Patterns of Comorbidity among Mental Disorders: A Person-Centered Approach

Participants

The NCS and NCS-R are two nationally representative surveys (Ns= 8098 and 9282, respectively) of mental health diagnoses conducted in the United States between 1990-1992 and 2001-2002, respectively, with response rates of 82.6% and 70.9%. Subjects in both surveys consist of noninstitutionalized participants (NCS age range: 15 – 54; NCS-R age range: 18 – 99). Further details regarding the recruitment, consent, and sampling strategy have been documented elsewhere [17-21].

Assessment of Mental Disorders

Lifetime diagnoses were utilized for all analyses of both the NCS and NCS-R datasets. Diagnoses for the two datasets are based on the World Mental Health Survey Initiative Version of the World Health Organization Composite International Diagnostic Interview (WMH-CIDI [22] a structured interview protocol. The version used in the NCS yielded DSM-IIIR [23] diagnoses and the version used in the NCS-R yielded DSM-IV [24] diagnoses. Further details regarding the assessment procedures for each sample are reported elsewhere [17, 19]. Given that the central aim of the study was to characterize patterns of comorbidity, non-hierarchical diagnoses were utilized.

Statistical Analyses

LCA models incorporating the design of the complex datasets including weighting, clustering, and stratification, were carried out using the Latent Gold 4.5 software package [25]. To account for the complex sampling design of the NCS and NCS-R datasets, models were estimated using weighted data, while controlling for stratification and clustering. Participants within each sample were assigned varying weights to adjust for survey nonresponding, variation in probability of selection within and between households, differences across successive phases of surveying, and to approximate the distribution of major demographic variables in the US population. This weighting procedure resulted in adjusted sample sizes (i.e., ‘weighted Ns’; see below) for each sample. Models specifying from 2 to 10 classes were compared. Model fit was assessed using the Bayesian Information Criterion (BIC) [26] and the Akaike Information Criterion (AIC) [27]. Both AIC and BIC are model selection indices that balance model fit and parsimony, but penalize model complexity to different degrees. When comparing models, generally, lowest values of both, or a scree-plot like test (i.e., where AIC and BIC values begin to level off) may be utilized to determine optimal model fit [28, 29]. Each model was run with 50 starting values to avoid problems with local maxima [30].

Four different sets of LCA models were examined including the following sets of disorders:

1. Mental disorders commonly used in structural models of epidemiological studies [1-3]: Disorders targeted for these LCA analyses consisted of conditions diagnosed most frequently in the general population: specific phobia, social phobia, agoraphobia, panic disorder, post-traumatic stress disorder (PTSD), major depressive disorder, generalized anxiety disorder (GAD), dysthymia, alcohol dependence, drug dependence, and conduct disorder. Separate models of these disorders were examined in both the NCS and NCS-R.

2. Next, LCA models for both samples were rerun after adding diagnoses of both bipolar disorderⁱ. In each case, profiles obtained from optimal LCA solutions were compared to the models obtained in the previous step.

Although the NCS and NCS-R contain 8098 and 9282 participants, respectively, targeted subsets of each were used to provide full representation of disorders of interest that were present in both samples, to verify if results replicated across both samples. In particular, PTSD, antisocial behavior in the form of conduct disorder, and alcohol and drug problems were available only for subsamples of the NCS and NCS-R. PTSD was assessed in 5877 and 5692 participants of the NCS and NCS-R, respectively. Similarly, in the NCS-R, alcohol and drug problems were assessed only in the subsample with 5692 subjects. Diagnoses of antisocial personality disorder were not a part of the publicly available dataset; however, lifetime diagnoses of conduct disorder were available for a subsample of the NCS-R (N = 3197), ranging from 18-44 years of age. In summary, to permit direct comparison of findings across the two samples, LCAs were performed using subsamples of participants from the NCS and NCS-R (unweighted/weighted Ns = 5,877/5,877 and 3,197/2,980, respectively) who had been assessed for conduct disorder, PTSD and substance use problems, along with other disorders of interest.

Following determination of the best-fitting LCA model in each subsample, parameter estimates of probability of diagnoses for specific disorders, along with standard errors, were compared across classes. Because estimated probabilities for diagnoses of certain disorders were zero or almost zero in some cases, it was not possible to calculate relative indices such as odds ratios or relative risk ratios. As an alternative to this, 95% confidence intervals (CI = estimated value ± [1.96 × standard error of estimate]) are presented for the estimated probabilities of being diagnosed with a particular disorder among individuals in each of the LCA classes.

