Key Points

Question听 Is there a common pattern of symptom combinations across mental disorders?

Finding听 This cross-sectional study found a specific pattern across 4 empirical samples (N鈥�=鈥�155鈥�474), with 41.7% to 99.8% of symptom combinations being reported by less than 1% of the sample, while the 1% most frequent combinations were highly prevalent in 33.1% to 78.6% of the corresponding sample. Because of the interdependence of a disorder鈥檚 symptoms, not all symptom combinations are equally likely.

Meaning听 Polythetic definitions lead to a common pattern of symptom heterogeneity: the presence of few prototypical and many atypical symptom combinations.

Importance听 DSM criteria are polythetic, allowing for heterogeneity of symptoms among individuals with the same disorder. In empirical research, most combinations were not found or only rarely found, prompting criticism of this heterogeneity.

Objective听 To elaborate how symptom-based definitions and assessments contribute to a distinct probability pattern for the occurrence of symptom combinations.

Design, Setting, and Participants听 This cross-sectional study involved a theoretical argument, simulation, and secondary data analysis of 4 preexisting datasets, each consisting of symptoms from 1 of the following syndromes: posttraumatic stress disorder, depression, schizophrenia, and anxiety. Data were obtained from various sources, including the National Institute of Mental Health Data Archive and Department of Veteran Affairs. A total of 155鈥�474 participants were included (individual studies were 3930 to 63鈥�742 individuals in size). Data were analyzed between July 2021 and January 2024.

Exposure听 For each participant, the presence or absence of each assessed symptom and their combination was determined. The number of all combinations and their individual frequencies were assessed.

Main Outcome and Measure听 Probability or frequency of unique symptom combinations and their distribution.

Results听 Among the 155鈥�474 participants, the mean (SD) age was 47.5 (14.8) years; 33鈥�933 (21.8%) self-identified as female and 121鈥�541 (78.2%) as male. Because of the interrelation between symptoms, some symptom combinations were significantly more likely than others. The distribution of the combinations鈥� probability was heavily skewed with most combinations having a very low probability. Across all 4 empirical samples, the 1% most common combinations were prevalent in a total of 33.1% to 78.6% of the corresponding sample. At the same time, many combinations (ranging from 41.7% to 99.8%) were reported by less than 1% of the sample.

Conclusions and Relevance听 This study found that within-disorder symptom heterogeneity followed a specific pattern consisting of few prevalent, prototypical combinations and numerous combinations with a very low probability of occurrence. Future discussions about the revision of diagnostic criteria should take this specific pattern into account by focusing not only on the absolute number of symptom combinations but also on their individual and cumulative probabilities. Findings from clinical populations using common diagnostic criteria may have limited generalizability to the large group of individuals with a low-probability symptom combination.

The systematic study of psychopathological symptoms and their co-occurrence has underpinned efforts to standardize the diagnosis and categorization of mental disorders. The DSM-5-TR¹ and International Classification of Diseases, 11th Revision (ICD-11),² the most used classification systems for mental disorders, use specific sets of co-occurring symptoms to define disorders. Because most disorder definitions are polythetic, many combinations of symptoms may fulfill a disorder鈥檚 criteria; individuals diagnosed with the same mental disorder can exhibit a wide range of symptoms. For example, the DSM-5-TR definition of posttraumatic stress disorder (PTSD) may be met in 636鈥�120 different ways.³

The large number of possible symptom combinations of mental disorders as defined by the DSM or the ICD has been subject of controversial debates.^3-6 This is because many differences, including the aim of a diagnosis or the underlying conceptualizations of mental disorders, manifest themselves in the discussion about the specific number and kind of symptoms included in a definition.^7,8 For example, some have argued to define disorders narrowly, ie, focusing on prototypical presentations, resulting in a minimal set of symptoms and high specificity of a diagnosis.^9,10 Others prefer an inclusive definition covering more possible combinations, resulting in higher sensitivity but more heterogeneity.^11,12 Similarly, there is a debate about whether mental disorders are better understood as distinct categories or as spectra of disordered states.¹³ The latter was considered in the latest revisions of the DSM and ICD (eg, in autism spectrum disorder in the DSM-5¹⁴ or personality disorders in the ICD-11¹⁵). While abstract in nature, these debates directly impact countless patients and health care professionals when the definition of a disorder is changed. It is therefore important to understand the assumptions and consequences of polythetic definitions. In this framework, mental disorders are considered latent, meaning they cannot be directly assessed, only their indicators (ie, symptoms) can be. Latent constructs cannot be defined by a single indicator (symptom) unless it is a trivial indicator (eg, depression is defined as having depression).³ Multiple indicators reflecting common as well as unique aspects of the latent construct are thus combined to better define the construct. However, no set of criteria will be sufficient to conclusively define a latent construct, and thus, a final or single correct definition does not exist.^3,16 It is therefore not surprising that the conceptual nature of polythetic definitions and the number and kind of criteria have been subject to an ongoing intense debate and that different panels of experts can come to different conclusions. For example, the ICD-11 definition of PTSD relies on fewer, more specific symptoms than the DSM-5-TR definition, resulting in fewer possible symptom combinations.^11,12 Such differences, as well as the large number of possible symptom combinations, have been repeatedly cited as evidence for the symptom heterogeneity of mental disorders^3,17-19 and have been blamed to hinder research into the underpinnings of mental disorders.⁴

