An Answer to “So What?” Implications of Network Theory for Research and Practice

Network theory posits that the set of psychological problems traditionally referred to as “mental illnesses” do not exist independently of their signs and symptoms. Instead, the relationship is mereological: whole to part. To illustrate, consider a school of fish. A school of fish does not exist independently from the fish. Whereas an independent disease entity (e.g., a cancerous tumor) can exist without the presence of any symptoms, a school of fish cannot exist without fish. In a similar way, network theory claims that mental disorders arise from a set of components (e.g., thoughts, emotions, behaviors, somatic experiences, and other relevant biological, psychological, or social factors) and the causal relationships among these components. The components of the network include the experiences we commonly refer to as symptoms. In other words, a pattern of pervasive worry is not caused by generalized anxiety disorder, the pattern of pervasive worry (and accompanying features) is generalized anxiety disorder.

This idea is illustrated in Figure 1. In this figure, there is no independent panic disorder node that causes panic attacks, avoidance, and worry. Instead, panic disorder is a state that the network can be in: a state characterized by the presence of these specific experiences. The stable, concurrent presence of panic attacks, avoidance, and worry is panic disorder. In other words, network theory provides an explanation for the observation that some symptoms tend to cohere as syndromes that obviates the need for an independent underlying cause. In doing so, it challenges the assumption that mental illnesses must necessarily exist independently of their signs and symptoms.

Notably, when taken literally, the term “symptom” itself is incompatible with network theory. The word “symptom” denotes a surface-level indicator of an underlying cause. Network theory asserts that most things that have traditionally been called symptoms (e.g., worry) are not indicators of underlying diseases, but are nodes in a causal web. This idea—that many “symptoms” are in fact causal players—is a key concept in network theory. Moreover, some things that have not traditionally been called symptoms (e.g., the belief that bodily sensations associated with arousal are dangerous) fit nicely as nodes within these networks (e.g., panic disorder) (13–14). Therefore, we prefer the terms “nodes” or “elements” rather than “symptoms” when speaking of the constituent parts of networks, because these terms clarify that these parts can be understood not as passive indicators, but rather as active causal components that constitute the mental disorder.

Many medical conditions can be summarized by a classification system that simultaneously describes both cause and consequence. When a certain medical symptom pattern arises (dizziness, shakiness, sweating), physicians are tasked with providing a differential diagnosis to determine which disease entity has led to the symptoms (influenza or hypoglycemia). By identifying the presence of the disease entity (which, together with the symptoms, constitutes a cause-consequence pair), the physician can typically rule out the presence of other disease entities. Such classification systems can be used to great effect when specific causes (i.e., diseases) reliably lead to the same general consequences (i.e., symptoms), when some consequences are unique to a given cause (e.g., pathognomonic symptoms, such as Koplik spots), when the different causes are mostly independent of one another, and when the consequences do not feed back into the causes (e.g., dizziness from hypoglycemia does not cause the flu). We will here refer to such cause-consequence classification systems as a medical model of classification, although we note that this phrase is fraught, defined in different ways by different sources, and an oversimplification of the actual practice of modern medicine.

The network perspective challenges the assumption that classification should follow this simple medical model. Most fundamentally, as we argued in the previous section, it does so by demonstrating that there need not be an independent underlying cause that gives rise to symptoms and, thus, there need not be an underlying disorder to which the symptoms can be attributed. However, there are other characteristics of psychopathology that further argue against a simple cause-consequence classification system. First, the boundaries between our diagnostic categories are fuzzy, with many symptoms appearing across multiple disorders. Second, equifinality (a given end state can be reached by many different starting points) and equipotentiality (many given end states can be reached from the same starting point) tend to be the rule, rather than the exception, in psychiatry. A given component in psychiatry (e.g., impaired sleep) can have diverse consequences (e.g., poor concentration, mania), and a given consequence may have many potential causes (e.g., poor concentration may arise not only from impaired sleep, but also from anhedonia, preoccupying worry, or intrusive memories of trauma). Third, symptoms in psychiatry are often causal agents with their own set of effects. In an idealized medical model, observable symptoms are the endpoint of the causal system: dizziness arising from hypoglycemia does not cause the flu, allowing diagnosticians to confidently rule out the flu if they attribute dizziness to hypoglycemia. In practice, things are perhaps not so simple (e.g., dizziness may not cause the flu, but it may lead to nausea, falls, or other health consequences), but this general framework holds well enough to make informative diagnostic decisions.

