R2. Who or what is the straw man?

Several commentators remark that our target article was not properly directed, in the sense that we attack an outdated straw man that nobody believes in anymore. For instance, Hur et al. suggest that we “use evidence against extreme reductionism and common causes to devalue clinical and translational neuroscience approaches – effectively throwing the baby out with the bathwater.” Similarly, Troisi states that the reductionist model as we defined it “is not illustrative of how biological explanations can improve our understanding of the origin of mental disorders.” Finally, Ward & Fisher argue that reductionism “is not seriously considered by contemporary cognitive neuroscientists, who agree that personal and subpersonal psychological levels of explanation provide unique and valuable insights into the mind.” The impression one gets from these remarks is that the actual research in biological approaches to clinical psychology and psychiatry is much more sophisticated than the simple search for a common biological cause of symptomatology, and does in fact already incorporate psychology and environment in integrative ways.

Surely, there exists research for which this picture is accurate, and we would in no way want to suggest that the crude form of reductionism we sketched in our target article is invariably entertained among researchers working on the biological factors involved in mental disorders. Still, it is hard to avoid the impression that neuroscience and genetics, in this respect, are Janus-faced. While the researchers who feature in the current discussion forum insist that nobody really believes in explanatory reductionism, at the very same time many of their colleagues are dominating the media with claims that clearly rest on the acceptance of the thesis that mental disorders are brain disorders in a literal and unqualified sense.

In our target article, we gave the example of NIMH leaders who, as late as 2015 and in full view of the absence of any strong evidence supporting their claim, felt confident declaring that mental disorders are brain disorders (Insel & Cuthbert Reference Insel and Cuthbert2015), and we discussed a number of other influential sources who are on the record as embracing full-fledged versions of explanatory reductionism. We think it is therefore rather remarkable that so many commentators feel confident in stating that this form of reductionism is a straw man. If this is really the case, why is it so easy to find examples of government-funded scientific organizations that communicate the message to the public that mental disorders are brain disorders? The Queensland Brain Institute, to give but one example, simply lists depression as a brain disorder on its website; and the first sentence under the header “What causes depression?” is “The neurobiology of depression is still poorly understood,” suggesting that the answer to the question, whatever it may be, must in fact turn on neurobiology.Footnote ¹ Similarly, in 2017, the Dutch National Institute of Public Health and the Environment published research concluding that one in every four citizens “suffers from a brain disorder.”Footnote ² The count, which was widely publicized, indiscriminately included all mental disorders for which data were available. Those who think this type of talk is limited to press releases and promotion material may consider some cutting-edge science. The most recent genetic work on depression, published in Nature Genetics, found that 44 genetic loci are significantly correlated with depression measures which, taken together, explain 1.9% of the variance in estimated depression liability (Wray et al. Reference Wray, Ripke, Mattheisen, Trzaskowski, Byrne and Abdellaoui2018). These results lead the authors of the otherwise excellent study to draw a stunning primary conclusion from the data: namely, that “major depression is a brain disorder.”

We are not cherry-picking. It is not at all hard to find these examples, as anyone can verify by searching the Internet for information on mental disorders. The notion that mental disorders are genetically encoded brain disorders is everywhere around us. It dominates the organization of research, it dominates teaching, and it dominates the media. The central problem is not even that the thesis is necessarily false – as we stated in our target article, in the future we may in fact witness the kinds of breakthroughs that would establish that mental disorders are brain disorders; this is, in our view, spectacularly unlikely but not impossible. The central problem is dogma: The reductionist hypothesis is not treated as a scientific hypothesis, but as an almost trivial fact.

Given that so many examples of outspoken reductionism exist, we feel hesitant to accept that explanatory reductionism is a straw man. Rather, we submit that the idea that mental disorders are brain disorders has somehow become a background assumption of our modern society – an undisputed member of the cabinet of scientifically respectable facts, on par with “the world is round” and “life developed through evolution.” Even though, as said, none of the commentators attempts to defend reductionism explicitly, reductionism as a background assumption clearly does shimmer through in several of the responses to our target article. Hur et al. state casually that “mental illness is undeniably based in brains and genes.” Redish, Kazinka, & Herman (Redish et al.) argue that “[b]ecause all behavior arises from brain function, neurobiology is obviously critical for understanding psychiatric phenomena.” Pernu declares that “[s]symptoms, however, are mere signs, marks of the underlying disorder or illness; and even in clinical practice, the ultimate aim is not just to remove the symptoms, but to cure the physiological condition they stem from.” One does not get the impression that these commentators, in their own view, are launching spectacular scientific hypotheses. If they looked at their statements in this way, the commentators would probably have felt compelled to provide arguments for these theses. Rather, they make the impression that they are stating the obvious.

Explanatory reductionism, however, is not obvious. It is not a fact but a hypothesis that mental disorders originate in the brain; it is not a fact but a hypothesis that there are genes “for” mental disorders; and it is not a fact but a hypothesis that finding out “what goes wrong in the brain” is a necessary condition for progress in the science of mental disorders. It is a realistic possibility that increased understanding of the neuroscience involved in mental disorders will in fact establish that they are not brain disorders. The network theory opens up a general and scientifically defensible prospect that shows how this could be the case, as we argued in our target article.

R3. Networks = notworks?

For some commentators, the network approach does not work, for various reasons and in various ways (Elbau, Binder, & Spoormaker [Elbau et al.]; Eronen & Bringmann; Hur et al.; Ward & Fischer). One returning theme concerns the supposed “thinness” of network models – in other words that, supposedly, the network framework “fails to provide a deeper explanation … of where those patterns come from” (Hur et al.). Ward & Fischer acknowledge the usefulness of a network approach as a model at the level of phenomenology, that can be employed to “detect patterns among symptoms,” but they argue that the approach does not represent “the mechanisms that constitute [symptoms].”

