Abstract
In this article, we briefly overview some current approaches to identifying targets of explanation in psychopathology. We suggest that DSM syndromes and symptom network models are too large to facilitate explanatory progress, while approaches couched in the Research Domain Criteria (RDoC) are too tightly focused, thereby overlooking the self-maintaining structures and processes that constitute mental disorders. We then present the Relational Analysis of Phenomena (RAP) as a complementary method. The RAP is a metamethodological explanatory framework that prescribes the repeated targeting of “phenomena complexes” (PCs). PCs are small systems of clinical phenomena and their relations. We describe a process by which PCs are targeted, richly described, and then explained. The RAP’s explanatory focus is on the relationships between clinical phenomena, which should help to reveal the constitutional structure of mental disorders. This framework is designed to be useful for theoreticians, research teams planning multidisciplinary investigatory projects, and clinicians reflecting on their explanatory methods.
The so-called forms of illness in their present-day delimitation have turned out to be too large, on the other hand the elementary symptoms, because they represent single phenomena, are less useful for distinguishing between the various conditions. Between these two ranges of phenomena would be the symptom complexes [emphasis added].
Mental disorders negatively affect the lives of hundreds of millions of individuals and represent a significant proportion of the global burden of disease (Kessler et al., 2009; Walker et al., 2015; Whiteford et al., 2015; World Health Organization, 2019). To treat and manage mental disorders more effectively it is first necessary to develop good explanations of them. There is now growing recognition that the status quo approach of launching research expeditions in and around current DSM constructs has not resulted in sufficient progress (e.g., Clark et al., 2017; Lilienfeld & Treadway, 2016). Recent responses to this challenge have been made in the form of several proposals regarding what mental disorders are and how we should go about explaining them. Key examples include the Research Domain Criteria (RDoC; Cuthbert & Insel, 2013; Insel et al., 2010) and Symptom Network Modeling (SNWM; Borsboom et al., 2019). These methods vary as to what they see as the most appropriate explanandum (i.e., the thing-to-be-explained; alternatively labeled “target of explanation”).
In the first part of this article, we briefly examine these approaches, focusing on how their explananda are conceptualized and considering the degree to which such targeting will support the timely development of good explanations in psychopathology. Due to space constraints, it is not our intention to review these approaches, merely to demonstrate the space they leave for our complementary approach. In the second half of this paper, we present our own method for developing explanations of psychopathology, which we label the Relational Analysis of Phenomena (RAP) approach. As well as stemming from the overview presented here, the justification for this approach has roots in our previous work on the nature of mental disorder when viewed through an enactive, embodied, and embedded lens, and ideas concerning the likely circular causal structures underlying the stability of disordered behavior (Fuchs, 2009, 2017; Nielsen & Ward, 2018, 2019). It is important to stress that we adopt a pluralistic perspective to the explanation of psychopathology, and as such view the RAP approach as providing an additional explanatory method, rather than being necessarily the “right way.”
Explananda in current approaches
The role of explanandum has two major competing requirements. First, a good explanandum should be a robust and reoccurring phenomenon (Haig, 2014). This means that the thing we are trying to explain needs to be reasonably similar (both in appearance and constitution) across different instances/persons—that is, it needs to demonstrate construct stability (Sullivan, 2014). At the same time, explanations have a motivating context within which they are sought, and this context forms the pragmatic landscape; rightfully influencing many of our decisions during the explanatory process (Potochnik, 2010, 2016, 2017; Thagard, 2017). Within science, this often takes the form of a research problem that motivates the enquiry process (Haig, 2014). The second major requirement is that a good explanandum must maintain its relevance to this reason for seeking an explanation. By this we mean that even a “perfect” explanation (in terms of purely explanatory values) can meaningfully be said to be a bad explanation if it explains something totally irrelevant to the problem at hand. We primarily seek explanations of psychopathological phenomena because these phenomena bring about harm and dysfunction in people’s lives. We want to know how to alleviate this harm as effectively as possible, and to do so quickly.
Diagnostic and statistical manual-based approaches
Historically and currently, Diagnostic and Statistical Manual (DSM) syndromes are commonly used to define the explanandum in psychopathology research (Berenbaum, 2013). There are many recognized problems with the DSM which have been extensively reviewed elsewhere (Karter & Kamens, 2019; Lilienfeld & Treadway, 2016; Zachar & Kendler, 2017). The issue most relevant here is that of heterogeneity (Lilienfeld, 2014). This is where individuals with the same diagnoses often have differing patterns of symptoms with differing levels of severity (i.e., symptomatic heterogeneity), suggesting that the diagnostic label in question may be capturing more than one underlying causal process (i.e., etiological heterogeneity). This concern is well evidenced; prototypical disorders such as posttraumatic stress disorder (PTSD), eating disorders, schizophrenia, and depression have all been shown to capture large and heterogenous populations (Contractor et al., 2017; Dickinson et al., 2018; Galatzer-Levy & Bryant, 2013; Hawkins-Elder & Ward, 2020; Monroe & Anderson, 2015). Ultimately this suggests that there are good reasons to doubt the etiopathological validity of the DSM’s diagnostic constructs—that is, that they pick out similarly constituted entities with common causal processes/structures. For someone seeking etiopathologically and constitutionally valid (rather than simply descriptively valid) entities, DSM-style syndromes seem artifactual and not the sorts of stable and relevant things we should seek to explain.
