| Summary: | High levels of comorbidity between the anxiety and depressive disorders have raised questions about whether the diagnostic boundaries between these disorders need to be redrawn, or even whether both should be considered as different facets of a single disease process. Accordingly, latent class analysis has been used in several attempts to find data-driven groupings of individuals based on the symptoms of anxiety and depression. However, the assumption of conditional independence in this approach risks the extraction of spurious ‘severity classes’, making findings difficult to interpret. Factor mixture analysis relaxes that assumption by incorporating a common factor within each class, thereby overcoming the problem. This project investigated whether factor mixture analysis can suggest a data-driven classification of individuals based on the symptoms of common mental disorders. The analysis was based on pooled symptom-level data from three national psychiatric morbidity surveys of adults living in Great Britain carried out in 1993, 2000 and 2007. A comparison of the fit from the various latent variable models indicated that factor mixture models provided the best fit to the data, both in terms of model parsimony and goodness-of-fit. However, subsequent investigations suggested that the classes did not represent true groups in the population, but were rather accommodating violations of key assumptions in the standard factor model. Therefore, the results provide little guidance for revising the psychiatric classification. This is the first study to carry out an in-depth investigation into the interpretation of the extracted classes after applying factor mixture models to investigate the latent structure of mental disorders; its findings highlight the difficulties of interpreting the results of these models. Consequently, the thesis questions whether factor mixture models are actually useful for exploring the true nature of psychiatric disorders, and whether the present heavy use of such models is justified. An investigation of previously published examples suggests that their results may be prone to misinterpretation. The thesis concludes with a set of recommendations for the reporting of these models that may help to minimise the risks of such misinterpretation.
|
|---|
