Latent Trait,Latent-Trait State,and a Network Approach to Mental Problems and Their Mechanisms of Change

Empirical illustration of the latent-trait approach

In one of the studies reviewed by Smits et al., Foa et al. (1996) investigated social cost as a mechanism of change in cognitive behavioral therapy (CBT) for SAD, measuring SAD symptoms with the Social Phobia subscale of the Social Phobia and Anxiety Inventory and social cost with the Social Cost Questionnaire. To test mediation, Social Phobia subscale total scores at posttreatment were regressed on (a) the residual-gain scores of Social Cost Questionnaire total scores from pretreatment to posttreatment and (b) pretreatment social-phobia subscale scores. Consistent with the hypothesis that latent cost causes latent SAD, the correlation was significant. In theory, this finding supports the proposition that the trait social cost is related to SAD. With the reservation that temporal precedence was not certain in this study, Foa et al. drew the possible clinical implication that reducing a patient’s trait level of social cost will lead to improvement of SAD.

Summary of the latent-trait explanation of mental problems

The latent-trait approach explains the stabile, multiple, and syndromal character of observable mental problems—in this perspective called symptoms—as effects of underlying trait-like disorders. The problems’ mechanisms are also conceptualized as latent traits.

Critical examination of the latent-trait approach

In the latent-trait approach, persons’ stable or usual levels of symptoms and mechanisms are addressed, whereas within-persons deviations from these levels are ignored. Consequently, only between-persons relations between mechanisms and outcome are examined and not their within-persons relations. We argue that therapists and therapy theories focus on within-persons relations between deviations from the usual on these variables and that inferences from between-persons relations obtained in empirical studies to within-persons relations in theory and therapy are seldom warranted.

A trait is considered to be constant for a certain period of time (Borsboom, Mellenbergh, & van Heerden, 2003). Any within-persons variation in measures of its indicators (e.g., symptoms) during this period is discarded as measurement error. Thus, the latent-trait approach concerns only between-persons differences and relations. That is, they concern where a person is placed on a latent dimension compared with other persons (Borsboom et al., 2003) and whether those who have higher scores compared with others on one latent dimension (e.g., SAD) also have higher scores than others on another latent dimension (e.g., social cost). Thus, in their analysis of mechanisms, Foa et al. (1996) examined the extent to which those patients who showed greater decrease in social-cost cognitions differed in SAD change from those who showed less decrease. In the measures used in such studies, subjects are explicitly or implicitly asked to report their usual, average level of symptoms and mechanism indicators for the period assessed (e.g., a period before and one toward the end of treatment). So it is the differences between patients in their average or usual level of symptoms that are the focus of the analyses.

However, in their direct work with a particular patient, therapists are not primarily interested in whether and how the patient differs in his or her usual level from other patients. Of more importance is whether and how the patient at a particular point in time deviates from his or her own usual or expected level. In particular, therapists are interested in how deviations in what they think are mechanistic variables relate to deviations in outcome variables. This interest is reflected in comments and questions such as the following: “I see from your completed questionnaires that you now believe less in your catastrophic cognitions than you use to. What has happened to your anxiety?” (Clark & Salkovskis, 2009, p. 21); “Now that you have got more insight into the conflict between your wish for affirmation and your fear of humiliation, what has happened to your symptoms?” (Leichsenring, Beutel, & Leibing, 2007, p. 60). Questioning may also be directed the other way around, from symptoms to mechanism. A therapist might say, “It seems that you become more restless now as you talk about your recent panic attack. What goes through your mind?” The usual or expected can also be a stable trend; in such a case, a therapist might say, “I gather from your completed forms that your steady reduction of symptoms has stopped. Has anything else happened in your life or to your thinking?”

Moreover, a therapeutic focus on within-persons relationships is shown by therapists promoting patients’ deviations from their usual behavior to explore or demonstrate mechanistic relationships. For instance, in the manipulation of self-focused attention and safety-behaviors exercise in cognitive therapy for SAD (Clark & Wells, 1995), patients are asked to role play a feared interaction under two conditions: one in which they focus attention on monitoring themselves and increase their use of their normal safety behaviors (e.g., avoid eye contact) and another in which they drop their safety behaviors and focus their attention on the other person or persons in the interaction and on what is being said. Although self-focus and the use of safety behaviors are intended to help them, patients almost invariably experience that these activities are associated with more anxiety and thoughts that their social fears (e.g., trembling) are more likely to have occurred. After this exercise, patients may be given the homework assignment of varying their attention and safety behaviors in everyday situations. For instance, a patient may be asked to one day do as usual, to monitor himself or herself (self-focus) and censor what he or she says during coffee breaks at work, and the next day, instead, to say whatever comes to mind and watch others’ reactions (external focus). If he or she again experiences that anxiety follows focus of attention, the patient is more likely to also try to be more externally focused in situations outside those planned in therapy. Thus, based on deviations from the usual, a change process is instigated: Self-focus and anxiety are reduced in an increasing number of situations. This process is reflected in continuously changing average levels of self-focus and anxiety (e.g., from week to week), as well as in momentary deviations from these average levels during particular social episodes. These episodic deviations are likely to be related to transient social (e.g., whether authority figures are present), psychological (e.g., mental fatigue), and/or biological (e.g., tiredness) variables (Hamaker, Nesselroade, & Molenaar, 2007) and may be used to further explore the relationship between these variables, “self-focus” and “anxiety.”

