Introduction
From Kraepelin to modern times, the course trajectory has been an important criterion for validating psychiatric syndromes. In 1980, course was incorporated into the diagnostic criteria of various subtypes of depressive disorders of the Diagnostic and Statistical Manual (DSM), resulting in the diagnostic categories of major depressive disorder (MDD) and dysthymic disorder (Dysth). Over the years, various other unipolar depressive syndromes or subtypes have been defined, including double depression (DD: the co-occurrence of MDD and Dysth).
This diversity among depressive disorders has been the topic of debate, with much of the controversy focusing on whether MDD and Dysth are qualitatively distinct disorders or represent different ends of a severity continuum (Shankman & Klein, Reference Shankman and Klein2002). The enormous overlap in diagnoses between MDD and Dysth (Klein et al. Reference Klein, Shankman and Rose2006) and the failure to demonstrate differences between MDD and Dysth on a broad range of demographic, clinical, psychosocial, family history and treatment response variables (Keller & Lavori, Reference Keller and Lavori1984; Angst & Wicki, Reference Angst and Wicki1991; McCullough et al. Reference McCullough, Klein, Borian, Howland, Riso, Keller and Banks2003) fuelled this debate. Judd et al. (Reference Judd, Schettler and Akiskal2002) concluded that Dysth is an integral component of the longitudinal structure of MDD, with each symptom level representing a different phase of illness intensity, activity and severity (Judd & Akiskal, Reference Judd and Akiskal2000). By contrast, there is evidence to support the existence of specific diagnostic subtypes within the larger group of affective disorders, such as chronic depression (McCullough et al. Reference McCullough, Klein, Borian, Howland, Riso, Keller and Banks2003).
To study the validity of various depressive subtypes, statistical methods designed to cluster heterogeneous outcomes naturally occurring in the data can be helpful. Free from any a priori assumptions, data-driven techniques, such as latent class growth analysis (LCGA), that cluster persons based on their trajectories of a given outcome may thus result in a more empirically based classification that can then be compared with the DSM classification of unipolar affective disorders. This method can be used to determine whether different depressive subtypes based on course trajectories can be distinguished and to what extent this corresponds with the current diagnostic categories. To date, several studies have used this approach to assess course trajectories in depressive disorders across adolescence (Stoolmiller et al. Reference Stoolmiller, Kim and Capaldi2005; Olino et al. Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010) and among mothers throughout the childrearing phase (Campbell et al. Reference Campbell, Morgan-Lopez, Cox and McLoyd2009; Skipstein et al. Reference Skipstein, Janson, Stoolmiller and Mathiesen2010), or among bereaved caregivers (Li Reference Li2005; Taylor et al. Reference Taylor, Ezell, Kuchibhatla, Ostbye and Clipp2008). Colman et al. (Reference Colman, Ploubidis, Wadsworth, Jones and Croudace2007) examined course trajectories over the life course among the general population. However, some of these studies were limited by the inclusion of men only (Stoolmiller et al. Reference Stoolmiller, Kim and Capaldi2005), use of different measures or informants over time (Stoolmiller et al. Reference Stoolmiller, Kim and Capaldi2005; Colman et al. Reference Colman, Ploubidis, Wadsworth, Jones and Croudace2007), or combined depressive and anxiety symptoms (Colman et al. Reference Colman, Ploubidis, Wadsworth, Jones and Croudace2007). Hence, comparison of latent course trajectories with current DSM categories for depressive subtypes is lacking.
In this study we aimed to identify clinically relevant course trajectories, derived from LCGA, in a large cohort of currently depressed subjects participating in the Netherlands Study of Depression and Anxiety (NESDA). To determine whether the identified classes represent different course trajectories, we briefly validated the classes by means of some external validators. Next, we compared the identified course trajectories with the current diagnostic categories of MDD, Dysth and DD. If DSM categories identify clinical course trajectories sufficiently, we would expect MDD to be the prominent diagnosis in a class with best course trajectory and, additionally, most persons with MDD should be allocated to the class with the best course trajectory. Likewise, Dysth should be the most prominent diagnosis in the class with a chronic course, albeit of moderate severity, and most persons with Dysth should be allocated to this class. Finally, we expected DD to be the most prominent diagnosis in the class, representing a chronic and severe course trajectory, and most persons with DD to be allocated to this particular class. In addition, because DSM categories suggest particular course trajectories, we examined to what extent DSM categories predict course trajectories. Finally, we explored the characteristics of each identified class and examined putative predictors of class membership, and hence course trajectory. In previous studies, sociodemographics, co-morbid anxiety disorders (Hayden & Klein, Reference Hayden and Klein2001), age of onset of depressive symptoms (Klein et al. Reference Klein, Schatzberg, McCullough, Keller, Dowling, Goodman, Howland, Markowitz, Smith, Miceli and Harrison1999; Angst et al. Reference Angst, Gamma, Rossler, Ajdacic and Klein2009), childhood adversity (Hayden & Klein, Reference Hayden and Klein2001; Rhebergen et al. Reference Rhebergen, Batelaan, De Graaf, Nolen, Spijker, Beekman and Penninx2011), somatic co-morbidity (Angst et al. Reference Angst, Gamma, Rossler, Ajdacic and Klein2009), neuroticism (Hayden & Klein, Reference Hayden and Klein2001; Rhebergen et al. Reference Rhebergen, Batelaan, De Graaf, Nolen, Spijker, Beekman and Penninx2011) and a family history of depression (Klein et al. Reference Klein, Schatzberg, McCullough, Keller, Dowling, Goodman, Howland, Markowitz, Smith, Miceli and Harrison1999; Stoolmiller et al. Reference Stoolmiller, Kim and Capaldi2005; Olino et al. Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010) have been demonstrated to predict course trajectories. However, some inconsistencies exist. In this study, we examine the predictive value of sociodemographics (age, gender, level of education), co-morbid anxiety disorder, age of onset, childhood adversity, number of somatic illnesses, neuroticism, extraversion, and the presence of a family member with depression for class membership.
