New insights into the correlation structure of DSM-IV depression symptoms in the general population v. subsamples of depressed individuals

Abstract

Aims.

Previous research failed to uncover a replicable dimensional structure underlying the symptoms of depression. We aimed to examine two neglected methodological issues in this research: (a) adjusting symptom correlations for overall depression severity; and (b) analysing general population samples v. subsamples of currently depressed individuals.

Methods.

Using population-based cross-sectional and longitudinal data from two nations (Switzerland, 5883 young men; USA, 2174 young men and 2244 young women) we assessed the dimensions of the nine DSM-IV depression symptoms in young adults. In each general-population sample and each subsample of currently depressed participants, we conducted a standardised process of three analytical steps, based on exploratory and confirmatory factor and bifactor analysis, to reveal any replicable dimensional structure underlying symptom correlations while controlling for overall depression severity.

Results.

We found no evidence of a replicable dimensional structure across samples when adjusting symptom correlations for overall depression severity. In the general-population samples, symptoms correlated strongly and a single dimension of depression severity was revealed. Among depressed participants, symptom correlations were surprisingly weak and no replicable dimensions were identified, regardless of severity-adjustment.

Conclusions.

First, caution is warranted when considering studies assessing dimensions of depression because general population-based studies and studies of depressed individuals generate different data that can lead to different conclusions. This problem likely generalises to other models based on the symptoms’ inter-relationships such as network models. Second, whereas the overall severity aligns individuals on a continuum of disorder intensity that allows non-affected individuals to be distinguished from affected individuals, the clinical evaluation and treatment of depressed individuals should focus directly on each individual's symptom profile.

Introduction

Major depression, which is characterised by the core symptoms of depressed mood and anhedonia (World Health Organization, 2012; American Psychiatric Association, 2013), is a leading contributor to the global burden of disease (Murray & Lopez, 1996; Bromet et al. 2011; Ferrari et al. 2013; Kessler & Bromet, 2013; Cuijpers et al. 2014). For years, debate has raged over the clinical presentation of depression (Baumeister & Parker, 2012; van Loo et al. 2012). Specifically, studies have been conducted seeking to identify symptom-based dimensions and subtypes via statistical analysis of symptom co-occurrence using factor analysis, principal component analysis and latent class analysis (Chen et al. 2000; Aggen et al. 2005; Shafer, 2006; Carragher et al. 2009; Aggen et al. 2011; Cole et al. 2011; Mezuk & Kendler, 2012; Hybels et al. 2013; Buhler et al. 2014; Li et al. 2014a, b; Rodgers et al. 2014; Fried et al. 2016). Tightly correlated symptom sets are important as they might constitute dimensions or subtypes that imply different aetiologies and/or treatment responses. However, as summarised in a recent systematic review, these studies have failed to generate replicable results (van Loo et al. 2012).

Two methodological issues have not yet been considered, however. First, previous studies failed to disentangle two distinct sources of correlation between any two symptoms. Such correlations may be: (1) due to differences in overall severity (i.e., individuals with more severe depression score higher for all symptoms than individuals with less severe depression; hence, symptoms A and B are correlated); or (2) due to a specific profile of symptom correlations (e.g., individuals scoring high for symptom A could typically score high for symptom B too, but not necessarily for symptom C, which is more closely linked to symptom D). Therefore, it may be more appropriate to study symptom correlations adjusted for overall depression severity. If a structure underlying the symptoms exists, it should be revealed more clearly in this way.

Second, most studies were conducted on samples of depressed individuals. Studies that examined the dimensions of depression in general population samples, however, consistently revealed one single dimension of depression severity (Muthén, 1989; Aggen et al. 2005; Aggen et al. 2011; Cole et al. 2011; Mezuk & Kendler, 2012; Familiar et al. 2015). Thus, in the general population, depression was found to be a uni-dimensional construct. Apparently, it makes a difference whether one studies the general population or depressed individuals only, but this issue remained unaddressed.

The present study's main objective was to examine the dimensional structure of the nine symptoms of depression listed in the DSM-IV (a) when adjusting symptom correlations for overall depression severity; and (b) in general-population samples v. subsamples of currently depressed individuals. We adopted a dimensional approach, since evidence increasingly suggests that many psychiatric syndromes, including depression, are continuous and hence dimensional rather than categorical (Slade & Andrews, 2005; Goldberg et al. 2009; Prisciandaro & Roberts, 2009; Markon et al. 2011; Haslam et al. 2012; Eaton et al. 2013).