Thus, although therapists’ and patients’ goals are to bring about change in patients’ usual levels of symptoms or problems and other wanted outcomes, the route to these overall changes goes through momentary changes: The patients need to react differently from how they usually do in the problem-relevant episodic cluster. Successful (problem-reducing) deviations from the usual are strengthened and likely to be repeated in future situations. Thus, with time, some deviations become the usual. Therefore, therapists look for and try to facilitate within-persons deviations from the usual in variables of the episodic cluster surrounding a mental problem; more specifically, in those variables that the therapists think are maintaining the patients’ problems and therefore operate as mechanisms of change. In addition, they tend to monitor whether within-persons deviations in a mechanistic variable (e.g., self-focus) is followed by within-persons deviations in the problems (e.g., social anxiety).

Finally, therapists check whether these within-persons deviations—if they are repeated—are followed by change in the usual level of the problem and whether this change generalizes to overall change in the usual levels of the other problems (symptoms). Thus, to attain overall changes in a patient’s usual or trait level of desired outcomes, therapists focus and intervene on within-persons relations between mechanism and outcome variables in episodic clusters. Therefore, therapy theories also focus on such within-persons relationships. For instance, cognitive models of SAD (e.g., Clark & Wells, 1995) propose that a patient will feel less socially anxious at times when he or she appraises negative social events as less costly than when he or she has higher cost appraisals. That the patient’s outcome level will improve by regular lessening of social-cost appraisals simply follows from this proposition, as explained above.

Although the within-persons relationship between a mechanism and a problem variable contributes to the usual levels on both variables and thus to their between-persons relationship, inferences the other way around are seldom warranted. Inferences from between- to within-persons relations are justified only under the strict statistical conditions of ergodicity (Hamaker, 2012; Molenaar, 2004). First, each subject in the population has to obey the same statistical model (homogeneity of the population). That is, the main features of a statistical model describing the data (e.g., means, variances, covariances, and lagged covariances) are stable across subjects. Second, these features must be stable over time. Given these conditions, between- and within-persons effects should be expected to be different (for examples, see Curran & Bauer, 2011; Hamaker, 2012; Hoffmann & Stawski, 2009; Wang & Maxwell, 2015). Thus, if one patient shows a large reduction in his or her usual level of social-cost appraisals and a large reduction in his or her SAD symptoms and another patient shows no cost reduction and no symptom change, this does not imply that either person will be less socially anxious when appraising less cost. Therefore, Foa et al.’s (1996) use of between-persons data to examine theories that focus on within-persons processes does not represent a direct and valid test of such theories (Curran & Bauer, 2011).

In conclusion, because therapy models and therapists focus on within-persons relations in episodic clusters, examination of between-persons relations—which is done within the latent-trait approach—represents an indirect and often misleading test of such models.

Empirical illustration of the latent-trait-state approach

In a study conducted by our group (Hoffart, Johnson, Nordahl, & Wells, 2018), the latent-trait-state model was adhered to in an examination of a metacognitive model of comorbid anxiety disorders (Wells, 2009). Seventy-four patients who met criteria for a principal disorder of posttraumatic stress disorder (PTSD), SAD, or panic disorder with or without agoraphobia as defined in the fourth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM–IV; APA, 1994) and had failed at least one structured psychological treatment were randomized to 8-week residential metacognitive therapy (MCT) or CBT (see the Supplemental Material available online for details of the sample and treatment). Every Monday they completed a measure of central mechanisms in the MCT model, the Cognitive Attentional Syndrome 1 (CAS-1; Wells, 2009); a measure of anxiety symptoms, the Beck Anxiety Inventory (BAI; Beck, Epstein, Brown, & Steer, 1988); and a measure of depressive symptoms, the Patient Health Questionnaire–9 (Kroenke, Spitzer, & Williams, 2001). The CAS-1 comprises subscales for coping activities to deal with negative feelings or thoughts (eight items), negative metacognitive beliefs (four items), and positive metacognitive beliefs (four items).

The variables were first disaggregated in their between- and within-persons variance components by use of person-mean centering (Wang & Maxwell, 2015). We proceeded with a series of mixed models using anxiety (BAI sum score) as a dependent variable and the person-mean-centered mechanism variables (state components) and the person-mean of the mechanism variables (trait components) as independent variables (for details, see the Supplemental Material). A separate analysis was conducted for each mechanism variable (coping activities, negative metacognitive beliefs, positive metacognitive beliefs). To establish a temporal sequence between mechanism and outcome, we lagged the mechanism variables.

Of the three CAS-1 variables, only positive metacognitive beliefs showed a significant within-persons relationship to anxiety. Higher-than-usual beliefs during a week was associated with more-than-usual anxiety in the subsequent week across the two treatments. There was a statistical trend that stronger negative metacognitive beliefs during a week predicted more anxiety in the subsequent week. Investigating reversed relationships revealed that within-persons changes in anxiety predicted subsequent changes in positive metacognitions, but only in CBT. With respect to theory, the obtained within-persons relationships support the central role of positive metacognitions in the MCT model, and more generally the important idea that positive beliefs may also lead to symptoms. A main clinical implication is that—in the treatment of comorbid anxiety disorders—therapists should look for and facilitate expression of positive metacognitive beliefs (e.g., “Worrying helps me cope”) and challenge these; that is, they should promote time-specific reductions of such beliefs with the expectation that these reductions will involve reinforcing experiences.