To analyze this situation, it is important to distinguish between a network as a theoretical model that aims to represent phenomena in the world (e.g., see Borsboom Reference Borsboom2017; Cramer et al. Reference Cramer, van Borkulo, Giltay, van der Maas, Kendler, Scheffer and Borsboom2016) and as a statistical model that one estimates on empirical data (e.g., see Epskamp & Fried Reference Epskamp and Fried2018; Van Borkulo et al. Reference Van Borkulo, Borsboom, Epskamp, Blanken, Boschloo, Schoevers and Waldorp2014). If the commentators, when noting the thinness of symptom networks, are referring to the fact that the statistical model does not churn out a set of empirical, mechanistic facts about the world – for example, the connection between depressed mood and suicidal thoughts is primarily driven by a lack of self-worth – then we fully agree. Statistical models contain variables connected by parameters. A model fitting exercise will return parameter estimates that describe relations between variables, but it will not generate mechanistic explanations for why these relations exist. Statistical models, in this respect, indeed should be seen as delivering evidence on the presence of causal processes rather than on the nature of these processes.

However, the conditional argument set up in our paper does not concern such statistical models; it instead concerns network theories that do address the nature of psychopathology – namely, that psychopathological conditions are alternative stable states of complex networks formed by symptoms and interactions between symptoms (Borsboom Reference Borsboom2017; Cramer et al. Reference Cramer, van Borkulo, Giltay, van der Maas, Kendler, Scheffer and Borsboom2016). For instance, Cramer et al. (Reference Cramer, van Borkulo, Giltay, van der Maas, Kendler, Scheffer and Borsboom2016) showed that, for major depression, we can fairly precisely explain how an episode of major depression can come about, namely by the dynamics in a strongly connected network in which symptoms cause and maintain one another. The connections between symptoms in this theoretical model do lend themselves well for coupling with real-world mechanisms, including biological ones (see also Ross). For example, the theoretical connection between insomnia and fatigue is supported by a real-world mechanism, namely the homeostatic processes instantiated by the biological clock; the connection between hallucinations and anxiety likely involves fear mechanisms deploying the amygdala (Hur et al.); the connection between gambling and money shortage undoubtedly deploys the mechanical specifications of the Roulette table. The network theory of psychopathology holds that, in psychopathology, these real-world mechanisms reinforce each other to such a degree that a symptom network becomes self-sustaining (Borsboom Reference Borsboom2017). Importantly, however, this theory is not necessarily tied to statistical network models; this can be seen from the fact that one can specify the theoretical idea independently of statistical modeling schemes whatsoever: it could, for instance, also be specified using deterministic models (e.g., dynamical systems models). Current statistical network modeling techniques should therefore not be seen as instantiations of this theory but as statistical devices that chart the structure of symptom-symptom connectivity (e.g., see Boschloo et al. Reference Boschloo, van Borkulo, Rhemtulla, Keyes, Borsboom and Schoevers2015); the hope in using such models is not that these analyses will magically return a substantive theory all by themselves, but that they will uncover patterns of network connectivity between symptoms that can subsequently be used to inform theoretical models. However, statistical models are not themselves substantive theories, and the network approach is not defined by or limited to one particular type of statistical model.

We also can see this clearly from the fact that the relation between network theory and network model is not one-to-one: Network theories may map into statistical network models, but they do not necessarily do so. Recently, for instance, it was shown that a Curie-Weiss model – a special case of the Ising model in which all connections have the same strength – is statistically equivalent to the Rasch model (an important latent variable model; Epskamp et al. Reference Epskamp, Maris, Waldorp, Borsboom, Irwing, Hughes and Booth2018; Marsman et al. Reference Marsman, Borsboom, Kruis, Epskamp, van Bork, Waldorp, van der Maas and Maris2017). This means that one can use latent variable models to specify statistical consequences of network theories, and one can use network models to specify statistical consequences of latent variable theories. As one example, it has been recently suggested that network models are particularly well suited to identify the dimensionality of the latent factor space (and even outperform factor analysis in this respect; Golino & Epskamp Reference Golino and Epskamp2017; Golino & Demetriou Reference Golino and Demetriou2017). These equivalence theorems also explain why, as Jayawickreme, Rasmussen, Karasz, Verkuilen, & Jayawickreme [Jayawickreme et al.] note, network models and factor models can lead to very similar results; when network models are applied to data of questionnaires that have been developed using psychometric models (e.g., item response theory of factor analysis), they will mirror the structure of these models (i.e., every factor will produce a clique in the network). In view of this, Jayawickreme et al. make the important methodological point that one should be careful in selecting which variables to include in network models if the findings are to inform network theories; for instance, if one uses sets of items that are selected to conform to a unidimensional factor model (in practice, this typically results in sets of items that all correlate with about the same strength), one's network analysis will be predictably homogeneous (i.e., all nodes will feature roughly equal interconnections so that all nodes are equally central; Marsman et al. Reference Marsman, Borsboom, Kruis, Epskamp, van Bork, Waldorp, van der Maas and Maris2017). However, this may not always be informative with respect to the issue of how disorders are structured.