Materials and methods

Study design

We used: (a) longitudinal data from the Cohort Study on Substance Use Risk Factors (C-SURF); and (b) cross-sectional data from the U.S. National Health and Nutrition Survey (NHANES). In total, we considered four samples: C-SURF baseline, C-SURF follow-up, NHANES men and NHANES women. Comparing C-SURF baseline and follow-up data permitted us to examine whether results were replicable across two time points in the same sample. Comparing the C-SURF and NHANES data allowed us to examine whether results were replicable in two different populations.

Each of the four samples was analysed twice: once in the full (general-population) sample, and once only in the subsample of participants within a current mild-to-severe depressive episode, generating eight analytical samples in total.

Participants

C-SURF is a large cohort study examining young men in Switzerland, for which details on sampling and non-response bias have been published elsewhere (Studer et al. 2013a, b). It was designed to be representative of young non-institutionalised Swiss men. The study protocol was approved by the Ethics Committee for Clinical Research at Lausanne University Medical School (protocol number 15/07), and all subjects consented to participate.

5990 men completed the baseline survey between September 2010 and March 2012. Of these, 107 (1.8%) were excluded for missing data on the depression items. Of these, 5155 (87.6%) answered all necessary items of the follow-up survey performed between January 2012 and April 2013. The mean time elapsed between baseline and follow-up was 1.3 years (standard deviation 0.2).

NHANES is a continuous cross-sectional survey released in 2-year cycles (Centers for Disease Control and Prevention (CDC). National Center for Health Statistics (NCHS)). It was designed to be representative of the non-institutionalised U.S. civilian population. NHANES study protocols were approved by the National Center for Health Statistics Research Ethics Review Board, and all participants consented.

We included NHANES data from NHANES cycles 2005–2012 for men and women from 18 to 28 years old, an age range chosen to resemble the C-SURF cohort. Men and women were analysed separately. Of the total 2371 men and 2542 women, 197 (8.3%) and 298 (11.7%) were excluded for missing depression items data.

Measures

C-SURF: Self-reported depressive symptoms were assessed via the Major Depressive Inventory – WHO-MDI (Bech et al. 2001; Olsen et al. 2003). This validated measure covers DSM-IV and ICD-10 depression symptoms over the past 14 days, using 12 items with six-point answer scales ranging from ‘never’ (0) to ‘all the time’ (5). Items were aggregated into the nine DSM-IV symptoms, as proposed previously (Bech et al. 2001) (Table 1). Subjects were classified as having ‘no’, ‘mild’, ‘moderate’ or ‘severe’ depression based on the MDI summation score (Olsen et al. 2003). For correlation and factor analyses, symptoms were dichotomised into present/absent, as per ICD-10 definitions (Bech et al. 2001) (Table 1).

Table 1.

Symptoms of depression in the ICD-10-based WHO-MDI and the DSM-IV-based PHQ-9

				Dichotomisation ruleb
Symptoms as assessed in the WHO-MDI (ICD-10)		Translation rulea	Symptoms as assessed in the PHQ-9 (DSM-IV)	WHO-MDI	PHQ-9
Depressed mood			Depressed mood	‘Most of the time’	‘More than half of the days’
Anhedonia			Anhedonia	‘Most of the time’	‘More than half of the days’
Lack of energy/fatigue			Lack of energy/fatigue	‘Most of the time’	‘More than half of the days’
Feelings of worthlessness	}	Highest score	Feelings of worthlessness and guilt	‘More than half of the time’	‘More than half of the days’
Feelings of guilt
Life not worth living			Life not worth living	‘More than half of the time’	‘More than half of the days’
Concentration problems			Concentration problems	‘More than half of the time’	‘More than half of the days’
Feeling restless	}	Highest score	Psychomotor disturbances	‘More than half of the time’	‘More than half of the days’
Feeling subdued or slowed down
Sleeping problems			Sleeping problems	‘More than half of the time’	‘More than half of the days’
Reduced appetite	}	Highest score	Appetite changes	‘More than half of the time’	‘More than half of the days’
Increased appetite

ICD-10, International Classification of Diseases 10th version; DSM-IV, Diagnostic and Statistical Manual of Mental Disorders 4th edition; WHO-MDI, World Health Organization Major Depression Inventory; PHQ, Patient Health Questionnaire.