In psychiatry, the complex web of causality is even more dense. For example, anxiety about social situations can (and frequently does) result in problematic patterns of alcohol use, which may give rise to feelings of worthlessness, loneliness, and depression. Similarly, intrusive memories following a trauma may evoke physiological reactivity and, in turn, fatigue, social disconnection, anhedonia, and irritability. In other words, symptoms are not the end of the causal story; they frequently result in consequences that do not respect diagnostic boundaries. Together, these commonly observed features of psychopathology present significant obstacles for a medical model of classification.

Viewing psychopathology from a network lens provides a new perspective on these obstacles. As depicted in Figure 2, the fuzziness among diagnostic categories arises because mental disorders are overlapping communities of causally interacting components, not discrete disease entities. For example, posttraumatic stress disorder symptoms overlap considerably with the symptoms of depression, a substantial challenge for diagnosticians working from a medical model, but an expected state of affairs when viewed from a network perspective, where individual symptoms can have a range of causal effects. Even in the absence of syndromic overlap, syndromes may bridge to each other if the symptoms of one disorder cause symptoms of another (e.g., persistent social anxiety may lead to problematic alcohol use). In addition, network theory allows that similar components may have divergent effects, thereby producing considerable heterogeneity across individuals in symptom presentation.

To be clear, our argument is not that classifications are useless in the context of network theory or that these classifications should be avoided altogether. Even rough, messy, and imprecise classification systems can be useful for quickly summarizing and communicating information in a lingua franca (15). For example, knowing that a patient has a diagnosis of panic disorder provides us information about potential causal relationships among panic attacks, worry, and avoidance that may produce this pattern of symptoms. However, it is important that the diagnoses in our classification systems are treated as rough shorthand and not as established independent disease entities.

In situations where classification is complex, it is often beneficial to have multiple systems of classification depending on one's goals (e.g., one classification system for treatment and another for etiology) (16). Indeed, multiple classification systems have been used productively in biology, and such a system could readily apply to clinical psychology (16, 17). As an example, consider that treatments for depression show different efficacy for different people, and this does not seem closely linked to the depression diagnosis as defined by its DSM signs and symptoms. Indeed, the treatments with the best evidence for depression (antidepressant medications and cognitive-behavioral therapy) are not even specific to depression at all (18–20). In developing a treatment-focused classification system, other variables might be more important than the same symptom patterns we have relied on for years. For instance, one study (21) has found that marital status (being married or cohabiting) predicts an advantage of cognitive-behavioral therapy over antidepressant medication among individuals with depression. In a treatment-focused classification system, marital status might be an important variable to consider (perhaps more important than the specifics of the symptom patterns). Classification systems could also rely on the within-person temporal dynamics of variables and relationships between variables, rather than just on the presence or absence of the variables (22). In sum, there may be advantages in departing from psychiatry’s traditional view of classification.

Perhaps most controversially, the network theory challenges the idea that psychological problems are caused by illnesses in the brain. Borsboom, Cramer, and Kalis (23) have presented three arguments for why a network view is incompatible with a reductive explanation of mental illnesses as brain disorders. Their first argument juxtaposes the “many causes” network model with a “common cause” brain disorder. If all symptoms are simultaneously dependent on a common latent variable, then one can accurately describe the system in terms of a common cause (in the brain). However, if a network model is correct, and there are many interacting causes, this type of reduction is blocked. Second, they argue that it is implausible that neurobiological causes could be a simultaneous explanation for all symptoms, because many symptoms are rational in the context of the specific content of other symptoms (e.g., hand washing is rational when one believes one has been contaminated). The specific content of thought is therefore necessary to explain the full symptom pattern. The fact that many symptoms are “intentional” (i.e., about something) means that there are an infinite variety of possible symptom manifestations. A small difference in thought content could lead to radical differences in behaviors or other symptoms. Third, they argue that specific connections in symptom networks rely on cultural and historical contexts, again blocking the idea that mental illnesses can be reduced to a single common cause in the brain invariant to context.

