Sage Journals: Discover world-class research

Abstract

We compare three theoretical frameworks for pursuing explanatory integration in psychiatry: a new dimensional framework grounded in the notion of computational phenotype, a mechanistic framework, and a network of symptoms framework. Considering the phenomenon of alcoholism, we argue that the dimensional framework is the best for effectively integrating computational and mechanistic explanations with phenomenological analyses.

Keywords

alcoholism computational phenotype dimensional psychiatry explanatory integration network of symptoms

Explanatory integration is one of the traditional aims of psychiatry, but it remains controversial how it should be pursued (Stephan et al., 2016). In this article, we examine three theoretical frameworks for pursuing explanatory integration in psychiatry: Kendler, Zachar, and Craver’s (2011) mechanistic framework, Borsboom and Cramer’s (2013) network of symptoms framework, and a dimensional framework we are going to develop and ground in the notion of computational phenotype, where a computational phenotype is “a measurable behavioural or neural type defined in terms of some computational model” (Montague, Dolan, Friston, & Dayan, 2012, p. 72). Considering alcohol use disorder (AUD), we show that, in comparison to the mechanistic and network of symptoms frameworks, the dimensional framework is more useful in a variety of clinical and research contexts, and more adequate for integrating computational and mechanistic explanations with phenomenological analyses.

The motivations for focusing on AUD are threefold. First, AUD is a prominent explanatory target of computational psychiatry. Computational psychiatry aims “to enable integration” of explanations of mental maladies across temporal and spatial scales—from genes to molecules, cells, circuits, brain systems, and individual and social behaviour—“by demonstrating, in a mathematically rigorous way, how phenomena on one level impact phenomena on another” (Kurth-Nelson et al., 2016, p. 79). Second, AUD has distinctive genetic, neurophysiological, behavioural, social, and cultural correlates, and a rich phenomenology. Its phenomenology involves a sense of impaired control over drinking, delirium with delusions and hallucinations, “blackouts,” craving, and suffering associated with hangovers, withdrawal, and social isolation (see Flanagan, 2013; Shinebourne & Smith, 2009; Smith, 1998). Third, given its numerous correlates and its rich phenomenology, and because computational psychiatry has been said to “honour the values and goals of those with lived experience of psychosis” (Powers, Bien, & Corlett, 2018, p. 640), AUD offers an ideal case for assessing how different theoretical frameworks can fruitfully integrate phenomenology, mechanism, and computation in psychiatry.

Alcohol use disorder and phenomenology

Alcohol is one of the most widely used psychoactive, dependence-producing substances in the world, and is associated with several mental maladies (Connor, Haber, & Hall, 2016). According to the Diagnostic and Statistical Manual of Mental Disorders (APA, 2013), the criteria for diagnosing alcohol use disorder (AUD) are the following:

Alcohol is often taken in larger amounts or over a longer period than was intended.

There is a persistent desire or unsuccessful efforts to cut down or control alcohol use.

A great deal of time is spent on activities necessary to obtain alcohol, use alcohol, or recover from its effects.

Craving, or a strong desire or urge to use alcohol.

Recurrent alcohol use resulting in a failure to fulfil major role obligations at work, school, or home.

Continued alcohol use despite having persistent or recurrent social or inter-personal problems caused or exacerbated by the effects of alcohol.

Important social, occupational, or recreational activities are given up or reduced because of alcohol use.

Recurrent alcohol use in situations in which it is physically hazardous.

Alcohol use is continued despite knowledge of having a persistent or recurrent physical or psychological problem that is likely to have been caused or exacerbated by alcohol.

Tolerance, as defined by either of the following:

A need for markedly increased amounts of alcohol to achieve intoxication or desired effect.

A markedly diminished effect with continued use of the same amount of alcohol.

Withdrawal, as manifested by either of the following:

The characteristic withdrawal symptoms for alcohol. . .

Alcohol [. . .] is taken to relieve or avoid withdrawal symptoms. (APA, 2013, pp. 490–491)

Over history, there have been several other diagnostic criteria for alcoholism (Tabakoff & Hoffman, 2013). The items on the DSM–5 list describe common psychological, behavioural, and social aspects of excessive alcohol consumption. Some of these items correlate with various risk factors, including: aggregate genetic risk factors, impairments in the frontal lobes and their connections with limbic regions in the brain, neuroticism and impulsivity, parental loss, peer alcohol use, and prices of alcoholic beverages (Kendler, 2012, pp. 12–14). Items 4–6 and 9 describe poor learning and decision-making as central features of AUD, and computational modelling of learning and decision-making in AUD has been growing in the last few years; items 2, 4, and 11 indicate AUD patients have expectations, perceptions, desires, moods, and thoughts infused with value, which can exert strong motivational power on their behaviours. Within philosophy, these types of mental states are the targets of phenomenological analyses.

Phenomenology is the study of structures of types of intentional mental states like perceptions, thoughts, emotions, desires, imaginations (i.e., mental states of, or about something) as they are experienced from a first-person point of view. Within this heterogeneous field of research, at least three methods have been adopted to study the structures of psychiatric patients’ experience. The first method consists in describing lived experience “as it is, without taking account of its psychological origin and the causal explanations which the scientist, the historian, or the sociologist may be able to provide” (Merleau-Ponty, 1945/1962, p. vii). Jaspers (1913/1997), for example, offers comprehensive descriptions of patients’ lived experiences, based on biographical information and patients’ accounts of their experiences, relationships, and condition. Another method consists in interpreting a type of experience by situating it in a social, material, and experiential context. For example, Laing (1960) interprets the psychotic experience of an individual with schizophrenia by situating it in the web of personal relationships of the patient. His interpretation of this type of experience is in terms of “ontological insecurity,” which is a fragile sense of self, impeding people’s taking for granted the “realness” and “meaning” of ordinary circumstances in everyday life. A third method consists in analysing the modal structure, and conditions of possibility, of embodied subjective experiences, distinguishing different features of different types of experiences.

Phenomenological methods in psychiatry involve complex descriptive, interpretative, and analytical processes, where patients’ experiences are organized on the basis of specific theoretical structures (see Carel, 2011; Fuchs, 2010). For example, Laing (1960) organized patients’ experiences on the basis of the existential structure of “ontological insecurity,” which, applied to the case of AUD, highlights that alcohol-dependent patients feel more unreal than real; they would feel so separated from the rest of the world that they experience their autonomy and identity as constantly threatened. The only escape from this existential despair is through the anaesthetization of alcohol. While triangulating between multiple data sources and methods can contribute to the trustworthiness of phenomenological descriptions, interpretations, or analyses, the ultimate criterion to evaluate these results is their capacity to make sense of out-of-the-ordinary, unexpected experiences that cannot be readily understood in terms of more familiar knowledge structures.

Phenomenological descriptions, analyses, or interpretations may not constitute explanations. After all, phenomenology is often characterized as a purely descriptive enterprise distinct from explanation. But they can still constrain and inform causal, computational, or other types of explanations of mental maladies. In fact, phenomenological results are often used to clarify the structure of the experiences involved in mental maladies, to interpret experimental results, and to inspire hypotheses for further research (Gallagher, 2004; Parnas & Sass, 2008).

