Contribution
The network approach to psychopathology had produced statistical methods and many exploratory studies but, by the paper’s own account, had “not yet been fully articulated as a scientific theory”. This paper supplies the theory: five principles about the structure and dynamics of symptom networks, and from them new definitions of mental health, disorder, resilience, vulnerability and liability, plus a classification of treatments. It is a theory paper with no new data.
Key insight
In general medicine a tumour is a common cause: remove it and the cough, the chest pain and the breathlessness go together. Psychiatry has looked for such causes and, for most disorders, not found them. The paper’s answer is that there are none to find. The symptoms cause one another, and “connections between symptoms are often prosaic. If you do not sleep, you get tired”. If those connections are strong enough the loop sustains itself, and the person stays ill after whatever started it has passed. A disorder is then a place the system can get stuck, not a thing inside the person.
Assumptions
- Interventionist causation (Woodward 2003, Pearl 2000). A causal connection between symptoms means that intervening on one would change the probability distribution of the other.
- Agnosticism about realisation. Symptom–symptom connections may be biological, psychological, homeostatic or social. The theory does not say which.
- Principle 2, symptom-component correspondence. The nodes are the symptoms of the diagnostic manuals, at the right grain. Anything else (neural states, genes, processes not in the manuals) must enter by constituting a symptom, constituting a connection, or acting from the external field. The paper calls this principle “quite strict”.
- The external field. Triggers such as life events, and also processes inside the person with no node in the network (inflammation, and possibly “abnormal brain functioning”), act on the network from outside it.
- Hysteresis in realistic parameterisations. The dynamic principle rests on Cramer’s thesis (2013) and Cramer et al. (in press), which the paper says “proved the existence” of hysteresis in “realistically parameterized symptom networks”. This reading takes that on trust.
Key results
There are no theorems or new data. The results are the principles and the definitions that follow from them.
- Principles 1–4 (section Network theory):
- P1 Complexity: disorders are best characterised by interactions among components of a psychopathology network.
- P2 Symptom-component correspondence (above).
- P3 Direct causal connections between symptoms generate the network.
- P4 Mental disorders follow network structure: the network has a non-trivial topology, and its tightly connected groups are what present as syndromes.
- Principle 5, hysteresis (section The dynamics of symptom networks): “Mental disorders arise due to the presence of hysteresis in strongly connected symptom networks, which implies that symptoms continue to activate each other, even after the triggering cause of the disorder has disappeared.” Development runs in four phases (Figure 3): dormant network; activation by an external event; spread through connections; self-sustaining activity after the event is gone.
- Definitions (Table 1). Mental health is “the stable state of a weakly connected network”, an equilibrium the system returns to, not an absence of symptoms. A mental disorder is “the (alternative) stable state of a strongly connected network”. Resilience is the disposition of a weakly connected network to return quickly to health. Vulnerability is the disposition of a strongly connected network to tip into disorder when perturbed. Liability differences between people are differences in network parameters. The table crosses connectivity with external stress: weak/low is resilient health, weak/high is elevated symptoms (normal grief), strong/low is vulnerable (possibly a remission), and strong/high is disorder.
- Comorbidity is intrinsic. Bridge symptoms such as insomnia, fatigue and concentration problems belong to several clusters, so activation crosses diagnostic borders (Figure 2).
- The DSM-5 is not theory-neutral with respect to networks. Its criteria require causal couplings between symptoms (obsession driving compulsion), context conditions and “negative causal specifications”.
- Treatment. Symptom interventions (an antipsychotic suppressing a delusion), external-field interventions (stopping precipitating drug use), and network interventions that change connections (CBT, so the delusion “no longer has the effect of causing social isolation”). Success will “generally require a combination of network interventions (in order to make the healthy state accessible) and symptom interventions (to knock the system into that healthy state)”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | For all but a few disorders no common pathogenic pathway has been found | moderate: a reading of the literature, cited | Kendler 2012, Kendler et al. 2011 |
| C2 | Symptoms are directly causally connected | weak-to-moderate: indirect evidence only | clinicians’ and patients’ causal reports; experience-sampling interactions; partial correlations among DSM symptoms (ref. 31), which the paper calls “indirect” |
| C3 | Syndromes are the tightly connected clusters of one symptom network, so factor analyses approximate network structure | interpretive | Principle 4; no test given |
| C4 | Comorbidity is an intrinsic consequence of bridge symptoms | moderate as a consequence of P1–P4; empirical size unstated | Figure 2, argument |
| C5 | Strongly connected symptom networks show hysteresis, and this is how disorders outlast their triggers | moderate for the model (cited simulations); weak as an empirical claim about patients | Cramer 2013; Cramer et al. in press; PTSD, bereavement and abuse offered as fitting cases |
| C6 | Health, disorder, resilience and vulnerability are states and dispositions of networks of differing connectivity | definitional, given C5 | Table 1 |
| C7 | The DSM-5 encodes causal network structure in its criteria | moderate: examples from the manual | OCD and substance-use criteria |
| C8 | Effective treatment generally needs a network intervention plus a symptom intervention | speculative | argument from the model |
| C9 | The theory fits episodic disorders with a clear onset better than slowly developing ones (autism, personality disorders, parts of schizophrenia) | the author’s own scope judgement | Conclusions |
Concepts
- symptom network: symptoms as nodes, with a connection wherever two symptoms directly activate each other.
