Major depressive disorder (MDD) is one of the most important challenges in global mental health [1, 2]. In research, a continuing challenge is the diversity in the symptoms and pathophysiology of patients classified as having the disorder. MDD patients vary considerably in clinical presentation, course, treatment response, genetics and neurobiology [3–7]. One explanation for this diversity is that MDD has a polythetic definition; that is, a patient needs to satisfy some but not all symptoms. For the diagnosis at least five of nine symptoms including at least one of the two core symptoms must be present [8]. It follows that there are 227 possible combinations of symptoms leading to this diagnosis. This is such a wide array of possibilities that two patients classified as having MDD might have only a single symptom in common. This diversity raises the question whether it makes sense for the purposes of comparisons in research to consider all the people who qualify for the diagnosis of MDD as having a single disorder.

To overcome the problem of symptom diversity, several attempts have been made to specify more homogenous subgroups within MDD. Subtypes have been proposed based on specific combinations of symptoms (for example, melancholic depression, psychotic depression), onset (seasonal affective disorder, postpartum, early versus late in life), course (single, recurrent, chronic), or severity [6]. Most subtyping schemes are based on pattern recognition and ordering using distinctions observed in clinical practice. For instance, 11 subtypes of MDD were proposed in the Research Diagnostic Criteria (RDC), the forerunner of the current Diagnostic and Statistical Manual of Mental Disorders (DSM), based on combinations of clinical characteristics, follow-up patterns, and findings from family studies [9–11]. However, the value of such distinctions has been called into question by the disappointing results of attempts to use these and subsequent subtyping distinctions in clinical practice [6, 12, 13].

A different approach to discern useful subtypes with similar symptom profiles would be one that is data-driven; that is, which uses any of several statistical techniques to recognize patterns in reported symptoms of a heterogeneous group of subjects. These kinds of models have in common the fact that they reduce a large number of data from individuals to smaller numbers of latent variables based on similarity. Two dominant types of latent variable models are latent factor models and latent class models [14]. Latent factor models, such as exploratory factor analysis (EFA), reduce originally correlated variables to fewer latent factors (which might be specified as either correlated or uncorrelated) based on the correlations between the original variables [15, 16]. By contrast, latent class models, such as cluster analysis (CA) and latent class analysis (LCA) cluster individuals rather than variables into relatively homogeneous subgroups. These subgroups are based on measures of similarity between each pair of individuals summed across all the variables considered in the analysis [17]. As both types of models are designed to discover structure in the absence of pre-existing hypotheses about subtypes, they provide useful approaches for examining heterogeneity based on distinctions that are not known beforehand [14].

Both latent factor and latent class models have been used to study the possible existence of useful MDD subtypes. However, findings in patients with MDD have not been systematically reviewed and thus the overall outcome is currently unclear. One of the questions, for instance, is whether a two-factor model applies to patients with MDD, as has been repeatedly found in patients with somatic illnesses. In somatically ill patients, latent factor analyses identified two main dimensions of depressive symptoms: one factor consisting of depressed mood, loss of interest, worthlessness, concentration problems and suicidality, and a second factor consisting of fatigue, appetite, sleep, and psychomotor disturbances [18, 19]. The first set of symptoms is typically referred to as 'cognitive' or 'cognitive/affective', whereas the second set is typically referred to as 'somatic' or 'somatic/affective'; however, whether a cognitive and somatic symptom dimension is generalizable to patients with MDD or is limited to patients with somatic illnesses and comorbid depressive complaints is unknown. Systematic reviews or meta-analyses of studies empirically investigating the depressive symptom profiles of patients with MDD have not yet been performed.

In search of data-driven subtypes of MDD, we performed a systematic review of published studies that used these latent variable models to distinguish symptomatic subgroups or symptom dimensions for patients with MDD. Our main question was whether those studies identify consistent subtypes or symptom dimensions. Second, we evaluated whether the characteristic symptoms of empirically derived subtypes resemble the descriptions of current specifiers as melancholic and atypical [8]. Third, we studied whether latent factor analyses in patients with depression reveals a cognitive and somatic symptom dimension, as has been found in patients with somatic illnesses.

Our results show that studies performed to date provide insufficient evidence to confirm the stable existence of clear data-driven symptomatic subtypes of depression. First, relatively few studies have been dedicated to the detection of data-driven MDD subtypes (20 of 1176 articles). Second, the outcomes of these few studies are conflicting. Latent class analyses mainly grouped patients on overall severity, but not in classes with qualitatively different symptom profiles, whereas latent factor analyses, most consistently identified a factor explaining the variance of a mixture of cognitive and somatic symptoms (s1, s2, s5b, s6), which seems in contradiction with a purely cognitive or somatic symptom dimension. However, the 13 identified factors differed to such an extent that generalizable conclusions are questionable. In short, the collected studies fail to give adequate evidence for the existence of qualitatively different subtypes or symptom dimensions of MDD. Thus, this lack of empirical support also holds for theoretical motivated subtypes such as as melancholic or atypical, or cognitive and somatic symptom dimensions in MDD.

Particularly notable is the great deal of diversity in the results. How can we explain this large diversity? From the collected results, it can be seen that all sorts of factors influence the outcomes of latent variable analyses; for instance, the included number of patients, and the severity and quantity of their symptoms obviously affect the resulting latent classes and factors. Presumably, there is considerable difference in symptom endorsement rates between the studies, as was reported from a recent mega-analysis of genome-wide association studies of patients with MDD [55]. Furthermore, the extensive number and differences of the questionnaires used, including all not recorded symptoms (weight gain, hypersomnia), is a likely contributor to the diverging results. On top of that, previous studies have found that in case of latent factor analyses, several items, including model choice (for example, component or common factor model), sample size, number and communalities of the variables, selection of the number of factors and model fit criteria, rotation method, and the degree of overdetermination, affect the stability and correspondence of the resulting factors [56–58]. An extra influence to take into account is the manifestly high correlation of symptoms within one questionnaire, as shown by some (not all) latent factor analyses on different questionnaires simultaneously. Thus, many theoretical choices preceding analysis are important determinants of the retrieved classes or dimensions.

