Dynamic Systems and Causal

Structures in Psychology
Connecting Data and Theory

Dynamische systemen en causale structuren in de psychologie:

De verbinding tussen data en theorie
(met een samenvatting in het Nederlands)

Proefschrift

ter verkrijging van de graad van doctor aan de

Universiteit Utrecht
op gezag van de
rector magnificus, prof.dr. H.R.B.M. Kummeling,
ingevolge het besluit van het college voor promoties
in het openbaar te verdedigen op

vrijdag 30 oktober 2020 des avond te 6.00 uur

door

Oisı́n Ryan

geboren op 23 november 1991

te Kilkenny, Ierland
 Promotor: Prof. dr. E.L. Hamaker
Copromotor: Dr. R.M. Kuiper

The studies in this thesis were funded by the Netherlands Organization for Sci-
entific Research (Onderzoekstalent Grant 406-15-128)
 Beoordelingscommissie:
Prof. dr. E. Ceulemans Katholieke Universiteit Leuven
Prof. dr. P. de Jonge Rijksuniversiteit Groningen
Prof. dr. I. Klugkist Universiteit Utrecht
Prof. dr. H. L. J. van der Maas Universiteit van Amsterdam
Prof. dr. M. C. Völkle Humboldt Universität zu Berlin

Dynamic Systems and Causal Structures in Psychology:

Connecting Data and Theory
Proefschrift Universiteit Utrecht, Utrecht.
- Met samenvatting in het Nederlands.

ISBN: 978-94-6416-145-8
Cover design: Luke Keeshan
Print: Ridderprint | www.ridderprint.nl
© Oisı́n Ryan 2020. All rights reserved.
 Contents

1 Introduction 1
1.1 Current Methodological Practice and Problems . . . . . . . . . . . 2
1.2 Alternative Methodological Frameworks . . . . . . . . . . . . . . . 3
1.3 Outline and Summary . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 The Challenge of Generating Causal Hypotheses Using Network

Models 7
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Using PMRFs to Generate Causal Hypotheses . . . . . . . . . . . . 15
2.4 Empirical Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

Appendices
2.A Moral-Equivalent DAGs: Violations of Sufficiency and Faithfulness 34
2.B The SE-set Algorithm: A Tool to Aid Causal Hypothesis Generation 35
2.C Empirical Illustration Details . . . . . . . . . . . . . . . . . . . . . 40

3 A Continuous-Time Approach to Intensive Longitudinal Data: What,

Why and How? 43
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 Two Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Why Researchers Should Adopt a CT Perspective . . . . . . . . . . 50
3.4 Making Sense of CT Models . . . . . . . . . . . . . . . . . . . . . . 51
3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Appendices
3.A Matrix Exponential . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.B Empirical Example Data Analysis . . . . . . . . . . . . . . . . . . . 67

4 Time to Intervene: A Continuous-Time Approach to Network Analysis

and Centrality 71
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2 Current Practice: DT-VAR Networks . . . . . . . . . . . . . . . . . 73
4.3 CT Network Analysis: Accounting for Continuity . . . . . . . . . . 81
4.4 Interventions and Centrality for CT Networks . . . . . . . . . . . . 88
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

i
CONTENTS

Appendices
4.A Centrality Measures as Summaries of Path-specific Effects . . . . . 101
4.B Centrality Values DT Stress-Discomfort System . . . . . . . . . . . 103
4.C The Matrix Exponential as Path-Tracing . . . . . . . . . . . . . . . 103
4.D Path-Tracing in CT models . . . . . . . . . . . . . . . . . . . . . . . 105
4.E Interventions and Path-Tracing in CT models . . . . . . . . . . . . 107

5 Recovering Bistable Systems from Psychological Time Series 113

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.2 Bistable Emotion System as Data-Generating Model . . . . . . . . 116
5.3 Recovering the Bistable System from Ideal Data . . . . . . . . . . . 124
5.4 Recovering the Bistable Systems from ESM Data . . . . . . . . . . 143
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

Appendices
5.A Determining Fixed Points . . . . . . . . . . . . . . . . . . . . . . . . 161
5.B Mean-Switching Hidden Markov Model . . . . . . . . . . . . . . . 163
5.C Data Generated from Estimated Models . . . . . . . . . . . . . . . 164
5.D Residual Partial Correlations TVAR(1) . . . . . . . . . . . . . . . . 166
5.E Differential Equation Model Building . . . . . . . . . . . . . . . . . 166
5.F Additional Results ESM Time Series . . . . . . . . . . . . . . . . . . 170

6 Modeling Psychopathology: From Data Models to Formal Theories 173

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
6.2 The Nature and Importance of Formal Theories . . . . . . . . . . . 174
6.3 Identifying Formal Theories from Data . . . . . . . . . . . . . . . . 180
6.4 An Abductive Approach to Formal Theory Construction . . . . . . 195
6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

Appendices
6.A Simulated Data from the Panic Model . . . . . . . . . . . . . . . . . 205
6.B Additional Details: The Panic Model and Statistical Dependencies 205
6.C Details Empirical vs Simulated Ising Model . . . . . . . . . . . . . 207

References 213

Nederlandse Samenvatting 239

About the Author 241

Publications & Working Papers 243

Acknowledgements 245

ii
 Chapter 1

Introduction
Psychological phenomena are best understood as dynamic processes: Political be-
liefs become more or less conservative as we age, the mathematics abilities of
children develop over the school year, and the duration and frequency of extreme
moods from day-to-day and week-to-week distinguish healthy from unhealthy
emotion regulation. As such, the key to understanding psychological phenomena
lies in understanding how behaviors, cognition, perceptions, emotions, disposi-
tions, abilities, and all other relevant facets of the mental world evolve, vary and
interact with each other over time, within an individual.
This process perspective has witnessed a tremendous growth in popularity,
bordering on consensus, in the psychological science literature in the past two
decades (Boker, 2002; Van Der Maas et al., 2006; Hamaker, 2012; Molenaar,
2004). In clinical psychology and psychiatry this idea has been particularly trans-
formative, with the traditionally static disease-based conceptualization of mental
disorder supplanted in recent years by the view that psychopathologies are in-
herently complex, multi-dimensional, and dynamic entities (Borsboom, Cramer,
Schmittmann, Epskamp, & Waldorp, 2011; Borsboom, 2017; Kendler, 2019; Nel-
son, McGorry, Wichers, Wigman, & Hartmann, 2017; Wichers, 2014). From this
perspective, to understand and eventually control and treat mental disorders we
must uncover the mechanistic relationships between psychological processes that
underlie psychopathology.
Researchers who subscribe to this theoretical perspective have collected a va-
riety of different types of empirical data with which they hope to study psycho-
logical processes. Due to the subject matter of clinical psychology, much of this
empirical data comes from non-experimental studies. Two popular categories of
non-experimental data can be distinguished. The first of these is cross-sectional
data, consisting of measurements of psychological processes across many indi-
viduals at a single point in time. The second is intensive longitudinal data, con-
sisting of repeated measurements of psychological processes over time, for one
or more individuals, typically in natural everyday settings. This latter type of
data has witnessed a surge in popularity in recent years, in part due to the ad-
vent of smartphone technology, and in part motivated by concerns over the diffi-
culty of making inferences from between-person data to within-person processes
(Molenaar, 2008; Hamaker & Wichers, 2017).
The core goal of this dissertation is to investigate how researchers can best
use non-experimental data on psychological processes to investigate the dynamic
mechanisms that give rise to psychopathology. Doing so will depend critically on
the methodology and more broadly the methodological framework that is used to
analyze these data.

1
1. Introduction

1.1 Current Methodological Practice and Problems

Researchers who wish to gain insights into psychological processes from non-
experimental data primarily do so through the estimation of relatively simple
statistical models from data. The current dominant methodological framework
for this type of modeling in psychology is the structural equation modeling (SEM)
framework. In the SEM framework, researchers begin by specify their theoretical
beliefs in the form of a graphical model describing which variables they believe
directly influence which others (Bollen, 1989). These graphical representations
imply a set of (usually linear) regression equations, the parameters of which can
be estimated directly from data. Two closely related statistical modeling tech-
niques are statistical network modeling (Borsboom & Cramer, 2013; Epskamp,
Rhemtulla, & Borsboom, 2017; Epskamp, Waldorp, Mõttus, & Borsboom, 2018;
Lauritzen, 1996), in which all variables are allowed relate to all others, and time-
series modeling (Hamaker, Dolan, & Molenaar, 2005; Hamilton, 1994; Molenaar &
Campbell, 2009), which is used to investigate relationships between time-lagged
variables from longitudinal data.
In each case, researchers working within these frameworks typically aim to
garner a greater understanding of the underlying mechanism through the in-
terpretation of estimated model parameters. Typically, dependencies between
variables are interpreted as indicative of direct causal effects between processes
(Granger, 1969; Bollen, 1987; Cole & Maxwell, 2003; Hamaker, Kuiper, & Gras-
man, 2015; Bulteel, Tuerlinckx, Brose, & Ceulemans, 2016), and the overall
mechanism is understood in terms of combinations or summaries of these param-
eters. However, two critiques of these approaches cast doubt on their suitability
for capturing the mechanisms underlying mental disorder.
First, these commonly used statistical models are limited in their ability to
capture or represent dynamic relationships, that is, the evolution of processes
over time. While we may expect this to be the case for models based on cross-
sectional data, even time-series models of longitudinal data typically treat the
passage of time only with respect to the ordering of measurements, ignoring
for instance the length of the time-interval that elapses between measurement
waves. Gollob and Reichardt (1987) describe a classic example of why this is
problematic. Imagine that we wish to study the effect of taking aspirin on re-
ducing headache pain. This effect may be zero two minutes after ingestion, sub-
stantial after thirty minutes, strong after two or three hours, reduced after five
hours and zero again twenty-four hours later. To truly understand the relation-
ship between aspirin and headache levels necessitates that we understand the
full dynamic picture of how that relationship evolves over time. Interpreting a
single parameter, exclusive to a particular interval, as the effect of one process on
another, as we would using common statistical modeling approaches, leaves us
with at best a critically incomplete understanding of the underlying mechanism.
This problem will necessarily be exacerbated when researchers wish to model
multidimensional systems of dynamic processes, and so the suitability of these
methods for that purpose is questionable.
The aspirin example is also helpful in highlighting the second problem with

