Reflective or formative?
On the often unspoken choice behind a diagnostic model, and why that choice can have consequences for treatment.
Suppose a patient scores high on various symptoms: sleep problems, fatigue, low mood, loss of interest and rumination. A factor analysis of the scales yields a common factor, which we call depression and which accounts for the association between the symptoms.
What, then, is depression? Two different answers to that question are possible, and they can lead to very different treatments.
Answer 1: The reflective model
In the reflective model, depression is an underlying entity (an illness, a state, a disposition) that causes the symptoms. The symptoms are then reflections or indicators of that underlying state, in the way that fever, headache and muscle pain together are the expression of influenza.
The implication for treatment is evident: interventions aimed at individual symptoms bring relief but do not cure. Paracetamol eases the headache, not the flu. Lasting improvement would then have to come from an intervention on the underlying cause.
As a diagram, it looks like this: one box in the middle (depression), with arrows pointing outwards (sleep problems, fatigue, low mood, and so on). A single common cause, in other words.
Answer 2: The formative model
In the formative model, depression is not an underlying entity but an index. The symptoms influence and reinforce one another, and what we call depression is a summary of that constellation. The factor then describes the configuration, but does not explain any causal mechanism.
Here the implication for treatment is different: an intervention on one symptom can set the whole system in motion. If symptoms reinforce one another, a change in one of them can carry through to the rest. Restoring sleep may thus break a spiral, not because sleep was "the cause", but because sleep is a node in the network.
As a diagram: no centre, but a network of symptoms pointing to one another with arrows, with a box drawn over them that we call "depression". That box is then a label rather than a causal entity.
Statistically indistinguishable
This is where a difficult issue arises. On the basis of correlational data (a questionnaire, a single measurement occasion, a factor analysis), the two models are statistically equivalent: they fit the data equally well. The choice between them is therefore in the first place a theoretical one rather than an empirical one.
"Underlying common causes are unsatisfactory if they cannot be identified independently of the observed relations they are supposed to explain." (after Van der Maas et al., 2006)
This idea is central to Denny Borsboom's argument against the reflective model in psychopathology. In medicine the reflective model works well because underlying entities can often be established independently of the symptoms. Influenza can be confirmed with a test; a heart valve can be seen on an ultrasound. There, the underlying cause has an existence of its own that can be pointed to.
In psychiatry this rarely succeeds. There is no test for depression that works independently of the symptoms, and no scan that shows a depression to be present. The "underlying factor" is inferred from precisely the symptoms it is supposed to explain. Strictly speaking, that amounts to a renaming rather than an explanation.
What this means in practice
What changes if we take the network model seriously? In our view, at least four things.
In diagnostics: we look not only at the severity score on a total scale, but also at the configuration of symptoms. Which symptoms are most central for this patient, which feed which others, and which are conspicuously absent?
In the case conceptualisation conversation: instead of a diagnosis, we draw a network together with the patient. This can become a tool with which she recognises her own pattern and sees for herself where change might take hold.
In treatment: there is no need to treat "the cause" first before doing anything about the symptoms. Working on a central symptom (often sleep, activation or social contact) is then not a second-rate treatment but a legitimate way of influencing a network. Interventions of proven effectiveness do not thereby become superfluous; they receive a different reading, as interventions on the network.
In the conversation with the patient: the reflective model carries a weight that the formative model carries less of. Saying "you have depression" readily suggests that something is wrong inside you that requires treatment. Saying "your sleep, mood, activity and rumination reinforce one another and are currently keeping one another in place" describes the same thing in more technical terms, and possibly in a way that leaves more room. The patient is then not the problem herself, but is caught in a pattern that feeds itself.
An honest caveat
The formative model is not automatically the better one. For some disorders, for example psychotic disorders with clear neurobiological correlates, the reflective model still has much to offer. And in depression or anxiety, too, there may be parts of the picture that do have an underlying cause, such as a trauma, a hormonal dysregulation or a social situation that is structurally constraining.
Network thinking is therefore a complement rather than a replacement: a second language alongside the first. For a large part of what we encounter in mental health care, it seems to us the more useful of the two.
For those who wish to explore it further, good methodological tools are now available: psychometric network analysis (qgraph, mgm, bootnet in R), longitudinal networks for within-person dynamics (for example mlVAR and graphicalVAR) and Ising models for binary symptom data. The methodology has developed rapidly and in some respects is even slightly ahead of the theory.
The choice between reflective and formative is no longer left unspoken, and that seems to us a good development. A choice we make explicitly is one we can better justify and, where necessary, revise.