Depression sum-scores don’t add up: why analyzing specific depression symptoms is essential

Despite extraordinary research expenditures and large genome-wide association studies, no pathognomonic biological markers of depression have been identified. This has been a major disappointment. In 1980, the DSM-III [10] preamble predicted that biomarkers associated with most diagnoses would be identified by the time the DSM-IV [11] appeared; 35 years and two DSM versions later, and with the exception of some neurological disorders, not one biological test for mental disorders was ready for inclusion in the criteria sets for the DSM-5, and not a single psychiatric diagnosis can be validated by laboratory or imaging biomarkers [12].

For depression research, results are specifically disappointing. In a recent large genome-wide association study with 34,549 subjects, no single locus reached genome-wide significance [13]. This is consistent with numerous other large genetic studies that have failed to identify any confirmed associations for MDD [14-17]. Studies predicting antidepressant response by common genetic variants have led to similarly disappointing results [18].

The analysis of specific symptoms offers opportunities to investigate biological factors that may be related to specific syndromes. Jang et al. [19] showed that 14 depression symptoms differ from each other in their degree of heritability (h² range, 0–35%). Somatic symptoms such as loss of appetite and loss of libido, as well as cognitions such as guilt or hopelessness (possibly reflecting heritable personality traits), showed higher heritability coefficients than other symptoms like negative affect or tearfulness. Another study [20] revealed differential associations of symptoms with specific genetic polymorphisms; for example, the symptom ‘middle insomnia’ assessed by the HRSD was correlated with the GGCCGGGC haplotype in the first haplotype block of TPH1. In addition, a recent report of 7,500 twins identified three genetic factors that exhibited pronounced differential associations with specific MDD symptoms [21]; the authors concluded that the “DSM-IV syndrome of MD[D] does not reflect a single dimension of genetic liability” (p. 599). Guintivano and Brown [22] analyzed several independent samples of post-mortem brains and blood samples from living subjects to document that 80% of the variation in one of the most relevant specific symptoms, suicidal behavior, could be explained by how polymorphisms of the gene SKA2 interacted with anxiety and stress.

Moving away from genes and gene expression to hormones, the hypothesis that depression can be caused by inflammation has received considerable attention in recent years [23,24]. However, evidence shows that less than half of the individuals diagnosed with depression exhibit elevated inflammatory markers [25], and elevated levels of cytokines are neither highly sensitive nor specific to MDD [26]. Furthermore, somatic symptoms such as sleep problems, appetite gain, and weight gain seem elevated in the context of inflammation [27-29], suggesting symptom specificity. A recent review acknowledges intragroup variability of MDD as main limitation of the research on inflammation and depression [26], and suggests that future analyses of distinct endophenotypes may move the field forward.

In summary, individual depression symptoms differ in their biological correlates. This underlines the heterogeneous nature of depression, which may in turn explain the lack of progress in validating depression diagnosis with biomarkers. Analyzing associations between symptom sum-scores and genetic markers can only capture the shared genetic variance of all symptoms, which may be low. A symptom-based approach offers opportunities for future research that could provide a potential partial explanation for the “mystery of missing heritability” [30] – the conundrum that specific genetic markers explain only small proportions of the variance even for mental disorders that are highly heritable. Specific markers may correlate better with specific symptoms independent of diagnostic categories – genes do not read the DSM [31]. Studies on symptom-polymorphism associations instead of syndrome-polymorphism associations, similar to the one conducted by Myung et al. [20], may prove insightful.

Several large meta-analyses of clinical trials have demonstrated that antidepressants outperform placebos in less than half of the trials, and that clinically relevant improvements can be documented only for a minority of severely depressed patients [32-34]. Part of the difficulty may be that measuring antidepressant efficacy via sum-scores conceals important effects on specific symptoms [35]. Little research has been conducted on the effect of antidepressants on individual depression symptoms compared to the mountain of literature on specific side effects.