Research domain criteria-based approaches
In response to the problems with the DSM mentioned above, the US National Institute of Mental Health (NIMH) launched the RDoC: a research framework with the goal of shifting attention from signs and symptoms to the underlying causal processes that generate them (Insel et al., 2010). In doing so, it assumes mental disorders to be disorders of “brain circuitry” (Insel et al., 2010; Morris & Cuthbert, 2012).
RDoC adopts an organizational matrix with a horizontal axis containing seven “units of analysis” (which specify structural “levels” of enquiry), and a vertical axis listing basic psychological functions (Cuthbert & Insel, 2013; Cuthbert & Kozak, 2013; Lilienfeld & Treadway, 2016; Morris & Cuthbert, 2012). The explanatory aim is to study how a phenomenon observed at a particular unit/level (e.g., higher levels of striatal dopamine, lower dendritic spine density in brain area X) affects the degree to which the basic functions are achieved (e.g., response to acute threat, approach motivation). The hope is that this process will uncover transdiagnostic mechanisms relevant to current diagnostic labels (Cuthbert & Insel, 2013; Hoffman & Zachar, 2017). Under RDoC, “transdiagnostic mechanisms” refer to neural circuit abnormalities that negatively affect the specified functional domains.
While many have concerns surrounding the potential neurocentricism and reductionism of the RDoC movement (Berenbaum, 2013; Hershenberg & Goldfried, 2015; Kirmayer & Crafa, 2014; Lilienfeld, 2014; Lilienfeld & Treadway, 2016; Nielsen & Ward, 2018; Wakefield, 2014), the shift to focusing on transdiagnostic mechanisms and their relation to specified functions represents a shrinking of explanatory targets towards more stable phenomena. This move seems an advisable response to the heterogeneity plaguing DSM-defined targets. Regarding this move, however, Hoffman and Zachar point out a concern that we share: “[t]he worry is that in order to achieve the fineness of grain needed for elucidation of causal mechanisms, we risk losing connection to the ‘coarse’ clinical phenomena of interest” (2017, p. 68). In other words, even if we managed to explain how some neural-level abnormality is associated with, say, differences in the initiation of motor action, there would still be a long way to go to understanding the role of this mechanism in currently recognized patterns of psychopathology such as depression.
Essentially then, there seems to be a possibility that RDoC represents an overcorrection in the grain size of the explanatory targets in psychopathology—in which targets do not maintain their relevance to the wider dysfunction and suffering that motivate our enquiries. Ultimately, this concern is probably outweighed by the sheer amount of basic research that RDoC will facilitate. But we need to be clear about what RDoC is doing. Research within the RDoC framework searches for (largely subpersonal) abnormalities that likely play constitutional and/or causal roles as components of psychopathology. This is vital work, as it discovers and confirms phenomena that can then be used to weave together an explanation—but such phenomena do not themselves constitute explanations of psychopathology.
Symptom network modeling-based approaches
SNWM is presented by its advocates as a new model of mental disorder that largely rejects the search for underlying cause/s of psychopathology. 1 Instead, SNWM assumes that many mental disorders are better understood as networks of symptoms, which can then be statistically modeled. Symptoms within these networks are hypothesized to cause each other, with recursive feedback resulting in the relative stability of the network over time (Borsboom et al., 2019; Cramer et al., 2010; McNally, 2016). Recent years have seen a significant increase in SNWM research, with many examples of it being used successfully in empirical studies (Fried et al., 2017).
The key strength that symptom network models (SNWMs) bring as descriptive models is specificity. We can ask more specific questions than we can with DSM syndromes such as: why particular symptoms predict others, or why certain networks of disorder predict the “activation” of other networks (e.g., the recognized comorbidity between depression and anxiety). We refer to this specificity regarding associations as horizontal detail. One potentially useful element of horizontal detail is the ability to measure the centrality of a symptom within a network, effectively mapping the strength of its associations with other relevant symptoms (Fried et al., 2017). For example, low energy and anhedonia are reasonably central to depression in that they predict the presence of other depression-related symptoms more strongly than less central symptoms such as agitation (Fried et al., 2016). Centrality, then, may potentially act as a guide as to where we should most efficiently focus our exploratory and explanatory work. 2 There are ways that horizontal detail could be improved in SNWM approaches, namely, shifting to the use of directed networks (these incorporate longitudinal data thereby allowing for causal inference), focusing on individual rather than group abstracted networks, and increasing the sample rate to produce greater temporal resolution/dynamicism (Bringmann & Eronen, 2018; Bringmann et al., 2013; Fried et al., 2017; McNally, 2016; Molenaar, 2010; Wichers, 2014).