2
 1.2. Alternative Methodological Frameworks

current practice, that is, the difficulty of using statistical models based on obser-
vational data to infer causal relationships. Ideally, we would study the aspirin-
headache relationship with an experiment. If we randomly assign headache-
suffering participants to take either an aspirin or a placebo, and we observe that
aspirin takers have lower headache levels two hours later, we can be reasonably
certain that this reduction was caused by the aspirin. However, if we are unable
to conduct such an experiment, we must instead rely on observations of aspirin
intake and headache levels in everyday life. Suppose that the analysis of this data
returns a negative statistical dependence between aspirin and headaches. When
(if ever) can we draw the same causal conclusion from this information as we
would have based on the experimental study? Unfortunately, the traditional sta-
tistical modeling literature provides few if any reasonable answers to this ques-
tion. Intuitive approaches which equate the fit of statistical models with their
causal veracity have variously been described as “radically suboptimal”, yield-
ing “garbage-can” models, and the very intuition behind this notion has been
labeled an urban-myth (Spirtes et al., 2000; Achen, 2005; Spector & Brannick,
2011; Rohrer, 2018). Without a principled way to link statistical with causal in-
formation, the utility of standard approaches for uncovering causal mechanisms
is left on somewhat uncertain footing.

1.2 Alternative Methodological Frameworks

In this dissertation I will explore how we can improve current statistical mod-
eling approaches in psychology by addressing the limitations set out above. To
achieve this, I will look to other disciplines that have grappled with similar is-
sues from different methodological perspectives. In particular the chapters of
this dissertation focus on exploring, borrowing and adapting ideas from two dis-
tinct methodological frameworks.
First, fields as diverse as physics, climatology, ecology, biology, chemistry
and engineering have used dynamical systems theory to understand and describe
phenomena that vary over time (Strogatz, 2015). In this framework, dynamic
mechanisms are described using the language of differential equations, breaking
down the evolution over time of processes into the fundamental building blocks
of moment-to-moment interactions. In so doing differential equations allow a
tremendous explanatory depth and breadth: Relatively simple differential equa-
tions can capture complicated patterns of dynamics, a modeling strategy that has
been successful in helping us to understand phenomena as diverse as the motion
of the earth around the sun (Newton, 1687), the relationship between predator
and prey populations in the wild (Volterra, 1931) and the reaction of the immune
system to the HIV virus (Ho et al., 1995). In comparison to the simple statistical
models described above, differential equations provide us with a rich and flexible
formalism for modeling dynamic processes, allowing us to describe and explore
how they evolve and vary continuously as a function of time.
Second, in fields such as epidemiology, econometrics and computer science,
researchers have long struggled with the issue of inferring causal relationships

3
1. Introduction

from observational data. In these fields, the different approaches which have
been developed to tackle this issue can broadly be described as the intervention-
ist causal inference framework, so called due to the definition of causal effects
in terms of hypothetical experimental interventions (Rubin, 1974; Greenland &
Robins, 1986; Angrist, Imbens, & Rubin, 1996; Pearl & Verma, 1991; Pearl, 2009).
This framework provides a mathematical language to describe causal structure,
formalizing the notion that statistical relationships (which describe patterns in
data) exist on a different level of explanation than causal relationships (which
describe the processes responsible for producing data). This in turn allows us
to answer the question of how and when statistical dependencies can be used to
infer causal structure, but moreover provides a new way to approach statistical
analyses when causal relationships are the target of inference.
Although these two approaches are distinct from one another, and indeed
tackle distinct shortcomings present in current practice, both the dynamical sys-
tems and interventionist causal inference frameworks offer promising avenues
by which we might improve our understanding of the mechanisms underlying
mental disorder. The chapters in my dissertation form an exploration of what we
can learn from these different approaches, and how we can use these lessons to
improve current practice.

1.3 Outline and Summary

The remainder of this dissertation consists of five chapters, each addressing some
problematic aspect of how current statistical modeling approaches are used to
investigate psychological processes, primarily based on non-experimental data,
and with a focus on the domains of clinical psychology and psychiatry. In each
chapter potential solutions to problems are offered inspired by either the method-
ological frameworks of dynamical systems theory, interventionist causal infer-
ence, or a mix of approaches borrowed from both traditions. The chapters them-
selves are organized with respect to the type of data and approach used as well
as the complexity with which the underlying mechanism is modeled.
Chapter 2 critically evaluates the use of statistical network models, fit on
cross-sectional data, in order to infer patterns of directed causal relationships
between variables. This practice is evaluated from the interventionist causal in-
ference perspective, revealing the inherent difficulties that are present in making
this type of inference even under ideal conditions. Moreover, this chapter in-
troduces a newly developed tool that aids researchers in exploring the different
graphical causal structures which may underlie a given statistical network model,
in principle aiding the generation of causal hypotheses. The discussion section of
this paper highlights the necessity of properly accounting for the time dimen-
sion if researchers wish to infer patterns of causal dependencies between dy-
namic processes. In the following chapters this is addressed by focusing largely
on methods of modeling dynamic relationships based on intensive longitudinal
data.
Chapter 3 and Chapter 4 examine how statistical models based on differen-

4
 1.3. Outline and Summary

tial equations, which are referred to throughout as continuous-time models, can

be used to improve the modeling of psychological processes in comparison to
standard time-series approaches. Chapter 3 focuses on a very simple continuous-
time model, providing a short treatment of the key practical and conceptual ad-
vantages of this model. Some core concepts from dynamical systems theory are
introduced, and the interpretation of the continuous-time model is illustrated
with a bivariate empirical example. Chapter 4 introduces a continuous-time ap-
proach to dynamical network analysis, and develops new centrality measures
specifically for these networks. Inspired by the interventionist causal inference
literature, these measures allow researchers to identify the optimal target for dif-
ferent types of interventions, either acute or continuous in nature.
The remaining chapters deal with the use of differential equations to model
more complex dynamics we may expect to operate between psychological pro-
cesses. Chapter 5 revisits the idea of using statistical models based on observa-
tional data to infer patterns of causal relationships between variables (as in Chap-
ter 2) when the structure of the underlying system is unknown. Here, however,
the data-generating mechanism takes the form of a bistable dynamical system, a
particular type of system which has received considerable theoretical attention
in clinical psychology, and which can be formalized by a set of non-linear differ-
ential equations. It is shown that, for highly idealized simulated data, some sta-
tistical models can be used to recover so-called global dynamics (i.e. more stable
characteristics of the system) but that it is difficult to correctly recover the micro-
dynamics (i.e. moment-to-moment relationships). Repeating the same analyses
for simulated data with a realistic sampling frequency showed that while global
dynamics may still be recovered, the recovery of micro-dynamics was unsuccess-
ful. These results highlight both the difficulty of making inferences from statisti-
cal models without a strong theory, as well as the fundamental role that sampling
frequency plays in statistical modeling of dynamic processes.
In Chapter 6, it is argued that formal theories are critically necessary to facil-
itate the study of mental disorders as complex dynamical systems. Formal the-
ories, expressed in the language of differential equations, are common in fields
that apply dynamical systems theory but almost absent from the clinical psychol-
ogy literature. The necessity of formal theories is supported by a short review of
the contemporary philosophy of science literature. Following this, three possi-
ble routes by which statistical models could be used to obtain formal theories
are investigated. The first route, interpreting statistical models directly as formal
theories (as in Chapters 3 and 4) and the second route, using common statistical
models to infer characteristics of the underlying mechanism (as in Chapters 2
and 5), are both shown to have fundamental shortcomings. The third route, us-
ing statistical models to help further develop an existing formal theory, is shown
to be most promising. The chapter concludes by proposing a framework for the
generation, development and testing of formal theories, detailing the role that
statistical models play at each step of this process.
This last chapter represents the culmination of developments in the disserta-
tion, synthesizing the different challenges and approaches described in the pre-
vious chapters, and utilizing these to inform a framework for formal theory con-

5
1. Introduction

struction. The dissertation ends with a clear message for the fields of clinical
psychology, psychiatry and the methodologists who work within these fields: If
we hope to understand the dynamic processes that give rise to psychopathology,
a radical reorientation of current research practice towards formal theory devel-
opment is urgently needed.