Significant side effects for both tricyclic antidepressants and selective serotonin reuptake inhibitors have prevalence rates of up to 27% in clinical trials [36,37], and common side effects include insomnia, hypersomnia, nervousness, anxiety, agitation, tremor, restlessness, fatigue, somnolence, weight gain or weight loss, increased or decreased appetite, hypertension, sexual dysfunction, dry mouth, constipation, blurred vision, and sweating [38,39] (Table 1). Side effects vary across drugs, and some have more benign effects in specific domains. For instance, certain atypical antidepressants have a superior sexual side effect profile [40], and individuals treated with bupropion and nortriptyline show decreased rates of weight gain [41].

Curiously, some of the common side effects reported by patients are the very symptoms that are used to measure depression (Table 1). This means that reductions in sum-scores thanks to reduced depression are concealed by increases in sum-scores caused by drug side effects. In addition, the instrument most commonly used in clinical trials is the HRSD which, compared to other depression scales such as the BDI, abounds in somatic symptoms that resemble the side effect profile caused by antidepressant treatment [42].

The presence of particular symptoms has been used to predict treatment response. Sleep problems, for instance, reduce the efficacy of depression treatment [43]; patients with persistent insomnia are more than twice as likely to remain depressed [44], and insomnia can become chronic despite successful resolution of depressive symptoms [45]. Other symptoms also moderate treatment efficacy: anxiety symptoms reduce depression remission rates, successful anxiety treatment prolongs depression remission [46-48], and loss of interest, diminished activity, and inability to make decisions predict poorer antidepressant response [49].

The overlap of antidepressant side effects and depression symptoms provides a compelling reason for analyzing symptoms such as weight problems, sleep problems, or sexual dysfunction separately from sum-scores. A detailed analysis of how different antidepressants influence specific symptoms may improve our ability to determine antidepressant efficacy.

Risk factors identified for depression include previous episodes of depression [50], demographic variables such as age and sex [51,52], and personality traits such as neuroticism [53]. Statistical models use these and other risk factors to predict the presence or absence of depression.

However, risk factors differ for different symptoms as first demonstrated by Lux and Kendler [54], who analyzed the associations of 25 risk factors on 9 different symptoms in a cross-sectional study of 1,015 individuals. The influence of risk factors differed substantially for different symptoms in a pattern the authors found difficult to reconcile with the general practice of summing symptoms. In another large prospective study, risk factors for depression in medical residents showed strong differential impact on changes of depression symptoms over time [55]. Restricting analyses to a sum-score suggested that women are at greater risk to develop depression during residency, but analyzing individual symptoms revealed that male residents were more likely to experience elevated levels of suicidal ideation under stress, whereas female study participants were more prone to develop increases in sleep, appetite, and concentration problems as well as fatigue.

Adverse life events are well-established risk factors for depression [56], and the depression symptoms individuals experience after a life event seem to depend on the nature of the event. In one experimental study, as well as different cross-sectional and longitudinal investigations of college students and adult samples [57-61], specific types of life events were associated with distinct patterns of depressive symptoms. For instance, after a romantic breakup, individuals mainly experienced depressed mood and feelings of guilt, whereas chronic stress was associated with fatigue and hypersomnia [59].

Overall, risk factors differ substantially for different depressive symptoms, and sum-scores obscure such insights. Studying the etiology of specific depression symptoms may enable the development of personalized prevention that focuses on specific problems and symptoms before they transition into a full-fledged depressive episode.

Most depressed individuals suffer from severe functional impairment in various domains of living such as home life, workplace, or family [2,62]. Their impairment is often long-lasting and equal to that caused by other chronic medical conditions such as diabetes or congestive heart failure [63,64]. The question of whether individual depression symptoms differentially impair psychosocial functioning is thus of great importance.

In a study of 3,703 depressed outpatients, DSM-5 criterion symptoms varied substantially in their associations with impairment [65]. Sad mood explained 20.9% of the explained variance of impaired functioning, but hypersomnia only contributed 0.9%. Symptoms also differed in their impacts across impairment subdomains. For example, interest loss had high impact on social activities, whereas fatigue most severely impacted home management. The overall findings are consistent with an earlier study documenting differential impact of DSM-III criterion symptoms of depression on functioning [66].

While these results require replication in different samples, they offer further evidence for the value of considering depression symptoms separately. Not all symptoms contribute equally to severity ratings, and two individuals with similar sum-scores may suffer from dramatically different levels of impairment.

