Commentary on Shand et al. (2011): Opioid use disorder as a condition of graded severity, similar to other substance use disorders


The upcoming transition to DSM-V has presented the substance abuse field with an opportunity to examine the diagnostic criteria for substance use disorders that have changed little since 1987, when DSM-III-R was published. DSM-III-R and DSM-IV described two substance use disorders, dependence and abuse, with dependence taking hierarchical priority over abuse. This hierarchy led to mistaken assumptions that abuse was mainly a prodromal condition to dependence, or that the criteria represented milder manifestations of the disorder than dependence criteria. The hierarchy also led to inconsistent and often low reliability and validity of the substance abuse diagnosis.

Several changes have been proposed for the DSM-V substance use disorders (available at; accessed 12 May 2010). One of the most important is elimination of the differentiation between abuse and dependence, with most criteria combined into a single disorder. While this has been examined extensively for alcohol [1–4], cannabis and other substances [5,6], little evidence has been available for opioids. Shand et al. have studied a large sample of a key clinical population to address this substance [7], filling an important gap in the literature.

Most studies addressing the structure of the abuse and dependence criteria have examined one of two questions: whether a class structure best fits the data (latent class analysis) or whether a single factor or dimension best fits the data (factor analysis). Latent class analyses have generally found classes differentiated by apparent severity. Factor analytical studies tended to find the best model fit for one- or two-factor solutions. Many investigators rejected two-factor solutions on the grounds of parsimony (fewer model parameters) and high between-factor correlations. After selecting unidimensional models of the abuse and dependence criteria, investigators then further explored important aspects of the individual criteria through item response theory analyses (e.g. item discrimination and severity, differential item functioning) or the criteria set as a whole (e.g. total information).

Shand and colleagues [7] pursued the issue of structure further. While they examined class and factor models, they additionally examined more complex structures involving both factors and classes, using factor mixture modelling (FMM) to investigate whether discrete classes or categories could be identified that had factors or dimensionality within the classes. Another model possibility not considered was a single factor with classes within the factor, a model also of potential interest that should be explored in future studies. Nevertheless, the FMM models studied by Shand et al. are essential to improving our understanding of the structure of the diagnostic criteria, and the authors are to be saluted for undertaking this work.

On the grounds of model fit indices, the authors selected one of the more complex models, consisting of two classes and a factor within the classes, and then investigated the relationship of this model to a set of covariates. Some questions could be raised concerning the authors' choice of the best model. Bayes' information criterion (BIC), the preferred model fit indicator, is actually slightly smaller for the one-factor, one-class (unidimensional) model than for the two-class, one-factor model, suggesting an equivalent or better fit for the unidimensional model. Further, the number of parameters is much lower for the one-factor, one-class model (22 parameters) than the one-factor, two-class model (34 parameters). Because model parsimony is often considered an important consideration, the advantage of the one-factor, two-class model here is not clear. Perhaps if the covariate analyses had indicated clearly qualitative differences in the classes, this would outweigh the parsimony disadvantage of the more complex, one-factor, two-class model. However, the classes seemed differentiated mainly on severity, not making an entirely convincing case for the added value of the classes given that each class also had a severity dimension within it.

The authors have also not been entirely clear about the implications of their results. In terms of treatment planning, should providers try to classify their patients into more and less severe groups, recognizing that severity differs within those groups, and assign them to different treatments? Should they assign different types of treatments to those in the different groups? For researchers, should the results be used in studies of aetiology or natural history? Informative answers to these questions would help to clarify the benefits of selecting a more complex model over a simpler, more parsimonious one.

Despite these concerns, the paper makes a very important contribution by bringing home convincingly the message that substance use disorders are graded in severity, even when the substance involved, an opioid, is almost always considered to indicate a severe problem, and when the clinical population is assumed commonly to be characterized by severe disorders. The results of the present paper for the one-factor, one-class (unidimensional) model suggest that this fairly simple, parsimonious approach can and should be used for opioids as well as for other substances.

Declaration of interest