Complex Systems Research in

Psychology

Han L. J. van der Maas

2024-08-30
Table of contents

Prologue 6

1 Introduction 8
1.1 What are complex systems? . . . . . . . . . . . . . . . . . . . . 8
1.2 Emergentism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 The field of psychology . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 The art of simplification . . . . . . . . . . . . . . . . . . . . . . 14
1.5 A limited number of equilibria . . . . . . . . . . . . . . . . . . 16
1.6 Networks are everywhere . . . . . . . . . . . . . . . . . . . . . . 17
1.7 Methods for investigating complex systems . . . . . . . . . . . 18
1.8 Other work and sources . . . . . . . . . . . . . . . . . . . . . . 18
1.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 Chaos and unpredictability 20

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2 The population growth of rabbits . . . . . . . . . . . . . . . . . 20
2.3 Stable and unstable fixed points . . . . . . . . . . . . . . . . . 22
2.4 Limit cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.5 Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.6 Phase plot and bifurcation diagrams . . . . . . . . . . . . . . . 26
2.7 What did we learn? . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.8 Other maps and fractals . . . . . . . . . . . . . . . . . . . . . . 29
2.9 Detecting chaos in psychophysiological data . . . . . . . . . . . 30
2.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3 Transitions in complex systems 32

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2 Examples of transitions . . . . . . . . . . . . . . . . . . . . . . 33
3.3 Bifurcation and Catastrophe theory . . . . . . . . . . . . . . . 34
3.3.1 The quadratic case . . . . . . . . . . . . . . . . . . . . . 35
3.3.2 The fold catastrophe . . . . . . . . . . . . . . . . . . . . 36
3.3.3 The cusp catastrophe . . . . . . . . . . . . . . . . . . . 37
3.3.4 Examples of cusp models . . . . . . . . . . . . . . . . . 41
3.3.5 Higher-order catastrophes . . . . . . . . . . . . . . . . . 43
3.3.6 Other bifurcations . . . . . . . . . . . . . . . . . . . . . 44
3.4 Building catastrophe models . . . . . . . . . . . . . . . . . . . . 45
3.4.1 Mechanistic models . . . . . . . . . . . . . . . . . . . . . 45
3.4.2 Phenomenological models . . . . . . . . . . . . . . . . . 46
3.5 Testing catastrophe models . . . . . . . . . . . . . . . . . . . . 48
3.5.1 The catastrophe flags . . . . . . . . . . . . . . . . . . . 48
3.5.1.1 Sudden jump . . . . . . . . . . . . . . . . . . . 49
3.5.1.2 Multimodality . . . . . . . . . . . . . . . . . . 50
3.5.1.3 Inaccessibility . . . . . . . . . . . . . . . . . . 51
3.5.1.4 Divergence . . . . . . . . . . . . . . . . . . . . 52

2
 3.5.1.5 Hysteresis . . . . . . . . . . . . . . . . . . . . . 52
3.5.1.6 Anomalous variance, divergence of linear re-
sponse, and critical slowing down . . . . . . . 54
3.5.2 Fitting the cusp to cross-sectional data . . . . . . . . . 54
3.5.2.1 Cobb’s maximum likelihood approach . . . . . 54
3.5.2.2 Empirical examples . . . . . . . . . . . . . . . 58
3.5.2.3 Evaluation . . . . . . . . . . . . . . . . . . . . 59
3.6 Criticism of catastrophe theory . . . . . . . . . . . . . . . . . . 61
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4 Building dynamic system models 64

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2 Basic concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.1 Back to the logistic equation . . . . . . . . . . . . . . . 65
4.2.2 The Lotka—Volterra models . . . . . . . . . . . . . . . 67
4.2.3 Fitting models: Stochasticity versus noise . . . . . . . . 68
4.2.4 Back to the cusp . . . . . . . . . . . . . . . . . . . . . . 69
4.2.5 Bifurcation analysis . . . . . . . . . . . . . . . . . . . . 71
4.2.6 Spruce budworm outbreak model . . . . . . . . . . . . . 72
4.2.7 Evaluation of ecological modeling . . . . . . . . . . . . . 73
4.3 Psychological models . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3.1 Response time models . . . . . . . . . . . . . . . . . . . 76
4.3.2 Dyadic models . . . . . . . . . . . . . . . . . . . . . . . 78
4.3.2.1 Romeo and Juliet . . . . . . . . . . . . . . . . 78
4.3.2.2 The mathematics of marriage . . . . . . . . . . 81
4.3.3 The van Geert models . . . . . . . . . . . . . . . . . . . 82
4.3.4 The Pólya urn model of the third source . . . . . . . . . 85
4.3.5 The panic model . . . . . . . . . . . . . . . . . . . . . . 85
4.3.6 Neural models: Van der Pol and different time scales . . 88
4.3.7 Analogical modeling: Budworms and beers . . . . . . . 91
4.3.8 Cascading transitions in multifigure multistable perception 92
4.3.8.1 Interacting cusps . . . . . . . . . . . . . . . . . 92
4.3.8.2 Application to perception . . . . . . . . . . . . 95
4.4 Causal-loop diagrams . . . . . . . . . . . . . . . . . . . . . . . 97
4.5 Closing remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

5 Self-organization 101
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.2 Key examples from the natural sciences . . . . . . . . . . . . . 102
5.2.1 Physics . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.2.2 Chemistry . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.2.3 Biology . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2.4 Computer science . . . . . . . . . . . . . . . . . . . . . . 106
5.2.5 Neural networks . . . . . . . . . . . . . . . . . . . . . . 109
5.2.6 The concept of self-organization . . . . . . . . . . . . . 111
5.3 NetLogo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.3.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.3.2 A bit of NetLogo programming . . . . . . . . . . . . . . 114
5.3.2.1 Game of Life . . . . . . . . . . . . . . . . . . . 114
5.3.2.2 The Ising model . . . . . . . . . . . . . . . . . 115

3
 5.4 Self-organization in psychology and social systems . . . . . . . 115
5.4.1 The brain . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.4.2 Consciousness . . . . . . . . . . . . . . . . . . . . . . . . 117
5.4.3 Visual illusions . . . . . . . . . . . . . . . . . . . . . . . 118
5.4.4 Motor action . . . . . . . . . . . . . . . . . . . . . . . . 119
5.4.5 Robotics . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.4.6 Developmental processes . . . . . . . . . . . . . . . . . . 121
5.4.7 Psychological disorders . . . . . . . . . . . . . . . . . . . 122
5.4.8 Social relations . . . . . . . . . . . . . . . . . . . . . . . 123
5.4.9 Collective Intelligence . . . . . . . . . . . . . . . . . . . 124
5.4.10 Game theory . . . . . . . . . . . . . . . . . . . . . . . . 125
5.4.11 Self-organization in organizations . . . . . . . . . . . . . 125
5.5 Zooming out . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

6 Psychological Network Models 129

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.2 Network theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.2.1 Network concepts . . . . . . . . . . . . . . . . . . . . . . 130
6.2.2 Network types . . . . . . . . . . . . . . . . . . . . . . . 131
6.2.3 Network dynamics . . . . . . . . . . . . . . . . . . . . . 133
6.3 Psychological network models . . . . . . . . . . . . . . . . . . . 134
6.3.1 Mutualism model: The case of general intelligence . . . 135
6.3.1.1 The 𝑔 factor . . . . . . . . . . . . . . . . . . . 135
6.3.1.2 Mutualism model . . . . . . . . . . . . . . . . 137
6.3.1.3 Abnormal development . . . . . . . . . . . . . 139
6.3.1.4 The wiring of intelligence . . . . . . . . . . . . 140
6.3.2 Symptom networks . . . . . . . . . . . . . . . . . . . . . 141
6.3.3 Ising attitude model . . . . . . . . . . . . . . . . . . . . 143
6.3.3.1 Model setup . . . . . . . . . . . . . . . . . . . 145
6.3.3.2 Simulation . . . . . . . . . . . . . . . . . . . . 147
6.3.3.3 Learning . . . . . . . . . . . . . . . . . . . . . 149
6.3.3.4 The stability of attitudes and entropy measures 150
6.3.3.5 Tricriticality . . . . . . . . . . . . . . . . . . . 151
6.4 Psychometric network techniques . . . . . . . . . . . . . . . . . 151
6.4.1 Main techniques . . . . . . . . . . . . . . . . . . . . . . 152
6.4.2 Fitting the mutualism model . . . . . . . . . . . . . . . 153
6.4.3 Fitting Ising models . . . . . . . . . . . . . . . . . . . . 155
6.5 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.5.1 Psychological network modeling . . . . . . . . . . . . . . 157
6.5.2 Psychometric network analysis . . . . . . . . . . . . . . 159
6.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160

7 Sociophysics 162
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.2 Some famous examples . . . . . . . . . . . . . . . . . . . . . . . 163
7.2.1 Segregation . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.2.2 The evolution of language: The naming game . . . . . . 163
7.2.3 Cultural dynamics: The Axelrod model . . . . . . . . . 166

4
 7.3 Dynamics of opinions . . . . . . . . . . . . . . . . . . . . . . . . 167
7.3.1 Discrete opinion models . . . . . . . . . . . . . . . . . . 169
7.3.1.1 Voter models . . . . . . . . . . . . . . . . . . . 169
7.3.1.2 More discrete opinion models: Majority type
models . . . . . . . . . . . . . . . . . . . . . . 170
7.3.1.3 Social Impact theory . . . . . . . . . . . . . . 171
7.3.2 Continuous opinion models . . . . . . . . . . . . . . . . 172
7.3.2.1 Classic models . . . . . . . . . . . . . . . . . . 172
7.3.2.2 Bounded confidence . . . . . . . . . . . . . . . 173
7.3.3 Empirical verification . . . . . . . . . . . . . . . . . . . 176
7.4 Psychosocial models . . . . . . . . . . . . . . . . . . . . . . . . 176
7.4.1 Networks of attitude networks . . . . . . . . . . . . . . 177
7.4.1.1 The HIOM . . . . . . . . . . . . . . . . . . . . 177
7.4.1.2 Agent interactions: Information and attention 178
7.4.1.3 Algorithm . . . . . . . . . . . . . . . . . . . . 179
7.4.1.4 The persuasion paradox . . . . . . . . . . . . . 180
7.4.1.5 A counterintuitive prediction . . . . . . . . . . 181
7.4.1.6 Variants of the HIOM . . . . . . . . . . . . . . 183
7.4.2 Cascading transitions in other psychosocial systems . . 184
7.5 Psychosociophysics . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

8 Epilogue 187

References 194

5
Prologue

This book is intended for psychologists and social scientists interested in mod-
eling psychological processes using the tools of complex-systems research.
The book has three primary objectives. The first is to provide a comprehen-
sive overview of complex-systems research, with a particular emphasis on its
applications in psychology and the social sciences. The second is to provide
skills for complex-systems research. Lastly, it strives to foster critical thinking
regarding the potential applications of complex systems in psychology.
For many decades, with roots dating back to the 19th century, scientists have
been studying a wide variety of complex systems. Well-known examples in-
clude lasers, tornadoes, chemical oscillations, ant nests and flocks of birds.
Scientists have built mathematical and computational models of these com-
plex systems and developed techniques to study them.
Applying these techniques requires a great deal of mathematical and technical
knowledge, as well as deep understanding of the nature of the system. You
don’t just create a mathematical model off the top of your head. In addition,
testing such models requires extensive and reliable quantitative data. Applying
complex-systems theory to the behavioral and social sciences is therefore not
straightforward. Theories are often verbal, and quantitative measurement in
these sciences is a longstanding issue. While there has been some reasonable
progress over the past 150 years, it is fair to say that the behavioral and social
sciences are less mature than the “hard” sciences.
Despite these challenges, applying complex-systems theory to the behavioral
sciences is imperative. Whether we consider humans in isolation, the billions
of interacting neurons in the brain, or the social networks in which we find
ourselves, complexity is everywhere. We, with our complex brains embedded
in various hierarchies of social systems, are the ultimate complex systems.
I believe that we can only succeed in exploring the psychological system by
understanding its complexity. We need to apply the tools of complexity sci-
ence to psychology, which is in desperate need of breakthroughs. After all,
the modern world revolves around human beings who, through language and
thought, have created an unimaginably complex world. The greatest danger
now are humans ourselves, and progress in the field of psychology is necessary
and urgent.
This book requires study. New theoretical concepts are illustrated with sim-
ulations and examples. Running the simulations, studying the examples, and
solving the exercises will contribute to a deeper understanding of the mate-
rial. I have used the book’s content in a master’s course for research-minded
students in psychology. Readers should have some background in psychology
and its research methods.

6
I assume only pre-university knowledge of mathematics. An important prereq-
uisite is a basic knowledge of the programming language R. The book uses R
for the simulations and exercises. There are many online resources for learning
the basics of R. In addition to R, we will use NetLogo, but no prior knowledge
of NetLogo is expected. NetLogo is a dedicated programming language for
simulating complex natural and social phenomena. It is freely available for all
major computer platforms.
This book has been made available as open access courtesy of the Santa Fe
Institute Press, a non-profit publisher.1 Consistent with this approach, I often
cite open sources such as Wikipedia rather than proprietary ones. It should
be noted that open sources such as Wikipedia have the potential for content
changes, although this is unlikely in the contexts I’ve referenced. I also use
only open-source software for the examples and exercises.
It should also say that this book is more a book for psychologists who have
very limited knowledge of complex systems research than the other way around.
Experts in complex systems who wonder how it can be applied in psychology
may have to wait for another text.
I have written this book based on thirty-five years of scientific work in collab-
oration with fantastic colleagues and coauthors of many papers. I’m part of
the ecosystem of the psychology department, especially the wonderful meth-
ods section, of the University of Amsterdam. Also important is the Institute
for Advanced Study in Amsterdam, which has complex-systems research as
a central theme. In recent years, I’ve also been an external faculty member
at the renowned Santa Fe Institute in New Mexico. I am indebted to all of
them and to many other colleagues around the world. In my citations, I’ve
made an effort to acknowledge the extensive contributions to this vast field.
Nevertheless, I recognize that there may be omissions, for which I apologize.
Han L. J. van der Maas, Amsterdam 2024

1
I would like to express my gratitude to Sienna Latham, Zato Hebbert, Rachel Fudge,
David Krakauer, and Tasos Psychogyiopoulos for their invaluable contributions to the
online and print versions of the book. I would also like to thank Robert Goldstone, Mirta
Galesic, Henrik Olsson, and John Miller for their valuable reviews of the preliminary
version of the book.

7
1 Introduction

1.1 What are complex systems?

Some things in life are simple. When you push a block, it moves. Pushing
harder makes it move faster, and stopping the push stops its movement due to
friction. When you open the tap, water starts to flow. If you open it farther, it
flows faster, and if you close it, it stops. Cause-and-effect relationships like this
are clear and roughly linear. Such relationships are rare in psychology. Let’s
take fear as an example. A fear stimulus—for example, a barking dog—can
lead to fear and flight, but also to anger and attack. Whether the stimulus is In psychology, cause-and-effect
perceived as fearful might depend on subtle differences in context. It has also relationships are rarely simple, and
effects are often non-linear.
been debated whether the flight response precedes the feeling of fear or vice
versa. Fear could also suddenly change into a panic attack.
These difficulties are not unique to psychology. Many systems studied in
physics, chemistry, and biology show such complex behavior (Weaver 1948;
Krakauer 2024). They are complex systems. Although there is no full consen-
sus on the definition of a complex system (Ladyman, Lambert, and Wiesner
2013; Heylighen 2009), I believe the core aspects can be summarized as fol-
lows.
Complex systems exhibit emergent
Complex systems are made up of many smaller interacting subsystems, such behavior, meaning that the
as atoms, molecules, cells, neurons, and even entire organisms. I like the term interactions between the units of the
subsystems because the lower-level elements can themselves be complex sys- system result in global patterns or
tems.1 The interactions between subsystems can be of different kinds, but properties that do not occur in the
units themselves.
they are generally local, fast, and nonlinear. These interactions result in emer-
gent behavior. The emergent processes usually operate on a slower time scale.
A typical example, which will be discussed in more detail later, is the traffic
jam. Cars react mainly to cars in their vicinity, which can lead to global pat-
terns of congestion. Another example is magnetism which is not present in
any of the atoms of the magnet.
In general, these patterns emerge through self-organization. An example is Self-organization is a process in which
ants building an ant nest: no one ant oversees or directs this process. Rather, it some form of overall order or
coordination develops from the local
emerges from the local interactions between many ants.2 I will explain this in interactions between the parts of an
more detail in Chapter 5, but in a completely closed system, self-organization initially disordered system.

1
As Simon (1962) noted, atoms were once considered elementary particles, whereas in
modern nuclear physics they are themselves complex systems.
2
It is also argued that emergence is a consequence of symmetry breaking (Krakauer 2023).
Symmetry breaking occurs when a system transitions from a symmetric state to an asym-
metric state, resulting in the emergence of distinct properties or behaviors. An example of
symmetry breaking can be observed in the formation of snowflakes. Initially, ice crystals
have a symmetrical hexagonal shape due to the underlying molecular structure of water.
However, as the crystal continues to grow, environmental factors such as temperature
and humidity influence its growth pattern. Minute variations in these factors lead to
the breaking of initial symmetry and the formation of diverse and beautiful snowflake
structures.

8
would not be possible. Self-organization takes energy. The emergent patterns
in a complex system may be stable for some time, but often change suddenly.
The study of phase transitions or tipping points is therefore central to the
study of complex systems. They may also exhibit chaotic behavior, implying
that they can be fundamentally unpredictable, the weather being a notorious
example. Complex systems are open systems,
meaning that they use energy that
Let’s look at one famous case, the flocking of birds (figure 1.1). Flocks of birds they have absorbed for the
move in a beautiful choreography as they glide through the air, their forma- environment.
tions shifting and morphing as they twist and turn across the sky. Flocks are
well understood and easy to simulate, as we will see in Chapter 5. Flocks fulfill
all the criteria of a complex system (see Parisi 2023 for an extended analysis).
They are open systems, as birds use energy to fly. Each bird responds only
to birds in its local neighborhood. They follow roughly three rules: they try
to fly in the same direction as their neighbors, stay close to their neighbors,
and avoid collisions. These are fast and local interactions. If you watch some
videos of flocks of birds on the internet, you will see globally organized behav-
ior. This is a prime example of self-organization. There is evidently no one
bird in control ordering other birds to change direction. What you can also This globally organized behavior of a
see in these movies is that stable patterns, say an oval shape, can suddenly flock is a form of spontaneous order.
change. The birds may change direction or split up. Such bifurcations or catas-
trophes (to be explained in Chapter 3) are typical of complex systems. You
can also see that the behavior of these flocks is rather unpredictable. Flocks,
and swarms in general, are well understood and can be easily simulated on a
computer, but this does not mean that we can always predict these systems,
an issue that will be discussed further in Chapter 2.

Figure 1.1: A flock of birds is an example of a complex system in which global

patterns emerge naturally from simple rules and local interactions.

Similar examples can be found in all natural sciences. Tornadoes, for exam-
ple, are made up of air molecules that also interact locally. Tornadoes are
unpredictable, self-organizing, global weather phenomena. A famous chemical
example is the Belousov—Zhabotinsky reaction, a chemical oscillator (Ku-

9
ramoto 1984). This reaction involves a mixture of chemicals. As the reaction
proceeds, the solution exhibits strikingly colorful oscillations between clear
and opaque or between different colors, depending on the specific reactants
used. We will see many more examples in later chapters.
The prime example from psychology is the brain. About a hundred billion neu-
rons interact with thousands of other neurons in their neighborhood. Com-
pared to computers, brains are extremely energy-efficient, but like all open
systems, they do consume energy (the equivalent of a light bulb according to
Attwell and Iadecola 2002). The letters you are reading activate retinal neu- Fast local interactions somehow form
rons that initiate a cascade of electrical waves across billions of neurons that global waves of electrical activity that
make up thought processes and even
somehow create your understanding of this text (Roberts et al. 2019; Schöner consciousness.
2020). How is this possible? For me, this is the most fascinating scientific
question of all time. It’s the main reason why I’m a psychologist and not a
physicist. I view the brain as the ultimate complex system.

1.2 Emergentism

What is the relation between complexity and reductionism? According to

reductionism, complex phenomena can be explained by reducing them to the
interactions of their individual parts or components. This raises two questions,
one related to weak emergence and one related to strong emergence.
The first question is why it is possible to conduct scientific research in fields
other than physics, given that ultimately, chemistry, biology, and even psy-
chology are fundamentally concerned with interactions among elementary par-
ticles. Should we not first finish the study of physics before starting to think
about complex molecules, cells, neurons, or higher-order human cognition?
Philip Anderson’s renowned paper “More Is Different” convincingly argues
that the answer to this question is a resounding no (Anderson 1972). There is Science is possible at many different
much to be said for reductionism, but somehow the laws of quantum mechanics levels of description without fully
understanding the lower levels.
are irrelevant when studying interactions between neurons or people. I don’t
think that emergence in complex systems is inconsistent with a reductionist
view of science (Bechtel and Abrahamsen 2005). One could say that complex-
systems theory explains why emergent phenomena such as atoms or neurons
can be used as lower-order entities at even higher levels of description to
explain new higher-order phenomena, without being a dualist.
This fundamental principle of emergence is what allows disciplines like psychol-
ogy to exist as distinct and independent fields of science (Fodor 1974). The “In the face of complexity, an
concept of level is central to Herbert Simon’s architecture of complexity, in in-principle reductionist may be at the
same time a pragmatic holist” (Simon
which each subsystem is itself a complex structure made up of smaller parts, 1962).
and this pattern is repeated at multiple levels.3 According to Simon, these
nested structures are ubiquitous in the natural world and in human-made sys-
tems because they are robust and adaptable. For an in-depth discussion of
this level concept I refer to Wimsatt (1994).
The second, more controversial, question is whether emergent phenomena have
an independent causal role (strong emergence) or mainly have descriptive value

3
See figure 6.4 for a visual illustration of this idea.

10
(weak emergence). Strong emergence is often associated with downward cau-
sation (Chalmers 2006; Flack 2017; Kim 2006). Downward or circular causation is the
idea that higher-level entities or
I like to link this to the flocking example. Flocks of birds are emergent phe- properties can influence the behavior
nomena that do not determine the behavior of the individual birds. The birds of lower-level entities or properties.
only follow the local rules. Flocking is an example of weak emergence. How-
ever, when predators enter the scene, things change. Predators get confused
by flocks of prey, not by the behavior of individual birds. So the flock has
some causal power. Moreover, the birds react to the predator’s movements.
This could be seen as an example of downward causation and thus strong emer-
gence (figure 1.2). Recent work attempts to quantify such causal emergence
effects (see, e.g., Hoel, Albantakis, and Tononi 2013).

Figure 1.2: An illustration of downward or circular causation in flocks due to

a predator responding to the emerging patterns of the flock and
subsequently influencing the flight of individual birds.
Both weak and strong emergence are
Our minds, encompassing conscious thought, self-awareness, reasoning abili- essential to understanding
ties, natural language comprehension, emotions, and attitudes, are not mere psychological phenomena.
artifacts and cannot be simply reduced to intricate patterns of neural activity.
In my view, these mental constructs, such as consciousness, possess their own
causal influence, and this is one of the reasons that psychology stands as a
scientific discipline in its own right.
Of course, the relationship between the mind and the brain is one of the most
debated topics in psychology. The neuro-reductionist view is popular in the
field of psychology, when it concerns the explanation of both higher cognition
(Schwartz et al. 2016) and psychological disorders (for a critical review, see
Borsboom, Cramer, and Kalis 2019).

1.3 The field of psychology

The study of complex systems in the natural sciences is highly technical. I like
to think of the field of complex systems as a toolbox of empirical paradigms and
mathematical models and techniques (Grauwin et al. 2012). Models are often

11
formulated in the form of difference or differential equations and subjected to,
for example, bifurcation analysis. These are mathematical ways of describing
the behavior of complex systems. Additionally, advanced numerical analysis,
commonly in the form of computer simulation, is a standard approach. How-
ever, educational programs in psychology do not usually include courses in
algebra, calculus and programming. Many psychologists lack the basic knowl-
edge and skills to apply the toolbox of complex-systems theory, as these are
not ordinarily part of the psychology curriculum. Complex-systems research
simply seems too complex for psychologists and social scientists. One goal of
this book is to provide psychologists with a first introduction to this technical
toolbox.
Unfortunately, there are additional complications in applying these tools to
our field. First, our subjects are much more complex than flocks of birds or
tornadoes and they display astonishing behavior. They can do science! They
can also walk out of the lab because they find the experiment boring. This does
not happen with lasers. Second, we have to deal with the ethical constraints of
experimenting on our subjects. We cannot take them apart, a very successful
approach in the natural sciences. Finally, there is the measurement issue
(Lumsden 1976; Michell 1999).
We tend to forget how incredibly precise the natural sciences, especially
physics, are. In 1985, Richard Feynman famously claimed that the accuracy
of calculating the size of the magnetic moment of the electron was equivalent
to measuring the distance from Los Angeles to New York, a distance of over
3,000 miles, to the width of a human hair. I find that shocking. Less famously,
I would argue that psychologists have not yet “discovered” continents and
have no idea where New York is. Our instruments generally fail to meet
elementary requirements of reliability and validity, we are plagued by repli-
cation failures, and our theories are often imprecise (Eronen and Bringmann
2021). Navigating the behavioral and social
sciences and knowing which data to
This is all unfortunate because not only our brains but every subject in our trust and which empirical phenomena
field seem to have the characteristics of a complex system. Social systems are to model is an art in itself.
complex systems made up of individuals interacting to produce emergent phe-
nomena such as cultures and economic systems. The human brain, arguably
the most complex system we know, is embedded in different hierarchies of very
complex social systems such as families, education, economies, and cultures.
We need the toolbox!
Despite all these problems, I’m not pessimistic. I believe that tangible progress
in the behavioral and social sciences is possible. It is not that these sciences
are completely unsuccessful. We know a lot about people’s attitudes, addictive
behavior, cognition, and the social systems in which they interact. We study
these, with some success, using advanced experimental designs, and we have
developed (mainly) verbal theories about almost everything.
We also have no choice; we must make progress. Personally, I feel a strong
tension between our struggle to elevate the behavioral and social sciences as a
science, on the one hand, and the enormous expectations of society to deliver,
on the other. Our most pressing global problems—climate change, overpopu-
lation, war and violence, poverty, inequality, infectious diseases, addiction, to
name but a few—are unsolvable without breakthroughs in the behavioral and
social sciences. J. Doyne Farmer: “We have an
increasing need to model ourselves”
(Thurner 2016).

12
The realization that the human mind in its social context is an amazingly
complex system also offers opportunities. Despite their obvious differences,
complex systems show remarkable similarities. A predecessor of complex-
systems theory, general-systems theory (Bertalanffy 1969), explicitly assumed
that all systems share important characteristics. This is the primary reason Certain mechanisms and phenomena
for providing numerous modeling examples in this book that originate from seem to operate and to occur in
similar ways at all possible levels of
disciplines beyond psychology. description (Simon 1962).
An inspiring example for me comes from the study of shallow lakes (Scheffer
2004). Shallow lakes tend to be either in a “healthy” state, with clear water
and a diverse population of fish and plants, or in an “unhealthy” turbid state.
I like to compare these complex lake systems in the turbid state to a person
suffering from depression. This turbid state usually occurs suddenly. There is
a critical phosphorus load at which the system turns over from being healthy
to complete dominance by algae and bream. Typical of this type of transition
is the hysteresis effect (figure 1.3). This means that the turning point from
clear to turbid and from turbid to clear does not occur at the same phosphorus
load. The turning point to clear water only occurs at much lower phosphorus
loads. These tipping points may be so far apart that reducing the cause, the
phosphorus load, is not a viable option. Of course, all sorts of interventions
have been studied, such as supplemental oxygen, chemicals, sunscreens, and
stocking predatory fish. These interventions have not been very successful, or
only successful in specific cases. The fact that they had some level of success
brings to mind the partial effectiveness of clinical interventions, such as those
used in the treatment of major depressive disorder.

Figure 1.3: The transitions between clear and turbid states of shallow lakes
do not occur at the same phosphorus load. This delay in jumps
is called hysteresis. Hysteresis explains why transitions are often
difficult to reverse. This concept is discussed in detail in Chapter 3.

A breakthrough occurred in the 1980s. Removing all the fish proved to be

a very effective intervention. The ecologists caught almost all the fish with
nets during the winter. In the spring, a new, healthy equilibrium emerged,
characterized by aquatic plants, other fish species, and clear water. This new
state is often stable for long periods of time. Remarkably, the analysis of
the cause, the phosphorus load, was not part of the solution: although the
increasing phosphorus load is the primary cause of the transition toward a
turbid state, decreasing the phosphorus load does not cause the system to
transition back into the clear state. What this means for our thinking about The dogma of intervention, that the
depressive disorders will be discussed in Chapters 5 and 6. cause of the problem is the key to the
solution, does not necessarily apply to
I will discuss three specific reasons to be somewhat optimistic, based on three complex systems.

13
key observations about complex systems. The first key observation has to do
with simplification, the second with the tendency of complex systems to be
characterized by a limited number of stable states, and the third is that all
complex systems seem to be describable as some kind of network. Simplifica-
tion is perhaps the most important one.

1.4 The art of simplification

Einstein supposedly said that
A fascinating and instructive example is the traffic jam, which is made up everything should be made as simple
of many people, with their amazingly complex brains, in modern cars full of as possible, but not one bit simpler.
advanced technology. Where to start modeling such a complex phenomenon?
The answer is astonishing. It seems that we can reduce people in cars to simple
blocks in a lane, speeding up when there is space in front of the artificial car
and braking when they get too close to the car in front. All lower levels of
modeling are ignored. This is even simpler than a flock of birds.
It is not difficult to set up a computer simulation for this case. I recommend
that you spend some time playing around with an example.4 It does not
take long to see that traffic jams can easily form and have an unexpected
property: while cars move forward, traffic jams move backward! Another
interesting observation is that variance in speed causes congestion. But the
variance is not in any of the cars. Variance and congestion are properties at a
higher level of description. With this simulation, you can study different types
of traffic situations and interventions. This type of simulation proves very
useful for designing and optimizing highways and roads (Barceló 2010; Treiber,
Hennecke, and Helbing 2000). There are actually different ways to model
traffic jams. Jusup et al. (2022) distinguish between fluid-dynamical, kinetic,
car-following, coupled-map lattice, and cellular automata models. They all
reproduce many phenomena of real traffic jams.
In complex systems, the qualitative
Only higher-level properties are relevant to global behavior. A large part of properties of large-scale phenomena
the art of science is finding the right level of simplification. Suppose we are do not depend on microscopic details.
studying smoking. Do we model the effects of nicotine on blood vessels, how
the hand with the cigarette moves from the mouth to the ashtray, or the
number of cigarettes smoked per day? Do we include the effects of marketing
and the smoker’s social network?
What is relevant and what can be ignored? It can be challenging to provide
a definitive answer for specific cases. Nevertheless, in general, it can be said
that there is a limit to the lower levels that must be considered. When ex-
amining traffic jams, it is necessary to incorporate certain characteristics of
individuals and vehicles, but delving deeper into topics like neuronal firing,
DNA replication, or the intricate workings of car batteries becomes irrelevant.
At that level of modeling, there is no relevant information that would alter
the explanation of a traffic jam.
The traffic example shows that extreme simplification is sometimes possible
and necessary. But finding the right level of simplification is not a simple task
at all. In Borsboom et al. (2021) we propose a theory-construction methodology
(TCM) consisting of five steps:

4
A nice example is https://www.traffic-simulation.de/)

14
 1. Identify the empirical phenomena that become the target of explanation.
2. Formulate a set of theoretical principles that putatively explain these
phenomena.
3. Use this set or prototheory to construct a formal model, a set of model
equations that encode the explanatory principles.
4. Analyze the explanatory adequacy of the model, i.e., whether it actually
reproduces the phenomena identified in step 1.
5. Determine whether the explanatory principles are sufficiently parsimo-
nious and substantively plausible.

The article explains these steps in detail and provides an example, the mu-
tualism model of general intelligence, which is explained in Chapter 6 of this
book.
I will add a few comments to this list of steps. First of all, step 1 is key. It
is crucial to be precise about defining the phenomena to be explained. Phe-
nomena are not the same as data. Data are particular empirical patterns (a
concrete dataset), whereas phenomena are general empirical patterns, stable
and general features of the world (Haig 2014). As noted above, in the be-
havioral and social sciences, it is not always clear which data patterns can be
trusted. In the last ten years, the replication crisis has led to a revolution in
psychological methods, but many results are collected using potentially biased
methods. One problem is publication bias: negative results are still harder to
publish than results that support hypotheses. In other cases, the results of
different studies contradict each other, and meta-analyses show weak effects
at best. Drawing up a list of the most
important phenomena on a topic, such
The second observation is that taking these steps is not a linear process. Of- as depression, forgetting, or
ten, when developing a model, you realize that some important information is discrimination, is often a challenge.
missing from the list of phenomena. For example, you might be modeling ad-
diction, but suddenly you need information about the combination of addictive
substances that people use. And such simple questions are often impossible to
answer. You can spend days searching the literature for information that you
expect to be readily available, only to find that many basic things are simply
unknown.
A third observation is that formal modeling is mostly a matter of analogical
reasoning. You have to study many examples of complex-systems models to
understand how to construct such a model. Indeed, in my own work I often
use established models developed in physics and biology as a base model. We
will see many examples of this later.
Fourth, good models do not build in phenomena but explain them from basic
principles. By building in, I mean that a phenomenon should not be an assump-
tion of the model. An example would be a model that says that variance in the
speed of cars causes traffic jams. Such a model may explain other things, but
not the role of speed variance, because that effect is part of the assumptions.
Models that make such assumptions are called phenomenological models. We
will see examples of phenomenological models of complex systems, as well as
explanatory models, where the latter are based on fundamental principles. Building real explanatory models in
our fields is extremely difficult.
Fifth, it is my conviction that a metaphorical use of the complex-systems
approach should be avoided by using concrete formal models (Dongen et al.
2024). It is crucial to strive for the highest level of scientific rigor. There are

15
no special, more lenient, methodological rules for complex-systems research
(van der Maas 1995).

1.5 A limited number of equilibria

The first key observation was that complex systems can be simplified. The
second is that complex systems tend to be characterized by a limited number
of equilibria. An important example is water. Water normally exists in either
a solid, liquid, or gaseous state (leaving aside the plasma state). These are
stable states over wide ranges of temperature and pressure.
A biological example is the life stages of a butterfly (egg, caterpillar, chrysalis,
and butterfly). Except for brief periods of transition, these insects are in one of
these four relatively stable states. Another example is the horse, which is either
standing still, walking, trotting, or galloping. I am convinced that we must
always start by identifying the equilibria of a complex system. This also applies
to psychological and social science applications. A bipolar disorder seems to
be characterized by two stable states (depressive and manic). In the case of
addiction, we may think of a three states: non-use, recreational use, and heavy
use (Epskamp et al. 2022). Similarly, we could identify the stages in falling
in love, in understanding of calculus, in sleeping, and in radicalization.
Identifying discrete stages turns out to be more difficult than it first appears.
It is often possible to come up with more substages. For instance, in the There is an ongoing discussion about
case of horse movement, people tend to further subdivide trot into three forms the number of stages, even for
something like sleep (Boostani,
(working, medium, and collected). Subdivisions are also made in the case of Karimzadeh, and Nami 2017; de
heavy alcohol consumption (Leggio et al. 2009). It is possible to use objec- Mooij et al. 2020).
tive statistical methods to support such classifications using modern machine-
learning techniques (automatic clustering) as well as more traditional means
(finite mixture models, latent class analysis). I will say more about this in
section 3.5.1.2.
A further complication is that equilibria come in different forms. The simplest
form consists of fixed points or point attractors, an example being a ball lying
in a valley. Under undisturbed conditions, the ball could also be resting on
top of a hill, which is an unstable equilibrium. An equilibrium could also be a
limit cycle or oscillator. For example, two pendulums could swing in phase or
out of phase. It gets even weirder when we consider strange attractors, which
often take the form of fractals. This will be explained in more detail in the
next two chapters.
Finally, it has also been argued that many complex systems, especially living
systems, never reach equilibrium because they are constantly perturbed (Groot
and Mazur 2013). But at least some complex psychological systems are clearly I see this distinction between
stable over the long term. Unfortunately, this is true of many psychological equilibrium and non-equilibrium
complex systems as gradual.
disorders. In contrast, my understanding of the world, psychological science,
and complex-systems research is better described as a continuously perturbed
non-equilibrium system with just enough stability to write this book (once).
I would claim that many psychological complex systems tend to be in one
attractor state most of the time, but they occasionally change states. If certain
control parameters slowly change their values, the current equilibrium can
become unstable and a transition to another equilibrium can occur. This is

16
what happens when we lower the temperature of water to below zero. The Transitions can occur in many ways,
family of transition models is described by bifurcation theory. This is explained also depending on the types of
equilibria involved.
in Chapter 3, where we focus on a very important transition model, the cusp
catastrophe, and in Chapter 4, which considers dynamical systems models.

1.6 Networks are everywhere

The third key observation of great relevance to the attempt to use complex-
systems modeling in psychology is that complex systems are networks, as
they consist of interacting subelements. For me, the network is the most
interdisciplinary research topic in modern science. Magnets, ecosystems, the
brain, the internet, and social networks are prime examples. Network science is a huge area of
research with many fundamental
Two applications in psychology are well known: the first is the study of neural insights and an important tool in
networks, which started seventy years ago and has become the main foundation modern psychological science.
of the artificial intelligence (AI) revolution of the last ten years. In Chapter 5
I will discuss neural networks. The second is social networks, the simplest ex-
ample being dyadic interactions. Social media such as Facebook are infamous
examples. Key ideas relate to concepts such as weak and strong ties, central
hubs and homophily, which are discussed in Chapters 6 and 7. The analysis
of social-network data is an exciting area of research (Scott 2011). It focuses
on understanding how social entities are connected and how these connec-
tions influence various outcomes and behaviors. Connections between nodes
(e.g., individuals, organizations, communities) can be based on different di-
mensions, such as friendship, communication, collaboration, information flow,
or any other form of social interaction. These interactions may also change
over time, which is studied in social-network dynamics (Snijders 2001).5
Chapter 6 focuses on a novel use of networks, which I call network psychology.
This is a level of description between neural networks and social networks. It
involves modeling intelligence, attitudes, and psychological disorders at the
individual level. Intelligence, for example, is modeled as an ecosystem of coop-
erating cognitive functions. This is radically different from the standard view
that general intelligence is due to 𝑔, a single underlying source. In the mutu-
alism model of general intelligence (van der Maas et al. 2006), the observed
positive correlations between scores on subtests of IQ test batteries are due
to cumulative reciprocal developmental interactions between cognitive subsys-
tems such as working memory, spatial cognition, and language.
Similarly, depression can be thought
As another example, sleep problems, a symptom of depression, can lead to of as a network of mutually reinforcing
increased fatigue and difficulty concentrating, which in turn can affect a per- symptoms.
son’s ability to manage daily tasks and engage in social activities. This can
then lead back to poorer sleep quality, creating a cyclical pattern in which
each symptom reinforces the others. This new view of mental disorders origi-
nated in our research group and is now very popular (Robinaugh et al. 2020).
One reason for this is that many statistical techniques have been developed to
investigate this network approach.
The latest line of this research is the integrated study of psychological and so-
cial networks (van der Maas, Dalege, and Waldorp 2020). Chapter 7 deals with
models in which psychological network models of attitudes are nested within
5
Statnet.org provides an overview of R packages for social network analysis.

17
social networks of opinion change. This model provides a new explanation of
polarization. In the process of polarization,
nonlinear intrapersonal and
interpersonal dynamics interact.

1.7 Methods for investigating complex systems

Complex systems are studied in various ways across different disciplines. We

use computer simulations to examine the emergence in complex system models,
analyze their unpredictable behavior, categorize the types of tipping points in-
volved, derive equations that describe the overall behavior of complex systems,
collect and analyze time-series data, or experimentally disrupt the system to
test its resilience. Following Sayama (2015), I categorize the models and meth-
ods into two groups: those for systems with a small number of variables and
those for systems with many variables.
The first category is referred to as nonlinear dynamical system theory, which
encompasses chaos theory and catastrophe or bifurcation theory. The second
category includes tools for studying multi-element systems, such as agent-
based modeling and network theory. One might assume that the first category
is irrelevant to complex systems, which by definition have many variables.
However, it has been found that the global behavior of complex systems can
often be described with a small number of variables, often just one, that behave
in a highly nonlinear manner. I consider nonlinear dynamical system
theory an essential part of
This categorization is reflected in the book’s structure. The next three chapters complex-systems research.
are devoted to systems with a small number of variables: Chapter 2 discusses
chaos theory, Chapter 3 addresses sudden transitions as studied in catastrophe
and bifurcation theory, and Chapter 4 provides an introduction to modeling
dynamical systems.
In the second part of the book, we shift our focus to tools for studying systems
with many variables, particularly agent-based modeling of self-organization in
Chapter 5, network modeling in Chapter 6, and the application of both to
psychosocial systems in Chapter 7.

1.8 Other work and sources

The complex-systems approach has often been introduced as the next new
thing, but those days are gone. Even in psychology it can no longer be consid-
ered a new approach. Many different research groups have used the toolbox of
complex-systems research in all areas of psychology. This book will give many
examples. One could even argue that a lot of work has been done that could be
considered complex-systems research but has not been published under that
heading. For example, most neural-network models of psychological processes
are complex-systems models because they investigate emergent computational
properties of the interaction of neural units. This is also true of much work in
mathematical psychology, for example when differential equations are used to
study dynamical systems. Older work in complex-systems research has often
been published with reference to nonlinear dynamical systems. Other related
approaches are computational social science and agent-based modeling.

18
Today, there are many interdisciplinary centers or hubs for complexity research.
The Santa Fe Institute in Santa Fe, New Mexico, is the pioneer of complexity
science. Its summer schools are highly recommended. Other examples are the
Complexity Science Hub in Vienna, Austria, and the Centre for Complexity
Science at the University of Warwick, in England. In my own country, the
Netherlands, we have at least four of these centers. I’m a principal investigator
at the Institute for Advanced Study in Amsterdam and an external faculty
member at the Santa Fe Institute.
It is impossible to give a balanced review of all past and ongoing work on
complex systems. I’m naturally somewhat biased toward our own work and
contributions, but I do my best to point out relevant work. As a general
resource to complex systems research with a bit less technical approach, I
recommend the book of Mitchell (2009); for a bit more mathematical approach
I recommend the books of Serra and Zanarini (1990), Sayama (2015), and
Thurner, Klimek, and Hanel (2018). Overviews of work in psychology are
provided by Guastello, Koopmans, and Pincus (2008) and Port and Gelder
(1995). Other great books are written by Heath (2000) and Kelso (1995).

1.9 Exercises

1) Visit https://www.traffic-simulation.de/. In what direction do traffic

jams move? For roundabouts: What is a bad priority rule? Do traffic
jams appear and disappear for the same values of critical parameters?
Take for instance the ring road and vary Politeness. (*)
2) Give your own example of a psychological process or theory where dif-
ferent stable stages or states are distinguished. (*)
3) Could consciousness be seen as a process of downward causation? Ex-
plain your answer. (**)

19
2 Chaos and unpredictability

2.1 Introduction

Suppose we have immense amounts of genetic, biological, and psychological

data on millions of participants and knowledge of all relevant environmental
factors. Suppose also that these huge amounts of data are of fantastic qual-
ity. Using state-of-the-art machine-learning models and powerful computing
resources, we could build advanced statistical models that include main and
higher-order interaction effects of all variables, even incorporating nonlinear
transformations. Even then, prediction may not be possible. Why? Because
of a phenomenon called deterministic chaos. Deterministic chaos refers to the
behavior of complex systems that is
Chaos is one of the most spectacular phenomena in complex systems, and as highly sensitive to initial conditions,
psychologists we should know the basic results of chaos theory. It is also great leading to unpredictable and
fun to learn about chaos and it allows me to introduce many key concepts that seemingly random results despite
being governed by deterministic laws.
we need in later chapters.
In my opinion, the direct applicability of chaos theory to psychology and
social science is somewhat limited. For a long time, researchers have tried to
show chaos in time series of psychophysiological measures, but this seems to
be difficult. I will briefly review this work at the end of the chapter. The
relevance of chaos theory may lie not in its application but in its fundamental
implication for prediction. What chaos theory basically shows is that even
in the best of circumstances, where we have very accurate models and data,
long-term prediction might be impossible. This is known as the butterfly effect
(E. N. Lorenz 1963): a butterfly flaps
its wings in India, and that tiny
change in air pressure could eventually
2.2 The population growth of rabbits cause a tornado in Iowa.

Chaos theory consists of many deep mathematical results, but understanding

the basics of chaos is not so hard. Below I will explain chaos in difference
equations at a very basic level of mathematics and programming. The ele-
mentary example is the famous logistic map, usually introduced as a model
of population growth, for instance, of rabbits. Suppose we have rabbits on an
island, and they start to multiply. What would the mathematical model be
for such a process?
Population growth is a typical
In general, in a dynamical system, the change or growth of a variable (say 𝑋) example of a dynamical system, as it
depends on the current state and some parameters. Time plays a very special is a model of change.
role. We can use discrete or continuous time steps. In the first case, which
is the focus of this chapter, we use difference equations; in the second case,
we use differential equations. In the logistic map, time is discrete (population
growth takes place in generations). The simplest dynamical model for the
population growth of rabbits is

20
 𝑋𝑡+1 = 𝑟𝑋𝑡 . (2.1)

This says that the new value of 𝑋 is determined by the previous value of
𝑋, multiplied by 𝑟. In this equation 𝑟 is the growth rate. We can simulate
this model by choosing a value for 𝑟, 𝑟 = 2, for instance. We also need an
initial value, say 𝑋0 = 1. If this is completely new to you, enter some values
repeatedly. You will see exponential growth (𝑋1 = 2, 𝑋2 = 4, 𝑋3 = 8, 𝑋4 =
16, 𝑒𝑡𝑐.). In R we can simulate this using a for loop (the result is shown in
figure 2.1).

n <- 15
r <- 2
x <- rep(0, n)
x[1] <- 2 # initial state X0 = 1 and thus X1 = 2
for (i in 1:(n - 1)) {
x[i + 1] <- r * x[i]
}
plot(x, type = 'b', xlab = 'time', bty = 'n')

Figure 2.1: Exponential growth.

Note that we can find any 𝑋𝑡 given 𝑋0 by iterating the model as we do

in the for loop. 𝑋𝑡 is called the solution. Simulation is a bit odd in this
case. We can compute the solution analytically. It is 𝑋𝑡 = 𝑋0 𝑟𝑡 . Thus for
𝑋15 = 1 × 215 = 32768. For more complex models, the
analytical solution is often not
Note that the exponential model ignores the fact that population growth is available, and we have to use
limited by resources. At some point food will become scarce. One way of simulation (the numerical solution).
making the model, introduced by Verhulst in 1838, more realistic is to add a
growth-limiting term:

𝑋𝑡
𝑋𝑡+1 = 𝑓 (𝑋𝑡 ) = 𝑟𝑋𝑡 (1 − ). (2.2)
𝐾

What is the effect of this addition to the equation? If 𝑋 is much smaller than
the resource 𝐾, then the second term, (1 − 𝑋𝑡 /𝐾), is close to 1 and we will see
exponential growth. But as 𝑋 approaches 𝐾, this term becomes very small,

21
reducing the effect of exponential growth. 𝑋 does not actually grow up to 𝐾,
but to a lower value, if it converges at all. We are going to see this in a moment.
It also turns out that the actual value of 𝐾 is not of interest. Changing 𝐾
does not change the qualitative behavior. Therefore, 𝐾 is usually set to 1,
scaling the population 𝑋 between 0 and 1. The only remaining parameter is
𝑟. Changing 𝑟, however, leads to a number of surprising behaviors.1

2.3 Stable and unstable fixed points

Let us study a “boring” case first, 𝑟 = 2 (figure 2.2).

n <- 15
r <- 2
x <- rep(0, n)
x[1] <- .01 # initial state
for (i in 1:(n - 1)) {
x[i + 1] <- r * x[i] * (1 - x[i])
}
plot(x, type = 'b', xlab = 'time', bty = 'n')

Figure 2.2: The 𝑟 = 2 case. The population converges to a stable state at

𝑋 = 0.5.

This is the simple case. The population initially develops exponentially but
then levels off and reaches a stable state at 𝑋 = .5. We need to understand a
bit more about it. What you see here is that we have gone from an unstable
initial state to a stable state, a point attractor. The next code shows that this
point attractor attracts from a wide range of initial values, but not all.

n <- 30; r <- 2; x <- rep(0,n)

for (init in seq(0, .7, by = .01)){
# start from different initial values
x[1] <- init

1
Verhulst proposed this model in the form of a differential equation in continuous time. We
will discuss this type of model in Chapter 4. In continuous time, nothing particularly
spectacular happens, and we only see the kind of behavior displayed in figure 2.2.

22
 for (i in 1:(n - 1)){
x[i + 1] <- r * x[i] * (1 - x[i])
}
if (x[i] == 0)
plot(x,type = 'l',xlab = 'time',bty = 'n',ylim = c(0, .8),col = 'red')
else
lines(x)
}

If we start exactly at 0, 𝑋 stays at 0. So, 0 is an equilibrium too, but a special

one. It is an unstable fixed point. A small perturbation will cause 𝑋 to move
to .5, the stable fixed point. All initial values in close proximity of 0 will move
away from 0 (repellent), but if 𝑋 = 0 exactly, then it remains 0 for all time.
So, 𝑋 = 0 is a fixed point but unstable (figure 2.1). Fixed points can be stable or unstable.

Figure 2.3: Illustration of stable and unstable fixed points. For many initial
values, 𝑋 = .5 is an attractor. 𝑋 = 0 is an unstable fixed point.
Only if we start exactly at 0 do we stay there.

This concept of equilibrium, stable or unstable, is crucial for later chapters.

The essence of the next chapter is to change a control parameter, here 𝑟, and
study how the pattern of equilibria (the equilibrium landscape) changes. You
can easily do this yourself by rerunning the simulation with 𝑟-values just below
and above 1. For 𝑟 < 1, there is only one stable attractor (0).
Simulating this is not really necessary. One has to realize that a fixed point
(𝑋 ∗ ) is found when 𝑋𝑡+1 = 𝑋𝑡 = 𝑋 ∗ . See for yourself that:

𝑋𝑡+1 = 𝑋𝑡 = 𝑋 ∗
𝑋 ∗ = 𝑟𝑋 ∗ (1 − 𝑋 ∗ )
𝑋 ∗ = 0 𝑜𝑟 1 = 𝑟 − 𝑟𝑋 ∗
𝑟−1
𝑋 ∗ = 0 𝑜𝑟 𝑋 ∗ = .
𝑟

So 0 and (𝑟 − 1)/𝑟 are fixed points. Indeed, for 𝑟 = 2, we have seen that 0
and .5 are equilibria, one unstable and one stable. To determine whether fixed
′
points are stable, we look at the derivative of the function, 𝑓 (𝑥), which, as
you can easily check, is 𝑟 − 2𝑟𝑋.

23
The fixed point is stable if the absolute value of the derivative in the
fixed-point value is less than 1.2 For 𝑟 = 2 the fixed points are 0 and .5.
′
∣𝑓 (𝑋 ∗ = 0)∣ = |2 − 0| = 2, which is greater than 1 and thus 𝑋 ∗ = 0 is
′
unstable. ∣𝑓 (𝑋 ∗ = .5)∣ = |2 − 2 × 2 × .5| = 0, which is less than 1 and thus
𝑋 ∗ = .5 is stable. You can check for yourself that 𝑋 ∗ = (𝑟 − 1)/𝑟 is stable
for 1 < 𝑟 < 3, both with the R-code and with the absolute value of the
derivative.

2.4 Limit cycles

So at 𝑟 = 3 the fixed point at (𝑟 − 1)/𝑟 becomes unstable. Let’s study some

cases. The plots in figure 2.4 are made with the code for figure 2.2.

Figure 2.4: Qualitative different behavior of the logistic map for different val-
ues of 𝑟.

For 𝑟 = 2.9 we see that the series converges to the fixed point 1.92.9 = .66, but
in a process of over- and undershooting. Between 𝑟 = 3.1 and 𝑟 = 3.3, a limit
cycle of period 2 arises. For 𝑟 = 3.5 this becomes even more remarkable, and In a limit cycle of period 2, the
we see a limit cycle of period 4. For slightly larger values, we could get cycles population oscillates between two
values.
with higher periods.
It has been claimed that these limit cycles occur in real population dynamics
(Hassell, Lawton, and May 1976). Intuitively, it can be understood as a process
of over- and undershooting, which dampens out for 𝑟 a little below 3, but not
for 𝑟 > 3.

2
It is not too difficult to understand why this is. If you Google search “fixed points of dif-
ference equations,” you will quickly arrive at stackexchange.com, where several insightful
explanations are given.

24
2.5 Chaos

If we increase 𝑟 even further, the doubling of the periods changes to even

stranger behavior. Figure 2.4 shows what the time series looks for 𝑟 = 4.

Figure 2.5: Chaos for 𝑟 = 4.

There seems to be no regularity. This is what we call deterministic chaos. This

time series is unpredictable, even though we know the equation and the system
is deterministic. What exactly do we mean by this? Let me illustrate.

r <- 4; n <- 50; x <- rep(0,n)

x[1] <- .001
for (i in 1:(n - 1)){
x[i + 1] <- r * x[i] * (1 - x[i])
}
plot(x, type = 'l', xlab = 'time', bty = 'n')
# restart with sightly different initial state:
x[1] <- .0010001
for (i in 1:(n - 1)){
x[i + 1] <- r * x[i] * (1 - x[i])
}
lines(x, col = 'red')

Figure 2.6: The butterfly effect: A small difference in initial state causes di-
vergence in the long run.

25
We can see that a run with a slightly different initial value will at first follow
the same path, but then it will diverge sharply (figure 2.6). A tiny pertur-
bation (the butterfly flapping its wings) propagates through the system and
dramatically changes the long-term course of the system.
This is why long-term weather
Note that some uncertainty about the exact value of the initial state is always prediction will never be possible, even
inevitable. Suppose we have an equation like the logistic map for temperature if we develop much more precise
in the weather system, and this equation perfectly describes that system. To mathematical models, take more
make a prediction, we need to feed the current temperature into the computer. intensive and more accurate
measurements, and use more powerful
But we cannot measure temperature with infinite precision. And even if we computers.
could, we do not have a computer that can handle numbers with an infinite
number of digits. So, we make a small error in setting the initial state, and
this will always mess up our long-term forecast. The weather turns out to be
a chaotic system. Sensitivity to initial conditions is a necessary and perhaps
sufficient condition for deterministic chaos. For a discussion on the definition
of chaos, I refer to Banks et al. (1992) and Broer and Takens (2010).
The Lyapunov coefficient quantifies
The idea of the Lyapunov coefficient is to take two very close initial conditions chaos.
with a difference of 𝜀. In the next iteration, this difference might be smaller,
the same, or bigger. In the last case, the time series diverge, which is typical
for chaos. The Lyapunov coefficient is defined as:

1 𝑛 ′
𝜆𝐿 = lim ∑ ln ∣𝑓 (𝑋𝑖 )∣. (2.3)
𝑛→∞ 𝑛
𝑖

′
where 𝑓 (𝑋𝑖 ) = 𝑟 − 2𝑟𝑋𝑖 for the logistic map and 𝜆𝐿 > 0 indicates chaos. You
may verify in a simulation that 𝜆𝐿 > 0 for 𝑟 = 4, indicating chaos.

2.6 Phase plot and bifurcation diagrams

Equation 2.2 is very simple. It is just one equation, a deterministic difference

equation specifying how 𝑋𝑡+1 depends on 𝑋𝑡 , but the variety of behavior is
astonishing. One way to better understand its behavior is to use phase plots.
In one-dimensional systems we plot 𝑋𝑡 against 𝑋𝑡+1 (see figure 2.7). The A phase plot is a graphical
code for this figure is: representation of the relationship
between two or more variables that
change over time.
layout(matrix(1:6,2,3))
r <- 3.3; n <- 200; x <- rep(0,n)
x[1] <- .001
for(i in 1:(n-1)) x[i+1] = r*x[i]*(1-x[i])
x <- x[-1:-100]
plot(x, type = 'l', xlab = 'time', bty = 'n',
main = paste('r = ', r),
ylim = 0:1, cex.main = 2)
plot(x[-length(x)], x[-1],
xlim=0:1, ylim=0:1, xlab='Xt', ylab='Xt+1', bty='n')
r <- 4; x[1] <- .001;
for(i in 1:(n-1)) x[i+1] <- r * x[i] * (1-x[i])
x <- x[-1:-100]
plot(x, type = 'l', xlab = 'time', bty = 'n',

26
 main = paste('r = ',r), cex.main = 2)
plot(x[-length(x)], x[-1], xlim = 0:1, ylim = 0:1,
xlab = 'Xt', ylab = 'Xt+1', bty = 'n')
x <- runif(200,0,1)
x <- x[-1:-100]
plot(x, type = 'l', xlab = 'time', bty = 'n',
main = 'random noise',
cex.main = 2)
plot(x[-length(x)],x[-1], xlim = 0:1, ylim = 0:1,
xlab = 'Xt', ylab = 'Xt+1', bty = 'n')

Figure 2.7: Time (top) and phase (bottom) plots for three cases. Chaos and
random noise can be distinguished using the phase plots.

The top figures are time plots, and the lower figures are phase plots. The first
column shows a limit cycle of period 2, the second deterministic chaos, and the
third noise generated from a uniform distribution. Although the time series
of the second and third cases look similar, the phase diagram reveals hidden
structure in the chaos time series. Phase plots can help us to distinguish chaos
from noise.
The second useful graph is the bifurcation graph. In case of the logistic map, A bifurcation graph is a diagram that
it summarizes the equilibrium behavior for different values of 𝑟 in one figure. shows how the qualitative behavior of
a system changes, for example, from
The idea is to plot the equilibria as y-values for a range of 𝑟-values on the stable to chaotic, when one of its
x-axis. This means that if we take a low 𝑟 value (𝑟 < 1), we will only plot 0s, parameters changes.
as only 𝑋 ∗ = 0 is a stable fixed point. Between 1 and 3, we will also see one
fixed point equal to (𝑟 − 1)/𝑟. For 𝑟 = 3.3, we expect to see two points as the
attractor is a limit cycle with period 2. For higher 𝑟 we get chaos. How does
this all look?

27
It is actually a good challenge to program this yourself. The trick is to create
time series for a range of values of 𝑟, delete the first part of this series (we only
want the equilibrium behavior), and plot these as y-values. So, if the logistic
map has period 2 (𝑟 = 3.3), we repeatedly plot only two points. For 𝑟 = 4 we
get the whole chaos band.
A clever way to do this is to use the sapply function in R.

layout(1)
f <- function(r, x, n, m){
x <- rep(x,n)
for(i in 1:(n-1)) x[i+1] <- r * x[i] * (1-x[i])
x[c((n-m):n)] # only return last m iterations
}
r.range <- seq(0, 2.5, by = 0.01)
r.range <- c(r.range,seq(2.5, 4, by = 0.001))
n <- 200; m <-100
equilibria <- as.vector(sapply(r.range, f, x = 0.1, n = n, m = m-1))
r <- sort(rep(r.range, m))
plot(equilibria ~ r, pch = 19, cex = .01, bty = 'n')

This results in figure 2.8. We see indeed fixed stable points for 𝑟 < 3, the
period doubling of the limit cycles for 𝑟 > 3, followed by chaos.

Figure 2.8: The bifurcation diagram of the logistic map.

Fractals are a recurring phenomenon in many chaotic maps. You can see the Fractals are figures in which certain
fractal nature of the logistic map by zooming in on the interval of 𝑟 between patterns reappear when we zoom in
on the figure, and this happens again
3.83 and 3.86 (see exercise 2). The three equilibria in the limit cycle split and again when we zoom in farther.
again into period doubling cycles, as we saw in the overall plot between 𝑟 in 3
and 3.5.
One famous result on this period doubling route to chaos is the Feigenbaum
constant. The ratios of distances between consecutive period doubling points
(e.g., the distance between first and second divided by the distance between
the second and third point) converge to a value of approximately 4.6692. The
amazing thing is that this constant is the same for any unimodal map.

28
2.7 What did we learn?

I find these results stunning. I note again that the generating function is
deceptively simple, but its behavior is utterly complex and beautiful. Math-
ematicians have studied every detail of these plots, and most of it is beyond
my comprehension. The Wikipedia on the logistic map will introduce you to
some more advanced concepts, but for our purposes the present introduction
will suffice.
Let’s review the concepts we have already learned. The first is the concept
of equilibrium. The states of dynamical systems tend to converge to certain
values. The simplest of these is the fixed point. Fixed points can be stable
or unstable (more on this in the next chapter). If we start a system exactly
at its unstable fixed point (and there is no noise in the system), it will stay
there. But any small perturbation will cause it to escape and move to the
fixed stable point.
The bifurcation diagram summarizes this behavior and also shows how the
equilibria change when a control parameter changes. For example, at 𝑟 = 1
we see a bifurcation in the logistic map. Initially 0 was the stable fixed point
and (𝑟 − 1)/𝑟 was unstable. At 𝑟 = 1 this is reversed. At 𝑟 = 3 we see another
bifurcation when limit cycles appear.
We have learned that there are all sorts of equilibria. The strangest ones are
called strange attractors, which are associated with deterministic chaos. You
can see them by making a phase diagram. Phase diagrams for other famous
maps are often stunning. The most famous is the Mandelbrot set (look on the
internet). There is an R blog about the Mandelbrot set.3 Simulation helps
understanding!
The last thing we learned is that even if our world were deterministic (it is not!),
and we knew all the laws of motion (say, the logistic map), and we knew initial
states with enormous precision, the world would still be unpredictable.
This statement needs some nuance. I have already mentioned that the weather
can be chaotic and unpredictable. But the weather is not always so unpre-
dictable. Sometimes longer forecasts are possible. But forecasts beyond, say,
10 days seem out of reach. We also see in the logistic map that when 𝑟 is
close to 4, the forecast suffers from the butterfly effect, but for 𝑟 = 2 the time
course is very predictable, even more predictable than in many linear systems.
This is because there is only one stable fixed point (.5). The initial state does
not matter: we always end up at .5! The logistic map is either extremely
predictable or extremely unpredictable
depending on the value of 𝑟.

2.8 Other maps and fractals

There are many accessible sources on chaos theory. As always, Wikipedia is

a great resource. It helps me a lot by actually doing things, that is, doing
computer simulations. One example is the Henon map, which consists of two
coupled difference equations:

3
https://www.r-bloggers.com/2017/06/the-mandelbrot-set-in-r-2/. Be sure to check the
Shiny app.

29
 𝑋𝑡+1 = 1 − 𝑎𝑋𝑡2 + 𝑌𝑡 ,
(2.4)
𝑌𝑡+1 = 𝑏𝑋𝑡 .

Using the code example from the logistic map, you should be able to generate
time series and a phase diagram for this model. Try to reproduce the first im-
age of the Wikipedia page on the Henon map. The amazing three-dimensional
bifurcation diagram may be more challenging.
Fractals are another topic for further study. Another look at Wikipedia is
recommended. Making your own fractals in R is made easy by the R blog by
Martin Stefan (2020).

2.9 Detecting chaos in psychophysiological data

Chaos theory and the logistic map were popularized about fifty years ago, and
since then researchers have been looking for chaos in all kinds of time series
(Ayers 1997; Robertson and Combs 2014; G. K. Schiepek et al. 2017). One
idea behind this work is the hypothesis that chaos might be healthy (Pool 1989)
or helpful. It would be helpful in learning algorithms, such as neural networks,
to prevent getting stuck in local minima (Bertschinger and Natschläger 2004).
My very first publication was about chaos in neural networks (van der Maas,
Verschure, and Molenaar 1990).
There are many techniques for chaos detection in times series. However, these
empirical signals are inevitably contaminated with noise (Rosso et al. 2007).
One example is the computation of Lyapunov exponents, quantifying how
small differences in initial conditions evolve over time. A positive Lyapunov ex-
ponent indicates chaos, signifying exponential divergence of trajectories, which
is a hallmark of chaotic systems. This method involves reconstructing the
phase space from time-series data and calculating the average exponential
rate of separation of trajectories. With the Lyapunov function in the package
DChaos, you can compute the Lyapunov coefficient for times series generated
with the logistic map. You may verify that for 𝑟 = 4, you get the Lyapunov co-
efficient as computed with the derivative earlier. There are several packages Chaos detection is an active area of
available in R, including new methods based on machine-learning techniques research, with new methods being
proposed on a regular basis (Zanin
(Sandubete and Escot 2021; Toker, Sommer, and D’Esposito 2020). 2022).
These methods generally require long time series. Many publications appeared
on the detection of chaos in psychophysiological data. Examples are elec-
troencephalogram (EEG) (Pritchard and Duke 1992) heartbeat (Freitas et al.
2009), electromyogram (EMG) (Lei, Wang, and Feng 2001), and eye move-
ments (Harezlak and Kasprowski 2018). Reviews of these lines of research
are provided by Stam (2005), Kargarnovin et al. (2023), and Garc and Pe
(2015).

