Network Psychiatry
Investigating models of mental constructs and disorders
with complex systems and Bayesian Artificial Intelligence
Giovanni Briganti
These defendue pour l’obtention du grade de Docteur en Sciences Medicales
Jury
Prof. Samuel Leistedt, MD, PhD, Promoteur, Universite de Mons
Prof. Paul Linkowski, MD, PhD, Co-Promoteur, Universite libre de Bruxelles
Prof. Alexandre Legrand, MD, PhD, President du Jury, Universite de Mons
Dr. Quoc Lam Vuong, PhD, Secretaire du Jury, Universite de Mons
Prof. Pierre Manneback, PhD, Universite de Mons
Prof. Christophe Lelubre, MD, PhD, Universite de Mons
Prof. Adelin Albert, PhD, Universite de Liege
Prof. Pierre Thomas, MD, PhD, Universite de Lille
Network Psychiatry
Investigating models of mental constructs and disorders with complex
systems and Bayesian Artificial Intelligence
Giovanni Briganti
Abstract
Inspired by the network approach to psychopathology, this thesis aims to investigate several
mental constructs and disorders with complex systems and Bayesian Artificial Intelligence
in order to model interactions among construct features and symptoms, evaluate their re-
spective importance in the determination of the network structure, as well as offer new
methodological perspectives to be used in network studies.
This work is organized as follows: first, an introduction to the network approach in psy-
chiatric research and potentially in clinical practice will be introduced. Second, a statistical
introduction to the network techniques, both frequentist and Bayesian will be offered to the
reader. Third, I will investigate from a network perspective several important psychiatric
constructs found in the general population. Fourth, I will translate the network approach to
psychiatric disorders using both nonclinical and clinical samples. Finally, I will discuss the
implications of this work as well as set further challenges based on my analyses.
1
List of publications
The publications resulting from this dissertation are listed below.
Briganti, G., Kempenaers, C., Braun, S., Fried, E. I., and Linkowski, P. (2018). Network
analysis of empathy items from the interpersonal reactivity index in 1973 young adults.
Psychiatry Research, 265:87–92.
Briganti, G., Fried, E. I., and Linkowski, P. (2019). Network analysis of Contingencies
of Self-Worth Scale in 680 university students. Psychiatry Research, 272:252–257.
Briganti, G. and Linkowski, P. (2019). Exploring network structure and central items
of the narcissistic personality inventory. International Journal of Methods in Psychiatric
Research, e1810.
Briganti, G. and Linkowski, P. (2019). Item and domain network structures of the Re-
silience Scale for Adults in 675 university students. Epidemiology and Psychiatric Sciences,
pages 1–9.
Briganti, G. and Linkowski, P. (2019). Network approach to items and domains from
thetoronto alexithymia scale. Psychological Reports, page 0033294119889586.
Briganti, G., Scutari, M., and Linkowski, P. (2020). Network structures of symptoms
from the Zung depression scale. Psychological Reports, page 0033294120942116.
Briganti, G. and Linkowski, P. (2019). Une nouvelle approche ontologique et statis-
tiquedes constructions et maladies mentales : introduction a la psychiatrie des networks.
PsyArXiv
Briganti, G., and Linkowski, P. (2020). A machine learning approach to alexithymia
2
components. Psychiatria Danubina, Sep;32(Suppl 1):180-187.
Briganti, G., Williams, D.R., Mulder, J., and Linkowski P. (2020). Bayesian network
structure and predictability of autistic traits. accepted for publication Psychological Reports.
Briganti, G., Kornreich, C., and Linkowski, P. (2020). A network structure of manic
symptoms. submitted.
Briganti, G., Hubain, P., Kornreich, C., and Linkowski, P. (2020). Investigating the het-
erogeneity of psychiatric symptomatology using community detection algorithms. accepted
in Acta Psychiatrica Belgica.
Briganti G., and Scutari M. (2020). An Introduction to Bayesian Artificial Intelligence
in Medicine. submitted.
3
List of abbreviations
A: Appearance
AC: Academic Competence
AI: Artificial Intelligence
ANN: Artificial Neural Network
AQ: Alexithymia Questionnaire (in Chapter 9); Autistic Spectrum Quotient (in Chapter 10).
ASD: Autism Spectrum Disorders
BAI: Bayesian Artificial Intelligence
BF: Bayes Factor
BGGM: Bayesian Gaussian Graphical Model
BN: Bayesian Network
BVAQ: Bermond-Vorst Alexithymia Questionnaire
C: Competition
CFA: Confirmatory Factor Analysis
CFI: Comparative Fit Index
CPM: Clique Percolation Method
CS: Centrality-stability coefficient
CSWS: Contingencies of Self-Worth Scale
DAG: Directed Acyclic Graph
DSM-5: Diagnostic and Statistical Manual of mental disorders (5th edition)
EBIC: Extended Bayesian Information Criterion
4
EC: Empathic Concern
EGA: Exploratory Graph Analysis
EI: Expected Influence
EFA: Exploratory Factor Analysis
FS: Fantasy Scale (in Chapter 4); Family Support (in Chapter 5)
GGM: Gaussian Graphical Model
GL: God’s Love
GVAR: Graphical Vector Autoregressive Model
HRSD: Hamilton Rating Scale for Depression
IC: Inductive Causation algorithm
ICD-10: International Classification of Diseases (10th revision)
IM: Ising Model
IRI: Interpersonal Reactivity Index
LASSO: Least Absolute Shrinkage and Selection Operator
MDD: Major Depressive Disorder
MGM: Mixed Graphical Model
ML: Machine Learning
NCT: Network Comparison Test
NPI: Narcissistic Personality Inventory
OA: Other’s Approval
PC: Peter & Clark algorithm
PD: Personal Distress
PT: Perspective Taking
PTSD: Post-Traumatic Stress Disorder
R: R for statistical computing
RMSEA: Root Mean Square Error of Approximation
RSA: Resilience Scale for Adults
5
SRMSR: Standardized Root Mean Square Residual
SDS: Self-rating Depression Scale
TAS: Toronto Alexithymia Scale
TSIA: Toronto Structured Interview for Alexithymia
V: Virtue
YMRS: Young Mania Rating Scale
6
Contents
I An Introduction to Network Psychiatry 14
1 Network psychiatry: state of the art and challenges 15
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2 Structure and composition of a psychiatric network . . . . . . . . . . . . . . 20
1.2.1 Elements composing a network . . . . . . . . . . . . . . . . . . . . . 20
1.2.2 Types of networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3 An introduction to the estimation of a psychiatric network . . . . . . . . . . 24
1.3.1 Conditional dependence and independence . . . . . . . . . . . . . . . 25
1.3.2 Partial correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.3.3 Gaussian Graphical Model, Ising Model, Mixed Graphical Model . . . 26
1.3.4 Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.3.5 Bayesian estimation of the Gaussian Graphical Model . . . . . . . . . 27
1.4 Network inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.4.1 Centrality and redundancy . . . . . . . . . . . . . . . . . . . . . . . . 27
1.4.2 Symptom prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.4.3 Community detection in Networks . . . . . . . . . . . . . . . . . . . . 30
1.4.4 Network of domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.5 Discussion: state of the art and challenges of the network approach in clinical
practice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
7
2 A primer on Bayesian Artificial Intelligence and networks 34
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2 Probabilistic reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3 Bayesian thinking in machine learning . . . . . . . . . . . . . . . . . . . . . 37
2.4 Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4.1 The use of graphs to represent interactions among entities . . . . . . 39
2.4.2 Directed Acyclic Graphs . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5 Difference between neural networks and Bayesian networks . . . . . . . . . . 43
2.6 Structure learning of Bayesian networks . . . . . . . . . . . . . . . . . . . . . 44
2.6.1 The Markov property . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.6.2 Markov blankets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.6.3 Bayesian networks and Causality . . . . . . . . . . . . . . . . . . . . 46
2.6.4 Structure learning algorithms . . . . . . . . . . . . . . . . . . . . . . 47
2.7 Discussion: applications and limitations of Bayesian Artificial Intelligence in
Medicine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 A primer on undirected network models 51
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 Estimating network models from cross-sectional data . . . . . . . . . . . . . 52
3.2.1 Network components in undirected graphical models . . . . . . . . . 52
3.2.2 The Gaussian Graphical Model . . . . . . . . . . . . . . . . . . . . . 53
3.2.3 Bayesian estimation of the Gaussian Graphical Model . . . . . . . . . 53
3.2.4 The Ising Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2.5 Interpreting undirected network models in cross-sectional data sets . 57
3.3 Estimating network models from several time points . . . . . . . . . . . . . . 57
3.3.1 Granger causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3.2 The Graphical Vector Autoregressive Model . . . . . . . . . . . . . . 58
3.4 Network inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
8
3.4.1 Centrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.4.2 Node predictability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.5 Stability and accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
II A network approach to mental constructs 63
4 A network model of empathy 64
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2.1 Data set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2.2 Network analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3.1 Empathy network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3.2 Network inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3.3 Network accuracy and stability . . . . . . . . . . . . . . . . . . . . . 77
4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5 A network model of self-worth 81
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2.2 Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.2.3 Network analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.3.1 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.3.2 Network stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.3.3 Network inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
9
5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6 A network model of resilience 97
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2.2 Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.2.3 Network analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3.1 Item network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3.2 Six-domain network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3.3 Network stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.3.4 Network inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.3.5 Four-domain network . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7 A network model of narcissism 114
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.2.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.2.2 Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.3 Network analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.3.1 Network estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.3.2 Network inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.3.3 Network stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.4.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.4.2 Network of narcissism . . . . . . . . . . . . . . . . . . . . . . . . . . 124
10
7.4.3 Network inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.4.4 Network stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8 A network model of alexithymia without fantasizing 130
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.3 Data set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.4 Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.4.1 Network analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
8.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.5.1 Item network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.5.2 Three-domain network . . . . . . . . . . . . . . . . . . . . . . . . . . 140
8.5.3 Network stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
8.5.4 Network inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
8.5.5 Four-domain network . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
8.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
9 A network model of alexithymia with fantasizing 146
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
9.2.1 Data set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
9.2.2 Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
9.2.3 Network analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
9.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
9.3.1 Partial correlation network . . . . . . . . . . . . . . . . . . . . . . . . 153
9.3.2 Directed Acyclic Graph . . . . . . . . . . . . . . . . . . . . . . . . . . 157
9.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
11
10 A network model of autistic traits 161
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
10.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
10.2.1 Data set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
10.2.2 Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
10.2.3 Network analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
10.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
10.3.1 Partial correlation network . . . . . . . . . . . . . . . . . . . . . . . . 167
10.3.2 Node predictability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
10.3.3 Sex differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
10.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
III A network approach to mental disorders 175
11 A network model of depressive symptoms in a student sample 176
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
11.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
11.2.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
11.2.2 Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
11.2.3 Network analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
11.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
11.3.1 Regularized partial correlation network . . . . . . . . . . . . . . . . . 183
11.3.2 DAG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
11.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
12 A network model of mania 188
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
12.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
12
12.2.1 Ethical approval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
12.2.2 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
12.2.3 Network Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
12.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
12.3.1 Cross-sectional networks . . . . . . . . . . . . . . . . . . . . . . . . . 197
12.3.2 Network inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
12.3.3 Temporal network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
12.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
IV Overcoming challenges in network psychiatry 208
13 Investigating the heterogeneity of psychiatric symptomatology using com-
munity detection algorithms 209
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
13.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
13.2.1 Ethical approval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
13.2.2 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
13.2.3 Network Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
13.2.4 Community detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
13.2.5 Bridge centrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
13.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
13.3.1 Community detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
13.3.2 Bridge centrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
13.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
14 Discussion 235
14.1 Future directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
14.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
13
Chapter 1
Network psychiatry: state of the art
and challenges
Abstract
The network approach is enjoying increasing success in psychiatry and computa-
tional neuroscience. This new methodology, resulting from the translation of models of
statistical physics into medicine, grants a new ontological view on mental disorders, by
defining them as complex systems emerging from the interconnection between different
symptoms. This work offers an introduction, both theoretical and methodological, to
the potential of this approach to improve the diagnosis and treatment of psychiatric
disorders, by providing some examples of application explored in clinical research.
1.1 Introduction
The word “symptom” is defined as a “subjective phenomenon which translates morbid states
and which is linked to the functional or lesional disorders which determine it”. In current
psychiatric practice, the complaints reported by a patient to the clinician taking care of
him have been categorized as “symptoms”, a definition exemplified in diagnostic reference
manuals, such as the ICD-10 or the DSM-5. In somatic medicine, the presence of a symptom
15
implies the presence of an underlying disease, due to a uni- or multi-factorial etiology most
often identified.
In psychiatry, however, the concept of “mental illness”, although still strongly repre-
sented in the forensic field (I am thinking in particular of the Belgian law of June 26, 1990
relating to the psychiatric hold of the mentally ill), gradually gives way to the concept of
“mental disorder”, as described by the DSM-V (American Psychiatric Association, 2013) and
referring to a “syndrome”. The syndrome is in turn defined as a “set of several symptoms
or signs related to a given pathological condition and allowing, by their grouping, to guide
the diagnosis“: it is therefore an ecosystem of entities detected by the clinician or declared
to the latter; this ecosystem carries an empirical meaning, but for an unknown reason. For
most psychiatric pathologies, however, the classic etiopathogenic approach used in medi-
cal research has proven unsuccessful (Borsboom and Cramer, 2013): in psychiatry, one can
rarely identify a cause which, completely eliminated, makes the entire the symptomatological
presentation of the patient disappear, such as the effect that administration of an antibiotic
would have on an infection caused by a particular germ.
However, classic psychometrics (the field common to psychiatry and psychology which
aims to measure mental entities, both normal and pathological) has nowadays a very large
majority of measurement tools that conceive of mental disorders as being the consequence
of a common cause. This approach, which has been widely used in recent decades and well
consolidated in the scientific literature, is known as the “latent variable model” (Marsman
et al., 2018). According to the latent variables model, the symptoms encountered in clinical
practice are elements which can be measured, but which constitute only a passive reflection
of the pathology in question; like a cough during pneumonia, the psychiatric symptom does
not represent a component which causes the pathology, but is the measured effect of the
pathology itself.
A decade ago, some work put forward the hypothesis that a central cause for mental
disorders cannot be found because there is no central cause; instead of being caused by an
16
Figure 1.1: Latent variable model. A latent variable (C) causes 9 symptoms.
invisible entity, psychiatric symptoms cause each other (Borsboom, 2008). For example, if
a patient has anxious ruminations, he may have trouble sleeping and therefore he will be
more tired; fatigue will generate some stress, which in turn will increase ruminations. With
such feedback activation, the patient’s mental state can degenerate, until it can be defined
as a mental disorder. This hypothesis has been formally defined as the “network” theory
of mental disorders (Borsboom, 2017) and has been applied in different fields of psychiatry
and clinical psychology, such as post-traumatic stress disorder (Fried et al., 2018), empathy
(Briganti et al., 2018), narcissistic personality (Briganti and Linkowski, 2019a), depression
(Mullarkey et al., 2018), alexithymia (Briganti and Linkowski, 2019c), self-esteem (Briganti
et al., 2019), autism (Deserno et al., 2017), and resilience (Briganti and Linkowski, 2019b).
From a mathematical point of view, the common cause model (illustrated in figure 1.1) is
equivalent to the network model, one example of which is indicated in figure 1.2.
Recent work has been able to propose an integrative model combining the approach of
17
Figure 1.2: The network model. Variables (nodes) interact through connections (edges).Positive edges are colored in blue, negative edges are colored in red. The correspondingthickness of an edge denotes its intensity (weight).
18
latent variables with the network approach: in this approach, called generalized network
psychometrics (Epskamp et al., 2017b) the symptoms can be caused by different latent
variables and interact with each other. This approach has proven useful in simplifying the
analysis of networks based on psychometric scales, where different dimensions are represented
by several redundant items. Once the redundancy is limited by representing the symptoms by
the latent variables that it measures, for example a particular domain of a mental construct,
such as the contribution of self-opinion in the construct of self-esteem (Briganti et al., 2019),
an interaction between the different latent variables is observed. Some work has been able
to demonstrate that new tools, such as exploratory graphical analysis (Golino and Epskamp,
2017), specifically designed for the detection of domains in networks (translate “factors” in
the approach by latent variables) is capable of detecting the correct number of dimensions
in a sample; if the number detected is not the same, a new model is proposed which is better
suited to the sample data.
The aim of this work is to offer a general introduction to network psychiatry, paying
particular attention to the implications that the development of this approach can have on
the diagnosis and treatment of mental disorders. I detail the challenges, opportunities and
main criticisms that have been described in the literature in this area, which has evolved
rapidly in the last decade. I will first describe the structure and composition of a psychiatric
network. Next, I will detail the different processes for estimating a psychiatric network.
Third, I will detail the different measures used to interpret the results of a network, as well
as their stability and accuracy. Finally, I will discuss the main applications that this approach
could develop in current clinical practice. The main criticisms of the network approach will
be integrated as the different methodologies are introduced.
19
Figure 1.3: An undirected network with three nodes (A, B, and C). The three nodes areconnected by two edges, A-B, and A-C. A-B is thicker than A-C; it has a greater weight.
1.2 Structure and composition of a psychiatric network
1.2.1 Elements composing a network
A network is composed of a set of nodes, connected through a set of edges (Boccaletti et al.,
2006).
A node represents a measured entity. In other areas, nodes represent people (social
network), stations, or cities. Networks can also explore brain regions via neuroimaging
(neuroanatomical networks). In psychiatry, nodes represent either observed symptoms/signs
(in the context of mental disorders) or other parts of mental constructs (such as empathy,
narcissism, or resilience). In graphic visualization, nodes are usually represented by circles,
squares or even triangles.
An edge represents a connection (or absence of connection, depending on the network
model that is chosen) between two nodes. An edge is usually interpreted as the presence (or
absence) of interaction, or morbidity, or causality between two entities. The interpretation
of the edge depends on the network model that is estimated. For example, if I choose to
explore the rail network of a region, an edge will represent the railway that connects two
stations. Likewise, a social network will use the edges to indicate a mutual relationship
between two people. The edges can be observed (the observer knows a priori that two nodes
20
are connected) or else unobserved (the presence or absence of a relationship between two
nodes must therefore be tested, by defining a null hypothesis and an alternative hypothesis).
1.2.2 Types of networks
There are different types of networks. Below I will detail the types of networks most used
in psychiatry.
Undirected network, weighted or unweighted
An undirected network is a structure with nodes connected by edges whose direction is
unknown. An example of an undirected network is illustrated in figure 1.3. The edge A-B
connecting nodes A and B, could, in an undirected network, assume 3 possible directions:
A to B, B to A, or symmetrically from A to B and vice-versa. The edges of an undirected
network can have a weight, reflecting the relative importance of one edge compared to others.
The weight of the edge is most often represented by a thicker link. Networks containing
weights are called weighted networks; weightless networks are unweighted.
To understand the importance of weighting, let’s take the example of three known symp-
toms of depression (Zung, 1965) that could underlie the figure 1.3: insomnia (A), fatigue (B),
and suicidal ideation (C). Insomnia could be linked to suicidal ideation as well as fatigue, but
its connection to fatigue is much more clinically important than to suicidal ideation. The
edge weights are used precisely to express this difference between the connections within a
network.
In weighted networks, edge weights can be positive (denoting a positive association be-
tween two nodes) or negative (representing a negative association). Figure 1.4 shows a
network with a positive association (A-B) and a negative association (A-C).
Weighted undirected networks are the most used networks in clinical and methodological
research (Epskamp and Fried, 2018). Edges most often represent partial correlations.
21
Figure 1.4: An undirected weighted network with three nodes (A, B, and C). The threenodes are connected by two edges, A-B and A-C. A-B denotes a positive connection. A-Cdenotes a negative connection. A-B has a greater weight than A-C.
Directed network, cyclic or acyclic
Some works use advanced statistical methods to determine a causal relationship between
two nodes (Moffa et al., 2017), even in cross-sectional data, using machine learning methods
(Scutari, 2010). These methodologies use so-called “directed” networks since the direction
of the edges is determined (Briganti et al., 2020b). There are two types of directed net-
works: cyclic and acyclic. Although their use is less popular than undirected networks, it is
important to know the basic structures of directed networks since a directed structure could
be the “ground truth” which underlie certain symptoms of mental disorders. If underlying
during the analysis of an undirected network, directed networks can generate unexpected
results that can be difficult to interpret.
Two examples of acyclic directed networks are proposed in the figures 1.5 and 1.6.
In figure 1.5, two nodes B and C cause the node A. This structure is also known under the
name of “V-structure”, or “collider” and constitutes the basis for the automatic learning of
the algorithms which will discover the structure underlying the analyzed data (Scutari, 2010).
The collider structure is useful since it stores temporal information at a given instant. For
example, if nodes B and C represent two police officers pointing a weapon against a criminal
(node A), I can, by determining the status of A (0 = dead, 1 = alive), and one of the two
22
Figure 1.5: An acyclic directed network composed of three nodes A, B, and C. B and Ccause A. This is commonly defined as a collider situation.
Figure 1.6: An acyclic directed network composed of three nodes A, B, and C. A causes B,B causes C.
23
Figure 1.7: A cyclic directed network composed of three nodes A, B, and C. A causes B, Bcauses C, and C causes A.
police officers (0 = did not fire, 1 = fired), infer the status of the other police officer.
In figure 1.6, A cause B which in turn causes C. C does not cause A, which leaves the
loop open.
The loop is however closed in the figure 1.7, representing a cyclic directed network.
Cyclic networks are most often encountered in time series, where repeated measurements of
psychiatric parameters are made to control the mental state. Cyclical directed networks could
be useful in order to show the existence of a “critical slowing down” of symptoms: during the
deterioration of a mental state, the prediction between the symptoms in a network becomes
more concrete (Wichers and Groot, 2016).
1.3 An introduction to the estimation of a psychiatric
network
This constitutes a brief and clinical introduction to the estimation of network models used in
clinical research. A statistical introduction of the methods used in this thesis can be found
in Chapters 2 and 3.
24
1.3.1 Conditional dependence and independence
The presence of an edge in a network can mean the presence or absence of a connection
between two nodes. This depends on whether the estimation of a network aims to detect the
presence of a conditional dependence or a conditional independence (Williams et al., 2019).
A conditional dependence is understood as an association between two variables when a
third variable is set. Conditional independence is understood as the absence of association
between two variables when a third variable is fixed at given levels (Epskamp et al., 2017b).
1.3.2 Partial correlations
Most often, the edges between the nodes are estimated in the form of partial correlations,
one of the statistical translation of the theoretical concept of conditional dependence relation
(Epskamp and Fried, 2018). A network of partial correlations will look like the models
proposed in the figures 1.2 and 1.4: a weighted network not directed. The partial correlations
will be positive or negative, reflecting a corresponding association between the nodes.
How to interpret the presence of a positive partial correlation between two variables X
and Y found in a network of psychiatric symptoms such as collected with self-administered
scales? From a statistical point of view, I can affirm that, if a partial correlation is present
between X and Y, that implies that, by controlling all the other variables of the network, a
connection exists between the two nodes (Briganti et al., 2018).
To translate this mathematical explanation into clinical practice, a possible interpretation
is as follows: by controlling all the other symptoms, if X and Y share a connection, this means
that in the observed sample the average response of the observed group to question X will
be able to predict that of Y and vice-versa, since the network is undirected (Briganti et al.,
2019). In practice, if I observe a partial correlation between two symptoms insomnia and
fatigue, I can deduce that if the observed group has on average significant insomnia, it will
also present significant fatigue, controlling the levels of other network symptoms.
25
1.3.3 Gaussian Graphical Model, Ising Model, Mixed Graphical
Model
All the partial correlations of a network are determined at the same time by the estimation
of a Gaussian Graphic Model (GGM), applicable for continuous data (or comparable to con-
tinuous) (Williams, 2018a). The GGM (the network itself) is estimated as the inverse of the
covariance matrix derived from the collected symptom data set. The alternative for binary
variable of the GGM is Ising’s model, derived from statistical physics and adapted in psychi-
atry (van Borkulo et al., 2014), but rarely used in empirical work (Briganti and Linkowski,
2019a). Depending on the type of variables included (for example, if a binary variable such
as the gender of the participants is added to a continuous database), Mixed Graphic Models
(MGM) can be used (Haslbeck and Waldorp, 2016). The GGM is nevertheless the model
used predominantly in the literature.
GGM is easy to understand, especially if it is used to study a mental disorder. Let
y = (y1, ..., yp)> be any normal multivariate vector. Without loss of information, the data are
considered to be centered with an average vector of 0 and a covariance matrix Σ. The inverse
of the covariance matrix Θ = Σ−1 is targeted in network analysis, since by standardizing
the off-diagonal elements of the matrix I find the partial correlation coefficients (ρ). The
GGM associated with y is an unsupervised netwok which is usually denoted G = (V,E). It
includes a set of nodes V = {y1, ..., yp} which corresponds to, say, a group of symptoms in
the questionnaires, as well as a set of edges E which includes the “connections”. Therefore,
there is an edge between two nodes yi and yj when they are determined as conditionally
dependent (for example, ρij 6= 0).
1.3.4 Regularization
In the vast majority of mental disorder networks studied to date, the GGM undergoes a
regularization process - `1 associated with an EBIC (Bayesian Extended Information Crite-
26
rion) called “lasso” (acronym for “least absolute shrinkage and selection operator “), which
calibrates the network and reduces small partial correlations to zero; it therefore makes
a “conservative” model (with as few connections as possible) (Epskamp and Fried, 2018),
considered easier to visualize and interpret.
1.3.5 Bayesian estimation of the Gaussian Graphical Model
It is important to note that the presence of a zero in the network matrix corresponds to
a partial correlation of zero, and therefore does not prove the existence of a conditional
independence relationship between the variables (symptoms) sharing a partial correlation of
zero; this implies that the absence of an edge between two nodes of a network estimated via
a conventional GGM cannot confirm the absence of connection.
A Bayesian GGM estimation method was recently introduced (Williams, 2018a). This
methodology uses the Bayes factor (directly reflecting a level of evidence) as a threshold
for detecting conditional / unconditional dependency relationships in a network, allows the
determination of the GGM on the basis of the posterior probabilities (therefore taking the
data into account), and offers the possibility for the researcher to formally test the condi-
tional relationships of dependence or independence within a network. The added value of
this approach is to provide information as to the level of evidence available in favor of the
hypothesis tested or the exploration of the network without a priori hypothesis.
This method, despite its recent introduction, has been able to show benefits in a network
(Williams et al., 2020) reanalysis, and studies are underway to test it empirically.
1.4 Network inference
1.4.1 Centrality and redundancy
Centrality is an measure imported from statistical physics, like the rest of the network
techniques : it is needed to represent the relative importance of a symptom compared to
27
Figure 1.8: A graph representing the centrality scores for each node (standardized z-scores).
the other symptoms of the network. The most used measure to report the centrality of a
symptom is strength. Strength is defined as the sum of the absolute values of the edges
that a given symptom shares with other symptoms in a network (Boccaletti et al., 2006).
Because some networks have negative edges, there is another measure of centrality that takes
into account the sum of the relative values of the edges; this measure is known as expected
influence (Briganti et al., 2019).
One issue is noticed with the use of centrality when applying networks to mental disorders
as measured in scales; in fact, most scales contain repeated measurements of the same type
of symptom, which implies the presence of redundant nodes within the same network. When
this happens, the redundant nodes are strongly connected, which considerably increases
their centrality, but at the cost of changing the nature of their connection: the latter must
therefore be interpreted as a variance shared between the two symptoms, which distorts
the interpretation of the centrality indices of the network (Briganti and Linkowski, 2019b).
28
Figure 1.9: A network presenting nodes with their respective predictability.
Another weak point of this technique is the relativity of the estimated index: within a
network, there will always be one symptom that is more central than another, even if the
most central symptom actually presents weak estimates.
An example of centrality graph is found in figure 1.8.
1.4.2 Symptom prediction
Node predictability was introduced as an alternative measure of centrality which is an abso-
lute measure of the connectivity of a symptom in a network. Node predictability represents
the shared variance of a given node with the surrounding nodes (Haslbeck and Fried, 2017).
Most often, node predictability is represented by a circular diagram filled up to the per-
centage, as illustrated in figure 1.9. Node predictability does not escape the problem of
redundant symptoms, since if two symptoms are similar, they will be highly predictable in
29
a network. Bayesian methods are now also applicable to the estimation of this parameter
(Williams and Mulder, 2019).
Node predictability has been defined as the upper limit of controllability: if a node A is
highly predictable and I consider that all the edges it shares with other nodes are directed
towards A, then I can consider that the predictability is the reflection of the control that the
other variables have on the node A (Briganti et al., 2019).
1.4.3 Community detection in Networks
Being able to structure symptoms or mental constructs into domains (or dimensions) that
can explain their clinical cohesion from an empirical point of view is a long-term research
challenge in psychiatry (Borsboom, 2017). Network psychiatry has its own domain detection
tools, such as the walktrap algorithm derived from machine learning and used to explore
dimensions (Briganti et al., 2018); this algorithm is based on the distance computed between
symptoms to establish groupings: the symptoms which are “close” indeed tend to form
a community in the network. It has been shown that this algorithm has good stability
for finding the right number of communities based on large data; he therefore inspired
the creation of the exploratory graphical analysis (Golino and Epskamp, 2017), used as
a systematic approach in network psychiatry to verify whether the expected number of
domains of a construct or mental disorder is found in a given sample. This methodology
made it possible to show that, in certain cases, one or more domains can be redundant in
a mental disorder, and that merging them makes it possible to obtain more relevant results
for the data (Briganti and Linkowski, 2019b,c).
1.4.4 Network of domains
I was able to show how the network analysis of scales representing a mental construct often
relates communities corresponding to the factors originally proposed (Briganti et al., 2018,
2019). Although the interactions between symptoms / items of the same community is
30
Figure 1.10: A network of domains regrouping nodes from figure 1.9
interesting, network analysis reveals its real plus value when I look at the interactions between
symptoms of different communities, in order to show the behavior of several facets of mental
constructs or disorders.
It is in this case that generalized network psychometrics has been developed (Epskamp
et al., 2017b); this approach makes it possible to combine the factorial approach (grouping
together as many variables as possible under the same label), then to analyze the connec-
tions between the remaining variables (i.e., corresponding truly to a network of fully-fledged
components). I have been able to show that this approach is rewarding when tackling mental
constructs where several nodes refer to the same aspect of behavior to highlight positive or
negative relationships between domains (Briganti and Linkowski, 2019b).
31
1.5 Discussion: state of the art and challenges of the
network approach in clinical practice
The aim of this chapter was to focus on the important subject of network psychiatry, enabled
by the development of the network theory of mental disorders (Borsboom, 2017). I have
introduced current issues into the current study and classification of constructs and mental
disorders, for which network analysis is posed as an interesting and new alternative. Then, I
introduced the state of the art on network theory and analysis as used for the vast majority
of studies in the literature as well as the main flaws and criticisms exposed for the methods.
When a large part of mental constructs and disorders have been redefined thanks to the
analysis of networks (Fried et al., 2018; Briganti and Linkowski, 2019a,b; Briganti et al.,
2020b), the next stages in the development of this new science take shape: I will discuss
some of those.
First, analyzing network inference has proved useful in predicting post-treatment out-
comes in the case of anorexia (Elliott et al., 2019): this reflects the potential usefulness of
measuring central symptoms in a network and its use it to formulate and test formal clinical
hypotheses.
Secondly, network analysis takes on its full meaning when it serves to redefine the psy-
chiatric theory underlying mental disorders: this is how, for example, new computational
models of “panic disorder” have been developed recently (Robinaugh et al., 2019); like the
latter, a large part of mental disorders could be redefined thanks to the analysis of networks,
which will allow in the future the birth of new classifications corresponding more to the
advances made in both fundamental and clinical research.
Third, we are witnessing the emergence of “hybrid” networks mixing symptoms and brain
regions in recent work (Hilland et al., 2020); in the future, we will be able to extend the
network approach to the different levels determining normal and abnormal human behavior,
in order to have an extended view on the dynamics controlling mental constructs and disor-
32
ders both on a symptomatic and electrophysiological, neuroanatomical, biological level. and
genetics.
Fourth, because the networks approach reconsiders the symptom as an active component
in mental disorder, an interventional approach is recommended in order to monitor how
therapeutic interventions affect given symptoms and their relationships in the network: this
approach, tested recently, has was defined as “interventional analysis of networks” (Blanken
et al., 2019). Network analysis is therefore also proposed as an objective monitoring tool for
therapies in psychiatry, a very interesting perspective from a clinical point of view.
Although relatively “young” (the field of research was born about ten years ago), network
psychiatry was quickly transformed from a basic research tool to a technique applied to
clinical practice. The perspectives of future studies are imbued with translational interest
necessary for the improvement of practices in psychiatry.
33
Chapter 2
A primer on Bayesian Artificial
Intelligence and networks
Abstract
Artificial Intelligence has become a topic of interest in the medical sciences in the
last two decades, but most studies of its applications in clinical studies have been
criticised for having unreliable designs and poor replicability: there is an increasing
need for literacy among medical professionals on this increasingly popular domain of
research. This work provides a short introduction to the specific topic of Bayesian
Artificial Intelligence: we introduce Bayesian reasoning, networks and causal discovery
as well as their (potential) applications in clinical practice.
2.1 Introduction
The origins of Artificial Intelligence (AI) can be traced back to the work of Alan Turing
in 1950 (Turing, 1950). He proposed a test (known today as the Turing Test) in which a
certain level of intelligence is required from a machine to fool a human into thinking he is
carrying on a conversation with another human (the Imitation Game). Although such level
of intelligence has not yet been attained, a simple conversation with a virtual assistant like
34
Siri serves as a clear example of how AI research has rapidly evolved in the past decades.
AI has become a topic of interest in the medical sciences in the last two decades, with
more and more applications being approved by health authorities worldwide. However, most
physicians and medical scientists have no formal training in the disciplines the world of smart
medical monitoring, diagnosis and follow-up has sprung from. This, along with historical
epistemological differences in the practice of medicine and engineering disciplines, is one
of the main obstacles to widespread collaboration between physicians and engineers in the
development of AI software. It may also be argued that that is in turn one of the root causes
of the ongoing replicability crisis in medical AI research (Briganti and Le Moine, 2020),
characterised by unreliable study designs and poor replication (Liu et al., 2019).
Although there is an increasingly rich literature on how AI can be used in the clinical
practice, few works aim to interest physicians in the fundamental concepts and terminology
of AI. The gradual shift towards quantitative methods in the last century has made them
familiar with such concepts from probability and statistics as correlation, regression and
confidence intervals; it is worthwhile to expand on these concepts and link them with modern
AI.
The most common AI approaches in the medical literature are neural networks (in the
early 2000s) and clustering (in the last decade) (Mintz and Brodie, 2019). Bayesian reasoning
and methods for AI are less known, although they can be used for medical decision making,
to study human physiology (Lucas et al., 2004), to identify interactions between symptoms
(Briganti et al., 2020b), and to investigate symptom recovery (Liew et al., 2019). Bayesian
models can facilitate reasoning in probabilistic terms when dealing with uncertainty, which is
omnipresent in the medical sciences since we cannot construct a complete mechanistic model
of diseases and of the physiological mechanisms they impact. One important characteristic
of Bayesian models, and in particular of Bayesian networks, is that the variables that are
used as inputs for the model and their relationships are direct representations of real-world
entities and of their interplay, not a purely mathematical construct as in neural networks
35
(Pearl and Russell, 2011). This is why such approach is interesting to the medical field:
Bayesian networks are the method of choice for explainable reasoning in AI.
This work aims to introduce physicians to Bayesian AI through a clinical perspective on
probabilistic reasoning and Bayesian networks. More detailed accounts on the topic may be
found in popular textbooks (Scutari and Denis, 2015; Korb and Nicholson, 2010).
2.2 Probabilistic reasoning
Bayes’ theorem describes the probability of an event based on prior knowledge of conditions
that may be related to that event: that is expressed with the well known mathematical
notation
Pr(A |B) =Pr(B |A) Pr(A)
Pr(B),
the probability Pr(A |B) of an event A given prior knowledge of a condition B (that is, B
has occurred). Pr(A |B) is a conditional probability, while Pr(A) and Pr(B) are known as
marginal probabilities, that is, the probabilities of observing the events A and B individually.
In the context of medical diagnosis, the goal is to determine the probability Pr(Di |Cp) of
presence of a particular disease or disorder Di given the clinical presentation of the patient
Cp (Miettinen and Caro, 1994), the prior probabilities P ′i of the disease in the patient’s
reference group, and the prior probabilities P ′j of other diseases Dj; that is
Pr (Di |Cp) =P ′i Pr (Cp |Di)
P ′i Pr (Cp |Di) +∑
j P′j Pr (Cp |Dj)
where Pr(Cp |Dj) is the probability of having the same clinical presentation given other
diseases.
Bayes’ theorem makes it possible to work with the distributions of dependent (or condi-
tionally dependent) variables. However, in order to reduce the number of variables we need
to observe simultaneously in probabilistic systems, it is also important to determine whether
36
two variables A and B are independent (A ⊥⊥ B),
Pr(A |B) = Pr(A),
or conditionally independent (A ⊥⊥ B |C) given the value of a third variable C,
Pr(A ∩B |C) = Pr(A |C) Pr(B |C).
Extracting conditional dependence tables is one of the building blocks upon which we can
build Bayesian probabilistic reasoning. It allows to take a set of variables (say, A, B and C
again) and to compute the conditional probabilities of some of them (say, A |C),
Pr(A = y |C = x) =
∑B∈{x,y} Pr(C = x,B,A = y)∑B,A∈{x,y} Pr(C = x,B,A)
.
Conversely, we can also take variables or set of variables that are (conditionally) independent
from each other and combine them to obtain their joint probability. This joint probability
will be structured as a larger conditional dependence table that is the product of smaller
tables associated with the original variables. For instance,
Pr(A = y,B = z |C = x) = Pr(A = y |B = z) Pr(B = z |C = x)
assuming A ⊥⊥ C. The ability of explicitly merging and splitting set of variables to separate
variables of interest we need for diagnostic purposes from redundant variables is one of the
reasons that makes Bayesian reasoning easy while at the same time mathematically rigorous.
2.3 Bayesian thinking in machine learning
Machine learning (ML) is the sub-field of AI that studies the algorithms and the statistical
tools that allow computer systems to perform specific and well-defined tasks without explicit
37
instructions (Murphy, 2012). Implementing machine learning requires four components.
First, we need a working model of the world that describes the tasks and their context
in a way that is understandable by a computer. In practice, this means choosing a class of
Bayesian models defined over the variables of interest, and over relevant exogenous variables,
and implementing it in software. Generative models such as Bayesian networks describe how
variables interact with each other, and therefore represent the joint probability
Pr (X1, . . . , XN) ;
while discriminative models such as random forests and neural networks only focus on how
a group of variables predicts a target variable by estimating
Pr (X1 |X2, . . . , XN) .
Clearly, depending on the application, a generative model may be a more appropriate choice
than a discriminative model or vice versa. If characterising the phenomenon we are modelling
from a systems perspective, generative models should be preferred. If we are only interested
in predicting some clinical event, as in the case of diagnostic devices, discriminative models
provide better predictive accuracy at the cost of being less expressive.
Second, we need to measure the performance of the model, usually by being able to
predict new events. Third, we must encode the knowledge of the world from training data,
experts or both into the model: this step is called learning. The computer system can
then learn what is the best model within the prescribed class by finding that that maximize
the chosen performance metric, drawing either from observational or experimental data or
expert knowledge available from practitioners. Fourth, the computer uses the model as a
proxy of reality and performs inference as new inputs come in, all while deciding if and how
to perform the assigned task.
Successfully implementing machine learning applications is, however, far from easy: it
38
requires large amounts of data, and it is difficult to decide how to structure the model from
a probabilistic and mathematical point of view. Principled software engineering also plays
an important role (Beam et al., 2020). For these reasons, machine learning models should
comprise a limited number of variables so that they are easy to construct and interpret. As a
side effect, compact models are unlikely to require copious amounts of computational power
to learn, unlike state-of-the-art neural networks (Beam et al., 2020).
Clinical settings limit our knowledge of the patients to what we can learn from them:
hence probability should be used to determine whether two variables are associated given
other variables. Formally, if the occurrence of an event in one variable affects the prob-
ability of an event occurring in another variable, we say the two are associated with or
probabilistically dependent on each other. Association is symmetric in probability; it does
not distinguish between cause and effects in itself. Bayesian network models, however, go
beyond probability theory to represent causal effects as arcs in a graph to allow for principled
causal reasoning.
2.4 Bayesian Networks
We introduce in this section the fundamental notions of Bayesian networks. Specialised text-
books have a more thorough review on the subject (Scutari and Denis, 2015), and software to
implement them is available from the bnlearn package (Scutari, 2009) for the R statistical
environment.
2.4.1 The use of graphs to represent interactions among entities
Graph theory is the field of mathematics that deals with the study of graphs: such structures
are meant to represent relationships between entities, like symptoms, signs or biological
markers in the medical sciences.
A graph G is understood as a set V of nodes (also known as vertices) representing variables
39
Figure 2.1: An undirected network of five nodes connected through edges.
(or other feature of the data) that are connected through a set A of edges (also known as
arcs). Let us consider a network of five nodes, such as that shown in Figure 2.1. In this
case, the set of nodes comprises
V = {v1, v2, v3, v4, v5},
and the set of edges is
A = {a12, a14, a23, a24, a25, a34},
where a12 represents the edges between node v1 and node v2. The network represented in
Figure 2.1 is called an undirected network, because edges are not directed in a particular
direction, and therefore (vi, vj) = (vj, vi). Undirected networks are commonly used to repre-
sent the pairwise interactions among psychopathological symptoms (Briganti et al., 2020b).
The edges can be unweighted, so that an edge can either be present aij = 1 or absent aij = 0
between two nodes; or they can be weighted, so that some edges can be stronger than others
in a network, and can have either a positive or negative sign. For instance, an edge weight
can represent a partial correlation estimate (Briganti et al., 2018) to convey the existence of
a conditional association relationship between two variables.
40
Figure 2.2: A Bayesian network, or Directed Acyclic Graph composed of five nodes. Edgesare directed from one node to another node.
2.4.2 Directed Acyclic Graphs
Bayesian networks, on the other hand, are based on directed acyclic graphs (DAGs). A DAG
contains only directed edges, hence (vi, vj) 6= (vj, vi) because the former is vi→ vj and the
latter is vj→ vi. An example is shown in Figure 2.2. Such arcs are often interpreted as
causal relationships in which the tail node is the cause and the head of the arrow is the
effect. Bayesian networks cannot contain loops (the effect of a node on itself) or cycles (for
instance, A goes to B, B goes to C, and C goes to A). The primary goal of a Bayesian
network is to express the conditional independence set of relationships among variables (that
is, variables that do not predict each other).
In addition to a DAG, Bayesian networks are defined by the global probability distribution
of X (with Xi being the variable that corresponds to the node vi in the network) with
parameters Θ,
Pr(X,Θ) =N∏i=1
Pr (Xi |ΠXi ; ΘXi)
where ΠXi represent the parent nodes of Xi. This factorisation derives from the Markov
property of Bayesian networks, that is, every variable Xi depends on its parents ΠXi (Korb
and Nicholson, 2010).
The three most common probability distributions for Bayesian networks are discrete,
41
DepMood
Sleep
Weight
FatigueIrritable
SuicideId
Anhedonia
DepMood: Depressive MoodSleep: Sleep ProblemsWeight: WeightFatigue: FatigueIrritable: IrritableSuicideId: Suicidal IdeationAnhedonia: Anhedonia
Figure 2.3: A Bayesian network, or Directed Acyclic Graph composed of seven depressionsymptoms from the Zung Depression Scale. Edges are directed from one node to anothernode.
Gaussian, conditional linear Gaussian (Scutari and Denis, 2015). Discrete Bayesian networks
have, for instance, been used in expert systems to differentiate between tuberculosis and
lung cancer (Lauritzen and Spiegelhalter, 1988). Gaussian Bayesian networks are common
is genetics and systems biology for reconstructing direct and indirect gene effects (Kruijer
et al., 2020), and together with conditional Gaussian Bayesian networks they have been used
to study various clinical treatments and conditions (Liew et al., 2019; Dao et al., 2016).
An example of Bayesian networks: psychiatric symptoms
Bayesian networks have been used to investigate causal relationships among psychiatric
symptoms, such as depression symptoms (Briganti et al., 2020b): an example of Bayesian
networks composed of seven depression symptoms from the Zung Depression Scale (Zung,
1965) is represented in figure 2.3.
42
Such Bayesian networks are computed from a data set of symptom scores: edges there-
fore represent admissible causal relationships among symptoms (Moffa et al., 2017). In the
example shown in figure 2.3, for instance, depressive mood has a causal relationship directed
towards weight loss, which in turn has a causal relationship directed towards irritability.
