Content
Summary
When analyzing correlations between continuous variables, the correlation coefficient r serves as the effect size index, but its variance depends strongly on the correlation value itself. To ensure stable variance properties for meta-analysis, researchers transform r to Fisher’s z, perform the synthesis, and then convert the results back to r for presentation. This approach avoids the instability inherent in using the raw correlation metric directly.
Intuition
Think of the correlation coefficient r as a raw measurement that behaves unpredictably when you try to combine it with other studies; its variance changes depending on how strong the correlation is. To fix this, we use a mathematical 'lens' called Fisher’s z that stabilizes this variance, making the data consistent and comparable across different studies. We perform all our statistical heavy lifting in this stabilized z-world, and only at the very end do we translate our findings back into the familiar language of correlation coefficients (r) so that readers can easily understand the strength of the relationship.
Key equations
The approximate variance of the sample correlation coefficient r, where n is the sample size.
The transformation from sample correlation r to Fisher’s z.
The variance of Fisher’s z, where n is the sample size.
The standard error of Fisher’s z.
The conversion of Fisher’s z back to the correlation coefficient r.
Derivation
- Identify the effect size: The correlation coefficient r is the index for studies reporting correlations between two continuous variables.
- Recognize the problem: The variance of r depends strongly on the correlation value itself, making direct synthesis unstable.
- Apply transformation: Convert the sample correlation r to Fisher’s z using the formula z = 0.5 * ln((1 + r) / (1 - r)).
- Calculate variance: Use the variance formula for z, which is V_z = 1 / (n - 3), ensuring the variance does not depend on the correlation value.
- Perform synthesis: Apply meta-analytic techniques to the z values to obtain a summary effect and confidence intervals in the z metric.
- Convert back: Transform the summary z values back to the correlation scale using r = (e^{2z} - 1) / (e^{2z} + 1) for final presentation.
Worked example
Calculate the Fisher’s z transformation and its variance for a study with a sample size of 100 and a correlation coefficient of 0.50.
- Identify the given values: r = 0.50 and n = 100.
- Calculate Fisher’s z using the formula z = 0.5 * ln((1 + r) / (1 - r)). Substituting the values: z = 0.5 * ln((1 + 0.50) / (1 - 0.50)) = 0.5 * ln(3) ≈ 0.5493.
- Calculate the variance of z using the formula V_z = 1 / (n - 3). Substituting the values: V_z = 1 / (100 - 3) = 1 / 97 ≈ 0.0103.
- Calculate the standard error .
- Verify the conversion back to r: r = (e^(2*0.5493) - 1) / (e^(2*0.5493) + 1) ≈ 0.5000.
Answer For r = 0.50 and n = 100, the Fisher’s z is 0.5493, the variance V_z is 0.0103, and the standard error SE_z is 0.1015.
Common mistakes
- Confusing Fisher’s z with the z-score used in significance tests.
- Using the variance for the correlation (V_r) when working with Fisher’s z; instead, one must use the Fisher’s z score and its variance (V_z).