6
 Chapter 2

The Challenge of Generating

Causal Hypotheses Using
Network Models
Abstract

The network approach to psychopathology is a theoretical framework

in which mental disorders are viewed as arising from direct causal inter-
actions between symptoms. To investigate such networks, researchers typ-
ically estimate undirected network models from empirical data, called Pair-
wise Markov Random Fields (PMRFs), or for normally distributed variables,
Gaussian Graphical Models (GGMs). In this paper, we critically evaluate
the use of PMRF-based methods to generate causal hypotheses about an un-
derlying directed causal structure. We argue that hypothesis generation is
critically dependent on the specification of a target causal structure: This
is generally absent from applications of PMRFs, researchers instead tak-
ing a causally-agnostic approach. We show that the agnostic approach is
fundamentally problematic, since the heuristics typically used for hypoth-
esis generation do not hold for all types of causal structure. The spec-
ification of a target structure, however, allows a principled approach to
hypothesis generation and yields novel insights. We illustrate this using
the (weighted) Directed Acyclic Graph (DAG) as the target structure. We re-
view the relationship between PMRFs and DAGs, showing that hypothesis
generation using heuristics alone is non-trivial: Many different DAGs can
result in the same PMRF, and these differ in their substantive interpreta-
tions. With an empirical example, we illustrate how the GGM can be used
to generate more informed causal hypotheses, by exploring the equivalence
set of weighted DAGs. This is aided by a novel tool implemented in an
R package. Finally, we discuss additional barriers to discovering causal
relationships in practice, and possible alternative formalisms for causal
structure.

This chapter has been adapted from: Ryan, O., Bringmann, L. F. & Schuurman, N. K. (un-
der review). The Challenge of Generating Causal Hypotheses Using Network Models. Pre-print:
https://psyarxiv.com/ryg69/. Author contributions: OR conceptualized the initial project, wrote
the paper and R code and ran the analyses. LFB and NKS helped further develop the ideas in the
project, discussed progress and provided textual feedback.

7
2. The Challenge of Generating Causal Hypotheses

2.1 Introduction
The network approach to psychopathology is a theoretical framework in which
mental disorders are viewed as arising from direct causal interactions between
symptoms (Borsboom & Cramer, 2013; Borsboom, 2017). In practice, researchers
often aim to uncover aspects of the underlying network structure by estimating
network (i.e. graphical) models, generally from cross-sectional data (Van Borkulo
et al., 2014; Epskamp, Waldorp, et al., 2018; Epskamp, Borsboom, & Fried, 2018).
In these instances, researchers typically estimate a Pairwise Markov Random Field
(PMRF); for normally distributed variables this takes the form of a Gaussian
Graphical Model (GGM). These are network models with undirected connections
between variables, representing their conditional relationship (i.e., partial corre-
lation) controlling for all other variables in the network.
The PMRF is often promoted as an exploratory method of generating causal
hypotheses about the underlying network structure (Borsboom & Cramer, 2013;
Epskamp, van Borkulo, et al., 2018). In many instances, this data-generating
structure is conceptualized as consisting of directed causal relationships, and the
PMRF is taken to reflect the causal skeleton, identifying the presence, but not the
direction, of direct causal links (e.g., van Borkulo et al., 2015; Boschloo, Scho-
evers, van Borkulo, Borsboom, & Oldehinkel, 2016; Haslbeck & Waldorp, 2018).
Typically researchers are agnostic regarding what specific form the underlying
causal structure takes, for instance, whether it consists of uni-directional or bi-
directional relationships, and whether the structure is cyclic or acyclic (Cramer,
Waldorp, van der Maas, & Borsboom, 2010; Epskamp, Waldorp, et al., 2018; Isvo-
ranu et al., 2016; McNally et al., 2015; Robinaugh, Millner, & McNally, 2016;
Van Borkulo et al., 2014; Costantini et al., 2015).
However, this agnostic approach to causal structure means that the task of
causal hypothesis generation is fundamentally intractable: If the class of casual
structure we wish to make inferences about is not clearly defined, it is impossible
to know how to go about generating hypotheses about that structure. In other
words, the causal information conveyed by the absence or presence of connec-
tions in a PMRF depends entirely on the precise mapping from that network to
the underlying directed causal structure.
The uncertainty resulting from this causal-agnosticism is typified by the dis-
cussion surrounding Directed Acyclic Graphs (DAGs) in the network analysis lit-
erature. DAGs are a popular approach to conceptualizing directed causal struc-
tures in the causal inference literature, and one for which the relationship with
PMRFs is both relatively simple and well-known (Spirtes et al., 2000; Pearl, 2009;
Lauritzen, 1996). Typically, the hypothesis generation heuristics suggested by
users of PMRFs in psychology are consistent with, and even seem to be derived
from, the DAG as an underlying structure (see Epskamp, Waldorp, et al., 2018
p.457-458; Epskamp, van Borkulo, et al., 2018 p.420, and Borsboom & Cramer,
2013 p.105). Simultaneously, however, the DAG is rejected as a plausible target
structure, for example due to the absence of cyclic effects. This represents a fun-
damental problem for researchers wishing to generate causal hypotheses: It is
unclear whether these same heuristics still apply when mapping PMRFs to some

8
 2.2. Background

other undefined class of causal structure.

In this paper, we critically evaluate the use of PMRFs as a causal hypothesis
generating tool. To this end, we outline the importance of specifying a target
causal structure when generating such hypotheses, and describe how critically
hypothesis generation relies on the relationship between the target structure and
the PMRF. Following this, we evaluate how PMRFs can be used to generate causal
hypotheses once a target structure is defined, using the DAG as an example of
such a target. In taking this approach, we show that even in the context of the
simple mapping between DAGs and PMRFs, generating causal hypotheses from
a PMRF is not trivial, as a variety of distinct DAG structures can lead to the same
PMRF (Lauritzen, 1996; Andersson, Madigan, Perlman, et al., 1997; Raykov &
Marcoulides, 2001; MacCallum, Wegener, Uchino, & Fabrigar, 1993). This com-
plicates the generation of causal hypotheses for two reasons. First, variables may
be connected in the PMRF even when there is no directed causal relationship between
them; Second, different hypotheses researchers make based on connections in the
PMRF may not be compatible with any one underlying structure.
The remainder of the paper is structured as follows. First, we introduce the
necessary background on the two graphical models, PMRFs and DAGs, which
will be discussed in the remainder of the paper. Second, we detail the necessity
of specifying a target causal structure by a) arguing in principle against taking a
causally-agnostic approach, and b) evaluating the utility of PMRFs when a target
structure is specified, using the DAG as that target. Third, we illustrate how
PMRF-based methods can be used to generate causal hypotheses once a target
structure is chosen, using an example based on a published GGM analysis. This
empirical illustration is aided by a novel tool which allows us to generate the
set of structures which map to a given GGM under ideal conditions. Finally, we
discuss some routes forward in using network models to explore causal structure.

2.2 Background
In this section we will give an overview of network models, also known as graph-
ical models, two terms which we will use interchangeably in the current article.
Specifically we will review two instances of graphical models under considera-
tion in the current paper: The Pairwise Markov Random Field (PMRF), which
is used in the network analysis literature to generate causal hypotheses, and the
Directed Acyclic Graph (DAG), which we will use as an example target causal
structure. The remainder of this paper will make use of this background to allow
for an informed evaluation of PMRFs as hypothesis-generating tools, under two
scenarios: First, when the target casual structure is unspecified (i.e., the causally-
agnostic approach); Second, when the target causal structure is a DAG.
For both the PMRF and DAG we describe a special case which can be ob-
tained when the variables in question have a joint Gaussian (normal) distribu-
tion: the Gaussian Graphical Model (GGM) and a weighted DAG based on linear
regression. Finally, we describe the relationship between PMRFs and DAGs, as
described in the graphical modeling literature (Lauritzen, 1996).

9
2. The Challenge of Generating Causal Hypotheses

Unweighted Weighted

Support Support
Undirected

Stress Worry Stress Worry

Pressure Pressure

(a) Pairwise Markov Random Field (PMRF) (b) Gaussian Graphical Model (GGM)

Support Support
Directed

Stress Worry Stress Worry

Pressure Pressure

(c) Directed Acyclic Graph (DAG) (d) Regression path model (weighted DAG)

Figure 2.1: Four different types of network (graphical model), arranged by their edge-characteristics.

2.2.1 Networks and Graphs: A Primer

A network (or graph) is made up of a set of nodes (or vertices), and a set of edges
which encode the connections between pairs of nodes (Lauritzen, 1996; Newman,
2018). In the current paper, we will focus on applications of empirical network
models in psychology, in which the nodes represent a set of p variables, denoted
X, and the edges in the network represent statistical relationships between these
variables. Typically these relationships are not directly observed, but must be
estimated from data. In Figure 2.1, the nodes in each network are four variables
related to burnout: levels of social support (Support), work pressure (Pressure),
Stress, and Worry.
Graphs can be described as either directed or undirected, and weighted or
unweighted, depending on the characteristics of the edges in the graph. In un-
weighted graphs, edges are either present or absent (as in Figure 2.1(a) and (c))
and in weighted graphs edges have a particular value attached to them. In Fig-
ure 2.1(b) and (d) the thickness of each edge represents the absolute value of
the weight, and the color represents the sign (blue for positive, red for negative).
The edges in a given graph are collected in an adjacency matrix (for unweighted
graphs) or weights matrix (for weighted graphs).
In each type of graphical model considered here, the edges represent a par-
ticular type of conditional dependence relationship between pairs of variables in
the network, that is, the relationship between two variables conditional on (also
referred to as “controlling for” or “keeping constant”) a particular subset of other
variables in the graph. The exact nature of the relationship described by the edges

10
 2.2. Background

depends on the graphical model under consideration.