Summary of the latent-trait-state explanation of mental problems

The latent-trait-state model represents a clear improvement in mechanism research compared with the latent-trait model. The trait part accounts for the stability, multiplicity and syndromal character of mental problems, whereas the state part accounts for their within-persons short-term fluctuations. In particular, it allows a study of within-persons relationships between mechanisms and problems, a focus that—as argued above—is more relevant to therapy than the between-persons relationships addressed in the latent-trait model. The isolated state component of a measured mechanism variable is, per definition, a part determined by transient factors and thus possibly by the actions of the therapist. In this way, the latent-trait-state approach appears to produce actionable propositions about therapy.

Critical examination of the latent-trait-state approach

The latent construct approaches abstract mechanism and symptom variables from the multiple symptom-related episodic clusters and aggregates these variables in separate and global measures. We argue that this procedure often leads to violations of the requirement of conceptual independence between mechanism and outcome variables and to a despecification of the causal relations within and between episodic clusters into global and indistinct theoretical propositions and clinical recommendations.

A demonstration of an empirical association between change in a proposed mechanism and change in symptoms (Kazdin, 2009) presupposes that these two variables are distinct. When two variables are not distinct, their relationship is tautological and not empirical (Edwards & Bagozzi, 2000). In the latent-trait approach, the two variables are usually estimated by sum scores of a number of items describing mechanistic activities and processes and of a number of items describing symptoms. In the latent-trait-state model, these sum scores are disaggregated in their within- and between-persons variance components—taken to represent two different latent constructs. The independence requirement demands no item overlap on the mechanism and symptom measures. Unfortunately, diagnostic symptom lists and questionnaires purported to measure latent disorders tend to include items often considered mechanisms. Conversely, mechanism measures tend to include items also figuring in diagnostic lists and questionnaires as symptoms. An example of overlap is the obvious similarity of the items “sounding dumb” and “sounding stupid” in Figure 1, the first used as a symptom indicator and the second as a mechanism indicator. A main source of these difficulties is that psychological reactions are defined as symptoms by their harmful qualities without consideration of their potential causal role. Thus, many of the DSM-defined symptoms are viewed as causal mechanistic variables in some therapy theory or other. Examples include the cognition “self-blame” (DSM–5: PTSD), which is seen as a maintaining factor in cognitive therapy of PTSD (Ehlers & Clark, 2000); the mental activity “worry” (DSM–5: generalized anxiety disorder) and the metacognition “finding it difficult to control worry” (DSM–5: generalized anxiety disorder), which are both seen as maintaining factors in MCT (Wells, 2009); and avoidance behaviors (e.g., avoid using public transportation; DSM–5: agoraphobia), which are viewed as maintaining factors in a number of CBT theories.

Moreover, with a closer look, propositions about within-persons latent-state relations may stand out as less actionable. In general, it appears challenging to influence a latent state. How do you do that? A latent variable is, by definition, something intangible (e.g., positive metabeliefs “themselves”) that cannot be directly observed (Edwards & Bagozzi, 2000). It is difficult to act on something you cannot observe, let alone on a statistically abstracted component of something unobservable. Moreover, in a reflective measurement model, the observable indicators of a latent variable are taken as effects and not causes of the variable. Thus, influencing the indicators will not change the latent state. The common-cause assumption of the relationship between latent variables and their indicators leads to the conclusion that it is not possible to improve, for instance, latent cost cognitions through communicative and instrumental action.

However, we could, as suggested by Bringmann and Eronen (2018), consider an alternative noncausal interpretation of the relationship between a latent variable and its indicators, viewing the latent variable not as a cause but as a common property of the indicators (McDonald, 2003). This would involve a substitution of the arrows from latent variables to their indicators in Figure 1 with lines. In this modified version of the latent variable approach, a latent variable would change with changes in its indicators. However, in this model, the causal influences between mechanism and symptom variables are also supposed to occur at a latent level. Effects from individual metacognitions to specific anxiety symptoms, for instance, are not considered (see Fig. 1). The individual metacognitions are considered to be equivalent and interchangeable, with no specific causal role. Thus, the clinical implication would be to target individual metacognitions to reduce latent metacognitions, reflected in the summary score of indicating metacognitions. With respect to causal influence, it would not matter which of them were targeted. The most efficient strategy would be to influence the cognitions with the largest potential for change; that is, those with the highest scores and loadings on the latent factor. Moreover, according to the assumption of local independence, there are no causal relations among the metacognitions or among the symptoms.

However, these assumptions and clinical implications may be unjustified. In the episodic clusters from which the items are abstracted, there may be direct causal influence from particular metacognitions to certain symptoms, instead of indirectly through the latent traits. For instance, the metacognition “I cannot control my thoughts” (CAS-1 item 11) may elicit fear of losing control (BAI item 14). If there are distinct causal dependencies between “indicators” across latent variables, the clinical strategy of targeting the strongest metacognitions should be replaced with targeting the ones with the strongest causal influence. The assumption of local independence may also be challenged. There could be causal relations between metacognitions, as well as between symptoms.

