| dc.description.abstract | Correlations and correlation matrices are common in quantitative research. As researchers move away from p-values to reporting effect sizes and confidence intervals, appropriate confidence interval techniques for multiple, dependent correlations become more important. In this study, I investigated the performance of several confidence interval techniques for correlations (parametric z, Spearman rank-order, univariate bootstrap, and multivariate bootstrap) as applied to correlations in correlation matrices. Using simulation, I varied conditions typical in simulation research for correlations (sample size, population correlations, variable distributions) and conditions unique to multivariate settings (number of variables in study, variability of correlations in the population matrix). Further, the two bootstrap procedures were implemented in two ways: simultaneously (matrixwise) and independently (pairwise). Results for the pairwise techniques were consistent with previous research on individual correlations. Contrary to expectations, the matrixwise multivariate technique did not perform differently than the pairwise multivariate technique. The matrixwise univariate technique, which was a proposed, novel extension of the pairwise technique, did not perform well,and suggestions for improvements are made. Overall, the two bootstrap techniques performed best over a range of simulation conditions, with the univariate technique performing better at smaller sample sizes and the multivariate technique performing betterat larger sample sizes. Recommendations for researchers and future directions are discussed. | |