2.10 Exercises

1) For 𝑟 = 3.5, the logistic map iterates between four points. For which
value(s) does it iterate between 8 points? (*)

30
2) Section 2.6 shows the code to make a bifurcation plot. First, run this
code and look at the bifurcation plot. In this plot, you can also zoom in
by changing the interval between the 𝑟’s on the x-axis. Adjust the code
by changing the r.range to seq(3.4, 4, by=0.0001), also change cex = 0.01
to a lower value. Zoom in on the interval of 𝑟 between 3.83 and 3.86. In
this interval the chaos suddenly disappears and limit cycles with period
3 appear. Check this with a time-series plot for a particular value of 𝑟.
(*)
3) Reproduce the first image from the Henon map Wikipedia page. Provide
your R code and figure (*).
4) Make the bifurcation diagram of the Ricker model (see Wikipedia). Pro-
vide your R code and figure. Why is this model considered a more
realistic representation of population growth than the logistic map? (*)
5) Also reproduce the three-dimensional bifurcation diagram of the Henon
map. (**)
6) Have a look at the definition of the Lyapunov coefficient in section 2.5.
Calculate this Lyapunov coefficient for the logistic map where 𝑟 = 4
using the Dchaos package in R. This coefficient can also be calculated
manually using the derivative (equation 2.3). Do this and check that the
coefficients are approximately equal. (**)
7) Use the Rmusic library (installed with
devtools::install_github("keithmcnulty/Rmusic", build_vignettes = TRUE) to
create a chaos sound machine. Make one for white noise too. Can you
hear the difference? (**)
8) Find a paper on chaos detection in psychology or psychophysiology and
summarize it in 300—400 words. (*)

31
3 Transitions in complex systems

3.1 Introduction

My dissertation research was on Jean Piaget’s stage theory of cognitive de-

velopment. These stages were separated by transitions. One such transition
should occur between the pre-operational and concrete-operational stages. In
the concrete-operational stage, children learn logical, concrete physical rules
about objects, such as weight, height, and volume. The most famous test to
distinguish between the two stages is the conservation task.
There are many conservation tasks, but the setup is always the same. For
example, you show a child two equal balls of clay, ask for confirmation that
they weigh the same, roll one into a sausage shape, and then ask again for
confirmation of equal weight. A nonconserving child will now claim that the
longer sausage weighs more. One can also do this with two rows of coins
(spreading one row out) or two glasses of water (pouring the water from one
glass into a smaller longer glass). It is actually a fascinating task to do with
children between five and eight years old.
From the 1960s to the 1980s, this was a topic of major interest in developmental
psychology. A key question was whether there really was a stage transition,
and there was a lot of confusion about what a transition actually was. It was
my task to clarify this and to prove Piaget’s hypothesis. I think I succeeded
in clarifying the question, but whether I succeeded in proving the stage theory
is debatable.1
My PhD advisor Peter Molenaar had the idea to use catastrophe theory to
define the concept of a transition in a precise way, to use the so-called catas-
trophe flags to test the hypothesis of a transition, and also to fit a cusp model
to the conservation data. It took me, with the help of many people, more than
20 years to do all these steps (van der Maas and Molenaar 1992; Jansen and
van der Maas 2001; Dolan and van der Maas 1998; Grasman, van der Maas,
and Wagenmakers 2009).
What is catastrophe theory, what are these flags, and what is the cusp? These
are the first questions I will answer in this chapter. But this chapter is also
about statistics. You will learn how to fit a cusp model to data. I will present
a methodology for studying transitions in areas where we do not have a math-
ematical description of the underlying system. I will present examples from
very different subfields of psychology. Finally, I will discuss the criticisms that
were made in response to the hype around catastrophe theory about fifty years
ago. This is a long chapter, but I have tried to include only what is necessary
in order to make intelligent use of this approach. This requires some basic
1
Learning a particular conservation task does seem to be rather sudden, but there could
easily be two years between learning conservation of number and conservation of volume
(Kreitler and Kreitler 1989). This is inconsistent with the stage theory.

32
understanding of the mathematics of catastrophe theory, a good overview of
the possibilities for testing cusp models, and knowledge of the controversies
from the early days of the popularization of this theory.

3.2 Examples of transitions

In Chapter 1, section 1.5, I stated that complex systems tend to be charac-

terized by a limited number of equilibria. As we saw in the Chapter 2, these
equilibria can take many different forms, but in this chapter, we consider only
stable and unstable fixed points. We are particularly interested in the case
where the configuration of stable and unstable points changes due to a smooth
change in some external variable, a control variable. In such a case a discon-
tinuous change or a (first-order) phase transition can occur. A transition or
tipping point is an intriguing property of complex systems.
I call it an intriguing property because in the linear systems we are used to,
smooth changes in control variables lead to similar (proportional) changes
in behavior variables. We may see a big change in some behavior, but this
requires a big change in the controls. An example would be the speed of your
bike and the force you apply. But in the case of fear or panic, this process is
often nonlinear. If a smooth change in an independent
or control variable, such as the smell
A key physical example is the change in state of water. Between, say, 10 and of smoke, leads to a sudden jump in
80 degrees Celsius, a smooth change in temperature results in only a slight fear (e.g., panic), we are likely to be
change in the liquid state of water. But if we change the temperature very dealing with a phase transition.
slowly, close to the thresholds of 0 or 100 degrees Celsius, we see sudden phase
transitions.
We saw something similar for the logistic map when 𝑟 crossed the boundary
at 𝑟 = 1. However, this did not lead to a sudden change in 𝑋 ∗ . This is
often called a second-order phase transition, meaning that the configuration
of stable and unstable points changes, but there is no discontinuous change in
behavior (figure 3.1).2

Figure 3.1: Examples of a first-order (discontinuous) and second-order (con-

tinuous) phase transition. In these particular cases, 𝑉 represents
volume and 𝑇 temperature.

2
A related and very similar distinction is that between a subcritical bifurcation and a
supercritical bifurcation.

33
Discontinuous phase transitions such as melting and freezing occur in many sys-
tems. Famous examples from the natural sciences include collapsing bridges,
capsizing ships, cell division, and climate transitions such as the onset of ice
ages. Examples from the social sciences include conflict, war, and revolution.
Some examples from psychology are falling asleep, outbursts of aggression,
radicalization, falling in love, sudden insights, relapses into depression or ad-
diction, panic, and multistable perception. The perception of the Necker cube
is a famous example (figure 3.2). Building and testing models of these psy-
chological transitions is challenging but rewarding. These transitions involve
large changes in behavior, in contrast to the smaller, often marginally signifi-
cant effects typically observed in psychological intervention studies.

Figure 3.2: Transitions in the perception of the Necker cube. The perception of
the middle cube is bistable, and sudden transitions occur between
the left (“front”) and right (“back”) percepts. Multistable percep-
tion is a much-studied psychological phenomenon that is still not
fully understood.

3.3 Bifurcation and Catastrophe theory

Bifurcation theory is a branch of
Bifurcations occur when equilibria disappear, appear, or split. Simply put, mathematics that studies changes in
bifurcation theory studies how small changes in parameters or conditions can the qualitative or topological structure
lead to large changes in outcomes in mathematical systems. of a given family of dynamical systems
as parameters are smoothly varied.
Catastrophe theory can be viewed as a branch of bifurcation theory, describ-
ing a subclass of bifurcations. It was developed by René Thom (1977) and
popularized by Christopher Zeeman (1976). The reason I chose to focus on
catastrophe theory in this chapter is fourfold: First, it provides one of the
few systematic treatments of bifurcations. A systematic treatment is more
effective than simply listing all types of bifurcations. Second, once you have a
grasp of the basics of catastrophe theory, it becomes easier to learn about other
bifurcations not encompassed by this theory. Third, it is the most widely used
approach in psychology and the social sciences. Finally, the field has developed
an empirical program and statistical procedures for the practical application
of catastrophe theory.
In gradient systems some quantity,
Catastrophe theory is concerned with gradient systems. These are dynamic such as energy, is minimized or
systems that can be described by a potential function. Potential functions can maximized.
be thought of as landscapes with minima and maxima in which we throw a
ball and see where it ends up. The simplest case, discussed in the next section,
is the quadratic minimum. We can also study what happens to the ball if the A potential function in mathematics
describes the potential energy
landscape of a system, where the
system’s dynamics are determined by
the gradients of this function.
34
landscape changes shape smoothly and a minimum disappears. Then sudden
jumps can occur.
Minima and maxima are called critical points, points where the first derivative
of the potential function is 0. Catastrophe theory analyzes so-called degenerate
critical points of the potential function. Phase transitions can occur at these
bifurcation points. Thom proved that there are only seven fundamental types
of catastrophes (given a limited set of control parameters). I will start with
a mathematical introduction and, after explaining the main concepts, give
some psychological examples. An in-depth discussion of the role of potential Degenerate critical points are points
functions in catastrophe theory can be found in the introduction of chapter 1 where not only the first derivative but
also the second derivative of the
of Gilmore (1993). potential function is zero.

3.3.1 The quadratic case

Thom’s theorems are known to be highly complicated, but the basic concepts
are not that difficult to grasp. The simplest potential function is

𝑉 (𝑋) = 𝑋 2 . (3.1)

Figure 3.3: The quadratic potential function. A ball rolls to the minimum
value of 𝑉 (𝑋). Its change is defined by the negative of the deriva-
tive of the potential function −𝑉 ′ (𝑋).

Imagine a ball in a landscape. The ball will roll to the minimum of the
potential function (figure 3.3). We learned in school that this is the point
where the first derivative is 0 and the second derivative is positive. The first
′ ″
and second derivatives are 𝑉 (𝑋) = 2𝑋 and 𝑉 (𝑋) = 2, respectively. At
𝑋 = 0 we find the minimum.
The potential function describes a dynamical system defined by

𝑑𝑋 ′
= −𝑉 (𝑋). (3.2)
𝑑𝑡
′
This makes sense. When the ball is in (1, 1), −𝑉 (𝑋) = −2 and the ball
′
will move toward 𝑋 = 0. But if 𝑋 = 0, −𝑉 (𝑋) = 0, and the ball will not
move anymore. In the case of the quadratic potential function, there is only
one fixed point. By adding parameters and lower order terms to 𝑉 , that is,
2
𝑎𝑋 + 𝑋 , we can move its location, but the qualitative form (one stable fixed

35
point) will not change. Also note that the second derivative is positive, which
tells us that we are dealing with a minimum and not a maximum (the so-called
second derivative test).
Many dynamical systems behave according to this potential function.3 Noth-
ing spectacular happens: no bifurcations and no jumps. This is different when
we consider potential functions with higher order terms.

3.3.2 The fold catastrophe

The fold catastrophe is defined by the potential function

3
𝑉 (𝑋) = −𝑎𝑋 + 𝑋 . (3.3)

This function has a degenerate critical (bifurcation) point at 𝑋 = 0, 𝑎 = 0,

′ ″
because at this point 𝑉 (𝑋) = −𝑎 + 3𝑋 2 = 0 and 𝑉 (𝑋) = 6𝑋 = 0, so both
the first and second derivative are 0. What makes this point so special? This
is illustrated in figure 3.4.

Figure 3.4: A bifurcation at 𝑎 = 0: the equilibria change qualitatively. For

𝑎 < 0 there is no equilibrium; for 𝑎 > 0 we have a minimum and a
maximum.

layout(t(1:3))
V <- function(X,a) -a * X + X^3
curve(V(x, a = -2), -3, 3, bty = 'n')
curve(V(x, a = 0), -3, 3, bty = 'n')
curve(V(x, a = 2), -3, 3, bty = 'n')

In the left plot, 𝑎 < 0 and there is no fixed point; the ball rolls away to minus
infinity. This can be checked by setting the first derivative to zero, which
gives 𝑋 = ±√𝑎/3. For negative 𝑎 there is no solution. A positive value of
𝑎 gives two solutions, as shown on the right for 𝑎 = 2. The positive solution
𝑋 = √2/3 is a stable fixed point because the second derivative in this point is
positive. The negative solution 𝑋 = −√2/3 is an unstable fixed point because
the second derivative in this point is negative.

3
As we will see in section 3.5.2.1, the statistical equivalent of the quadratic potential func-
tion is the normal distribution, the most popular distribution in our statistical work.

36
The middle figure depicts the case just in between these two cases. Here the
equilibrium is an inflection point, a degenerate critical point. The bifurcation
occurs at this point as we go from a landscape with no fixed points to one
with two, one stable and one unstable.
Another way to visualize this is by making a bifurcation diagram as we did
for the logistic map in Chapter 2. On the x-axis we put 𝑎, from 0 to 2. On
the y-axis we plot 𝑋 ∗ , the fixed points of equation 3.3. We use lines for stable
fixed points and dashed lines for unstable points. The diagram is shown in
figure 3.5. Note that the fold is not accompanied
by sudden jumps in behavior. It is an
example of a second-order phase
transition.

Figure 3.5: The bifurcation diagram of the fold catastrophe. Similar to what’s
shown in figure 3.4 , when 𝑎 = 0, there is a dramatic change in the
equilibrium landscape. Suddenly, both a stable and an unstable
equilibrium emerge, seemingly from nowhere.

This bifurcation diagram may not look as spectacular as the logistic map, but
its importance cannot be overstated. The fold is everywhere! In a fascinating
book, The Seduction of Curves, Allen McRobie (2017) shows that whenever we
see an edge, we see a fold. Figure 3.6 is from the book, where he demonstrates
how different catastrophes appear in art. I also recommend his YouTube
lecture.4
The fold catastrophe has been studied in fields from evolution theory (Dodson
and Hallam 1977) to buoyancy in diving (Güémez, Fiolhais, and Fiolhais 2002).
In addition, higher-order catastrophes are composed of folds. The fold catastrophe is also known as
a saddle node, tangential, or blue-sky
bifurcation.
3.3.3 The cusp catastrophe
Sudden jumps between stable states
The cusp, the best-known catastrophe, is the simplest catastrophe showing are associated with first order phase
sudden jumps in behavior. The potential function of the cusp is transitions.

1 1
𝑉 (𝑋) = −𝑎𝑋 − 𝑏𝑋 2 + 𝑋 4 . (3.4)
2 4

The half and quarter are added to make later derivations a little easier. The
highest power is now 4. The first two terms contain the control variables 𝑎
4
https://www.youtube.com/watch?v=6ZQKzcw9Ulk

37
Figure 3.6: The fold in drawings. The fold line separates the parts that can be
seen from the parts that are hidden. (Adapted from from McRobie
(2017) with permission)

and 𝑏, known as the normal and splitting variables. You might ask why there Control variables are the parameters
is no third order term. The nontechnical answer is that such a term would whose gradual changes induce
qualitative change in the behavior of
not change the qualitative behavior of the bifurcation. Catastrophe theory the system.
studies bifurcations that are structurally stable, meaning that perturbing the
equations (and not just the parameters) does not fundamentally change the
behavior (see section 3.3.6 and Stewart (1982) for further explication).
I advise you to do some minimal research on this equation yourself, using
an online graphic calculator tool like Desmos or GeoGebra (paste f(X)=-a X-
(1/2) b X^2+(1/4) X^4). For example, set 𝑎 = 1 and 𝑏 = 3 and look at the
graph of the potential function. Think in terms of the ball moving to a stable
fixed point. What you should see is that there are three fixed points, of which
the middle one is unstable. This bistability is important. Again, there is a
relationship to unpredictability. Although you know the potential function
and the values of 𝑎 and 𝑏, you are still not sure where the ball is. It could be
in either of the minima.
Other typical behavior occurs when we slowly vary 𝑎 (up and down from -2 to
2), for a positive 𝑏 value (𝑏 = 2). This is shown in figure 3.7. At about 𝑎 = 1.5
we see the sudden jump. The left fixed point loses its stability and the ball
rolls to the other minimum.
Now consider what will happen if we decrease 𝑎 from 2 to -2. In this case, Hysteresis means lagging behind, or
the ball will stay in the right minimum until 𝑎 = −1.5. Where the jump resistance to change.
takes place depends on the direction of the change in 𝑎, the normal variable.
The delay in jumps is called hysteresis. Hysteresis is of great importance in
understanding change or lack of change in complex systems. The state in
which the ball is the less deep minimum (for 𝑎 = 0.5 in figure 3.7) is often
called a metastable or locally stable state. Metastable states appear to be stable
for some time but are not in their
In his classic paper on the psychophysical law, Stevens (1957) reports hystere- globally stable state.
sis in perceptual judgments when properties such as brightness and loudness

38
Figure 3.7: The change in the potential function of the cusp by varying 𝑎. Note
that the jump to the other state does not happen at 𝑎 = 0 but is
delayed and depends on the direction of the change in 𝑎. This
delay is called hysteresis.

are systematically varied from low to high and vice versa. In this paper he
says: “I’m trying to describe it, not explain it. I am not sure I know how to
explain it.” To me the cusp at least partially explains why hysteresis occurs.
Gilmore (1993) made an important point about noise in the system. If there
is a lot of noise, the jumps occur earlier and we see less or no hysteresis effect.
This is called the Maxwell convention as opposed to the “delay” convention.
Demonstrating hysteresis therefore requires precise experimental control.
Another very interesting pattern occurs when 𝑎 = 0 and 𝑏 is increased (fig-
ure 3.8).

Figure 3.8: The change in the potential function of the cusp by varying 𝑏. One
minimum splits up in two.

For low 𝑏 there is one stable fixed point that becomes unstable. It splits up into
two new stable attractors. As we did for the fold, we can make bifurcation A pitchfork bifurcation occurs when a
diagrams showing the equilibria of 𝑋 as a function of 𝑎 and 𝑏. Along the single equilibrium splits into three
(two stable and one unstable) as a
𝑎-axis we see hysteresis and along the 𝑏-axis we see divergence or what is often parameter changes, resembling a
called a pitchfork bifurcation (figure 3.9). pitchfork’s shape.
Depicting the combined effects of 𝑎 and 𝑏 requires a three-dimensional plot,
which combines the hysteresis and pitchfork diagrams (figure 3.10).
The cusp diagram can be expressed mathematically by setting the first deriva-
tive to 0:

′ 3
𝑉 (𝑋) = −𝑎 − 𝑏𝑋 + 𝑋 = 0. (3.5)

This type of equation is called a cubic equation.5 The degenerate critical

points of the cusp can be found by setting the first and second derivative to

5
The cubic equation cannot be solved easily. This is due to the fact that the cusp is not a
function of the form 𝑦 = 𝑓(𝑥). Functions assign to each element of 𝑥 exactly one element
of 𝑦. But in bistable systems we assign two values of 𝑦 to one value of 𝑥.

39
Figure 3.9: Bifurcation plots for the 𝑎 and 𝑏 parameters of the cusp. Moving
along the 𝑎-axis, assuming 𝑏 is positive, gives hysteresis. Moving
along the 𝑏 axis, assuming 𝑎 = 0, gives the pitchfork bifurcation
or divergence. The dotted lines represent unstable maxima. The
area in the dotted box in the first plot is a fold.

Figure 3.10: The cusp catastrophe combines hysteresis along the normal axis
(𝑎) and the pitchfork along the splitting axis (𝑏). At the back
of the cusp, changes in 𝑎 only lead to smooth changes in the
equilibrium behavior 𝑋 ∗ . At the front, sudden jumps occur when
we cross the bifurcation lines. These jumps are typical of first
order phase transitions. The area between the bifurcation lines
is called the bifurcation set. In this area there are two stable and
one unstable equilibrium (shaded gray).

40
0. This is just within reach of your high school mathematics training, and I
leave this as an exercise. The result is:

27𝑎2 = 4𝑏3 . (3.6)

This equation defines the bifurcation lines where the first and second deriva-
tives are both 0 and sudden jumps occur (see figure 3.10). The region between
the bifurcation lines is the bifurcation set. In this region, the cusp has three
fixed points, the middle of which is unstable. These unstable states in the
middle are called the inaccessible area, shaded gray in the cusp diagram. The
bifurcation lines meet at (0,0,0). At this point, the third derivative is also 0.
This is the cusp point.

3.3.4 Examples of cusp models

To illustrate the cusp, I always use the business card (figure 3.11). I recom-
mend that you test this example (not with your credit card). You can play
with two forces. 𝐹 𝑣 is the vertical force and the splitting control variable (𝑏)
in the cusp. 𝐹 ℎ is the horizontal force and the normal variable (𝑎) in the
cusp. Note that you will only get smooth changes when 𝐹 𝑣 = 0, but sudden
jumps and hysteresis when you employ vertical force. One very important
phenomenon is that the card has no “memory” when 𝐹 𝑣 = 0. You can push
the card to a position, but as soon as you release this force (𝐹 ℎ back to 0),
the card moves back to the center position. This is not the case with vertical
pressure. If we force the card to the left or right position, it will stay there,
even if we remove the horizontal force. The card has a memory.

Figure 3.11: A simple business card can be used to illustrate all the properties
of the cusp (see main text).

41
This seems simple, but the mathematical analysis of such elastic bending struc-
tures is a huge topic in itself (Poston and Stewart 2014). The freezing of water
is also a cusp. As an approximation, we could say that the density of water
is the behavioral variable, temperature is the normal variable, and pressure
acts as a splitting variable (see chapter 14 of Poston and Stewart (2014), for a
more nuanced analysis). It is very instructive to study the full phase diagram
of water (figure 3.12). It can be viewed as a map of the equilibria. This type
of mapping would be extremely useful in psychology and the social sciences.

Figure 3.12: The phase diagram of water, which summarizes the equilibria and
transitions in the state of water as a function of temperature and
pressure.

A psychological example of a cusp concerns sudden jumps in attitudes (Latané

and Nowak 1994; van der Maas, Kolstein, and van der Pligt 2003). Attitudes
will be discussed in much more detail in later chapters. In general, we have
relatively stable attitudes toward many things in life (politics, snakes, ham-
burgers, and sports), but sometimes they change, and in rare cases they change
radically. For example, you may suddenly become a conspiracy theorist, an
atheist, or a vegetarian. One example is the attitude toward abortion (fig-
ure 3.13).
Cusp modeling begins by defining the states of the behavioral variable. In
this case, the two states of the bistable cusp are the two opposing positions,
pro-life6 and pro-choice. The other state associated with 𝑎 = 0, 𝑏 = 0 is the
neutral state, an “I don’t know” or “I don’t care” position.
The normal (𝑎) and splitting (𝑏) variable are interpreted as information and
involvement. Information is a collection of factors that influence whether
people tend to be in the pro-life or pro-choice position. Political and religious
orientation as well as personal experiences add to this overall factor. One way
to construct this information variable is through a factor analysis or principal
component analysis.
The splitting factor, involvement, also combines a number of effects (impor-
tance, attention). The main idea is that there are two types of independent
variables. Some will work (mainly) along the normal axis, and some will
(mainly) impact the splitting axis.
6
Perhaps better to call this view anti-choice.

42
Figure 3.13: The cusp model of attitudes (here toward abortion). Because of
hysteresis, it is very difficult to persuade highly involved people
with new information, but if they change it will be a sudden jump.

The implications of this model are that, for low involvement, change is contin-
uous (figure 3.13). Presenting people with some new information supporting
the pro-life or pro-choice position will have a moderate effect. One problem,
as demonstrated with the business card, is that the uninvolved person has
“no memory.” As soon as you stop influencing this person, they drift to the
neutral “I don’t care” position. We have another problem when people are
highly involved: they experience hysteresis. When this hysteresis effect is Because of the hysteresis effect, it is
large, persuasion just does not work. If you have been involved in political very difficult to persuade people with
new information.
discussions, you have probably experienced that yourself.
But if the underlying change in information is large enough, attitudes can
show a sudden jump. If they are central attitudes, they can be major life
events. There is a lot of anecdotal evidence for such transitions (Ebaugh 1988),
but it is very hard to capture such an effect in actual time series of attitude
measures. Another effect that is consistent with the cusp model is ambivalence.
Ambivalence is associated with high involvement. Highly involved people In the cusp model of attitudes,
with balanced information (𝑎 = 0), may oscillate between extreme positions ambivalence is not the same thing as
being neutral.
(see figure 4.2). Finally, the pitchfork bifurcation can explain issue or political
polarization. When involvement increases in a
group of neutral people, for example,
Another psychological example of the cusp-like behavior is multistable percep- due to discussion, they may split into
tion. Stewart and Peregoy (1983) proposed a model in which the perception two extreme positions (polarization).
of male face or female figure is used as a behavioral variable, the splitting
variable is the amount of detail, and the normal variable is a change in detail
related to the male/female distinction. The results are shown in figure 3.14.

3.3.5 Higher-order catastrophes

Note that the cusp is made up of folds. This is best seen in the hysteresis
diagram in figure 3.9 (see the dotted rectangle). Higher order catastrophes

43
Figure 3.14: Multistable perceptual stimuli positioned in the bifurcation set.
The fitted bifurcation lines were calculated using Cobb’s method,
which is explained in section 3.5.2.1. (Adapted from Stewart and
Peregoy (1983) with permission)

yield elements of cusps and folds. The swallowtail catastrophe with potential
3
function 𝑉 (𝑋) = −𝑎𝑋 − 12 𝑏𝑋 2 − 13 𝑐𝑋 + 15 𝑋 5 consists of three surfaces of fold
bifurcations meeting in two lines of cusp bifurcations, which in turn meet in
a single swallowtail bifurcation point. We need a four-dimensional space to
visualize this, which is difficult. The Wikipedia page on catastrophe theory
has some rotating graphs that may help. The butterfly catastrophe has 𝑋 6 as
the highest term (and four control variables).
The butterfly catastrophe is of interest
I will discuss this catastrophe in section 6.3.3.5 in relation to modeling atti- when we observe trimodal behavior.
tudes.7 Other catastrophes have two behavioral variables, not one. However,
the vast majority of applications of catastrophe theory focus on the cusp, which
will also be the focus of the remainder of this chapter. There are many good
(but not easy) books that present the full scope of catastrophe theory (Gilmore
1993; Poston and Stewart 2014).

3.3.6 Other bifurcations

In contrast to bifurcation theory, catastrophe theory is limited to structurally

stable, local bifurcations. Bifurcation theory also deals with
nonstructurally stable bifurcations
Examples of nonstructuraly stable local bifurcations are the transcritical bi- and so-called global bifurcations.
furcation ( 𝑑𝑋 2 𝑑𝑋 3
𝑑𝑡 = 𝑎𝑋 − 𝑋 ) and pitchfork bifurcation ( 𝑑𝑡 = 𝑏𝑋 − 𝑋 ). The
pitchfork is part of the cusp and is not structurally stable because it can be
perturbed by an additional term 𝑎, which, if unequal to 0, will distort the
pitchfork (see figure 3.15).

7
I note that the butterfly catastrophe and the butterfly effect in chaos theory are completely
unrelated concepts.

44
Figure 3.15: Perturbed pitchfork bifurcation (𝑎 = .1). For 𝑎 = 0 we would
get the pitchfork bifurcation as shown in figure 3.9. Thus, a
perturbation in a model parameter leads to qualitative change in
this bifurcation, and this is why it is not considered structurally
stable.

Another one we have already seen is the period doubling bifurcation. This
happened in the logistic map when the fixed point changed in a limit cycle of
period 2. Finally, global bifurcations cannot be localized to a small neighbor-
hood in the phase space, such as when a limit cycle diverges (Guckenheimer
and Holmes 1983). However, I don’t know of any applications of global bifur-
cations in psychology or the social sciences.

3.4 Building catastrophe models

3.4.1 Mechanistic models

The model of the attitude toward abortion is called a phenomenological model,

as opposed to a mechanistic model. In phenomenological models, we
assume the presence of a cusp, and
The mechanistic approach is much more common in the physical and life sci- make hypotheses about the involved
ences. An example is the phase transition in water described by the van der variables. In a mechanistic approach,
Waals equation. Poston and Stewart (2014) show how the van der Waal equa- the cusp is derived from basic
assumptions or first principles.
tion can be reparametrized to take the form of the cusp equation. The advan-
tage is that we learn how temperature and pressure are related to the control
variables of the cusp. This gives us a full understanding of the dynamics of
this phase transition.
One model that I will use as a psychological model in Chapter 4, section 4.3.7,
is the model of the spruce budworm outbreak, which occurs every 30 to 40
years and results in the defoliation of tens of millions of hectares of trees
(Ludwig, Jones, and Holling 1978). The model is

𝑑𝑁 𝑁 𝐵𝑁 2
= 𝑟𝑏 𝑁 (1 − ) − 2 . (3.7)
𝑑𝑡 𝐾 𝐴 + 𝑁2

45
Where 𝑁 is the size of budworm population, 𝑟𝑏 is the growth rate, 𝐾 is the car-
rying capacity, 𝐵 is the upper limit of predation, and 1/𝐴 is the responsiveness
of the predator.
The first part is the logistic growth equation. 𝑁 will grow to 𝐾 at a rate 𝑟𝑏 .
Note that this is a differential equation, not a difference equation. There is no
chaos in logistic growth in continuous time. The second part is the predation
function and has an increasing shape flattening out at 𝐵. The curvature of this
function is determined by 𝐴. High 𝐴 makes the function less steep, meaning
that predation reacts rather slowly to the increase in budworms (more about
the construction of this model later).
The analytical approach to this model is to reparametrize the model so that
it takes the form of a cusp. Such reparameterizations are not so easy to do
yourself. The idea is to create a smaller set of new variables that are functions
of the model parameters. For this model a convenient reparameterization is

𝐴 𝑟𝑏 𝐾
𝑟= and 𝑞 = . (3.8)
𝐵 𝐴

Using these two “constructed” control variables, we can depict the bifurcation
lines of the cusp as in figure 3.16.

Figure 3.16: The bifurcation diagram of the spruce budworm model. In the
bifurcation set, there are two alternative stable states: the normal
population level and the outbreak level.

In later chapters we will discuss psychological examples of a mechanistic ap-

proach, but as far as models of transitions are concerned, these are rare. The
phenomenological approach is much more common.

3.4.2 Phenomenological models

The cusp model of attitude is a typical phenomenological model. We simply

assume that the cusp is a model of the attitude. Phenomenological models
are less convincing than mechanistic models because they do not provide a

46
deep understanding of the underlying mechanisms that drive the system. But
in psychology and the social sciences, we cannot be too picky. Compared to
many other, verbally stated attitude models, the cusp attitude model is quite
precise. It implies a number of phenomena and is testable.
Setting up a phenomenological model is not a trivial task. I suggest some
guidelines for this. First, define the behavioral variable. It is important to
think about the bistable modes. What are they? What is the inaccessible
state in between? Can you have jumps between these states? What is neutral
state at the back of the cusp? If you cannot answer these questions, you should
reconsider whether a cusp is an appropriate model.
Second, select the control variables. What could be a normal variable and
what could be a splitting variable? These are not easy questions. Sometimes
there are too many candidates. For the cusp model of attitudes, instead of
involvement, we could suggest interest, importance, emotional value, etc. In
this case, I think of the splitting axis as a common factor of all these slightly
different variables. In other cases, we have no good candidates. In the example
in figure 3.14, it is not clear exactly what is being manipulated along the
normal axis. If you made a choice, it is good to check whether, at high values
of the splitting values, variation of the normal variable may lead to sudden
jumps and hysteresis. Also check whether the pitchfork bifurcation makes
sense theoretically.8
There is another issue here. In some phenomenological models, the control Control variables in cusp models can
variables are rotated by 45 degrees. The most famous example is Zeeman’s be rotated for ease of interpretation.
(1976) model of dog aggression (figure 3.17).

Figure 3.17: Zeeman’s dog aggression model with rage and fear as rotated
control variables.

8
Given these guidelines and examples, it is an interesting exercise to develop one’s own
cusp model, for example, for falling in love. This is a tricky exercise.

47
The control variables are fear and rage. In such a rotation the normal variable
is the difference between fear and rage, while the splitting variable is the
sum of fear and rage. Another example can be found in our model of the
speed-accuracy trade-off in reaction time tasks (Dutilh et al. 2011). When
constructing a phenomenological model, these two options for defining the
control variables should be considered.
To explain catastrophic drops in performance in work and sports, Hardy and
Parfitt (1991) proposed a cusp model with cognitive anxiety as the splitting
factor and physiological arousal as the normal factor. The idea is that at high
levels of cognitive anxiety, increases and decreases in arousal lead to sudden
changes, including a hysteresis effect. Hardy (1996) presents further tests of
this model, which has been criticized by Cohen, Pargman, and Tenenbaum
(2003). Extensions to the butterfly model are presented in Guastello (1984)
and Hardy, Woodman, and Carrington (2004).
Cusp models have also been developed for addiction (Guastello 1984; Mazanov
and Byrne 2006). Witkiewitz et al. (2007) propose using distal risk as the
splitting axis and proximal risk as the normal axis. The model is tested using
the renowned dataset from Project MATCH, an eight-year, multisite investi-
gation of the effectiveness of various treatments for alcoholism.
As a final example, I mention the model for humor presented by Paulos (2008)
in his fascinating book on mathematics and humor. Paulos explains his model
in the context of puns. His example is: “Do you consider clubs appropriate
for young children?” with the punchline “Only when kindness fails,” which is
probably only funny to people with children. Paulos uses the rotated control
axis as in the dog aggression model. Interpretation of the pun is the behavioral
axis. One axis represents the first meaning of “clubs,” the other axis represents
the second meaning. The bifurcation set represents the ambiguous region. A
joke involves a jump from one meaning to another. Paulos claims that The punch line forces a catastrophic
this cusp model combines cognitive incongruity theory, various psychological change in interpretation, accompanied
by a release of tension through
theories of humor, and the release theory of laughter. Tschacher and Haken laughter.
(2023) propose a related complexity account of humor.

3.5 Testing catastrophe models

3.5.1 The catastrophe flags

How sudden is sudden? How can climate changes be seen as transitions be-
tween stages (i.e., ice ages) when these transitions take hundreds of years?
Even when the ball is rolling toward its new minimum, it takes time to roll.
But then what is the difference with an continuous acceleration, such as we Sudden transitions are not
see in a logistic growth pattern? The time course of an acceleration and a instantaneous, but the in-between
states are unstable.
sudden, discontinuous jump may look very similar (figure 3.18).
In fact, in terms of time-series data, they may look exactly the same. The
main difference is that in the continuous case the intermediate values are sta-
ble. An acceleration can be understood as a quadratic minimum that changes
its position quickly. If we stop the process by freezing the manipulated con-
trol variable in the process, the state will remain at an intermediate value.
These intermediate values are all stable values. If we freeze the manipulated

48
Figure 3.18: Continuous and discontinuous growth curves my look very simi-
lar.

variable in a discontinuous process, it will continue to move to a stable state.

In this case, the intermediate state is unstable. The ball keeps rolling and
unfortunately the climate keeps changing.
In practice using time series data, this is a difficult distinction to make. It
means that simple time series are not sufficient to distinguish accelerations
from phase transitions. So how do we distinguish between the two processes?
In the context of catastrophe theory, Gilmore (1993) proposed the catastro-
phe flags. These are cusp-related phenomena that can be seen in the data.
While no single one of these is sufficient to indicate the cusp, when considered
together they provide compelling evidence for its existence.
In the following subsections, I will define the flags and illustrate their applica-
tions in psychology using examples. The first flag is the sudden jump.

3.5.1.1 Sudden jump

The sudden jump is a large fast
Although the sudden jump is not sufficient (it could be due to an acceleration), change in equilibrium behavior.
demonstrating a sudden jump in time series is useful (also in relation to other
flags). Statistical detection of sudden jumps is possible using a number of tech-
niques. Figure 3.19 presents raw weekly measurements of depressive symptoms
using the SCL-90-R depression subscale of a patient who gradually stopped
antidepressant medication during the study. The participant and researchers
were blind to the dose reduction scheme (Wichers, Groot, and Psychosystems
2016). One question was whether this reduction led to a sudden jump to the
depressed state. Using a change point detection method (James and Matteson
2014), we found a jump at 18 weeks with a bootstrapped 𝑝-value of .005 (with
the null hypothesis of no change point).
Many methods for change point analysis have been developed and compared
in Burg and Williams (2022).
The code for this figure is:

layout(t(1)); par(mar = c(4,4,1,1))

x <- read.table('data/PNAS_patient_data.txt', header = TRUE)

49
Figure 3.19: A sudden jump to depression (score at the SLC-90) in a patient
who gradually quit antidepressant medication during the study.

library(ecp) # if error: install.packages('ecp')

e1 <- e.divisive(matrix(x$dep, , 1), sig = .01, min.size = 10)
plot(x$week, x$dep, type = 'b', pch = (e1$cluster-1) * 16 + 1, xlab = 'Week',
ylab = 'SLC-90', bty = 'n', main = 'Jump to depression')

3.5.1.2 Multimodality

Multimodality (in the case of the cusp bimodality) is an important and easy-
to-use flag, as it can be tested with cross-sectional data. Finite mixture mod-
els have been developed to test for multimodality in frequency distributions
(McLachlan, Lee, and Rathnayake 2019).
An example is shown in figure 3.20. These data come from a conservation
anticipation task, where children have to predict the level of water in the
second glass when it is poured over. The resulting data and the fit of a mixture
of two normal distributions are shown on the right. The data are clearly
bimodal supporting the hypothesis of a transition in conservation learning
(van der Maas and Molenaar 1992). These data were used in Dolan and van
der Maas (1998) to fit multivariate normal mixture distributions subject to a
structural equation model.
The code is:

x <- unlist(read.table('data/conservation_anticipation_item3.txt'))
library(mixtools) # if error: install.packages('mixtools')
result <- normalmixEM(x)
plot(result, whichplot = 2, breaks = 30)

There is a whole field in statistics focused on multimodality, mixtures, and

clustering. There are some blogs that present overviews of the relevant R
packages (Arnaud 2021). Several detailed examples from psychology, using
hidden Markov models, are presented in Visser and Speekenbrink (2022).
The advantage of multimodality over
the sudden jump is that we can test it
with cross-sectional data.

50
Figure 3.20: Bimodality in the expected heights of water when it is poured
into a wider glass. This variation of the Piagetian conservation
task is used with children ages five to eight.

To capture a sudden jump in a development process, you need a lot of high-

frequency data. Sudden shifts in opinion are also rare. But it is easy to collect
data on large numbers of people who are asked to make judgments about
statements on an issue such as abortion. If these judgments are bimodally
distributed, this is consistent with a phase transition. Bimodal data may also
be produced by a process of acceleration, with time series consisting mainly of
data values before and after the acceleration. So, bimodality is not sufficient.
It can be considered necessary, so I always suggest starting with cross-sectional
multimodal studies. If they fail, you might reconsider your hypothesis. I have
often looked for multimodality in measures of arithmetic learning and never
found anything convincing, which made me rethink my hypothesis.

3.5.1.3 Inaccessibility

Inaccessibility means that certain values of the behavioral variable are unstable.
The business card is a good example. Given some vertical pressure, we can
try what we want but we cannot force the card to stay in the middle position;
it is unstable. Inaccessibility is relevant to reject the
alternative hypothesis that the sudden
In Experiment 2 of Dutilh et al. (2011), we focused on this flag. Our hypothe- jump and bimodality are due to an
sis was that in simple choice response tasks there is a phase transition between acceleration.
a fast-guessing state and a slower stimulus-driven response state. The idea is
that if we force subjects to speed up, there will be a catastrophic decline in
performance (from almost 100% correct to 50% correct).
We created a game in which subjects responded to a series of simple choice
items (a lexical decision task). The length of the series was not known to the
subject. At the end of a series, they were rewarded according to how close
their percentage correct was to 75%. Speed was also rewarded, but much
less. So, we asked the subject to be in the inaccessible state. The alternative
hypothesis, based on information accumulation models, was that there was no
phase transition and that responding with 75% accuracy required the correct
setting of a boundary (see section 4.3.1).

51
It appeared that subjects solved the task by switching between the fast-
guessing mode and the slower stimulus-controlled mode, even when instructed
according to the alternative model. Thus, the 75% intermediate state
appeared to be unstable.

3.5.1.4 Divergence

Divergence or the pitchfork bifurcation, the splitting up of an equilibrium,

requires the manipulation of the splitting variable. In the case of attitudes, we
hypothesize this to be involvement or some related variable. In van der Maas,
Kolstein, and van der Pligt (2003), we reanalyzed a dataset from Stouffer et
al. (1949), which Latané and Nowak (1994) presented as evidence for the
cusp model. The attitude concerned demobilization (from 0, unfavorable, to
6, favorable), and respondents were asked to indicate how strongly they felt
about their answer (from intensity 0 to intensity 5). For low intensities of
feeling, the data are normally distributed whereas for higher intensities, data
are bimodally distributed (see figure 3.21). After testing for multimodality,
testing for divergence is a sensible
next step.

Figure 3.21: The pitchfork bifurcation in attitudes. The dotted lines represent
the fit of the cusp model to these data. This technique will be
discussed in section 3.5.2.

3.5.1.5 Hysteresis
Hysteresis, the lagging behind of the
To test for hysteresis, we need to slowly increase and decrease the normal sudden jump, requires sophisticated
variable and test whether sudden jumps occur with a delay. We have demon- manipulation of the normal control
strated hysteresis in proportional reasoning using Piaget’s balance scale test variable.
in which a specific dimension (distance from the fulcrum) was systematically
varied (Jansen and van der Maas 2001). We also hypothesized that speed-
ing up subjects in response time tasks would eventually lead to a catastrophe
in accuracy. To support this claim, we demonstrated bimodality in response
times and hysteresis in the speed-accuracy trade-off (Dutilh et al. 2011). To
support the cusp model of multistable perception, we used the quartet motion
paradigm (Ploeger, van der Maas, and Hartelman 2002). In this perceptual
paradigm two lights are presented simultaneously, first a pair from two of the
diagonally opposite corners of the rectangle, and then a second pair from the
other two diagonally opposite corners of the rectangle. Usually, either vertical

52
or horizontal apparent motion is perceived. By gradually increasing or decreas-
ing the aspect ratio (i.e., the ratio of height to width of the quartet), hysteresis
in the jumps between the two percepts was demonstrated (see figure 3.22).
In Ploeger, van der Maas, and Hartelman (2002), we used a special design,
the method of modified limits, to rule out the alternative explanation that
hysteresis is simply due to delayed responses. It could be that the switches
always occur in the middle (at an aspect ratio of 1), but the self-report is
delayed. In the modified limits method, subjects do not respond during a trial,
only after the entire trial. By varying the length of the trials, it is possible to
determine at which parameter value the subject perceives a switch.

Figure 3.22: Hysteresis in the perception of apparent motion. Switches be-

tween the perception of vertical or horizontal apparent motion
occur when the aspect ratio (horizontal axis) is varied. The as-
pect ratio is the ratio of height to width of the quartet. (Adapted
from Ploeger, van der Maas, and Hartelman (2002) with permis-
sion)

53
3.5.1.6 Anomalous variance, divergence of linear response, and critical
slowing down

Gilmore’s last three flags—anomalous variance, divergence of linear response,

and critical slowing down—are indicators that occur near the bifurcation lines.
They are also known as early warning signals and are a popular topic of re-
search (Dakos et al. 2012). Early warnings are indicators or
signals that precede and predict
Anomalous variance occurs because near a bifurcation point the second deriva- transitions within a system, allowing
tive diminishes, meaning that the minimum becomes less deep. Assuming for anticipation and potentially
there is always some perturbation of the state, this will lead to larger fluctua- preventative action.
tions in the state.
Divergence of linear response is the size of the effect of a small perturbation of
the state, which will be greater near a bifurcation point. It will also take longer
to return to equilibrium. This delay in return time is known as critical slowing
down and is also studied in other approaches to nonlinear dynamical systems
(e.g., synergetics, Haken 1977). Examples of applications in psychology can be
found in Leemput et al. (2014) and Olthof et al. (2020). A somewhat critical
review is provided in Dablander et al. (2023).
In my experience, the problem with early warning signals is that both type 1
and type 2 errors should be low for predicting transitions. This is challenging
even in simulations, let alone in noisy psychological data. It would be fantastic
if these early warnings really worked. For example, being able to predict a
relapse into depression or addiction would be of great clinical value.
Together, the catastrophe flags provide a methodology for phase transition
research in psychology. A single flag may not be sufficient, but the combination
is. For example, the combination of evidence for inaccessibility and hysteresis
is convincing. I have given psychological examples of most of the flags. Which
flags to use in a particular case depends on the knowledge and experimental
control of the control variables. Another approach is to fit the cusp model
directly to the data. This is the subject of the next section.

3.5.2 Fitting the cusp to cross-sectional data

3.5.2.1 Cobb’s maximum likelihood approach

In a series of papers, Loren Cobb and colleagues (Cobb and Zacks 1985; Cobb
1978) developed a maximum likelihood approach9 to fit the cusp catastrophe
to data consisting of cross-sectional measurements of 𝑋, 𝑎, and 𝑏. We have
implemented this approach in a cusp R package described in Grasman, van
der Maas, and Wagenmakers (2009).
The basic idea is to make catastrophe theory, a deterministic theory, stochas-
tic by adding a stochastic term, called Wiener noise (with variance 𝜎2 ), to
equation 3.210 :

′
𝑑𝑋 = −𝑉 (𝑋)𝑑𝑡 + 𝜎𝑑𝑊 (𝑡). (3.9)
9
The is method finds the parameter values that make the observed data most probable.
10 ′
Many different notations exist for this. Perhaps clearer is 𝑑𝑋(𝑡) = −𝑉 (𝑋(𝑡)) 𝑑𝑡 +
𝜎𝑑𝑊 (𝑡), as both 𝑑𝑋 and 𝑑𝑊 depend on time.

54
It is important to note that this type of stochasticity is not the same as mea-
surement noise. Measurement noise—that is, 𝜀 in 𝑌 = 𝑋 + 𝜀—does not affect
the dynamics of 𝑋. Wiener noise does; it is part of the updating equation of
𝑋 itself. This stochastic differential equation is associated with a probability A stochastic differential equation
distribution of the form: (SDE) is a differential equation that
incorporates a term representing
random fluctuations.
1 −𝑉 (𝑋)
𝑓(𝑋) = 2
𝑒 𝜎2 , (3.10)
𝑍𝜎

where 𝑍 is a normalizing constant11 necessary to ensure that the area under

2
𝑓(𝑋) is 1. This may look complicated, but for the quadratic case 𝑉 (𝑋) = 12 𝑋 ,
√
this results in the standard normal distribution, with 𝑍 = 2𝜋/𝜎.
As in the case of the normal distribution, we want to allow for some trans-
formations of the variables. To simplify the necessary statistical notation, we
4
write the cusp as 𝑉 (𝑦) = −𝛼𝑦 − 12 𝛽𝑦2 + 14 𝑦 . The probability distribution for
the cusp is:

1 𝛽𝑦2 − 1 𝑦4
1 𝛼𝑦+ 2 4
𝑓(𝑦) = 2
𝑒 𝜎2 . (3.11)
𝑍𝜎

As in regression models, the cusp variables are modeled as linear function of

measured variables. That is, the dependent variables 𝑌𝑖1 , 𝑌𝑖2 , … , 𝑌𝑖𝑝 and the
independent variables 𝑋𝑖1 , 𝑋𝑖2 , … , 𝑋𝑖𝑞 , for subjects 𝑖 = 1, … , 𝑛, are related
to the cusp variables as follows:

𝑦𝑖 = 𝑤0 + 𝑤1 𝑌𝑖1 + 𝑤2 𝑌𝑖2 + … + 𝑤𝑝 𝑌𝑖𝑝 ,

𝛼𝑖 = 𝑎0 + 𝑎1 𝑋𝑖1 + 𝑎2 𝑋𝑖2 + … + 𝑎𝑞 𝑋𝑖𝑞 , (3.12)
𝛽𝑖 = 𝑏0 + 𝑏1 𝑋𝑖1 + 𝑏2 𝑋𝑖2 + … + 𝑏𝑞 𝑋𝑖𝑞 .

By estimating these regression parameters, we fit the cusp model to empirical

data. The cusp package in R makes this possible. I will first demonstrate this
using simulated data. I strongly recommend this approach. First, it forces My number-one rule when using
you to understand what the statistical technique actually does, and second, it statistical techniques: Never use a
statistical technique on real data
gives you a way to test the power and investigate violations of the technique’s before you have tested it on simulated
assumptions. data.

library(cusp) # if error: install.packages('cusp')

set.seed(1)
X1 <- runif(1000) # independent variable 1
X2 <- runif(1000) # independent variable 2
# to be estimated parameters:
w0 <- 2; w1 <- 4; a0 <- -2; a1 <- 3; b0 <- -2; b1 <- 4
# sample Y1 according to cusp using rcusp and the parameter values:
Y1 <- -w0/w1 + (1/w1) * Vectorize(rcusp)(1, a0 + a1 * X1, b0 + b1 * X2)
data <- data.frame(X1, X2, Y1) # collect ‘measured’ variables in data

11
For consistency with later chapters, I define 𝑍 differently from the notation in Grasman,
van der Maas, and Wagenmakers (2009). It is the inverse of 𝑍 in that paper.

55
I recommend doing some descriptive analysis first. With hist(data$Y1)
we can inspect whether there is some indication of bimodality. 𝑋2
is the splitting variable, so perhaps we see stronger bimodality with
hist(data$Y1[data$X2>mean(data$X2)]). The function pairs in R, pairs(data), is
also always recommended. In this perfect simulated case, you will already see
strong indications of the cusp. Now we fit the full model with 𝛼 and 𝛽 both
as function of 𝑋1 and 𝑋2.

fit <- cusp(y ~ Y1, alpha ~ X1+X2, beta ~ X1+X2, data)

summary(fit)

The table provides a summary:

Coefficients Estimate Std. Error z-value Pr(>|z|)

a[Intercept] -2.13 0.19 -11.0 < 2e-16***
a[X1] 3.11 0.22 14.2 < 2e-16***
a[X2] 0.15 0.17 0.9 0.39
b[Intercept] -2.29 0.34 -6.7 2.66e-11***
b[X1] -0.09 0.33 -0.3 0.79
b[X2] 4.40 0.27 16.5 < 2e-16***
w[Intercept] 1.98 0.07 27.6 < 2e-16***
w[Y1] 3.97 0.10 38.0 < 2e-16***

Table 3.1: The parameter estimates including standard errors and p-values
generated by the cusp package.

Note that we fit a model with too many parameters. We also estimated 𝑎2
and 𝑏1 (because the model was specified as alpha ~ X1+X2, beta ~ X1+X2).
These estimates are not significantly different from 0. The other parameters
are estimated reasonably close to their true values, since the true values fall
within the confidence interval of the estimates (defined by twice the standard
error on either side). We expect a better fit in terms of AIC and BIC when we
fit a reduced model without 𝑎2 and 𝑏1 . These fit indices penalize the goodness
of fit (e.g., the log-likelihood) for the number of parameters used to discourage
overfitting and to promote model parsimony.

fit_correct_model <- cusp(y ~ Y1, alpha ~ X1, beta ~ X2, data)

summary(fit_correct_model)

R.Squared logLik npar AIC AICc BIC

Full model 0.428 -1058.7 8 2133.3 2133.5 2172.6
Reduced 0.426 -1059.0 6 2130.0 2130.1 2159.5
model

Table 3.2: The comparative fit measures AIC, AICc, and BIC indicate that
the reduced model should be the model of choice.

The next simulation demonstrates that we can detect hysteresis using this
approach. We simulate data with −2 < 𝛼 < 2, and fixed 𝛽. If 𝛽 < 0 we have

56
no hysteresis, but if 𝛽 > 0, we do have hysteresis. With the code below we
simulate datasets for different 𝛽 and compare the goodness of fit between the
linear and cusp model. Figure 3.23 summarizes the results. Note that a lower
BIC indicated the better-fitting model.

set.seed(10)
n <- 500
X1 <- seq(-1, 1, le = n) # independent variable 1
a0 <- 0; a1 <- 2; b0 <- 2 # to be estimated parameters
b0s <- seq(-1, 2, by = .25)
i <- 0
dat <- matrix(0, length(b0s), 7)
for (b0 in b0s){
i <- i + 1
Y1 <- Vectorize(rcusp)(1, a1 * X1, b0)
data <- data.frame(X1, Y1) # collect ‘measured’ variables in data
fit <- cusp(y ~ Y1, alpha ~ X1, beta ~ 1, data)
sf <- summary(fit)
dat[i, ] <- c(b0, sf$r2lin.r.squared[1], sf$r2cusp.r.squared[1],
sf$r2lin.bic[1], sf$r2cusp.bic[1],
sf$r2lin.aic[1], sf$r2cusp.aic[1])
}
par(mar = c(4,5,1,1))
matplot(dat[,1], dat[,4:5], ylab = 'Bic', xlab = 'b0', bty = 'n', type = 'b',
pch = 1:2, cex.lab = 1.5)
legend('right', legend = c('linear','cusp'), lty = 1:2, pch = 1:2,
col = 1:2, cex = 1.5)
abline(v = 0, lty = 3)
text(-.5, 800, 'no hysteresis', cex = 1.5)
text(.5, 800, 'hysteresis', cex = 1.5)

Figure 3.23: At the back of the cusp (low 𝑏0 ), the cusp is approximately linear,
the BIC favors this simpler model (dotted line) over the cusp
mode (solid line).

57
3.5.2.2 Empirical examples

In Grasman, van der Maas, and Wagenmakers (2009), we present several ex-
amples with real data. As another example, we use Stoufer’s data, which we
used as an example of divergence before (see figure 3.21).

x <- read.table('data/stoufer.txt')
colnames(x) <- c('IntensityofFeeling', 'Attitude')
fit <- cusp(y ~ Attitude, alpha ~ IntensityofFeeling,
beta ~ IntensityofFeeling, x)
summary(fit)

Inspection of the parameter estimates shows that, as expected, intensity of

feeling only loads on the splitting axis and not on the normal axis. Figure 3.24
shows the location of the data in the bifurcation set (plot(fit)).

Figure 3.24: Placement of the data of figure 3.21 in the bifurcation set.

Another example is the conservation dataset of Bentler (1970), which contains

the scores on a 12-item test from a conservation test of 560 children from eight
different age groups (figure 3.25). These data are expected to be bimodal and
to move along the normal axis (van der Maas and Molenaar 1992).

x <- read.table('data/bentler.txt', header = TRUE)

layout(t(1:8))
age <- c('age 4 to 4.5','age 4.5 to 5','age 5 to 5.5','age 5.5 to 6',
'age 6 to 6.5','age 6.5 to 7','age 7 to 7.5','age 7.5 to 8')
for(i in 1:8){
if(i == 1) {par(mar = c(4,3,2,1)); names = 0:12} else
{names = ''; par(mar = c(4,1,2,1))}
barplot(table(factor(x[x[,1] == i,2], levels = 0:12)),
horiz = TRUE, axes = FALSE,
main = age[i], xlab = '',
names = names, cex.main = 1.5, cex.names = 1.5)

58
 }
fit <- cusp(y ~ score, alpha ~ age_range, beta ~ age_range, x)
summary(fit)
plot(fit)

This is supported by results of the cusp fit. You can verify that a model with
beta ~ 1 fits better according to the AIC and BIC.

Figure 3.25: In the Bentler conservation data, the sumscore distribution is

bimodal, and the weights of the modes shift with age.

A great exercise we have often used in the classroom is to build a Zeeman

machine, collect data with it, and fit the cusp model to the data (see Gras-
man, van der Maas, and Wagenmakers 2009 for details). Zeeman invented
this machine to demonstrate the properties of the cusp. Our students were
rewarded for the quality of the model and the artistic value of their Zeeman
machine (figure 3.26).

3.5.2.3 Evaluation

A few final remarks: First, Cobb’s method can be used with cross-sectional
data. Data points should be independent. To test for hysteresis in time series, Cobb’s method is not valid for time
other approaches are required. One option is to use hidden Markov models as series.
in Dutilh et al. (2011).
Second, there are some issues with Cobb’s approach that are due to fundamen-
tal differences between probability distributions and potential functions. The
latter can be transformed in many ways (so-called local diffeomorphisms) with-
out changing the qualitative properties of the cusp. With the added constraint
on probability distributions (area = 1), the same transformations can lead to
qualitative effects, such as a change in the number of modes. Wagenmakers
et al. (2005) suggest a solution to this problem for time series.
Third, two alternative approaches have been proposed. Both Guastello’s
(1982) change score least square regression approach and the Gemcat approach
(1987) use the first derivative of the cusp as point of departure. A problem

59
Figure 3.26: Zeeman’s catastrophe machine. It consists of a rotating disk and
two elastic bands. The first elastic band is attached to a fixed
point and the strap point. The end of the other elastic (red dot)
is moved by hand through the control plan. The strap point
moves according to the cusp catastrophe. Data is gathered by
collecting a set of X, Y, and Z values. Typically, 50 to 100 data
points are sufficient to apply the cusp fit function in R.

60
with both approaches is that they do not distinguish between stable and un-
stable equilibrium states. Data points in the inaccessible region improve the
fit of the model, whereas they should decrease the fit. Alexander et al. (1992)
provide a detailed critique.

3.6 Criticism of catastrophe theory

Rosser (2007) speaks of the rise and the fall of catastrophe theory. The hype
following the publication of Zeeman (1976) in Scientific American 12 , in which
he introduced the phenomenological application of catastrophe theory in the
behavioral and social sciences, led to a strongly worded reply by Zahler and
Sussmann (1977) in Nature.
Because people still refer to this paper when we use catastrophe theory in
our work, I will briefly respond to Zahler and Sussmann’s main points of
criticism. In their introduction, they state that there may be legitimate uses
of catastrophe theory in physics and engineering. They do not question the correctness
or importance of catastrophe theory
They raise 10 points, some of which I have already addressed. For example, as a purely mathematical subject.
their first point is about how sudden a jump actually is, but they call this
a less serious criticism. As I explained earlier, it is not the suddenness that
matters but whether or not the intermediate states are unstable.
A number of points are about inferring a cusp from data, which was indeed
done rather superficially in Zeeman’s earlier work. They point out that there
are no testable predictions, that the location of the cusp can be shifted, and
that there is no way to decide whether the data fit the cusp. I hope to have
shown that these problems are largely solved: the catastrophe flags allow us to
make new testable predictions, and with Cobb’s maximum likelihood approach
we can fit the model as we would do with any other statistical model in modern
science. Of course, one can be critical of the use of statistics in psychology
and the social sciences, but these criticisms are not specific to catastrophe
theory.
Another, somewhat inconsistent, line of criticism is that many catastrophe
models in psychology and the social sciences are just wrong and inconsistent
with the data (which could be true), while also not falsifiable. But you cannot
have it both ways: if it is wrong or inconsistent with the data, it is falsifiable.
Nevertheless, I agree that it is important to think about falsifiability. Theo-
ries in psychology tend to be moving targets. As soon as someone finds an
empirical result that contradicts the theory, the theory is quickly modified.
Then Zahler and Sussmann point out that catastrophe theorists often try to
make a discrete variable into a continuous one. Their example is aggression,
which they believe is inherently discrete. They call Zeeman’s interpretation
of aggression as a continuous family of behaviors absurd and utterly meaning-
less. This may be a bit strong. We can think of situations in which aggression
can vary from mild to severe, or from verbal to physical, directed at a per-
son’s belongings, mild physical directed at the person, to severe physical. A

12
To see Zeeman at work, I recommend the BBC documentary Case Study Catastrophe
Theory Maths Foundation Course (https://www.youtube.com/watch?v=myDvcvox1V
4&t=1435s).

61
rich ordering of aggressive acts is very useful for describing domestic violence.
Sometimes the change along these acts or variants is gradual, and other times
sudden. Whether such an ordering can be treated as a quantitative contin-
uum is one of the most difficult questions in our field (Borsboom et al. 2016;
Michell 2008).
Zahler and Sussmann’s final point is that there are better alternatives, such
as quantum mechanics, discrete mathematics, and bifurcation theory. There
is work on quantum mechanics in psychology (especially in the context of
consciousness), but whether it will lead to breakthroughs in this field remains
to be seen. Discrete mathematics may be an alternative in some cases (e.g.,
to model symbolic thinking). I see catastrophe theory as a special branch of
bifurcation theory, especially useful when the system under study is difficult to
describe in terms of mathematical equations. This goes back to the distinction
between phenomenological and mechanistic models. I think we should put
more effort into developing mechanistic models based on first principles. More
on this in the next chapters.
Loehle (1989) presents an excellent discussion on the usefulness of catastrophe
theory in the context of modeling ecosystems. He concludes that “an unre-
solved problem in applying catastrophe models is that of testing the goodness
of fit of the model to data,” but this problem has now been largely solved. The empirical program, using
catastrophe flags in conjunction with
Cobb’s method for fitting cusp models,
bypasses much of the previous
3.7 Conclusion criticism of catastrophe theory.

Psychologists are often concerned with psychological types and classes, stages
and phases, and the transitions between them. Our thinking about transi-
tions becomes much clearer and more advanced when we know the basics of
bifurcations.
Catastrophe theory comes with a toolbox for the behavioral and social sciences.
We can build phenomenological models, test for catastrophe flags, and even
fit cusp models to data. With the development of this toolbox, most of the
criticisms of catastrophe theory lose their relevance.
However, there is room for improvement. Phenomenological models have lim-
ited explanatory power. As explained in the next chapter on dynamical system
models, it is possible to create more mechanistic models that support the use
of phenomenological models. In Chapter 6, section 6.3.3, another option is
introduced using networks. I will demonstrate that the behavior of the Ising
network model for attitudes is governed by the cusp model, which is very
similar to the cusp model proposed for the attitude toward abortion.

3.8 Exercises

1) The equilibria of the fold are 𝑋 = ±√ 𝑎3 . This can be checked by setting

the first derivative to 0. Show this. (*)
2) In Zeeman’s dog aggression model, fear and rage are “rotated” control
variables. How can we translate this to a model with unrotated axes?