2.2.2 The Pairwise Markov Random Field and Gaussian

Graphical Model
A PMRF is an undirected, unweighted graph in which the absence of an edge
between a pair of variables Xi and Xj denotes that this pair is independent when
conditioning on the set of all other variables in the network X −(i,j)
⊥ Xj |X −(i,j)
Xi ⊥ (2.1)
where ⊥ ⊥ denotes independence between a pair of random variables (Dawid,
1979; Lauritzen, 1996).1 An example of a PMRF is shown in Figure 2.1(a). In
this graph, Worry is only connected to Stress. This denotes that Worry is inde-
pendent of Pressure when we condition on Stress and Support (Worry ⊥ ⊥ Pressure
| Stress, Support). Similarly, we can say Worry is dependent on Stress when we
condition on Pressure and Support (Worry 6⊥ ⊥ Stress | Pressure, Support).
The GGM is a particular type of weighted PMRF for variables following a
Gaussian distribution, where the edge weights indicate the strength of the linear
conditional relationship between a pair of variables, controlling for all others
(Dempster, 1972; Cox & Wermuth, 1996; Lauritzen, 1996). An example of a
GGM is shown in Figure 2.1(b). The positive edge connecting Stress and Worry
denotes that, keeping Support and Pressure constant, high scores for Stress tend
to co-occur with high scores for Worry. Typically the weights matrix of the GGM
is given by the matrix of partial correlations.
To obtain a GGM from a set of p variables with a joint Gaussian distribution
X ∼ N (µ, Σ), researchers use the precision matrix, the inverse of the variance co-
variance matrix
Ω = Σ −1 (2.2)
The precision matrix describes the conditional dependency relationships in
Equation (2.1) for a set of Gaussian variables in the following way: If an element
of this matrix is zero (ωij = 0) this means that the two variables pertaining to that
element are conditionally independent, given all other variables in the network
(Xi ⊥ ⊥ Xj |X −(i,j) ), while non-zero off-diagonal elements mean two variables are
not conditionally independent (ωij , 0 ⇒ Xi 6⊥ ⊥ Xj |X −(i,j) ). The matrix of partial
correlations is found by standardizing the precision matrix and multiplying the
off-diagonal elements by −1 (for further details see Dempster, 1972; Epskamp et
al., 2017).2
In psychological applications, the GGM is most often estimated from cross-
sectional data (although it is sometimes also fit on the residuals from a time se-
ries model, cf. Epskamp, Waldorp, et al., 2018). Typically the precision matrix
1 PMRFs are also sometimes referred to as Conditional Independence Graphs (e.g., M. Kalisch &
Bühlmann, 2007). As we deal with two different types of graphs which both describe particular
conditional (in)dependence statements, we use the terminology of PMRF throughout.
2 The relationship between a precision matrix and a partial correlation matrix is the similar to the
relationship between a covariance matrix and a correlation matrix. The partial correlation matrix is
easier to interpret directly, and so is more often used to visualize the results of a GGM analysis.

11
2. The Challenge of Generating Causal Hypotheses

is estimated directly from the data using regularization techniques such as the
graphical lasso (Friedman, Hastie, & Tibshirani, 2008; Epskamp & Fried, 2018).
These techniques introduce bias in the parameter estimates in order to avoid
over-fitting. The graphical lasso technique returns an estimate of the precision
matrix Ω̂ in which small values are set exactly to zero (Friedman et al., 2008). The
presence of exact zero values in this matrix means that the estimated precision
matrix, and thus the corresponding GGM, is often simpler (more sparse) than the
true population precision matrix.

2.2.3 The PMRF in Practice

Epskamp, Waldorp, et al. (2018) outline three potential uses for PMRF models
in psychology. The first is to use the PMRF to investigate purely predictive re-
lationships: For example, in Figure 2.1(a), Support is conditionally dependent
on both Stress and Pressure, but independent of Worry conditional on those two
variables. This tells us that, in order to optimally predict Support, we would need
information on Stress and Pressure, but that information on Worry is not needed.
Second, the PMRF can be interpreted directly as an undirected data-generating
mechanism. This approach has been used for the binary-variable Ising model
both in the statistical physics literature (Murphy, 2012) and in the context of
theoretical toy models in the psychology literature (Cramer et al., 2016; Dalege
et al., 2016; Borsboom, 2017). Third, Epskamp, Waldorp, et al. suggest that the
PMRF can be used as a causal hypothesis generating tool: That is, taking the undi-
rected edges in the PMRF as indicative of the presence of a directed causal effect
in the underlying causal structure (e.g., Borsboom & Cramer, 2013; Boschloo et
al., 2016; Deserno, Borsboom, Begeer, & Geurts, 2017; Knefel, Tran, & Lueger-
Schuster, 2016; Fried, Boschloo, et al., 2015). It is this use of the PMRF that we
will critically evaluate in the remainder of this paper.
It seems intuitive that the conditional dependencies estimated by the PMRF
are somehow informative about the underlying directed causal structure. How-
ever, to evaluate how well PMRFs perform in generating causal hypotheses, we
need to specify the form of that underlying causal structure, and understand how
the PMRF relates to it. Epskamp, Waldorp, et al. (2018) primarily discuss the re-
lationship between the PMRF and the Directed Acyclic Graph (DAG). DAGs (also
referred to as Bayesian networks) are directed graphical models consisting of uni-
directional edges Xi → Xj , and have proven widely popular as a way of conceptu-
alizing causal structures in formal approaches to causal inference (Spirtes et al.,
2000; Pearl, 2009; Dawid, 2002; Richardson & Robins, 2013; Galles & Pearl, 1995;
Dawid, 2010). However, it is evident from the literature that many researchers
use the PMRFs for hypothesis generation while taking what can be described
as a causally-agnostic approach: Rejecting the notion of a DAG as the underly-
ing structure, but without specifying an alternative formalism (e.g., Cramer et
al., 2010; Isvoranu et al., 2016; McNally et al., 2015; Robinaugh et al., 2016;
Van Borkulo et al., 2014; Costantini et al., 2015).
In the following sections we will evaluate the use of the PMRF to generate
causal hypotheses, both in the causally-agnostic setting, and when a DAG is taken

12
 2.2. Background

as the target causal structure. To facilitate this, in the following sub-section we

introduce the necessary background on DAGs, and the relationship between the
DAG and PMRF.

2.2.4 Directed Acyclic Graphs

The structure of a DAG describes which variables in X are conditionally depen-
dent and independent from one another. The DAG structure is often described
in “familial” terms: Directed edges Xi → Xj connect parent nodes or causes (Xi )
to children nodes or effects (Xj ). Nodes which share a common child, but are not
directly connected to one another, are termed unmarried parents. An example
of a DAG is shown in Figure 2.1(c), where Support and Pressure are unmarried
parents of Stress, and Stress is a parent of Worry. Both children and children of
children (and so forth) are called descendants, and parents and parents of parents
(and so forth) are called ancestors. In the graph shown in Figure 2.1(c) Worry is a
descendant of its ancestors Support, Pressure and Stress.
A DAG describes the conditional dependencies present in a set of random
variables X according to the causal Markov condition (Spirtes et al., 2000). This
condition states that each variable Xi is conditionally independent of its non-
descendants, given its parents

⊥ X −de(i) |X pa(i) ,
Xi ⊥ (2.3)

which can be used to read off conditional (in)dependencies between pairs of vari-
ables from the DAG. For example, in Figure 2.1(c), we can derive that Support
and Pressure are marginally independent (Support ⊥ ⊥ Pressure |∅), because Pres-
sure is a non-descendant of Support, and Support has no parents in this graph
(pa(Support) = ∅). Any two nodes connected by an edge are dependent condi-
tional on any subset of other variables.
Further conditional (in)dependency relationships between any pair of vari-
ables in the graph can be derived from the structure of a DAG using so called
d-separation rules (Pearl, 2009). These rules allow us to relate DAG structures to
other graphical models such as the PMRF. The most important of the d-separation
rules for the current paper relates to situations in which two variables share a
common child, also known as a common effect or collider structure Xi → Xk ← Xj .
According to d-separation rules, the parent variables Xi and Xj are dependent
conditional on the collider variable Xk . For example in Figure 2.1(c), although
Support and Pressure are marginally independent, they are dependent when
conditioning on Stress (Support 6⊥ ⊥ Pressure | Stress). For substantive exam-
ples applying d-separation rules in social science settings, readers are referred
to Glymour (2006).
A weighted DAG can be obtained from a set of Gaussian variables X, assum-
ing linear relationships between these variables, using a linear regression path
model in which child nodes are predicted by their parents

X = a + BX + e (2.4)

13
2. The Challenge of Generating Causal Hypotheses

where B is a p × p matrix of regression weights, a represents a p × 1 vector of in-

tercepts, and e represents a p × 1 vector of residuals, which are assumed to have
a Gaussian distribution with mean zero and diagonal variance-covariance Ψ (i.e.
the residual terms are uncorrelated). This type of weighted DAG is exactly equiv-
alent to a linear structural equation or path model. In a weighted DAG, as shown
in Figure 2.1(c), the matrix of regression coefficients B serves as the weights ma-
trix. Here we can see for instance that Support has a moderate negative effect on
Stress, and Pressure has a stronger positive effect on Stress. Crucially, while the
GGM can be estimated directly from data, a unique weighted DAG can typically
only be obtained from data if the structure of the unweighted DAG is known
(Levina, Rothman, Zhu, et al., 2008; Shojaie & Michailidis, 2010).