Method
Study sample
The NESDA is designed to investigate the course and consequences of depressive and anxiety disorders. It is a multisite naturalistic cohort study of adults (aged 18–65 years) recruited from the general population, general practices, and mental health organizations. The method of recruitment was extensively described elsewhere (Penninx et al. Reference Penninx, Beekman, Smit, Zitman, Nolen, Spinhoven, Cuijpers, De Jong, Van Marwijk, Assendelft, Van der Meer, Verhaak, Wensing, De Graaf, Hoogendijk, Ormel and Van Dyck2008). The study protocol was approved by the Ethical Review Board of the VU University Medical Center and written informed consent was obtained from all participants. For the present analysis, we selected respondents from the NESDA baseline assessment with a 6-month MDD (n=1115) and/or a 6-month Dysth (n=305) (total n=1158). Of the 1158 persons at baseline, 946 (81.7%) could be followed up after 2 years. Attrition in the study population was higher among those with younger age [odds ratio (OR) 1.02], lower educational level (OR 1.08) and those with DD (OR 1.47), but was not associated with sex, MDD or Dysth. Attrition for the NESDA has been described more extensively elsewhere (Lamers et al. Reference Lamers, Hoogendoorn, Smit, van Dyck, Zitman, Nolen and Penninx2011). To exclude persons already in remission, we restricted the sample to 804 subjects with a 6-month depressive diagnosis who confirmed symptoms in the month prior to baseline at either the Composite Interview Diagnostic Instrument (CIDI) recency questions or the Life Chart Interview (LCI; Lyketsos & Nestadt, Reference Lyketsos and Nestadt1994) recorded at baseline. Of this final sample, 65.4% were female, with a mean age of 42.0±12.2 years and 11.6±3.3 years of education. Diagnoses were established using CIDI version 2.1, according to DSM-IV criteria (WHO, 1998). The CIDI is a structured interview with acceptable reliability and validity (Wittchen, Reference Wittchen1994; Wittchen et al. Reference Wittchen, Zhao, Abelson, Abelson and Kessler1996). It was administered by specially trained research staff.
Measurements
Symptoms of depression
The LCGA was based on depressive symptoms derived from the LCI (Lyketsos & Nestadt, Reference Lyketsos and Nestadt1994). The LCI was assessed both at baseline and after 2 years of follow-up by a trained interviewer. Baseline LCI served to compute a measure of duration of depressive symptomatology. LCI at the 2-year follow-up provided the data for the LCGA. First, the occurrence of life events in a particular period was explored to refresh memory and then the presence of depressive symptoms was assessed during each month of the successive years. In addition, for each month with reported depressive symptoms, the self-reported burden of the symptoms was assessed on a five-point scale, ranging from no or minimal severity, to mild, moderate, severe, or very severe. As the LCI at follow-up retrospectively assessed the burden of depressive symptoms over 2 years, this resulted in 24 scores of perceived burden of depressive symptoms. None of the 804 respondents had missing data on the LCI at follow-up.
External validators
To assess the validity of the identified course trajectories, we examined the distribution of various putative validators across the identified classes. We examined the total scores on the 30-item Inventory of Depressive Symptomatology – Self Report version (IDS-SR, www.ids-qids.org; Rush et al. Reference Rush, Giles, Schlesser, Fulton, Weissenburger and Burns1986, Reference Rush, Gullion, Basco, Jarrett and Trivedi1996), both at baseline and after 2 years. In addition, the percentage of decline in IDS score during 2 years of follow-up was calculated. Next, level of functioning was measured by total scores on the World Health Organization Disability Assessment Schedule II (WHO-DAS; Chwastiak & Von Korff, Reference Chwastiak and Von Korff2003), both at baseline and after 2 years of follow-up. Again the percentage of improvement on the WHO-DAS over 2 years of follow-up was calculated. The WHO-DAS queries difficulties in various domains of life during the past 30 days. We computed total scores based on five domains of the WHO-DAS (cognition, mobility, self-care, interpersonal interactions, participation in society). Finally, we examined the presence of a 6-month CIDI diagnosis of depression and/or Dysth at follow-up.