2.5 Difference between neural networks and Bayesian
networks
Neural networks, often called Artificial Neural Networks (ANNs), are in widespread use
in the AI field. They represent networks of interconnected artificial neurons that change
their state with the external or internal information that flows through the network during
a learning phase (Hecht-Nielsen, 1990). Neural networks are usually organised in several
layers: an input layer, several intermediate layers of latent variables, and an output layer.
Their aim is to identify a relationship between the input and the output. Hence they are
discriminative models, and do not provide any insight into the interplay of the variables nor
a semantic representation of causes and effects. They are known to be difficult to interpret
(Correa et al., 2009), to the point that post hoc methods to improve their interpretability
and explainability are now a challenging new avenue for research(Holzinger et al., 2019). The
key advantage of Bayesian networks is that they model of the real world: the phenomenon
under investigation is understood by the machine, dissected, and clearly represented as a
set of causal relationships. This allows for a predictive reasoning much needed in medicine:
Bayesian networks can answer diagnostic and prognostic questions of the form “how will
symptom/sign/disease A change if we act upon symptom/sign/disease B?” in a way that is
understandable by both patients and clinicians. This is possible because of the reversibility
of Bayes’ theorem (Pr(A |B) Pr(B) = Pr(A,B) = Pr(B |A) Pr(A)).
43
Figure 2.4: A Bayesian network of six nodes to illustrate graphical separation (meaning twonodes are not connected in the network). For instance, 1 is separated from 4 and 5 through3; 2 is separated from 4 and 5 through 3, and 3 is separated from 6 through 5.
2.6 Structure learning of Bayesian networks
In this section we will introduce the concepts of graphical separation and probabilistic inde-
pendence.
2.6.1 The Markov property
In Bayesian networks, if two nodes are unconnected (that is, they do not share an edge),
that means that they are also conditionally independent: this is called the Markov property
(Korb and Nicholson, 2010). Graphical separation implies probabilistic independence,
A ⊥⊥G B |C =⇒ A ⊥⊥P B |C.
making the network itself is a clear representation of the conditional independence rela-
tionships between nodes. For this reason, the DAG is called an independence map of the
variables. The Markov property makes it possible to write
Pr(X,Θ) =N∏i=1
Pr (Xi |ΠXi ; ΘXi) ,
44
decomposing the larger model Pr(X,Θ) into a set of smaller models Pr (Xi |ΠXi ; ΘXi) that
are easier to understand. This decomposition is only possible because of the absence of loops
and cycles in the graph.
Figure 2.4 represents a Bayesian network with six nodes. Two nodes, say v1 and v4, are
graphically separated by node v3, and are therefore conditionally independent given node v3:
v1 ⊥⊥G v4 | v3 =⇒ Pr(v1, v4 | v3) = Pr(v1 | v3) Pr(v4 | v3).
Figure 2.4 also shows a specific kind of relationship in a Bayesian network, that is the one
among nodes v1, v2 and v3. Both v1 and v2 have an edge pointing to v3 and the two share
no connection. This kind of motif is commonly known as a v-structure, or a collider, and
it is often considered as one of the building blocks of Bayesian networks. In a collider, the
two causes are known to be negatively correlated, which is counter-intuitive; conditioning
on the common effect in the collider (that is, studying the associations while manipulating
the effect) leads to different estimates compared to studying the two causes on their own.
This phenomenon is known as collider bias or Berkson’s bias and it is an important source
of bias in the medical sciences (Berkson, 1946).
2.6.2 Markov blankets
A DAG is understood as an independence map of the probability distribution of the variables
Xi; retrieving such a map means testing which nodes are conditionally (in)dependent. d-
separation is a useful instrument to algorithmically determine whether two nodes in a network
are (in)dependent or conditionally (in)dependent (Geiger et al., 1990): two nodes A and B
are d-separated by a conditioning set of nodes S if conditioning on all members of S blocks
all paths (sequence of nodes and edges with A as starting node and B as ending node)
between A and B. A collider is known to block all paths it overlaps. The set S is known as
the Markov blanket of node A in the graph G. By definition, the Markov blanket includes
45
a node’s parents (nodes with an edge directed towards A), children (nodes that receive a
directed edge from A), and spouses, that is, children’s other parents. The Markov blanket
is useful in investigating a target node of interest while forgetting about the rest of the
Bayesian network; all nodes outside of the Markov blanket are independent from the node
of interest.
2.6.3 Bayesian networks and Causality
Since Bayesian networks are based on DAGs, the relationships among variables are easily
interpreted as causal relationship. However, three assumptions should be made before inter-
preting an edge as a causal effect. First, each variable (node) must conditionally independent
of its indirect and direct non-effects given its direct causes (this is the causal translation of
the Markov property). Second, there must exist a DAG faithful to the probability distribu-
tion of X so that the only dependencies in the probability distribution are those that arise
from d-separations in the DAG. The third assumption descends from the first two: there
must be no latent variables that act as confounding factors (therefore developing causal ef-
fects on one or several nodes in the network without the DAG reporting such relationships).
The third assumption is particularly important in clinical settings: to safely interpret a di-
rected connection as a causal effect, the experimental design should be set as to block any
confounding factors. A common device to achieve that is randomisation, which severs any
incoming causal link between the randomised variables and possible exogenous effects.
It is important to make a distinction between the probabilistic and causal interpretations
of Bayesian networks. From a causal perspective, the direction of arcs is uniquely identified
by the asymmetry between cause and effect: if we act on the cause, we may influence
the effect; but if we act on the effect, the cause remains unaffected. From a probabilistic
perspective, this is not true because of the reversibility of Bayes’ theorem. For instance, if
46
we consider again the DAG in Figure 2.2 we can write
Pr(v1, v2, v3, v4, v5) = Pr(v1) Pr(v2 | v1) Pr(v3 | v2) Pr(v4 | v2, v3) Pr(v5 | v2, v4)
where each node has a distribution conditional on its parents. However, for the nodes v1 and
v2 we have that Pr(v1) Pr(v2 | v1) = Pr(v2) Pr(v1 | v2). This implies that the DAG in which
the arc v1→ v2 is reversed into v2→ v1 encodes the same probability distribution as that in
Figure 2.2, despite having different arcs. The only arcs whose direction cannot be changed
in this way, and is thus uniquely identified even without making causal assumptions, are
those that are part of a collider; or those that would be part of a new collider or introduce
a cycle if they were reversed.
Another consequence of the duality between the probabilistic and the causal interpreta-
tion of Bayesian networks is that we can compute the conditional probability of any pair
of variables regardless of how we construct the DAG. Depending on the application, it may
make more sense to use a prognostic DAG in which arcs point from diseases to symptoms,
or a diagnostic DAG in which arcs point from symptoms to diseases. From a purely proba-
bilistic perspective, we note that for every diagnostic DAG there is a prognostic DAG that
represents the same probability distribution and vice versa. Clearly, one will be easier to
interpret than the other because the DAG will be easier to read. However, any conditional
probability we may wish to compute will be identical for both.
2.6.4 Structure learning algorithms
Structure learning algorithms learn the structure of Bayesian networks from data, given the
assumptions of Bayesian networks (such as the Markov property). Constraint-based algo-
rithms use statistical tests to learn conditional independence relationships (the constraints
themselves); score-based algorithms rank candidate DAGs based on some goodness-of-fit
criterion; hybrid algorithms use conditional independence tests to exclude the vast majority
47
of candidate DAGs, and then perform a score-based search on the few that are still under
consideration.
The Inductive Causation algorithm and its implementation
The Inductive Causation algorithm (IC) (Pearl and Verma, 1995) is the simplest example of
structure learning constraint-based algorithm.
Firstly, for each pair of variables A and B in X, the algorithm searches for a set SAB
such that A and B are independent given SAB and such that A and B are not part of it
(A,B /∈ SAB). Secondly, for each pair of variables that are not connected by an edge but that
are both connected to a common neighbour C, the algorithm checks whether C ∈ SAB: if
C /∈ SAB, then the direction of edge A−−C becomes A→C and that of edge C −−B becomes
C←B. Thirdly, the direction of the edges that are still undirected is set following two rules:
if A is adjacent to B and there is a strictly directed path from a to B, then A−−B becomes
A→B; if A and B are not adjacent but A→C and C −−B, then C −−B becomes C→B.
This step ends the algorithm which then returns the partially directed graph in which only
those arc directions that can be uniquely identified from the data are represented. The IC
algorithm is implemented by the Peter & Clark algorithm (PC) (Spirtes et al., 1993), which
starts from a saturated network and then performs tests that gradually increases the number
of conditioning nodes.
2.7 Discussion: applications and limitations of Bayesian
Artificial Intelligence in Medicine
Bayesian AI and networks are in use in several areas in the medical science and are a tool
of interest for physicians.
Applications of Bayesian AI in medicine stem from four main domains: diagnostic
reasoning—establishing a diagnosis in a target patient given clinical evidence; prognostic
48
reasoning—making predictions about what might happen in the future, since Bayesian net-
works encode a temporal information even in cross-sectional data; treatment selection—
making predictions about the possible effects for a treatment; studying functional interac-
tions among clinical evidence such as symptoms, signs and biomarkers.
Several examples from the four main domains described above illustrate the vast poten-
tial of Bayesian AI. First, recovering clinical evidence from the electronic medical record
is a substantial starting block for making inference: systems to construct clinical Bayesian
networks from electronic medical records have been developed (Shen et al., 2018). Second,
prognostic Bayesian networks are used to predict mortality in patients (Verduijn et al., 2007).
Third, Bayesian networks are also used for clinical decision support and treatment selection
in complicated cases (Sesen et al., 2013). Fourth, studying functional interactions among
symptoms in the medical domain of psychiatry shows great promise: because the classifica-
tion of mental disorders is rapidly shifting paradigms, the new approach of mental disorders
as networks of mutually influencing components (Borsboom and Cramer, 2013) is a promis-
ing setting for the application of Bayesian reasoning. Works have already endeavoured to
represent interacting symptoms for disorders like depression as DAGs in cross-sectional data,
therefore retrieving the possible causal relationships among them (Briganti et al., 2020b).
Future work in this area may for instance include different types of variables (other than
symptoms) in networks.
The use of Bayesian networks is limited by the assumptions required to correctly learn and
perform inference on their structure. It is up to the researchers to design studies accordingly:
blocking confounding variables is by far the most difficult task in this respect (Briganti et al.,
2020b).
In conclusion, Bayesian artificial intelligence captures uncertain reasoning in medicine
through the promising model of Bayesian networks: they can be learned automatically from
data and combine graphs and probabilities in a rigorous way, with algorithms that automate
reasoning and use the graphical part of the model to guide a computer system in computing
49
Chapter 3
A primer on undirected network
models
Abstract
In this work the estimation, inference and stability of network models with undi-
rected edges for the study of mental disorders are reviewed. Gaussian Graphical Models
and Ising Models capture the statistical relationships among variables of interest such
as symptoms in a network of mutually influencing items in continuous and binary data
sets. Cross-sectional networks as well as temporal networks can be estimated. The
main measures of inference in such networks, such as centrality, as well as the impor-
tance of stability analyses are also reviewed. The clinical implications of the study of
mental disorders as networks are discussed.
3.1 Introduction
The network theory of mental disorders considers that the psychiatric disease arises from
the set of interactions among its symptoms (Borsboom, 2008). Network theory comes with
a specific set of statistics called network analysis which includes methods allowing for the
study of network models from data sets (Epskamp and Fried, 2018).
51
Several steps are necessary when studying mental disorders as network structures. First
the network structure itself must be estimated from the data set: this first step implies a
choice for the network model itself, which can be estimated from continuous data set through
a Gaussian Graphical Model (GGM), or from a binary data set through an Ising Model
(IM), with or without a regularization procedure (Epskamp, 2019). Second, the relative
or absolute importance, or centrality of a network component, such as a symptom, can be
inferred from the network structure through several measures, such as strength, expected
influence, or bridge centrality. Third, the stability of the parameters that are estimated
within the network structure is evaluated through bootstrapping analyses (Epskamp and
Fried, 2018).
This work aims to review the fundamental concepts introduced in the emerging field of
network analysis: first, the GGM and its binary counterpart, the IM in cross-sectional data-
sets with both frequentist and Bayesian methods; second, the GGM in temporal (panel) data
sets; third, the most common inference techniques; third, the main stability analyses. The
clinical implications of the models are also discussed throughout this work.
3.2 Estimating network models from cross-sectional data
3.2.1 Network components in undirected graphical models
A node is defined as an entity, such as a part of a psychiatric construct or a symptom
(Briganti et al., 2018). An edge (a connection between two nodes) is understood as a pairwise
association between two variables. The network G = (V,A) is therefore composed of a set of
nodes V = {v1, v2, . . . , vp} and a set of edges A. In undirected graphical models, the edges
are weighted, that is, some edges are stronger than others in the network structures.
52
3.2.2 The Gaussian Graphical Model
Let y be a a normal multivariate vector y = (y1, ..., yp) with mean vector µ and a variance-
covariance matrix Σ. For all subjects,
y ∼ N(µ,Σ).
Let Θ be the inverse of Σ,
Θ = Σ−1
that is known as the precision matrix or a Gaussian Graphical Model (GGM). The
elements of of the GGM encode the partial correlation coefficients θij of two variables yi and
yj given all other variables in y, that is, y−(i,j):
Cor(yi, yj|y−(i,j)
)= − θij√
θii√θjj,
therefore, the GGM represents the network itself.
The partial correlation θij between yi and yj is used as the edge weight, that is, the
strength of the connections between nodes vi and vj in the network. Edge weights can be
positive (usually represented as blue connections) or negative (usually represented as red
connections) depending on the sign of θij. The presence of an edge between two nodes vi
and vj in the network can be interpreted as a conditional dependence relationship: node
Vi predicts (or is predicted by) node vj, after controlling for all other nodes in the network
v−(ij). GGMs can be estimated in R with the bootnet package (Epskamp and Fried, 2018).
3.2.3 Bayesian estimation of the Gaussian Graphical Model
The estimation of a GGM with Bayesian methods allows for providing evidence for the
hypothesis that best predicts the observed data. For instance, when testing for conditional
53
dependence relationships among nodes, providing a Bayes Factor between 3 and 20 as a
cut-off value is associated with positive evidence, while a Bayes Factor > 20 is associated
with strong evidence (Kass and Raftery, 1995). In this section an overview is provided to
Bayes Factor and the Wishart distribution, both necessary for the estimation of a Bayesian
GGM. They can be computed in R with the BGGM package (Williams and Mulder, 2019).
Bayes Factor
Bayes factor is used for the selection of a statistical model by quantifying the support for
a model over another (Lambert, 2018): it is understood as the radio of the conditional
probabilities of the data given the first model and the second model
Bf =p (data | model1)
p (data | model2),
that is, the ratio of the marginal likelihoods for each model p(data | modeli). Given a
variable X = (x1, x2, . . . , xp) with a probability distribution of parameters Θ, the marginal
likelihood is the probability p(X | model) where Θ has been integrated out (marginalized
out).
p(X | model) =
∫Θ
p(X | Θ)p(Θ | model)dΘ
Bayes factor chooses more parsimonious models and therefore penalizes complexity.
Wishart distribution
Given the multivariate data set with a normal distribution with mean 0, X ∼ (0,Σ), an
m× p matrix of , subjects and p variables X = (x1, x2, . . . , xp), the Wishart distribution
Wprior ∼ (Q, h)
is the conjugate prior (that is, the prior probability distribution) of the inverse-covariance
54
matrix or precision matrix, where Q is the scale matrix and h the degrees of freedom such
as h > p− 1. The posterior distribution is also a Wishart distribution
Θ | X ∼ Wposterior
(h+ n, (Sq + εIp)
−1)where Ip is the identity matrix, and Sq the sums of squares matrix X ′X. (Williams,
2018b). The Wishart probability distribution function follows
fΘ(θ) =|θ|(n−p−1)/2e− tr(Q−1θ)/2
2np2 |Q|n/2Γp
(n2
)where Γp is the multivariate gamma function (generalization of the factorial function for
complex numbers in multivariate statistics), and tr is the trace function (the sum of elements
on the main diagonal) of the square matrix.
3.2.4 The Ising Model
The Ising Model (IM) is the binary equivalent of the Gaussian Graphical Model (van Borkulo
et al., 2014; Marsman et al., 2018) used to estimate a network of partial correlations in
continuous datasets. From an Ising Model perspective, each variable in the network is
influences by all other variables, and this is represented in the following distribution for a
given variable X:
Z =∑x
exp
(∑i
τixi +∑<ij>
ωijxixj
).
With the Ising Model it is possible to determine the conditional probability distribution
of a variable Xi as it is predicted by other variables in the network given that their values
are known X(−i),
55
Pr(Xi|X(−i) = x(−i)
)=
Pr(X = x)
Pr(X(−i) = x(−i)
) =exp
(xi
(τi +
∑j ωijxj
))∑
xiexp
(xi
(τk +
∑j ωijxj
))where τ represents the threshold parameter which determines whether the variable Xi
prefers to be in state +1 (if τ is higher than 0) or -1 (if τ is lower than 0), ωij represents
the pairwise interaction between two variables Xi and Xj, and∑
xitakes the sum over both
possible outcomes of xi, that is the logistic regression model. Neural networks are used to
fit a non-regularized multinomial logistic regression as a loglinear binary logistic regression
in the data set (Venables and Ripley, 2002).
This is achieved by modeling the logarithm of the probability of seeing a given output
for the predicted variable Xi using the linear predictor X(−i) (that is, all other variables in
the network) and a normalization factor, the logarithm of the partition function:
ln Pr (Xi = K) = βK ·X(−i) − lnZ
where K is the given output, βK the regression coefficient and lnZ is the logarithm of
the partition function, which in turn can be estimated as
Z =K∑k=1
exp(βk ·X(−i))
,
and represents a constant, in terms that it does not depend on the predicted variable
Xi. The partition function, compared to the Ising distribution, generates the following
equivalency for the linear predictor
∑i
τixi +∑<ij>
ωijxixj = βK ·X(−i)
in the multinomial logistic regression approach to the Ising Model.
56
Because of the need to sum many terms to estimate the Ising Model, the computation
without using additional techniques is only possible for data sets where the number of vari-
ables is inferior or equal to 10. Regularized versions of the Ising Model, such as used in
recent applied studies (Briganti and Linkowski, 2019a) allow for computation in data sets
with a higher number of variables. Ising Models can be computed in R with the bootnet
package (Epskamp and Fried, 2018).
3.2.5 Interpreting undirected network models in cross-sectional
data sets
The interpretation of the network model depends on the model itself and the variables in-
cluded. The interpretation of a connection in a network is somewhat more straightforward
for symptoms: if symptom A and symptom B are connected in a network, than the two pre-
dict each other. It is less straightforward however to interpret connections among two items
from a psychometric scale that measures a construct: one can interpret that a connection
still means that the traits that are measured predict each other, and its clinical implication
lies in the better understanding of psychiatric constructs themselves.
3.3 Estimating network models from several time points
Panel data are interesting to monitor the evolution of symptoms over time, for instance on
admittance, at the middle and the end of the hospital stay of a patient. To see whether
a symptom at one time point influences another symptom at the subsequent time point, a
specific network model exists that allow for the encoding of temporal effects. Such temporal
predictions are known as Granger causality. In this section, both Granger causality and the
temporal network models are reviewed.
57
3.3.1 Granger causality
Granger causality is a statistical test (Granger, 1969) for the hypothesis that a variable (or
a set of variables) predicts another (Lutkepohl, 2005).
Let X(t) ∈ Rd×1 for t = 1, . . . , T be a d-dimensional panel data set. Granger causality is
obtained by estimating a Vector Auto-Regressive (VAR) model with L time lags
X(t) =L∑τ=1
AτX(t− τ) + ε(t)
where τ = 1, . . . , L, ε(t) a Gaussian random vector, and Aτ a matrix for every τ . A
variable Xi is said to Granger-cause another variable Xj if the element Aτ (ji) is greater
than 0.
3.3.2 The Graphical Vector Autoregressive Model
To model the dynamics of manic symptoms with a pharmacological intervention in panel
data, a panel Graphical Vector Autoregressive Model (GVAR). This model was first intro-
duced in recent works with its own package for computation (Epskamp, 2019) to translate
time-series methods to panel data. GVAR can be seen as a multivariate multiple regression
on the previous measurement occasion.
For a set of symptoms y = (y1, ..., yp) measured in a given individual, GVAR is expressed
as
yt1 |yt0 ∼ N(µ+B(yt0 − µ),Σζt0),
where B represents a p × p matrix of temporal effects, µ the vector of means, Σt0 the
variance-covariance matrix on measurement occasion t0, and ζ a vector of normally dis-
tributed innovations. Because B encodes temporal prediction, a nonzero matrix element bij
means that yt1 is predicted by yt0 : this prediction is known as Granger causality (Granger,
1969) because the condition of cause preceding the effect is fulfilled.
58
3.4 Network inference
Network inference is the domain of network analysis that deals with reaching conclusions
about the network structure that has been estimated from the data. Often, network re-
searchers want to study which nodes are more important than others: in psychiatry, that
translates with identifying symptoms that could be chosen as targets for a clinical interven-
tion (either with psychopharmacology or with psychotherapy).
3.4.1 Centrality
Network centrality is a relative approach to identifying important symptoms or items in a
network: it is relative because there will always be a more central item, even in a poorly
connected network (Briganti et al., 2019). Network centrality is however important in identi-
fying which symptoms connect more than others, and can therefore aid in poorly connected
networks.
Strength (sometimes called degree) is the absolute sum of the edge weights w of a node
Vi,
CwS (i) =
N∑j
| wij |
A variant of strength is expected influence (Jones, 2017), which is the sum of edge weights,
and therefore accounts for negative edges in a network
CwEI(i) =
N∑j
wij.
Betweenness is understood as the shortest paths that go through the node under investi-
gation, while closeness measures the sum of shortest paths from the node under investigation
to all other nodes in the network (Boccaletti et al., 2006); betweenness and closeness are
seldom used in symptom networks.
59
3.4.2 Node predictability
Node predictability is used to have an absolute estimate of a node’s connectedness: the
measure of R2 (coefficient of determination) is used to convey this information.
Let us consider a regression model of outcomes y and predictors X with predicted values
E(y | X, θ), fit to data (X, y)n, n = 1, . . . , N. Ordinary least squares yields an estimated
parameter vector θ with predicted values yn = E(y | Xn, θ
)and residual variance VarNn=1 yn,
where I am using the notation,
VarNn=1 zn =1
N − 1
N∑n=1
(zn − z)2 , for any vector z
Node predictability, that is, the proportion of shared variance, is
R2 =VarNn=1 yn
VarNn=1 yn
and is computed as an R function in the package mgm (Haslbeck and Waldorp, 2016).
A Bayesian approach to the estimation of node predictability also exists as it was in-
troduced recently in the literature (Gelman et al., 2019). Expected values conditional on
unknown parameters are used
ypredn = E (yn | Xn, θ)
where yn represents a future observation from the model with predictors Xn. However,
because several posterior distributions are drawn θs, s = 1, . . . , S the vector of predicted
values is written as ypredsn = E (y | Xn, θ
s).
The residual variance varres is defined as
varres = E(V Nn=1
(yn − ypred
n
)| θ)
60
and the Bayesian R2 is expressed as
Bayesian R2s =
VarNn=1 ypred sn
VarNn=1 ypred sn + varsres
,
and is computed as a function in the R package BGGM (Williams and Mulder, 2019).
Although it is interpreted as an absolute measure of connectedness, it is also understood
as the upper bound of controllability : if one assumes that all the edges of a node are directed
towards it, then node predictability reflects how a node can be controlled; on the other hand,
if one assumes that all the edges are directed towards other nodes, than node predictability
reflects how a node can control others (Briganti et al., 2019).
3.5 Stability and accuracy
The question of stability and accuracy of results obtained in the estimation of psychiatric
networks quickly became important in this new field of research. Networks have shown
similar results in several populations in the context of post-traumatic stress disorder (Fried
et al., 2018) and this adds a supporting argument for the replication of the results obtained
via these analyzes.
The most used instrument to analyze the stability and precision of network parameters
is bootstrapping (Epskamp and Fried, 2018). This method re-calculates the parameters that
make up the network to answer specific questions. As far as the edges are concerned, the
two questions are as follows: 1) is edge A between two nodes X and Y really larger than
edge B between two nodes V and W? and 2) is edge A accurately estimated?
To answer the first question, confidence intervals are estimated around the difference
between edges A and B as determined by a number of re-estimates of choice; generally, 2000
network re-estimates are performed (Briganti et al., 2018), if this test proves positive, the
observer can interpret that an edge A is significantly larger than an edge B (Briganti et al.,
2019). To answer the second question, confidence intervals are calculated around several
61
thousand re-estimates from all edges of the network. If these confidence intervals are not
very wide and do not overlap between different edges, then the observer can better appreciate
the difference between different edges of the network.
The same process can be applied for centrality indicators: in that case, subset bootstrap-
ping is applied, that is, a bootstrapping process that involves gradually reducing the number
of subjects in the sample and re-estimating the network parameters so as to test how stable
is the centrality order (the order of which symptoms rank in centrality estimates).
3.6 Conclusion
This work aimed to review the introductory concepts of network analysis of undirected graph
structures and how to conduct such analyses in cross-sectional and time-series. Although
specific to the field of networks of mental disorders, these statistical techniques can be applied
in a wide range of specialties and are highly versatile for both low and high dimensional data.
62
Chapter 4
A network model of empathy
Abstract
The aim of this work is to perform a network analysis on the French adaptation
of the Interpersonal Reactivity Index (IRI) scale from a large Belgian database and
provide additional information for the construct of empathy. I analyze a database of
1973 healthy young adults who were queried on the IRI scale. A regularized partial
correlation network is estimated. In the visualization of the model, items are displayed
as nodes, edges represent regularized partial correlations between the nodes. Central-
ity denotes a node’s connectedness with other nodes in the network. The spinglass
algorithm and the walktrap algorithm are used to identify communities of items, and
state-of-the-art stability analyses are carried out. The spinglass algorithm identifies
four communities, the walktrap algorithm five communities. Positive edges are found
among nodes belonging to the same community as well as among nodes belonging to
different communities. Item 14 (“Other people’s misfortunes do not usually disturb
me a great deal”) shows the highest strength centrality score. The network edges and
node centrality order are accurately estimated. Network analysis highlights interesting
connections between indicators of empathy; how these results impact empathy models
must be assessed in further studies.
64
4.1 Introduction
Empathy is a main component of short-term as well as long-term human interactions. Despite
its importance and because of its complexity, a unified definition is yet to be found. For some
authors, empathy incarnates the ability to perceive and be sensitive to others’ emotions and
the desire for their well-being (Decety et al., 2016). It is not to be confused with sympathy,
which is considered to be a part of empathy and defined as the consciousness of another’s
emotions and feelings without sharing them, together with a feeling of pity (Wispe, 1986).
Empathy is a key item to mental health professionals because it belongs to a collection of
indicators of good outcomes in psychotherapy (Elliott et al., 2011).
In 1980, Mark H. Davis presented a self-report empathy questionnaire, the Interpersonal
Reactivity Index (IRI), where he identified the construct as built upon two dimensions (Davis,
1980). The first one represents the cognitive dimension, or the tendency to adopt others’
perspectives and feelings; the second one represents an affective dimension reflecting one’s
feeling of another’s emotional state (Decety and Jackson, 2004).
Out of these two dimensions Davis identified four components in his model of empathy:
(1) Fantasy (belonging to the cognitive dimension), or the tendency to get involved in the
actions and feelings of one or more fictional characters in movies, books or plays (e.g.,
item 23—“When I watch a good movie, I can very easily put myself in the place of a
leading character”); (2) Perspective taking (also belonging to the cognitive dimension), or
the tendency to comprehend others’ point of view (e.g., item 25—“When I am upset at
someone, I try to put myself in his shoes for a while”); (3) Empathic concern (belonging
to the affective dimension), the feeling of concern and sympathy for people in distress (e.g.,
item 9—“When I see someone being taken advantage of, I feel kind of protective toward
them”); (4) Personal distress (also belonging to the affective dimension), or the feeling of
unease in difficult, tense or emotional situations (e.g., item 10—“I sometimes feel helpless
when I am in the middle of a very emotional situation”).
Even though the two-dimension model is frequently accepted (Bohart and Greenberg,
65
1997; Davis, 1980; Decety and Jackson, 2004; Reniers et al., 2011), further models were pro-
posed, such as Blair’s (Blair, 2005), which distinguished three components (motor, cognitive
empathy and emotional). Cliffordson proposed a hierarchical model putting the empathic
concern factor at the top of the pyramid (Cliffordson, 2002). Empathy is an important issue
for psychiatrists. Its dysfunctioning is part of major psychiatric diseases such as psychopathy
and autism (Blair, 2005) and is perceived by patients as a key element to treatment (Ross
and Watling, 2017).
In the last few years, a new way of analyzing data in psychology and psychiatry has
arisen: network analysis. In this conceptual model (Borsboom and Cramer, 2013), pairwise
interactions among symptoms represent a network of mutually influencing elements. This
model has affirmed itself as a way of analyzing mental disorders such as depression (Beard
et al., 2016; Boschloo et al., 2016; Fried and Cramer, 2017), posttraumatic stress disorder
(Bryant et al., 2017), as well as autism and obsessive–compulsive disorder (Ruzzano et al.,
2015) by focusing on the interaction between symptoms, attributes, emotions, and behaviors
(Fried and Cramer, 2017).
Network analysis provides a new opportunity to conceive psychological constructs not as
the consequence of an underlying disease as in the latent variable model, but instead as con-
stituted by the mutual interaction of its items. While largely applied to research on mental
illness, network models have been used in other psychological sciences such as personality
(Costantini et al., 2015), health-related quality of life (Kossakowski et al., 2016), intelligence
(van der Maas et al., 2006), and attitudes (Dalege et al., 2017). Network models have also
been used to specifically investigate the structure of multivariate data in psychology, for
instance to identify the number of item clusters: this is the case of recent papers concerning
PTSD (Gluck et al., 2017) and development (Demetriou et al., 2017).
This paper extends this conceptual framework to the psychological construct of empathy.
Network analysis facilitates the identification of interactions between psychological variables
such as items on self-report questionnaires; allows for the estimation of item communities
66
(i.e. clusters of items that are closely related with each other); and can give insights into the
connectedness or importance of items within the network, often referred to as ‘centrality’
(Boccaletti et al., 2006).
According to Davis’ model (1980), I might expect significant positive relations between
items from the Empathic concern scale and items from the Perspective taking and the Fan-
tasy scales. Inspired by network analysis in other fields of psychological science, I apply
network models for the first time to the domain of empathy research, specifically, to the 28-
item French version of the IRI (Braun et al., 2015). This paper highlights potential insights
that network analysis can offer — as a complementary tool to factor modeling that is more
established in the field — to empathy research. The primary aim of the paper is to explore
empathy items and their relationships in an empathy network, and the secondary aim is to
build up on prior factor modeling work in this dataset. Braun and colleagues used confir-
matory or exploratory factor analysis (CFA and EFA) to investigate the factor structure in
the present data (Braun et al., 2015), and I want to use community detection algorithms to
see whether the results align with prior work, and to discuss why the identified communities
have a radically different interpretation (Demetriou et al., 2017; Golino and Epskamp, 2017).
4.2 Method
4.2.1 Data set
The database for this study was composed of 1973 French-speaking students in several uni-
versities or schools for higher education in the following fields: engineering (31%), medicine
(18%), nursing school (16%), economic sciences (15%), physiotherapy, (4%), psychology
(11%), law school (4%) and dietetics (1%). The subjects were 17 to 25 years old (M = 19.6
years, SD = 1.6 years), 57% were females and 43% were males. Even though the full data set
was composed of 1973 participants, only 1270 answered the full questionnaire: I dealt with
missing data by using pairwise complete observations in estimating a Gaussian Graphical
67
Model, meaning that I used all available information from every subject.
The IRI is composed of 28 items meant to assess the four following components: fan-
tasy, perspective taking, empathic concern and personal distress. In the questionnaire, the
items are mixed; reversed items (items 3, 4, 7, 12, 13, 14, 15, 18, 19) are present. Items
are scored from 0 to 4, where “0” means “Doesn’t describe me very well” and “4” means
“Describes me very well”; reverse-scoring is calculated afterwards. The IRI questionnaires
were anonymized. The reanalysis of the database in this retrospective study was approved
by the ethical committee of the Erasmus Hospital.
Table 4.1: The Interpersonal Reactivity Index.
Item Item Label Domain
color
Item meaning
1 1FS Green I daydream and fantasize, with some regularity, about
things that might happen to me.
2 2EC Purple I often have tender, concerned feelings for people less
fortunate than me.
3 3PT R Yellow I sometimes find it difficult to see things from the “other
guy’s” point of view. (Reversed)
4 4EC R Purple Sometimes I don’t feel very sorry for other people when
they are having problems. (Reversed)
5 5FS Green I really get involved with the feelings of the characters
in a novel.
6 6PD Red In emergency situations, I feel apprehensive and ill-at-
ease.
7 7FS R Green I am usually objective when I watch a movie or play, and
I don’t often get completely caught up in it. (Reversed)
68
8 8PT Yellow I try to look at everybody’s side of a disagreement before
I make a decision.
9 9EC Purple When I see someone being taken advantage of, I feel
kind of protective towards them.
10 10PD Red I sometimes feel helpless when I am in the middle of a
very emotional situation.
11 11PT Yellow I sometimes try to understand my friends better by
imagining how things look from their perspective.
12 12FS R Green Becoming extremely involved in a good book or movie
is somewhat rare for me. (Reversed)
13 13PD R Red When I see someone get hurt, I tend to remain calm.
(Reversed)
14 14EC R Purple Other people’s misfortunes do not usually disturb me a
great deal. (Reversed)
15 15PT R Yellow If I’m sure I’m right about something, I don’t waste
much time listening to other people’s arguments. (Re-
versed)
16 16FS Green After seeing a play or movie, I have felt as though I were
one of the characters.
17 17PD Red Being in a tense emotional situation scares me.
18 18EC R Purple When I see someone being treated unfairly, I sometimes
don’t feel very much pity for them. (Reversed)
19 19PD R Red I am usually pretty effective in dealing with emergencies.
(Reversed)
20 20FS Green I am often quite touched by things that I see happen.
69
21 21PT Yellow I believe that there are two sides to every question and
try to look at them both.
22 22EC Purple I would describe myself as a pretty soft-hearted person.
23 23FS Green When I watch a good movie, I can very easily put myself
in the place of a leading character.
24 24PD Red I tend to lose control during emergencies.
25 25PT Yellow When I’m upset at someone, I usually try to “put myself
in his shoes” for a while.
26 26FS Green When I am reading an interesting story or novel, I imag-
ine how I would feel if the events in the story were hap-
pening to me.
27 27PD Red When I see someone who badly needs help in an emer-
gency, I go to pieces.
28 28PT Yellow Before criticizing somebody, I try to imagine how I
would feel if I were in their place.
4.2.2 Network analysis
The software used for the analysis is R (version 3.4.0, open source, available at https:
//www.r-project.org/). I used the packages qgraph, (Epskamp et al., 2012) and glasso
(Friedman et al., 2014a) for network estimation and visualization, mgm, for node predictabil-
ity (Haslbeck and Waldorp, 2016), igraph, (Csardi and Nepusz, 2006) for the spinglass algo-
rithm, walktrap algorithm and bootnet, (Epskamp and Fried, 2018) for stability.
Network estimation
I estimated Spearman correlations for the 28 ordinal items, which was the input to estimate
a Gaussian Graphical Model (GGM), a regularized partial correlation network (Epskamp
70
and Fried, 2018). I used Spearman correlations instead of polychoric correlations because
of low variability between items that can lead to zeroes in the marginal crosstables. The
graphical lasso (least absolute shrinkage and selection operator) was used to regularize the
edge weight parameters resulting from the GGM, which ensures avoiding the estimation of
spurious edges.
Nodes represent items from the French adaptation of IRI. Edges are connections between
two nodes: they are regularized partial correlations between two items of the questionnaire.
An edge between two items therefore means that there is an association after controlling
for all other nodes in the network. Statistically speaking, an edge between items in the IRI
network can be interpreted as following: when two nodes A and B are strongly connected
and the observed group scores high on A, the observed group is more likely to also score
high on B, controlling for all other nodes in the network.
Nodes are placed in the network using the Fruchterman–Reingold algorithm, which de-
termines the position of the node based on the sum of connections it has with other nodes
(Fruchterman and Reingold, 1991). Each edge has a sign: blue edges represent positive
regularized partial correlations whereas red edges represent negative regularized partial cor-
relations. The corresponding thickness and saturation of an edge denote its weight (i.e. the
strength of association).
Network inference
The centrality plot illustrates the centrality of a node in connection with other nodes. Boc-
caletti et al. described three types of centrality: strength, betweenness, and closeness (Boc-
caletti et al., 2006). One can understand strength centrality as the sum of direct connections
a given node has in the network; betweenness is understood as the shortest paths that go
through the node under investigation; closeness measures the sum of shortest paths from the
node under investigation to all other nodes in the network. Since centrality represents the
relative importance of a node in a network, Freeman conceptualized three possible interpre-
71
tations to a central item (Freeman, 1978): control, independence or activity.
Statistically speaking, a central item shares the most variance with all other items. Con-
ceptually, and in case of IRI, which is a self-administered scale, I suggest that the answer of
a subject to a central item might predict the way the subject answers to other items which
share a connection with it in the network. Centrality estimates are standardized with a mean
of 0 and a standard deviation of 1, and strength centrality is the main metric used in this
paper since it is the most robustly estimated centrality metric described in the literature
(Epskamp and Fried, 2018). However, centrality measures are relative metrics, since the
centrality of each node is estimated in comparison with other nodes (there is always a highly
central node, no matter how weak the edges in the network are). I therefore also estimated
node predictability. Node predictability represents the shared variance of each node with
all its neighbors, which constitutes an absolute measure of its interconnectedness (Haslbeck
and Fried, 2017).
The spinglass algorithm was used to identify communities of items in the GGM. It is
based on the principle that edges should connect nodes of the same community, whereas
nodes belonging to different communities should not be connected (Yang et al., 2016). It
is important to note that an item can only be part of one community using this procedure.
Since the spinglass algorithm can give different results in the same sample, I assessed the
stability of the solution by running the algorithm 100 times and extracted the number of
communities with the highest frequency. To complement the results, I also used the walktrap
algorithm, which is based on the principle that adjacent nodes tend to belong to the same
community (Yang et al., 2016). The walktrap algorithm is shown to have high accuracy in
simulation studies (Golino and Epskamp, 2017; Demetriou et al., 2017).
Network accuracy and stability
I tested the accuracy of edge weights and the stability of the order of centrality estimation
through bootstrapping (Epskamp et al., 2017a, I used 2000 bootstraps). I bootstrapped 95%
72
confidence intervals of all edge weights, followed by the edge-weights comparison test and an
edge weight difference test to see which edges differ from each other in size significantly (to
answer the question is edge A significantly larger than edge B). I used the subsetting boot-
strap procedure that re-estimates the network with a dropping percentage of participants to
determine the stability of centrality estimation, and results in a centrality-stability coeffi-
cient (CS-coefficient) that should not be lower than 0.25 and preferably above 0.5. Finally,
I performed a centrality difference test to see which centrality estimates differ statistically
from each other (to answer the question is node A significantly more central than node B).
4.3 Results
4.3.1 Empathy network
Figure 4.1 illustrates the estimated network of the 28-item IRI. Overall, most items are
positively connected within the network. Item 16 (“After seeing a play or movie, I have felt
as though I were one of the characters”) is strongly connected to item 23 (“When I watch a
good movie, I can very easily put myself in the place of a leading character”) (weight 0.38).
Item 4 (“Sometimes I don’t feel very sorry for other people when they are having problems”)
has a wide edge to item 14 (“Other people’s misfortunes do not usually disturb me a great
deal”) (weight 0.29). Other strong edges include item 10 (“I sometimes feel helpless when I
am in the middle of a very emotional situation”) and item 17 (“Being in a tense emotional
situation scares me”), item 24 (“I tend to lose control during emergencies”) and item 27
(“When I see someone who badly needs help in an emergency, I go to pieces”), item 25
(“When I’m upset at someone, I usually try to “put myself in his shoes” for a while”) and
item 28 (“Before criticizing somebody, I try to imagine how I would feel if I were in their
place”).
The spinglass algorithm identifies a mean of four communities of items corresponding to
the four factors of the IRI as proposed originally and confirmed by Braun et al. (Braun
73
1
2
3
4
5
6
7
8
910
11
12
13
14
15
16
17
18
19
2021
22
23
24
25
26
27
28
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
A1: 1FS5: 5FS7: 7FS_R12: 12FS_R16: 16FS23: 23FS26: 26FS
B3: 3PT_R8: 8PT11: 11PT15: 15PT_R21: 21PT25: 25PT28: 28PT
C2: 2EC4: 4EC_R9: 9EC14: 14EC_R18: 18EC_R20: 20EC22: 22EC
D6: 6PD10: 10PD13: 13PD_R17: 17PD19: 19PD_R24: 24PD27: 27PD
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
A1: 1FS5: 5FS7: 7FS_R12: 12FS_R16: 16FS23: 23FS26: 26FS
B3: 3PT_R8: 8PT11: 11PT15: 15PT_R21: 21PT25: 25PT28: 28PT
C2: 2EC4: 4EC_R9: 9EC14: 14EC_R18: 18EC_R20: 20EC22: 22EC
D6: 6PD10: 10PD13: 13PD_R17: 17PD19: 19PD_R24: 24PD27: 27PD
Figure 4.1: Network composed of the 28-item IRI. Each item is represented by a node(1 to 28) and belongs to a different community of empathy, indicated by a code in thecolumn on the right: Fantasy Scale (FS), Perspective Taking (PT), Empathic Concern (EC)and Personal Distress (PD). Reversed items are marked with an R (e.g. 7FS R indicatesa reversed item). Blue lines are positive connections, red lines are negative connections.The thickness of the line represents the connection strength. Colored areas in the ringssurrounding the nodes represent the node predictability (percentage of variance of a givennode explained by surrounding nodes).
74
et al., 2015). Cluster A is composed of items 1, 16, 23, 26, 5, 12, 7, forming the Fantasy
component (FS). Cluster B is formed by items 25, 28, 21, 8, 11, 15, 3, all of which constitute
the Perspective-taking component (PT). Cluster C is formed by items 22, 20, 2, 14, 18, 4, 9
and reflects the Empathic concern component (EC). Cluster D is formed by items 10, 17, 6,
24, 27, 13, 19 and represents the Personal distress component (PD).
The walktrap algorithm identifies 5 communities of items. Most items belong to the same
communities in the spinglass solution above, whereas items 6 (“In emergency situations, I
feel apprehensive and ill-at-ease”), 10 (“I sometimes feel helpless when I am in the middle
of a very emotional situation”) and 17 (“Being in a tense emotional situation scares me”)
form a new community of items (community 5).
Furthermore, in some cases, two items from different communities (as identified by the
spinglass algorithm) have a positive connection: for example, this is the case of item 1 (“I
daydream and fantasize, with some regularity, about things that might happen to me”) and
item 10 (“I sometimes feel helpless when I am in the middle of a very emotional situation”),
item 23 (“When I watch a good movie, I can very easily put myself in the place of a leading
character”) and item 22 (“I would describe myself as a pretty soft-hearted person”), item 8
(“I try to look at everybody’s side of a disagreement before I make a decision”) and item 9
(“When I see someone being taken advantage of, I feel kind of protective towards them”).
Mean node predictability is 0.27, which means that on average, 27% of the variance of each
node is explained by its neighbors: assuming that all edges go to the node under investigation
from its neighbors, I can see how well the given node can be predicted by the other nodes
surrounding it (Haslbeck and Fried, 2017).
4.3.2 Network inference
In Figure 4.2, I illustrate the strength centrality estimates for the 28 questionnaire items.
Item 14 (“Other people’s misfortunes do not usually disturb me a great deal”) has the
highest standardized strength centrality in the network. Other central items include node
75
−2
0
2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28
Cen
tral
ity
Figure 4.2: Strength centrality estimates for the 28-item IRI. The Y-axis represents thecentrality indices as standardized z-scores (the greater the estimate the more central theitem is), and the X-axis represents the 28 IRI items.
76
10 (“I sometimes feel helpless when I am in the middle of a very emotional situation “), and
node 26 (“When I am reading an interesting story or novel, I imagine how I would feel if the
events in the story were happening to me”). Items 1 (“I daydream and fantasize, with some
regularity, about things that might happen to me”) and 15 (“If I’m sure I’m right about
something, I don’t waste much time listening to other people’s arguments”) show the lowest
strength centrality values.
