Network Analysis: An Integrative Approach to the Structure of Psychopathology - summary of an article by Borsboom and Cramer (2013)

Critical thinking
Article: Borsboom, D. and Cramer, A, O, J. (2013)
Network Analysis: An Integrative Approach to the Structure of Psychopathology
doi: 10.1146/annurev-clinpsy-050212-185608

Introduction

The current dominant paradigm of the disease model of psychopathology is problematic.
Current handling of psychopathology data is predicated on traditional psychometric approaches that are the technical mirror of of this paradigm.
In these approaches, observables (clinical symptoms) are explained by means of a small set of latent variables, just like symptoms are explained by disorders.

• From this psychometric perspective, symptoms are regarded as measurements of a disorder, and in accordance, symptoms are aggregated in a total score that reflects a person’s stance on that latent variable.
• The dominant paradigm is not merely a matter of theoretical choice, but also of methodological and pragmatic necessity.

In this review, we argue that complex network approaches, which are currently being developed at the crossroads of various scientific fields, have the potential to provide a way of thinking about disorders that does justice to their complex organisation.

• In such approaches, disorders are conceptualized as systems of causally connected symptoms rather than as effects of a latent disorder.
• Using network analysis techniques, such systems can be represented, analysed, and studied in their full complexity.
• In addition, network modeling has the philosophical advantage of dropping the unrealistic idea that symptoms of a single disorder share a single causal background, while it simultaneously avoids the realistic consequence that disorders are merely labels for an arbitrary set of symptoms.
  • It provides a middle ground in which disorders exists as systems, rather than as entities

Symptoms and disorders in psychopathology

We know for certain that people suffer from symptoms and that these symptoms cluster in a non-arbitrary way.
For most psychopathological conditions, the symptoms are only empirically identifiable causes of distress.

• Mental disorders are themselves not empirically identifiable in that they cannot be diagnosed independently of their symptoms.
  • It is impossible to identify any of the common mental disorders as conditions that exists independently of their symptoms.

In order for a disease model to hold, it should be possible to conceptually separate conditions from symptoms.

• It must be possible (or at least imaginable) that a person should have a condition/disease without the associated symptoms.

This isn’t possible for mental disorders.
As an important corollary, this means that disorders cannot be causes of these symptoms.
This strongly suggests that the treatment of disorders as causes that exists independently of the symptoms used to identify them involves an unwarranted reification.

The relation between symptoms and disorders has to be conceptualized differently.

• Mental disorders are not identifiable as separate disease entities
• There appear to be many direct relations between symptoms
• Thee symptom-symptom relations are not only likely to produce a considerable part of the empirical covariance between symptoms, but may also play an important generative role in the etiology of a disorder.
• In psychopathology, research suggests that mental disorders may be caused by the direct activation of symptoms.

In sum, only do we not know that symptoms are caused by mental disorders, but it is in fact extremely unlikely that they are.

Complex psychopathology networks

The foundation of the network approach is simple.
Instead of interpreting symptoms as a function of a set of underlying/latent disorders, the network approach conceptualizes symptoms as mutually interacting, often reciprocally reinforcing, elements of a complex network.

• Rather than interpreting symptoms as measurements of a latent disorder, symptoms are viewed as part of a causal system.
• The relation between symptoms and disorder becomes one of mereology (a part-whole relation) rather than measurement (a causal relation).
• Psychopathology symptoms are not symptoms in the strict sense of the word.
  • Symptoms are causally active ingredients of the mental disorders themselves

The move from latent disorders to networks of causally connected symptoms is in itself a simple and straightforward matter.
It does not involve the acceptance of any particular theory about psychopathology.
It merely results from accepting two simple propositions

• Given the current evidence, we should forestall the conclusions that symptoms of the same disorder are uniformly caused by a single psychological or biological condition
• Psychopathology symptoms causally influence one another

The consequences of accepting these propositions are potentially radical

• If it is indeed the case that direct and possibly reciprocal interactions exist between symptoms, then it becomes unclear whether the disorder itself is at all required as a separate entity to make sense of the empirical correlational structure of symptoms
• If one accepts that symptoms and causal connections between them are are what constitutes a mental disorder, the term ‘comorbidity’ gathers a different meaning.
  • No longer can comorbidity be meaningfully explained as a correlation between two disorders, nor as the result of a common underlying dysfunction or ‘super disorder’.
  • Instead, the causal relations between symptoms constitute pathways that can connect different disorders, for example via bridge symptoms

Bridge symptoms: symptoms that are part of both disorders.