In sum, because mental problems (symptoms) are parts of episodic clusters and may or may not have causal effects on other elements in these clusters, mental problems and their mechanisms are difficult to aggregate in separate classes. Therefore, the assumption of independence of constructs is often violated. Moreover, with its focus on global latent states, findings attained according to the latent-trait-state model can only support or not support general theoretical propositions and lead only to global clinical recommendations. Although such global recommendations may sensitize the therapist for certain phenomena (e.g., “Look for positive metacognitions”), they provide no guidance for what particular mechanisms (e.g., a specific metacognition’s effect on a specific anxiety symptom)—described on a concrete, observable, and thus actionable level—to target. Thus, this line of reasoning suggests that the focus should shift from disorders (wholes) to single symptoms (parts) and from latent to observable mechanistic variables.

The network approach

A shift of focus from the level of disorders, wholes, and latent variables to the level of symptoms (mental problems), parts, and observables is exactly what is achieved in the network perspective on psychopathology, a new alternative to the latent-trait and latent-trait-state perspectives (Borsboom, 2017; Borsboom, Cramer, Schmittmann, Epskamp, & Waldorp, 2011; Bringmann & Eronen, 2018; Hayes, Yasinski, Barnes, & Bockting, 2015; Hofmann, 2014; Schmittmann et al., 2013). Here, concepts of clinical conditions such as SAD are taken to refer to a causal network of observable mental problems or symptoms (observables). The observable symptoms that have traditionally been considered indicators of latent traits (disorders) and states are taken to be directly related causal entities.

In Figure 2, the network conceptualization of SAD is contrasted with the one of the latent-trait-state approach. From the latter perspective, represented in the upper part of the figure, the latent disorder SAD is a common cause of stable dispositions to experience symptoms. These dispositions and a latent social-anxiety state, elicited by transient factors, explain the momentary symptom states. The symptom states are viewed as causally independent and interchangeable effects of the latent disorder SAD and of the latent state of social anxiety.

OPEN IN VIEWER

Fig. 2.

From the latent-trait-state perspective, represented in the upper part of the figure, a latent disorder (social anxiety disorder) is a common cause of stable symptom dispositions. These dispositions and a latent social anxiety state, elicited by transient factors, explain the momentary symptom states. From a causal-network perspective, on the other hand, causal relations between the momentary symptom states are considered, shown by arrows between the momentary states in the lower part of the figure. Networks result in the trait-like character of mental problems or symptoms because symptoms repeatedly activate each other when transient factors in the external field activate one or more of them.

From a causal network perspective, on the other hand, causal relations between the momentary symptom states are considered, shown by arrows between the momentary states in the lower part of the figure. For instance, the experience that one’s voice sounds unsteady may strengthen anxiety as well as the urge to avoid. Moreover, an increase of anxiety may strengthen the experience of unsteadiness. Thus, networks result in the trait-like character of mental problems or symptoms because symptoms repeatedly activate each other (e.g., fear of showing signs of anxiety elicits anxiety) when transient factors in the external field (e.g., strangers) activate one or more of them. Furthermore, symptoms are multiple because they elicit each other and clustered in syndromes because some are usually more strongly connected than others.

Thus, the disorder results from the interaction of symptoms rather than being their cause. Cramer et al. (2012, pp. 423–424) showed how the interaction of nodes (i.e., the components of the construct or phenomenon) in a network leads to stability at the aggregate node level. At a random time point, whether a node is active depends on the status (0 or 1) of each of its neighboring nodes times the relevant connection weight, which results in a total incoming effect. With regard to the network in Figure 2, a person with SAD will have strong connections, leading to a high probability that all three symptoms will be activated when the external node is activated (a social situation is entered). This translates to a high stable (usual) level of aggregated SAD symptoms. A shy person, in contrast, may have moderate connections between other peoples’ presence and experiencing and disliking an unsteady voice and between unsteady voice and anxiety—and zero connection to avoidance (never flees from the situation). This network structure results in a moderate stable level of aggregated symptoms. An outgoing person may have zero connections all over and therefore have a stable level of no symptoms over time. Thus, the interactions of symptoms are responsible for the stable trait-like character of SAD and contribute to the between-persons differences in stable symptom aggregates.

Empirical illustration of the network approach

Bringmann, Lemmens, Huibers, Borsboom, and Tuerlinckx (2015) investigated the temporal network among patients with DSM–IV (APA, 1994) major depressive disorder who received either cognitive or interpersonal therapy. Patients completed the 21-item Beck Depression Inventory–II (Beck, Steer, & Brown, 1996) before each session for an average of 14 sessions. The data were analyzed with the multilevel vector autoregession method (mlVAR; see Supplemental Material). A common network for both treatment groups was estimated. This network provided information of specific connections (e.g., feeling like a failure strongly predicted subsequent worthlessness feelings) and self-sustaining loops (worthlessness and guilty feelings predicted each other) between symptoms. Moreover, the centrality of the symptoms was estimated. Suicidal thoughts and loss of pleasure had the highest out-degree centrality, meaning that they had the highest number of connections going out from them (Opsahl, Agneessens, & Skvoretz, 2010). Thus, when one reports more than usual suicidal thoughts in one session, one is likely to also report an increase in other symptoms in the next session. In contrast, indecisiveness and loss of interest had high in-degree centrality, meaning that they received a high number of connections from other symptoms. Theoretically, the findings suggest that suicidality and loss of pleasure have the strongest causal impact on other depressive symptoms and are therefore promising first targets of clinical intervention. The findings also suggest that indecisiveness and loss of interest are the symptoms most influenced by other symptoms and are therefore hard to change. Thus, they are less promising first targets.

