| Previous investigations of tests of mediation have primarily focused on single mediator models with mediated effects comprised of two paths. Using a more complex path model, a simulation study was conducted to evaluate several single sample and resampling (bootstrap) methods for testing mediation and contrasts of mediated effects. Mediated effects with two and three paths were tested, as were contrasts between pairs of two-path effects and contrasts of a two-path and a three-path effect. Confidence intervals were used to evaluate the power and Type I error rate of each method, and intervals were tested for coverage and balance of coverage. The single sample methods examined were the standard z test for mediated effects and a method based on the mathematical distribution of the product (the M test). The two resampling methods were the percentile bootstrap and the bias-corrected bootstrap. Using a path model with a single independent variable, three mediators and two outcome variables, all methods were applied to two-path indirect effects, and all but the M test were applied to three-path mediated effects and tests of contrasts of pairs of effects. The simulation study varied sample size (values: 50, 100, 200), number of paths in the mediated effects (2, 3), test used to evaluate effects (z, M test, percentile bootstrap, bias-corrected bootstrap), the combination of path coefficient effect sizes that made up each effects (effect sizes for each path: 0, .14, .39, .59), and the value of the contrast. Results indicated that the standard z test was the most conservative method for both types of mediated effects as well as contrasts. The M test was superior to the z and comparable to the percentile bootstrap but was limited to two-path effects. Both the percentile bootstrap and bias-corrected bootstrap had greater power and more accurate overall Type I errors for three-path effects and contrasts, but the bias-corrected bootstrap had too high Type I error in certain situations. The percentile bootstrap offered the best balance of flexibility, power, and Type I error accuracy. Confidence intervals were unbalanced for mediated effects, as in previous studies. |
