Article summary with an Network Analysis Transform Psychopathology? by McNally - 2016

Summary with the article: Can Network Analysis Transform Psychopathology? - McNally (2016)

Highlights

• Latent variable approaches to mental disorder are conceptually flawed.
• Network analysis views disorders as causal systems of interacting symptoms.
• Network analysis has clinical implications.

One purpose of this article is to contrast network analysis with traditional approaches, and consider its strengths and limitations. A second purpose is to review novel computational methods that could enable researchers to discern the causal structure of disorders.

1. Can network analysis transform psychopathology?

Three years ago the American Psychiatric Association (APA) released the DSM-5 amidst a storm of controversy. Right after this unveiling, the director of the National Institute of Mental Health (NIMH) announced that they would be re-orienting its research away from DSM categories, and favouring grant proposals targeting transdiagnostic mechanisms (and not the DSM).

Around this time, was also the debate about whether disorders should be conceptualized categorically or dimensionally. Do mental disorders differ by kind or degree?

Categorical and dimensional approaches attempt to explain a fundamental observation of our field: psychiatric symptoms don’t occur randomly; some are more likely to co-occur than others.

• The categorical account suggests a discrete entity as the common cause of symptom emergence and covariance.
• The dimensional model supposes a common underlying cause of symptom emergence and coherence.

One reason for the persistence of this debate is that the strengths and weaknesses of the categorical model are the mirror images of those of the dimensional model leaving few psychopathologists fully satisfied with ether approach.

2. The network approach to psychopathology

According to Denny Borsboom’s and his colleagues’ network model of psychopathology, an underlying latent variable is not the common cause of symptom covariance. It rather emerges from the casual interactions among symptoms themselves. Symptoms are not reflective of underlying mental disorders; they are constitutive of them.

According to the network perspective, an episode of disorder occurs whenever the requisite number of symptoms become activated for a sufficient duration. Recovery from a disorder happens when symptoms deactivate and/or when the links between them disappear. Thus, a mental disorder constitutes a causal system of dynamically interacting and possibly self-reinforcing symptoms.

Clinical researchers have applied network methods to explain the causal interactions between symptoms constituting depression, PTSD, OCD, schizophrenia, childhood disorders, social anxiety disorder, substance abuse/dependence, and persistent complex bereavement disorder. Purpose of this article is to review the network analysis of psychopathology, and consider its strengths and weaknesses.

3. Key concepts in network analysis

Networks consist of nodes and edges. Nodes representing objects of study, and edges representing the connections between them. In psychopathology, nodes represent symptoms, and edges the associations between symptoms.

Networks can have weighted or unweighted edges. An unweighted edge shows that two symptoms are connected, and a weighted edge shows the magnitude of the connection (shown through thickness of the edge). Association between symptoms can be positive or negative (green or red).

Edges can be undirected or directed. Undirected networks consist of edges that connect pairs of symptoms, signify an association but not causal or allowing the direction of prediction to go both ways. Directed networks consist of edges with arrow tips, indicating direction of prediction and perhaps causation.

4. Node centrality metrics

Traditional categorical approaches to psychiatric diagnosis emphasize distinctive symptoms that are strongly associated with a single disorder, but rarely with other disorders. Some proposed purifying diagnostic criteria sets of nonspecific symptoms that appear in many disorders, and leave only those strongly associated with the syndrome.

But nonspecific symptoms appearing in many diagnostic criteria sets may be especially important. They may serve as bridges linking two syndromes e.g. major depression and generalized anxiety disorder. Activation issuing form a bridge symptom can spread to both syndromes – producing diagnostic comorbidity.

Network analysis computes metrics of node centrality as opposed to focusing on distinctive symptoms unique to a disorder. Highly central nodes are most important in the network. Five measure of centrality are degree, strength, expected influence, closeness, and betweenness.