2.2.5 From DAG to PMRF: Moral Graphs and Skeletons

So far we have reviewed two types of graphical models, PMRFs and DAGs, and
the conditional dependency relationships described by each. From the graphical
modeling literature we know that there is a straightforward relationship between
the structure of a DAG and the structure of a PMRF (Lauritzen, 1996; Spirtes et
al., 2000). Specifically, when the underlying causal structure is a DAG, then the
corresponding PMRF is equivalent to the moral graph of that DAG. The moral
graph is an undirected graph obtained by first “marrying” (i.e. drawing an edge
between) all “unmarried parents” in the DAG, and then replacing all directed
edges with undirected edges. The moral graph therefore contains an undirected
edge if either a) these two nodes are connected by a directed edge in the DAG or
b) these two nodes share a collider (Wermuth & Lauritzen, 1983; Lauritzen, 1996;
Spirtes et al., 2000).3 The equivalence between PMRFs and moral graphs can be
seen when comparing the DAG and the PMRF in Figures 2.1(a) and (c). Here, the
PMRF contains an undirected version of all of the edges in the DAG, in addition
to an edge connecting the unmarried parents Support and Pressure, identical to
the moral graph of that DAG.
Crucially, the moral graph of a DAG should not be confused with the skeleton
of a DAG. The skeleton of a DAG is the undirected graph obtained by replacing the
directed edges in a DAG with undirected edges: The skeleton contains an edge
between two nodes if and only if the underlying DAG contains an edge between
those nodes (Spirtes et al., 2000). The skeleton describes exactly which variables
do and do not share a connection, but does not contain information on the direc-
tionality of that connection. In general the moral graph will contain more edges
than the skeleton, with additional edges induced in the moral graph whenever
there is a collider with unmarried parents.
The difference between the moral graph (PMRF) and the skeleton is shown
for three example DAGs in Figure 2.2. Each of the three DAGs shown in the first
column share the same skeleton, as shown in the second column. The third col-
3 A collider structure X → X ← X , which does not contain an edge directly connecting the
i k j
parents (Xi 9 Xj and Xi 8 Xj ) is called an open v-structure. If there are any open v-structures in the
DAG, then the moral graph must contain an undirected edge between the relevant parents Xi − Xj .
Thus, the moral graph “marries” unmarried parents.

14
 2.3. Using PMRFs to Generate Causal Hypotheses

Moral Graph
DAG Skeleton
(PMRF)

A B A B A B
C C C

D D D

A B A B A B
C C C

D D D

A B A B A B
C C C

D D D

Figure 2.2: Three examples of DAGs which all have the same skeleton, but result in different moral
graphs depending on the orientation of the edges in the DAG.

umn of Figure 2.2 shows the moral graph for each DAG. We see that although all
three DAGs share the same skeleton, each moral graph is distinct. Furthermore,
the moral graph contains additional edges connecting variables which are not con-
nected in the DAG, apart from the DAG in the second row of Figure 2.2, in which
there are no collider structures.

2.3 Using PMRFs to Generate Causal Hypotheses

Having reviewed the exact relationship between DAGs and PMRF models in the
previous section, we can now evaluate the practice of using PMRF-based models
to generate hypotheses about an underlying directed causal structure. First, we
outline the fundamental problem of generating causal hypotheses from PMRFs
when the form of the target causal structure is unspecified (i.e., the causally-
agnostic approach). This discussion motivates our focus on DAGs as candidate
causal structures. In the remainder, we make use of DAGs to outline how PMRFs

15
2. The Challenge of Generating Causal Hypotheses

can be used to generate causal hypotheses once the target structure is specified.
Specifically, we review what the mapping between DAGs and PMRFs described
above means for hypothesis generation, address the particular challenges which
arise when using the PMRF to infer multivariate patterns of relationships in an
underlying DAG, and describe some additional assumptions which are needed to
make causal hypothesis generation feasible in practice.

2.3.1 Causal Hypotheses and Unspecified Causal Structures

As outlined previously, in the psychology literature PMRF-based models are of-
ten interpreted as indicative of some true underlying causal structure, with the
presence of an edge taken as a necessary but not sufficient condition for a causal
relationship between two variables. The underlying causal structure itself is of-
ten described in terms of directed causal relationships between variables (e.g. in-
somnia → fatigue, support → stress) and the PMRF is promoted as a tool by
which to generate hypotheses about this directed structure (Borsboom & Cramer,
2013; Epskamp, Waldorp, et al., 2018; Epskamp, van Borkulo, et al., 2018). As-
suming that all variables involved in the causal system have been observed (i.e.
no unobserved common causes), the heuristics which are supplied for this hy-
pothesis generation task can be broadly summarized as follows:
Heuristic 1: An edge between two variables in the PMRF (Xi − Xj ) indicates
that two variables share either a direct causal link (Xi → Xj or Xi ← Xj ) or
a common effect (Xi → Xk ← Xj )
Heuristic 2: The absence of an edge between two variables (Xi X − Xj ) indi-
cates that these two variables do not share a direct causal link (Xi 9 Xj and
Xi 8 Xj )

Notably, these heuristics are consistent with treating the PMRF as the moral graph
of an underlying DAG, as described in the previous section. Furthermore, these
heuristics are often derived explicitly with reference to relationship between PM-
RFs and DAGs (Borsboom & Cramer, 2013; Epskamp, Waldorp, et al., 2018).
However, these same heuristics are typically described and applied by re-
searchers who simultaneously reject the possibility that the underlying structure
is a DAG, for instance due to the hypothetical presence of causal “loops” Xi  Xj
(e.g., Cramer et al., 2010; Isvoranu et al., 2016; McNally et al., 2015; Robinaugh
et al., 2016; Van Borkulo et al., 2014; Costantini et al., 2015). Moreover, beyond
the presence of such hypothetical loops, an agnostic approach to the underly-
ing causal structure is typically taken, in that the precise form of this alternative
structure is left unspecified. This represents a fundamental contradiction: With-
out specifying the form of the underlying causal structure, it is impossible to
verify whether these heuristics apply outside of the case used to derive them in
the first place.
In fact, we know for certain that for some types of causal structure, the heuris-
tics relating conditional dependencies (in the form of PMRF edges) to causal de-
pendencies described above do not hold. For instance, take it that the target

16
 2.3. Using PMRFs to Generate Causal Hypotheses

causal structure takes the form of directed relationships between dynamic, time-
varying processes, such as described by a Local Independence Graph (Schweder,
1970; Aalen, 1987; Didelez, 2000). This type of structure allows us to specify
causal loops in the form of time-forward relationships between processes linked
by a system of differential equations. For this type of causal structure, it is well
known that both heuristics described above can fail even when there are no un-
observed common causes: 1) observations of two causally independent processes
can be conditionally dependent (Aalen, Røysland, Gran, Kouyos, & Lange, 2016;
Maxwell & Cole, 2007) and 2) observations of two causally dependent processes
can be conditionally independent under certain conditions (Kuiper & Ryan, 2018).
This counter-example shows the critical necessity of specifying a target causal
structure. First, it shows that we cannot simply assume that the heuristic rules
described above hold for any and all types of causal structure. Second, it shows
that for an intuitive time-forward interpretation of causal loops, these heuris-
tics cannot be applied in an out-of-the-box fashion. This means that if a re-
searcher wishes to simultaneously apply these hypothesis generation heuristics,
while rejecting the notion of an underlying DAG, the burden of proof is on that
researcher to provide an alternative formalism for the underlying structure, and
to prove that the heuristics described above hold for that formalism. Finally, the
counter-example highlights the profound advantage of using the DAG as a target
causal structure, as we know for certain that these heuristics can be applied to
learn about an underlying DAG structure. Given that the DAG is, to our knowl-
edge, the only formalized causal structure for which we can currently be certain
these heuristics apply, the best case scenario for using these heuristics to generate
causal hypotheses is if the underlying causal structure is a DAG.4
In the following, we will examine the difficulties which remain in this hy-
pothesis generation task, even when the underlying structure is a (faithful) DAG
consisting only of observed variables (with these additional conditions for now
left implied, and discussed at the end of the section).