Summary of the network explanation of mental problems

According to the network view, the trait-like character of symptoms and their aggregates into syndromes (e.g., SAD), as well as their short-term fluctuations, arise from the same network structure. The density of this structure explains why symptoms are multiple and clustered in syndromes. SAD and other disorders are viewed as maintained by the causal connections among symptoms and external triggers. Consequently, the mechanisms of change are to be found among the symptom–symptom connections and possibly among processes mediating these connections (e.g., the biological clock involved in the insomnia–fatigue relation; Borsboom, 2017). The placement of mechanisms within the observable elements of the disorders contrasts with the latent-variable approaches, in which mechanistic variables are viewed as latent and external to the disorders. Centrality analyses provide hypotheses about the overall causal impact of symptoms and thus identify symptoms that are promising targets of intervention.

Critical examination of the network approach

The network approach has so far emphasized the syndromal cluster of mental problems or symptoms and not their fifth feature; that is, their episodic clusters. Consequently, it has missed mechanistic relations explaining each symptom as well as important aspects of such relations between symptoms. It has also missed access to and guidance of clinical theories of such episodic clusters.

In the seminal article introducing network theory to the field of psychopathology, Borsboom (2008) used panic disorder as an example by putting the symptoms that define this syndrome in DSM–IV (APA, 1994) into a causal chain: People have a panic attack in a public place, which causes them to worry about the implications of that event and that they may have another one in a public place, as a result of which they do not get out of the house anymore. Although this is a cogent and clinically useful account, the symptoms included in it represent only parts of episodic clusters of reactions, namely, those parts that are experienced as (most) distressing and/or impairing. Thus, each symptom is embedded in a causal chain of cognitive, affective, and behavioral components, and these other components together explain the occurrence of an individual symptom. For instance, according to the cognitive model of panic disorder (Clark & Salkovskis, 2009), the symptom panic attack emerges as a result of a vicious cycle of bodily sensations, catastrophic interpretations of these sensations, and anxiety.

In contrast, the symptom of worrying about new attacks has a somewhat different constitution and dynamics, the persistent worry leading to a more stable level of heightened anxiety, which is also maintained by body scanning and a focus on safety behaviors to prevent the occurrence of attacks. The attack affects the worry by the catastrophic belief strength being transferred from the attack. Thus, the syndromal cluster consists of reactions that are abstracted from their natural context and subsequently brought into relation. However, to understand their mechanisms, mental problems and symptoms have to be brought back to the context in which they are situated, that is, to their typical episodic cluster. Therefore, the network nodes should reflect all causally relevant activities and processes occurring in the episodic clusters and not only the reactions appraised as problems and symptoms (see Jones, Heeren, & McNally, 2017, for a similar quest). Still, the majority of existing dynamic network studies of clinical disorders have focused on the syndromal cluster of symptoms or the subset of affective symptoms (e.g., depressed mood) and their counterparts (e.g., positive affect; Fried et al., 2017).

Thus, network studies have so far addressed the symptom-related episodic clusters to only a limited extent and, consequently, not made use of clinical theories that explain the dynamics within and between symptom-related episodic clusters. Such models could have informed the selection of variables as well as the measurement design.

An extended network approach to mental problems

As shown above, the practice of psychotherapy is a dynamic process in which within-persons deviations from the usual on certain key mechanistic variables are promoted to produce change in the usual level of desired outcome variables. The network approach reflects this practice well by accounting for the relationship between the strength of the connections of temporal and contemporaneous networks—which in effect is the size of the relationships between the within-persons deviations from the usual on the variables—and the overall level of the aggregate of variables included in the network (Cramer et al., 2012). By contrast, the latent-trait-state approach does not conceptualize the connections between the components of a construct and therefore cannot explain the aggregate level by such connections.

Because network studies so far have, for the most part, addressed the syndromal cluster of symptoms, we propose an extended network approach to mental problems by widening the network perspective to the symptom-related episodic clusters. The dynamics of these episodic clusters is explained in some clinical theories of the mechanistic property cluster type, in which clinical features are supposed to cause each other (Hoffart & Johnson, 2017; Kendler, Zachar, & Craver, 2011). Thus, by focusing on episodic clusters, access to relevant clinical theories is attained. For many disorders and their symptoms, even alternative cluster theories exist. For instance, MCT provides an alternative to the cognitive model of panic disorder depicted above. Here, more importance is assigned to the symptom of worrying about new attacks, posited to consist not only of worry and attention to threatening bodily sensations but also of negative metacognitive beliefs (“I must control my thoughts or I could die”) and positive metacognitive beliefs about the importance of negatively interpreting bodily sensations (e.g., “Thinking the worst about sensations could save me”; Wells, 2009, pp. 252–253). These beliefs are supposed to drive the worry and the inward threat monitoring. Thus, use of cluster theories may provide several contrasting hypotheses about predictive relationships to be tested in network studies.

