Effective use of Pearson's product-moment correlation coefficient

The calculation of correlation coefficients is widespread in biological research. Often, the null hypothesis of zero correlation is tested and/or confidence intervals for the correlation are computed. There are several different methods for this purpose; we compare the performance of different methods. According to our results the standard t test approach does offer generally reasonable performance even when the underlying distribution departs from bivariate normality. However, for non-normal data alternative techniques, especially the permutation test and using the RIN (rank-based inverse normal) transformation, offer better control of type I error and good power. With regard to confidence intervals, all investigated methods perform similarly; and there is no consistent pattern with which to strongly recommend one method over another. However, we show that two easy-to-calculate methods based on asymptotic results often perform tolerably well even for small sample sizes.