Consider AUD. One recurrent feature of phenomenological descriptions of AUD is the sense of powerless, helpless suffering that accompanies alcoholics’ drinking behaviour. Smith (1998), for example, conducted in-depth interviews with six alcohol-dependent patients, between 42 and 61 years of age, in a clinic in Scotland. Using interpretative and descriptive phenomenological methods, he puts into focus how patients’ “suffering is lived as an insidious process, a movement of ever decreasing circles, whose momentum accelerates you into a rapid, spiralling decline. This vortex is a spinning vicious circle, full of energy, yet symbolising powerlessness” (p. 216).

Phenomenological descriptions seem to suggest “learning to drink” consists in the acquisition of a rigid habit, of a behavioural pattern that is so ingrained as to make heavy drinkers insensitive to their motivation state and to the bodily, psychological, and social consequences of drinking. In acquiring a habit of drinking, goal-directed processes are likely to be involved too (Everitt & Robbins, 2016). But, in acquiring and enacting drinking habits, the flexibility of alcohol-dependent individuals’ goal-directed decision processes is often reduced, as alcohol-dependent individuals’ repeated drinking behaviour often biases their expectations about the goodness of consuming alcohol. One 31-year-old alcohol-dependent patient interviewed by Shinebourne and Smith (2009) explains how excessive alcohol consumption reconfigured her sense of self:

Some big wave, you know, you just get caught with it, that’s what it used to be like, this kind of like helpless feeling, just having to go and get drunk almost, you know, not even particularly wanting to, just feeling like there’s no other way when you are in that situation. I was very much at sea, really, and, I didn’t feel grounded. . . just this flux and thought, when am I ever going to go on land. . . and even if you were sitting on the beach, you know, you’d get caught back in. (pp. 155–156)

This feeling ungrounded, helplessly “caught back in by some big wave,” is also one of the motifs in Flanagan’s (2013) memoir of his own alcohol dependence and recovery. Flanagan discusses the myriad ways in which the habit of drinking blends into one’s sense of self. For at least some alcohol-dependent patients, alcohol does not completely hijack their capacity for goal-directed control, but biases it towards drinking. Alcohol becomes part of the way they are:

Their personhood, their character, is constituted, in part, by a history of drinking, by a set of identifications and practices that involve alcohol, and that make these individuals who and what they are. Alcoholism, of this sort, at any rate, is a wide ecological phenomenon; it involves the deep-self. (pp. 885–886)

Because drinking may be partly constitutive of one’s sense of self, “undoing alcoholism as a form of life, and not more narrowly as just a drinking problem, involves fairly radical undoing and then redoing of oneself” (p. 886). It involves acquiring new habits that fill with meaning one’s understanding of their subjective experiences and social condition.

In summary, standard diagnostic criteria of AUD indicate that bad decision-making and impaired learning are central features of AUD. Some phenomenological descriptions and analyses highlight alcohol-dependent people typically experience powerlessness, suffering, and self-stigma; these experiences routinely accompany their drinking habits. For many patients, alcohol is constitutive of their form of life. Undoing alcoholism would require “redoing oneself,” by acquiring new goals and expectations, and developing habits that may give a new meaning to one’s lived experiences.

Explanatory integration beyond reduction

Integrating two or more accounts of a phenomenon consists in combining the concepts, evidence, results, or methods involved in those accounts into one integral explanatory account. The resulting integrated account explains different aspects of the phenomenon, displaying how such aspects are logically, probabilistically, constitutively, or causally related. To the extent an explanatory account is integrated, it yields understanding of the phenomenon as a multifaceted whole.

An integrated explanatory account of a target phenomenon requires some kind of dependence—logical, probabilistic, constitutive, or causal dependence—between the accounts to be integrated. If two accounts do not possess any logical, statistical, evidential, causal, or conceptual dependence, they cannot be integrated in an explanatory account. If they are fully independent, the two accounts are mutually irrelevant, and each would be unconstrained by the evidence, concepts, results, or methods on which the other account relies. Two accounts are independent to the extent they each enjoy many epistemic autonomies to a great degree. An account of a given phenomenon can enjoy different kinds of epistemic autonomies with respect to another account of the same phenomenon. Specifically, (a) autonomy in the selection and use of taxonomic categories, (b) autonomy in the selection of theoretical vocabulary, (c) autonomy in the choice of methods of investigation, and (d) autonomy in the selection of and weight given to relevant evidence. If an account does not enjoy any of these autonomies with respect to another, then the relationship between the two accounts is one of full dependence.

Full dependence between accounts in the science of mind and brain is generally understood in terms of a “classical” notion of reduction. Different strategies for scientific reduction have been developed in the philosophy of science (e.g., Bickle, 2006; Nagel, 1961; Schaffner, 1993). Although these strategies aim at establishing that a certain phenomenon (entity, property, or process) is identical to or fully explained by another, more basic phenomenon (entity, property, or process), or that a certain account (concept, model, or theory) can be logically derived from another more basic account, they all share at least the assumptions that phenomena and their scientific study belong on different “levels,” and that the concepts of a reduced explanatory approach should be connectable to the concepts of the reducing explanatory approach.

While it is often claimed that “reductionism has dominated both research directions and funding policies in clinical psychology and psychiatry” (Borsboom, Cramer, & Kalis, 2018, p. 1),¹ actual attempts at reducing particular psychiatric phenomena to lower-level neural, molecular, or genetic underpinnings are sparse and patchy. As Schaffner (2013) explains, “in the past fifty years, a reductionistic approach in the biomedical sciences and in psychology has become far less imperialistic and considerably more fragmented and tentative” (p. 1018). So, because reductionism does not offer an adequate framework for integrating different accounts in psychiatry in a way that comports with successful inter-field explanatory practices, we leave it on the side.

Dimensional computational phenotypes

Between full independence and dependence are several intermediate positions, which involve relationships of partial dependence or mutual constraint (Kaplan, 2017). One increasingly popular framework in psychiatry (Heinz, 2017; Montague et al., 2012) appeals to David Marr’s (1982) three-level framework for analysing information-processing systems (Colombo, 2015, Sec. 4). The computational level specifies what input-output function the system computes and why that type of system ought to compute that function. The algorithmic level specifies the effective procedures and representations employed by the system. The implementation level specifies how those procedures and representations are physically realized in the system.

Reinforcement Learning (RL) approaches to computational modelling have been flourishing in psychiatry in the last few years (Maia & Frank, 2011). Within RL, the computational level specifies the problem of learning what to do in an unfamiliar environment so as to maximize a numerical reward signal. Model-free control and model-based control are two families of RL algorithms that can be used to solve this problem. Model-based control algorithms learn a model of the environment, which they use to compute the expected value of possible actions by simulating their consequences. Model-based control produces more accurate and flexible decisions than model-free control, but is also computationally expensive, since it requires the agent to simulate future possibilities. Model-free algorithms do not exploit and search any model of the environment; they store the long-run expected value of each action, computing them online, on the basis of a reward prediction error, the difference between predictions about the reward obtained by taking a certain action in a given state and the rewards actually received. Model-free control is less computationally costly than model-based control, but produces relatively inflexible decisions, which are similar to habits. If the agent’s motivational state changes, or the structure of the environment changes, then the values “cached” by a model-free algorithm may be outdated and produce maladaptive choices. At the level of implementation, a wealth of neurobiological evidence suggests that the phasic activity of dopaminergic neurons in the basal ganglia encode prediction error signals that are recruited by the cortical-basal ganglia circuit for model-free (Colombo, 2014), and model-based control too (Langdon, Sharpe, Schoenbaum, & Niv, 2018).

