Calculate a Fisher Interval for Pearson Correlation

Identify the correlation and its sample

Record the two numerical variables, the population, and the number of complete paired observations used to calculate Pearson's r. The relevant n is the number of pairs contributing to that correlation, not the number enrolled or the sum of two variable-specific sample sizes.

Confirm that the coefficient is Pearson correlation rather than Spearman correlation, an intraclass correlation, or a partial correlation. These measures answer different questions and do not automatically share the same interval formula. Preserve the reported method and original precision when extracting a published result.

Inspect the data before the formula

Review the scatterplot for curvature, influential observations, subgroups, and dependence. The usual Fisher interval is an approximate procedure for Pearson correlation under suitable assumptions, including independent paired observations and an appropriate bivariate-normal model. It is not a guarantee for every dataset with an r value.

The official R correlation-test documentation describes the asymptotic Fisher transformation interval for Pearson correlation with at least four complete pairs. A summary-input calculator cannot check the scatterplot or diagnose repeated-measure dependence for you.

Recompute a fictional example

Suppose a fictional sample has r = 0.50 from 30 independent complete pairs. Fisher's transformed value is atanh(0.50), approximately 0.549306. Its approximate SE is 1/sqrt(30 − 3), approximately 0.192450.

For a two-sided 95% interval, subtract and add 1.959964 times that SE on the transformed scale, then apply tanh to both endpoints. The resulting interval for the correlation is approximately 0.1704 to 0.7290. Use the Fisher-correlation tool to reproduce these illustrative numbers.

Keep the uncertainty and limitations visible

The correlation audit worksheet should retain r, complete-pair n, confidence level, variable definitions, and assumption checks. Do not round r aggressively before conversion: a coarsely rounded input can change the reported interval, particularly near the boundaries.

The transform requires r strictly between minus one and one and n greater than three. A reported perfect correlation needs investigation rather than substitution of a nearby value just to make the formula run. Small samples can still give fragile approximate inference even when the arithmetic is defined.

Report the interval with the correlation and sample size, and distinguish association from causation or measurement agreement. A positive coefficient can reflect confounding, selection, or a restricted sample. A wide interval should not be disguised by describing the point estimate with a confident adjective. Finish with the supporting plot and a concise account of missing pairs, transformations, and any sensitivity analysis so readers can judge whether the interval answers the intended question.

Source checked on 2 November 2026; fictional Fisher interval independently recalculated.