Therefore, we should take care in distinguishing network theories from network models. We think that complaints on the thinness of networks, or the idea that they are “merely phenomenological,” are relevant to statistical network models, which indeed only represent conditional associations between symptoms. As Van Loo & Romeijn argue, such models can be interpreted in an instrumentalist fashion. And if one engages in such an interpretation, it is indeed somewhat hard to see exactly what is meant in a conditional argument that proceeds from the subjunctive conditional “if the network theory of psychopathology were true,” because truth is a property of semantically interpreted theories, not of statistical models per se. However, such an interpretation is not suited to network theories, interpreted as formalized theories that represent actual problems people have (i.e., the problems encoded as symptoms in diagnostic manuals such as DSM-IV) and causal processes by means of which these problems maintain, promote, or inhibit each other. Interpreted in this way, network models are not susceptible to arguments that they are too thin in the sense of being merely about phenomenology. Symptoms, and relations between them, allow any depth of theoretical explanation one might envisage, including those ventured in learning theories, cognitive schema accounts, psychodynamic approaches, neuroscientific theories, and, ultimately, evolutionary theories (O & Brüne). For this reason, network theories do not rule out hierarchical explanatory accounts, as Oller suggests. However, what is very clear when one considers plausible candidate accounts that connect symptoms, is that it is spectacularly unlikely that all of these relations will be amenable to a neuroscientific account, as they will require reference to the world external to the body as well as the (partly culturally loaded) content of mental states, of which several commentators provided important examples (e.g., see Crafa & Nagel; Field et al.; Jayawickreme et al.). Network theories therefore almost by necessity lead to pluralism, as different symptoms (and relations between symptoms) require analyses at different levels.

R4. Intentionality, multiple realizability, and extended systems

Many commentators agree with the claim that symptoms often bear intentional content and that this constitutes a challenge for reductionists. As Ross rightly remarks, intentionality is a general obstacle to reductionist explanation, and not one that is specific for the explanation of mental disorder. Nevertheless, intentionality pops up as a specific challenge for reductionist explanation of mental disorders when it is claimed (as we do) that mental disorders should be understood in terms of causally interrelated symptoms. In other words, it is precisely the network approach to mental disorders that brings intentionality to the fore as a challenge for reductionism. Eronen & Bringmann are similarly correct to claim that “defending anti-reductionism does not require taking a network perspective,” but this has no consequences for our argument because our claim rests on the sufficiency, but not the necessity, of the premise that network theory correctly describes psychopathology: We argue that the network approach constitutes a specific obstacle to reductionist explanation, because it locates the causal nexus at the level of symptoms, which are often inherently intentional. This has little to do with whether other obstacles to reductionism also exist (which is undeniably the case).

To avoid this difficulty, Eronen & Bringmann and Van Loo & Romeijn both bring up the idea that one could adopt an instrumentalist understanding of network models. Doing so would discharge us of the obligation to argue for the causal relevance of intentional symptoms: We could just show that modeling mental disorders as causally related symptoms with intentional content is a useful and predictive explanatory tool. However, we believe instrumentalism is not a serious option here. Adopting instrumentalism about models implies that the models one proposes cannot be said to be either true or false – and this means that what we attempt in the target article would not even be possible: One cannot develop a conditional argument on the truth of a model that cannot be said to be either true or false. More importantly, network models get their “bite” precisely from their ambition to provide an account of what mental disorders are; therefore, retreating to an instrumentalist position would render the approach toothless.

The network approach suggests that the intentional content itself figures in causal explanation. For example, Field et al. mention that addicts who believe that they suffer from a chronic disease are less likely to recover than those who believe they suffer from an unhealthy habit. Their implicit suggestion is that the specific content of their beliefs about their addiction causally affects their chance of recovery. But how should we understand this? According to Eronen & Bringmann, the hard-core reductionist could “accept the importance of intentional contents and their meaningful relationships, but nevertheless argue that the real causal work is done by brain states.” To show that this is a dead end, Eronen & Bringmann argue that we might have to resort to an interventionist understanding of causation (a suggestion we already hinted at in the original article), and subsequently propose to start testing the hypothesis that interventions on the level of intentional content indeed have causal effects, a point also made by Müller.

While the emphasis on interventions is certainly worthwhile, Eronen & Bringmann seem to overlook the fact that we already have a substantial amount of empirical evidence for the thesis that manipulating the content of mental states (fears, beliefs) has content-specific causal effects in psychopathology. In particular, the research program surrounding cognitive behavioral therapy (CBT) offers a rich trove of findings that evaluate exactly this thesis. CBT, which is probably the most extensively researched and empirically supported psychological treatment in existence (Hofmann et al. Reference Hoffmann, Asnaani, Vonk, Sawyer and Fang2012), is based on the premise that psychopathology is at least in part caused and maintained by maladaptive cognitions (general schemas about the self and the world): intentional states par excellence. CBT explicitly targets the content of these cognitive representations through techniques such as cognitive restructuring (Beck Reference Beck1970; Ellis Reference Ellis1962).

Panic Disorder (PD) is an important example of a case where interventions on the content of cognitive states are explicitly used. In PD, intentional states play a central role. Primary theories on the etiology of PD hold that panic attacks arise from a feedback loop between physiological arousal (e.g., increased heart rate) and a cognitive schema that specifies a particular intentional content, that is, about the arousal (e.g., “My heart is racing, so I must be having a heart attack”). The interpretation of the bodily signal as heralding an impending catastrophe then reinforces the arousal itself, which in turn reinforces the cognitive representation, resulting in a runaway feedback process that culminates in a panic attack (Clark Reference Clark1986). CBT interventions in use involve a variety of techniques aimed at modifying these cognitive representations, for instance through controlled exposure, which teaches the person that the impending catastrophe in fact need not happen, and by training the person to replace the cognitive representation with another one. (For example, CBT may teach the person to consider at least one other hypothesis, apart from “I am having a heart attack,” that could explain the physiological arousal when it arises.) Such interventions have been shown to be effective (e.g., Hofmann et al. Reference Hoffmann, Asnaani, Vonk, Sawyer and Fang2012; Mitte Reference Mitte2005). Similar techniques are used in a variety of other cases ranging from eating disorders to depression and from somatoform disorders to psychosis (see Hoffmann et al. [Reference Hoffmann, Asnaani, Vonk, Sawyer and Fang2012] for an overview).