Relying on RL modelling, psychiatrists have started to identify possible computational phenotypes of mental maladies. Computational phenotypes are measurable behavioural, psychological, and neural types defined in terms of specific parameters extracted from specific computational models of a given task on the basis of behavioural, psychological, and neurophysiological data (Montague et al., 2012; Patzelt, Hartley, & Gershman, 2018). As we will explain later, one clinically relevant² computational phenotype is a parameter that controls the trade-off between model-based and model-free processes in humans (Daw, Gershman, Seymour, Dayan, & Dolan, 2011).³ A range of psychiatric symptoms central to both AUD and many compulsive disorders has been associated with values of this phenotype, where model-based control is reduced in favour of model-free control (Gillan, Kosinski, Whelan, Phelps, & Daw, 2016; Sebold et al., 2014; Sebold et al., 2017; Voon, Reiter, Sebold, & Groman, 2017).

Computational phenotypes can ground a dimensional framework for explanatory integration in psychiatry. Similar to the dimensional approach taken by the Research Domain Criteria (RDoC) initiative of the National Institute of Mental Health (2010), dimensional approaches like ours assert that mental maladies should be understood as quantitatively, rather than qualitatively different from non-pathological psychological functions. Unlike the RDoC, our proposal does not assume that all mental maladies must have a localizable neurophysiological correlate; their organic correlate might be widely distributed and have diffuse effects at multiple spatial and temporal scales. Unlike the RDoC, we do not subscribe to the idea that different levels of psychological function should be defined on the basis of genetic or neurophysiological dysfunction (Insel et al., 2010, p. 749). We propose, instead, that different levels of a psychological function should be defined more abstractly, in terms of different levels of a computational phenotype, and that mental maladies should be conceived as regions of the space defined by a set of clinically relevant computational phenotypes (e.g., balance between model-based and model-free control, delay discounting, learning rate, sensitivity to other agents’ mental states). In so doing, our dimensional approach avoids challenges to the RDoC concerning its clinical fruitfulness, or its ability to remain neutral on the distinction between health and malady (Varga, 2019).

In summary, our dimensional framework conceives of mental maladies as regions of the space defined by the computational phenotypes, understands levels as Marr’s (1982) levels of analysis, and pursues explanatory integration by uncovering the common computational structure of apparently different maladies.

Mechanisms

Within the mechanistic framework, explanatory integration proceeds by revealing multi-level mechanisms responsible for phenomena. Mechanistic levels are levels of organization (not analysis). The mechanism and its causal activities are at a higher level than the mechanism’s constitutive component parts and operations; and, in turn, the mechanism’s component parts and operations are at a higher level than their sub-components. The relationship between mechanistic levels is one of physical constitution, not causation; and talk of levels in this framework refers to part–whole relationships within mechanisms.

Mechanistic integration of different levels proceeds by decomposing a system believed to be responsible for a phenomenon into its functionally relevant components, and by localizing which function is performed by which physical component when the mechanism produces the phenomenon (Bechtel & Richardson, 1993/2010). According to this approach, integrating different mechanistic levels of explanation consists in decomposing, localizing, and recomposing a mechanism with the aim of displaying how entities and operations at many different levels are related to one another and contribute to the production of the target phenomenon to be explained.

Within the mechanistic framework, Kendler et al. (2011) individuate mental maladies with mechanistic property clusters, that is: clusters of properties underlain, produced, or maintained by a mechanism. This view entails that explanatory integration of computational accounts and phenomenological analyses is successful to the extent that such accounts can each reveal structures that produce, underlie, or maintain mental maladies. Kaplan (2011) captures this commitment in terms of a “model-mechanism-mapping constraint,” whereby a model of a phenomenon has explanatory power to the extent that:

(a) the variables in the model correspond to identifiable components, activities, and organizational features of the target mechanism that produces, maintains, or underlies the phenomenon, and (b) the (perhaps mathematical) dependencies posited among these (perhaps mathematical) variables in the model correspond to causal relations among the components of the target mechanism. (p. 347)

Compliance with this constraint would guarantee that different accounts of a target mental malady combine concepts, results, and methods to uncover a single mechanism responsible for the malady.

In summary, Kendler et al.’s (2011) mechanistic framework conceives of mental maladies as property cluster mechanisms, understands levels as levels of physical organization within a mechanism, and pursues explanatory integration by combining concepts, results, and methods from different fields in the service of discovering the mechanism responsible for a malady.

Networks of symptoms

Borsboom and Cramer’s (2013) network of symptoms framework is another prominent approach for explanatory integration in psychiatry. Unlike the mechanistic framework, the network of symptoms framework understands psychiatric maladies as alternative, stable states of networks of strongly connected symptoms. These networks of symptoms need not have a common mechanism that is causing them. Symptoms here are not indicators of some underlying condition that causes them, but are understood as interconnected variables that are constitutive of mental maladies. The network of symptoms framework pursues explanatory integration in psychiatry by constructing networks of symptoms that reflect interdependencies between various neurobiological, psychological, behavioural, social, and cultural symptoms (Borsboom et al., 2018).

Symptoms can be activated by external conditions, for example by the presence of empty bottles of beer in the environment; they can also be triggered by internal states, for example by steroids’ interference with long-term potentiation in the hippocampus (e.g., Wetherill & Fromme, 2016). As the network strategy understands symptoms as statistically and causally connected variables that can change over time, activation of one symptom can cause activation (or suppression) of some other symptom. When certain symptoms in a network are co-active, and have the appropriate causal strength, a mental malady emerges.

In summary, Borsboom and Cramer’s (2013) network of symptoms framework conceives of mental maladies as stable, interacting symptoms without an underlying common cause, posits no levels of organization or analysis, and pursues explanatory integration by constructing networks of symptoms, and by holistically capturing the structure of their relationships.

How to pursue explanatory integration in psychiatry

We now examine some of the theoretical and practical virtues and limitations of the three frameworks we have outlined. Our overall conclusion is that a dimensional framework grounded in the notion of computational phenotype is the best for effectively integrating computational and mechanistic explanations with phenomenological analyses of mental maladies, in a way that accords with successful practices in psychiatry.

Mechanistic integration of computation and phenomenology

Kendler et al.’s (2011) mechanistic framework has several attractions. It denies that mental maladies can be adequately understood as natural kinds defined in terms of necessary and sufficient conditions. They argue mental maladies should instead be understood as mechanistic property clusters. These clusters consist of sets of varying symptoms that are produced, stabilized, and maintained by some mechanism. Mechanisms individuate mental maladies; and different properties of a mental malady—say, biological, psychological, and behavioural properties—would be properties of a single mechanism, at different levels of organization, where no single level of organization enjoys epistemic or practical privileges. In fact, although we are “far from being able to define plausible stability-producing mechanisms for most psychiatric disorders” (Kendler et al., 2011, p. 1148), one goal of current psychiatric practice is to discover and localize multiple, causal factors at different spatial and temporal scales that might constitute the mechanism of AUD. Such factors make a difference to whether a person compulsively seeks and takes alcohol, loses control in limiting alcohol intake, and tends to have negative emotional states associated with craving and withdrawal.