^aTranslation rule to combine ICD-10 symptoms into DSM-IV symptoms. The rule is to take the highest value of the relevant ICD-10 symptoms to represent the corresponding DSM-IV symptom.

^bThreshold for scoring the symptom as ‘present’.

NHANES 2005–2012: Self-reported depressive symptoms were assessed via the Patient Health Questionnaire (PHQ-9) (Kroenke et al. 2001). This validated measure covers the nine DSM-IV depression symptoms over the past 14 days (Table 1). Four answer options are provided, ranging from ‘not at all’ (0) to ‘nearly every day’ (3). Participants were classified as having ‘no’, ‘mild’, ‘moderate’ or ‘severe’ depression based on the PHQ9 summation score (Kroenke et al. 2001). Note that the threshold score we used to denote ‘mild’ depression was termed ‘moderate’ depression by Kroenke et al. This threshold resembled most closely the threshold for ‘mild’ depression that we used in the C-SURF sample, in terms of the percentage summation score required for the diagnosis (40% in C-SURF, 37% in NHANES). For correlation and factor analyses, symptoms were dichotomised into present/absent, as per DSM-IV definitions (Kroenke et al. 2001) (Table 1).

Statistical analysis

First, we examined tetrachoric correlations of the depression symptoms. To compare the correlations of each general-population sample with those of the corresponding subsample of currently depressed subjects, we calculated the ratio of the squared correlations for each symptom pair and used Steiger's test to formally examine the hypothesis that the two correlation matrices differed (Steiger, 1980). Steiger's test sums the squared differences of the Fisher transformed correlations of the two matrices and tests this sum against the chi-square distribution. Second, we assessed the dimensionality of the depression symptoms in three steps, each step conducted separately for each of the eight samples to determine whether the results were replicable.

Step 1: We first performed one-factor confirmatory factor analysis (CFA) and exploratory factor analysis (EFA). CFA consisted of the nine symptoms as indicators of an underlying depression factor, thereby modelling overall depression severity (model 1). With EFA, we tested one- to seven-factor models to determine which best fit the data. Both CFA and EFA were estimated using mean and variance-adjusted weighted least squares (WLSMV) estimations, which is the standard for categorical indicators (Barendse et al. 2014; Li et al. 2014a). Model fit was evaluated via standard criteria for good model fit (Aggen et al. 2005; Li et al. 2014a): root-mean-square error of approximation (RMSEA) ≤0.05; comparative fit index (CFI) ≥0.95; and Tucker–Lewis index (TLI) ≥0.95. For EFA, the model with the lowest number of factors achieving these criteria was adopted (model 2).

Step 2: From model 1, we derived the modification indices for the residual symptom co-variances as indicators of symptom pairs correlated beyond the general factor (i.e., as indicators of substantial severity-adjusted symptom correlations). Modification indices estimate the degree of improvement in model fit if the corresponding parameter is included in the model (Brown & Moore, 2012). Consequently, the modification index of a residual co-variance indicates whether the model would fit better if this co-variance was included in the CFA model.

We considered a modification index ≥3.84 statistically significant (Brown & Moore, 2012). We then re-fitted the one-factor CFA, this time including the residual symptom co-variances revealed by the modification indices (model 3). The residual symptom correlations derived from this CFA model generated an estimate of symptom correlations corrected for overall depression severity. If there is a dimensional structure beyond overall depression severity, these correlations: (a) should be replicable across the samples and (b) form interpretable symptom clusters. Because adopting CFA models based on modification indices is associated with a high risk of overfitting (MacCallum et al. 1992), we used the median of each modification index across 5000 case-based bootstrap samples.

Step 3: Finally, we estimated a series of bifactor models that, by definition, consist of one general factor and several group factors. Each indicator variable loads simultaneously on the general factor and one of the group factors (Reise et al. 2010). Thus, bifactor models allow for estimating group factors controlled for a general factor (Reise et al. 2010) and, hence, correspond directly to our notion of assessing depression dimensions (the group factors) controlled for overall depression severity (the general factor). If there is a replicable dimensional structure underlying the depression symptoms, at least one of the bifactor models should either converge with the residual correlations revealed in step 2, or provide an alternative model that is replicable across samples.

