Bootstrap Procedures for Testing Homogeneity Hypotheses

Before pooling data on eect sizes (a generic term for parameters of interest in the context of meta-analysis) from dierent studies, it is important to test for homogeneity of the eect sizes. A well known test for homogeneity is based on Cochran’s chisquare statistic. Our recent investigation showed that when the eect size of interest is a pairwise correlation, Cochran’s homogeneity test is inaccurate; it has a highly inated type I error probability, and hence cannot be recommended for practical use. However, we also noted that the homogeneity test for the correlations, performed after Fisher’s variance stabilizing z transformation, is quite accurate. In general, such a transformation is not known for every parameter of interest. A natural approach to try is then to use the bootstrap. We propose to investigate the accuracy of the bootstrap for testing the homogeneity hypothesis of two natural problems for which the Cochran’s test is known to be inaccurate.