Type I error rates in multivariate behavioral genetic models

Numerical Type I error rates for certain commonly used multivariate twin models do not align with theoretical Type I error rates. This issue can have profound implications for hypothesis testing and statistical inference. ... [ view full abstract ]

Numerical Type I error rates for certain commonly used multivariate twin models do not align with theoretical Type I error rates. This issue can have profound implications for hypothesis testing and statistical inference.  Here we present two simulation studies to compare the Type I error rates for commonly used parameterizations of multivariate twin models.  In particular, both the Cholesky decomposition and correlated factors models show markedly reduced power to reject the null hypothesis when it is false.  By contrast, a model where the covariance matrices for additive genetic, common environment and specific environment variance components are estimated directly and without constraints yields Type I error rates that are consistent with expectations derived from hypothesis testing theory.   The loss of power is considerable and increases with the number of observed variables.  It appears that each model implied boundary, whether explicit or implicit, increases the discrepancy between the numerical and theoretical Type I error rates. Furthermore, imposing boundary conditions truncates the sampling distribution, and induces bias in the expected parameters. Implications for published research and limitations associated with directly estimating the relevant covariance matrices are discussed.