4.3.3 Network accuracy and stability
The edge weight bootstrap revealed relatively small CIs, which indicates a more precise
estimation. The edge weight difference test reveals that the empathy network is accurately
estimated and that the strongest edges are significantly stronger than other edges. The
subset bootstrap shows that the order of item strength centrality is more stable than the
other kinds of centrality values, which is consistent with numerous prior papers (Armour
et al., 2017; Epskamp and Fried, 2018). CS-values obtained are 0.44 for node betweenness,
0.67 for node closeness and 0.75 for node strength. CS-values should preferably be above 0.5
and should not in any case be lower than 0.25: my results are above 0.5 and are therefore very
stable. The centrality difference test shows that highest centrality estimates are statistically
different from lowest centrality estimates, even though a statistical difference is not shown
among nodes with the highest strength centrality estimates.
4.4 Discussion
The network analysis I presented is, to my knowledge, the first one applied to empathy
research. This study highlights connections between empathy components and provides new
insights on how they might interact: some items are more interconnected than others, items
differ in centrality, and interactions exist between items from different empathy components.
Positive connections are found throughout the network, confirming that, in my sample,
77
most items from the IRI share some variance and are connected. However, some items present
weak with others; this means that some nodes are conditionally independent of all other items
in the network. The spinglass algorithm identifies on average four communities of items in the
network, corresponding to the four a priori components of Davis’ construct: Fantasy (cluster
A), perspective taking (cluster B), empathic concern (cluster C) and personal distress (cluster
D). The identification of these four node communities supports, using a different method, the
results of Braun et al.’s confirmatory factor analysis study (Braun et al., 2015). This is not
necessarily surprising, given that network and factor models are, under certain conditions,
mathematically interchangeable (Kruis and Maris, 2016).
Even though communities might be mathematically close to factors, from a network
perspective they mean something entirely different: they are clusters of interrelated items
that stem from mutual dynamics; they actively contribute to the construct of empathy itself.
However, the walktrap algorithm identified five communities, describing a fifth community
formed by items 6, 10 and 17. This is a consequence of the strong connection and clustering
of these three items, which nonetheless share two important connections with the rest of the
Personal Distress cluster (6-24 and 10-24). Some items belonging to a given community are
connected to those from different communities, suggesting — from a network perspective —
that empathy communities interact with each other in the network through specific items.
For example, there is a connection between items 8 and 9, respectively belonging to the
Perspective taking and Empathic concern subscales.
Items belonging to the Empathic concern community (9, 14 and 20) have high centrality
values; this finding supports Cliffordson’s theory that puts Empathic concern at the basis (in
this case, at the center) of empathy. Items from the Empathic concern cluster are connected
to all the other communities in the network. Item 14 shows the highest centrality value: to
interpret this finding, one must associate the statistical meanings of centrality and network
connections (edges). First, strength centrality (the main subtype used) means that the sum
of all edges of item 14 to all other nodes is the highest in the IRI network; second, a connection
78
between item 14 and another item means for instance that a high-score answer to item 14
(which is reversed) lets us guess a high-score answer to all the items item 14 is connected
to, controlling all other nodes. I can then interpret the high centrality of item 14 as the one
that might influence and/or might be influenced by most answers of the IRI. However, when
I look at the centrality difference test, I understand that the strength centrality of node 14
is not statistically different than strength centralities of nodes 10, 26, 20, 23 and 24, but is
statistically different from that of all other nodes: this means that these nodes are roughly
equivalent in their centrality. Node predictability, especially when focusing on the average,
is somewhat more straightforward to interpret: on average, if I influence a group of nodes
surrounding a given node, and assume that all edges go towards this node, I can influence
27% of its variance (Haslbeck and Waldorp, 2016).
Stability analysis shows that both centrality and edge weight estimates were reasonably
stable. My results must be interpreted in the light of a number of limitations. First, my
empathy network is estimated from a sample of young adults, which likely limits the general-
izability of my results; further studies should investigate networks structures across different
samples. Second, because I used cross-sectional data to carry out the analyses, I cannot
determine the direction of edges. For instance, I cannot interpret whether the most central
item activates other items, is activated by other items, or both. Third, similar to many other
statistical models such as factor models, the network model used here estimates between-
subjects effects on a group level. This means that network properties such as structure or
centrality may not replicate in the same way in single individuals. Fourth, Marshall et al.
provided evidence that the order in which items are presented in a questionnaire may influ-
ence their relationships (Marshall et al., 2013). Again, this is a limitation for any statistical
model based on the correlation matrix among items, such as factor models, and not a spe-
cific shortcoming over network models, but important enough to warrant mentioning. Fifth,
network analysis, in which I interpret edges as putative causal connections, is based on the
premise that nodes differ from each other meaningfully: if two nodes represent the same
79
aspect of a construct, an edge is not a putative causal connection, but simply represents
shared variance (Fried and Cramer, 2017). IRI might in some cases have this problem, for
instance item 7 (“I am usually objective when I watch a movie or play, and I don’t often get
completely caught up in it”) and item 12 (“Becoming extremely involved in a good book or
movie is somewhat rare for me”) seem to measure the same concept.
Future research may also endeavor to apply empathy networks in people with psy-
chopathology.
80
Chapter 5
A network model of self-worth
Abstract
This study investigates the Contingencies of Self-Worth Scale (CSWS) in a sample
of 680 university students from a network perspective. I estimated regularized partial
correlations among seven CSWS domains: family support, competition, appearance,
God’s love, academic competence, virtue and other’s approval. Competition - academic
competence and competition – appearance represent the strongest connections in the
network. Mean node predictability (shared variance with surrounding nodes) is 0.25.
Appearance and academic competence were the most central (i.e. interconnected)
domains in the network. Future studies should explore the network structure of self-
worth in other healthy adult samples, and also in people with psychopathology.
5.1 Introduction
The human desire to feel worthy is an important constituent of human behavior (Pyszczynski
et al., 2004). A troubled self-esteem has been shown to contribute to several psychiatric
disorders such as eating disorders (Pearl et al., 2014), substance abuse (James, 2011), and
schizophrenia (Xu et al., 2013).
The Contingencies of Self-Worth Scale (CSWS) is a psychometric tool proposed by
81
Crocker and colleagues (Crocker et al., 2003) to assess seven domains of self-esteem: (1)
family support measures the influence of perceived approval, support and love from family
members on the feeling of self-worth (e.g. item 7 “Knowing that my family members love
me makes me feel good about myself”); (2) competition evaluates how self-worth is influ-
enced by feeling better than others (e.g. item 12 “Knowing that I am better than others on
a task raises my self-esteem”); (3) appearance quantifies how physical traits influence the
way people evaluate themselves (e.g. item 1 “When I think I look attractive, I feel good
about myself”); (4) God’s love measures the association between religiosity and self-esteem
(e.g. item 2 “My self-worth is based on God’s love”); (5) academic competence evaluates
the impact of grades on self-esteem (e.g. item 20 “Doing well in school gives me a sense
of self-respect”); (6) virtue measures the connection between self-worth and the adherence
to a moral code (e.g. item 5 “Doing something I know is wrong makes me lose my self-
respect”); (7) other’s approval measures the influence of perceived approval from others on
self-esteem (e.g. item 9 “I can’t respect myself if others don’t respect me”). This model of
self-esteem has already undergone structural validation (Crocker et al., 2003) which makes
it an interesting tool for exploring the construct of self-esteem (Geng and Jiang, 2013).
A common understanding of self-esteem is that the seven domains are all observable in-
dicators of self-esteem, that is, the domains of the questionnaire do not actively contribute
to the construct — they are effects of the construct. In the last decade, a new way of
conceptualizing psychological constructs has been proposed: network theory, which hypoth-
esizes psychological constructs as interacting systems. Network models are related statistical
models that can be used to try to uncover such structures in data: a network is formed by
pairwise interactions of its components (Borsboom and Cramer, 2013) usually calculated as
regularized partial correlations (Epskamp and Fried, 2018). Components of a network mu-
tually influence each other to actively participate in the emergence of a construct. Mental
disorders such as depression (Beard et al., 2016; Boschloo et al., 2016; Fried and Cramer,
2017; Mullarkey et al., 2018), schizophrenia (Galderisi et al., 2018), posttraumatic stress dis-
82
order (Fried et al., 2018), autism, and obsessive-compulsive disorder (Ruzzano et al., 2015)
have been conceptualized and analyzed statistically from a network perspective. Network
structures of psychological constructs such as empathy (Briganti et al., 2018), personality
(Costantini et al., 2015), health-related quality of life (Kossakowski et al., 2016), intelligence
(van der Maas et al., 2006), and attitudes (Dalege et al., 2017) have also been studied.
Researchers usually analyze constructs as network composed of items — answers of the
observed group to a given questionnaire such as the Interpersonal Reactivity Index (Briganti
et al., 2018). However, scales in psychology are usually constructed to assess one underlying
dimension; this means that they often feature several highly similar items that might measure
the same thing, which has been discussed as a challenge for network models (Fonseca-Pedrero
et al., 2018). In that case, the meaning of the connection between items changes: an associa-
tion between X and Y simply reflects the shared variance of the two items, and not a genuine
mutual relation (Fried and Cramer, 2017). This limitation also holds for the Contingencies
of Self-Worth Scale where a common cause is plausible: items in a given domain might mea-
sure the same construct, and can therefore also be explored with factor models. My work
thus aims to apply network modeling to the construct of contingent self-worth as described
originally (Crocker et al., 2003) while addressing the challenge of items measuring the same
variable, using both structural equation models and network models. The primary goal is to
explore connections between domains of the CSWS; it is plausible to conceptualize the con-
struct of self-worth as a network and consider that its various domains interact and influence
each other instead of being separate consequences with the same origin. Second, I want to
estimate the expected influence (EI) of domains in the network, which can be thought of as
the importance of a domain in the network. EI is calculated as the sum of all connections
of a domain (Robinaugh et al., 2016). Finally, I want to estimate domain predictability
(Haslbeck and Fried, 2017), which reflects the percentage of shared variance of a domain
with surrounding domains in the network. Although I expected EI and predictability to be
related (i.e. domains high on either are likely high on the other), EI provides a measure of
83
the relative importance of a construct, whereas predictability provides insights into absolute
value (Haslbeck and Fried, 2017).
5.2 Method
5.2.1 Participants
This study is based on a data set composed of 680 French-speaking university students: 59%
of them were women and 41% men. The subjects were 17 to 25 years old (M = 19 years, SD
= 1.5 years).
5.2.2 Measurement
The CSWS is composed of 35 items meant to assess self-worth contingency in the following
seven domains: family support, competition, appearance, God’s love, academic competence,
virtue and other’s approval. The items are shuffled in the questionnaire. Item score ranges
from 1 (strongly disagree) to 7 (strongly agree); some reverse-scored items are included
(items 4, 6, 10, 13, 15, 23 and 30).
The data set was anonymized, and its analysis was approved by the Ethical Committee
of the Erasme university hospital.
Table 5.1: The Contingencies of Self-Worth Scale by
Crocker et al. (Crocker et al., 2003).
Item Domain
color
Item meaning Domain
1 Dark
yellow
When I think I look attractive, I feel good
about myself.
Appearance
84
2 Light yel-
low
My self-worth is based on God’s love. God’s love
3 Orange I feel worthwhile when I perform better than
others on a task or skill.
Competition
4 Dark
yellow
My self-esteem is unrelated to how I feel
about the way my body looks. (Reversed)
Appearance
5 Blue Doing something I know is wrong makes me
lose my self-respect.
Virtue
6 Dark blue I don’t care if other people have a negative
opinion about me. (Reversed)
Other’s approval
7 Red Knowing that my family members love me
makes me feel good about myself.
Family support
8 Light yel-
low
I feel worthwhile when I have God’s love. God’s love
9 Dark blue I can’t respect myself if others don’t respect
me.
Other’s approval
10 Red My self-worth is not influenced by the quality
of my relationships with my family members.
(Reversed)
Family support
11 Blue Whenever I follow my moral principles, my
sense of self-respect gets a boost.
Virtue
12 Orange Knowing that I am better than others on a
task raises my self-esteem.
Competition
13 Light blue My opinion about myself isn’t tied to how
well I do in school. (Reversed)
Academic competence
85
14 Blue I couldn’t respect myself if I didn’t live up to
a moral code.
Virtue
15 Dark blue I don’t care what other people think of me.
(Reversed)
Other’s approval
16 Red When my family members are proud of me,
my sense of self-worth increases.
Family support
17 Dark
yellow
My self-esteem is influenced by how attrac-
tive I think my face or facial features are.
Appearance
18 Light yel-
low
My self-esteem would suffer if I didn’t have
God’s love.
God’s love
19 Light blue Doing well in school gives me a sense of self-
respect.
Academic competence
20 Orange Doing better than others gives me a sense of
self- respect.
Competition
21 Dark
yellow
My sense of self-worth suffers whenever I
think I don’t look good.
Appearance
22 Light blue I feel better about myself when I know I’m
doing well academically.
Academic competence
23 Dark blue What others think of me has no effect on
what I think about myself. (Reversed)
Other’s approval
24 Red When I don’t feel loved by my family, my
self- esteem goes down.
Family support
25 Orange My self-worth is affected by how well I do
when I am competing with others.
Competition
26 Light yel-
low
My self-esteem goes up when I feel that God
loves me.
God’s love
86
27 Light blue My self-esteem is influenced by my academic
performance.
Academic competence
28 Blue My self-esteem would suffer if I did some-
thing unethical.
Virtue
29 Red It is important to my self-respect that I have
a family that cares about me.
Family support
30 Dark
yellow
My self-esteem does not depend on whether
or not I feel attractive. (Reversed)
Appearance
31 Light yel-
low
When I think that I’m disobeying God, I feel
bad about myself.
God’s love
32 Orange My self-worth is influenced by how well I do
on competitive tasks.
Competition
33 Light blue I feel bad about myself whenever my aca-
demic performance is lacking.
Academic competence
34 Blue My self-esteem depends on whether or not I
follow my moral/ethical principles.
Virtue
35 Dark blue My self-esteem depends on the opinions oth-
ers hold of me.
Other’s approval
5.2.3 Network analysis
Data were analyzed with R software (open source, available at https://www.r-project.
org/). Packages used to carry out the analysis include qgraph (Epskamp et al., 2012), and
glasso (Friedman et al., 2014b) for network estimation and visualization, mgm for node
predictability (Haslbeck and Waldorp, 2016), igraph (Csardi and Nepusz, 2006) and bootnet
(Epskamp and Fried, 2018) for stability.
87
Sum score vs factor analysis
Items from CSWS subdomains tend to measure the same construct, which is a situation
which a network of all individual items can be problematic because different nodes measure
the same underlying psychological construct (Fried and Cramer, 2017). Therefore, I chose
to estimate a network of 7 domains instead of a network of 35 items. The preferred way for
doing so is using generalized network psychometrics framework (Epskamp et al., 2017b) via
the R-package lvnet. Unfortunately, the method does not currently scale well, and was not
applicable to the current datasets due to the large number of items.
Instead, I studied the network structure of self-worth domains with nodes reflecting sum
scores of the 7 CSWS domains, and used these factor scores as variables in the Gaussian
Graphical Model (GGM), a regularized partial correlation network (Epskamp and Fried,
2018). As an additional sensitivity analysis, I also estimated a factor model using confir-
matory factor analysis for each of the 7 CSWS domains, and then used these factor scores
in a GGM. I expected somewhat stronger relations, because factor scores are disattenuated
for measurement unreliability and therefore likely increase the relations among variables
(Spearman, 1904).
Network estimation
A network structure is composed of nodes and edges: nodes represent, in this case, domains
from the CSWS, and edges are connections between two domains. A regularized partial
correlation network was estimated on the correlation matrix of the 7 domains; as described
above, sumscores for each participant were used. Edge weight parameters that resulted from
the GGM were regularized by using the graphical lasso (least absolute shrinkage and selec-
tion operator): this procedure avoids the estimation of spurious edges (Tibshirani, 2011).
The estimation procedure selected the network (out of 100 networks) with the lowest lambda
value (lambda being the tuning parameter for this procedure); in these situations, it is rec-
ommended to lower the tuning parameter to 0.001, and I followed this recommendation. For
88
the GGM, an edge represents the regularized partial correlation (or conditional dependence
relation) between two domains, controlling for all other domains. If two nodes are connected,
this means they are conditionally dependent, given all other nodes in the network. When
visualizing the model output as graph, blue edges indicate positive relations, and red edges
negative relations. The corresponding thickness of an edge represents its weight (i.e. the
strength of association between nodes, ranging from -1 to 1). The Fruchterman-Reingold
algorithm was used to place nodes in a network (Fruchterman and Reingold, 1991).
Network stability
Stability tests are necessary to safely interpret network inference results from a network
analysis. To answer the question “is edge X significantly stronger than edge Y?”, 95%
confidence intervals of the edge weights were estimated through bootstrapping (Epskamp and
Fried, 2018), 2000 bootstraps were used, and the edge weight difference test was performed.
To answer the question “is the EI of node X stronger than the EI of node Y”, I performed
the centrality difference test.
Network inference
To investigate the network structure of self-worth, I computed two different local inference
measures: node predictability and EI. Expected influence is the sum of a node’s connections
and represents the relative importance of a node in a network (Robinaugh et al., 2016) –
relative because even in weakly connected networks (with overall low edge weights), there
will always be a node with a high expected influence in case of standardized results. Node
predictability is an absolute measure of the interconnectedness of a given node in the net-
work and represents its shared variance with surrounding nodes (Fried et al., 2018). Node
predictability can be interpreted as the upper bound of controllability: if one assumes that
all edges for node X are directed towards that node, predictability provides an estimate of
how much influence I can have on X via all other nodes.
89
5.3 Results
5.3.1 Descriptive statistics
Means range from 12.4 (God’s love) to 26.6 (academic competence). Standard deviations
range from 5.2 (appearance) to 9 (God’s love). God’s love has the lowest mean as well as
the highest standard deviation in the network.
Network of self-worth
Figure 5.1 illustrates the estimated the seven-domain network of self-worth. The network
is composed of domains that connect with each other. Each domain is represented with a
different color. Competition and academic competence share the strongest connection in
the network; other’s approval also shares a strong edge with appearance. Competition and
appearance as well as academic competence and family support are also positively connected.
Family support is positively connected with most domains. God’s love is only connected
to virtue. Appearance and virtue share a negative connection. The factor score network
resulted in considerably stronger associations, which can be expected due to disattenuation.
The adjacency matrices were correlated 0.95 between both methods (sum score and factor
score).
5.3.2 Network stability
Edge weight bootstrap reports relatively small CIs, as is expected from a network with
several hundred participants and only 7 nodes; this means that edge weight estimation is
precise. The edge weight difference test reveals that stronger edges in the network are
significantly stronger than the other edges. In other words, stronger edges in Figure 1 can
be interpreted as being considerably stronger than weaker edges. For instance, the edges
between competition and academic competence and competition and appearance represent
the statistically strongest edge coefficients in the network, significantly stronger than all
90
FS
C
A
GL
AC
V
OA
FS: Family SupportC: CompetitionA: AppearanceGL: God's LoveAC: Academic competitionV: VirtueOA: Other's Approval
Figure 5.1: 7-domain CSWS network. Each node represents a domain: FS is “Familysupport”, C is “Competition”, A is “Appearance”, GL is “God’s love”, AC is “Academiccompetence”, V is “Virtue” and OA is “Other’s approval” cluster. Blue edges representpositive connections and red edges represent negative connections; the thicker the connec-tion the stronger it is. The pie chart surrounding the node represents node predictability(percentage of shared variance with all connected nodes).
91
−2
0
2
A AC C FS GL OA V
CSWS_Domain
Exp
ecte
d In
fluen
ce
Figure 5.2: 7-domain CSWS network. Each node represents a domain: FS is “Familysupport”, C is “Competition”, A is “Appearance”, GL is “God’s love”, AC is “Academiccompetence”, V is “Virtue” and OA is “Other’s approval” cluster. Blue edges representpositive connections and red edges represent negative connections; the thicker the connec-tion the stronger it is. The pie chart surrounding the node represents node predictability(percentage of shared variance with all connected nodes).
other edges but not statistically different from one another. EI difference tests show that EI
estimates in nodes with high EI are statistically different from EI estimates in nodes with
low EI. .
5.3.3 Network inference
Expected influence
Figure 5.2 illustrates the EI estimates for the self-worth network. Academic competence and
family support domains have the highest EI values. This means from a statistical point of
view that these are the most connected domains in the network. On the other hand, God’s
92
love has the lowest EI value. This means that it is the domain that least influences the rest
of domains in the network. Correlation between EI and predictability is 0.96.
Node predictability
Mean node predictability ranges from 0.06 to 0.40, with an average of 0.25. This means that
on average, 25% of the variance of the node in the network can be explained by its neighbors.
God’s love is the domain with the lowest node predictability: it shares 6% of variance with
its surrounding nodes. Academic competence has the highest node predictability: it shares
40% of its variance with its surrounding nodes. Competition has the second highest node
predictability (0.35).
5.4 Discussion
To my knowledge, I have conducted the first network analysis of the psychological construct
self-worth contingencies. Overall, the seven domains of self-worth form a heterogeneous
system in which domains are not uniformly positively connected with each other. This is
interesting, because a homogeneous network with uniformly positive connections would be
expected if all domains are passive and interchangeable measures of one latent variable:
self-worth. Below, I discuss the findings in more detail.
Academic competence and competition share the strongest connection in the network:
it is reasonable to consider that the impact on self-esteem of competing with others and
obtaining good grades are connected while following a university curriculum. The same
kind of connection is found between appearance and other’s approval; this means that, while
considering self-worth, if physical appearance is important to an individual, so is the approval
of others, and vice versa. Competition also shares a strong connection with appearance,
which means that physical appearance might be important for individuals competing with
others (and vice-versa). Family support and other’s approval share connections with most
93
domains in the network. Appearance showed a negative connection to virtue: that means
that people that base their self-worth upon acting and living by a moral code might not draw
self-worth from physical appearance (and vice-versa), controlling for all other associations in
the network. While there is no prior work on partial correlations, previous work on zero-order
correlations found a positive association between the two domains (Crocker et al., 2003).
Negative edges have not been observed commonly in the psychopathology network lit-
erature, which calls for an explanation. In this case, the negative association between ap-
pearance and virtue might be plausible from a theoretical perspective. Since both subscales
are positively associated with academic competence and family support, the finding implies
that in individuals whose self-worth is simultaneously contingent on academic competence
and family support, knowing that self-esteem is more contingent on virtue allows predicting
that their self-esteem is less likely to be also contingent on physical appearance (and vice
versa). Two other possibilities also come to mind. First, negative connections in Gaussian
Graphical Models can arise when dealing with small samples and/or when estimating poly-
choric correlations (Epskamp and Fried, 2018), which I can rule out as explanation here.
Second, collider structures in conditional dependence networks can induce spurious negative
relations between two nodes in case they both cause a third node (Greenland et al., 1999).
God’s love is a relatively disconnected node: it shares only one positive connections
with virtue: this is not surprising, since in the original work (Crocker et al., 2003) God’s
love showed its strongest correlation with virtue. From a network perspective, this means
that God’s love is largely conditionally independent from other domains in the network.
From a network perspective, one plausible interpretation of the conditional independence
of God’s Love in the self-esteem network is that people may derive a sense of self-worth
from their religious belief (in this case, feeling that they have the love of God) regardless
of the other contingencies; this may highlight religious belief as an independent source of
self-esteem in people. Another possible interpretation of this finding is statistical, i.e. a floor
effect or a ceiling effect: because of a low or high parameter, the domain might share few
94
connections with other domains. This may be applicable to my findings, since God’s love
has the highest standard deviation among all domains in the network, as well as the lowest
mean. I identified strong differences in predictability, ranging from 6% (God’s love) to 40%
(academic competence). Average node predictability is 0.25, which means that on average,
25% of the variance of the nodes is explained by other nodes in the network. From a network
perspective, I can infer that some domains such as academic competence are well explained
by its surrounding domains. Academic competence and God’s love are respectively the most
and least predictable nodes in the network.
The analysis of EI shows that academic competence, and family support have the highest
values in the network: this means that these two domains share the strongest connections
in the network and therefore may influence or be influenced by other domains of contingent
self-worth the most. Node predictability is therefore simpler to interpret than EI and gives
us a clear information about how a node is influenced by surrounding nodes, assuming all
edges are directed towards this node.
This study should be interpreted in the light of some limitations. First, my network is
estimated from a sample of university students. While the CSWS was originally developed
based on a similar sample (Crocker et al., 2003), it is worth noting that results of my study
may not generalize to other samples. Second, the current cross-sectional data set does not
allow for causal or even Granger-causal inference. For instance, I cannot interpret whether
a given domain causes or is caused by domains sharing a connection with it. This requires
temporal follow-up studies, which would be most interesting across important developmental
periods such as adolescents and early adulthood. Third, the network model I estimated is
a between-subjects models. This also means that inferences from the study should only be
drawn for a group of people, and it is unclear if and how well the present between-subjects
network structure describes individuals’ networks of self-worth contingency.
Future research may endeavor to apply self-worth contingency networks in other kinds
of samples, both healthy and with different kinds of psychopathology, as to analyze possible
95
Chapter 6
A network model of resilience
Abstract
The Resilience Scale for Adults (RSA) is a questionnaire that measures protective
factors of mental health. The aim of this paper is to perform a network analysis of the
Resilience Scale for Adults (RSA) in a dataset composed of 675 French-speaking Belgian
university students, to identify potential targets for intervention to improve protective
factors in individuals. I estimated a network structure for the 33-item questionnaire
and for the six domains of resilience: perception of self, planned future, social com-
petence, structured style, family cohesion and social competence. Node predictability
(shared variance with surrounding nodes in the network) was used to assess the con-
nectivity of items. An Exploratory Graph Analysis (EGA) was performed to detect
communities in the network: the number of communities detected being different than
the original number of factors proposed in the scale, I estimated a new network with
the resulting structure and verified the validity of the new construct which was pro-
posed. The network composed of items from the RSA is overall positively connected
with strongest connections arising among items from the same domain. The domain
network reports several connections, both positive and negative. The EGA reported
the existence of four communities that I propose as an additional network structure.
Node predictability estimates show that connectedness varies among the items and
domains of the RSA. Network analysis is a useful tool to explore resilience and identify
97
targets for clinical intervention. In this study, the four domains acting as components
of the additional four-domain network structure may be potential targets to improve
an individual’s resilience. Further studies may endeavor to replicate my findings in
different samples.
6.1 Introduction
Resilience is understood as a positive adaptation despite significant adversities or trauma
(Luthar, 2006). Resilience is a psychological construct which has been proven to be related
to psychiatric disorders, such as anxiety, depression, substance abuse obsessive-compulsive
disorder (Hjemdal et al., 2011a; Bonfiglio et al., 2016).
In recent years, the construct of resilience has been conceived as an outcome rather than a
trait, which highlights the ability to improve an individual’s protective factors against mental
illness (Chmitorz et al., 2018). In this framework, protective factors composing resilience
compete with risk factors, for instance, adverse events (such as traumatic experiences, loss
or neglect) which have been shown to be present in up to 50% of individuals under the age
of 18 (Fritz et al., 2018). Other important factors influencing the framework of resilience
involve age, social status and education (Aburn et al., 2016).
The Resilience Scale for Adults (RSA) is a psychometric questionnaire that assesses
protective factors of mental health (Friborg et al., 2003). The RSA has been defined as one of
the best resilience questionnaires with regard to psychometric ratings (Windle et al., 2011).
Largely validated in Norwegian samples, the construct has undergone in the last decade
cross-cultural validation in different countries, such as Belgium (Hjemdal et al., 2011b), Iran
(Jowkar et al., 2010), Italy (Bonfiglio et al., 2016) and Peru (Morote et al., 2017). The
RSA measures six domains of resilience: (1) perception of self represents the confidence in
oneself, one’s own capabilities, judgment and decision-making (e.g. item 17 “My judgment
and decisions I trust completely”); (2) planned future identifies goal-oriented individuals (e.g.
item 32 “My goals for the future are well thought through”); (3) social competence represents
98
the ability to adapt in social environments (e.g. item 21 “Meeting new people is something I
am good at”); (4) structured style identifies with organized individuals who follow routines
(e.g item 23 “When I start on new things/projects, I prefer to have a plan”); (5) family
cohesion measures the loyalty, support, optimism, mutual understanding and appreciation
among family members (e.g. item 3 “My family understanding of what is important in life
is very similar”); (6) social resources identifies the availability of social support from friends
and family (e.g. item 6 “I can discuss personal issues with friends/family members”). These
six domains are commonly understood as being effects of the construct of resilience itself,
since they are measurable indicators of the construct.
However, in recent years, network theory has emerged as a way of studying psychological
constructs as interacting entities (Borsboom, 2017). Such entities are uncovered in real-world
data using network models, usually composed of pairwise interactions of its elements, and
the constructs emerge from these connections (Borsboom and Cramer, 2013). Interactions
between elements composing a network are often statistically represented as regularized
partial correlations (Epskamp and Fried, 2018). Several mental disorders have been analyzed
using a network perspective, such as posttraumatic stress disorder (Fried et al., 2018; Phillips
et al., 2018), depression (Mullarkey et al., 2018), schizophrenia (Galderisi et al., 2018) and
obsessive-compulsive disorder (Ruzzano et al., 2015). Network analysis has also been applied
to several psychological constructs, such as personality (Costantini et al., 2015), empathy
(Briganti et al., 2018), attitudes (Dalege et al., 2017), intelligence (van der Maas et al., 2006)
and self-worth (Briganti et al., 2019). Other studies used innovative methods, including
networks to harmonize rating scales (Gross et al., 2018; Purgato et al., 2018; Haroz et al.,
2016).
Learning the network structure of a given construct (such as resilience) or mental disor-
der (such as PTSD) is particularly relevant in clinical practice since it highlights potential
clinical target that may affect multiple symptoms or elements composing the network (Fried
et al., 2018); for instance, intervening on the connection between two components of the
99
network is likely to modify the clinical presentations of said components (such as symp-
toms). In the specific case of resilience, which is considered a protection against mental
disorders, learning the network structure of resilience components may highlight potential
targets to strengthen the overall mental health of a given individual. In recent years, several
intervention methods to foster resilience have been studied worldwide, but their efficiency is
variable because of limited comprehension of this relevant psychological construct (Chmitorz
et al., 2018). A network analysis of resilience factors has also been proposed in two sample
of adolescent subjects with and without childhood adversities (Fritz et al., 2018) and showed
that childhood adversities impact the degree of connectivity of resilience factors.
Network components are usually answers of an observed group to items of a questionnaire,
such as the RSA. A current challenge in network models when dealing with self-report
scales is the redundancy of several items of a given questionnaire in measuring the same
aspect of a construct (Fonseca-Pedrero et al., 2018); while addressing the meaning of a
given connection between two items, their interaction will represent shared variance (and
not a pairwise relationship) if they tend to measure the same thing (Fried and Cramer,
2017). In the case of the RSA, this challenge goes beyond the notion of a single items of the
questionnaire and may apply to entire domains of the RSA: for instance, questions from both
perception of self and planned future refer to one’s own dispositional attributes and internal
source of resilience and were original part of the same factor, which was called personal
competence (Friborg et al., 2003). The same line of reasoning applies to family cohesion
and social resources, even though originally distinct factors, since they represent an external
source of resilience – that is, the support that the individual feels both within and without
the family nucleus: furthermore, several items from social resources include the concept of
family support (e.g. item 6 “I can discuss personal issues with friends/family). Exploratory
graph analysis (EGA) has emerged as a highly effective and reliable tool in network analysis
when addressing the issue of recovering the number of factors (CFA) in datasets (Golino and
Epskamp, 2017). An optimum solution proposed in the literature is to first explore the basic
100
dimensionality of an instrument with an EGA then authenticate the suggested structure by
performing a confirmatory factor analysis (Golino and Demetriou, 2017).
I aim to extend the conceptual framework of network analysis to the construct of re-
silience such as represented by the RSA and address the challenge of domain redundancy
using both network models and structural equation models. First, I want to explore the con-
nectivity of the RSA as a network composed of its items, then study the connections arising
among resilience domains, such as performed in recent network papers (Briganti et al., 2019).
Second, I want to apply community detection algorithms and the EGA to the item network,
explore then verify the suggested structure with CFA and network analysis. Third, I want
to measure node predictability which an absolute measure of interconnectedness (Haslbeck
and Fried, 2017) of a node in a network.
Statistically speaking, node predictability represents the shared variance of a network
component with surrounding components. Although performed on university students, ex-
ploring a network structure that shows how domains of resilience interact may have mean-
ingful clinical implications as it highlights potential target to improve the overall protective
factors of a given individual; also, it may serve as basis for future replication studies designed
to identify the network structure of RSA in other samples.
6.2 Method
6.2.1 Participants
The analyses in this paper are performed on a dataset composed of 675 university students
from the French-speaking region of Belgium. 59% of the students were women and 41% were
men; subjects were 17 to 25 years old (M=19 years, SD=1.5 years).
101
6.2.2 Measurement
The RSA is composed of 33 items that measure resilience in 6 domains: perception of self,
planned future, social competence, structured style, family cohesion and social competence.
The items are shuffled in the questionnaire. Item scoring is semantic and differential-based
(Friborg et al., 2006a): for instance, when scoring item 13 “My family is characterized by”,
a minimum score of 1 is represented by the answer “Disconnection” and a score of 7 is
represented by the answer “Healthy cohesion”. Reversed-score items are included in the
scale.
This study was approved by the Ethical Committee of the Erasme university hospital.
6.2.3 Network analysis
I used the software R (open source, available at https://www.r-project.org/). Pack-
ages and functions to carry out the analysis include qgraph (Epskamp et al., 2012), glasso
(Friedman et al., 2014b) for network estimation and visualization, mgm (Haslbeck and Wal-
dorp, 2016) for node predictability, EGA (Golino and Epskamp, 2017) and igraph (Csardi
and Nepusz, 2006) for community detection, and bootnet (Epskamp and Fried, 2018) for
stability.
Item network
I calculated correlations for the 33 RSA items and used the resulting correlation matrix as
an input to estimate a Gaussian Graphical Model (GGM), a regularized partial correlation
network (Epskamp and Fried, 2018). Graphical lasso (least absolute shrinkage and selection
operator) was used to regularize the parameters resulting from the GGM, therefore avoiding
the estimation of spurious connections (non-existing connections that may be present due
to noisy information). In the item network, nodes represent resilience items from the RSA
questionnaire. Each node is surrounded by a pie chart representing node predictability.
Connections between nodes are called edges. An edge in a network is interpreted as the
102
existence of an association between two nodes, controlling for all other nodes in the network.
An edge between two items of the RSA may be statistically interpreted as following: if
two given nodes X and Y share an edge XY in the network, and the observed group of
subjects scores high on X, then the observed group is also more likely to score high on Y
(Briganti et al., 2018). Each edge in the network represents either positive regularized partial
correlations (visualized as blue edges) or negative regularized partial correlations (visualized
as red edges). The thickness and color saturation of an edge denotes its weight (the strength
of the association between two nodes). The Fruchterman-Reingold algorithm places the
items in the network based on the sum of connections of a given node with other nodes
(Fruchterman and Reingold, 1991).
Six-domain network
To assess the overall connectedness of the domains of resilience as described in the RSA
(Hjemdal et al., 2011a) I used the methodology described in recent papers (Briganti et al.,
2019) and estimated a factor model using CFA for each of the six RSA domains. I then used
the factor scores obtained to estimate an additional GGM.
Network stability
Network stability is composed of several state-of-the-art analyses which are necessary to
safely interpret results from a network analysis. I estimated 95% confidence intervals (CI)
of the edge weight through bootstrapping (Epskamp and Fried, 2018), 2000 bootstraps were
used and performed an edge weight difference test to answer the question “is edge A signifi-
cantly stronger than edge B?”.
Network inference
I estimated node predictability for the 33 RSA items and for the six domains. Node pre-
dictability (Haslbeck and Fried, 2017) represents shared variance of a given node with sur-
103
rounding nodes in a network. Node predictability is an absolute measure of the interconnect-
edness of network nodes (Fried et al., 2018). Other measures of inference frequently used in
network literature such as strength centrality (Boccaletti et al., 2006) or expected influence
(Robinaugh et al., 2016) can only address the relative importance of nodes (Briganti et al.,
2019) and are therefore less informative when it comes to address the issue of interconnect-
edness; that is why I decided not to use these measures in this paper. One interpretation
of node predictability that has been previously described in the literature (Briganti et al.,
2019) is that of the upper bound of controllability: this measure can provide an estimate
of how much a node X can be influenced by all other nodes if I assume that all edges that
node X shares with other nodes are directed towards X. To explore the dimensionality of
the RSA in my sample I performed an EGA on the item network. EGA uses the walktrap
algorithm to detect communities. This algorithm is based on the principle that adjacent
nodes tend to belong to the same community (Yang et al., 2016), was shown to have high
accuracy in simulation studies (Golino and Epskamp, 2017) and used in empirical network
papers (Briganti et al., 2018).
Four-domain network
Because I detected a different structure – composed of four domains instead of six as the
one originally proposed (Hjemdal et al., 2011a), I used CFA to estimate a four-factor model
and used the resulting factor scores to estimate a four-domain network; because the network
estimation procedure selected the network (out of a 100 networks) with the lowest tuning
parameter (called lambda value), I lowered the lambda value to follow standard recommen-
dations. I also calculated node predictability for the four nodes of the resulting network.
104
F1 F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14 F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25F26
F27
F28
F29
F30F31
F32 F33
Figure 6.1: 33-item RSA network. Each node represents an item from the questionnaire.Blue edges represent positive connections and red edges represent negative connections; thethicker the connection the stronger it is. The pie chart surrounding the node represents nodepredictability (percentage of shared variance with all connected nodes).
105
6.3 Results
6.3.1 Item network
Figure 6.1 represents the item network: to render the visualization more readable, I hid
all edges smaller than 0.05 (one tenth of the value of the maximum edge weight). Overall,
the 33 items from the RSA form a network of positively connected nodes. The strongest
connection (0.5) is the edge between node 18 (“New friendships are something I make easily”)
and node 21 (“Meeting new people is something I am good at”), both belonging to social
competence. Two other examples of strong connections are edge 9-15 belonging to social
resources (“Those who are good at encouraging me are some close friends/family; “I get
support from friends/family members”) and edge 4-32 belonging to planned future (“I feel
that my future looks very promising”; “My goals for the future are very thought through”).
These examples of highly connected nodes reflect the challenge of items representing the same
aspect of a construct and are discussed in section 4. However, several edges connect different
domains of resilience. For instance, edge 11-19 connects perception of self and social resources
(“My personal problems I know how to solve”; “When needed, I have always someone who
can help me”), and edge 17-26 connects perception of self and social competence (“My
judgment and decisions I trust completely”; “For me, thinking of good topics of conversation
is easy”).
6.3.2 Six-domain network
Figure 6.2 illustrates the six-domain network of resilience. This network reports considerably
stronger connections because of disattenuation due to measurement unreliability. This issue
is to be expected when dealing with a GGM based on correlations between factor scores and
has been previously described in the literature (Briganti et al., 2019).
The strongest connections are found between social resources and family cohesion (0.59),
and between planned future and perception of self, (0.58). Two negative connections are
106
Future
SelfFamily
SocRes
SocComp
Style
Future: Planned FutureSelf: Perception of SelfFamily: Family CoherenceSocRes: Social ResourcesSocComp: Social CompetenceStyle: Structured Style
Figure 6.2: Six-domain RSA network. Each node represents a domain from the RSA.
found between structured style and perception of self (0.19) and between family cohesion
and social competence (0.21). Several domains present no direct connection with each other,
such as structured style and social competence or social resources and perception of self;
from a network perspective, that means that the two domains are conditionally independent
from each other.
I performed a CFA to assess the validity of the six-domain structure. Root Mean Square
Error of Approximation (RMSEA) is 0.047 (cut-off for good fit < 0.06) and the Standardized
Root Mean Square Residual (SRMSR) is 0.058 (cut-off for good fit <0.08); Cronbach’s alpha
is 0.64 (>0.8 for good fit); Comparative Fit Index (CFI) is 0.87 (>0.9 for good fit) and the
p-value for the chi-square fit test is 0 (> 0.05 for good fit; Schreiber, 2017).
6.3.3 Network stability
Bootstrapped 95% edge weight CI show that the edge weights are accurately estimated, and
the edge weight difference tests report that stronger edges can be safely interpreted as to
be stronger that other edges in both the item network and the six-domain network, but do
not statistically differ from each other in the six-domain network. For instance, one cannot
safely interpret the edge between social resources and family cohesion to be stronger than
107
the edge between perception of self and planned future.
6.3.4 Network inference
Node predictability
Node predictability was estimated in both the item network and the six-domain network.
In the item network, the two nodes with the highest node predictability are node 18 (“New
friendships are something I make easily”; 0.54) and 21 (“Meeting new people is something
I am good at”; 0.53), which both belong to social competence and also share the strongest
edge in the network. The node with the lowest node predictability is node 33 which belongs
to perception of self (“Events in my life that I cannot influence I manage to come to terms
with”; 0.13). Mean node predictability is 0.32, which means that on average, items from
the RSA share 32% of their variance with surrounding nodes. In the six-domain network,
planned future shows the highest node predictability (0.67) and structured style is the least
predictable node (0.36). The mean node predictability is 0.55, which means that on average,
domains present 55% of shared variance.
Community detection
The EGA and walktrap algorithm applied to the item network report four communities of
items instead of six as proposed in the original scale. To visualize the differences between
communities, I first reestimate the network with a different color palette, each color indicating
a community of items as detected by the algorithm as shown in Figure 6.3.
Overall, items from perception of self and planned future form a new community, that
I identify as personal competence, referring to one of the first versions of the RSA (Friborg
et al., 2006b); the same process applies to items from social resources and family cohesion,
forming a new community that I identify as support since it is an aspect of resilience that
the two domains represent. Items 10 (“The bonds among my friends are strong”) and 19
(“When needed, I have always someone who can help me”) switch communities, going from
108
F1 F2
F3
F4
F5
F6
F7
F8
F9
F10
F11
F12
F13
F14 F15
F16
F17
F18
F19
F20
F21
F22
F23
F24
F25F26
F27
F28
F29
F30F31
F32 F33
Figure 6.3: 33-item RSA network with communities assigned by the Exploratory GraphAnalyses
social resources to social competence.
6.3.5 Four-domain network
As suggested in the literature (Golino and Demetriou, 2017), I performed a CFA to assess
the validity of the proposed structure. RMSEA is 0.064 (cut-off for good fit < 0.06) and
the SRMSR is 0.074 (cut-off for good fit <0.08); Cronbach’s alpha is 0.64 (>0.8 for good
fit); CFI is 0.74 (>0.9 for good fit) and the p-value for the chi-square fit test is 0 (>0.05 for
good fit). Figure 6.4 represents the four-domain network. Stability analyses carried out in
this network show that stronger edges are significantly stronger than other edges. Personal
competence is the most interconnected node, which is represented with the strongest positive
connections with the three other domains (0.5 with social competence, 0.37 with structured
style and 0.32 with support), and with a node predictability of 0.54. The node with the
lowest node predictability is structured style (0.22). The mean node predictability for the
109
PersComp
Support
SocComp
Style
PersComp: Personal CompetenceSupport: SupportSocComp: Social CompetenceStyle: Structured Style
Figure 6.4: Four-domain network of resilience. Each node represents a domain as detectedby the Exploratory Graph Analysis.
four-domain network is 0.37. A negative edge is found between structured style and social
competence.
6.4 Discussion
This paper is to my knowledge the first work to report a network analysis of the psychological
construct of resilience as conceived in RSA. The different analyses carried out bring new
and interesting information on the construct, reporting overall that resilience is formed of
interacting components which are not mere consequences of a latent variable. If the network
structures presented in this work were to replicate in different samples, interventions to
improve protective factors in individuals may become more efficient by acting on meaningful
targets, such as two highly connected nodes in the resilience network.
The item network shows that the strongest edges are shared between items representing
overall the same aspect of a domain: such connections must therefore be interpreted as shared
variance between items, such as reported in recent papers that further address the issue (Fried
and Cramer, 2017) and propose solutions such as estimating a network of domains instead of
a network of items (Briganti et al., 2019). However, in the case of the RSA, the item network
110
sheds light on the connectivity between items from different subscales: items from the RSA
in my sample therefore form a complex system of mutual interactions that actively contribute
to the construct of resilience itself. From a network perspective, this means the observed
group is likely to similarly answer items that present a connection in the resilience network,
after controlling for all other items in the network (Briganti et al., 2018). Items from the RSA
also have different levels of importance in the network; this information is provided by node
predictability, which is an absolute measure of the interconnectedness of a node (Haslbeck
and Fried, 2017). In the item network, two nodes from the social competence domain (18
and 21) show the highest predictability, sharing over 50% of variance with surrounding nodes
in the network structure: however, as addressed in the section 3, the two nodes with the
highest node predictability are also the nodes sharing the strongest edge in the network; the
high predictability is therefore largely influenced by the presence of one very strong edge,
which is also a known feature influencing centrality criteria.