Critically, this does not mean that psychological problems cannot result from insults to or impairments in the brain—history is replete with such examples, from syphilitic insanity to postconcussive syndrome. Rather, it provides a plausible alternative to this model that may apply to some specific types of problems. Equipped with this alternative account, we can begin to ask which type of theoretical framework offers the best explanation for a particular syndrome. Given the lack of progress in identifying reliable brain markers of most mental problems over the past century, we believe there is good reason for us to devote greater energy to alternative accounts, such as network theory.

We would again emphasize that our call to further investigate alternative accounts need not be an inflammatory position—network theory does not assert that structural or functional abnormalities in the brain are irrelevant. Within the framework of network theory, brain processes can be viewed as a separate level of analysis that can help explain the connections among nodes (6). Alternatively, these processes could be directly incorporated as components that play an important role within the network, providing a biopsychosocial network that accounts for the complex interplay among components across separate levels of analysis (13). Ultimately, we suspect that an integration of network theory with research on the genetic and neurobiological factors that contribute to psychopathology will be the most fruitful path toward meaningful progress in psychiatric research and practice.

Classifications of disorders are unlikely to be precise, and diagnostic labels should not be reified and treated as established disease entities (24). Because the network theory challenges the view that disorders are discrete entities, future network research should work toward elucidating more flexible systems of classifying mental problems. The presence and absence of specific symptoms are useful for creating classifications of mental problems, but other factors (including network properties, such as the relationships between symptoms) may also be useful.

Analyses that focus on disorders (e.g., the joint occurrence of major depressive disorder and social anxiety disorder) rather than on individual components of disorders (e.g., the joint occurrence of depressed mood and social avoidance behavior) are useful descriptively, but they are unlikely to produce substantial progress in our understanding of psychopathology on their own, because they entirely overlook the important component-level associations posited to give rise to these disorders and their joint occurrence. Analyses that focus on individual psychological problems (e.g., social avoidance) and the relationships among them (e.g., the effect of social avoidance on social fear) are likely to provide critical new information, helping us to understand not only how these problems interact as a causal system, but also how other genetic, neurobiological, or social components may affect this system (25).

A focus on symptoms typically implies a focus on the specific symptoms encoded in the DSM. Yet DSM symptoms do not appear to have a privileged role in networks (26, 27). Symptoms are typically identified as such because they are highly distressing to the patient, highly visible, or because they seem especially tied to a specific syndrome (i.e., hallmark symptoms). But there is no reason to assume that the causal feedback loops generating mental disorders will exclusively contain experiences with these qualities. Accordingly, although symptoms may be a tractable starting point for considering which components may be present in the network, it will likely be important for networks to expand to encompass more than just the symptoms enumerated in diagnostic manuals (28, 29). Other well-researched factors in psychopathology (e.g., anxiety sensitivity) (30) can also be incorporated into network models. In addition, more complex models, such as hybrid networks, should be used to incorporate variables that are causally relevant but not necessarily part of the feedback loops that maintain a disorder over time (e.g., genetic factors or trauma) (28).

Ultimately, empirical network research aims to advance our understanding of precisely how a given disorder may operate as a causal system. That is, it aims to facilitate well-developed theories that can represent the system that gives rise to a disorder (e.g., a vicious cycle between arousal-related bodily sensations and perceived threat that gives rise to panic attacks). Recently, some (31) have argued that the causal structure of even the relatively simple systems thought to give rise to many anxiety and stress-related disorders will be difficult to estimate directly from empirical data. Instead, these researchers (32) have proposed that it may be necessary to formalize initial theories about how a disorder may operate as a causal system by expressing those theories as mathematical or computational models. Those models can be used to critically examine what the theory can explain and what it cannot, including the patterns of association identified in empirical network studies. In doing so, these initial formal theories can identify a theory’s shortcomings, and thus opportunities to improve the explanatory breadth or precision of the theory, and move us toward greater understanding of the system that gives rise to the disorder.