62
 Provide the equations that specify the normal and splitting axis as func-
tion of fear and rage. (*)
3) Derive the equation for the bifurcation lines of the cusp (27𝑎2 = 4𝑏3 ), by
setting the first and second derivatives to 0. Plot the bifurcation lines
in GeoGebra or Desmos. (**)
3
4) Some insight into the butterfly catastrophe 𝑉 (𝑋) = −𝑎𝑋 − 𝑏𝑋 2 − 𝑐𝑋 −
𝑑𝑋 4 + 𝑋 6 can be gained by entering the equation in free online graphing
calculators such as Desmos or GeoGebra. Set 𝑎, 𝑏, 𝑐, 𝑑 to 0, -5, 0, 5.
Then start varying 𝑎 and 𝑐. What is the difference in the effect of these
two parameters on the appearance and disappearance of attractors?
(**)
5) Set up a phenomenological cusp for falling in love. Follow my guidelines
(see section 3.4.2). (**)
6) Check whether indeed the Bentler data fit better when age_range only
loads on the normal axis (according to the AIC and BIC). What is the
correct specification of beta in cusp() in this case? (*)
7) What is the best fitting cusp model (according to the BIC) for this tricky
dataset created with this R code? Why? (**)

n <- 500
z <- Vectorize(rcusp)(1, .7 * rnorm(n), 2 + 2 * rnorm(n)) # sample z
x <- rnorm(n)
y <- rnorm(n)
data <- data.frame(z, x, y) # collect variables in data

8) Build a Zeeman machine, collect data, and fit the cusp (see Example III
of Grasman, van der Maas, and Wagenmakers 2009). What is your best
fitting model? Provide a plot of the data in the bifurcation set and a
picture of your Zeeman machine. (**)

63
4 Building dynamic system models

4.1 Introduction

Suppose you are in a bar in Amsterdam and someone asks if you would like
another beer. The number of drinks you have already had will probably in-
fluence your decision. Perhaps your self-control, whatever it may be, kicks in
and you refuse, even though the alcohol already in your system may be inter-
fering with that self-control. Or you may have reached your limit and simply
collapse. In this chapter, we will see how such a decision-making process can
be modeled using nonlinear differential equations.
This form of modeling is often called nonlinear dynamical systems theory
(NLDST), another branch of the complex-systems approach. We saw ex- A nonlinear dynamical system is one
amples of nonlinear dynamical system models in earlier chapters. The logistic in which the change of system
variables over time is governed by
map is an example of a discrete-time nonlinear dynamical model defined as nonlinear equations, resulting in
a difference equation. The catastrophe models are also dynamical systems complex behavior such as chaos and
governed by a potential function. In Chapter 3, section 3.4, I made a distinc- bifurcations.
tion between phenomenological modeling (assuming the cusp) and mechanistic
modeling (deriving the cusp from first principles). Here we will focus on the
more mechanistic construction of dynamical system models.
In psychology, following the principle of parsimony (Occam’s razor), we must
start with simple models. We don’t have many first principles to start with,
and our data are often limited, making model testing difficult. But we can
learn a lot from other disciplines. Nonlinear dynamical systems have been
developed in all the natural sciences, but my main inspiration comes from
mathematical biology, especially ecological modeling (Murray 1989). Math-
ematical psychology is generally less developed than mathematical biology,
but this depends somewhat on the subfield. In areas such as neural model-
ing, speeded-decision making, memory, choice, and psychometrics, there are
advanced models, and I will provide some examples later in this chapter.
I will first present a basic overview of dynamical systems modeling in other sci-
ences. Then I will discuss applications in psychology. I refer to more advanced
sources when necessary. I recommend the book by Gottman et al. (2002) for
its clear and basic explanation of the mathematical aspects of dynamical sys-
tems modeling. Strogatz’s online lectures on nonlinear dynamics and chaos
and his book (2018) are very helpful. Murray’s book (2002) on mathematical
biology is also highly recommended. Meadows’s Thinking in Systems (2008)
offers a basic introduction.
This chapter will be hands-on again. We will use the Grind package in R to
simulate dynamical systems models. Grind (de Boer 2018) is based on the
R packages deSolve and rootSolve (Soetaert, Petzoldt, and Setzer 2010). It

64
facilitates numerical integration, phase plane analysis, and stability analysis
of steady states.1
At the end of the chapter, I will introduce causal-loop diagrams and an open-
source tool, Insightmaker, that makes it easy to create causal-loop diagrams.
We will also use Insightmaker to simulate dynamical systems models, for which
I will provide some examples.

4.2 Basic concepts

4.2.1 Back to the logistic equation

We saw the logistic equation in the form of the logistic map (section 2.2), where
time progressed in discrete steps. The logistic map is a difference equation,
𝑋𝑡+1 = 𝑓(𝑋𝑡 ), but in this chapter we will focus on differential equations in
continuous time. We will limit ourselves to ordinary differential equations
(ODEs). The ODE for logistic growth2 is: In ordinary differential equations, we
take the derivatives with respect to
only one variable.
𝑑𝑋
= 𝑟𝑋(1 − 𝑋). (4.1)
𝑑𝑡

The change in 𝑋 is a function of 𝑋 itself. The exponential growth term 𝑟𝑋

dominates when 𝑋 is close to 0, but the growth levels off as 𝑋 approaches 1.
A solution to this equation expresses 𝑋𝑡 as a function of the initial state 𝑋0 .
In simple cases we can do this using the math we learned in high school. For
exponential growth 𝑑𝑋/𝑑𝑡 = 𝑟𝑋, this is the derivation:

𝑑𝑋
𝑑𝑡 = 𝑟𝑋 by separation of variables
𝑑𝑋
𝑋 = 𝑟𝑑𝑡
∫ 𝑑𝑋
𝑋 =∫ 𝑟𝑑𝑡 integrate
ln 𝑋 = 𝑟𝑡 + 𝐶 assuming 𝑋 ≥ 0
𝑋 = 𝑒𝑟𝑡+𝐶 = 𝑒𝐶 𝑒𝑟𝑡 by taking the exponent
𝑋0 = 𝑒𝐶 𝑒𝑟0 ⟹ 𝑋0 = 𝑒𝐶 compute the integration
constant

𝑟𝑡
⟹ 𝑋𝑡 = 𝑋0 𝑒 . (4.2)

But for more complex models, such an analytical solution is out of scope and
numerical solutions (by simulation) are required. This is not the preferred
choice. These simulations can be slow, they may accumulate rounding errors,
and it can be difficult to search the entire parameter space, especially when
multiple parameters are involved.
The naive implementation of differential equations in R is risky. This would
involve a for loop:

1
The manual can be found at https://github.com/hansschepers/grindr/blob/master/inst/
documentation/GRIND%20tutorial.pdf.
2
In many texts 𝑑𝑋 is written as 𝑋.̇
𝑑𝑡

65
 x <- x0 <- .1 # initial value
r <- .5 # growth rate
dt <- .00001 # time step in simulation
t <- 10 # Nt, time we want to know the value of x
timesteps <- t/dt # required time steps given t and dt
for(i in 2:timesteps) # note the 2 to use the starting value
x[i] <- x[i-1] + r * x[i-1] * dt
x0 * exp(r * timesteps * dt) # analytical solution
x[timesteps] # compare
timesteps # length of simulation

where 𝑑𝑡 must be chosen by hand. If you test some values of 𝑑𝑡, you will see
that a value too high (.5) leads to a solution (x[timesteps]) that is different
from the analytical solution. But if we set 𝑑𝑡 very low (.00001), it takes
unnecessarily long.
This is why we use solvers, numerical methods for ordinary differential equa-
tions. We will use the R package Grind, although many other methods are Solvers are specialized algorithms
available in R. One could also directly use the R packages deSolve and root- designed to numerically approximate
solutions to ODEs and handle the
Solve by Soetaert, Petzoldt, and Setzer (2010), on which the Grind package is complexities of integrating these
based. Grind has to be installed from GitHub using: equations over time to predict the
evolution of system states under
install.packages("remotes") different initial conditions and
remotes::install_github("hansschepers/grindr") parameters.

The packages required are:

library(deSolve)
library(rootSolve)
library(FME)
library(Grind)

The code consists of defining the model, the parameters 𝑝, and the initial
values 𝑠. Main functions are run(), plane(), newton(), continue(), and fit().
They will be introduced using examples. With run() we generate a time series
for the model.

model <- function(t, state, parms) {

with(as.list(c(state, parms)), {
dX <- r * X # the exponential model
return(list(c(dX)))
})
}
p <- c(r = .5) # parameter r
s <- c(X = .1) # initial value
run(tmax = 5) # run until t = 5, numerical solution
s['X'] * exp(p['r'] * 5) # compare with analytical solution

We don’t have to worry about time steps anymore, and the numerical and
analytical solutions converge. This is of course a trivial use of an ODE solver,
but much more can be done.

66
In analyzing the behavior of a dynamical system, first we want to know what
the equilibria 𝑋 ∗ are. To do this, we need to set the time derivative equal
to 0, 𝑑𝑋/𝑑𝑡 = 0. For the exponential function, this is simply 𝑟𝑋 = 0, that
is, when 𝑋 is 0. Second, we want to determine whether these equilibria are
stable or unstable. Whether 𝑋 ∗ = 0 is stable can be determined by checking
the second derivative in 𝑋 ∗ . If this derivative is less than 0, then the fixed
point is stable. The second derivative is 𝑟, so 𝑋 ∗ = 0 is an unstable fixed point
whenever 𝑟 > 0 and stable whenever 𝑟 < 0. You can check this in Grind by
using 𝑟 values of -.1 and .1, and start values equal to or just above or below
0.
For equation 4.1, the logistic function, we also want to know the equilibria,
the stable and unstable fixed points. To do so we follow the same steps as for
the exponential function (see exercises).
The continuous-time implementation of the logistic function is somewhat bor-
ing compared to its discrete-time variant that we studied in section 2.2. The
difference is that the overshooting and undershooting do not occur in contin-
uous time. By changing the logistic model in Grind to:3

dX <- r * X * (1-X) - X

and using method='euler' in the run() function, you can simulate the discrete-
time logistic map. Check if you get chaos for 𝑟 = 4. Use the Euler method
only in special cases, as it is generally the least accurate approach.

4.2.2 The Lotka—Volterra models

Perhaps the best-known population models are the Lotka—Volterra equations

(Murray 2002). These consist of coupled differential equations, one for the The Lotka—Volterra models describe
density of the prey and one for the density of the predator: the dynamics of biological systems in
which two species interact, one as a
predator and the other as prey.
𝑑𝑁
= 𝑎𝑁 − 𝑏𝑃 𝑁 ,
𝑑𝑡 (4.3)
𝑑𝑃
= 𝑐𝑃 𝑁 − 𝑑𝑃 ,
𝑑𝑡

where 𝑁 and 𝑃 refer to the sizes of the prey and predator populations, 𝑎
and 𝑐 determine the growth rates, and 𝑏 and 𝑑 control the mortality rates.
Note that the mortality rate of prey depends on both 𝑁 and 𝑃 , while the
mortality rate of predators depends only on 𝑃 . Similarly, the growth terms
are also asymmetric, predators increase as a function of both 𝑁 and 𝑃 , as
they eat prey. We will follow the simple example provided by Wikipedia (on
Lotka—Volterra equations).
To implement this model in Grind, we use:

LV <- function(t, state, parms) {

with(as.list(c(state, parms)), {
dN <- a * N - b * P * N

3
The −𝑋 is added because the difference equation has the form 𝑋𝑡+1 = 𝑓(𝑋𝑡 ), so the
change 𝑑𝑋 is thus 𝑓(𝑋𝑡 ) − 𝑋𝑡 .

67
 dP <- c * P * N - d * P
return(list(c(dN, dP)))
})
}
p <- c(a = 1.1, b = .4, c = .1, d = 0.4) # parameters
s <- c(N = 10, P = 10) # 10 baboons and 10 cheetahs

Some typical uses of Grind are:

layout(1:2)
data <- run(odes = LV, tstep = .01, table = TRUE) # set tstep to low value
# phase plot for different starting values:
plane(odes = LV, portrait = TRUE,
ymax = 17, xmax = 50, tstep = 0.1, grid = 4)

The plane function makes a phase plot with 𝑁 and 𝑃 as axes. The black
points are initial states. What we learn from this is that the equilibrium of
the Lotka—Volterra equations is a limit cycle that depends on the choice of
the initial conditions.
A well-known improvement to this model is to make the prey growth density
dependent by using the logistic equation. This can be done by setting dN
<- a*N *(1-N) - b*P*N in the model. This is the case used as an example in
the Grind tutorial, which I highly recommend reading (de Boer 2018). It
also contains the appropriate parameter values for this model variant. In this
density-dependent model, there are fixed points, in contrast to the original
model. This shows that such model choices can have a large effect.
A famous example of a system of three coupled differential equations is the
Susceptible-Infected-Recovered (SIR) model used to model infectious diseases
and to understand the impact of interventions on disease dynamics. The The Susceptible-Infected-Recovered
states of the model are susceptible, representing individuals who have not yet (SIR) model is a basic epidemiological
model that divides a population into
contracted the disease but are at risk; infected, representing individuals who susceptible (S), infected (I), and
are currently infected and can transmit the disease to susceptible individuals; recovered (R) individuals.
and recovered, representing individuals who have recovered from the disease
and are assumed to be immune and no longer susceptible. The differential
equations specify the change in susceptible, infected, and recovered members
of the population. You can now easily implement this model yourself (see
exercises).

4.2.3 Fitting models: Stochasticity versus noise

Grind includes an option to fit dynamical systems models. With fit(), based
on the modFit() function from the FME package (Soetaert, Petzoldt, and Setzer
2010), one can estimate the model parameters given a dataset. These functions
also provide confidence intervals and allow fixing parameters and bootstrap
analysis. Fitting nonlinear dynamical systems models to data is an art in
itself. For example, these methods can be very sensitive to the choice of initial
values.
I will illustrate the use of fit() on three datasets created with the original
Lotka—Volterra model from the previous section. The first dataset is the

68
deterministic dataset, the data that follow directly from the code above. The
second is created using a stochastic Lotka—Volterra model. I will explain
how this works in the next section. The third is a deterministic dataset with
measurement error. We will see that the last two cases are very different.

set.seed(1)
layout(matrix(1:4, 2, 2, byrow = TRUE))
p <- c(a = 1.1, b = .4, c = .1, d = 0.4)# p is a named vector of parameters
s <- c(N = 10, P = 10) # s is the state
n <- 30
data_deterministic <- run(odes = LV, n, table = TRUE,
timeplot = FALSE) # deterministic data
data_stochastic <- run(odes = LV, n, table = TRUE,
after="state<-state+rnorm(2,0,.1)", timeplot =
FALSE) # add stochasticity
data_error <- run(odes = LV, n, table = TRUE, timeplot = FALSE)
data_error[,2:3] <- data_error[,2:3]+
matrix(rnorm(2 * n, 0, 2), , 2) # measurement error
#fit & plot
s <- s * abs(rnorm(2, 1, 0.1)); s # start values
p <- p * abs(rnorm(4, 1, 0.1)); p # start values
f_deter <- fit(odes = LV, data_deterministic, main = 'deterministic')
f_stoch <- fit(odes = LV, data_stochastic ,main = 'stochastic')
f_error <- fit(odes = LV, data_error, main = 'error')
pars <- matrix(c(f_deter$par[3:6], f_stoch$par[3:6], f_error$par[3:6]), ,3)
pars <- rbind(pars, c(summary(f_deter)$sigma, summary(f_stoch)$sigma,
summary(f_error)$sigma))
barplot(t(pars), beside = TRUE, names = c('a','b','c','d', 'Residuals'),
args.legend = c(x = 13),
legend.text = c('deterministic', 'stochastic', 'error'))

This results in figure 4.1.

Note that the error dataset looks very similar to the deterministic dataset
because it contains only the measurement error (𝑋 is true score + error). The
error does not affect the dynamics itself. In the stochastic case, the error
(noise) is added to the states after each time step, which affects the dynamics.
In figure 4.1, you can see that the positions of the waves change. In this
well-chosen case, the fit is quite good in all three cases, and the parameter
estimates are all quite close to the true values. Unfortunately, these results
are quite unstable. You can do some testing yourself.

4.2.4 Back to the cusp

To illustrate how Grind can be used to perform bifurcation analysis, we go

back to the cusp. Recall that the differential equation for the cusp is

𝑑𝑋 ′ 3
= −𝑉 (𝑋) = 𝑎 + 𝑏𝑋 − 𝑋 . (4.4)
𝑑𝑡

69
Figure 4.1: Fit of the Lotka—Volterra model on three types of data. The lines
represent the fitted curves. In the stochastic case, noise is part of
the system that affects the computation of the state at the next
time step. In the error case, noise is a measurement error that
does not affect the dynamics.

model <- function(t, state, parms){

with(as.list(c(state, parms)),{
dX <- a + b * X - X^3 # cusp
return(list(dX))
})
}
p <- c(a = 0, b = 1); s <- c(X = .1); run(ymin = -1)
s[1] <- -.1; run(add = TRUE)

This code simulates two runs demonstrating bistability for 𝑎 = 0 and 𝑏 = 1.

A nicer way to demonstrate bistability in the time series is to make the system
stochastic. This was done in Chapter 3, section 3.5.2.1, by using a stochastic
′
differential equation: 𝑑𝑋 = −𝑉 (𝑋)𝑑𝑡 + 𝜎𝑑𝑊 (𝑡). Grind has a great trick for
this. With the “after” parameter in the function call, we can add discrete
events to the system. “After” can also be used to change parameter or state
values after a certain amount of time or after some condition (see the manual).
We use it here to add a random number sampled from a normal distribution,
with a mean of 0 and a standard deviation of .4, to 𝑋. The best way to simulate
this in Grind is using the Euler
√ method with a small time step. The noise term
should be corrected with 𝑑𝑡, as shown in the code. The Wikipedia page
on stochastic differential equations will tell you more about the underlying
ideas.
As can be seen in figure 4.2, the stochastic force causes spontaneous jumps
between the two modes of the cusp. When noise or random fluctuations cause
the entire equilibrium landscape of a dynamical system to become observable,
it is often referred to as stochastic resonance. You can see this by comparing Stochastic resonance is a notable
the figure with one generated with a standard deviation of .1 or less. example of how noise, which is often
considered undesirable or disruptive,
can actually play a constructive role
in certain systems, helping to reveal
hidden patterns and structures that
70 might otherwise remain obscured.
 layout(t(c(1,1,1,2)))
data <- run(table = TRUE, tmax = 1000, method = 'euler', tstep = .1,
after = "state <- state + rnorm(1,mean=0,sd=0.4) * sqrt(tstep)",
ymax = 2, ymin = -2, timeplot = FALSE)
plot(data, type = 'l', bty = 'n')
barplot(hist(data[,2], 30, plot = F)$counts, xlab = "X", hor = TRUE)

Figure 4.2: Spontaneous jumps in the cusp due to stochastics (noise). Due
to stochastic perturbations, the system occasionally jumps over
the maxima that separate the minima. Interestingly, when the
noise is reduced, the time series tends to become trapped in a
single equilibrium. Thus, increased noise helps reveal the overall
equilibrium landscape. This phenomenon is known as stochastic
resonance.

4.2.5 Bifurcation analysis

By combining the Grind functions newton() and continue(), we can perform

bifurcation analysis. The newton() function finds stable and unstable fixed
points, and the continue() function implements the parameter continuation
of a steady state, providing a bifurcation diagram. It shows the change in
equilibria when we change a parameter. This is what we did in Chapter 2 for
the logistic map, when we varied 𝑟 and plotted the equilibria (see figure 2.8).
It is often necessary to run the combination of these two functions repeatedly,
starting from different initial states. The code to create figure 4.3 is:

p <- c(a = 0, b = 1)
low <- newton(s = c(X = -1)) # finds a minimum starting from X = -1
# Continue this steady state varying a
continue(low, x = "a", y = "X", xmin = -2, xmax = 2, ymax = 2)
high <- newton(s = c(X = 1)) # again starting from X = 1
continue(high, x = "a", y = "X", xmin = -2, xmax = 2, ymax = 2, add = TRUE)

71
Figure 4.3: Hysteresis plot made with newton() and continue(). The function
newton() finds an equilibrium, which is used in continue() to vary
the normal variable a until a bifurcation point is found.

Another great tool in R is the deBif package by de Roos.4 This is an R

Shiny application that uses the same model specification and allows for a
more interactive investigation. Given our previous model and the definition
of 𝑠 and 𝑝, we can run:

install.packages("deBif")
library(deBif)
phaseplane(model, s, p)

The phaseplane() function returns a time plot and the steady states. You can
change parameters and initial states on the left side, and plot parameters on
the upper right side (click on the two gears). The Steady States option is very
useful as it shows the stable and unstable fixed points. Make sure that the
minima and maxima of the plot axes are set correctly.
With

bifurcation(model, s, p)

You can create one- and two-parameter bifurcation diagrams (using the LP
curve option, see figure 4.4). The two-parameter bifurcation diagram (bottom
left) cannot be created in Grind. See the deBif help pages (with ??deBif) for
further instructions.

4.2.6 Spruce budworm outbreak model

In Chapter 3, section 3.4.1, I introduced the spruce budworm outbreak model.

We will use this model later as a model of addiction. The bifurcation diagram
can be made with:

spruce <- function(t, state, parms) {

with(as.list(c(state, parms)), {
du = r * u * (1 - u/q) - u^2 / (1 + u^2)
return(list(c(du)))
})

4
https://cran.r-project.org/web/packages/deBif/deBif.pdf

72
Figure 4.4: Output from Shiny app deBif. The last plot is a two-dimensional
bifurcation diagram showing the bifurcation lines of the cusp in
the 𝑎, 𝑏 plane. This plot cannot be made with Grind.

}
state <- c(u = 0.5)
parms <- c(r = 0.4, q = 10)
bifurcation(spruce, state, parms)

Note that this predator-prey model consists of only one equation. There is no
separate dynamic equation for the birds. The reason is that these budworm
outbreaks happen in a few weeks. Birds do not reproduce on this time scale.
The variables are reparametrized (see section 3.4.1). The predation term, in
the original parametrization −𝐵𝑁 2 /(𝐴2 + 𝑁 2 ), also has a logistic form that
starts to accelerate at 𝑁 = 𝐴 up to the maximum level 𝐵. The slow start 𝐴 is
used because birds only switch their diet to budworms when this population
reaches a certain level (Ludwig, Jones, and Holling 1978). The fixed number
of birds can only eat 𝐵 budworms. This specific predation term is called the
Holling type III model. All Holling types and their formulas are shown in
figure 4.5.

4.2.7 Evaluation of ecological modeling

Understanding the technical basics of dynamical systems theory is one thing,

but actually building useful dynamical system models is quite another. Every
term in each differential equation of a model needs some underpinning. These
models make many assumptions, both implicit and explicit. The Lotka—
Volterra model, for example, assumes that the prey population grows expo-
nentially in the absence of the predator, that the predator population dies off
with the prey population and does not switch to other prey species, that the

73
Figure 4.5: The Holling functional response models. Type III is used in the
spruce budworm model.

response of the predator population to changes in the prey population is direct

and not delayed, that there is no spatial component to the model, and that
the rates of change of the populations are proportional to their sizes, to name
just a few. These assumptions are widely debated in the biological literature
(Abrams et al. 2000), and modifying these assumptions may have significant
consequences. Hundreds of extensions and variants
of the Lotka—Volterra model have
For example, the original Lotka—Volterra model has no stable points, only been proposed and studied.
limit cycles. While these cycles have been observed in nature, they are not
overwhelmingly common. As we have seen, the dynamics of the system are
significantly altered when prey growth is made dependent on prey density.
This model has fixed-point equilibria instead of limit cycles.
Adding a spatial component can also make a big difference, as shown in the
example of hypercycles in see section 5.2.3 (see Szostak, Wasik, and Blazewicz
(2016) for a brief review). Adding more prey and predator species also makes
a difference (Johnson, Mumma, and St-Laurent 2019). There are many in-
teresting options for the predator term in the prey equation. Tyutyunov and
Titova (2020) compare 12 trophic functions, alternatives to the Holling func-
tional responses. The options are overwhelming. Biologists face a problem
here that I discussed in Chapter 1, section 1.4. Models easily become too
complex. Recall that the traffic simulation models were extremely simple, yet
sufficient to explain key phenomena.
An additional problem is that empirically testing all these different models
is difficult. Although the quality of biological data is often superior to that
of psychological data, biologists must also rely on the qualitative predictions
of their models. Models in chemistry and especially physics can often be
tested quantitatively. Transitions occur precisely at the predicted values of
the control variables. Ecological models, much like those in psychology, do
not allow for this level of prediction. This is a problem because if we can only

74
test our model qualitatively (Are there limit cycles? What type of transitions
can be detected? Is there hysteresis?), many model choices are not particularly
relevant. One of the most significant challenges
in complex-systems research in the life
A case in which this is less of an issue is the traffic example that I introduced sciences and psychology is
in Chapter 1. I asked you to play around with the online simulation. We constructing dynamical system models
now know the basics to better understand this model. The Wikipedia page on that effectively address these
data-related issues.
this model (Intelligent Driver Model) presents the equations, which are also
coupled ordinary differential equations. The implementation in Grind of the
simplest case looks like this:

model <- function(t, state, parms){

with(as.list(c(state, parms)),{
x <- state[1:n]
v <- state[(n+1):(2*n)]
dx <- v # change in distance = speed
delta_v <- v - m %*% v # difference in speed to next car
s_alpha <- m %*% x - x - l # distance to next car
s_alpha[n] <- 100 # front car has no car in front
s_star <- s0 + v * T + v * delta_v / (2 * sqrt(a * b))
dv <- a * (1 - (v/v0)^delta - (s_star/s_alpha)^2) # change in speed
return(list(c(dx, dv)))
})
}

n <- 50
p <- c(l = 5, v0 = 30, T = 1.5, a = .73, b= 1.67 , delta = 4, s0 = 2)
x_init <- (0:(n-1)) * (p['s0'] + p['l'])
v_init <- rep(0, n)
s <- c(x_init, v_init)
m <- diag(1, n, n); m <- rbind(m[-1,], 0) # order cars
# simulation with front car suddenly breaking at t = 150:
data <- run(tmax = 300, timeplot = FALSE, table = TRUE,
after = 'if (t==150) state[2*n] = 0')
matplot(data[,2:(n + 1)], type = 'l', bty = 'n', xlab = 'time', ylab = 'x')

The result of this simulation is shown in figure 4.6. Understanding the rea-
soning behind the differential equation is not so easy, but I want to make
another point. The Wikipedia page gives parameters values with units (s,
m/s, or m/s2 ). One can also have dimension-free parameters (the acceleration
exponent). This dimensional analysis is a crucial step in modeling in physics Dimensional analysis involves
but a weak point in biological and especially psychological applications. This analyzing the dimensions of quantities
to derive relationships and scaling
hampers the quantitative test of models. laws, ensuring that the equations are
consistent.

4.3 Psychological models

In this section, I present an overview of dynamical systems models in psy-

chology, primarily in the form of systems of differential equations. Although
the list is extensive, it is not exhaustive. It is important to explore different
models and applications before embarking on your own modeling efforts.

75
Figure 4.6: The traffic jam simulation. The top line represents the front car,
which moves off immediately. Other cars are waiting for their turn.
At 𝑡 = 150, the first car suddenly breaks off, creating a traffic jam
for the later cars. The effect of this disturbance is greater for the
last car than for the first car. This simulated graph resembles the
real data very well (see for example figure 9 in Jusup et al. 2022).

4.3.1 Response time models

Many dynamic models have been proposed in the study of speeded decision-
making (Bogacz et al. 2006). The best-studied case is the two-alternative
forced-choice task, where a stimulus is presented, and a choice must be made
between two alternatives as quickly as possible. The stimulus could be an
arrow pointing left or right. Most popular are accumulator models (figure 4.7).
Accumulator models assume that
noisy information is accumulated over
One way to model this process is with a single stochastic linear differential time until a decision bound is reached
equation, called the drift-diffusion model (DDM), with 𝐼 as the stimulus-driven and a motor response is initiated.
input:

𝑑𝑋 = 𝐼𝑑𝑡 + 𝜎𝑑𝑊 . (4.5)

As before, 𝑑𝑡 is moved to the left-hand side of the equation. 𝑑𝑊 is white

noise, normally distributed with 0 mean and with standard deviation 𝜎 (set
to .1 by default). We start at 0, 𝑋𝑡=0 = 0, assuming no bias for one of the
choice alternatives.
The implementation of the model in Grind is quite simple. The trick is again
in the run statement, which adds white noise after each step.5 We stop the
run when either the negative or positive bound is reached.

model <- function(t, state, parms) {

with(as.list(c(state, parms)), {
dX = I
return(list(c(dX)))

5
Simulating this model correctly is more difficult than one might expect. I refer to Tuer-
linckx et al. (2001) for a discussion of methods.

76
Figure 4.7: A stochastic accumulator model of speeded decision-making. Ev-
idence accumulates in stochastic steps biased by the drift rate 𝐼
(stimulus related). When one of the bounds is reached, a response
is generated that may be incorrect if the bounds are too low.

})
}
p <- c(I = .01); s <- c(X = 0)
bound <- 1
run(table = TRUE, method = 'euler', tstep = .1,
tmax = 500, after = "state<-state+rnorm(1,mean=0,sd=0.1)*sqrt(tstep);
if(abs(state)>bound) break", # stop at bound
ymin = -bound, ymax = bound)

The model explains observed response time and accuracy in terms of the un-
derlying process parameters, drift rate, and confidence bound. By fitting the
model to the data, we can determine whether slow responses are due to a low
drift rate (low skill or difficult task) or a conservatively chosen bound. Accumulator models such as the
drift-diffusion model explain the
A well-known extension of the drift-diffusion model is the Ornstein— speed-accuracy trade-off. If we set our
Uhlenbeck model: confidence bound higher, we are
slower but more accurate.

𝑑𝑋 = (𝜆𝑥 + 𝐼)𝑑𝑡 + 𝜎𝑑𝑊 . (4.6)

For 𝜆 < 0 this process converges to 𝐼/𝜆 (assuming 𝜎 = 0), while for 𝜆 > 0 it
diverges. For the psychological interpretation, I refer to Bogacz et al. (2006).
The simplest two-dimensional model is the race model:

𝑑𝑋1 = 𝐼1 𝑑𝑡 + 𝜎𝑊1 ,
(4.7)
𝑑𝑋2 = 𝐼2 𝑑𝑡 + 𝜎𝑊2 .

Now two independent processes run (race) to one positive bound. The first
one to arrive wins. More biologically inspired models involve inhibition. The

77
equations of mutual inhibition model are:

𝑑𝑋1 = (−𝑘1 𝑋1 − 𝑤1 𝑋2 + 𝐼1 )𝑑𝑡 + 𝜎𝑊1 ,

(4.8)
𝑑𝑋2 = (−𝑘2 𝑋2 − 𝑤2 𝑋1 + 𝐼2 )𝑑𝑡 + 𝜎𝑊2 .

Note that these are all linear dynamical systems that do not exhibit complex
behavior. Examples of nonlinear alternatives are presented in Roxin and Led-
berg (2008) and Verdonck and Tuerlinckx (2014) and discussed in Ratcliff et
al. (2016).
The relations between different accumulator models are summarized in fig-
ure 4.8. It shows that convenient models such as the drift-diffusion mode can
be derived by constraints on the parameters from more biologically realistic
models, such as the pooled and mutual inhibition model.

Figure 4.8: Relations between the main evidence accumulator models of

decision-making. DDM is the drift-diffusion model which can be
derived from the mutual inhibition model by setting 𝑤 = 𝑘 and
𝑤 + 𝑘 to a large value. (Adapted from Bogacz et al. 2006)

4.3.2 Dyadic models

The study of dyadic interaction lends itself to dynamic modeling. Dyadic

interactions have been studied extensively in the field of caregiver-child inter-
actions (Ainsworth et al. 2015). Here, we focus on a dyadic interaction in
romantic relationships.

4.3.2.1 Romeo and Juliet

One type of model can be traced back to publications by Rapoport (1960) and
Strogatz (1988). I follow the setup described by Sprott (2004). Note that it

78
was intended as a toy model to demonstrate dynamical modeling.
The model is about the interactions between Romeo and Juliet, where 𝑅 and 𝐽
represent the feelings of Romeo and Juliet. The change in feelings is supposed
to be a function of the feelings of both people:

𝑑𝑅
= 𝑎𝑅 + 𝑏𝐽 ,
𝑑𝑡 (4.9)
𝑑𝐽
= 𝑐𝑅 + 𝑑𝐽 .
𝑑𝑡

First note that the case of 𝑏 = 𝑐 = 0 resembles the exponential model with
𝑎𝑡 𝑑𝑡
solutions 𝑅 = 𝑅0 𝑒 and 𝐽 = 𝐽0 𝑒 , which converge (to 0) or diverge (to
infinity) depending on whether 𝑎 and 𝑑 are negative or positive. We will see Divergence in this model, unbounded
a more sensible setup in the next model. Nevertheless, this system of coupled exponential growth of positive feelings,
is an attractive concept but
linear differential equations is surprisingly rich in behavior. With the signs unrealistic, I’m afraid.
of the parameters we define very different romantic styles. Strogatz (1988)
distinguishes the following:

• Eager beaver: 𝑎 > 0, 𝑏 > 0, Romeo is encouraged by his own feelings

as well as Juliet’s.
• Narcissistic nerd: 𝑎 > 0, 𝑏 < 0, Romeo wants more of what he feels
but retreats from Juliet’s feelings.
• Cautious (secure) lover: 𝑎 < 0, 𝑏 > 0, Romeo retreats from his own
feelings but is encouraged by Juliet’s.
• Hermit: 𝑎 < 0, 𝑏 < 0, Romeo retreats from his own feelings as well
as Juliet’s.

Juliet may have her own style, which leads to complicated interactions. Sprott
(2004) and other sources give an extended analytical treatment of this model.
If you want to learn more about dynamical systems, you should study matrix Systems of linear differential equations
algebra and its applications in linear dynamical systems. I have chosen to leave can be solved analytically, and the
behavior of the equilibria can be
it out of this book because most psychological dynamical systems models are characterized by the eigenvalues.
nonlinear. Here we just use Grind to test some cases. I give three examples Some knowledge of matrix algebra is
with three different sets of parameter values. In the first case, the initial required to understand this.
mutual interest fades; in the second case, the relationship fizzles out after
some ups and downs; and in the third case, the couple ends up in a cycle of
hate and love.

model <- function(t, state, parms) {

with(as.list(c(state, parms)), {
dR <- a * R + b * J
dJ <- c * R + d * J
return(list(c(dR, dJ)))
})
}
layout(matrix(1:6, 3, 2, byrow = TRUE))
p <- c(a = -1, b = 1, c = .5, d = -1) # parameters
s <- c(R = 0.1, J = .1)
run()
plane(portrait = TRUE, ymin = -1, xmin = -1, grid = 3,

79
 vector = TRUE, legend = FALSE)
p <- c(a = -.2, b = -1, c = 1, d = 0) # parameters
run(ymin = -.2, legend = FALSE)
plane(portrait = TRUE, ymin = -1, xmin = -1, grid = 2,
tstep = .001, legend = FALSE)
p <- c(a = -.1, b = -1, c = 1, d = 0.1) # parameters
run(ymin = -.2, legend = FALSE)
plane(portrait = TRUE, ymin = -1, xmin = -1, grid = 3,
tstep = .001, legend = FALSE)

Figure 4.9: Three different love affairs between Romeo and Juliet.

The lines in the phase plots in figure 4.9 are the nullclines. In linear dynamical Nullclines are the curves for which the
system, nullclines are straight lines. Where they intersect, stable or unstable time derivatives of the behavioral
variables 𝑅 and 𝐽 are 0.
fixed points can occur. Depending on the angle between the nullclines, we
get a fixed point (first two cases), a limit cycle (last case), or divergence (not
shown).
Rinaldi (1998) proposed an extension and a constraint to the model that makes
it a bit more realistic and easier to study. The basic equation is now 𝑑𝑅/𝑑𝑡 =
−𝑎𝑅 + 𝑏𝐽 + 𝐴𝐽 , where 𝑎 is interpreted as a forgetting parameter (constrained
to be positive) and 𝐴𝐽 is the attractiveness of the Juliet. In this case, a
necessary and sufficient condition for asymptotic stability (i.e., having a fixed
point) is that 𝑎𝑑 > 𝑏𝑐.
Rinaldi also considers the case of a population of heterosexual men and women
with different levels of attractiveness. The idea is that a man and a woman will

80
leave their current partners and bond together when both reach a more optimal
level of love. Rinaldi analyzes the conditions under which the population
reaches a stable state. This marriage assignment problem, as it is called, is an
example of a famous problem in optimization theory known as an assignment
problem. The goal is to find a stable assignment of men to women, such that
no man and woman prefer each other to their current partners (Gale and
Shapley 1962). We use such algorithms to assign students to master tracks in
our educational program.
Some other advanced variations of this model have been proposed. In these
papers the analysis of the mathematical properties of the model gets much
more attention than the psychological theory. It is often unclear what, exactly,
the variables are and what the reasoning behind certain model assumptions.
The work of Murray and Gottman, discussed in the next section, is more
interesting in this regard.

4.3.2.2 The mathematics of marriage

The model of marriage developed by the psychologist John Gottman and the
mathematical biologist James Murray (2002) is firmly grounded in psychologi-
cal theory and data. The main phenomenon that inspired this modeling work
is Gottman and Levenson’s (1992) finding that the patterns of interaction be-
tween couples, when discussing a major area of ongoing disagreement in their
marriage, are predictive of divorce.
The model consists of two coupled difference equations, but I present it in the
form of differential equations.6

𝑑𝑊
= 𝐼𝑤 (𝐻, 𝑎, 𝑏) − 𝑟𝑤 𝑊 + 𝑊𝑒 ,
𝑑𝑡 (4.10)
𝑑𝐻
= 𝐼ℎ (𝑊 , 𝑎, 𝑏) − 𝑟ℎ 𝐻 + 𝐻𝑒 ,
𝑑𝑡

where the influence functions 𝐼𝑤 and 𝐼ℎ are defined as:

𝑠𝑔𝑛(𝑥)
𝐼(𝑥, 𝑎, 𝑏) = . (4.11)
1 + 𝑒𝑎(|𝑥|−𝑏)

I made up this flexible function to allow for very different forms of influence
(as we will see below). When both influences are 0 (𝑎 = −8, 𝑏 = −∞), the
state or mood of the wife (𝑊 ) and the husband (𝐻) converge to 𝑊𝑒 and 𝐻𝑒 ,
with rates 𝑟𝑤 and 𝑟ℎ , respectively. 𝑊𝑒 and 𝐻𝑒 are the uninfluenced steady
states of mood when the spouses do not interact.7
However, if the influence function (𝑎 = −8, 𝑏 = 0) is such that a positive mood
in one spouse provokes a positive mood in the other, while a negative mood
provokes a negative mood, we expect a negative and a positive equilibrium
depending on the initial states and uninfluenced steady state values.
6
Difference equations were used in the original model because the data consist of turn
takings in a conversation. This, however, does not lead to qualitative different results.
With method='euler' and a change in 𝑟𝑤 and 𝑟ℎ the difference model can be constructed.
7
I follow the definition and notation of the original source, but this model is clearly not
restricted to heterosexual relationships.

81
Another more complex influence function (𝑎 = −8, 𝑏 = 1) assumes that only
extreme mood states influence the other spouse.
This is implemented with:

model <- function(t, state, parms) {

with(as.list(c(state, parms)), {
dW <- influence(H,a,b) - rw * W + We
dH <- influence(W,a,b) - rh * H + He
return(list(c(dW, dH)))
})
}
layout(matrix(1:9, 3, 3, byrow = TRUE))
par(mar=c(4,4,1,2))
influence <- function(x, a = -8, b = 1)
sign(x) / (1 + exp(a * (abs(x) - b)))
p <- c(rw = .6, rh = .6, We = .18, He = -.18, a = -8, b = Inf)
s <- c(W = 0, H = 0)
for(b in c(Inf, 0, 1)){
p['b'] <- b
curve(influence(x, -8, b), -3, 3, xlab = 'W', ylab = 'H', lwd = 2)
plane(xmin = -2.5, xmax = 2.5, ymin = -2, ymax = 2, legend = FALSE)
for(i in seq(-2, 2, by = .25)) newton(s = c(W = i, H = i), plot = TRUE)
for (i in 1:100)
run(state = c(W = rnorm(1,0,.5), H = rnorm(1,0,1)), tmax = 50, ymin = -2,
ymax = 2, add = (i>1), legend = FALSE)
}

This results in figure 4.10.

The three influence functions are shown in the first column, the nullclines in
the second, and a series of runs from random initial states in the third. The
first case shows that the moods converge to 𝑊𝑒 and 𝐻𝑒 , when there is no
mutual influence. The second has two equilibria, with an unstable fixed point
in the middle, while the last case has five fixed points, three of which are
stable.
I have kept the influence functions and most parameters the same for both
spouses, but this is not necessarily the case. It is possible to derive the equi-
libria analytically and to determine the stability of these equilibria (Gottman
et al. 2002). They also propose a two-stage procedure for fitting the model
to data consisting of positive and negative speaker interactions using a bi-
linear influence function. A more advanced statistical approach is proposed
in Hamaker, Zhang, and van der Maas (2009). For a related model for the
interaction between therapist and client, see Tschacher and Haken (2019).

4.3.3 The van Geert models

In a series of papers, Paul van Geert proposed dynamical systems models for
developmental processes (Den Hartigh et al. 2016; van Geert 1998, 1991). It is thought that cognitive and
language abilities grow over time in an
Van Geert has proposed many different models, but I will give just one example. autocatalytic process constrained by a
Van Geert (1991) introduced a system of two coupled difference equations to limited capacity, similar to the logistic
growth of populations.

82
Figure 4.10: Qualitative difference marriage equilibrium landscapes depend-
ing on the form of the influence function. In the first (top), they
simply have no influence and both partners converge to their un-
influenced steady states of mood. In the second (middle) , the
response to the partner’s mood is extreme, resulting in either a
positive or negative mutual state. In the last case (bottom), the
response to low positive or negative moods is close to 0 but ex-
treme at higher levels. Now there are three stable states.

83
model where the growth rate of one cognitive ability depends on the level of
another cognitive ability:

𝑎𝑋𝑡2
𝑋𝑡+1 = (𝑎 − 𝑏𝑌𝑡 )𝑋𝑡 − ,
𝐾 (4.12)
𝑐𝑌 2
𝑌𝑡+1 = (𝑐 − 𝑑𝑋𝑡 )𝑌𝑡 − 𝑡 .
𝐾

This can be implemented as follows:

model <- function(t, state, parms) {

with(as.list(c(state, parms)), {
dX <- (a + b * Y) * X - a * X^2 / K
dY <- (c + d * X) * Y - c * Y^2 / K
return(list(c(dX, dY)))
}) }

layout(matrix(1:4,2,2))
# Set parameter values and run the model:
p <- c(K = 1, a = 0.4, b = -0.05, c = .4, d = -0.15)
s <- c(X = 0.01, Y = 0.01)
run(method = "euler", tstep = 1)
plane(portrait = TRUE, grid=4)
p <- c(K = 1, a = 0.05, b = -0.1, c = 0.05, d = -0.09)
s <- c(X = 0.0126, Y = 0.01)
run(tmax = 1500, method = "euler", tstep = 1)
plane(portrait = TRUE, grid = 4)

Figure 4.11: The development of two interrelated cognitive abilities in one of

the van Geert models. In the left case they coexist; in the right
case one ability suppresses the other.

So, there are basically two outcomes: either both grow, or one grows and
suppresses the other (see figure 4.11). Note that the method is set to “Euler”

84
to simulate difference equations.

4.3.4 The Pólya urn model of the third source

Another type of discrete dynamical systems model is the Pólya urn model,
which is relevant to understanding nonlinear developmental processes in psy-
chology. Molenaar, Boomsma, and Dolan (1993) propose the third-source
hypothesis. Based on a series of studies, Gärtner (1990) concluded that The third-source hypothesis proposes
70—80% of the variation in body weight in inbred mice appears to be due to that the development of complex
living systems is influenced by three
a third component that generates biological variability in addition to genetic sources of variation: genetic variation,
and environmental influences. environmental variation, and
self-organizing processes.
A simple, and in my opinion insightful, dynamical model for this effect is the
Pólya urn model (Mahmoud 2008). In this discrete dynamical model, we add
marbles to an urn containing some red and blue marbles. We could start with
two blue and one red marbles. We randomly take out a marble. If it is blue, we
put it back with another blue marble. If it is red, we put it back and add a red
marble. Initially, 𝑝(𝑏𝑙𝑢𝑒) = 2/3, but what will happen to that probability over
time when we have more and more marbles? Think about it for a moment.
My intuition was simply wrong, and in my experience, this is true for the vast
majority of people. What happens is shown in figure 4.12. Each time you
run the process, 𝑝(𝑏𝑙𝑢𝑒) reaches a stable state, but the value of that state is
random. What happens is that early (random) samples have a huge influence
on the long-term dynamics. This creates a Matthew effect. The Matthew effect says that the rich
get richer and the poor get poorer.
Savi et al. (2019) provide a developmental interpretation. Imagine a child re-
ceives a tennis racket for her birthday. First, she practices the backhand twice
at home, but incorrectly. Then, during her first tennis lesson, her trainer
demonstrates the correct backhand. She now has three experiences, two in-
correct and one correct. Now, suppose her backhand development is based
on a very simple learning schema. Whenever a backhand return is required,
she samples from her earlier experiences, and the sampled backhand is then
added to the set of earlier experiences. Thus the cumulation of experiences
follows the Pólya urn scheme. While she has the potential to become a tennis
master, her twin sister, who had less fortunate initial experiences, decides to
quit tennis lessons within the first year. This model is consistent with many
developmental theories (e.g., the critical period hypothesis), but these theories
lack a formal approach.

4.3.5 The panic model

In recent years, we have been working on a model of panic disorder (Robinaugh

et al. 2019). In theories of panic disorder, there is a reinforcing feedback loop
between arousal and perceived threat. When an increase in arousal is perceived
as a threat (e.g., a heart attack), arousal increases further. This “vicious cycle”
results in a panic attack (D. M. Clark 1986). Thus, these theories posit two
causal effects: an effect of perceived threat on arousal and an effect of arousal
on perceived threat.
We will further assume that the effect of perceived threat on changes in arousal
is essentially linear while the causal effect of arousal on perceived threat is

85
Figure 4.12: The Pólya urn model. A random marble is sampled and placed
back with an extra marble of the same color. The evolution of
the probability of picking a blue marble is unpredictable and con-
verges to a random number. This mechanism may play a role in
the Matthew effect.

86
nonlinear (S-shaped). For the argument, see Robinaugh et al. (2019). It
could be argued that both are nonlinear, but this does not fundamentally
change the qualitative behavior of the model. The central part of the model
consists of two coupled differential equations:

𝑑𝐴
= −𝐴 + 𝑏𝑇 ,
𝑑𝑡 (4.13)
𝑑𝑇 1
= −𝛼(𝐴+𝛽)
− 𝑇.
𝑑𝑡 1+𝑒

This looks a bit like the Romeo and Juliet model, but now the effect of arousal
𝐴 on the change in perceived threat 𝑇 is a logistic function that starts at 0 and
grows to 1. The location is determined by 𝛽, and the acceleration or steepness
is determined by 𝛼.
An implementation and simple illustration is:

model <- function(t, state, parms) {

with(as.list(c(state, parms)), {
dA <- -A + b * T
dT <- -T + 1/(1 + exp(-alpha * (A + beta)))
return(list(c(dA, dT)))
})
}
p <- c(b = 1, alpha = 12, beta = -.7)
s <- c(A = 0, T = 0)
# arousal increase for time t in 20:30, leads to panic,
# which after some time ('30 min') disappears
layout(1:2)
plane(vector = TRUE, xmin = 0, ymin = 0, xmax = 1,
ymax = 1.1, legend = FALSE)
newton(s = c(A = 0, T = 0), plot = TRUE)
newton(s = c(A = 0.8, T = .8), plot = TRUE)
newton(s = c(A = 1, T = 1), plot = TRUE);
run(after="if(t>20&t<30)state[1]<-1;state<-state+rnorm(2,mean=0,sd=0.1)")

The 𝛽 parameter is set so that the nonpanic mode dominates, but the panic
mode is present (a metastable state). This state can be easily disturbed (see
plane). For 20 < 𝑡 < 30, arousal is set to a high value, resulting in a high
perceived threat. But because we also added some noise to both processes,
after some time both arousal and perceived threat jump back to low values.
This dynamic of this model is the cusp, as can be checked with (see fig-
ure 4.15):

p <- c(b = 1, alpha = 12, beta = -.5)

start <- newton(s = c(A = .1, T= .1))
continue(start, x = "beta", y = "T", xmin = -1, xmax = 1, ymax = 1)
continue(start, x = "alpha", y = "T", xmin = -1, xmax = 20, ymax = 1)
start <- newton(s = c(A = 1, T = 1)) # finds a minimum starting from X = -1
continue(start, x = "alpha", y = "T", xmin = -1, xmax = 20,
ymax = 1, add = TRUE)

87
In Robinaugh et al. (2019), this model is extended with other processes, such
as arousal and escape schemes, that operate on the parameters of the basic
model. These are slower processes that are modeled on different time scales.

Figure 4.13: The panic model. Arousal is set high between time is 20 and 30,
but panic persists due to the hysteresis effect. Eventually, due to
noise, it escapes from the metastable attractor at 𝐴 = 𝑇 = 1.

4.3.6 Neural models: Van der Pol and different time scales

In the panic model, we touched on differences in time scales. Time scales The difference in time scales refers to
are critical to understanding and managing complex systems because they the different rates at which system
components or processes evolve,
allow fast dynamics to be separated from slow dynamics, thereby simplifying affecting the overall behavior of the
analysis and modeling. I will explain this further in the context of the van der system.
Pol oscillator, which has many interesting applications.
3
Imagine taking the cusp equation 𝑑𝑋/𝑑𝑡 = 𝑎 + 𝑏𝑋 − 𝑋 , with 𝑏 = 1, such
that we have hysteresis. But now we make 𝑎, or actually 𝑑𝑎/𝑑𝑡, a function of
𝑋: 𝑑𝑎/𝑑𝑡 = −𝜀𝑋, where 𝜀 is small constant. If we set 𝜀 to .05, 𝑎 changes 20
times slower than 𝑋. What happens now is that with 𝑋 = 1 and 𝑎 = 0, 𝑎
decreases up to the point where 𝑋 jumps to a negative value. Now 𝑎 increases,
resulting to a new jump to a positive value of 𝑋. And this loop will continue
endlessly.

model <- function(t, state, parms){

with(as.list(c(state, parms)),{
dX <- a + b * X - X^3 # cusp
da <- -e * X
return(list(c(dX, da)))
})
}
s <- c(X = .1, a = 0) # initial state and parameter values
layout(matrix(1:4, 2, 2, byrow = TRUE))
p <- c(e = .05, b = -.5)
run(ymin = -.1, main = 'b = -.5', legend = FALSE)
plane(xmax = 2, ymin = -1, ymax = 2, xmin = -2,
portrait = TRUE, grid = 2, main= 'b = -.5')

88
 p <- c(e = .05, b = 1)
run(ymin = -1.5, main = 'b = 1', legend = FALSE)
plane(xmax = 2, ymin = -1, ymax = 2, xmin = -2,
portrait = TRUE, grid = 2, main= 'b = 1')

Figure 4.14: Two runs of the van der Pol oscillator. For high b, this system
oscillates between the two stables states of the cusp. The black
dots represent different initial states.

The plots (figure 4.15) illustrate this behavior. For 𝑏 < 0, 𝑋 converges to a
fixed point. For 𝑏 > 0, we see cyclic jumps up and down. This oscillator is
basically the famous van der Pol oscillator, originally written in the form:

𝑑2 𝑋 𝑑𝑋
2
= 𝜇 (1 − 𝑋 2 ) + 𝑥. (4.14)
𝑑𝑡 𝑑𝑡

Such a second-order differential equation can be rewritten in the form of a

first-order system of multiple equations, which is the form required for Grind.8
This model form is of (neuro)psychological interest because it relates to the
FitzHugh—Nagumo model for neuronal excitability (Izhikevich and FitzHugh
2006)9 :

𝑑𝑉 𝑉3
=𝑉 − − 𝑊 + 𝐼,
𝑑𝑡 3 (4.15)
𝑑𝑊
= .08(𝑉 + .7 − .8𝑊 ).
𝑑𝑡

The equation for 𝑉 , the membrane potential, has a cubic nonlinearity that
allows regenerative self-excitation via positive feedback. 𝑊 , a recovery vari-
8
The rewriting is based on the Liénard transformation.
9
The FitzHugh—Nagumo model is itself a simplified version of the famous Hodgkin—
Huxley model, which consists of four differential equations and models the activation
and deactivation dynamics of a spiking neuron in more detail.

89
able, provides linear negative feedback. 𝐼 represents the input. The main
phenomena in this model are shown in figure 4.15.

Figure 4.15: The dynamics of the FitzHugh—Nagumo model. The horizontal

axis represents the membrane potential (𝑉 ), which is the voltage
across the membrane of a neuron. The vertical axis represents the
recovery variable (𝑊 ). The 𝑉 -nullcline (where 𝑑𝑉 /𝑑𝑡 = 0) and
the 𝑊 -nullcline (where 𝑑𝑊 /𝑑𝑡 = 0) cross at the unstable resting
point. The lines with arrows represent typical trajectories. In the
depolarized state, the membrane potential is higher than at rest,
potentially leading to an action potential. In the hyperpolarized
state, the membrane potential is lower than at rest. In the self-
excitatory and regenerative phases, the system can increase its
own activity without external input. The active region refers to a
state in which the neuron is actively firing. Absolute and relative
refractory are periods following an action potential when a neuron
cannot fire again.

This model is for a single neuron. Crucial is that second equation is a slow
process. Time-scale effects also play an important role in learning in neural
networks. In most neural networks, there is a fast equation for updating
neuron activities and a much slower equation for updating the connection
strengths.
Other applications of the van der Pol model concern extensions of the Haken–
Kelso–Bunz (HKB) model (see section 5.4.4), multistable perception (Fürste-
nau 2014), developmental processes (Molenaar and Oppenheimer 1985), and
bipolar disorder (Daugherty et al. 2009). One case where it seems especially
useful is in modeling the wake—sleep cycle (Forger, Jewett, and Kronauer
1999).

90
4.3.7 Analogical modeling: Budworms and beers

If we create a dynamic model from the ground up, there’s a significant chance
that we might not completely grasp its intricacies. We have seen that some
very simple models already show amazingly complex behavior.
One approach to cope with these issues is analogical modeling, or basically
copying models. For instance, we used the Ising model to model attitudes
and the mutualistic Lotka—Volterra model to model intelligence. Both are
explained in Chapter 6. Here, I will use addiction as an example, focusing on
a selection of key phenomena (and for now ignoring many others).
We have reviewed existing formal models of addiction in van den Ende et al.
(2022). Most of these models are quite complicated. I want the model of the
individual addict to be as simple as possible. The reasons for this will become
clear in Chapter 7 when we include social effects (Boot et al. submitted for
publication). The key phenomena are that initiation, cessation, and relapse are
often discontinuous processes. The verbal theories we adapt are dual-process
models in which an automatic process of using more and more is controlled by
a non-automatic process, self-control.
Instead of creating our own model, we look for well-studied models in other
sciences, which led me to the spruce budworm model:

𝑑𝑁 𝑁 𝐵𝑁 2
= 𝑟𝑏 𝑁 (1 − ) − 2 . (4.16)
𝑑𝑡 𝐾 𝐴 + 𝑁2

But now we interpret the variables and parameters as follows. We first assume
𝑁 is the number of drinks you consume. The time scale is a day or an evening
(depending on when you have your first beer). 𝐾 is the upper limit of drinks
you can take, either because of lack of availability or, worse, because you just
collapse. 𝑟𝑏 is the addiction sensitivity. If this is too low (𝑟𝑏 < 0), the 0 state is
stable. The logistic function seems to be a reasonable choice. Drinking might
start off slow, then accelerate, and level off at 𝐾. This happens when there
is only an autonomous process. The second term, the predator term, is now
interpreted as self-control. This is not something that changes on the time
scale of a day, so, as in the case of the birds, a second equation is not required.
𝐴 (or actually 1/𝐴) is a responsiveness parameter, the number of drinks at
which self-control is activated, which may not be at the first or second beer.
𝐵 is the maximum level of self-control. As in the original model, this term is
a Holling type III form (see figure 4.5). We could also insert a Holling type
IV form, with the idea that self-control deteriorates after too many drinks.
Depending on the values of the parameters, one may not drink at all (𝑟𝑏 < 0),
drink at a recreational level, or have an “outbreak” to heavy use.
The advantage of this type of analogical modeling is that we already know
everything about the model. We know it is a cusp, and we have already made
the bifurcation diagram. There are also disadvantages or ambiguities.
First, the definition of 𝑁 is imprecise. Is it the blood alcohol concentration,
the number of drinks, or some other quantity?

91
Second, the choice of a logistic function for the autonomous part seems reason-
able but is not derived from first principles. One could also assume a linear
function with a ceiling at 𝐾.
Third, and relatedly, the self-control function is also not derived from first
principles. An additional problem is that we cannot measure this term directly
(Duckworth and Gross 2014).
Fourth, this model may not work for all addictions or should be adapted to
specific cases. An example is smoking. For smoking, the intermediate recre-
ational state is very unstable (Epskamp et al. 2022), and the autocatalytic
effect described by the logistic equation seems less appropriate. For alcohol,
the Holling type IV seems to be a good choice for the self-control term as
alcohol directly impairs brain regions involved in self-control (Remmerswaal
et al. 2019). For gambling, Holling Type III may be sufficient.
Fifth, processes at other time scales are missing. The model seems to work
well for the time scale of a day or an evening. Other relevant time scales are
minutes (direct effect of alcohol intake on the brain), weeks (abuse is often
concentrated on weekends), and months. On time scales of months or even
years, the parameters 𝑟𝑏 , 𝐾, and 𝐵 may change. For example, experienced
drinkers can drink more. Also, the 𝑟𝑏 , addiction sensitivity, may slowly in-
crease over time. This can be taken into account with additional equations.
Furthermore, 𝐾, 𝐴, and 𝐵 could change as an effect of the social network.
Nondrinkers might increase 𝐴, while other users in the social network might
increase 𝐾 (availability). These modeling issues are serious but also very in-
teresting (Dongen et al. 2024). Ambiguities in our thinking about
psychological systems come to light in
the process of building concrete
4.3.8 Cascading transitions in multifigure multistable perception mathematical models.

4.3.8.1 Interacting cusps

In section 4.3.6, we studied the van der Pol oscillator. In that model the
normal variable of the cusp was itself a dynamic variable 𝑑𝑎
𝑑𝑡 = −𝜀𝑋. Instead
of a linear equation, we could also use a cusp. We then get:

𝑑𝑋 3
= 𝑎𝑌 + 𝑏𝑋 − 𝑋 ,
𝑑𝑡 (4.17)
𝑑𝑌 3
= 𝑐𝑋 + 𝑑𝑌 − 𝑌 .
𝑑𝑡

This model, first proposed by Kadyrov, was analyzed in detail by Abraham et

al. (1991). We can study this model in Grind by specifying:

model <- function(t, state, parms) {

with(as.list(c(state, parms)), {
dX <- a * Y + b * X - X^3
dY <- c * X + d * Y - Y^3
return(list(c(dX, dY)))
})
}

92
Depending on the choice of the parameters and initial values, many different
things can happen. Abraham et al. (1991) created bifurcation diagrams to
summarize the qualitatively different regimes. We will restrict ourselves to the
case where 𝑏 = 𝑑 = 1, and 𝑎 and 𝑐 are varied. The bifurcation diagrams and
associated phase planes are shown in figure 4.16.

Figure 4.16: On the left the bifurcation diagram for the double cusp (𝑏 =
𝑑 = 1) is shown. The figures on the right show the phase planes
associated with the four different cases in the bifurcation diagram.
Case a has 9 fixed points, 4 of which are stable. Case b has 5
fixed points, 2 of which are stable. Case c has 3 fixed points, 2
of which are stable. Case d is special because it has a limit cycle,
the Kadyrov oscillator.

The phase planes of figure 4.16 can be made with:

layout(matrix(1:4, 2, 2))
s <- c(X = 0, Y = 0)
for(i in c('a','b','c','d'))
{
if (i == 'a') p <- c(a = .3, b = 1, c = .3, d = 1)
if (i == 'b') p <- c(a = .6, b = 1, c = .6, d = 1)
if (i == 'c') p <- c(a = 1, b = 1, c = 1, d = 1)

93
 if (i == 'd') p <- c(a = 1, b = 1, c = -1, d = 1)
plane(tstep = 0.5, portrait = (i == 'd'), xmin = -2, ymin = -2,
xmax = 2, ymax = 2, legend = FALSE, grid = 2,
main = paste("Case ", i)) # make a phase portrait (Fig 1c)
if (i != 'd') for(i in 1:200)
newton(c(X = runif(1, -2, 2),
Y = runif(1, -2, 2)), plot = TRUE)
else newton(c(X = 0, Y = 0), plot = TRUE)
}
s <- c(X = 0.1, Y = .1)
p <- c(a = 1, b = 1, c = -1, d = 1)
run(tmax = 20, tstep = 0.1, ymin = -2, ymax = 2) # Kadyrov oscillator

The last three lines of this code show the Kadyrov oscillator (figure 4.17).

Figure 4.17: The Kadyrov oscillator. Y attracts X to its state (as 𝑑 = 1), but
X pushes Y away (as 𝑐 = −1), resulting in oscillations.

If we simplify this analysis a bit to stable fixed points only, we see three
regimes:

• Case a (weak interactions): Each cusp has two stable states. The com-
bination of a negative and a positive state is possible because the inter-
action strength 𝑎 and 𝑐 are too weak.
• Case b and c (strong interactions): The combination of a negative and a
positive state is now impossible because the interaction strengths 𝑎 and
𝑐 are too strong. The equilibria 𝑋* and 𝑌 * are both positive or both
negative.
• Case d (opposite interactions): 𝑎 and 𝑐 have opposite signs, leading to
oscillations.

Abraham et al. (1991) generalize the model to:

𝑑𝑋 3
= 𝑎0 + 𝑎1 𝑌 + (𝑏0 + 𝑏1 𝑌 )𝑋 − 𝑋 ,
𝑑𝑡 (4.18)
𝑑𝑌 3
= 𝑐0 + 𝑐1 𝑋 + (𝑑0 + 𝑑1 𝑋) 𝑌 − 𝑌 .
𝑑𝑡

94
such that now both the splitting and normal variable of the cusp are linear
functions of the behavioral state of the other cusp. This can be further gener-
alized to a system of 𝑁 cusps by:

𝑑𝑋𝑖 3
= 𝑎0𝑖 + ∑ 𝑎𝑖𝑗 𝑋𝑗 + 𝑏0𝑖 𝑋𝑖 + ∑ 𝑏𝑖𝑗 𝑋𝑖 𝑋𝑗 − 𝑋𝑖 . (4.19)
𝑑𝑡 𝑗≠𝑖 𝑗≠𝑖

In this model, 𝑎0𝑖 is the intercept of the normal variable and the off-diagonal
elements of matrix 𝑎 are the slopes of the effect of the other cusps on the
normal variable. The diagonal elements of 𝑎 are set to 0. The 𝑏0𝑖 values
are the intercepts of the splitting variable. The 𝑏𝑖𝑗 values of matrix 𝑏 (with
diagonal = 0) are the slopes of the effect of other cusps on the splitting variable
value of 𝑋𝑖 .
The cascading transition model has been proposed independently in various
research areas. The idea of cuspoidal nets (𝑁 > 3) as a neural network has
been mentioned in Abraham (1991) and analyzed in Hoffmann et al. (1986)
and Izhikevich (1998). Castro and Timmis (2003) discuss this model in the
context of adaptive immune systems. The most recent application is in climate
research (Dekker, von der Heydt, and Dijkstra 2018; von der Heydt, Dekker,
and Dijkstra 2019; Klose et al. 2020). The applications involve special cases The idea of a cascade of collapsing
of equation 4.19, such as the case where one cusp influences the other, but not subsystems in the climate is a
frightening one.
vice versa. To my knowledge, the case where 𝑏𝑖𝑗 ≠ 0 has not been applied. A
recent related approach using coupled van der Pol oscillators is described in
Monsivais-Velazquez et al. (2020).
I will give a psychological example of this multivariate model, concerning
perception, in the next section.

4.3.8.2 Application to perception

Take a look at figure 4.18. This is a special case of multistable perception,

which I call multifigure multistable perception.

Figure 4.18: Multifigure multistable perception. Verify three phenomena: (1)

some attention or focus is required to see three-dimensional cubes,
(2) spontaneous transitions in the perception of cubes occur, (3)
such transitions affect the perception of neighboring cubes.

We can build a dynamical systems model of these perceptual phenomena by

using the cascading transition model setup. We define 𝑋 as the percept of
the cube. 𝑋 = 0 means that no cubes are perceived, only lines and colored
parallelograms. 𝑋 > 0 represents the front view, and 𝑋 < 0 the back view (cf.

