a straight-line cloud and there are no extreme outliers.
- Use Spearman when the relationship is monotonic but curved, when the data are ranks, or when outliers are present.
- Use Kendall for small samples with many ties; it is slower to compute but its p-values are more reliable there.
- Always draw the scatter plot first. Very different datasets can share the same r.