SNWMs also have significant weaknesses as descriptive models. SNWMs are often generated from group-level data with large sample sizes and, currently, there are no established “goodness-of-fit” measures that assess the reliability with which the group-level abstracted network matches the pattern of associations within individuals (Beard et al., 2016). This is problematic because patterns that emerge at a group level are not always present at an individual level, yet ultimately it is the individual that we are most interested in when developing explanations and treatments (Barlow & Nock, 2009; Beltz et al., 2016; Blampied, 2017). The lack of an appropriate measure of group-level to individual-level reliability brings into question SNWM’s ability to meet the stability requirement we have outlined. The suggestion of shifting to the measurements of symptom networks in individuals across time mentioned earlier may go a long way in addressing this issue (Bringmann et al., 2013; Fried & Cramer, 2017; Wichers, 2014).
The biggest weakness of SNWMs as phenomenal models, however, is that they lack vertical detail. By this we are referring to the fact that SNWMs operate purely at the “symptom level” (Ward & Clack, 2019). This may seem an odd criticism to make—they are symptom networks after all. However, Borsboom et al. (2019) claim that SNWM as an idea is inspired by, or at least conceptually related to, the concept of a mechanistic property cluster (MPC); “A research program that has put . . . [mental disorders as MPCs] to work is the network approach to mental disorders” (p. 12). An MPC is a concept applied to mental disorders by Kendler et al. (2011). Under the MPC view, mental disorders are explicitly multi-scale clusters of mutually reinforcing causal mechanisms. Given this grounding in the MPC concept, it is unclear why SNWMs should be limited to the symptomatic scale. For SNWMs to act as phenomenal models of disorders that facilitate mechanistic insight, they need to map rich multi-scale detail, including the constitution of the symptoms themselves (Ward & Clack, 2019).
Summary
To summarize, DSM-defined targets seem too unstable to serve as explananda productively. SNWM and the RDoC, meanwhile, both perform complementary but distinct roles relative to the method we will present in the latter half of this article. SNWMs are useful for describing functional relationships between symptoms, and the RDoC will help uncover (largely subpersonal) differences in those who experience disorder; phenomena which may then play a role in our explanations. The method we will present is designed to focus explanatory attention on the relationships between symptoms (or rather, clinical phenomena). We will now provide an overview on why we think such a focus will be productive before presenting the method itself. The intention is to isolate explanatory targets that are more stable and more richly described than both DSM syndromes and SNWMs, while also being more immediately relevant to individual suffering than RDoC’s transdiagnostic abnormalities.
Groundwork for an alternative proposal
Recent conceptual models concerning the ontic structure of mental disorders highlight the possibility of emergent stability playing a key role in their maintenance (Borsboom et al., 2019; Fuchs, 2009, 2017; Kendler et al., 2011). Emergent stability refers to the idea that something (e.g., a mental disorder) may persist due to the causal relationships between its parts cycling back and resulting in a stable pattern or state. Primary among these views is the MPC view, which, as we mentioned earlier, is independent from but underlies the SNWM approach (Borsboom et al., 2019). This view holds that many mental disorders may be constituted by mutually reinforcing causal mechanisms that cross scales of analysis (Kendler et al., 2011). Our own conceptual work—grounded in an embodied, embedded, and enactive understanding of human functioning—also highlights emergent stability (Nielsen & Ward, 2018). We have argued that at a conceptual level it is this emergent stability that explains why sufferers continue to act contrary to their own functional norms 3 (Nielsen & Ward, 2019; see also Nesse & Stein, 2012, on disorders from control system failures).
The common thread to these views is the highly circular process structure of the mechanisms seen to constitute disorders (Fuchs, 2009, 2017). In effect, this circularity can be seen as a basic form of self-maintenance. 4 As one example, there is now converging evidence that nonsuicidal self-injury (NSSI) such as cutting, scratching, punching objects, and so forth, self-perpetuates due to it serving an emotion regulation function. Short-term, engagement with NSSI has been shown to alleviate emotional distress. In the long-term, it fails to relieve distress and discourages the use of alternative regulation strategies. This then seems to lead to continued engagement with NSSI despite its significant risks (Chapman et al., 2006; Robinson et al., 2019). 5
To optimize relevance then, what we need to understand about mental disorders is this self-maintaining dynamic. As reviewed, the RDoC is targeted at a grain size that is likely inappropriate for the purposes of capturing this circular causality and the ensuant maintenance of dysfunctional behavior. SNWM is of an appropriate grain size for this purpose but is focused on modeling the relationships between the parts (symptoms) of disorder rather than explaining them. How then should we best seek to understand the relational structure—the diachronic constitution—of mental disorders? We see room here for a method that focuses on the relationships between the parts of a disorder, as a way of developing an understanding of how mental disorders self-maintain. One way that this could be achieved is to repeatedly select out small systems of interacting parts from the wider disorder for closer analysis. This is the core idea of the RAP approach, which we will now present.