Depressive categories
To compare the identified course trajectories with DSM-IV categories and also to examine the predictive value of DSM-IV categories on class membership, we examined the relative distribution of depressive disorders at baseline per class and across classes (i.e. the percentages of baseline disorders across the rows). A depressive disorder at baseline consisted of a 6-month CIDI MDD and/or Dysth at baseline. Three depression groups were defined: major depressive disorder (MDD) (n=573) consisted of those fulfilling an MDD and no Dysth diagnosis; dysthymic disorder only (Dysth) (n=34) consisted of those with a dysthymic disorder but not an MDD diagnosis; and DD (n=197) was defined as Dysth with a co-morbid MDD, ignoring the sequence of onset of Dysth and MDD.
Putative predictors of class membership
Although DSM-IV depressive categories are partly based on course and severity measures, we examined additional course and severity markers. This enabled us to perform multivariate multinomial logistic regression analyses to test whether DSM-IV categories have some additional value in the prediction of class membership over other baseline severity and duration measures. As a measure of severity, we used IDS-SR at baseline. This measure was also used as a putative external validator for the identified course trajectories. As a measure of duration, we calculated the number of months with depressive symptoms prior to baseline, using LCI data from the past 5 years prior to baseline, recorded at baseline. Next, we examined the distribution of several other characteristics across the identified course trajectories and their predictive value on class membership (age, gender, level of education, co-morbid anxiety disorder, age of onset, childhood adversity, number of somatic illnesses, neuroticism, extraversion, and the presence of a family member with depression). Age was included as a continuous variable. Level of education was measured by years of education. Co-morbid anxiety disorder included a 6-month diagnosis of panic disorder (with or without agoraphobia); agoraphobia without history of panic disorder; social phobia and/or generalized anxiety disorder, as recorded by the CIDI. Age of onset of the index disorder was assessed in the CIDI. Earliest age was used for those with co-morbid disorders. Childhood adversity was assessed using a structured interview from the Netherlands Mental Health Survey and Incidence Study (NEMESIS; Bijl et al. Reference Bijl, Van Zessen, Ravelli, De Rijk and Langendoen1998), in which respondents were asked to retrospectively recall whether they had experienced emotional neglect, psychological abuse, physical abuse or sexual abuse before the age of 16 years. A childhood index was computed (range 0–4), as used previously (Wiersma et al. Reference Wiersma, Hovens, van Oppen, Giltay, van Schaik, Beekman and Penninx2009). The presence of somatic conditions was assessed in the baseline interview. A count of the number of somatic diseases (including cardiovascular disease, diabetes, lung disease, osteoarthritis, cancer, gastrointestinal disease, liver disease, epilepsy, thyroid disease) for which medical treatment is received was constructed. Neuroticism and extraversion were assessed by means of the neuroticism and extraversion dimension of the NEO (Neuroticism–Extroversion–Openness) personality questionnaire, a 60-item questionnaire measuring five personality domains (Costa & McCrae, Reference Costa and McCrae1995). The presence of a depressive disorder in first-degree family members was assessed using the family tree method (Fyer & Weissman, Reference Fyer and Weissman1999).
Statistical analyses
To identify distinct classes based on different course trajectories, we performed LCGA, which is used to identify different classes of persons showing a comparable course of symptoms (Nylund et al. Reference Nylund, Asparouhov and Muthén2007). It starts with one class, suggesting one type of course fitting for all persons. Then successively more classes are added to determine the best-fitting model. Because LCGAs and other mixture models are susceptible to converging on local rather than global solutions (Nylund et al. Reference Nylund, Asparouhov and Muthén2007), we used multiple random starting values for the estimated models (500 repeats with 20 final optimizations). We increased the number of random starting values when necessary to avoid local solutions. We first fitted standard LCGA with intercept and linear effect of time. Then we successively added a quadratic and cubic slope factor to allow for curved trajectories. These additional growth factors will add computational burden and may add to convergence problems and model instability. Therefore, we fixed the variance and covariance estimates in all models to zero to aid convergence (Jung & Wickrama, Reference Jung and Wickrama2008). The LCI score was treated in the LCGA as a continuous variable. The LCGAs were conducted using M-plus version 5 (Muthén & Muthén, Reference Muthén and Muthén2007).
To determine which model best fitted the data, we combined the smallest Bayesian information criterion (BIC), the Lo–Mendell–Rubin (LMR) likelihood ratio test, entropy and proportions of the classes with interpretability and clinical relevance of the latent trajectory classes. Of the traditionally used information criteria, such as Aikaike's Information Criterion (AIC) and BIC, the latter performs best (Nylund et al. Reference Nylund, Asparouhov and Muthén2007). The accuracy of the AIC decreases as the sample size increases because there is no adjustment for sample size (Nylund et al. Reference Nylund, Asparouhov and Muthén2007). Lower BIC values indicate better model fit. The LMR test provides a p value that indicates whether the k–1 class model is rejected in favor of the k class model. Finally, entropy, as a measure of the quality of classification, is presented for models with more than one class, in addition to the proportions of individuals in each class. To identify clinically relevant classes, we only considered models with classes greater than 5%.