The six-domain network further helps us explore the connectivity and importance of the
protective factors as described in the most recent version of the RSA. Domains of the RSA
form a heterogenous system with positive and negative connection: this further supports the
theory that domains of resilience are not interchangeable measures of resilience; the construct
arises from the connections among domains. For instance, two negative connection exists, the
first between structured style and perception of self, and the second between family cohesion
and social competence. Negative edges are a rare finding when dealing with a network
approach of psychological constructs; a recent paper (Briganti et al., 2019) addressed the
issue of interpreting negative edges in the case of a domain network such as the six-domain
network of RSA estimated in this paper. From a theoretical point of view, I may interpret
the negative edge between social competence and family cohesion as follows: knowing that
an individual’s resilience is strongly drawn from the ability to rapidly adapt in different social
context, that individual’s resilience is less likely to be drawn from the support originating
from a cohesive family (and vice versa). The same line of reasoning applies to the negative
111
connection between structured style and perception of self: knowing that an individual’s
resilience is drawn from routines and structure, his/her resilience is less likely drawn from
confidence in own capabilities/decisions (and vice versa).
In the six-domain network, the strongest connections are found between social resources
and family cohesion, and between planned future and perception of self. As mentioned in
section 1, these two couples of domains theoretically overlap, with several items measuring
the same source of resilience; it is therefore not surprising that these domains are highly
connected in a network structure. Domains of resilience predict each other well, with mean
node predictability indicating 55% of shared variance on average. Planned future is the most
important node in the resilience network according to the node predictability estimates (it
has 67% of shared variance with surrounding nodes).
The EGA reported the existence of four communities in the item network, with a first
new community, personal competence, emerging from perception of self and planned future,
and a second new community, support, emerging from social resources and family cohesion.
Personal competence (adding up items from perception of self and planned future) is, from
a psychometric point of view, not a surprising finding: the two communities composing the
new domain were originally a single factor (Friborg et al., 2003). However, social resources
and family cohesion were originally proposed as distinct factors since the first published
version of the scale, which makes this analysis an interesting finding.
In the four-domain network, the new personal competence community is the most in-
terconnected node, sharing the three strongest connections in the network and 54% of its
variance with the three surrounding domains. A negative edge is found between structured
style and social competence, two domains unconnected in the six-domain network: from a
theoretical point of view, it is plausible to consider that in people whose resilience depends on
a structured life based on routines, being able to adapt in social situations is a less important
source of resilience, controlling for all other sources (and vice versa). On average, nodes in
the four-domain network are less predictable then domains in the six-domain network, with
112
37% of shared variance.
However, when comparing results from the CFA of both the six-factor model and the
four-factor model such as suggested in the literature (Golino and Demetriou, 2017), the
six-factor model presents with better indicators than the four-factor model. This being the
first network approach to this particular scale of resilience, future papers may endeavor to
replicate these findings in other samples while comparing the original six-factor structure
with structures proposed from EGA.
My analyses should be interpreted in the light of several limitations. First, my dataset
is composed of university students, which may likely limit the generalization of my findings
to different samples. Second, because this is a cross-sectional study, I cannot infer whether
a given node (item or domain) causes or is caused by another node to which it is connected;
determination of causality requires time-series which may be interesting in follow-up studies
of, for instance, young individuals with and without childhood adversities (Fritz et al., 2018).
113
Chapter 7
A network model of narcissism
Abstract
The aim of this work is to explore the Narcissistic Personality Inventory (NPI) using
network analysis in a dataset of 942 university students from the French-speaking part
of Belgium. I estimated an Ising Model for the forty items in the questionnaire and
explored item interconnectedness with strength centrality. The NPI is presented as
an overall positively connected network with items from entitlement, authority and
superiority reporting the highest centrality estimates. Network analysis highlights new
properties of items from the NPI. Future studies should endeavor to replicate my
findings in other samples, both clinical and non-clinical.
7.1 Introduction
Narcissism has been defined as the ability to maintain a positive self-image despite various
internal and external processes. Narcissistic subjects have a need for self-enhancing experi-
ences from their social environment (Pincus et al., 2009). Narcissism has been theorized to
possess both normal and pathological aspects, which have been considered by some authors
as two different personality constructs (Von Kanel et al., 2017) and as a continuum by oth-
ers (Paulhus, 1998). Grandiosity and vulnerability are considered as the two expressions of
114
narcissism (Cain et al., 2008): grandiose narcissism is associated with the predisposition to
exploit others, a lack of empathy and one’s feelings of entitlement and superiority, whether
vulnerable narcissism is associated with a depleted self-image, social withdrawal and suici-
dality (Miller et al., 2013). The current gold-standard models of narcissism, the trifurcated
model (Miller et al., 2016) and the narcissism spectrum model (Krizan and Herlache, 2018)
postulate that grandiosity and vulnerability are two largely independent factors that are tied
together by a core of entitlement.
The main tool used to study the construct of narcissism is the Narcissistic Personality
Inventory or NPI (Raskin and Hall, 1979), which represents grandiose narcissism (Krizan
and Herlache, 2018). The NPI consists of forty dichotomous items composed of both a
narcissistic and a non-narcissistic statement. The authors of the questionnaire propose seven
domains of narcissism: authority reflects one’s need for authority and success (e.g., item 33
“I would prefer to be a leader”); exhibitionism represents one’s need to be the center of
attention in a social context (e.g., item 30 “I like to be the center of attention”); superiority
measures one’s belief of being better than other people (e.g., item 40 “I am an extraordinary
person”); entitlement reflects one’s desire to receive respect and wield power (e.g., items 14
“I insist upon getting the respect that is due me” and 27 “I have a strong will to power”);
exploitativeness represents one’s capacity to manipulate other people (e.g., item 13 “I find it
easy to manipulate people”); self-sufficiency measures one’s autonomy and belief in oneself
(e.g., items 22 “I rarely depend on anyone else to get things done” and 34 “I am going to
be a great person”); vanity measures one’s admiration of one’s own physical appearance
(e.g., item 19 “I like to look at my body”). However, this seven-domain structure of the
NPI is controversial; several studies report different structures of the questionnaire, such as
a four-factor model (Emmons, 1987) and a three-factor model (Boldero et al., 2015).
Despite inconsistent results in the exploration of dimensionality (Ackerman et al., 2011;
Corry et al., 2008; Kubarych et al., 2004), narcissism is commonly understood as composed
of domains that are interchangeable measures of the construct proposed. In the last decade,
115
a new way of analyzing psychological constructs as complex systems has been proposed:
the network approach (Borsboom, 2017). Such complex systems are uncovered in empirical
studies with network models, that represent a given construct as emerging from mutual
interactions of its components (Borsboom and Cramer, 2013).
The network approach has been used to analyze a number of mental disorders, such as
depression (Mullarkey et al., 2018), posttraumatic stress disorder (Fried et al., 2018; Phillips
et al., 2018). Psychological constructs such as personality (Costantini et al., 2015), empathy
(Briganti et al., 2018) and self-worth (Briganti et al., 2019) have also been proposed as
network structures. The Pathological Narcissism Inventory has been recently investigated
through the lens of network analysis (Di Pierro et al., 2019), which identified Contingent self-
esteem, Grandiose Fantasies and Entitlement Rage to be important traits of the constructs.
A network can be composed of items of a questionnaire such as the NPI. In the case of a
network of self-reported questions, several items tend to be redundant and represent the
same aspect of a construct; this has been described in the network literature as a delicate
challenge, since the meaning of a connection between two redundant elements changes and
simply represent shared variance between the two corresponding questions that measure the
same thing (Fried and Cramer, 2017).
This challenge applies to the NPI: for instance, items 19 (“I like to look at my body”)
and 29 (“I like to look at myself in the mirror”) are two very similar measurements from
vanity. This is the case for several other items in the questionnaire, including items from
different domains, such as items 12 (“I like to have authority over other people”) and 27
(“I have a strong will to power”) that respectively belong to authority and entitlement. In
a network structure, I would expect these items to be strongly associated. It is plausible
to consider narcissism as a network of components (in this case, items from a self-reported
questionnaire indicating an individual’s perspective on narcissistic traits) that mutually in-
fluence each other instead of being passive consequences of the same construct. The network
approach to narcissism is relevant because it might allow in clinical samples the identification
116
of meaningful targets for intervention, even more so if considered that normal and patholog-
ical narcissism form a continuum. The aim of this work is to explore for the first time NPI
items and their relationship in a network of narcissism, therefore applying network analysis
to the items of the questionnaire. Network analysis has been shown to offer substantial in-
sight as a complementary tool to factor analysis, which is a more established technique in the
field of personality assessment (Briganti et al., 2018): as mentioned, modeling a construct or
mental disorder as a network can highlight connections between items or symptoms which
can therefore be used for intervention (Blanken et al., 2019). First, I want to explore the
connectivity of the NPI network. Second, I want to explore the importance of each item in
the questionnaire using strength centrality, which is the absolute sum of connections of a
given node in the network (Boccaletti et al., 2006).
7.2 Method
7.2.1 Participants
The data set used for this study is composed of 942 university students from the French-
speaking region of Belgium. The participants were first-year students in several Belgian
universities and in different undergraduate courses and they volunteered to fill a set of
questionnaires which included a French version of the NPI.
7.2.2 Measurement
The NPI (Table 7.1) contains 40 items that are meant to assess seven domains of narcis-
sism: authority, exhibitionism, superiority, entitlement, exploitativeness, self-sufficiency and
vanity. Items from different domains are shuffled in the questionnaire, and their scoring is
dichotomous: each item possesses both a narcissistic and a non-narcissistic statement. The
protocol of this study was approved by the Ethical Committee of the Erasme university
hospital.
117
Table 7.1: The Narcissistic Personality Inventory (Raskin
and Hall, 1979).
N 0 1 Domain Label
1 I am not good at influ-
encing people
I have a natural talent for
influencing people
Authority A1
2 I am essentially a modest
person
Modesty doesn’t become
me
Exhibitionism Exh2
3 I tend to be a fairly cau-
tious person
I would do almost any-
thing on a dare
Exhibitionism Exh3
4 When people compli-
ment me I sometimes get
embarrassed
I know that I am good
because everybody keeps
telling me so
Superiority S4
5 The thought of ruling the
world frightens the hell
out of me
If I ruled the world it
would be a better place
Entitlement En5
6 I try to accept the conse-
quences of my behavior
I can usually talk my way
out of anything
Exploitativeness Exp6
7 I prefer to blend in with
the crowd
I like to be the center of
attention
Exhibitionism Exh7
8 I am not too concerned
about success
I will be a success Authority A8
9 I am no better or worse
than most people
I think I am a special per-
son
Superiority S9
10 I am not sure if I would
make a good leader
I see myself as a good
leader
Authority A10
118
11 I wish I were more as-
sertive
I am assertive Authority A11
12 I don’t mind following or-
ders
I like to have authority
over other people
Authority A12
13 I don’t like it when I
find myself manipulating
other people
I find it easy to manipu-
late people
Exploitativeness Exp13
14 I usually get the respect
that I deserve
I insist upon getting the
respect that is due me
Entitlement En14
15 I don’t particularly like
to show off my body
I like to show off my body Vanity V15
16 People are sometimes
hard to understand
I can read people like a
book
Exploitativeness Exp16
17 If I feel competent, I am
willing to take respon-
sibility for making deci-
sions
I like to take responsibil-
ity for making decisions
Self-sufficiency SS17
18 I just want to be reason-
ably happy
I want to amount to
something in the eyes of
the world
Entitlement En18
19 My body is nothing spe-
cial
I like to look at my body Vanity V19
20 I try not to be a show off I will usually show off if I
get the chance
Exhibitionism Exh20
21 Sometimes I am not sure
of what I am doing
I always know what I am
doing
Self-sufficiency SS21
119
22 I sometimes depend on
people to get things done
I rarely depend on any-
one else to get things
done
Self-sufficiency SS22
23 Sometimes I tell good
stories
Everybody likes to hear
my stories
Exploitativeness Exp23
24 I like to do things for
other people
I expect a great deal from
other people
Entitlement En24
25 I take my satisfactions as
they come
I will never be satisfied
until I get all that I de-
serve
Entitlement En25
26 Compliments embarrass
me
I like to be complimented Superiority S26
27 Power for its own sake
doesn’t interest me
I have a strong will to
power
Entitlement En27
28 I don’t care about new
fads and fashions
I like to start new fads
and fashion
Exhibitionism Exh28
29 I am not particularly in-
terested in looking at my-
self
I like to look myself in the
mirror
Vanity V29
30 It makes me uncomfort-
able to be the center of
attention
I really like to be the cen-
ter of attention
Exhibitionism Exh30
31 People can’t always live
their lives in term of
what they want.
I can live my life in any
way I want to
Self-sufficiency SS31
120
32 Being an authority
doesn’t mean that much
to me
People always seem to
recognize my authority
Authority A32
33 It makes little difference
to me whether I am a
leader or not
I would prefer to be a
leader
Authority A33
34 I hope I am going to be
successful
I am going to be a great
person
Self-sufficiency SS34
35 People sometimes believe
what I tell them
I can make anybody be-
lieve anything I want
them to
Exploitativeness Exp35
36 Leadership is a quality
that takes a long time to
develop
I am a born leader Authority A36
37 I don’t like people to pry
into my life for any rea-
son
I wish somebody would
someday write my biog-
raphy
Superiority S37
38 I don’t mind blending
into the crowd when I go
out in public
I get upset when people
don’t notice how I look
when I go out in public
Exhibitionism Exh38
39 There is a lot that I can
learn from other people
I am more capable than
other people
Self-sufficiency SS39
40 I am much like every-
body else
I am an extraordinary
person
Superiority S40
121
7.3 Network analysis
The software R (open source, available at https://www.r-project.org/) was used to carry
out the analyses. I used the R-packages “qgraph” (Epskamp et al., 2012) and “glasso”
(Friedman et al., 2014b) and IsingFit (van Borkulo et al., 2014) for network estimation and
visualization, and “bootnet” (Epskamp and Fried, 2018) for stability analyses.
7.3.1 Network estimation
An Ising Model (IM) was estimated from my dataset. An IM (van Borkulo et al., 2014;
Marsman et al., 2018) is the binary equivalent of the Gaussian Graphical Model used for
continuous datasets (Epskamp and Fried, 2018). A lasso (least absolute shrinkage and se-
lection operator) was used to provide a conservative network structure (Epskamp and Fried,
2018). I used the default eLasso procedure which combines an l1-regularized logistic re-
gression with an Extended Bayesian Information Criterion or EBIC (Chen and Chen, 2008)
which reports relevant connections between variables. The lasso procedure provides a neigh-
borhood (set of nodes that interact) and decides the best set of regression coefficients given
the data, based on EBIC (which is in turn based on log likelihood); the set of regression
coefficients with the lowest EBIC is the best fit. To construct the final network, a connec-
tion is drawn between two nodes A and B if node A has node B in its set of neighbors and
vice-versa.
The default eLasso procedure was used in bootnet and IsingFit (van Borkulo et al.,
2014). The hyperparameter gamma (to select how many edges the model recovers) was set
by default at 0.25; the optimal tuning parameter lambda (used to select the model with
the best fit) was automatically chosen by the eLasso procedure. The network structure
resulting from this estimation contains items from the NPI represented as nodes. An edge
is a connection between two nodes in the network, which is interpreted as the existence of a
connection between two nodes controlling for all other nodes in the network.
122
While estimating a network structure from items of a questionnaire, a connection between
two nodes means that the observed group answers on average in a similar way to both items
of the questionnaire (Briganti et al., 2018). Each edge in the network represents either a
positive (visualized as blue edges) or a negative connection (visualized as red edges). The
thickness and color saturation of an edge denotes its weight (the strength of the connection
between two nodes). The Fruchterman-Reingold algorithm places the items in the network
based on the inverse of the sum of connections of a given node with other nodes (Fruchterman
and Reingold, 1991): this means that strongly connected nodes are put closer in the network
visualization.
7.3.2 Network inference
I estimated strength centrality (Boccaletti et al., 2006) for the 40 items in the questionnaire.
Strength centrality represents the absolute sum of the edges of a given node and therefore
informs us of the connectedness of items in the network (Briganti et al., 2018).
7.3.3 Network stability
Stability analyses (Epskamp and Fried, 2018) were carried out through bootstrapping, which
is a repeated estimation of a model under sampled data: I used 2000 bootstraps in this pa-
per. An edge weight difference test was performed to compare all edges against all other
edges and to answer the question “is edge A significantly stronger than edge B?”. Centrality
stability analyses for strength centrality were also carried out to answer the question “is the
centrality order stable?”. Centrality difference test was performed to answer the question
“is the centrality estimate of node A statistically different from that of node B?”. I used the
subsetting bootstrap procedure that re-estimates the network with a dropping percentage of
participants to determine the stability of centrality estimation, and results in a centrality-
stability coefficient (CS-coefficient) that should not be lower than 0.25 and preferably above
0.5. Both difference tests (edge weight and centrality) are carried out by estimating con-
123
fidence intervals around the difference of two elements A and B (which are bootstrapped
edge weights or bootstrapped centrality estimates, depending on the test): if 0 belongs in
the confidence interval then there is no difference between A and B.
7.4 Results
7.4.1 Participants
Participants were 17 to 25 years old (M=20 years, SD=1.7 years), 55% of them were female
and 45% were male. 25.4% of students studied engineering, 20% medicine, 17.7% economics,
11.3% sciences, 4.7% psychology and 2% law. The average NPI score of the participants of
this study was 13 (out of 40), and the standard deviation was 6.4.
7.4.2 Network of narcissism
Figure 7.1 illustrates the estimated network of the 40-item NPI. Overall, items from the
NPI form a positively connected network. The strongest connections in the network are
found between nodes belonging to the same domain of narcissism: for instance, item 10
(“I see myself as a good leader”) is strongly associated to item 33 (“I would prefer to be
a leader”) and both belong to the authority domain; item 7 (“I like to be the center of
attention”) presents the second strongest connection in the network to item 30 (“I really
like to be the center of attention”) and both belong to exhibitionism; item 9 (“I think
I am a special person”) shares the strongest edge of the network with item 40 (“I think
I am an extraordinary person”) and both belong to the superiority cluster. In the case
of these three connections, the items involved in an edge measure the same aspect of the
construct. Several connections are found between items belonging to different domains, and
I want to illustrate some of these connections. Domains superiority and self-sufficiency are
connected through items 9 (“I think I am a special person”) and 39 (“I am more capable than
other people”); domains authority and entitlement connect through items 12 (“I like to have
124
A1
Exh2
Exh3
S4
En5
Exp6
Exh7
A8
S9
A10
A11
A12
Exp13
En14
V15
Exp16
SS17
En18
V19Exh20
SS21SS22
Exp23
En24
En25
S26
En27
Exh28
V29
Exh30
SS31
A32
A33
SS34
Exp35
A36
S37
Exh38
SS39
S40
Figure 7.1: Network composed of the 40 items from the NPI (Table 7.1). Each item isrepresented by a node (1 to 40) and belongs to a different domain of the NPI (indicatedby a color code). The name of each node is composed as following: an abbreviation of thedomain to which the item belongs to followed by the item number.
125
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
0 10 20 30 40
01
23
45
67
Index
grap
h1.c
$InD
egre
e
Figure 7.2: Strength centrality estimates for the 40 items of the NPI. The Y-axis repre-sents centrality indices (the higher the estimate the more central the item), and the X-axisrepresents the 40 NPI items.
authority over other people”) and 27 (“I have a strong will to power”); domains authority and
exploitativeness connect through items 1 (“I have a natural talent for influencing people”)
and 35 (“I can make anybody believe anything I want them to”). These domains also tend to
measure the same thing, even though belonging to different domains. Some small, negative
edge are also found in the network, such as the one between items 11 (“I am assertive”) and
24 (“I expect a great deal from other people”).
7.4.3 Network inference
Figure 7.2 shows strength centrality estimates for the 40-item NPI. Item 27 from entitlement
(“I have a strong will to power”) presents the highest strength estimate, which means that
it is the most interconnected node in the network. Other strong items include item 33
from authority (“I would prefer to be a leader”) and item 40 from superiority (“I am an
extraordinary person”). Several items present with a strength centrality of 0, which means
that they are not connected with any item in the network.
126
7.4.4 Network stability
The edge weight bootstrap shows relatively narrow CIs, which indicates a precise estimation
of the edge weights in the network. The edge-weight difference test performed shows that
stronger edges are significantly stronger than other edges in the network; however, edges
9-40 and 7-30 are not statistically different from each other, which means that, even though
edge 9-40 reports a stronger connection in the network, I cannot safely interpret it to be
statistically stronger than edge 7-30. Strength centrality stability analyses report that the
centrality order is relatively stable, with a centrality stability coefficient (CS-coefficient)
of 0.67. Strength centrality difference test reports that stronger centrality estimates are
significantly stronger than other estimates but are not significantly different from each other;
for instance, I cannot infer whether the centrality of item 27 is really stronger than that of
item 33. I obtained a CS-coefficient of 0.67, which indicates stable results.
7.5 Discussion
This study is to my knowledge the first application of network analysis to the NPI. Con-
nections are shown between narcissistic domains and shed light on how they interact. Items
from the NPI are overall positively connected and some items are more connected than
others. Most items from the NPI share some variance and are connected. However, some
items present weak connections with others; this means that some nodes are conditionally
independent of all other items in the network. Connections exist both between items from
the same domain and between items from domain, and stability analyses show that I can
safely interpret connections in this study.
Several strong connections between items from the NPI are found in the network. In
the case of the three connections between items 10-33, 7-30 and 9-40 belonging to the same
domains (respectively authority, exhibitionism and superiority) as described in the Results
section, the interpretation of an edge changes (Fried and Cramer, 2017), and the resulting
127
connection simply represents shared variance between the two questions (since they measure
the same thing). In some cases, items from different domains also tend to represent the same
construct, such as items 12 and 27 that connect authority and entitlement in the network.
These items can be considered as “bridge items”, since they can transfer information from
one domain to another and vice-versa; however, bridge items as the examples described in
the Results section also tend to represent the same aspect of narcissism.
Centrality analysis shows that items from entitlement, authority, and superiority present
the highest strength centrality estimates: that means that items from these domains connect
well to a greater number of nodes in the network, therefore identifying these 3 domains as
containing specific items that are important in this NPI network. From a network point
of view, it is also not surprising to find entitlement to contain central items, as this find-
ing supports my current gold standard models of narcissism, the trifucated model (Miller
et al., 2016) and the narcissism spectrum model (Krizan and Herlache, 2018) that describe
entitlement as a connection between grandiosity and vulnerability. My finding also supports
the recent network study of pathological narcissism (Di Pierro et al., 2019), which reported
high centrality values for Entitlement Rage. In the network approach, if the observed group
scores high on a highly central node, then the observed group is also more likely to score high
on a relevant number of nodes in the network. The identification of central items may help
in identifying potential targets for clinical intervention in people suffering from narcissism.
My findings should be interpreted in the light of several limitations. First, my dataset is
composed of university students, which limits the potential generalization of my findings to
different samples. Second, because this is a cross-sectional study, I cannot infer whether a
given node (item or domain) causes or is caused by another node to which it is connected.
Third, redundancy among items that measure the same thing is an important issue that
has yet to be solved in psychological networks of self-reported questionnaires; in the case
of the NPI network, several items can be considered as redundant, which would alter the
connectivity with other items (such as reported with strength centrality values).
128
Further studies may endeavor to replicate my findings in different samples, both non-
clinical and clinical, to identify central features of narcissism.
129
Chapter 8
A network model of alexithymia
without fantasizing
Abstract
The aim of this paper is to explore network structures of the Toronto Alexithymia
Scale (TAS) in a large sample of 1925 French-speaking Belgian university students
and compare results with previous studies from different samples and tools to identify
potential targets for clinical intervention. I estimated network models for the twenty
items of the TAS and for its three domains difficulty identifying feelings, difficulty de-
scribing feelings and externally-oriented thinking. I explored item connectivity through
node predictability (shared variance with other network components). I performed an
Exploratory Graph Analysis (EGA) to explore the dimensionality of my dataset and
compare results with the original three-factor model; because a different model was pro-
posed, I estimated an additional network structure on the new structure. Items from
the TAS connect both within and between domains. The three-domain network iden-
tifies difficulty describing feelings as the most connected domain. The EGA reported
that three items from externally-oriented thinking form a new domain, distraction. In
the new four-domain network, difficulty describing feelings remains the most intercon-
nected domain; however, two negative connections are found. My findings support the
130
relative importance of identifying and describing feelings as a meaningful target for
intervention.
8.1 Introduction
The construct of alexithymia has been around for more than half a century: it was named
from Greek meaning “lack of word for emotion” (Sifneos, 1972). Alexithymia is understood
as a personality construct with four main features: 1) difficulty identifying and distinguishing
emotions from bodily sensations; 2) difficulty describing and verbalizing emotions; 3) poverty
of fantasy life; and 4) externally-oriented thinking (Loas et al., 2017).
It is a subject of interest in psychiatric research so as to better understand the physi-
ological basis of mental disorders associated with emotions, such as mania, addiction and
depression. Extensive neuroimaging research has been conducted on alexithymia, report-
ing alterations in affective arousal to external stimuli, voluntary cognitive functioning; an
impaired activation of the insula during cognitive processes has indicated a potential over-
lap between alexithymia and other psychiatric disorders presenting with impaired empathy
such as psychopathy and autism (Moriguchi and Komaki, 2013). Because of such clinical
implications, some researchers recommend totranspose the concept of alexithymia to that of
“affective agnosia” (Lane et al., 2015). The Toronto Alexithymia Scale (TAS) is the most
commonly used measure of alexithymia in empirical research (Bagby et al., 1994) and is
composed of twenty items designed to assess three domains: difficulty identifying feelings
(e.g. item 1 “I am often confused about what emotion I am feeling”); difficulty describing
feelings (e.g. item 4 “It is difficult for me to find the right words for my feelings”), and
externally-oriented thinking (e.g. item 8 “I prefer to just let things happen rather than to
understand why they turned out that way”). These three domains as originally described
are to be interpreted as passive and interchangeable consequences of the construct itself, and
therefore are not active contributors to alexithymia (Van Bork et al., 2017).
131
In the last decade, network analysis has affirmed itself as new way of analyzing data
in psychiatry and psychology, which allows to conceive constructs or mental disorders as a
complex system of mutually influencing elements (Borsboom and Cramer, 2013). In this
framework, the unobserved interactions between psychological components (for instance
symptoms, or items from a questionnaire) are often computed as regularized partial cor-
relations (Epskamp and Fried, 2018). Network analysis has been used to explore several
mental disorders, such as posttraumatic stress disorder (Fried et al., 2018; Phillips et al.,
2018), depression (Mullarkey et al., 2018), and autism (Ruzzano et al., 2015), but also psy-
chological constructs such as empathy (Briganti et al., 2018), self-worth (Briganti et al.,
2019) and resilience (Fritz et al., 2018; Briganti and Linkowski, 2019b).
When considering a network model of a psychological construct (such as represented by
a given questionnaire), a known challenge is that of redundancy of items in addressing the
measure of the same aspect of a given construct. This may result in the case of an empirical
study in a network where redundant items share very strong connections whose meaning
is different (it represents shared variance rather than a true interaction between the two
components). A solution to address this problem is the estimation of a network model of
the domains of the construct instead of a network model of items (Briganti et al., 2019)
to facilitate the process of inference (e.g. identifying relevant interactions between network
components).
Network analysis is also useful to recover the true number of item communities (or fac-
tors) in a dataset; this can be achieved to the highly effective and reliable Exploratory Graph
Analysis or EGA (Golino and Demetriou, 2017); in this case, a confirmatory factor analysis
(CFA) is useful to authenticate the structure suggested by EGA, as suggested in a simula-
tion study (Golino and Epskamp, 2017). A recent empirical study (Briganti and Linkowski,
2019b) which applied the recommended EGA procedure showed that the problem of redun-
dancy in psychological networks is not exclusive to questionnaire items but can extend to
construct domains as well; community detections in psychological networks may therefore
132
also be useful to re-evaluate the construct themselves.
Network models of a construct such as alexithymia are particularly interesting to inte-
grate in clinical practice since relevant components may serve as targets for intervention
(Fried et al., 2018); in the case of alexithymia, finding and acting upon potentials tar-
gets is particularly interesting since it may attenuate the neurocognitive alterations that
have been described in the literature. The construct of alexithymia as represented by the
Bermond-Vorst Alexithymia Questionnaire (BVAQ) and the Toronto Structured Interview
for Alexithymia (TSIA) has already been explored through the lenses of network analysis
(Watters et al., 2016a,b); in the case of the TSIA whose factors closely resemble those of the
TAS, the domains difficulty identifying feelings and difficulty describing feelings formed one
community and provided the most connected items.
I therefore aim to extend the conceptual framework of psychological networks to the
construct of alexithymia as represented by the TAS. In this study, I want first to explore
alexithymia as both a network of items (as reported in the questionnaire) and domains.
Second, I want to apply EGA to recover the number of communities in the network, compare
the outcome with previous studies (Watters et al., 2016a,b), and authenticate the suggested
structure with CFA; if the result differs from the original three-domain structure, I want
to estimate the network of the structure proposed by the community detection. Third, I
want to explore the connectedness of items and domains with node predictability, which
statistically represent the shared variance of a given network component with surrounding
components (Haslbeck and Fried, 2017). These exploratory analyses may have meaningful
clinical implications and serve as basis for future replication studies to identify and act upon
potential clinical targets in people with alexithymia.
133
8.2 Method
8.3 Data set
My dataset is composed of 1925 university students from the French-speaking region of
Belgium. Subjects were 17 to 25 years old (M = 19 years; SD = 1.6 years); 58% of them
were women and 42% were men.
8.4 Measurement
The TAS (Table 8.1) is composed of 20 shuffled items that measure alexithymia in three
domains: difficulty identifying feelings, difficulty describing feelings and externally-oriented
thinking. The minimum score for each item is 1 (“I completely disagree”) and the maximum
score is 5 (“I completely agree”). The questionnaire contains reverse-scored items whose
score is reversed before data analysis (Table 8.1). The protocol for this study was approved
by the ethical committee of the Erasme teaching hospital.
Table 8.1: The Toronto Alexithymia Scale (Bagby et al.,
1994)
N Item Domain WC
1 I am often confused about what emotion I
am feeling
Difficulty identifying
feelings
1
2 It is difficult for me to find the right words
for my feelings
Difficulty describing
feelings
4
3 I have physical sensations that even doctors
don’t understand
Difficulty identifying
feelings
1
4 I am able to describe my feelings easily (re-
versed)
Difficulty describing
feelings
4
134
5 I prefer to analyse problems rather than just
describe them (reversed)
Externally-oriented
thinking
3
6 When I am upset, I don’t know if I am sad,
frightened, or angry
Difficulty identifying
feelings
1
7 I am often puzzled by sensations in my body Difficulty identifying
feelings
1
8 I prefer to just let things happen rather than
to understand why they turned out that way
Externally-oriented
thinking
3
9 I have feelings that I can’t quite identify Difficulty identifying
feelings
1
10 Being in touch with emotions is essential (re-
versed)
Externally-oriented
thinking
3
11 I find it hard to describe my feelings more Difficulty describing
feelings
4
12 People tell me to describe my feelings more Difficulty describing
feelings
4
13 I don’t know what’s going on inside me Difficulty identifying
feelings
1
14 I often don’t know why I am angry Difficulty identifying
feelings
1
15 I prefer talking to people about their daily
activities rather than their feelings
Externally-oriented
thinking
2
16 1 prefer to watch “light” entertainment
shows rather psychological dramas
Externally-oriented
thinking
2
17 It is difficult for me to reveal my innermost
feelings, even to close friends
Difficulty describing
feelings
4
135
18 I can feel close to someone, even in moments
of silence (reversed)
Externally-oriented
thinking
3
19 I find examination of my feelings useful in
solving personal problems (reversed)
Externally-oriented
thinking
3
20 Looking for hidden meanings in movies or
plays distracts from their enjoyment.
Externally-oriented
thinking
2
8.4.1 Network analysis
Software
I used the software R (open source, available at https://www.r-project.org/). Packages
and functions to carry out the analysis include qgraph (Epskamp et al., 2012), glasso (Fried-
man et al., 2014b) for network estimation and visualization, mgm (Haslbeck and Waldorp,
2016) for node predictability, EGA (Golino and Epskamp, 2017) and igraph (Csardi and Ne-
pusz, 2006) for community detection, lavaan (Rosseel, 2012) for CFA and bootnet (Epskamp
and Fried, 2018) for stability.
Item network
A correlation matrix was calculated for the 20 TAS items and used as input to estimate a
Gaussian Graphical Model (GGM) which is a regularized partial correlation network (Ep-
skamp and Fried, 2018). I regularized the parameters of the GGM with a graphical lasso
(least absolute shrinkage and selection operator). Although recent relevant discoveries in
the network field (Williams et al., 2019) showed that regularized partial correlation network
estimation may provide an anti-conservative network model, my regularized TAS network
model reports fewer edges (and is therefore easier to interpret) than the nonregularized
partial correlation network.
The item network is composed of nodes representing alexithymia questions from the
136
TAS, and each node is surrounded by a pie chart representing its predictability (shared
variance with surrounding nodes). The position of nodes in the network is determined by
the Fruchterman-Reingold algorithm (Fruchterman and Reingold, 1991), which runs on the
sum of connections that a given node has with other nodes. Nodes are connected through
edges, which are interpreted as an association between two components, controlling for all
other nodes in the network. In the alexithymia item network, if two nodes A and B are
connected, it means for instance that if the observed group scored high on A, then the
observed group is also more likely to score high on B and vice versa (Briganti et al., 2018).
Each edge in the network has a weight which is defined as the importance of association
between two nodes and is denoted by thickness and color saturation; edges can therefore
be positive (blue edges, denoting a positive association) or negative (red edges, denoting a
negative association). For this work, I lowered the lambda parameter (tuning parameter).
Three-domain network
To better explore the interactions between alexithymia domains from the TAS, I estimated
a factor model for the three domains and obtained factor scores to estimate a GGM and
therefore a domain network as described in recent empirical papers (Briganti et al., 2019;
Briganti and Linkowski, 2019b).
Network stability
To safely interpret results from network analysis several stability tests were performed as
recommended in the literature (Briganti et al., 2018; Epskamp and Fried, 2018). To answer
the question “are edge weights accurately estimated”, 95% confidence intervals (CI) were
estimated through bootstrapping (2000 bootstraps were used). To answer the question “is
edge A significantly stronger than edge B” I performed an edge weight difference test.
137
Network inference
Node predictability (Haslbeck and Fried, 2017) represents shared variance that a given node
has with surrounding nodes; it has been defined as the upper bound of controllability: if
one assumes that all edges for a given node are directed toward that node, then node pre-
dictability provides an estimate of how much influence one can have on that node via all
other nodes (Briganti et al., 2019). Node predictability was estimated for all nodes in the
item and domain network and is represented as a pie chart surrounding each node.
To detect the number of communities in my TAS dataset, an EGA was performed (Golino
and Epskamp, 2017), which uses the walktrap algorithm. The algorithm is based on the
principle that adjacent nodes tend to belong to the same community (Yang et al., 2016).
Following the methodology of a recent study (Briganti and Linkowski, 2019b) and because
I found a different number of dimensions than the original model, I estimated an additional
network structure based on the results from EGA.
Four-domain network
I used CFA to estimate a four-factor model in the data and used the scores to estimate an
additional GGM. Node predictability and stability analyses were also carried out.
8.5 Results
8.5.1 Item network
Figure 8.1 shows the 20-item alexithymia network. Nodes connect both within and between
communities. I will hereby detail some of the connections. Item 2 (“It is difficult for me to
find the right words for my feelings”) is highly connected to item 4 (“I am able to describe
my feelings easily”, reversed) and both belong to difficulty describing feelings. Item 3 (“I
have physical sensations that even doctors don’t understand”) shares a connection with item
7 (“I am often puzzled by sensations in my body”) and both belong to difficulty identifying
138
TAS1
TAS2 TAS3TAS4
TAS5
TAS6
TAS7
TAS8
TAS9
TAS10
TAS11
TAS12
TAS13
TAS14
TAS15
TAS16
TAS17
TAS18
TAS19
TAS20
Figure 8.1: 20-item alexithymia network. Each node represents an item from the TAS ques-tionnaire (Table 8.1). The pie chart surrounding each node represents node predictability.
feelings. In these two cases, items strongly resemble each other and are therefore redundant;
the meaning of edges in this case shifts and should be interpreted as shared variance. Item 1
(“I am often confused about what emotions I am feeling”) from difficulty identifying feelings
and 2 (“It is difficult for me to find the right words for my feelings”) from difficulty describing
feelings are also strongly connected. Item 8 (“I prefer to just let things happen rather than to
understand why they turned out that way”) from externally-oriented thinking is associated
with item 9 (“I have feelings that I can’t quite identify”) from difficulty identifying feelings.
Item 15 (“I prefer talking to people about their daily activities rather than their feelings”)
from externally-oriented thinking is associated with item 17 (“It is difficult for me to reveal
my innermost feelings, even to close friends”) from difficulty describing feelings. Items 15,
16 and 20 from externally-oriented thinking are detached from the rest of the items from the
same factor.
139
ID
DES
Think
ID: Difficulty identifying feelingsDES: Difficulty describing feelingsThink: Externally−oriented thinking
Figure 8.2: Three-domain network. Each node represents a domain from the TAS.
8.5.2 Three-domain network
Figure 8.2 illustrates the three-domain alexithymia network. The domain difficulty describing
feelings is connected to the two other domains difficulty identifying feelings and externally-
oriented thinking which share no connection with each other. The connection between
difficulty describing feelings and difficulty identifying feelings is stronger than the connection
between difficulty describing feelings and externally-oriented thinking.
8.5.3 Network stability
Edges are overall accurately estimated in both the item network and the three-domain net-
work, and I can safely interpret stronger edges in both networks to be significantly stronger
than weaker edges.
140
8.5.4 Network inference
Node predictability
In the item network, the most predictable node is item 2 (“It is difficult for me to find
the right words for my feelings”), which shares 54% of variance with surrounding nodes.
Item 2 belongs to difficulty describing feelings, which is also the most predictable domain
in the three-domain network and shares 72% of variance with the two other domains. The
least predictable node in the item network is item 18 (“I can feel close to someone, even
in moments of silence”) which belongs to externally-oriented thinking and shares only 8%
of variance with other nodes. In the three-domain network, the least predictable node is
externally-oriented thinking which shares 12% of variance with the two other domains. The
mean node predictability is 24% in the item network and 51% for the three-domain network.
In both networks, difficulty describing feelings shows the most connectivity and externally-
oriented thinking the least connectivity.
Community detection
Figure 8.3 shows the twenty-item TAS network with colors corresponding to the communities
as detected by the EGA. Four communities are reported by the EGA: a new community
emerges from items 15, 16 and 20 from externally-oriented thinking. I defined the new
community as distraction, and factor analyses for both the three-factor and the four-factor
model report comparable and satisfactory construct validity.
8.5.5 Four-domain network
Figure 8.4 shows the four-domain network. In the four-domain network, two strong edges
exist: the first shared between distraction and externally-oriented thinking, and the second
between difficulty identifying feelings and difficulty describing feelings. Distraction and dif-
ficulty describing feelings share a weaker connection, such as externally-oriented thinking
141
TAS1
TAS2 TAS3TAS4
TAS5
TAS6
TAS7
TAS8
TAS9
TAS10
TAS11
TAS12
TAS13
TAS14
TAS15
TAS16
TAS17
TAS18
TAS19
TAS20
Figure 8.3: Community detection through Exploratory Graph Analysis.
ID
DES
Think
DST
ID: Difficulty identifying feelingsDES: Difficulty describing feelingsThink: Externally−oriented thinkingDST: Distraction
Figure 8.4: Four-domain network. Blue edges represent positive connections, red edgesrepresent negative connections.
142
and difficulty identifying feelings. However, two weak negative connections are reported: the
first between difficulty identifying feelings and distraction and the second between externally-
oriented thinking and difficulty describing feelings. The highest node predictability (73%) is
reported by difficulty describing feelings, and externally-oriented thinking reports the lowest
node predictability (54%). The mean node predictability is 63%. Stability analyses show
that edges are accurately estimated and each edge in the network is significantly different
from other edges in the network.
8.6 Discussion
This is to my knowledge the first network analysis of the psychological construct of alex-
ithymia as conceived in the TAS with state-of-the-art methods. The different analyses carried
out bring new and interesting information on the construct and build on previous findings.
The TSA item network sheds light on the connectivity between items within the same domain
and items from different domains of alexithymia. Several items that are strongly connected
(for instance item 2,3,4 and 7), report a certain redundancy in measuring the same aspect of
alexithymia with other items; other items share meaningful mutual interactions while mea-
suring a different aspect of alexithymia. Items from the alexithymia questionnaire report
different levels of interconnectedness. The EGA reported the existence of four communities;
the domain externally-oriented thinking loses three of its items to a new component that I
defined as distraction.
Because of the redundancy issue and the difficulty on extracting meaningful clinical
implications from a network of items (Briganti and Linkowski, 2019b), I estimated network
models of three and four domains: these models help us better explore the connections
between different domains of the TAS. Difficulty describing feelings being the most connected
domain in both network models, it may become of interest for future longitudinal studies
associated with clinical interventions. From a three-domain network perspective, answers
143
of the observed group to items from difficulty describing feelings can help us better predict
answers to items from the two other domains, and vice versa.
In both models, the most interconnected domain shares the strongest connections with
difficulty identifying feelings, which has similar connectivity; this may be due to the strong
similarity between these two aspects of alexithymia, which have already been shown to
merge in a single community in other network studies (Watters et al., 2016a,b); in my
dataset however, the two domains remain separate, as shown by EGA. Future studies may
also endeavor to explore the relationship between these two domains in other samples, as
well as the evolution of their association when an intervention is carried out.
Externally-oriented thinking splits in two separate communities: this finding, along with
the small negative edges found in the four-domain network should however first be replicated
in other different samples before being considered as clinically meaningful. This study reports
meaningful, clinical implications even if carried out on university students, more so because
it can build on previous findings.
I found similar findings as well as differences with previous network models of alexithymia
measurement tools (Watters et al., 2016a,b). For instance, in my TAS network models,
difficulty describing feelings is the most interconnected domain and contains node 2, the most
interconnected item in the questionnaire; this constitutes a similarity with the previous TSIA
and BVAQ studies. However, this domain remains a separate community from difficulty
identifying feelings, which constitutes a difference with previous studies. Because of these
findings, I support the arguments made by the authors of previous alexithymia network
analyses in stating that in a clinical setting, priority should be given in intervening on affect
awareness to better identify and describe feelings. However, in my sample, because items
from externally-oriented thinking and the domain itself do not show particularly strong
connectivity (even so because the domain splits into two distinct communities in the EGA),
it should not be given a higher priority in intervention than other components of alexithymia;
future studies may endeavor to replicate and compare findings to further understand if it is
144
a viable clinical target or not.
My findings should be met with a number of limitations. First, my dataset is composed of
university students, which likely limits the generalization of my findings to different samples.
Second, because this is a cross-sectional study, I cannot infer causality from connections
between nodes; this can be obtained, for instance, with time-series analyses.
Network analysis is a powerful set of tools to explore alexithymia in both clinical and
nonclinical samples. Future studies should endeavor to replicate findings from different
measurement tools and also address clinical samples of individuals with psychopathology.
This may help to confirm targets for clinical interventions (such as highly connected nodes
from the domain networks) and respective outcome in different types of mental illness.
145
Chapter 9
A network model of alexithymia with
fantasizing
Abstract
The aim of this paper is to explore the network structures of alexithymia compo-
nents in a large sample of university students and compare results with relevant prior
literature. Undirected and directed network structures of items from the Bermond
Vorst Alexithymia Questionnaire form B are estimated with state of the art network
analysis and structure learning tools. Centrality estimates are used to address the
topic of item redundancy and select relevant alexithymia components to study. The
undirected network structure of alexithymia components reports new features with re-
spect to prior literature, and the directed network structures offers new insight on the
construct.
9.1 Introduction
Alexithymia is named from the Greek words “a”, “lexis”, and “thymos”, meaning “lack of
word for emotion” (Sifneos, 1972): initially, it was used to describe emotional deficiencies
in patients suffering from classic psychosomatic disorders and epilepsy; those patients were
146
unaware of their feelings and were unable to fantasize about their inner thoughts, feelings,
and attitudes. This personality construct has been described more than half a century ago
and is characterized by several key features: difficulty identifying, verbalizing and analyzing
emotions, poor fantasy life and poor insight (Loas et al., 2017); these features have been
found to be constant over time, in contrast to what was initially observed.
