Background. Expert committees of clinicians have chosen diagnostic criteria for psychiatric disorders with little guidance from measurement theory or modern psychometric methods. The DSM-III-R criteria for major depression (MD) are examined to determine the degree to which latent trait item response models can extract additional useful information.

Method. The dimensionality and measurement properties of the 9 DSM-III-R criteria plus duration are evaluated using dichotomous factor analysis and the Rasch and 2 parameter logistic item response models. Quantitative liability scales are compared with a binary DSM-III-R diagnostic algorithm variable to determine the ramifications of using each approach.

Results. Factor and item response model results indicated the 10 MD criteria defined a reasonably coherent unidimensional scale of liability. However, person risk measurement was not optimal. Criteria thresholds were unevenly spaced leaving scale regions poorly measured. Criteria varied in discriminating levels of risk. Compared to a binary MD diagnosis, item response model (IRM) liability scales performed far better in (i) elucidating the relationship between MD symptoms and liability, (ii) predicting the personality trait of neuroticism and future depressive episodes and (iii) more precisely estimating heritability parameters.

Conclusions. Criteria for MD largely defined a single dimension of disease liability although the quality of person risk measurement was less clear. The quantitative item response scales were statistically superior in predicting relevant outcomes and estimating twin model parameters. Item response models that treat symptoms as ordered indicators of risk rather than as counts towards a diagnostic threshold more fully exploit the information available in symptom endorsement data patterns.