Confidence Intervals for Correlation Coefficients
What You'll Learn
You'll learn how to build a confidence interval around Pearson's r to express the uncertainty in your correlation estimate—not just whether it's significant, but the plausible range of its true value.
Why a Confidence Interval for r?
When you calculate r = 0.65 from a sample, that's just a point estimate. The true population correlation could reasonably be 0.55 or 0.72. A confidence interval gives you a range that likely contains the true correlation, helping you understand estimation uncertainty.
The Problem: r Is Not Normally Distributed
Here's the catch: the sampling distribution of r is skewed, especially when the true correlation is far from zero. This makes standard confidence interval methods unreliable.
Fisher's z-Transformation to the Rescue
The solution is Fisher's z-transformation, which converts r into a value z' that is approximately normally distributed:
z' = 0.5 × ln[(1 + r) / (1 – r)]
Once transformed:
- Calculate the standard error: SE = 1 / √(n – 3)
- Build a confidence interval around z' using the normal distribution
- Transform back to the correlation scale using the inverse transformation
Example: For r = 0.60 with n = 50 and 95% confidence:
- Transform: z' ≈ 0.693
- SE = 1/√47 ≈ 0.146
- 95% CI for z': 0.693 ± 1.96(0.146) = [0.407, 0.979]
- Transform back: CI for r ≈ [0.39, 0.75]
The true correlation likely falls between 0.39 and 0.75—much more informative than just r = 0.60!
Key Takeaway: Fisher's z-transformation lets you construct valid confidence intervals for correlation coefficients by temporarily converting to a normally distributed scale, then transforming back to give a realistic range of plausible correlation values.