Latent Class Analysis of Lifetime Depressive Symptoms in the National Comorbidity Survey

The National Comorbidity Survey (NCS) was a probabalistic survey of the U.S. population with a response rate of 82.4% (17). The NCS had two phases: a part 1 diagnostic interview (N=8,098) and a part 2 risk factor interview administered to a subsample (N=5,877) because of budgetary constraints. The part 2 interview was administered to all part 1 subjects aged 15–24 years, to all other subjects in part 1 with a lifetime diagnosis, and to a random subsample of the remaining subjects in part 1. The sampling and weighting procedures are described in detail elsewhere (17, 18). After complete description of the study, written informed consent was obtained.

In the NCS, depressive symptoms from the participants’ worst lifetime episodes were assessed in the part 1 interview with the Composite International Diagnostic Interview (19, 20). Because of our interest in replication of our findings in the Virginia Twin Registry study, we took an approach as similar as possible to the approach in our previous report (16). To be included in the latent class analysis (LCA) in our present study, the worst lifetime episodes from the NCS (N=2,836) must have lasted at least 2 weeks, been associated with help-seeking or impairment, and contained one or more contemporaneous depressive symptoms. These depressive syndromes may or may not have met DSM-III-R criteria for major depression. To delineate typical and atypical symptoms, we “disaggregated” the nine DSM-III-R “A” criteria for major depression into 14 individual symptoms (table 1).

We used latent class analysis (LCA) (21-23) to determine empirically the typologies of depressive symptoms in the NCS. LCA is a “categorical analogue” to factor analysis and is particularly appropriate for data on the presence or absence of symptoms. Using a program by Eaves et al. (22), we applied LCA to a 2,836×15 data matrix in an iterative manner. The rows corresponded to the 2,836 NCS participants with one depressive symptom or two or more co-occurring depressive symptoms. The 15 columns corresponded to responses of each subject to the 14 disaggregated major depression criteria and the subject’s gender. We included gender to allow for the possibility of gender-related differences in symptoms. We first fit a one-class solution followed by a two-class solution, and so on, until we reached the “best” solution. We determined the best solution with reference to two criteria. First, the solution had to fit the data significantly better than had the previous class. The difference between the log-likelihood of the previous and the current class approximates a chi-square distribution; if this difference was greater than the critical chi-square statistic (χ²=26.3 for alpha=0.05, df=16), then the current class provided a better fit to the data than had the previous class. Second, the Numerical Algorithms Group subroutine (E04UCF) (24) used by the LCA program for maximum likelihood minimization had to reach a valid solution (i.e., the E04UCF IFAIL parameter had to return as zero). We had relatively clear evidence in favor of a six-class solution as it was significantly better than all lower-order solutions and no seven-class solution was superior. Individual subjects were then assigned class membership based on the likelihood of their particular response profile.

The next step was to determine whether there were important differences across the LCA classes. We used six sets of external validators not entered into the LCA: characteristics of the depressive disorder, symptom consequences, demography, comorbid psychiatric disorders, personality and attitudinal self-report data, and parental psychiatric history.

The first four sets of validators were assessed in the part 1 interview. The Comprehensive International Diagnostic Interview/DSM-III-R diagnoses available were major depression, bipolar disorder, conduct disorder, antisocial personality, panic disorder, agoraphobia without panic disorder, generalized anxiety disorder, social phobia, simple phobia, alcohol dependence, and non-alcohol/non-nicotine drug dependence (17). The part 2 interview (N=5,877) yielded the diagnosis of posttraumatic stress disorder (PTSD) (18) and empirically abbreviated forms of several personality and attitudinal measures: neuroticism, extraversion, and openness to experience (25); internal and external locus of control (26); self-esteem and self-reliance (27); fatalism (28); and interpersonal dependency and interpersonal conformity (29). All scores were standardized before analysis.

We evaluated a history of depressive disorder, alcoholism, and drug use disorder for the biological parents of the respondent in the part 2 interview by using standardized criteria (30, 31).