The relational analysis of phenomena approach: A three-phase metamethodological framework
The RAP framework is designed for use by researchers when attempting to develop explanations in the field of psychopathology, by research teams planning multidisciplinary investigatory projects, and may also be useful for clinicians reflecting on their explanatory methods (although it is not intended to be applicable to clinical practice wholesale). It is particularly focused on the development of explanations of the maintenance of disorder, in that it is primarily designed to produce constitutional explanations of the dysfunctional behavioral pattern (e.g., why the components hang together/continue to be engaged in) rather than etiological ones (e.g., what led to the development of the disorder in the first place).
According to the RAP, it is not the objects of our classification system that we seek to explain. Classification systems are simultaneously ontological lists, diagnostic tools, and sociopolitical documents (Zachar, 2019). Each of these tasks bring their own biases and constraints. This issue requires much deeper analysis than space allows, but, suffice to say that classification systems in psychopathology will always be subject to such competing purposes. Instead of recommending the complete separation of psychopathology research from DSM categories as per the RDoC, the RAP allows DSM syndromes to point out potential areas of exploration, while not allowing for DSM syndromes to define either the local explanandum nor the wider disorder one is trying to understand. We henceforth refer to the wider disorder as the problem space for the purposes of clarity, and to highlight this decoupling.
The three phases of the RAP are: (a) phase 1: list and map, (b) phase 2: focus and enrich, and (c) phase 3: explain and evaluate. The first two phases are descriptive in nature and serve to isolate a target of explanation for the third explanatory phase. As we will describe, these three phases are designed to allow investigators to go back and improve their descriptions and explanations over time (see Figure 2 for a guide to this structure). The RAP is also iterative in a larger sense. Cycling back and seeking to explain different overlapping PC structures should produce an understanding of the wider problem space; an understanding of the constitutional structure of the disorder.
Phase 1: List and map
List (1a)
The key task at this stage is to develop a list of reliable phenomena within the problem space. A comprehensive literature review is called for, identifying and evaluating the reliability of possible phenomena (e.g., checking for replicability, multimethod triangulation, lack of conceptual overlap with other phenomena). The objective is a listing of clinical phenomena.
We have used the term “phenomena” throughout this article; it is worth specifying our meaning here. The concept of central importance during this stage of investigation is the data/phenomena distinction. This distinction is made by Bogen and Woodward (1988) and discussed further in Haig’s (2014) Abductive Theory of Method. According to this distinction, data are observable things such as recordings or reports about the state of the world. Unfortunately, data are inherently noisy and often biased. A phenomenon, meanwhile, is an apparent fact about the world, inferred from the data based on reliable patterns therein. A reliable phenomenon will be inferred from multiple and replicable sources of data (e.g., self-report, observation, behavioral tests). On this view, scientific theories do not explain the data, rather they explain the phenomena. This explanation is achieved through the postulation of causal or constitutional mechanisms. Generally speaking, phenomena can take many forms such as objects, states, processes, events, and effects (Haig, 2014).
When we speak of clinical phenomena, we refer to phenomena that are relatively specific to the target population compared to the wider population, or that otherwise seem to be playing a role in the problem space. Within the RAP approach we explicitly use “clinical phenomena” to refer to behavioral and phenomenological instances found reliably within the problem space. This is not to say that these phenomena only exist at these scales/perspectives (phenomena are usually observable at multiple scales), only that within this method, phase 1 should be limited to detecting phenomena within these domains.
At phase 1a, the focus is on states (e.g., moods and emotions, levels of awareness), events/actions (e.g., self-harm, outbursts of anger, bodily sensations or other perceptual experiences), and tendencies/dispositions (e.g., thought–action fusion, apparent perceptual biases, anhedonia, paranoia). Effect and process type phenomena are of interest but are incorporated in phase 1b (e.g., that purging often follows binging, that anhedonia often increases with chronic stress).
Phenomena that occur within other disorders should be included on this list. While it may be tempting at this stage to simply import the DSM criteria, which are often taken to describe the recognized problem, this will not be a fruitful approach. We want to eventually explain the disorder as it actually occurs, not as it is idealized in our diagnostic manuals which have been heavily biased for diagnostic reliability and other purposes (Zachar, 2019). Phenomena measured by psychometric tests may also be unexhaustive and should be supplemented by comprehensive literature trawling and observation.