We used two approaches to identify the group factors:

• (a)We examined three theoretically-derived groupings (models 4.a1–3):
  • (1)Three genetic factors revealed by (Kendler et al. 2013).
  • (2)The common distinction of a cognitive/affective factor v. a somatic factor, as defined in the systematic review by van Loo et al. (2012);
  • (3)The symptoms most consistently found on a single factor in the review (van Loo et al. 2012) v. the remaining symptoms;
• (b)A non-rotated EFA that comprises several factors can be rotated into a bifactor structure (Jennrich & Bentler, 2011, 2012), resulting in an exploratory bifactor analysis (EBFA) that can then undergo confirmatory analysis. We derived the EBFA from the EFA calculated in step 1 and re-fitted it as a CFA model (model 4.b). However, only one EFA model revealed a sufficient number of factors to be rotated into a bifactor structure. We therefore used this bifactor model across all samples, rather than assessing a separate bifactor solution for each sample.

Bifactor models were estimated using WLSMV estimation and model fit was evaluated as in step 2.

Analyses were performed using R-software version 3.1.2 (R Core Team, 2014), particularly using the packages ‘psych’ (Revelle, 2013), ‘semTools’ (Pornprasertmanit et al. 2013) and ‘lavaan’ (Rosseel, 2012). R-scripts are available at https://osf.io/a6tuw/.

Results

Participants’ baseline characteristics are summarised in Table 2. Prevalence rates for current depression of at least mild degree ranged from 4.9 to 7.6%.

Table 2.

Baseline characteristics of study participants

C-SURF sample	Baseline	Depression (%)a	Follow-up	Depression (%)a
Total	5883	6.1	5155	7.0
Age (M ± SD)	20.0 ± 1.2	–	21.3 ± 1.2	–
Below median	2940 (50.0%)	5.4	2583 (50.1%)	5.8
Above median	2937 (50.0%)	6.7	2570 (49.9%)	8.3
Education
Primary school	2842 (48.5%)	5.9	358 (7.0%)	13.4
Secondary vocational education	1870 (31.9%)	6.0	2871 (55.9%)	6.4
Secondary school education	1054 (18.0%)	6.5	1659 (32.3%)	7.1
Above secondary	99 (1.7%)	6.1	248 (4.8%)	4.4
Linguistic region
German	2658 (45.2%)	5.3	2335 (45.3%)	5.9
French	3225 (54.8%)	6.7	2820 (54.7%)	8.0
NHANES 2005–2012 samples	Men	Depression (%)a	Women	Depression (%)a
Total	2174	4.9b	2244	7.6b
Age (M ± SD)	22.2 ± 3.3	–	22.4 ± 3.3	–
Below median	1218 (56.0%)	5.6b	1195 (53.3%)	7.6b
Above median	956 (44.0%)	4.7b	1049 (46.7%)	8.9b
Education
Primary school	650 (29.9%)	7.8b	531 (23.7%)	15.1b
Secondary education	607 (27.9%)	6.0b	575 (25.6%)	10.1b
Above secondary	915 (42.1%)	3.8b	1138 (50.7%)	6.2b
Race
Non-Hispanic white	764 (35.1%)	4.2b	778 (34.7%)	6.9b
Non-Hispanic black	552 (25.4%)	6.7b	548 (24.4%)	14.5b
Mexican American	501 (23.0%)	5.2b	503 (22.4%)	9.6b
Other	357 (16.4%)	8.3b	415 (18.5%)	9.1b

M, mean; SD, standard deviation.

^aPrevalence of current depression of at least mild-moderate degree.

^bAverage prevalence rate calculated across the NHANES cycles 2005–2012. The prevalence rates within each cycle were calculated using weighted data.

Symptom correlations

Substantial symptom correlations were revealed in the general-population samples (median correlations ranging from r = 0.55–0.74), while correlations in the depressed samples were surprisingly weak (median correlations from r = 0.04–0.24, Table 3). Correlations were greater in the general-population samples, and these differences were pronounced: in average correlations were higher by a factor ranging between 8.4 and 30.9 across samples (Table 4). Steiger tests confirmed that all general-population sample correlation matrices differed significantly from their counterparts in the depressed samples (Table 4). Only one correlation among women (‘life not worth living’ and ‘appetite changes’) was slightly higher in the depressed sample (ratio of squared correlation = 0.8, Table 3).

Table 3.