The mechanistic framework shows some theoretical and practical limitations in helping psychiatrists to pursue explanatory integration of mechanisms with computational and phenomenological accounts. Consider phenomenological methods. One of their aims is to elucidate the structure of subjective experiences, which cannot be detached from the whole circumstance of an embodied, ecologically situated patient. Elucidating the structure of subjective experiences would provide patients, psychiatrists, and therapists with a holistic account of the experiences involved in a mental malady, which can play distinctive roles in successful practices of diagnosis, sense-making, and therapy (Parnas & Sass, 2008). Instead, the mechanistic strategies suggested by Kendler et al. (2011) aim at explaining mental maladies in terms of the decomposable and localizable parts and operations of a mechanism that produce and maintain a cluster of symptoms. If the mechanistic strategies and phenomenological methods make inconsistent assumptions about mental maladies—the former assume mental maladies can be spatially decomposed and localized, while the latter assume they cannot—they enjoy at least some degree of autonomy.

This leads us to the second limitation of Kendler et al.’s (2011) mechanistic framework: at best, within a mechanistic framework, conscious experiences contribute to focusing on what needs to be explained with the vocabulary and taxonomies of the sciences aimed at discovering the mechanisms of mental maladies. Phenomenological taxonomies—for example, subjective descriptions of the “ontological insecurity” experienced by alcohol-dependent patients—can be considered subjective, personal-level descriptions of pathological, sub-personal mechanisms giving rise to cravings (see Colombo, 2013). These descriptions can contribute to elucidating the nature of the malady to be explained, but they do not provide constraints on the adequacy of mechanistic accounts, which ultimately appeal to different taxonomies from genetics, neuroscience, psychology, and ethology. After all, phenomenological descriptions do not generally map onto the types of entities and processes posited by mechanistic accounts. For example, some alcohol-dependent patients experience time and temporal relationships as “circular”—as long as drinking continues, the future is a re-enactment of the past, and future outcomes are as valuable as past outcomes (Thune, 1977). But this experienced circularity does not map in any meaningful way onto the temporal relationships displayed by their neural processes. Because phenomenological descriptions and mechanistic accounts enjoy a relatively high degree of taxonomic autonomy, the mechanistic framework is theoretically inadequate for pursuing explanatory integration of mechanisms with phenomenological analyses.

If, within Kendler et al.’s (2011) mechanistic framework, phenomenology enjoys a high degree of methodological and conceptual autonomy, computational accounts enjoy too little autonomy, in a way that does not comport with successful practices in psychiatry and other sciences of mind and brain. Computational models are conceived of as “elliptical or incomplete mechanistic explanations” (Piccinini & Craver, 2011, p. 284) within the mechanistic framework. Mechanist philosophers focus their attention on the level of biological implementation, on which the explanatory value of accounts at Marr’s (1982) algorithmic and computational levels would depend. But successful practices in psychiatry, as well as in other sciences of mind and brain, show that computational models need not map onto biological mechanisms to be practically useful in a variety of clinical and research contexts, to answer counterfactual questions and unify different phenomena under the same computational description (Chirimuuta, 2018; Weiskopf, 2018). This suggests computational models should not be understood as mechanistic sketches: their explanatory value does not reduce to their capacity to uncover mechanisms; it also suggests that the taxonomies employed at the computational and algorithmic levels of analysis should enjoy “soft” constraints with respect to the details of physical implementation, as computational analyses and algorithmic models do not make any claim about the spatial localization and organization of the components they posit, which may be implemented in multiple physical mechanisms (Elber-Dorozko & Shagrir, 2019).

Voon et al. (2017) offer an example of the explanatory and taxonomic autonomies computational models should enjoy with respect to mechanism. Voon and collaborators review several lines of evidence that indicate the clinical and translational relevance of RL model-based control across multiple psychiatric disorders with different underlying mechanisms, including binge eating disorder, obsessive compulsive disorder, and AUD. Specifically, self-reported severity of alcohol use has been found to be associated with impairments in model-based control (Gillan et al., 2016) and treatment outcome abstinence duration (Voon et al., 2015); the interaction between reduced model-based control and high expectations about the positive effects of alcohol has been found to predict risk of relapse (Sebold et al., 2017). Furthermore, computational modelling of the balance between model-free and model-based control provides a theoretical foundation for therapeutic interventions that aim to increase model-based control and inhibit model-free processes underlying temptation and societal pressure, such as training at cognitive bias modification,⁴ which has been found to improve treatment outcome (Heinz et al., 2017; Wiers, Eberl, Rinck, Becker, & Lindenmeyer, 2011; see also Moutoussis, Shahar, Hauser, & Dolan, 2018).

This body of evidence indicates a specific computational phenotype—viz. balance between model-based and model-free control in learning tasks—can unify apparently different compulsive disorders. It also indicates that computational models can support some counterfactual predictions about treatment outcome and risk of relapse. While model-based control has been associated with ventromedial prefrontal cortex and caudate activity, and model-based computation shares a dopaminergic foundation with model-free control (Deserno et al., 2015), the success of computational phenotypes in unifying apparently different disorders and supporting counterfactual predictions does not obviously depend on their mapping onto specific neural structures. If Kendler et al.’s (2011) mechanistic framework does not acknowledge the relative degree of explanatory and taxonomic autonomies of computational models, then it cannot adequately integrate distinctively explanatory computational results in psychiatry.

Network integration of computation and phenomenology

Borsboom and Cramer’s (2013) network of symptoms framework presents attractive features too. It eschews ill-defined talk of levels, as well as the simplistic identification of mental maladies with neurobiological states. Although the covariance between symptoms in a network can warrant causal conclusions about a target mental malady, and can thus uncover variables for intervention, these conclusions don’t assume the covariance between symptoms arises from some common latent causes (Borsboom et al., 2018). In this sense, Borsboom and Cramer’s (2013) network of symptoms framework is “flat”: it does not seek to uncover causes of symptoms at different levels of organization. Different features of a mental malady—say, craving, positive expectations about the outcomes of drinking, or reduced dopaminergic firing—are integrated as different interconnected nodes in a “flat” network structure. This kind of integration highlights statistical (or causal) patterns of heterogeneous variables that characterize a mental malady, while it can offer plausible accounts of comorbidity, in a way that accords with the more holistic methods of phenomenology.

Borsboom and Cramer’s (2013) network of symptoms framework doesn’t adequately integrate computational accounts and phenomenology. Within this framework, computational phenotypes, mechanisms, and phenomenological analyses enjoy a relatively high degree of conceptual, evidential, and methodological autonomy. Consider phenomenological descriptions. Though Borsboom and Cramer (2013) and Borsboom et al. (2018) do not address this point, phenomenology might only constrain networks of symptoms indirectly. In fact, Borsboom et al. (2018) argue that the covariation between symptoms in a network “can be seen to make sense” (p. 20), because symptoms often correspond to intentional mental states, that is, to mental states that are about something. For example, the desire to drink alcohol is about drinking alcohol. Since intentional mental states display “a rational relation,” Borsboom and colleagues suggest that networks of symptoms allow us to understand the lived experience involved in a malady. While this suggestion is a promising start, it falls short of providing a convincing answer to the question of how phenomenological descriptions or analyses fit or constrain a network of symptoms structure. After all, many phenomenological analyses of AUD experiences are focused on “pre-intentional” mental states like moods that need not be about any specific object in the world. Moods can be understood as providing subjects of experience with a background sense that structures their engagements with the environment infusing them with meaning (Heidegger, 1962, p. 176; Jaspers, 1913/1997, p. 688ff).