After the identification of the classes, persons were assigned to their most likely class based on model probabilities. Then, we examined the validity of the identified classes by testing the distribution of the potential validators across classes, and we compared the characteristics of participants across the identified classes using two-tailed χ2 statistics for categorical variables and one-way analysis of variance (ANOVA) statistics for continuous variables. Additional pairwise comparisons were performed to test for differences between pairs of classes. Multivariate, multinomial logistic regression analyses were conducted to examine determinants of class membership. First, the association between DSM categories and class membership was examined (model 1). Second, the association between severity and duration measures and class membership was examined (model 2). Third, both DSM categories and severity and duration measures were combined in one model to test any additional predictive value of DSM categories above measures of severity and duration in predicting class membership (model 3). Fourth, the association between other characteristics and class membership was examined. All comparisons were conducted using SPSS version 15.0 (SPSS Inc., USA).
Results
In Table 1 the maximum likelihood values, the BIC, the entropy, the LMR likelihood ratio tests and the proportion of individuals in each class of each LCGA model are presented. The standard linear model provided a good representation of the various course trajectories. Addition of cubic and quadratic growth factors resulted in non-fitting models (data not shown). Non-convergence and solutions with inadmissible estimates occurred. Therefore, we altered the random starts to 2000/750, but this still did not yield adequate estimates. To test the assumption that cubic or quadratic slopes might be present in some but not all classes, we tried to constrain each respective class to follow a linear model, while allowing the other classes cubic or quadratic trajectories. This also did not yield fitting models. Hence, both cubic and quadratic models were considered as non-fitting. Although the BIC values continued to diminish in the linear model, the LMR test failed to reach significance, from the six-class model onwards. Hence, the five-class linear model provided the best fit for the observed data.
Parameters of fit of latent class growth analysis (LCGA)
BIC, Bayesian Information Criterion; LMR, Lo–Mendell–Rubin; LL, log likelihood.
Figure 1 shows the course trajectories of the five latent classes identified. The first class (prevalence 44.3%) was characterized by a rapid remission of symptoms. The second class (prevalence 24.1%) showed a decline of symptoms, albeit slower than class 1. In addition, the burden of depressive symptoms at baseline was higher compared to the first class. The third class (prevalence 9.8%) showed a gradual decline of symptoms, but initial depression severity was higher than both previous classes. The fourth class (prevalence 15.0%) had a chronic course trajectory, of moderate depression severity. Finally, the fifth class (prevalence 6.7%) had a chronic course trajectory, with the highest burden of depressive symptoms.
Identified course trajectories, derived from latent class growth analysis (LCGA). To facilitate interpretation, class numbering corresponds to increasing severity and/or duration of depressive symptoms.
Validation of the course trajectories
We tested the validity of the five-class model by examining the distribution of severity measures (IDS score, level of functioning) and measures of change (decline in IDS score, improvement in functioning) across the course trajectory classes identified (Table 2). As expected, both extremes, class 1 (remission) and class 5 (chronic severe), were markedly distinct from each other on all parameters. Next, we expected that classes of similar severity at baseline (classes 2 and 4, classes 3 and 5) would not differ on baseline severity measures but would differ on measures of change. This was confirmed. In addition, we expected no differences between chronic course trajectories (classes 4 and 5) on measures of change (decline in IDS score, improvement in functioning), whereas difference in severity measures were expected. This was also confirmed. The course trajectory over 2 years of class 3 (decline, severe) could only partly be confirmed by external validators. Despite expectations that class 3 (decline, severe) would improve over time, severity measures of class 3 after 2 years were more like those of class 4 (chronic, moderate) (no significant differences, except baseline severity measures) than class 2 (decline, moderate). In addition, we expected classes 3 (decline, severe) and 4 (chronic, moderate) to differ on all parameters. However, they only differed significantly in baseline functioning. To conclude, our LCGA-identified model could largely be confirmed by external validators.
Distribution of putative validators across the identified latent course trajectories
IDS, Inventory of Depressive Symptomatology; s.d., standard deviation; df, degrees of freedom.
Significant differences across classes (p<0.05): Class 1 versus 2: all parameters significantly different; Class 1 versus 3: all parameters significantly different; Class 1 versus 4: all parameters significantly different; Class 1 versus 5: all parameters significantly different; Class 2 versus 3: IDS at baseline, IDS at 2 years (functioning baseline p=0.06), functioning 2 years, improvement in functioning; Class 2 versus 4: decline in IDS, CIDI disorder after 2 years (IDS at 2 years p=0.06); Class 2 versus 5: all parameters significantly different; Class 3 versus 4: (IDS at baseline p=0.07), functioning baseline; Class 3 versus 5: IDS at 2 years, functioning at 2 years, improvement in functioning, CIDI disorder after 2 years; Class 4 versus 5: all parameters significantly different, except change in IDS.