- external field: conditions that act on symptoms from outside the network. It is “external relative to the psychopathology network, but not relative to the physical boundaries of the person”.
- bridge symptom: a symptom in more than one disorder’s cluster, through which activation and comorbidity spread.
- hysteresis: removing a trigger does not undo the activation it caused in a strongly connected network.
- mental health: the stable state of a weakly connected network, an equilibrium, not the absence of symptoms.
- mental disorder: the alternative stable state of a strongly connected network, separated from health by hysteresis.
- network intervention: a treatment that changes symptom–symptom connections rather than the symptoms or their triggers.
Connections
- The medical model. The paper’s opening contrast is with the common-cause model of general medicine. That is the biomedical model Engel (LIT-548) challenged in 1977, though Borsboom’s objection is different: not that the model omits the psychological and social, but that psychiatry has found no common causes for it to name. Engel is not cited.
- Kendler. The premise C1 is cited to Kendler (2012) and Kendler, Zachar and Craver (2011), and Kendler is thanked for comments. The integrative ambition in the Conclusions (biological, psychological and sociological levels in one framework) is close to the explanatory pluralism of Kendler’s 2005 paper (LIT-552, unread here).
- Process-based therapy (LIT-544). Hofmann and Hayes treat networks as an extension of functional analysis and an “analytic alternative to the latent disease model”. Borsboom’s network interventions and idiographic network detection (through experience sampling or perceived causal relations) are the diagnostic half of what they want.
- CBT’s maintenance models (LIT-538). Salkovskis and colleagues’ maintaining “vicious circles”, in which behaviour sustains the negative meaning that drives it, are loops of exactly this kind. The difference is in what the nodes are. Borsboom’s nodes are manual symptoms by Principle 2. In the cognitive model the loop runs through appraisals and safety-seeking behaviours, which the manuals do not list. By Principle 2 those would have to be constituents of a symptom or of a connection.
- Emergence (LIT-150). Rizi treats phase transitions and order parameters as the mark of emergence. Borsboom borrows hysteresis from that physics (citing van der Maas and Molenaar’s catastrophe-theory account of cognitive development) without an order parameter or a measured transition.
Bearing on the record
- No THEORY is indicated by this reading alone. A candidate, “mental disorders are alternative stable states of strongly connected symptom networks, maintained by feedback after their triggers have gone”, would be Proposed at most. Promotion would need idiographic time-series evidence of hysteresis or of early warning signals (ref. 15, Wichers et al. on critical slowing down, is cited but not read here).
- A convergence across the batch. This paper, the cognitive maintenance account of LIT-538 and the process-based programme of LIT-544 all say that disorder is maintained by self-reinforcing loops rather than by a single latent cause. They come from three traditions (psychometrics, Beckian CBT, contextual behavioural science). See the batch report for a proposed THEORY.
- No ML instruction. The network estimation methods are statistics read for psychiatry. Nothing here belongs in the anthology.
- Filed, with LIT-538 and LIT-539, as THEORY-058, Proposed.
Limitations
- The evidence for direct causation is indirect, as the paper says: conditional associations and people’s reports of causal relations among their symptoms.
- Principle 2 is a stated idealisation. If the important variables are not manual symptoms, the networks are drawn over the wrong nodes.
- Slow-developing disorders are outside the clear fit, by the author’s own account; a “Russian doll of networks within networks” is suggested for them, not built.
- No mathematical model is given here. The paper says the theory “can be represented in mathematical form” and points to Epskamp et al.’s network psychometrics and Marsman et al.’s low-rank Ising networks for it.
- The treatment taxonomy is untested. No study is cited in which network interventions are distinguished from symptom interventions by their effect on network structure.
Open questions
- Is hysteresis observable in individual patients’ time series, and does its presence predict relapse after the trigger is gone?
- Does CBT in fact work by weakening connections (a network intervention), as the paper’s example assumes, or by moving symptoms (a symptom intervention)? This is the mechanism question LIT-538 says the field has not answered.
- What replaces Principle 2 if processes outside the manuals (appraisals, avoidance, rumination) carry the loops?
Corrections
- none to a seeded skim (there was no seed)
- Reference slip. Ref. 3 dates Kendler’s “dappled nature” paper to 2013 (Mol Psychiatry 17:377–88). Crossref gives volume 17, issue 4, published online 10 January 2012.