SIGNIFICANCE TESTING OF INTERCLASS CORRELATIONS FROM FAMILIAL DATA

In the estimation of interclass correlations from familial data, the pairwise estimator has been recently shown to be most effective. Significance testing of this parameter has been based upon either (1) the classical method, whereby one degree of freedom is attributed to each pair, or (2) the conservative or (3) sib mean methods, whereby one degree of freedom is attributed to each family. In this report, we show that use of the classical pairwise method of significance testing, although the most widely used procedure, provides overstated type I errors yielding estimated significance levels two to five times larger than nominal levels. We further show that the conservative test procedure provides drastically understated type I errors with a resulting loss of power against many alternatives of interest. We propose a new test procedure, viz. the adjusted pairwise procedure, where the effective number of degrees of freedom in a family are estimated as a function of both the number of siblings in the family and the estimated sib-sib correlation. The aggregate degrees of freedom are then summed up over all families. Finally, this new method is shown by Monte Carlo simulation to give appropriate type I errors which are similar to those obtained by the sib mean method. The adjusted pairwise procedure, however, is shown to be more powerful than the sib mean method and therefore is the method of choice in testing interclass correlations from familial data.