Empirical research found that many symptom combinations are not or only rarely present in real-world samples. In a study of 3703 individuals from the STAR*D cohort with major depressive disorder (MDD),¹⁷ symptoms of depression were operationalized using a 12-symptom questionnaire rather than DSM criteria, thereby relying on symptoms that differ from the strict diagnostic criteria. Of the 4096 possible combinations, 1030 were found, with 48.6% of them being reported by only a single participant. Although all individuals had confirmed MDD, the most common combination was endorsement of no symptoms, displayed by 1.8% of the sample. The authors鈥� main conclusion of this highly cited article was that 鈥渄epression is not a consistent syndrome.鈥�¹⁷ However, the study did not question polythetic definitions of mental disorders. A similar study focused on DSM-IV criteria of MDD and identified 170 of 227 possible combinations among 1566 patients with MDD.²⁰ The most common combination was endorsement of all symptoms reported by 10.0% of the sample. In addition, 9 other combinations were relatively common as well, leading the authors to conclude that 鈥渄iagnostic heterogeneity may be more theoretical than actual鈥� and that 鈥渁 relatively small number of combinations could be considered as diagnostic prototypes.鈥�²⁰

Similar findings were reported in a study of 3511 individuals diagnosed with PTSD (as defined by the DSM-5), which found 2181 different symptom combinations, with 88.7% of them being reported only by a single participant.¹⁸ The most common symptom combination was the presence of all symptoms, which was reported by 13.4% of the individuals. These findings corroborate those from a previous study of 3810 active military personal meeting DSM-IV criteria for probable PTSD in which 1837 different combinations were observed, 83.5% of them endorsed by a single individual, while 25.1% of the individuals reported the same most common combination.¹⁹ The authors concluded that this is 鈥溾��raising concerns about the classification [of PTSD] itself鈥�.鈥� While reviewing the complete empirical literature is beyond the scope of this article, the studies described above exemplify diverse symptom combinations, and many of these studies are often cited as evidence highlighting the problematic nature of the heterogeneity of mental disorders.

The mechanism underpinning these empirical findings has not been elucidated conclusively. In this study, we attempt to examine why and how any symptom-based definition or assessment of mental disorders necessarily leads to differences in the probability of individual symptom combinations. Furthermore, we elaborate on the distribution of these probabilities, specifically noting that a small minority of the combinations will have a high probability, while the probability of most combinations is very low. These findings highlight an important consequence of polythetic definitions (but should not necessarily be considered an argument against them). We accomplish this by presenting a theoretical argument, a simulation study, and 4 empirical examples to underscore the generalizability of our findings. Finally, we discuss the relevance of our findings for future revisions of the DSM and the ICD-11, as well as for research based on polythetic definitions of mental disorders.

This study consists of 3 parts: a theoretical argument, a simulation, and secondary analyses of 4 preexisting empirical datasets. These samples were chosen to cover various symptoms, definitions, and assessments of disorders to increase our findings鈥� robustness. Samples 1, 2, and 4 contain electronic health record (EHR) data from the US Department of Veteran Affairs (VA). The West Haven VA鈥檚 institutional review board (IRB) waived informed consent. Sample 3 was gathered from the National Institute of Mental Health Data Archive (NDA). The NDA also ensured compliance with IRB approvals and that informed consent was obtained or waived. More details are provided in eMethods 1 and eTables 1 and 2 in Supplement 1.The results are reported following the Strengthening the Reporting of Observational Studies in Epidemiology () guidelines.