The many theories explaining each symptom and disorder seem to lead to an infinite number of relevant variables. However, Harvey, Watkins, Mansell, and Shafran (2004) examined the commonalities among the processes delineated in models for specific disorders and identified a number of transdiagnostic maintaining processes (e.g., suppressing thoughts, focusing excessively on self, recurrent negative thinking). These findings suggest that a more generic and parsimonious clinical model could be developed, serving as a general guide to what variables should be included in networks. Hofmann (2014) actually suggested a generic CBT model intended to serve as a guide for variable selection.

Within the network approach, the temporal character of clinical observables is largely unresolved (Fried et al., 2017). The distinction between the syndromal and the episodic clusters has implications for the measurement schedules and time lags to be studied. Assessing the relevant predictive relationships requires the registration of the momentary changes in the reactions within the episodic clusters as well as the activation and deactivation of the symptoms (with their episodic clusters). Examples of these relationships are how bodily sensations and catastrophic beliefs predict each other during a panic attack, and how these and other variables predict the variables of the subsequent worry, as well as of the next attack. Thus, these variables must be measured on both a very short and a longer time lag. Intensive methods, such as experience sampling (several observations a day), frequently used in network studies (Geschwind, Peeters, Drukker, van Os, & Wichers, 2011), and electronic measures continuously ² monitoring behavior and psychophysiology, should be used. Noninvasive methods for monitoring mental experience on a moment-by-moment basis are not yet developed and are called for.

Empirical illustration of the extended-network approach

We illustrate the elaborated network approach with a second study (published elsewhere; Johnson & Hoffart, 2018) of the data from our study presented above. Rather than analyzing the within-persons relationships between the latent constructs of this model, in the second study we focused on the network of connections between some of the observables used as indicators of these constructs. Because of the high numbers of parameters to be estimated in network analysis, we had to limit the number of items. We selected central symptoms and variables from the episodic clusters of these symptoms that CBT and MCT theories propose are causal determinants of them (see Fig. 3 for an overview of the items).

OPEN IN VIEWER

Fig. 3.

Temporal networks for cognitive-behavioral therapy (CBT) and metacognitive therapy (MCT). Positive relationships are marked with blue lines and negative relationships are marked with red lines. The strength of the relationship between nodes is represented by the thickness of the arrows and the relative distance of the nodes. The thicker the arrow between two nodes and the closer the nodes, the stronger the relationship. Nodes of items taken from the Cognitive Attentional Syndrome 1: Wor = worry or dwelling on your problems; Att = focusing attention on threatening things; and Nmu = I cannot control my thoughts. Nodes of items taken from the Beck Anxiety Inventory: Con = fear of losing control; Die = fear of dying; Num = numbness or tingling; Hea = heart pounding or racing; and Sha = shaky. Nodes of items taken from the Patient Health Questionnaire–9: Int = little interest or pleasure in doing things; Dep = feeling down, depressed, or hopeless; and Sle = trouble falling or staying asleep, or sleeping too much. Adapted from Johnson and Hoffart (2018).

Network structures were generated from the weekly data provided by our 74 patients using the mlVAR package (Version 0.3.2; Epskamp, Deserno, & Bringmann, 2019) for the R software environment (Version 3.5.3; R Core Team, 2019; see Supplemental Material), producing a temporal, contemporaneous, and between-persons network for each treatment condition. The contemporaneous and between-persons networks (see Figs. S1 and S2 in the Supplemental Material) are not further discussed here because they are considered to be of less clinical consequence. The temporal networks are presented in Figure 3. They represent the average lagged within-persons relationships between the weekly scores on the variables, controlled for each other and unaffected by the between-persons differences in within-persons means. Note that this study had a number of limitations. For instance, having at least 20 measurements per subject is recommended for mlVAR (Epskamp, Deserno, & Bringmann, 2019), but the patients in this study had a mean of only 9.4 measurements (range = 7–14) measurements. Short time series may lead to biased estimates, most notably to negative self-loops (Nickell, 1981). Thus, the negative self-loop of the node Att in the temporal network for CBT in Figure 3 may reflect this methodological artifact. Other limitations are listed in Note 3.

Several differences from the latent-trait-state approach to the present data should be noted. At odds with the latent construct assumption that variables are interchangeable, centrality is estimated. In the CBT network, heart pounding or racing, sleep problems, worry, and attention to threat had high out-strength centrality, measured by the sum of all outgoing connection weights (Opsahl et al., 2010). In the MCT network, worry had the highest out-strength centrality. Also at odds with the interchangeability assumption, there were predictive relationships between a specific metacognition (uncontrollability belief) and a specific coping activity (attention to threat), and between a coping activity (worry) and an anxiety symptom (shaky). At odds with the latent construct assumption of local independence, there were predictive relationships between two coping activities (worry and attention to threat) and between two anxiety symptoms (shaky and heart pounding or racing). In contrast to the failure to find an expected within-persons predictive relationship between coping activities and anxiety symptoms when the latent-trait-state approach was applied, such relationships appeared when these constructs were broken down to their individual components. Thus, worry predicted shaky in MCT and attention to threat predicted depressed mood in CBT.