Even if a phenomenological description or analysis of AUD includes only intentional states, it is far from obvious how their “rational” relationships within a network of symptoms should be specified. Specifying them in terms of logical or semantic relationships might be a promising route, but it is likely these relationships will not always track the causal relationships uncovered by network analysis. For example, some phenomenological descriptions like Smith’s (1998) and Flanagan’s (2013) highlight several alcohol-dependent patients truly believe drinking will not make their suffering disappear and genuinely desire to stop drinking. Yet, an overwhelming majority of alcohol-dependent patients will relapse within their first year of sobriety (Beck et al., 2012). Unless a network of symptoms includes other mental states that could explain and rationalize the apparent inconsistency between relapse and the conscious belief and desire that one should remain abstinent, rational and causal relationships in the network will present a mismatch. These types of mismatches hinder the “sense-making” that a network of symptoms can provide of the lived experiences of alcohol-dependent patients.

The network of symptoms framework can include nodes corresponding to computational phenotypes. But, because Borsboom and Cramer’s (2013) network of symptoms framework is meant to be “flat,” it does not take into account the organizational relationships between the neurobiological components and causal activities that physically realize computational parameters and algorithmic transformations. Nor does the network of symptoms framework specify how the transformations posited at Marr’s (1982) algorithmic level relate to what a system is computing and why the system is computing that function instead of another. So, this framework—to the extent it pits itself against latent variable (or common cause) models, which it need not do (Bringmann & Eronen, 2018)—cannot integrate computational accounts of mental maladies.

Consider variables that correspond to entities and causal activities that realize a computational phenotype like the balance between model-based and model-free control. Some of these variables (e.g., level of dopamine release in the ventral striatum) are likely to be common causes of psychiatric symptoms (e.g., inflexible learning and craving for alcohol in certain environments). But one of the assumptions of Borsboom et al. (2018) is that a network approach is incompatible with a causal modelling approach aimed at inferring common (latent) causes of observed correlations between symptoms; they claim: “If a network model is correct. . . there exists no common cause” (p. 17). Despite this claim, however, network and latent variable models should not be seen as providing competing accounts, and instead should be considered as complementary strategies for understanding and treating mental maladies. Networks of symptoms can gain “depth” and allow for the integration of information about Marr’s (1982) different levels of analysis, by using representations of networks that encompass latent variable structures (see, e.g., Epskamp, Rhemtulla, & Borsboom, 2017), and that show how causal transactions between organized sets of variables systematically relate to computational transformations of information.

Dimensional integration of computation and phenomenology

Conceptualizing mental maladies in terms of dimensional computational phenotypes allows us to unify apparently different diagnostic categories on the basis of their common dimensional computational structure. Within a space of computational phenotypes, we can also answer counterfactual questions concerning how social behaviour, neural activity, and subjective experience would change, had the value of a certain computational phenotype defining that space changed. These are two reasons why computational phenotypes have explanatory power. Let us now focus on the notion of a computational phenotype, and consider how a dimensional framework grounded in this notion can help psychiatrists pursue the integration of mechanistic and computational explanations with phenomenological accounts of mental maladies.

As we already mentioned, computational phenotypes are types of parameters defined within a computational model of a task (Montague et al., 2012; Patzelt et al., 2018). Computational phenotypes include such model parameters as rate of learning, which controls the extent to which new information overrides old information, delay discounting, which determines the extent to which the present value of a reward is discounted with delay of its receipt, loss aversion, which controls the preference to avoid losses to acquiring equivalent gains, and depth of reasoning, which controls to what extent one works out the thoughts of other people in strategic reasoning.

Computational phenotypes are continuous parameters and define types of continuous (or dimensional) psychological functions. A set of computational phenotypes can be used to define an abstract space of human phenotypes, that is, a space of types of individuals who share modes of behaviour and information processing for a wide range of decision and learning scenarios. For example, AUD might correspond to a region of the space defined by delay discounting, learning rate, and trade-off between model-based and model-free control. The choice of computational phenotypes most relevant to define a certain dimensional space for a target malady depends on evidence available about an individual’s psychological and neurobiological dysfunctions, on the individual’s level of social participation, and on the individual’s affective life. It also depends on the practical clinical needs, and on clinicians’ phenomenological insights into the condition of a patient.

Now, how does a space defined over dimensional computational phenotypes exactly promote explanatory integration in a way that is more theoretically adequate and practically useful than Kendler et al.’s (2011) mechanistic framework and Borsboom and Cramer’s (2013) network of symptoms framework?

Let’s start from mechanism, clarifying the main differences between our own proposal and Kendler et al.’s (2011). First, we do not assume an account of a mental malady is adequate to the extent it uncovers its neurobiological mechanism; unlike the mechanistic framework, we assume an integrative explanatory account should be judged in terms of practical success, not in terms of its ability to latch onto mechanisms that exist independently of human theorizing. Second, computational modelling enjoys a relatively higher degree of autonomy within our framework; though it constrains and it is constrained by available mechanistic evidence, computational analyses and algorithmic models are not sketches of neurobiological mechanisms. Third and finally, on our view, types of mental maladies are in part individuated on the basis of human classificatory practices, in particular practices involving computational and phenomenological analyses.

Computation and neural mechanisms can obviously be integrated within our dimensional framework. In keeping with Marr’s (1982) three levels of analysis, computational phenotypes are realized by neurobiological mechanisms that transform exteroceptive, proprioceptive, or interoceptive inputs into behavioural, emotional, or cognitive outputs. Given that a computational phenotype such as balance between model-based and model-free control is extracted from a computational model on the basis of behavioural and neural data, different values of this phenotype will be associated with different environmental stimuli, but also different levels of activity in certain neural circuits in the medial prefrontal cortex, also in alcohol-dependent patients (Daw et al., 2011; Deserno et al., 2015; Sebold et al., 2017). More importantly, the trade-off between model-based and model-free processes towards model-free control can be a computational phenotype of a range of disorders underlain by different mechanisms but all involving compulsion or drug abuse (Gillan et al., 2016; Voon et al., 2017). In this way, dimensional computational phenotyping can ground a unified explanation of several kinds of addictions beyond AUD, displaying their common computational structure within a certain space of computational phenotypes.

Consider phenomenology. Berrios and Marková (2013) argue that a dimensional approach to psychiatry is misguided and cannot integrate phenomenological analyses or descriptions. Their argument is that a dimensional approach to some phenomenon entails the possibility of measuring that phenomenon by concretely interacting with it. Because, according to Berrios and Marková, mental symptoms have “abstract” or “ideal attributes (meanings),” which cannot be measured, “mental symptoms can only be evaluated (not measured)” (p. 78).

