Karl G Jöreskog1, Stefan Mattson1
Different parameterisations of the ACE model are considered and their advantages and dissadvantages are discussed. It is shown that the commonly used latent variable (LV) parameterisation often give rise to computational problems in LISREL because of its non-unique solution. By contrast, the variance component (VC) parameterisation avoids these problems, is easier to apply, and gives correct standard errors of the estimated variance components due to genes, common and unique environment. The ACE model is often estimated by the maximum likelihood (ML) method assuming a normally distributed phenotype. However, this method can lead to incorrect standard errors and chi-square goodness-of-fit measures when the phenotype is non-normal. We report results of a simulation study that demonstrate the superiority of the asymptotically distribution free (ADF) method when the phenotype is non-normal. The case when the phenotype is measured only on an ordinal scale is also considered. This leads to complicated identification problems. It is shown how these can be resolved and that the relative variance components of genes, common and unique environment can be estimated by maximising simultaneous multinomial likelihoods of two or more observed contingency tables. The methods discussed are illustrated on real data.
Address: Department of Statistics, Uppsala University, P O Box 513, S-75120 Uppsala, Sweden, Phone +46 18 4711165, Fax +46 18 554422, E.mail Karl.Joreskog@Statisik.uu.se
1Department of Statistics, Uppsala University, Sweden