Several differences from the common network approach to the syndromal cluster of symptoms should also be noted. First, the theory-derived mechanistic variables supposed to be active in the symptom-related episodic clusters will explain much of the variance of a symptom otherwise explained by other symptoms both in these clusters and in the syndromal clusters. Thus, the explanatory power of the symptoms will be greatly reduced. This is reflected in the MCT temporal network, in which only one symptom (depressed mood) was predictive. Second, as the selection of network variables is informed by clinical cluster theories and not only by diagnostic lists, analyses provide tests of the causal chains and feedback processes proposed by these theories. Thus, the centrality findings and the patterns of specific connections did not support the CBT model, as the cognitions were not central and did not predict symptoms in either of the therapies. The findings were more consistent with the MCT model, as worry was central in both therapies and—in MCT—worry predicted an anxiety symptom and uncontrollability belief predicted attention to threat. Third, the causal variables proposed by cluster theories are typically behaviors or internal activities that are or may ultimately be brought under intentional control, such as safety behaviors, worry, and self-focus. Thus, variables that may be more easily and directly modified are suggested. By contrast, symptom lists consist, for the most part, of passive reactions such as emotions and bodily sensations that in principle cannot be directly controlled. Finally, the use of cluster models gives access to interventions designed to influence the mechanistic variables. For instance, worry stands forth as a variable with high out-strength centrality in both therapies, and the MCT model proposes worry postponement, that is designating a single time and place for worrying (Wells, 2009). Moreover, the MCT model suggests the intervention detached mindfulness—leaving the worrying thoughts alone—to break the connection between worry and shakiness in the MCT network (Wells, 2009). The CBT network suggests that sleep problems are a central target, and Harvey’s (2002) cluster theory of insomnia provides a range of interventions.

Summary and contextualization of the extended-network approach

To provide a causal network explanation of a symptom, as well as this symptom’s relationship to other symptoms, variables (nodes) of the symptom’s typical episodic cluster need to be included. Thus, beyond the syndrome-focused network approach, the extended-network approach presented here accounts for dynamics in the episodic clusters around problems and symptoms. Established property cluster models of episodic clusters should be used to guide the inclusion of variables. Thus, a cross-fertilization of the domains of clinical cluster theories and network analysis may occur. The cluster models provide the hypotheses to be examined, whereas network analyses provide methods to visualize and examine these hypotheses.

The extended-network approach is consistent with and is provided a wider reference frame by the quest for a process-based therapy (Hofmann & Hayes, 2019). As an alternative to the medical-illness model, which assumes that latent disease entities are targeted with specific therapy protocols, process-based therapy emphasizes therapeutic processes. A therapeutic process is defined as a set of theory-based, dynamic, progressive, and multilevel changes that occur in predictable, empirically established sequences oriented toward the desirable outcomes (Hofmann & Hayes, 2019, p. 38). The elaborated-network approach is based on clinical cluster theories, which lead to testable predictions about relationships between observables. It is dynamic by way of proposing temporal within-persons relationships between deviations from patients’ own usual levels on mechanism and symptom variables. It is progressive by way of specifying how related within-persons changes on these variables lead to subsequent change in the usual levels on them. It is multilevel by way of distinguishing the episodic and the syndromal clusters. Finally, it has immediate and long-term treatment goals by way of seeking repeated deviations from the usual on important mechanistic variables and changes in the usual personal level on symptom and other outcome variables.

Parsimony

The latent-variable views lead to simple models. They provide three possible causal pathways between a potential mechanism and disorders, for instance, between cost cognitions and SAD. The indicating observables are taken to be equivalent and interchangeable. The extended-network view, by contrast, leads to complex models. They consist of multiple relations exemplified by the networks in Figure 3. The latent-variable approaches also differ from the network approach with regard to the amount of data needed to examine models. Analyses of summaries of multiple observables instead of the observables themselves require many fewer data units. Thus, the latent-variable approaches are more parsimonious than the network approach, in terms of both the simplicity of conceptual models and amount of data needed.

Conceptual-independence requirement

The latent-variable approaches rely on a grouping (e.g., into sum scores, factor scores, or structural equations) of variables (items) taken from problems’ episodic clusters of observable reactions. Such grouping requires that the observables can be unambiguously defined as mechanisms or symptoms. As elaborated above, many mental problems are also considered causal mechanisms, precluding their unambiguous ordering into these classes. Although nonharmful mechanistic variables (e.g., mindfulness) are separate from problems and symptoms, many central cognitions, mental activities, metacognitions, and behaviors are alternately considered mechanisms and symptoms. Thus, the requirement of independence between mechanisms and symptoms is often violated, making the propositions about mechanistic relationships not meaningful to test in these cases. Because the focus of the network approach is on the observable, activities, and states themselves and not on latent entities, the problems of classification and consequent dependence between constructs do not arise.

Precision of theory tests

The latent construct approaches abstract mechanism and symptom variables from the multiple symptom-related episodic clusters and aggregates these variables in separate and global measures. Thus, a double abstraction activity is conducted: First, the variables are removed from their natural episodic context and, second, they are made homogeneous within two classes. Consequently, only global and general theoretical propositions can be tested, propositions with an unknown relationship to the causal influences actually occurring in the episodic clusters. The network analyses on the level of individual items allow a much more differentiated pattern of relationships, not assuming interchangeability and local independence of the items. Thus, the multiple causal chains and feedback processes proposed by clinical cluster theories are examined simultaneously.