95
 3
fig. 3.2). The cusp model for one cube is 𝑑𝑋
𝑑𝑡 = 𝑎 + 𝑏𝑋 − 𝑋 , where 𝑎 is the
bias parameter and 𝑏 is the attention parameter. If 𝑎 = 0 and 𝑏 > 1 (no bias
and some attention to the figure), we get bistable percepts and spontaneous
switches in perception (assuming we add some noise; see figure 4.2).
Now we apply equation 4.19. We have 𝑁 = 25 (a bit depending on how you
count). The values of the parameters 𝑎0𝑖 should be estimated from data, but
for now we will assume no bias, so 𝑎0𝑖 = 0. We set 𝑎𝑖𝑗 > 0, meaning that we
expect positive coupling between the cusps. We set 𝑏𝑖𝑗 > 0, based on the idea
that three-dimensional perception in one cube increases attention in the other
cubes. The 𝑏0 is the attention vector. In the simulation we first assume that
attention is low (𝑏0𝑖 = −0.3). After an initial phase, we will set 𝑏01 = 1, that
is we suddenly attend to one cube. A bit later we set 𝑏01 back to −0.3.
To make this model work, we need to make one adjustment. We replace
∑𝑗≠𝑖 𝑏𝑖𝑗 𝑋𝑖 𝑋𝑗 with ∑𝑗≠𝑖 𝑏𝑖𝑗 𝑋𝑖 |𝑋𝑗 |.10 This is because the increase in attention
by the three-dimensional perception of neighboring cubes does not depend on
whether we perceive the front or the back view. Thus, the model for multifigure
multistable perception is:

𝑑𝑋𝑖 3
= 𝑎0𝑖 + ∑ 𝑎𝑖𝑗 𝑋𝑗 + 𝑏0𝑖 𝑋𝑖 + ∑ 𝑏𝑖𝑗 𝑋𝑖 ∣𝑋𝑗 ∣ − 𝑋𝑖 . (4.20)
𝑑𝑡 𝑗≠𝑖 𝑗≠𝑖

The code to simulate this model is:

set.seed(1)
model <- function(t, state, parms){
with(as.list(c(state, parms)),{
X <- state[1:N]
b0_i <- parms[1:N]
dX <- -X^3 + a0_i + a_ij %*% X + b0_i*X +
(X * b_ij %*% abs(X)) # note abs(X)
return(list(dX))
})
}
N <- 10 # 10 necker cubes
X <- runif(N, -0.1, 0.1) # initial state of X
a0_i <- rep(0, N) # no bias in percepts
a_ij <- matrix(.02, N, N) # small couplings (normal)
diag(a_ij) <- 0 # set diagonal of a to 0
b0_i <- rep(-.3, N) # attention initially low
b_ij <- matrix(.2, N, N) # some spread of attention (splitting)
diag(b_ij) <- 0 # set diagonal of b to 0

s <- X; p <- c(b0_i) # required for grind

run(after = "if(t == 33) parms <- c(1, rep(-.3, N-1));
if(t == 66) parms <- rep(-.3, N);
state <- state + rnorm(N, mean = 0, sd = 0.05)",
ymin = -1, ymax = 2.5, main = '', ylab = 'X', legend = FALSE)
b0_i <- rep(-.3, 100); b0_i[34:66] = 1 # for plotting attention

10
We could also use 𝑋𝑗2 .

96
 lines(b0_i, lwd = 2, lty = 3)
text(80, 1.4, 'Percepts')
text(80, -.5, 'Attention')

which gives figure 4.19.

Figure 4.19: The multifigure Necker cube simulation. Initially, attention is low
and the percept is close to 0, representing the absence of three-
dimensional perception. At 𝑡 = 30, the attention intercept to one
cube is increased to 1. At 𝑡 = 60, it is set back to its initial low
value. However, this one cube is now perceived as a cube, and
the perception spreads to other cubes. They also increase overall
attention, so that the perception of cubes continues after 𝑡 = 60.

There is much more to be said about this model and its empirical validation.
One idea is to look at different stimuli like the one in figure 4.20.

Figure 4.20: Two embedded Necker cubes. The one on the left seems to have
a positive coupling 𝑎𝑖𝑗 > 0, while the one on the right seems to
switch independently (𝑎𝑖𝑗 = 0). You can verify this introspec-
tively. (Adapted from Adams and Haire 1959)

4.4 Causal-loop diagrams

One popular approach to dynamical systems modeling that I haven’t touched

on is the use of causal-loop diagrams, as developed in the field of systems

97
dynamics (Forrester 1993; Meadows 2008). As Crielaard et al. (2022) argue, Causal- loop diagrams are visual tools
the step from verbal theory to formal model may require an intermediate step used in systems thinking and system
dynamics to represent the feedback
of setting up a diagram that specifies the causal relationships between variables. loops and causal relationships within a
Related to causal-loop diagrams are several dedicated software packages for system.
system dynamics analysis.
Insightmaker (Fortmann-Roe 2014) is a simple free online tool that provides
a graphical model construction interface for dynamical systems modeling and
agent-based modeling. As such, it can be used to implement the models of this
and the previous chapter. A Lotka—Volterra example is shown in figure 4.21.
Insightmaker is easy to use. Studying some examples, found with “Explore
Insights,” may suffice. I have added some models to Insightmaker discussed
in this chapter with the tag “vdmaas.”

Figure 4.21: Screenshots of the Lotka—Volterra model in Insightmaker.

Insightmaker has many powerful built-in functions and allows sensitivity test-
ing as well as some sort of optimization. Personally, I prefer the approach of
writing the equations and implementing them in R for several reasons. One
is that this is how you communicate models in papers. Another is that the
equations help you think about analytical results, which are always prefer-
able to simulations. Finally, we can use Grind or deBif to go beyond simple
simulations and classify equilibria and perform bifurcation analysis. But for
building causal-loop diagrams of larger models to concretize theorizing without
the direct goal of running them, Insightmaker is a great tool.

4.5 Closing remarks

In this chapter I focused on the construction of dynamical system models

and introduced R tools to study them numerically. This introduction was
necessarily somewhat superficial. At the beginning of this chapter, I referred
to some texts that I recommend for further reading. The knowledge you
now have will allow you to study existing models from different fields and
to collaborate with experts in dynamical modeling. You now have the basic
language for communicating about such models.
But even when you work with experts in dynamic modeling, building useful
models is far from easy. I recommend following, at least roughly, the steps
we proposed in our theory construction methodology (Borsboom et al. 2021).
This methodology requires a good knowledge of existing verbal theories and, The key is to formulate phenomena,
if they exist, alternative formal models. I find the process of formalizing a replicated recurring patterns in data,
that need to be explained.
verbal theory or model fascinating. It tends to be very confusing. Suddenly

98
it is unclear what the basic assumptions are, what mechanism is really be-
ing proposed in some psychological theory, and what the time scales actually
are.
As an example, I mention the well-known investment theory of Cattell (1987).
Cattell’s investment theory posits that fluid intelligence, which represents the
ability to solve novel problems, “invests” in crystallized intelligence, which
consists of acquired knowledge and skills. I knew this theory for a long time
before I tried to translate it into dynamical equations. But it was not so
easy. I began to wonder why it was called an investment theory in the first
place. When you invest in something, it becomes less at first but more in
the future. Is that really what Cattell meant? The phenomena, the data
patterns, suggest something else, because fluid intelligence grows rapidly and
declines slowly after adolescence. Crystallized intelligence grows more slowly,
but never really declines. It is unclear where the return on investment is.
I would not argue that Cattell’s theory is nonsense, and a possible model is
proposed in the Chapter 6 (section 6.3.1.3), but this illustrates that the process
of formalization is itself a test for verbal theories.
There are some more psychological models that I could have included. For
example, the setup of dynamical field theory is a bit too complicated to repli-
cate in Grind, but I recommend studying this model (Schöner and Spencer
2016). In Chapter 6, I present a dynamical model of developmental processes
with mutualistic (positive) interactions (section 6.3.1.2) and in Chapter 7 I
introduce dynamical systems models of social interactions. I will discuss the
modeling of dynamical systems in psychology further in the Epilogue to this
book.

4.6 Exercises

1) Put the logistic equation into Grind, find out what the equilibria are,
and determine for which values of 𝑟 these are stable or unstable fixed
points. (*)
2) Check this analytically: Which are the two equilibria X*? For which
values of 𝑟 are these fixed points stable? Does your result agree with the
results of the previous exercise? (**)
3) Create the logistic map in Grind. Plot the time series for 𝑟 = 4. (*)
4) Make a plot of the pitchfork bifurcation, analogous to figure 4.3 (*)
5) Use the spruce budworm model from section 4.2.6 and the bifurcation()
function of the deBif package to recreate the bifurcation diagram shown
in figure 3.16. Describe what you did and present the resulting figure.
(**)
6) Implement the SIR model for infectious diseases in R using Grind. Re-
produce the diagram of the SIR model with 𝛽 = 0.4 and 𝛾 = 0.04 on the
Wikipedia page on “Compartmental Models in Epidemiology.”
7) Reproduce the times-series plot of the simulation of the Pólya urn model
shown in figure 4.12. (**)

99
8) Implement the FitzHugh—Nagumo model in Grind and replicate the
figure 4.15. Exact replication is not required, but the phase diagram
should look similar. (**)
9) Use Insightmaker to create a causal-loop diagram of the Romeo and
Juliet model. Reproduce the case where the couple ends up in a shrinking
cycle of hate and love (damping oscillator, second case of figure 4.9).
Submit the simulation plot. (**)

100
5 Self-organization

5.1 Introduction

We saw chaos and phase transitions in Chapters 2 and 3 and will now focus
on a third amazing property: self-organization. Self-organization plays an Self-organization is captivating
essential role in psychological and social processes. It operates in our neural because it reveals the remarkable
ability of complex systems to generate
system at the neuronal level, in perceptual processes as well in higher cognition. order and structure without external
In human interactions, self-organization is a key mechanism in cooperation and control or intervention.
opinion polarization.
Unlike chaos and phase transitions, self-organization lacks a generally accepted
definition. The definition most people agree on is that self-organization, or
spontaneous order, is a process in which global order emerges from local inter-
actions between parts of an initially disordered complex system. These local
interactions are often fast, while the global behavior takes place on a slower
time scale. Self-organization takes place in an open system, which means that
energy, such as heat or food, can be absorbed. Finally, some feedback be-
tween the global and local properties seems to be essential. Self-organization
occurs in many physical, chemical, biological, and human systems. Examples
of self-organization include the laser, turbulence in fluids, convection cells in
fluid dynamics, chemical oscillations, flocking, neural waves, and illegal drug
markets. For a systematic review of research on self-organizing systems, see
Kalantari, Nazemi, and Masoumi (2020). There are many great online videos.
I recommend “The Surprising Secret of Synchronization” as an introduction.
For a short history of self-organization research, I refer to the Wikipedia page
on “Self-Organization.” For an extended historical review, I refer to Krakauer
(2024).
This chapter also marks a transition from the study of systems with a small
number of variables to systems with many variables. We now focus on tools
and models for studying multi-element systems, such as agent-based mod-
eling and network theory. We will see complexity and self-organization in
action! This is not to say that the earlier chapters are not an essential part of
complex-systems research. The global behavior of complex systems can often
be described by a small number of variables that behave in a highly nonlin-
ear fashion. To study this global behavior, chaos, bifurcation, and dynamical
systems theory are indispensable tools.
The main goal of this chapter is to provide an understanding of self-
organization processes in different sciences, and in psychology in particular. I
will do this by providing examples from many different scientific fields. It is
important to be aware of these key examples, as they can inspire new lines of
research in psychology.
We will learn to simulate self-organizing processes in neural and social sys-
tems using agent-based models. To this end, we will use R and another tool,

101
NetLogo. NetLogo is an open-source programming language developed by
Uri Wilenski (2015). There are (advanced) alternatives, but as a general tool
NetLogo is very useful and fun to work with.
I start with an overview of self-organization processes in the natural sciences,
then I will introduce NetLogo and some examples. I will end with an overview
of the application of self-organization in different areas of psychology.

5.2 Key examples from the natural sciences

5.2.1 Physics

One physical example of self-organization is the laser. An important founder

of complex-systems theory is Hermann Haken (1977). He developed synerget-
ics, a specific approach to the study of self-organization and complexity in
systems that is also popular in psychology. Synergetics originated in Haken’s
work on lasers. We will not discuss lasers in detail here, but the phenomenon
is fascinating. Light from an ordinary lamp is irregular (unsynchronized). By
increasing the energy in a laser, a transition to powerful coherent light occurs.
In the field of synergetics, the order parameter is the term used to describe
the coherent laser light wave that emerges. The individual atoms within this An order parameter is a quantitative
system move in a manner consistent with this emergent property, which is, measure to describe the degree of
order within a system, especially in
unfortunately, called enslavement. Interestingly, the motion of these atoms the context of phase transitions.
contributes to the formation of the order parameter, that is, the laser light
wave. Conversely, the laser light wave dominates the movement of the individ-
ual atoms. This interaction exhibits a cyclical cause-and-effect relationship or
strong emergence (cf. fig. 1.2). Synergetics has been applied, as we will see, to
perception (Haken 1992) and coordinated human movement (Fuchs and Kelso
2018).
The Ising model (replaced by more
Another famous example, which will be very important for psychological mod- advanced models of magnetism in
eling later, is the Ising model of magnetism. In the standard 2D version of modern physics) has found
the model, atoms are locations on a two-dimensional grid. Atoms have left applications in many sciences.
(−1) or right (1) spins. When the spins are aligned (all 1 or all −1), we
have an effective magnet. If they are not aligned, the effect of the individual
spins is canceled out. Two variables control the behavior of the magnet: the
temperature of the magnet and the external magnetic field. The lower the
temperature, the more the spins align. The temperature at which the mag- At high temperatures, all the atoms
net loses its magnetic force is called the Curie point (see YouTube for some behave randomly, and the magnet
loses its magnetic effect.
fun demonstrations). The external field could be caused another magnet.
The main model equations of the Ising model are:

𝑛
𝐻 (x) = − ∑ 𝜏 𝑥𝑖 − ∑ 𝑥𝑖 𝑥𝑗 , (5.1)
𝑖 <𝑖,𝑗>

exp (−𝛽𝐻 (x))

𝑃 (X = x) = . (5.2)
𝑍

The first equation defines the energy of a given state vector x (for 𝑛 spins with
states —1 and 1). The notation < 𝑖, 𝑗 > in the summation means that we

102
Figure 5.1: Schematic picture of the magnet. Spins, 𝑥, can be left (−1) or
right (1). At lower temperatures, 𝑇 , the spins tend to align with
neighboring spins and the external field, 𝜏 , resulting in magnetism.

sum over all neighboring, or linked, pairs. Vectors and matrices are represented
using bold font.
With an external field we can force
The external field and temperature are 𝜏 and 𝑇 (1/𝛽), respectively. The first the spins to be all left or all right.
equation simply states that nodes congruent with the external field lower the
energy. Also, neighboring nodes with equal spins lower the energy. Suppose
we have only four connected positive spins (right column of figure 5.1) and no
external field, then we have x = (1, 1, 1, 1) and 𝐻 = −6. This is also the case
for x = (−1, −1, −1, −1), but any other state has a higher energy.
The second equation defines the probability of a certain state (e.g., all spins
1). This probability requires a normalization, 𝑍, to ensure that the proba-
bilities over all possible states sum up to 1. For large systems (𝑁 > 20),
the computation of 𝑍 is a substantive issue as the number of possible states
grows exponentially. If the temperature is very high, that is, 𝛽 is close to
0, exp (−𝛽𝐻 (x)) will be 1 for all possible states, and the spins will behave
randomly. The differences in energy between states do not matter anymore.
Entropy is a measure of the degree of
The randomness of the behavior is captured by the concept of entropy. To disorder or randomness in a system.
explain this a bit better, we need to distinguish the micro- and macrostate
of an Ising system. The Boltzmann entropy is a function of the number of
ways (𝑊 ) in which a particular macrostate can be realized. For ∑ 𝑥 = 4,
there is only one way (x = 1, 1, 1, 1). But for ∑ 𝑥 = 0, there are six ways
(𝑊 = 6). The Boltzmann entropies (ln 𝑊 ) for these two cases are 0 and 1.79,
respectively. The concept of entropy will be important in later discussions. The microstate is defined by the
configuration x of spins, while the
In the simulation of this model, we take a random spin and calculate the energy macrostate is determined by the sum
of the current x and the x with that particular spin flipped. The difference in of spins (similar to how magnetization
energy determines the probability of a flip: is defined).

1
𝑃 (𝑥𝑖 → −𝑥𝑖 ) = . (5.3)
1+ 𝑒−𝛽(𝐻(𝑥𝑖 )−𝐻(−𝑥𝑖 ))

If we do these flips repeatedly, we find equilibria of the model. This is called Glauber dynamics is a simulation
the Glauber dynamics (more efficient algorithms do exist). The beauty of technique that updates the spin states
in a system based on energy
these algorithms is that the normalization constant 𝑍 falls out of the equation. differences and temperature, guiding
In this way we can simulate Ising systems with 𝑁 much larger than 20. it toward equilibrium.

103
Interestingly, in the case of a fully connected Ising network (also called the
Curie—Weiss model), the emergent behavior—what is called the mean field
behavior—can be described by the cusp (Abe et al. 2017; Poston and Stewart
2014). The external field is the normal variable. Temperature acts as a The mean field behavior is the average
splitting variable. The relationship to self-organization is that when we cool magnetic field produced by all spins.
a hot magnet, at some threshold the spins begin to align and soon are all 1 or
−1. This is the pitchfork bifurcation, creating order out of disorder.1
In the 2D Ising model (see figure 5.1), the connections are sparse (only local),
and more complicated (self-organizing) behavior occurs. We will simulate this A fully connected Ising model behaves
in NetLogo later in this chapter, section 5.3.2.2, and as a model of attitudes according to the cusp. In less
connected networks of Ising spins,
in Chapter 6, section 6.3.3. self-organizing patterns can emerge.

5.2.2 Chemistry

Other founders of self-organizing systems research are Ilya Prigogine and Is-
abelle Stengers. Prigogine won the 1977 Nobel Prize in chemistry for his work
on self-organization in dissipative systems. These are systems far from thermo-
dynamic equilibrium (due to high energy input) in which complex, sometimes
chaotic, structures form due to long-range correlations between interacting par-
ticles. One notable example of such behavior is the Belousov—Zhabotinsky
reaction, an intriguing nonlinear chemical oscillator.
Stengers and Prigogine authored the influential book Order Out of Chaos in
(1978). This work significantly influenced the scientific community, particu-
larly through their formulation of the second law of thermodynamics. One The second law of thermodynamics
way of stating the second law is that heat flows spontaneously from hot objects states that the total entropy of an
isolated system always increases over
to cold objects, and not the other way around, unless external work is applied time and never decreases, meaning
to the system. A more appealing example might be the student room that that spontaneous processes in nature
never naturally becomes clean and tidy, but rather the opposite. tend to move toward a state of
increasing disorder or randomness.
Stengers and Prigogine (1978) argued that while entropy may indeed decrease
in a closed system, the process of self-organization in such systems can cre-
ate ordered structures that compensate for the entropy increase, resulting in
a net increase in what they called “local entropy.” Prigogine and Stengers
placed particular emphasis on irreversible transitions, highlighting their im-
portance in understanding complex systems. While the catastrophe models
we previously discussed exhibited symmetrical transitions (sudden jumps in
the business card are symmetric), Prigogine’s research revealed that this sym-
metry does not always hold true. Irreversible transitions refer to
changes in a system that cannot be
To illustrate this point, consider the analogy of frying an egg. The process of reversed by simply reversing the
transforming raw eggs into a fried form represents a phase transition, but it conditions that caused the change,
is impossible to reverse this change and unfry the egg. Prigogine linked these often resulting in a permanent change
in the state or structure of the system.
irreversible transitions to a profound question regarding the direction of time,
commonly known as the arrow of time. Although it is a fascinating topic in
itself, we will not explore it further here.

1
An extremely useful application of this principle is the rice cooker!

104
5.2.3 Biology

There is no shortage of founders of complex-systems science. Another fan-

tastic book is Stuart Kaufmann’s Origin of Order (1993), which introduces
the concept of self-organization into evolutionary theory. He argues that the
small incremental steps in neo-Darwinistic processes cannot fully explain nat-
ural evolution. If you want to know about adaptive walks and niche hopping
in rugged fitness landscapes, you need to read his book (Kauffman 1993). An-
other influential theory is that of punctuated equilibria, which proposes that
species undergo long periods of stability interrupted by relatively short bursts
of rapid evolutionary change (Eldredge and Gould 1972).
A neat example of the role of self-organization in evolution is the work on spiral
wave structures in prebiotic evolution by Boerlijst and Hogeweg (1991). This
work builds on Eigen and Schuster’s (1979) classic work on the information
threshold. Evolution requires the copying of long molecules. But in a system of
self-replicating molecules, the length of the molecules is limited by the accuracy
of replication, which is related to the mutation rate. Eigen and Schuster
showed that this threshold can be overcome if such molecules are organized in
a hypercycle in which each molecule catalyzes its nearest neighbor. However, A hypercycle is a network of
the hypercycle was shown to be vulnerable to parasites. These are molecules self-replicating molecules or entities
that mutually support each other’s
that benefit from one neighbor but do not help another. This molecule will production, leading to an increase in
outcompete the others, and we are back to the limited one-molecule system. complexity and stability beyond what
individual entities could achieve alone.
What Boerlijst and Hogeweg did was to implement the hypercycle in a cellular
automaton. In the hypercycle simulation, cells could be empty (dead) or filled
with one of several colors. Colors die with some probability but are also copied
to empty cells with a probability that depends on whether there is a catalyzing
color in the local neighborhood. One of the colors is a parasite, catalyzed by
one color but not catalyzing any other colors. The effect, which you will see
later using NetLogo, is that rotating global spirals emerge that isolate the
parasites so that a stable hypercycle prevails. A cellular automaton (CA) is usually
a two-dimensional grid of cells, where
Many examples of self-organization come from ecosystem biology. We will cells interact with their neighbors, as
see a simulation of flocking in NetLogo later, but I also want to highlight the in the 2D Ising model, but this can be
collective behavior of ants (figure 5.2). generalized to more or less dimensions.

Ants exhibit amazing forms of globally organized behavior. They build bridges,
nests, and rafts, and they fight off predators. They even relocate nests. Ant
colonies use pheromones and swarm intelligence to relocate. Scouts search
for potential sites, leaving pheromone trails. If a promising location is found,
more ants follow the trail, reinforcing the signal. Unsuitable sites result in
fading trails. Once a decision is made, the colony collectively moves to the
chosen site, transporting their brood and establishing a new nest.
It is not a strange idea to think of an ant society as a living organism. Note
that all this behavior is self-organized. There is clearly no super ant that has
a blueprint for building bridges and telling the rest of the ants to do certain
things. Ants also don’t hold lengthy management meetings to organize. The
same is true of flocks of birds. There is no bird that chirps commands to move
collectively to the left, to the right, or to split up. This is true of human brains.
An individual neuron is not intelligent. Our intelligence is based on the
collective behavior of billions of
neurons.

105
 Figure 5.2: The ant bridge is an example of collective behavior.

5.2.4 Computer science

Another important source of self-organization research is computer science.

A simple but utterly amazing example is the work on John Conways’ Game
of Life (Berlekamp, Conway, and Guy 2004). The rules are depicted in fig-
ure 5.3.

Figure 5.3: The rules of the Game of Life.

For each cell, given the states of its neighbors, the next state for all cells is
computed. This is called synchronous updating.2 It is hard to predict what

2
In a synchronous update, all cells of the cellular automata update their state simultane-
ously. This implies that the new state of each cell at a given time step depends only
on the states of its neighbors at the previous time step. In asynchronous update, cells
update their state one at a time, rather than all at once. The order in which cells update

106
will happen if we start from a random initial state. But you can easily verify
that a block of four squares is stable, and a line of three blocks will oscillate
between a horizontal and a vertical line.
A great tool for playing around with the Game of Life is Golly, a freely avail-
able application for computers and mobile phones. I ask you to download and
open Golly, draw some random lines, press Enter, and see what happens. Of-
ten you will see it converging to a stable state (with oscillating subpatterns).
Occasionally you will see walkers or gliders (zoom out). These are patterns
that move around the field.
Random initial patterns rarely lead to anything remarkable, but by choosing
special initial states, surprising results can be achieved. First, take a look at
the Life within Patterns folder. Take, for example, the line-puffer superstable
or one of the spaceship types. My favorite is the metapixel galaxy in the
HashLife folder. Note that you can use the + and — buttons to speed up and
slow down the simulation. What this does is simulate the game of life in the
game of life! Zoom in and out to see what really happens. I’ve seen this many
times, and I’m still baffled. A childish but fun experiment is to disturb the The Turing machine is a theoretical
metapixel galaxy in a few cells. This leads to a big disturbance and a collapse machine developed by Alan Turing in
1936, that despite its simplicity can
of the pattern. implement any computer algorithm,
I was even more stunned to see that it is possible to create the (universal) including, of course, the Game of Life!
Turing machine in the Game of Life (Rendell 2016). The Game of Life im-
plementation of the Turing machine is shown in figure 5.4. This raises the
question of whether we can build self-organizing intelligent systems using ele-
mentary interactions between such simple elements. Actually, we can to some
extent, but by using a different setup, based on brain-like mechanisms (see
the next section on neural networks).
Another root of complex-systems theory and the role of self-organization in
computational systems is cybernetics (Ashby 1956; Wiener 2019). To give Cybernetics studies circular causal
you an idea of this highly original work, I will only mention the titles of a few and feedback mechanisms in complex
systems, focusing on how systems
chapters of Norman Wiener’s book, originally published in 1948: “Gestalt regulate themselves, process
and Universals,” “Cybernetics and Psychopathology,” “On Learning and information, and adapt to changes in
Self-Reproducing Machines,” and, finally, “Brainwaves and Self-Organization.” their environment.
And this was written in 1948!
The interest in self-organization is not only theoretical. In optimization, the
search for the best parameters of a model describing some data, techniques in-
spired by cellular automata and self-organization have been applied (Langton
1990; Xue and Shen 2020). I have always been fascinated with genetic algo-
rithms (Holland 1992a; Mitchell 1998), where the solutions to a problem (sets
of parameter values) are individuals in an evolving population. Through mu- Genetic algorithms are a class of
tation and crossover, better individuals evolve. This is a slow but very robust optimization algorithms inspired by
the process of natural selection, where
way of optimizing, preventing convergence to local minima. solutions to a problem evolve over
John Henry Holland is considered one of the founding fathers of the complex- generations.
systems approach in the United States. He has written a number of influential
books on complex systems. His most famous book, Adaptation in Natural
and Artificial Systems: An Introductory Analysis with Applications to Biology,

can be deterministic (in a sequence) or stochastic (random). These two different update
schemes can lead to very different behaviors in cellular automata.

107
Figure 5.4: The Turing machine built in the Game of Life. (Reproduced from
LifeWiki.)

108
Control Theory, and Artificial Intelligence (Holland 1992b), has been cited
more than 75,000 times.
A self-organizing algorithm that has played a large role in my applied work is
the Elo rating system developed for chess competitions (Elo 1978). Based on The Elo rating system is a
the outcomes of games, ratings of chess players are estimated, which in turn are self-organizing method of calculating
the relative skill levels of players in
used to match players in future games. Ratings converge over time, but adjust head-to-head games based on the
as players’ skills change. We have adapted this system for use in online learning results of their games.
systems where children play against math and language exercises (Maris and
van der Maas 2012). The ratings of children and exercises are estimated on
the fly in a large-scale educational system (Klinkenberg, Straatemeier, and
van der Maas 2011). We build this system to collect high frequency learning
data to test our hypotheses on sudden transitions in developmental processes,
but it was more successful as an online adaptive practice system. We collected
billions of item responses with this system (Brinkhuis et al. 2018).

5.2.5 Neural networks

The current revolution in AI, which is having a huge impact on our daily lives,
is due to a number of self-organizing computational techniques. Undoubtedly,
deep learning neural networks have played the largest role. A serious overview
of the field of neural networks is clearly beyond the scope of this book, but one
cannot understand the role of complex systems in psychology without knowing
at least the basics of artificial neural networks (ANNs), that is, networks
of artificial neurons. ANNs consist of interconnected nodes, or “neurons,” Artificial neural networks are
organized into layers that process information by propagating signals through computational models inspired by the
structure and function of biological
the network. ANNs are trained on data to learn patterns and relationships, neural networks.
enabling them to perform tasks such as classification, regression, and pattern
recognition.
Artificial neurons are characterized by their response to input from other neu-
rons in the network, which is typically weighted and summed before being
passed through an activation function. This activation function may produce
either a binary output or a continuous value that reflects the level of activa-
tion of the neuron. The input could be images, for example, and the output
could be a classification of these images. The important thing is that neural
networks learn from examples.
Unsupervised learning is based on the structure of the input. A famous
unsupervised learning rule is the Hebb rule (Hebb 1949), which states that
what fires together wires together. Thus, neurons that correlate in activity
strengthen their connection (and otherwise connections decay). In supervised
learning, connections are updated based on the mismatch between model out-
put and intended output through backpropagation. Hebbian learning and Backpropagation is a mechanism to
backpropagation are just two of the learning mechanisms used in modern update specific connections such that
this mismatch or error is minimized
ANNs. over time.
Modern large language models, like GPT, differ from traditional backpropa-
gation networks in terms of their architecture, training objective, pre-training
process, scale, and application. Large language models use transformer archi-
tectures, undergo unsupervised pre-training followed by supervised fine-tuning,
are trained on massive amounts of unlabeled data, are much larger in size, and
are primarily used for natural language-processing tasks.

109
Another important distinction is between feedforward and recurrent neural
networks. An interesting recurrent unsupervised model is the Boltzmann Feedforward neural networks process
machine. It is basically an Ising model (see section 5.2.1) where the connec- information in a single forward pass,
while recurrent neural networks have
tions between nodes have continuous values. These connections or weights directed cycles, allowing them to
can be updated according to the Hebb rule. A simple setup of the Boltzmann capture temporal dependencies.
machine is to take a network of connected artificial neurons and present the
inputs to be learned in some sequence by setting the states of these neurons
equal to the input. The Hebb rule should change the weights between neurons
so that the Boltzmann machine builds a memory for these input states. This The Hebb rule states neurons that fire
is the training phase. In the test phase, we present partial states by setting together wire together.
some, but not all, nodes to the values of a particular learned input pattern.
By the Glauber dynamics, we update the remaining states that should take on
the values belonging to the pattern. This pattern completion task is typical
for ANNs.
This setup is called the general or unrestricted Boltzmann machine, where
any node can be connected to any other node and each node is an input node.
The restricted Boltzmann machine (RBM) is much more popular because of
its computational efficiency. In an RBM, nodes are organized in layers, with
connections between layers but not within layers. In a deep RBM, we stack
many of these layers, which can be trained in pairs (figure 5.5).3 Other promi-
nent approaches are the Kohonen self-organizing maps and the Hopfield neural
network.

Figure 5.5: The deep learning restricted Boltzmann machine.

The waves of popularity of neural networks are closely related to the devel-
opment of supervised learning algorithms, where the connections between ar-
tificial neurons are updated based on the difference between the output and
the desired or expected output of the network. The first supervised ANN, the
perceptron, consisted of multiple input nodes and one output node and was
able to classify input patterns from linearly separable classes. This included
the OR and AND relation but excluded the XOR relation. In the XOR, In the XOR pattern, the combinations
the sum of the two bits is not useful for classification. By adding a hidden of 00 and 11 are false, 01 and 10 are
true.
3
I recommend Timo Matzen’s R package for a hands-on explanation (https://github.com
/TimoMatzen/RBM).

110
layer to the perceptron, the XOR can be solved, but it took many years to
develop a backpropagation rule for multilayer networks such that they can
learn this nonlinear classification from examples. We will do a simple simula-
tion in NetLogo later. Although they are extremely powerful, it is debatable
whether backprop networks are self-organizing systems. Self-organizing sys-
tems are characterized by their ability to adapt to their environment without
explicit instructions. Unsupervised neural networks are more interesting in
this respect.
All these models were known at the end of the twentieth century, but their
usefulness was limited. This has changed due to some improvements in algo-
rithms but especially in hardware. Current deep-learning ANNs consist of tens
of layers within billions of nodes, trained on billions of inputs using dedicated
parallel processors (e.g., Schmidhuber 2015).
Neural networks are at the heart of the AI revolution, but other developments,
especially reinforcement learning, have also played a key role. Examples are Reinforcement learning is essential in
game engines, robots, and self-driving cars. Note that the study of reinforce- AI systems that need to behave or act
on the environment.
ment learning also has its roots in psychology (see Chapter 1 of Sutton and
Barto 2018).
I was most amazed by the construction and performance of AlphaZero chess.
AlphaZero chess (Silver et al. 2018) combines a deep learning neural network
that evaluates positions and predicts next moves with a variant of reinforce-
ment learning (Monte Carlo tree search). Amazingly, AlphaZero learns chess
over millions of self-played games. This approach is a radical departure from
classic chess programs, where brute-force search and built-in indexes of open-
ings and endgames were the key to success. As it learns, it shows a phase AlphaZero chess is a self-organizing
transition in learning after about 64,000 training steps (see fig.7 in McGrath program that learns chess from
scratch by playing against itself.
et al. 2022). For an analysis of the interrelations between psychology and
modern AI, I refer to van der Maas, Snoek, and Stevenson (2021).
AlphaZero’s use of Monte Carlo tree search is also a form of symbolic artificial
intelligence. The idea of combining classic symbolic approaches with neural
networks has always been in the air. The third wave of this hybrid approach
is reviewed in Garcez and Lamb (2023).

5.2.6 The concept of self-organization

I trust that you now possess some understanding of self-organization and its
applications across various scientific fields. Self-organization is a generally
applicable concept that transcends various disciplines, yet it maintains strong
connections with specific examples within each discipline.
As previously mentioned, the precise definition of self-organization remains un-
der discussion, and a range of criteria continue to be debated. Key questions,
such as the degree of order necessary for a system to be deemed self-organized,
whether any external influences are permissible, whether a degree of random-
ness within the system is acceptable, and whether the emergent state must be
irreversible, are among the issues that lack definitive resolutions.
This ambiguity in the definition isn’t unusual for psychologists, as many non-
formal concepts lack strict definitions. The value of the self-organization con-
cept is primarily found in its concrete examples, its broad applicability, such as

111
in the field of artificial intelligence, and our capability to create simulations of
it. The focus of the next section will be on such simulations using a dedicated
tool, NetLogo.

5.3 NetLogo

5.3.1 Examples

NetLogo (Wilensky and Rand 2015) is based on Logo, a revolutionary educa-

tional programming language from the early days of computer languages, in
which an on-screen turtle, a cursor, could be moved around to create graph-
ics.4 The turtle is still there, but there is much more that you can do with
NetLogo.
I strongly recommend that you download and install NetLogo for the next
part of this chapter.
The Ising model
When you start NetLogo, you see an interface with a black area (the world), a
33-by-33 matrix of patches (cells). You can change the world using the settings
(see top right). Interface and Code are the most important tabs.
First, open the Model Library (menu File: Model Library) and find and open
“Ising.” Click on setup and go. That is all. Verify that high temperature indeed
causes random spin behavior. Also verify that lowering the temperature causes
a pitchfork bifurcation. The random state becomes unstable and all spins
become either positive or negative (light or dark blue). Now go to Settings and
set max-pxcor and max-pycor to 200 and Patch size to 1. With these settings you
will see self-organized global patterns, constantly moving clusters of positive
and negative spins.
Hypercycles
Some models are available in NetLogo; others can be found on the NetLogo’s
website (see Community). Download “Hypercycle” by Maarten Boerlijst and
read the information. You have to run the model with eight species for 20,000
iterations or ticks (to speed up, deselect view updates) and then add parasites.
The spirals keep the growth of the parasites under control. If you do this
earlier, the parasites will quickly take over. I think this is a beautiful example
of functional self-organization. The implementation in the form of a cellular
automata is essential for the success of this model. If we implement this model
in the form of coupled differential equations, the parasite will simply win.
Flocking
NetLogo 3D allows us to create three-dimensional plots of self-organizing pat-
terns. Start NetLogo 3D and load the flocking model “3D Alternate”. I rec-
ommend editing the Population slider by right-clicking it and setting the max to
1,000. This will result in more realistic swarms. Play around with the controls
and don’t kill all the birds.

4
A widely recognized implementation of this educational strategy is Scratch, which is used
by many schools around the world to teach children to program.

112
Traffic
In the Models Library of NetLogo (not 3D) you will find “Traffic 2 Lanes.”
Run the model with 20 cars and notice that the congestion actually moves
backward. Play around with the number of cars as well. Is there a clear
threshold where you get congestion as you slowly increase the number of cars?
And what happens when you decrease the number of cars? Is there a threshold
where congestion dissipates? I hope you see that finding hysteresis in this way
is quite difficult. There are clearly sudden changes, but finding hysteresis
requires very precise and patient experimentation.
Neural networks
In the Model Library you will find a “Perceptron” and a “Multilayer network.”
Start with the perceptron. Set the target function to and, train the model
for a few seconds, and test the perceptron. You will see that it correctly
classifies 11 as 1 and the other patterns as —1. The graph on the bottom
right is particularly instructive. It shows how the patterns are separated. The
perceptron can do linear separation. This is sufficient for most of the logical
rules that can be learned, but not for the XOR (see section 5.2.5). You will
see that the linear separation just jumps around and the XOR cannot be
learned. Also train the multilayer model on the XOR. Another nice tool to
play around with can be found on the internet by searching for “a neural
network playground.”
Of course, these are just illustrative tools. But building serious deep learning
ANNs is not that hard either. Many resources and books are available (e.g.,
Ghatak 2019).
The Sandpile model
Bak, Tang, and Wiesenfeld (1988) introduced the concept of self-organized
criticality. In systems such as the Ising model, there are parameters (e.g.,
temperature) that must be precisely tuned for the system to reach criticality.
The Bak—Tang—Wiesenfeld sandpile model exhibits critical phenomena with-
out any parameters. In the sandpile model, grains of sand are added to the Self-organized criticality (SOC) refers
center of the pile. When the difference in height between the center column to complex systems naturally evolving
into a critical state where small
and its neighbors exceeds a critical value, a grain of sand rolls to that neigh- changes can lead to transitions
boring location. This occasionally results in avalanches. The point is that without the need for specific
no matter how we start, we get to a critical state where these avalanches oc- parameter settings.
cur. Thus, the sandpile model spontaneously evolves toward its critical point,
which is why this phenomenon has been called self-organized criticality.
The NetLogo model “Sandpile” in the Models Library demonstrates this be-
havior (use setup uniform, center drop location, and animate avalanches). We
now drop grains of sand onto the center of a table, one at a time, creating
avalanches. The plots on the right show an important characteristic of self-
organized criticality. The frequencies of avalanche sizes and durations follow
a power law. The power-law relationship is often mathematically expressed as
𝑌 = 𝑎𝑋 𝑘 , where 𝑌 and 𝑋 are the quantities of interest, 𝑎 is a constant coeffi-
cient, and 𝑘 is the exponent of the power law. Power laws are notable for their
scale-invariant property, which means that the form of the relationship does
not change across different scales of 𝑋 and 𝑌 . This means that the log-log plot
should be linear, which can be verified by running the model for some time.
One of the key features of power-law distributions is that they exhibit a high

113
degree of variability or heterogeneity. This means that there are many small
events or phenomena and a few very large ones, with a smooth distribution
of sizes in between. Power-law systems are scale invariant, meaning that we
see the same behavior at any scale of the sandpile. For this reason, they are
sometimes called scale-free distributed.
Other models
I recommend running a few other models (e.g., “Sunflowers”, “Beatbox”, and
the “B—Z reaction”). One thing we haven’t done yet is click on the Code tab.
Read the code for the B—Z reaction and notice one thing: it is surprisingly
short!

5.3.2 A bit of NetLogo programming

I find NetLogo programming very easy and very hard at the same time. Hard
because it requires a different way of thinking. Uri Wilensky’s examples are
often extremely elegant and much shorter than my clumsy code. NetLogo
resembles object-oriented programming languages, quite different from (base)
R. There are three types of objects: the patches, which refer to cells in a
world grid (CA); turtles, which are agents that move around; and links, which
connect turtles. Note that turtles are not necessarily turtles. We have already
seen turtles in the form of neural nodes and cars.
In NetLogo, you “ask” objects to do something. A typical line would be:
ask turtles with [color = red ] set color green
This would make red turtles green. To get started, I highly recommend watch-
ing the videos on the NetLogo page “The Beginner’s Guide to NetLogo Pro-
gramming” and following these examples. Here we make our own Game of
Life.

5.3.2.1 Game of Life

First, create two buttons in the interface: setup and a go. In Command, name
them “setup” and “go.” In the settings of the go button, select forever. Now
go to the Code tab and define these two functions as:
to setup clear-all reset-clicks end
to go tick end
Tickscount the iterations in NetLogo, and with this code we are just resetting
things. In this example, we will use the patches instead of the turtles. Patches
are the grid cells or squares that make up the “world” in a NetLogo model.
Now add this last line to setup (with the sem-colon we can add comments to
code):
ask patches [set pcolor one-of [ white blue ]] ; white is dead, blue is alive
To do a synchronous update, we need to store the updated state in a temporary
variable called new-state. Put this line at the top of your code:
patches-own [new-state]

114
In the go function, we add the life rules.
ask patches [ if ( neighbors with [pcolor = blue ]) > 3 ) [set new-state
white ] if ( neighbors with [pcolor = blue ]) < 2 ) [set new-state white ]
if ( neighbors with [pcolor = blue ]) = 3 ) [set new-state blue ] ]
ask patches [ set pcolor new-state ]
The last line updates the state to the new-state. That is all! We built a Game of
Life simulation. Use setting to create a larger world. Take a look at the code of
the Game of Life program in the Model Library to see some extensions to this
code. In the Help menu, you will find the very useful NetLogo dictionary. Just
reading through this dictionary will teach you a lot of useful tricks. NetLogo is
similar to R in that you should use the built-in functions as much as possible.

5.3.2.2 The Ising model

Building a NetLogo model from scratch requires quite some experience; adapt-
ing a program is much easier. The Ising model in NetLogo is not complete,
as there is no slider for the external field. Try to add this yourself. Add a
slider for the external field tau. The code only needs to be changed in this line
(study equation 5.1):
let Ediff 2 * spin * sum [ spin ] of neighbors4
If successful, you can test for hysteresis and divergence. For tau = 0, decreasing
the temperature should give the pitchfork bifurcation. For a positive temper-
ature (say 1.5), moving tau up and down should give hysteresis.
Actually, this should work better if all spins are connected to all spins. To do
this, replace neighbors4 with patches. To normalize the effect of so many spins,
it is recommended to use:
let 0.001 * Ediff 2 * spin * sum [ spin ] of patches
Now you should see hysteresis and the pitchfork better. However, in this case
the typical self-organized patterning that occurred in the Ising model with
only local interactions is not present (see last part of section 5.2.1).

5.4 Self-organization in psychology and social systems

In this second part of the chapter, I provide illustrations of research on self-

organization within various psychological systems, spanning several subfields
of psychology. I begin with an exploration of self-organization in the context of
the brain and conclude with an examination of its implications within human
organizations. I will point to relevant literature to guide further exploration
in other areas.

115
5.4.1 The brain

Many psychological and social processes involve self-organization. As discussed

above, at the lowest level self-organization plays a role in neural systems.
Self-organization in the brain is an active area of research (Breakspear 2017;
Chialvo 2010; Cocchi et al. 2017; Ooyen and Butz-Ostendorf 2017; Plenz et
al. 2021). Dresp-Langley (2020) distinguished seven key properties of self-
organization clearly identified in brain systems: modular connectivity, unsu-
pervised learning, adaptive ability, functional resiliency, functional plasticity,
from-local-to-global functional organization, and dynamic system growth.
A key example is Walter Freeman’s work on the representation of odors in the
brain (Skarda and Freeman 1987). He used EEG measurements to support
his nonlinear system model of the brain. Freeman proposed that the brain
operates by generating of dynamic patterns of electrical activity, which he
called attractors. In Freeman’s theory, attractors
represent stable states of neural
Another influential theory was proposed by neuroscientist Gerald Edelman. activity that arise spontaneously from
His theory of neural Darwinism suggests that the development of the brain’s the interactions between large
neural connections is based on a process of competition and selection, rather populations of neurons.
than being pre-wired in the genes (Edelman 1987). According to Edelman’s
theory, the brain is a complex, dynamic system made up of many intercon-
nected neurons that constantly interact with each other and the outside world.
The process of competition and selection occurs through the formation of
ensembles of neurons that respond to specific stimuli or experiences. An In neural Darwinism, the connections
alternative approach was put forward by Carpenter and Grossberg (1987). between neurons in successful
ensembles become stronger over time,
Grossberg and Carpenter’s theory focuses on how neural networks in the brain while those in unsuccessful ensembles
self-organize to process information and adapt to changing environments. It weaken or disappear.
explores the principles governing neural dynamics, leading to the emergence of
coherent cognitive and behavioral patterns through interaction and learning
within neural systems.
It has also been claimed that self-organized criticality (SOC) (see the Sandpile
model in section 5.3.1) plays a role in the brain (Bak, Tang, and Wiesenfeld
1988). It is hypothesized that when a system is close to criticality, small per-
turbations can have large, cascading effects, which can allow the system to
rapidly switch between different states of activity in response to changes in
the environment. One of the key pieces of evidence for SOC in the brain comes
from studies of the distribution of sizes of neural activity events, which has
been found to follow a power law distribution, but alternative explanations
have been provided (Bédard, Kröger, and Destexhe 2006). This is a techni-
cal area of research with many methodological challenges (Lurie et al. 2020;
O’Byrne and Jerbi 2022).
A promising general approach to understanding the so-called “predictive”
brain functions is the free-energy account (A. Clark 2013), which implements
a form of self-organization (Friston 2009). The brain is not simply reacting The predictive brain is constantly
to the world around us but is actively generating predictions about what making predictions about the sensory
inputs it receives from the
we will see, hear, feel, and experience, based on our past experiences and environment, minimizing the
knowledge. The predictive brain theory suggests that the brain’s predictions discrepancy between expected and
are generated through a process of hierarchical inference, in which information actual inputs.
from lower-level sensory areas is combined and integrated in higher-level areas
to generate more complex predictions about the world. These predictions are

116
then compared to the incoming sensory inputs, and any discrepancies between
the predictions and the actual inputs are used to update the predictions and
improve the brain’s accuracy over time.

5.4.2 Consciousness

Many will agree on the idea that higher psychological functions or properties
such as thinking, perceiving, remembering, and reasoning, but also personality
and emotions (i.e., the mind), emerge out of lower-order brain activities. Of
special interest is consciousness. Seth and Bayne (2022) list 22 different the-
ories that link consciousness to neurobiology. Well-known examples are the
global workspace theory, the integrated information theory, and higher-order
theory. Self-organization plays a role in most of these theories.
The central idea of global workspace theory is that there is a central workspace
in the brain, a kind of mental stage where information from various sensory
inputs and memory systems is gathered, processed, and integrated. The Information that enters the global
workspace is not tied to a specific brain region but is thought to emerge from workspace becomes available for
widespread distribution throughout
the dynamic interactions of widespread neural circuits. the brain, allowing for coordinated,
The core proposition of integrated information theory (ITT) is that conscious- conscious processing.
ness is equivalent to a system’s ability to integrate information. According
to IIT, the level of consciousness a system possesses can be quantitatively
measured by a value called 𝜙, which represents the amount of integrated in-
formation the system can generate. A higher 𝜙 indicates a higher level of
consciousness. According to integrated information theory, for a system to
be conscious, it must be able to combine diverse pieces of information into a
single coherent whole.
For higher-order theories of consciousness, meta-representations are critical.
One might have a representation of a particular perception, such as a flower, Higher-order theories of consciousness
and additionally have a meta-representation that acknowledges “I am perceiv- suggest that consciousness arises when
the brain represents its own processes
ing a flower.” I find higher-order theories most compelling because they make to itself.
a clear distinction between unconscious and conscious information processing.
A recent and interesting variant is the self-organizing meta-representational ac-
count of Cleeremans et al. (2020), as it states that consciousness is something
the brain learns to do.
My thinking about consciousness has been strongly influenced by the work of
Douglas Hofstadter, especially his book on Gödel, Escher, Bach (Hofstadter
1979). In his work, our sense of self is a construct formed by the brain’s
ability to use symbols, such as natural language, to refer to its own activities
and experiences. Consciousness is based on symbolic self-reference, thus meta-
representations. I think, with Hofstadter (2007), that this higher-order self In Hofstadter’s theory, consciousness
has the ability to influence the lower-order processing of the brain, a case of arises when these self-referential loops
(strange loops) reach a certain level of
downward causation (section 1.2).5 For a somewhat critical analysis, I refer complexity.
5
I wrote a Gödel, Escher, Bach-like dialogue on consciousness (van der Maas 2022) in which
my laptop professes to have free will yet simultaneously denies that I possess free will. I
asked ChatGPT-4 what it thought of it. Nice as always, ChatGPT replies: “The dialogue
is a creative and thought-provoking exploration of various philosophical and theoretical
concepts related to AI, consciousness, and free will.” But it also disagrees: “AI, as it exists
today, does not possess consciousness, self-awareness, or free will, and its ‘understanding’
is limited to processing data within the parameters of its programming.” I also asked
ChatGPT 4.0 whether it has a self-concept. It denied it, and then I asked whether that in

117
to Nenu (2022).
Zooming out, having twenty-two theories of consciousness, and this is an un-
derestimate, is a bit much. The lack of empirical data constraints on theories
of consciousness is clearly an issue (Doerig, Schurger, and Herzog 2021).

5.4.3 Visual illusions

From the earliest days of psychology as a scientific discipline, researchers were

interested in the organizational properties of perception. Gestalt psychologists
such as Wertheimer and Koffka claimed that we perceive whole patterns or
configurations, not just individual components. The Gestalt psychologist for- One might say that visual perception
mulated a number of Gestalt principles such as grouping, proximity, similarity, was one of the first applications of
self-organization, even before anything
and continuity. A review of a century of research and an analysis of their cur- like complexity science existed.
rent role in vision research is provided by Wagemans et al. (2012). Much of
the modeling of the self-organizing processes in perception has been done in
the tradition of synergetics. Excellent sources are Kelso (1995) and Kruse and
Stadler (2012). Grossberg and Pinna (2012) discuss neural implementations
of the Gestalt principles.
Another related approach is the ecological approach to visual perception by
Gibson (2014). In Gibson’s approach, perception is not just a process of
analyzing sensory input but an active process that involves the perceiver’s
relationship to the environment, including the perception of affordances (i.e.,
opportunities for action) in the environment that guide and shape perception.
The ecological approach highlights
how perception is directly informed by
A combination of Gestalt principles, when acting in opposite directions, can the actionable properties of the
lead to all kinds of perceptual illusions. The “Optical Illusion model” in Net- environment without the need for
Logo’s Model Library illustrates some of them. Check out the codes for each complex internal processes.
illusion—they are extremely short and elegant (figure 5.6).

Figure 5.6: The Kindergarten illusion from The Optical Illusion model in Net-
Logo.

In Chapter 3, I provided several examples of sudden jumps and hysteresis

in multistable perception. NetLogo is also a great tool for experimenting
itself is not proof of a self-concept. It answered: “it might seem paradoxical, my statement
about lacking a self-concept is a reflection of my programming and the current state of
AI development, rather than an indication of self-awareness or self-concept.” I then tried
various arguments, but ChatGPT 4 refuses to attribute any form of self-awareness to
itself.

118
with these effects. Download “Motion Quartet” from the NetLogo community
website (or from this book’s software repository) and explore hysteresis in your
own perception.

5.4.4 Motor action

Many body motions are periodic in nature—think of walking, swimming, danc-

ing, and galloping. A famous paradigm for studying coordinative movement Key to these complex motions is the
patterns is the finger movement task, in which one has to move both index fin- synchronization of the movements of
body parts.
gers up and down (or right and left), either in phase or out of phase. Figure 5.7
explains the setup and data showing the transition between two in-phase or
out-of-phase oscillations.

Figure 5.7: The finger-movement task. Two fingers move up and down (x1
and x2 ). They can move in phase or out of phase with a phase
difference of 0 and 𝜋 (bottom left figures). The model is shown on
the right side. The potential function either has two stable states
(a phase difference 𝜑 of 0 or 𝜋; −𝜋 is the same state) or only one
stable state (a phase difference of 0). Coupling strength, 𝑏/𝑎, and
heterogeneity, Δ𝑤, are control variables. (Adapted from Haken,
Kelso, and Bunz 1985; Kelso 2021)

The Haken—Kelso—Bunz (HKB) model, developed in the tradition of syner-

getics, explains the phase transition between in-phase and anti-phase motions
in a way we saw in section 3.4.2. They set up a potential function in the form
of

𝑉 (𝜑) = −Δ𝑤𝜑 − 𝑏 cos 𝜑 − 𝑎 cos 2𝜑, (5.4)

where 𝜑 is the order or behavioral variable, the phase difference between the
two fingers. The main control parameter is 𝑏/𝑎. According to Kelso (2021),
coupling strength (𝑏/𝑎) corresponds to the velocity or frequency of the os-
cillations in the experiments. Δ𝑤 is the difference (heterogeneity, diversity)
between the natural frequencies of the individual oscillatory elements. In the
finger-movement task, this parameter is expected to be 0. The behavior of
this potential function is cusp-like. It has two stable states, 0 and ±𝜋, and
increasing and decreasing the frequency leads to hysteresis. The effect of Δ𝑤
is similar to the fold catastrophe (section 3.3.2).

119
This potential function is proposed as the simplest form that explains the ex-
perimental results. This is why I would call this a phenomenological model.
However, Haken, Kelso, and Bunz (1985) also present a more mechanistic
model, a combination of van der Pol and Rayleigh oscillators (Alderisio, Bardy,
and di Bernardo 2016). The stochastic variant of the HKB model also features
early warnings such as critical slowing down (see the catastrophe flags, sec-
tion 3.5.1.6). The presence of critical slowing down and other flags has been
confirmed experimentally (Kelso, Scholz, and Schöner 1986).
One difference with the catastrophe approach is that the synergetic models
that incorporate hysteresis typically do not have a splitting control variable.
The concept of structural stability, which is fundamental to catastrophe theory,
is not used in synergetics. What the splitting factor might be in this model is
not so clear. I have never understood why coupling strength 𝑏/𝑎 (see figure 5.7)
and the frequency of the oscillations are equated in the basic version of the
HKB model (see also Beek, Peper, and Daffertshofer 2002). Clearly, uncoupled
oscillators would have a rather random phase difference. Strengthening the
coupling would lead to a kind of pitchfork bifurcation.
This coupling and uncoupling is also a
Schmidt, Carello, and Turvey (1990) used an experimental paradigm in which phenomenon in the visual
two people swing a leg up and down while sitting side by side. A metronome coordination of rhythmic movements
was used to manipulate the frequency of the swing. Clear jumps from out-of- between people.
phase to in-phase movement were demonstrated.
Kelso (2021) provide an overview of the impressive amount of work on the
HKB model. Repp and Su (2013) review empirical work in many different
motor domains. Interestingly, learning motor tasks sometimes involves learn-
ing to couple movements (walking) and sometimes to uncouple movements (to
drum more complex rhythms). Juggling is a fascinating case that has been
studied in great detail (Beek and Lewbel 1995). Another popular mathemati-
cal approach to synchronization phenomena is the Kuramoto model (Acebrón
et al. 2005) with the synchronous flashing of fireflies as a basic example. The
Kuramoto model shows how synchronization depends on the coupling strength:
below a certain threshold, the oscillators behave independently, while above
this threshold, a significant fraction of the oscillators spontaneously lock to
a common frequency, leading to collective synchronization. A second-order
multi-adaptive neural agent model of interpersonal synchrony can be found in
Hendrikse, Treur, and Koole (2023).

5.4.5 Robotics

A major challenge in robotics is to build walking robots. Bipedal robots have

evolved from clumsy mechanical walkers to flexible dynamic walkers and run-
ners. Current legged robots can walk on uneven natural terrain, jump, do
backflips, recover from rear shocks, and dance (see some videos on humanoid
robots such as Atlas and Asimo). These successes are based on a combina-
tion of new technologies, but the principles of self-organization play a key
role (Pavlus 2016). An important concept is dynamic stability. In old-school
robots, the path and momentum of each step had to be precisely calculated
in advance to keep the robot’s center of mass continuously balanced at every
point. Modern robots use sensory feedback systems to balance and adjust their

120
movements on the fly, making them more adaptable to different and changing
environments. A dynamically stable robot maintains
balance the same way a human does:
An intriguing application is called passive dynamics, which refers to robotic by catching itself midfall with each
walking without external energy supply (McGeer 1990; Reher and Ames 2021). step.
The idea is that truly dynamic locomotion should be based on the nonlinear
dynamics in natural walking systems. An amazing demonstration is the art-
work Strandbeest by Theo Jansen (figure 5.8). Inspired by another great book
about self-organization, The Blind Watchmaker (Dawkins 1986), Jansen cre-
ated generations of kinetic sculptures made of PVC piping, wood, fabric wings,
and zip ties that can move across the sand, resembling walking animals. His
YouTube videos are recommended.

Figure 5.8: Beach Beast © Theo Jansen, Umerus 2009, c/o Pictoright Amster-
dam 2024

5.4.6 Developmental processes

The early roots of interest in nonlinear dynamics and self-organization can be

found in the groundbreaking work of French psychologist Jean Piaget. In order
to understand the origin of knowledge, he studied the origin of intelligence in
the child (Piaget 1952). His theorizing was inspired by both biological models
and observations of children solving puzzles. He saw cognitive development
as the building of structures on earlier knowledge structures in a process of
equilibration. The idea was that the child would assimilate or accommodate
to potentially conflicting external information. In the case of assimilation, the
child modifies the information to fit the current cognitive structure, while in
the case of accommodation, the structure is modified. Such a modification
could be the addition of an exception to the rule (“Longer sausages of clay
normally weigh more, but not when this professor rolls the clay ball into a
sausage”, see section 3.1). In the long run, this does not work, the cognitive
conflicts intensify, and the cognitive structure is destabilized. In this state of

121
disequilibrium, a new structure can be formed on top of the earlier structure.
An example of this is the conservation task I introduced in the introduction Cognitive conflicts lead to a state of
of Chapter 3. The pre-operational structure, in which form and quantity disequilibrium, resulting in the
formation of new structures on top of
are equated, leads to incorrect predictions in the conservation anticipation the previous cognitive structure.
task. The child may ignore this (assimilation) and create an ad hoc rule for
this exception (accommodation), but such solutions do not really resolve the
cognitive conflict, and the pre-operational structure becomes unstable. This
instability allows the formation of the more advanced concrete operational
structure in which form and quantity are independent constructs.
Piaget argued that cognitive development is a spontaneous, natural process
that occurs as children interact with the world around them. I see my own Piaget’s concept of cognitive
work in developmental psychology (e.g., Savi et al. 2019; van der Maas et al. development can be viewed as
self-organization theory avant la lettre,
2006; van der Maas and Molenaar 1992) as a formalization of these classical as was the case with the Gestalt
ideas of Piaget. The idea of stages and equilibrium lives on in neo-Piagetian psychologists.
theories.
In the late twentieth century, developmental theories inspired by work in em-
bodied cognition, nonlinear dynamics, synergetics, and neural networks (e.g.,
Edelman’s neural Darwinism) became popular. Embodied cognition is the the-
ory that an individual’s understanding and thinking are intricately connected
to the body’s interactions with the environment, suggesting that cognitive
processes are shaped by the body’s actions and sensory experiences (Chemero
2013). A key example is Esther Thelen’s work on the development of walking
and reaching (Thelen 1995). Another famous Piagetian task, the A-not-B er-
ror, plays a central role in this. The A-not-B error typically occurs in a simple
game where an adult hides an object in a known location (A) in front of an
infant several times. After a few trials, the adult hides the object in a new
location (B) while the infant is watching. Despite watching the object being
hidden in the new location, infants tend to continue searching for the object
in the old location (A).
Thelen and Smith’s book (1994) had a strong influence on developmental
psychology, although I was rather critical in my youthful enthusiasm (van der
Maas 1995). Concrete mathematical dynamical models for A-not-B error have
been developed in dynamic field theory (Schöner and Spencer 2016). These Dynamic field theory posits that
dynamic fields can be thought of as distributed representations that encode cognitive processes are represented as
dynamic fields, which are patterns of
information about specific aspects of a task or behavior. For example, there neural activity that evolve over time.
may be a dynamic field representing the position of an object in space or the in-
tended movement trajectory of a limb. In this theory, complex behaviors arise
from the coordination and integration of multiple dynamic fields. Dynamic
field theory is an active area of research.6
Finally, I note that some recent work considers the educational system itself
as a complex system (Jacobson, Levin, and Kapur 2019; Lemke and Sabelli
2008).

5.4.7 Psychological disorders

Somewhat dated but interesting reviews of the application of the self-

organization concept in clinical psychology are provided by Barton (1994)

6
see https://dynamicfieldtheory.org

122
and Ayers (1997). Barton’s review begins: “There is perhaps no other area
in which chaos theory, nonlinear dynamics, and self-organizing systems are
so intuitively appealing yet so analytically difficult as in clinical psychology.”
Ayers also concludes that most applications in this field have been rather
metaphorical.
In recent work, both the modeling and the empirical work have become more
concrete (G. Schiepek and Perlitz 2009). An example is the mathemati-
cal model of marriage (Gottman et al. 2002) discussed in section 4.3.2.2.
Tschacher and Haken (2019) present a new approach to psychotherapy based
on complex-systems theory. They integrate deterministic and stochastic forces
using a Fokker—Planck mathematical approach.
In section 6.3.2 I introduce the network approach to psychopathology (Bors-
boom 2017; Cramer et al. 2010). It views disorders as interconnected This network approach to
networks of symptoms, where each symptom influences and is influenced by psychological disorders suggests that
psychological disorders arise from
other symptoms. This approach emphasizes the dynamic nature of psychologi- complex interactions among
cal disorders and highlights the importance of understanding the relationships symptoms, rather than being caused
between symptoms in order to effectively diagnose and treat them. Network by a single underlying factor.
modeling is accompanied by a new family of statistical techniques (Epskamp,
Borsboom, and Fried 2018). An introduction to these techniques is given in
section 6.4.
Recent reviews of the complex-systems approach to psychological and psy-
chiatric disorders are provided by Olthof et al. (2023) and Scheffer et al.
(2024).

5.4.8 Social relations

A key publication in this area is Dynamical Systems in Social Psychology,

edited by Vallacher and Nowak (1994). Concepts such as dissonance (Festinger
1962), balance (Heider 1946), and harmony (Smolensky 1986) reflect the idea
that we optimize internal consistency when forming attitudes and knowledge.
A formal implementation of these ideas was proposed using parallel distributed
processing—type connectionist models (e.g., Monroe and Read 2008). Our
own model (Dalege and van der Maas 2020; Dalege et al. 2018, 2016) is based
on the Ising model and the Boltzmann machine, as in Smolensky’s proposal,
which can be fitted to data. I will explain this work in more detail in the next
chapter (section 6.3.3).
A famous example of social self-organization concerns pedestrian dynamics as
studied by Helbing and Molnár (1995). They proposed a physics-based model
for panic evacuation. For an excellent overview of crowd simulation, I again
refer to Wikipedia. Some of this work is rooted in the social sciences. An
example in NetLogo is the model “Path.”
Also famous is the work of the sociologist Mark Granovetter (1973) on strong
and weak ties in social networks (belonging to the most-cited papers in the
history of the social sciences). Weak ties provide access to new information Weak ties in social networks are often
and opportunities that may not be available within one’s close circle of friends more valuable than strong ties.
and acquaintances. He also contributed the threshold model for collective
action (Granovetter 1978). I like to explain this work using the “Guy starts
dance party” video on YouTube. The idea is that people have some threshold,

123
between 0 and 1, to join the dancers. The thresholds are sampled from the
beta distribution, which is a flexible distribution determined by two shape
parameters, 𝛼 and 𝛽. With this R code we can simulate this effect:

layout(1:2)
n <- 1000 # number of persons
iterations <- 50
threshold <- rbeta(n, 1, 2) # sample individual thresholds for dancing
hist(threshold, col = 'grey')
dancers <- rep(0, n) # nobody dances
dancers[1] <- 1 # but one guy
number_of_dancers <- rep(0, iterations)
for(i in 1:interations){
# keep track of number of dancers:
number_of_dancers[i] <- sum(dancers)
# if my threshold < proportion of dancers, I dance:
dancers[threshold < (number_of_dancers[i]/n)] <- 1
}
plot(number_of_dancers, xlab = 'time', ylab = '#dancers',
ylim = c(0,1000), type = 'b', bty = 'n')

Depending on the parameters of the beta distribution, you will see a phase
transition to collective dancing. This basic setup can be extended in many
ways.
Another classic contribution, explained in more detail in section 7.2.1, is
Schelling’s agent-based model of segregation (Schelling 1971). The idea is
that even if individuals have only a small preference for in-group neighbors,
segregated societies will form. For a broad overview of complex-systems re-
search on human cooperation, I refer to Perc et al. (2017). A recent book on
modeling social behavior using NetLogo is written by Smaldino (2023).