Thus, the evidence for the causal relevance of intentional states and the effectiveness of intervening on these states is overwhelming. However, according to Eronen & Bringmann, this would still leave open a reductionist rebuttal: The causal relevance of symptoms-with-content could still be brought about by neural or biological causes. Here we can point to the argument developed in Ross’s comment, which aims to show that the fact that neural processes are causally involved does not guarantee a successful reduction. What is needed for a reduction to succeed is to show that these neural processes can explain why a person has a belief with a certain content and not with another.

One of the arguments we bring forward for the claim that intentionality blocks reductionism, is that it is widely held that mental states with intentional content are multiply realizable and therefore not type identical with brain states. Some of the commentators argue that this is not convincing. Pernu, for example, claims that many apparent cases of multiple realizability can be dissolved by “kind splitting.” This solution is based on the idea that one should analyze mental states at a level that is just as “fine-grained” as the level at which one analyzes physical states (Bechtel & Mundale Reference Bechtel and Mundale1999) – and that if one insists on a fine-grained analysis of physical states, one might have to conclude that two people who believe “my neighbor is a secret agent for the CIA” on the basis of different underlying physical processes, do not actually have the same belief at all (Polger Reference Polger2009).

However, this response makes sense only if one already assumes that mental states such as beliefs and fears are, in fact, completely constituted by physical states in the brain. As said, it might be that in the future this will turn out to be true, but at present this is still very much an open issue. In fact, there is a growing chorus of voices in contemporary philosophy of mind arguing that we should not think about intentional states in this way. According to these approaches, what makes John's belief about “his neighbor being a secret agent for the CIA” is not that he is in a certain brain state, but that some coherent set of counterfactuals is true of him (Baker Reference Baker1995). His belief is characterized through a set of counterfactual conditionals such as: If the neighbor were to start a friendly talk, John would respond nervously; if John passed the CIA headquarters, he would expect his neighbor to be there; and so on. What makes a state a state with a certain specific content, is thus thought to depend on the occurrence of a set of meaningfully related phenomena, embedded in a certain context.

Building on the by now almost mainstream assumption (explicitly or implicitly endorsed by several commentators: Crafa & Nagel; Field et al.; Jayawickreme et al.) that cognition is embedded and extended (see Clark & Chalmers Reference Clark and Chalmers1998; Menary Reference Menary2010), the content of one's mental state is thought to be at least partly determined by one's environment. This raises another barrier to purely biological explanations of mental disorders, for two reasons: First, the simple but profoundly important and easily forgotten fact that the environment is not in the brain. Second, the more complicated fact that the relevant features of the environment are unlikely to admit an effective characterization in terms of their purely physical features, more or less for the same reason that behaviorist analyses hardly ever managed to characterize classes of stimuli in purely physical terms, without indirect reference to mental states. The physical constitution of the Roulette table certainly matters to its function, but that physical constitution realizes a gambling apparatus only from the point of view of the human being, not from the point of view of physics; it is, in other words, unlikely that the Roulette table will ever become a “kind” of physics (or of neuroscience; Fodor Reference Fodor1974). Perhaps for this reason, the relevant environmental features in psychopathology are themselves almost always multiply realizable. One need only consider the fact that debts and poverty are crucial factors in the maintenance, and probably also in the genesis, of several disorders; these properties come down to a lack of money, and money is the quintessential example of a multiply realizable phenomenon.

Pernu, however, holds that multiple realizability does not offer sufficient grounds to reject reductionism, and that a more successful argument for the failure of reductionist explanation would be if something like inverse multiple realizability would be true. So, if it were true that one and the same brain state could realize different symptoms in different circumstances, this would show that even an ideally complete neural description of a symptom would not allow us to identify which symptom it is. Pernu argues that the truth of inverse multiple realizability would undermine even the basic supervenience thesis that he ascribes to us, turning such an argument into what he calls a “Pyrrhic victory.” However, in our reference to supervenience in the target article, we merely suggested that “the strongest viable position that is still available would be non-reductive materialism along the lines of the supervenience thesis in the philosophy of mind,” that is the thesis defended by Kim (Reference Kim1982; Reference Kim1984). At the same time, as becomes clear in our argument, any form of non-reductive materialism that our position allows for would clearly need to take a broader spectrum of physical states into account than mere brain states (including, for example, states of the environment). In fact, we consider it highly plausible that something like inverse multiple realizability for brain states is correct, and thus that supervenience about neural states is wrong. As discussed above, Ross (who suggests that our tolerance for supervenience is an unwarranted concession – about which he might be right) convincingly argues that the context is often what makes a disorder a disorder.

This means that it might not always be possible, even in science-fiction scenarios in which one would have access to the complete physical description of brain states, to determine purely on the basis of these brain states whether a person has a certain symptom or not. Indeed, such a conclusion would seem correct for many of the symptoms listed in diagnostic manuals such as DSM-IV. To start with, symptoms that involve truth conditions in the external world would seem to defy supervenience of psychological states on neural states more or less by definition. Elizabeth and Bob may both believe that they are persecuted by the CIA, and this belief may be instantiated in the exact same way in their brains. Depending on the external circumstances, however, this belief may count as a symptom or not – for instance, when the belief is veridical for Elizabeth (who is actually a Russian spy) but finds no grounding in reality for Bob.

A second important class of symptoms that is likely to violate supervenience with respect to neural states involves the many symptoms that code behaviors as “excessive” or “out of proportion” with respect to the circumstances: For example, persistent handwashing is not a symptom of Obsessive Compulsive Disorder (OCD) in a situation where there is a nontrivial risk of infection.

A third class of symptoms that will likely defy supervenience with respect to the brain refers to social norms. Two people may both exhibit the same level of systematic violence, with all of the same neural states that come with it, but if one is a professional boxer while the other is a choir boy, we may justifiably consider the behavior as a sign of psychopathology in the latter but not the former.

Fourth, symptoms that explicitly involve relations with the environment (e.g., having debts, having been in contact with law enforcement, etc.) will not satisfy supervenience with respect to brain states because in these cases what counts is what happened to the person rather than what brain states the person has.