Summary of tetrachoric correlations of the nine DSM-IV depression symptoms across general-population samples and subsamples of currently depressed subjects

Sample	Symptom correlations
	M	IQR	Range
General population
C-SURF baseline	0.74	0.65–0.78	0.52–0.85
C-SURF follow-up	0.69	0.64–0.78	0.49–0.85
NHANES 2009–2012 men	0.55	0.51–0.60	0.44–0.73
NHANES 2009–2012  women	0.58	0.52–0.61	0.31–0.75
Depressed
C-SURF baseline	0.24	0.09–0.34	−0.05–0.69
C-SURF follow-up	0.22	0.12–0.38	−0.06–0.65
NHANES 2009–2012 men	0.04	−0.04–0.19	−0.24–0.47
NHANES 2009–2012  women	0.09	−0.03–0.17	−0.34–0.38

M, median; IQR, inter-quartile range.

Table 4.

Comparison of tetrachoric correlations of the nine DSM-IV depression symptoms in general-population samples v. subsamples of currently depressed subjects

		Steiger testa		Ratios of squared correlationsb
Samples compared		χ2 (df)	p-value	M	IQR	Range
C-SURF baseline	General-population/depressed	5840.5 (36)	<0.0001	8.4	4.9–51.8	1.5–221100.0
C-SURF follow-up	General-population/depressed	4777.6 (36)	<0.0001	9.4	3.7–28.8	1.5–561.7
NHANES 2009–2012 men	General-population/depressed	1362.7 (36)	<0.0001	30.9	6.7–128.8	2.5–22460.0
NHANES 2009–2012 women	General-population/depressed	2292.1 (36)	<0.0001	14.2	8.1–46.2	0.8–2472.0

M, median; IQR, inter-quartile range; df, degrees of freedom.

^aTests the hypothesis that two correlation matrices differ from each other.

^bFor each symptom pair, its squared correlation in the general-population sample was divided by its squared correlation in the corresponding sample of depressed. Ratios >1.0 indicate that the correlation was higher in the general population than among depressed.

Factor analyses

Step 1 revealed that the one-factor model fit the data very well in all general-population samples. This was revealed by both CFAs and EFAs (Table 5, models 1 and 2). In contrast, in depressed samples, no replicable dimensional structure was identified. Specifically, the one-factor CFAs failed to achieve good model fit in three of four samples and the EFAs revealed different numbers of factors across samples.

Table 5.

Summary of exploratory and confirmatory factor and bifactor analyses of the nine DSM-IV depression symptoms in general-population samples and subsamples of currently depressed subjects

	Model 1 (1-factor CFA)			Model 2 (EFA)				Model 3 (model 1 including residual co-variances)			Model 4.a1 (theoretical bifactor model 1)
Sample	CFI	TLI	RMSEA	Number of factors extracted	CFI	TLI	RMSEA	CFI	TLI	RMSEA	CFI	TLI	RMSEA
General population
C-SURF baseline	0.994	0.993	0.026	1	0.994	0.993	0.026	1.000	1.000	0.000	–b	–b	–b
C-SURF follow-up	0.993	0.991	0.027	1	0.993	0.991	0.027	1.000	1.001	0.000	–b	–b	–b
NHANES 2005–2012 men	0.983	0.977	0.029	1	0.983	0.977	0.029	0.994	0.990	0.018	–b	–b	–b
NHANES 2005–2012 women	0.980	0.973	0.039	1	0.980	0.973	0.039	1.000	1.002	0.000	–b	–b	–b
Depressed
C-SURF baseline	0.867	0.822	0.084	4	0.991	0.944	0.047	0.989	0.978	0.029	–b	–b	–b
C-SURF follow-up	0.925	0.900	0.064	2	0.968	0.939	0.050	0.995	0.992	0.018	–b	–b	–b
NHANES 2005–2012 men	1.00	1.107	0.000	1	1.00	1.107	0.000	1.000	1.107	0.000	–b	–b	–b
NHANES 2005–2012 women	0.556	0.407	0.059	–a	–a	–a	–a	0.944	0.916	0.022	–b	–b	–b
	Model 4.a2 (theoretical bifactor model 2)				Model 4.a3 (theoretical bifactor model 3)			Model 4.b (exploratory bifactor model)
Sample	CFI	TLI	RMSEA		CFI	TLI	RMSEA	CFI	TLI	RMSEA
General population
C-SURF baseline	–b	–b	–b		–b	–b	–b	–b	–b	–b
C-SURF follow-up	–b	–b	–b		–b	–b	–b	–b	–b	–b
NHANES 2005–2012 men	1.000	1.004	0.000		–b	–b	–b	0.989	0.976	0.029
NHANES 2005–2012 women	0.997	0.994	0.019		–b	–b	–b	–b	–b	–b
Depressed
C-SURF baseline	0.959	0.917	0.057		–b	–b	–b	1.000	1.009	0.000
C-SURF follow-up	–b	–b	–b		0.975	0.951	0.045	–b	–b	–b
NHANES 2005–2012 men	1.000	1.545	0.000		1.000	1.432	0.000	–b	–b	–b
NHANES 2005–2012 women	–b	–b	–b		–b	–b	–b	–b	–b	–b

CFA, confirmatory factor analysis; EFA, exploratory factor analysis; CFI, comparative fit index; TLI, Tucker–Lewis index; RMSEA, root-mean-square error of approximation.

