Abstract
Most analyses of continuous human twin data assume that twin values have a bivariate normal distribution, an assumption that is not always valid. Nonparametric and robust methods have been proposed for analysis of twin data, but the power of these tests has not been compared to corresponding parametric tests. We compare several tests for the simplified problem in which all tests are based on the absolute twin differences computed from each twin pair. We compare the performance of three likelihood-based tests, a method-of-moments-based test, two nonparametric tests, and a robust test for detecting genetic variance based on the absolute intratwin differences for the monozygotic and dizygotic twin groups. The performance of these tests is compared under three non–normal distributions for the original twin data in addition to the assumed bivariate normal distribution. Results indicate that the best test statistic depends heavily upon the relative sample sizes in the two zygosity groups. Generally, the Wald test is best for samples with less MZ twins, the likelihood-ratio test does best for equal sample sizes, and the score test is best for samples with more MZ twins. Several situations arise in which a method of moments-based test outperforms all likelihood-based tests.