Such multiple pathways from one disorder to another might exist in such a way that there is no objective or ‘true’ point at which to carve the symptom network in two, with each part representing a separate disorder.
Boundaries between disorders are fuzzy. This is a result of the intrinsic structure of disorders.
The reason that we have been unable to find true boundaries is simply that there are no true boundaries.
Although, in the network approach, one may still define disorders as a set of more densely connected symptoms that show synchronized behaviour, these disorders are literally intertwined with one another and cannot be neatly separated.

• A final consequence of accepting the premises of the network approach is that, with a shifting focus of scientific attention, the target of therapeutic interventions may change.
  • Instead of some ephemeral ‘latent disorder’, therapeutic interventions target symptoms and the relations between symptoms
  • In a network approach, interventions are optimally targeted at the symptoms themselves or at the causal relations that connect them.

Constructing and analysing psychopathology networks

At its core, a network is simply a set of elements (nodes) that are connected through a set of relations.
Elements as well as relations between elements ca be virtually anything.
The construction and analysis of networks are highly accessible in the sense that the application of network models does not require extensive prior knowledge. All one needs is a set of elements and an idea of how these elements are connected.

Psychopathology networks can be constructed in several ways, each of which may yield important information about the structure of disorders.

• One can use the information in diagnostic systems themselves, as these often contain clues about the causal constitution of disorders
• One can use the assessment of (causal) relations between symptoms, as rated by clinicians or patients
• One may use data on symptom endorsement frequencies to extract empirical patterns of association that can serve as input for network structures

Networks based on diagnostic systems

Diagnostic systems like the DSM-V can be considered to partly reflect the structure of psychopathology through patterns of symptom overlap.
A straightforward way of studying such patterns is by representing individual symptoms as nodes in a network and connecting them whenever they feature as symptoms of the same disorder.
This type of network reveals the structure of the diagnostic system itself.
A giant component: a large group of nodes that are all connected to one another, either directly or via intermediary nodes
A small world in the network analysis literature: on average, paths from one node to another are short and there is a large degree of clustering
Clustering: the extent to which nodes tend to form a connected group

Small world structure implies that, even though a network may be very large and feature strong clustering, any node can be reached from any other node in only a few steps.

• For the DSM network, the small world property means that comorbidity appears to be, at least partially, encoded in the structure of the diagnostic criteria themselves.

It is not entirely unreasonable to suspect that the network may harbour relevant causal information.

• The DSM itself frequently mentions (or even requires) causal relations between symptoms of the same disorders.
  • Causal links that are not explicated in the system may occasionally be highly likely.

Perceived causal relations

A second way of gaining insight into the causal organization of disorders is by asking experts or patients to report causal relations between symptoms.

Causal relations scaling

• A person indicates which of a set symptoms is present
• Each combination of presented symptoms (q,j) is combined in a question that assesses whether q caused j (reciprocal causal relations are typically allowed).

In this way, one essentially builds a self-reported adjacency matrix for all symptom-symptom relations.
That matrix defines a network that represents the cognitive representation of causal structure of disorders.

Extended psychopathology networks

Networks for psychopathology feature relations between symptoms.
Typically, we see these symptoms as interacting with one another at the level of the individual person.
However, in some cases, one person’s symptom may infect another person.

Extended psychopathology networks: a network in which the activation of one person’s symptom not only has produced other symptoms within his own system, but also in the system of another person.
Extended psychopathology networks may be studied in more or less the same way as ordinary psychopathology networks, but are especially useful when time information is present, so that one can estimate person-specific networks as well as the way they interact.
Such methodology could be used to chart the interaction between symptoms of different people in various social situations.

In almost any mental disorder, significant social effects of that kind exist.
In general, prolonged severe problems lead to a greater degree of social isolation.
This means that the way in which one person’s symptom network interacts with other people’s networks leads to the alteration of that person’s social network.
Reciprocal influence is likely to be the norm, rather than the exception, for the development of social isolation.

