Genetic analyses using a structural equation modelling approach and full information maximum likelihood estimation with raw ordinal data were undertaken with the software package Mx [38]. Conventional tests of fit are not appropriate for this type of analysis; however, the relative difference in the fit of minus twice the log-likelihood between submodels is approximately distributed as χ2. Thus, it is possible to compare more complex models with simpler models, for example a model including gender differences in aetiology with one that includes no gender differences, and assess on the grounds of parsimony which is the best-fitting model. To test whether a model represents a good fit of the data, the model is compared to a ‘perfect’ fitting model, referred to as a saturated model, in which the variance and covariance of the data is freely modelled. If the model is not significantly different from the saturated model, it represents a good fit to the data [39].