^aNone of the admissible models reached the criteria for good model fit.

^bModel inadmissible due to negative residual variance of at least one symptom.

Step 2 indicated that 50 of 288 (8 samples × 36 symptom pairs) possible residual co-variances (17.4%) were substantial. Including these residual co-variances in the CFAs improved the fit of all models and resulted in good-fitting models, except for the NHANES sample of depressed women (Table 5, model 3). Both positive and negative correlations were revealed, positive correlations ranging from 0.10 to 0.48 (median: r = 0.29) and negative correlations from  − 0.46 to −0.07 (median: r = −0.26, Table 6). However, the correlations failed to exhibit any replicable pattern across the samples, and 17 of the 50 correlations (34.0%) were not statistically significant (Table 6).

Table 6.

Correlations of the nine DSM-IV depression symptoms adjusted for overall depression severity as estimated by confirmatory factor analysis (model 3)

		General-population				Depressed
Symptom pair		C-SURF baseline	C-SURF follow-up	NHANES 2005–2012 men	NHANES 2005–2012 women	C-SURF baseline	C-SURF follow-up	NHANES 2005–2012 men	NHANES 2005–2012 women
Depressed mood	– Anhedonia	0.40	0.32		0.10
Depressed mood	– Fatigue/energy			−0.17	−0.14
Depressed mood	– Worthlessness/guilt			0.26	0.22	0.48			0.34
Depressed mood	– Life not worth		0.29				0.24
Depressed mood	– Concentration					−0.32
Depressed mood	– Sleep problems				−0.22
Depressed mood	– Appetite changes				−0.39
Anhedonia	– Fatigue/energy	0.32	0.40		−0.26	0.35	0.39
Anhedonia	– Concentration						−0.23
Anhedonia	– Psychomotor changes	−0.09				−0.33
Anhedonia	– Sleep problems	−0.08	−0.10		−0.27
Anhedonia	– Appetite changes	−0.11	−0.07
Fatigue/energy	– Worthlessness/guilt				−0.17
Fatigue/energy	– Life not worth	−0.35	−0.29			−0.15	−0.31
Fatigue/energy	– Sleep problems			0.34	0.39				0.33
Worthlessness/guilt	– Psychomotor changes				−0.35				−0.29
Life not worth	– Concentration					0.23
Life not worth	– Sleep problems				−0.31
Life not worth	– Appetite changes				−0.46
Concentration	– Psychomotor changes	0.20				0.28	0.26
Psychomotor changes	– Sleep problems	0.22				0.26
Psychomotor changes	–Appetite changes						−0.25
Sleep problems	– Appetite changes	0.22	0.23			0.38	0.17

Note. Correlations printed in bold are statistically significant with p < 0.05.

In Step 3, no bifactor model was replicable across the samples (Table 5, models 4.a1–4.b). The most stable model was model 4.a2, which achieved good model fit in three of four NHANES samples and one C-SURF sample. Note that all models were inadmissible in at least four of the eight samples, due to negative residual variances.

A closer look at the bifactor models revealed two issues (see Supplementary material available at https://osf.io/a6tuw/). First, 18 of 24 models that were inadmissible were inadmissible because at least one of the group factors consisted of one large factor loading, with all other loadings being virtually zero, thereby leading to a model that was empirically underidentified (Kline, 2011). Furthermore, this pattern of one very large loading with otherwise negligible loadings is indicative of overfactoring (i.e., the inclusion of unnecessary factors) (Rindskopf, 1984). In the remaining six inadmissible models, the majority or all of the loadings were non-significant for at least one group factor. Second, among the admissible models, six had at least one group factor with only non-significant factor loadings, and only two models had significant factor loadings across both the general and group factors. Thus, bifactor analysis provided no evidence for any dimensional structure existing beyond the general severity factor.