Thus, the complexity of psychopathology not only involves complex reciprocal relations between symptoms but also between networks of symptoms and social networks.

Association and concentration networks

Another way of exploring the causal organisation of mental disorders is by studying empirical associations between symptom reports in patient or community samples.

Association networks are very useful for seeing at first glance which clusters or symptoms tend to be strongly connected or not.
However, if one is interested in knowing which of these symptoms are truly related, then correlations may not provide optimal information.
That is because a high correlation between any tow symptoms might be the result of:

• A true direct (possibly reciprocal) relation between these symptoms
• A third variable that causes both symptoms
• Selection on a common effect of the symptoms

How can one figure out which correlations are indicative of direct causal relations and which are not?

• Obtain the matrix of partial correlations
  • Partial correlations: the correlations between pairs of symptoms that remain when all other symptoms are controlled for
  • This may be considered to provide clues about the causal skeleton of a network
• A matrix of this is a concentration graph

In the traditional disease model, the most interesting individual differences are to be found at the level of risk factors/dysfunctions that cause a particular disease.
Individual differences exist at the level of symptoms.

A network perspective predicts that relevant differences arise at the level of the symptoms and the relations between them rather than at the level of the disorder.
Concentration graphs in particular are useful for an assessment of which pathways between symptoms appear to be common.

• Strong partial correlations in a between-subjects weighted network may indicate that these pathways reflect real causal relations that are relatively common in the sample in which the network representation is based.

Directed networks

Association and concentration graphs provide clues about possible causal relations between variables, but they do not provide information about the direction of causal relations (if these relations are unidirectional in the first place).
Unidirectional causal relations between nodes are typically represented by arrows.
Causal analysis is easiest when the pattern of causal relations among variables creates a directed acyclic graph.

• In such a graph, all connections between nodes are directed, and it is not possible to visit any node more than once when traversing the edges along the direction of the arrows in the graph (there are no feedback loops).

Under a (strict) set of statistical assumptions, the causal network structure can be deduced from a set of observational data by exploiting the connection between causal relations and certain patterns of conditional independence.

The many roads to disorder: individual networks

From a network perspective, each individual may have his or her own network, which comes with specific vulnerabilities or risk factors.

Time series, time series, and time series

When the aim of network analysis is to construct disorder networks for individuals, cross-sectional data will be of little use.

In networks of individuals, an arrow between any two symptoms is indicative of a process that takes place over time.
As such, querying a person about his or her symptomatology at one point in time is simply not enough to extract the causal information necessary to build a network of this person’s symptom space.
It is possible to ask people to draw their own causal scheme, but of course the success of such a method relies on the ability of people to accurately report on their symptom development retrospectively, which may not be equally accurate in all circumstances.

A viable alternative is to collect time-series data.
Time-series data: one asks individuals to report on various aspects of their physiological and psychological well-being at least once a day for many consecutive days.
Advantages:

• One is able to collect time-intensive data
• One can collect data on the relation between events happening in a person’s life and the subsequent ripple effects of that event in the symptomatology of this person
• Once and collect data from people without psychopathology who might be progressing toward developing a mental disorder.

Another possibility to learn about the intraindividual behaviour displayed by a given network structure is by simulating time-intensive intraindividual data.

Risk in individual networks

In disease, risk is defined at the level of disease entity, which is not present in a network, at least not as an entity that is separable form its symptoms.
From a network perspective, there at (at least) two ways in which a network can harbour risk of developing a certain mental disorder.

• The structure of a particular network might be risky.
  When the connections between symptoms are strong.
  (Like domino’s placed close together).
• There might be symptoms that, when developed in a particular person, have stronger causal influence on the rest of the network compared to other symptoms.
  Different symptoms pairs will have different connection strengths, which determine the extent to which symptoms causally influence one another.
  Central nodes in someone’s network are most dangerous.
  • If a central symptom is developed in someone, then the probability of that symptom causing the development of other symptoms is high.

Instead of defining risk or liability at the level of the disease, the network perspective offers two concrete explanations of why certain people are at risk while others are not.

• It potentially gives therapists specific targets of where to intervene either to prevent the development of a full-blown disorder or to treat a person who already has developed a disorder.