Depressive diagnoses
As presented in Table 3, we tested the distribution of the baseline DSM-IV categories of MDD, Dysth and DD across the trajectory classes identified. As expected, comparisons of the relative distribution of depressive disorders per class showed that the percentage of MDD gradually and significantly declined when course trajectories worsened, whereas the percentage of DD gradually rose. In addition, the fourth class (chronic, moderate) had the highest relative percentage of persons with Dysth (9.1%) and the fifth class (chronic, severe) had a relatively high percentage of persons with DD (44.4%). However, when DSM categories represent clinically relevant course types, persons with MDD should mainly be allocated to the most favorable course trajectory, whereas persons with Dysth and DD should mainly be allocated to the chronic classes. To examine this assumption, we compared the distribution of baseline depressive disorders across classes (i.e. the percentages of baseline disorders across the rows). As expected, MDD was most prevalent among the rapidly improving, mild class (class 1, 50%). However, almost two-thirds (61.7%) of the persons with Dysth at baseline were allocated to the best course trajectory classes (38.2% class 1; 23.5% class 2). Similarly, more than half of the persons with DD were allocated to the best course trajectory classes (28.4% class 1; 26.4% class 2), a significantly higher percentage compared to other classes. This is in contrast to our a priori assumptions.
Distribution of putative risk factors across the identified latent course trajectories
MDD, Major depressive disorder; Dysth, dysthymia; DD, double depression (=MDD+Dysth); IDS, Inventory of Depressive Symptomatology; s.d., standard deviation; df, degrees of freedom.
Significant differences across classes (p<0.05): Class 1 versus 2: MDD, DD, IDS baseline, months depressed, age, neuroticism, extraversion; Class 1 versus 3: MDD, DD, IDS baseline, months depressed, childhood adversity, neuroticism, extraversion; Class 1 versus 4: MDD, Dysth, DD, IDS baseline, months depressed, age, gender, childhood adversity, somatic illnesses, neuroticism, extraversion; Class 1 versus 5: MDD, DD, IDS baseline, months depressed, age of onset, childhood adversity, neuroticism, extraversion; Class 2 versus 3: (Dysth p=0.07), IDS baseline, months depressed; Class 2 versus 4: MDD (Dysth p=0.07), months depressed, somatic illnesses; Class 2 versus 5: MDD, DD, IDS baseline, months depressed, family with depression; Class 3 versus 4: Dysth, age; Class 3 versus 5: Age of onset; Class 4 versus 5: IDS baseline, age of onset, family with depression.
To further test the association between DSM categories and course trajectories, we conducted multinomial multivariate logistic regression analyses. First, the association between DSM categories and class membership was examined (model 1) (reference class 1; rapid remission and mild severity). As shown in Table 4, the association between DD and class membership showed an increase in ORs across the classes. Dysth was associated with class 4 (chronic, moderate) only. Next, the association between severity and duration measures and class membership was examined (model 2). IDS score was associated with class membership, with the highest odds for the severe classes [class 3: OR 2.03, 95% confidence interval (CI) 1.56–2.65; class 5: OR 2.56, 95% CI 1.86–3.52]. Number of months was associated with classes 3, 4 and 5, with similar ORs. Finally, both DSM categories and severity and duration measures were examined in one multivariate, multinomial, logistic regression model to test whether DSM categories would have additional predictive value over duration and severity measures only, in predicting class membership (model 3). Whereas the IDS score continued to predict class membership, DD was only associated with class 5 (chronic, severe). Dysth still predicted class membership for class 4.
Results from multivariate, multinomial, logistic regression analyses associating baseline diagnosis, severity and duration and other characteristics with the identified course trajectories, compared to the most favorable course trajectory (class 1: remission, mild severity; reference)
MDD, Major depressive disorder; Dysth, dysthymia; DD, double depression (=MDD+Dysth); IDS, Inventory of Depressive Symptomatology; s.d., standard deviation.
a Reported per s.d. increase: IDS score at baseline s.d.=11; months depressed s.d.=17; age s.d.=12; education s.d.=3; childhood adversity s.d.=1; somatic illnesses s.d.=1; neuroticism s.d.=7; extraversion s.d.=7.
Values given as odds ratio (95% confidence interval).
Bold is significant association (p<0.05).
Characteristics of the classes and their predictive value on course trajectories
Whereas class 1 significantly differed on a broad range of characteristics (lower age, more females, less childhood adversity, less somatic illnesses, lower neuroticism, higher extraversion) from the other classes, differences among the other four classes were less profound (Table 3). The fourth class (chronic, moderate) was significantly older than classes 1 and 3 and had a significantly higher number of somatic illnesses compared to classes 1 and 2. In addition, the fifth class (chronic, severe) had a lower age of onset compared to classes 1, 3 and 4, and a higher number of family members with depression compared to classes 2 and 4.
