Structural equation modelling of twin data is performed in the flexible matrix algebra program Mx [28], which employs maximum likelihood modelling procedures. We fitted models to the raw ordinal data, where it is assumed that a normally distributed continuum of liability underlies the observed categories of the variables. Age and sex effects on the phenotypes were accounted for. The goodness-of-fit of a model is summarised by a statistic distributed as a chi-square (χ2). By testing the change in model fit (Δχ2) against the change in degrees of freedom (Δdf), we tested whether constraining parameters to zero or constraining them to be equal across groups, significantly worsened the model fit. This way we tested the significance of each pathway in the model.
