Abstract
Experimental psychology has long embodied the quest to identify the causes of psychopathology. This venerable tradition has been joined in this quest by network theory, a novel approach to conceptualizing episodes of disorder as emerging from complex systems characterized by dynamic interactions of symptoms. Although issuing from the correlational, psychometric tradition rather than the experimental one, it nevertheless offers methods for identifying symptom targets for clinical experimental intervention. The purpose of this article is to sketch the points of contact between network psychometrics and experimental psychopathology.
Keywords
The network perspective conceptualizes mental disorders as phenomena that emerge from the causal interactions of their constitutive symptomatic elements (e.g., Borsboom, 2017; Borsboom & Cramer, 2013). This perspective differs drastically from the traditional categorical approach that views a mental disorder as a discrete disease entity causing the emergence and covariance of the signs and symptoms presumptively reflective of its underlying presence (e.g., Klerman, 1978; cf. Borsboom, 2008). This traditional view, implicitly embodied in the Diagnostic and Statistical Manual of Mental Disorders (DSM-5; American Psychiatric Association [APA], 2013), likens syndromes such as major depression, bulimia nervosa, and so forth to conditions such as strep throat (McNally, 2021). Hence, consider a physician assessing a patient who complains of pain upon swallowing, fever, fatigue, and who has white patches at the back of his throat. She will suspect that the common cause of his symptoms is Streptococcus bacteria. Prior to a confirmatory throat culture, this conjectured cause is latent—unobserved, but not unobservable. But once the culture confirms its presence, the physician will prescribe an antibiotic to abolish the causal source of symptoms, not merely prescribe aspirin for fever, bed rest for fatigue, and an anesthetic spray for throat pain. The antibiotic strategy works because once one conditionalizes on the presence of strep bacteria, the covariance among symptoms vanishes. That is, the symptoms are locally independent; they are not causing one another. They cluster because each is an independent effect of a common cause. Moreover, the hypothesized latent cause has an existential referent—strep bacteria—that is potentially observable and distinct from the symptomatic effects it produces.
These conditions seldom hold for most psychiatric disorders (Borsboom, 2008). Depression is not a latent, potentially observable cause of symptoms, and symptoms are not locally independent. Indeed, direct causal relations abound in psychopathology. Insomnia causes fatigue, obsessions trigger compulsions, auditory hallucinations motivate delusions, and so forth. In fact, Down syndrome is a rare exception where an originally latent, but potentially observable, entity—trisomy 21—is the common cause of locally independent signs and symptoms. Unlike the cause of Down syndrome, depression does not have an existential referent akin to trisomy 21. Rather, depression denotes an emergent phenomenon issuing from causal interactions among its symptomatic elements. The disorder is not the cause of the symptoms as in the case of strep throat. Rather, the relation between symptoms and disorder is mereological, or parts to whole (Borsboom, 2008).
Bulimia nervosa exemplifies how causal relations among features of the disorder may interact, resulting in the emergence of a disorder. Consider a young woman who acquires beliefs about desirable standards of body shape and size from her ambient culture. If she deems that she falls short of such standards—the “thin ideal”—she may take certain steps to rectify matters. She may restrain her eating, becoming carefully attuned to the caloric content of her daily food intake. She may vow to avoid otherwise highly desirable, palatable items whose consumption may interfere with her goal of thinness. Unfortunately, drastically reducing intake renders her chronically hungry. Sooner or later, she yields to her hunger and rapidly consumes a large quantity of “forbidden” food. After doing so, she experiences guilt, shame, and anxiety about what she has just done. To relieve her distress about her binge, she self-induces vomiting to eliminate the calories she has just consumed.
Note that the signs, symptoms, and associated features of the diagnosis of bulimia nervosa are best construed as a complex, dynamic network of interacting elements rather than the causally unrelated consequences of an underlying common cause akin to Streptococcus bacteria (e.g., Meier et al., 2020). The strength of the relations between elements in this complex system can vary. For example, highly restrictive eating can be especially strongly related to hunger and thus the motivation to eat, whereas endorsing a desire to be thin need not be especially strongly related to restrained eating. Indeed, many obese people may express a desire to be thin but never take sufficient steps toward losing weight.