Discussion

Main findings

We sought to examine the dimensions underlying the nine DSM-IV depressive symptoms in young adults while adjusting symptom correlations for overall depression severity, and while comparing general-population samples v. subsamples of currently depressed individuals. Analyses revealed three main results. First, adjusting symptom correlations for overall depression severity left little substantial correlation between the symptoms, and we failed to find any evidence to support a replicable dimensional structure when correcting symptom correlations for overall depression severity. Second, in the general-population samples, symptoms correlated substantially and were uni-dimensional. Third, among depressed individuals, symptom correlations were mostly weak and there was no evidence of any replicable dimensional structure, regardless of whether or not correlations were adjusted for overall severity.

Our finding that depressive symptoms were uni-dimensional in the general population is totally consistent with results from previous studies that analysed combined samples of healthy and affected individuals. These studies included general population samples from the USA and Mexico, and youths ages 5–18 in the USA and UK (Muthén, 1989; Aggen et al. 2005; Aggen et al. 2011; Cole et al. 2011; Mezuk & Kendler, 2012; Familiar et al. 2015). Our results replicate these results among young adults in the USA and extend them to young Swiss men. Furthermore, they resemble recent results reported by Fried et al. who found that, as samples of American and Dutch depression patients became more heterogeneous with respect to overall depression severity, average symptom correlations increased and the factor structures became simpler (Fried et al. 2016).

Our finding indicates that, within the general population, depression can be described by a single dimension of severity, the main reason being that depressed individuals form a comparably homogeneous group, relative to the large majority of individuals who are mostly or completely symptom-free (data not shown). The sizeable symptom correlations found in the general population samples mainly reflected this difference between depressed and non-depressed individuals. As such, the common set of ICD-10 and DSM-IV depression symptoms has diagnostic utility identifying individuals suffering from depression within the general population, and the listed symptoms seem to capture the basic scope and severity of the syndrome well.

Conversely, the uni-dimensionality of depression symptoms was not present among depressed individuals and we found no evidence of any replicable dimensional structure. Our failure to uncover such a structure is totally consistent with a recent systematic review that failed to identify any conclusive evidence that data-driven dimensions or subtypes of depression exist (van Loo et al. 2012, see also Chen et al. 2000; Aggen et al. 2005; Shafer, 2006; Carragher et al. 2009; Aggen et al. 2011; Cole et al. 2011 Mezuk & Kendler, 2012; Hybels et al. 2013; Buhler et al. 2014; Li et al. 2014a, b; Rodgers et al. 2014; Fried et al. 2016). Furthermore, the factor structure of depression changes over time among depressed patients (Fried et al. 2016). It therefore seems unlikely that a dimensional structure underlies the symptoms of depression. Consequently, previous literature reporting and making use of depression dimensions should be considered cautiously.

The mostly-weak symptom correlations among depressed individuals were particularly surprising. Symptom correlations have seldom been reported in the literature and, hence, this phenomenon seems to have gone unnoticed. Nonetheless, Cramer et al. reported average symptom correlations among American adults with a ‘dysphoric episode’ (defined as an episode with at least two depressive symptoms), and Fried et al. reported average symptom correlations among American and Dutch depression patients. Consistent with our results, these authors reported average correlations ranging from r = 0.17 to 0.23 (Cramer et al. 2012) and from r = 0.12 to 0.39 (Fried et al. 2016). Additionally, previous studies failed to detect substantial stability of depression symptoms and subtypes between successive depressive episodes (Coryell et al. 1994; Lewinsohn et al. 2003; Melartin et al. 2004; Oquendo et al. 2004). Thus, symptom correlations seem to be rather weak, both within and between depressive episodes.

That symptom correlations were so weak implies that, even if a replicable dimensional model existed, it would be based on an average correlation of r ≈ 0.20; the vast majority of symptom variance would remain unexplained, as correlation-based models cannot explain symptom variance beyond these correlations. This agrees with two recent studies that uncovered highly-diverse symptom profiles among depression patients (Fried & Nesse, 2015; Zimmerman et al. 2015). For example, Fried and Nesse identified 1030 unique profiles of depression symptoms in a sample of 3703 depressed American outpatients, with the most frequent profile only occurring in 1.8% of patients (Fried & Nesse, 2015). One explanation of how such diverse profiles develop is that adverse life events and other risk factors exerted differential impacts on depressive symptoms (Keller & Nesse, 2006; Keller et al. 2007; Lux & Kendler, 2010; Fried et al. 2014; Fried et al. 2015) and appeared to change the symptoms’ correlation patterns (Cramer et al. 2012). Thus, an individual's symptom profile depends at least partially on the aetiological factors that provoked the depressive episode. Furthermore, these different aetiologies are likely to imply differential responses to various treatment options. For example, evidence indicates that depression related to negative life events and trauma is more responsive to psychotherapy than to medication, whereas depressed individuals with maladaptive personality traits may respond better to selective serotonin-reuptake inhibitors (Simon & Perlis, 2010).