Identification of Comorbidity Patterns using Latent Class Analysis

For the LCA models with just internalizing and externalizing disorders, in both the NCS and NCS-R samples, a 5-class model fit best according to BIC and AIC (see Table 1; note that AIC values level off after 5 classes for both models). For the LCA models with bipolar disorder, while the NCS revealed a 5-class solution again, in the NCS-R, the BIC indicated a 4-class model, while the AIC suggested a 5-class model. To reconcile this discrepancy, we examined bivariate residuals among disorders in this model (BVRs) [25], and employed the bootstrap likelihood ratio test (BLRT) [32] to formally assess the statistical significance of adding an extra class. BVRs indicate the strength of association between pairs of disorders, after accounting for those modeled by the latent classes; the presence of large BVRs in a model indicates that the model is not adequately capturing such associations. The BLRT evaluates the difference in log-likelihoods between two models by comparing it to an empirically estimated difference distribution generated by bootstrapping. An examination of BVRs among disorders in the 4-class model revealed several large significant associations among the internalizing disorders, suggesting that the 4-class model was not adequately accounting for all relationships among these disorders. Similarly, the BLRT indicated that the additional fifth class significantly improved model fit to data (p < .0001 with 500 replication samples) ⁱⁱ. Thus, given the weight of this evidence, the 5-class model was chosen as the best model in this set of LCA models as well.

Characterizing the Latent Classes

Inspection of the profiles for the NCS and NCS-R five-class LCA models (see figures 1, 2, 3, and 4) revealed the following. Class 1, the largest of the five classes, showed extremely low probabilities of diagnosis of any disorder. Class 2 was composed of individuals with modal diagnoses of phobias (all types) and panic disorder (to a lesser extent), in conjunction with depression. Class 3 consisted predominantly of subjects diagnosed with depression in conjunction with dysthymia, and also GAD. Class 4 was distinguished by a high prevalence of externalizing disorders (conduct disorder, alcohol dependence, drug dependence), in conjunction with lesser elevations on certain other disorders including depression. The fifth and final class was composed of subjects who demonstrated high levels of both internalizing (fear, distress) and externalizing (antisocial, substance-related) psychopathology, in addition to bipolar I disorder. Given these disorder profiles, we labeled the five classes as follows: few-disorders, fear, distress, externalizing, and multimorbid.

Disorders Common Across Classes

As Figures 1 - 4 illustrate, besides showing marked elevations in disorders most distinct to their class, individuals in the fear, distress, and externalizing classes also showed elevated prevalence of certain other disorders. In particular, across the two participant samples, depression, PTSD, and social phobia occurred in all three of these classes at high rates (> 10% probability, in all cases). Notably, marked elevations in prevalence rates for all disorders were evident in the highly multimorbid class, with the prevalence of some disorders (panic disorder and PTSD, in particular) higher in this class than in any other class across both samples.

Overlap between Internalizing and Externalizing Domains of Psychopathology

In addition to providing converging evidence for distinct domains of psychopathology marked by excessive fear or pervasive distress (internalizing disorders) or by deficient impulse control (externalizing disorders), the current work also sheds new light on a key unanswered question emerging from dimensional-structural studies: What accounts for the overlap between internalizing and externalizing disorders, reflected in moderate-to-high correlations among factors and subfactors in dimensional models of comorbidity?

The current findings point to two systematic sources for this overlap. One of these pertains to elevated rates within the fear, distress, and externalizing classes of what might be termed cross-class disorders—that is, disorders evident at elevated rates (exceeding estimated probabilities of diagnosis of .1) in each of these classes. This set of disorders included major depression, social phobia, and PTSD. The other source consists of a distinctive subgroup of individuals exhibiting heightened rates of all disorders from both the internalizing and externalizing domains—i.e., the multimorbid class. Across the two study samples, individuals in this class showed: (1) high rates of all phobic disorders (agoraphobia as well as social and specific phobias) – comparable to or higher than those evident in the fear class – along with markedly higher rates of panic disorder; (2) rates of depression, dysthymia, and GAD comparable to those observed in the distress class, along with the highest rates of PTSD compared to all classes; and (3) rates of conduct disorder, alcohol dependence, and drug dependence that substantially exceeded rates in other classes aside from the externalizing class. These observations, by highlighting major ways in which individuals in the fear, distress, and externalizing classes manifest particular disorders characteristic of other classes, suggest ways to account for the overlap that exists between internalizing and externalizing domains.

Homogeneity and Heterogeneity of Disorders

The current findings have implications for traditional conceptions of mental disorders as discrete, etiologically coherent entities. In line with findings from previous structural modeling studies [1-3], common mental disorders were found to co-occur reliably in distinctive patterns, consistent with the idea of shared etiologic underpinnings to differing disorders [37, 38]. As noted above, however, for the four comorbid classes in both the NCS and NCS-R, heightened rates of disorders characteristic of other classes were also evident, with one particular set of disorders (i.e., depression, social phobia, and PTSD) elevated across classes. Although somewhat speculative, this raises the possibility that mechanisms governing manifestations of particular disorders vary according to what class they occur in – i.e., as a function of the overall pattern of affiliated comorbidity.

Consider, for example, major depression, which emerged as one of the three most prevalent diagnoses in all five classes across both samples. While at odds with the traditional idea of depression as a coherent entity, further consideration reveals deeper ties to the extant literature. Longstanding debates exist regarding depression subtypes, and a number of schemes have been proposed including unipolar versus bipolar, episodic versus chronic, and reactive versus endogenous. The occurrence of depression in the context of differing patterns of comorbidity may indicate variability in the nature and etiology of depressive syndromes. Thus, whereas recurrent depression in the context of dysthymia and GAD may reflect a core dysphoric liability, episodes of depression in the context of chronic antisocial behavior and substance abuse may reflect the cumulative impact of adverse consequences over time. Indeed, in this context, it is worth noting that there have been recent calls to demarcate melancholia as a distinct mood disorder from major depression in the upcoming version of the DSM [39]. A similar argument can be made with regard to the syndrome of PTSD, which has been conceptualized as including both internalizing and externalizing variants [40, 41].