Map (1b)
The key question at this stage is, “what are the known/apparent relationships between the clinical phenomena?” Technically, these relationships (when reliably detected) are themselves phenomena. For clarity we therefore refer to them as relational phenomena. Investigators should seek to map the clinical phenomena listed in phase 1a into a network of relations. Here we are drawing on the SNWM approach. This can be done using directed symptom network modeling (dSNWM), some other form of dynamical modeling, or (in lieu of such tools), a time-sensitive conceptual sketch. The relational phenomena that emerge should themselves be subject to the requirements of replicability and multimethod triangulation to ensure their status as phenomena.
At this stage, awareness needs to be drawn to the fact that relational phenomena in psychopathology exist at varying time scales. For example, panic disorder is defined by the presence of panic attacks leading to persistent worry and/or maladaptive behavioral changes in response to the panic attacks (commonly taking the form of agoraphobia). The development of fear and avoidance strategies, and the possible resulting low moods and other secondary impacts, occur on a time scale ranging from days to months. Compare this to panic attacks themselves: a collection of physiological/experiential phenomena which occur over a timescale of minutes. In managing this temporal complexity, it may be necessary to produce multiple maps of associations between phenomena at different time scales. The tighter frame–rate associations may then be nested into the wider time–frame network as a composite phenomenon.
A tangential but important task at this stage is to perform a validity check of sorts. Investigators should consider whether the behaviors under study are genuinely problematic. The question here is, “what is it that makes this a problem for individuals within their physical and cultural context?” Creating a list of the functional norms typically being impinged on by this network of behavioral and phenomenological phenomena may be helpful, and—alongside the centrality of clinical phenomena—can also be used to guide the targeting process in phase 2.
Phase 2: Focus and enrich
Focus (2a)
The key task at this stage is to select a cluster of two to four clinical phenomena and their relations from within the now-mapped problem space, and to model them as a small system of interacting phenomena (i.e., temporarily ignoring their relation to clinical phenomena outside this selected system). This idealized model is referred to as the phenomena complex (PC). At least two phenomena are needed so that there is a relation to explain. The suggested upper limit of four phenomena within the PC is chosen simply to support manageability on the part of the explainer. Once richly described (in phase 2b), this PC will take the role of explanandum.
The selection of phenomena at this stage is not arbitrary, but at the same time PCs are not intended to be “real” things in the sense of being naturally separable parts of disorders. Instead, they are pragmatically defined abstractions (idealizations in the model buildings sense; Potochnik, 2017) that try to balance relevance, fertility, and manageability. 6 Accordingly, there are certain considerations that should inform the selection. First, early in the project, investigators should prioritize core phenomena and relations that appear to be doing a lot of work in the network produced during phase 1b (SNWM measures such as centrality may be useful here), or phenomena that seem important because of their particularly negative impact on people’s lives. Second, PCs are primarily identified pragmatically, however, targeting of apparent natural clusters within the mapped problem space is a good option. The limiting factor is that the PC should be limited in size so as to keep the task of explanation manageable.
Finally, ideal PCs will feature a circular organizational structure. This circular process structure is conceptually what allows for the self-maintenance of the dysfunctional behavioral pattern. Selecting circular structures as PCs, then, effectively balances the two key explananda requirements of stability and relevance. The competing need to keep the task of explanation cognitively manageable, however, should not be undervalued. If capturing this circularity is not possible while keeping the number of constituent phenomena low, then ignoring the possibility of capturing circularity within the current PC and focusing on the selected relational phenomena is perfectly valid. As we will show later, the iterative nature of RAP’s design allows for some exploration here—there is no one correct selection of phenomena.
These PCs, then, are small systems of two to four clinical phenomena. The relations between the constituent clinical phenomena are seen to be potentially causal in that there is good evidence for a causal link between the constituent phenomena, but the exact mechanism is unknown. By conceptualizing the explanandum in this way, explicit attention is drawn to the process structure of the disorder space and how this supports the organism’s continued engagement with dysfunctional behavior. What we seek to explain (in phase 3), is the nature of the relationships between the constituent phenomena.
Prototypical examples of ideal PCs already exist, such as the binge–purge cycle or self-starvation spiral in eating disorders (Hawkins-Elder & Ward, 2020), the experiential avoidance cycle in OCD, or escalation cycles in the families of children with conduct problems. Note that these are ideal examples in that they are all highly circular structures, metaphorically acting as “engines of distress.” To restate, PCs do not always have to feature this circular organization. Readers may protest at this point that these examples seem to be theories rather than descriptions, and they would be in part correct. These examples seem to foster a degree of understanding as to why individuals continue to binge and purge, starve themselves, perform bizarre rituals, or consistently misbehave respectively. But as “theories,” they are remarkably thin. They rely on intentional and empathetic inferences on the part of the person using them to understand someone’s behavior. Beyond this, the mechanisms remain largely unknown. It therefore seems more accurate to consider them as phenomenal models (Hochstein, 2012, 2013, 2016a), or as cyclical mechanism sketches waiting to be filled out (Bechtel, 2011; Piccinini & Craver, 2011).