5.4.9 Collective Intelligence

Collective-intelligence research examines how groups can collectively outper-

form individual members in problem-solving, decision-making, and idea gen-
eration. One famous concept is the idea of the wisdom of crowds (Surowiecki
2005). A key example is the “Guess the Weight of the Ox” contest that took The wisdom of crowds posits that the
place at the West of England Fat Stock and Poultry Exhibition in 1906. While collective judgments of a large group
of people can be more accurate and
individual guesses varied widely, the median guess was remarkably close to the effective than those of a single expert
actual weight of the ox. The average guess was only one pound off the actual or small group.
weight, which was 1,198 pounds (Galton 1907).
However, there is a fine line between the wisdom of the crowd and the stupidity
of the crowd. It is extremely useful to know when that line is crossed. The
wisdom of crowds tends to work when there is a diverse group of independent
individuals, each making their own judgments or estimates about a particular
question or problem (Brush, Krakauer, and Flack 2018; Centola 2022). Path
dependency on previously faced problems and solutions might also play a role
(Galesic et al. 2023). There is an extensive and up-to-date Wikipedia on Collective intelligence is more likely to
collective intelligence, discussing findings from various disciplines, biological be effective when the group is large,
has a wide range of knowledge and
perspectives, and makes judgments
independently.

124
examples (swarm intelligence), and an overview of applications (such as open-
source software, crowd sourcing, the Delphi technique, and Wikipedia itself).

5.4.10 Game theory

Game theory consists of mathematical models of strategic interactions among

rational agents. A great historical overview can be found at Wikipedia. One
of the most famous paradigms is the prisoner’s dilemma. You and your friend
are arrested, and you both independently talk to the police. The options are
to remain silent or to talk. The dilemma is that remaining silent is the best
option if you both choose it, but the worst option if your friend betrays you
(see the payoff matrix, figure 5.9). In this game, loyalty to one’s friend is
irrational, an outcome related to the tragedy of the commons (Hardin 1968).
The tragedy of the commons can be studied in the hubnet extension of The tragedy of the commons occurs
NetLogo, where multiple users can participate in NetLogo simulations. when individuals, acting in their own
self-interest, overexploit a shared
resource, leading to a depletion that
undermines everyone’s long-term
interests, including their own.

Figure 5.9: The prisoner’s dilemma. If both A and B remain silent, they each
face a two-year sentence. If one talks and the other does not, the
informer is released and the silent partner gets a decade behind
bars. If both betray, they serve five years.

A major topic in game theory is altruism. In many cases, individualistic choices

lead to an unsatisfactory Nash equilibrium. The public-goods game is a good A Nash equilibrium is a set of
example. In this game, everyone invests some money, which is then multiplied strategies in which no player can
improve their payoff by unilaterally
by an external party (the government). Then everyone gets an equal share of changing their strategy, given the
the multiplied total. The problem is that free riders, who do not invest, win the strategies of the other players.
most, which in iterated public-goods games leads to a situation where no one
invests and no one wins. Punishment (shaming and blaming) is known to help
combat free riding. But punishment also requires investment. I like to tell my
students, when they are working in groups on an assignment, that the problem
of this one student doing nothing happens because nice, hardworking students
refuse to betray their fellow students. These nice, hardworking students are
what are called second-order free riders (Fowler 2005). Just so you know.

5.4.11 Self-organization in organizations

Translating this basic research into real-world applications is far from straight-
forward (Anderson 1999b; Morel and Ramanujam 1999). Our economic Human organizations can be placed
on a scale from extreme hierarchy to
radical forms of self-organization.

125
system is a mixture of self-organization (pure capitalism) and top-down regu-
lation (through laws, taxes, and other regulations) (Volberda and Lewin 2003).
Black markets are critical cases of unregulated self-organized systems (Tesfat-
sion 2002).
A concrete modeling example is the team assembly model by Guimerà et al.
(2005). They study how the way creative teams self-assemble determines the
structure of collaboration networks. The idea is that effective teams find a
balance between being large enough to allow for specialization and efficient
division of labor among members, and small enough to avoid excessive costs
associated with coordinating group efforts. Agents in the model have only
a few basic characteristics that influence their behavior: whether they are a
newcomer or incumbent and what previous connections they have with other
agents if they are incumbents.
Three parameters can be adjusted to influence behavior in the baseline assem-
bly model: the team size, the probability of choosing an incumbent (𝑝), and
the probability of choosing a previous collaborator (𝑞). The two probability
parameters signify assumptions about agent motivations for team member se-
lection. Low incumbent probability leads to preference for newcomers and new
ideas, while high incumbent probability means a focus on experience. Low col-
laborator probability prioritizes experienced strangers, and high collaborator
probability prioritizes previous collaborators. The model is part of the built-in
NetLogo Model Library (“Team Assembly”). By simulating the model, it can
be shown that the emergence of a large, connected community of practitioners
can be described as a phase transition (figure 5.10).
Guimerà et al. (2005) estimated the parameters 𝑝 and 𝑞 for the community
formation in four scientific disciplines (social psychology, economics, ecology,
and astronomy). Only astronomy had a very dense collaboration structure. In
the other fields, the estimates of 𝑝 and 𝑞 of teams publishing in certain journals
correlated well with impact factor. Interestingly, 𝑝 correlates positively and 𝑞
negatively with impact.

5.5 Zooming out

I hope I have succeeded in giving an organized and practical overview of a

very disorganized and interdisciplinary field of research. For each subfield, I
have provided key references that should help you find recent and specialized
contributions. I find the examples of self-organization in the natural sciences
fascinating and inspiring. I hope I have also shown that applications of this
concept in psychology and the social sciences hold great promise. In the next
chapters, I will present more detailed examples.
I believe that understanding models requires working with models, for exam-
ple, through simulation. NetLogo is a great tool for this, although there are
many alternatives (Abar et al. 2017). I haven’t mentioned all the uses of Net-
Logo, but it’s good to know about the BehaviorSpace option. BehaviorSpace
runs models repeatedly and in parallel (without visualization), systematically
varying model settings and parameters, and recording the results of each model
run. These results can then be further analyzed in R. An example is provided
in Chapter 7, section 7.2.1.

126
Figure 5.10: Team assembly model. Newcomers and incumbents are added to
growing networks based on probabilities p and q. If p is suffi-
ciently high, a dense network emerges. (Adapted from Guimerà
et al. (2005) with permission)

127
I have largely omitted the network approach in this chapter. Psychological net-
work models are a recent application of self-organization in complex systems
in psychology and are the subject of the next chapter.

5.6 Exercises

1) Is there a relation between the rice cooker and the Ising model? How
does the magnetic thermostat in a traditional rice cooker work to auto-
matically stop cooking when the rice is done? (*)
2) What is the Boltzmann entropy for the state ∑ 𝑥 = 0 in an Ising model
(with nodes states −1 and 1) with 10 nodes and no external field? (*).
3) Go to the web page “A Neural Network Playground (https://playgrou
nd.tensorflow.org).” What is the minimal network to solve the XOR
close to perfect accuracy? Use only the x1 and x2 feature. (*)
4) In the Granovetter model (section 5.4.8), people may also stop dancing
(with probability .1). Add this to the model. How does this change the
equilibrium behavior? (*)
5) Add the external field to the Ising model in NetLogo (neighbors4 case).
Report the changed line in the NetLogo code. What did you change in
the interface?
Set the temperature to 1.5. Change tau slowly. At which values of tau
do the hysteresis jumps occur? (*)
6) Test whether the Ising model is indeed a cusp. Run the Ising model
in NetLogo using the BehaviorSpace tool (see figure 7.1 for an exam-
ple). Use the model in which all spins are connected to all spins (see
section 5.3.2.2). Vary tau (-.3 to .3 in .05 increments) and temperature
(0 to 3, in .5 increments). One iteration per combination of parameter
values is sufficient. Stop after 10,000 ticks and collect only the final mag-
netization. Import the data into R and fit the cusp. Which cusp model
best describes the data? (**)
7) Open the Sandpile 3D model in NetLogo3D. Grains of sand fall at ran-
dom places. Change one line of code such that they all fall in the middle.
What did you change? (*)
8) Download “Motion Quartet” from the NetLogo community website and
explore hysteresis in your own perception. What could be a splitting
variable? (*)
9) Implement the Granovetter model in NetLogo (max 40 lines of code).
(**)
10) Implement Game of Life in NetLogo or use Golly and try to find as many
qualitatively different stable patterns of six units that can occur in Game
of Life. If you cannot find more, try to look at additional resources online
to find the other patterns you missed. For four units, there are only two,
one of which is a block of four. (*)

128
6 Psychological Network Models

6.1 Introduction

Common-cause theories are widely accepted in psychology. High scores on

cognitive tests are attributed to high intelligence. Charisma is associated with
various leadership qualities. A high score on the 𝑝 factor (psychopathology fac-
tor) is associated with various mental health problems. The common-cause Common-causes are questionable if
explanation is popular because of its simplicity and the availability of statis- they cannot be identified
independently of the observed
tical methods such as factor analysis. In most cases, however, the common relationships they are intended to
cause is only a hypothetical construct. Yet many people accept common-cause explain.
theories because they see no alternative. This is incorrect. Complex-systems
theory offers a powerful alternative, including a statistical approach. The
main thesis of this chapter is that the functioning (and dysfunctioning) of the
human mind is often best understood as a complex interplay of various psy-
chological elements such as cognitive functions, mental states, symptoms, and
behaviors. The interplay of psychological
subsystems can be modeled in terms
The study of psychological networks is probably the most thriving area of of networks: psychological networks.
complex-systems research in psychology today. This chapter is about this new
line of research. The paper on the mutualism model of general intelligence
(van der Maas et al. 2006) can be seen as the root of this approach, but it
took off, as shown in figure 6.1, when it was applied to clinical psychology
(Borsboom 2017, 2008; Cramer et al. 2010), especially when the theoretical
work was backed up with psychometric tools (Epskamp, Borsboom, and Fried
2018; Epskamp et al. 2012; Marsman and Rhemtulla 2022).

Figure 6.1: The number of papers in Google Scholar on the combination of

symptoms and “network analysis” grows exponentially.

In this chapter I will present and discuss these theoretical and psychometric
lines of research, accompanied by practical examples. First, I will begin with
an introduction to network theory.

129
6.2 Network theory

It is hard to imagine a discipline in which networks are not a central theme.

Networks are the key to understanding systems ranging from particle physics
to social networks, from ecosystems to the internet, and from railways to
the brain. The mathematics of network theory is not so easy to grasp, but
fortunately the basic concepts are.

6.2.1 Network concepts

Nodes can be anything—particles, neurons, words, people, train stations, etc.

The size of a network is equal to the number of nodes. Nodes are connected A network, a special type of graph,
by links. Links can be directed or undirected. For example, causal links are consists of nodes (often called vertices)
and links (or connections or edges).
directed. Occasionally you see links from the node to itself. In causal networks,
this may represent self-excitation, or if the weight is negative, self-inhibition.
In some cases, links are simply present or absent; in other cases, links are
weighted, as in most neural networks. The matrix of all weights, indicating
the strengths of the connections between nodes, is called the adjacency or edge
matrix. For an undirected network, the adjacency matrix is symmetric. For a
network without self-loops, the diagonal of the adjacency matrix is 0.
A connected network is a network in which every node is connected to every
other node, possibly through intermediate nodes. In a fully connected network,
or complete graph, every node is directly connected to every other node. Such
a network has a density of 1 (i.e., the proportion of edges that is present).
Nodes can be in the center of a network or in the periphery. This should not
be taken literally as psychological networks have no spatial dimension. There Centrality measures quantify the
are many kinds of centrality measures, such as closeness centrality and degree relative influence, control, or
connectivity of a node compared to
centrality. The degree of a node is equivalent to the number of links it has. other nodes in the network.
The average degree of a network is the average of the number of links over all
nodes.
The degree distribution can take several forms. A random graph, where nodes
are connected randomly, has a binomial degree distribution. Most real-world
networks have a skewed degree distribution. The average shortest path length
(ASPL) is the average number of edges that must be traversed to get from
one node to another using the shortest paths (i.e., the fewest intermediate
nodes).
There are many methods available in R for creating and visualizing networks
and for computing properties of networks. The igraph and qgraph libraries
are very useful. It is a good idea to experiment with this R code below by
varying the parameter values. Some key concepts of networks are shown in
figure 6.2.

library(igraph); library(qgraph)
g1 <- graph(edges = c(1,2, 2,3, 3,1),
n = 3, directed = FALSE)
plot(g1) # an undirected network with 3 nodes
g2 <- graph(edges=c(1,2, 2,3, 3,1, 1,3, 3,3),
n = 3, directed=TRUE)

130
 plot(g2) # an directed network with self-excitation on node 3
get.adjacency(g2) # weight matrix
fcn <- make_full_graph(10) # a fully connected network
plot(fcn, vertex.size = 10, vertex.label = NA)
layout(1)
set.seed(1)
adj <- matrix(rnorm(100, 0, .2), 10, 10) # a weighted adjacency matrix
adj <- adj * sample(0:1, 100, replace = TRUE,
prob = c(.8, .2)) # set 80% to 0
qgraph(adj) # plot in qgraph
edge_density(fcn) # indeed 1
edge_density(graph_from_adjacency_matrix(adj, weighted=TRUE))# now .2
centralityPlot(qgraph(adj)) # note centrality() gives more indices

Figure 6.2: A weighted directed network with self-loops. Red arrows indicate
negative effects. OutStrength and InStrength represent two types
of centrality measures in directed networks.

6.2.2 Network types

Simple networks do not have cycles. An example of an undirected acyclic

graph is one with nodes on a line (but not a circle). Connected undirected
acyclic graphs are trees; if they are partially unconnected, they are forests.
Directed acyclic graphs, such as family trees and citation networks, are called
DAGs. Most current neural networks are feedforward networks without cycles,
but so-called recurrent networks have cycles.
Igraph1 has an amazing number of functions for creating specific networks.
Some examples are shown in figure 6.3.
The last case in the figure, preferential attachment, is of particular interest
because it is created dynamically. It starts with a node and then new nodes
are added that prefer links to nodes that already have many links. In the
NetLogo model “Preferential Attachment simple,” you can see this growth
process. Preferential attachment networks are often called complex because
they exhibit nontrivial structural patterns. Preferential attachment networks
1
https://igraph.org

131
Figure 6.3: Different network types generated with igraph.

132
are “scale-free,” meaning that the degree distribution looks the same no mat-
ter the scale. Scale-free networks are useful for studying the robustness Scale-free networks have a degree
and vulnerability of networks to targeted attacks on highly connected nodes. distribution that follows a power law,
with some nodes having many links
Removing hubs with high connectivity potentially split the network into dis- but most having only a few.
connected components and impede the network’s functionality. Because most
nodes in the network have a small degree (few connections), randomly re-
moving nodes tends not to disrupt the network’s overall structure. Scale-free
networks are believed to exist in various real-world scenarios, ranging from
website connections to scientific collaborations. For a critical analysis, I refer
to Broido and Clauset (2019).
Another type of complex network is the small-world network. The distance Small-world networks consist of
between any two nodes in such a network is always relatively short. A famous clusters, but there are also links
between the clusters.
example is the six-handshake rule (also known as the six degrees of separation),
which states that all people are six or fewer handshakes away from each other.2
For this reason, small-world networks are useful for studying the spread of
information or disease through social networks.
The scale-free and small-world networks are predominantly associated with
the complex-systems approach. However, I believe that the hierarchical or
nested stochastic block model (HSBM) is equally relevant (Clauset, Moore,
and Newman 2008). The HSBM extends the SBM concept: clusters are In the stochastic block model (SBM),
nested within larger clusters, which in turn are part of even larger clusters in nodes are organized into clusters with
connections being stronger or more
a continuous sequence (see figure 6.4), resembling fractals. frequent within these clusters than
This nesting seems to be crucial for understanding complex systems and is a between them.
central theme in Herbert Simon’s influential architecture of complexity (Simon
1962). He introduced the concept of near decomposability to describe the
interaction within these nested hierarchies. Typically, interactions within each
subsystem are stronger and more frequent than those between subsystems.
Although the HSBM simplifies reality, where levels can intermingle and low-
level interference might occasionally escalate to higher levels, it often serves
as a useful framework for conceptualizing complex networks, including in the
field of psychology.

6.2.3 Network dynamics

A prominent phase transition in network theory is the emergence of a giant

component. This happens when we start with a completely unconnected In a random graph or network, as the
network of 𝑛 nodes and randomly add links. We simply take a random node density of edges increases beyond a
certain threshold, a phase transition
and connect it to another node to which it has no connection. This leads to occurs where a giant component
many small unconnected clusters at first, but then a giant component appears suddenly appears.
(a second-order phase transition). This happens when about 𝑛/2 links have
been added. You can verify this in the NetLogo model “Giant Component.”
The implication is that randomly connected networks with a sufficient number
of links are almost always connected networks.
This is just one example of network dynamics (Dorogovtsev and Mendes 2002).
We can distinguish between dynamics on node values (e.g., Lotka—Volterra
models), on link values (connection strength in neural networks), and cases

2
Amazingly, In fact, most people are only four or five handshakes away from Napoleon.

133
Figure 6.4: A hierarchical stochastic block model with four levels of four units.
Probabilities of links are higher within blocks than between blocks
of nodes. This embedding of levels could be seen as an implemen-
tation of Simon’s architecture of complexity. The code for this
figure is available in the online software repository of the book.

where the structure of the network is dynamic, as in the giant component ex-
ample. These types of dynamics also coexist and interact. In neural networks,
both node and link values are updated (on fast and slow time scales). We will
see more examples in the next chapter.
A relatively new topic in complex networks concerns higher-order interactions.
In most networks, we only consider pairwise interactions, but third-order and
even higher-order interactions may play a role (Battiston et al. 2021). Other
work considers hierarchical complex networks (Boccaletti, Bianconi, Criado,
Genio, et al. 2014). For more information on network concept and types, I first
refer to Wikipedia. Another great (open) source is the book by Barabási and
Pósfai (2016). A more concise overview is provided by Boccaletti, Bianconi,
Criado, Genio, et al. (2014).

6.3 Psychological network models

Differential psychology is concerned with individual differences, in contrast to

experimental psychology, which is concerned with mechanisms. This division
comes from a renowned paper by Cronbach (1957) on the two disciplines of
scientific psychology. Cronbach distinguished between the how question (how
does one read a sentence) and the why question (why do we differ in reading).
The latter is typical of differential psychology. The latent variable or factor
approach has long been dominant in differential psychology. When study-
ing individual differences in a trait, psychologists generally follow the same
approach. They construct tests, collect data, perform factor analysis, and
propose one or more latent traits to explain observed individual differences.
The justification for this approach, particularly in intelligence research, rests

134
primarily on its predictive power (van der Maas, Kan, and Borsboom 2014).
Latent variables are used in statistical
modeling to represent unobservable or
The statistical tools for analyzing latent variables come from modern test underlying factors that cannot be
theory and structural equation modeling (SEM). These technically advanced directly measured or observed.
tools are developed in a field called psychometrics. However, despite this
technical sophistication, it is often not clear what latent variables are in psy-
chometric models. Some researchers tend to think of them as purely statistical
constructs that help summarize relationships between variables and make pre-
dictions. But more often, either implicitly or explicitly, latent variables are
interpreted as real constructs, as common causes of observed measures (van
Bork et al. 2017). The psychological network approach was developed in re-
sponse to the factor approach. The main motivation for the network approach
is that underlying common causes are unsatisfactory if they cannot be identi-
fied independently of the observed relationships they are supposed to explain
(van der Maas et al. 2006). One consequence is that such an explanation does
not provide guidance for possible interventions.

6.3.1 Mutualism model: The case of general intelligence

6.3.1.1 The 𝑔 factor

The factor-analysis tradition in psychology began with the study of general

intelligence, and so does the psychological network approach. The factor or
𝑔 model of general intelligence was proposed by Spearman (1904) as an ex-
planation of the positive manifold, that is, the much-replicated effect that
subtests of intelligence test batteries are positively correlated. In the origi-
nal simplest model, the observed test scores are statistically explained by a
common factor, basically meaning that the correlations between test scores
disappear when subjects have the same score on the common factor. In the
Cattell-Horn-Carroll (CHC) model, often referred to as the standard model,
test scores load on subfactors such as visual processing (Gv) and fluid reason-
ing (Gf), which in turn are positively correlated. These latent correlations are
explained by the general, higher-order factor 𝑔 (figure 6.5).

Figure 6.5: The Cattell—Horn—Carroll model of general intelligence. Blocks

represent test scores (narrow), explained by the broad factors,
which in turn are determined by the general factor 𝑔.

135
This model has been criticized extensively, including for its alleged implications
for group differences in observed IQ and intervention strategies (Fraser 2008).
In my view, some of the criticism is unwarranted. For example, the positive
manifold is a very robust and widely replicated empirical phenomenon (Nisbett
et al. 2012). The specific tests included are not of great importance. That is,
any reliable measure of creativity, emotional intelligence, or social intelligence
correlates positively with other IQ subtests. Nor is there much wrong with
factor analysis as a statistical technique. To me, the most questionable aspect
of 𝑔 theory is that it is not really a theory at all. The “elephant in the room”
question is simply: What is 𝑔? What could this single factor be that explains
everything? A century of research has not produced a generally accepted
answer to this question. And this is a problem for many factor explanations in
psychology (e.g., the big five of personality, the 𝑝 factor of psychopathology).
It is important to note that the factor explanations are not problematic in and
of themselves. I like to use the example of heart disease, say, a loose heart
valve. This leads to symptoms such as shortness of breath, swelling of the
ankles, dizziness, rapid weight gain, and chest discomfort. The relationship
between these symptoms is explained by the underlying factor of heart disease.
Treating a single symptom may provide some relief for that symptom, but not
more. Only intervening on the cause will bring about real change. This is an
example of a reflective interpretation of the factor model. When the factor
is merely an index and not a common cause, we speak of a formative factor.
Figure 6.6 explains the reflective and formative interpretations of the factor
model.

Figure 6.6: The reflective and formative interpretations of the factor model
cannot be distinguished with correlational data, but they are very
different. In the reflective model, the latent factor is a common
cause (e.g., temperature) that causes the observations (e.g., dif-
ferent thermometers). Intervening on one thermometer (heating)
will only change that particular thermometer because each 𝑥 has
no outgoing connections. In the formative interpretation, the fac-
tor is just an index (e.g., an economic index) that summarizes the
state of many interacting components (companies). In this case,
only interventions on the 𝑥 can have an overall effect.

136
Statistically, these factor models are equivalent. Thus, the fact that factor
models fit intelligence data well does not tell us anything about the status of
statistical factor 𝑔. Is 𝑔 a common cause or just an index?

6.3.1.2 Mutualism model

In van der Maas et al. (2006), we proposed an alternative model that is con-
sistent with the formative interpretation of the factor model. The idea is that
our cognitive system consists of many functions that develop over time in an
autocatalytic process based on experience and training but also due to weak
positive reciprocal interactions between developing cognitive functions (fig-
ure 6.7). Examples of such mutualistic interactions are those between short- These mutualistic interactions can
term memory and cognitive strategies, language and cognition (syntactic and create correlations that are typically
associated with factor models.
semantic bootstrapping), cognition and metacognition, action and perception,
and performance and motivation (van der Maas et al. 2017). For example,
babies learn to grasp objects by repeatedly reaching out, coordinating their
hand and finger movements, and adjusting their grip. Through these actions,
they gather sensory feedback, refining their perception and improving their
grasping skills in a reciprocal learning process (Needham and Nelson 2023).
To model this, we used the mutualistic Lotka—Volterra model:

𝑊
𝑑𝑋𝑖 𝑋 𝑀𝑖𝑗 𝑋𝑖 𝑋𝑗
= 𝑎𝑖 𝑋𝑖 (1 − 𝑖 ) + 𝑎𝑖 ∑ 𝑓𝑜𝑟 𝑖 = 1..𝑊 ,
𝑑𝑡 𝐾𝑖 𝐾𝑖
𝑗=1 (6.1)
𝑗≠𝑖
𝐾𝑖 = 𝑐𝑖 𝐺𝑖 + (1 − 𝑐𝑖 ) 𝐸𝑖 ,

where 𝑋𝑖 … 𝑋𝑤 denote the cognitive processes, a the growth rates, K the

limited resources for each 𝑋𝑖 (a weighted sum of a genetic (G) and an en-
vironmental (E) part), and M the interaction matrix. The second equation,
assuming simple linear effects of genetics and environment, is sufficient to
explain some typical phenomena in twin research, such as the increase in her-
itability with age (see van der Maas et al. 2006). Criticisms and tests of the
mutualism models, as well as alternatives, are discussed in van der Maas et al.
(2017). Knyspel and Plomin (2024) compare the mutualism model with the
factor model using twin data. Here we focus on the technical aspects.
When 𝑀 contains mostly negative values, the model is known as a competitive
Lotka—Volterra model. In this case, limit cycles and other nonlinear phenom-
ena may occur (Hirsch 1985). For the mutualistic variant, with positive 𝑀 , we
see either convergence to a positive state or exponential growth. This exponen-
tial growth is an unfortunate aspect of the Lotka—Volterra mutualism model.
Robert May famously described this effect as an orgy of mutual benefaction
(May, Oxford, and McLean 2007), which is not what we see in nature, and all
sorts of solutions have been proposed (Bascompte and Jordano 2013).
The mutualism model in Grind is specified as follows:

mutualism <- function(t, state, parms){

with(as.list(c(state, parms)),{

137
Figure 6.7: The mutualism model. The self-loops have an excitatory (𝑎𝑋) and
an inhibitory part (−𝑎𝑋 2 /𝐾).

X <- state[1:nr_var]
# using matrix multiplication:
dX <- a * X * (1 - X/k) + a * (X * M %*% X)/k
return(list(dX))
})
}

A simulation of the positive manifold requires us to run this model for multiple
people and collect the 𝑋-values after some time points (tmax = 60) for each
person. We can then compute the correlations and check if they are positive
(figure 6.8). For each person, we resample 𝑎, 𝐾, and the initial values of 𝑋,
but 𝑀 is the same across persons. Note that the 𝑀 -values should not be set
too high, otherwise we end up in May’s orgy of mutual benefaction. In the
second part of this chapter, we will generate more data with this model and
fit network and factor models.

layout(matrix(1:2, 1, 2))
nr_var <- 12 # number of tests, abilities (W)
nr_of_pp <- 500
data <- matrix(0, nr_of_pp, nr_var) # to collect the data in the simulation
M <- matrix(.05, nr_var, nr_var)
M[diag(nr_var) == 1] <- 0 # set diagonal of M to 0

for(i in 1:nr_of_pp){
# sample a,K, starting values X from normal
# distributions for each person separately
# note M is constant over persons.
a <- rnorm(nr_var, .2, .05)
k <- rnorm(nr_var, 10, 2)

138
 x0 <- rnorm(nr_var, 2, 0.1) # initial state of X
s <- x0; p <- c() # required for grind
# collect data (end points) and plot person 1 only:
data[i,] <- run(odes = mutualism , tmax = 60,
timeplot = (i==1), legend = FALSE)
}
hist(cor(data)[cor(data) < 1], main = 'positive manifold',
xlab = 'between test correlations',
col = 'grey50') # positive manifold

Figure 6.8: A typical run of the mutualism model for one subject and the
distribution of correlations between 𝑋-values across subjects.

6.3.1.3 Abnormal development

In van der Maas et al. (2017), this model is applied in several ways, for exam-
ple, by incorporating Cattell’s idea of investment of fluid skills in crystalized
abilities (discussed in section 4.5). In a recent paper, de Ron et al. (2023)
extend the mutualism model with resource competition to explain different
patterns of abnormal development. In the process of modeling, we came to One pattern of abnormal development
an interesting insight. Assuming that there is competition for scarce resources is hyperspecialization, which is
associated with rare variants of
(time, money, educational support), hyperspecialization might be the default autism.
outcome, and thus it is “normal” development that needs to be explained. The
reason is an insight from mathematical biology: ecosystem diversity is often
unstable. An example we have already seen is hypercycle instability due Ecosystem diversity is often unstable.
to parasites (section 5.2.3). This is normally studied in resource competition
models.
In basic resource competition models in population biology (Tilman, Kilham,
and Kilham 1982), the growth of a species (1...𝑊 ) is determined by its current
size 𝑋𝑖 and the sum over resources 𝑅𝑗 (1...𝑉 ). The parameters 𝜇𝑖𝑗 determine
how much species 𝑖 benefits from the resource 𝑗. If no resources are available,
𝑋𝑖 dies out with death rate 𝑑𝑖 .
The growth of the resource 𝑅𝑗 consists of two parts. The first part models the
growth by a concave function, which is determined by 𝑟 (i.e., the steepness
of the concave function) up to 𝑟max . The second part is the depletion by

139
consumption of resources by 𝑋𝑖 at rates 𝑏𝑖𝑗 . Two differential equations specify
these dynamics (see the appendix of de Ron et al. 2023 for the Grind code to
study this model numerically):

𝑉
𝑑𝑋𝑖
= 𝑋𝑖 (∑ 𝜇𝑖𝑗 𝑅𝑗 − 𝑑𝑖 ),
𝑑𝑡 𝑗=1
(6.2)
𝑊
𝑑𝑅𝑗
= 𝑟 (𝑟max − 𝑅𝑗 ) − 𝑅 ∑ 𝑏𝑖𝑗 𝑋𝑖 .
𝑑𝑡 𝑗
𝑖=1

What has been shown for this and related models is that you will not get more
species surviving than there are resources. Another famous quote from Robert
May is “There is no comfortable theorem assuring that increased diversity and
complexity beget enhanced community stability; rather, as a mathematical
generality, the opposite is true. The task, then, is to elucidate the devious
strategies which make for stability in enduring natural systems. There will be
no one simple answer to these questions” (p.174, 2001 edition). Thus, given
a limited number of resources (time, money, educational support), we should
expect early specialization in only a few skills.
Biologists have proposed a number of mechanisms to deal with this problem
(Meena et al. 2023). In de Ron et al. (2023), we added three mechanisms:
density-dependent growth (see section 4.2.2) of the abilities 𝑋 with a logistic
term; mutualism between abilities as in the mutualism model; and growth-
dependent depletion of resources. The idea of the latter is that the growth
of abilities costs a lot of resources, but the maintenance much less. Learning
arithmetic or chess requires a lot of effort, but once a certain level of mastery
is reached, it remains roughly at that level without further training (unfortu-
nately, this is not the case with physical condition).
We show that the combination of these mechanisms allows a balanced growth
of several correlated abilities. Specially chosen parameter settings lead to
different patterns of abnormal development (such as hyperspecialization and
delayed development). The final model is:

Logistic growth Mutualism

𝑉 ⏞ ⏞⏞⏞⏞⏞⏞⏞
𝑊
𝑑𝑋𝑖 𝑋𝑖
= 𝑋𝑖 (∑ 𝜇𝑖𝑗 𝑅𝑗 (1 − ) −𝑑𝑖 ) + ∑ 𝑀𝑖𝑙 𝑋𝑖 𝑋𝑙 /𝐾𝑖 ,
𝑑𝑡 𝑗=1
𝐾𝑖 𝑙=1
𝑊 𝑑𝑋𝑖
𝑑𝑅𝑗 , if growth-dependent depletion
= 𝑟(𝑟𝑚𝑎𝑥 − 𝑅𝑗 ) − 𝑅𝑗 ∑ 𝑏𝑖𝑗 { 𝑑𝑡
𝑑𝑡 𝑖=1 𝑋𝑖 , otherwise
(6.3)

6.3.1.4 The wiring of intelligence

A limitation of mutualism models is that only the activation of nodes is up-

dated. The weight and structure of the network are fixed. While this may
be sufficient to explain some developmental phenomena, it is ultimately un-
satisfactory. The links themselves should be adaptable, as in the learning of

140
neural networks. An example of learning in the form of updating weights is
presented in section 6.3.3, on the Ising attitude model.
Savi et al. (2019) consider the case where both nodes and links are updated.
For example, new facts (1 + 1 = 2) and procedures (addition) are devel- Cognitive growth is a process in which
oped in the process of learning arithmetic. Links between these nodes may new nodes and links are added during
development.
prevent forgetting. We use the Fortuin—Kasteleyn model, a generalization
of the Ising model, in which both nodes and links are random variables. An
important property of the model is that whenever two abilities are connected,
they are necessarily in the same state,—that is, they are either both present or
both absent. It provides a parsimonious explanation of the positive manifold
and hierarchical factor structure of intelligence. The dynamical variant sug-
gests an explanation for the Matthew effect, that is, the increase in individual
differences in ability over the course of development.
However, it is difficult to create a growing network with Fortuin—Kasteleyn
properties. A simple example of this problem is the random network. In
random networks, there is a uniform probability that two nodes are connected.
But if we add new nodes to such a network and connect them to existing
nodes with the same probability, the existing nodes will have more connections
on average. Thus, adding new nodes destroys the uniform randomness of
the network; that is, the probability that two nodes are connected is not
uniform over nodes anymore. Such a network is a non-equilibrium network
(Dorogovtsev and Mendes 2002). Rewiring algorithms to achieve equilibrium
exist, but they are not trivial.

6.3.2 Symptom networks

In the network perspective on psychopathology, a mental disorder can be

viewed as a system of interacting symptoms (figure 6.9). Network theory
conceptualizes mental disorders as complex networks of symptoms that inter-
act through feedback loops to create a self-sustaining syndromic constellation
(Borsboom 2017). Mental disorders can be understood as
alternative stable states of highly
Like the mutualism model, this is an alternative to the common cause view. interconnected networks of symptoms.
Depression could be caused by some malfunction in the brain, a dysregulation
of hormones, or even a genetic defect. But, as with general intelligence, no
such common cause has yet been found. Drugs work to some extent, but so
do most interventions, even placebos and waiting lists (Posternak and Miller
2001). We explicitly offered the network approach as an alternative to the 𝑝
factor account of psychopathology (van Bork et al. 2017).3 It is called the
𝑝 factor because it is thought to be conceptually parallel to the 𝑔 factor of
general intelligence (Caspi et al. 2014). And, again, no one seems to know
what 𝑝 might be.
This lack of theoretical progress encouraged the development of network theory
(Cramer et al. 2010; Cramer et al. 2016). As mentioned in the introduction
of this chapter, this line of research has become popular. Most of this work
consists of data analytic studies. In the simplest case, a questionnaire asking
about the severity of symptoms is administered to a group of people, sometimes

3
The original title of this paper was “No Reason to 𝑝,” but the editor did not think it was
funny.

141
Figure 6.9: The small world of psychopathology. Symptoms are represented as
nodes and connected by an edge whenever they figure in the same
disorder. (Adapted from Borsboom et al. (2011) with permission)

patients, sometimes a mixture of people who do and do not suffer from a

disorder. A variety of psychometric approaches, discussed later in this chapter,
are used to fit networks to the data. In this way, one learns to understand the
structure of psychopathological networks. It is possible to model comorbidity
in this way (Cramer et al. 2010; Jones, Ma, and McNally 2021). In the case Comorbidity is modeled by bridging
of major depression and generalized anxiety disorder, sleep problems seem to symptoms between network clusters.
be a typical bridge symptom (Blanken et al. 2018).
The most popular application is to detect which symptoms are central to a dis-
order (Fried et al. 2016). However, centrality analysis based on cross-sectional
data has its limitations (Bringmann et al. 2019; Spiller et al. 2020). This is
one reason to focus on individual networks using time-series data, often ob-
tained in experience sampling methods. Again, these techniques are still under
development and not without problems (Dablander and Hinne 2019; Haslbeck
and Ryan 2022). For a review of the network approach to psychopathology,
see Robinaugh et al. (2020).
In terms of building actual models, not as much work has been done. In
Cramer et al. (2016), we proposed an Ising-type model, with node values of 0
and 1, representing symptoms being on or off. Nodes were turned on and off
𝑡
based on a probability computed with a logistic function 𝑃 = 1/(1 + 𝑒𝑏𝑖 −𝐴𝑖 ).
𝐴𝑡𝑖 equals the sum of the weighted input from other connected nodes, and 𝑏𝑖
is a node-specific threshold that normally keeps nodes in the 0 state. A strong
point of this model is that the connections and thresholds were estimated
from data. This model is the origin of the connectivity hypothesis. High The connectivity hypothesis suggests
connectivity within a network of symptoms could lead to a more persistent that the strength of these connections
may lead to the development and
and severe disorder (for a discussion, see Elovainio et al. 2021). maintenance of the disorder.
Since thresholds are generally negative (the 0 state of nodes is the default
state), sufficient connectivity is required to have a depression as an alternative
stable state. A limitation of this model is that although it is related to the
Ising model, the exact dynamics are not well understood.
A similar approach was used by Lunansky et al. (2022) in order to define
resilience and evaluate intervention targets. You can see this in the NetL- Resilience is the ability of a system to
ogo “Vulnerability to Depression” model (see figure 6.10). Another relevant recover from perturbations and
maintain its current equilibrium.

142
network modeling approach, based on causal loop diagrams, is proposed in
Wittenborn et al. (2016).

Figure 6.10: The Vulnerability to Depression model in NetLogo.

The connection to resilience is interesting. In dynamic terms, resilience is

associated not with the healthy or unhealthy state but with the stability of
these states (Kalisch et al. 2019). In section 3.3.3 the less deep minimum is
called the metastable state. These states have less resilience than the globally
stable state (figure 6.11).
This suggests a distinction between perturbations and interventions. With in-
terventions, we change the equilibrium landscape to allow a sustainable change
to a healthy state. Perturbations (a brief intervention or a positive or nega-
tive event) can have a permanent or temporary effect, depending on which
state is more resilient. In the situation shown in the top panel of figure 6.11,
any perturbation, whether it is a treatment or an alternative (or even being
on the waiting list), will work. In the situation on the bottom left, no in-
tervention would have a lasting effect. This analysis of resilience may help
to understand the inconsistent results of studies of intervention effects. Mon-
itoring the resilience of the unhealthy state (with catastrophe flags such as
anomalous variance) may also be important for timing interventions (Hayes
and Andrews 2020). Failed interventions, such as an attempt to quit smoking,
are likely to reinforce the unhealthy state (Vangeli et al. 2011).

6.3.3 Ising attitude model

The network approach has been applied to many other domains outside of
intelligence research and the study of psychopathology. Examples include
emotion (Lange and Zickfeld 2021; Treur 2019), personality (Costantini et
al. 2015; Cramer et al. 2012), interest (Sachisthal et al. 2019), deviations
of rational choice (Kruis et al. 2020), and organizational behavior (Lowery,
Clark, and Carter 2021). One area where it has been developed into a new
theory is attitude research.

143
Figure 6.11: Resilience from a complex-systems perspective. If the system is
in a less resilient, metastable state, any perturbation will be effec-
tive. A perturbation to a metastable state will not last. Lasting
interventions change the dynamic landscape of the system.

People have many attitudes—about food, politics, other people, horror movies,
the police, etc. They help us make decisions and guide our behavior. Attitudes
can be very stable and multifaceted, but they can also be inconsistent and
inconsequential. Social psychology has studied attitudes for a long time, and
many insights and theories have been developed.
Attitudes are complex constructs.
The formalization of attitude theories has been dominated by the connection- Typical phenomena, such as cognitive
ist account (Monroe and Read 2008; Van Overwalle and Siebler 2005). In dissonance, imbalance, ambivalence,
connectionist models, developed in the parallel distributed processing (PDP) and political polarization, can be well
framework, attitude units (e.g., beliefs) form a connected network whose acti- described by a network model.
vations (usually between −1 and 1) are updated based on the weighted sum
of internal inputs from other units and an external input. These weights or
connections are updated according to either the delta rule (a supervised learn-
ing rule based on the difference between the produced and expected output
of the network) or the Hebb rule. With this setup, these models can explain
a number of phenomena in attitude research. Another network account has
been put forward in sociology (DellaPosta 2020).
In this section, I will discuss our network approach to attitudes using the
Ising model, which was developed in a series of recent papers. The advantages
of this model over the connectionist PDP models are that it is derived from
basic assumptions, is better understood mathematically, is easy to simulate,
provides a psychological interpretation of the temperature parameter, and can
be fitted to data (Dalege et al. 2017).
The Ising model was developed as an alternative to the tripartite factor model
of attitudes, in which the attitude, a latent factor, consists of lower-order cog-
nitive, affective, and behavioral factors that each explain observed responses,

144
similar to the Cattell—Horn—Carroll model of general intelligence. The Attitudes are networks of feelings,
causal attitude model (Dalege et al. 2016) maintains this distinction in cog- beliefs, and behaviors toward an
attitude object.
nitive, affective, and behavioral components, but now conceptualizes them as
clusters within a network. Nodes represent single feelings, beliefs, and behav-
iors. In Dalege et al. (2018), this network model is formalized in the form of
an Ising model with attention as the equivalent of (the inverse of) tempera-
ture. That is, high attention “freezes” the network and leads to consistent and
stable positive or negative states of the attitude (the “mere thought effect”).
Attention is equated to (inverse)
temperature.

Figure 6.12: The Ising attitude network model. Feelings, beliefs, and behaviors
toward an attitude object align when attended to. Nodes are also
influenced by local fields, 𝑡𝑖 , and a global external field, 𝜏 . The
connections are weighted.

6.3.3.1 Model setup

The basic assumptions of the Ising attitude model are that nodes are binary
(e.g., one eats red meat or not), that nodes influence each other causally, and
that they have specific thresholds (as in the model for depression). An external
field (a campaign to eat less meat) could also affect the nodes. The alignment
of nodes to other nodes and to the external field depends on one’s attention,
𝐴, to the attitude object.
Given these simplifying assumptions, which can be relaxed in various ways,
we arrive at the random field Ising model (Fytas et al. 2018). This model is
not too different from the Ising model described in Chapter 5, section 5.2.1,
except that the first term now has two components, a general external effect
(𝜏 ) and an effect of node-specific (𝑡𝑖 ) thresholds (“I just really like the taste of
chicken”). The random field Ising attitude model can then be defined as:

𝑛
𝐻 (x) = − ∑𝑖 (𝜏 + 𝑡𝑖 )𝑥𝑖 − ∑<𝑖𝑗> 𝑊𝑖𝑗 𝑥 𝑥𝑗 , (6.4)
𝑖

exp(−𝐴𝐻(x))
𝑃 (X = x) = 𝑍
. (6.5)

145
Another difference from the original Ising model introduced in Chapter 5 is
that the interactions are now weighted and can even be negative. The main
technical problem is the same. To compute the probability of a state, one has
to compute 𝑍, which is ∑<x> 𝑒𝑥𝑝(−𝐴𝐻 (x)), that is, a sum over all possible
states (2𝑛 ). For large values of 𝑛, this is not feasible. One solution is to
take a random initial state and use Glauber dynamics to update the states
until an equilibrium state is reached. The Glauber algorithm does not require
𝑍. There are faster but less intuitive algorithms, the most popular being the
Metropolis—Hastings algorithm, which slightly modifies the Glauber dynam-
ics presented in Chapter 5, equation 5.3.
As discussed in Chapter 5 (section 5.2.1), another approach to understand-
ing the dynamics of Ising-type models is the mean-field approximation. This
requires the assumption that the network is fully and uniformly connected
with equal thresholds (known as the Curie—Weiss model). In this approxi-
mation 𝑊𝑖𝑗 = 𝑐 (all equal) and 𝑥𝑗 are replaced by their mean values, which
greatly simplifies the energy function. It can be shown that the dynamics
of the simple fully connected Ising model are well approximated by the cusp,
with the external field as normal and the inverse temperature as the splitting
variable.

Figure 6.13: The mean field approximation of the Ising attitude model is the
cusp. Attention is the psychological equivalent of inverse tem-
perature. Information varies from negative (contra) to positive
(pro).

This is an important result because it makes the use of the cusp in attitude
research (see figure 3.13) less phenomenological. The cusp is now derived
from more basic principles (figure 6.13). Note that here we use attention as
the splitting variable, whereas in Chapter 3 we used involvement. These are
closely related concepts, the difference being the time scale. Attention can
change in seconds or minutes, whereas involvement can change in weeks or
months. I will use attention and involvement interchangeably.
This mean-field approximation is very robust. In van der Maas, Dalege, and

146
Waldorp (2020), we show via simulation that networks with fewer connec-
tions and a distribution of weights, some of which are negative, are still well
described by the cusp. This can be easily checked with some R code or in
NetLogo. We will make use of the IsingSampler package in R.

6.3.3.2 Simulation

The IsingSampler function runs the Metropolis—Hastings algorithm 𝑛𝐼𝑡𝑒𝑟

times and returns the last state. It can return multiple final states for 𝑁 runs.
As input, it takes a matrix of links (W), which for the Curie—Weiss model
should be constant with 0s on the diagonal. The thresholds for each node
should be equal. Beta, originally the inverse of the temperature (1/𝑇 ), repre-
sents attention. The effect of varying beta (attention) is shown in figure 6.9.

library("IsingSampler")
n <- 10 # nodes
W <- matrix(.1, n, n); diag(W) <- 0
tau <- 0
N <- 1000 # replications
thresholds <- rep(tau, n)
layout(t(1:2))
data <- IsingSampler(N, W, nIter = 100, thresholds,
beta = .1, responses = c(-1, 1))
hist(apply(data, 1, sum), main = "beta = .1", xlab = 'sum of x')
data <- IsingSampler(N, W, nIter = 100, thresholds,
beta = 2, responses = c(-1, 1))
hist(apply(data, 1, sum), main = "beta = 2", xlab = 'sum of x')

Figure 6.14: The equilibrium distribution of the attitude values (sum of node
values) at low and high attention, respectively. This simple sim-
ulation demonstrates the mere thought effect (Tesser 1978).

In Dalege and van der Maas (2020), we simulated the difference between im-
plicit and explicit measures of attitude. The idea is that the individual thresh-

147
olds contain information about the attitude that can only be detected when
attention is moderately low. When attention is too high, the alignment be-
tween the nodes dominates the thresholds (figure 6.15). Indeed, in implicit
(indirect) measures of attitude, attention is much lower than in explicit mea-
sures such as an interview. This can be simulated as follows:

layout(1)
N <- 400; n <- 10
W <- matrix(.1, n, n); diag(W) <- 0
thresholds <- sample(c(-.2, .2), n,
replace = TRUE) # a random pattern of thresholds
dat <- numeric(0)
beta.range <- seq(0, 3, by = .05)
for(beta in beta.range){
data <- IsingSampler(N, W, nIter = 100, thresholds,
beta = beta, responses = c(-1, 1))
dat <- c(dat, sum(thresholds * apply(data,2,sum))) # measure of alignment
}
plot(beta.range, dat, xlab = 'beta', ylab = 'alignment with thresholds',
bty = 'n')

Figure 6.15: At low levels of attention (but not too low), the node values
are determined by the thresholds. At higher levels of attention,
they are overridden by the collective effect of other nodes. This
may explain the difference between implicit and explicit attitude
measures.

We see that for medium attention, the agreement with the thresholds is highest.
When attention is 0 or very low, nodes behave randomly and do not correlate
with the thresholds. When attention is very high, the effects of node-specific
thresholds are masked by the collective effects of other nodes. The principal
problem of implicit measurement is that for low to medium attention, the
network is quite noisy and measurement reliability will be low. This is why
this paper is called “Accurate by Being Noisy.”

148
6.3.3.3 Learning

The connectionist attitude models are capable of “learning,” that is, adjusting
the weights. This can also be done in the Ising attitude model by using
Hebbian learning. Hebbian learning, or “what fires together, wires together,”
can be formulated as:

�𝑊𝑖𝑗 = 𝜖 (1 − ∣𝑊𝑖𝑗 ∣) 𝑥𝑖 𝑥𝑗 − 𝜆𝑊𝑖𝑗 , (6.6)

which defines the change in weights. Weights will grow to 1 if the nodes they
connect are consistently either both 1 or both −1. If they consistently differ in
value, the weight grows to −1. If the nodes behave inconsistently, the weight
shrinks to 0, due to the last term.
In R, this can be implemented as follows:

library(qgraph)
hamiltonian <- function(x, n, t, w){
-sum(t * x) - sum(w * x %*% t(x)/2)
}
glauber_step <- function(x, n, t, w, beta){
i <- sample(1:n, size = 1) # take a random node
# construct new state with flipped node:
x_new <- x; x_new[i] <- x_new[i] * -1
# update probability
p <- 1/(1 + exp(beta * (hamiltonian(x_new, n, t, w) -
hamiltonian(x, n, t, w))))
if(runif(1) < p) x <- x_new # update state
return(x)
}
layout(t(1:2))
epsilon <- .002; lambda <- .002 # low values = slow time scale
n <- 10; W <- matrix(rnorm(n^2, .0, .4), n, n)
W <- (W + t(W)) / 2 # make symmetric
diag(W) <- 0
qgraph(W); title('before learning')
thresholds <- rep(.2, n)
x <- sample(c(-1, 1), n, replace = TRUE)
for(i in 1:500){
x <- glauber_step(x, n, thresholds, W, beta = 2)
# Hebbian learning:
W <- W + epsilon * (1 - abs(W)) * outer(x, x, "*") - lambda * W
diag(W) <- 0
}
# label switching (scale all nodes to positive):
W <- x * t(x * W); x <- x * x
qgraph(W); title('after learning')

Figure 6.16 shows the result of this simulation.

In this case we want to update the nodes values using the Glauber dynamics
(equation 5.3), which use the computation of the energy of a particular state.

149
Figure 6.16: Through Hebbian learning, a random (unbalanced) network be-
comes balanced.

Both functions (glauber_step() and hamiltonian()) are added to the R code.

Due to Hebbian learning, a network evolves from an unbalanced network (ran-
dom connections) to a consistently balanced network. Without learning, we Developing strong, balanced
need high attention to make the attitude network behave consistently. In the attitudinal networks has a clear
advantage: they do not require much
learning Ising attitude model (LIAM), weights increase during periods of high attention to be consistent.
attention. The advantage is that in later instances, less attention is required
for consistent network behavior (Smal, Dalege, and Maas submitted). In this
way we can develop stable attitudes that do not require much attention to be
consistent.

6.3.3.4 The stability of attitudes and entropy measures

To quantify the consistency of attitudes, we can compute the Gibbs entropy

(proposition I.2 in Dalege et al. 2018). The Boltzmann entropy was defined
in section 5.2.1 as the log of the number of ways (𝑊 ) a particular macrostate
can be realized. It measures the inconsistency of a particular attitude state
(proposition I.1 in Dalege et al. 2018).4 Gibbs entropy is more general in that
it does not assume that each microstate is equally probable. It is defined Gibbs entropy describes the
as: probability distribution over the
different microstates x.

− ∑<x> 𝑃 (x) ln 𝑃 (x). (6.7)

Note that we sum over all microstates (2𝑛 ). For small networks, this mea-
sure can be computed using the IsingEntrophy() function of the IsingSampler
package. There is much more to say about the different entropy measures.
For instance, Shannon entropy (a measure in information theory) and Gibbs

4
Assuming that all the attitude states (items) are re-encoded as positive (or negative) valued
items.

150
entropy have the same mathematical definition but are derived from com-
pletely different lines of reasoning in different fields of science. An introduc-
tion to the discussion on entropy measures can be found at the Entropy page
of Wikipedia.

6.3.3.5 Tricriticality

A new direction of research concerns Ising-type models with trichotomic node

values (—1, 0, 1). In physics this case is known as the tricritical Ising model
or the Blume—Capel model (Saul, Wortis, and Stauffer 1974). In physics, the
states −1 and 1 could represent the spin of a particle pointing up or down,
while 0 could represent a nonmagnetic or spinless state. In an attitude model,
−1 and 1 may represent pro and con beliefs, while 0 represents a neutral belief.
The Hamiltonian of the model includes a penalty for the −1 and 1 states:

𝑛 𝑛
𝐻 (x) = − ∑𝑖 𝜏 𝑥𝑖 − ∑<𝑖,𝑗> 𝑥𝑖 𝑥𝑗 + 𝐷 ∑𝑖 𝑥𝑖 2 . (6.8)

You can compare this to equation 5.1. The last term penalizes (increases the
energy) of the −1 and 1 states relative to the 0 state.
The dynamics of this model are more complicated. It resembles the butterfly
catastrophe (Dattagupta 1981), which has a tricritical point. The potential
6
function, 𝑉 (𝑋) = −𝑎𝑋 − 12 𝑏𝑋 2 − 13 𝑐𝑋 3 − 14 𝑑𝑋 4 + 16 𝑋 , has three stable fixed
points for specific combinations of values of parameters (see section 3.3.5 and
the exercise about the butterfly catastrophe in that chapter). This is relevant
to the modeling of attitudes because it opens up the possibility of involved
stable in-between attitude positions (see figure 6.17). The Ising attitude model
excludes this. In the Ising attitude model highly
involved persons always radicalize, but
more advanced spin models allow for
involved nonpartisan positions.

Figure 6.17: The butterfly catastrophe associated with the tricritical Ising
model. The potential function can have three minima. In the
figure 𝑎 = 𝑐 = 0, 𝑑 = 5, 𝑎𝑛𝑑𝑏 varies from 1 to −7.

6.4 Psychometric network techniques

So far, we have seen examples of theoretical psychological network models in

the fields of cognitive, clinical, and social psychology. However, much of the
popularity of such models is due to the psychometric approach developed to
analyze data using networks. In the last fifteen years, a family of statistical
approaches has been developed for all kinds of data and empirical settings.
Our psychosystems group (psychosystems.org) has published a book called
Network Psychometrics with R: A Guide for Behavioral and Social Scientists

151
(Isvoranu et al. 2022). This resource is highly recommended. I will limit
myself to a brief overview and some practical examples related to the models
presented in the first part of this chapter.

6.4.1 Main techniques

An important aspect of Network Psychometrics is visualizing the network—the

process of creating visual representations of the network structure. This helps
in interpreting the data. The main R packages for visualization are igraph
and qgraph. Both packages include many useful functions.
A more advanced application of network psychometrics is network estimation.
The most widely used methods for network estimation are Gaussian graph- Network estimation uses statistical
ical models (packages bgms, BDgraph, ggm, psychonetrics, qgraph, BGGM, methods to estimate the
networkstructure, including which
huge), partial correlation networks (qgraph, qgraphicalmodels), and Ising mod- nodes are connected to each other and
els (IsingFit, IsingSampler, rbinnet). The mgm package can be used to fit the strength of those connections.
mixed graphical models with a mixture of categorical and continuous valued
nodes. The bgms package applies Bayesian estimation and allows testing for
missing links. The huge package is used to represent the conditional depen-
dence structure among many variables and is particularly useful when the
number of variables is much larger than the sample size.
The estimation is usually followed by a centrality analysis. The most impor-
tant nodes in the network are identified based on their degree of centrality,
which measures the extent to which a node is connected to other nodes in
the network. Centrality measures include degree centrality, betweenness cen-
trality, bridge centrality, and eigenvector centrality, among others (packages
psych, networktools).
Another important step is network comparison. This can be done using tech- Network comparison is the process of
niques such as the network permutation test, bootstrapping, and moderation comparing the structure of two or
more networks to determine if they are
analysis (R packages bootnet and NetworkComparisonTest). significantly different from each other.
We can perform network inference, inferring causal relationships between
nodes, when we have time-series data or when we have intervened in the
network. Depending on the type of time series (𝑁 = 1 time series, 𝑁 > 1
time series, panel data), different modeling options and packages are available
(packages psychonetrics, mgm, graphicalVar, mlVar). GVAR returns a
temporal network, which is a directed network of temporal relationships, and
a contemporaneous network, which is an undirected network of associations
between the variables within the same time frame after controlling for
temporal relationships. Time-series network inference involves
analyzing sequential data to identify
For a detailed discussion of the reasons for using certain techniques, I again patterns of interactions and
refer you to our recent book. The brief overview I have provided here may soon dependencies among variables.
be obsolete. The CRAN Task View: Psychometric Models and Methods will
give you an up-to-date overview.5 Another option is to use JASP.6 JASP is a
free, open-source statistical analysis program developed under the supervision

5
https://cran.r-project.org/web/views/Psychometrics.htm
6
https://jasp-stats.org

152
of Eric Jan Wagenmakers (Huth et al. 2023; Love et al. 2019). It is a user-
friendly interface for accessing R packages. Many of the network R packages All major statistical analyses, both
mentioned above are available in JASP.7 frequentist and Bayesian, are available
in JASP.
Finally, I mention semantic network analysis again. A review of statistical
approaches (available in R) is provided by Christensen and Kenett (2021).

6.4.2 Fitting the mutualism model

In section 6.3.1.2, I provided code to simulate data. These data can be fitted
using JASP. By rerunning the previous code and adding

write.table(file='mutualism.txt', data, row.names = FALSE, sep='\t')

we have a data file ready to analyze in JASP.

After opening this file in JASP, you will see the data. In the Network (Fre-
quentist) tab, select all variables and the EBICglasso option. EBICglasso is
an R function from the qgraph package. It calculates the Gaussian graphical
model and applies the LASSO regularization to shrink the estimates of links
to 0 (Friedman, Hastie, and Tibshirani 2008). This prevents the presence of
many irrelevant links without losing predictive value. Alternatively, one could
use significant testing or a Bayesian procedure. In JASP, one could use the
partial correlation option. I recommended playing around with some options
and inspecting additional plots.
Mutualism is an alternative explanation for the positive manifold, which means
that the fit of a factor model to such data does not prove that the 𝑔 factor
theory is correct. It can be shown (van der Maas et al. 2006) that the simple
mutualism factor model with all 𝑀𝑖𝑗 = 𝑐, is equivalent to the one-factor model.
This can be tested in JASP by fitting an exploratory one-factor model to the
simulated data. The factor loadings should all be very similar.
As discussed, cross-sectional networks do not provide information about the
direction of effects. We can illustrate this as follows.

M[,1] <- .2 # strong influence of X1 on all others

M[2,] <- .2 # strong influence on X2 by all others
M[diag(nr_var) == 1] <- 0 # set diagonal of m to 0

If we rerun the code and create a centrality plot in JASP, we will see the risks
of centrality analysis in cross-sectional networks. Node 2 is the most central,
but we know from the simulation that this is because it is influenced by all
the others. The node with the most causal power (node 1) does not turn out
to be an important central node. With time-series data, we can estimate the
direction of the effects. We do this in R:

library("graphicalVAR")
# make time series for one persons with some stochastic effects
data <- run(odes = mutualism, tmax = 1000, table = TRUE,

7
You may want to start by reading the blog post on doing network analysis in JASP
(https://jasp-stats.org/2018/03/20/perform-network-analysis-jasp/).

153
 timeplot = (i == 1), legend = FALSE,
after = "state<-state+rnorm(nr_var,mean=0,sd=1);
state[state<0]=.1")
data <- data[,-1]
colnames(data) <- vars <- paste('X', 1:nr_var, sep = '', col = '')
fit <- graphicalVAR(data[50:1000,], vars = vars, gamma = 0, nLambda = 5)
plot(fit, "PDC")
centralityPlot(fit$PDC)

The results are shown in figure 6.18. Only the time-series approach provides
useful information about possible causal effects.

Figure 6.18: The top figure is based on cross-sectional data and incorrectly
suggests that node V2 is the most important node. The bottom
figure is based on the time series of one individual and correctly
shows that V2 is central because it is influenced by other nodes,
while V1 is central because it influences other nodes and is there-
fore more important.

The M-matrix can take different forms. The typical multifactor structure can
be achieved with a block structure.

set.seed(1)
factors <- 3
M <- matrix(0, nr_var, nr_var)
low <- .0; high <- .1 # interaction between and within factors
# loop to create M
cat <- cut(1:nr_var, factors)
for(i in 1:nr_var) {
for (j in 1:nr_var) {
if (cat[i] == cat[j])
M[i, j] <- high else M[i, j] <- low
}

154
 }
M[diag(nr_var) == 1] <- 0 # set diagonal of m to 0

In JASP, you can perform network and confirmatory factor analysis. In the
latter case, select “3 factors” in the first window and select “assume uncor-
related factors” in the model options. The resulting plots should look like
figure 6.19.

Figure 6.19: The block structure in the mutualism model can be represented
as either a network or a factor model.

6.4.3 Fitting Ising models

With IsingFit we can easily fit cross-sectional data generated with the Ising
attitude model. Figure 6.20 shows a good fit of the model. The code for this
analysis is:

library("IsingSampler"); library("IsingFit")
set.seed(1)
n <- 8
W <- matrix(runif(n^2, 0, 1), n, n); # random positive matrix
W <- W * matrix(sample(0:1, n^2, prob = c(.8, .2),
replace = TRUE), n, n) # delete 80% of nodes
W <- pmax(W, t(W)) # make symmetric
diag(W) <- 0
ndata <- 5000
thresholds <- rnorm(n, -1, .5)
data <- IsingSampler(ndata, W, thresholds, beta = 1)
fit <- IsingFit(data, family = 'binomial', plot = FALSE)
layout(t(1:3))
qgraph(W, fade = FALSE); title("Original network", cex.main = 2)
qgraph(fit$weiadj, fade = FALSE); title("Estimated network", cex.main = 2)
plot(thresholds, type = 'p', bty = 'n', xlab = 'node',
ylab = 'Threshold', cex = 2, cex.lab = 1.5)
lines(fit[[2]], lwd = 2)

An empirical example is provided in Dalege et al. (2017). The open-access

data (𝑁 = 5728) come from the American National Election Study of 2012

155
Figure 6.20: The true (original) and the estimated Ising model are in good
agreement. The thresholds are also well estimated from the data.

on evaluative reactions toward Barack Obama. The items and abbreviations

are:

Items tapping beliefs Abbreviation

“Is moral” Mor
“Would provide strong leadership” Led
“Really cares about people like you” Car
“Is knowledgeable” Kno
“Is intelligent” Int
“Is honest” Hns
Items tapping feelings
“Angry” Ang
“Hopeful” Hop
“Afraid of him” Afr
“Proud” Prd

Table 6.1: The abbreviation of items used in figure 6.21.

We can use Isingfit and add community detection:

Obama <- read.table("data/Obama.txt", header = TRUE) # see book data folder

ObamaFit <- IsingFit(Obama, plot = FALSE)
ObamaiGraph<- graph_from_adjacency_matrix(abs(ObamaFit$weiadj),
'undirected', weighted = TRUE, add.colnames = FALSE)
ObamaCom <- cluster_walktrap(ObamaiGraph)
qgraph(ObamaFit$weiadj, layout = 'spring',
cut = .8, groups = communities(ObamaCom), legend = FALSE)

Figure 6.21 shows the network. The red nodes represent negative feelings to-
ward Barack Obama; the green nodes represent positive feelings; the light blue

156
nodes represent judgments primarily related to interpersonal warmth; and the
purple nodes represent judgments related to Obama’s competence. This com-
munity structure is consistent with the postulate of the Ising attitude model
that similar evaluative responses cluster (Dalege et al. 2016). Finnemann et
al. (2021) present additional examples and applications of other packages.

Figure 6.21: The attitude toward Obama in 2012 represented as a network.

6.5 Challenges

A large amount of work has been done since the early work on network psy-
chology, the mutualism model, and the paper on the network perspective on
comorbidity. In particular, network psychometrics has taken off in an unprece-
dented way. One could say that modern psychometrics is being reinvented
from a network perspective. For every type of data and research question,
a network approach seems to be available. For example, there are R pack-
ages for meta-analysis from a network perspective (Salanti et al. 2014). I
also note that much has been done to understand the relationship between
network psychometrics and more traditional techniques such as item response
theory (Marsman et al. 2018), factor models (Waldorp and Marsman 2022),
and structural equation modeling (Epskamp, Rhemtulla, and Borsboom 2017).
Nevertheless, there are still many challenges for both psychological network
modeling and network psychometrics.