Berrios and Marková’s (2013) argument is inconclusive. A dimensional approach to a certain phenomenon does not entail that a phenomenon must be measurable via interaction or must be a concrete object. Sets are not concrete objects; yet, their cardinality can be measured. In fact, measurement and dimensionality often involve the representation of ideal systems, such as the consumption of alcohol in the average household in a certain neighbourhood in a country. Furthermore, measurement theory is a heterogeneous field, where different authors with different epistemic commitments understand the nature of the relata of an act of quantitative measurement differently. Regardless of the nature of the relata of measurements, Berrios and Marková’s (2013) argument is at odds with the fruitfulness of psychometric and dimensional approaches to understanding mental maladies (cf. Hägele et al., 2015; Heinz, Schlagenhauf, Beck, & Wackerhagen, 2016). What is correct in Berrios and Marková’s (2013) suggestion is that psychiatric research often involves phenomenological description, analysis, and interpretation of subjective experiences of suffering that cannot be measured only with questionnaires, scales, experimental tasks, or bodily measurements.

Our dimensional computational framework, however, is responsive to phenomenological descriptions and analyses of mental maladies in two ways. First, Marr (1982) says that “the most abstract is the level of what the device does and why” (p. 22). What a system does and why it does that instead of something else contribute to delineating the phenomenon to be explained (Shagrir, 2010). Within our framework, phenomenology contributes to specifying what a system is meant to accomplish within a certain ecological context, in a way that demonstrates the aptness of what the system does in that ecology. For example, Laing’s (1960) analysis of ontological insecurity displays psychoses as bound up with one’s sense of “ontological insecurity,” where one feels they are losing their sense of self, reality, and meaningful social relationships. While ontological insecurity can usher in anxiety, withdrawal, and avoidance, this concept can illuminate the phenomenon to be explained and its ecological constraints. A computational-level hypothesis informed by Laing’s phenomenological analysis of ontological insecurity is that alcohol-dependent patients may fail to integrate afferent interoceptive and exteroceptive representations with self-referential representations. Couched in mathematical terms, this hypothesis can be specified algorithmically and tested in the light of behavioural and neural data.

Second and more generally, phenomenological analyses and descriptions can provide patients and clinicians with narrative glue that may help patients make sense of the relationship between their suffering and their computational phenotypes. The abstract, non-biological taxonomies of computational models can be reinterpreted more easily than mechanistic accounts in terms of phenomenological categories. These categories may help one see how different computational phenotypes might be related and may reflect one’s lived experience of choices and perceptions of reality. They may help patients and their loved ones answer “existential” questions about the point of the suffering involved in their malady (see Roberts, 2000).

One important objection to our proposal is that rather than offering an alternative explanatory framework, what we are proposing only changes the topic: unlike the mechanistic and network of symptoms frameworks, our dimensional framework only redefines the explanandum (i.e., mental maladies); it does not explain why or how mental maladies come about.

To address this objection, it is helpful to draw an analogy. Different quantities suffice to physically characterize a system. For example, if you want to characterize a spring undergoing simple harmonic motion, its mass, period, and the acceleration of gravity suffice (plus a constant k determined by Hooke’s law). These quantities have dimensions. The dimension of the period of a pendulum is time [T]; the dimension of mass is [M]; the dimension of the acceleration of gravity is length divided by the square of a time [L/T²]; and the dimension of the constant k is [M/T²]. If we want to know why the spring has a certain period of oscillation, then we can derive the dimensional structure of the spring by working out an equation that gives us one quantity of interest as a function of all the quantities on which that quantity depends. From knowledge of the dimensional structure of the spring, we can conclude that the period P is proportional to the square root of mass divided by k.

This type of dimensional analysis is commonplace in physics, and it allows us to find the functional relationships between a set of quantities. These functional relationships can provide us with information about why apparently different systems behave similarly by considering the common dimensional structure they share (Lange, 2009). It also allows us to gain modal information about the behaviour of a system by allowing us to answer counterfactual questions about how change in some quantity of a system would influence change in some other quantity of that system. As Pexton (2014) puts it, this type of

dimensional explanation is not simply about reading off dimensions naturalistically from a system and combining them to get functional forms of dependence between variables. Rather it implicitly involves picking a conceptualisation of a target system that in part creates a perspective from which the dimensional architecture is constructed. (p. 2350)

In the current state of research in computational psychiatry, we are far from being able to specify a plausible set of computational phenotypes for most mental maladies. And computational phenotypes do not have any obvious dimension we are familiar with from physics; their dimensions need to be clarified within computational psychiatry and value theory. Yet, by using a dimensional structure defined by a set of computational phenotypes relevant to a target mental malady, we can represent what the malady might consist of. We can also see that apparently different mental maladies present a common dimensional computational structure. In this way, computational phenotyping can help psychiatrists understand what clusters of symptoms are produced by the same type of processes, and to what extent these processes are realized by common types of neurobiological mechanisms.

Furthermore, using a dimensional structure defined by a set of computational phenotypes allows us to gain clinically relevant, modal information about a mental malady. In particular, it can give us information about clinical heterogeneity and about possible targets for treatment. For example, Heinz et al. (2017) suggest that if impaired model-based control is a key computational phenotype of AUD that predicts relapse, then interventions aimed to enhance model-based vs. model-free control with behavioural and cognitive training, or pharmacological manipulations, are particularly promising therapeutic strategies for treating AUD (see also Moutoussis et al., 2018). In this way, computational phenotyping involves models connecting change-relating variables that allow psychiatrists to answer counterfactual questions generated by an explanandum mental malady.

Conclusion

One of the aims of psychiatry is explanatory integration. How can different concepts, sources of evidence, and methods used in different fields be integrated to adequately explain why a certain mental malady emerges and how it can be effectively treated? In this paper, we have started to articulate a dimensional theoretical structure based on the notion of computational phenotypes of mental maladies to pursue explanatory integration in psychiatry. Examining the case of AUD, we have shown how our dimensional framework can structure the search for tailored treatments targeting patients’ expectations, social environment, computational modes of control, and neurophysiology. Our proposal is compatible with attractive aspects of alternative frameworks for explanatory integration in psychiatry, like RDoC (Insel et al., 2010), Kendler et al.’s (2011) mechanistic framework, and Borsboom and Cramer’s (2013) network of symptoms frameworks; but, unlike these frameworks, our dimensional proposal allows us to more adequately integrate mechanism, computation, and phenomenology in the pursuit of general explanatory accounts of mental maladies.

Footnotes

Acknowledgements

We thank Roberto Fumagalli, Alexander Genauck, Miriam Sebold, and two anonymous referees for helpful conversations and comments on previous versions of this paper.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Work on this paper was financially supported by the Alexander von Humboldt Foundation and by the Deutsche Forschungsgesellschaft (Grant DFG FOR 1617/2).

Notes

Author biographies

Matteo Colombo is Associate Professor at the Tilburg Center for Logic, Ethics, and Philosophy of Science, Tilburg University, the Netherlands, and a Humboldt Research Fellow at the Clinic for Psychiatry and Psychotherapy, Charité University Clinic Berlin, Germany. His research focuses on questions about evidence and explanation in the computational, cognitive, and brain sciences. His recent publications include “First Principles in the Life Sciences: The Free-Energy Principle, Organicism, and Mechanism” (Synthese) and The Routledge Handbook of the Computational Mind (with M. Sprevak, Eds., Routledge, 2019).