Enrich (2b)
The task at phase 2b is to develop constitutional descriptions of each constituent clinical phenomenon. The constitution of the selected clinical phenomena must be described across scales of analysis both below and above the behavioral and phenomenological.
Here we draw on the ideas of Ward and Clack (2019) and Hochstein (2016b), in that the constitution of each clinical phenomena should be described via a set of friendly models at varying scales of analysis. This method is required given the constitutional complexity of clinical phenomena. The term “friendly” refers to the fact that the descriptive models should be reasonably coherent, but not necessarily integrated or reducible to each other. The reason for this use of pluralism is that explanations at different scales make different idealizations, that is, different models of the same phenomenon or mechanism are designed to abstract away from certain elements and to focus on different elements (Hochstein, 2016b). Consider the phenomenon of anhedonia. One popular neurological model of anhedonia, postulated by Ferenczi et al. (2016), focuses on activations/modulations of different brain areas/neurochemical systems and their ensuant effects of reward-seeking behavior. In doing so, it abstracts away from individual differences and contexts, and indeed genetic factors that may be playing a role in the wider phenomenon. Compare this model to behavioral models that may focus on wider contextual factors (e.g., stress) and map the behavior in finer detail: phenomenological models that attempt to richly describe the difficulties with feeling pleasure from a first-person perspective 7 or cultural models that try to capture how different kinds of positive emotions may be more important across different cultures, thus changing the impact of anhedonia in different contexts. A truly rich understanding of the constitution of a clinical phenomenon requires description through a plurality of models across scales of analysis (see Ward & Clack, 2019, and Hochstein, 2016b, for further discussion of this pluralistic method of description).
Incorporating externalism?
As a guide for structuring pluralistic description, Ward and Clack (2019) suggest the possibility of utilizing the RDoC units of analysis. The use of an organizing structure in which to nest the set of descriptions is a useful one, however, using the RDoC units does risk importing its neurocentricism. The mechanism’s “underlying” behaviors are not necessarily within the individual, but often span the environment. When seeking to understand complex systems such as humans, we must—in Bechtel’s (2009) words—look not only down, but also up and around.
Consequently, we support Ward and Clack’s (2019) suggestion that the RDoC units may provide a helpful structure to support multiscale description, but suggest that investors add to this heuristic structure in a manner that prompts consideration of the situational, developmental, historical, and cultural contexts in which the phenomenon occurs, fails to occur, or occurs in a different form. There is room here for flexibility in response to the theoretical orientations of the explainers, for example, particular self-concepts and/or sociocultural variables are very reasonable things to incorporate into the descriptions when relevant. Table 1 gives a hypothetical example of pluralistic multiscale modeling of the clinical phenomenon of hypervigilance.
Hypothetical Example of Multiscale Description Looking at the Phenomenon of Hypervigilance
Note. In actual practice this description would be a lot more detailed, and based on detailed literature review.
Endpoint of phase 2b
Once the multiscale models for each constituent clinical phenomena have been collected, the PC is seen to be complete and ready to serve as a pragmatically defined explanandum in the explanatory phase. By this stage, PCs should: (a) be composed of parts that reliably correlate, or better yet, parts that have a longitudinally or experimentally evidenced directional relationship; (b) be more (mereologically) simple than current diagnostic constructs, yet much more richly described (in terms of constitution); (c) be thought to play a role within one or more recognized disorders (not necessarily DSM-recognized).
Further to these requirements, ideal PCs may be highly circular in their organization (Fuchs, 2017)—for example, they already have a simple form of self-maintenance. It is this causal structure that represents work against the self-maintenance and adaptivity of the individual (Nielsen & Ward, 2019). This causal structure may not be present at the PC level, but instead may emerge as investigators cycle back and develop explanations of other PC structures in the problem space. Figure 1 offers a visualization of a PC structure.
Visualization of a PC
Phase 3: Explain and evaluate
Explain (3a)
The task at this point is to make an inference to a constitutional explanation of the PC. Effectively, the aim is to utilize the rich understanding of the constitution of the clinical phenomena (developed in phase 2) to infer explanations of the relational phenomena, thus explaining the internal structure of the PC.
The inference here is an abductive one. At its simplest, abductive inference is the postulation/recognition of some state of the world that serves to make another state of the world (the explanandum) less surprising (Haig, 2014). A vital point here is that investigators should be looking for both potential causal links between the constituent clinical phenomena, but also be looking for potential constitutional overlap (analogous to a latent variable approach). Within the explanations generated, no scale of analysis should be given a priori preference.