Measuring depression severity by sum-scores of symptoms ignores a plethora of information pertaining to the intra-individual development of depression, including the power of individual symptoms to cause other symptoms.

Insomnia, for example, leads to psychomotor impairment [67], cognitive impairment [68], fatigue [69], low mood [70], and suicidal ideation or actual suicide [71] – symptoms that closely resemble DSM symptomatic criteria for depression (psychomotor problems; fatigue; diminished ability to think or concentrate, or indecisiveness; suicidal ideation). A meta-analysis of laboratory-based sleep loss studies documented the strength of these effects: sleep-deprived subjects performed 0.87 standard deviations (SD) lower than the control group on psychomotor tasks, 1.55 SD lower on cognitive tasks, and reported mood 3.16 SD lower than the control group. Collapsing over all three measures, performance of sleep-deprived subjects at the 50^th percentile in their group was equivalent to subjects at the 9^th percentile in the control group [72]. Another recent meta-analysis revealed that psychiatric patients with sleep disturbances are about twice as likely to report suicidal behaviors compared to patients without sleep problems, a finding that generalized across various conditions including MDD, post-traumatic stress disorder (PTSD), and schizophrenia [73].

Hopelessness describes negative expectancies about the future [74]. Although not part of the DSM-5 MDD criteria, it plays a major role in the cognitive triad originally described by Beck [75], performs more strongly than some DSM symptoms in distinguishing depressed from healthy individuals [76], and is assessed in various scales. Numerous studies have confirmed the predictive role of hopelessness for suicidal ideation and suicide [71]. The effects are long-reaching: hopelessness predicted suicidal thoughts, attempts, and actual suicide up to 13 years into the future in a large community sample [77], and was identified as a predictor of suicide among psychiatric patients followed for up to 20 years [78]. The association of hopelessness and suicide generalizes from depressed individuals to patients with other psychiatric conditions [79,80], once more underlining symptom specificity irrespective of a given diagnosis. Hopelessness predicts suicide better than the sum-score from an inventory assessing multiple depressive symptoms [80], and mediates the effect of rumination on suicidal ideation and other depressive symptoms in children and undergraduates [81,82]. In adolescents, rumination predicts the development of subsequent symptoms of depression, bulimia, and substance abuse, while depression and bulimia symptoms in turn predict increases in rumination [82,83]. Symptoms are associated in complex dynamic networks that can form vicious circles which transcend any specific diagnosis, a notion that is also supported by recently developed self-report methods demonstrating complex interactions among symptoms [84,85].

In contrast to longitudinal studies that span months or years, experience sampling methods that allow for the analysis of a large number of timepoints over a comparably short timeframe have consistently revealed short-term associations among depression symptoms (for a review, see [86]). For example, sleep quality predicted affect during the next day in a sample of 621 women, while daytime affect was not related to subsequent night-time sleep quality [70], implying a clear direction of causation. Complementing such group-level analyses with longitudinal idiographic studies is likely to contribute important information. Bringmann et al. [87] documented differences among depressed patients in the way their emotions impacted each other across time; for instance, they found the autoregressive coefficient of rumination to vary substantially across participants – rumination at a given timepoint strongly predicted rumination at the next timepoint for some individuals but not for others. Another study identified heterogeneity in the direction of causation between depression symptoms and physical activity [88]. Overall, a growing chorus of voices advocates the study of inter-individual differences [89-91] which may pave the way towards the development of more personalized treatment approaches. Heterogeneity may also help to resolve controversies about how some symptoms cause others. Sleep deprivation, for instance, has rapid mood-enhancing effects in some depressed patients [92], but other reports suggest that sleep difficulties cause low mood [70].

The notion that symptoms trigger, influence, or maintain other symptoms is widely recognized in clinical practice. A major goal in cognitive therapy is trying to break causal links between different MDD symptoms [75] and approaches like mindfulness-based cognitive therapy suggest that stopping rumination prevents it from causing other depression symptoms [93]. Kim and Ahn [94] demonstrated that causally central depression symptoms (symptoms that trigger many other symptoms) are judged to be more typical symptoms of depression by clinicians, are recalled with greater accuracy than peripheral symptoms, and are more likely to result in an MDD diagnosis. The authors concluded that clinicians think about causal networks of symptoms in ways far more sophisticated than the atheoretical DSM approach of counting symptoms.