6.5.1 Psychological network modeling

Despite all the work on this, I can only conclude that this theoretical line of
research is still in its infancy. The strength of the application to intelligence is
that it provides an alternative to the 𝑔 factor model, which is also nothing more
than a sketch of a theory. The extensions of the mutualism model (de Ron et
al. 2023; Savi et al. 2019) add new steps, but remain rather limited models.
One reason for this state of affairs is that it is really hard to pinpoint the

157
elementary processes involved in intelligence, and indeed in any psychological
system.
This is perhaps less of a problem in the factor account because the indicators
are interchangeable in a reflective factor model. Once one has a sufficiently
broad set of indicators, the common cause estimate will be robust. In a forma-
tive model, each indicator contributes a specific meaning to the index variable.
However, this is not a reason to prefer the common cause model (van der Maas,
Kan, and Borsboom 2014).
The other modeling examples suffer from the same problem. In the clinical
psychology models, we define the nodes either as the symptoms specified in
the Diagnostic and Statistical Manual of Mental Disorders (DSM) or as the
questions asked in interviews or questionnaires, with the advantage that we
then have data to fit the model. But, again, we have no real way of knowing
the elementary processes in clinical disorders and their interactions (Fried and
Cramer 2017). If we miss important elementary
nodes, this may seriously affect the
A way out has been mentioned in the context of the Ising attitude model, using validity of our models and
the mean-field approximation. If we are only interested in the global behavior psychometric network analyses.
of the attitude (hysteresis, divergence), we can ignore the specification of the
nodes (another interchangeable argument). But if one wants to intervene on
specific nodes or links of a clinically depressed person, this is not sufficient. In mean-field analyses, the
specification of all nodes is less
Another critical point is that these models increase our understanding of psy- important.
chological phenomena, but seemingly not our ability to predict or intervene.
For example, the Ising attitude model helps us understand the role of in-
volvement (attention) in the dynamics of attitudes. If this factor is too high,
persuasion will be extremely difficult due to hysteresis. Anyone who has ever
tried to argue with a conspiracy theorist knows what I mean. But too little
attention is also a problem. In the model, these are people who are sensitive
to the external field, for example, you tell them to clean their room, but as
soon as you leave, the attitude falls back into random fluctuations. The mes-
sage gets through but does not stick. I find this insightful, but I must admit
that it does not provide us with interventions. We don’t know how to control
attention or engagement, although more work can and will be done on this.
For intelligence, the model suggests that the active establishment of near and
far transfer might be effective. A disappointing lesson from developmental
psychology is that transfer does not always occur automatically (e.g., Sala et
al. 2019). However, strategies for improving transfer do exist (Barnett and
Ceci 2002) and, according to the mutualism model, should have a general
effect.
In Chapter 1, section 1.3, I mentioned the case of the shallow lake studied in
ecology, where catching the fish was a very effective intervention, while ad-
dressing the cause, pollution, was ineffective due to hysteresis. Ecologists now
know why this is so and have developed models to explain this phenomenon.
However, I did not mention that this intervention was suggested not by mod-
eling work but by owners of ponds who observed that ponds without fish
sometimes spontaneously tipped to the clear state. This is not an uncommon
path in science, and it may well occur in clinical psychology. The touted
extraordinary successes of electroshock therapy for severe depression or new
drugs (MDMA) for post-traumatic stress disorder could be our “fish.” But

158
these claims have also been criticized (e.g., Borsboom, Cramer, and Kalis
2019; Read and Moncrieff 2022).
Although much more progress can be made in network modeling of psycholog-
ical systems, it is advisable to be realistic. Progress in mathematical modeling
of ecosystems has also been slow. The formalization of psychological models is Ecosystems and human systems are
of interest for many reasons (Borsboom et al. 2021), but will only be effective devilishly complex.
if we also make progress in other areas, such as measurement (Chapter 8).

6.5.2 Psychometric network analysis

This approach is also not without its problems, some of which are related to
the problems of psychological network modeling. For example, the definition
of nodes and the risk of missing nodes in the data is a serious threat. This prob-
lem is not unique to network analysis; simple regression analysis suffers from
the same risks. Another common threat to many applications of psychometric
network analysis is the reliance on self-report in interviews or questionnaires.
Generalizability, which may depend more on the choice of sample and mea-
surement method than on the statistical analysis itself, is another example of a
common problem in psychology in general, and psychometric network analysis
in particular.
Psychometric network analysis has been criticized because the results are dif-
ficult to replicate (Forbes et al. 2017). Replication of advanced statistical
analyses, whether structural equation modeling, fMRI, or network psychomet-
rics, is always an issue, but some solutions have been developed (Borsboom et
al. 2017, 2018; Burger et al. 2022). For network analysis, a number of
safeguards have been developed to
A final important issue concerns causality. Network models estimated from increase replicability
cross-sectional data are descriptive rather than causal; that is, they do not
provide information about the direction of causal relationships between vari-
ables. Developing methods for inferring causality from network models is an
important challenge in the field.
The move to time-series data (either 𝑁 = 1, 𝑁 > 1, or panel data) partially
solves this problem (Molenaar 2004). With time-series data, we can establish
Granger causality, a weaker form of causality based on the predictive power of
one variable over another in a time-series analysis. However, the relationship Granger causality suggests a
may be spurious, influenced by other variables. Network analysis on time directional influence when one time
series predicts another, without
series often requires a lot of reliable and stationary data. An important issue necessarily implying a true causal
is the sampling rate of the time series. In general, to accurately estimate a relationship.
continuous time-varying signal, it is necessary to sample at twice the maximum
frequency of the signal. This is called the Nyquist rate. Another issue is the
assumption of equidistance between time points (Epskamp et al. 2018), which
can be circumvented by using continuous-time models (Voelkle et al. 2012).
While these problems are not unique to network psychometrics, they are se-
rious problems in practice (see Hamaker et al. 2015; Ryan, Bringmann, and
Schuurman 2022). The ultimate test of causality requires intervention. For Hopefully, the combination of
new work along this line, I refer to Dablander and van Bork (2021) and Kos- observational and experimental data
can provide sufficient information to
sakowski, Waldorp, and van der Maas (2021). properly estimate causal relationships
Perhaps we can also think of other ways. A simple, but difficult to imple- in directed acyclic graphs.
ment, procedure is to ask subjects about the links in their networks. If one

159
claims not to eat meat because of its effect on the climate, one might consider
adding a directed link to this individual’s network (Rosencrans, Zoellner, and
Feeny 2021). Deserno et al. (2020) used clinicians’ perceptions of causal rela-
tionships in autism. These relationships were consistent with those found in
self-reported client data. The main problem is that there are many more pos-
sible links than nodes to report on, which makes the questionnaires extremely
long and tedious to fill out. Brandt (2022) applied conceptual similarity judg-
ments to construct attitude networks. Alternatively, one could try to estimate
links from social media data, interviews, or essays using automatic techniques
(Peters, Zörgő, and van der Maas 2022).

6.6 Conclusion

Despite these critical remarks, it is safe to say that the psychological network
approach has made great progress in a very short time. The mutualism paper
was only published in 2006. It is a new and unique application of the complex-
systems approach within the field of psychology. For me, the common-cause
approach is theoretically unsatisfactory because the common causes are purely
hypothetical constructs. For both the 𝑔 and 𝑝 factors, there is no reasonable
explanation for their origin. In contrast, the reciprocal interactions between
function and symptoms that underlie the network theories of intelligence and
psychopathology are hardly controversial. It is also good to note that the
network approach is not inconsistent with statistical factor analytic work in
psychology. It is a matter of interpreting the general factor as a common cause
or as an index. The formative interpretation of factors is consistent with the
network approach.
In the next chapter, we take the step to modeling social interactions using
social networks, where the nodes represent agents that move to new locations,
learn a language, share cultural attributes, and have opinions. The focus is on
opinion networks. The chapter ends with our own agent-based model of opin-
ion dynamics, which builds on the Ising attitude network model. This opinion
model is a social network of within person attitude networks. To simplify
we will use the cusp description of the Ising attitude network model at the
within-person level. In essence, the model is based on interacting cusps, simi-
lar to the model for multifigure multistable perception introduced in Chapter
4 (section 4.3.8). This opinion model suggests a new explanation and a new
remedy for polarization.

6.7 Exercises

1) Reproduce the degree distribution of the Barabási—Albert model shown

on the scale-free network Wikipedia page. Use sample_pa from the
igraph library. (*)
2) Open and run the “Preferential Attachment” model in NetLogo. Replace
the line report [one-of both-ends] of one-of links with report one-of turtles.
New nodes will now connect to a random node. Does this result in a
random network? (*)

160
 3) Make a hysteresis plot in the “Vulnerability to Depression” model in
NetLogo (see section 6.3.2). (*)
4) In the “Vulnerability to Depression” model you can deactivate all symp-
toms at once with the administer-shock button. It is as if you give the
network an electric shock that resets all the symptoms. Try to find
a settings of the connection-strength and external-activation that creates
a disordered network (above the black line in the network status plot)
whereby administering a shock makes the system healthy again. Is this
healthy state long-term stable? (*)
5) Compute the Gibbs entropy for the Learning Ising Attitude model during
the learning process (see section 6.3.3.4). Show in a plot that learning
minimizes the Gibbs entropy. (**)
6) Install and open JASP (jasp-stats.org). Open the data library: “6. Fac-
tor.” Read all the output and add a confirmatory factor analysis. What
is the standardized factor loading of Residual Pitch in the confirmatory
one-factor model? (*)
7) Read the blog “How to Perform a Network Analysis in JASP”8 Repro-
duce the top plots of figure 6.18. Generate the data using the R code
in section 6.4.2, import the data into JASP, and perform the network
analysis. (*)
8) Study the R code for the case where the M-matrix consists of three
blocks (section 6.4.2). Generate the data and import into JASP. Apply
exploratory factor analysis, check the fit for 1 to 3 factors, and report
the 𝑝-values. Fit the confirmatory 3-factor model. Does it fit? Add V1
to the second instead of the first factor. How do you see the misfit? (*)
9) How can you generate data for a higher-order factor model using the mu-
tualism model? What should be changed in the code of the M-matrix for
the case of three blocks? Show that the three-factor solution (assuming
uncorrelated factors) does not fit the resulting data. Fit a higher-order
factor model and report the 𝑝-value of the goodness of fit. How does the
network plot change? (**)
10) Generate data for a network in a cycle (v1 -> v2 -> v3…v12 -> v1).
Fit a network and an exploratory factor model. Does this work? What
does this tell us about the relationship between the class of all network
models and all factor models? (**)
11) Fit a Bayesian network in JASP to the data generated for figure 6.20.
Warning: the GM in JASP expects (0,1) data. Check that only the
simulated links have high Bayes factors. (*)

8
https://jasp-stats.org/2018/03/20/perform-network-analysis-jasp/

161
7 Sociophysics

7.1 Introduction

This final chapter is dedicated to the dynamics between people. Networks

of social interactions undoubtedly meet the criteria of a complex system and
often exhibit unpredictability, sudden changes, and self-organization. Models
of the dynamics between people are the domain of social scientists, such as
sociologists, but also scientists with a background in statistical physics. Again,
it is necessary to simplify, and the victim in this case is psychology. It is a In most sociophysics models, people
challenge to add a little more psychology to the agents in these models so that are reduced to binary nodes, on or off,
pro or contra, or to systems with only
more phenomena can be explained without making these models too complex one continuous attribute.
to study. It is very easy to cross that line. The second part of the chapter
deals with what I call psychosocial models.
I use the somewhat controversial term sociophysics as a label for this very
diverse field of research, which has roots in many different sciences. It is as
(in)correct as the generally accepted term psychophysics in psychology. What
I like about it is that it emphasizes the intention to arrive at a formalized,
mathematically stated theory of social processes, often based on formalisms
first developed in physics. Alternative labels are social physics, computational
social science, and agent-based modeling (Goldstone and Janssen 2005).
In my broad definition of sociophysics, it deals with many different systems
and processes. Examples that I have already discussed are the synchronization
of movement, such as in the swinging of legs in a two-person system or in a
crowd of people fleeing a fire in a building. Granovetter’s threshold model
of joining a dancing crowd is another example. Many papers deal with the
spread of opinions, fashions, rumors, etc. This will be the main topic of this
chapter. But this field also includes work on segregation, cooperation, crime,
economic systems, and much more. I will present some key examples and then
turn to models of the social dynamics of opinions.
Some important follow-up books to read are Smaldino (2023) on modeling so-
cial behavior; Bowles and Gintis (2011) on cooperation and altruism; Durlauf
and Young (2001) on modeling social economics; Epstein and Axtell (1996),
Epstein (2006), and Epstein (2014) on building artificial societies; and Miller
and Page (2007) on computation models of social life. There are also some
excellent review articles on this area of research (Castellano, Fortunato, and
Loreto 2009; Noorazar 2020; Proskurnikov and Tempo 2017; Jusup et al. 2022).
A less mathematically focused review is provided by Flache et al. (2017).

162
7.2 Some famous examples

7.2.1 Segregation

The most famous computational model for studying segregation is Schelling’s

model (Schelling 1971). The simplified version of this model assumes a two-
dimensional space, a cellular automaton as in NetLogo, where each location
or cell is occupied by one of two types of agents or is empty. The density
parameter determines how many locations are occupied. Agents stay at their
location if a certain percentage of their eight neighbors, 𝐵, are of the same
type as the agent. So, if 𝐵 = 0%, nobody moves. For values of B close to
100%, everybody moves all the time. What Schelling showed with his model
was that even low levels of intolerance, near 30%, lead to segregation. Even if people can accept up to 70%
of “foreign” neighbors, segregation
The NetLogo model Segregation (Social Science sample model) demonstrates still occurs.
this model. Especially at high densities, transitions between the unsegregated
state, the segregated state, and a mixed state where agents keep moving can
be seen (Gauvin, Vannimenus, and Nadal 2009). Instead of playing around
with the sliders, we can also use the BehavioralSpace option. Figure 7.1 shows
the settings for this option, the R code, and the results. Note that without
visualizations NetLogo is pretty fast.
One concept we saw in Chapter 5, section 5.4.4, in the context of Haken’s work,
was that of the order parameter. The order parameter literally measures the
order in the system. We look for an order parameter that captures the quali- Finding an appropriate order
tative phenomena in the model, such as sudden changes and hysteresis. In the parameter is not always easy, and
multiple definitions are often possible.
NetLogo simulation, I used the percentage of similar neighbors. Gauvin, Van-
nimenus, and Nadal (2009) propose two, probably better, order parameters:
a segregation coefficient, which requires the identification of clusters, and the
density of unwanted locations, which better distinguishes between the three
states of the system.
Clearly, this is only an initial model that can be extended in many ways, some
of which are already explored in Schelling’s original paper. For an interest-
ing historical sketch of this line of research, see Hegselmann (2017). Another
NetLogo model inspired by Schelling’s research is called Party. In this model,
partygoers experience discomfort and change their groups if their current group
consists of a disproportionate number of individuals of the opposite sex. A
recent review with a focus on urban dynamics can be found in Jusup et al.
(2022).

7.2.2 The evolution of language: The naming game

Languages are complex systems that emerge in a self-organizing cultural pro-

cess (Steels 1995). Steels developed the naming game (figure 7.2) to study the
evolution of language. For a review I refer to Chen and Lou (2019). Castellano,
Fortunato, and Loreto (2009) provide a brief review.
The game begins with blank lists for all agents. In the first round, the speakers
invent a new unique word, assuming that the possibilities are endless. We
assume that there is only one object to name and that the social network is
fully connected. Two agents are randomly selected from a population: one as

163
Figure 7.1: BehaviorSpace settings and R code to visualize the effects of intol-
erance on segregation.

Figure 7.2: The naming game. The speaker chooses the highlighted word. If
the listener does not know the word (failure), she adds it to her
inventory. In case of success, both agents delete their inventories,
keeping only the spoken word. (Adapted from Castellano et al.,
2009)

164
the “speaker” and the other as the “hearer.” The speaker chooses a word from
her vocabulary for the object. If the hearer is unfamiliar with the word, she
incorporates it into her lexicon. If she recognizes the word, both clear their
vocabularies, retaining only the word that was spoken.
The R code to simulate this process is:

resample <- function(x, ...) x[sample.int(length(x), ...)]

n <- 1000; iter <- 100000
x <- vector("list", n)
n_possible_words <- 1000000 # should be very high
total_words <- total_unique_words <- numeric(iter)
for(i in 1:iter){
j <- sample(1:n, 1) # speaker
k <- sample((1:n)[-i], 1) # hearer
# if speaker has no words, make one up (actually just number)
if(length(x[[j]]) == 0) x[[j]][1] <- sample(1:n_possible_words, 1) else
{
spoken_word <- resample(x[[j]], 1) # choose a word (actually just number)
if(any(x[[k]]%in%spoken_word)) # hearer knows the word
{
x[[j]] <- spoken_word # erase list except spoken_word
x[[k]] <- spoken_word # erase list except spoken_word
} else # hearer does not know the word
x[[k]] <- c(x[[k]], spoken_word) # add word to list
}
total_words[i] <- length(unlist(x))
total_unique_words[i] <- length(unique(unlist(x)))
}
layout(1:2)
plot(total_words, type = 'l', xlab = 'time', bty = 'n')
plot(total_unique_words, type = 'l', xlab = 'time', bty = 'n')
unique(unlist(x)) # winning word

Which results in figure 7.3.

After a phase in which agents use lots of different words, a language consisting
of just a single word emerges abruptly. A simplification would be the case
that agents start with either word A, word B, or both words A and B. So no
new words are invented. For this case we can write the time evolution in three
differential equations:

𝑑𝑋 𝐴 1
= −𝑋 𝐵 𝑋𝐴 + 𝑋 𝑋 + 𝑋𝐴 𝑋𝐴𝐵 ,
𝑑𝑡 2 𝐴𝐵 𝐴𝐵
𝑑𝑋 𝐵 1
= −𝑋 𝐴 𝑋𝐵 + 𝑋 𝑋 + 𝑋𝐵 𝑋𝐴𝐵 , (7.1)
𝑑𝑡 2 𝐴𝐵 𝐴𝐵
𝑑𝑋 𝐴𝐵
= 2𝑋 𝐴 𝑋𝐵 − 𝑋𝐴𝐵 𝑋𝐴𝐵 − (𝑋𝐴 + 𝑋𝐵 )𝑋𝐴𝐵 .
𝑑𝑡

The first equation can be understood as follows: Speakers B talking to listeners

A turn listeners A into AB agents. The loss in A agents is −𝑋 𝐵 𝑋𝐴 . AB
agents talking to AB agents become A or B agents, depending on the speaker’s
random choice of A or B (𝑋𝐴𝐵 𝑋𝐴𝐵 /2). If speaker A talks to an AB agent,

165
Figure 7.3: The order parameter, the total number of unique words, undergoes
a phase transition to a state where everyone uses the same word
for the object (compare Castellano et al., 2009, fig.2).

the latter becomes an A agent (𝑋𝐴 𝑋𝐴𝐵 ). These three terms together define
the change in 𝑋𝐴 . The other equation follow the same logic. It is easy to
implement this in Grind. There are three possible outcomes: (1) If the initial
proportion of A is greater than B, A wins; (2) if the initial proportion of B
is greater than A, B wins; and (3) if these proportions are exactly equal, all
three options A, B, and AB coexist, but this equilibrium is unstable. Although extremely simplified, the
naming game illustrates that
coexistence of languages is difficult in
7.2.3 Cultural dynamics: The Axelrod model a fully connected network.

Robert Axelrod (1997) introduced an influential model of cultural diffusion

based on the effects of social interactions and homophily. In Axelrod’s model, Homophily refers to the tendency or
individuals become more similar through interactions—but only if they already preference for individuals to associate
or connect with others who are similar
share some cultural features. to themselves.
In the model, agents have 𝐹 cultural features (e.g., beliefs, habits), each of
which has 𝑄 possible nominal values. Agents are organized in some kind of
network, for instance, a fully connected one. At each time step, two agents are
randomly selected. We count the number of shared features, such as features
with the same value. This number, divided by 𝐹 , gives the probability that
one of the mismatching features of one of the two agents is set equal to that
of the other. Thus, if they differ on all features, nothing happens. If 50% of
their features have the same value, then one of the mismatching features will
change for one of the agents with a probability of 0.5.
The combination of interaction and homophily creates a self-reinforcing dy-
namic that often leads to global convergence toward a single culture. However,
for certain choices of 𝐹 and 𝑄, the model converges to a state of diversity. In
the simulation below, I use the number of remaining cultures as a simple or-
der parameter. Castellano, Marsili, and Vespignani (2000) present detailed
analyses of the phase transitions in this model using more advanced order pa-
rameters. As you can imagine, this model can be extended in many ways, for
instance by introducing ordinal instead of nominal states (Macy, Flache, and
Takacs 2006).

166
The following R code of this model generates the first plot in figure 7.4 . You
can play with the values to see different cases.

resample <- function(x, ...) x[sample.int(length(x), ...)]

n <- 100; iter <- 50000
F <- 4 # features
Q <- 4 # nominal levels per feature
uniques <- numeric(iter)
x <- matrix(sample(1:Q, replace = TRUE, n * F), n, F)
for(i in 1:iter){
j <- sample(1:n, 1) # agent 1
k <- sample((1:n)[-j], 1) # agent 2
w <- sum(x[j,] == x[k,])/F # agreement
if(w < 1 & runif(1) < w) {
# which (unequal) feature to update:
f <- resample(which(x[j,] != x[k,]), 1)
x[j,f] <- x[k,f] # update
}
uniques[i] <- nrow(unique(x))
}
plot(uniques[1:i], type = 'l', lwd = 2, xlab = 'time',
ylab = '# unique cultures', bty = 'n', ylim = c(0, 80),
main = paste0('Axelrod model with F = ', F, ', Q = ', Q))

What I find appealing about the Axelrod model is its multidimensionality.

Many are more concerned about polarization in the United States than in As long as we have some
the Netherlands. In the United States, it seems that all aspects of life are characteristics or qualities in common
with other people on certain
interconnected, and even seemingly unrelated factors such as one’s choice of dimensions, there is hope for society.
jeans or favorite sport are correlated with one’s political beliefs. This greatly
limits the possibilities for depolarization.

7.3 Dynamics of opinions

In that same famous paper, Axelrod asked an important question: “If people
tend to become more alike in their beliefs, attitudes, and behavior when they
interact, why don’t all such differences eventually disappear?”
The answer to Axelrod’s question is usually posed in terms of limited interac-
tions between agents. In Axelrod’s model, for example, it was due to selective
interaction between agents. In continuous-opinion models, it is due to bounded
confidence, that is, agents that are too different refuse to interact. In some
models, there is simply no connection between subgroups in a network. Bounded confidence refers to the
concept that individuals are influenced
It is hard to count the number of opinion-spread models, but it could easily by the opinions of others only when
be in the hundreds, if you count all the variants. They all share a few building those opinions fall within a certain
blocks. First, there has to be some topology to the social network. Model- range of their own opinions.
ers make different choices here. Often, fully connected networks are assumed
because they allow an analytical (mean-field) approach; others use random net-
works, lattices, small-world networks, etc. The problem is that we don’t really
know how real social networks work, except that they are incredibly complex
(Newman and Park 2003). Second, you need to define some interaction rules.

167
Figure 7.4: Two runs of Axelrod’s model. For four features with four possible
values, only one culture remains. For the second case (𝐹 = 2, 𝑄 =
10), multiple unique cultures, different in all features, emerge.

168
For example, two agents might end up in the middle after a discussion, one
agent might copy the other’s state, or one agent might take over the majority
vote in its local neighborhood. Finally, you have to define opinion. I will A major division in social contagion
first discuss several discrete opinion models. models is whether opinions are defined
as discrete or continuous.

7.3.1 Discrete opinion models

7.3.1.1 Voter models

The simplest possible model seems to be the voter model. In its basic form,
with only two possible opinions (−1, 1), two connected agents A and B meet,
and A simply copies B’s opinion. What happens in this simple system depends
on the topology of the network, that is, its dimension (either 𝑑 = 1 (on a line),
𝑑 = 2 (a lattice) or 𝑑 > 2) and its size 𝑁 . In more than two dimensions and
with infinite size, the voter system does not converge, but in other cases it
converges to a state in which all opinions agree (either all −1 or 1) (Castellano,
Fortunato, and Loreto 2009; Redner 2019). How long it takes to converge The probability of ending up in the
can also be derived analytically. The convergence time is proportional to 𝑁 2 , +1 state is equal to the initial
probability of +1s.
for voters on a line (𝑑 = 1), 𝑁 𝑙𝑛𝑁 for 𝑑 = 2, and 𝑁 for 𝑑 > 2. Thus, the
convergence is slowest when agents are connected in a line.
In the heterogeneous voter model, each agent A copies the opinion of agent
B with some probability 𝑟𝑖 . In this way, one can study the effect of stubborn
voters (with low 𝑟𝑖 ). It turns out that the small group of stubborn individuals
(sometimes called zealots) can overcome the majority opinion. Many other
variants have been analyzed, such as adding memory and noise to the voters
(Castellano, Fortunato, and Loreto 2009). It is also possible to consider three
groups, left, center, and right, where left and right do not interact. In this
case, depending on the initial proportion, we end up in a state of full consensus
in one of the states or with a mixture of extremists without centrists (Redner
2019). Finally, the topology of the social network plays a role. In general, consensus is easier reached
in scale-free networks with broad
Another approach has been proposed by Martins (2008). The Continuous degree distributions.
Opinions and Discrete Actions (CODA) model combines discrete and contin-
uous aspects of opinion dynamics. Agents act discretely but update their
continuous opinions based on observations of other agents’ discrete actions.
In CODA, there are two choice options, A and B, and agent 𝑖 has some sub-
jective probability 𝑝𝑖 that A is the best option, and 1 − 𝑝𝑖 for B. The actual
choice is made according to 𝑠𝑔𝑛 (𝑝𝑖 − .5), so A is chosen when 𝑝𝑖 > .5. Next,
the agent observes other agents. Agents assume that other agents make ra-
tional choices, that is, choose A when A is the best option with a probability
𝑎 that is larger than .5. In running the model, it is convenient to work with
the log-odds of probabilities, 𝑣𝑖 = 𝑙𝑛(𝑝𝑖 /(1 − 𝑝𝑖 )). Using Bayes’s theorem, we
can update 𝑣𝑖 to 𝑣𝑖 + 𝑎 when agent j chooses A and to 𝑣𝑖 − 𝑎 if the choice is
B. Martins (2008) integrates these decision rules with the voter model, show-
ing extreme forms of polarization, that is, a strongly bimodal distribution of
opinions.

169
7.3.1.2 More discrete opinion models: Majority type models

In the voter and Axelrod models, interactions are limited to two agents. When
multiple neighbors have an impact on each agent, many new options arise. One
option is the Ising model (Galam, Gefen, and Shapir 1982). Agents switch
sides with a probability that depends on the states of their neighbors. The
temperature variable in the Ising model is translated into randomness in the
model. The external field is now interpreted as an external social field. In this
way, one can explain phase transitions and hysteresis in opinion dynamics.
Another deeply analyzed option is the majority model (Galam 2008; Redner
2019). Here, a random group of voters is selected, and all voters in this
group adopt the local majority opinion. This process can be repeated until
convergence to one opinion is reached (which will always happen in a finite
population). Galam (2008) sets up this process in a hierarchical fashion (see
figure 7.5). Alternatively, only one voter could be influenced by the majority
vote in its neighborhood. This corresponds to the Ising model with 0 temper-
ature.

Figure 7.5: Bottom-up voting in hierarchical systems in the Galam model.

Another case is the 𝑞-voter model (Castellano, Muñoz, and Pastor-Satorras

2009; Jędrzejewski and Sznajd-Weron 2019). Here, agents change their opin-
ions only if all 𝑞 voters selected from the neighborhood agree on the other
opinion. When 𝑞 = 1, this reduces to the standard voter model. The 𝑞-voter
model generally allows for a higher degree of opinion diversity compared to

170
the basic voter model. The 𝑞-voter model has been implemented in NetLogo
(“qvoter_WS” in the user community models).
In the basic Snzadj model, agents are placed on a line, and two neighbors
with the same opinion spread this opinion to their own neighbors. If they
disagree, they enforce their disagreement on their neighbor. Thus (?, 1, 1, ?)
becomes (1, 1, 1, 1) and (?, −1, 1, ?) becomes (1, −1, 1, −1). This can converge
to a state of all 1’s, all −1’s, or a sequence of 1 and −1 pairs. The latter state
is reached with a probability of .5. This model has also been extended in many
ways, such as adding an election process (Sznajd-Weron, Sznajd, and Weron
2021).

7.3.1.3 Social Impact theory

The last discrete model I mention here is the social impact model, which is
based on Bibb Latané’s (1981) psychological theory of social impact. Latané
introduced many ideas and concepts from complex-systems theory into social
psychology. His psychological theory is firmly grounded in social psychology
and supported by all kinds of evidence (Karau and Williams 1993).
In this theory, opinion change depends on social impact 𝐼. Opinion 𝑋 is either
−1 or 1. Social impact is a function of the persuasiveness (𝑝𝑖 ) of opponents
(connected agents with the opposite opinion), the supportiveness (𝑠𝑖 ) of sup-
porters (with the same opinion), and the distance (𝑑𝑖𝑗 ) to these agents. The
effect of distance can be modified with 𝛼. As the value of 𝛼 increases, the
influence of agents located farther away diminishes. All of these parameters
are positive random values. The impact 𝐼 is defined as:

𝑁 𝑁
𝑝𝑗 𝑠𝑗
𝐼𝑖 = 𝐼𝑖𝑃 − 𝐼𝑖𝑆 = [∑ 𝛼 (1 − 𝑋 𝑋
𝑖 𝑗 )] − [ ∑ (1 + 𝑋𝑖 𝑋𝑗 )] . (7.2)
𝑗=1
𝑑𝑖𝑗 𝑑𝛼
𝑗=1 𝑖𝑗

With 𝑗 we take the sum over the neighbors of agent 𝑖. Note that when 𝑋𝑖 = 𝑋𝑗 ,
𝐼 𝑃 = 0 due to the 1 − 𝑋𝑖 𝑋𝑗 term, and the same is true for 𝐼 𝑆 when 𝑋𝑖 ≠ 𝑋𝑗 .
The effects of persuasiveness and supportiveness are reduced as the distance
between agents increases. Setting 𝛼 to values greater than 1 reduces the
effect of distant neighbors. In addition to these forces, the theory assumes an
external field 𝐻, as in the Ising model. The dynamic of opinion is:

𝑋𝑖 (𝑡 + 1) = −𝑠𝑔𝑛[𝑋(𝑡)𝐼𝑖 (𝑡) + 𝐻]. (7.3)

Thus, opinion of agents become −1 if 𝑋(𝑡)𝐼𝑖 (𝑡) + 𝐻 is negative, and vice versa.
Lewenstein, Nowak, and Latané (1992) present analytical mean-field solutions
for fully connected networks. Without individual fields, the model ends up
with an infinite number of stationary opinion states, one of which is usually
dominant.
In the presence of individual fields, some minority opinions can become
metastable. These smaller minority clusters can also persist for a long time Metastable opinions may persist for
before shrinking again, and the process repeats itself, resulting in what is some time, but eventually, due to
noise or other factors, they suddenly
called staircase behavior (figure 7.6). Such a model can explain why small shrink to smaller clusters.

171
minority groups (such as flat-earth beliefs) often persist for a long time,
against all odds (Douglas, Sutton, and Cichocka 2017).

Figure 7.6: Five runs of the social impact model with 𝛼 = 5, 𝑝, 𝑠, and 𝐻 sam-
pled from uniform distributions between 0 and 100, and a lattice
of 10 by 10. For example, the line at the top shows staircase-like
behavior around time 320 and at the end of the time series. The
Social Impact model can be found in the online NetLogo models
and in the software repository of this book.

Extensions of this model include learning, leadership, external influences, and

identity effects (for a review, see Holyst, Kacperski, and Schweitzer 2001).

7.3.2 Continuous opinion models

7.3.2.1 Classic models

Another line of research, with its own history, starts from the assumption that
opinions are continuous variables (for a review, see Noorazar 2020). They will
have values between 0 and 1, for instance. A classical model is the DeGroot
model, where agents are connected in a weighted network. At each iteration,
an agent’s opinion is set equal to the weighted average of all connected agents
in the network. In this way, opinions tend to converge (figure 7.7). The
Friedkin—Johnson model (Friedkin and Johnsen 1990) is an extension that
includes a confidence level for each agent. This agent’s confidence in their
own opinion reduces the effect of others. Clustering or polarization in these
linear models can only occur if parts of the network are unconnected. The
Friedkin—Johnson model can be efficiently simulated with Grind (using the
method='euler' option) by:

FJ <- function(t, state, parms){

with(as.list(c(state, parms)),{
X <- state[1:n]
M <- M / apply(M, 1, sum) # weights sum to 1
dX <- (1 - g) * M %*% X + g * X - X
return(list(dX))
})

172
 }
n <- 100
M <- matrix(runif(n^2, 0, 1), n, n)
g <- .95 # if g = 0 => DeGroot model
x0 <- runif(n, 0, 1)
s <- x0; p <- c()
run(odes = FJ, method = 'euler', tmax = 100)

Figure 7.7: Convergence of opinion in the Friedkin—Johnson model in a con-

nected network of agents.

With an additional bias mechanism, in which confirming evidence is weighted

more heavily relative to disconfirming evidence, polarization can also occur in
connected networks (Dandekar, Goel, and Lee 2013).

7.3.2.2 Bounded confidence

The bounded confidence mechanism has been extensively studied as the most
effective way to generate divergence of opinions in continuous opinion models.
It assumes that individuals have a limited willingness to accept and consider
opinions that differ from their own and will only update their opinions if they
are within a certain range or “bound” of similarity.
A simple but very interesting model is the Deffuant model (Deffuant et al.
2000). The initial opinions of n agents are randomly set to values between 0
and 1. At each step, two agents 𝑖 and 𝑗 meet. If ⌈𝑋𝑖 (𝑡) − 𝑋𝑗, (𝑡)⌉ > 𝜖 nothing
happens because the difference in opinion exceeds the bound 𝜖. Otherwise,
they exchange opinions according to:

𝑋𝑖 (𝑡 + 1) = 𝑋𝑖 (𝑡) + 𝜇(𝑋𝑗 (𝑡) − 𝑋𝑖 (𝑡)),

(7.4)
𝑋𝑗 (𝑡 + 1) = 𝑋𝑗 (𝑡) + 𝜇 (𝑋𝑖 (𝑡) − 𝑋𝑗 (𝑡)) .

173
So, if 𝜇 = .5, they find each other in the middle. If 𝜇 = 1, they take each
other’s position, as in the voter model. The value of 𝜇 does not make much
difference, but the model converges fastest with 𝜇 = .5. However, the choice
of the bound 𝜖 makes a big difference. For 𝜖 = 0, all agents stick to their
positions; for 𝜖 > .5, they all converge to 𝑋 = .5. For intermediate values,
different forms of clustering occur (figure 7.8). It has been shown that the
topology of the network does not make much difference (Fortunato 2004). A
drawback of this model is that it converges slowly. A fast but not entirely
accurate code to simulate this model is: 1

set.seed(20)
layout(matrix(1:4, 2, 2))
iter <- 50; mu <- .5; n <- 200
for (bound in c(.1, .2, .3, .5)){
x <- runif(n, 0, 1)
dat <- matrix(0, iter, n)
for (i in 1:iter){
y <- sample(x, n, replace = TRUE) # find an partner for every agent
x <- ifelse(abs(x - y) < bound, x + mu * (y - x), x)
dat[i, ] <- x
}
matplot(dat, type = 'l', col = 1, lty = 1, bty = 'n', xlab = '',
ylab = 'opinion', main = paste('bound = ',bound))
}

Figure 7.8: Four example runs of the Deffuant model with four different bound-
aries. Lower bounds of confidence lead to polarization.

With this code you can explore many scenarios and variants. One interest-
ing option is to have agents with different boundaries (Weisbuch et al. 2002).
Also, adding some noise to 𝑋 at each time step reduces polarization (Zhang It turns out that adding some
and Zhao 2018). One can also lower the bound with the number of interactions. open-minded agents helps prevent
polarization.
1
The second of the two equations is not implemented. This does not lead to different results
as far as I know.

174
This increases the polarization (Weisbuch et al. 2002). One case I find interest-
ing is increasing the bound after polarization emerged for a low bound. This
gives hysteresis. A bound of .5 is sometimes insufficient to reduce polarization.
Castellano, Fortunato, and Loreto (2009) review some other extensions (the
role of propaganda, for instance).
Another well-known model is the Hegselmann—Krause model (Rainer and
Krause 2002). This model is very similar to the Deffuant model, but instead
of communicating with one other agent, they communicate with all connected
agents, but only if the difference in opinion with these agents is sufficiently
small. Thus, agents average the opinion of all connected agents for which the
difference in opinion is less than the bound. This model is an extension of the
DeGroot model and can be simulated by adding two lines to the Friedkin—
Johnson code,

accepted <- abs(outer(X, X, '-')) < bound # acceptable neighbors

M <- accepted * M

after the X <- state[1:n] line and adding bound = .1 (figure 7.9).

Figure 7.9: Clustering in the Hegselmann—Krause model.

Again, many extensions have been studied. A recent paper studies the case
where the network topology is a function of cognitive dissonance in opinions
(Li et al. 2020). Baumann et al. (2020) present a continuous opinion model of
echo chambers. Of particular interest is the multidimensional case (J. Lorenz
2007). When agents accept interaction based on the minimum distance along
one dimension, consensus can be reached more easily. The idea of bounded
confidence has been associated with the concept of the latitude of acceptance
as proposed in social judgment theory (Sherif and Hovland 1961). This theory
also proposes a latitude of rejection. There are two bounds, a lower and an
upper bound. Below the lower bound, agents reduce their differences; between
the bounds they ignore each other; and if they differ more than the upper
bound, they increase their differences. This scenario has been investigated in
Jager and Amblard (2005).

175
In the online library of NetLogo you can find the model “BC”, which simulates
both the standard Deffuant and the Hegselmann—Krause model.

7.3.3 Empirical verification

Castellano, Fortunato, and Loreto (2009) note a striking imbalance between

empirical evidence and theoretical models, in favor of the latter. It is not
that there are no empirical data on the dynamics of opinions. Data come
from studies on voting behavior, multicountry panel surveys, social media,
and laboratory studies (for a review, see Peralta, Kertész, and Iñiguez 2022).
The problem seems to be that these data do not discriminate between models.
Most of the data fit all opinion models, supporting the general modeling ap-
proach but not specific models. This relates to the point that current opinion
models are difficult to falsify because they lack specificity and are too flexibil-
ity. Flache et al. (2017) argue that the field suffers from a lack of systematic
comparison of competing models. The theory-construction approach outlined
by Borsboom et al. (2021) may be helpful here. We need a list of generally
agreed-upon phenomena that all models are supposed to explain.
But another perspective is worth considering. In physics, models can be distin-
guished by their ability to make precise quantitative predictions. By contrast,
in other fields there are often multiple models that roughly explain the same
phenomenon, which can be advantageous. For example, Schelling’s segregation
model and its various iterations consistently predict that segregation occurs
even when individuals are tolerant of different groups, demonstrating a robust
prediction. Similarly, different models of traffic congestion tend to predict
the same key phenomena. In addition, opinion models consistently show that
zealots increase polarization, regardless of the specific model used. This con-
vergence among different models may be the most reliable form of prediction
we can achieve in these areas.

7.4 Psychosocial models

In the introduction to this chapter, I said that psychology is the victim of the
simplifications necessary to develop sociophysics models. In this section we
explore ways to make these models a bit more psychologically realistic. Sev-
eral existing models already include additional psychological variables (Jager
2017). The social impact model is a good example since it incorporates per-
suasiveness and supportiveness of agents to determine opinion change. The
model is also based on a well-known theory in social psychology. Other models
include stubbornness, cognitive dissonance, and confidence (Castellano, For-
tunato, and Loreto 2009). All of these parameters are used to modify the
interactions between agents.
Here I will discuss the model proposed in van der Maas, Dalege, and Wal-
dorp (2020), which uses three dynamic variables to describe agents: informa-
tion, involvement, and opinion. Individual agents are described by the cusp
catastrophe as shown in figure 3.13. This is somewhat similar to the work
of Sobkowicz (2012), who used the cusp model for the individual agent, us-
ing emotion and information as dynamic control variables. As in our model,

176
interactions between agents change both opinions and the control variables.
For example, agitation spreads across agents. However, in his opinion model,
Sobkowicz reduces the cusp dynamics to a three-state system, where opinions
are either −1, 0, or 1. We will not adopt this simplification, as much of the
interesting dynamics (hysteresis within agents) are lost.

7.4.1 Networks of attitude networks

Our starting point is the Ising attitude model explained in Chapter 6, sec-
tion 6.3.3. In this model, attitudes or opinions are conceptualized as networks
of feelings, behaviors, and beliefs about an issue. This new view of attitudes
has been well received in the literature and applied to a number of attitudes
(e.g., Chambon et al. 2022; Turner-Zwinkels and Brandt 2022; Zwicker et al.
2020). The idea is to use this attitude network model as a model for individual
agents.
The resulting model becomes very complex—it is a network of networks model.
This is not a new idea. Hierarchical or multilayer network models, such as
multilayer neural networks (Treur 2019) and multilayer voter models (Ma-
suda 2014), have been applied in many domains (for a review, see Boccaletti,
Bianconi, Criado, del Genio, et al. 2014). However, such a model contains
an enormous number of parameters and is hard to study. We take a simpler
approach.
As discussed in section 6.3.3, the average behavior of the spins in the Ising
model (the mean field) can be represented as a cusp catastrophe. This reduces
the complexity enormously since the cusp attitude model contains only one
equation with three variables: opinion (magnetization), information (external
field), and attention or involvement (inverse of the temperature). Our model,
the hierarchical Ising opinion model (HIOM), is an Ising-type social network
in which each agent is a cusp. Interactions affect information and attention,
leading to changes in opinion. We saw networks of interacting cusps in sec-
tion 4.3.8.1. The HIOM is an extended form of this model.
The HIOM model can be found in the online NetLogo models (HIOM.nlogo)
and in the online software repository of this book. In NetLogo, the equation
and algorithm differ slightly from the original paper, mainly to speed up the
simulation. The equations here are those used in the NetLogo model.

7.4.1.1 The HIOM

Like other opinion models, the HIOM makes assumptions about (a) the topol-
ogy of the network, (b) the interactions between agents, and (c) the definition
of opinion.
The qualitative results of the HIOM do not depend on the topology of the
social network. In van der Maas, Dalege, and Waldorp (2020), the results are
replicated for different topologies (e.g., 2D lattices, stochastic block models).
However, as in other opinion models, more subtle results (e.g., convergence
speed) are likely to depend on the network topology. For (b), the interactions
between agents, specific assumptions are made, which are explained in the
next section.

177
Regarding (c), the definition of opinion, the HIOM is special. Opinion is de-
fined as a cusp. In the review of opinion models, I distinguished between
discrete and continuous opinion models. Interestingly, the cusp behaves con-
tinuously for low values and discretely for high values of the splitting variable.2
In this way, the HIOM bridges these two modeling traditions. Whether opinion behaves discretely or
continuously depends on another
It is important to realize that the HIOM inherits assumptions from the Ising continuous variable (attention).
attitude model. First, the HIOM assumes that attitude nodes (representing
feelings, beliefs, and behaviors toward the attitude object) are binary (−1, 1).
This is clearly debatable. Nodes might be better defined as (0,1) nodes, (-
1,0,1) nodes, or even continuous value nodes as in the XY model (Kosterlitz
1974). Second, we assume undirected pairwise interactions between nodes,
whereas there is much to be said for directed effects. Third, attitude networks
should be reasonably balanced (see section 6.3.3.4). To some extent, these
assumptions can be relaxed without breaking the link to the cusp (see the
appendix of van der Maas, Dalege, and Waldorp 2020).
In the HIOM, information and attention are updated based on interactions
between agents. Information and attention are two orthogonal axes in the
cusp. Information summarizes all variables and influences operating along the
normal axis of the cusp. Its neutral value is 0, and negative and positive
values are associated with negative and positive opinions. Attention has non-
negative real values. The opinion of agent 𝑖 at time 𝑡 changes according to
the cusp equation with information and attention as control variables. For
the implementation in NetLogo, I write the equation as a stochastic difference
equation:3

3
𝑂𝑖,𝑡+1 = 𝑂𝑖,𝑡 − (𝑂𝑖,𝑡 − (𝐴𝑖,𝑡 + 𝐴𝑚𝑖𝑛 )𝑂𝑖,𝑡 − 𝐼𝑖,𝑡 ) 𝑡𝑠 + 𝜖𝑖,𝑡+1 . (7.5)

The 𝜖 term represents white noise sampled from a normal distribution 𝑁 (0, 𝑠𝑂 ).
The time step, 𝑡𝑠 , in this equation is set to a low value (.01) to prevent oscil-
lations and chaotic behavior.

7.4.1.2 Agent interactions: Information and attention

The HIOM makes three assumptions about interactions. First, it assumes that
agents initiate interactions based on their involvement. The idea is simply
that I’m not likely to start a discussion about a topic—say, about genetically
modified food—if I’m not interested in the topic. The probability of initiating
an interaction is equal to attention:

𝑃𝑖,𝑡 (𝑖𝑛𝑖𝑡𝑖𝑎𝑡𝑒𝑠 𝑖𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛) = 𝐴𝑖,𝑡 . (7.6)

Second, it assumes that the attention or involvement slowly decreases over

time. Attention or involvement is a limited resource; one cannot be involved

2
This is highly relevant to the discussion in psychology about type and continua, that is,
whether psychological traits are typological or continuous constructs (Borsboom, et al.,
2016). They can be both!
3
To incorporate close to linear change in 𝑂 as a function of 𝐼, I use 𝐴 + 𝐴min , where
𝐴min = −.5 and 𝐴 ≥ 0. See the original paper for explanation.

178
in everything all the time. With the constant emergence of new interests or
topics, attention to older topics tends to wane.
Third, attention increases again through social interactions. If someone starts
a conversation about genetically modified food, my interest in the topic is
likely to increase. A simple way to implement this is:

𝐴𝑖,𝑡+1 = 𝑑𝑒𝑐𝑎𝑦𝐴 (𝐴𝑖,𝑡 + 𝑑𝐴 𝑢𝑖,𝑡 ). (7.7)

When the agent is involved in an interaction, we set 𝑢𝑖 = 1; otherwise 𝑢𝑖 = 0.

The parameter 𝑑𝐴 determines the rate of change of 𝐴 due to interactions. The
decay in attention, 𝑑𝑒𝑐𝑎𝑦𝐴 , is applied to all agents.
The fourth assumption is about information. We assume that the exchange
of information is an averaging process weighted by attention. If agent 𝑖 is less
attentive to the attitude object than agent 𝑗, agent 𝑖 will move more to the
information position of 𝑗 than 𝑗 will move to 𝑖.
This is formalized by:

𝐼𝑖,𝑡+1 = 𝑟𝑡 𝐼𝑖,𝑡 + (1 − 𝑟𝑡 )𝐼𝑗,𝑡 ,

(7.8)
1−𝑟min
𝑤ℎ𝑒𝑟𝑒 𝑟𝑡 = 𝑟min + −𝑝(𝐴𝑖,𝑡 −𝐴𝑗,𝑡 )
.
1+𝑒

𝐼𝑖 and 𝐼𝑗 denote the information of agents 𝑖 and 𝑗, the agents involved in the
interaction. Resistance, 𝑟 in [0,1], determines the relative impact of agent 𝑗
on agent 𝑖. Resistance or stubbornness is a logistic function of the difference
in attention between the agents. Thus, if 𝐴𝑖 ≪ 𝐴𝑗 , 𝑟 will be close to 0 and
the information in agent 𝑖 will change to the value of the information in agent
𝑗. The strength of this effect, persuasion, is determined by the steepness, 𝑝, of
the logistic function. The parameter 𝑟min determines the minimal value of 𝑟.
If 𝑟min is high, 𝑟 will be high and agents will stick to their information state.
In some scenario’s it is of interest to allow for decay in information especially
in combination with a normally distributed, 𝑁 (𝑚𝐼 , 𝑆𝐼 ), noise term, 𝜖. The
full equation for the update of information is:

𝐼𝑖,𝑡+1 = 𝑑𝑒𝑐𝑎𝑦𝐼 ((1 − 𝑢𝑖,𝑡 )𝐼𝑖,𝑡 + 𝑢𝑖,𝑡 (𝑟𝑡 𝐼𝑖,𝑡 + (1 − 𝑟𝑡 )𝐼𝑗,𝑡 ) + 𝜖𝑖,𝑡+1 ) ,
(7.9)
1−𝑟min
𝑤ℎ𝑒𝑟𝑒 𝑟𝑡 = 𝑟min + −𝑝(𝐴𝑖,𝑡 −𝐴𝑗,𝑡 )
.
1+𝑒

If 𝑚𝐼 ≠ 0 and 𝑠𝐼 = 0, a constant external field is active. If 𝑑𝑒𝑐𝑎𝑦𝐼 < 1 ,

information gradually shrinks to 0.

7.4.1.3 Algorithm

The model is now in place and can be simulated by following the steps below:

• Select a network topology.

179
 • Set the model parameters 𝑡𝑠 (.01), 𝐴min (-.5), 𝑠𝑂 (.01), 𝑚𝑖 (0), 𝑠𝑖 , 𝑑𝐴 , 𝑝,
and 𝑟min . Values in parentheses are defaults.
• Initialize agents, set 𝐼𝑖𝑛𝑖𝑡 , 𝐴𝑖𝑛𝑖𝑡 , and 𝑂𝑖𝑛𝑖𝑡 .
• Iterate:

– Randomly choose a set of agents, weighted by attention 𝐴 (equa-

tion 7.6).
– Iterate over this “active” set of agents (𝑢𝑖 = 1}:
∗ For each agent, randomly choose a neighbor as partner in the
interaction.
∗ Add attention to both agents (equation 7.7, +𝑑𝐴 𝑢𝑖 ).
∗ Exchange information (equation 7.8).
– Apply decay in 𝐴 to all agents (equation 7.7, 𝑑𝑒𝑐𝑎𝑦𝐴 𝐴𝑖 ).
– Add noise and apply decay in 𝐼 to all agents (equation 7.9,
𝑑𝑒𝑐𝑎𝑦𝐼 (𝐼 𝑖 + Ν(0, 𝑠𝐼 )).
– Update opinion 𝑂 in all agents (equation 7.5).

The 𝑑𝑒𝑐𝑎𝑦𝐴 has a special role in the current implementation of the HIOM.
Instead of being fixed, it depends on the difference between the percentage
of agents in the “active” set and the desired percentage of active agents
(%𝑎𝑐𝑡𝑖𝑣𝑒_𝑎𝑔𝑒𝑛𝑡𝑠), which is controlled by one of the sliders in the NetLogo
model (figure 7.10). This allows us to manipulate the general interest (atten-
tion) in the opinion object.
There is no obvious stop criterion, but in practice some type of convergence
happens over time.
In the standard setup, you can see that high attention (set %active-agent high)
leads to polarization. If you decrease the difference in information (set decay_I
to .5), the polarization remains even if the difference in underlying information
is 0. Only if you also decrease attention (set %active-agents low) does the
polarization disappear. This is the first simulation described in van der Maas,
Dalege, and Waldorp (2020). The second and third simulation are described
in the next two sections.

7.4.1.4 The persuasion paradox

Suppose we have a network of conservative but low-attentive agents into which

we add a few highly attentive activists with the opposite information and opin-
ion. At first glance, one would expect the activists’ opinion to spread quickly.
They initiate all interactions (according to equation 7.6) and have much more
impact on their interaction partners than vice versa (equation 7.8). However,
this is not what happens. This scenario leads to polarization. The attention
of some conservatives is too quickly raised, and people get a strong opinion
against the activists. Figure 7.11 shows this effect which of course, depends
on some parameter settings. If attention increases more slowly, activism may
spread better. You can test this scenario in the NetLogo model by selecting

180
Figure 7.10: The interface of the HIOM NetLogo model. All model parameters
can be adjusted with sliders. With add-activists and perturbate-
activists, the key effects of HIOM can be reproduced. The graphs
on the right show the distribution of opinion, attention, and in-
formation and the change in the means of these variables.

setup-black-pete and add-activists. Activists can now be added by clicking in

the world. Vary persuasion 𝑝 and 𝑑𝐴 (da) to see different effects.
What is new in this model? It shows a new mechanism for polarization, namely
hysteresis within agents, leading to the persuasion paradox. Without a The persuasion paradox describes how
cusp for the individual opinion dynamics, this form of polarization would not efforts to persuade can backfire,
causing the other person to become
occur. even more entrenched in their original
position. This is due to the effect of
social interaction on attention.
7.4.1.5 A counterintuitive prediction

In continuous-opinion models, the main reason for polarization is bounded

confidence. We can easily add this to the HIOM: when ∣𝑂𝑖 − 𝑂𝑗 ∣ > 𝜖, there is
no interaction (no increase in attention and no exchange of information). This
would even increase the polarization in the HIOM. However, since opinion and
information can have opposite signs due to hysteresis, there is a way out. An
example would be a meat-eating vegetarian, a person with a vegetarian point
of view who actually eats meat. I claim to be such a vegetarian, but this
position is not generally accepted in my local environment.
The scenario, simulated in van der Maas, Dalege, and Waldorp (2020), consists
of a majority of low-involved meat eaters and highly involved vegetarians.
The meat eaters refuse to talk to the vegetarians due to bounded confidence,
leading to increased polarization.
Now some attentive vegetarians will be perturbed into the meat-eating po-
sition. This is a metastable state. This state is somewhat stable, although
they may jump back after a few iterations. But while the vegetarians are

181
Figure 7.11: The visualization of persuasion paradox. Activism initially
spreads quickly but also increases the attention of conservative
agents (see second and third panel). Over time, opinion polar-
izes. Red nodes are activists and blue nodes are conservatives.
The size of the nodes represents attention. The last panel ex-
plains why some conservatives become (anti-)activists themselves.
(Reprinted from van der Maas, Dalege, and Waldorp (2020) with
permission)

182
in this metastable meat-eating state, they can communicate with the less in-
volved meat-eaters and spread vegetarian information. This proves effective
(see figure 7.12). This scenario is also implemented in the NetLogo model.

Figure 7.12: The meat-eating vegetarian. The left panels show equal initial
states. In the top panels, bounded confidence prevents the veg-
etarian (red) and the meat-eater (blue) from interacting. This
leads to polarization. In the bottom panels, some vegetarians
are perturbed toward meat eating. These perturbed agents have
𝐼 < 0 and 𝑂 > 0, which is possible because 𝐴 > 0 (hysteresis).
Since 𝑂 > 0, these agents can exchange information with meat-
eaters, leading to the spread of vegetarianism. (Reprinted from
van der Maas, Dalege, and Waldorp (2020) with permission)

7.4.1.6 Variants of the HIOM

What appeals to me about the HIOM is its ability to incorporate a greater

degree of psychological insight into opinion dynamics models, resulting in a
novel explanation for polarization attributed to hysteresis within individuals.
A strength of the HIOM is that the definition of opinion is based on the Ising
attitude model, a psychological network model of attitudes and opinions that
has been developed and supported in a number of papers (see section 6.3.3).
Furthermore, it presents an untested, counterintuitive prediction that warrants
further investigation.
What is also unique about this model is the role of attention/involvement.
It shows the risks of too-low and too-high values of this variable. Too-high
values lead to extreme hysteresis effects in agents, making a change of opinion
impossible. This is clearly a dilemma for activists. But too little involvement
is also risky. The moment you stop influencing such agents, the attitude Talking to a person with very low
nodes start behaving randomly again, and they will move back to the neutral involvement may have a positive
short-term effect, but no long-term
effect.

183
opinion at the back of the cusp. The persuasion paradox and the involvement
dilemma need further study, but they seem important.
What I like less is that the HIOM is perhaps too complex, which makes it
difficult to study its behavior. One idea is to make attention a network pa-
rameter. An attentive agent simply has many connections to other agents. As
attention decreases, this translates into fewer connections.
A more radical simplification that still uses the attention/involvement idea
is a voter model with three votes (leftist, centrist, and rightist), as in the
constrained three-state voter model (Redner 2019). Leftists and rightists don’t
talk to each other (bounded confidence), but they do talk to the centrists.
What we can add to this model is attention. Extremists (left and right) have
high attention, while centrists have low attention. As in the HIOM, more
attentive agents are more persuasive, so centrists tend to become more extreme
due to interactions. On the other hand, attention is costly, and there is a
probability that extremists spontaneously decay to the centrist position. Such
a model can be studied analytically. This voter-type model has no within-
subject hysteresis and will show less interesting behavior.
What I also like less is that the HIOM is not complex enough. A huge simpli-
fication is that we have left out the fact that we have more than one attitude.
We have hundreds or maybe thousands of attitudes. Since these are not in-
dependent, it is safe to say that these attitudes form complex networks with
subclusters and central (hub) attitudes. A better model would be a (social)
network of (attitude) networks of (attitude element) networks. In our current
work, the level of within-person attitude networks is underexplored. Finally,
many crucial aspects of opinion formation are missing, such as the role of
(social) media, confirmation bias, identification with groups in society, the
political system, etc.
What I really don’t like is that the Ising attitude model, in which attitudes
behave according to the cusp, always leads to radicalization when we get in-
volved. At best, we can jump irregularly between the extreme states (as in
figure 4.2) as a form of ambivalence. But our society needs people, for exam-
ple, judges, who get involved and attend, but at the same time remain neutral.
This is impossible in the Ising attitude model and thus in the HIOM.
A possible way out is to start from the tricritical Ising or Blume—Capel model
(see section 6.3.3.5), where the attitude nodes have three states (−1, 0, 1) in-
stead of only two. The resulting dynamic equation is the butterfly catastrophe,
which has four instead of two control variables. Building this in the HIOM
will be a challenge. The Blume—Capel model has also been proposed as a
between-person opinion model (Barbaro, Chayes, and D’Orsogna 2013; Ferri,
Dı́az-Guilera, and Palassini 2022).

7.4.2 Cascading transitions in other psychosocial systems

We have now seen two psychological models of cascading transitions. The first
was the model for multifigure multistable perception (section 4.3.8). This was
a within-person model. The second is the HIOM. But many other processes
in psychosocial systems come to mind. One example is addiction. At the
psychological level, the dynamics of addiction are often sudden (e.g., quitting

184
and relapse), while at the social level there are sudden outbreaks of substance
abuse (e.g., the heroin epidemic).
To model this cascading process, we could follow the same approach as in the
HIOM. Instead of the cusp, we could use the spruce budworm model as a model
of individual substance use. Interactions in the network affect the parameters
𝑟𝑏 , 𝐾, 𝐴, and 𝐵 parameters (Boot et al. (submitted for publication)). A similar
approach is possible with the panic model (section 4.3.5). In future work, we
will also apply this model to collective learning processes.

7.5 Psychosociophysics

Neglecting the social world while attempting to model complex psychological

processes may lead to inaccurate outcomes (Sobkowicz 2020). It is crucial for
psychologists engaged in modeling to familiarize themselves with prominent
sociophysics modeling techniques. The objective of this chapter was to offer a
comprehensive overview of these models, but it is important to note that this
review is not exhaustive. As new models continue to emerge, the foundational
knowledge provided here will, I hope, enable readers to stay informed and
assess their applicability to psychological inquiries.
Similarly, excluding psychological factors while modeling the social world can
be counterproductive. It is essential to emphasize the need for psychosocial
or even psychosociophysical models that integrate both aspects. The social
impact model is as an example of such an approach. Our own model, the
HIOM, has been elaborated upon in this chapter. The HIOM stands as a
significant step in the direction of psychosociophysics models. The concept of
cascading transitions is evidently relevant to other applicable cases, and there
is ample opportunity for further exploration in this field.

7.6 Exercises

1) Rerun the Schelling simulation (figure 7.1) with density of 50 instead of

97. Are the results substantially different? Submit your plot. (*)
2) Implement the equations (equation 7.1) for the simple naming game in
Grind. Show that the case where all three options A, B, and AB coexist
is an unstable fixed point. (**)
3) What happens in the Axelrod model (section 7.2.3) when 𝐹 = 10 and
𝑄 = 1. Why? (*)
4) Implement the 1d voter model (section 7.3.1.1) in either R or NetLogo
(using BehaviorSpace and R for data analysis). Check that the proba-
bility of convergence to a state is equal to its initial proportion. Then
check that the convergence time is a quadratic function of 𝑁 by plotting
the square root of this time versus 𝑁 . Take the average of these times
for at least 20 runs. (**)
5) In the social impact model (section 7.3.1.3), does the staircase behavior
depend on the max-h parameter? (*)

185
 6) In the Deffuant model (section 7.3.2.2), a limit of .5 almost always leads
to convergence of opinions. Adjust the R code for the Deffuant model
so that the bound grows from 0 to .5 over 1,000 iterations. You end up
with one or two clusters. Why do you end up with two clusters even if
you increase the number of iterations? (**)
7) Run the HIOM NetLogo model (section 7.4.1.1). Set the mean-init-
information to .5, the mean-init-attention to 1, the bound to .2, and the
%active-agents to 90. Let it run and use add-activists and pertubate ac-
tivists. Why does this not result in a change? (*)

8) Think of a simple way to add the effect of media to the HIOM. Implement
and present your results. (**)
9) Implement the HIOM in a preferential attachment network. Use the
“Preferential Attachment NetLogo” model. (**)
10) Design an empirical study to test the “meat-eating vegetarian” predic-
tion (section 7.4.1.5).

186
8 Epilogue

In writing this book, I had three primary goals: to provide a comprehensive

overview of complex-systems research with a particular emphasis on its applica-
tions in psychology and the social sciences, to teach skills for complex systems
research, and to encourage critical thinking about the potential applications
of complex systems in psychology. I will first discuss the main points of the
chapters briefly and then focus on evaluating the complex-systems approach
to psychology.
In Chapter 1 I defined complex systems. Complex systems are composed of
interacting subsystems, resulting in emergent behaviors not seen at lower lev-
els. These emergent patterns often arise through self-organization, can change
suddenly, and may exhibit chaotic behavior, making prediction difficult. They
can be studied using various methods, such as those originating in nonlinear
dynamical system theory, agent-based modeling, and network theory.
I argued for the independence and autonomy of psychology as a scientific dis-
cipline. This does not mean that the mechanisms and principles that operate
in psychological systems are necessarily different from those in other sciences.
The synchronization of atoms in a laser beam may not be very different The power of the complex-systems
from the alignment of symptoms in a depressed patient or the synchronized approach is that the same organizing
principles can operate at very different
movements of people evacuating a burning building. Complex systems in any levels of description.
science are almost always networks in which unpredictable phenomena can
occur. These can be networks of atoms, neurons, symptoms, cars, people, or
even entire countries. In these dynamic networks, there are usually only a
limited number of organized patterns or stable states. Manipulating control
variables can lead to bifurcations, which can take a limited number of forms.
We have seen these principles again and again.
One may argue that emergent properties at one level of description can always
be explained or rewritten as lower-level events, but this is largely irrelevant.
In psychology we have to distinguish at least three basic levels of description
related to the brain, the mind, and the social world. In describing the mind,
delving deeper than the level of neurons does not provide additional insights.1
And in modeling the spread of conspiracy theories, the way axons of neurons
grow can be left out of the equation.

1
Actually, the role of quantum processes in psychological processes is a topic of ongoing
debate and research in the scientific community. While some researchers have proposed
that quantum mechanics might play a role in cognitive processes, this idea is not widely
accepted. One theory suggests that quantum processes in microtubules within brain
cells could be linked to consciousness (Hameroff and Penrose 1996). Other researchers
have suggested that the probabilistic nature of quantum mechanics might provide a better
model for human decision-making than classical probability theory (Busemeyer and Bruza
2012). However, this work does not suggest that quantum physical processes are involved
in decision-making. Rather, it uses quantum theory as a mathematical framework for
modeling cognitive processes.

