Network

Axioms

Psychological network systems map symptoms (variables) on a network structure according to three axioms

  1. Interconnected symptoms directly or indirectly
  2. Causal dynamic relationship between symptoms
  3. Factors outside the system influence the system as a whole or its elements

Validity

Instrument measures what is supposed to measure.

Valid or not valid?

  • Apply an instrument designed for identifying PTSD to determine whether someone is depressed
  • Apply an instrument designed for identifying anxiety in people suspected they may be anxious
  • Evaluate bi-polar disorder with an instrument designed to assessment quality of life

Reliability

Instrument will give the same result every time when used to measure.

Reliable or unreliable?

  • An instrument gives different results every time it is applied
  • An instrument changes unit periodically and needs to be adjusted before using
  • An instrument is new and not yet fielded

Validity and reliability in real life

Accurate test results in the healthcare domain is critical!

Accurate test results in the healthcare domain is critical!

Interventions for mental health and wellbeing must draw on valid and accurate assessments!

PHQ-9 for depression assessment

PHQ instrument; see Kroenke & Spitzer (2002)

PHQ instrument; see Kroenke & Spitzer (2002)

An (incomplete) list of assessment instruments for depression from the APA:

Children and adolescents

  • Behavior Assessment System for Children (BASC)
  • Child Behavior Checklist (CBCL)

Older adults

  • Geriatric Depression Scale (GDS)
  • Life Satisfaction Index (LSR)

Can you think of reasons why some instruments were developed in specific age groups?

Chance or a fact?

Edge 1

Edge 2

Edge 3

Network

How to make sure that the network estimation is stable such that edges (i.e. relationship between symptoms) is mapped together as a network that is more likely to be present in the data than a chance result?

Parsimony as a goal

Network estimation from no connected nodes to all nodes connected.

Network estimation from no connected nodes to all nodes connected.

Regularization technique

Such as the graphical GLASSO, to determine an optimal balance between keeping edges that are most likely zero or insignificant and maintaining a simple network for interpretation purposes.

Confidence in estimations

Estimation uses mathematics to determine the probability that the observed data or a more extreme one is indeed the observed one under the null hypothesis as opposed to being a chance result.

(In)famously better known as the p- value significance level.

The null hypothesis states that there is no effect/ no difference.

  • 90% significance level = more liberal; greater room for accepting chance results
  • 95% significance level = standard threshold in social sciences
  • 99% significance level = conservative; leaves minimal room for accepting chance results

Which significance level would be appropriate to use in developing clinical interventions?

Confidence in estimations

Confidence is how confident are we allowed to be that the found effect is the observed one.

(In)famously better known as the confidence interval.

95% CI [lower bound; upper bound]

If contains 0 (zero), the result is statistically not significant and it is more likely to be a chance result.

Summary

Psychological network systems map symptoms on a network structure according to three axioms

  1. Interconnected symptoms directly or indirectly
  2. Causal dynamic relationship between symptoms
  3. Factors outside the system influence the system as a whole or its elements

Three methodological criteria must hold in mapping symptoms on a network system:

  1. Valid and reliable assessment instrument
  2. Stable estimated network
  3. Accurate estimated results

Homomorphism

A structure existent in the real world can be reproduced accurately in a geometrical space with the help of numbers to which we attach a meaning.

A group of friends is a social network.

A group of friends is a social network.
  • What might we want to examine between persons? And why?
  • What might we want to examine within persons? And why?
  • What might we want to examine between and within persons? And why?

Depressive symptoms.

Depressive symptoms.
  • What might we want to examine between persons? And why?
  • What might we want to examine within persons? And why?
  • What might we want to examine between and within persons? And why?

Typical research approach

See Borsboom et al. (2021)

See Borsboom et al. (2021)

PNS publications over the years

See Robinaugh et al. (2020)

See Robinaugh et al. (2020)

Re-defining abnormal mental health

What if …?

abnormal mental health as a long-term activation of a network of causally interconnected symptoms whereby each can influence another either directly or indirectly whereas external factors may impact the resulting system entirely or specific formative elements

See Robinaugh et al. (2020)

See Robinaugh et al. (2020)

Symptoms are the network nodes.

Associations between symptoms are the network edges.

Through network analysis the nodes and edges can be systematically examined.

In systematic examinations, chance results or randomness is accounted for.

Panick attack re-visited

See Robinaugh et al. (2020)

See Robinaugh et al. (2020)

See Robinaugh et al. (2020)

See Robinaugh et al. (2020)

From theory through data to numbers and statistical models.

See Robinaugh et al. (2020)

See Robinaugh et al. (2020)

From theory through data to numbers and statistical models.

…philosophical yet mathematical

  • Variables, which define the mental disorder that we examine
  • People, whose experience of a disorder we want to understand
  • Time, which can mean that the experience of a disorder differs at different moments in time

Cross-section

One problem with cross-sectional data is that the shape of a cross-section dictates what and how much one sees from the bigger picture.

One advantage however is that it is practical and cost effective. Fairly accurate too if limitations are acknowledged.

Cross-sectional networks

Depict relations between pairs of nodes (symptoms) involving many people assessed at a single point in time.