Fifth, a large class of symptoms is defined in terms of a specific trigger in the environment; for instance, the fear response of a phobic who sees a spider may be physiologically indistinguishable of the panic attack in a patient with panic disorder. What makes the difference between these conditions could lie in how they are triggered by the environment.

It should be noted that the above examples putatively refute supervenience of symptoms with respect to brain states, but not with respect to the physical world at large. However, if supervenience of the mental on the physical is meant as supervenience on the physical state of the world in its totality, this would engender a form of non-reductive materialism with even fewer teeth than the thesis that mental states supervene on brain states.

So how should we deal with accounting for the intentional content of symptoms? Slors, Francken, & Strijbos [Slors et al.] suggest that making sense of intentional content requires more than “just” assumptions of rationality. Instead, they argue that what is needed is “a range of further, non-rationalizing interpretive strategies, including simulation, empathizing, and using clinical knowledge and experience.” They rightly note that, for example, delusions are precisely delusions insofar as they do not manifest the kind of rational relations to other states that we observe in regular beliefs. For example, whereas beliefs are at least to some extent susceptible to correcting evidence, delusions are generally not. In the target article, we acknowledge that pathological intentional states differ from non-pathological ones in terms of rationality. We tried to account for the inherent irrationality of psychopathology by pointing out that psychopathology stands out as, so to say, an irrational figure against a “minimally rational” background. Slors et al. turn the picture around, and speak of the search for “islands of reason in a sea of confusion.” Which picture best reflects the “rational status” of someone suffering from psychopathological symptoms will clearly depend on the individual case; however, we believe our notion of rationality is much more minimal and bounded than Slors et al. have taken it to be. Our point in speaking of a background of rationality was to indicate that almost all persons suffering from symptoms still take part in the social practice of exchanging reasons: They take part in conversations about treatment options or what they want for dinner, and often even participate in complex practices such as “talking therapies” like cognitive behavioural therapy. It is in this sense that we claim that almost all persons suffering from psychopathology, still manifest a substantial amount of background rationality. So, whereas we fully agree that “making sense” requires a wider interpretative spectrum than just naive assumptions of full rationality, we do not think our arguments offer such a restricted view on interpretation.

R5. Beam me up, Scotty!

As Hur et al. state, “Clinical and translational neuroscience has historically been oversold and under-delivered.” These are true words, except that overselling neuroscience is not a thing of the past but of the present as well. Even if the prospects for a neuroscientific analysis of mental disorders seem gloomy upon a sober evaluation of the evidence, many researchers keep pointing to the future, presumably under the assumption that, sooner or later, the biological ship must sail in and their utopic Reductionia will materialize.

For example, Elbau et al. keep their hopes up and state that “there is not yet reason to abandon the effort of biological reduction that has been most fruitful in all other fields of medicine.” Oller thinks that “recognition of the abstract nature of mental/linguistic events does not diminish the importance of the neural impulses that, at another level, form the infrastructure for those events” (emphasis Oller's). And Eronen & Bringmann suggest that the reductionist could maintain that “it is pragmatically useful to describe and predict human behavior in terms of beliefs and desires, but this is consistent with the idea that the real causes of behavior are biological or neural.” Perhaps the most adventurous commentators are Perlovsky, who envisages a “physics of mind,” and Pernu, who hopes for an entirely new categorization scheme which, in contrast to the current set of disorders, would map neatly onto as yet unknown neurobiological explanations: “Therefore, the idea is that we should aim to abandon the superficial DSM classifications and replace them with more valid classifications based on physiological aetiologies.” As a result, Pernu argues, “our current understanding of mental disorders – ‘folk psychiatry’ – will be fundamentally transformed, and the symptomatically defined notions of mental disorders (M) will give way to new notions, aligned with their neural-level realisers (M1 and M2).”

Of course, one can hope for such transformations to materialize, and one is free to pursue research that might deliver such beautifully delineated categories. We do insist, however, that we all keep our eyes on the ball. In the current scheme of things, explanatory reductionism is a remote possibility, not a realistic research target. We do not have biomarkers that are sufficiently reliable and predictive for diagnostic use. We have not identified genes that are specific to disorders and explain an appreciable amount of variance. We have not obtained insight into pathogenetic pathways in the brain that are sufficiently secure to inform treatment. If anything, we should wonder why the massive investments in research, that should have uncovered these factors, have not pushed back the prevalence of common mental disorders by a single percentage point. Of course, everybody wants a penicillin for psychopathology, but we do not do anyone a favor by supposing that we are almost there, or even on the way. Therefore, despite the strong scientific image that is built up by brain scanners and genome sequencing machines, the situation as sketched by Insel and Cuthbert (Reference Insel and Cuthbert2015), for example, and echoed by Pernu should not be understood as science but as science fiction.

If the role of biology is not to teleport us into explanatory reductionism, what exactly is it? Ward & Fisher aim in a different and, in our view, more plausible direction, stating that “the primary causes of mental disorders could be biological in nature without being disease entities.” Their idea is that, while network models may identify the relevant connections between symptoms, “mechanistic models shed light on the way different symptoms exert their effects and influence each other” — and some of these mechanistic models may well be neurobiological in nature. This is certainly true and can be easily seen from some of the prosaic examples we discussed: Disruptions in the sleep-wake cycle cause concentration problems and fatigue; lack of appetite causes weight loss and lack of energy; prolonged use of drugs often causes tolerance and dependence. There is no doubt whatsoever that these mechanisms are grounded in neurobiology, and that the explanation of why one network structure obtains rather than another will have to rely on such neurobiology as well.