The goal of these analyses was to attempt to validate LCA class membership. We were particularly interested in whether the classes found in the Virginia Twin Registry (VTR) sample (16) could be replicated. We first compared the six LCA classes in terms of depressive symptoms, characteristics, and symptom consequences. We then focused on demography, psychiatric comorbidity, personality and attitudes, and parental psychiatric history in reference to a comparison group of the 5,262 NCS subjects not entered into the LCA because they denied lifetime depressive symptoms. We used logistic regression and multiple regression models (32) including the NCS weighting variables and controlling for the effects of age, race, and gender.

These analytic aims necessitated a large number of statistical comparisons and, given the large number of validators, the overall type I error was inflated. We employed a two-tailed significance level of 0.01 to compensate partially for the number of comparisons. More importantly, as the major aim was to validate the empirically derived LCA classes, the pattern of differences among the validators (rather than an isolated difference) had to support its existence as a valid entity.

Table 1 depicts the observed class membership and endorsement frequencies of the 14 disaggregated DSM-III-R depressive symptoms in the best-fitting LCA solution. These data are from the 2,836 participants (35.0%) in the NCS with one or more contemporaneous depressive symptoms in the lifetime worst episode. Class 1 (4% of total NCS participants) was characterized by a very high lifetime occurrence of major depression (98%) and high endorsement rates for most of the 14 disaggregated symptoms of major depression. We termed class 1 “severe typical” because of the nearly universal presence of major depression and the preponderance of classical depressive symptoms (i.e., weight and appetite loss, insomnia, psychomotor retardation, anergia, and poor concentration). Class 2 (6% of total NCS participants) was also characterized by classical depressive symptoms, although fewer subjects (63%) met criteria for major depression. We thus termed class 2 “mild typical.” Like class 1, class 3 (2% of total NCS participants) was characterized by many depressive symptoms, and 100% of this class met criteria for major depression; however, the symptom pattern was more characterized by appetite increase and weight gain. Both insomnia and hypersomnia were prevalent. We termed class 3 “severe atypical.” Like class 3, class 4 (4% of total NCS participants) was characterized by atypical depressive symptoms with 63% meeting criteria for major depression. We termed class 4 “mild atypical.” Class 5 (labeled “intermediate”) was the most prevalent (11% of total NCS participants) and was characterized by intermediate occurrence of major depression (52%) with prominent symptoms of low mood, loss of interest, insomnia, loss of energy, and thoughts of death. Class 6 (8% of total NCS participants) had low endorsement rates for all depressive symptoms except the stem question (whether the subject felt depressed) and a very low occurrence of major depression (1%, labeled “minimal symptoms”). The first four classes were predominantly female, whereas the gender ratios for the last two classes were equal.

Table 2 depicts the syndrome characteristics for the six empirically derived LCA classes. Age at onset was similar across groups, although the severe typical and severe atypical classes tended to be associated with recurrent and persistent episodes of mood disturbance accompanied by considerable behavioral and syndrome consequences.

Table 3 presents descriptive data for demography, comorbidity, personality/attitudes, and parental psychiatric history of members of the six LCA classes in comparison with NCS subjects who denied depressive symptoms (N=5,262). There were highly significant differences across these seven groups for all of these variables except age at interview and fatalism.

The six empirically derived LCA classes can be considered as the four combinations of atypical/typical symptoms and severe/mild symptoms as well as the intermediate and minimal symptoms classes.

In the following analysis, we focused on four LCA classes (severe atypical, mild atypical, severe typical, and mild typical) to understand more clearly the influence of atypicality and severity. We created three variables that were used in the subsequent analyses. “Atypicality” coded the presence/absence of atypical symptoms (1=severe and mild atypical classes, 0=otherwise), “severity” coded the presence/absence of severe symptoms (1=severe atypical and typical classes, 0=otherwise), and the interaction term was the product of the atypicality and severity variables (1=severe atypical, 0=otherwise).

Table 4 summarizes these analyses. The main effect of severity had considerable impact across the majority of the validators. The severe atypical and severe typical classes were associated with more and longer episodes; more syndrome consequences; substantial comorbidity; increased conformity, dependency, neuroticism, and external locus of control; lower self-esteem; and increased parental alcohol/drug use disorders.