Types of networks that use cross-sectional data:

  1. Association = undirected associations
  2. Concentration (partial correlation) = eliminates spurious associations by adjusting for the influence of remaining nodes
  3. Relative importance = each edge signifies the relative importance of a node as a predictor of another node
  4. Regularized partial correlation = adjust partial correlations for the influence of the remaining nodes based on parsimony principles (our focus)
  5. Bayesian (Directed Acylic Graphs) = an ambitious method to “make believe” causality from cross-sectional data

The final choice depends on available data and research question.

Association network

See McNally (2016)

See McNally (2016)

Network depicts PTSD symptoms in adults reporting histories of childhood sexual abuse.

Note the cloud of edges suggesting indiscrimination between what is statistically relevant and what is not.

Regularized partial correlation network

See McNally (2016)

See McNally (2016)

Network depicts PTSD symptoms in adult survivors of the Wenchuan, China earthquake.

Note the fewer edges suggesting that only statistically relevant edges had been mapped.

Relative importance network

See McNally (2016)

See McNally (2016)

Network depicts the strength of a PTSD symptom as a predictor of another PTSD symptom in adults reporting histories of childhood sexual abuse.

Note, again, the cloud of edges suggesting indiscrimination between what is statistically relevant and what is not.

This time, however, note the arrows of varying thickness suggesting that some directionality may be stronger than other and thus we gain one further step toward assessing causality.

Directed Acylic network

See McNally (2016)

See McNally (2016)

Network depicts PTSD symptoms in adult survivors of the Wenchuan, China earthquake.

Note the different mapping style.

Note the top-down structure as signaled by the arrow directions suggesting there is only one direction possible: from the upper situated node to the lower situated nodes.

And, thus, we are one further step closer to assessing causality.

Comorbidity

Comorbidity is a natural consequence of partially overlapping symptomatic clusters. (Read more McNally (2021))

For example, insomnia and concentration impairment in generalized anxiety disorder (GAD) and major depressive disorder (MDD).

How to accurately assess the bridge symptoms?

Replicability

Something may or may not intervene.

Something may or may not intervene.

If results are not replicated in distinct cross-sectional data, does this mean that

  • the method is flawed
  • the phenomenon unfolds differently in different populations?

See the network stability and edge estimation accuracy (Epskamp et al., 2018).

Ontology

Galileo’s heliocentric theory changed everything

Galileo’s heliocentric theory changed everything

Seeing mental disorders as latent constructs and as resulting from dynamics of network systems are statistically exchangeable.

But it has clinical importance whether a mental disorder is the common cause of symptoms or emerges from a causal dynamic network of associations between symptoms.

Centrality

How to define the importance of a symptom?

For a diagnosis, identify symptoms occuring in one but not other disorders.

For a treatment, identify the most severe or dangerous (e.g., suicidal thoughts) symptoms.

For a network approach, identify those symptoms that score high on node centrality indicators.

But, can the identified highly central symptom help interventions if the data are cross-sectional?

Further reading McNally (2021)

Temporality

Time and space, Salvador Dali

Time and space, Salvador Dali

At least two measurement points required to test the therapeutic promise of a centrality metric.

The node centrality indicator “expected influence centrality” is a promising metric.

Overview: Between persons

What can one do and cannot do with cross-sectional network data?

Can

  • Cost-effective data
  • Explore
  • Network structure, nodes and edges at one point in time and/ or a population
  • Group level

Cannot

  • Temporal dependency
  • Frequency of patterns at the individual level

Possible questions

  • How are depression symptoms mapped in group X?
  • Are there group differences concerning network structure?

What other questions can you think of?

References

https://adrianstanciu.eu/

Borsboom, D., Deserno, M. K., Rhemtulla, M., Epskamp, S., Fried, E. I., McNally, R. J., Robinaugh, D. J., Perugini, M., Dalege, J., Costantini, G., Isvoranu, A.-M., Wysocki, A. C., Borkulo, C. D. van, Bork, R. van, & Waldorp, L. (2021). Network analysis of multivariate data in psychological science. Nature Reviews Methods Primer, 1(58), 1–58. https://doi.org/10.1038/s43586-021-00055-w
Epskamp, S., Borsboom, D., & Fried, E. I. (2018). Estimating psychological networks and their accuracy: A tutorial paper. Behavior Research Methods, 50, 195–212. https://doi.org/10.3758/s13428-017-0862-1
Kroenke, K., & Spitzer, R. L. (2002). The PHQ-9: A new depression diagnostic and severity measure. Psychiatric Annals, 32(9), 509–515. https://doi.org/10.3928/0048-5713-20020901-06
McNally, J. R. (2016). Can network analysis transform psychopathology? Behaviour Research and Therapy, 86, 95–104. https://doi.org/10.1016/j.brat.2016.06.006
McNally, J. R. (2021). Network analysis of psychopathology: Controversies and challenges. Annual Review of Clinical Psychology, 17, 31–53. https://doi.org/10.1146/annurev-clinpsy-081219-092850
Robinaugh, D. J., Hoekstra, R. H. A., Toner, E. R., & Borsboom, D. (2020). The network approach to psychopathology: A review of the literature 2008–2018 and an agenda for future research. Psychological Medicine, 50(3), 353–366. https://doi.org/10.1017/S0033291719003404