Clinical cluster theories often include propositions about the relative importance of causal variables in the episodic clusters. For instance, the cognitive model of SAD holds that the reciprocal relationship between perceived social danger and self-focus is the “central engine” in this disorder, claiming that perceived social danger is much more determined by self-focus than by the actual social situation (Clark & Wells, 1995, p. 74). In the latent-variable views, centrality of observables is not formulated, as the observables are postulated to be interchangeable. By contrast, network analysis includes an estimation of centrality derived from the pattern of network connections. Thus, in our illustrative study, worry was a central variable in both MCT and CBT, as expected from the MCT model.

It should be noted, however, that the centrality measures used are developed for social networks, and it is unclear to what extent they reflect the relationships between psychological variables (Bringmann et al., 2019). Within the symptom-related episodic clusters, one reaction is assumed to lead to several other reactions in a relatively fast and automatic way. For instance, social-danger cognitions are assumed to immediately influence self-focus, anxiety symptoms, and impulses to engage in safety behaviors (Clark & Wells, 1995). Thus, it is plausible that a node affects those that it is directly connected to in a parallel way. Such a one-lag parallel flow can be modeled with strength centrality, which we have used in our empirical example. It is more questionable, however, whether the importance of variables is validly reflected in the unique and unshared explained variance of a node produced by VAR models (Bringmann et al., 2019). Therefore, improved measures of the importance of psychological variables in networks need to be developed.

Specification of mechanisms

A description of a mechanism should also include a further explanation of how it operates through an account of the lower-level activities or processes involved (Kazdin, 2009). From a latent-variable perspective, such specificity is attained through introducing latent subfactors. For instance, an obtained relationship between threat monitoring—attention to danger cues—and anxiety could be further studied by examining the influence of subfactors of attention. Still, supposed latent entities are addressed, requiring the ordering and aggregation of often-ambiguous observables.

The focus of the extended-network approach, however, is the observable reactions themselves and their relationships within the episodic clusters. In virtue of being concrete, these activities can be further divided into subactivities. For instance, self-focus—threat monitoring of one’s own reactions—is proposed to be a central mechanism in SAD. To investigate how self-focus exerts its effects, Heeren and McNally (2016) included various components of attentional control and bias—measured by reaction time in laboratory tasks—in a network study of SAD. ⁴ They found that the orienting component of control—the selective engagement and disengagement with certain stimuli rather than others—was one of the central variables of the SAD network, whereas the other components—alertness and executive control—played more peripheral roles. These findings support the use of attention training techniques (Wells, 2009) to alleviate social anxiety. Thus, network findings obtained between observables naturally observed and reported in daily life can be a vantage point for laboratory-based studies of subprocesses not available to introspection or unaided behavioral observation. Studies identifying the activities and processes actually producing the symptoms within the episodic clusters are possible, providing direct tests of clinical theories.

Clinical guidance

The simplicity of latent-variable models prevails at the cost of multiplicity and specificity. Whereas these models lead to general recommendations such as “modify positive metacognitions to reduce anxiety,” the extended-network approach leads to multiple and specific recommendations based on several kinds of network features (specific connections, centrality of variables). The specific recommendations are formulated at the level of what can be observed. This difference in specificity of targets transfers to the associated interventions: strategies that apply to all positive metacognitions (e.g., go through the advantages and disadvantages of the belief) versus particular interventions tailored to address specific beliefs. For instance, the belief “attention to threat makes me safe” could be handled by a specific behavioral experiment in which attention is manipulated. A PTSD patient could be instructed to focus on neutral or safety signals in the environment, learning that the strategy of looking for danger leads to ignorance of the actual features of the situation (Wells, 2009, p. 69).

As a caveat, that intervention on central and predictive variables will be effective is only a hypothesis that needs to be tested. A temporal link may arise for reasons other than causality; for instance, because a third variable is causing the two variables involved in the link (Epskamp, Waldorp, Mõttus, & Borsboom, 2018). Moreover, most research reports only the average within-persons connection between variables, and particular patients may deviate more or less from this average. ⁵

Personalized psychotherapy

There have been increasing calls for personalized psychotherapy (e.g., Insel, 2009). From repeated data in a pretreatment observational period, personal models can be constructed from both a latent variable (Fisher, 2015) and a network (Epskamp, van Borkulo, et al., 2019) perspective. Latent dimensions of symptom and process variables can be developed through person-specific dynamic factor analysis. However, the application of semantic labels to the factors may have the effect that indicators with lesser factor loadings may be overlooked in treatment planning (Fernandez, Fisher, & Chi, 2017). For instance, concentration difficulties loaded moderately on a factor called muscle tension in one participant of Fisher’s (2015) study, but treatment would probably target muscle tension. Such problems do not arise within the network approach, as no classification of observables is involved. Based on personal networks of observables in symptom-related episodic clusters, personalized treatment may be constructed, consisting of modules to address the most important processes and symptoms in the optimal sequence (Fried et al., 2017).