The network perspective holds that the strength of association between elements in such a system may determine whether the emergence of one or more symptoms may activate other elements in the network, thereby triggering an episode of disorder. Hysteresis denotes a stable phase transition whereby a person shifts from a healthy state to a persistent, pathological one (Cramer et al., 2016). When hysteresis occurs, symptomatic elements tend to sustain one another, thereby qualifying a person for a diagnosis when impairment in everyday life results.
Cross-Sectional Networks
Networks comprise nodes (circles) and edges (lines) that signify an association between two nodes. In the psychopathology field, nodes usually represent symptoms and edges signify an association between two nodes. Edge thickness represents the magnitude of association, interpretable as the likelihood of co-activation.
Most research in network psychopathology has involved cross-sectional studies that involve a “snapshot” of the relations among symptoms in a large number of individuals at a single point in time, such as people experiencing symptoms of posttraumatic stress disorder (PTSD) following exposure to a serious stressor (e.g., McNally et al., 2015). Standardized questionnaires and structured diagnostic interviews usually provide measures of symptom severity and frequency and thus the input for network analyses. Epskamp et al.’s (2012) R package, qgraph, has been the indispensable tool for visualizing psychopathology in the emerging field of network psychometrics (Epskamp, Maris, et al., 2018). Although a correlation between two symptoms does not signify a causal relation, it represents a first step in an effort to ascertain causality.
A second step is to compute a network of regularized partial correlations (Epskamp & Fried, 2018). This procedure accomplishes two things. First, it adjusts for the influence of all other nodes in the network, thereby testing whether an association between two symptoms is direct or merely a spurious association induced by the influence of other symptoms in the network. Second, regularizing a partial correlation network eliminates edges of small magnitude, retaining the largest ones most likely to be genuine.
Although a regularized partial correlation between two symptoms signifies a direct, nonspurious association between them, one cannot ascertain whether one symptom is activating the other, or vice versa, or whether bidirectional influences occur. That is, these undirected networks lack arrowheads that point in the direction of prediction and possible causation.
However, additional information can sometimes furnish disambiguating clues to directional influence. For example, a regularized partial correlation between insomnia and fatigue is more likely attributable to sleep deprivation causing fatigue than vice versa. In other cases, the direction of influence may work both ways. For example, hypervigilance and exaggerated startle are commonly associated in PTSD symptom networks, and it is likely that people hypervigilant for threat will be prone to startle easily and having been startled may render one hypervigilant for threat (e.g., McNally et al., 2017).
Temporal (Time Series) Networks
Cross-sectional networks involve one data point per node for many participants; time-series networks involve many data points for participants collected over time (Epskamp, Waldorp, et al., 2018). These studies often involve pinging participants on their smartphones and having them rate moods or symptoms at regularly scheduled intervals throughout the day on slider scales ranging from 0 to 100. Resultant directed networks can reveal whether a symptom at time 1 predicts another symptom at time 2 as well as visualize undirected associations between pairs of symptoms within a temporal window (i.e., a contemporaneous network, e.g., Aalbers et al., 2019). Edges in a temporal network have an arrowhead at one end, signifying the direction of prediction and potentially causation.
Temporal networks bring us closer still to a causal model as they disclose directional predictive relations between symptoms at one time point and symptoms at the next one. In fact, one can compute a temporal network in time series data points from a single person (Epskamp, van Borkulo, et al., 2018). These data, in turn, may furnish insights on what symptoms should be targeted for therapeutic intervention for specific patients and may also be useful for tracking progress during a course of therapy.
Bayesian Networks
Bayesian network analysis provides the most ambitious framework for modeling a causal system from cross-sectional data (Pearl, 2011; Pearl & Mackenzie, 2018; Scutari, 2010)
Clinical investigators have computed two types of DAGs from symptom datasets (e.g., McNally, Mair, et al., 2017). Both DAG types are structurally identical except that edge thickness can vary because thickness indicates a different attribute in each DAG type. In one type, edge thickness signifies the importance of an edge to model fit. The thicker a connection between two symptoms, the more important it is.
In the other type of DAG, edge thickness represents the level of confidence (probability) that the direction of prediction depicted in the graph is correct. For example, a DAG is more likely to have a thick directed edge issuing from the node representing depressed mood and incident on the node representing suicidal ideation than vice versa. A thin edge in a direction-of-prediction DAG means that the arrowhead was often pointing in the opposite direction for a large minority of the bootstrapped graphs computed during the Bayesian iterative process.