Sample 1 included 41鈥�543 veterans with a PTSD diagnosis in the EHR who also met DSM-5 criteria assessed with the PTSD Checklist for DSM-5 (PCL-5). The PCL-5 contains 20 items, each assessing the severity of a symptom in the past month. Individual items are rated on a 5-point scale ranging from 0 (not at all) to 4 (extremely). As suggested by the developers, scores of 3 or above defined symptom presence.²¹

Sample 2 included 46鈥�259 veterans with an MDD diagnosis in the EHR who also met the corresponding DSM-5 criteria as assessed with the Patient Health Questionnaire (PHQ-9). The PHQ-9 assesses 9 symptoms each rated on a 4-point Likert scale ranging from 0 (not at all) to 3 (nearly every day). A score of 2 or higher defined symptom presence.

Sample 3 was assessed using the Positive and Negative Syndrome Scale (PANSS) and rated on a 7-point scale from 1 (absent) to 7 (extreme). Symptom presence was defined as scores of 2 or higher, indicating at least minimal symptom presence.²² PANSS symptoms were mapped onto DSM-5 criteria, and 3930 individuals with a probable diagnosis of schizophrenia were included.

Sample 4 consisted of 63鈥�742 veterans with an anxiety disorder in the EHR whose symptoms of anxiety were assessed with the Generalized Anxiety Disorder questionnaire (GAD-7). The GAD-7 assesses 7 symptoms each rated on a 4-point scale from 0 (not at all) to 3 (nearly every day). A score of 2 or higher defined symptom presence. Only individuals who endorsed item 2 and 3 and had a total score of 10 or higher, indicative of clinically relevant symptom severity,²³ were included.

In every sample, we defined a symptom combination based on the presence or absence of each evaluated symptom. To assess the specific combination for each participant, symptoms were initially dichotomized using the previously mentioned cutoffs. The number of possible symptoms combinations was based on the logic of the DSM-5, except for sample 4 because it aimed to assess symptoms of anxiety across multiple disorders. For each sample, we assessed the number of observed combinations, their frequency, and the share of individuals reporting 1 of the 1% most common combinations. All analyses were conducted between July 2021 and January 2024 in the R environment (R Foundation).

When mental disorders are defined polythetically, not every symptom combination is equally common.²⁰ The probability of a combination is equal to the joint probability of the individual symptom and their interaction. Therefore, differences in symptom prevalence and interrelations, which are both well known, necessarily result in differences in the probability of combinations. For example, in cases of MDD, experiencing anhedonia is more common than suicidal ideation, and insomnia and concentration difficulties do co-occur more often than suicidal ideation and weight gain.

Based on the argument above, a simulation was conducted to calculate the probability distribution of combinations in a fictitious mental disorder. The disorder is unifaceted and defined by 5 symptom criteria of which at least 2 must be present to consider the disorder to be present, allowing for a total of 32 combinations. To simplify the simulation, no gating conditions were assumed. A population with 500 individuals was simulated with symptom ratings assumed to be normally distributed with different means and standard deviations for different symptoms and with symptoms related to each other. Then each symptom combination鈥檚 probability was calculated. The simulation was repeated 100 times (eMethods 2 in Supplement 1). The results are presented in Figure 1A (visualization) as well as in eMethods 2 in Supplement 1 (individual combinations and their frequency) and show a highly skewed distribution of the combinations鈥� probabilities with few highly probable combinations and a majority with much lower probabilities. An interactive version of the simulation, depicting the results numerically and visually, can be found online.²⁴

Among the 155鈥�474 participants, the mean (SD) age was 47.5 (14.8) years; 33鈥�933 (21.8%) self-identified as female and 121鈥�541 (78.2%) as male. The Table summarizes participant demographic data and occurrence of observed symptom profiles. In the PCL-5 sample, we found more than 3000 different combinations constituting only a small fraction of the theoretically possible. In the 3 remaining samples, almost all possible combinations were observed. Congruent with the simulation, stark differences regarding the frequency of the combinations within each sample were observed. In each sample, most symptom combinations occurred rarely, and less than 10% of combinations were observed frequently (Figure 1B). In numbers, the share of each combination that was reported by less than 1% of the sample varied between 41.7% (PANSS sample) and 99.8% (PTSD sample), with the share being higher the more combinations being theoretically possible. One of the 1% most common combinations was reported by 33.1% (PANSS sample) to 78.6% (PTSD sample) of the corresponding sample. The most common combination alone had a prevalence of 33.1% (PANSS sample) to 46.2% (PTSD sample). The 10 most common combinations of each sample are detailed in eTables 4-7 in Supplement 1. Correlations between individual symptoms are presented in eFigures 1-4 in Supplement 1.