As a simple example, we will look at two hypothetical relationships and their possible explanation. Let’s say we have mapped out the problem space of “depression” (phase 1), and from this mapping we have isolated a PC containing three phenomena: high stress, sudden waking during sleep, and weight gain (phase 2a). At phase 2b, we have described these three phenomena at multiple scales. When describing “high stress” and “night-time waking” at a biological scale, we may notice that both of these phenomena commonly involve some kind of dysregulation of the cortisol system. Investigating this link further—through the review of relevant literature and/or empirical exploration—we may find it plausible that this issue with the cortisol system is underlying both phenomena. This then is an explanation of the relationship between stress and night-time waking by noting constitutive overlap—that is, both phenomena are underpinned by the same mechanism—rather than the relationship being causal. Comparatively, when looking at our collected descriptions of weight gain and stress, we may note that a common reaction to stress is “stress-snacking,” which seems to provide temporary relief but also causes weight gain. Weight gain, in turn, is often associated with fear of negative social evaluation, plausibly increasing stress. In explaining this relation then, we may propose a mutually reinforcing causal relationship. In actual practice, explanations would be more detailed and rigorous than in this example.
Evaluate (3b)
The task at this final stage is to evaluate the explanations generated at phase 3a. 8 There may well be multiple possible explanations for each relational phenomenon, so the job here is to choose the best ones. This selection should be made on the basis of the competing explanation’s epistemic values. Epistemic values are qualities of explanations that we value because they make the explanations more likely to be accurate (Haig, 2014). Epistemic values include: external coherence (whether the explanation fits well with our other systems of knowledge, e.g., biological plausibility), internal coherence (that the explanations postulated don’t conflict with each other or their own internal postulations), predictive utility (that the explanation produces predictions that withstand empirical testing—pointing to the important role of experimental work in this phase), and parsimony (which can be thought of as simplicity divided by the scope of the explanation).
As part of the evaluation process, investigators are free to cycle back to an earlier phase (see the blue arrows on the left of Figure 2). Making different choices along the way and then comparing resulting problem spaces, descriptions, or explanations allows for continual refinement of the outputs at each stage. Returning to phase 1 allows for refining of the problem space. This may include the removal or addition of clinical phenomena, the merging or splitting of phenomena, or even the splitting or lumping of entire problem spaces as evidence emerges. For example, as evidence is uncovered, investigators may decide that it is better to split the classic depression-rated phenomenon of “anhedonia” into separate phenomena (e.g., avolition towards pleasurable activities, diminished experience of pleasure when the activities are engaged in, reduced focus on pleasure when remembering activities that were enjoyed at the time). Alternatively, if initially splitting the phenomena up this way, investigators may decide these phenomena occur so regularly together that it is better to think of them as one phenomenon (see Ward & Clack, 2019, for further discussion of this example).
Schematic Representation of the RAP Process
Returning to phase 2b or 3a will produce different explanations of the same PC which can then be compared. Repeatedly cycling back to 2a and selecting a different set of phenomena will (over time and with different research groups) produce a network of models that explain overlapping PCs. An important idea here is that these overlapping explanations will eventually populate the problem space with a rich understanding of the relationships between the clinical phenomena that constitute the disorder under investigation.
Summary
To summarize the RAP, there are three overall phases (see Figure 2). First, investigators list all the behavioral and phenomenological (i.e., clinical) phenomena present within the problem space and map out their relations. Second, investigators artificially select two to four phenomena and idealize them as a small system which we have referred to as a phenomena complex (PC). Each constituent clinical phenomenon is then richly described using a selection of friendly models from across different scales of analysis. Third, investigators infer the nature of the relations observed between the constituent phenomena (i.e., the relational phenomena), before evaluating these explanations. At any stage necessary, investigators are free to return to an earlier phase to improve and refine their explanations, descriptions, or mapping of the problem space. As a corpus of explanations of PC structures develops, this knowledge will represent greater understanding of the internal causal structure of the problem space.
Limitations and counter-arguments
Generalizability
The main limitation of our proposal is that, as we have argued for it here, it is grounded in a particular view of what mental disorders are (i.e., complex multiscale entities with fuzzy boundaries and circular process structures). We have presented the RAP as an idea worth pursuing across psychopathology, but the reality is that, while we tentatively have different kinds of mental disorders (depression, anxiety, schizophrenia, etc.), these kinds likely fall into broader types (e.g., organic diseases, socially constructed disorders, dysfunctional variances on a normally distributed trait, etc.; Haslam, 2002, 2014; Kendler et al., 2011). It is likely that the RAP is more relevant for certain types of mental disorders than others.