Alexithymia is considered a subject of interest in psychiatric research, because it allows
for a deeper understanding of the physiological basis of mental disorders that are associated
with emotions, such as bipolar disorder, addiction and depression (Briganti and Linkowski,
2019c). Important clinical implications include this construct, such as the potential overlap
between alexithymia and other psychiatric disorders that present with a lack of empathy,
such as psychopathy and autism, as reported by extensive neuroimaging research that has
been conduced on the topic (Moriguchi and Komaki, 2013). Some authors recommend to
transpose the construct of alexithymia to that of “affective agnosia” (Lane et al., 2015).
Several psychometric tools have been validated to measure the construct of alexithymia.
One of the most well-known and widespread tools is the Bermond Vorst Alexithymia Ques-
tionnaire (BVAQ, shortened to AQ in this manuscript) which describes the construct of
alexithymia as a composed of five domains (Vorst and Bermond, 2001): difficulty identifying
emotions, difficulty analyzing emotions, difficulty verbalizing emotions, lack of emotional
insight, and poverty of fantasy life. This last domain of alexithymia is what sets the AQ
apart from its main counterpart, the Toronto Alexithymia Scale (Bagby et al., 1994), since
the latter psychometric scale does not contain any items that reflect the difficulty in fan-
tasizing. The AQ has been defined as a reliable tool for the study of alexithymia, and it
has been heavily investigated with exploratory and confirmatory factor analyses in several
populations (de Vroege et al., 2018).
From an ontological point of view, alexithymia as represented by psychometric tools such
as the AQ, developed and validated through the lenses of factor analyses, is a common cause
that can be measured via the items in the questionnaire; those items are a reflection of a
147
given factor (such as poor fantasy life), and each factor is itself a consequence of the common
cause that is the personality construct at hand (“the latent variable“). Hence, the observable
variables (the items themselves) only represent passive and interchangeable elements of the
latent variable. However, previous work on alexithymia highlighted the opportunity that is
the study of relationships between observable variables as a complementary tool to factor
analysis (Watters et al., 2016a,b).
The study of relationships between observable variables is allowed by network analysis,
which is a new way of analyzing psychiatric constructs as complex systems arising from inter-
actions between symptoms or components (Borsboom and Cramer, 2013). Such systems are
conceived as networks of nodes (the variables themselves) and edges (undirected connections
among variables): often, the unobserved connection among items are computed as partial
correlations, either regularized (Epskamp and Fried, 2018) or non-regularized (Williams and
Mulder, 2019); the latter has been shown to be a good fit for psychological data, since it is
low dimensional (the number of subjects exceeds by far the number of variables) (Williams
et al., 2019).
Network analysis is becoming more and more established in the field of psychometrics
and has been used to explore several mental disorders, such as posttraumatic stress disorder
(Fried et al., 2018; Phillips et al., 2018), depression (Mullarkey et al., 2018), autism Ruzzano
et al. (2015) and also psychological constructs, such as empathy (Briganti et al., 2018), self-
worth (Briganti et al., 2019), resilience (Fritz et al., 2018; Briganti and Linkowski, 2019b),
and narcissism (Briganti and Linkowski, 2019a). Alexithymia has been analyzed three times
with network analysis, two with the Toronto Alexithymia Questionnaire variants (Briganti
and Linkowski, 2019c; Watters et al., 2016b) and the AQ (Watters et al., 2016a): all these
studied in depth the connections between the items from alexithymia scales as well as the
regrouping of variables in domains.
The modeling of network structures of constructs such as alexithymia is particularly
interesting to integrate in clinical practice, since relevant components may serve as targets
148
for clinical intervention (Fried et al., 2018); in the case of alexithymia, finding and acting upon
relevant components may attenuate the neurocognitive alterations that have been described
in the literature.
However, the identification of the central components of a construct is complicated be-
cause of the redundancy of items in questionnaires (Briganti and Linkowski, 2019d): the
more redundancy exists in a questionnaire, the more the redundant items will be heavily
connected, which in turn will boost their relative importance: this can be called “central-
ity corruption” (Briganti and Linkowski, 2019c). Several strategies have been proposed to
overcome redundancy, such as exploring networks of domains instead of networks of items
(Briganti et al., 2019) and topological overlap. The latter has been proposed as a way of
dropping (or regrouping) the items that repeat the same aspect of a construct as other items
in a scale (Fried and Cramer, 2017), but it has not been used experimentally to regroup
items from a psychometric tool.
Moreover, because of the clinical relevance of alexithymia, it would be useful to uncover
causal relationships (directed connections among items) between meaningful items in order to
gain further information on the nature of connections among them. Such causal relationships
can be identified with specific tools in network science, such as Directed Acyclic Graphs
(DAGs). DAGs are the foundation of probabilistic models such as Bayesian networks and
other machine learning approaches that are capable of learning the underlying causal graphs
from data (Moffa et al., 2017), compute and represent such relationships. DAGs are well
established at the crossroads of machine learning and network science literature (Scutari
and Denis, 2015) and have been previously used in empirical research to explore depression
(McNally et al., 2017) and psychosis (Moffa et al., 2017).
Inspired by recent works in the field of both structure learning and alexithymia, I aim
to explore several network structure of the AQ. First, I will estimate a partial correlation
network of items from the AQ and infer their relative importance with established measures
in the field (Briganti et al., 2018). Second, I will tackle the problem of centrality corruption
149
by reducing the network to a group of five items from the scale based on their belonging to
a given domain and their relative importance in the network. Third, I will apply a structure
learning algorithm to construct a DAG of the main alexithymia components and therefore
explore causal pathways.
9.2 Method
9.2.1 Data set
The data set is composed of 537 university students attending programs from academic
institutions in the French-Speaking region of Belgium. Subjects were 17 to 25 years old
(M=20 years; SD=1.7 years). 71% of students were women and 29% were men.
9.2.2 Measurement
The AQ (Vorst and Bermond, 2001) is composed of items assessing alexithymia in five
domains: difficulty identifying emotions, difficulty analyzing emotions, difficulty verbalizing
emotions, lack of emotional insight, and poverty of fantasy life. In this study, the form B
of the questionnaire, which is composed of 20 items, was used. The data set for this study
was anonymized before analysis, and the protocol for this study was approved by the ethical
committee of the Erasme teaching hospital.
9.2.3 Network analysis
Software and packages
I used the software R for statistical computing (version 3.6.1, open source, available at
https://www.r-project.org/). The package used to carry out the analysis include qgraph
(Epskamp et al., 2012) for the undirected network estimation and visualization, bootnet
150
(Epskamp and Fried, 2018) for stability analyses and bnlearn (Scutari, 2010) for DAG esti-
mation.
Partial correlation network
Estimation of the partial correlation network I estimated a Gaussian Graphical
Model (GGM), that is, a partial correlation network for the items in the AQ. The GGM is
calculated as the inverse-covariance matrix: it is a network that includes a set of nodes that
correspond to the alexithymia items in the AQ and a set of edges that connect the nodes in
the network. If two nodes are connected, that means they are conditionally dependent given
all other nodes in the network (i.e their partial correlation is nonzero).
In the network of alexithymia components, if two nodes A and B are connected, it means
for instance that if the observed group scored high on component A, then the observed group
is also more likely to score high on component B, and vice versa, controlling for other nodes
in the network (Briganti et al., 2018). Each edge in the network has a weight representing
the strength of association between two alexithymia components; edges can be positive (and
therefore represent a positive association) or negative (denoting a negative association). In
the network the edge weight is represented as a combined thickness and saturation of the
edge; positive edges are shown in blue, and negative edges in red. Nodes are placed in the
network by the Fruchterman-Reingold algorithm, based on the sum of the connections a
given node has with other nodes (Fruchterman and Reingold, 1991).
Network inference To find comparatively important items in the partial correlation net-
work, I used strength centrality, which represents the absolute sum of the edges that nodes
in the network share with other nodes (Boccaletti et al., 2006).
Network accuracy and stability Accuracy and stability analyses were carried out fol-
lowing state-of-the-art methods (Epskamp and Fried, 2018) that were applied in previous
empirical papers (Briganti et al., 2018).
151
Accuracy analyses were carried out to answer the question: “is edge X accurately esti-
mated?”; 95% confidence intervals (CI) were estimated through bootstrapping (i.e., repeated
re-sampling from the original dataset to re-estimate network parameters; 2000 bootstraps
were used). Edge weight difference tests were carried out to answer the question: “is edge
X significantly stronger than edge Y?”.
Stability analyses were carried out to answer the question “is the centrality order stable?”
with the same bootstrapping method. Centrality difference tests were carried out to answer
the question “is the centrality of node A significantly stronger than the centrality of node
B?”.
Topological overlap To address the important topic of redundancy (i.e., items in the
questionnaire measuring the same aspect of the construct of alexithymia), the approach
proposed by Fried and Cramer (Fried and Cramer, 2017) of topological overlap was used.
Because items from the same domain strongly resemble each other, the most central item
from each domain (i.e items reporting the highest strength centrality score) was selected to
represent the corresponding facet of the construct in a five-item network.
Five-item network structure A five-item network structure was constructed with the
same methods described in the section “Estimation of the partial correlation network”, and
it was studied with inference analysis (that is, strength centrality computation) as well as
stability and accuracy analyses.
Directed Acyclic Graphs
In Bayesian networks an edge may represent a causal pathway between two nodes. The
structure of a Bayesian network can be estimated using constraint-based algorithms, which
analyze conditional independence relations among the nodes in the network. Constraint-
based algorithms produce a network model that can be interpreted as a causal model even
from observational data, under assumptions that in clinical terms exclude confounding and
152
sampling bias. In this paper the PC-algorithm is used, which is a constraint-based algorithm
(Spirtes et al., 1993).
To estimate a network model, the PC-algorithm first estimates an undirected network
model in which all pairs of nodes are connected, then deletes edges between conditional
independent pairs of variables, and directs edges starting with v-structures (two disconnected
nodes causing a third node). The estimated network was investigated using stability analysis
through bootstrapping.
Afterwards, a network is reported with a minimum connection strength (% of fitted
networks in which a given connection appears) of 85 and a minimum connection direction
(% of fitted networks in which a given connection has a given direction) of 50. This resulting
network therefore reports connections that are presents in more than 85% of the fitted
networks. Moreover, these connections present a direction (for instance, from node A to node
B) which is found in more than half of the fitted networks resulting from the bootstrapping
procedure. By default, with the software I used all edges are represented as red, and their
strength are represented as a combination of thickness and color saturation in the edges.
9.3 Results
9.3.1 Partial correlation network
Central items in the AQ
The five items showing the highest strength centrality values in each of the five domains are
reported in table 9.1.
“Difficulty identifying emotions” is represented by item 22, “I rarely let myself go to
my imagination”; “difficulty analyzing emotions” is represented by item 30, “I think one
should stay in touch with one’s feelings” (reversed item); “difficulty verbalizing emotions”
is represented by item 26, “When I am upset by something, I tell others about how I feel”
153
Table 9.1: Central items from the Bermond Vorst Alexithymia QuestionnaireN Item Domain22 I rarely let myself go to my imagination. Difficulty Identifying Emotions26 When I am upset by something, I tell others
about how I feel.Difficulty Verbalizing Emotions
29 I often get upset by unexpected events. Poor Emotional Insight30 I think one should stay in touch with one’s
feelings.Difficulty Analyzing Emotions
33 When I’m tired of myself, I can’t know if I’msad, afraid, or unhappy.
Poor Fantasy
(reversed item); “lack of emotional insight” is represented by item 29, “I often get upset by
unexpected events” (reversed item), and “poverty of fantasy life” is represented by item 33,
“When I’m tired of myself, I can’t know if I’m sad, afraid, or unhappy”. Three of the five
items with the highest centrality in the AQ twenty-item network have reversed scores.
Partial correlation network structure
The five-item network is represented in figure 9.1. As opposite as most construct networks
reported in the literature (Briganti et al., 2018, 2019; Briganti and Linkowski, 2019b,a,c),
the AQ network is not overall positively connected, as it includes several positive as well as
negative edges. Some of the connections in the network are described in the following. The
domains “difficulty identifying emotions” and “poor fantasy” show a positive connection, as
well as the domains “difficulty verbalizing emotions” and “emotional insight”, “poor emo-
tional insight” and “difficulty analyzing emotions”, and “difficulty verbalizing emotions” and
“difficulty analyzing emotions”. The domains “poor emotional insight” and “poor fantasy”,
are negatively connected in the network.
Network inference
The centrality estimates for the five-item network are reported in figure 9.2. Item 33 from
the domain “poor fantasy” and item 29 from the domain “poor emotional insight” report
the highest strength centrality estimates.
154
A22
A26
A29
A30
A33
A22: Difficulty Identifying EmotionsA26: Difficulty Verbalizing EmotionsA29: Poor Emotional InsightA30: Difficulty Analyzing EmotionsA33: Poor Fantasy
Figure 9.1: Partial correlation network. Each item represents one of the five domains in theAQ. Blue connections represent positive edges, red connections represent negative edges.
155
●
●
●
●
●
Strength
−1.0 −0.5 0.0 0.5 1.0
A22
A26
A29
A30
A33
Figure 9.2: Strength centrality estimates for the five-item network. The x-axis reports thestandardized z-scores and the y-axis reports the corresponding item.
156
A22
A26
A29
A30
A33
A22: Difficulty Identifying EmotionsA26: Difficulty Verbalizing EmotionsA29: Poor Emotional InsightA30: Difficulty Analyzing EmotionsA33: Poor Fantasy
Figure 9.3: Directed Acyclic Graph of alexithymia components obtained with the constraint-based PC-algorithm. Relationships between nodes (arrows) can be understood as causalpathways under certain assumptions.
Network accuracy and stability
Edges are overall accurately estimated in both the five-item and twenty-item network. I can
safely interpret stronger edges to be significantly stronger than weaker edges in the network.
Centrality estimates of nodes 29 and 33 in the five-item network are not significantly different,
which means I cannot say which of the two items is the most central.
9.3.2 Directed Acyclic Graph
The DAG of alexithymia components is reported in figure 9.3. Item 29 from the domain
“poor emotional insight” has three incoming edges from “poor fantasy” (item 33), “difficulty
verbalizing emotions” (item 26) and “difficulty analyzing emotions” (item 30). Item 22 from
157
the domain “difficulty identifying emotions” has an outgoing edge to item 33 from the domain
“poor fantasy”.
9.4 Discussion
This is to my knowledge the first network analysis of alexithymia components that com-
bines the classic partial correlation network approach with the Bayesian network approach.
Several of the resulting analyses bring new and interesting information on the construct of
alexithymia.
The partial correlation network for the twenty-item AQ reports a structure with an
overall mixture of positive and negative connections among items. Because items from the
same domain tend to measure the same aspect of the construct of alexithymia (that is, the
domain they belong to), the solution of topological overlap is applied: the original twenty-
item network is translated to a network of the five items in the AQ that have the highest
centrality values in their respective domains and that are reported in table 9.1. In this work,
the exploratory analyses on the twenty-item network (representing the full questionnaire) was
important to highlight with network inference methods relevant alexithymia components so
as to analyze a more simple and non-redundant network structure.
The five-item network reports, similarly to the twenty-item network, a set of positive
and negative edges: alexithymia components therefore present a heterogeneous connectivity.
Items from domains “poor emotional insight”, “difficulty verbalizing emotions”, “difficulty
analyzing emotions” share a set of positive connections: this means that the average score
of the observed group on one of these three questions can be predicted based on the score
on the other two questions. The same phenomenon is observed with items representing
the domains “difficulty identifying emotions” and “poor fantasy”. However, “poor fantasy”
and “poor emotional insight” are negatively connected, which from an undirected network
perspective, can be interpreted as follows: given all other alexithymia components in the
158
network, if the average score on one component is high, I may expect that the average score
of the observed group on the other component is low, and vice versa. My findings differ
slightly in that respect from the recent work of Watters et al (Watters et al., 2016b), in
which the two domains sharing a negative connection are “difficulty identifying emotions”
and “poor fantasy”.
The inference analyses show that items 29 and 33 share a negative connection in the
five-item network and are also the two items with the highest centrality estimates. However,
stability analyses show how the two centrality indices are not statistically different from each
other, hence I cannot say whether item 29 (that reports the highest centrality estimate) is
really the most central item.
In this work, the study of the undirected interplay among relevant alexithymia compo-
nents was an important preliminary step before entering the realm of causal inference through
the lenses of Bayesian networks. The DAG structure derives from the constraint-based PC-
algorithm. Directed connections between alexithymia components can be interpreted as
causal pathways under assumptions that in clinical terms exclude confounding and sampling
bias. The DAG reports that “poor emotional insight”, the lack of ability to fantasize, is
essentially a consequence of a poor ability to fantasize, a difficulty in verbalizing emotions
and a difficulty in analyzing emotions.
The information from the undirected five-item network inference analyses and the Bayesian
network analyses can be combined to obtain some interesting insights. First, in the partial
correlation network, “poor emotional insight” has a high estimated centrality, and the overall
item connectivity to other items in the network can be interpreted as predictability. Second,
the DAG shows that all edges that item 29 shares with other nodes point towards “poor
emotional insight”, which means that not only is “poor emotional insight” is a highly “pre-
dictable” domain of alexithymia, but that it is also a highly “controllable” domain of the
construct. This notion can be interpreted as that the aspect of alexithymia that deals with
the lack of insight regarding emotional arousal can be controlled through other aspects of
159
the construct.
The two network models proposed in this paper present similarities as well as differences:
for instance, both the partial correlation network and the DAG share several edges between
the same nodes; however, some nodes that are connect in the partial correlation network
are not connected in the DAG. The reason is the different definitions of the two models: in
a DAG two nodes are not automatically connected when they share a common child node
(such as item 29), but they will be connected in the corresponding partial correlation network
because of the indirect dependence conditional on that child node.
My results must be interpreted in light of several limitations. First, my data set is com-
posed of university students, which may limit the generalization of my findings to different
samples. Second, DAG structures do not involve loops: if in a three-node network a com-
ponent A causes component B and a component C, component C cannot cause component
A (the structure is therefore acyclic). However, it is plausible to consider that in the case
of alexithymia components, certain items may activate each other in a loop. Third, causa-
tion may be inferred from a connection between two nodes in a DAG assuming there are
no confounding or sampling bias. Further studies may endeavor to replicate my findings of
network structures (both Bayesian and non-Bayesian) of alexithymia components in different
samples, both non-clinical and clinical.
160
Chapter 10
A network model of autistic traits
Abstract
The aim of this work is to explore the construct of autistic traits through the
lens of network analysis with recently introduced Bayesian methods. A conditional
dependence network structure was estimated from a data set composed of 649 university
students that completed an autistic traits questionnaire. The connectedness of the
network is also explored, as well as sex differences among female and male subjects
in regard to network connectivity.The strongest connections in the network are found
between items that measure similar autistic traits. Traits related to social skills are the
most interconnected items in the network. Sex differences are found between female and
male subjects. The Bayesian network analysis offers new insight on the connectivity
of autistic traits as well as confirms several findings in the autism literature.
10.1 Introduction
For the past two decades, there has been a growing interest in psychiatric research for the
classification and measurement of autistic traits (Volkmar et al., 2009). The Diagnostic and
Statistical Manual of Mental Disorders (DSM-V) has defined a framework for autistic traits:
the Autism Spectrum Disorder (ASD), that includes child autism, Kanner’s infantile autism,
161
high-functioning autism, and Asperger Syndrome.
To diagnose a patient with ASD as defined in the classification, the clinician must use
diagnostic criteria that associate persisting deficits in communication and social interactions
with narrow and repetitive behaviors, interests or activities. These two features manifest in
childhood, affect social, academic or professional functioning, and are not better explained by
an intellect developmental disorder or a global developmental delay (American Psychiatric
Association, 2013). Although initially considered as an abnormal condition, several works
have showed how the distribution of autistic traits across the population is continuous (Wing,
1988; Constantino and Todd, 2003), and people diagnosed with autism score at the extreme
end of the distribution when autistic traits are measured (Baron-Cohen, 2010).
The Autism-Spectrum Quotient (AQ) is a widely used self-administered questionnaire
that provides a quantified evaluation of the degree to which an adult with a normal intelli-
gence quotient presents with signs of ASD. This measurement tool assesses the respondent’s
behaviors, preferences and cognition based on five domains of autistic functioning: social
skill, attention switching, attention to detail, communication, and imagination; autistic traits
are considered when the respondent shows poor social skill, strong attention focus, excep-
tional attention to detail, poor communication and poor imagination (Baron-Cohen et al.,
2001). The AQ was translated in several languages, such as Italian (Ruta et al., 2012),
Dutch (Hoekstra et al., 2008), Chinese (Lau et al., 2013), Japanese (Wakabayashi et al.,
2006), Turkish (Kose et al., 2013), Polish (Pisula et al., 2013), and Persian (Mohammadi
et al., 2012), and French (Kempenaers et al., 2017), and its structural validity was studied
using exploratory and confirmatory factor analysis (EFA and CFA) in the different trans-
lations. Although initial evidence supporting the structural validity of AQ, several studies
reported doubts regarding the factorial aspects of the tool (Hoekstra et al., 2011; Hurst et al.,
2007), with multiple works not obtaining the five-factor original model to fit in student data
(Hoekstra et al., 2011; Kloosterman et al., 2011; Lau et al., 2013); most of which argued for
shorter versions of the questionnaire. An abridged ten-item version of the AQ was therefore
162
created including “red flag” items for autism screening in general practice (Allison et al.,
2012). Despite criticism, the AQ is the most commonly used measurement tool for autistic
traits detection and is widely cited in scientific research (Ruzich et al., 2015).
The AQ was conceived based on the assumption that autistic traits are measurable con-
sequences of an underlying cause – that is, ASD. There is growing evidence pointing towards
defects in neuronal migration and the consequent malformation and malfunction of various
brain circuits as etiology of several brain disorders such as ASD (Pan et al., 2019), which
then presents itself through several autistic traits; as understood in the latent variable model
of mental illness, however, autistic traits do not actively participate to the clinical manifesta-
tion in subjects, as they are effects of the construct they stem from. Recent works, however,
have shown how behaviors in the autism spectrum reinforce each other (i.e. present a mutual
influence): learning strategies can therefore be adopted as to correct the social behavior of
people with autistic traits, and brain responses to social stimuli can be altered, especially
if an early intervention is conducted (Schuetze et al., 2017). From an ontological point of
view, this changes the attribute of autistic traits, that evolve from measurable consequences
of ASD to participate in its clinical presentation.
In the last decade, network analysis has affirmed itself as a new way of analyzing data in
psychiatry and psychology, which allows the conception of of mental disorders or constructs
as emerging from a complex system of mutually influencing components (Borsboom and
Cramer, 2013). This novel theoretical and methodological approach has been widely used
to explore a variety of constructs including depression (Mullarkey et al., 2018), obsessive
compulsive disorder (McNally et al., 2017), empathy (Briganti et al., 2018), personality
(Costantini et al., 2015), alexithymia (Briganti and Linkowski, 2019c), self-worth (Briganti
et al., 2019), posttraumatic stress disorder (Fried et al., 2018; Phillips et al., 2018), resilience
(Fritz et al., 2018; Briganti and Linkowski, 2019b), and narcissism (Di Pierro et al., 2019;
Briganti and Linkowski, 2019a).
Researchers usually analyze mental disorders and constructs as network composed of
163
items – answers of the observed group to a given questionnaire, such as the AQ. Conceiving
mental disorders as networks is interesting in clinical practice, since relevant components
in the proposed model could serve as a target for intervention (Fried et al., 2018). Most
of the recent empirical works using network analysis compute the unobserved interactions
between psychological components as regularized partial correlations (Epskamp and Fried,
2018): this is commonly achieved with `1-regularization (Friedman et al., 2014b), which
is also known as the “least absolute shrinkage and selection operator” or “lasso”, which
pushes the smaller estimates in the network structure to zero and therefore renders a sparse
(or conservative) network. The dominant lasso procedure used in network papers is the
graphical lasso that associates the `1-regularization with an Extended Bayesian Information
Criterion to determine a tuning parameter for the Gaussian Graphical Model (i.e. the
network) λ (Chen and Chen, 2008). However, recent work (see for example Williams et al.,
2019) reported how the regularization of network estimates can be inconsistent for model
selection, and does not provide evidence for the null hypothesis (i.e. evidence for no effect);
the latter consequence is of great importance because one of the core ideas of estimating
network structure is to uncover the conditional independence structure of a construct, and
therefore zero partial correlations among the variables composing the network.
Recently, Bayesian methodology has been introduced for the estimation of network struc-
tures. This new methodological framework allows for the estimation of Gaussian Graphical
Models with posterior probabilities that can assess the conditional dependent and indepen-
dent relations among components of a network with a decision rule that can be calibrated to
a desired level of specificity (Williams, 2018a). The Bayesian approach to the estimation of
network structures shows advantages when determining conditional independence relations
based on a network estimation where the threshold for selection depends on Bayes factors
(Williams and Mulder, 2019): this technique allows for studying the amount of evidence for
no effect in a given structure, which is useful to assess the uncertainty of estimates. Despite
offering such advantages, Bayesian methodology for the estimation of Gaussian Graphical
164
Models has never been used in empirical works concerning network analysis in psychological
or psychiatric research.
Inspired by these new advances in network analysis that has been applied in many fields
of psychiatry and psychology, I apply the network methodology to autistic traits. The aim of
the present work is therefore to perform a network analysis of the autistic traits as presented
in the 50-item version of the AQ with the Bayesian inference methods that have been recently
introduced (Williams, 2018a), in a sample of 649 university students.
Autism has recently been analyzed through the lens of network analysis in relationship
with depression (van Heijst et al., 2019); however, to my knowledge, this is the first time
that the AQ in its full 50-item version is explored with the network methodology, therefore
expanding this conceptual and methodological framework to the construct of autistic traits.
In this work, I will first estimate the conditional dependence structure of the autistic traits
network. Second, I will estimate node predictability, i.e. point-estimate and confidence
intervals for variance explained in all items, to measure their overall connectedness to other
items in the questionnaire. Third, I will explore sex differences among female and male
subjects from this study. The protocol for this study was approved by the ethical committee
of Hopital Erasme (Erasme Teaching Hospital) in Brussels, Belgium.
10.2 Method
10.2.1 Data set
My data set is composed of 649 university students from the French-speaking region of
Belgium. Subjects were 17 to 25 years old (M = 19.3, SD= ± 1.49); 58% were female and
42% were male.
165
10.2.2 Measurement
The AQ is composed of 50 items that assess autistic traits based on five domains: social skill,
attention switching, attention to detail, communication, and imagination. The minimum
score for each item is 1 (“Definitely disagree”) and the maximum score is 4 (“Definitely
agree”); approximately half of the items in the questionnaire are to be reverse-scored (Baron-
Cohen et al., 2001).
10.2.3 Network analysis
Software and packages
I used the software R for statistical computing (open source, available at https://www.
r-project.org/). The package used to carry out the analysis is BGGM (Williams and
Mulder, 2019).
Network estimation
I estimated a Gaussian Graphical Model (GGM) with Bayesian methods (Williams, 2018a),
that is, a partial correlation network for the 50 items in the AQ. The GGM is calculated as
the inverse-covariance matrix: it is a network that includes a set of nodes that correspond
to the autistic traits in the AQ and a set of edges that connect the nodes in the network. If
two nodes are connected, that means they are conditionally dependent given all other nodes
in the network (i.e their partial correlation is nonzero). In the network of autistic traits, if
two nodes A and B are connected, it means for instance that if the observed group scored
high on trait A, then the observed group is also more likely to score high on trait B, and
vice versa, controlling for other nodes in the network (Briganti et al., 2018). Each edge
in the network has a weight representing the strength of association between two autistic
traits; edges can be positive (and therefore represent a positive association) or negative
(denoting a negative association). The estimation of network with Bayesian methods allows
166
for providing evidence for the hypothesis that best predicts the observed data. For instance,
when testing for conditional dependence relationships among nodes, providing a Bayes Factor
(BF) between 3 and 20 as a cut-off value is associated with positive evidence, while a BF >
20 is associated with strong evidence (Kass and Raftery, 1995). For the network estimation
in this paper, a BF of 20 was used.
Node predictability and sex differences
I computed Bayesian R2 for the fifty autistic traits in the network, which represents the
percentage of variance explained of a given autistic trait with all other autistic traits in the
network (Williams and Mulder, 2019): this measure is commonly defined “node predictabil-
ity” in the network literature (Haslbeck and Fried, 2017) as it can be interpreted as how
well a node connects to other nodes in the network, or the self-determination of the network.
Node predictability can be understood as the upper bound of controllability: if one assumes
that all edges for a given node are directed toward that node, predictability provides an
estimate of how much influence I can have on the given node via all other nodes (Briganti
et al., 2019). Because the estimation of node predictability results in a distribution, I com-
pared the predictability estimates of females and male subjects in my data set to detect sex
differences in the network structures.
10.3 Results
10.3.1 Partial correlation network
The autistic traits partial correlation network is reported in figure 10.1. The strongest
connections are found between items that measure the same (or similar) autistic traits: for
example item 38 (“I am good at social chit-chat”, reversed) is strongly connected to item 17
(“I enjoy social chit-chat”, reversed) and item 26 (“I frequently find that I don’t know how
to keep a conversation going“). Two other examples of this phenomenon can be found in
167
1 23
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
2324
25262728
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
4849
50
Figure 10.1: Autistic traits (N = 649) partial correlation network with the 50 items from theAQ. Each edge (connection between nodes) is denoted by a weight represented by thicknessand color saturation.
168
the connections between item 8 (“When I’m reading a story, I can easily imagine what the
characters might look like”, reversed) and 20 (“When I’m reading a story, I find it difficult to
work out the characters’ intentions”) as well as between item 9 (“I am fascinated by dates”)
and 19 (“I am fascinated by numbers”).
However, connections between items that measure different autistic traits can also be
found in the network: for instance, 27 (“I find it easy to read between the lines when
someone is talking to me”, reversed) and 36 (“I find it easy to work out what someone
is thinking or feeling just by looking at their face”, reversed), 11 (“I find social situations
easy”, reversed) and 38 (“I am good at social chit-chat”, reversed), 35 (“I am often the last
to understand the point of a joke”) and 46 (“New situations make me anxious“).
Several negative connections can also be found in the network, such as the ones between
items 18 (“When I talk, it isn’t always easy for others to get a word in edgeways”) and
32 (“I find it easy to do more than one thing at once”, reversed), 21 (“I don’t particularly
enjoy reading fiction”, reversed) and 49 (“I am not very good at remembering people’s date
of birth”, reversed), 12 (“I tend to notice details that others do not”, reversed) and 31 (“I
know how to tell if someone listening to me is getting bored”).
10.3.2 Node predictability
The computed values for node predictability in the network are reported in figure 10.2. Item
38 (“I am good at social chit-chat”, reversed) has the highest R2 score (54%) in the network,
while item 41 (“I like to collect information about categories of things”) has the lowest R2
score (9%) in the network. The average R2 score is 20%.
10.3.3 Sex differences
Figure 10.3 represents sex differences in node predictability. The autistic trait network struc-
tures are statistically different in female and male subjects, with a Kullback-Leibler diver-
gence value (that translates the difference between the two node predictability distributions)
169
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
4124484
154237133
1623311
33352843467
50342040212
1130271429106
39228
442532185
453649129
1947261738
0.2 0.4 0.6bayes_R2
Nod
epost.pred
Figure 10.2: Node predictability (Bayesian R2) for the 50 nodes in the partial correlationnetwork.
●Yg1 vs Yg2
4.0 4.5 5.0 5.5Predictive Check
Con
tras
t
Figure 10.3: Sex differences in the network of autistic traits.
170
5049484746454443424140393837363534333231302928272625242322212019181716151413121110987654321
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50Alternative Hypothesis Matrix
1.10
5.57log(BF)
Figure 10.4: Heat map reporting the edges that are statistically different in regard to thesex of participants. The darker the color, the higher the evidence.
of 5.8, and a p-value of 0.
A heat map is reproduced in figure 10.4 that represents which specific edges are different
in the female and male network structure (and the respective amount of evidence supporting
the difference, represented by a variation in BF). For instance, there is a moderate amount
of evidence supporting sex differences for the specific edges between items 14 (“I find making
up stories easy”) and 22 (“I find it hard to make new friends“), with a Bayes Factor of 115,
item 17 (“I enjoy social chit-chat”) and 34 (“I enjoy doing things spontaneously”), with a
Bayes Factor of 90.
The number of edges that are statistically different in regard to the sex of participants
variates among the items in the AQ and are reported in figure 10.5. Item 17 (“I enjoy social
chit-chat”) presents the highest number of edges in that regard.
171
●
● ●
● ● ●
● ●
●
● ●
●
●
●
● ●
●
●
●
●
●
● ●
●
●
●
● ●
●
●
●
●
● ●
● ●
●
●
●
●
●
●
●
● ● ●
●
●
●
●
0
1
2
3
4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50Item
Diff
eren
ces
Figure 10.5: Plot reporting for each node the number of edges that are statistically differentin regard to the sex of participants.
172
10.4 Discussion
This is the first work tackling the exploration of autistic traits through the lenses of Bayesian
network analysis and using the full 50-item version of the largely validated AQ (Baron-Cohen
et al., 2001). New methods introduced in recent Bayesian network literature (Williams,
2018a) allow for the estimation of conditional relationships among nodes with Bayesian
methods, including adopting a cut-off value to provide consistent evidence of the existence
of such relationships. Studying how items from the AQ connect sheds new information
on the important construct such as the one representing autistic traits. Even in a student
sample, investigating autistic traits through the lenses of network analysis is relevant because
“non-autistic” individuals may still present autistic traits, which are continuously distributed
across the general population.
The strongest connections in the network are found between items that measure similar
autistic traits: this is a recurrent phenomenon in network analysis of psychometric scales
(Briganti et al., 2018, 2019; Briganti and Linkowski, 2019b,a), which substantially influences
the network predictability.
Additional insight to the construct is mostly brought by connections between items that
measure different autistic traits, which are also found in the network, but their connection
strength is proportionally lower. From a network perspective, a connection between different
autistic traits such as the one between item 35 (“I am often the last to understand the point
of a joke”) and 46 (“New situations make me anxious”) is interesting if interpreted on a
clinical level. For instance, if a patient with autistic traits is bothered that he often lacks
the ability to understand jokes, it might be interesting for the clinician to investigate if and
how new situations make that same patient anxious, while taking into account other autistic
traits.
The presence of several negative connections in the network of autistic traits is an inter-
esting finding: the model is not overall positively connected, which is unusual from a latent
variable model perspective in which all items are considered as interchangeable measures of
173
the latent variable (in this case, the autism spectrum). This means several autistic traits
are inversely correlated. Negative connections occur between items from different domains
(i.e. different facets of the autism spectrum). For instance and in my sample, the average
individual presenting the autistic trait of being good at remembering people’s date of birth
(item 49) is less likely to present the autistic trait of enjoying reading fiction (item 21).
The highest R2 score in the network belongs to item 38, which measures an autistic
trait related to social skills. From a network point of view this can be interpreted as the
social skills trait being the domain that best predicts (or is predicted by) the rest of autistic
traits, since it is the most interconnected trait. On the other hand, item 41, an autistic trait
related to the domain of imagination, has the lowest R2 score in the network, which means it
is poorly connected to other autistic traits. The average R2 score is 20%, which means that
on average, nodes in the network of autistic traits share 20% of variance with other nodes.
Sex differences are found between the network of female and male subjects. This finding
supports extensive research done in the past few years which points to high AQ scores being
preponderant in male subjects (Ferri et al., 2018; Baron-Cohen et al., 2014); this study
however reports a difference in network connectivity, which is an additional information
never explored before.
My findings must be interpreted in light of certain limitations: two of them are hereby
described. First, the data set used in this study is composed of Belgian university students,
which likely limits the replicability of my findings in other sample, including samples with
people diagnosed with autism. Second, the uncertainty regarding the true network structure
of autistic traits must be taken into account: whereas it is possible to detect edges (non-zero
effects) with large uncertainty, inferring null effects (zero effects or the lack of connection
between two nodes) requires a much greater sample, especially in my case, where 50 items
are analyzed.
Further studies may endeavor to replicate my findings in other samples, including people
diagnosed with autism.
174
Chapter 11
A network model of depressive
symptoms in a student sample
Abstract
The Self-rating Depression Scale (SDS) is a psychometric tool composed of 20 items
used to assess depression symptoms. The aim of this work is to perform a network
analysis of this scale in a large sample composed of 1090 French-speaking Belgian uni-
versity students. I estimated a regularized partial correlation network and a Directed
Acyclic Graph for the 20 items of the questionnaire. Node predictability (shared vari-
ance with surrounding nodes in the network) was used to assess the connectivity of
items. The network comparison test was performed to compare networks from female
and male students. The network composed of items from the SDS is overall positively
connected, although node connectivity varies. Item 11 (“My mind is as clear as it used
to be”) is the most interconnected item. Networks from female and male students did
not differ. DAG reported directed edges among items. Network analysis is a useful tool
to explore depression symptoms and offers new insight as to how they interact. Further
studies may endeavor to replicate my findings in different samples, including clinical
samples to replicate the network structures and determine possible viable targets for
clinical intervention.
176
11.1 Introduction
Depression (also known as Major Depressive Disorder, MDD) is a common psychiatric disor-
der (Goldberg, 2011), and the top cause of disability worldwide (Lopez et al., 2006). In the
fifth edition of the Diagnostic and Statistical Manual of Mental Disorders or DSM (American
Psychiatric Association, 2013), depression is characterized as a combination of five symptoms
in a list of nine, including depressed mood or anhedonia (this symptom must be included in
each combination), increase or decrease in weight or appetite, fatigue, sleep problems and
suicidal ideation. Depression is then defined as a heterogenous disorder, which leads to an
important symptom variability (Fried and Nesse, 2015).
In most depression scales, such as the Hamilton Rating Scale for Depression (HRSD),
sum scores are used to explain the severity of the disorder (Hamilton, 1960). Zung created
in 1965 a self-rating depression scale (SDS) using 20 items representing symptoms from
three conceptual domains (Zung, 1965): pervasive effects (e.g. item 1: “I feel downhearted
and blue”), physiological equivalents (e.g. item 4: I have trouble sleeping at night”) and
psychological equivalents (e.g. item 19: “I feel that others would be better off if I were
dead”). In both these approaches, symptoms are considered as interchangeable measures of
depression.
In recent years, the network approach to psychopathology has been proposed as a novel
way of analyzing mental disorders as complex systems. This approach considers that a men-
tal disorder arises from connections among its symptoms, that can cause each other (Bors-
boom, 2017): this conceptualization differs from the latent variable theory, which states that
symptoms are passive consequences of an underlying cause. In psychiatry, the latter is un-
likely to identify a specific mechanism that causes a mental disease (Borsboom, 2008), which
makes the network approach an interesting alternative. Five principles of network theory
proposed by Borsboom are likely to apply in the case of depression: complexity (mental
disorders are characterized as interactions between the components of a network, such as
symptoms), symptom-component correspondence (components in a network are problems
177
defined as symptoms), direct causal connections (symptoms cause each other in a network),
network structure of mental disorders (certain symptoms are more connected than others
and symptoms from the same network often arise together) and hysteresis (symptoms acti-
vate each other even after the trigger cause has disappeared). Hysteresis has been proven
to be a dynamic present in depression (Cramer et al., 2016): a dormant (asymptomatic)
network of depression is present in the healthy individual: a trigger cause may activate one
or more symptoms which, in turn, activate the rest of the network in a “pathological” con-
dition. Mental health can then be defined as a stable state of a weakly connected network
(Borsboom, 2017).
Network analysis is a set of statistical techniques developed in the conceptual framework
of network theory to identify network structures in datasets. Network analysis has been
used in various fields, such as empathy (Briganti et al., 2018), posttraumatic stress disorder
(Fried et al., 2018), schizophrenia (Galderisi et al., 2018) and self-worth (Briganti et al.,
2019). Depression symptoms have already been analyzed with network analysis and several
networks structures have been proposed in the literature (Beard et al., 2016; McNally et al.,
2017).
Researchers usually analyze constructs as undirected networks composed of nodes (symp-
toms) and undirected connections representing regularized partial correlations (Epskamp and
Fried, 2018). However, in the case of depression, identifying a causal pathway might help
to gather information about the sequence of symptom activation in a previously healthy
individual. Such causal pathways can be identified in Bayesian networks and more precisely
with Directed Acyclic Graphs (DAGs), which are probabilistic models. DAGs are capable
of learning the underlying causal graphs from the data (Moffa et al., 2017). DAGs are well
established in the network literature (Scutari and Denis, 2015) and have been recently used
to explore depression (McNally et al., 2017).
My work therefore aims to build on previous papers and apply network modeling to
depression symptoms as described in the SDS using both directed and undirected network
178
models. First, I will estimate a regularized partial correlation network composed of symptoms
from the SDS and explore its overall connectivity. Second, I will explore the connectedness of
items from the SDS with well-established centrality measures (Briganti et al., 2018). Third, I
will explore causal pathways between SDS symptoms using DAG constraint-based structure
learning algorithms.
11.2 Method
11.2.1 Participants
This study is carried out on a dataset composed of 1090 French-speaking Belgian university
students (54% females, 46% males), aged 17 to 25 years old (M=20; SD=2).
11.2.2 Measurement
The Self-Rating Depression Scale (SDS) is composed of 20 items (Table 11.1) meant to assess
depression symptoms. Item scores range from 1 (“a little of the time) to 4 (“most of the
time”). Items are presented in their original order in the questionnaire. Certain items are
reverse scored (items 2, 5, 6, 11, 12, 14, 16, 17, 18 and 20), such as item 6 (“I still enjoy
sex”).
Table 11.1: The Self-Rating Depression Scale (Zung,
1965)
N Item
1 I feel down hearted and blue
2 Morning is when I feel the best
3 I have crying spells or feel like it
4 I have trouble sleeping at night
179
5 I eat as much as I used to
6 I still enjoy sex
7 I notice that I am losing weight
8 I have trouble with constipation
9 My heart beats faster than usual
10 I get tired for no reason
11 My mind is as clear as it used to be
12 I find it easy to do the things that I used to
13 I am restless and cannot keep still
14 I feel hopeful about the future
15 I am more irritable than usual
16 I find it easy to make decisions
17 I feel that I am useful and needed
18 My life is pretty full
19 I feel that others would be better off if I were dead
20 I still enjoy the things I used to
11.2.3 Network analysis
Network analysis is carried out in R (open source, available at https://www.r-project.
org/). Packages used to carry out the analysis include qgraph (Epskamp et al., 2012) and
glasso (Friedman et al., 2014b) for network estimation and visualization, mgm for node
predictability (Haslbeck and Waldorp, 2016), bootnet (Epskamp and Fried, 2018) for stabil-
ity, NetworkComparisonTest (van Borkulo et al., 2014), bnlearn (Scutari, 2010) and pcalg
(Kalisch et al., 2012) for the estimation of directed acyclic graphs.
180
Regularized partial correlation network
Network estimation A network is composed of nodes (representing items from the SDS)
and edges (connections between items). A correlation matrix is used as input to estimate a
Gaussian Graphical Model (GGM), which is a regularized partial correlation network. Each
edge has a weight (parameter resulting from the GGM) which is regularized using a graphical
lasso (least absolute shrinkage and selection operator) to avoid the estimation of spurious
edges and therefore provides a conservative model (Epskamp and Fried, 2018). An edge
therefore represents a regularized partial correlation between two symptoms, controlling for
all other symptoms in the network. The thickness and color saturation of an edge represent
its weight (the strength of the association between two nodes); edges can therefore be positive
(blue) or negative (red). In the case of the SDS, a self-report scale which measures depression
symptoms, a positive edge can be statistically interpreted as following: if two given nodes
X and Y share an edge XY in the network, and the observed group of subjects scores high
on X, then the observed group is also more likely to score high on Y (Briganti et al., 2018).
On the other hand, a negative edge implies that if an observed group of subjects scores
high on X, then the observed group is less likely to score high on Y. Two nodes will be
disconnected if they are conditionally independent. Nodes are placed in the network by the
Fruchterman-Reingold algorithm, based on the sum of connections a given node has with
other nodes (Fruchterman and Reingold, 1991).
Network inference Node predictability is an absolute measure of the interconnectedness
of a node (Haslbeck and Waldorp, 2016) and is the percentage of variance of a given node
explained by surrounding nodes. Node predictability has been described as the upper bound
of controllability: “if one assumes that all edges for node A are directed towards that node,
predictability provides an estimate of how much influence I can have on A via all other
nodes” (Briganti et al., 2019). Node predictability is represented in the network as a pie
chart surrounding the node.
181
Network Comparison Test I performed a Network Comparison Test (NCT) to compare
networks from female and male subjects in my dataset (van Borkulo et al., 2016). The NCT
rearranges the samples to test whether two networks are invariant with respect to global
strength (sum of all edge weights), network structure and edge values.
Network stability State-of-the-art stability analyses (Epskamp and Fried, 2018) were
carried out using the same methodology used in my previous studies (Briganti et al., 2018).
An edge weight difference test was performed to answer the question “is edge A significantly
stronger than edge B?”.