Andreas Heinz is Medical Director, Department of Psychiatry and Psychotherapy at Campus Charité Mitte. His research group is active in the field of monoaminergic dysfunction in psychiatric disorders, particularly in alcoholism and schizophrenia, combining genetic research and multimodal brain imaging techniques. His recent publications include A New Understanding of Mental Disorders: Computational Models for Dimensional Psychiatry (MIT Press, 2017).

References

American Psychiatric Association (APA). (2013). Diagnostic and statistical manual of mental disorders (DSM–5). Washington, DC: Author.

Bechtel

Richardson

R. C.

(2010). Discovering complexity: Decomposition and localization as strategies in scientific research (2nd ed.). Cambridge, MA: MIT Press/Bradford Books. (Original work published 1993)

Beck

Wustenberg

Genauck

Wrase

Schlagenhauf

Smolka

M. N.

. . . Heinz , (2012). Effect of brain structure, brain function, and brain connectivity on relapse in alcohol-dependent patients. Archives of General Psychiatry, 69(8), 842–852.

Berrios

G. E.

Marková

I. S.

(2013). Is the concept of “dimension” applicable to psychiatric objects? World Psychiatry, 12(1), 76–78.

Bickle

(2006). Reducing mind to molecular pathways: Explicating the reductionism implicit in current cellular and molecular neuroscience. Synthese, 151, 411–434.

Borsboom

Cramer

A. O. J.

(2013). Network analysis: An integrative approach to the structure of psychopathology. Annual Review of Clinical Psychology, 9, 91–121.

Borsboom

Cramer

A. O. J.

Kalis

(2018). Brain disorders? Not really: Why network structures block reductionism in psychopathology research. Behavioral and Brain Sciences, 42, 1–54. doi: 10.1017/S0140525X17002266

Bringmann

L. F.

Eronen

M. I.

(2018). Don’t blame the model: Reconsidering the network approach to psychopathology. Psychological Review, 125, 606–615.

Carel

(2011). Phenomenology and its application in medicine. Theoretical Medicine and Bioethics, 32(1), 33–46.

10.

Chirimuuta

(2018). Explanation in computational neuroscience: Causal and non-causal. The British Journal for the Philosophy of Science, 69, 849–880.

11.

Colombo

(2013). Constitutive relevance and the personal/subpersonal distinction. Philosophical Psychology, 26(4), 547–570.

12.

Colombo

(2014). Deep and beautiful: The reward prediction error hypothesis of dopamine. Studies in History and Philosophy of Science: Part C. Studies in History and Philosophy of Biological and Biomedical Sciences, 45, 57–67.

13.

Colombo

(2015). For a few neurons more: Tractability and neurally informed economic modelling. The British Journal for Philosophy of Science, 66, 713–736.

14.

Connor

J. P.

Haber

P. S.

Hall

W. D.

(2016). Alcohol use disorders. The Lancet, 387(10022), 988–998.

15.

Daw

N. D.

Gershman

S. J.

Seymour

Dayan

Dolan

R. J.

(2011). Model-based influences on humans’ choices and striatal prediction errors. Neuron, 69, 1204–1215.

16.

Deserno

Beck

Huys

Q. J.

Lorenz

R. C.

Buchert

Buchholz

H. G.

. . . Heinz

(2015). Chronic alcohol intake abolishes the relationship between dopamine synthesis capacity and learning signals in the ventral striatum. European Journal of Neuroscience, 41(4), 477–486.

17.

Elber-Dorozko

Shagrir

(2019). Computation and levels in the cognitive and neural sciences. In Sprevak

Colombo

(Eds.), Routledge handbook of the computational mind (pp. 205–222). Abingdon, UK: Routledge.

18.

Epskamp

Rhemtulla

Borsboom

(2017). Generalized network psychometrics: Combining network and latent variable models. Psychometrika, 82(4), 904–927.

19.

Everitt

B. J.

Robbins

T. W.

(2016). Drug addiction: Updating actions to habits to compulsions ten years on. Annual Review of Psychology, 67, 23–50.

20.

Flanagan

(2013). Identity and addiction: What alcoholic memoirs teach. In Fulford

K. W. M.

Davies

Gipps

R. G. T.

Graham

Sadler

J. Z.

Stanghellini

Thornton

(Eds.), The Oxford handbook of philosophy and psychiatry (pp. 865–888). Oxford, UK: Oxford University Press.

21.

Fuchs

(2010). Phenomenology and psychopathology. In Schmicking

Gallagher

(Eds.), Handbook of phenomenology and cognitive science (pp. 546–573). Dordrecht, the Netherlands: Springer.

22.

Gallagher

(2004). Neurocognitive models of schizophrenia: A neurophenomenological critique. Psychopathology, 37(1), 8–19.

23.

Gillan

C. M.

Kosinski

Whelan

Phelps

E. A.

Daw

N. D.

(2016). Characterizing a psychiatric symptom dimension related to deficits in goal-directed control. eLife, 5, e11305.

24.

Hägele

Schlagenhauf

Rapp

Sterzer

Beck

Bermpohl

. . . Heinz

(2015). Dimensional psychiatry: Reward dysfunction and depressive mood across psychiatric disorders. Psychopharmacology, 232(2), 331–341.

25.

Heidegger

(1962). Being and time ( Macquarrie

Robinson

, Trans.). Oxford, UK: Blackwell.

26.

Heinz

(2017). A new understanding of mental disorders: Computational models for dimensional psychiatry. Boston, MA: MIT Press.

27.

Heinz

Deserno

Zimmermann

U. S.

Smolka

M. N.

Beck

Schlagenhauf

(2017). Targeted intervention: Computational approaches to elucidate and predict relapse in alcoholism. Neuroimage, 151, 33–44.

28.

Heinz

Schlagenhauf

Beck

Wackerhagen

(2016). Dimensional psychiatry: Mental disorders as dysfunctions of basic learning mechanisms. Journal of Neural Transmission, 123(8), 809–821.

29.

Insel

Cuthbert

Garvey

Heinssen

Pine

D. S.

Quinn

. . . Wang

(2010). Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders. American Journal of Psychiatry, 167(7), 748–751.

30.

Jaspers

(1997). General psychopathology (7th ed.; Hoenig

Hamilton

M. W.

, Trans.). Baltimore, MD: Johns Hopkins University Press. (Original work published 1913)

31.

Kaplan

D. M.

(2011). Explanation and description in computational neuroscience. Synthese, 183(3), 339–373.

32.

Kaplan

D. M.

(2017). Integrating mind and brain science: A field guide. In Kaplan

(Ed.), Explanation and integration in mind and brain science (pp. 1–28). Oxford, UK: Oxford University Press.

33.

Kendler

K. S.

(2012). Levels of explanation in psychiatric and substance use disorders: Implications for the development of an etiologically based nosology. Molecular Psychiatry, 17, 11–21.

34.

Kendler

K. S.

Zachar

Craver

(2011). What kinds of things are psychiatric disorders? Psychological Medicine, 41(6), 1143–1150.

35.