The major insight from this review is that we should improve our research techniques considerably to find data-driven subtypes of depression. Of course, it is an open-ended question as to whether such a pattern exists. It is possible that there simply are no symptomatic subtypes or dimensions, and that latent variable techniques have failed to show consistent subtypes because of this fact. One could argue that if there were clear symptomatic subtypes or dimensions, a more consistent pattern would have emerged out of the data, regardless of the theoretical and measurement choices involved. The other possibility is that symptomatic subtypes and dimensions do exist, but that the techniques used to date did not succeed in identifying them. If the second possibility is true, a crucial question is how study methods could be improved to detect patterns in the data that have not yet been detected. Careful consideration of all theoretical and modeling choices that influence study outcomes will be required to answer that question.

One possible strategy to improve study methods is data enrichment, because it is clear that the quality of data crucially determines the quality of study outcomes. First, some of the studies reviewed above indicate the possible benefit of dynamic measurements, as they showed that differentiated symptom profiles might be clearer with more rather than less severe cases or for patients earlier rather than later in the treatment process. Therefore it would seem prudent to study changes in symptom structure and severity over the course of treatment, preferably also distinguishing the influence of medication. A second set of choices would involve the scales used to assess these symptoms. Some diagnostic instruments use dichotomous (yes/no) measures for each symptom, whereas others have gradual assessments. Gradual assessments, would of course be expected to provide textured differentiation, and in this way possibly lead to greater precision in detecting meaningful subtypes. Another concern is that the frequently used rating scales are primarily designed to be sensitive to change as opposed to capturing the detailed phenomenological picture of MDD. To address the potential heterogeneity in patients with depression, at least all DSM criteria, including all disaggregated symptoms (s3-s5), should be measured in a standardized fashion. A third possibility to enrich data would be the inclusion of other variables in addition to depressive symptoms in analyses. Concerning the symptoms to include in the evaluation, it seems clear that this set of symptoms should be a broad one, so as to allow for the possibility of detecting subtypes associated with symptoms beyond those in the current DSM and ICD systems. For example, there is some suggestion in the literature that there might be value in differentiating irritable from non-irritable MDD [59] and in assessing the influence of anxiety [60, 61]. Moreover, apart from symptoms, other indicators, such as hormone status, genetic profile, sex, and age, could also be included as variables in future analyses. The inclusion of these non-symptomatic variables possibly contributes to the unraveling of causal mechanisms of depressive subtypes, although we firmly believe that symptoms are the basis on which to start. Symptoms are non-invasive measures of disease; clinicians are trained to recognize and classify patients based on symptoms; and data on symptoms have been collected world-wide.

A second strategy to improve study methods involves the statistical approaches used to uncover depressive dimensions and subtypes. In addition to the latent factor and latent class approaches, we should consider complex CA models, especially those that use a canonical formulation to predict diverse outcomes [62, 63], and mixture models that combine features of latent class and item response theory models [64] or latent class and latent factor models [14]. Some recent studies have used factor mixture analyses to identify subtypes in other psychiatric disorders, such as attention-deficit/hyperactivity disorder [65], post-traumatic stress disorder [13], and schizophrenia [66]. To date, these approaches have not yet been used to search for subtypes of MDD, but are attractive alternatives in light of their success in detecting useful subtypes of other disorders. Obviously, in accordance with the described latent variable models, many theoretical factors should be considered before applying those new techniques [67]. However, if subtypes of MDD do exist, using different statistical methods to reveal their structure could be worthwhile.

Thus, future analyses should ideally explore several advanced statistical techniques on enriched datasets. An investigation of the possibilities and limitations of different modeling techniques seems more reasonable than adhering exclusively to the latent factor and latent class models used up to now. Mega-analyses of the MDD symptoms of different samples could be worthwhile, as combined data may have a positive effect on robustness and generalizability of the results [55, 68]. However, when performing mega-analyses, it is even more important to have rich datasets and to apply sophisticated modeling techniques to the data to accommodate inter-study heterogeneity [55]. Experiences with those new symptom-based classification attempts might inform other data-driven classification attempts that go beyond the DSM, such as the Research Domain Criteria [69, 70].

Finally, what is the value of searching for data-driven subtypes of MDD? We started our review with the observation that patients with MDD differ considerably in their symptomatic presentation, with over 200 possible symptom combinations. To date, theoretical motivated subtypes have not resolved the substantial population heterogeneity of MDD. Empirical discernment of subtypes with similar symptoms could give an impetus to research on etiology, course, and treatment. Improved statistical tools are available to discover patterns in rich datasets, therefore, data-driven subtyping of depression is a valuable approach to be explored.

There is no conclusive evidence for the existence of depressive symptom dimensions or symptomatic subtypes in adults with MDD. Many theoretical and modeling choices affect the results of latent variable analyses. If there is any structure to be discovered in the current heterogeneity, consideration of those choices would provide an useful starting point for future studies.

HMvL is a medical doctor and doctoral scholar at the University Center for Psychiatry (UCP), University Medical Center Groningen (UMCG), the Netherlands. PdJ is a professor at the UCP, UMCG and Center of Research on Psychology in Somatic Diseases at Tilburg University. JWR is professor at the Faculty of Philosophy of the University of Groningen. RCK is McNeil Family Professor of Health Care Policy at the Department of Health Care Policy at Harvard Medical School. RAS is professor and head of the UCP, UMCG.