Directed Acyclic Graph
In Bayesian networks an edge may represent a causal pathway between two nodes. The
structure of a Bayesian network can be estimated using constraint-based algorithms, that
analyze conditional independence relations among the nodes in the network. Constraint-
based algorithms generate a network model that can be interpreted as a causal model even
in observational data, under assumptions that in clinical terms exclude confounding and
sampling bias.
In this paper I used the PC algorithm, a constraint-based algorithm (Spirtes et al.,
1993). To estimate a network model, the PC algorithm first estimates an undirected network
model in which all pairs of nodes are connected, then deletes edges between conditional
independent pairs of variables, and directs edges starting with v-structures (two disconnected
nodes causing a third node).
The estimated network was investigated using stability analysis through bootstrapping.
Afterwards, a network is reported with a minimum connection strength (% of fitted networks
in which a given connection appears) of 85 and a minimum connection direction (% of
fitted networks in which a given connection has a given direction) of 50. This resulting
network therefore reports connections that are presents in more than 85% of fitted networks;
182
Blue
Morning
Crying
SleepP
Eat
Sex
Weight
Constipated
HeartR
Tired
Mind
Things
Restless
Hope
Irritable
Decision
Useful
LifeFull
Dead
Enjoy
Blue: I feel down hearted and blueMorning: Morning is when I feel the bestCrying: I have crying spells or feel like itSleepP: I have trouble sleeping at nightEat: I eat as much as I used toSex: I still enjoy sexWeight: I notice that I am losing weightConstipated: I have trouble with constipationHeartR: My heart beats faster than usualTired: I get tired for no reasonMind: My mind is as clear as it used to beThings: I find it easy to do the things I used toRestless: I am restless and cannot keep stillHope: I feel hopeful about the futureIrritable: I am more irritable than usualDecision: I find it easy to make decisionsUseful: I feel that I am useful and neededLifeFull: My life is pretty fullDead: I feel that others would be better off if I were deadEnjoy: I still enjoy the things I used to do
Figure 11.1: Regularized partial correlation network of items from the SDS. Each noderepresents an item from the SDS; positive connections are blue, negative connections arered. The pie chart surrounding each node represents node predictability.
moreover, these connections present a direction (for instance, from node A to node B) which
is found in more than half of the fitted networks resulting from the bootstrapping procedure.
11.3 Results
11.3.1 Regularized partial correlation network
Figure 11.1 represents the glasso network of items from the SDS. The network is overall
positively connected. The strongest connection in the network (0.26) is found between items
17 (“I feel that I am useful and needed”) and 18 (“My life is pretty full”). Other strong
connections include edge 1-3 (“I feel downhearted and blue”; “I have crying spells or feel
like it”), edge 11-12 (“My mind is as clear as it used to be”; “I find it easy to do the things
I used to”) and edge 16-17 (“I find it easy to make decisions”; “I feel that I am useful and
needed”). Items 2 (“Morning is when I feel best”) and 8 (“I have trouble with constipation”)
are poorly connected with the rest of the nodes in the network. One negative connection is
183
found between item 7 (“I notice that I am losing weight”) and 17 (“I feel that I am useful
and needed”).
Network inference
Mean node predictability is 0.23, which means that on average items from the SDS share
23% of variance with surrounding items. The most predictable node is item 11 (“My mind
is not as clear as it used to be”), which shares 38% of variance with surrounding items. The
least predictable node is item 7 (“I notice that I am losing weight”) and it shares 10% with
surrounding items.
Network Comparison Test
The networks estimated from both female and male students were substantially similar and
did not differ statistically regarding global strength, network structure and edge values.
Network stability
The edge weight difference test reports that the strongest edges (17-18; 1-3; 11-12; 16-17; 13-
15; 9-10) are significantly stronger than other edges in the network but are not significantly
different from each other; therefore, I cannot safely interpret which edge is the strongest.
11.3.2 DAG
Figure 11.2 shows the DAG estimated with PC algorithm and reporting stable connections
resulting from bootstrapping. The DAG shows some interesting directed connections; I will
describe a few examples of such edges. Item 1 (“I feel downhearted and blue”) has two
outgoing edges to items 9 (“My heart beats faster than usual”) and 14 (“I feel hopeful about
the future”). Item 2 (“Morning is when I feel best”) also has two outgoing edges to items
14 (“I feel hopeful about the future”) and 19 (“I feel that others would be better off if I
were dead”). Item 4 (“I have trouble sleeping at night”) has three outgoing edges to items
184
Blue
Morning
Crying
SleepP
Eat
Sex
Weight
Constipated
HeartR
Tired
Mind
Things
Restless
HopeIrritable
Decision
Useful
LifeFull
Dead
Enjoy
Blue: I feel down hearted and blueMorning: Morning is when I feel the bestCrying: I have crying spells or feel like itSleepP: I have trouble sleeping at nightEat: I eat as much as I used toSex: I still enjoy sexWeight: I notice that I am losing weightConstipated: I have trouble with constipationHeartR: My heart beats faster than usualTired: I get tired for no reasonMind: My mind is as clear as it used to beThings: I find it easy to do the things I used toRestless: I am restless and cannot keep stillHope: I feel hopeful about the futureIrritable: I am more irritable than usualDecision: I find it easy to make decisionsUseful: I feel that I am useful and neededLifeFull: My life is pretty fullDead: I feel that others would be better off if I were deadEnjoy: I still enjoy the things I used to do
Figure 11.2: Directed Acyclic Graph.
12 (“I find it easy to do the things I used to”), 16 (“I find it easy to make decisions”) and
17 (“I feel that I am useful and needed”). Item 15 (“I am more irritable than usual”) has
three incoming connections from items 2 (“Morning is when I feel the best”), 11 (“My mind
is as clear as it used to be”), and 17 (“I feel that I am useful and needed”). Item 19 (“I feel
that others would be better off if I were dead”) has two incoming connections from items
2 (“Morning is when I feel best”) and 16 (“I find it easy to make decisions”). Regression
coefficients for nodes in the DAG are positive, which means that nodes from the DAG are
overall positively connected.
11.4 Discussion
This is to my knowledge the first application of network analysis to the SDS. This study,
conducted in a non-clinical large sample of French-speaking university students, sheds light
on the network structure of depression items, which may replicate in clinical samples. The
regularized partial correlation network shows that items from the SDS are overall positively
connected, but the degree of connectivity of a given node from the SDS to other nodes (also
185
defined as node degree in network science) varies; for instance, I identified two items from
the SDS (2 and 8) which are overall poorly connected to the rest of the network.
From a network point of view, disconnected items may be considered to be independent
controlling for other nodes in the network. Some connections are stronger than others as
reported in the stability analyses. Some strong edges however are shared between two items
that might measure the same symptom, such as edge 1-3 (“I feel downhearted and blue”;
“I have crying spells or feel like it”); in this case, this connection should be interpreted as
shared variance between the two items (Fried and Cramer, 2017). In most cases in this
network, however, connections offer novel insight as to how symptoms interact; I will detail
one example to show such insight offered by my study. For instance, the strongest connection
in the network between items 17 (“I feel that I am useful and needed”) and 18 (“My life is
pretty full”) may be interpreted as such: at the level of the observed group and controlling
for other nodes in the network, it is more likely not to feel useful and needed if one’s life
is empty, and vice-versa; if this same edge were to replicate at the individual network of a
person diagnosed with depression, intervening on one’s feeling of having a “full life” may
alleviate the feeling of being useless, and vice-versa.
Node predictability was used to assess the connectivity of items from the SDS. Item 11
(“My mind is not as clear as it used to be”) shows the highest node predictability value, which
means it shares the most variance with surrounding nodes. As reported in the DAG structure,
this item receives many incoming edges from different surrounding items; I may therefore
consider item 11 as the most controllable item in the network. Item 11 may therefore not
be a viable target for clinical intervention if a similar network structure were to replicate in
a clinical sample, since it may be considered as the consequence of surrounding symptoms.
Future studies may also endeavor to translate the causal meaning of node predictability in
other psychometric tools. The DAG offers additional insight as to how symptoms from the
SDS may cause each other. I described several symptoms that are mainly causing other
symptoms in the network, such as item 1 (“I feel downhearted and blue”), item 2 (“Morning
186
is when I feel best”) and item 4 (“I have trouble sleeping at night”). These symptoms might
be considered as viable targets for clinical intervention if the same network structure were
to replicate in clinical samples.
The two network models proposed in this paper present similarities as well as differences:
for instance, both the regularized partial correlation network and the DAG contain positive
edges; however, nodes that connect well in the regularized partial correlation network do not
connect in the DAG such as the redundant items discussed in the results. The reason the
different definitions of the two models: in a DAG two nodes are not automatically connected
when they share a common child node, but they will be connected in the corresponding
partial correlation network because of the indirect dependence conditional on that child
node.
My results must be interpreted in light of several limitations. First, my dataset is com-
posed of university students, which may limit the generalization of my findings to different
samples. Second, DAG structures do not involve loops: if in a three-node network a symptom
A causes symptom B and a symptom C, symptom C cannot cause symptom A (the struc-
ture is therefore acyclic); however, it is plausible to consider that in the case of depression
symptoms, certain symptoms may activate each other in a loop. Third, causation may be
inferred from a connection between two nodes in a DAG assuming there are no confounding
or sampling bias.
Further studies may endeavor to replicate my findings in different samples, both non-
clinical and clinical, to replicate network structure (both Bayesian and non-Bayesian) of
symptoms described in the SDS, and look for similarities as well as differences between
different models of depression components.
187
Chapter 12
A network model of mania
Abstract
The aim of this study is to explore mania as a network of its symptoms, inspired by
the network approach to mental disorders. Network structures of both cross-sectional
and temporal effects were measured at three time points (admittance, middle of hos-
pital stay, discharge) in a sample of 100 involuntarily committed patients diagnosed
with bipolar I disorder with severe manic features and hospitalised in a specialised
psychiatric ward. Elevated mood is the most interconnected symptom in the network
at admittance, while aggressive behavior and irritability are highly predictive of each
other, as well as language-thought disorder and delusions. Elevated mood influences
many symptoms in the temporal network. Network analysis is a useful tool to model
and explore the interconnectedness and relative importance of manic symptoms, as
well as monitor their evolution over time in patients under treatment.
12.1 Introduction
The word “mania” directly derives from two Greek words: “mania”, which can be translated
as “madness”, “enthusiasm”, or “passion”, and “mainomai”, which is the verb for “mania”
and translates the concepts “to be mad”, “to be furious”, and “to rage”. The notion of
188
mania can be traced back to Sophocles and his tragedy “Aias” first played almost 2500
years ago, based on the Homeric character Ajax from the Iliad and Odyssey. Ajax became
“mainomenos” (De Jong and Rijksbaron, 2006), meaning “berserk”, because he was cast a
spell from Athena, and killed a flock of sheep believing it was a group of enemy soldiers. 500
years later, Aretaeus of Cappadocia, a physician from Alexandria, first described mania as a
“worsening of melancholia”, therefore connecting the two states in the very first description
of bipolar disorder (Angst and Marneros, 2001).
Mania has come a long way since its first descriptions. In the fifth version of the Di-
agnostic and Statistical Manual of Mental Disorders (DSM V), a manic episode is defined
as a “distinct period of abnormally and persistently elevated, expansive or irritable mood
and abnormally and persistently increased goal-directed activity or energy, lasting at least
1 week and present most of the day, nearly every day (or any duration if hospitalization is
necessary)”, which may include a subset of three or more of symptoms including grandiosity,
decreased need for sleep, speech abnormalities, flight of ideas, distractibility, psychomotor
agitation, and involvement in activities with a high potential for painful consequences (e.g.,
engaging in unrestrained buying sprees, sexual indiscretions, or foolish business investments).
A manic episode causes marked impairment in social functioning and is not attributable to
the effects of a substance or another medical condition. (American Psychiatric Association,
2013). At least one manic episode is required to diagnose a bipolar I disorder. Correctly
diagnosing mania therefore has a meaningful impact on the treatment plan and follow-up for
the patient with bipolar I disorder: however, because of all the possible 5040 combinations
of the seven manic symptoms (and 35 different subsets of three symptoms out of seven) such
as described in the DSM V, diagnosis can be challenging for the psychiatrist.
The Young Mania Rating Scale (YMRS) is clinical tool for diagnosing mania and its
severity. It is composed of eleven features meant to assess different symptoms of mania
(Young et al., 1978), namely “Elevated Mood”, “Increased Motor Activity-Energy”, “Sexual
Interest”, “Sleep”, “Irritability”, “Speech (Rate and Amount)”, “Language-Thought Disor-
189
der”, “Content”, “Aggressive Behavior”, “Appearance”, and “Insight”. The items are used
to calculate a sum score, which is supposed to indicate the severity of the manic episode. The
YMRS is largely known and it has been validated in several languages, because it showed
good reliability and validity (Young et al., 1978).
Apart from reporting similar symptoms, the YMRS and the DSM V approach to mania
have another thing in common: they consider the symptoms of mania to be interchangeable
measures of the manic episode; in other words, they do not actively contribute to it, but
are rather consequences of it. From this categorical perspective of mental illness, manic
symptoms have no contributing role in the disease itself.
The categorical approach to mental disorders (such as the one used in the DSM V) has
been heavily criticized in recent years: for instance, it has been pointed out that it is unlikely
for one to be able to pinpoint a cause for a mental disorder that, when corrected, makes
the disorder disappear completely (as it would in a somatic disease); instead, psychiatric
symptoms are likely to influence each other (Kendler et al., 2011), and the mental disorder
arises from the set of interactions among the symptoms that are related to it (Borsboom,
2017). This approach is known as the network theory of mental disorders (Borsboom and
Cramer, 2013). Many fields in psychiatric research have translated network theory into em-
pirical studies for both behavioral constructs and mental disorders: for instance schizophrenia
(Galderisi et al., 2018), depression, (Mullarkey et al., 2018), posttraumatic stress disorder
(Fried et al., 2018; Phillips et al., 2018), autism (Ruzzano et al., 2015), empathy (Briganti
et al., 2018), self-worth (Briganti et al., 2019), resilience (Fritz et al., 2018; Briganti and
Linkowski, 2019b), alexithymia (Briganti and Linkowski, 2019c) and narcissism (Briganti
and Linkowski, 2019a). Predictors of lithium response (Scott et al., 2020) and relationships
between positive and negative affects (Curtiss et al., 2019) in bipolar disorder have been
studied with the network approach. The network approach comes with a set of statistical
methods called network analysis (Epskamp and Fried, 2018): mental disorders are usually
represented as partial correlation networks in which the network components (nodes) are
190
symptoms, and the connections (edges) are the partial correlations among symptoms. When
two nodes are connected in the network, the state of one node can be used to predict the
state of the other node, and vice-versa (Epskamp et al., 2017b).
It is reasonable to consider mania as a network of mutually influencing features, that
is, symptoms reinforce each other and determine the clinical presentation of the patient
(Borsboom, 2017). For instance, it is plausible to think that, when mood is elevated, the
patient is unable to sleep: insomnia could make the patient irritable, and the irritability
may in turn cause the patient to become aggressive. Although symptoms can influence each
other, it is also plausible to imagine that a given symptom can influence itself over time:
this may be the case for insomnia.
In network analysis, important nodes can be identified based on their quality to predict
(or be predicted by) other nodes in the network, and therefore their connectivity (Haslbeck
and Fried, 2017): such important nodes, which are customarily defined as “predictive” or
“predictable”, and could be considered as good targets for clinical intervention (Fried and
Cramer, 2017). Because these parameters can be determined, modeling mania as a network
of manic features can have considerable impacts on treatment decision: one could for instance
pharmacologically act upon one highly connected node to change the state of other nodes
that are connected to it.
In the case of a network approach to mania, it is reasonable to hypothesize that elevated
mood would be a highly connected node, because of the nature of bipolar I disorder itself:
however, it is also important to study what features of mania (besides the mood alteration)
are relatively important compared to others. Another research question that is important
to address is how network connectivity evolves in patients when they get pharmacological
treatment: in other words, “do symptoms show stronger connections before treatment or
after treatment?”.
While treating a severe manic episode and targeting specific features of mania (e.g. mood
stabilizers, neuroleptics) it is also important to know how the patient will respond: in other
191
terms, how do the states of given variables at time t0 predict the state of other variables at
time t1 (with the help of treatment). This principle is driven by Granger causality in panel
data (Wild et al., 2010): one variable in t0 predicts another in t1. The study of temporal
networks constitutes an important field in network analysis (Epskamp, 2019), along with
contemporaneous or cross-sectional networks, and can be applied in the case of mania. For
instance, it is interesting from a clinical perspective to investigate how stabilizing mood
has an effect on other features, or whether there are some symptoms that stay unaffected
after treatment (e.g. they reinforce each other or themselves). Inspired by what the network
approach can offer to the construct of mania, this study aims to explore mania as a network of
its symptoms as described in the YMRS (Young et al., 1978) in a sample of 100 involuntarily
committed patients hospitalized with severe bipolar I disorder. This work is organized as
follows: first, the cross-sectional network structures of mania at the start, middle and end
of the involuntary commitment are estimated. Second, node predictability (shared variance
of a given node with surrounding nodes) is studied as a measure of absolute connectivity
and therefore importance of symptoms in the network structure (Haslbeck and Waldorp,
2016). Third, the differences between the cross-sectional networks (start, middle and end of
commitment) are estimated regarding the strength of connections (van Borkulo et al., 2014).
Fourth, a temporal network is estimated between the time points to investigate temporal
effects.
12.2 Method
12.2.1 Ethical approval
This study was approved from the Ethical Committee of the Brugmann Teaching Hospital
in Brussels (CHU Bruxelles Brugmann).
192
12.2.2 Data sets
Participants
My data set is composed of 100 patients, hospitalized in the context of an involuntary
commitment in a secure psychiatric unit. To be included in this study, patients needed to
be diagnosed with bipolar I disorder and to present with a manic episode at admission.
An abnormal blood or urine analysis in regards to toxicology as well as a concomitant
personality or somatic disorder that could account for the presence of manic symptoms were
considered as exclusion criteria. All patients were involuntarily committed for a period which
lasted for 40 days on average (following the Belgian law for involuntary commitment).
All manic patients were treated with a standard set of drugs following the local protocol
for manic patients: in the first stage of treatment (when the clinical presentation is severe),
an association of a typical and atypical antipsychotic drugs is administered, with a mood
stabilizer as well as soporific and anxiolytic drugs when necessary and depending on the
symptoms presented. In the second stage of treatment (when the clinical presentation is
stable), the patient is left with an atypical antipsychotic, a mood stabilizer, as well as a
soporific and/or anxiolytic when necessary.
Three time points were collected for each patient: on admission (t0), halfway through the
commitment period t1, and on discharge (t2). The data sets were anonymized by default.
Patients were 20 to 72 years old (M = 44.5, SD = 14.5); 47 of them were female, and 53 of
them were male.
Measurement
The YMRS (Young et al., 1978) was used to assess manic symptoms, namely “Elevated
Mood”, “Increased Motor Activity-Energy”, “Sexual Interest”, “Sleep”, “Irritability”, “Speech
(Rate and Amount)”, “Language-Thought Disorder”, “Content”, “Aggressive Behavior”,
“Appearance”, and “Insight”. Symptoms were scored 0 to 4, depending on the severity of
193
Table 12.1: Symptoms from the Young Mania Rating Scales (Young et al., 1978)N symptom Symptom1 Elevated Mood2 Increased Motor Activity-Energy3 Sexual Interest4 Sleep5 Irritability6 Speech (Rate and Amount)7 Language-Thought Disorder8 Content9 Aggressive Behavior10 Appearance11 Insight
the clinical presentation, both at t0, t1 and t2. The symptoms are illustrated in table 12.1.
12.2.3 Network Analysis
Software
The software used for the analyses carried out in this study is R (version 3.6.3, available at
https://r-project.org). The packages needed for the analyses were bootnet (Epskamp
and Fried, 2018) and qgraph (Epskamp et al., 2012) for network estimation, visualization and
stability, psychonetrics (Epskamp, 2019) for temporal network estimation, mgm (Haslbeck
and Waldorp, 2016) for network inference, and NetworkComparisonTest (van Borkulo et al.,
2016) for comparing network structures at different measurement occasions.
Network estimation
Cross-sectional networks Let y be a a normal multivariate vector y = (y1, ..., yp) with
mean vector µ and a variance-covariance matrix Σ. For all subjects,
y ∼ N(µ,Σ). (12.1)
194
Let Θ be the inverse of Σ,
Θ = Σ−1 (12.2)
that is known as the precision matrix or a Gaussian Graphical Model (GGM). The
elements of of the GGM encode the partial correlation coefficients θij of two variables yi and
yj given all other variables in y, that is, y−(i,j):
Cor(yi, yj|y−(i,j)
)= − θij√
θii√θjj, (12.3)
therefore, the GGM represents the network itself.
The partial correlation θij between yi and yj is used as the edge weight, that is, the
strength of the connections between nodes Vi and Vj in the network. Edge weights can be
positive (usually represented as blue connections) or negative (usually represented as red
connections) depending on the sign of θij. The presence of an edge between two nodes Vi
and Vj in the network can be interpreted as a conditional dependence relationship: node
Vi predicts (or is predicted by) node Vj, after controlling for all other nodes in the network
V−(ij).
Three separate GGMs were estimated for t0, t1 and t2 data sets to study cross-sectional
effects among manic symptoms. For the estimation of the network structures, Spearman ρ
correlation was used as an input parameter because of the structure of the data (high scores
at t0, low scores at t2).
Further details about the GGM can be found in recent state of the art methodological
works (Epskamp et al., 2018).
Temporal network To model the dynamics of manic symptoms with a pharmacological
intervention in three time points t0, t1 and t2, a panel Graphical Vector Autoregressive
Model (GVAR). This model was first introduced in recent works with its own package for
computation (Epskamp, 2019) to translate time-series methods to panel data. GVAR can
195
be seen as a multivariate multiple regression on the previous measurement occasion.
For a set of symptoms y = (y1, ..., yp) measured in a given individual, GVAR is expressed
as
yt1 |yt0 ∼ N(µ+B(yt0 − µ),Σζ), (12.4)
where B represents a p × p matrix of temporal effects, µ the vector of means, Σt0 the
variance-covariance matrix on measurement occasion t0, and ζ a vector of normally dis-
tributed innovations. Because B encodes temporal prediction, a nonzero matrix element bij
means that yt1 is predicted by yt0 : this prediction is known as Granger causality (Granger,
1969) because the condition of “cause preceding the effect” is fulfilled. Non significant tem-
poral edges in the network are recursively removed by pruning (Pearl, 1984). More details
about GVAR can be found in the recent work that first introduces it in the psychonetrics
package (Epskamp, 2019).
Network inference
Node predictability was estimated for the 11 symptoms in the three data sets. Node pre-
dictability (Haslbeck and Fried, 2017) represents shared variance of a given node with sur-
rounding nodes in a network, that is, what percentage of variance of a given node can be
explained by other nodes. It can be interpreted as an absolute measure of how well a node
is connected in the network (Fried et al., 2018).
The Network Comparison Test (NCT) was performed to compare global strength in
the three networks (van Borkulo et al., 2016). Three tests were performed, respectively to
compare t0 and t1, t1 and t2, and t0 and t2. Because the samples were not independent, the
paired version of the NCT was used.
196
Network accuracy and stability
Accuracy analyses for the cross-sectional network structures were carried out following guide-
lines in network methodology (Epskamp and Fried, 2018). For the cross-sectional networks,
the accuracy of edge weights through bootstrapping (Epskamp et al., 2017a, I used 2000
bootstraps). I bootstrapped 95% confidence intervals of all edge weights (to answer the
question “is edge A accurately estimated?“), followed by an edge weight difference test to
see which edges differ from each other in size significantly (to answer the question “is edge
A significantly larger than edge B?”).
Similar stability analyses were carried out for the temporal network: a random subset
constituted of 10% of the data were dropped and the model was re-estimated.
12.3 Results
12.3.1 Cross-sectional networks
Table 12.2: Descriptive statistics for the eleven symptoms in the three time points.Mood0 Motor0 Sexual0 Sleep0 Irritable0 Speech0 LgTtAbn0 Content0 Aggressive0 Appearance0 Insight0
Mean 3.630 2.870 0.810 2.650 2.560 2.860 2.330 2.580 2.090 1.620 3.190Std. Deviation 0.787 1.203 1.440 1.218 1.282 1.198 1.264 1.304 1.712 1.556 1.412Minimum 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000Maximum 4.000 4.000 4.000 4.000 4.000 4.000 4.000 4.000 4.000 4.000 4.000
Mood1 Motor1 Sexual1 Sleep1 Irritable1 Speech1 LgTtAbn1 Content1 Aggressive1 Appearance1 Insight1
Mean 2.360 1.720 0.470 1.860 1.640 1.750 1.260 1.540 1.130 0.910 2.530Std. Deviation 0.644 0.766 0.870 0.853 0.772 0.702 0.799 0.809 0.884 0.889 1.185Minimum 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000Maximum 3.000 3.000 3.000 3.000 3.000 3.000 3.000 3.000 3.000 3.000 4.000
Mood2 Motor2 Sexual2 Sleep2 Irritable2 Speech2 LgTtAbn2 Content2 Aggressive2 Appearance2 Insight2
Mean 1.110 0.560 0.140 1.050 0.430 0.520 0.320 0.420 0.120 0.220 1.600Std. Deviation 0.737 0.625 0.427 0.757 0.573 0.627 0.548 0.622 0.409 0.524 1.356Minimum 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000Maximum 2.000 2.000 3.000 3.000 2.000 2.000 2.000 2.000 2.000 2.000 4.000
The three network structures of manic symptoms at t0, t1 and t2 are illustrated in figures
12.1, 12.2 and 12.3. For easier comparison, the three network structures are combined in
figure 12.4 (while only reporting edges with a weight greater than 0.1).
Overall, the networks present both positive and negative connections, however, the rel-
ative strength of connections (compared to other connections in the same network) in the
197
Mood
Motor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Mood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual InterestSleep: SleepIrrit: IrritabilitySpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: ContentAggr: Aggressive BehaviorApp: AppearanceInsg: Insight
Figure 12.1: Network structure of manic symptoms at t0. Each node represents one ofthe eleven items from the YMRS. Blue connections represent positive edges, red connectionsrepresent negative edges. The pie chart surrounding each node represents node predictability.
198
Mood
Motor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Mood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual InterestSleep: SleepIrrit: IrritabilitySpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: ContentAggr: Aggressive BehaviorApp: AppearanceInsg: Insight
Figure 12.2: Network structure of manic symptoms at t1. Each node represents one ofthe eleven items from the YMRS. Blue connections represent positive edges, red connectionsrepresent negative edges. The pie chart surrounding each node represents node predictability.
199
Mood
Motor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Mood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual InterestSleep: SleepIrrit: IrritabilitySpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: ContentAggr: Aggressive BehaviorApp: AppearanceInsg: Insight
Figure 12.3: Network structure of manic symptoms at t2. Each node represents one ofthe eleven items from the YMRS. Blue connections represent positive edges, red connectionsrepresent negative edges. The pie chart surrounding each node represents node predictability.
200
Mood
Motor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Mood
Motor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Mood
Motor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Figure 12.4: Network structure of manic symptoms at t0, t1 and t2. Each node representsone of the eleven items from the YMRS. Blue connections represent positive edges, redconnections represent negative edges. The pie chart surrounding each node represents nodepredictability. Only edges with a weight greater than 0.1 are reported.
201
−2
0
2M
ood
Mot
or
Sexu
al
Slee
p
Irrit
Spee
ch
LgTA
b
Con
t
Aggr
App
Insg
Pre
dict
abili
ty Datasets
1
2
3
Figure 12.5: Network predictability values for the t0 (red line), t1 (green line) and t2 (blueline) networks a standardized z values.
respective networks varies over time. Hereby, some of the connections are described.
Aggressive behavior shares a strong connection to irritability at t0 and t1: however, at t2,
it shares a connection with Insight instead. Language-thought disorder is highly connected to
content at all time points; content is also connected with speech at all time points. Elevated
mood shares connections with motor activity and insight.
Network accuracy
On average, stronger connections in the network are significantly stronger than weaker con-
nections, but weaker connections are not on significantly different from each other.
202
Table 12.3: Node predictability estimates for the eleven symptoms at the three time points.R2t0 R2t1 R2t2
Mood 0.403 0.217 0.339Motor 0.374 0.138 0.168Sexual 0.000 0.000 0.000Sleep 0.312 0.177 0.077Irritable 0.352 0.240 0.215Speech 0.173 0.120 0.252LgTtAbn 0.291 0.278 0.485Content 0.224 0.220 0.443Aggressive 0.307 0.268 0.164Appearance 0.201 0.177 0.000Insight 0.236 0.211 0.228
12.3.2 Network inference
Network predictability estimates for the three time points are reported in 12.3 and repre-
sented in figure 12.5. Elevated mood is the most interconnected node at t0, and language-
thought disorder is the most interconnected node at t1 and t2.
The NCT reported a statistically significant difference between the global strength of
t0 and t2 (p = 0.0018); however, t0 and t1, as well as t1 and t2 do not present significant
differences in global strength (p = 0.49 and p = 0.98, respectively).
12.3.3 Temporal network
The temporal network, estimated through the GVAR adapted for panel data is represented
in figure 12.6. All variables have an effect on themselves over time. However, there are
several variables in the network that are Granger-caused by other variables: for instance,
mood Granger-causes aggressive behavior (which in turn also has a temporal effect on mood),
appearance, increased motor activity, irritability and speech disorder. Other variables, such
as sleep, receive few or no temporal effects.
203
Stability of the temporal effects
The correlation between the temporal effects before and after dropping 10% of subjects was
0.94 and points toward a certain stability of the network structure.
12.4 Discussion
This work tackled the important issue of the study of mania as a network of symptoms.
Several measures were used to study the network structure of manic symptoms.
The strong connection between aggressive behavior and irritability t0 and t1 means that,
when mania is untreated, the two symptoms are highly predictive of one another: if patients
are irritable, then there will be a high chance for them to be aggressive as well, and vice
versa. However, when mania is treated at t1, aggressive behavior becomes predictive of
insight instead: this can be interpreted as a hint that treated manic patients that tend to
have no insight, also tend to show greater signs of aggressive behavior (even though the
mean value for aggressive behavior at t2 is close to 0 in my sample), controlling for all other
symptoms in the network.
Language-thought disorder and content are highly connected at all time points: from
a network point of view, this means that the presence of a an abnormal communication
with the patient (due to, for instance, flight of ideas or echolalia) may suggest that there
is an ongoing delusional or hallucinatory process (and vice versa), controlling for all other
symptoms. The association of language-thought disorder and delusions or hallucinations is
recurrent in the literature: recent works on psychosis also suggest that the less patients are
in control of their thought process, the more likely they suffer from hallucinations (Hartley
et al., 2015). Language-thought disorder may therefore serve as a “proxy” for psychiatrists
to better investigate hallucinatory or delusional patterns. Another possible red flag linked to
delusions or hallucinatory processes may be found in speech alterations, such as logorrhea,
because of the association between speech and content at all time points.
205
Elevated mood is connected with an increased motor activity and insight: although this
is not very informative during severe manic episode which is the case at t0 (because an
elevated mood is the core feature of mania), it informs the clinician that, when the patient
is stable (t2) a loss of insight (if not present before) or an increased energy may be the sign
of a rising mood: this finding is supported by recent literature (Silva et al., 2018).
It is not surprising that at t0 elevated mood is the most interconnected symptom in the
manic network. Because edges are not directed in the partial correlation networks, elevated
mood can be interpreted as the symptom that best predicts (or is predicted by) all other
symptoms in the network at t0. However, when patients becomes more stable, language-
thought disorder becomes the most predictable node in the network: this is likely due to
the strong connection that it shares with content, which also presents high estimates at t1
and t2. This phenomenon is described as “centrality corruption” (Briganti and Linkowski,
2019d): that is, because two nodes share one strong network connection, they rapidly become
important in the self-determination of the network. For this reason and because it shows a
high estimate when patients have a severe mental state, the connectivity of elevated mood at
t0 is more straightforward to interpret, as it shares connections with several nodes: elevated
mood can be considered as the core feature of mania in my sample of manic patients.
The NCT reported a statistically significant difference between the global strength of
networks at t0 and t2. This means that the network present a different connectivity at the
two time points: it is worthy of note that the networks compared by NCT are slightly different
than the ones estimated, because they have regularised partial correlation estimated with a
Pearson’s rho input. Although other comparisons between network structures are described
in this work, the results from NCT offer supplementary arguments in favor of a difference in
network structures on admittance and discharge.
Granger causal effects were explored in the temporal network. From a network point of
view, it is not surprising to see how elevated mood (which is supposed to be the direct reflec-
tion of the manic episode) has a temporal effect on many other manic symptoms; however,
206
in the temporal network, several other symptoms influence each other over time, therefore
supporting a complex system view of the concept of mania. However, temporal effects be-
tween variables seem to be much weaker than the temporal effects that each variables has
on itself.
The results of this work should be interpreted in light of three limitations. First, the
assumption of stationarity (the variables have the same mean and the same standard devi-
ation over time) is likely violated in my sample, because patients present high scores at t0
and low scores at t2: this leads to a less good model fit. Second, and for the same reason
regarding patients scores, Spearman’s ρ correlation was used instead of Pearson’s as an input
for the GGM because the distribution of the data is skewed. Third, although GVAR can be
estimated from three measurement occasions, it would be optimal to obtain a sample with
many more time points, and with more subjects to obtain more accurate estimates.
This study expanded the network theory of mental disorders to mania. Manic symptoms
were interconnected in a network structure, and specific associations between symptoms,
both static (cross-sectional) and dynamic (temporal) nature were explored, as well as the
importance of symptoms in the self-determination of the network. Future work may endeavor
to replicate my results in other population, as well as in patient with a less severe condition.
207
Chapter 13
Investigating the heterogeneity of
psychiatric symptomatology using
community detection algorithms
Abstract
This work aims to challenge current state-of-the-art practices to retrieve and in-
terpret the number of domains of psychiatric symptoms while modeling disorders as
networks by comparing the performances of different community detection algorithms
in a sample of 100 severe manic patients diagnosed with bipolar disorder type I at
three time points (at the start, middle and end of hospital stay). The performance
of three known algorithms (walktrap algorithm, spinglass algorithm and clique perco-
lation) are compared to retrieve different domains of mania in the sample, while the
bridge centrality measure is used to interpret the membership of symptoms to a given
community. The spinglass algorithm is the only algorithm able to retrieve the same
number of communities at all time points. The clique percolation algorithm is useful
to retrieve symptoms that connect multiple communities. I formulate the recommen-
dation to use the spinglass algorithm in addition to the more established walktrap
algorithm to study the number of communities.
209
13.1 Introduction
Uncovering the way symptoms and signs of mental illness co-occur in a given patient or
group of patients has always been a domain of interest in psychiatric research. In the past
decades, the fields of clinical psychiatry and psychiatric epidemiology have tried through
statistical tools to formalize theories around such interactions. This has allowed theoretical
and empirical works to move away from a categorical approach to mental disorder which
dominates currently used diagnostic manuals like the Diagnostic and Statistical Manual of
Mental Disorders (DSM) (American Psychiatric Association, 2013), and to embrace a more
dimensional approach. Such approach has been driven in the last decades by the widely
known common cause framework, which considers that psychiatric symptoms are caused by
a common cause, that is, the mental disorder itself.
Some psychiatric disorders, however, are known to be heterogeneous, that is, symptoms
may occur in specific subsets, and therefore, two patients with the same disorder may present
with different combinations of symptoms: this is for instance the case of depression (Fried and
Nesse, 2015). The subsets of co-occurring symptoms in data sets have been named factors
in the common cause literature using structural equation modeling as a statistical approach
(Kempenaers et al., 2017): the common cause (the disorder itself) causes factors, which in
turn cause the symptoms, which can therefore be understood as passive and interchangeable
consequences of the disorder, and does not actively contribute to it.
A decade ago, some put forward the hypothesis that one cannot identify a single cause
that, if eliminated, can make the clinical presentation of a psychiatric disorder entirely dis-
appear because there is no such cause; instead, psychiatric symptoms mutually cause each
other, and the psychiatric disorder arises from the set of interactions among such symp-
toms. This hypothesis, which translates the science of complex systems to medicine, has
been formally introduced as the network theory of mental disorders (Borsboom, 2017), and
has enjoyed an increasing popularity in the past years. Network theory is accompanied by
a set of statistical tools, called network analysis, which allows researchers to compute disor-
210
ders as a network of nodes (symptoms) that interact with each other through edges (which
represent conditional dependence relationships), often computed as partial correlations. A
series of psychiatric constructs and disorders have been studied using network analysis, such
as empathy (Briganti et al., 2018), self-worth (Briganti et al., 2019), resilience (Briganti and
Linkowski, 2019b), narcissism (Briganti and Linkowski, 2019a), alexithymia (Briganti and
Linkowski, 2019c), post-traumatic stress disorder (Fried et al., 2018), depression (Fried et al.,
2016) and autism (Deserno et al., 2017).
The network framework is interesting because it allows for the study of individual con-
nections among symptoms: network inference methods can put forward symptoms that are
more important than others in the self-determination of the network: in other terms, such
symptoms can better predict (or be predicted by) other symptoms in the network than oth-
ers, and can therefore be considered as prime candidates for clinical intervention (Fried et al.,
2017). Studying network structures is also interesting because they tend to replicate across
populations (Fried et al., 2018): it is therefore probable that a connection discovered between
two symptoms in a given data set may replicate in other data sets, and the more network
structures replicate, the more intervention studies can effectively tackle relevant symptoms.
The measure of network centrality (Boccaletti et al., 2006) disposes of several measures
to capture the relative or absolute importance of symptoms in the network. Widely used
centrality measures are strength, which is the absolute sum of connections of a given node (a
measure of relative importance because there will always be a more central node even though
the network is poorly connected), and node predictability, which is the shared variance (R2)
of a given node with surrounding nodes (an absolute measure of connectivity). One reported
problem with centrality measures is that if two symptoms present a very strong connection
to one another, although poorly connected to the rest of the network, they tend to receive
very high centrality estimates: this phenomenon, which has been referred to as centrality
corruption (Briganti and Linkowski, 2019d), occurs with symptoms that are redundant, that
is they tend to represent the same thing. However, redundancy should not be exclusively
211
understood as a negative property, because it is a way of showing how two or more symptoms
connect more closely to each other than to other symptoms in the network, and therefore
highly influence each other.
However, although the way a given symptom connects to the rest of symptoms in the
network is very important, it should be considered in parallel with the way that a given
symptoms connects to neighboring symptoms; that is, how symptoms cluster together in
communities, or domains (understood as factors from a common cause point of view). This
is particularly important in the case of redundant symptoms, because they tend to form
communities.
Community detection algorithms are a popular way to retrieve the number of communities
in networks (Briganti et al., 2018): for instance, exploratory graph analysis is an increasingly
used tool that has been shown to have high accuracy (Golino and Epskamp, 2017) and
applied in empirical papers (Briganti et al., 2018, 2019; Briganti and Linkowski, 2019b).
The exploratory graph analysis (EGA) relies on the walktrap algorithm, which is based
on the principle that adjacent nodes tend to belong to the same community (Yang et al.,
2016). For an optimal use of EGA, it is recommended that, once the number of communities
retrieved empirically, fitness of the model should be checked with a traditional confirmatory
factor analysis (CFA) (Golino and Demetriou, 2017), even more so if the model proposed by
the algorithm is somewhat different from the one that was expected. These guidelines were
followed and supported by some recent empirical works (Briganti et al., 2019; Briganti and
Linkowski, 2019b). However, two other community detection algorithms are optimized for
use in symptom network: the spinglass algorithm and clique percolation.
The spinglass algorithm is based on the principle that edges should connect nodes of the
same community, whereas nodes belonging to different communities should not be connected
(Yang et al., 2016).
Clique percolation is an algorithm recently adopted in the psychiatric network framework:
it considers the network as a set of k-cliques (communities), that are understood subsets
212
of k nodes that are fully connected (Palla et al., 2005). Clique percolation accepts two
communities as adjacent if they share k−1 nodes, which means nodes can belong to multiple
communities, therefore establishing a meaningful bridge between them.
The notion of bridge has mainly been analyzed in network settings when studying comor-
bidity (Jones et al., 2019). A set of statistics, grouped under the name bridge centrality have
therefore been developed to assess which network nodes better connect different communities
within the network; it is currently used in networks composed of symptoms from different
disorders to detect which symptoms connect different disorders.
Except from the guidelines proposed within the framework of EGA (Golino and Epskamp,
2017), there is to this date no standardized way to explore communities in network structures.
This is an important issue to address, since retrieving how symptoms closely cluster together
is crucial when choosing a target for therapeutic intervention in the framework of mental
disorders as complex systems. It is also important to correctly interpret the membership of
a given symptom to a given community (or multiple communities), and how it impacts the
way the symptom connects to other items in the network. If a network of symptoms from
the same mental disorder is divided into multiple communities, than it is plausible to assume
that if a patient presents a given symptoms, the patient will more likely present symptoms
that belong to the same community than symptoms that belong to other communities. This
is likely to greatly impact how the patient is treated, even more so in severe situations, such
a psychotic or manic episode.
Bridge centrality can be helpful as a measure to interpret the connectivity of a given
symptom while taking into account the presence of community: it is a much needed measure
that circumvents the problem of centrality corruption and can therefore provide an estimate
of how a symptom connects two different communities, that is, how the symptom brings
different parts of the network together: in clinical terms, how the symptom, if activated,
may cause many other symptoms in the network to co-occur. This phenomenon has been
described as hysteresis in the network approach (Borsboom and Cramer, 2013).
213
An additional point of interest which has not been studied in the network field is how
communities of symptoms evolve over time when treatment has been administered. This
is a research question that can be difficult to answer in disorders such as depression or
PTSD, but may be easier to tackle in disorders such as bipolar disorder: for instance, the
administration of neuroleptic, anxiolytic and soporific treatment in acute and severe mania
can have a radical and rapid effect on behavior, and may change how symptoms co-occur.
The aim of this work is to delve into the complicate issue of community detection in psy-
chiatric network and its interpretation. In this work, I will first compare the performances of
the walktrap algorithm, clique percolation and spinglass algorithm in detecting communities
in a sample of 100 severe manic inpatients in three time points – start, middle and end of
hospital stay; I will also estimate bridge centrality parameters and interpret the results in
light of symptom membership.
It is interesting to choose a sample of severe manic inpatients, because there has been
interest in the past years into the hypothesis that there exists several types of mania. Three
types of mania were for instance described by Hanwella et al. (Hanwella and de Silva,
2011): elated mania (with elements of elevated mood and sexual interest), irritable mania
(mainly driven by irritability and aggressive behavior), and psychotic mania (with psychotic
symptoms). It is therefore expected that manic symptoms in my sample will also cluster
into different communities, and it is plausible to expect that one of several symptoms may
switch communities over time.
This work is organized as follows: first, I will introduce the reader to the statistical
methods associated with community detection and bridge centrality in network structures.
Second, I will compare the performance of the three algorithms chosen in the case of my
sample of manic inpatients and estimate bridge centrality parameters; third, I will test the
fitness of the average model proposed by the different algorithms through confirmatory factor
analysis; fourth, I will discuss my results and their implication in clinical practice.
214
13.2 Method
13.2.1 Ethical approval
This study was approved from the Ethical Committee of the Brugmann Teaching Hospital
in Brussels (CHU Bruxelles Brugmann).
13.2.2 Data sets
Participants
My data set is composed of 100 patients, hospitalized in the context of an involuntary
commitment in a secure psychiatric unit. To be included in this study, patients needed to
be diagnosed with bipolar I disorder and to present with a manic episode at admission.
An abnormal blood or urine analysis in regards to toxicology as well as a concomitant
personality or somatic disorder that could account for the presence of manic symptoms were
considered as exclusion criteria. All patients were involuntarily committed for a period which
lasted for 40 days on average (following the Belgian law for involuntary commitment).
All manic patients were treated with a standard set of drugs following the local protocol
for manic patients: in the first stage of treatment (when the clinical presentation is severe),
an association of a typical and atypical antipsychotic drugs is administered, with a mood
stabilizer as well as soporific and anxiolytic drugs when necessary and depending on the
symptoms presented. In the second stage of treatment (when the clinical presentation is
stable), the patient is left with an atypical antipsychotic, a mood stabilizer, as well as a
soporific and/or anxiolytic when necessary.
Three time points were collected for each patient: on admission (t0), halfway through the
commitment period t1, and on discharge (t2). The data sets were anonymized by default.
Patients were 20 to 72 years old (M = 44.5, SD = 14.5); 47 of them were female, and 53 of
them were male.