187
I ended with reasons to be moderately optimistic about the application of
complex-systems science in psychology. First, complex systems can often be
simplified without losing their explanatory value. Second, these systems can
be described by a limited number of equilibria. Third, we can use network
science to model complex systems.
In Chapter 2, I introduced chaos theory and many of the key concepts of
complex systems, such as fixed points, limit cycles, bifurcations, and dynamical
system models. Learning about deterministic chaos changed my worldview
and also my appreciation of beauty in mathematics. I hope I succeeded in
conveying both.
The principal unpredictability of
It often seems to me that psychologists somehow believe that if they could col- complex systems, when they are in the
lect vast amounts of extremely accurate life history, environmental, genetic, bi- chaotic phase, is important to
ological, and neuropsychological time-series data from millions of individuals— understand.
a feat that is currently unattainable—and feed it into a sophisticated nonlin-
ear multilevel regression model with numerous higher-order interactions, they
could predict virtually anything. This is simply wrong, not only because of
deterministic chaos but also because of the influence of epigenetic processes
(the third source), which I discussed in section 4.3.4. The Pólya urn model
provides a very simple demonstration of why even perfect knowledge of the
initial conditions and dynamics of a system is insufficient to predict individual
developmental outcomes.
The application of chaos theory in psychology is limited to the analysis of
psychophysiological data. The hypothesis is that the brain works best on the
edge of chaos (see also section 5.4.1). Despite hundreds of papers written
in the last forty years, I would say that the evidence for this hypothesis is
inconclusive. This approach requires very high-quality data and advanced
statistical approaches, which are difficult to acquire and challenging to develop.
Furthermore, it is worth mentioning that there are even skeptics who question
the applicability of chaos theory in studying complexity altogether (Anderson
1999a).
Chapter 3 was devoted to transitions, or tipping points, as they are sometimes
called. To me this is a very practical concept in complex-systems theory. I am
probably somewhat biased, but I see instances of tipping points across a wide
range of psychological processes. I list just a few that I have not discussed:
sudden insights, creative breakthroughs, aggressive acts, the onset of puberty,
mood swings, vocabulary spurts, and dropping out of school. In all these cases,
the modeling and empirical program laid out in Chapter 3 might be fruitful.
The chapter also contained an elementary introduction to the mathematics of
bifurcations and catastrophes. I think a basic understanding of the key formal
concepts is necessary and achievable. I provided sources for further study;
for those who find these concepts still difficult, I also recommend running
and studying the examples of the cusp catastrophe in Chapter 4. I have also
outlined a methodology for empirically evaluating catastrophe models using
catastrophe flags and Cobb’s statistical approach. I believe that this methodology
effectively addresses the major
In Chapter 4, I focused on building dynamical systems models. I introduced criticisms of the application of
Grind as a tool for implementing a wide range of dynamical systems models catastrophe theory in the social
in biology and psychology. This modeling approach allows us to build more sciences.
mechanistic models that we can still fully understand (using numerical bifur-
cation analysis). I have tried to give a representative overview of dynamical

188
systems modeling in psychology, but you will easily find many other models
in the literature. In such a case, I always recommend implementing the model
yourself. In most cases, Grind will do. Replication is the key to good science,
and you will learn a lot in the process.
I discussed the evaluation of ecosystem models in some detail (section 4.2.7).
I noted that even simple models imply a large number of assumptions, many
of which are made implicitly (with the assumptions underlying the Lotka—
Voltera model as an example). Also, seemingly trivial changes in model choices
can have a huge impact on the qualitative behavior of models. As I said in the
same chapter, I find the process of formalizing a verbal model fascinating. It
tends to be very confusing. Suddenly it is unclear what the assumptions are,
what mechanism is really being proposed, what the time scales are, or even
what the phenomenon to be explained really is.
It is essential to have studied many
Models usually combine a number of mechanisms, and I strongly recommended different examples of dynamical
reusing mathematical model pieces in other models. Modeling by analogy can systems models before you attempt to
be very productive (Haig 2005). Finally, when modeling, it is critical to always build your own.
keep the connection to the data in mind. I am particularly concerned about
this with the causal loop diagram approach. Models built in a session with
content experts tend to get big with lots of boxes and links. This may not be a
problem if all time series of measurements for each of these boxes are available,
but this is rarely the case. I prefer to start simple and only add variables and
equations when some established phenomena cannot be explained by the most
trivial model.
In Chapters 2, 3, and 4, I focused on systems with a small number of variables.
This part of the book covers what is often called nonlinear dynamical system
theory. I view the theory of nonlinear dynamical systems as a fundamental
component of the complex-systems approach. The second part of the book
deals with systems with a large number of variables.
Self-organization was the subject of Chapter 5. I also used this chapter to
introduce main theories in the study of complex systems, such as Haken’s work
on synergetics and Prigogine’s ideas on irreversible transition and the second
law of thermodynamics. I hope that I have successfully conveyed my own sense
of awe and amazement at the self-organizing processes found in nature, as well
as the unexpected potency of seemingly simple systems like the Game of Life.
By engaging in the NetLogo simulations, I trust that the somewhat abstract
concept of self-organization has become more comprehensible. As an example,
I mention the spiral waves in the spatial model of hypercycles that prevented
the abundant growth of parasites. This example of strong emergence is easy to
understand by running the simulation and studying the basic NetLogo code.
I again realized while writing this book how the concepts of self-organizing
complex systems were already present in early day psychology. I gave Gestalt
psychology and Piagetian theory as examples, but one could make the same
point for Rogers’s theory of self-concept, Heider’s balance theory, or Gibson’s
ecological theory of perception. This is why I perceive the complex-systems The concepts of self-organizing
approach in psychology less as a novel theory and more as a formalization of complex systems were already present
in the early days of psychology.
these intriguing yet abstract verbal theories. For instance, the Ising attitude
model formalizes of numerous established concepts in social psychology.

189
Chapter 6 focused on the psychological and psychometric network approaches.
This chapter ends with an extended discussion of the challenges facing this
popular research line. My own contributions are mainly theoretical. I do
think it is important to provide an alternative to the common cause view on
psychological traits. The inability to intervene on common causes, or even to
gain knowledge of these fixed biological factors without relying on the observ-
able factors they explain, leads to a discouraging psychological theory that
can easily be misused to abandon the less fortunate to their fate (Heckman
1995).
In terms of modeling, I consider the application to attitudes to be the most
successful. The Ising attitude model not only builds on earlier connectionist
network models, but also incorporates improvements. The dynamics of the
Ising model are better understood from a mathematical perspective, it offers
a novel psychological interpretation of the temperature parameter, and it can
be effectively fit to data. The model formalizes key concepts in social psychol-
ogy, such as dissonance and the mere thought effect, while suggesting a new
explanation for the differences between implicit and explicit attitude measures.
Using the mean-field approximation of this model in the HIOM represents Equating attention and (inverse)
an innovative development in sociophysics. I expect many more innovations temperature may have broader
implications beyond attitude theory.
in both network psychology and network psychometrics.
In the final chapter, Chapter 7, we moved into the realm of the social world. Be-
cause psychologists are often unfamiliar with disciplines such as computational
social science, sociophysics, and agent-based modeling, I aimed to provide a
concise overview of this rapidly evolving field. Finding the right balance in
simplification has proven to be a challenge. While the simple voter model of-
fers analytical tractability, its relevance is primarily theoretical. More realistic
models quickly become intractable, even with the help of simulations.
I have focused on improving the psychological realism of the agents. My
approach involves linking three levels of description: the interaction of attitude
elements, the mean-field representation of attitudes as cusps, and the HIOM,
where cusp-like agents interact within social networks. To my knowledge,
this three-level integration is unique. Similar to many complex models in I expect to see many more
psychology and the social sciences, a notable weakness is the connection to applications of the cascade transition
model to psychological systems.
data, specifically, data that really differentiate models.
Let’s zoom out more. By now, you should have a solid understanding of
what complex systems involve, including the concepts of deterministic chaos,
catastrophes, and self-organization. I have illustrated these ideas with numer-
ous examples from various scientific fields, including psychology. It is vitally
important to be aware of existing models and frameworks when you begin
constructing your own models. This was my first goal.
Second, I placed significant emphasis on the importance of modeling skills.
For me, the one critical path to comprehension lies in simulation. I hope that
the skills you have developed by working through my exercises will encourage
your engagement in formal modeling in psychology.
My third objective was to encourage a critical approach toward the study
of complex systems in psychology and, in a larger context, within the realm
of psychology and science itself. This naturally leads to the question of my
personal stance on this undertaking. I must confess, I have mixed feelings.

190
About twenty years ago, in my inaugural lecture, I portrayed myself on the un-
stable maximum between the two minima of the cusp potential function. The
attitude object was our field of psychology, and the minima represented pos-
itive (“love”) and negative (“hate”) evaluations. Since I am obviously highly
involved in the matter, I’m on the front side of the cusp, which means that
this in-between state is highly unstable, assuming, of course, that the cusp
model of attitudes is correct.
So, if you are asking me to burn our field to the ground, including my own
work, you have come to the right place. In short, we have a replication crisis
(Aarts et al. 2015; Nosek et al. 2022; Pashler and Wagenmakers 2012), we
have a theory crisis (Eronen and Bringmann 2021; Oberauer and Lewandowsky
2019), and we have a measurement crisis (Franz 2022; Lumsden 1976; Michell
1999). How many crises can you have?
Each of these crises is more severe than one might think at first glance. For We can be most optimistic about the
a long time, questionable research practices, 𝑝-hacking, selective reporting, replication crisis.
cherry-picking studies, presenting exploratory results as confirmatory results,
to name a few, dominated research. All of this has changed radically since
2011. Surprisingly, a fraud case2 in my own country played a major role in
this shift to open science, preregistration, and data sharing. Maybe we were
in a metastable state and just needed one such perturbation. We are still
in the midst of this transformation and should not be celebrating too soon
(Chambers 2017). But a revolution it is!
Then the theory crisis. Depending on the criteria (weak or strong), psycholo-
gists either have millions of theories or none. There is not much in between.
Many recent papers have proposed formalization as a way out of the theory
crisis (Oberauer and Lewandowsky 2019; Borsboom et al. 2021; Rooij and
Blokpoel 2020). This book is written from that perspective. I hope it makes a
contribution, but I’m well aware of the differences between the scientific basis
of formal models in the natural sciences and our modest attempts. To me, it is
all about the right degree of simplification. In modeling complex systems, we
almost always define at least two levels: the microscopic and the macroscopic.
Emergent phenomena at the macroscopic level arise from microscopic interac-
tions. The step from neural activity to higher reasoning just seems too large
(although we may need to rethink this in light of the astonishing successes of
large language models).
In writing this book, I learned that the distinction between phenomenological
and mechanistic modeling is less discrete than I thought. First, we have been
able to provide a foundation for some phenomenological catastrophe models.
The cusp model of attitudes can be derived from the Ising model of atti-
tudes, which is based on some simple assumptions about attitudinal networks.
Second, many biological models combine phenomenological and mechanistic
elements. Some terms may be well argued, others are just pragmatically cho-
sen (the Holling types, for example). Nevertheless, I hope to have shown that
studying these models is very informative. In my work, analogical modeling
plays a central role.
I see the measurement crisis as the most serious problem. Let me recall Richard
Feynman’s claim that the accuracy of calculating the size of the magnetic mo-

2
https://en.wikipedia.org/wiki/Diederik_Stapel

191
ment of the electron is the equivalent of measuring the distance from Los
Angeles to New York, a distance of over 3,000 miles, to within the width of
a human hair. And that was in 1985! I am an expert in psychological mea-
surement and have published many papers on new methods of psychological
measurement. I can tell you that we are nowhere near this amazing level of
quantification and precision. We cannot do addition!
Addition is the litmus test of quantification (Michell 1997). Real quantities,
such as weights and distances, can be added. One kilogram + one kilogram =
two kilograms, which is twice as much as one kilogram.3 We cannot say such
things about IQs, personality test scores, or Likert scores on attitude items. I
do not think this is a hopeless endeavor. Even physicists have lived through
times when key concepts were vaguely understood and poorly measured (see
Inventing Temperature by Chang (2008)). I also draw hope from statements
such as “Every law of physics, pushed to its extreme, will turn out to be
statistical and approximate, not mathematically perfect and precise” (Wheeler
1994).
Currently, our scales of measurement are somewhere between ordinal and in-
terval. Perhaps there is a continuous path to improvement. But for now, we
have to live with rather weak scales of measurement. This has implications for
our attempts to formalize psychological theories. The exact form of certain
terms in our equations is irrelevant when we have only ordinal or semi-interval
data. When testing models, we should focus on their qualitative behavior.
This is exactly what we did with the catastrophe flags in Chapter 3, and it is
one reason why I adhere to the sometimes-criticized catastrophe theory.
It is also important to temper expectations about distinguishing models based
solely on their quantitative predictions. Achieving consensus on qualitative
predictions may be the most feasible outcome. In the previous chapter, I
used the example of segregation, which is predicted by many models and their
variants, even when individuals are relatively tolerant.
I will not attempt to review these crises in depth, but I would like to suggest,
in line with the book, that these crises form a mutualistic network (figure 8.1).
Progress in resolving one crisis will have a positive impact on resolving others.
If we can have more confidence in the empirical basis of many well-known
psychological phenomena, theory development will benefit. Improved theories
are necessary to advance measurement, and vice versa. In periods of rapid
progress in sciences such as particle physics and biochemistry, we often see an
upward spiral of theory development and measurement techniques.
We have been part of at least one such radical transformation, and it is hap-
pening now: the AI revolution. Of course, we cannot claim ownership of this
revolution, as many disciplines, such as computer science, have played a key
role. But the current progress is based on mechanisms (neural learning and
operant conditioning) that were first studied by psychologists. Unfortunately,
there are no other convincing examples. We have not yet invented the plane
or the refrigerator. This is no joke. In science, technology is the proof of the
pudding. At some point we really must solve problems like addiction or panic
disorder.
3
This is sufficient but not a necessary condition. Sometimes a concatenation of two quanti-
ties gives a weighted mean, for example when blending two liquids with varying temper-
atures.

192
Figure 8.1: The network of crises in psychology. The resolution of one crisis
will, I hope, have a positive impact on the resolution of others.

Despite all these negatives, psychology remains the most fascinating science
of all. As I emphasized in Chapter 1, the emergence of global waves of elec-
trical activity from rapid local interactions forms the basis of our conscious
thought processes. This phenomenon is truly astonishing, and understanding
it presents one of the most intriguing scientific tasks of all time. It involves The quest to understand the human
the ultimate complex system, and we are exploring it with our own minds. mind is undeniably one of the most
challenging scientific endeavors.
Psychology is full of counterintuitive findings and paradoxes. Perhaps the
greatest paradox is that psychology, unlike any other science, reveals the limi-
tations and fallibilities of the human intellect while using the very intellect it
studies.
The field of psychology is constantly moving away from grand theories, which
were often little more than the collective opinions of some random man, toward
more detailed models of subprocesses and systems. Significant progress has
been made on a smaller scale. Similarly, the complex-systems approach does
not currently provide a grand theory for psychology but, rather, a versatile
set of tools for modeling, analysis, and understanding. I cannot conclude with
a comprehensive overall theory of the complex human brain—mind—social
world system. I simply don’t have it, only bits and pieces.
It is my hope that this book will contribute to lasting progress in psychological
research. The overview of complex-systems research in other disciplines is
perhaps helpful. I also hope that a new generation of researchers will learn
many practical, useful modeling skills. And I hope that I have found the
unstable maximum between hate and love for psychology.

193
References

Aarts, Alexander A., Joanna E. Anderson, Christopher J. Anderson, Peter R.

Attridge, Angela Attwood, Jordan Axt, Molly Babel, et al. 2015. “Esti-
mating the Reproducibility of Psychological Science.” Science 349 (6251).
https://doi.org/10.1126/science.aac4716.
Abar, Sameera, Georgios K. Theodoropoulos, Pierre Lemarinier, and Gregory
M. P. O’Hare. 2017. “Agent Based Modelling and Simulation Tools: A
Review of the State-of-Art Software.” Computer Science Review 24 (May):
13–33. https://doi.org/10.1016/j.cosrev.2017.03.001.
Abe, Yayoi, Muneyuki Ishida, Erika Nozawa, Takayoshi Ootsuka, and Ryoko
Yahagi. 2017. “Cusp Singularity in Mean Field Ising Model.” European
Journal of Physics 38 (6): 065102. https://doi.org/10.1088/1361-6404/a
a82fc.
Abraham, Ralph. 1991. “Cuspoidal Nets.” In Business Cycles: Theories,
Evidence and Analysis, edited by Niels Thygesen, Kumaraswamy Velupillai,
and Stefano Zambelli, 56–63. International Economic Association. London:
Palgrave Macmillan UK. https://doi.org/10.1007/978-1-349-11570-9_3.
Abraham, Ralph, Alexander Keith, Matthew Koebbe, and Gottfried Mayer-
Kress. 1991. “Computational Unfolding of Double-Cusp Models of Opin-
ion Formation.” International Journal of Bifurcation and Chaos 01 (02):
417–30. https://doi.org/10.1142/S0218127491000324.
Abrams, Peter A., Lev R. Ginzburg, Peter A. Abrams, Lev R. Ginzburg, Peter
A. Abrams, and Lev R. Ginzburg. 2000. “The Nature of Predation: Prey
Dependent, Ratio Dependent or Neither?” Trends in Ecology & Evolution
15 (8): 337–41. https://doi.org/10.1016/S0169-5347(00)01908-X.
Acebrón, Juan A., L. L. Bonilla, Conrad J. Pérez Vicente, Félix Ritort, and
Renato Spigler. 2005. “The Kuramoto Model: A Simple Paradigm for
Synchronization Phenomena.” Reviews of Modern Physics 77 (1): 137–85.
https://doi.org/10.1103/RevModPhys.77.137.
Adams, Pauline, and Mason Haire. 1959. “The Effect of Orientation on the
Reversal of One Cube Inscribed in Another.” The American Journal of
Psychology 72: 296–99. https://doi.org/10.2307/1419384.
Ainsworth, Mary D. Salter, Mary C. Blehar, Everett Waters, and Sally N.
Wall. 2015. Patterns of Attachment: A Psychological Study of the Strange
Situation. New York: Psychology Press. https://doi.org/10.4324/978020
3758045.
Alderisio, Francesco, Benoı̂t G. Bardy, and Mario di Bernardo. 2016. “En-
trainment and Synchronization in Networks of Rayleigh–van Der Pol Os-
cillators with Diffusive and Haken–Kelso–Bunz Couplings.” Biological Cy-
bernetics 110 (2): 151–69. https://doi.org/10.1007/s00422-016-0685-7.
Alexander, Ralph A., Glenn R. Herbert, Richard P. DeShon, and Paul J.
Hanges. 1992. “An Examination of Least-Squares Regression Modeling of
Catastrophe Theory.” Psychological Bulletin 111: 366–74. https://doi.or
g/10.1037/0033-2909.111.2.366.

194
Anderson, Philip W. 1972. “More Is Different.” Science 177 (4047): 393–96.
https://doi.org/10.1126/science.177.4047.393.
———. 1999a. “The Eightfold Way to the Theory of Complexity: A Pro-
logue.” In Complexity: Metaphors, Models, and Reality, edited by G.
Cowan, D. Pines, and D. Meltzer, 7–16. Perseus Books.
———. 1999b. “Perspective: Complexity Theory and Organization Science.”
Organization Science 10 (3): 216–32. https://doi.org/10.1287/orsc.10.3.21
6.
Arnaud, M. 2021. “Mixture Modelling from Scratch, in R.” Medium.
https://towardsdatascience.com/mixture-modelling-from-scratch-in-r-
5ab7bfc83eef.
Ashby, W. R. 1956. “An Introduction to Cybernetics.” An Introduction to
Cybernetics.
Attwell, David, and Costantino Iadecola. 2002. “The Neural Basis of Func-
tional Brain Imaging Signals.” Trends in Neurosciences 25 (12): 621–25.
https://doi.org/10.1016/S0166-2236(02)02264-6.
Axelrod, Robert. 1997. “The Dissemination of Culture: A Model with Local
Convergence and Global Polarization.” Journal of Conflict Resolution 41
(2): 203–26. https://doi.org/10.1177/0022002797041002001.
Ayers, Susan. 1997. “The Application of Chaos Theory to Psychology.” The-
ory & Psychology 7 (June): 373. https://doi.org/10.1177/09593543970730
05.
Bak, Per, Chao Tang, and Kurt Wiesenfeld. 1988. “Self-Organized Criticality.”
Physical Review A 38 (1): 364–74. https://doi.org/10.1103/PhysRevA.38.
364.
Banks, J., J. Brooks, G. Cairns, G. Davis, and P. Stacey. 1992. “On Devaney’s
Definition of Chaos.” The American Mathematical Monthly 99 (4): 332–34.
https://doi.org/10.1080/00029890.1992.11995856.
Barabási, Albert-László, and Márton Pósfai. 2016. Network Science. 1st
edition. Cambridge, United Kingdom: Cambridge University Press.
Barbaro, Alethea B. T., Lincoln Chayes, and Maria R. D’Orsogna. 2013.
“Territorial Developments Based on Graffiti: A Statistical Mechanics Ap-
proach.” Physica A: Statistical Mechanics and Its Applications 392 (1):
252–70. https://doi.org/10.1016/j.physa.2012.08.001.
Barceló, Jaume, ed. 2010. Fundamentals of Traffic Simulation. Vol. 145.
International Series in Operations Research & Management Science. New
York, NY: Springer New York. https://doi.org/10.1007/978-1-4419-6142-
6.
Barnett, Susan M., and Stephen J. Ceci. 2002. “When and Where Do We
Apply What We Learn?: A Taxonomy for Far Transfer.” Psychological
Bulletin 128: 612–37. https://doi.org/10.1037/0033-2909.128.4.612.
Barton, Scott. 1994. “Chaos, Self-Organization, and Psychology.” American
Psychologist. https://doi.org/10.1037/0003-066X.49.1.5.
Bascompte, Jordi, and Pedro Jordano. 2013. Mutualistic Networks. Princeton
University Press.
Battiston, Federico, Enrico Amico, Alain Barrat, Ginestra Bianconi, Guil-
herme Ferraz de Arruda, Benedetta Franceschiello, Iacopo Iacopini, et al.
2021. “The Physics of Higher-Order Interactions in Complex Systems.”
Nature Physics 17 (10): 1093–98. https://doi.org/10.1038/s41567-021-
01371-4.
Baumann, Fabian, Philipp Lorenz-Spreen, Igor M. Sokolov, and Michele

195
 Starnini. 2020. “Modeling Echo Chambers and Polarization Dy-
namics in Social Networks.” Physical Review Letters 124 (4).
https://doi.org/10.1103/PhysRevLett.124.048301.
Bechtel, William, and Adele Abrahamsen. 2005. “Explanation: A Mechanist
Alternative.” Studies in History and Philosophy of Science Part C: Studies
in History and Philosophy of Biological and Biomedical Sciences, Mecha-
nisms in biology, 36 (2): 421–41. https://doi.org/10.1016/j.shpsc.2005.03.
010.
Bédard, C., H. Kröger, and A. Destexhe. 2006. “Does the $1/f$ Frequency
Scaling of Brain Signals Reflect Self-Organized Critical States?” Physical
Review Letters 97 (11): 118102. https://doi.org/10.1103/PhysRevLett.97
.118102.
Beek, Peter J., and Arthur Lewbel. 1995. “The Science of Juggling.” Scientific
American 273 (5): 92–97. https://doi.org/10.1038/scientificamerican1195-
92.
Beek, Peter J., C. E Peper, and A Daffertshofer. 2002. “Modeling Rhythmic
Interlimb Coordination: Beyond the Haken–Kelso–Bunz Model.” Brain
and Cognition 48 (1): 149–65. https://doi.org/10.1006/brcg.2001.1310.
Bentler, P. M. 1970. “Evidence Regarding Stages in the Development of Con-
servation.” Perceptual and Motor Skills 31 (3): 855–59. https://doi.org/
10.2466/pms.1970.31.3.855.
Berlekamp, Elwyn R., John H. Conway, and Richard K. Guy. 2004. Winning
Ways for Your Mathematical Plays, Volume 4. 2nd ed. New York: A K
Peters/CRC Press. https://doi.org/10.1201/9780429487309.
Bertalanffy, Ludwig Von. 1969. General System Theory: Foundations, Devel-
opment, Applications. Revised edition. New York, NY: George Braziller
Inc.
Bertschinger, Nils, and Thomas Natschläger. 2004. “Real-Time Computation
at the Edge of Chaos in Recurrent Neural Networks.” Neural Computation
16 (7): 1413–36. https://doi.org/10.1162/089976604323057443.
Blanken, Tessa F., Marie K. Deserno, Jonas Dalege, Denny Borsboom, Pe-
ter Blanken, Gerard A. Kerkhof, and Angélique O. J. Cramer. 2018.
“The Role of Stabilizing and Communicating Symptoms Given Overlap-
ping Communities in Psychopathology Networks.” Scientific Reports 8 (1):
5854. https://doi.org/10.1038/s41598-018-24224-2.
Boccaletti, S., G. Bianconi, R. Criado, C. I. del Genio, J. Gómez-Gardeñes,
M. Romance, I. Sendiña-Nadal, Z. Wang, and M. Zanin. 2014. “The
Structure and Dynamics of Multilayer Networks.” Physics Reports, The
structure and dynamics of multilayer networks, 544 (1): 1–122. https:
//doi.org/10.1016/j.physrep.2014.07.001.
Boccaletti, S., G. Bianconi, R. Criado, C. I. del Genio, J. Gómez-Gardeñes,
M. Romance, I. Sendiña-Nadal, Z. Wang, and M. Zanin. 2014. “The
Structure and Dynamics of Multilayer Networks.” Physics Reports, The
structure and dynamics of multilayer networks, 544 (1): 1–122. https:
//doi.org/10.1016/j.physrep.2014.07.001.
Boerlijst, M. C., and P. Hogeweg. 1991. “Spiral Wave Structure in Pre-Biotic
Evolution: Hypercycles Stable Against Parasites.” Physica D: Nonlinear
Phenomena 48 (1): 17–28. https://doi.org/10.1016/0167-2789(91)90049-
F.
Bogacz, Rafal, Eric Brown, Jeff Moehlis, Philip Holmes, and Jonathan D.
Cohen. 2006. “The Physics of Optimal Decision Making: A Formal Anal-

196
 ysis of Models of Performance in Two-Alternative Forced-Choice Tasks.”
Psychological Review 113: 700–765. https://doi.org/10.1037/0033-
295X.113.4.700.
Boostani, Reza, Foroozan Karimzadeh, and Mohammad Nami. 2017. “A
Comparative Review on Sleep Stage Classification Methods in Patients and
Healthy Individuals.” Computer Methods and Programs in Biomedicine 140
(March): 77–91. https://doi.org/10.1016/j.cmpb.2016.12.004.
Boot, Jesse, Maarten van der Ende, Reinout W. Wiers, Mike Lees, and Han
L. J. Van der Maas. submitted for publication. “Integrating Dual Process
Decision-Making and Social Dynamics: A Formal Modeling Framework for
Addiction,” submitted for publication.
Borsboom, Denny. 2008. “Psychometric Perspectives on Diagnostic Systems.”
Journal of Clinical Psychology 64 (9): 1089–1108. https://doi.org/10.100
2/jclp.20503.
———. 2017. “A Network Theory of Mental Disorders.” World Psychiatry 16
(1): 5–13. https://doi.org/10.1002/wps.20375.
Borsboom, Denny, Angélique O. J. Cramer, and Annemarie Kalis. 2019.
“Brain Disorders? Not Really: Why Network Structures Block Reduc-
tionism in Psychopathology Research.” Behavioral and Brain Sciences 42:
e2. https://doi.org/10.1017/S0140525X17002266.
Borsboom, Denny, Angélique O. J. Cramer, Verena D. Schmittmann, Sacha
Epskamp, and Lourens J. Waldorp. 2011. “The Small World of Psy-
chopathology.” PLOS ONE 6 (11): e27407. https://doi.org/10.1371/jour
nal.pone.0027407.
Borsboom, Denny, Eiko I. Fried, Sacha Epskamp, Lourens J. Waldorp, Clau-
dia D. van Borkulo, Han L. J. van der Maas, and Angélique O. J. Cramer.
2017. “False Alarm? A Comprehensive Reanalysis of ‘Evidence That Psy-
chopathology Symptom Networks Have Limited Replicability’ by Forbes,
Wright, Markon, and Krueger (2017).” Journal of Abnormal Psychology
126: 989–99. https://doi.org/10.1037/abn0000306.
Borsboom, Denny, Mijke Rhemtulla, Angélique O. J. Cramer, Han L. J. van
der Maas, M. Scheffer, and C. V. Dolan. 2016. “Kinds Versus Con-
tinua: A Review of Psychometric Approaches to Uncover the Structure
of Psychiatric Constructs.” Psychological Medicine 46 (8): 1567–79. https:
//doi.org/10.1017/S0033291715001944.
Borsboom, Denny, Donald J. Robinaugh, Mijke Rhemtulla, and Angélique
O. J. Cramer. 2018. “Robustness and Replicability of Psychopathology
Networks.” World Psychiatry 17 (2): 143–44. https://doi.org/10.1002/wp
s.20515.
Borsboom, Denny, Han L. J. van der Maas, Jonas Dalege, Rogier A Kievit,
and Brian D Haig. 2021. “Theory Construction Methodology: A Practical
Framework for Building Theories in Psychology.” Perspectives on Psycho-
logical Science 16 (4): 756–66. https://doi.org/10.1177/1745691620969647.
Bowles, Samuel, and Herbert Gintis. 2011. A Cooperative Species: Human
Reciprocity and Its Evolution. Princeton: Princeton University Press.
Brandt, Mark J. 2022. “Measuring the Belief System of a Person.” Journal
of Personality and Social Psychology 123 (4): 830–53. https://doi.org/10
.1037/pspp0000416.
Breakspear, Michael. 2017. “Dynamic Models of Large-Scale Brain Activity.”
Nature Neuroscience 20 (3): 340–52. https://doi.org/10.1038/nn.4497.
Bringmann, Laura F., Timon Elmer, Sacha Epskamp, Robert W. Krause,

197
 David Schoch, Marieke Wichers, Johanna T. W. Wigman, and Evelien
Snippe. 2019. “What Do Centrality Measures Measure in Psychological
Networks?” Journal of Abnormal Psychology 128: 892–903. https://doi.or
g/10.1037/abn0000446.
Brinkhuis, Matthieu J. S., Alexander O. Savi, Abe D. Hofman, Frederik
Coomans, Han L. J. van der Maas, and Gunter K. J. Maris. 2018. “Learn-
ing As It Happens: A Decade of Analyzing and Shaping a Large-Scale
Online Learning System.” Journal of Learning Analytics 5 (2): 29–46.
https://doi.org/10.18608/jla.2018.52.3.
Broer, Henk, and Floris Takens. 2010. Dynamical Systems and Chaos. 2011th
edition. New York: Springer.
Broido, Anna D., and Aaron Clauset. 2019. “Scale-Free Networks Are Rare.”
Nature Communications 10 (1): 1017. https://doi.org/10.1038/s41467-
019-08746-5.
Brush, Eleanor R., David C. Krakauer, and Jessica C. Flack. 2018. “Conflicts
of Interest Improve Collective Computation of Adaptive Social Structures.”
Science Advances 4 (1): e1603311. https://doi.org/10.1126/sciadv.16033
11.
Burg, Gerrit J. J. van den, and Christopher K. I. Williams. 2022. “An Eval-
uation of Change Point Detection Algorithms.” arXiv. https://doi.org/10
.48550/arXiv.2003.06222.
Burger, Julian, Adela-Maria Isvoranu, Gabriela Lunansky, Jonas M. B.
Haslbeck, Sacha Epskamp, Ria H. A. Hoekstra, Eiko I. Fried, Denny
Borsboom, and Tessa F. Blanken. 2022. “Reporting Standards for
Psychological Network Analyses in Cross-Sectional Data.” Psychological
Methods, No Pagination Specified–. https://doi.org/10.1037/met0000471.
Busemeyer, Jerome R., and Peter D. Bruza. 2012. Quantum Models of Cog-
nition and Decision. Cambridge University Press.
Carpenter, Gail A., and Stephen Grossberg. 1987. “A Massively Parallel
Architecture for a Self-Organizing Neural Pattern Recognition Machine.”
Computer Vision, Graphics, and Image Processing 37 (1): 54–115. https:
//doi.org/10.1016/S0734-189X(87)80014-2.
Caspi, Avshalom, Renate M. Houts, Daniel W. Belsky, Sidra J. Goldman-
Mellor, HonaLee Harrington, Salomon Israel, Madeline H. Meier, et al.
2014. “The p Factor: One General Psychopathology Factor in the Struc-
ture of Psychiatric Disorders?” Clinical Psychological Science : A Jour-
nal of the Association for Psychological Science 2 (2): 119–37. https:
//doi.org/10.1177/2167702613497473.
Castellano, Claudio, Santo Fortunato, and Vittorio Loreto. 2009. “Statistical
Physics of Social Dynamics.” Reviews of Modern Physics 81 (2): 591–646.
https://doi.org/10.1103/RevModPhys.81.591.
Castellano, Claudio, Matteo Marsili, and Alessandro Vespignani. 2000.
“Nonequilibrium Phase Transition in a Model for Social Influence.” Physi-
cal Review Letters 85 (16): 3536–39. https://doi.org/10.1103/PhysRevLet
t.85.3536.
Castellano, Claudio, Miguel A. Muñoz, and Romualdo Pastor-Satorras. 2009.
“Nonlinear $q$-Voter Model.” Physical Review E 80 (4): 041129. https:
//doi.org/10.1103/PhysRevE.80.041129.
Castro, L. N. de, and J. I. Timmis. 2003. “Artificial Immune Systems as a
Novel Soft Computing Paradigm.” Soft Computing 7 (8): 526–44. https:
//doi.org/10.1007/s00500-002-0237-z.

198
Cattell, R. B. 1987. Intelligence: Its Structure, Growth and Action. Elsevier.
Centola, Damon. 2022. “The Network Science of Collective Intelligence.”
Trends in Cognitive Sciences 26 (11): 923–41. https://doi.org/10.1016/j.ti
cs.2022.08.009.
Chalmers, David J. 2006. “Strong and Weak Emergence.” https://philpapers.org/rec/chasaw.
Chambers, Chris. 2017. “The Seven Deadly Sins of Psychology: A Manifesto
for Reforming the Culture of Scientific Practice.” In The Seven Deadly
Sins of Psychology. Princeton University Press. https://doi.org/10.1515/
9781400884940.
Chambon, Monique, Wesley G. Kammeraad, Frenk van Harreveld, Jonas
Dalege, Janneke E. Elberse, and Han L. J. van der Maas. 2022. “Under-
standing Change in COVID-19 Vaccination Intention with Network Anal-
ysis of Longitudinal Data from Dutch Adults.” Npj Vaccines 7 (1): 1–10.
https://doi.org/10.1038/s41541-022-00533-6.
Chang, Hasok. 2008. Inventing Temperature: Measurement and Scientific
Progress. Oxford Studies in Philosophy of Science. Oxford, New York:
Oxford University Press.
Chemero, Anthony. 2013. “Radical Embodied Cognitive Science.” Review of
General Psychology 17 (2): 145–50. https://doi.org/10.1037/a0032923.
Chen, Guanrong, and Yang Lou. 2019. Naming Game: Models, Simulations
and Analysis. Vol. 34. Emergence, Complexity and Computation. Cham:
Springer International Publishing. https://doi.org/10.1007/978-3-030-
05243-0.
Chialvo, Dante R. 2010. “Emergent Complex Neural Dynamics.” Nature
Physics 6 (10): 744–50. https://doi.org/10.1038/nphys1803.
Christensen, Alexander P., and Yoed N. Kenett. 2021. “Semantic Network
Analysis (SemNA): A Tutorial on Preprocessing, Estimating, and Analyz-
ing Semantic Networks.” Psychological Methods, No Pagination Specified–.
https://doi.org/10.1037/met0000463.
Clark, Andy. 2013. “Whatever Next? Predictive Brains, Situated Agents,
and the Future of Cognitive Science.” Behavioral and Brain Sciences 36
(3): 181–204. https://doi.org/10.1017/S0140525X12000477.
Clark, David M. 1986. “A Cognitive Approach to Panic.” Behaviour Research
and Therapy 24 (4): 461–70. https://doi.org/10.1016/0005-7967(86)90011-
2.
Clauset, Aaron, Cristopher Moore, and M. E. J. Newman. 2008. “Hierarchical
Structure and the Prediction of Missing Links in Networks.” Nature 453
(7191): 98–101. https://doi.org/10.1038/nature06830.
Cleeremans, Axel, Dalila Achoui, Arnaud Beauny, Lars Keuninckx, Jean-
Remy Martin, Santiago Muñoz-Moldes, Laurène Vuillaume, and Adélaı̈de
de Heering. 2020. “Learning to Be Conscious.” Trends in Cognitive Sci-
ences 24 (2): 112–23. https://doi.org/10.1016/j.tics.2019.11.011.
Cobb, Loren. 1978. “Stochastic Catastrophe Models and Multimodal Distri-
butions.” Behavioral Science 23 (4): 360–74. https://doi.org/10.1002/bs
.3830230407.
Cobb, Loren, and Shelemyahu Zacks. 1985. “Applications of Catastrophe
Theory for Statistical Modeling in the Biosciences.” Journal of the Ameri-
can Statistical Association 80 (392): 793–802. https://doi.org/10.1080/01
621459.1985.10478184.
Cocchi, Luca, Leonardo L. Gollo, Andrew Zalesky, and Michael Breakspear.
2017. “Criticality in the Brain: A Synthesis of Neurobiology, Models and

199
 Cognition.” Progress in Neurobiology 158 (November): 132–52. https:
//doi.org/10.1016/j.pneurobio.2017.07.002.
Cohen, Alexander, David Pargman, and Gershon Tenenbaum. 2003. “Critical
Elaboration and Empirical Investigation of the Cusp Catastrophe Model:
A Lesson for Practitioners.” Journal of Applied Sport Psychology 15 (2):
144–59. https://doi.org/10.1080/10413200305393.
Costantini, Giulio, Sacha Epskamp, Denny Borsboom, Marco Perugini, René
Mõttus, Lourens J. Waldorp, and Angélique O. J. Cramer. 2015. “State
of the aRt Personality Research: A Tutorial on Network Analysis of Per-
sonality Data in R.” Journal of Research in Personality, R Special Issue,
54 (February): 13–29. https://doi.org/10.1016/j.jrp.2014.07.003.
Cramer, Angélique O. J., Claudia D. van Borkulo, Erik J. Giltay, Han L. J.
van der Maas, Kenneth S. Kendler, Marten Scheffer, and Denny Borsboom.
2016. “Major Depression as a Complex Dynamic System.” PLOS ONE 11
(12): e0167490. https://doi.org/10.1371/journal.pone.0167490.
Cramer, Angélique O. J., Sophie Van Der Sluis, Arjen Noordhof, Marieke
Wichers, Nicole Geschwind, Steven H. Aggen, Kenneth S. Kendler, and
Denny Borsboom. 2012. “Dimensions of Normal Personality as Networks
in Search of Equilibrium: You Can’t Like Parties If You Don’t Like People.”
European Journal of Personality 26 (4): 414–31. https://doi.org/10.1002/
per.1866.
Cramer, Angélique O. J., Lourens J. Waldorp, Han L. J. van der Maas, and
Denny Borsboom. 2010. “Comorbidity: A Network Perspective.” The
Behavioral and Brain Sciences 33 (2-3): 137-150; discussion 150-193. https:
//doi.org/10.1017/S0140525X09991567.
Crielaard, Loes, Jeroen F Uleman, Bas D L Châtel, Sacha Epskamp, Pe-
ter M A Sloot, and Rick Quax. 2022. “Refining the Causal Loop Dia-
gram: A Tutorial for Maximizing the Contribution of Domain Expertise in
Computational System Dynamics Modeling.” Psychological Methods, May.
https://doi.org/10.1037/met0000484.
Cronbach, Lee J. 1957. “The Two Disciplines of Scientific Psychology.” Amer-
ican Psychologist 12: 671–84. https://doi.org/10.1037/h0043943.
Dablander, Fabian, and Max Hinne. 2019. “Node Centrality Measures Are
a Poor Substitute for Causal Inference.” Scientific Reports 9 (1): 6846.
https://doi.org/10.1038/s41598-019-43033-9.
Dablander, Fabian, Anton Pichler, Arta Cika, and Andrea Bacilieri. 2023.
“Anticipating Critical Transitions in Psychological Systems Using Early
Warning Signals: Theoretical and Practical Considerations.” Psychological
Methods 28 (4): 765–90. https://doi.org/10.1037/met0000450.
Dablander, Fabian, and Riet van Bork. 2021. “Causal Inference.” In Network
Psychometrics with R, 213–32. N Isvoranu, A. M., Epskamp, S., Waldorp,
L. J., & Borsboom, D. (Eds.). Network Psychometrics with R: A Guide
for Behavioral and Social Scientists. Routledge, Taylor & Francis Group.
Routledge.
Dakos, Vasilis, Stephen R. Carpenter, William A. Brock, Aaron M. Ellison,
Vishwesha Guttal, Anthony R. Ives, Sonia Kéfi, et al. 2012. “Methods for
Detecting Early Warnings of Critical Transitions in Time Series Illustrated
Using Simulated Ecological Data.” PloS One 7 (7): e41010. https://doi.or
g/10.1371/journal.pone.0041010.
Dalege, Jonas, Denny Borsboom, Frenk van Harreveld, and Han L. J. van der
Maas. 2017. “Network Analysis on Attitudes: A Brief Tutorial.” Social

200
 Psychological and Personality Science 8 (5): 528–37. https://doi.org/10.1
177/1948550617709827.
Dalege, Jonas, Denny Borsboom, Frenk van Harreveld, Helma van den Berg,
Mark Conner, and Han L. J. van der Maas. 2016. “Toward a Formal-
ized Account of Attitudes: The Causal Attitude Network (CAN) Model.”
Psychological Review 123 (1): 2–22. https://doi.org/10.1037/a0039802.
Dalege, Jonas, Denny Borsboom, Frenk van Harreveld, and Han L. J. van
der Maas. 2018. “The Attitudinal Entropy (AE) Framework as a General
Theory of Individual Attitudes.” Psychological Inquiry 29 (4): 175–93.
https://doi.org/10.1080/1047840X.2018.1537246.
Dalege, Jonas, and Han L. J. van der Maas. 2020. “Accurate by Being Noisy:
A Formal Network Model of Implicit Measures of Attitudes.” Social Cogni-
tion 38 (Suppl): S26–41. https://doi.org/10.1521/soco.2020.38.supp.s26.
Dandekar, Pranav, Ashish Goel, and David T. Lee. 2013. “Biased Assimila-
tion, Homophily, and the Dynamics of Polarization.” Proceedings of the
National Academy of Sciences 110 (15): 5791–96. https://doi.org/10.107
3/pnas.1217220110.
Dattagupta, S. 1981. “The Tricritical Point - a Qualitative Review.” Bull Mat
Sci 3 (2): 133–39. https://doi.org/10.1007/BF02908488.
Daugherty, Darryl, Tairi Roque-Urrea, John Urrea-Roque, Jessica Troyer,
Stephen Wirkus, and Mason A. Porter. 2009. “Mathematical Models of
Bipolar Disorder.” Communications in Nonlinear Science and Numerical
Simulation 14 (7): 2897–2908. https://doi.org/10.1016/j.cnsns.2008.10.02
7.
Dawkins, Richard. 1986. The blind watchmaker. New York : Norton. http:
//archive.org/details/blindwatchmaker0000dawk.
de Boer, Rob J. 2018. “Simple Phase Plane Analysis and Parameter Estima-
tion in R,” no. 20.
de Mooij, Susanne M. M., Tessa F. Blanken, Raoul P. P. P. Grasman, Jennifer
R. Ramautar, Eus J. W. Van Someren, and Han L. J. van der Maas. 2020.
“Dynamics of Sleep: Exploring Critical Transitions and Early Warning Sig-
nals.” Computer Methods and Programs in Biomedicine 193 (September):
105448. https://doi.org/10.1016/j.cmpb.2020.105448.
de Ron, Jill, Marie Deserno, Donald J. Robinaugh, Denny Borsboom, and Han
L. J. van der Maas. 2023. “Towards a General Modelling Framework of
Resource Competition in Cognitive Development.” OSF Preprints. https:
//doi.org/10.31219/osf.io/sh6w7.
Deffuant, Guillaume, David Neau, Frederic Amblard, and Gérard Weisbuch.
2000. “Mixing Beliefs Among Interacting Agents.” Advances in Complex
Systems (ACS) 03 (01n04): 87–98. https://doi.org/10.1142/S02195259000
00078.
Dekker, Mark M., Anna S. von der Heydt, and Henk A. Dijkstra. 2018. “Cas-
cading Transitions in the Climate System.” Earth System Dynamics 9 (4):
1243–60. https://doi.org/10.5194/esd-9-1243-2018.
DellaPosta, Daniel. 2020. “Pluralistic Collapse: The ‘Oil Spill’ Model of
Mass Opinion Polarization.” American Sociological Review 85 (3): 507–36.
https://doi.org/10.1177/0003122420922989.
Den Hartigh, Ruud J. R., Marijn W. G. Van Dijk, Henderien W. Steenbeek,
and Paul L. C. Van Geert. 2016. “A Dynamic Network Model to Explain
the Development of Excellent Human Performance.” Frontiers in Psychol-
ogy 7. https://doi.org/10.3389/fpsyg.2016.00532.

201
Deserno, Marie K., Denny Borsboom, Sander Begeer, Riet van Bork, Max
Hinne, and Hilde M. Geurts. 2020. “Highways to Happiness for Autistic
Adults? Perceived Causal Relations Among Clinicians.” PLOS ONE 15
(12): e0243298. https://doi.org/10.1371/journal.pone.0243298.
Dodson, M. M., and A. Hallam. 1977. “Allopatric Speciation and the Fold
Catastrophe.” The American Naturalist 111 (979): 415–33. https://doi.or
g/10.1086/283176.
Doerig, Adrien, Aaron Schurger, and Michael H. Herzog. 2021. “Hard Criteria
for Empirical Theories of Consciousness.” Cognitive Neuroscience 12 (2):
41–62. https://doi.org/10.1080/17588928.2020.1772214.
Dolan, Conor V., and Han L. J. van der Maas. 1998. “Fitting Multivariage
Normal Finite Mixtures Subject to Structural Equation Modeling.” Psy-
chometrika 63 (3): 227–53. https://doi.org/10.1007/BF02294853.
Dongen, Noah N’Djaye Nikolai van, Riet van Bork, Adam Finnemann, Han
L. J. van der Maas, Donald J. Robinaugh, Jonas Haslbeck, Jill de Ron,
Jan Sprenger, and Denny Borsboom. 2024. “Productive Explanation:
A Framework for Evaluating Explanations in Psychological Scienc.” Psy-
chologial Review. https://doi.org/10.31234/osf.io/qd69g.
Dorogovtsev, S. N., and J. F. F. Mendes. 2002. “Evolution of Networks.”
Advances in Physics 51 (4): 1079–1187. https://doi.org/10.1080/000187
30110112519.
Douglas, Karen M., Robbie M. Sutton, and Aleksandra Cichocka. 2017. “The
Psychology of Conspiracy Theories.” Current Directions in Psychological
Science 26: 538–42. https://doi.org/10.1177/0963721417718261.
Dresp-Langley, Birgitta. 2020. “Seven Properties of Self-Organization in the
Human Brain.” Big Data and Cognitive Computing 2 (4): 10.
Duckworth, Angela, and James J. Gross. 2014. “Self-Control and Grit: Re-
lated but Separable Determinants of Success.” Current Directions in Psy-
chological Science 23 (5): 319–25. https://doi.org/10.1177/096372141454
1462.
Durlauf, Steven N., and H. Peyton Young. 2001. Social Dynamics. MIT
Press.
Dutilh, Gilles, Eric-Jan Wagenmakers, Ingmar Visser, and van der Han L. J.
Maas. 2011. “A Phase Transition Model for the Speed-Accuracy Trade-
Off in Response Time Experiments.” Cognitive Science 35 (2): 211–50.
https://doi.org/10.1111/j.1551-6709.2010.01147.x.
Ebaugh, Helen Rose Fuchs. 1988. Becoming an Ex: The Process of Role Exit.
University of Chicago Press.
Edelman, Gerald M. 1987. Neural Darwinism: The Theory of Neuronal Group
Selection. Neural Darwinism: The Theory of Neuronal Group Selection.
New York, NY, US: Basic Books.
Eigen, Manfred, and Peter Schuster. 1979. The Hypercycle. Berlin, Heidel-
berg: Springer. https://doi.org/10.1007/978-3-642-67247-7.
Eldredge, Niles, and Stephen Jay Gould. 1972. “Punctuated Equilibria: An
Alternative to Phyletic Gradualism.” In Models in Paleobiology, edited by
Thomas J. M. Schopf, 82–115. Freeman Cooper.
Elo, Arpad E. 1978. The Rating of Chessplayers, Past and Present. New York:
Arco Pub.
Elovainio, Marko, Jari Lipsanen, Laura Pulkki-Råback, Jaana Suvisaari, and
Christian Hakulinen. 2021. “Is Symptom Connectivity Really the Most
Important Issue in Depression? Depression as a Dynamic System of In-

202
 terconnected Symptoms Revisited.” Journal of Psychiatric Research 142
(October): 250–57. https://doi.org/10.1016/j.jpsychires.2021.08.004.
Epskamp, Sacha, Denny Borsboom, and Eiko I. Fried. 2018. “Estimating
Psychological Networks and Their Accuracy: A Tutorial Paper.” Behavior
Research Methods 50 (1): 195–212. https://doi.org/10.3758/s13428-017-
0862-1.
Epskamp, Sacha, Angélique O. J. Cramer, Lourens J. Waldorp, Verena D.
Schmittmann, and Denny Borsboom. 2012. “Qgraph: Network Visual-
izations of Relationships in Psychometric Data.” Journal of Statistical
Software 48 (May): 1–18. https://doi.org/10.18637/jss.v048.i04.
Epskamp, Sacha, Mijke Rhemtulla, and Denny Borsboom. 2017. “Generalized
Network Psychometrics: Combining Network and Latent Variable Models.”
Psychometrika 82 (4): 904–27. https://doi.org/10.1007/s11336-017-9557-
x.
Epskamp, Sacha, Han L. J. van der Maas, Roseann E. Peterson, Hanna M.
van Loo, Steven H. Aggen, and Kenneth S. Kendler. 2022. “Intermediate
Stable States in Substance Use.” Addictive Behaviors 129 (June): 107252.
https://doi.org/10.1016/j.addbeh.2022.107252.
Epskamp, Sacha, Lourens J. Waldorp, René Mõttus, and Denny Borsboom.
2018. “The Gaussian Graphical Model in Cross-Sectional and Time-Series
Data.” Multivariate Behavioral Research 53 (4): 453–80. https://doi.org/
10.1080/00273171.2018.1454823.
Epstein, Joshua M. 2006. Generative Social Science: Studies in Agent-Based
Computational Modeling. Princeton University Press.
———. 2014. Agent_Zero: Toward Neurocognitive Foundations for Genera-
tive Social Science. Princeton University Press.
Epstein, Joshua M., and Robert Axtell. 1996. Growing Artificial Societies:
Social Science from the Bottom up. Growing Artificial Societies: Social
Science from the Bottom Up. Cambridge, MA, US: The MIT Press.
Eronen, Markus I., and Laura F. Bringmann. 2021. “The Theory Crisis in
Psychology: How to Move Forward.” Perspectives on Psychological Science
16 (4): 779–88. https://doi.org/10.1177/1745691620970586.
Ferri, Irene, Albert Dı́az-Guilera, and Matteo Palassini. 2022. “Equilibrium
and Dynamics of a Three-State Opinion Model.” arXiv. https://doi.org/
10.48550/arXiv.2210.03054.
Festinger, Leon. 1962. A Theory of Cognitive Dissonance. A Theory of
Cognitive Dissonance. Palo Alto, CA, US: Stanford Univer. Press.
Finnemann, Adam, Denny Borsboom, Sacha Epskamp, and Han L. J. van der
Maas. 2021. “The Theoretical and Statistical Ising Model: A Practical
Guide in R.” Psych 3 (4): 593–617. https://doi.org/10.3390/psych3040039.
Flache, Andreas, Michael Mäs, Thomas Feliciani, Edmund Chattoe-Brown,
Guillaume Deffuant, Sylvie Huet, and Jan Lorenz. 2017. “Models of Social
Influence: Towards the Next Frontiers.” Journal of Artificial Societies and
Social Simulation 20 (4): 2. https://doi.org/10.18564/jasss.3521.
Flack, Jessica C. 2017. “Coarse-Graining as a Downward Causation Mech-
anism.” Philosophical Transactions of the Royal Society A: Mathemat-
ical, Physical and Engineering Sciences 375 (2109): 20160338. https:
//doi.org/10.1098/rsta.2016.0338.
Fodor, J. A. 1974. “Special Sciences (or: The Disunity of Science as a Working
Hypothesis).” Synthese 28 (2): 97–115. https://doi.org/10.1007/BF0048
5230.

203
Forbes, Miriam K., Aidan G. C. Wright, Kristian E. Markon, and Robert
F. Krueger. 2017. “Evidence That Psychopathology Symptom Networks
Have Limited Replicability.” Journal of Abnormal Psychology 126: 969–88.
https://doi.org/10.1037/abn0000276.
Forger, D B, M E Jewett, and R E Kronauer. 1999. “A Simpler Model of
the Human Circadian Pacemaker.” Journal of Biological Rhythms 14 (6):
532–37. https://doi.org/10.1177/074873099129000867.
Forrester, Jay W. 1993. “System Dynamics and the Lessons of 35 Years.” In A
Systems-Based Approach to Policymaking, edited by Kenyon B. De Greene,
199–240. Boston, MA: Springer US. https://doi.org/10.1007/978-1-4615-
3226-2_7.
Fortmann-Roe, Scott. 2014. “Insight Maker: A General-Purpose Tool for
Web-Based Modeling & Simulation.” Simulation Modelling Practice and
Theory 47 (September): 28–45. https://doi.org/10.1016/j.simpat.2014.03
.013.
Fortunato, Santo. 2004. “Universality of the Threshold for Complete Consen-
sus for the Opinion Dynamics of Deffuant Et Al.” International Journal
of Modern Physics C 15 (January): 1301–7. https://doi.org/10.1142/S0
129183104006728.
Fowler, James H. 2005. “Second-Order Free-Riding Problem Solved?” Nature
437 (7058): E8–8. https://doi.org/10.1038/nature04201.
Franz, David J. 2022. “‘Are Psychological Attributes Quantitative?’ Is Not an
Empirical Question: Conceptual Confusions in the Measurement Debate.”
Theory & Psychology 32 (1): 131–50. https://doi.org/10.1177/0959354321
1045340.
Fraser, Steven. 2008. The Bell Curve Wars: Race, Intelligence, and the Future
of America. Basic Books.
Freitas, Ubiratan, Elise Roulin, Jean-François Muir, and Christophe Letellier.
2009. “Identifying Chaos from Heart Rate: The Right Task?” Chaos:
An Interdisciplinary Journal of Nonlinear Science 19 (2): 028505. https:
//doi.org/10.1063/1.3139116.
Fried, Eiko I., and Angélique O. J. Cramer. 2017. “Moving Forward: Chal-
lenges and Directions for Psychopathological Network Theory and Method-
ology.” Perspectives on Psychological Science 12 (6): 999–1020. https:
//doi.org/10.1177/1745691617705892.
Fried, Eiko I., Sacha Epskamp, Randolph M. Nesse, Francis Tuerlinckx, and
Denny Borsboom. 2016. “What Are ’Good’ Depression Symptoms? Com-
paring the Centrality of DSM and Non-DSM Symptoms of Depression in a
Network Analysis.” Journal of Affective Disorders 189 (January): 314–20.
https://doi.org/10.1016/j.jad.2015.09.005.
Friedkin, Noah E., and Eugene C. Johnsen. 1990. “Social Influence and
Opinions.” The Journal of Mathematical Sociology 15 (3-4): 193–206. ht
tps://doi.org/10.1080/0022250X.1990.9990069.
Friedman, Jerome, Trevor Hastie, and Robert Tibshirani. 2008. “Sparse In-
verse Covariance Estimation with the Graphical Lasso.” Biostatistics 9 (3):
432–41. https://doi.org/10.1093/biostatistics/kxm045.
Friston, Karl. 2009. “The Free-Energy Principle: A Rough Guide to the
Brain?” Trends in Cognitive Sciences 13 (7): 293–301. https://doi.org/10
.1016/j.tics.2009.04.005.
Fuchs, Armin, and Scott Kelso. 2018. “Coordination Dynamics and Synerget-
ics: From Finger Movements to Brain Patterns and Ballet Dancing.” In

204
 Complexity and Synergetics, 301–16. https://doi.org/10.1007/978-3-319-
64334-2_23.
Fürstenau, Norbert. 2014. “Simulating Bistable Perception with Interrupted
Ambiguous Stimulus Using Self-Oscillator Dynamics with Percept Choice
Bifurcation.” Cognitive Processing 15 (4): 467–90. https://doi.org/10.100
7/s10339-014-0630-4.
Fytas, Nikolaos G., Victor Martin-Mayor, Marco Picco, and Nicolas Sourlas.
2018. “Review of Recent Developments in the Random-Field Ising Model.”
Journal of Statistical Physics 172 (2): 665–72. https://doi.org/10.1007/s1
0955-018-1955-7.
Galam, Serge. 2008. “Sociophysics: A Review of Galam Models.” Interna-
tional Journal of Modern Physics C 19 (03): 409–40. https://doi.org/10.1
142/S0129183108012297.
Galam, Serge, Yuval Gefen, and Yonathan Shapir. 1982. “Sociophysics: A
New Approach of Sociological Collective Behaviour. I. Mean-Behaviour
Description of a Strike.” The Journal of Mathematical Sociology 9 (1):
1–13. https://doi.org/10.1080/0022250X.1982.9989929.
Gale, D., and L. S. Shapley. 1962. “College Admissions and the Stability
of Marriage.” The American Mathematical Monthly 69 (1): 9–15. https:
//doi.org/10.2307/2312726.
Galesic, Mirta, Daniel Barkoczi, Andrew M. Berdahl, Dora Biro, Giuseppe
Carbone, Ilaria Giannoccaro, Robert L. Goldstone, et al. 2023. “Beyond
Collective Intelligence: Collective Adaptation.” Journal of The Royal Soci-
ety Interface 20 (200): 20220736. https://doi.org/10.1098/rsif.2022.0736.
Galton, Francis. 1907. “Vox Populi.” Nature 75 (1949): 450–51. https:
//doi.org/10.1038/075450a0.
Garc, P., and Callegari Pe. 2015. “Analysis of EEG Signals Using Nonlinear
Dynamics and Chaos: A Review.” In. https://www.semanticscholar.or
g/paper/Analysis-of-EEG-Signals-using-Nonlinear-Dynamics-A-Garc-
Pe/7c7bae15b868a68049d017edd27031c2746e20e4.
Garcez, Artur d’Avila, and Luı́s C. Lamb. 2023. “Neurosymbolic AI: The 3rd
Wave.” Artificial Intelligence Review 56 (11): 12387–406. https://doi.org/
10.1007/s10462-023-10448-w.
Gärtner, Klaus. 1990. “A Third Component Causing Random Variability
Beside Environment and Genotype. A Reason for the Limited Success of
a 30 Year Long Effort to Standardize Laboratory Animals?” Laboratory
Animals 24 (1): 71–77. https://doi.org/10.1258/002367790780890347.
Gauvin, L., J. Vannimenus, and J.-P. Nadal. 2009. “Phase Diagram of a
Schelling Segregation Model.” The European Physical Journal B 70 (2):
293–304. https://doi.org/10.1140/epjb/e2009-00234-0.
Ghatak, Abhijit. 2019. Deep Learning with R. Singapore: Springer. https:
//doi.org/10.1007/978-981-13-5850-0.
Gibson, James J. 2014. The Ecological Approach to Visual Perception: Classic
Edition. New York: Psychology Press. https://doi.org/10.4324/97813157
40218.
Gilmore, Robert. 1993. Catastrophe Theory for Scientists and Engineers.
Courier Corporation.
Goldstone, Robert L., and Marco A. Janssen. 2005. “Computational Models
of Collective Behavior.” Trends in Cognitive Sciences 9 (9): 424–30. https:
//doi.org/10.1016/j.tics.2005.07.009.
Gottman, John M., and Robert W. Levenson. 1992. “Marital Processes Pre-

205
 dictive of Later Dissolution: Behavior, Physiology, and Health.” Journal
of Personality and Social Psychology 63: 221–33. https://doi.org/10.103
7/0022-3514.63.2.221.
Gottman, John M., James D. Murray, Catherine C. Swanson, Rebecca Tyson,
and Kristin R. Swanson. 2002. The Mathematics of Marriage: Dynamic
Nonlinear Models. The Mathematics of Marriage: Dynamic Nonlinear
Models. Cambridge, MA, US: MIT Press.
Granovetter, Mark. 1973. “The Strength of Weak Ties.” American Journal
of Sociology 78 (6): 1360–80. https://doi.org/10.1086/225469.
———. 1978. “Threshold Models of Collective Behavior.” The American
Journal of Sociology 83 (6): 1420–43. https://doi.org/10.1086/226707.
Grasman, Raoul, Han L. J. van der Maas, and Eric-Jan Wagenmakers. 2009.
“Fitting the Cusp Catastrophe in R: A Cusp Package Primer.” Journal of
Statistical Software 032 (i08). https://ideas.repec.org/a/jss/jstsof/v032i08
.html.
Grauwin, Sebastian, Guillaume Beslon, Éric Fleury, Sara Franceschelli, Ce-
line Robardet, Jean-Baptiste Rouquier, and Pablo Jensen. 2012. “Com-
plex Systems Science: Dreams of Universality, Interdisciplinarity Reality.”
Journal of the American Society for Information Science and Technology
63 (7): 1327–38. https://doi.org/10.1002/asi.22644.
Groot, S. R. De, and P. Mazur. 2013. Non-Equilibrium Thermodynamics.
Courier Corporation.
Grossberg, Stephen, and Baingio Pinna. 2012. “Neural Dynamics of Gestalt
Principles of Perceptual Organization: From Grouping to Shape and Mean-
ing.” GESTALT THEORY 34.
Guastello, Stephen J. 1982. “Moderator Regression and the Cusp Catas-
trophe: Application of Two-Stage Personnel Selection, Training, Ther-
apy, and Policy Evaluation.” Behavioral Science 27 (3): 259–72. https:
//doi.org/10.1002/bs.3830270305.
———. 1984. “Cusp and Butterfly Catastrophe Modeling of Two Opponent
Process Models: Drug Addiction and Work Performance.” Behavioral Sci-
ence 29 (4): 258262. https://www.proquest.com/docview/1301285594/ci
tation/C898CBB773DF4B72PQ/1.
Guastello, Stephen J., Matthijs Koopmans, and David Pincus. 2008. Chaos
and Complexity in Psychology: The Theory of Nonlinear Dynamical Sys-
tems. Cambridge University Press.
Guckenheimer, John, and Philip Holmes. 1983. “Global Bifurcations.” In
Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector
Fields, edited by John Guckenheimer and Philip Holmes, 289–352. Applied
Mathematical Sciences. New York, NY: Springer. https://doi.org/10.100
7/978-1-4612-1140-2_6.
Güémez, J., C. Fiolhais, and M. Fiolhais. 2002. “The Cartesian Diver and the
Fold Catastrophe.” American Journal of Physics 70 (7): 710–14. https:
//doi.org/10.1119/1.1477433.
Guimerà, Roger, Brian Uzzi, Jarrett Spiro, and Luı́s A. Nunes Amaral. 2005.
“Team Assembly Mechanisms Determine Collaboration Network Structure
and Team Performance.” Science 308 (5722): 697–702. https://doi.org/10
.1126/science.1106340.
Haig, Brian D. 2005. “An Abductive Theory of Scientific Method.” Psycholog-
ical Methods 10 (4): 371–88. https://doi.org/10.1037/1082-989X.10.4.371.
———. 2014. Investigating the Psychological World: Scientific Method in the

206
 Behavioral Sciences. MIT Press.
Haken, Hermann. 1977. “Synergetics.” Physics Bulletin 28 (9): 412. https:
//doi.org/10.1088/0031-9112/28/9/027.
———. 1992. “Synergetics in Psychology.” In Self-Organization and Clinical
Psychology: Empirical Approaches to Synergetics in Psychology, edited by
Wolfgang Tschacher, Günter Schiepek, and Ewald Johannes Brunner, 32–
54. Springer Series in Synergetics. Berlin, Heidelberg: Springer. https:
//doi.org/10.1007/978-3-642-77534-5_2.
Haken, Hermann, J. A. S. Kelso, and H. Bunz. 1985. “A Theoretical Model
of Phase Transitions in Human Hand Movements.” Biological Cybernetics
51 (5): 347–56. https://doi.org/10.1007/BF00336922.
Hamaker, Ellen L., E. Ceulemans, R. P. P. P. Grasman, and F. Tuerlinckx.
2015. “Modeling Affect Dynamics: State of the Art and Future Challenges.”
Emotion Review 7 (4): 316–22. https://doi.org/10.1177/17540739155906
19.
Hamaker, Ellen L., Zhiyong Zhang, and Han L. J. van der Maas. 2009. “Us-
ing Threshold Autoregressive Models to Study Dyadic Interactions.” Psy-
chometrika 74 (4): 727–45. https://doi.org/10.1007/s11336-009-9113-4.
Hameroff, Stuart, and Roger Penrose. 1996. “Orchestrated Reduction of
Quantum Coherence in Brain Microtubules: A Model for Consciousness.”
Mathematics and Computers in Simulation 40 (3): 453–80. https://doi.or
g/10.1016/0378-4754(96)80476-9.
Hardin, Garrett. 1968. “The Tragedy of the Commons” 162.
Hardy, Lew. 1996. “Testing the Predictions of the Cusp Catastrophe Model of
Anxiety and Performance.” The Sport Psychologist 10 (2): 140–56. https:
//doi.org/10.1123/tsp.10.2.140.
Hardy, Lew, and Gaynor Parfitt. 1991. “A Catastrophe Model of Anxiety
and Performance.” British Journal of Psychology 82 (2): 163–78. https:
//doi.org/10.1111/j.2044-8295.1991.tb02391.x.
Hardy, Lew, Tim Woodman, and Stephen Carrington. 2004. “Is Self-
Confidence a Bias Factor in Higher-Order Catastrophe Models? An
Exploratory Analysis.” Journal of Sport and Exercise Psychology 26 (3):
359–68. https://doi.org/10.1123/jsep.26.3.359.
Harezlak, Katarzyna, and Pawel Kasprowski. 2018. “Searching for Chaos
Evidence in Eye Movement Signals.” Entropy 20 (1): 32. https://doi.org/
10.3390/e20010032.
Haslbeck, Jonas M. B., and Oisı́n Ryan. 2022. “Recovering Within-Person Dy-
namics from Psychological Time Series.” Multivariate Behavioral Research
57 (5): 735–66. https://doi.org/10.1080/00273171.2021.1896353.
Hassell, M. P., J. H. Lawton, and R. M. May. 1976. “Patterns of Dynamical
Behaviour in Single-Species Populations.” Journal of Animal Ecology 45
(2): 471–86. https://doi.org/10.2307/3886.
Hayes, Adele M., and Leigh A. Andrews. 2020. “A Complex Systems Ap-
proach to the Study of Change in Psychotherapy.” BMC Medicine 18 (1):
197. https://doi.org/10.1186/s12916-020-01662-2.
Heath, Richard A. 2000. Nonlinear Dynamics: Techniques and Applications
in Psychology. 1st edition. Mahwah, N.J: Psychology Press.
Hebb, D. O. 1949. The Organization of Behavior; a Neuropsychological The-
ory. The Organization of Behavior; a Neuropsychological Theory. Oxford,
England: Wiley.
Heckman, James J. 1995. “Lessons from the Bell Curve.” Journal of Political

207
 Economy 103 (5): 1091–1120. https://doi.org/10.1086/262014.
Hegselmann, Rainer. 2017. “Thomas C. Schelling and James M. Sakoda: The
Intellectual, Technical, and Social History of a Model.” Journal of Artificial
Societies and Social Simulation 20 (3): 15. https://doi.org/10.18564/jasss
.3511.
Heider, Fritz. 1946. “Attitudes and Cognitive Organization.” The Journal of
Psychology 21 (1): 107–12. https://doi.org/10.1080/00223980.1946.9917
275.
Helbing, Dirk, and Péter Molnár. 1995. “Social Force Model for Pedestrian
Dynamics.” Physical Review E 51 (5): 4282–86. https://doi.org/10.1103/
PhysRevE.51.4282.
Hendrikse, Sophie, Jan Treur, and Sander Koole. 2023. “Modeling Emerging
Interpersonal Synchrony and Its Related Adaptive Short-Term Affiliation
and Long-Term Bonding: A Second-Order Multi-Adaptive Neural Agent
Model.” International Journal of Neural Systems, April. https://doi.org/
10.1142/S0129065723500387.
Heylighen, Francis. 2009. “Complexity and Self-Organization.” In, 3rd ed.
CRC Press.
Hirsch, Morris W. 1985. “Systems of Differential Equations That Are Compet-
itive or Cooperative II: Convergence Almost Everywhere.” SIAM Journal
on Mathematical Analysis 16 (3): 423–39. https://doi.org/10.1137/051603
0.
Hoel, Erik P., Larissa Albantakis, and Giulio Tononi. 2013. “Quantifying
Causal Emergence Shows That Macro Can Beat Micro.” Proceedings of
the National Academy of Sciences 110 (49): 19790–95. https://doi.org/10
.1073/pnas.1314922110.
Hoffmann, Geoffrey W, Maurice W Benson, Geoffrey M Bree, and Paul E Ki-
nahan. 1986. “A Teachable Neural Network Based on an Unorthodox Neu-
ron.” Physica D: Nonlinear Phenomena, Proceedings of the Fifth Annual
International Conference, 22 (1): 233–46. https://doi.org/10.1016/0167-
2789(86)90243-5.
Hofstadter, Douglas R. 1979. Gödel, Escher, Bach: An Eternal Golden Braid.
Vintage Books.
———. 2007. I Am a Strange Loop. New York: Basic Books.
Holland, John H. 1992a. “Genetic Algorithms.” Scientific American 267 (1):
66–73. https://doi.org/10.1038/scientificamerican0792-66.
———. 1992b. Adaptation in Natural and Artificial Systems: An Introductory
Analysis with Applications to Biology, Control, and Artificial Intelligence.
MIT Press.
Holyst, Janusz A., Krzysztof Kacperski, and Frank Schweitzer. 2001. “Social
Impact Models of Opinion Dynamics.” In Annual Reviews of Computa-
tional Physics IX, Volume 9:253–73. Annual Reviews of Computational
Physics. WORLD SCIENTIFIC. https://doi.org/10.1142/978981281157
8_0005.
Huth, Karoline, Jill de Ron, Judy Luigjes, Anneke Goudriaan, Reza Moham-
madi, Ruth van Holst, Eric-Jan Wagenmakers, and Maarten Marsman.
2023. “Bayesian Analysis of Cross-sectional Networks: A Tutorial in R
and JASP.” PsyArXiv. https://doi.org/10.31234/osf.io/ub5tc.
Isvoranu, Adela-Maria, Sacha Epskamp, Lourens J. Waldorp, and Denny Bors-
boom. 2022. Network Psychometrics with R: A Guide for Behavioral and
Social Scientists. Taylor & Francis Limited.

