Confidence intervals for correlations when data are not normal

Abstract

With nonnormal data, the typical confidence interval of the correlation (Fisher z') may be inaccurate. The literature has been unclear as to which of several alternative methods should be used instead, and how extreme a violation of normality is needed to justify an alternative. Through Monte Carlo simulation, 11 confidence interval methods were compared, including Fisher z', two Spearman rank-order methods, the Box-Cox transformation, rank-based inverse normal (RIN) transformation, and various bootstrap methods. Nonnormality often distorted the Fisher z' confidence interval-for example, leading to a 95% confidence interval that had actual coverage as low as 68%. Increasing the sample size sometimes worsened this problem. Inaccurate Fisher z' intervals could be predicted by a sample kurtosis of at least 2, an absolute sample skewness of at least 1, or significant violations of normality hypothesis tests. Only the Spearman rank-order and RIN transformation methods were universally robust to nonnormality. Among the bootstrap methods, an observed imposed bootstrap came closest to accurate coverage, though it often resulted in an overly long interval. The results suggest that sample nonnormality can justify avoidance of the Fisher z' interval in favor of a more robust alternative. R code for the relevant methods is provided in supplementary materials.