215
Table 13.1: Symptoms from the Young Mania Rating Scales (Young et al., 1978)N symptom Symptom1 Elevated Mood2 Increased Motor Activity-Energy3 Sexual Interest4 Sleep5 Irritability6 Speech (Rate and Amount)7 Language-Thought Disorder8 Content9 Aggressive Behavior10 Appearance11 Insight
Measurement
The YMRS (Young et al., 1978) was used to assess manic symptoms, namely Elevated
Mood, Increased Motor Activity-Energy, Sexual Interest, Sleep, Irritability, Speech (Rate
and Amount), Language-Thought Disorder, Content, Aggressive Behavior, Appearance, and
Insight. Symptoms were scored 0 to 4, depending on the severity of the clinical presentation,
both at t0, t1 and t2. The symptoms are illustrated in table 13.1.
13.2.3 Network Analysis
Software
The software used for the analyses carried out in this study is R (version 3.6.3, available at
https://r-project.org). The packages needed for the analyses were bootnet (Epskamp
and Fried, 2018) and qgraph (Epskamp et al., 2012), igraph (Csardi and Nepusz, 2006)
and CliquePercolation for community detecton, networktools (Jones et al., 2019) and lavaan
(Rosseel, 2012) for confirmatory factor analysis.
216
Network estimation
Let y be a a normal multivariate vector y = (y1, ..., yp) with mean vector µ and a variance-
covariance matrix Σ. For all subjects,
y ∼ N(µ,Σ). (13.1)
Let Θ be the inverse of Σ,
Θ = Σ−1 (13.2)
that is known as the precision matrix or a Gaussian Graphical Model (GGM). The
elements of of the GGM encode the partial correlation coefficients θij of two variables yi and
yj given all other variables in y, that is, y−(i,j):
Cor(yi, yj|y−(i,j)
)= − θij√
θii√θjj, (13.3)
therefore, the GGM represents the network itself.
The partial correlation θij between yi and yj is used as the edge weight, that is, the
strength of the connections between nodes Vi and Vj in the network. Edge weights can be
positive (usually represented as blue connections) or negative (usually represented as red
connections) depending on the sign of θij. The presence of an edge between two nodes Vi
and Vj in the network can be interpreted as a conditional dependence relationship: node
Vi predicts (or is predicted by) node Vj, after controlling for all other nodes in the network
V−(ij).
Three separate GGMs were estimated for t0, t1 and t2 data sets to study cross-sectional
effects among manic symptoms. For the estimation of the network structures, Spearman ρ
correlation was used as an input parameter because of the structure of the data (high scores
at t0, low scores at t2). Further details about the GGM can be found in recent state of the
217
art methodological works (Epskamp et al., 2018).
In a network of symptoms, if two nodes A and B are connected, it means for instance
that if the observed group scored high on symptom A, then the observed group is also
more likely to score high on symptom B, and vice versa, controlling for other nodes in the
network (Briganti et al., 2018); if two nodes are connected, that means they are conditionally
dependent given all other nodes in the network (i.e their partial correlation is nonzero).
Each edge in the network has a weight representing the strength of association between
two symptoms; edges can be positive (and therefore represent a positive association) or
negative (denoting a negative association). In the network the edge weight is represented as a
combined thickness and saturation of the edge; positive edges are shown in blue, and negative
edges in red. Nodes are placed in the network by the Fruchterman-Reingold algorithm, based
on the sum of the connections a given node has with other nodes (Fruchterman and Reingold,
1991).
13.2.4 Community detection
Defining a community
Let us define an undirected graph G = (V,E) from a set of nodes V = {V1, . . . , Vn} that
share a set of edges E. A network community C = {V1, . . . , Vk} of a graph G is a subset of
nodes V1, . . . , Vk that share a higher proportion of edges among them within the community
than they share with other nodes in the network.
The walktrap algorithm
The walktrap algorithm was introduced by Pons and Latapy (Pons and Latapy, 2005) and
is based on the principle that a random walk on a graph (from one node to another) gets
trapped into communities because they are composed of densely connected nodes. The graph
is characterized by a partition P = {C1} ∪ {C2} ∪ . . . ∪ {Cn} of all communities.
The algorithm starts by dividing the graph of communities composed of only one node.
218
In a random walk of length t the probability that a walker transitions from node Vi to node
Vj is denoted P tij: this probability depends on the connection between Vi and Vj, which is
denoted Aij, and the number of connections Vi has, also called degree d(i).
Pij =Aijd(i)
P ti• designs the ith row of the transition matrix P t. D is the diagonal matrix that encodes
degrees d(i) for all nodes.
The walktrap algorithm merges two communities according to a criterion based on the
distance between two nodes. The distance between two nodes Vi and Vj is defined as
rij =
√√√√ n∑k=1
(P tik − P t
jk
)2
d(k)= ‖D−
12P t
i• −D−12P t
j•‖
where ‖.‖ is the Euclidean Norm.
If two communities minimize the mean of squared distances of each node and its com-
munity σk, they are merged.
σk =1
n
∑C∈Pk
∑i∈C
r2iC
At each step, the distance between communities is updated, but only if the communities
are adjacent –that is, they share at least one edge.
The walktrap algorithm has been used in several empirical studies (Briganti et al., 2018,
2019; Briganti and Linkowski, 2019b) and has been shown to have high accuracy (Golino
and Epskamp, 2017).
The Clique Percolation Method
The Clique Percolation Method (CPM, or Clique Percolation) was introduced for weighted
networks (CPMw) over a decade ago (Palla et al., 2005; Farkas et al., 2007). For weighted
219
networks, a community, or k-clique C is composed of k nodes that form a subgraph if the
geometric mean of the edge weights w within the community, called intensity, or I, is greater
than a threshold. Each k-clique C has k(k − 1)/2 edges among its nodes. Therefore, I can
be determined as
I(C) =
∏i<ji,j∈C
wij
2/k(k−1)
which means that edges weaker than the intensity threshold can be included in the
community. Each node can belong to multiple communities, and therefore has a community
membership number (the number of communities the node belongs to), as well as community
neighbours, that is nodes that belong to one of the communities that the node belongs to.
Clique Percolation therefore has two parameters, k (which can never be lower than three)
and I. Although it is plausible to set a minimum k of three in psychiatric networks, it is
important to find the optimal I threshold: if too high, no communities are found; if too low,
then one giant community will include all nodes in the network. The optimal I threshold is
low enough so a sufficient number of k-cliques can participate in the partition of communities,
and yet high enough so that it impedes giant communities to appear. In big networks (with
many nodes), the algorithm starts by setting I as equal to the biggest edge in the network,
and then progressively lowers it until the two largest communities retrieved (the number of
nodes belonging in communities is defined as n1 and n2) have a ratio n1
n2= 2. However,
psychiatric networks are often considered as small and therefore another equations sets the
threshold as
χ =∑
nα 6=nmax
n2α/
(∑β
nβ
)2
.
The algorithm first starts by finding a parent k-clique that fulfills the criterion IC > I,
then tests whether children k-cliques exist that also fulfill the same criterion.
220
Setting, for a given k, an optimal I allows for the approximation of the transition point
pC(I), which is the point below which a giant community appears (and therefore other
communities cannot be retrieved); for k = 3, that
pC(I)
pC(0)
∣∣∣∣k=3,4
'
[1− In
n−1∑i=0
(−n ln I)i
i!
]−1/(k−1)
.
Clique Percolation therefore offers the possibility of having a node belong to multiple
communities, which is an interesting feature for the study of psychiatric networks. In this
work, I was able to retrieve an optimal value I = 0.18 for t0, however, no optimal value
were retrieved for t1 and t2: two suboptimal values were chosen, respectively I = 0.16 and
I = 0.21 that retrieved a reasonable number of communities while leaving the lowest number
of nodes as their own community.
The spinglass algorithm
The spinglass algorithm, by Reichardt and Bornholdt (Reichardt and Bornholdt, 2006) and
adapted for weighted networks by Traag and Bruggeman (Traag and Bruggeman, 2009) is un-
derstood as an optimization method relying on an analogy between the statistical mechanics
of complex networks and physical spin glass models.
The properties of the spinglass algorithm are straightforward: the algorithm rewards the
presence of edges between nodes of the same community and the absence of edges between
nodes of different communities, as well as penalizes the absence of edges between nodes of
the same community and the presence of edges between nodes of different communities. The
properties of the spinglass algorithm are retrieved in its function, which is a derivation of
the Hamiltonian (Reichardt and Bornholdt, 2006):
H({σ}) = −∑i 6=j
aijAijδ (σi, σj) +∑i 6=j
bij (1− Aij) δ (σi, σj)
221
+∑i 6=j
cijAij [1− δ (σi, σj)]−∑i 6=j
dij (1− Aij) [1− δ (σi, σj)]
where σi represents the spin state or community membership of a node Vi, and aij, bij, cij, dij
individual contributions to the adjacency matrix Aij, that is, partial correlation in a psychi-
atric network. the equation is split into four parts; the first represents the presence of edges
between nodes of the same community, the second the absence of edges between nodes of
the same community, the third the presence of edges between nodes of different communities
and the fourth the absence of edges between nodes of different communities. To determine
the community membership of a node, the algorithm adds a node to the community that
will raise its adhesion coefficient –that is, the coefficient that sets apart a given community
from the rest of the network; the adhesion coefficient depends on the edge distribution, and
in the case of psychiatric networks, edge weight parameters are retrieved with either a GGM
or an Ising Model (the GGM equivalent for binary data) (Ising, 1925; van Borkulo et al.,
2014).
In this work I follow the methodological guidelines introduced in empirical papers (Brig-
anti et al., 2018) and estimate the spinglass algorithm 100 times to increase stability: the
node membership is therefore a mean over the hundred estimations.
Application of community detection algorithms in a clinical data set
I applied the three community detection algorithms described to the data set of manic
patients at the three time points: when doing so, each symptom (a node in the network) is
assigned to a community, and each community is designed by a color. In the specific case of
Clique Percolation, several nodes can have multiple communities and will therefore be split
with multiple colors. A supplementary analysis is done to follow recent guidelines (Golino
and Demetriou, 2017) recommending to fit the overall model retrieved using community
detection algorithms with CFA; good fit indices mirror a good model. Because I used three
different algorithms at three time points, I fit an average model. I report in this work the
Comparative Fit Index (CFI; should be higher than 0.95), the Root Mean Square Error of
222
Approximation (RMSEA; should be lower than 0.06) and Standardized Root Mean Square
Residual (SRMR; should be lower than 0.08).
13.2.5 Bridge centrality
Jones et al. (Jones et al., 2019) introduced bridge centrality to measure a the connectivity
of a node outside its community. Bridge strength is the sum of all edge weights of a given
node to other nodes that do not belong in the same community C:
bridge strength =∑
b∈(N(i)−C)
|wij|
bridge centrality is therefore useful to investigate which symptoms are capable of acti-
vating other symptoms, and be considered as targets for clinical intervention.
13.3 Results
13.3.1 Community detection
Symptom membership for all three algorithms at the three time points are represented in
table 13.2.
The walktrap algorithm
Figure 13.1 shows the networks at t0, t1, and t2 with the communities retrieved by the walk-
trap algorithm. It detects five communities at t0 and t2, as well as eleven communities of
one symptom each. The five communities retrieved at t0 and t2 are identical: the first com-
munity is composed of Elevated Mood, Increased Motor Activity, Appearance, and Insight;
the second community is composed of Sexual interest and Aggressive Behavior; the third
community is composed of Sleep, Irritability and Language-Thought Disorder; Speech and
Content make for a one-symptom community each.
223
MoodMotor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Mood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual InterestSleep: SleepIrrit: IrritabilitySpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: ContentAggr: Aggressive BehaviorApp: AppearanceInsg: Insight
At t0 At t1 At t2
Mood
Motor
Sexual
Sleep
Irrit
SpeechLgTAb
Cont
Aggr
App
Insg
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−EnergyApp: AppearanceInsg: Insight
BSexual: Sexual InterestAggr: Aggressive Behavior
CSleep: SleepIrrit: IrritabilityLgTAb: Language−Thought Disorder
DSpeech: Speech (Rate and Amount)
ECont: Content
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−EnergyApp: AppearanceInsg: Insight
BSexual: Sexual InterestAggr: Aggressive Behavior
CSleep: SleepIrrit: IrritabilityLgTAb: Language−Thought Disorder
DSpeech: Speech (Rate and Amount)
ECont: Content
MoodMotor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−EnergyApp: AppearanceInsg: Insight
BSexual: Sexual InterestAggr: Aggressive Behavior
CSleep: SleepIrrit: IrritabilityLgTAb: Language−Thought Disorder
DSpeech: Speech (Rate and Amount)
ECont: Content
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−EnergyApp: AppearanceInsg: Insight
BSexual: Sexual InterestAggr: Aggressive Behavior
CSleep: SleepIrrit: IrritabilityLgTAb: Language−Thought Disorder
DSpeech: Speech (Rate and Amount)
ECont: Content
Figure 13.1: Communities detected by the walktrap algorithm at t0 (left), t1 (middle) andt2 (right). Eleven communities of one symptom each were detected at t1. Each communityis denoted by a different color.
Clique Percolation
Figure 13.2 shows the networks at t0, t1, and t2 with the communities retrieved by Clique
Percolation. The first of the two biggest communities at t0 involve Language-Thought disor-
der, Content and Speech: this could be customarily defined as Psychotic mania; the second
includes Elevated Mood, Increased Motor Activity, Irritability and Aggressive Behavior,
which could be customarily called Exalted Mania. The third, smaller community involves
Appearance and Insight: this could be customarily called Careless Mania. Sleep belongs to
all three communities. At t1 only two communities emerge: although Psychotic Mania stays
mostly unchanged as a community, Careless Mania disappears and its symptoms join the
Exalted Mania cluster. At t2, one giant community emerges regrouping Exalted and Psy-
chotic Mania, while Sleep, Irritability and Increased Motor Activity and Sleep form however
a different community.
224
Mood
Motor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Mood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual InterestSleep: SleepIrrit: IrritabilitySpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: ContentAggr: Aggressive BehaviorApp: AppearanceInsg: Insight
Mood
Motor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Mood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual InterestSleep: SleepIrrit: IrritabilitySpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: ContentAggr: Aggressive BehaviorApp: AppearanceInsg: Insight
Mood
Motor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
Mood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual InterestSleep: SleepIrrit: IrritabilitySpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: ContentAggr: Aggressive BehaviorApp: AppearanceInsg: Insight
Figure 13.2: Communities detected by Clique Percolation at t0 (left), t1 (middle) and t2(right). Each community is denoted by a different color. Symptoms can belong to differentcommunities: they will therefore be denoted by multiple colors. If a node is white, then itconstitutes its own community.
225
Mood
Motor
Sexual
Sleep
Irrit
SpeechLgTAb
Cont
Aggr
App
Insg
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−Energy
BSexual: Sexual InterestIrrit: IrritabilityAggr: Aggressive Behavior
CSleep: SleepApp: AppearanceInsg: Insight
DSpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: Content
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−Energy
BSexual: Sexual InterestIrrit: IrritabilityAggr: Aggressive Behavior
CSleep: SleepApp: AppearanceInsg: Insight
DSpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: Content
MoodMotor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual Interest
BIrrit: IrritabilityAggr: Aggressive Behavior
CSleep: SleepApp: AppearanceInsg: Insight
DSpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: Content
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual Interest
BIrrit: IrritabilityAggr: Aggressive Behavior
CSleep: SleepApp: AppearanceInsg: Insight
DSpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: Content
MoodMotor
Sexual
Sleep
Irrit
Speech
LgTAb
Cont
Aggr
App
Insg
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual Interest
BSleep: SleepIrrit: Irritability
CAggr: Aggressive BehaviorInsg: Insight
DSpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: ContentApp: Appearance
●
●
●
●
●
●
●
●
●
●
●
AMood: Elevated MoodMotor: Increased Motor Activity−EnergySexual: Sexual Interest
BSleep: SleepIrrit: Irritability
CAggr: Aggressive BehaviorInsg: Insight
DSpeech: Speech (Rate and Amount)LgTAb: Language−Thought DisorderCont: ContentApp: Appearance
At t0 At t1 At t2
Figure 13.3: Communities detected by the spinglass algorithm at t0 (left), t1 (middle) and t2(right). Each community is denoted by a different color. Symptoms can belong to differentcommunities: they will therefore be denoted by multiple colors.
The spinglass algorithm
Figure 13.3 shows the networks at t0, t1, and t2 with the communities retrieved by the
spinglass algorithm.
The results of community detection are fairly stable across the three time points. Four
communities are detected. The first relates to Exalted Mania and contains Elevated Mood,
Increased Motor Activity and Sexual Interest, except for t0 where the latter is included in the
second community, customarily defined as Irritable Mania with the symptoms Aggressive Be-
havior and Irritability. The third community relates to Careless Mania and includes Sleep,
Appearance and Insight The fourth community is Psychotic Mania and includes Speech,
Language-Thought Disorder and Content. Psychotic Mania contains the symptom Appear-
ance at t2.
226
Table 13.2: Symptom membership retrieved by the three algorithms at each time point.Symptom WalktrapT0 CliqueT0 SpinglassT0Mood 3 1 & 2 1Motor 3 1 1Sexual 1 NA 3Sleep 2 1 & 2 & 3 2Irritable 2 1 3Speech 4 3 4LgTtAbn 2 3 4Content 5 3 4Aggressive 1 1 3Appearance 3 2 2Insight 3 2 2
Symptom WalktrapT1 CliqueT1 SpinglassT1Mood 1 1 1Motor 2 1 1Sexual 3 1 1Sleep 4 1 2Irritable 5 NA 3Speech 6 1 & 2 4LgTtAbn 7 2 4Content 8 2 4Aggressive 9 NA 3Appearance 10 1 2Insight 11 1 2
Symptom WalktrapT2 CliqueT2 SpinglassT2Mood 3 2 2Motor 3 1 2Sexual 1 2 2Sleep 2 1 3Irritable 2 1 3Speech 4 2 4LgTtAbn 2 2 4Content 5 2 4Aggressive 1 2 1Appearance 3 2 4Insight 3 2 1
227
Table 13.3: Fit indices reported by CFA for the three time points: the Comparative FitIndex (CFI; should be higher than 0.95), the Root Mean Square Error of Approximation(RMSEA; should be lower than 0.06) and Standardized Root Mean Square Residual (SRMR;should be lower than 0.08).
Time Point CFI RMSEA SRMRt0 0.990 0.021 0.067t1 0.993 0.014 0.065t2 0.814 0.087 0.089
Table 13.4: Bridge centrality of manic symptoms at the three time points.Symptom T0 T1 T2Mood 0.536553547572247 0.71117320254293 1.39122331371955Motor 0.557759501813567 0.554662780688992 0.969490158197589Sexual 0.585202022768328 0.671285105896003 0.848914759652126Sleep 1.02399980413521 0.717178621701378 0.64818785130062Irritable 0.919669074419216 0.428980471315663 0.76748772793819Speech 0.454110034973616 0.699339618436966 0.707135221780921LgTtAbn 0.642989557041043 0.526637276371529 0.907471762302611Content 0.642647813143483 0.528031950690735 1.02242089051355Aggressive 0.371135439017514 0.167709813268731 0.782824607734004Appearance 0.594002911110612 0.747660398297268 0.569317636625077Insight 0.44246399256873 0.518188886854173 1.27973007896221
Confirmatory Factor Analysis for the overall mean model
I estimated fit indices for the following model, which is strongly inspired by the one re-
trieved by the spinglass algorithm and to which overall converged the models retrieved by
the walktrap algorithm and Clique Percolation: Exalted Mania, composed by Mood, In-
creased Motor Activity, and Sexual Interest; Irritable Mania, composed by Irritability and
Aggressive Behavior; Careless Mania composed of Sleep, Appearance and Insight; Psychotic
Mania composed of Speech, Language-Thought Disorder and Content. The fit indices are
reported in table 13.3.
Fit indices indicate that a model is a good fit at t0 and t1 but not at t2.
228
●
●
●
●
●
●
●
●
●
●
●
Bridge Strength
−1 0 1 2
Agg
Ins
Spc
Mod
Mtr
Sxl
App
Cnt
LTA
Irr
Slp
●
●
●
●
●
●
●
●
●
●
●
Bridge Strength
−2 −1 0 1
Agg
Irr
Ins
LTA
Cnt
Mtr
Sxl
Spc
Mod
Slp
App
●
●
●
●
●
●
●
●
●
●
●
Bridge Strength
−1 0 1 2
App
Slp
Spc
Irr
Agg
Sxl
LTA
Mtr
Cnt
Ins
Mod
At t0 At t1
At t2
Figure 13.4: Bridge centrality estimates for the eleven manic symptoms at t0 (top left), t1(top right), and t2 (bottom).
13.3.2 Bridge centrality
Table 13.4 and figure 13.4 report the bridge centrality estimates for the eleven manic symp-
toms at t0, t1, and t2. At t0, Sleep is the most interconnected symptom, and it stays inter-
connected at t1, before becoming one of the poorly connected symptoms at t2. Appearance
is the most interconnected symptom at t1, and Mood at t2.
13.4 Discussion
This work tackled the important topic related to the study of the heterogeneity of psychi-
atric symptomatology using community detection algorithms within the network analysis
framework for mental disorders. I aimed to compare the performances of well-known com-
munity detection algorithms, as well as provide a possible solution to the phenomenon of
redundancy and centrality corruption among network symptoms. The conceptual framework
229
of community detection under examination in this work has important clinical implications
that I will also discuss in this section.
It is worth commenting on the choice of my study settings to contextualize the discussion
of the performance of the community detection algorithms itself. First, I deliberately chose
to compare the performance of community detection algorithms in a setting characterized
by a clinical sample of 100 patients and with a relatively small network of eleven symptoms,
because it is one of the most common settings in clinical research: this is important because
most of the network methodology was initially conceived for relatively big networks and for
relatively big samples. Second, the chosen data set is composed of severe manic patients
diagnosed with bipolar I disorder at three relatively relevant time points: on admission
(t0) and on discharge (t2), the data set likely presents a ceiling and floor effect –that is,
symptoms score very high on average at t0 (because patients are affected with severe mania)
and very low on average at t2 (because patients are treated and have present on average with
an improved mental examination); the low variability in the data set that comes with the
ceiling or floor effect is the reason why Spearman ρ was used instead of Pearson’s correlation
coefficient (Briganti et al., 2019). These are important parameters to take into account when
both developing or validating methods for network psychiatry and applying them in empirical
studies, because they are likely present in many data sets that deal with patients in clinical
settings: when admitted, symptoms are usually high, and are therefore likely to reinforce
each other: in network terms, their connection is stronger and they better predict each
other. The same happens when discharging patients, since their symptoms are either mild
or absent, but there is no reinforcement on a clinical basis, because although the connection
is strong, it is justified by a conditional dependence between variables that are lowly scored:
it is therefore crucial to take the data into account when interpreting connections among
symptoms in a network. Between admission and discharge (t1), however, there is a time
frame where certain symptoms are on the way to recovery because a psychiatric treatment
is effectively acting on those specific symptoms, and others are still unchanged because they
230
are either poorly connected, or because there has not been enough time for the effect to
disseminate to such entities; from a network point of view, this translates to an overall
desynchronized network at t1 (because symptoms present with very different scores)1.
The latter point greatly influences how community detection algorithms work, and in par-
ticular the performance of the walktrap algorithm and Clique Percolation, since the strength
of connections greatly impacts the determination of symptom membership in both of them.
It is likely to partially explain why the walktrap algorithm detects eleven communities of one
symptom each (and therefore no community at all) at t1. The same does not however happen
with Clique Percolation, although the low connectivity may influence the appearance of two
bigger communities instead of the three detected at the first time point where connectivity
is higher.
One of the main interest in using Clique Percolation lies in an exploratory approach to
detect symptoms that connect multiples communities, as a basis for an inference using bridge
centrality. However, its use and outcomes heavily rely on the choice of the two parameters
k and I which can vary. For instance, an optimal I was not retrieved for t1 and t2, and
therefore suboptimal thresholds had to be chosen.
The spinglass algorithm is the only algorithm able to retrieve the same number of com-
munities, with an overall stable symptom membership at all time points. Although running
the spinglass algorithm a limited number of times can lead to obtaining different community
memberships, it has been reported that retrieving the communities with highest frequency
over several runs leads to a certain stability of the results (Briganti et al., 2018). Such sta-
bility is likely influenced by the straightforwardness of the spinglass function that rewards
or penalizes the presence of edges within and between communities and focuses less on the
strength of edges: this is important because, when considering the principle of hysteresis
(Borsboom and Cramer, 2013) in the network theory of mental disorders 2, although the
1The difference in the overall connectivity of the network at the three time points was explored in aseparate paper (Briganti et al., 2020a). The overall connectivity is statistically different at t0 and t2.
2Hysteresis: symptoms stay connected over time but the strength of connection can vary if the patient iswell, (poorly connected in the case of this work), or unwell (strongly connected in the case of this work).
231
connectedness may vary, the network structure (and its partition into communities) is sup-
posed to remain unchanged, since symptoms are likely to be related by their nature (Kendler
et al., 2011). Therefore, although initially poorly introduced in the psychiatric network field,
I recommend the use of the spinglass algorithm as a complementary tool to the more es-
tablished walktrap algorithm to detect communities in network structures because of its
reliability and ease of use (since the algorithm does not depend like Clique Percolation on
the determination of optimal parameters).
It is also crucial to check for model fitness with CFA. Although CFA conceptually implies
a model where symptoms do not form a complex system but are instead a consequence of
a latent variable, it is plausible to interpret symptoms that are highly associated within the
same community as highly predictive of each other, because they convey a similar clinical
meaning. For instance, it is not surprising that Aggressive Behavior and Irritability are con-
nected at all three time points and form a community, since they can easily be interpreted
to clinically represent similar aspects of mania: this reason justifies the act of validating the
model found through exploratory community detection with CFA (Golino and Epskamp,
2017). However, validating through CFA the model retrieved does not necessarily mean
applying a reductionist view onto the model itself: the complex systems approach to mental
disorder implies that, if a network is more than the sum of its communities, than the com-
munity is more than the sum of its symptoms: within the community, there are symptoms
that interact with the neighboring network in different ways, and therefore contribute in
different ways to the emergence of the network and the heterogeneous nature of the clinical
presentation in psychiatric disorder; this means that one cannot reduce a network to a set
of communities either. From a probabilistic view, for instance, it is plausible to assume that
there exist sub-clinical entities within the manic presentation, such as the four communities
belonging to the overall recurring model retrieved in the present analysis (Exalted, Irritable,
Careless and Psychotic) that lead to different clinical presentation in patients, but all com-
munities within a network interact together through specific symptoms, and this is likely to
232
further increase the clinical heterogeneity.
From this perspective, bridge centrality is a good way forward to measure how commu-
nities interact together and which symptoms are the main responsible for such interactions.
Although initially introduced to tackle comorbidity in networks composed of symptoms from
different disorders (Jones et al., 2019), I showed how bridge centrality can be a very inter-
esting approach in networks of symptoms from one mental disorder that is susceptible of
be heterogeneous (that is most if not all mental disorders). My findings further show how
Clique Percolation and bridge centrality work well together and are complementary: Clique
Percolation reports which communities the interconnected item belongs to, and bridge cen-
trality quantifies the interaction itself among the different communities. For instance, Sleep
is identified to be the most interconnected network with both analyses: the former shows how
sleep disorders are symptoms that belong to multiple communities, and the latter reports
how they connect the multiple communities they belong to.
From a clinical point of view, symptoms that belong to different communities can be
considered as prime candidates when choosing a target for an intervention: because bridge
symptoms are predictive of symptoms in different communities, the effect of clinical interven-
tion on bridge symptoms is likely to greatly affect the overall clinical presentation by affecting
multiples communities. This is the case of Sleep, and its connectedness in my study justifies
further studies that explore it as a risk factor in bipolar disorder the (la Cour Karottki et al.,
2020).
The results of this work should be interpreted in light of a number of limitation. I hereby
describe three. First, my data set is composed of 100 patients over three time points: this
is likely to limit the replication of my results in other samples. Second, although symptoms
that are connected in a GGM can be interpreted to be highly predictive of one another,
I do not know whether there is a directed causal effect among symptoms: for instance, a
symptom with high bridge centrality score and belonging to different communities (such as
Sleep) could in fact be the common effect of different variables. Third the methodological
233
world of network analysis is based on software that is rapidly evolving: the computation
could be further optimised in the future, and this may slightly change the results retrieved.
Community detection is an important tool when addressing the heterogeneity of mental
disorders: the study of how symptoms co-occur and cluster together can lead to a better
theoretical and clinical framework for the definition and treatment of psychiatric illness.
234
Chapter 14
Discussion
In each chapter of this dissertation, I introduced improvements on the way psychiatric en-
tities are analyzed as networks. This work heavily relied on the use of Bayesian Artificial
Intelligence to uncover the underlying causal structure in the symptom data sets, which is
a new and interesting advance in the medical sciences that will help clinicians better pre-
vent, diagnose and treat mental disorders. The works in this dissertation therefore have
two scopes: first, to have a meaningful clinical impact through expanding the knowledge
of psychiatric entities while studying them as complex systems; second, to introduce and
discuss improvements of existing methods to reach such meaningful clinical impact.In this
discussion, I will detail how those aims were reached.
The first part of works constituting this dissertation is composed of the analyses of
psychiatric constructs of high clinical importance.
First, I analyzed empathy, which is a crucial construct in psychiatry, since its absence is
associated with several disorders, and I used the network approach to explore the connec-
tivity of empathy components such as they are represented in the Interpersonal Reactivity
Index (Briganti et al., 2018). I wanted to recover if some components of empathy are more
important than others, since several models put the affective components of empathy on top
of the others. This hypothesis was supported in my findings, since the affective component
235
of empathy was more important in the self-determination of the empathy network than other
components. In this paper, I showed how network methods for recovering the underlying
clustering structure of a data set are as effective as more established methods such as factor
analysis, while keeping a complex system view to the construct at hand. I also proposed
a way to interpret the connections among components of a psychiatric construct (that is,
two traits predict each other in the observed group and could therefore predict each other
in the individual), which could be interpreted to be fundamentally different from symptoms.
The most important caveat I observed in analyzing a full psychometric scale of a psychiatric
construct as a network is that traits from the same community tend to resemble each other
and are therefore redundant: it is therefore more difficult to infer from the network structure
with many redundant variables.
Second, I analyzed the construct of self-worth: I proposed a way forward for analyzing
constructs from psychometric scales such as analyzing domains instead of items (Briganti
et al., 2019). A domain is understood a community or cluster of items (and therefore is
defined with a data-driven method such a community detection algorithm). I found that
several domains of self-worth, including the self-worth drawn from religious beliefs, can be
unconnected, and therefore self-worth itself can be a very heterogeneous construct. This is
an important added value of analyzing domains, since the complex view allows for demon-
strating how theoretically uniform psychiatric entities can in fact be heterogenous: on an
interpersonal level, this can be interpreted as the possibility for different individuals to ex-
hibit different parts of a same psychiatric entity (a construct or a mental disorder).
Third, I analyzed resilience, an important psychiatric construct at the heart of therapy
outcomes (Briganti and Linkowski, 2019b): the psychometric tool at hand having had mul-
tiple revisions regarding the domain structure of the scale, it was an interesting opportunity
to test whether domains themselves can be redundant : that is why I adopted a data-driven
method to compare to choose a more parsimonious domain subdivision and test whether
using conventional factor analysis method such subdivision is a better fit for the data.
236
Fourth, I studied narcissistic personality, a topic of ever-growing interest in psychiatric
research. I performed a network analysis to detect which components of narcissistic person-
ality were the most important: entitlement, authority and superiority were found to be the
most contributing components to narcissistic personality. In this work I heavily relied on the
use of centrality measures (Briganti and Linkowski, 2019a) to infer the relative importance
of components and discussed their use.
Fifth, I tackled the important construct of alexithymia as represented in the two most
widely used measures, the Toronto Alexithymia Scale (which does not include the component
of fantasizing) and the Bermond-Vorst Alexithymia Questionnaire (which includes fantasiz-
ing) using both frequentist and Bayesian approaches (Briganti and Linkowski, 2019c). In
both constructs, the difficulty describing feelings was found to be the most important com-
ponent of the construct. These findings greatly contribute to support the fact that inference
from network structures is highly replicable across samples and different psychometric tools
(Fried et al., 2018). In the latter of the two alexithymia papers, I used centrality measures
to tackle the issue of redundancy and choose the most important item in each domain before
constructing a network structure: this is an additional method that can be used in large
scales to avoid having to interpret too many nodes.
Sixth, I used the Bayesian approach to Gaussian Graphical Models to investigate autistic
traits. Traits related to social skills are the most interconnected items in the network.
Sex differences were found between female and male subjects: using Bayesian methods to
estimate network structures lead to an easier interpretation of results, mainly because of
tools such as Bayesian factors that allow the researcher to dispose of supporting evidence for
a given parameter.
The second part of works composing this dissertation tackles the domain of affective
disorders in both a healthy sample and a sample of severe inpatients.
The construct of depressive symptoms such as represented in the Zung Depression Scale
was analyzed through the lenses of both undirected and directed network using both fre-
237
quentist and Bayesian approaches (Briganti et al., 2020b). Lack of focus was identified to
be the most interconnected symptom in the depression network in a healthy sample. In this
work I introduced the comparison between frequentist and Bayesian methods in the case of
network structures. Although stemming from different concepts, the two methods can be
combined in order to study different aspects of a mental disorder: the frequentist approach
is better suited to study the connectivity among symptoms, while the Bayesian approach is
better suited for (causal) inference.
In a sample of severe manic patients, I analyzed the connectivity of manic symptoms in
three time points (start, middle and end of hospital stay) during a psychiatric commitment.
Elevated mood was identified to be the most interconnected symptoms at the three time
points. I also estimated the network of temporal effects (Granger-causal effects), where
Elevated Mood emerged as an item causing many other symptoms over time. I used the
optimised Graphical Vector Autoregressive Model for panel data to infer the temporal effects
among manic symptoms.
In the last part, I proposed a methodological way forward for the important problem
of community detection in networks –a crucial way of addressing the heterogeneity of psy-
chiatric symptomatology. This methodological work stems from the previous works of the
dissertation and answers the relevant clinical question of identifying sub-types of disorders
(e.g. a manic patient may present with only a subset of symptoms and still be considered
a manic patients), and this particular complex view of the mental disorder challenges cur-
rent psychiatric diagnosis, which is based on summing the presence of several categories of
symptoms.
14.1 Future directions
I formulate the following recommendation for the future studies of psychiatric entities as
networks. Although sometimes too technical (Fried, 2020), cross-sectional reports of net-
238
work structures are useful to expand the current knowledge of psychiatric entities which
is necessary to build better models (Robinaugh et al., 2019). Cross-sectional networks of
observed groups are therefore the preferable way of studying the connection in a population
before trying to address the individual. To support the study of temporal data (to answer
the question which symptom influences which symptom over time), however, we need more
established models that stem from existing and widely used psychometric tools for clinical
practice.
It is however in the individual patient that lie the most promising advances of network
psychiatry: with temporal data (obtained either at the bedside or in between consultations)
psychiatrists will be able to monitor the evolution of individual symptoms and how they
affect others over time. The efficacy of identifying targets for clinical intervention will in the
future have to be tested with clinical trials. This will serve as a fundamental basis to achieve
what is called precision psychiatry –that is, optimize the right treatment for the right patient
based on a series of parameters, including its dynamic symptomatology.
14.2 Conclusion
Inspired by the ever-growing development of network psychometrics in the scientific liter-
ature, this dissertation aimed to translate and optimize the use of a selected number of
methods in complex systems and Bayesian Artificial Intelligence proposed by the network
approach to a number of mental constructs and disorders relevant to psychiatric practice.
It is worthy of note that my reports of network structures of the constructs and disorders
were realised using validated tools that are widely used in clinical practice, and were the first
to be studied as network structures in the literature. First, I introduced the methods that
the reader needs to understand this dissertation; second, I studied psychiatric constructs
(psychiatric entities that are not defined as mental disorders) as networks; third, I studied
unipolar depression and mania as network structures; fourth, I introduced a way forward for
239
the study of communities in networks by comparing the rationale and the performance of
several community-detection algorithm on a clinical sample.
Network psychiatry is a promising field that will be able to translate fundamental re-
search that aims to better understand mental disorders to clinical practice: its complex
view on mental disorders will in the future be able to integrate the study of symptoms with
other important variables, such as genes, neuroanatomy and environmental factors to achieve
precision psychiatry.
240
Bibliography
Aburn, G., Gott, M., and Hoare, K. (2016). What is resilience? An Integrative Review of
the empirical literature. Journal of Advanced Nursing, 72(5):980–1000.
Ackerman, R. A., Witt, E. A., Donnellan, M. B., Trzesniewski, K. H., Robins, R. W., and
Kashy, D. A. (2011). What does the narcissistic personality inventory really measure?
Assessment, 18(1):67–87.
Allison, C., Auyeung, B., and Baron-Cohen, S. (2012). Toward brief “red flags” for autism
screening: the short autism spectrum quotient and the short quantitative checklist in
1,000 cases and 3,000 controls. Journal of the American Academy of Child & Adolescent
Psychiatry, 51(2):202–212.
American Psychiatric Association (2013). Diagnostic and Statistical Manual of Mental Dis-
orders. American Psychiatric Association, fifth edition.
Angst, J. and Marneros, A. (2001). Bipolarity from ancient to modern times:: conception,
birth and rebirth. Journal of affective disorders, 67(1-3):3–19.
Armour, C., Fried, E. I., Deserno, M. K., Tsai, J., and Pietrzak, R. H. (2017). A network
analysis of DSM-5 posttraumatic stress disorder symptoms and correlates in U.S. military
veterans. Journal of Anxiety Disorders, 45:49–59.
Bagby, R. M., Parker, J. D. A., and Taylor, G. J. (1994). The twenty-item Toronto Alex-
241
ithymia scale—I. Item selection and cross-validation of the factor structure. Journal of
Psychosomatic Research, 38(1):23–32.
Baron-Cohen, S. (2010). Progress in brain research: Sex differences in the human brain,
their underpinnings and implications. Progress in Brain Research, 186:167–175.
Baron-Cohen, S., Cassidy, S., Auyeung, B., Allison, C., Achoukhi, M., Robertson, S., Pohl,
A., and Lai, M.-C. (2014). Attenuation of Typical Sex Differences in 800 Adults with
Autism vs. 3,900 Controls. PLoS ONE, 9(7).
Baron-Cohen, S., Wheelwright, S., Skinner, R., Martin, J., and Clubley, E. (2001).
The autism-spectrum quotient (AQ): evidence from Asperger syndrome/high-functioning
autism, males and females, scientists and mathematicians. Journal of Autism and Devel-
opmental Disorders, 31(1):5–17.
Beam, A. L., Manrai, A. K., and Ghassemi, M. (2020). Challenges to the reproducibility
of machine learning models in heath care. Journal of the American Medical Association,
323(4):305–306.
Beard, C., Millner, A. J., Forgeard, M. J. C., Fried, E. I., Hsu, K. J., Treadway, M. T.,
Leonard, C. V., Kertz, S. J., and Bjorgvinsson, T. (2016). Network analysis of depres-
sion and anxiety symptom relationships in a psychiatric sample. Psychological Medicine,
46(16):3359–3369.
Berkson, J. (1946). Limitations of the application of fourfold table analysis to hospital data.
Biometrics Bulletin, 2(3):47–53.
Blair, R. J. R. (2005). Responding to the emotions of others: Dissociating forms of empathy
through the study of typical and psychiatric populations. Consciousness and Cognition,
14(4):698–718.
242
Blanken, T. F., Van Der Zweerde, T., Van Straten, A., Van Someren, E. J., Borsboom, D.,
and Lancee, J. (2019). Introducing Network Intervention Analysis to Investigate Sequen-
tial, Symptom-Specific Treatment Effects: A Demonstration in Co-Occurring Insomnia
and Depression. Psychotherapy and Psychosomatics, 88(1):52–54.
Boccaletti, S., Latora, V., Moreno, Y., Chavez, M., and Hwang, D. U. (2006). Complex
networks: Structure and dynamics. Physics Reports, 424(4):175–308.
Bohart, A. C. and Greenberg, L. (1997). Empathy Reconsidered: New Directions in Psy-
chotherapy. APA.
Boldero, J. M., Bell, R. C., and Davies, R. C. (2015). The Structure of the Narcissistic
Personality Inventory With Binary and Rating Scale Items. Journal of Personality As-
sessment, 97(6):626–637.
Bonfiglio, N. S., Renati, R., Hjemdal, O., and Friborg, O. (2016). The resilience scale for
adults in italy: A validation study comparing clinical substance abusers with a nonclinical
sample. Psychology of Addictive Behaviors: Journal of the Society of Psychologists in
Addictive Behaviors, 30(4):509–515.
Borsboom, D. (2008). Psychometric perspectives on diagnostic systems. Journal of Clinical
Psychology, 64(9):1089–1108.
Borsboom, D. (2017). A network theory of mental disorders. World Psychiatry, 16(1):5–13.
Borsboom, D. and Cramer, A. O. (2013). Network Analysis: An Integrative Approach to
the Structure of Psychopathology. Annual Review of Clinical Psychology, 9(1):91–121.
Boschloo, L., Borkulo, C. D. v., Borsboom, D., and Schoevers, R. A. (2016). A Prospective
Study on How Symptoms in a Network Predict the Onset of Depression. Psychotherapy
and Psychosomatics, 85(3):183–184.
243
Braun, S., Rosseel, Y., Kempenaers, C., Loas, G., and Linkowski, P. (2015). Self-report of
empathy: a shortened French adaptation of the Interpersonal Reactivity Index (IRI) using
two large Belgian samples. Psychological Reports, 117(3):735–753.
Briganti, G., Fried, E. I., and Linkowski, P. (2019). Network analysis of Contingencies of
Self-Worth Scale in 680 university students. Psychiatry Research, 272:252–257.
Briganti, G., Kempenaers, C., Braun, S., Fried, E. I., and Linkowski, P. (2018). Network
analysis of empathy items from the interpersonal reactivity index in 1973 young adults.
Psychiatry Research, 265:87–92.
Briganti, G., Kornreich, C., and Linkowski, P. (2020a). A network structure of manic symp-
toms. submitted.
Briganti, G. and Le Moine, O. (2020). Artificial intelligence in medicine: Today and tomor-
row. Frontiers in Medicine, 7:27.
Briganti, G. and Linkowski, P. (2019a). Exploring network structure and central items of
the narcissistic personality inventory. International Journal of Methods in Psychiatric
Research, n/a:e1810.
Briganti, G. and Linkowski, P. (2019b). Item and domain network structures of the Resilience
Scale for Adults in 675 university students. Epidemiology and Psychiatric Sciences, pages
1–9.
Briganti, G. and Linkowski, P. (2019c). Network approach to items and domains from the
toronto alexithymia scale. Psychological Reports, page 0033294119889586.
Briganti, G. and Linkowski, P. (2019d). Une nouvelle approche ontologique et statistique des
constructions et maladies mentales : introduction a la psychiatrie des networks. PsyArXiv.
Briganti, G., Scutari, M., and Linkowski, P. (2020b). Network structures of symptoms from
the zung depression scale. Psychological Reports, page 0033294120942116.
244
Bryant, R. A., Creamer, M., O’Donnell, M., Forbes, D., McFarlane, A. C., Silove, D., and
Hadzi-Pavlovic, D. (2017). Acute and Chronic Posttraumatic Stress Symptoms in the
Emergence of Posttraumatic Stress Disorder: A Network Analysis. JAMA Psychiatry,
74(2):135–142.
Cain, N. M., Pincus, A. L., and Ansell, E. B. (2008). Narcissism at the crossroads: phenotypic
description of pathological narcissism across clinical theory, social/personality psychology,
and psychiatric diagnosis. Clinical Psychology Review, 28(4):638–656.
Chen, J. and Chen, Z. (2008). Extended Bayesian information criteria for model selection
with large model spaces. Biometrika, 95(3):759–771.
Chmitorz, A., Kunzler, A., Helmreich, I., Tuscher, O., Kalisch, R., Kubiak, T., Wessa, M.,
and Lieb, K. (2018). Intervention studies to foster resilience - A systematic review and
proposal for a resilience framework in future intervention studies. Clinical Psychology
Review, 59:78–100.
Cliffordson, C. (2002). The hierarchical structure of empathy: Dimensional organization and
relations to social functioning. Scandinavian Journal of Psychology, 43(1):49–59.
Constantino, J. N. and Todd, R. D. (2003). Autistic traits in the general population: a twin
study. Archives of general psychiatry, 60(5):524–530.
Correa, M., Bielza, C., and Pamies-Teixeira, J. (2009). Comparison of bayesian networks
and artificial neural networks for quality detection in a machining process. Expert systems
with applications, 36(3):7270–7279.
Corry, N., Merritt, R. D., Mrug, S., and Pamp, B. (2008). The factor structure of the
Narcissistic Personality Inventory. Journal of Personality Assessment, 90(6):593–600.
Costantini, G., Richetin, J., Borsboom, D., Fried, E. I., Rhemtulla, M., and Perugini, M.