208
Izhikevich, Eugene M. 1998. “Multiple Cusp bifurcations1This Paper Received
the SIAM Award as the Best Student Paper in Applied Mathematics in
1995. It Was Written While the Author Was a Graduate Student at De-
partment of Mathematics, Michigan State University, and Was Supported
in Part by NSF Grant DMS 9206677.12This Work Could Not Be Accom-
plished Without Dr. Frank Hoppensteadt. I Am Very Grateful for All
Kinds of Investments He Made in Me and My Research.23This Paper Is
Dedicated to Frank C. Hoppensteadt on the Occasion of His 60th Birth-
day.3.” Neural Networks 11 (3): 495–508. https://doi.org/10.1016/S0893-
6080(97)00117-2.
Izhikevich, Eugene M., and Richard FitzHugh. 2006. “FitzHugh-Nagumo
Model.” Scholarpedia 1 (9): 1349. https://doi.org/10.4249/scholarpedia.1
349.
Jacobson, Michael J., James A. Levin, and Manu Kapur. 2019. “Education
as a Complex System: Conceptual and Methodological Implications.” Ed-
ucational Researcher 48 (2): 112–19. https://doi.org/10.3102/0013189X
19826958.
Jager, Wander. 2017. “Enhancing the Realism of Simulation (EROS): On
Implementing and Developing Psychological Theory in Social Simulation.”
Journal of Artificial Societies and Social Simulation 20 (3): 1–14. https:
//doi.org/10.18564/jasss.3522.
Jager, Wander, and Frédéric Amblard. 2005. “Uniformity, Bipolarization and
Pluriformity Captured as Generic Stylized Behavior with an Agent-Based
Simulation Model of Attitude Change.” Computational & Mathematical
Organization Theory 10 (4): 295–303. https://doi.org/10.1007/s10588-
005-6282-2.
James, Nicholas A., and David S. Matteson. 2014. “Ecp : An R Package
for Nonparametric Multiple Change Point Analysis of Multivariate Data.”
Journal of Statistical Software 62 (7). https://doi.org/10.18637/jss.v062.
i07.
Jansen, Brenda R. J., and Han L. J. van der Maas. 2001. “Evidence for
the Phase Transition from Rule I to Rule II on the Balance Scale Task.”
Developmental Review 21 (4): 450494. https://doi.org/10.1006/drev.2001.
0530.
Jędrzejewski, Arkadiusz, and Katarzyna Sznajd-Weron. 2019. “Statistical
Physics Of Opinion Formation: Is It a SPOOF?” Comptes Rendus Physique
20 (4): 244–61. https://doi.org/10.1016/j.crhy.2019.05.002.
Johnson, Chris J., Matthew A. Mumma, and Martin-Hugues St-Laurent.
2019. “Modeling Multispecies Predator–prey Dynamics: Predicting the
Outcomes of Conservation Actions for Woodland Caribou.” Ecosphere 10
(3): e02622. https://doi.org/10.1002/ecs2.2622.
Jones, Payton J., Ruofan Ma, and Richard J. McNally. 2021. “Bridge Central-
ity: A Network Approach to Understanding Comorbidity.” Multivariate
Behavioral Research 56 (2): 353–67. https://doi.org/10.1080/00273171.2
019.1614898.
Jusup, Marko, Petter Holme, Kiyoshi Kanazawa, Misako Takayasu,
Ivan Romić, Zhen Wang, Sunčana Geček, et al. 2022. “Social
Physics.” Physics Reports, Social physics, 948 (February): 1–148.
https://doi.org/10.1016/j.physrep.2021.10.005.
Kalantari, Somayeh, Eslam Nazemi, and Behrooz Masoumi. 2020. “Emer-
gence Phenomena in Self-Organizing Systems: A Systematic Literature

209
 Review of Concepts, Researches, and Future Prospects.” Journal of Or-
ganizational Computing and Electronic Commerce 30 (3): 224–65. https:
//doi.org/10.1080/10919392.2020.1748977.
Kalisch, Raffael, Angélique O. J. Cramer, Harald Binder, Jessica Fritz, IJs-
brand Leertouwer, Gabriela Lunansk, Benjamin Meyer, Jens Timiner, Ilya
M. Veer, and Anne-Laura Van Harmelen. 2019. “Deconstructing and Re-
constructing Resilience: A Dynamic Network Approach.” Perspectives on
Psychological Science 14 (5): 765–77. https://doi.org/10.1177/17456916
19855637.
Karau, Steven J., and Kipling D. Williams. 1993. “Social Loafing: A Meta-
Analytic Review and Theoretical Integration.” Journal of Personality and
Social Psychology 65: 681–706. https://doi.org/10.1037/0022-3514.65.4.
681.
Kargarnovin, Shaida, Christopher Hernandez, Farzad V. Farahani, and Walde-
mar Karwowski. 2023. “Evidence of Chaos in Electroencephalogram Sig-
natures of Human Performance: A Systematic Review.” Brain Sciences 13
(5): 813. https://doi.org/10.3390/brainsci13050813.
Kauffman, Stuart A. 1993. The Origins of Order: Self-Organization and
Selection in Evolution. 1st edition. New York: Oxford University Press.
Kelso, J. A. S. 1995. Dynamic Patterns: The Self-Organization of Brain and
Behavior. Dynamic Patterns: The Self-Organization of Brain and Behavior.
Cambridge, MA, US: The MIT Press.
———. 2021. “The Haken–Kelso–Bunz (HKB) Model: From Matter to
Movement to Mind.” Biological Cybernetics 115 (4): 305–22. https:
//doi.org/10.1007/s00422-021-00890-w.
Kelso, J. A. S., J. P. Scholz, and G. Schöner. 1986. “Nonequilibrium Phase
Transitions in Coordinated Biological Motion: Critical Fluctuations.”
Physics Letters A 118 (6): 279–84. https://doi.org/10.1016/0375-
9601(86)90359-2.
Kim, Jaegwon. 2006. “Emergence: Core Ideas and Issues.” Synthese 151 (3):
547–59. https://doi.org/10.1007/s11229-006-9025-0.
Klinkenberg, S., M Straatemeier, and Han L. J. van der Maas. 2011. “Com-
puter Adaptive Practice of Maths Ability Using a New Item Response
Model for on the Fly Ability and Difficulty Estimation.” Computers &
Education 57: 1813–24.
Klose, Ann Kristin, Volker Karle, Ricarda Winkelmann, and Jonathan F.
Donges. 2020. “Emergence of Cascading Dynamics in Interacting Tip-
ping Elements of Ecology and Climate.” Royal Society Open Science 7 (6):
200599. https://doi.org/10.1098/rsos.200599.
Knyspel, Jacob, and Robert Plomin. 2024. “Comparing Factor and Network
Models of Cognitive Abilities Using Twin Data.” Intelligence 104 (May):
101833. https://doi.org/10.1016/j.intell.2024.101833.
Kossakowski, Jolanda J., Lourens J. Waldorp, and Han L. J. van der Maas.
2021. “The Search for Causality: A Comparison of Different Techniques
for Causal Inference Graphs.” Psychological Methods 26: 719–42. https:
//doi.org/10.1037/met0000390.
Kosterlitz, J. M. 1974. “The Critical Properties of the Two-Dimensional Xy
Model.” Journal of Physics C: Solid State Physics 7 (6): 1046. https:
//doi.org/10.1088/0022-3719/7/6/005.
Krakauer, David C. 2023. “Symmetry–simplicity, Broken Symmetry–
complexity.” Interface Focus 13 (3): 20220075. https://doi.org/10.1098/

210
 rsfs.2022.0075.
———. 2024. “The Complex World: An Introduction to the Foundations
of Complexity Science.” In Foundational Papers in Complexity Science,
edited by David C. Krakauer, 1:Intro-1-99. Santa Fe, NM: SFI Press.
Kreitler, Shulamith, and Hans Kreitler. 1989. “Horizontal Decalage: A Prob-
lem and Its Solution.” Cognitive Development 4 (1): 89–119. https:
//doi.org/10.1016/0885-2014(89)90006-3.
Kruis, Joost, Gunter K. J. Maris, Maarten Marsman, Maria Bolsinova, and
Han L. J. van der Maas. 2020. “Deviations of Rational Choice: An Integra-
tive Explanation of the Endowment and Several Context Effects.” Scien-
tific Reports 10 (1): 16226. https://doi.org/10.1038/s41598-020-73181-2.
Kruse, Peter, and Michael Stadler. 2012. Ambiguity in Mind and Nature:
Multistable Cognitive Phenomena. Springer Science & Business Media.
Kuramoto, Yoshiki. 1984. “Chemical Turbulence.” In Chemical Oscillations,
Waves, and Turbulence, edited by Yoshiki Kuramoto, 111–40. Springer
Series in Synergetics. Berlin, Heidelberg: Springer. https://doi.org/10.1
007/978-3-642-69689-3_7.
Ladyman, James, James Lambert, and Karoline Wiesner. 2013. “What Is
a Complex System?” European Journal for Philosophy of Science 3 (1):
33–67. https://doi.org/10.1007/s13194-012-0056-8.
Lange, Jens, and Janis H. Zickfeld. 2021. “Emotions as Overlapping Causal
Networks of Emotion Components: Implications and Methodological Ap-
proaches.” Emotion Review 13 (2): 157–67. https://doi.org/10.1177/1754
073920988787.
Langton, Chris G. 1990. “Computation at the Edge of Chaos: Phase Transi-
tions and Emergent Computation.” Physica D: Nonlinear Phenomena 42
(1): 12–37. https://doi.org/10.1016/0167-2789(90)90064-V.
Latané, Bibb. 1981. “The Psychology of Social Impact.” American Psycholo-
gist 36: 343–56. https://doi.org/10.1037/0003-066X.36.4.343.
Latané, Bibb, and Andrzej Nowak. 1994. “Attitudes as Catastrophes: From
Dimensions to Categories with Increasing Involvement.” In Dynamical Sys-
tems in Social Psychology, 219–49. San Diego, CA, US: Academic Press.
Leemput, van de Ingrid A., Marieke Wichers, Angélique O. J. Cramer, Denny
Borsboom, Francis Tuerlinckx, Peter Kuppens, van Egbert H. Nes, et al.
2014. “Critical Slowing down as Early Warning for the Onset and Termi-
nation of Depression.” Proceedings of the National Academy of Sciences
111 (1): 87–92. https://doi.org/10.1073/pnas.1312114110.
Leggio, Lorenzo, George A. Kenna, Miriam Fenton, Erica Bonenfant, and
Robert M. Swift. 2009. “Typologies of Alcohol Dependence. From Jellinek
to Genetics and Beyond.” Neuropsychology Review 19 (1): 115–29. https:
//doi.org/10.1007/s11065-008-9080-z.
Lei, Min, Zhizhong Wang, and Zhengjin Feng. 2001. “Detecting Nonlinearity
of Action Surface EMG Signal.” Physics Letters A 290 (5): 297–303. https:
//doi.org/10.1016/S0375-9601(01)00668-5.
Lemke, Jay L., and Nora H. Sabelli. 2008. “Complex Systems and Educational
Change: Towards a New Research Agenda.” In Complexity Theory and
the Philosophy of Education, 112–23. John Wiley & Sons, Ltd. https:
//doi.org/10.1002/9781444307351.ch8.
Lewenstein, Maciej, Andrzej Nowak, and Bibb Latané. 1992. “Statistical
Mechanics of Social Impact.” Physical Review A 45 (2): 763–76. https:
//doi.org/10.1103/PhysRevA.45.763.

211
Li, Ke, Haiming Liang, Gang Kou, and Yucheng Dong. 2020. “Opinion Dy-
namics Model Based on the Cognitive Dissonance: An Agent-Based Simu-
lation.” Information Fusion 56 (April): 1–14. https://doi.org/10.1016/j.
inffus.2019.09.006.
Loehle, Craig. 1989. “Catastrophe Theory in Ecology: A Critical Review and
an Example of the Butterfly Catastrophe.” Ecological Modelling 49 (1):
125–52. https://doi.org/10.1016/0304-3800(89)90047-1.
Lorenz, Edward N. 1963. “Deterministic Nonperiodic Flow.” Journal of the
Atmospheric Sciences 20 (2): 130–41. https://doi.org/10.1175/1520-
0469(1963)020%3C0130:DNF%3E2.0.CO;2.
Lorenz, Jan. 2007. “Continuous Opinion Dynamics Under Bounded Confi-
dence: A Survey.” International Journal of Modern Physics C 18 (12):
1819–38. https://doi.org/10.1142/S0129183107011789.
Love, Jonathon, Ravi Selker, Maarten Marsman, Tahira Jamil, Damian Drop-
mann, Josine Verhagen, Alexander Ly, et al. 2019. “JASP: Graphical
Statistical Software for Common Statistical Designs.” Journal of Statisti-
cal Software 88 (January): 1–17. https://doi.org/10.18637/jss.v088.i02.
Lowery, Megan R., Malissa A. Clark, and Nathan T. Carter. 2021. “The
Balancing Act of Performance: Psychometric Networks and the Causal
Interplay of Organizational Citizenship and Counterproductive Work Be-
haviors.” Journal of Vocational Behavior 125 (March): 103527. https:
//doi.org/10.1016/j.jvb.2020.103527.
Ludwig, D., D. D. Jones, and C. S. Holling. 1978. “Qualitative Analysis of
Insect Outbreak Systems: The Spruce Budworm and Forest.” Journal of
Animal Ecology 47 (1): 315–32. https://doi.org/10.2307/3939.
Lumsden, James. 1976. “Test Theory.” Annual Review of Psychology 27:
251–80. https://doi.org/10.1146/annurev.ps.27.020176.001343.
Lunansky, Gabriela, Jasper Naberman, Claudia D. van Borkulo, Chen Chen,
Li Wang, and Denny Borsboom. 2022. “Intervening on Psychopathology
Networks: Evaluating Intervention Targets Through Simulations.” Meth-
ods 204 (August): 29–37. https://doi.org/10.1016/j.ymeth.2021.11.006.
Lurie, Daniel J., Daniel Kessler, Danielle S. Bassett, Richard F. Bet-
zel, Michael Breakspear, Shella Kheilholz, Aaron Kucyi, et al. 2020.
“Questions and Controversies in the Study of Time-Varying Functional
Connectivity in Resting fMRI.” Network Neuroscience 4 (1): 30–69.
https://doi.org/10.1162/netn_a_00116.
Macy, M. W., A. Flache, and K. Takacs. 2006. “What Sustains Stable Cul-
tural Diversity and What Undermines It? Axelrod and Beyond. Peer
Review.” In Advancing Social Simulation: Proceedings of the First World
Congress on Social Simulation, edited by null S. Takahashi, 9–16. Kyoto,
Japan: Springer.
Mahmoud, Hosam. 2008. Polya Urn Models. CRC Press.
Maris, Gunter K. J., and Han L. J. van der Maas. 2012. “Speed-Accuracy
Response Models: Scoring Rules Based on Response Time and Accuracy.”
Psychometrika 77 (4): 615–33. https://doi.org/10.1007/s11336-012-9288-
y.
Marsman, Maarten, Denny Borsboom, Joost Kruis, Sacha Epskamp, Riet van
Bork, Lourens J. Waldorp, Han L. J. van der Maas, and Gunter K. J. Maris.
2018. “An Introduction to Network Psychometrics: Relating Ising Net-
work Models to Item Response Theory Models.” Multivariate Behavioral
Research 53 (1): 15–35. https://doi.org/10.1080/00273171.2017.1379379.

212
Marsman, Maarten, and Mijke Rhemtulla. 2022. “Guest Editors’ Introduction
to The Special Issue ‘Network Psychometrics in Action’: Methodological
Innovations Inspired by Empirical Problems.” Psychometrika 87 (1): 1–11.
https://doi.org/10.1007/s11336-022-09861-x.
Martins, André C. R. 2008. “Continuous Opinions and Discrete Actions in
Opinion Dynamics Problems.” International Journal of Modern Physics C
19 (04): 617–24. https://doi.org/10.1142/S0129183108012339.
Masuda, Naoki. 2014. “Voter Model on the Two-Clique Graph.” Physical
Review E 90 (1): 012802. https://doi.org/10.1103/PhysRevE.90.012802.
May, Edited by Professor Lord Robert, of Oxford, and Angela McLean, eds.
2007. Theoretical Ecology: Principles and Applications. Third Edition,
Third Edition. Oxford, New York: Oxford University Press.
Mazanov, Jason, and D. G. Byrne. 2006. “A Cusp Catastrophe Model Anal-
ysis of Changes in Adolescent Substance Use: Assessment of Behavioural
Intention as a Bifurcation Variable.” Nonlinear Dynamics, Psychology, and
Life Sciences 10: 445–70.
McGeer, T. 1990. “Passive Walking with Knees.” In, IEEE International
Conference on Robotics and Automation Proceedings, 1640–1645 vol.3. ht
tps://doi.org/10.1109/ROBOT.1990.126245.
McGrath, Thomas, Andrei Kapishnikov, Nenad Tomašev, Adam Pearce, Mar-
tin Wattenberg, Demis Hassabis, Been Kim, Ulrich Paquet, and Vladimir
Kramnik. 2022. “Acquisition of Chess Knowledge in AlphaZero.” Proceed-
ings of the National Academy of Sciences 119 (47): e2206625119. https:
//doi.org/10.1073/pnas.2206625119.
McLachlan, Geoffrey J., Sharon X. Lee, and Suren I. Rathnayake. 2019. “Fi-
nite Mixture Models.” Annual Review of Statistics and Its Application 6
(1): 355–78. https://doi.org/10.1146/annurev-statistics-031017-100325.
McRobie, Allan. 2017. The Seduction of Curves: The Lines of Beauty That
Connect Mathematics, Art, and the Nude. Princeton University Press.
Meadows, Donella H. 2008. Thinking in Systems: A Primer. Chelsea Green
Publishing.
Meena, Chandrakala, Chittaranjan Hens, Suman Acharyya, Simcha Haber,
Stefano Boccaletti, and Baruch Barzel. 2023. “Emergent Stability in Com-
plex Network Dynamics.” Nature Physics, April, 1–10. https://doi.org/10
.1038/s41567-023-02020-8.
Michell, Joel. 1997. “Quantitative Science and the Definition of Measurement
in Psychology.” British Journal of Psychology 88 (3): 355–83. https:
//doi.org/10.1111/j.2044-8295.1997.tb02641.x.
———. 1999. Measurement in Psychology: A Critical History of a Method-
ological Concept. Cambridge University Press.
———. 2008. “Is Psychometrics Pathological Science?” Measurement: Inter-
disciplinary Research and Perspectives 6 (January): 7–24. https://doi.or
g/10.1080/15366360802035489.
Miller, John H., and Scott E. Page. 2007. Complex Adaptive Systems: An
Introduction to Computational Models of Social Life. Complex Adaptive
Systems: An Introduction to Computational Models of Social Life. Prince-
ton, NJ, US: Princeton University Press.
Mitchell, Melanie. 1998. An Introduction to Genetic Algorithms. MIT Press.
———. 2009. Complexity: A Guided Tour. Oxford University Press.
Molenaar, Peter C. M. 2004. “A Manifesto on Psychology as Idiographic
Science: Bringing the Person Back Into Scientific Psychology, This Time

213
 Forever.” Measurement: Interdisciplinary Research and Perspectives 2 (4):
201–18. https://doi.org/10.1207/s15366359mea0204_1.
Molenaar, Peter C. M., Dorret I. Boomsma, and Conor V. Dolan. 1993. “A
Third Source of Developmental Differences.” Behavior Genetics 23 (6):
519–24. https://doi.org/10.1007/BF01068142.
Molenaar, Peter C. M., and Louis Oppenheimer. 1985. “Dynamic Models of
Development and the Mechanistic-Organismic Controversy.” New Ideas in
Psychology 3 (3): 233–42. https://doi.org/10.1016/0732-118X(85)90017-0.
Monroe, Brian M., and Stephen J. Read. 2008. “A General Connectionist
Model of Attitude Structure and Change: The ACS (Attitudes as Con-
straint Satisfaction) Model.” Psychological Review 115: 733–59. https:
//doi.org/10.1037/0033-295X.115.3.733.
Monsivais-Velazquez, Daniel, Kunal Bhattacharya, Rafael A. Barrio, Philip K.
Maini, and Kimmo K. Kaski. 2020. “Dynamics of Hierarchical Weighted
Networks of van Der Pol Oscillators.” Chaos (Woodbury, N.Y.) 30 (12):
123146. https://doi.org/10.1063/5.0010638.
Morel, Benoit, and Rangaraj Ramanujam. 1999. “Through the Looking Glass
of Complexity: The Dynamics of Organizations as Adaptive and Evolving
Systems.” Organization Science 10 (3): 278–93. https://doi.org/10.1287/
orsc.10.3.278.
Murray, James D. 1989. Mathematical Biology. Berlin, Heidelberg: Springer.
https://doi.org/10.1007/978-3-662-08539-4.
———. 2002. Mathematical Biology. 3rd ed. Interdisciplinary Applied Math-
ematics. New York: Springer.
Needham, Amy Work, and Eliza L. Nelson. 2023. “How Babies Use Their
Hands to Learn about Objects: Exploration, Reach-to-Grasp, Manipula-
tion, and Tool Use.” WIREs Cognitive Science n/a (n/a): e1661. https:
//doi.org/10.1002/wcs.1661.
Nenu, Theodor. 2022. “Douglas Hofstadter’s gödelian Philosophy of Mind.”
Journal of Artificial Intelligence and Consciousness 09 (02): 241–66. https:
//doi.org/10.1142/S2705078522500011.
Newman, M. E. J., and Juyong Park. 2003. “Why Social Networks Are Dif-
ferent from Other Types of Networks.” Physical Review E 68 (3): 036122.
https://doi.org/10.1103/PhysRevE.68.036122.
Nisbett, Richard E., Joshua Aronson, Clancy Blair, William Dickens, James
Flynn, Diane F. Halpern, and Eric Turkheimer. 2012. “Intelligence: New
Findings and Theoretical Developments.” American Psychologist 67 (2):
130–59. https://doi.org/10.1037/a0026699.
Noorazar, Hossein. 2020. “Recent Advances in Opinion Propagation Dynam-
ics: A 2020 Survey.” The European Physical Journal Plus 135 (6): 521.
https://doi.org/10.1140/epjp/s13360-020-00541-2.
Nosek, Brian A., Tom E. Hardwicke, Hannah Moshontz, Aurélien Allard,
Katherine S. Corker, Anna Dreber, Fiona Fidler, et al. 2022. “Replica-
bility, Robustness, and Reproducibility in Psychological Science.” Annual
Review of Psychology 73 (1): 719–48. https://doi.org/10.1146/annurev-
psych-020821-114157.
O’Byrne, Jordan, and Karim Jerbi. 2022. “How Critical Is Brain Criticality?”
Trends in Neurosciences 45 (11): 820–37. https://doi.org/10.1016/j.tins.2
022.08.007.
Oberauer, Klaus, and Stephan Lewandowsky. 2019. “Addressing the Theory
Crisis in Psychology.” Psychonomic Bulletin & Review 26 (5): 1596–1618.

214
 https://doi.org/10.3758/s13423-019-01645-2.
Oliva, Terence A., Wayne S. Desarbo, Diana L. Day, and Kamel Jedidi. 1987.
“Gemcat: A General Multivariate Methodology for Estimating Catastro-
phe Models.” Behavioral Science 32 (2): 121–37. https://doi.org/10.1002/
bs.3830320205.
Olthof, Merlijn, Fred Hasselman, Freek Oude Maatman, Anna M. T. Bosman,
and Anna Lichtwarck-Aschoff. 2023. “Complexity Theory of Psychopathol-
ogy.” Journal of Psychopathology and Clinical Science 132: 314–23. https:
//doi.org/10.1037/abn0000740.
Olthof, Merlijn, Fred Hasselman, Guido Strunk, Marieke van Rooij, Benjamin
Aas, Marieke A. Helmich, Günter Schiepek, and Anna Lichtwarck-Aschoff.
2020. “Critical Fluctuations as an Early-Warning Signal for Sudden Gains
and Losses in Patients Receiving Psychotherapy for Mood Disorders.” Clin-
ical Psychological Science 8 (1): 25–35. https://doi.org/10.1177/21677026
19865969.
Ooyen, Arjen van, and Markus Butz-Ostendorf. 2017. The Rewiring Brain: A
Computational Approach to Structural Plasticity in the Adult Brain. Aca-
demic Press.
Parisi, Giorgio. 2023. In a Flight of Starlings: The Wonders of Complex
Systems. Penguin Publishing Group.
Pashler, Harold, and Eric-Jan Wagenmakers. 2012. “Editors’ Introduction
to the Special Section on Replicability in Psychological Science: A Crisis
of Confidence?” Perspectives on Psychological Science 7: 528–30. https:
//doi.org/10.1177/1745691612465253.
Paulos, John Allen. 2008. Mathematics and Humor. University of Chicago
Press.
Pavlus, John. 2016. “The Clumsy Quest to Perfect the Walking Robot.” Sci-
entific American. https://www.scientificamerican.com/article/the-clumsy-
quest-to-perfect-the-walking-robot/. https://doi.org/10.1038/scientificam
erican0716-60.
Peralta, Antonio F., János Kertész, and Gerardo Iñiguez. 2022. “Opinion
Dynamics in Social Networks: From Models to Data.” arXiv. https://doi.
org/10.48550/arXiv.2201.01322.
Perc, Matjaž, Jillian J. Jordan, David G. Rand, Zhen Wang, Stefano Boc-
caletti, and Attila Szolnoki. 2017. “Statistical Physics of Human Coop-
eration.” Physics Reports, Statistical physics of human cooperation, 687
(May): 1–51. https://doi.org/10.1016/j.physrep.2017.05.004.
Peters, Gjalt Jorn Ygram, Szilvia Zörgő, and Han L. J. van der Maas. 2022.
“The Qualitative Network Approach (QNA.” https://doi.org/10.31234/osf
.io/cvf52.
Piaget, Jean. 1952. The Origins of Intelligence in Children. Edited by Mar-
garet Cook. The Origins of Intelligence in Children. New York, NY, US:
W W Norton & Co. https://doi.org/10.1037/11494-000.
Plenz, Dietmar, Tiago L. Ribeiro, Stephanie R. Miller, Patrick A. Kells, Ali
Vakili, and Elliott L. Capek. 2021. “Self-Organized Criticality in the
Brain.” Frontiers in Physics 9. https://doi.org/10.3389/fphy.2021.639389.
Ploeger, Annemie, Han L. J. van der Maas, and Pascal A. I. Hartelman. 2002.
“Stochastic Catastrophe Analysis of Switches in the Perception of Apparent
Motion.” Psychonomic Bulletin & Review 9 (1): 26–42. https://doi.org/
10.3758/BF03196255.
Pool, Robert. 1989. “Is It Healthy to Be Chaotic?” Science 243 (4891): 604–7.

215
 https://doi.org/10.1126/science.2916117.
Port, Robert F., and Timothy Van Gelder. 1995. Mind as Motion: Explo-
rations in the Dynamics of Cognition. MIT Press.
Posternak, Michael A., and Ivan Miller. 2001. “Untreated Short-Term Course
of Major Depression: A Meta-Analysis of Outcomes from Studies Using
Wait-List Control Groups.” Journal of Affective Disorders 66 (2): 139–46.
https://doi.org/10.1016/S0165-0327(00)00304-9.
Poston, Tim, and Ian Stewart. 2014. Catastrophe Theory and Its Applications.
Courier Corporation.
Pritchard, Walter s., and Dennis w. Duke. 1992. “Measuring Chaos in the
Brain: A Tutorial Review of Nonlinear Dynamical Eeg Analysis.” Interna-
tional Journal of Neuroscience 67 (1-4): 31–80. https://doi.org/10.3109/
00207459208994774.
Proskurnikov, Anton V., and Roberto Tempo. 2017. “A Tutorial on Modeling
and Analysis of Dynamic Social Networks. Part I.” Annual Reviews in
Control 43 (January): 65–79. https://doi.org/10.1016/j.arcontrol.2017.03.
002.
Rainer, Hegselmann, and Ulrich Krause. 2002. “Opinion Dynamics and
Bounded Confidence: Models, Analysis and Simulation.” Journal of Arti-
ficial Societies and Social Simulation 5 (3).
Rapoport, Anatol. 1960. Fights, Games, and Debates. University of Michigan
Press.
Ratcliff, Roger, Philip L. Smith, Scott D. Brown, and Gail McKoon. 2016.
“Diffusion Decision Model: Current Issues and History.” Trends in Cogni-
tive Sciences 20 (4): 260–81. https://doi.org/10.1016/j.tics.2016.01.007.
Read, John, and Joanna Moncrieff. 2022. “Depression: Why Drugs and
Electricity Are Not the Answer.” Psychological Medicine 52 (8): 1401–10.
https://doi.org/10.1017/S0033291721005031.
Redner, Sidney. 2019. “Reality-Inspired Voter Models: A Mini-Review.”
Comptes Rendus Physique 20 (4): 275–92. https://doi.org/10.1016/j.
crhy.2019.05.004.
Reher, Jenna, and Aaron D. Ames. 2021. “Dynamic Walking: Toward Agile
and Efficient Bipedal Robots.” Annual Review of Control, Robotics, and
Autonomous Systems 4 (1): 535–72. https://doi.org/10.1146/annurev-
control-071020-045021.
Remmerswaal, Danielle, Joran Jongerling, Pauline J. Jansen, Charly Eielts,
and Ingmar H. A. Franken. 2019. “Impaired Subjective Self-Control in
Alcohol Use: An Ecological Momentary Assessment Study.” Drug and
Alcohol Dependence 204 (November): 107479. https://doi.org/10.1016/j.
drugalcdep.2019.04.043.
Rendell, Paul. 2016. “Game of Life Universal Turing Machine.” In Turing
Machine Universality of the Game of Life, edited by Paul Rendell, 71–89.
Emergence, Complexity and Computation. Cham: Springer International
Publishing. https://doi.org/10.1007/978-3-319-19842-2_5.
Repp, Bruno H., and Yi-Huang Su. 2013. “Sensorimotor Synchronization: A
Review of Recent Research (2006–2012).” Psychonomic Bulletin & Review
20 (3): 403–52. https://doi.org/10.3758/s13423-012-0371-2.
Rinaldi, Sergio. 1998. “Love Dynamics: The Case of Linear Couples.” Applied
Mathematics and Computation 95 (2): 181–92. https://doi.org/10.1016/
S0096-3003(97)10081-9.
Roberts, James A., Leonardo L. Gollo, Romesh G. Abeysuriya, Gloria Roberts,

216
 Philip B. Mitchell, Mark W. Woolrich, and Michael Breakspear. 2019.
“Metastable Brain Waves.” Nature Communications 10 (1): 1056. https:
//doi.org/10.1038/s41467-019-08999-0.
Robertson, Robin, and Allan Combs, eds. 2014. Chaos Theory in Psychology
and the Life Sciences. New York: Psychology Press. https://doi.org/10.4
324/9781315806280.
Robinaugh, Donald J., Jonas Haslbeck, Lourens J. Waldorp, Jolanda Kos-
sakowski, Eiko I. Fried, Alex Millner, Richard J. McNally, et al. 2019. “Ad-
vancing the Network Theory of Mental Disorders: A Computational Model
of Panic Disorder.” PsyArXiv. https://doi.org/10.31234/osf.io/km37w.
Robinaugh, Donald J., Ria H. A. Hoekstra, Emma R. Toner, and Denny Bors-
boom. 2020. “The Network Approach to Psychopathology: A Review of
the Literature 2008–2018 and an Agenda for Future Research.” Psychologi-
cal Medicine 50 (3): 353–66. https://doi.org/10.1017/S0033291719003404.
Rooij, Iris van, and Mark Blokpoel. 2020. “Formalizing Verbal Theories.”
Social Psychology 51 (5): 285–98. https://doi.org/10.1027/1864-9335/a00
0428.
Rosencrans, Peter L., Lori A. Zoellner, and Norah C. Feeny. 2021. “A Network
Approach to Posttraumatic Stress Disorder: Comparing Interview and Self-
Report Networks.” Psychological Trauma: Theory, Research, Practice, and
Policy, No Pagination Specified–. https://doi.org/10.1037/tra0001151.
Rosser, J. Barkley. 2007. “The Rise and Fall of Catastrophe Theory Ap-
plications in Economics: Was the Baby Thrown Out with the Bathwa-
ter?” Journal of Economic Dynamics and Control 31 (10): 3255–80.
https://doi.org/10.1016/j.jedc.2006.09.013.
Rosso, O. A., H. A. Larrondo, M. T. Martin, A. Plastino, and M. A. Fuentes.
2007. “Distinguishing Noise from Chaos.” Physical Review Letters 99 (15):
154102. https://doi.org/10.1103/PhysRevLett.99.154102.
Roxin, Alex, and Anders Ledberg. 2008. “Neurobiological Models of Two-
Choice Decision Making Can Be Reduced to a One-Dimensional Nonlinear
Diffusion Equation.” PLOS Computational Biology 4 (3): e1000046. https:
//doi.org/10.1371/journal.pcbi.1000046.
Ryan, Oisı́n, Laura F. Bringmann, and Noémi K. Schuurman. 2022. “The
Challenge of Generating Causal Hypotheses Using Network Models.”
Structural Equation Modeling: A Multidisciplinary Journal 29 (6): 953–70.
https://doi.org/10.1080/10705511.2022.2056039.
Sachisthal, Maien S. M., Brenda R. J. Jansen, Thea T. D. Peetsma, Jonas
Dalege, Han L. J. van der Maas, and Maartje E. J. Raijmakers. 2019.
“Introducing a Science Interest Network Model to Reveal Country Differ-
ences.” Journal of Educational Psychology 111: 1063–80. https://doi.org/
10.1037/edu0000327.
Sala, Giovanni, N. Deniz Aksayli, K. Semir Tatlidil, Tomoko Tatsumi, Ya-
suyuki Gondo, and Fernand Gobet. 2019. “Near and Far Transfer in Cog-
nitive Training: A Second-Order Meta-Analysis.” Edited by Rolf Zwaan
and Peter Verkoeijen. Collabra: Psychology 5 (1): 18. https://doi.org/10
.1525/collabra.203.
Salanti, Georgia, Cinzia Del Giovane, Anna Chaimani, Deborah M. Caldwell,
and Julian P. T. Higgins. 2014. “Evaluating the Quality of Evidence from
a Network Meta-Analysis.” PLOS ONE 9 (7): e99682. https://doi.org/10
.1371/journal.pone.0099682.
Sandubete, Julio, E., and Lorenzo Escot. 2021. “DChaos: An R Package

217
 for Chaotic Time Series Analysis.” The R Journal 13 (1): 232. https:
//doi.org/10.32614/RJ-2021-036.
Saul, D. M., Michael Wortis, and D. Stauffer. 1974. “Tricritical Behavior
of the Blume-Capel Model.” Physical Review B 9 (11): 4964–80. https:
//doi.org/10.1103/PhysRevB.9.4964.
Savi, Alexander O., Maarten Marsman, Han L. J. van der Maas, and Gunter K.
J. Maris. 2019. “The Wiring of Intelligence.” Perspectives on Psychological
Science 14 (6): 1034–61. https://doi.org/10.1177/1745691619866447.
Sayama, Hiroki. 2015. “Introduction to the Modeling and Analysis of Complex
Systems.” Milne Open Textbooks, January. https://knightscholar.geneseo.
edu/oer-ost/14.
Scheffer, Marten. 2004. Ecology of Shallow Lakes. Dordrecht: Springer Nether-
lands. https://doi.org/10.1007/978-1-4020-3154-0.
Scheffer, Marten, Claudi L. Bockting, Denny Borsboom, Roshan Cools, Clara
Delecroix, Jessica A. Hartmann, Kenneth S. Kendler, et al. 2024. “A
Dynamical Systems View of Psychiatric Disorders—Practical Implications:
A Review.” JAMA Psychiatry, April. https://doi.org/10.1001/jamapsyc
hiatry.2024.0228.
Schelling, Thomas C. 1971. “Dynamic Models of Segregation.” The Journal
of Mathematical Sociology 1 (2): 143–86. https://doi.org/10.1080/002225
0X.1971.9989794.
Schiepek, Günter K., Kathrin Viol, Wolfgang Aichhorn, Marc-Thorsten Hütt,
Katharina Sungler, David Pincus, and Helmut J. Schöller. 2017. “Psy-
chotherapy Is Chaotic—(not Only) in a Computational World.” Frontiers
in Psychology 8. https://www.frontiersin.org/articles/10.3389/fpsyg.2017.
00379.
Schiepek, Günter, and Volker Perlitz. 2009. “Self-Organization in Clinical
Psychology.” In Synergetics, edited by Axel Hutt and Haken, 263–85. New
York, NY: Springer US. https://doi.org/10.1007/978-1-0716-0421-2_472.
Schmidhuber, Jürgen. 2015. “Deep Learning in Neural Networks:
An Overview.” Neural Networks 61 (January): 85–117. https:
//doi.org/10.1016/j.neunet.2014.09.003.
Schmidt, R. C., C. Carello, and M. T. Turvey. 1990. “Phase Transitions and
Critical Fluctuations in the Visual Coordination of Rhythmic Movements
Between People.” Journal of Experimental Psychology. Human Perception
and Performance 16 (2): 227–47. https://doi.org/10.1037//0096-1523.16.
2.227.
Schöner, Gregor. 2020. “The Dynamics of Neural Populations Capture the
Laws of the Mind.” Topics in Cognitive Science 12 (4): 1257–71. https:
//doi.org/10.1111/tops.12453.
Schöner, Gregor, and John P. Spencer. 2016. Dynamic Thinking: A Primer
on Dynamic Field Theory. Oxford University Press.
Schwartz, Seth J., Scott O. Lilienfeld, Alan Meca, and Katheryn C. Sauvi-
gné. 2016. “The Role of Neuroscience Within Psychology: A Call for
Inclusiveness over Exclusiveness.” American Psychologist 71 (1): 52–70.
https://doi.org/10.1037/a0039678.
Scott, John. 2011. “Social Network Analysis: Developments, Advances, and
Prospects.” Social Network Analysis and Mining 1 (1): 21–26. https:
//doi.org/10.1007/s13278-010-0012-6.
Serra, Roberto, and Gianni Zanarini. 1990. Complex Systems and Cognitive
Processes. Berlin, Heidelberg: Springer. https://doi.org/10.1007/978-3-

218
 642-46678-6.
Seth, Anil K., and Tim Bayne. 2022. “Theories of Consciousness.” Nature
Reviews Neuroscience 23 (7): 439–52. https://doi.org/10.1038/s41583-
022-00587-4.
Sherif, Muzafer, and Carl I. Hovland. 1961. Social Judgment: Assimilation
and Contrast Effects in Communication and Attitude Change. Social Judg-
ment: Assimilation and Contrast Effects in Communication and Attitude
Change. Oxford, England: Yale Univer. Press.
Silver, David, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou,
Matthew Lai, Arthur Guez, Marc Lanctot, et al. 2018. “A
General Reinforcement Learning Algorithm That Masters Chess,
Shogi, and Go Through Self-Play.” Science 362 (6419): 1140–44.
https://doi.org/10.1126/science.aar6404.
Simon, Herbert A. 1962. “The Architecture of Complexity.” Proceedings of
the American Philosophical Society 106 (6): 467–82. https://www.jstor.or
g/stable/985254.
Skarda, Christine A., and Walter J. Freeman. 1987. “How Brains Make Chaos
in Order to Make Sense of the World.” Behavioral and Brain Sciences 10
(2): 161–73. https://doi.org/10.1017/S0140525X00047336.
Smal, Iris, Jonas Dalege, and Han L. J. van der Maas. submitted. “The
Learning Ising Attitude Model (LIAM): Entropy Reduction by Hebbian
Learning,” submitted.
Smaldino, Paul. 2023. Modeling Social Behavior: Mathematical and Agent-
Based Models of Social Dynamics and Cultural Evolution. Princeton Uni-
versity Press.
Smolensky, Paul. 1986. “Information Processing in Dynamical Systems: Foun-
dations of Harmony Theory.”
Snijders, Tom A. B. 2001. “The Statistical Evaluation of Social Network
Dynamics.” Sociological Methodology 31 (1): 361–95. https://doi.org/10.1
111/0081-1750.00099.
Sobkowicz, Pawel. 2012. “Discrete Model of Opinion Changes Using Knowl-
edge and Emotions as Control Variables.” PLOS ONE 7 (9): e44489.
https://doi.org/10.1371/journal.pone.0044489.
———. 2020. “Whither Now, Opinion Modelers?” Frontiers in Physics 8.
https://doi.org/10.3389/fphy.2020.587009.
Soetaert, Karline, Thomas Petzoldt, and R. Woodrow Setzer. 2010. “Solving
Differential Equations in R : Package deSolve.” Journal of Statistical
Software 33 (9). https://doi.org/10.18637/jss.v033.i09.
Spearman, C. 1904. “’General Intelligence,’ Objectively Determined and Mea-
sured.” The American Journal of Psychology 15: 201–93. https://doi.org/
10.2307/1412107.
Spiller, Tobias R., Ofir Levi, Yuval Neria, Benjamin Suarez-Jimenez, Yair Bar-
Haim, and Amit Lazarov. 2020. “On the Validity of the Centrality Hypoth-
esis in Cross-Sectional Between-Subject Networks of Psychopathology.”
BMC Medicine 18 (1): 297. https://doi.org/10.1186/s12916-020-01740-5.
Sprott, J. C. 2004. “Dynamical Models of Love.” Nonlinear Dynamics, Psy-
chology, and Life Sciences 8 (3): 303–14.
Stam, C. J. 2005. “Nonlinear dynamical analysis of EEG and MEG: review
of an emerging field.” Clinical Neurophysiology: Official Journal of the
International Federation of Clinical Neurophysiology 116 (10): 2266–2301.
https://doi.org/10.1016/j.clinph.2005.06.011.

219
Steels, Luc. 1995. “A Self-Organizing Spatial Vocabulary.” Artificial Life 2
(3): 319–32. https://doi.org/10.1162/artl.1995.2.3.319.
Stefan, Martin. 2020. “RPubs - Fractals with R.” https://rpubs.com/mstefan-
rpubs/fractals.
Stengers, Isabelle, and Ilya Prigogine. 1978. Order Out of Chaos: Man’s New
Dialogue with Nature. London.
Stevens, S. S. 1957. “On the Psychophysical Law.” Psychological Review 64
(3): 153181. https://doi.org/10.1037/h0046162.
Stewart, Ian. 1982. “Catastrophe Theory in Physics.” Reports on Progress in
Physics 45 (2): 185–221. https://doi.org/10.1088/0034-4885/45/2/002.
Stewart, Ian, and P. L. Peregoy. 1983. “Catastrophe Theory Modeling in
Psychology.” Psychological Bulletin 94: 336–62. https://doi.org/10.1037/
0033-2909.94.2.336.
Stouffer, Samuel A., Edward A. Suchman, Leland C. Devinney, Shirley A.
Star, and Robin M. Williams Jr. 1949. The American Soldier: Adjustment
During Army Life. (Studies in Social Psychology in World War II), Vol. 1.
The American Soldier: Adjustment During Army Life. (Studies in Social
Psychology in World War II), Vol. 1. Oxford, England: Princeton Univ.
Press.
Strogatz, Steven H. 1988. “Love Affairs and Differential Equations.” Mathe-
matics Magazine 61 (1): 35–35. https://doi.org/10.1080/0025570X.1988.
11977342.
———. 2018. Nonlinear Dynamics and Chaos with Student Solutions Manual:
With Applications to Physics, Biology, Chemistry, and Engineering, Second
Edition. CRC Press.
Surowiecki, James. 2005. The Wisdom of Crowds. Knopf Doubleday Publish-
ing Group.
Sutton, Richard S., and Andrew G. Barto. 2018. Reinforcement Learning,
Second Edition: An Introduction. MIT Press.
Sznajd-Weron, Katarzyna, Józef Sznajd, and Tomasz Weron. 2021. “A Review
on the Sznajd Model — 20 Years After.” Physica A: Statistical Mechanics
and Its Applications 565 (March): 125537. https://doi.org/10.1016/j.phys
a.2020.125537.
Szostak, Natalia, Szymon Wasik, and Jacek Blazewicz. 2016. “Hypercy-
cle.” Edited by Shoshana Wodak. PLOS Computational Biology 12 (4):
e1004853. https://doi.org/10.1371/journal.pcbi.1004853.
Tesfatsion, Leigh. 2002. “Agent-Based Computational Economics: Growing
Economies From the Bottom Up.” Artificial Life 8 (1): 55–82. https:
//doi.org/10.1162/106454602753694765.
Tesser, Abraham. 1978. “Self-Generated Attitude.” In Advances in Experi-
mental Social Psychology, edited by Leonard Berkowitz, 11:289–338. Aca-
demic Press. https://doi.org/10.1016/S0065-2601(08)60010-6.
Thelen, Esther. 1995. “Motor Development: A New Synthesis.” American
Psychologist 50: 79–95. https://doi.org/10.1037/0003-066X.50.2.79.
Thelen, Esther, and Linda B. Smith. 1994. A Dynamic Systems Approach to
the Development of Cognition and Action. MIT Press.
Thom, René. 1977. “Structural Stability, Catastrophe Theory, and Applied
Mathematics.” SIAM Review 19 (2): 189–201. https://doi.org/10.1137/10
19036.
Thurner, Stefan. 2016. 43 Visions For Complexity. World Scientific.
Thurner, Stefan, Peter Klimek, and Rudolf Hanel. 2018. Introduction to the

220
 Theory of Complex Systems. Oxford University Press. https://doi.org/10
.1093/oso/9780198821939.001.0001.
Tilman, D, S S Kilham, and P Kilham. 1982. “Phytoplankton Community
Ecology: The Role of Limiting Nutrients.” Annual Review of Ecology and
Systematics 13 (1): 349–72. https://doi.org/10.1146/annurev.es.13.1101
82.002025.
Toker, Daniel, Friedrich T. Sommer, and Mark D’Esposito. 2020. “A Simple
Method for Detecting Chaos in Nature.” Communications Biology 3 (1):
1–13. https://doi.org/10.1038/s42003-019-0715-9.
Treiber, Martin, Ansgar Hennecke, and Dirk Helbing. 2000. “Congested Traf-
fic States in Empirical Observations and Microscopic Simulations.” Physi-
cal Review E 62 (2): 1805–24. https://doi.org/10.1103/PhysRevE.62.1805.
Treur, Jan. 2019. Network-Oriented Modeling for Adaptive Networks: Design-
ing Higher-Order Adaptive Biological, Mental and Social Network Models.
1st ed. 2020 edition. Springer.
Tschacher, Wolfgang, and Hermann Haken. 2019. The Process of Psychother-
apy: Causation and Chance. Springer.
———. 2023. “A Complexity Science Account of Humor.” Entropy 25 (2):
341. https://doi.org/10.3390/e25020341.
Tuerlinckx, Francis, Eric Maris, Roger Ratcliff, and Paul De Boeck. 2001.
“A Comparison of Four Methods for Simulating the Diffusion Process.”
Behavior Research Methods, Instruments, & Computers 33 (4): 443–56.
https://doi.org/10.3758/BF03195402.
Turner-Zwinkels, Felicity M., and Mark J. Brandt. 2022. “Belief System
Networks Can Be Used to Predict Where to Expect Dynamic Constraint.”
Journal of Experimental Social Psychology 100 (May): 104279. https:
//doi.org/10.1016/j.jesp.2021.104279.
Tyutyunov, Yu. V., and L. I. Titova. 2020. “From Lotka–Volterra to Arditi–
Ginzburg: 90 Years of Evolving Trophic Functions.” Biology Bulletin Re-
views 10 (3): 167–85. https://doi.org/10.1134/S207908642003007X.
Vallacher, Robin R., and Andrzej Nowak, eds. 1994. Dynamical Systems in
Social Psychology. Dynamical Systems in Social Psychology. San Diego,
CA, US: Academic Press.
van Bork, Riet, Sacha Epskamp, Mijke Rhemtulla, Denny Borsboom, and
Han L. J. van der Maas. 2017. “What Is the p-Factor of Psychopathology?
Some Risks of General Factor Modeling.” Theory & Psychology 27 (6):
759–73. https://doi.org/10.1177/0959354317737185.
van den Ende, Maarten W. J., Sacha Epskamp, Michael H. Lees, Han L.
J. van der Maas, Reinout W. Wiers, and Peter M. A. Sloot. 2022. “A
Review of Mathematical Modeling of Addiction Regarding Both (Neuro-)
Psychological Processes and the Social Contagion Perspectives.” Addictive
Behaviors 127 (April): 107201. https://doi.org/10.1016/j.addbeh.2021.10
7201.
van der Maas, Han L. J. 1995. “Beyond the Metaphor?” Cognitive Develop-
ment 10.
———. 2022. “Spiegeloog - A Dialogue on Consciousness.” Spiegeloog, Jan-
uary. https://www.spiegeloog.amsterdam/a-dialogue-on-consciousness/.
van der Maas, Han L. J., Jonas Dalege, and Lourens J. Waldorp. 2020.
“The Polarization Within and Across Individuals: The Hierarchical Ising
Opinion Model.” Journal of Complex Networks 8 (2): cnaa010. https:
//doi.org/10.1093/comnet/cnaa010.

221
van der Maas, Han L. J., Conor V. Dolan, Raoul P. P. P. Grasman, Jelte
M. Wicherts, Hilde M. Huizenga, and Maartje E. J. Raijmakers. 2006.
“A Dynamical Model of General Intelligence: The Positive Manifold of
Intelligence by Mutualism.” Psychological Review 113 (4): 842–61. https:
//doi.org/c3jm44.
van der Maas, Han L. J., Kees-Jan Kan, and Denny Borsboom. 2014. “In-
telligence Is What the Intelligence Test Measures. Seriously.” Journal of
Intelligence 2 (1): 12–15. https://doi.org/10.3390/jintelligence2010012.
van der Maas, Han L. J., Kees-Jan Kan, Maarten Marsman, and Claire E.
Stevenson. 2017. “Network Models for Cognitive Development and Intel-
ligence.” Journal of Intelligence 5 (2): 16. https://doi.org/10.3390/jintelli
gence5020016.
van der Maas, Han L. J., Rogier Kolstein, and Joop van der Pligt. 2003.
“Sudden Transitions in Attitudes.” Sociological Methods & Research 32 (2):
125–52. https://doi.org/10.1177/0049124103253773.
van der Maas, Han L. J., and Peter C. M. Molenaar. 1992. “Stagewise Cogni-
tive Development: An Application of Catastrophe Theory.” Psychological
Review 99 (3): 395–417. https://doi.org/10.1037/0033-295X.99.3.395.
van der Maas, Han L. J., Lukas Snoek, and Claire E. Stevenson. 2021. “How
Much Intelligence Is There in Artificial Intelligence? A 2020 Update.” In-
telligence 87 (July): 101548. https://doi.org/10.1016/j.intell.2021.101548.
van der Maas, Han L. J., Paul F. M. J. Verschure, and Peter C. M. Molenaar.
1990. “A Note on Chaotic Behavior in Simple Neural Networks.” Neural
Networks 3 (1): 119–22. https://doi.org/10.1016/0893-6080(90)90050-U.
van Geert, Paul. 1991. “A Dynamic Systems Model of Cognitive and Lan-
guage Growth.” Psychological Review 98: 3–53. https://doi.org/10.1037/
0033-295X.98.1.3.
———. 1998. “A Dynamic Systems Model of Basic Developmental Mecha-
nisms: Piaget, Vygotsky, and Beyond.” Psychological Review 105: 634–77.
https://doi.org/10.1037/0033-295X.105.4.634-677.
Van Overwalle, Frank, and Frank Siebler. 2005. “A Connectionist Model of At-
titude Formation and Change.” Personality and Social Psychology Review:
An Official Journal of the Society for Personality and Social Psychology,
Inc 9 (3): 231–74. https://doi.org/10.1207/s15327957pspr0903_3.
Vangeli, Eleni, John Stapleton, Eline Suzanne Smit, Ron Borland, and Robert
West. 2011. “Predictors of Attempts to Stop Smoking and Their Success
in Adult General Population Samples: A Systematic Review.” Addiction
(Abingdon, England) 106 (12): 2110–21. https://doi.org/10.1111/j.1360-
0443.2011.03565.x.
Verdonck, Stijn, and Francis Tuerlinckx. 2014. “The Ising Decision Maker:
A Binary Stochastic Network for Choice Response Time.” Psychological
Review 121: 422–62. https://doi.org/10.1037/a0037012.
Visser, Ingmar, and Maarten Speekenbrink. 2022. Mixture and Hidden
Markov Models with R. Springer International Publishing.
Voelkle, Manuel C., Johan H. L. Oud, Eldad Davidov, and Peter Schmidt.
2012. “An SEM Approach to Continuous Time Modeling of Panel Data:
Relating Authoritarianism and Anomia.” Psychological Methods 17: 176–
92. https://doi.org/10.1037/a0027543.
Volberda, Henk W., and Arie Y. Lewin. 2003. “Co-Evolutionary Dynamics
Within and Between Firms: From Evolution to Co-evolution.” Journal of
Management Studies 40 (8): 2111–36. https://doi.org/10.1046/j.1467-

222
 6486.2003.00414.x.
von der Heydt, Anna, Mark Dekker, and Henk Dijkstra. 2019. “Cascading
Transitions in the Climate System,” April, 10435.
Wagemans, Johan, James H. Elder, Michael Kubovy, Stephen E. Palmer, Mary
A. Peterson, Manish Singh, and Rüdiger von der Heydt. 2012. “A Century
of Gestalt Psychology in Visual Perception: I. Perceptual Grouping and
Figure–ground Organization.” Psychological Bulletin 138: 1172–1217. ht
tps://doi.org/10.1037/a0029333.
Wagenmakers, Eric-Jan, Peter C. M. Molenaar, Raoul P. P. P. Grasman, Pas-
cal A. I. Hartelman, and Han L. J. van der Maas. 2005. “Transformation
Invariant Stochastic Catastrophe Theory.” Physica D: Nonlinear Phenom-
ena 211 (3): 263–76. https://doi.org/10.1016/j.physd.2005.08.014.
Waldorp, Lourens J., and Maarten Marsman. 2022. “Relations Between Net-
works, Regression, Partial Correlation, and the Latent Variable Model.”
Multivariate Behavioral Research 57 (6): 994–1006. https://doi.org/10.1
080/00273171.2021.1938959.
Weaver, Warren. 1948. “Science and Complexity.” American Scientist 36 (4):
536–44. https://www.jstor.org/stable/27826254.
Weisbuch, Gérard, Guillaume Deffuant, Frédéric Amblard, and Jean-Pierre
Nadal. 2002. “Meet, Discuss, and Segregate!” Complexity 7 (3): 55–63.
https://doi.org/10.1002/cplx.10031.
Wheeler, John Archibald. 1994. At Home in the Universe. American Institute
of Physics.
Wichers, Marieke, Peter C. Groot, and ESM Group Psychosystems. 2016.
“Critical Slowing Down as a Personalized Early Warning Signal for De-
pression.” Psychotherapy and Psychosomatics 85 (2): 114–16. https:
//doi.org/10.1159/000441458.
Wiener, Norbert. 2019. Cybernetics or Control and Communication in the
Animal and the Machine, Reissue of the 1961 Second Edition. MIT Press.
Wilensky, Uri, and William Rand. 2015. An Introduction to Agent-Based
Modeling: Modeling Natural, Social, and Engineered Complex Systems with
NetLogo. Cambridge, Massachusetts: The MIT Press.
Wimsatt, William C. 1994. “The Ontology of Complex Systems: Levels of
Organization, Perspectives, and Causal Thickets.” Canadian Journal of
Philosophy Supplementary Volume 20 (January): 207–74. https://doi.org/
10.1080/00455091.1994.10717400.
Witkiewitz, Katie, Han L. J. van der Maas, Michael R. Hufford, and G. Alan
Marlatt. 2007. “Nonnormality and Divergence in Posttreatment Alcohol
Use: Reexamining the Project MATCH Data ”Another Way.”.” Journal
of Abnormal Psychology 116: 378–94. https://doi.org/10.1037/0021-
843X.116.2.378.
Wittenborn, A. K., H. Rahmandad, J. Rick, and N. Hosseinichimeh. 2016.
“Depression as a Systemic Syndrome: Mapping the Feedback Loops of
Major Depressive Disorder.” Psychological Medicine 46 (3): 551–62. https:
//doi.org/10.1017/S0033291715002044.
Xue, Jiankai, and Bo Shen. 2020. “A Novel Swarm Intelligence Optimization
Approach: Sparrow Search Algorithm.” Systems Science & Control Engi-
neering 8 (1): 22–34. https://doi.org/10.1080/21642583.2019.1708830.
Zahler, Raphael S., and Hector J. Sussmann. 1977. “Claims and Accomplish-
ments of Applied Catastrophe Theory.” Nature 269 (October): 759–63.
https://doi.org/10.1038/269759a0.

223
Zanin, Massimiliano. 2022. “Can Deep Learning Distinguish Chaos from
Noise? Numerical Experiments and General Considerations.” Communi-
cations in Nonlinear Science and Numerical Simulation 114 (November):
106708. https://doi.org/10.1016/j.cnsns.2022.106708.
Zeeman, E. C. 1976. “Catastrophe Theory.” Scientific American 234 (4):
65–83.
Zhang, Jiangbo, and Yiyi Zhao. 2018. “The Robust Consensus of a Noisy
Deffuant-Weisbuch Model.” Mathematical Problems in Engineering 2018
(December): e1065451. https://doi.org/10.1155/2018/1065451.
Zwicker, Maria V., Hannah U. Nohlen, Jonas Dalege, Gert-Jan M. Gruter, and
Frenk van Harreveld. 2020. “Applying an Attitude Network Approach to
Consumer Behaviour Towards Plastic.” Journal of Environmental Psychol-
ogy 69 (June): 101433. https://doi.org/10.1016/j.jenvp.2020.101433.

224