In multinomial multivariate logistic regression analyses (reference class 1; rapid remission and mild severity), lower extraversion was associated with class 2 (decline, severe) (OR 1.33, 95% CI 1.08–1.67) (Table 4). Higher neuroticism was associated with class 3 (decline, severe). For the two chronic classes more predictors could be identified. A chronic course trajectory of moderate severity (class 4) was associated with being male (OR 1.59, 95% CI 1.01–2.63), having more somatic illnesses (OR 1.26, 95% CI 1.02–1.55) and a higher level of neuroticism (OR 1.40, 95% CI 1.05–1.86). Similarly, a higher age (OR 1.45, 95% CI 1.03–2.04), an earlier age of onset (OR 1.79, 95% CI 1.23–2.63) and a lower level of extraversion (OR 1.49, 95% CI 1.04–2.13) predicted poorest course trajectory (chronic, severe). First-degree family members with depression resulted in high odds for the poorest course trajectory (class 5) (OR 2.57, 95% CI 0.75–8.76), albeit not reaching significance. In a post-hoc analysis comparing both chronic classes, a family history of depression was significantly associated with class 5 (chronic severe) membership (OR 4.00, 95% CI 1.14–14.29) in addition to an earlier age of onset (reference class 4).
Discussion
The aim of our study was to identify clinically relevant course trajectories in a large cohort of depressed subjects participating in the NESDA and to compare these with DSM categories of MDD, Dysth and DD. By implementing data-driven techniques, we identified five different course trajectories, ranging from mild severity and rapid remission to high severity and chronic course trajectory. LCGA enabled classification of course trajectories beyond the scope of DSM and produced detailed course trajectories, in which heterogeneity during course trajectories was taken into account.
More than 40% of the respondents had a rapid remission of the depressive symptoms (class 1) and differed significantly on a broad range of characteristics from the other classes, whereas 20% of the respondents had a chronic course trajectory (classes 4 and 5) that was associated with specific putative predictors of course. Spijker et al. (Reference Spijker, De Graaf, Bijl, Beekman, Ormel and Nolen2002) found similar percentages of chronic course trajectories of depressed persons in a large 3-year follow-up population-based study, although they did not use a data-driven technique. A review by Nandi et al. (Reference Nandi, Beard and Galea2009) described studies that identified distinct course trajectories of depressive symptoms in the general population and their determinants, based on data-driven techniques; whereas most studies included children, adolescents or young adults, six studies assessed depressive trajectories among adults, including specific samples, such as mothers (Campbell et al. Reference Campbell, Morgan-Lopez, Cox and McLoyd2009) or bereaved caregivers (Li, Reference Li2005; Taylor et al. Reference Taylor, Ezell, Kuchibhatla, Ostbye and Clipp2008). Despite differences in sample and design, most studies identified several distinct course trajectories, with the most prevalent classes differing in severity of symptoms and often classifying a chronic course trajectory. Some studies also identified classes with increasing symptoms over time (Colman et al. Reference Colman, Ploubidis, Wadsworth, Jones and Croudace2007; Olino et al. Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010; Skipstein et al. Reference Skipstein, Janson, Stoolmiller and Mathiesen2010). As can be seen in Fig. 1, a worsening class is lacking in our model. As we limited our model to clinically relevant classes, representing at least 5% of the respondents, this class was not identified. In a higher-class model we indeed identified a small worsening class of less than 5% (data not shown).
Whereas in clinical practice depressive disorders are classified by DSM-IV categories, we now present a more empirically based classification of course trajectories, identified by data-driven techniques. The question is raised: are the DSM-IV depressive categories associated with meaningful clinical course trajectories, or are the course trajectories derived from LCGA, based on severity and duration measures, clinically more relevant? Our aim was to compare DSM categories with the identified classes. Although the distribution of MDD across classes was largely as expected, considerable numbers of Dysth (61.7%) and DD (54.8%) were allocated to the two most favorable classes (classes 1 and 2). These results are in line with our previous results on course trajectories of depressive disorders in a large population-based study (Rhebergen et al. Reference Rhebergen, Beekman, De Graaf, Nolen, Spijker, Hoogendijk and Penninx2009). In that study, we found that 50.0% of persons with Dysth at baseline were free from any diagnosis after 1 year, rising to 70.8% after 3 years. Similarly, 46.4% of persons with DD were free from any depressive diagnosis after 1 year, rising to 69.1% after 3 years. These results indicate that the current diagnostic categories do not accurately distinguish or predict clinically relevant course trajectories. Although Dysth and DD are by definition chronic disorders, both our previous work and the current results suggest heterogeneous course trajectories within these diagnostic DSM categories. To our knowledge, previous comparisons of identified course trajectories with DSM categories are limited. Olino et al. (Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010) investigated the prevalence of lifetime DSM diagnoses in the identified course trajectories of depression and anxiety from adolescence to adulthood in a general population sample. They found a high percentage of persons with Dysth in the persistent depressed class. All persons in this class had a lifetime diagnosis of MDD and half had a lifetime diagnosis of Dysth. They concluded that this class represents chronic or DD. We found a similar number of persons with Dysth in our chronic, severe class: Dysth was present in 48.1% of the respondents. However, although percentages of Dysth and DD within classes were as expected, the division of baseline disorders across classes differed from expectations. The majority of persons with Dysth or DD at baseline were allocated to the milder classes. Because Olino et al. (Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010) recruited high school students with a history of psychopathology and controls without psychopathology, they did not investigate whether persons with Dysth were allocated as expected to the chronic classes or not. In addition, they did not find a class with decreasing probability of depressive disorder. Comparison to our decreasing symptom classes is therefore hampered. Likewise, probably due to limited follow-up, we did not identify a class with increasing depressive symptoms. In the class with increasing probability of depressive disorder, Olino et al. (Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010) identified only a small minority with dysthymia. They suggest that this class represents persons with brief, recurrent depressive episodes. A more extended follow-up would enable us to investigate the prevalence of Dysth in other course trajectories. Stoolmiller et al. (Reference Stoolmiller, Kim and Capaldi2005) investigated lifetime depressive diagnoses in the identified classes. None in the very low group, 5.7% in the moderate-decreasing group, 20.0% in the high-decreasing group and 29.2% in the high-persistent group were diagnosed with a lifetime depressive disorder. No distinction was made between MDD, Dysth or DD. In addition, this study only included men, hence limiting generalizability to the general population.