Polythetic definitions or assessments of mental disorders allow for substantial heterogeneity.^3,16,25 Here, we detailed that because of the interdependence of individual symptoms, most of the possible symptom combinations have a low probability of occurrence, and at the same time, few combinations have an exceedingly high probability. Our empirical analyses showed generalizability of this pattern across 4 different settings, including symptom combinations of PTSD as defined by the DSM-5-TR as well as combinations of symptoms of anxiety assessed with the GAD-7 in a population with any anxiety disorder. This indicates that the observed pattern is inherent to symptom-based systems of mental disorders and should be considered when relying on such definitions.

Our findings underscore that symptom heterogeneity is not random but structured in a specific way. The low probability of most combinations may help explain why previous empirical studies often witnessed many, but not all possible, symptom combinations.^17,20,26 Many of these low-probability combinations are situated at the borderline between disorders and healthy states; ie, small changes in the severity of 1 symptom alone can result in a different classification of the combination. This reflects the often criticized 鈥渇uzziness鈥� of polythetic definitions of mental disorders.^27,28 Furthermore, these findings are also in accordance with descriptive findings in the existing literature, reporting that a majority of assessed individuals display a unique symptom combination.^17-20

Despite the presence of many low-probability combinations, prototypical presentations of mental disorders exist.⁹ Our simulation predicted that few combinations have a remarkably high probability of occurring, in fact an order of magnitude higher than most other combinations. Moreover, such combinations were characterized by the presence of all or almost all symptoms specified in the definition. This was confirmed by our empirical samples, with the most common combinations being those with endorsement of (almost) all symptoms, which is also consistent with previously reported results.^18-20

While our findings cannot and do not aim to resolve debates about heterogeneity in the definitions of mental disorders, they expand the knowledge about the assumptions and consequence of polythetic definitions. First, the presence of a few but highly prevalent symptom combinations and many low-probability combinations within a population with the same diagnosis resonates well with the clinical experience of encountering both kinds of patients in a clinical context: 鈥渢extbook鈥� presentations as well as diagnostically challenging cases with highly individual symptoms. Second, focusing not only on the number of possible combinations but also on their individual and combined probabilities helps to understand the effect of modifying diagnostic criteria. For example, how adding or omitting a specific symptom changes the probability of each individual symptom combination. This can be especially helpful when comparing the impact of different concepts, for example, when contrasting prototypical combinations with broader definitions or comparing categorical and dimensional interpretations of psychopathology. In such a context, an analysis similar to the one presented here could be used to determine minimum probability that a symptom combination must exceed to be considered relevant.

Research relying on symptom-based definitions of mental disorders must be aware that random samples are only selectively representative of the true clinical population. Because of the stark differences in probabilities of individual symptom combinations, 2 random samples will include the same prototypical combinations but different low-probability ones (Figure 2). Because symptoms, rather than diagnostic status, are associated with biological correlates (eg, inflammation markers²⁹ or genetics³⁰) and treatment response,^31-33 individuals with the same diagnosis but different symptoms cannot necessarily be assumed to share the same underlying biological or psychological mechanisms.⁴ Because low-probability combinations phenomenologically differ starkly from prototypical combinations, such correlates (eg, biomarkers) could differ across these 2 populations. Consequently, sampling from such a population can jeopardize the validity and generalizability of findings. For example, if most of the sample consists of prototypical combinations, the results likely generalize well to other samples comprising prototypical combinations. Yet it cannot be assumed that these findings also apply to individuals with low-probability combinations.⁹ Thus, if an object of interest (eg, a biomarker) is mainly associated with prototypical combinations, a high cumulative prevalence of low-probability cases will reduce the power to infer this object. Similarly, an object of interest associated with specific low-probability combinations might be inferred in a specific sample, but replication in a sample with a similar share but different specific low-probability cases might fail. Therefore, research must take the symptom heterogeneity and its structure into account to balance power and replicability.