• 4.1 Degree Centrality: a node’s degree is the number of edges connected to it, the higher the degree, the more central it is to the network. E.g. social network with individuals (nodes) and friendships (edges) between pairs.
• 4.2 Strength Centrality: weighted networks enable computation of node strength. The edge thickness represents the magnitude of the association and thus the probability that activation of one node will be associated with activation of nodes connected to it. Strength centrality is the sum of the weights (e.g. correlation coefficients) of the edges connected to a node. Especially important for psychopathology network as it reflects the likelihood that activation of a certain symptom will be followed by activation of other symptoms. Symptoms scoring high on strength centrality may require early intervention to prevent episodes of disorder. For directed networks, one can compute the in- and out-strength centrality of nodes. High out-strength node is source of activation for nodes receiving its edges and high in-strength is the recipient of activation emanating from other nodes.
• 4.3 Expected Influence: computation of strength centrality concerns a node’s incident edges weighted by their absolute magnitude, which is fine as long as there are no negative edges. Expected influence takes into account both negative and positive edges. Matches performance of strength centrality when there are only positive edges, but outperforms it when there are also negative edges.
• 4.4 Closeness Centrality: the closeness of a node is the average distance from that node to all the other nodes in the network.
• 4.5 Betweenness Centrality: to determine betweenness centrality of a node, one has to calculate the shortest path length between each pair of nodes in the network first. The betweenness of anode is the number of times that node lies on the shortest path between to other nodes. Symptoms shared by two often-comorbid disorders are high on betweenness centrality and serve as bridges between the two disorders.

5. Types of networks

Psychopathologists have computed several types of networks, most concerning cross-sectional, observational symptom data. Network analysts have devised methods that can bring us closer to characterizing mental disorders as causal systems. (check article for visual examples of each network)

• 5.1 Association Networks: the most basic type (qgraph). Can only infer association and not causation. It is weighted and undirected. Places nodes with stronger correlations near the center.
• 5.2 Partial Correlation Network: computes partial correlation between a pair of nodes/symptoms after adjusting for the influence of all other nodes in the network. So edges appearing in an association network that remain after adjustment are good candidates for counting as causal connections.
• 5.3 Relative Importance Networks: partial correlation networks only depict direct associations between symptoms, but the network itself is not directed. In a relative importance network, each edge depicts the relative importance of a symptom as a predictor of another symptom. Describes strength and direction of prediction, but still not causation.
• 5.4 Bayesian Networks: a parametric method that produces directed acyclic graphs (DAGs). A directed network whereby each edge has an arrow tip, signifying direction of prediction and possibly causation. A DAG lacks cycles. DAGs also reveal the percentage of bootstrapped samples in which the edge appeared in the direction depicted in the final, averaged Bayesian network. Finally, it also computes a BIC value for each edge, which indicates how damaging it would be to model fit if one were to remove the edge form the network. High absolute BIC values = how important the edge is to the model.

6. Strengths and limitations of network analysis?

A potentially fatal objection to latent variable approaches to psychopathology is their failure to satisfy the axiom of local independence requisite for justifying an inference to an underlying entity as the common cause of symptom emergence and covariance.

Most network studies rely on single self-report measures of specific symptoms; this raises concerns that such measures may not fully capture a clinical phenomenon. To try and fix this, network theorists could integrate multiple measures of a symptom. Which may reintroduce latent variables, but network theorists have distinguished latent variables that have a natural referent versus abstract ones that don’t.

Focusing on symptoms – “symptomics”. Symptoms can vary in their risk factors, the kinds of life stressors triggering them, and their impact on different aspects of psychosocial functioning.

Even though some theorists believe that latent variable research is revelatory of ‘core psychopathological processes’ it is questionable whether these variables tell us anything about the processes operating within individuals.

Bayesian network analysis has potential limitations. DAGS tend to emphasize causal relations as if they all occur at once but causal relations in psychopathology occur over different time scales. DAGS also prohibits looping effects whereby a node early in the chain events activates other nodes that wind up influencing the state of the originating node – cycles can’t happen.

7. Future Directions

The aim of research in abnormal psychology is to discover the causes of mental health problems and enhance the efficacy of prevention and treatment. But many questions concerning causality are unanswerable with experimental methods. Network analysis aims to clarify the causal processes among symptoms that lead to episodes of disorders.

Researchers have begun to introduce a temporal dimension to network analysis that moves us closer to discerning causality.

Network analysis has identified markers of impending tipping points whereby people shift suddenly from a health to depressed state.

Network analysis may transform the field of psychopathology in important ways. Advances in quantitative methods, computational power, and mobile technology will pay if clinical researchers can use idiographic network methods to guide therapeutic intervention. No single method in the field of psychopathology is likely to provide answers to all the questions. Yet network analysis holds promise as both a scientific and practical approach to conceptualizing and guiding treatment of these conditions.