Kurth-Nelson

O’Doherty

J. P.

Barch

D. M.

Denève

Durstewitz

Frank

M. J.

. . . Tost

(2016). Computational approaches for studying mechanisms of psychiatric disorders. In Redish

A. D.

Gordon

J. A.

(Eds.), Computational psychiatry: New perspectives on mental illness (pp. 77–99). Cambridge, MA: MIT Press.

36.

Laing

R. D.

(1960). The divided self: An existential study in sanity and madness. Harmondsworth, UK: Penguin.

37.

Langdon

A. J.

Sharpe

M. J.

Schoenbaum

Niv

(2018). Model-based predictions for dopamine. Current Opinion in Neurobiology, 49, 1–7.

38.

Lange

(2009). Dimensional explanations. Nous, 43(4), 742–775.

39.

Maia

T. V.

Frank

M. J.

(2011). From reinforcement learning models to psychiatric and neurological disorders. Nature Neuroscience, 14(2), 154–162.

40.

Marr

(1982). Vision: A computational investigation into the human representation and processing of visual information. New York, NY: Freeman.

41.

Merleau-Ponty

. (1962). The phenomenology of perception ( Smith

, Trans.). New York, NY: Routledge & Kegan Paul. (Original work published 1945)

42.

Montague

P. R.

Dolan

R. J.

Friston

K. J.

Dayan

(2012). Computational psychiatry. Trends in Cognitive Sciences, 16(1), 72–80.

43.

Moutoussis

Shahar

Hauser

T. U.

Dolan

R. J.

(2018). Computation in psychotherapy, or how computational psychiatry can aid learning-based psychological therapies. Computational Psychiatry, 2, 50–73.

44.

Nagel

(1961). The structure of science: Problems in the logic of explanation. New York, NY: Harcourt, Brace & World.

45.

National Institute of Mental Health. (2010). From discovery to cure: Accelerating the development of new and personalized interventions for mental illnesses. Report of the National Advisory Mental Health Council’s Workshop: 31. Bethesda, MD: Author.

46.

Parnas

Sass

L. A.

(2008). Varieties of “phenomenology”: On description, understanding, and explanation in psychiatry. In Kendler

Parnas

(Eds.), Philosophical issues in psychiatry: Explanation, phenomenology, and nosology (pp. 239–277). Baltimore, MD: Johns Hopkins University Press.

47.

Patzelt

E. H.

Hartley

C. A.

Gershman

S. J.

(2018). Computational phenotyping: Using models to understand individual differences in personality, development, and mental illness. Personality Neuroscience, 1(e18), 1–10. doi:10.1017/pen.2018.14

48.

Pexton

(2014). How dimensional analysis can explain. Synthese, 191(10), 2333–2351.

49.

Piccinini

Craver

(2011). Integrating psychology and neuroscience: Functional analyses as mechanism sketches. Synthese, 183(3), 283–311.

50.

Powers

A. R.

Bien

Corlett

P. R.

(2018). Aligning computational psychiatry with the hearing voices movement: Hearing their voices. JAMA Psychiatry, 75(6), 640–641.

51.

Roberts

G. A.

(2000). Narrative and severe mental illness: What place do stories have in an evidence-based world? Advances in Psychiatric Treatment, 6(6), 432–441.

52.

Schaffner

(1993). Theory structure, reduction and disciplinary integration in biology. Biology and Philosophy, 8(3), 319–347.

53.

Schaffner

K. F.

(2013). Reduction and reductionism in psychiatry. In Fulford

K. W. M.

Davies

Gipps

R. G. T.

Graham

Sadler

J. Z.

Stanghellini

Thornton

(Eds.), The Oxford handbook of philosophy and psychiatry (pp. 1003–1022). Oxford, UK: Oxford University Press.

54.

Sebold

Deserno

Nebe

Schad

D. J.

Garbusow

Hägele

. . . Rapp

M. A.

(2014). Model-based and model-free decisions in alcohol dependence. Neuropsychobiology, 70(2), 122–131.

55.

Sebold

Nebe

Garbusow

Guggenmos

Schad

D. J.

Beck

. . . Heinz

(2017). When habits are dangerous: Alcohol expectancies and habitual decision making predict relapse in alcohol dependence. Biological psychiatry, 82(11), 847–856.

56.

Shagrir

(2010). Marr on computational-level theories. Philosophy of Science, 77, 477–500.

57.

Shinebourne

Smith

J. A.

(2009). Alcohol and the self: An interpretative phenomenological analysis of the experience of addiction and its impact on the sense of self and identity. Addiction Research & Theory, 17(2), 152–167.

58.

Smith

B. A.

(1998). The problem drinker’s lived experience of suffering: An exploration using hermeneutic phenomenology. Journal of Advanced Nursing, 27(1), 213–222.

59.

Stephan

K. E.

Binder

E. B.

Breakspear

Dayan

Johnstone

E. C.

Meyer-Lindenberg

. . . Flint

(2016). Charting the landscape of priority problems in psychiatry: Part 2. Pathogenesis and aetiology. The Lancet Psychiatry, 3(1), 84–90.

60.

Tabakoff

Hoffman

P. L.

(2013). The neurobiology of alcohol consumption and alcoholism: An integrative history. Pharmacology Biochemistry and Behavior, 113, 20–37.

61.

Thune

C. E.

(1977). Alcoholism and the archetypal past: A phenomenological perspective on Alcoholics Anonymous. Journal of Studies on Alcohol, 38(1), 75–88.

62.

Varga

(2019). Challenges to the dimensional approach. Philosophy, Psychiatry, & Psychology, 26(1), 77–79.

63.

Voon

Derbyshire

Rück

Irvine

M. A.

Worbe

Enander

. . . Robbins

T. W.

(2015). Disorders of compulsivity: A common bias towards learning habits. Molecular Psychiatry, 20(3), 345–352.

64.

Voon

Reiter

Sebold

Groman

(2017). Model-based control in dimensional psychiatry. Biological Psychiatry, 82(6), 391–400.

65.

Weiskopf

D. A.

(2018). The explanatory autonomy of cognitive models. In Kaplan

D. M.

(Ed.), Explanation and integration in mind and brain science (pp. 44–69). Oxford, UK: Oxford University Press.

66.

Wetherill

R. R.

Fromme

(2016). Alcohol-induced blackouts: A review of recent clinical research with practical implications and recommendations for future studies. Alcoholism: Clinical and Experimental Research, 40(5), 922–935.

67.

Wiers

R. W.

Eberl

Rinck

Becker

E. S.

Lindenmeyer

(2011). Retraining automatic action tendencies changes alcoholic patients’ approach bias for alcohol and improves treatment outcome. Psychological Science, 22(4), 490–497.

Explanatory integration,computational phenotypes,and dimensional psychiatry: The case of alcohol use disorder

Abstract

Keywords

Alcohol use disorder and phenomenology

Explanatory integration beyond reduction

Dimensional computational phenotypes

Mechanisms

Networks of symptoms

How to pursue explanatory integration in psychiatry

Mechanistic integration of computation and phenomenology

Network integration of computation and phenomenology

Dimensional integration of computation and phenomenology

Conclusion

Footnotes

Acknowledgements

Declaration of conflicting interests

Funding

Notes

Author biographies

References