Some common statistical models (i.e., [reflective] factor analysis) assume that a single underlying latent construct (e.g., a mental disorder) causes the emergence of diverse psychological problems. This assumption is broadly at odds with network theory (28, 33), which emphasizes the interactions among components. To best inform network theories, it is likely to be more fruitful to focus on statistical models that aim to examine the interrelationships among components. Although network psychometric models may serve this purpose in some cases, we would more fundamentally argue that researchers should carefully select the statistical models that will best inform their theories. In our experience, it is often the practice of researchers in psychiatry to focus on statistical models that are popular, well-accepted, or familiar to the researcher. However, this practice can result in choosing statistical models ill-suited to addressing the research questions of interest. Statistical tests and their interpretations should be justified by the theoretical aims of the researchers. Researchers should specify their aims and then choose statistical models accordingly; they should not choose statistical models and then specify their aims accordingly.

Most empirical research in the network literature has focused on cross-sectional network analyses that use regularized, partial correlations between pairs of nodes (7). Within-person analyses have also been conducted, typically by calculating lagged partial correlations, lagged regression parameters, or similar statistical parameters (7). Both types of analyses can provide valuable information about interrelationships among components. However, both are also extremely limited in their ability to inform our understanding of causal interactions among components (10).

Causal interactions among components are at the heart of network theory. Thus, it is essential that network researchers develop and integrate randomized experiments into their investigations, because such experiments provide insights into causality that cannot be matched by cross-sectional or within-person observational studies. Such experiments are not the only tool for causal inference, but we believe they have been especially underutilized in the network approach literature.

Here again, we would stress that diagnosis does not reveal an underlying disease that exists within a patient. Classifying patient conditions into subgroups (e.g., as social anxiety or obsessive-compulsive disorder [OCD]) may be useful for descriptive and communicative purposes, but these classifications should be interpreted as rough descriptions of common patterns, not as reified disease entities. The value of these descriptions through the lens of network theory is that they may point us toward the plausible systems driving maintenance of these syndromes. For example, knowing that an individual has received a diagnosis of OCD suggests a causal feedback loop between obsessions and compulsions that leads these symptoms to persist together over time.

Therapists should continue to use functional analysis and case conceptualization (i.e., process-based therapy) (34–36) throughout the course of treatment, attempting to understand the relationships between patients' thoughts, behaviors, and emotions as a fundamental approach to psychotherapy.

One productive approach may be to focus on a set of targeted symptoms jointly defined by the clinician and patient. This clinical approach fits with the idea of self-reinforcing feedback loops. By thwarting the activation of specific nodes in feedback loops, problematic patterns can be efficiently disrupted. Choosing the ideal symptoms to target is not a trivial task and differs for individual patients. In choosing symptoms to target, clinicians should carefully consider which symptoms are present, directly malleable, causally linked to other downstream symptoms (considering both generalized scientific evidence and personalized case conceptualization), and relevant to the patient’s goals. To measure the success of therapy, clinicians and patients should jointly track progress on specific goals.

A physician who establishes hypoglycemia as the cause of dizziness is probably justified in ruling out the presence of other potential causes, such as influenza. This is because these conditions follow a traditional disease-symptom pattern with minimal causal influence from the symptom (i.e., dizziness) to other conditions (e.g., the flu). Network theory predicts the opposite. From the perspective of network theory, one should not rule out one diagnosis on the basis of the presence of another (e.g., ruling out OCD because generalized anxiety is present and more prominent). Rather, network theory predicts that the presence of one problematic pattern of nodes increases the likelihood of the presence of another problematic pattern. That is, the presence of one disorder should lead us to expect that other disorders may also be present.