In some correlational studies it is not reasonable to assume that bivariate observations are uncorrelated. An example would be a configural analysis in which two individuals are correlated across several variables (e.g. Q-technique). The present study was a Monte Carlo investigation of the robustness of techniques used in judging the magnitude of a sample correlation coefficient when observations are correlated. Empirical distributions of r, t, and Fisher's z were generated. Patterns of correlation were found which caused error rates to be as high as 20 when the nominal alpha was 05. A technique for controlling error rates in certain situations is suggested.