DAGs can model causal relations among variables if two assumptions are met. First, all important variables figure in the computation of a network. This is a tall order to fill, but this assumption is not confined to Bayesian network analysis; it holds for causal modeling in any science.
Second, as its name implies, a DAG prohibits cycles among nodes, such as bidirectional influence (i.e., symptom X > symptom Y, and symptom Y > symptom X). However, a thin edge between two nodes in a direction-of-prediction graph suggests such a possibility. Moreover, DAGs also prohibit looping cycles (i.e., symptom X > symptom Y > symptom Z > symptom X). For example, this limitation renders it difficult to model Clark’s (1986) model of panic (i.e., detection of certain bodily sensations > catastrophic misinterpretation > increased fear > increased bodily sensations). The inability to model cyclic causal relations is a notable limitation, but as Garcia-Velazquez and her colleagues observed, “Cyclic relations are inherently difficult to analyze, and thus it seems wise to start with robust findings on directional dependence and only then build toward more complex models” (García-Velázquez et al., 2020, p. 247).
Cross-sectional networks and Bayesian directed acyclic graphs provide complementary insights just as the mean and the median provide two distinct measures of central tendency for a distribution. The former can accommodate direct, bidirectional associations between symptoms, such as those between exaggerated startle and hypervigilance in PTSD, but do not disclose the direction of prediction and potentially causation. The latter point to directional relations of conditional dependence and independence but prohibit cycles. High-centrality symptoms in the former and parent symptoms in the latter suggest targets for intervention by clinical experimental psychopathologists, and this is especially true when both types of networks identify the same symptom targets (e.g., McNally et al., 2022). And temporal networks provide yet further clues to causation.
Points of Contact Between the Two Traditions
Many years ago, Cronbach (1957) bemoaned that practitioners of the two disciplines of psychology—the experimental and the correlational—seldom had much to do with one another. He noted how experimentalists regarded individual differences among subjects exposed to the same experimental treatment as undesirable error variance, whereas the correlationists regarded such variance as their stock-in-trade. He hoped that someday psychologists would probe the interactions of experimental treatments and individual difference variables, thereby unifying the two disciplines.
Experimental psychopathology embodies the realization of Cronbach’s hope. Subjects who vary on risk factors for mental disorders or who do or do not meet criteria for a disorder are exposed to different experimental treatments and their responses recorded. For example, a vast body of research on cognitive biases in people with anxiety and mood disorders has illuminated potential mechanisms implicated in the etiology and maintenance of these conditions (e.g., Williams et al., 1997).
Although network psychometrics emerged within the correlational wing of psychology, it shares causal aspirations with experimental psychopathology. Both aim to identify the causes of mental illness as the road to reducing suffering. Causal accounts can be formulated at different levels of analysis. Experimental psychopathologists often sought to identify mechanisms producing symptoms (e.g., attentional bias for threat increasing anxiety), whereas network analysts have often sought to identify causal relations among symptoms themselves whereby the occurrence of one symptom (e.g., fatigue) may activate another (e.g., difficulty paying attention). And each symptom itself may comprise a network of elements exemplified by the neural circuitry producing exaggerated startle responses elucidated by experimentalists. Despite the emphasis of network analysts on observational, not experimental, studies, these investigators have sought to identify the most important symptoms in psychopathology networks that figure in the emergence of episodes of disorder.
The network analytic approach to psychopathology regards symptoms high on certain measures of centrality as especially important and potentially relevant to a causal account. There are a variety of centrality metrics that measure a node’s importance in a network (Freeman, 1978–1979), but many (e.g., node closeness centrality, node betweenness centrality) are less relevant to psychopathology than to political science, sociology, epidemiology of infectious disease, and other fields where each node represents a person, and the centrality metric measures relations among individual persons. In our field, each node typically signifies a symptom, and the centrality metric represents how interconnected the symptom is with other symptoms within a person.