Comparison of DSM categories with measures of severity and duration, by means of multivariate, multinomial, logistic regression analyses, revealed that the IDS score was persistently identified as a predictor for class membership, duration was associated with class 3 and 4 membership, whereas the predictive value of DSM categories was only marginal. Dysth only predicted class 4 membership (chronic, moderate) and DD only predicted class 5 membership (chronic, severe). These results suggest the importance of a combination of severity and duration measures in predicting course trajectories. However, both Dysth and DD additionally predicted chronic course trajectories (class 1 reference), which suggests that these DSM-IV categories do have some additional predictive power above severity and duration measures only. In post-hoc analyses, comparing both chronic classes (class 4 reference), DSM-IV categories were no longer associated with course. We may conclude that DSM-IV categories predict chronic depression, compared to non-chronic depression, but they do not differentiate between chronic depressions. These findings are in line with previous research (Spijker et al. Reference Spijker, De Graaf, Bijl, Beekman, Ormel and Nolen2002; Ormel et al. Reference Ormel, Oldehinkel, Nolen and Vollebergh2004; Vuorilehto et al. Reference Vuorilehto, Melartin and Isometsa2009). An earlier (albeit non-longitudinal) latent class analysis (LCA) study in the general population also identified a differentiation in subclasses based on severity (Sullivan et al. Reference Sullivan, Kessler and Kendler1998). In this context, McCullough et al. (Reference McCullough, Klein, Keller, Holzer, Davis, Kornstein, Howland, Thase and Harrison2000, Reference McCullough, Klein, Borian, Howland, Riso, Keller and Banks2003) and Klein et al. (Reference Klein, Shankman and Rose2006) suggested that the chronic versus non-chronic distinction with a cross-cutting dimension of severity constitutes an important source of heterogeneity in depression. Our results, based on data-driven techniques, support the suggestion of McCullough et al. (Reference McCullough, Klein, Borian, Howland, Riso, Keller and Banks2003) to characterize depressive disorders on cross-cutting categories or dimensions, based on chronicity and severity of the index disorder. Continuous measurements, such as the LCI, might contribute to a more thorough classification of depressive disorders that better reflects clinical course trajectories.
The final aim was to examine the association between various characteristics and class membership. In multinomial regression analyses, we observed a preponderance of males, somatic illnesses and a higher level of neuroticism in class 4 (chronic, moderate), and a higher age, an earlier age of onset and lower extraversion in class 5 (chronic, severe) (class 1 reference). Of note, we found male sex to be associated with the poorest course trajectory. Previously, this was also demonstrated by Olino et al. (Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010) in a sample of adolescents and young adults. They suggested that being female was more associated with transient and episodic forms of internalizing psychopathology. Both somatic co-morbidity (Angst et al. Reference Angst, Gamma, Rossler, Ajdacic and Klein2009) and an earlier age of onset of depression were previously associated with chronic depressions, compared to non-chronic depressions (Mondimore et al. Reference Mondimore, Zandi, Mackinnon, McInnis, Miller, Crowe, Scheftner, Marta, Weissman, Levinson, Murphy-Ebenez, Depaulo and Potash2006; Angst et al. Reference Angst, Gamma, Rossler, Ajdacic and Klein2009). Although most studies to date have not used data-driven techniques, Olino et al. (Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010) did used LCGA and concluded that the persistent depression class they identified may reflect an early-onset chronic form of depression (>70% chance of being depressed at each time interval), whereas the second depression class they identified may reflect an adult-onset, episodic form of depression with <25% chance of being depressed at each time interval. Because Olino et al. (Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010) included adolescents, comparison with our results is hampered. However, a considerable number of persons in our study reported the onset of their depressive episodes prior to 18 years (29.4%). The mean age of onset was 27.5±12.7 years (median=25.0 years). Hence, many persons had their first depressive episode in adolescence. Klein et al. (Reference Klein, Schatzberg, McCullough, Keller, Dowling, Goodman, Howland, Markowitz, Smith, Miceli and Harrison1999) also demonstrated that the early-onset subtype was associated with a poorer course trajectory and also greater family history of mood disorders. In our model, family history of depression showed high odds (OR 2.57, 95% CI 0.75–8.76) for the poorest course trajectory (class 5; class 1 reference), but this did not reach significance. Similarly, Angst et al. (Reference Angst, Gamma, Rossler, Ajdacic and Klein2009) could not demonstrate significant differences between chronic and non-chronic depressions related to family history of depression. Some other studies based on data-driven techniques identified family history of depressive symptomatology as a predictor for poor course trajectories. Stoolmiller et al. (Reference Stoolmiller, Kim and Capaldi2005) found that parental depressive symptoms contributed to class discrimination, discriminating the high-persistent class from the other three classes. This finding was replicated by Olino et al. (Reference Olino, Klein, Lewinsohn, Rohde and Seeley2010), who concluded that this pattern suggests that persistent depression may represent a more familial form of the disorder.
