Despite its prominence in clinical research, the concept of comorbidity is heavily contested (e.g., Lilienfeld et al. Reference Lilienfeld, Waldman and Israel1994). In the target article, Cramer et al. contend that the major problem with comorbidity research stems from its widespread adoption of standard latent variable theory, which holds that latent variables are common causes that explain the correlations between the manifest variables to which they give rise. On this view, comorbidity is a bidirectional relation between latent variables that causally produce correlated indicator variables or symptoms. The problem arises from the model's acceptance of the statistical assumption of local independence, which says that the manifest indicators are solely due to the influences of the common causes; they cannot relate to one another causally. We are unsure how widespread in comorbidity research this latent variable theory actually is, but we agree with Cramer et al. that its inability to allow manifest variables or symptoms to be causally related is a major limitation of its conceptualization of comorbidity.
Cramer et al.'s alternative conceptualization of comorbidity as complex networks of causally related symptoms holds considerable promise for resolving some troubling issues in the field, and the authors point to several advantages that this perspective holds. A major achievement of the target article is its demonstration that network theory has the resources to help resolve several syndromal questions about comorbidity (see also Meehl Reference Meehl2001).
Cramer et al.'s network depiction of comorbidity has no place for latent variables as common causes. In fact, as it is currently formulated, their model has no place for latent variables at all. The result is a “flat” model of comorbidity that embraces only manifest symptoms in causal relation to one another. However, this particular feature of the model should be seen as a contingent feature of its initial formulation, not as an expression of an unyielding empiricist commitment to focus on the phenomenology of psychopathology only. The authors make clear that their network characterization of comorbidity is necessarily incomplete. They explicitly acknowledge that relations between symptoms are likely to be mediated by other unobservable (i.e., latent) factors in causal chains. Of course, to embrace latent factors in an extended network model would not be to admit common cause latent variable theory back into the fold. But it is perhaps worth noting here that there are alternative nonstandard models within latent variable theory that permit the specification of different relations between latent and/or manifest variables. For example, Keith Markus has suggested to us in personal communication that there are no good theoretical and statistical reasons why one could not insert direct causal paths between indicator variables in a structural equation model.
We think that further research on network formulations of psychopathologies is justified on strategic methodological grounds, and should now be extended to explicitly incorporate reference to etiological factors at whatever level of specificity can be achieved. Despite the use of latent variable models in research on psychopathology, there is a general distrust of theorizing about latent causes by researchers in the field. The largely atheoretical nature of the DSMs III and IV (Diagnostic and Statistical Manual of Mental Disorders, 3rd and 4th editions; American Psychiatric Association 1987; 1994), for example, came about because their task forces believed that good causal theories of psychopathology were seldom to be had. We think the construction of explanatory theories of comorbid conditions should be pursued with vigour. Good theories are to be valued because they are our primary vehicles for understanding the relevant syndromal facts. However, this will require the implementations of sound theory construction strategies. Psychology, with its penchant for the weak testing of austere theories via tests of statistical significance, and an overreliance on goodness-of-fit measures of empirical adequacy, has been somewhat remiss in this regard. Relatedly, there has been a general failure to acknowledge that good theories in science are often generated by abductive or explanatory means in order to explain established phenomena, developed through a strategy of analogical modeling, and evaluated on multiple criteria, some of which have to do with the explanatory worth of theories (Haig Reference Haig2005; for an application of abductive theory construction methods to clinical reasoning and case formulation, see Vertue & Haig Reference Vertue and Haig2008).
In a preliminary test of their hypothesis about the importance of bridge principles in explaining comorbidity, Cramer et al. use the Akaike information criterion to judge the fit of competing models. This can be seen as one approach to the widespread practice in science of testing hypotheses about models for their predictive accuracy. However, the inclusion of latent factors in a network model of comorbidity raises the question of how one should evaluate those models when they contain explanatory hypotheses. Cramer et al. rightly point out that common cause latent variable theories have modest explanatory power. However, network theories that appeal to latent causes more generally have the potential to offer more powerful explanations. Their evaluation will have to combine information about postulated causes (preferably in the form of causal mechanisms), correlations, and competing causal accounts. An approach to theory evaluation that can do this is known as inference to the best explanation (see Haig Reference Haig2009). With inference to the best explanation, the ideas of explanation and evidence come together, and explanatory reasoning becomes the basis for evaluating theories: The explanatory goodness of theories counts in their favour. Conversely, the explanatory failings of theories detract from their credibility. According to Thagard (Reference Thagard1992), inference to the best explanation is essentially a matter of establishing relations of explanatory coherence between propositions within a theory. To infer that a theory is the best explanation is to judge it as more explanatorily coherent than its rivals. Theories depicted as networks of propositions, including network models of comorbidity, lend themselves naturally to evaluation in terms of considerations of explanatory coherence.
In conclusion, we think the long-term prospects of network theory in psychopathological research are good precisely because it can exploit the considerable conceptual and investigative resources of dynamical systems theory. Cramer et al.'s dynamical systems conceptualization of comorbidity can study change in individuals over time. It therefore promises to feature in a future clinical psychology that has moved beyond the strictures of the present DSM model to embrace, among other things, an idiographic, process-oriented approach to scientific research on comorbidity and other psychopathologies.
