The article by Cooke et al (Reference Cooke, Michie and Skeem2007) contains a number of fundamental modelling errors. First, the authors continue to present an over-factored model (i.e. hierarchical three-factor model with testlets), which results in negative variances. This 13-item model actually contains 10 factors: 6 first-order factors/testlets, 3 second-order factors and 1 third-order factor (simply count the number of circles/factors in Fig. 1). Any model can achieve good fit when it is as complex as the data it attempts to summarise. We have shown that this testlet model results in untenable parameters in four separate studies (Reference Neumann, Kosson and ForthNeumann et al, 2006). One author of the Cooke et al paper has also suggested that the testlet model is over-factored (Reference Skeem, Mulvey and GrissoSkeem et al, 2003). Cooke does not acknowledge this problem of an over-factored model, even though it is evident in his published work (see Reference Cooke and MichieCooke & Michie, 2001, Figs 2 and 3, which contain zero variance terms that the EQS program sets to zero when estimating negative variances). Cooke et al (Reference Cooke, Michie and Skeem2007) mention that we have criticised their use of testlets but they do not dispute that it creates a misspecified model with untenable parameters. Our analysis of the testlet model is available upon request.

Cooke et al provided a polychoric correlation matrix, ostensibly to give investigators the opportunity to replicate their findings. However, as noted in the EQS program manual, robust procedures can only be conducted with the raw items. Thus, the results reported by Cooke et al appear to be transparent but in reality no one will be able to unambiguously verify their analyses. When one analyses their published correlation matrix using a non-robust procedure, very different findings result. Also, Cooke et al relied upon a maximum likelihood procedure for estimating model parameters, despite the fact that it is well known that this procedure underestimates model parameters and model fit when used with ordinal data (Reference Everitt and DunnEveritt & Dunn, 2001) such as the items of the Psychopathy Checklist – Revised. There was no serious discussion on why robust maximum likelihood with polychoric correlations was employed, except that it is recommended in the manual for EQS version 6. None the less, the verisimilitude of this new approach is currently unknown. A program such as Mplus, which employs a robust weighted least-squares procedure for ordinal data is an accepted approach (Reference Neumann, Kosson and ForthNeumann et al, 2006). Cooke et al's use of Mplus was limited. Our Mplus analyses of the UK data along with our previously published findings can be found online (http://bjp.rcpsych.org/cgi/eletters/190/49/s39).

Contrary to Cooke et al, the four-factor model clearly fits as well or better than a viable three-factor model. Moreover, our recent research indicates that the four first-order factors are explained by a cohesive superordinate factor (Neumann et al, Reference Neumann, Kosson and Forth2006, Reference Neumann, Hare and Newman2007).