First, the assumption of categorical differences between symptom combinations does not account for symptom overlap between combinations. We relied on symptom combinations because they have been very prominent in the existing literature on the heterogeneity of mental disorders.^3,17,18,26 Second, the dichotomization of symptoms (present or absent) is an imperfect simplification of their severity. Generally, a symptom combination can be considered a probability density in a multidimensional space with each symptom and its severity being represented by 1 dimension. Dichotomizing symptom ratings limits the number of possible states on a dimension to 2. While this simplifies the complexity of the real-world data, it does not change any of the mechanisms detailed in this study. Third, although this study aimed to outline a mechanism that is inherent to any symptom-based definition of a mental disorder, the investigated symptom combinations were not necessarily congruent with DSM-5-TR criteria, and were collected using self-report questionnaires, and 3 samples comprised veterans, with the same individual potentially included in more than 1 sample. Further research with additional, nonveteran samples assessed for DSM-5-TR symptoms and by trained interviewers is needed to evaluate the generalizability of the observed heterogeneity of specific disorders. Fourth, there are more possible symptom combinations of PTSD than participants in our corresponding sample, and thus larger samples are needed to estimate an empirical probability of all these combinations.

Our findings demonstrate that in a population with a specific, polythetic mental disorder, a few combinations of symptoms have an exceedingly high probability to occur while this probability is markedly lower for most of the possible symptom combinations. This specific structure of symptom heterogeneity is a direct consequence of the polythetic symptom-based definition of mental disorders, with differences in the individual symptoms鈥� prevalence as well as their interrelations resulting in the described differences in the probability of symptom combinations. Finally, consequences of changes of the definition of a disorder, for example, for its prevalence, can be better anticipated by focusing not only on the total number of possible symptom combinations but rather their individual and cumulative probability. These probabilities should also inform upcoming reviews of mental health classification systems.

Accepted for Publication: May 21, 2024.

Published Online: August 7, 2024. doi:10.1001/jamapsychiatry.2024.2047

Correction: This article was corrected on September 25, 2024, to fix an error in the text and missing elements in Supplement 1.

Corresponding Author: Tobias R. Spiller, MD, Department of Psychiatry, Yale School of Medicine, 333 Cedar St, New Haven, CT 06510 (tobias.spiller@yale.edu).

Author Contributions: Dr Spiller had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Drs Spiller and Duek contributed equally to this work as co鈥揻irst authors.

Concept and design: Spiller, Duek, von K盲nel, Harpaz-Rotem.

Acquisition, analysis, or interpretation of data: Spiller, Duek, Helmer, Murray, Fielstein, Pietrzak, Harpaz-Rotem.

Drafting of the manuscript: Spiller, Duek, Harpaz-Rotem.

Critical review of the manuscript for important intellectual content: Duek, Helmer, Murray, Fielstein, Pietrzak, von K盲nel, Harpaz-Rotem.

Statistical analysis: Spiller, Duek.

Obtained funding: Harpaz-Rotem.

Administrative, technical, or material support: Fielstein, Pietrzak, Harpaz-Rotem.

Supervision: von K盲nel, Harpaz-Rotem.

Other 鈥� methodology: Helmer.

Conflict of Interest Disclosures: Dr Duek reported serving on the advisory board of Madrigal Mental Health. Dr Helmer reported grants from Yale during the conduct of the study, being an employee of Manifest Technologies outside the submitted work, and having patents for USPTO (63/533,888 and 63/565,397) pending. Dr Murray reported personal fees from Manifest Technologies for consulting on the topic of computational neuroimaging outside the submitted work. Dr von K盲nel reported honoraria from Heel and Vifor Inc Switzerland outside the submitted work. Dr Harpaz-Rotem reported grants for investigator-initiated research from Boehringer Ingelheim International outside the submitted work. No other disclosures were reported.

Funding/Support: This study received support from the Swiss National Science Foundation, Early Postdoc Mobility Fellowship (P2ZHP3_195191, Dr Spiller) and National Institutes of Health (R01MH112746). This material is the result of work supported with resources from and the use of facilities at the West Haven VA Medical Center.

Role of the Funder/Sponsor: The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.

Disclaimer: The views expressed in this article are solely those of the authors and may not reflect the opinions or views of the National Institutes of Health or of the submitters submitting original data to the National Institute of Mental Health (NIMH) Data Archive (NDA). Furthermore, they do not reflect an endorsement by or the official policy or position of the US Department of Veterans Affairs or the US government.

Data Sharing Statement: See Supplement 2.

Additional Information: Some of the data and/or research tools used in the preparation of this article were obtained from the NDA. NDA is a collaborative informatics system created by the National Institutes of Health to provide a national resource to support and accelerate research in mental health (dataset identifier doi:10.15154/1522637).