You have
Access
Despite its prominence in clinical research, the concept of comorbidity is heavily contested (e.g., Lilienfeld et al. Reference Lilienfeld, Waldman and Israel1994). In the target article, Cramer et al. contend that the major problem with comorbidity research stems from its widespread adoption of standard latent variable theory, which holds that latent variables are common causes that explain the correlations between the manifest variables to which they give rise. On this view, comorbidity is a bidirectional relation between latent variables that causally produce correlated indicator variables or symptoms. The problem arises from the model's acceptance of the statistical assumption of local independence, which says that the manifest indicators are solely due to the influences of the common causes; they cannot relate to one another causally. We are unsure how widespread in comorbidity research this latent variable theory actually is, but we agree with Cramer et al. that its inability to allow manifest variables or symptoms to be causally related is a major limitation of its conceptualization of comorbidity.
Cramer et al.'s alternative conceptualization of comorbidity as complex networks of causally related symptoms holds considerable promise for resolving some troubling issues in the field, and the authors point to several advantages that this perspective holds. A major achievement of the target article is its demonstration that network theory has the resources to help resolve several syndromal questions about comorbidity (see also Meehl Reference Meehl2001).
Cramer et al.'s network depiction of comorbidity has no place for latent variables as common causes. In fact, as it is currently formulated, their model has no place for latent variables at all. The result is a “flat” model of comorbidity that embraces only manifest symptoms in causal relation to one another. However, this particular feature of the model should be seen as a contingent feature of its initial formulation, not as an expression of an unyielding empiricist commitment to focus on the phenomenology of psychopathology only. The authors make clear that their network characterization of comorbidity is necessarily incomplete. They explicitly acknowledge that relations between symptoms are likely to be mediated by other unobservable (i.e., latent) factors in causal chains. Of course, to embrace latent factors in an extended network model would not be to admit common cause latent variable theory back into the fold. But it is perhaps worth noting here that there are alternative nonstandard models within latent variable theory that permit the specification of different relations between latent and/or manifest variables. For example, Keith Markus has suggested to us in personal communication that there are no good theoretical and statistical reasons why one could not insert direct causal paths between indicator variables in a structural equation model.
We think that further research on network formulations of psychopathologies is justified on strategic methodological grounds, and should now be extended to explicitly incorporate reference to etiological factors at whatever level of specificity can be achieved. Despite the use of latent variable models in research on psychopathology, there is a general distrust of theorizing about latent causes by researchers in the field. The largely atheoretical nature of the DSMs III and IV (Diagnostic and Statistical Manual of Mental Disorders, 3rd and 4th editions; American Psychiatric Association 1987; 1994), for example, came about because their task forces believed that good causal theories of psychopathology were seldom to be had. We think the construction of explanatory theories of comorbid conditions should be pursued with vigour. Good theories are to be valued because they are our primary vehicles for understanding the relevant syndromal facts. However, this will require the implementations of sound theory construction strategies. Psychology, with its penchant for the weak testing of austere theories via tests of statistical significance, and an overreliance on goodness-of-fit measures of empirical adequacy, has been somewhat remiss in this regard. Relatedly, there has been a general failure to acknowledge that good theories in science are often generated by abductive or explanatory means in order to explain established phenomena, developed through a strategy of analogical modeling, and evaluated on multiple criteria, some of which have to do with the explanatory worth of theories (Haig Reference Haig2005; for an application of abductive theory construction methods to clinical reasoning and case formulation, see Vertue & Haig Reference Vertue and Haig2008).
In a preliminary test of their hypothesis about the importance of bridge principles in explaining comorbidity, Cramer et al. use the Akaike information criterion to judge the fit of competing models. This can be seen as one approach to the widespread practice in science of testing hypotheses about models for their predictive accuracy. However, the inclusion of latent factors in a network model of comorbidity raises the question of how one should evaluate those models when they contain explanatory hypotheses. Cramer et al. rightly point out that common cause latent variable theories have modest explanatory power. However, network theories that appeal to latent causes more generally have the potential to offer more powerful explanations. Their evaluation will have to combine information about postulated causes (preferably in the form of causal mechanisms), correlations, and competing causal accounts. An approach to theory evaluation that can do this is known as inference to the best explanation (see Haig Reference Haig2009). With inference to the best explanation, the ideas of explanation and evidence come together, and explanatory reasoning becomes the basis for evaluating theories: The explanatory goodness of theories counts in their favour. Conversely, the explanatory failings of theories detract from their credibility. According to Thagard (Reference Thagard1992), inference to the best explanation is essentially a matter of establishing relations of explanatory coherence between propositions within a theory. To infer that a theory is the best explanation is to judge it as more explanatorily coherent than its rivals. Theories depicted as networks of propositions, including network models of comorbidity, lend themselves naturally to evaluation in terms of considerations of explanatory coherence.
In conclusion, we think the long-term prospects of network theory in psychopathological research are good precisely because it can exploit the considerable conceptual and investigative resources of dynamical systems theory. Cramer et al.'s dynamical systems conceptualization of comorbidity can study change in individuals over time. It therefore promises to feature in a future clinical psychology that has moved beyond the strictures of the present DSM model to embrace, among other things, an idiographic, process-oriented approach to scientific research on comorbidity and other psychopathologies.