An important question, however, is how far one wants to move down this explanatory ladder and whether it would make equal sense to do so for different symptom-symptom connections. The cases of insomnia → fatigue, lack of appetite → weight loss, and substance use → tolerance are obvious enough, but it is no coincidence that these particular connections do not typically invoke the content of mental states. For connections that do rest on the content of mental states, it is much more difficult to see how they are to be explained through neurobiological mechanisms. For instance, John and Jane are both depressed and convinced they are a burden to their environment; more specifically, John thinks he makes his family unhappy, while Jane feels that her friends despise her behind her back. John and Jane both have feelings of worthlessness, but the content of their ideations differs, as will their brain states. How could information on those states possibly help us understand why John and Jane have the feelings and thoughts they have, or in what sense they both instantiate feelings of worthlessness despite the different contents and neural realizations? Similarly, what about the distinct neural mechanisms that must underlie compulsive hand-washing (obviously engaging the motor cortex) versus checking locks (probably engaging the visual cortex)? What about gambling addicts who are preoccupied by Roulette tables versus those who cannot help thinking about slot machines? Does anybody seriously expect to be able to identify these phenomena as being feelings of worthlessness, compulsive behaviors, or preoccupations with gambling, and to successfully differentiate them from each other and group them in the right categories, merely by looking at the brain?

One can hope for advances in mind-reading neuroscience that would allow one to detect suicidal ideations without having to ask the person what they think. But even if such science-fiction devices were to obtain, it would seem that primacy for identifying the content of mental states will remain with the intentional level of description rather than with the neurological one. One is hard-pressed to accept a situation in which the mind-reading machine detects the presence of persistent suicidal ideation, whereas one subjectively experiences only persistent thinking about how to fix a leaking tap. The success of any mind-reading machine will have to be measured against the ordinary folk psychological level of intentional description (after all, it's supposed to be reading the mind). In that sense, the intentional level will always have epistemic priority – at best, a mind-reading machine might become exceptionally well calibrated to that intentional level. As a result, in cases that involve the content of mental states, whatever we expect to find at the level of neurobiology must be epistemically slaved to the mental states we already identify at the intentional level. If this is correct, we can never expect the identification of the relevant mental states and behavioral patterns to be successfully executed at the level of neuroscience itself, and this means that whatever neuroscientific explanation we will get is going to be parasitic on “folk psychology” rather than an alternative to it.

R6. A little less observation, a little more experiment

Müller contests the testability of network models, stating that: “Empirical network approaches simply purport testability by bioinformatic large data analysis.” Although network models can imply testable constraints on conditional independence relations in the data, we agree that in most applications of statistical network models the results should be taken as exploratory, and that it is important to develop ways of critically testing complexity approaches. The question of how exactly network theories should be empirically interrogated is, however, not trivial; the same holds for the question of how to decide whether a given disorder is amenable to a network theory. As Tabb suggests, the assumption that a network model is broadly correct may not hold for all mental disorders (see also Fried & Cramer Reference Fried and Cramer2017), and as such, it is important to develop ways of deciding which approach is most likely to be fruitful; for instance, whether a common cause model or a network model would be more appropriate. Simply pitting the statistical models against one another is difficult because that different explanatory models can have very similar consequences at the level of correlations in an observational dataset (Jayawickreme et al.). How should we then tackle the challenge of picking the most suitable candidate model for a given disorder?

One potentially fruitful strategy is increased reliance on experimental instead of observational data. In clinical psychology and psychiatry, network analyses have frequently relied on observational data in which no single variable is manipulated, for example, data from major population surveys in which a large number of people are asked once about the presence/absence or severity of psychopathological symptoms (for an overview, see Fried et al. Reference Fried, van Borkulo, Cramer, Boschloo, Schoevers and Borsboom2017). While useful in many respects (e.g., as a means to obtain prevalence and comorbidity estimates, or to develop hypotheses on network structure), such designs are limited when it comes to understanding the nature of mental disorders. We can fit ever more complex statistical models to these data, but given the data-driven, non-confirmatory nature of current network models, and their possible statistical equivalence with latent variable models, this may not be the most productive way of deciding between different generative theories of psychopathology. We emphasize that this is so regardless of whether or not one includes physiological and neuroscientific data into the network model, except in the rare cases where one has a very strong argument for the etiological primacy of factors measured at such a level. A biological correlate in principle remains just that: a correlate.

Experiments might, however, offer fruitful avenues for distinguishing between different frameworks. Suppose, for example, that a researcher is able to show in the lab that people, somehow manipulated to have depressed mood, subsequently suffer from more thoughts of death than people whose mood was not manipulated. That is, the researcher has experimentally revealed a connection between two symptoms of major depression. This finding would be consistent with a full network perspective (as advocated in the target article) and with a hybrid perspective in which both a (local) common cause and network dynamics play a role (Fried & Cramer Reference Fried and Cramer2017). However, importantly, such findings would be inconsistent with a common cause conceptualization of major depression (and a fortiori with the reductionist interpretation of that common cause). In a common cause model, symptoms are merely effects of a latent condition, and manipulation of effects does not change their causes. Just as we cannot manipulate the behavior of one thermometer to induce changes in another – only the common cause temperature can change readings on both thermometers – so it does not make sense that symptoms can cause one another if their covariance is truly due to an underlying abnormality – whether it be biological or psychological (Nephew, Febo, & Santos [Nephew et al.]) – that causes the overt symptomatology of psychopathology. Experimental data in which symptoms are locally manipulated thus should be able to shed light on which disorders are amenable to network theory and which are not. This may also illuminate whether current hierarchical factor models of the covariance between disorders (Kotov et al. Reference Kotov, Krueger, Watson, Achenbach, Althoff, Bagby, Brown, Carpenter, Caspi, Clark, Eaton, Forbes, Forbush, Goldberg, Hasin, Hyman, Ivanova, Lynam, Markon, Miller, Moffitt, Morey, Mullins-Sweatt, Ormel, Patrick, Regier, Rescorla, Ruggero, Samuel, Sellbom, Simms, Skodol, Slade, South, Tackett, Waldman, Waszczuk, Widiger, Wright and Zimmerman2017), which typically include putative common causes in the form of latent variables, reflect causal order or are merely descriptive, as Bornstein suggests.