245
(2015). Development of Indirect Measures of Conscientiousness: Combining a Facets
Approach and Network Analysis. European Journal of Personality, 29(5):548–567.
Cramer, A. O. J., Borkulo, C. D. v., Giltay, E. J., Maas, H. L. J. v. d., Kendler, K. S.,
Scheffer, M., and Borsboom, D. (2016). Major Depression as a Complex Dynamic System.
PLOS ONE, 11(12):e0167490.
Crocker, J., Luhtanen, R. K., Cooper, M. L., and Bouvrette, A. (2003). Contingencies of
self-worth in college students: theory and measurement. Journal of Personality and Social
Psychology, 85(5):894–908.
Csardi, G. and Nepusz, T. (2006). The igraph software package for complex network research.
InterJournal, Complex Systems:1695.
Curtiss, J., Fulford, D., Hofmann, S. G., and Gershon, A. (2019). Network dynamics of
positive and negative affect in bipolar disorder. Journal of affective disorders, 249:270–
277.
Dalege, J., Borsboom, D., van Harreveld, F., and van der Maas, H. L. J. (2017). Network
Analysis on Attitudes. Social Psychological and Personality Science, 8(5):528–537.
Dao, M. C., Everard, A., Aron-Wisnewsky, J., Sokolovska, N., Prifti, E., Verger, E. O.,
Kayser, B. D., Levenez, F., Chilloux, J., Hoyles, L., MICRO-Obes Consortium, Dumas,
M.-E., Rizkalla, S. W., Dore, J., Cani, P. D., and Clement, K. (2016). Akkermansia
muciniphila and improved metabolic health during a dietary intervention in obesity: rela-
tionship with gut microbiome richness and ecology. Gut, 65(3):426–436.
Davis, M. H. (1980). A multidimensional approach to individual differences in empathy.
JSAS Catalog of Selected Documents in Psychology, (85).
De Jong, I. J. and Rijksbaron, A. (2006). Sophocles And the Greek Language: Aspects on
Diction, Syntax and Pragmatics, volume 269. Brill.
246
de Vroege, L., Emons, W. H., Sijtsma, K., and van der Feltz-Cornelis, C. M. (2018). Psycho-
metric properties of the bermond–vorst alexithymia questionnaire (bvaq) in the general
population and a clinical population. Frontiers in psychiatry, 9:111.
Decety, J., Bartal, I. B.-A., Uzefovsky, F., and Knafo-Noam, A. (2016). Empathy as a driver
of prosocial behaviour: highly conserved neurobehavioural mechanisms across species.
Phil. Trans. R. Soc. B, 371(1686):20150077.
Decety, J. and Jackson, P. L. (2004). The functional architecture of human empathy. Be-
havioral and Cognitive Neuroscience Reviews, 3(2):71–100.
Demetriou, A., Spanoudis, G., Kazi, S., Mouyi, A., Zebec, M. S., Kazali, E., Golino, H.,
Bakracevic, K., and Shayer, M. (2017). Developmental Differentiation and Binding of
Mental Processes with g through the Life-Span. Journal of Intelligence, 5(2):23.
Deserno, M. K., Borsboom, D., Begeer, S., and Geurts, H. M. (2017). Multicausal systems ask
for multicausal approaches: A network perspective on subjective well-being in individuals
with autism spectrum disorder. Autism, 21(8):960–971.
Di Pierro, R., Costantini, G., Benzi, I. M. A., Madeddu, F., and Preti, E. (2019). Grandiose
and entitled, but still fragile: A network analysis of pathological narcissistic traits. Per-
sonality and Individual Differences, 140:15–20.
Elliott, H., Jones, P. J., and Schmidt, U. (2019). Central Symptoms Predict Posttreatment
Outcomes and Clinical Impairment in Anorexia Nervosa: A Network Analysis. Clinical
Psychological Science.
Elliott, R., Bohart, A. C., Watson, J. C., and Greenberg, L. S. (2011). Empathy. Psy-
chotherapy (Chicago, Ill.), 48(1):43–49.
Emmons, R. A. (1987). Narcissism: theory and measurement. Journal of Personality and
Social Psychology, 52(1):11–17.
247
Epskamp, S. (2019). Psychometric network models from time-series and panel data.
Epskamp, S., Borsboom, D., and Fried, E. I. (2017a). Estimating psychological networks
and their accuracy: A tutorial paper. Behavior Research Methods, pages 1–18.
Epskamp, S., Cramer, A. O. J., Waldorp, L. J., Schmittmann, V. D., and Borsboom, D.
(2012). qgraph: Network Visualizations of Relationships in Psychometric Data. Journal
of Statistical Software, 48(4).
Epskamp, S. and Fried, E. I. (2018). A tutorial on regularized partial correlation networks.
Psychological Methods, 23(4):617–634.
Epskamp, S., Maris, G. K. J., Waldorp, L. J., and Borsboom, D. (2018). Network Psycho-
metrics. arXiv:1609.02818 [stat]. arXiv: 1609.02818.
Epskamp, S., Rhemtulla, M., and Borsboom, D. (2017b). Generalized Network Psychomet-
rics: Combining Network and Latent Variable Models. Psychometrika, 82(4):904–927.
Farkas, I., Abel, D., Palla, G., and Vicsek, T. (2007). Weighted network modules. New
Journal of Physics, 9(6):180.
Ferri, S. L., Abel, T., and Brodkin, E. S. (2018). Sex Differences in Autism Spectrum
Disorder: A Review. Current psychiatry reports, 20(2):9.
Fonseca-Pedrero, E., Ortuno, J., Debbane, M., Chan, R. C. K., Cicero, D., Zhang, L. C.,
Brenner, C., Barkus, E., Linscott, R. J., Kwapil, T., Barrantes-Vidal, N., Cohen, A.,
Raine, A., Compton, M. T., Tone, E. B., Suhr, J., Inchausti, F., Bobes, J., Fumero,
A., Giakoumaki, S., Tsaousis, I., Preti, A., Chmielewski, M., Laloyaux, J., Mechri, A.,
Aymen Lahmar, M., Wuthrich, V., Larøi, F., Badcock, J. C., Jablensky, A., Isvoranu,
A. M., Epskamp, S., and Fried, E. I. (2018). The Network Structure of Schizotypal
Personality Traits. Schizophrenia Bulletin.
248
Freeman, L. C. (1978). Centrality in social networks conceptual clarification. Social Networks,
1(3):215–239.
Friborg, O., Hjemdal, O., Rosenvinge, J. H., and Martinussen, M. (2003). A new rating scale
for adult resilience: what are the central protective resources behind healthy adjustment?
International Journal of Methods in Psychiatric Research, 12(2):65–76.
Friborg, O., Hjemdal, O., Rosenvinge, J. H., Martinussen, M., Aslaksen, P. M., and Flaten,
M. A. (2006a). Resilience as a moderator of pain and stress. Journal of Psychosomatic
Research, 61(2):213–219.
Friborg, O., Martinussen, M., and Rosenvinge, J. H. (2006b). Likert-based vs. semantic
differential-based scorings of positive psychological constructs: A psychometric comparison
of two versions of a scale measuring resilience. Personality and Individual Differences,
40(5):873–884.
Fried, E. I. (2020). Lack of theory building and testing impedes progress in the factor and
network literature.
Fried, E. I., Borkulo, C. D. v., Cramer, A. O. J., Boschloo, L., Schoevers, R. A., and
Borsboom, D. (2017). Mental disorders as networks of problems: a review of recent
insights. Social Psychiatry and Psychiatric Epidemiology, 52(1):1–10.
Fried, E. I. and Cramer, A. O. J. (2017). Moving Forward: Challenges and Directions
for Psychopathological Network Theory and Methodology. Perspectives on Psychological
Science: A Journal of the Association for Psychological Science, 12(6):999–1020.
Fried, E. I., Eidhof, M. B., Palic, S., Costantini, G., Huisman-van Dijk, H. M., Bockt-
ing, C. L. H., Engelhard, I., Armour, C., Nielsen, A. B. S., and Karstoft, K.-I. (2018).
Replicability and Generalizability of Posttraumatic Stress Disorder (PTSD) Networks: A
Cross-Cultural Multisite Study of PTSD Symptoms in Four Trauma Patient Samples.
Clinical Psychological Science.
249
Fried, E. I., Epskamp, S., Nesse, R. M., Tuerlinckx, F., and Borsboom, D. (2016). What are
’good’ depression symptoms? comparing the centrality of dsm and non-dsm symptoms of
depression in a network analysis. Journal of affective disorders, 189:314–320.
Fried, E. I. and Nesse, R. M. (2015). Depression is not a consistent syndrome: An investi-
gation of unique symptom patterns in the STAR*D study. Journal of Affective Disorders,
172:96–102.
Friedman, J., Hastie, T., and Tibshirani, R. (2014a). glasso: Graphical lasso- estimation of
Gaussian graphical models.
Friedman, J., Hastie, T., and Tibshirani, R. (2014b). glasso: Graphical lasso- estimation of
Gaussian graphical models.
Fritz, J., Fried, E. I., Goodyer, I. M., Wilkinson, P. O., and van Harmelen, A.-L. (2018).
A Network Model of Resilience Factors for Adolescents with and without Exposure to
Childhood Adversity. Scientific Reports, 8(1):15774.
Fruchterman, T. M. J. and Reingold, E. M. (1991). Graph drawing by force-directed place-
ment. Software: Practice and Experience, 21(11):1129–1164.
Galderisi, S., Rucci, P., Kirkpatrick, B., Mucci, A., Gibertoni, D., Rocca, P., Rossi, A.,
Bertolino, A., Strauss, G. P., Aguglia, E., Bellomo, A., Murri, M. B., Bucci, P., Carpiniello,
B., Comparelli, A., Cuomo, A., Berardis, D. D., Dell’Osso, L., Fabio, F. D., Gelao, B.,
Marchesi, C., Monteleone, P., Montemagni, C., Orsenigo, G., Pacitti, F., Roncone, R., San-
tonastaso, P., Siracusano, A., Vignapiano, A., Vita, A., Zeppegno, P., and Maj, M. (2018).
Interplay Among Psychopathologic Variables, Personal Resources, Context-Related Fac-
tors, and Real-life Functioning in Individuals With Schizophrenia: A Network Analysis.
JAMA Psychiatry, 75(4):396–404.
Geiger, D., Verma, T., and Pearl, J. (1990). d-separation: From theorems to algorithms. In
Machine Intelligence and Pattern Recognition, volume 10, pages 139–148. Elsevier.
250
Gelman, A., Goodrich, B., Gabry, J., and Vehtari, A. (2019). R-squared for bayesian regres-
sion models. The American Statistician, 73(3):307–309.
Geng, L. and Jiang, T. (2013). Contingencies of Self-Worth Moderate the Effect of Specific
Self-Esteem on Self-Liking Or Self-Competence. Social Behavior and Personality: an
international journal, 41(1):95–107.
Gluck, T. M., Knefel, M., and Lueger-Schuster, B. (2017). A network analysis of anger,
shame, proposed ICD-11 post-traumatic stress disorder, and different types of childhood
trauma in foster care settings in a sample of adult survivors. European Journal of Psy-
chotraumatology, 8(sup3):1372543.
Goldberg, D. (2011). The heterogeneity of “major depression”. World Psychiatry, 10(3):226–
228.
Golino, H. F. and Demetriou, A. (2017). Estimating the dimensionality of intelligence like
data using Exploratory Graph Analysis. Intelligence, 62:54–70.
Golino, H. F. and Epskamp, S. (2017). Exploratory graph analysis: A new approach for esti-
mating the number of dimensions in psychological research. PLOS ONE, 12(6):e0174035.
Granger, C. W. (1969). Investigating causal relations by econometric models and cross-
spectral methods. Econometrica: journal of the Econometric Society, pages 424–438.
Greenland, S., Pearl, J., and Robins, J. M. (1999). Causal Diagrams for Epidemiologic
Research. Epidemiology, 10(1):37.
Gross, A. L., Tommet, D., D’Aquila, M., Schmitt, E., Marcantonio, E. R., Helfand, B.,
Inouye, S. K., Jones, R. N., and BASIL Study Group (2018). Harmonization of delirium
severity instruments: a comparison of the DRS-R-98, MDAS, and CAM-S using item
response theory. BMC medical research methodology, 18(1):92.
251
Hamilton, M. (1960). A rating scale for depression. Journal of Neurology, Neurosurgery,
and Psychiatry, 23:56–62.
Hanwella, R. and de Silva, V. A. (2011). Signs and symptoms of acute mania: a factor
analysis. BMC psychiatry, 11(1):137.
Haroz, E. E., Bolton, P., Gross, A., Chan, K. S., Michalopoulos, L., and Bass, J. (2016). De-
pression symptoms across cultures: an IRT analysis of standard depression symptoms us-
ing data from eight countries. Social Psychiatry and Psychiatric Epidemiology, 51(7):981–
991.
Hartley, S., Haddock, G., e Sa, D. V., Emsley, R., and Barrowclough, C. (2015). The influence
of thought control on the experience of persecutory delusions and auditory hallucinations
in daily life. Behaviour research and therapy, 65:1–4.
Haslbeck, J. M. B. and Fried, E. I. (2017). How predictable are symptoms in psychopatholog-
ical networks? A reanalysis of 18 published datasets. Psychological Medicine, 47(16):2767–
2776.
Haslbeck, J. M. B. and Waldorp, L. J. (2016). mgm: Structure Estimation for Time-Varying
Mixed Graphical Models in high-dimensional Data. arXiv preprint:1510.06871v2.
Hecht-Nielsen, R. (1990). On the algebraic structure of feedforward network weight spaces.
In Advanced Neural Computers, pages 129–135. Elsevier.
Hilland, E., Landrø, N. I., Kraft, B., Tamnes, C. K., Fried, E. I., Maglanoc, L. A., and
Jonassen, R. (2020). Exploring the Links between Specific Depression Symptoms and
Brain Structure: A Network Study. Psychiatry and Clinical Neurosciences.
Hjemdal, O., Friborg, O., Braun, S., Kempenaers, C., Linkowski, P., and Fossion, P. (2011a).
The Resilience Scale for Adults: Construct Validity and Measurement in a Belgian Sample.
International Journal of Testing, 11(1):53–70.
252
Hjemdal, O., Vogel, P. A., Solem, S., Hagen, K., and Stiles, T. C. (2011b). The relationship
between resilience and levels of anxiety, depression, and obsessive-compulsive symptoms
in adolescents. Clinical Psychology & Psychotherapy, 18(4):314–321.
Hoekstra, R. A., Bartels, M., Cath, D. C., and Boomsma, D. I. (2008). Factor struc-
ture, reliability and criterion validity of the Autism-Spectrum Quotient (AQ): a study in
Dutch population and patient groups. Journal of Autism and Developmental Disorders,
38(8):1555–1566.
Hoekstra, R. A., Vinkhuyzen, A. A. E., Wheelwright, S., Bartels, M., Boomsma, D. I.,
Baron-Cohen, S., Posthuma, D., and van der Sluis, S. (2011). The construction and
validation of an abridged version of the autism-spectrum quotient (AQ-Short). Journal of
Autism and Developmental Disorders, 41(5):589–596.
Holzinger, A., Langs, G., Denk, H., Zatloukal, K., and Muller, H. (2019). Causability and
explainability of artificial intelligence in medicine. WIREs Data Mining and Knowledge
Discovery, 9:e1312.
Hurst, R. M., Mitchell, J. T., Kimbrel, N. A., Kwapil, T. K., and Nelson-Gray, R. O. (2007).
Examination of the reliability and factor structure of the Autism Spectrum Quotient (AQ)
in a non-clinical sample. Personality and Individual Differences, 43(7):1938–1949.
Ising, E. (1925). Beitrag zur Theorie des Ferromagnetismus. Zeitschrift fur Physik, 31(1):253–
258.
James, R. (2011). Correlates of sexual self-esteem in a sample of substance-abusing women.
Journal of Psychoactive Drugs, 43(3):220–228.
Jones, P. (2017). networktools: Tools for identifying important nodes in networks. R package
version, 1(0).
253
Jones, P. J., Ma, R., and McNally, R. J. (2019). Bridge centrality: A network approach to
understanding comorbidity. Multivariate behavioral research, pages 1–15.
Jowkar, B., Friborg, O., and Hjemdal, O. (2010). Cross-cultural validation of the Resilience
Scale for Adults (RSA) in Iran. Scandinavian Journal of Psychology, 51(5):418–425.
Kalisch, M., Machler, M., Colombo, D., Maathuis, M. H., and Buhlmann, P. (2012). Causal
Inference Using Graphical Models with the R Package pcalg. Journal of Statistical Soft-
ware, 47(1):1–26.
Kass, R. E. and Raftery, A. E. (1995). Bayes Factors. Journal of the American Statistical
Association, 90(430):773–795.
Kempenaers, C., Braun, S., Delvaux, N., and Linkowski, P. (2017). The assessment of
autistic traits with the Autism Spectrum Quotient: Contribution of the French version to
its construct validity. European Review of Applied Psychology, 67(6):299–306.
Kendler, K. S., Zachar, P., and Craver, C. (2011). What kinds of things are psychiatric
disorders? Psychological medicine, 41(6):1143–1150.
Kloosterman, P. H., Keefer, K. V., Kelley, E. A., Summerfeldt, L. J., and Parker, J. D. A.
(2011). Evaluation of the factor structure of the Autism-Spectrum Quotient. Personality
and Individual Differences, 50(2):310–314.
Korb, K. B. and Nicholson, A. E. (2010). Bayesian artificial intelligence. CRC press.
Kose, S., Bora, E., Erermis, S., Ozbaran, B., Bildik, T., and Aydın, C. (2013). Broader
autistic phenotype in parents of children with autism: Autism Spectrum Quotient–Turkish
version. Psychiatry and Clinical Neurosciences, 67(1):20–27.
Kossakowski, J. J., Epskamp, S., Kieffer, J. M., Borkulo, C. D. v., Rhemtulla, M., and
Borsboom, D. (2016). The application of a network approach to Health-Related Quality
254
of Life (HRQoL): introducing a new method for assessing HRQoL in healthy adults and
cancer patients. Quality of Life Research, 25(4):781–792.
Krizan, Z. and Herlache, A. D. (2018). The Narcissism Spectrum Model: A Synthetic View
of Narcissistic Personality. Personality and Social Psychology Review, 22(1):3–31.
Kruijer, W., Behrouzi, P., Bustos-Korts, D., Rodrıguez-Alvarez, M. X., Mahmoudi, S. M.,
Yandell, B., Wit, E., and van Eeuwijk, F. A. (2020). Reconstruction of networks with
direct and indirect genetic effects. Genetics, 214(4):781–807.
Kruis, J. and Maris, G. (2016). Three representations of the Ising model. Scientific Reports,
6:34175.
Kubarych, T. S., Deary, I. J., and Austin, E. J. (2004). The Narcissistic Personality Inven-
tory: Factor structure in a non-clinical sample. Personality and Individual Differences,
36(4):857–872.
la Cour Karottki, N. F., Coello, K., Stanislaus, S., Melbye, S., Kjærstad, H. L., Sletved, K.
S. O., Kessing, L. V., and Vinberg, M. (2020). Sleep and physical activity in patients with
newly diagnosed bipolar disorder in remission, their first-degree unaffected relatives and
healthy controls. International Journal of Bipolar Disorders, 8:1–10.
Lambert, B. (2018). A student’s guide to Bayesian statistics. Sage.
Lane, R. D., Weihs, K. L., Herring, A., Hishaw, A., and Smith, R. (2015). Affective agnosia:
Expansion of the alexithymia construct and a new opportunity to integrate and extend
Freud’s legacy. Neuroscience & Biobehavioral Reviews, 55:594–611.
Lau, W. Y.-P., Gau, S. S.-F., Chiu, Y.-N., Wu, Y.-Y., Chou, W.-J., Liu, S.-K., and Chou,
M.-C. (2013). Psychometric properties of the Chinese version of the Autism Spectrum
Quotient (AQ). Research in Developmental Disabilities, 34(1):294–305.
255
Lauritzen, S. L. and Spiegelhalter, D. J. (1988). Local computations with probabilities
on graphical structures and their application to expert systems. Journal of the Royal
Statistical Society: Series B (Methodological), 50(2):157–194.
Liew, B. X. W., Scutari, M., Peolsson, A., Peterson, G., Ludvigsson, M. L., and Falla,
D. (2019). Investigating the causal mechanisms of symptom recovery in chronic whiplash
associated disorders using bayesian networks. The Clinical Journal of Pain, 35(8):647–655.
Liu, X., Faes, L., Kale, A. U., Wagner, S. K., Fu, D. J., Bruynseels, A., Mahendiran, T.,
Moraes, G., Shamdas, M., Kern, C., Ledsam, J. R., Schmid, M. K., Balaskas, K., Topol,
E. J., Bachmann, L. M., Keane, P. A., and Denniston, A. K. (2019). A comparison of deep
learning performance against health-care professionals in detecting diseases from medical
imaging: a systematic review and meta-analysis. The Lancet Digital Health, 1(6):e271–
e297.
Loas, G., Braun, S., Delhaye, M., and Linkowski, P. (2017). The measurement of alexithymia
in children and adolescents: Psychometric properties of the Alexithymia Questionnaire
for Children and the twenty-item Toronto Alexithymia Scale in different non-clinical and
clinical samples of children and adolescents. PloS One, 12(5):e0177982.
Lopez, A. D., Mathers, C. D., Ezzati, M., Jamison, D. T., and Murray, C. J., editors (2006).
Global Burden of Disease and Risk Factors. World Bank, Washington (DC).
Lucas, P. J., Van der Gaag, L. C., and Abu-Hanna, A. (2004). Bayesian networks in
biomedicine and health-care. Artificial intelligence in medicine, 30(3):201–214.
Luthar, S. S. (2006). Resilience in development: A synthesis of research across five decades.
In Developmental psychopathology: Risk, disorder, and adaptation, Vol. 3, 2nd ed, pages
739–795. John Wiley & Sons Inc, Hoboken, NJ, US.
Lutkepohl, H. (2005). New introduction to multiple time series analysis. Springer Science &
Business Media.
256
Marshall, G. N., Schell, T. L., and Miles, J. N. (2013). A multi-sample confirmatory factor
analysis of ptsd symptoms: what exactly is wrong with the dsm-iv structure? Clinical
psychology review, 33(1):54–66.
Marsman, M., Borsboom, D., Kruis, J., Epskamp, S., Bork, R. v., Waldorp, L. J., Maas, H.
L. J. v. d., and Maris, G. (2018). An Introduction to Network Psychometrics: Relating
Ising Network Models to Item Response Theory Models. Multivariate Behavioral Research,
53(1):15–35.
McNally, R. J., Mair, P., Mugno, B. L., and Riemann, B. C. (2017). Co-morbid obses-
sive–compulsive disorder and depression: a Bayesian network approach. Psychological
Medicine, 47(7):1204–1214.
Miettinen, O. S. and Caro, J. J. (1994). Foundations of medical diagnosis: what actually
are the parameters involved in bayes’ theorem? Statistics in medicine, 13(3):201–209.
Miller, J. D., Gentile, B., Wilson, L., and Campbell, W. K. (2013). Grandiose and vulnerable
narcissism and the DSM-5 pathological personality trait model. Journal of Personality
Assessment, 95(3):284–290.
Miller, J. D., Lynam, D. R., McCain, J. L., Few, L. R., Crego, C., Widiger, T. A., and
Campbell, W. K. (2016). Thinking Structurally About Narcissism: An Examination of the
Five-Factor Narcissism Inventory and Its Components. Journal of Personality Disorders,
30(1):1–18.
Mintz, Y. and Brodie, R. (2019). Introduction to artificial intelligence in medicine. Minimally
Invasive Therapy & Allied Technologies, 28(2):73–81.
Moffa, G., Catone, G., Kuipers, J., Kuipers, E., Freeman, D., Marwaha, S., Lennox, B. R.,
Broome, M. R., and Bebbington, P. (2017). Using Directed Acyclic Graphs in Epidemiolog-
ical Research in Psychosis: An Analysis of the Role of Bullying in Psychosis. Schizophrenia
Bulletin, 43(6):1273–1279.
257
Mohammadi, M. R., Zarafshan, H., and Ghasempour, S. (2012). Broader Autism Phenotype
in Iranian Parents of Children with Autism Spectrum Disorders vs. Normal Children.
Iranian Journal of Psychiatry, 7(4):157–163.
Moriguchi, Y. and Komaki, G. (2013). Neuroimaging studies of alexithymia: physical, affec-
tive, and social perspectives. BioPsychoSocial Medicine, 7(1):8.
Morote, R., Hjemdal, O., Martinez Uribe, P., and Corveleyn, J. (2017). Psychometric
properties of the Resilience Scale for Adults (RSA) and its relationship with life-stress,
anxiety and depression in a Hispanic Latin-American community sample. PloS One,
12(11):e0187954.
Mullarkey, M. C., Marchetti, I., and Beevers, C. G. (2018). Using Network Analysis to Iden-
tify Central Symptoms of Adolescent Depression. Journal of Clinical Child & Adolescent
Psychology, 0(0):1–13.
Murphy, K. P. (2012). Machine Learning: A Probabilistic Perspective. The MIT Press.
Palla, G., Derenyi, I., Farkas, I., and Vicsek, T. (2005). Uncovering the overlapping commu-
nity structure of complex networks in nature and society. Nature, 435(7043):814–818.
Pan, Y.-H., Wu, N., and Yuan, X.-B. (2019). Toward a better understanding of neuronal
migration deficits in autism spectrum disorders. Frontiers in Cell and Developmental
Biology, 7:205.
Paulhus, D. L. (1998). Interpersonal and intrapsychic adaptiveness of trait self-enhancement:
a mixed blessing? Journal of Personality and Social Psychology, 74(5):1197–1208.
Pearl, J. (1984). Intelligent search strategies for computer problem solving. Addision Wesley.
Pearl, J. and Russell, S. (2011). Bayesian networks.
Pearl, J. and Verma, T. S. (1995). A theory of inferred causation. In Studies in Logic and
the Foundations of Mathematics, volume 134, pages 789–811. Elsevier.
258
Pearl, R. L., White, M. A., and Grilo, C. M. (2014). Overvaluation of shape and weight as
a mediator between self-esteem and weight bias internalization among patients with binge
eating disorder. Eating Behaviors, 15(2):259–261.
Phillips, R. D., Wilson, S. M., Sun, D., VA Mid-Atlantic MIRECC Workgroup, and Morey,
R. (2018). Posttraumatic Stress Disorder Symptom Network Analysis in U.S. Military
Veterans: Examining the Impact of Combat Exposure. Frontiers in Psychiatry, 9:608.
Pincus, A. L., Ansell, E. B., Pimentel, C. A., Cain, N. M., Wright, A. G. C., and Levy,
K. N. (2009). Initial construction and validation of the Pathological Narcissism Inventory.
Psychological Assessment, 21(3):365–379.
Pisula, E., Kawa, R., Szostakiewicz, L., Lucka, I., Kawa, M., and Rynkiewicz, A. (2013).
Autistic traits in male and female students and individuals with high functioning autism
spectrum disorders measured by the Polish version of the Autism-Spectrum Quotient.
PloS One, 8(9):e75236.
Pons, P. and Latapy, M. (2005). Computing communities in large networks using random
walks. In International symposium on computer and information sciences, pages 284–293.
Springer.
Purgato, M., Gross, A. L., Betancourt, T., Bolton, P., Bonetto, C., Gastaldon, C., Gordon,
J., O’Callaghan, P., Papola, D., Peltonen, K., Punamaki, R.-L., Richards, J., Staples,
J. K., Unterhitzenberger, J., van Ommeren, M., de Jong, J., Jordans, M. J. D., Tol, W. A.,
and Barbui, C. (2018). Focused psychosocial interventions for children in low-resource
humanitarian settings: a systematic review and individual participant data meta-analysis.
The Lancet. Global Health, 6(4):e390–e400.
Pyszczynski, T., Greenberg, J., Solomon, S., Arndt, J., and Schimel, J. (2004). Why do
people need self-esteem? A theoretical and empirical review. Psychological Bulletin,
130(3):435–468.
259
Raskin, R. N. and Hall, C. S. (1979). A narcissistic personality inventory. Psychological
Reports, 45(2):590.
Reichardt, J. and Bornholdt, S. (2006). Statistical mechanics of community detection. Phys-
ical review E, 74(1):016110.
Reniers, R. L. E. P., Corcoran, R., Drake, R., Shryane, N. M., and Vollm, B. A. (2011).
The QCAE: A Questionnaire of Cognitive and Affective Empathy. Journal of Personality
Assessment, 93(1):84–95.
Robinaugh, D., Haslbeck, J. M. B., Waldorp, L., Kossakowski, J. J., Fried, E. I., Millner, A.,
McNally, R. J., van Nes, E. H., Scheffer, M., Kendler, K. S., and Borsboom, D. (2019).
Advancing the Network Theory of Mental Disorders: A Computational Model of Panic
Disorder.
Robinaugh, D. J., Millner, A. J., and McNally, R. J. (2016). Identifying highly influential
nodes in the complicated grief network. Journal of Abnormal Psychology, 125(6):747–757.
Ross, J. and Watling, C. (2017). Use of empathy in psychiatric practice: Constructivist
grounded theory study. BJPsych Open, 3(1):26–33.
Rosseel, Y. (2012). lavaan: An R Package for Structural Equation Modeling. Journal of
Statistical Software, 48(1):1–36.
Ruta, L., Mazzone, D., Mazzone, L., Wheelwright, S., and Baron-Cohen, S. (2012). The
Autism-Spectrum Quotient–Italian version: a cross-cultural confirmation of the broader
autism phenotype. Journal of Autism and Developmental Disorders, 42(4):625–633.
Ruzich, E., Allison, C., Smith, P., Watson, P., Auyeung, B., Ring, H., and Baron-Cohen,
S. (2015). Measuring autistic traits in the general population: a systematic review of the
Autism-Spectrum Quotient (AQ) in a nonclinical population sample of 6,900 typical adult
males and females. Molecular Autism, 6:2.
260
Ruzzano, L., Borsboom, D., and Geurts, H. M. (2015). Repetitive Behaviors in Autism and
Obsessive–Compulsive Disorder: New Perspectives from a Network Analysis. Journal of
Autism and Developmental Disorders, 45(1):192–202.
Schuetze, M., Rohr, C. S., Dewey, D., McCrimmon, A., and Bray, S. (2017). Reinforcement
learning in autism spectrum disorder. Frontiers in Psychology, 8:2035.
Scott, J., Bellivier, F., Manchia, M., Schulze, T., Alda, M., and Etain, B. (2020). Can
network analysis shed light on predictors of lithium response in bipolar i disorder? Acta
Psychiatrica Scandinavica.
Scutari, M. (2009). Learning bayesian networks with the bnlearn r package. arXiv preprint
arXiv:0908.3817.
Scutari, M. (2010). Learning Bayesian Networks with the bnlearn R Package. Journal of
Statistical Software, 35(1):1–22.
Scutari, M. and Denis, J.-B. (2015). Bayesian Networks: With Examples in R.
Sesen, M. B., Nicholson, A. E., Banares-Alcantara, R., Kadir, T., and Brady, M. (2013).
Bayesian networks for clinical decision support in lung cancer care. PloS one, 8(12):e82349.
Shen, Y., Zhang, L., Zhang, J., Yang, M., Tang, B., Li, Y., and Lei, K. (2018). Cbn:
Constructing a clinical bayesian network based on data from the electronic medical record.
Journal of biomedical informatics, 88:1–10.
Sifneos, P. E. (1972). Short-term Psychotherapy and Emotional Crisis. Harvard University
Press.
Silva, R. d. A. d., Mograbi, D. C., Camelo, E. V., Amadeo, L. N., Santana, C. M., Landeira-
Fernandez, J., and Cheniaux, E. (2018). The relationship between insight and affective
temperament in bipolar disorder: an exploratory study. Trends in psychiatry and psy-
chotherapy, 40(3):210–215.
261
Spearman, C. (1904). Measurement of association, part ii. correction of ‘systematic devia-
tions’. Am J Psychol, 15:88–101.
Spirtes, P., Glymour, C., and Scheines, R. (1993). Discovery Algorithms for Causally Suf-
ficient Structures. In Spirtes, P., Glymour, C., and Scheines, R., editors, Causation,
Prediction, and Search, Lecture Notes in Statistics, pages 103–162. Springer New York,
New York, NY.
Tibshirani, R. (2011). Regression shrinkage and selection via the lasso: a retrospective.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 73(3):273–
282.
Traag, V. A. and Bruggeman, J. (2009). Community detection in networks with positive
and negative links. Physical Review E, 80(3):036115.
Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236):433.
Van Bork, R., Wijsen, L. D., and Rhemtulla, M. (2017). Toward a Causal Interpretation of
the Common Factor Model. Disputatio, 9(47):581–601.
van Borkulo, C. D., Borsboom, D., Epskamp, S., Blanken, T. F., Boschloo, L., Schoevers,
R. A., and Waldorp, L. J. (2014). A new method for constructing networks from binary
data. Scientific Reports, 4:5918.
van Borkulo, C. D., Epskamp, S., and Millner, A. J. (2016). NetworkComparisonTest:
Statistical Comparison of Two Networks Based on Three Invariance Measures.
van der Maas, H. L. J., Dolan, C. V., Grasman, R. P. P. P., Wicherts, J. M., Huizenga,
H. M., and Raijmakers, M. E. J. (2006). A dynamical model of general intelligence: the
positive manifold of intelligence by mutualism. Psychological Review, 113(4):842–861.
van Heijst, B. F., Deserno, M. K., Rhebergen, D., and Geurts, H. M. (2019). Autism and
262
depression are connected: A report of two complimentary network studies. Autism, page
1362361319872373.
Venables, W. N. and Ripley, B. D. (2002). Modern Applied Statistics with S. Statistics and
Computing. Springer-Verlag, New York, 4 edition.
Verduijn, M., Peek, N., Rosseel, P. M., de Jonge, E., and de Mol, B. A. (2007). Prog-
nostic bayesian networks: I: Rationale, learning procedure, and clinical use. Journal of
Biomedical Informatics, 40(6):609–618.
Volkmar, F. R., State, M., and Klin, A. (2009). Autism and autism spectrum disorders:
diagnostic issues for the coming decade. Journal of Child Psychology and Psychiatry, and
Allied Disciplines, 50(1-2):108–115.
Von Kanel, R., Herr, R. M., Van Vianen, A. E. M., and Schmidt, B. (2017). Association
of adaptive and maladaptive narcissism with personal burnout: findings from a cross-
sectional study. Industrial Health, 55(3):233–242.
Vorst, H. C. and Bermond, B. (2001). Validity and reliability of the bermond–vorst alex-
ithymia questionnaire. Personality and individual differences, 30(3):413–434.
Wakabayashi, A., Baron-Cohen, S., Wheelwright, S., and Tojo, Y. (2006). The Autism-
Spectrum Quotient (AQ) in Japan: A cross-cultural comparison. Journal of Autism and
Developmental Disorders, 36(2):263–270.
Watters, C. A., Taylor, G. J., and Bagby, R. M. (2016a). Illuminating the theoretical
components of alexithymia using bifactor modeling and network analysis. Psychological
Assessment, 28(6):627–638.
Watters, C. A., Taylor, G. J., Quilty, L. C., and Bagby, R. M. (2016b). An Examination
of the Topology and Measurement of the Alexithymia Construct Using Network Analysis.
Journal of Personality Assessment, 98(6):649–659.
263
Wichers, M. and Groot, P. (2016). Critical Slowing Down as a Personalized Early Warning
Signal for Depression. Psychotherapy and Psychosomatics, 85(2):114–116.
Wild, B., Eichler, M., Friederich, H.-C., Hartmann, M., Zipfel, S., and Herzog, W. (2010).
A graphical vector autoregressive modelling approach to the analysis of electronic diary
data. BMC medical research methodology, 10(1):28.
Williams, D., Briganti, G., Linkowski, P., Mulder, J., and Rhemtulla, M. (2020). On testing
the null hypothesis of conditional independence: A bayesian reanalysis of four psycholog-
ical networks. in preparation.
Williams, D. R. (2018a). Bayesian estimation for gaussian graphical models: Structure
learning, predictability, and network comparisons.
Williams, D. R. (2018b). Bayesian Estimation for Gaussian Graphical Models: Structure
Learning, Predictability, and Network Comparisons.
Williams, D. R. and Mulder, J. (2019). BGGM: A R Package for Bayesian Gaussian Graph-
ical Models.
Williams, D. R., Rhemtulla, M., Wysocki, A. C., and Rast, P. (2019). On Nonregularized
Estimation of Psychological Networks. Multivariate behavioral research, 54(5):719–750.
Windle, G., Bennett, K. M., and Noyes, J. (2011). A methodological review of resilience
measurement scales. Health and Quality of Life Outcomes, 9:8.
Wing, L. (1988). Aspects of autism: Biological research. Royal College of Psychiatrists.
Wispe, L. (1986). The distinction between sympathy and empathy: To call forth a concept,
a word is needed. Journal of Personality and Social Psychology, 50(2):314–321.
Xu, Z.-Y., Zu, S., Xiang, Y.-T., Wang, N., Guo, Z.-H., Kilbourne, A. M., Brabban, A., King-
don, D., and Li, Z.-J. (2013). Associations of self-esteem, dysfunctional beliefs and coping
264
style with depression in patients with schizophrenia: a preliminary survey. Psychiatry
Research, 209(3):340–345.
Yang, Z., Algesheimer, R., and Tessone, C. J. (2016). A Comparative Analysis of Community
Detection Algorithms on Artificial Networks. Scientific Reports, 6.
Young, R. C., Biggs, J. T., Ziegler, V. E., and Meyer, D. A. (1978). A rating scale for mania:
reliability, validity and sensitivity. The British journal of psychiatry, 133(5):429–435.
Zung, W. W. (1965). A self-rating depression scale. Archives of General Psychiatry, 12:63–70.
265
Glossary
Alexithymia: difficulty identifying, analyzing or verbalizing emotions.
Autism: developmental disorder characterized with difficulties in social interactions and
communication.
Bayesian Network: a sub-type of network structure, composed of a Directed Acyclic Graph
and a probability distribution.
Bootstrapping: a procedure that uses random resampling with replacement to attribute a
measure of accuracy to sample estimates. It is used in stability analyses in network studies.
Centrality: a set of measures of the relative importance of a node in a network.
Community: a set of nodes that are strongly related, also called domain or cluster in net-
work science.
Construct: a set of variables used to express psychiatric phenomena.
Depression: also known as major depressive disorder or major depression, it is a mental
disorder mainly characterized with a pervasive low mood.
Directed Acyclic Graph: a network with directed edges and without loops (A −→ A) or
cycles (A −→ B −→ C).
Edge: also called, arc. It is a connection between two nodes in a network. It can be undi-
rected (A−B) or directed (A −→ B), and have a weight denoting the strength of association.
Empathy: the ability to understand someone’s feelings.
Exploratory Graph Analysis: a set of techniques to detect and investigate communities
in a network.
266
Gaussian Graphical Model: a partial correlation network, with undirected edges.
Graphical Vector Autoregressive Model: a model used to depict the dynamics of symp-
toms in panel data. Edges represent a temporal effect from one node to another (this is called
Granger causality).
Ising Model: a network structure estimated from binary data.
Item: a component in a questionnaire, usually presented as a sentence or question and
meant to measure a mental construct or disorder.
Mania: a period of abnormally elevated arousal, affect and energy. It is a set of symptoms
occurring within the context of a bipolar I disorder.
Markov blankets: a set of nodes that blocks all paths between two nodes A and B if
conditioned on.
Mental disorder: pattern of behavior that causes significant distress or danger to the in-
dividual.
Mixed Graphical Model: a graphical model in which variables of different kinds (contin-
uous, ordinal, binary) can be used as input.
Narcissism: the pursuit of gratification from vanity.
Network: a system of components (nodes) interacting with each other through connections
(edges).
Network analysis: a set of statistical techniques operating within the framework of net-
work theory.
Network theory: a theory that considers mental disorders to arise from a set of mutually
influencing symptoms.
Node: an entity, such as a symptom, interacting in a network.
Predictability: a measure of a node’s shared variance with surrounding nodes. It Is un-
derstood as an absolute measure of connectedness in a network.
Redundancy: the act of repeating information across several items that measure the same
thing.
267
Regularization: in the case of network analysis, a process of adding information to provide
a conservative network structure (with few edges).
Resilience: the ability to recover from difficulties.
Self-worth: an individual’s evaluation of their own worth.
Sign: an objective indication of the presence of an abnormal physiological state.
Symptom: a subjective phenomenon indicating the presence of a morbid state.
Syndrome: a set of symptoms or signs related to a morbid state and that can indicate or
orientate a diagnosis.
Topological overlap: the act of reducing the number of items in a psychometric question-
naire based on the resemblance between items.
268
Acknowledgments
I started working on my PhD during my fourth year of medical school. I was 21 years old.
I knew I was always going to be a doctor and a scientist since I was 10 years old, but I
never ended up in the right environment to do so. In December 2016, I met Professor Paul
Linkowski (or simply The Professor), and he invited me to a talk about networks. I knew
then that I had found that right environment for my scientific development. Professor, thank
you for providing me with all the motivation and close follow up a young person needs to
carry on such a big project from the start. Thank you for carrying me through four years of
research even though I never had funds for it. Thank you for all the never ending phone calls,
meet ups, thank you for sending me off to Amsterdam in order to learn network statistics,
thank you for guiding me towards the right path in my career: it has been an honour to
meet that one person that is never wrong. You are the scientific father that every student
should have, and that I hope I will one day be. Twenty years ago you already knew that big
data would be the biggest thing of the 21st century in science and medicine, and you made
sure I had all it took to translate such a challenge to psychiatry, the discipline you admire
and made me admire so much. I will always be grateful for these four incredible years.
My path, as a clinician interested in artificial intelligence that Professor Linkowski has
started, has provided me an exciting adventure over the past four years, but such an ad-
venture would not have been the same had I not had some other great protectors : the word
protector is not an overstatement, since I am sure I would have faced many more difficulties
had I not been protected by some other great mentors. Professor Olivier Le Moine, thank
269
you for educating me to medical informatics and giving me the possibility to teach medical
students at such an early moment in my career. Professors Lucio Scanu and Karolien Haese,
thank you for teaching me that research is not enough: one must always look at the big-
ger picture, which includes the patients themselves, healthcare structures and society as a
whole. Philippe Marchal, thank you for introducing me to the world of digital medicine and
giving me a space to discuss my ideas. Professor Michel Goldman, thank you for pushing
me to always think about how I can use my research to innovate in healthcare. Professor
Gustave Moonen, thank you for having challenged me on the impact of artificial intelligence
in medicine: my work would not have been the same if our debate at the Royal Academy
of Medicine never happened. Professors Daniele Marinazzo and Yves Rosseel, thank you
for our discussions over the implications of synergy and redundancy in networks. Donald
Williams, our countless discussions and your Bayesian mind have greatly contributed to this
dissertation: thank you for those and collaborating with me on such an interesting project.
Chantal Kempenaers, Stephanie Braun and Professor Joris Mulder, thank you for your pre-
cious contributions to my work. Professor Charles Kornreich and doctor Georgios Persefonis,
thank you for setting a framework in my workplace so I could finish this dissertation while
working full time. Professor Jerome Lechien, thank you for being my friend since the very
beginning and showing me the ropes of research when I was just 18 years old. Professor
Eiko Fried, thank you for having spent countless hours training me and reviewing my first
network papers: your talk about networks is why I chose this research path. Professor Marco
Scutari, thank you for training me in Bayesian artificial intelligence, since I had the opportu-
nity to learn from the very best. Professor Sandrine Ansermet, thank you for giving me the
opportunity to teach college students and granting me my very own first course in Lausanne.
Professor Pasquale Nardone, thank you for having supervised my master thesis on this topic.
Professor Alain Leveque, thank you for your support and having me included in the won-
derful scientific environment that is the Research Center for Epidemiology, Biostatistics and
Clinical Research at the School of Public Health of ULB. Professor Michele Dramaix-Wilmet,
270
thank you for coaching me in statistics during those very interesting afternoons for several
years: your help was essential when I was getting started.
Professor Alexandre Legrand, thank you for believing in my research project and giving it
the beautiful home of my alma mater, the University of Mons. Thank you for tolerating my
lack of patience. You probably do not know this, but you greatly contributed to my interest
in research when I listened to you for hundreds of hours teaching medical physiology.
Professor Pierre Manneback, thank you for all your support and advice in the past years.
Thank you for pushing me in creating bridges between the medical and the engineering
worlds.
Professor Samuel Leistedt, thank you for being the promotor of this work, going through
countless revisions of this long dissertation. You greatly inspired me towards the world of
psychiatry since my early days as a medical student, and you saw me through the end of
this journey.
Finally, I thank those whom I love and cherish (they will know who they are): this work
is just as yours as it is mine.
271