The main effect of atypicality was more complex. First, there were fewer significant differences for atypicality than for severity, suggesting that the latter was a more potent discriminator. Second, independent of severity, the atypical classes were characterized by decreased syndrome consequences, increased risk for conduct disorder and social phobia but decreased risk for PTSD, personality traits of higher interpersonal dependency and lower self-esteem, and a parental history of alcohol/drug use disorder. Finally, the interaction of atypicality and severity was associated with increased interference and total symptom consequences, decreased social phobia, and increased PTSD.

The intermediate and minimal symptom classes were the most prevalent of the LCA classes, although their symptom endorsement patterns were quite different (table 1). Statistical comparisons of these two classes for the external validators in tables 2 and 3 revealed essentially no differences between them (data not shown). However, both of these classes had a surprisingly high degree of syndrome consequences (table 2) which were equivalent to or, occasionally, exceeded those of the mild typical and mild atypical classes. In regard to comorbidity (table 3), conduct disorder, antisocial personality, alcohol dependence, and drug dependence were particularly prevalent in the intermediate and minimal symptom classes.

Application of LCA to the lifetime depressive symptom data from the NCS yielded six interpretable classes. Four of these classes were defined by the combinations of severity (severe versus mild symptoms) and atypicality (atypical versus typical symptoms); the remaining 2 classes were distinctive. Identification of depressive subtypes defined by mild and severe symptoms is neither profound nor surprising and has been widely documented in prior clinical and epidemiological studies by use of a host of methodological approaches.

The particularly interesting result from this study was the identification of depressive classes defined principally by the atypicality of the symptoms reported. Depressive symptoms with an atypical character or a “reversed functional shift” (5) (i.e., appetite increase, weight gain, hypersomnia, and psychomotor agitation) constituted one side of the atypicality dichotomy. The symptoms that contributed to the other side of this dichotomy were typical, or vegetative, in nature (i.e., appetite decrease, weight loss, insomnia, and psychomotor retardation) resembling a “functional shift” (33).

Our principal interest in conducting this study was whether our prior finding in the VTR (16) could be reproduced in a different sample by a similar analytic approach (i.e., the use of LCA and similar item definitions). The NCS and the VTR are similar in that both are population based and used structured diagnostic instruments. However, these similarities are overshadowed by many differences:

1. The NCS was a probability sample of the continental U.S., whereas the VTR consisted solely of twins born in Virginia.

2. The NCS consisted of male and female subjects of any ethnicity aged 15–54 years, whereas the VTR report was on Caucasian women aged 22–59 years.

3. The NCS depressive symptoms were for the lifetime worst episode of at least 2 weeks’ duration, whereas the VTR data were for a period in the prior year that had lasted at least 5 days.

4. Depressive symptoms were coded for all VTR subjects but only for a subset of the NCS participants.

5. Although the NCS interview (based on the Composite International Diagnostic Interview) and VTR interview (based on the Structured Clinical Interview for DSM-III-R [34]) both assessed the DSM-III-R criteria for major depression, the two interviews were considerably different in structure, item wording, and format.

Thus, given that the dissimilarities were greater than the similarities between the VTR and the NCS, it is notable that we were able to reproduce the existence of the atypical classes, one of the key findings from our prior report (16).

In the depression treatment literature, the clinical importance of identifying a subset of depressed patients with atypical symptoms is supported by expert opinion (8, 9) and clinical trials (10–1235), although not in all studies (5). Patients with atypical depression may constitute a distinct neurobiological subset (36). However, given the many different uses of the term atypical over time (6, 7), a critical question is the degree to which the atypical subtype of depressive symptoms identified in this study and in our prior report (16) overlaps with the Columbia/DSM-IV definitions.

We possess insufficient information to answer this question as we lacked data on all elements of the Columbia/DSM-IV criteria for atypical depression (i.e., mood reactivity, leaden paralysis, and enduring rejection sensitivity). The atypical subtypes we identified were characterized by appetite increase, weight gain, hypersomnia, and, to some extent, psychomotor retardation, which overlap with the Columbia/DSM-IV criteria. Of note, the focus on mood reactivity in the Columbia/DSM-IV criteria is not found in other definitions of atypicality that highlight instead symptoms of a reversed functional shift (4, 5, 8, 9).