However, current experimental work – of which the randomized controlled trial (RCT) is its most prominent example – can (and perhaps should) be modified in the coming years to better incorporate a non-reductionist, multifactorial perspective on mental disorders. We agree with Ioannidis that, for example, RCTs could (and perhaps should) be redesigned so that they better capture non-biological outcomes over a longer period of time. We add to these helpful suggestions the idea that RCTs may test the efficacy of an intervention that targets either a symptom (e.g., targeting the most central node insomnia with a sleep hygiene protocol) or a network connection (e.g., cognitive restructuring to weaken the connection between the symptoms depressed mood and suicidal thoughts); as a result of which one may be able to deduce whether network approaches, as a possible intervention strategy, are fruitful or not.

Including time series data on dynamics as dependent variables in the experimental design (Snippe et al. Reference Snippe, Viechtbauer, Geschwind, Klippel, de Jonge and Wichers2017; Wichers et al. Reference Wichers and Groot2016) and/or modeling experimental manipulations as treatment factors in a statistical network models (Bekhuis et al. Reference Bekhuis, Schoevers, de Boer, Peen, Dekker, Van and Boschloo2018; Blanken et al. Reference Blanken, Van Der Zweerde, Van Straten, Van Someren, Borsboom and Lancee2019) could be two important starting points of such analyses. Another possible starting point would be to develop dynamical network models that could be used to model treatment effects through surgical (or not-so-surgical, i.e. “fat hand”) interventions in the model structure, perhaps analogous to how this is done in modern approaches in causal inference (Pearl Reference Pearl2000). This would involve extensions of the current causal modeling apparatus (most importantly the inclusion of nonlinearity and feedback), but there seems to be no principled reason why suitably adapted interventionist models of causation (Woodward Reference Woodward2003) could not be employed in, say, a dynamical systems framework. The often-used Ising model (van Borkulo et al. Reference Van Borkulo, Borsboom, Epskamp, Blanken, Boschloo, Schoevers and Waldorp2014) is a simple toy example that would seem to support models for causal interventions (Marsman et al. Reference Marsman, Borsboom, Kruis, Epskamp, van Bork, Waldorp, van der Maas and Maris2017); for example, if the Ising model is true, then intervening on one of its nodes may be modeled using the conditioning operation, which induces a new Ising model (Epskamp et al. Reference Epskamp, Maris, Waldorp, Borsboom, Irwing, Hughes and Booth2018, Eq. 30.6). However, many other approaches are conceivable, too. If we were able to simulate dynamical networks well enough to deliver reasonable predictions on what should happen under various interventions to the model, that would be a significant advance. Clearly, current models and theories cannot handle even such comparably easy questions, as Müller correctly suggests, and there is accordingly a huge space of opportunity for the development of network theories and analysis techniques that optimally model experimental interventions, in addition to the extensions of RCTs suggested by Ioannidis.

R7. A system of … what, actually?

Many commentators (Crafa & Nagel; Hens et al.; Jayawickreme et al.; Redish et al.) noted, correctly, that a systems perspective on mental disorders necessitates having a clear sense of what the elements of such a system are. In this respect, it is obvious that network researchers have not yet developed systematic approaches to assess this issue. In our target article we have focused on (DSM) symptoms as the key nodes of a psychopathological system (see also Borsboom Reference Borsboom2017), but it should be an important part of future research endeavors to investigate the extent to which this is the most optimal characterization. This matters, as Redish et al. point out, because one's conclusion based on network analysis (or any statistical analysis, for that matter) – for example, the most central node is X – heavily relies on the variables that feature as nodes in the network structure. When using DSM depression symptoms, the most central node could be depressed mood while, when using RDoC factors, the most central node might be perception and understanding of the self.

The notion of a “symptom” itself is also problematic. To many, it suggests the idea that a symptom is an indication of something else – a disease/disorder – which exists independently of the symptom. This may not be the right way to think about symptoms in the context of network models, however. In network theory, the phenomena that we call symptoms can also be guides to diagnosis, but they do not signify the presence of a distinct entity, as is often the case in medicine; rather, psychopathology symptoms should be seen as signalling the disturbance of the network as a whole, just like abnormal amounts of algae signal the alternative stable state of a turbid lake (Scheffer et al. Reference Scheffer, Carpenter, Foley, Folke and Walker2001). One could wonder, however, whether symptoms (traditionally evaluated for their epistemic quality as indicators of an underlying abnormality) are the only and/or most crucial elements of a psychopathological system from a causal point of view – that is, whether they are the crucial drivers of network dynamics that result in a disordered state (Fried & Cramer Reference Fried and Cramer2017; Jones et al. Reference Jones, Heeren and McNally2017). Even without a network approach in mind, there is reason to be critical regarding the capacity of symptoms to capture the essential features of a disorder. For example, Hens et al. convincingly argue that ASD symptoms do not capture “what it actually means to have such a disorder.” In addition, as Hyland argues, the covariation of somatic symptoms in the case of functional disorders “cannot be explained in terms of symptom-to-symptom causality.” So, what other variables may be relevant for psychopathology networks? Many candidate factors could (and should) be considered, primarily within the context of explicit theoretical models (i.e., formalized dynamical systems); Crafa & Nagel even advocate including social flexibility in psychopathology networks, as it may underpin the degree to which a person is susceptible to intentionality itself – that is, the degree to which “exposure to new social information will reinforce or undermine existing processes.” Jones et al. (Reference Jones, Heeren and McNally2017) similarly provide a motivation for a number of cases where interplay between symptoms may not suffice to characterize disorders.