Psychometric techniques such as factor analysis (grouping symptoms) and latent class analysis (grouping individuals) are commonly used to address heterogeneity of MDD. In a more detailed discussion of these methods we draw two general conclusions, both of which support the study of individual symptoms [5].

First, extensive efforts to identify specific forms of treatment effective for specific depression subtypes have been disappointing. There has been little agreement about the number and nature of depression subtypes [95-98], and limited success in identifying external validators for subtypes [99-102]. A recent systematic review that compared the results of 34 factor and latent class analyses concluded that they did not provide evidence for valid subtypes of MDD [95], suggesting the analysis of individual symptoms.

Second, most rating scales for depression are multifactorial and do not measure one underlying factor [103-105]. However, individual symptoms are often at least moderately inter-correlated [106], and the first factor – often a general mood factor or higher-order factor – explains substantially more variance than subsequent factors [103,107]. This means that sum-scores certainly carry information about the general psychopathological load of a particular person, but that the approximation may be fairly rough and that summing symptoms may ignore important information [5,108] (for instance, because MDD symptoms are differentially impairing [65] and because sum-scores do not take into account reciprocal interactions of symptoms [108]).

Applying psychometric tools such as item response theory (IRT) and structural equation modeling (SEM) can yield important insights on the level of individual symptoms because they allow the examination of exact relationships between symptoms and underlying dimensions. One example technique that helps to understand such relations is differential item functioning; a prior study testing for this revealed that different MDD risk factors, such as neuroticism or adverse life events, impact on specific depression symptoms, implying that symptoms are ‘biased’ towards certain risk factors [55]. A second practical application is research on residual dependencies. A major assumption of IRT and SEM models is that the underlying latent variables fully explain the correlation of the manifest indicators. This is rarely the case [109], and especially unlikely in the context of MDD, seeing that symptoms influence each other directly [86,110]. Ignoring such residual dependencies unaccounted for by the latent variables, however, can substantially bias inferences [109,111].

The first group of research implications is for the measurement of depression symptoms. After reviewing many depression rating scales, Snaith [42] concluded that “The measurement of ‘depression’ is as confused as the basic construct of the state itself” (p. 296). Below we explain why this is the case, and suggest several important steps that could reduce confusion.

While the more traditional SEM and IRT models assume that all depression symptoms share a common cause and are locally independent (i.e., uncorrelated beyond the common cause; see [109]), a growing number of studies have shown that symptoms can trigger other symptoms. A recently developed framework – the network approach to psychopathology – allows the study of such dynamic interactions. Network models estimate the relationships among symptoms within or across time [106,109,110], and offer a new perspective on why symptoms cluster. While latent variable models explain symptom covariation by a latent factor that is viewed as the common cause of all symptoms, network models suggest that syndromes are constituted by the connections among symptoms. This perspective encourages consideration of how vicious circles of symptoms can fuel each other, an alternative to the schema in which all symptoms arise from a single brain disorder.

We anticipate fundamental advances from researchers who report and analyze information about specific symptoms. For instance, inconsistent reports about the efficacy of antidepressants may result from samples with different symptom patterns that may respond differently to different agents. A meta-analysis to test this hypothesis requires data on individual symptoms that is not available in the Food and Drug Administration database of depression studies.

A recent study by Uher et al. [132] suggests the available opportunities. The authors found that individuals with high baseline levels of systemic inflammation exhibited increased depression recovery under nortriptyline, while low inflammation levels were associated with superior depression improvement under escitalopram, supporting earlier work on the topic [133]. These results are especially interesting considering that inflammation levels are particularly elevated among depressed individuals with somatic symptoms [28], specifically appetite and weight gain [27]. If patients with high and low baseline inflammation levels exhibit different symptoms, it should be possible to select study participants who will respond to a particular drug. Finding biological markers for specific depressive symptoms will open new research vistas.