If the overlap between the clinically and epidemiologically defined atypical depressive symptoms is reasonable, these convergent findings—epidemiological dissection of atypical classes in several distinct studies—combined with the clinical utility of atypical depressive symptoms, constitute a compelling rationale for the existence of an atypical subtype and its inclusion in any typology of unipolar depression.

To facilitate comparison, table 5 summarizes the results of this NCS-based investigation and the VTR report (16). The VTR report described severe and mild typical LCA classes that were fairly similar to their NCS counterparts in symptoms, mean number of symptoms, the proportion with major depression, and the proportion reporting role interference. The VTR atypical class was most similar to the NCS mild atypical class. The NCS severe atypical class had no counterpart in the VTR, perhaps because the NCS was larger or used a longer time frame.

Kendler et al. (16) noted several differences between the VTR atypical class and the other classes—i.e., the atypical class was less likely to have comorbid general affective disorder and panic disorder and had decreased extraversion. With the identification in the NCS of a severe atypical class as well as a mild atypical class (similar to the VTR atypical class), the differences in the VTR appear to be more related to severity than to atypicality. The NCS results also differ from those of the VTR in that neuroticism was elevated in the NCS classes, particularly in the severe classes. The VTR report also noted that the atypical class had increased body mass index, risk of bulimia, moderate risk of future major depression, moderate symptom stability over time, and high monozygotic twin concordance. These data were not available in the NCS.

Perhaps the most informative comparisons are those between atypicality and severity (tables 3 and 4). Most differences across the four classes defined by atypicality and severity were related to severity. The main effect of severity (table 4) was associated with worse episodes, more syndrome consequences, increased comorbidity, more deviant personality traits, and a parental history of alcohol/drug use disorders. These differences were in the anticipated direction and validated the severe versus mild LCA class distinction.

The main effect of atypicality was less frequently significant, although the atypical classes were characterized by decreased syndrome consequences, comorbid conduct disorder and social phobia (but decreased PTSD), higher interpersonal dependency and lower self-esteem, and a parental history of alcohol/drug use disorder. These results suggest that atypicality—although unexpectedly similar to typicality—may have several distinct characteristics. The association of atypicality with parental alcohol/drug use disorders is reminiscent of Winokur’s depression spectrum disease (37).

We expected to find that the atypical classes were characterized by a distinct personality profile—perhaps akin to neurotic depression (3). Although atypicality had little impact on many personality/attitudinal attributes, the atypical classes had greater interpersonal dependency and lower self-esteem suggesting that there may be a specific personality typology associated with atypical depression.

From some prior studies of atypical depression (2, 4, 36), we expected to find an overrepresentation of women in the severe and mild atypical classes. As in other reports (13, 14), we identified no significant gender differences.

The occurrence of these “subthreshold” classes (together comprising 19% of the subjects in the NCS) and the rather high proportion of each class that reported recurrent and enduring symptoms along with help-seeking behavior or marked symptom consequences were surprising. We know little about the so-called minor depressive disorders. The characteristics of these classes—particularly the equal gender ratio—would argue for further specific study.

We put forward three speculations about these classes. First, studies from the NCS (38) and the VTR (39) suggest that depressive states that do not quite meet the DSM-III-R criteria differ from more definitive major depression quantitatively and not qualitatively. These classes might represent formes frustes of major depression. In addition, subsyndromal depressive symptoms are associated with considerable social morbidity (40, 41) and an increased risk for first-onset major depression (42). Second, as the diagnosis of major depression in epidemiological samples has modest reliability (43, 44), these classes might contain individuals who in fact had major depression but who were coded as having subthreshold symptoms. Finally, these two classes were notable for their increased prevalence of alcohol and drug dependence; depressive symptoms could have represented substance-induced mood disorders.

The depressive symptoms we studied were during the worst lifetime episode. We had no information about whether subsequent or prior episodes had a similar or different pattern of typical and atypical symptoms. Moreover, some of our findings may have been due to chance given the large number of statistical comparisons performed.