Those disorders that roughly seem to be more “organic” in nature, such as schizophrenia or autism spectrum disorder, may not sit as well with the RAP approach, because the descriptive phases are oriented around the behavioral and symptomatic level. However, regarding the association with dysfunction at the level of the person in their environment—which, in our view, is essential to the disordered status—some degree of this circular process structure seems necessarily present (Nielsen & Ward, 2019). Because of this, even if the RAP approach does not assist with uncovering the underlying causes of these more “organic” syndromes, it may still be useful in analyzing how these disorders are associated with distress and dysfunction. Such a focus will thereby still be fruitful for the development of management strategies.
Is this pluralism sustainable?
The externalism and multiscale pluralism prescribed at phase 2b presents a potential challenge. Without limits, this risks unsustainable and unproductive holism (Nassir Ghaemi, 2009). It is simply not an achievable task to model a phenomenon at every conceivable scale from chemical through anatomical, to economic and cultural. The use of pluralistic modeling in this phase then, seems to require the instantiation of reasonable limits to make it sustainable. Without such limits it could be claimed that investigators will not know when to stop describing the constituent phenomena and when to move on to inferring an explanation of the PC. But which scales should investigators restrict themselves to? This is a reasonable concern and providing a definitive solution to this issue is an area of potential future development for the RAP framework.
Put simply, the issue here is “how much detail is required before moving to phase 3?” When presenting his distributed theory of mechanistic explanation from which we—along with Ward and Clack (2019)—draw this pluralistic model of phenomena description, Hochstein (2016b) presents the limits of nonredundancy and relevance. Effectively, if an investigator is considering adding a model to her set of descriptors, she should make sure it makes a meaningful contribution to her understanding of the phenomenon, tracking novel difference makers to the occurrence or form of the explanandum (see Craver & Kaplan, 2018, for further development of these limits).
Further to these limits, pragmatic guidance can be found in the work of Potochnik (2010, 2017), who highlights the need to consider the wider purpose for seeking an explanation. Within the RAP, the purpose of richly describing the constituent phenomena is to inspire the creative exercise of abductive inference to an explanation of the PC (phase 3a). The number of models required to do so, and which particular scales of analysis will be fruitful, will change with every investigation and investigator. While the RAP represents the shift from phase 2b to phase 3 as a definitive step, the reality is that, in practice, this shift in the investigatory process will be iterative and gradual. Investigators are free to go backwards and forwards between phases 2b and 3 while exploring their theoretical speculations, adding to and removing descriptive models as required. Our general suggestion then is to start with three or four different models, going back and adding more or removing them as required. We accept that this is quite a rough and ready answer and welcome any suggestions as to a more definitive solution.
Conclusions and summary
A vital element in any explanatory endeavor is the selection and depiction of the thing you are trying to explain. DSM syndromes are likely too heterogenous to serve this role. SNWMs have great potential in their ability to map the wider problem space that a disorder reflects. Currently though, SNWMs seem too unstable and thin to act as good models of the explananda by themselves. Regarding the RDoC, we suggested that it will support the discovery of relevant (largely subpersonal) phenomena rather than explanations of disorder in toto. While knowledge of these phenomena will be vital, they are too distant to the disorder-as-a-whole to serve as ideal targets of explanation.
We have proposed the RAP as a complementary method to these approaches. With its focus on developing deep explanations of the relationships between clinical phenomena, the RAP is designed to fulfill a separate purpose to either the SNWM approach or the RDoC. The RAP is a metamethodological framework that conceives of its targets as PCs, so named in reference to Hoche’s (1912/1991) work within classification that argues for a similar shift to this middling level of complexity.
By focusing on PC structures, the RAP isolates explananda that are more manageable (and likely less heterogenous) than DSM syndromes or SNWMs, yet more directly relevant to the perpetuation of dysfunctional behavior than RDoC-derived targets. Our review suggested that RDoC is largely focused on uncovering dysfunctional neural mechanisms. Comparatively, the RAP is focused on uncovering the wider mechanisms of disorder in people’s lives. This makes the RAP a useful tool if researchers are interested in understanding the self-maintaining process structures of a mental disorder. If interest is in a different facet of mental disorder, then a different tool may be better suited.
We have developed the RAP with an intent for it to be used by individual theoreticians, interdisciplinary research teams, and potentially as a framework for encouraging co-ordination within the wider sciences of psychopathology. Our hope is that, further to its use, our presenting of this framework will begin a dialogue concerning how to co-ordinate our investigatory and explanatory efforts (Sullivan, 2017). Targeting at this level should facilitate the timely development of explanations as to the maintenance of mental disorder and, with some luck, more efficacious treatments.
Footnotes
Acknowledgements
The authors would like to thank Samuel Clack and Daniel Wegerhoff for their comments through drafting, and the rest of the EPC Lab at Victoria for their support. Thank you to Emma Ashcroft for coming up with the RAP acronym.
Declaration of conflicting interests
The authors declare that there is no conflict of interest.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