The two most relevant metrics are node strength centrality (Bringmann et al., 2019) and node expected influence centrality (Robinaugh et al., 2016). One computes a symptom’s strength centrality by summing the absolute values of the magnitudes of the edges connected to symptom (e.g., partial correlations). One computes a symptom’s expected influence centrality by considering the sign of the edge weights before summing them. Node strength centrality and node expected influence centrality are identical when all associations in the network are positively correlated, but the latter is the more accurate metric when there is at least one edge connecting a pair of symptoms that signifies a negative association (i.e., the activation of one symptom is associated with diminished activation in the other).
These centrality metrics suggest the possibility that a symptom highly interconnected with other symptoms may be key to the emergence and maintenance of an episode of psychopathology. Accordingly, if such a high centrality symptom is the source of activation for other symptoms, then it might be a prime target for therapeutic intervention. If so, then “deactivating” it might very well produce a beneficial cascade by diminishing the severity of symptoms downstream from it, hastening the patient’s recovery.
The therapeutic promise of centrality metrics sparked enthusiasm, especially among behavioral and cognitive-behavioral therapists who have long focused on functional relations among symptoms (See Davison, 2019; e.g., fear > avoidance, obsessions > compulsions, binging > purging) integral to psychological disorders (McNally, 2016). Network psychometrics formalized and quantified these relations, hence integrating network theory and the statistical methods of network analysis.
However, several requirements must first be satisfied to determine the causal importance of a central node. First, a high-centrality symptom must be the source of activation for other symptoms, not its recipient. In a cross-sectional network, a symptom having many undirected edges (especially of high magnitude) connected to it would be an ideal target for intervention only if it is the source of activation.
Second, although a time-series network has directed edges signifying the direction of prediction, a predictive relation between a symptom at time 1 and activation of another symptom at time 2 need not confirm a causal relation even though it is consistent with that interpretation.
Third, successful intervention to deactivate a symptom having high centrality presupposes that we have efficacious treatments available. Consider PTSD. We have methods for reducing psychophysiological reactivity to reminders of one’s trauma, whereas specific treatments for emotional numbing are underdeveloped.
Fourth, ideally, an intervention would diminish the activation of a single symptom in the network, and then subsequently one would observe alleviation of symptoms “downstream” from the target symptom. But it can be difficult to intervene only on a single symptom, akin to deactivating a single gene in a gene knockout procedure. Often the best we can do is to ensure that our intervention deactivates the target symptom regardless of whatever other symptoms may also be deactivated. For example, a clinician might prescribe a benzodiazepine for sleep onset insomnia—an intervention also likely to simultaneously reduce anxiety. Similar reasoning holds for “parent” symptoms in Bayesian networks.
A point of contact between network psychometrics and experimental psychopathology would be to identify high-centrality symptoms and parent symptoms and then randomly assign patients to conditions whereby these symptoms are targets for initial intervention or not. For example, if sleep disturbance rather than persistent sad mood turns out to be central in patients with major depressive disorder, then patients randomly assigned to have their sleep problems targeted first would recover faster than patients randomly assigned to have their persistent mood targeted first.
Another point of contact is to broaden the kinds of nodes beyond diagnostic symptoms to include variables identified in research on experimental psychopathology (e.g., relevant to cognitive biases for threat; Williams et al., 1997). Network analyses incorporating data from the laboratory within networks themselves (Jones et al., 2017) has only just begun (e.g., Bernstein et al., 2017; Heeren & McNally, 2016).
Conclusions
Despite their originating in each of the two distinct disciplines once causing dismay in Cronbach, the waning taboo against rigorous causal reasoning and analyses in observational studies (e.g., Grosz et al., 2020; Pearl & Mackenzie, 2018) bodes well for the fruitful interchange between network psychometrics and the venerable tradition of experimental psychopathology. The field of network psychometrics is developing very rapidly (Borsboom et al., 2021, Borsboom, Deserno et al., 2022) including its interface with clinical practice (Bringmann et al., 2022). Considerable empirical advances have occurred (for a review, see Robinaugh et al., 2020).
The frontier may turn out to be rigorous computational modeling—a qualitative advance in network analysis (Robinaugh et al., 2019). Indeed, just as computational sciences have devised models of climate change and perhaps less successfully economics, clinicians, computational modelers, and psychopathologists will likely work together to model how interventions may affect the complex systems of psychopathology (Borsboom, Haslbeck et al., 2022).
Footnotes
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