2.3.2 Causal Hypotheses and DAGs: The Moral Graph

Given that we choose an underlying DAG as our target causal structure, how can
the PMRF be used to help us generate hypotheses about the directed causal re-
lationships in that (unknown) DAG? As discussed in the previous section, the
equivalence of PMRFs and moral graphs means that the main information the
PMRF can give us about an underlying DAG structure is which edges are absent.
As stated in the second heuristic above, and illustrated in the examples in Fig-
ure 2.1 and Figure 2.2, the absence of an edge between two variables in the PMRF
implies the absence of a directed edge between those two variables in the under-
lying DAG. Critically, however, the presence of an edge between two variables in
4 It is possible, in theory, that these heuristics also apply equally well for some other type of causal
structure, possibly including causal loops, of which the current authors are not aware, and which is
not discussed by any of the researchers applying PMRF models in this way. However, we deem this
unlikely, at least without specification of a wider array of additional assumptions. We address the
issue of alternative causal models in the discussion.

17
2. The Challenge of Generating Causal Hypotheses

the PMRF cannot be taken to imply the presence of a directed edge between those
two variables in the underlying DAG: This latter statement is a property of the
DAG skeleton, and not the moral graph.
It is pertinent to note here an overlap in terminology used in the DAG and net-
work analysis literature. In the network analysis literature, multiple researchers
refer to PMRF models explicitly as representing the causal skeleton of some di-
rected causal structure, encoding the presence but not the direction of a causal
relationship (van Borkulo et al., 2015; Borsboom & Cramer, 2013; Isvoranu et al.,
2016; Haslbeck & Waldorp, 2018; Fried, Epskamp, Nesse, Tuerlinckx, & Bors-
boom, 2016; Armour, Fried, Deserno, Tsai, & Pietrzak, 2017). We reiterate that,
if the underlying causal structure is a DAG, then in general the PMRF will not be
equivalent to the skeleton of that DAG.
This point bears repeating because it lies at the heart of the difficulties which
arise when using PMRFs to hypothesize about DAGs. The value of PMRFs for
discovering causal structure is often justified by a comparison to networks where
edges represent marginal dependencies (for instance, full rather than partial cor-
relations). In contrast to these marginal dependency graphs, by conditioning
on all variables, the PMRF omits what Costantini et al. (2015) and Heeren and
McNally (2016) (amongst others) refer to as “spurious” edges: That is, connec-
tions between variables which do not reflect the presence of a direct connection
in the causal data-generating structure (but which reflect either a common cause
or indirect relationship). However, what is perhaps under-appreciated by users
less familiar with the graphical modeling literature is that, when the underlying
causal structure is a DAG, conditioning on all other variables in the network also
has the effect of inducing additional “spurious” connections between variables
which are unconnected in the DAG, resulting from conditioning on collider vari-
ables (as discussed by, amongst others, Epskamp, Waldorp, et al., 2018; Borsboom
& Cramer, 2013).

2.3.3 Challenges in Using PMRFs to Generate DAG Hypotheses

This uncertainty regarding the causal status of edges in the PMRF means that
there is, in general, a one-to-many mapping from a single PMRF to a set of DAGs:
While any given DAG only has one corresponding PMRF, multiple different DAGs
typically lead to the same PMRF.5 This means that, while patterns of connected
variables in a part of the PMRF (i.e. a local structure) may seem to imply a par-
ticular sequence of directed causal relationships, there are often many different
directed structures which could have led to any such pattern. This represents
a fundamental challenge for causal hypothesis generation: Although the heuris-
tics described above may hold for any one edge, it is typically less clear from the
PMRF alone whether two or more of these hypothetical directed relationships can
be co-present in any one DAG.
For example, in the psychological network literature chains of dependent vari-
ables in PMRFs, of the form Xi − Xj − Xk are often interpreted as indicative of a
5 Note that the PMRF can also represent conditional dependency structures which cannot be rep-
resented by a DAG, such as A − B − C − D − A

18
 2.3. Using PMRFs to Generate Causal Hypotheses

directed mediation structure, Xi → Xj → Xk . Such interpretations are present

in, for example, Deserno et al. (2017) (Xi = social contacts, Xj = social satisfac-
tion, Xk = feeling happy), Isvoranu et al. (2016) (Xi = sexual abuse, Xj = anxiety
and depression, Xk = psychosis) and Fried, Bockting, et al. (2015) (Xi = bereave-
ment, Xj = lonely, Xk = sad and happy). While these are potentially correct causal
hypotheses, there are typically many other valid possibilities that are consistent
with the PMRF and an underlying DAG. These possibilities may include DAGs
in which one or more of the connections of interest are absent, and are only in-
duced in the PMRF due to a collider structure. Researchers may be at risk of
drawing misleading conclusions about the underlying causal structure if these
possibilities are disregarded, or dismissed without a justification for why one
particular structure is more plausible than the next. Even more problematic is
that combinations of different causal hypotheses based on local structures may
be incompatible with one another, as they may imply for example, a new collider
structure, or imply some (in)dependence relationships which contradict those in
the PMRF. The plausibility of any given hypothesis regarding a part of the DAG,
must be assessed by taking the global structure of this system into account, that
is, considering the hypothetical DAG structure as a whole.
The challenging nature of using PMRFs can be illustrated with an example.
Figure 2.3 shows a four-variable PMRF, and in addition, all of the different DAG
structures which can generate that PMRF.6 In this case we can see that 13 distinct
DAG structures result in the given PMRF: We will call such a set of DAGs the
moral-equivalent set. It is immediately clear that the the one-to-many mapping
of a PMRF to different DAG structures precludes us from making any definitive
statement regarding the directionality of any particular relationships. However,
exploring the moral-equivalent set can still provide quite useful information. In-
specting the commonalities and differences in the moral-equivalent set gives us a
more complete idea of the causal structures that the PMRF may reflect, and aids
in generating and assessing causal hypotheses.
For example, numbers one through five of these DAGs contain one less edge
than is present in the PMRF, indicating that there is some uncertainty regarding
whether the variables A, B and C are all linked to each other by directed causal
relationships. Furthermore, we can see that the edge C − D is oriented in one
direction more frequently than another, as C → D in 9 out of 13 DAGs and C ← D
in the remaining 4. If we assume that all 13 DAG structures are equally plausible,
this gives some indication that C → D is more likely than C ← D. Furthermore,
in those DAGs in which either A and/or B are parents of C, D cannot be a parent
of C: This would result in a collider structure (A → C ← D and/or B → C ← D)
which is not allowed by the PMRF. This means we have some indication that C
is not a common child of A, B and D, but must be a parent of at least one other
variable.
By using Figure 2.3 we can re-evaluate the plausibility of different hypotheses
that researchers may be tempted to make from local structures in the PMRF. For
example, as A, B and C are connected to one another in the PMRF, one may be
6 Assuming sufficiency and faithfulness.

19
2. The Challenge of Generating Causal Hypotheses

PMRF DAGs
1) 2) 3) 4) 5)
A B A B A B A B A B
C C C C C

A B D D D D D

6) 7) 8) 9)

C
A B A B A B A B
C C C C

D D D D

D 10)
A B
11)
A B
12)
A B
13)
A B
C C C C

D D D D

Figure 2.3: A Pairwise Markov Random Field (PMRF) and each DAG structure which generates that
PMRF.

tempted to conclude that A has a direct effect on B (A → B), as well as an indirect

effect through C (A → C → B). Let us call this hypothesis I. To take another ex-
ample, a researcher may be tempted to hypothesize that, as there is no connection
between D and B, this is indicative that the effect of D on B is fully mediated by
the variable C, that is D → C → B. Let us call this hypothesis II.
Inspecting Figure 2.3 we can see that hypothesis I holds in only one of thirteen
DAG structures, DAG number 10. Furthermore, hypothesis II holds in exactly
three DAG structures, numbers 2, 12 and 13. Strikingly, hypothesis I and II are
totally incompatible with one another: although both hypotheses are reasonable
explanations of two different local structures, and in fact hypothesize the same
direction for one edge C → B, it is impossible for both to be true in any underlying
DAG structure.
This treatment highlights the difficulties which remain in generating causal
hypotheses from the PMRF alone, even in the best case scenario of taking the
DAG as the target causal structure. However, we have also seen that a lot of in-
formation which can be used in causal hypothesis generation is gained simply
by generating different directed graphs which are consistent with a given PMRF.
This entire procedure is made possible by the fact that we have specified the
form of our target causal structure, and know the mapping from the PMRF to
that causal structure. Only when this is done can we go about the task of causal
hypothesis generation in a principled way, as illustrated here. It is pertinent to
note here that in Figure 2.3 we depict a rather optimistic scenario for causal hy-
pothesis generation - as the number of nodes and/or the number of connections
in the graph grow, so too does the size of the moral-equivalent set. However,
more informed causal hypotheses still can be generated by taking into account
the size and sign of the relationships between variables, and by making some
distributional assumptions: For instance, by relating GGMs to weighted DAGs in

20
 2.3. Using PMRFs to Generate Causal Hypotheses

the form of linear SEM models. We illustrate the generation of causal hypotheses
from GGMs in the next section.

2.3.4 On Faithfulness, Sufficiency and Other Obstacles to

Causal Hypotheses
Once a target structure is identified, in order to use the PMRF to generate hy-
potheses about that structure, at least three simplifying assumptions must be
made. These additional assumptions are necessary to ensure that the data-
generating DAG is contained in the moral-equivalence set, given above. But
more generally, these assumptions are necessary for the two hypothesis gener-
ation heuristics outlined above to be valid, even in the best-case scenario that
the underlying structure is a DAG. These assumptions bear further consideration
here, as they pose additional barriers to the generation of causal hypotheses from
PMRF models in practice.

