Statistical & Mathematical Formulas
Fisher z-Transformation & 95% Confidence Interval for Correlation (r)
Stabilize skewed Pearson correlation sampling distributions via Ronald Fisher’s logarithmic transform z = 0.5 * LN((1 + r) / (1 - r)) and back-transform 95% confidence bounds r = (EXP(2*z) - 1) / (EXP(2*z) + 1).
Why raw Pearson r cannot use symmetric r +/- 1.96 * SE confidence intervals
The Fisher transform z = 0.5 × LN((1 + r) / (1 - r)) is useful for an approximate confidence interval for a Pearson correlation. Under suitable assumptions, including independent pairs and a bivariate normal model, the transformed estimate is approximately normal for large samples, with standard error approximately 1 / SQRT(N - 3). This is not an exact normal distribution or an exact 95% interval. Transform r, add or subtract 1.96 / SQRT(N - 3), and apply tanh to return to the correlation scale. For r = 0.65 and N = 28, z = 0.7753 and the approximate lower bound is 0.3656.
Build asymmetric 95% confidence intervals in three universal steps that require zero add-ins: (1) Transform r into Fisher z = 0.5 * LN((1 + r) / (1 - r)), (2) Shift z_low = z - 1.96 / SQRT(N - 3) and z_high = z + 1.96 / SQRT(N - 3), and (3) Back-transform via hyperbolic tangent r = (EXP(2 * z) - 1) / (EXP(2 * z) + 1). For r = 0.65 across N = 28 subjects (SE_z = 1 / SQRT(25) = 0.2000), Fisher z = 0.7753, z_low = 0.3833, and the exact 95% lower confidence bound is r_low = 0.3656 (staying strictly bounded inside [-1, 1] at both ends).
| Sample Correlation r (N = 28, SE_z = 0.20) | Fisher z = 0.5 * LN((1+r)/(1-r)) | 95% Lower Bound r_low | 95% Upper Bound r_high |
|---|---|---|---|
| Moderate Correlation (r = 0.65) | 0.7753 | 0.3656 | 0.8234 |
| Strong Correlation (r = 0.80) | 1.0986 | 0.6085 | 0.9034 (Never > 1.00) |
| Boundary Singularity (|r| >= 1 or N <= 3) | Logarithm / DoF Guard | "Invalid r or Sample N <= 3" | Requires -1 < r < 1 and N >= 4 |
Exact Logarithmic Fisher z-Transformation = 0.5 * LN((1 + r) / (1 - r))
Evaluates Fisher’s variance-stabilizing transform using pure LN so the formula works in every spreadsheet engine.
=IF(COUNT(B2:B3) <> 2, "Invalid r or Sample N <= 3", IF(OR(B2 <= -1, B2 >= 1, B3 <= 3, B3 <> INT(B3)), "Invalid r or Sample N <= 3", ROUND(0.5 * LN((1 + B2) / (1 - B2)), 4)))Approximate 95% Lower Confidence Bound for Pearson Correlation r
Subtracts 1.96 / SQRT(N - 3) in Fisher z-space and back-transforms via (EXP(2 * z_low) - 1) / (EXP(2 * z_low) + 1).
=IF(COUNT(B2:B3) <> 2, "Invalid r or Sample N <= 3", IF(OR(B2 <= -1, B2 >= 1, B3 <= 3, B3 <> INT(B3)), "Invalid r or Sample N <= 3", ROUND((EXP(2 * ((0.5 * LN((1 + B2) / (1 - B2))) - (1.96 / SQRT(B3 - 3)))) - 1) / (EXP(2 * ((0.5 * LN((1 + B2) / (1 - B2))) - (1.96 / SQRT(B3 - 3)))) + 1), 4)))Try the example
Use the original values or change them to test your understanding.