Potentially moving away from symptoms means relying less on existing data and collecting more new data with the specific purpose of capturing networks and their dynamics. This is something Jayawickreme et al. also explicitly advocate, but for another reason: Many of the items that feature in existing datasets were specifically developed with the intent of accurately measuring one and the same latent variable. This usually means that such items tap into similar aspects of the construct targeted. Take, for example, the items “I like a desk without clutter” and “I sort my socks by color.” Such items are, by necessity and design, highly correlated; and in a network analysis, this will show up as a relatively strong connection. Importantly, such a strong connection may not stem from an actual direct relation – for example, liking a desk without clutter causing someone to sort their socks by color – but, rather, from the fact that these two items really do measure the same thing, namely liking order (or not). This will prove an additional challenge in future research: telling network fact (actual relations) from fiction (spurious correlations). Clearly, if many variables that overlap too strongly in their semantics are included in a network structure, this may yield inadequate solutions (Costantini Reference Costantini2014).

Successful modeling of complex systems therefore requires a judicious choice of the key variables in the system. This has proven true for dynamical systems modeled in fields ranging from ecosystems to meteorology, and is undoubtedly true for psychopathology as well. As more theoretically informed network theories are developed, we hope it will become clear which variables are essential and how the dynamic interaction between them unfolds in time. Current symptomatology is likely to include some key variables already, but is unlikely to contain a definitive list; hence, we expect considerable progress to unfold in this respect in both substantive and methodological directions.

R8. Mereological matryoshkas? Integrating multiple levels of analysis

As we have stressed throughout our target paper, blocking explanatory reductionism by means of a network perspective is not analogous to blocking biology-oriented research altogether. That is, although searching for a biological common cause (e.g., etiological brain pathways as mentioned by Ioannidis) is not fruitful if a mental disorder is the outcome of symptom-symptom network dynamics, searching for biological processes that are implicated in symptoms (e.g., sclerotic plaque as the process that gives rise to the symptom chest pains; Elbau et al.) and connections between them (e.g., homeostatic mechanisms that give rise to interindividual differences in connectivity between insomnia and fatigue) is by no means a pointless exercise. Symptoms in particular that either have what McNally refers to as “formal” features (e.g., auditory hallucinations), or that “appear to lack intentional content altogether” (e.g., exaggerated startle and emotional numbing), are suitable candidates for research programs that aim at biological elucidation. That is, there are interesting biological/psychological correlates to be found that may partially account for the non-intentional component of certain symptoms and connections between them. The main difference with existing research programs from a network perspective is that these correlates are not the holy grail, as in Tabb's “triumphalist reductionism,” but instead would be modeled and analyzed as making up or informing the structure and dynamics of the complex system that drives psychopathology.

A challenge for the coming years is to develop sensible ways in which to integrate various levels of network analysis – for example, combining biological with more psychological variables (as suggested by Baran, Hyland, Oller, Pasqualotto, and Pessoa). Addressing this challenge properly will be anything but simple. The most straightforward solution – that is, just estimating a network structure for all of these biological and psychological variables simultaneously – is suboptimal for various reasons. For instance, it is well known that correlations between data coming from different sources will typically be low. As a result, using current state-of-the-art network estimation methods that use regularization in order to avoid false positives will result in very small or even absent conditional dependence relations – while, in fact, such a relation between a biological and psychological variable may be important. One way to deal with this problem is to find statistical solutions (e.g., by locally relaxing penalty parameters of the statistical regularization procedure), but most of these solutions naturally come at the expense of increasing the probability of finding spurious relations. A similar way of including biological variables in a symptom network is by treating the biological variables as moderator variables that determine the strengths of connections between symptoms.

It is certainly important to develop and try out such techniques, but it should be realized that the above strategies analyze distinct variables (e.g., psychological and biological ones) as if they are functionally distinct entities, which may not be appropriate. Another approach would be to assume that the “biological” and the “psychological” represent different network structures, as Hyland argues for example, particularly in the case of functional disorders such as fibromyalgia. Such a solution, while highly interesting, is not trivial. Two major questions are how these network structures relate to one another at a theoretical level, and which statistical model, if any, best captures this relation. Under the assumption that psychological variables are not simply higher-level realizations of lower-level biological processes, as we have argued in our target paper, relating “biological” to “psychological” network structures may operate in at least two ways. First, they may be related through a mereological structure, where the biological parts do not cause a particular psychological variable but rather form that variable, just as parts of a tangerine form a tangerine yet do not cause it. In this case, subsets of the biological variables could be modeled psychometrically as formative indicators of psychological variables (Kievit et al. Reference Kievit, Romeijn, Waldorp, Wicherts, Scholte and Borsboom2011). Second, psychological and biological networks may be related through a Russian doll structure, where biological network structures are “nested” in a psychological network without one causing the other, just as one smaller matryoshka does not cause a larger matryoshka. One should, in this case, find a sensible way of relating the state and architecture of elements in the psychological network structure to those of the biological network structures. This perhaps could be done by making the value of each psychological variable a function of the state of the embedded biological network, while the state of the embedded biological network is a function of the relations in the symptom network; however, as far as we know, there is currently no modeling framework to encode this idea.

Interestingly, as soon as one moves from verbal descriptions of relations between “levels of analysis” to formalized models of these relations, it becomes unclear exactly how the formalization should be done (apart from the most simple models; see also Kievit et al. Reference Kievit, Romeijn, Waldorp, Wicherts, Scholte and Borsboom2011). Perhaps it would be useful to construct a “simplest non-trivial case” in the form of a simulation model that explicitly codes relations between psychological and biological networks. Such a model could also be used to focus the debate. In any case, it is clear that the question of how to integrate biological and psychological levels of analysis is wide open in psychopathology research, and that there are considerable opportunities for progress in addressing this issue.