Our findings on personality dimensions are in line with previous reports. Both high neuroticism and low extraversion were previously found to be associated with chronicity (Robison et al. Reference Robison, Shankman and McFarland2009; Wiersma et al. Reference Wiersma, van Oppen, van Schaik, van der Does, Beekman and Penninx2011).
Previous studies failed to demonstrate significant differences between various chronic depressions on a variety of characteristics (Yang & Dunner, Reference Yang and Dunner2001; McCullough et al. Reference McCullough, Klein, Borian, Howland, Riso, Keller and Banks2003). Klein et al. (Reference Klein, Shankman and Rose2006) suggested that the nature of chronic depression varies over time within individuals, indicating that the various manifestations of chronic depression in DSM-IV do not represent distinct disorders. In post-hoc analyses, comparing both chronic classes (class 4, chronic, moderate and class 5, chronic severe) we found that age of onset and a family history of depression differentiated between the two chronic classes (4 and 5). Lyons et al. (Reference Lyons, Eisen, Goldberg, True, Lin, Meyer, Toomey, Faraone, Merla-Ramos and Tsuang1998) and, more recently, Kendler et al. (Reference Kendler, Fiske, Gardner and Gatz2009) suggested that early age of onset does typify a subtype of depressive disorders, with poorer course trajectories. Hence, our fifth class (chronic, severe), characterized by greater severity, an early age of onset and a higher familial loading, may typify a distinct subtype and may be informative for etiological research.
Strengths and limitations
The present study is one of the largest to date to longitudinally investigate the existence of depressive course types in a large cohort of depressed subjects by data-driven methods. Furthermore, a wide range of putative predictors was tested. However, it should be noted that LCGAs draw upon mathematical models that cannot account for a great variety of fluctuations in symptom levels. As Judd and colleagues (Judd & Akiskal, Reference Judd and Akiskal2000; Judd et al. Reference Judd, Schettler and Akiskal2002) state, there is strong evidence during the long-term course of illness that major, minor, dysthymic and subsyndromal symptoms wax and wane within the same patient and that these symptomatic periods are interspersed in the overall course within times when patients are in remission and symptom free. This pleiomorphic course cannot be described completely by LCGA. In addition, by excluding persons with a remitted depression, insight into the course trajectories of probable recurrent depression is not provided. This requires further research but falls outside the scope of this study. Likewise, LCI assessment over a more extended time period might identify course trajectories that could not be captured in the rather short time-frame of the current study, for example relapse symptomatology. Furthermore, it should be noted that classification of persons based on the most likely class membership may mean that for some persons classification is highly accurate, whereas for others the posterior probabilities lie much closer to each other, making classification less accurate. In subsequent analyses of class differences, these differences between persons are not taken into account, which may lead to distorted estimates and standard errors. However, in our model, mean posterior probabilities ranged from 0.96 to 0.97 across classes, indicating that assigning individuals to their class was performed with precision. Finally, LCIs retrospectively assess depressive symptoms and perceived burden, which is prone to recall bias.
To conclude, MDD, Dysth and DD did not adequately match empirically derived course trajectories for unipolar depression. For the future classification of unipolar depression, it may be wise to retain the larger, heterogeneous category of unipolar depression, adopting cross-cutting dimensions of severity and duration to further characterize patients. LCI information may be helpful in clinical practice because it combines both severity and course information. In addition, the identified putative predictors facilitate identification of persons at risk for chronic course trajectories in clinical care.
Acknowledgements
The infrastructure for the NESDA (www.nesda.nl) is funded through the Geestkracht program of the Netherlands Organization for Health Research and Development (ZonMw, grant no. 10-000-1002) and is supported by participating universities and mental health care organizations: VU University Medical Center, GGZ inGeest, Arkin, Leiden University Medical Center, GGZ Rivierduinen, University Medical Center Groningen, Lentis, GGZ Friesland, GGZ Drenthe, IQ Healthcare, Netherlands Institute for Health Services Research (NIVEL) and Netherlands Institute of Mental Health and Addiction (Trimbos).
Declaration of Interest
None.