2.3.4.1 Sufficiency
The first of these assumptions is known as sufficiency, which essentially entails
that the underlying DAG consists of only the observed variables X (in the exam-
ples above, A, B, C and D) in the sense that there are no unobserved common
causes of two or more observed variables (Lauritzen, 1996; Pearl, 2009; Spirtes et
al., 2000).7 This means that the conditional dependencies captured by the PMRF
of X are assumed to come about due to directed causal relationships between the
observed variables X, and not due to relationships the observed variables share
with unobserved variables (such as an unobserved common cause, or an unob-
served collider variable on which we have unwittingly conditioned).
Relaxing the sufficiency assumption means that the task of generating causal
hypotheses becomes much more difficult, as, depending on the number and type
of missing variables we are willing to consider, there are many more DAGs which
may have generated the given PMRF. For example, if we consider there to be a
single unobserved variable, E, which acts as a common cause of A and B, then
we must consider the DAG in panel (a) of Figure 2.4 as a plausible underlying
causal structure: This DAG produces the same conditional dependencies present
between the observed variables A − D as the PMRF in Figure 2.3. If we consider
only this exact type of unobserved variable, we must now consider 9 additional
causal structures in which at least one of the connections in the PMRF is induced
by this common cause (see Appendix 2.A).
If we relax the sufficiency assumption even further, for example allowing E
to be a common cause of other variables, as in panel (b) of Figure 2.4 or allowing
more than one unobserved common cause as in panel (c) of Figure 2.4, deriving
a complete set of possible underlying DAGs quickly becomes infeasible. In gen-
eral, without the sufficiency assumption, the causal information conveyed by the
presence of an edge (as stated in Heuristic 1) becomes less certain, as there are
7 More precisely we could say that there are no unblocked back-door paths, based on d-separation
rules, passing through unobserved variables, that connect any pair of observed variables (Pearl, 2009).

21
2. The Challenge of Generating Causal Hypotheses

(a) (b) (c) (d)

E E E A B
A B A B A B
C
C C C
F
D D D D
Figure 2.4: Four DAGs which generate the same PMRF between the variables A − D (as shown in
Figure 2.3), but which violate some assumption which is made in deriving the moral-equivalence set.
The DAGs in panels (a), (b) and (c) represent sufficiency violations, with unobserved common cause(s)
E (and F). The weighted DAG in panel (d) violates faithfulness, as the directed positive relationship
between B and D is exactly canceled out by conditioning on C.

typically any number of possibilities involving unobserved variables which may

explain a particular conditional dependency. In practice then, for any statistical
analysis which hopes to capture causal relationships in some form, some assump-
tions regarding sufficiency and unobserved variables are necessary to make any
hypothetical statements at all.

2.3.4.2 Faithfulness
The second major assumption needed to generate causal hypotheses on the ba-
sis of the PMRF is called faithfulness (Spirtes et al., 2000; Pearl, 2009). A DAG
and associated probability distribution P meet the faithfulness condition if every
conditional (in)dependence relation in P is entailed by the causal Markov condi-
tion in Equation 2.3 (Spirtes et al., 2000). For example, if two variables Xi and Xj
are marginally independent, then by faithfulness the corresponding DAG should
have no directed paths which can be traced from Xi to Xj , e.g. Xi → Xk → Xj .
This means that we assume away the possibility that Xi and Xj are connected by
two different directed pathways, which when combined in the marginal relation-
ship between Xi and Xj , exactly cancel one another out. This would happen if,
for example, there was a negative direct pathway Xi → Xj as well as a positive
indirect pathway of equal size through Xk .
In panel (d) of Figure 2.4 we show a weighted DAG which would result in a
violation of faithfulness. In this DAG, both B and D have a positive direct effect
on C, making it a collider between them. Conditioning on this collider induces
a negative conditional relationship between B and D. However, simultaneously,
B has a positive direct effect on D: When combined, this negative and positive
relationship cancel one another out, and so B and D appear to be conditionally
independent given C - the partial correlation of B and D given C is zero, so they
are unconnected in the PMRF (see Appendix 2.A for details). In general, with-
out the faithfulness condition, the causal information conveyed by the absence
of an edge (as in Heuristic 2) becomes less certain. In theory, the assumption of
faithfulness is often justified in the context of relating the true probability dis-
tribution to a DAG; the probability of the true DAG having two pathways which
exactly cancel out is said to be negligible (Spirtes et al., 2000).

22
 2.4. Empirical Illustration

2.3.4.3 Population vs Estimated Conditional Dependence

However, the faithfulness assumption bears further consideration in combination
with the third simplifying assumption: Namely, that the given PMRF captures
exactly the true conditional dependencies in the population, and not some esti-
mate thereof. In other words, to enable a straightforward use of any estimated
PMRF, we must assume that any edge represents that two variables are condi-
tionally independent in the population. In practice of course, we typically don’t
have access to population statistics, and instead must try to account for uncer-
tainty about these estimates in some way. In the context of sampled data Uhler,
Raskutti, Bühlmann, and Yu (2013) have shown that violations of faithfulness
have a non-negligible probability, as conditional dependence relationships are
inferred on the basis of estimated parameters, rather than directly observed. In
the context of the mediation example given above, it may be the case that while in
the population the direct and indirect effects between Xi and Xk do not perfectly
cancel out, the estimated marginal dependency (i.e. total effect) from a given
sample may not meet some decision criteria to be considered non-zero (e.g., may
not be significantly different from zero based on a t-test). This is a well-known
phenomena in the SEM literature known as a suppression effect (Tzelgov & Henik,
1991).
It is pertinent to note again that, in practice, regularization techniques are
typically used to estimate weighted PMRFs such as the GGM in psychology
(Epskamp & Fried, 2018). These techniques introduce bias in parameter esti-
mates, setting parameters representing weak conditional dependencies exactly
to zero in the interest of parsimony. As such, our ability to make confident state-
ments about the absence of an edge in such a regularized PMRF implying the
absence of a direct causal relationship in the underlying DAG is on even less sure
footing.
We wish to note here that, although these assumptions are framed with re-
spect to the DAG, at least some similar assumptions are likely to hold no matter
what the target causal structure we wish to make some inference about. One ben-
efit of choosing the popular DAG as the target causal structure is that the types
of assumptions which need to be met to infer DAG structures from conditional
dependency information in different settings are well documented and well un-
derstood (Pearl, 2009; Spirtes et al., 2000; Dawid, 2010; Robins, 1999). By de-
lineating these assumptions, it becomes possible to study their validity, enabling
researchers to make better and more informed causal hypotheses, and facilitating
the accumulation of knowledge. In contrast, taking a causally-agnostic approach
precludes us in principle from understanding under what conditions and what
assumptions causal hypotheses can be made.

2.4 Empirical Illustration

To illustrate the value of specifying a target causal structure in practice, we
will make use of an empirical example taken from Hoorelbeke, Marchetti,
De Schryver, and Koster (2016). This example will allow us to examine in more

23
2. The Challenge of Generating Causal Hypotheses

detail the type of inferences which researchers typically aim to make based on
an estimated PMRF, and how specifying a target structure influences these infer-
ences.
In this illustration, we will make use of a novel tool which takes a precision
matrix, as estimated in the GGM method, and produces the set of statistically
equivalent weighted DAGs (hereby referred to as the SE-set) under ideal condi-
tions. This can be considered the weighted equivalent of the moral-equivalent
set, described in the previous section. In deriving the SE-set, we make use of the
sufficiency assumption, outlined above, but we do not need to invoke the faith-
fulness assumption. Since we are deriving a weighted DAG equivalent to a linear
path model, we also assume linear relationships between variables (as captured
by the GGM), and uncorrelated error terms. Full details on the SE-set algorithm
can be found in Appendix 2.B and it can be downloaded as an R-package from
the github page of the first author.8
Hoorelbeke et al. (2016) analyzed the network structure of cognitive risk
and resilience factors in a cross-sectional sample of 69 remitted depression pa-
tients. This network consisted of six variables: self-reported cognitive con-
trol (BRIEF WM); performance on a behavioral cognitive control measures
(PASAT ACC); adaptive emotion regulation strategies (Adapt ER), maladaptive
emotion regulation strategies (Maladapt ER); resilience levels (Resilience); and
residual depression symptoms (Resid Depres).
In the original article, the authors estimated a GGM, and additionally a di-
rected relative importance network. The relative importance network is another
way of encoding how well one variable is predicted by another, based on ex-
plained variance, and is often used in conjunction with an estimated PMRF
model in order to further inform the generation of causal hypotheses (Robinaugh,
LeBlanc, Vuletich, & McNally, 2014; McNally et al., 2015; Heeren & McNally,
2016). We begin by first reproducing their analysis and examining the conclu-
sions they make on the basis of this analysis. Full details of the re-analysis is
given in Appendix 2.C. Following this, we use the SE-set algorithm to re-examine
and expand upon the causal hypotheses generated from the original analysis.9