Two final issues concern the recent emergence of network models as an alternative account of mental disorders (Bringmann et al. 2013; Goekoop & Goekoop, 2014; van Borkulo et al. 2014; Boschloo et al. 2015; Bringmann et al. 2015; van Borkulo et al. 2015; Beard et al. 2016). Network models are based on the premise that symptom inter-relationships reflect direct causal influences between symptoms, rather than underlying latent factors, as in the factor analysis framework. The exact relationship between factor and network models remains unclear, however (Molenaar, 2010; Ross, 2010), and various authors disagreed with the network proponents’ critique of the latent variable approach (Belzung et al. 2010; Danks et al. 2010; Haig & Vertue, 2010; Humphry & McGrane, 2010; Markus, 2010). Most importantly, no empirical comparison of these two approaches has yet been reported (Krueger et al. 2010). Thus, how and to what degree one would draw different conclusions when applying factor analysis v. network modelling to one and the same sample is unclear. Future research needs to address this issue.

Second, our results are likely of importance to network research, since they indicate that the choice of sample type can impact the strength of symptom relationships considerably. Since networks are also based on symptom relationships, this should be an issue in network research, too. Indeed, depression-related network studies have been based on all sorts of samples (Bringmann et al. 2013; Goekoop & Goekoop, 2014; van Borkulo et al. 2014; Boschloo et al. 2015; Bringmann et al. 2015; van Borkulo et al. 2015; Beard et al. 2016). Even more intriguing, it was recently found that global network connectivity increased as disorder severity decreased over time (Beard et al. 2016).

Limitations

Our study had several limitations. First, it was restricted to young adults, so the results’ generalisability must be re-examined in demographically-broader samples. Second, symptom lists that are more differentiated than the nine DSM-IV criteria might be required, especially considering the weak correlations we detected in our depressed samples. More differentiated symptoms might be needed to capture depression subtypes in patient samples. Note, however, that studies using more comprehensive symptom sets have thus far also failed to uncover replicable dimensions (van Loo et al. 2012). Third, we used dichotomised symptom scores to facilitate comparisons against previous research. Doing so, some information might have been lost. Future studies should evaluate more finely-grained symptom scales. Fourth, step 2 of our analysis was exploratory and included multiple testing. Note, however, that we used a bootstrap procedure and replicated our analyses across different samples to safeguard against this. Finally, contrary to subtype research using latent class analysis, a dimensional approach could not detect subtypes that are based on only one or two symptoms (if a subtype is defined by several symptoms, however, these symptoms would be tightly correlated and, hence, emerge as a dimension). Thus, whereas our results rule out a dimensional structure of depression, there might still be subtypes of depression characterised by the presence of one or two specific symptoms. Note, however, that previous research focusing on statistically-derived subtypes has also failed to reveal replicable results (van Loo et al. 2012).

Implications

Given prior research findings, our results have two implications. First, caution is warranted when considering studies assessing dimensions of depression because general population-based studies and studies of depressed individuals generate different data that can lead to different conclusions. This problem likely generalises to other models based on the symptoms’ inter-relationships (e.g., network models). Second, it appears that the two dominant aspects of depression are its overall severity and each individual's symptom profile. Whereas the overall severity aligns individuals on a continuum of disorder intensity that allows non-affected individuals to be distinguished from affected individuals, the clinical evaluation and treatment of depressed individuals should focus directly on each individual's symptom profile, since it seems to convey most clinically-relevant information.

Financial Support

This work was supported by the Swiss National Science Foundation (grant number FN33CS30_148493). The funder had no role in study design, data collection or analysis, decision to publish or preparation of the manuscript.

Conflicts of Interest

None.

Ethical Standards

All procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975 revised in 2008.

Availability of Data and Materials

The raw data of the C-SURF-cohort study and the NHANES study are available at http://www.c-surf.ch/en/30.html and at http://www.cdc.gov/nchs/nhanes/nhanes_questionnaires.htm.