More broadly, these cross-class disorders may reflect characteristic output pathways along which psychological discomfort tends to be expressed in most individuals. In support of this notion, rates for these particular disorders were elevated even within the few disorders class in both samples. This suggests that these disorders may be more indicative of general distress or difficulty in dealing with adverse life events. For example, symptoms of depression represent a natural reaction to loss or defeat, symptoms associated with PTSD (e.g., persisting distress and wariness) are common following exposure to extreme stressors, and discomfort in situations involving social performance or evaluation is normative in the population. Thus, though our findings indicate groups of individuals with a preponderance of problems in distinct domains (possibly implying specific underlying vulnerabilities or diatheses along each of these lines), such individuals may be prone to greater levels of psychological discomfort in general, resulting in elevated rates of these common disorders across all classes identified in the current study.

Bipolar I Disorder and Internalizing and Externalizing Psychopathology

An interesting feature of our results pertains to the multimorbid class in the current study, which exhibited high rates of all relatively common disorders utilized in the LCA models, as well as bipolar I disorder. These results fit with data indicating that bipolar I disorder shows substantial comorbidity with internalizing and externalizing disorders of differing types [8, 9]. However, what is particularly noteworthy about this class is that it suggests that the presence of high levels of internalizing and externalizing psychopathology appears to be associated with more than just either.

Understanding Comorbidity from Multiple Perspectives

A final important point concerns how to interpret and understand the phenomenon of comorbidity utilizing results from previous factor analytic work and from work in the current study. As noted earlier, the purpose of the current study was an attempt to provide a complementary descriptive perspective on psychopathology by examining how comorbidity could be understood, when examined as clusters of persons occurring along dimensions of psychopathology. Results from the current study suggest that certain forms of psychopathology – i.e., major depression, social phobia, and PTSD – appear to “link” the dimensions, in addition to a multimorbid class. Thus, these findings could be interpreted from a purely mathematical and statistical point of view, without attributing a deeper meaning to the latent classes we found in the current study.

Alternatively, if one were to interpret the model used in the current study in the spirit of the latent variable model that was utilized, then the results might support the existence of latent classes of mental disorders. For example, as proposed by Jablensky, it could be hypothesized that each of the classes obtained in the current study are “syndromes” reflecting “facets of the same clinical entity” [42]. If one were to draw analogies from this perspective to medical disorders, applying a structural model to symptoms such as fever and headaches that are common to differing disorders such as bacterial meningitis and influenza, would yield correlated “fever” and “headache” factors. This would not mean that meningitis and influenza are the same, nor would it mean that the fever and headache symptoms are unimportant. On the contrary, it is valuable to know both what general syndrome is present so that the appropriate broadband treatment can be applied (e.g., antibiotics, antivirals, or other medication), as well as what individual symptoms (e.g., fever) require in terms of focal treatment. Thus, information from both levels (i.e., both the individual symptoms, as well as how they cluster together) is important to understand and treat the disorder. A similar case can be made for psychopathological syndromes as well.

A third alternative, as noted by Jaspers [43], and more recently suggested by Maj, is that psychopathology may indeed be “intrinsically heterogeneous”, with “ ‘true diseases’ (such as general paresis), which have clear boundaries among themselves and with normality; ‘circles’ (such as manic-depressive insanity and schizophrenia), which have clear boundaries with normality but not among themselves; and ‘types’ (such as neuroses and abnormal personalities), which do not have clear boundaries either among themselves or with normality” [44]. Indeed, results from the current study support such an interpretation as well in that bipolar disorder only appears to intersect with extremes of internalizing and externalizing psychopathology (i.e., the multimorbid class). Future research is doubtless required to understand what the true nature of psychopathology may be [45].

In this context of attempting to understand interrelations among psychopathology from differing perspectives, it is worth noting that there has been a recent effort to develop a “metastructure” of mental disorder clusters for the upcoming version of the DSM (see Psychological Medicine, December 2009 special issue [16, 46-50]). However, a point acknowledged by contributing authors [46] and highlighted in accompanying commentaries [51-53] was that further empirical evidence is required to support a move in this direction. Results from the current study could contribute to such efforts from the perspective of latent variable modeling. However, alternative methods to understanding comorbidity rather than latent variable modeling [54] have been proposed as well. Thus, it remains to be seen which approach will prove most useful with regard to nosological considerations of psychopathology.

In conclusion, our analyses importantly extend prior work by suggesting some possible reasons for overlap among differing dimensions of psychopathology. Ultimately, research that incorporates findings from multiple perspectives of interrelations among differing mental disorders is likely to contribute most to understanding, remediation, and prevention, of disorders of this type.