2.4.1 GGM Analysis

The precision matrix was estimated with the qgraph package (Epskamp, Cramer,
Waldorp, Schmittmann, & Borsboom, 2012), using the marginal correlations re-
ported by Hoorelbeke et al. (2016). The corresponding GGM is shown in Fig-
ure 2.5(a). From this GGM, we see that Resilience has strong negative connections
to both working memory and residual depressive symptoms; a weak negative
connection to maladaptive emotion regulation; and a weak positive connection
to adaptive emotion regulation. Furthermore, we see that the cognitive control
8 https://github.com/ryanoisin/SEset
9 The code to reproduce this analysis is available at https://github.com/ryanoisin/
CausalHypotheses

24
 2.4. Empirical Illustration

Maladapt ER Adapt ER

Adapt ER

Resilience BRIEF_WM
Maladapt ER Resilience BRIEF_WM

PASAT_ACC

Resid Depress Resid Depress

(a) GGM using EBICglasso (b) Relative Importance Network

Figure 2.5: Networks of cognitive risk and resilience factors based on re-analysis of Hoorelbeke et al.
(2016)

measure is unconnected to the rest of the variables.10

2.4.1.1 Relative Importance and Predictability

Relative importance is a measure of unique predictive strength, based on ex-
plained variance. The relative importance network is a directed, weighted net-
work, in which each edge Xi → Xj represents the unique variance explained in Xj
by Xi in a multiple regression where Xj is the outcome variable and X −j are the
predictors. For each node in the relative importance network, the sum of incom-
ing pathways (i.e., in-strength centrality) is related to the predictability of that
node. Predictability is calculated as the total variance explained (R2 ) in a given
outcome variable from the above regression. This is measure which, in network
analysis, is often interpreted with respect to an underlying directed causal struc-
ture: The predictability of a node Xi is typically interpreted as an upper-bound on
controllability, that is, the variance explained in Xi by its causes (parents) in the
true underlying causal structure (Haslbeck & Fried, 2017; Haslbeck & Waldorp,
2018; Fonseca-Pedrero et al., 2018). Predictability is typically depicted as a ring
around each node, with the proportion of the ring filled denoting the R2 .
Relative importance was calculated using the lmg metric in the R package
relaimpo (Grömping, 2006). As the PASAT ACC item is unrelated to the other
variables in the GGM, it is omitted from this analysis. The resulting relative im-
portance network is shown in Figure 2.5(b). Hoorelbeke et al. note that Resilience
10 Due to the different methods used to create a GGM, there are slight differences in the GGM net-
work depicted in Hoorelbeke et al. (2016) and the current paper. Apart from small numerical differ-
ences in parameters, the main difference is the presence of a weak positive partial correlation between
working memory and residual depressive symptoms. Further details can be found in Appendix 2.C

25
2. The Challenge of Generating Causal Hypotheses

shows the highest level of betweenness, closeness and strength centrality, and
is the only node with higher out-strength values (0.77) than in-strength values
(0.64). This means that Resilience appears to have a greater value in predicting
other variables than vice versa. However, the predictability of each node in Fig-
ure 2.5(b) shows that Resilience is the most predictable node in absolute terms,
followed by working memory and residual depressive symptoms. In compari-
son, the variables relating to emotion regulation strategies have relatively lower
predictability.

2.4.1.2 Interpretation of GGM Analysis

In the original paper, Hoorelbeke et al. (2016) interpret the GGM as showing that
Resilience is the “main hub” of the network, connecting all other variables. Based
on the relative importance network, it is tempting to draw further conclusions
about the underlying causal structure. For instance, it is tempting to conclude
that the undirected edges in the GGM reflect the presence of causal effects from
Resilience to other variables, or that a causal structure where Resilience is the
common cause of the other factors is more likely than a structure where Resilience
is a common effect (child) of all other variables.
However, it is well known that prediction is not necessarily indicative of cau-
sation; neither is it clear why an edge in the relative importance network is evi-
dence that a causal relationship is oriented in a particular direction. Although the
original authors do note early in the paper that the weights of the directed edges
here represent predictive relationships and do not imply causality, they later in-
terpret the greater out-strength of Resilience as implying that Resilience “exerted
a larger influence on the rest of the network than vice versa” (Hoorelbeke et al.,
2016, p. 100). Although these interpretations feel intuitive, we will show in the
following that inferring possible underlying structures in this way can be mis-
leading.

2.4.2 Analysis of Causal Hypotheses: The SE-set

As the researchers are explicitly interested in the directionality of causal effects,
this is a typical scenario in which the specification of some target causal structure
can aid greatly in the exploration and generation of causal hypotheses. Here, we
will take the target causal structure to be a weighted DAG. That is, we assume the
data-generating structure takes the form of a linear path model. This allows us
to explicitly account for the sign and strength of relationships estimated in the
GGM, rather than only the presence and absence of conditional dependencies, as
we did when mapping a PMRF to an unweighted DAG in the previous section.
In order to explore different possible underlying causal structures, we will
make use of the SE-set algorithm. This tool allows us to generate a set of weighted
DAGs which reproduce the given GGM. In doing so, we will assume sufficiency,
in that we only allow DAG structures between observed variables, but we will
not impose the assumption of faithfulness. The SE-set algorithm was run on
the estimated precision matrix consisting of 6 variables, which would produce a

26
 2.4. Empirical Illustration

Maladapt ER Maladapt ER

Adapt ER Adapt ER

Resilience BRIEF_WM Resilience BRIEF_WM

PASAT_ACC PASAT_ACC

Resid Depress Resid Depress

(a) DAG #12 (b) DAG #52

Maladapt ER Maladapt ER

Adapt ER Adapt ER

Resilience BRIEF_WM Resilience BRIEF_WM

PASAT_ACC PASAT_ACC

Resid Depress Resid Depress

(c) DAG #14 (d) DAG #20

Figure 2.6: Networks representing frequency of occurrence of edges in the DAG equivalence set.

maximum of 6! = 720 different weighted DAGs, one for every possible ordering of
the six variables from cause to effect. After rounding and the removal of duplicate
weights matrices, we are left with only nDAG = 58 unique weighted linear DAG
structures. We can explore the SE-set by inspecting some examples of DAGs in
the set and summarizing different characteristics of members of the set.

2.4.2.1 Examples of Equivalent DAGs

Four examples of weighted DAGs in the SE-set are shown in Figure 2.6. We can
see that all four DAGs represent quite distinct causal structures. In Figure 2.6(a)
residual depression has a negative causal effect on Resilience and a positive ef-
fect on working memory. Resilience also fully mediates the relationship between
these two variables and the two emotion regulation variables. In Figure 2.6(b)
the sign (positive/negative) of these relationships remains the same, but the di-
rection of each relationship is exactly the opposite. Furthermore the size of the
relationship between working memory and residual depression has decreased.
Figure 2.6(c) represents a DAG in which Resilience is the common cause of all

27
2. The Challenge of Generating Causal Hypotheses

Maladapt ER Maladapt ER

Adapt ER Adapt ER

Resilience BRIEF_WM Resilience BRIEF_WM

PASAT_ACC PASAT_ACC

Resid Depress Resid Depress

(a) Any edge relative frequency (b) Directed edge relative frequency

Figure 2.7: Networks representing frequency of occurrence of edges in the SE-set.

of the other variables in the network, as may have been hypothesized on the basis
of the relative importance network. Finally, Figure 2.6(d) represents an example
where Resilience is the common effect, i.e. caused by all other variables in the
network. Note that in this example, there are a number of edges present which
are not present in the estimated GGM, even though this weighted DAG does re-
produce the input precision matrix (for a discussion on how this may arise, see
the discussion of the faithfulness assumption in the previous section and Ap-
pendix 2.A).
By examining different elements of the SE-set we can see that there is a variety
of distinct directed structures that may have resulted in the estimated GGM. Re-
searchers can use substantive expertise to rule out some elements of the SE-set:
For instance, based on previous research we may know that depressive symp-
toms are a cause of working memory problems, and not vice versa, which would
rule out DAGs 52 and 20 (Figure 2.6(b) and 2.6(d)) respectively as plausible data-
generating structures. In the absence of such substantive knowledge, in the cur-
rent paper we treat each of the 58 DAGs in the SE-set as equally plausible data-
generating structures.

2.4.2.2 Relative Frequency of Edges

Rather than examining each weighted DAG individually, we can summarize the
information in the SE-set. One way to do this is by calculating in what pro-
portion of DAGs in the SE-set an edge of any direction (i.e. A → B or B → A)
connects two nodes. This information is shown as an undirected weighted net-
work in Figure 2.7(a). In this network, the presence of an edge indicates that in
at least one member of the equivalence set, there is a directed arrow connecting
the two nodes. The thickness of the arrow indicates how frequently such an edge

28
 2.4. Empirical Illustration

is present. In this particular instance, the thick arrows denote edges which are
always present in the SE-set, with the less thick edges indicating that these edges
were present in only 52 percent of the SE-set.
From Figure 2.7(a) we see that every edge which is present in the estimated
GGM is present in every member of the SE-set. We can further see that none
of the weighted DAGs contain edges between Adapt ER and Maladapt ER, or
between PASAT ACC and any other variable. From this we can say that, for this
particular example, an edge in the GGM appears to correspond to the presence of
an edge in an underlying weighted DAG; none of the edges in the GGM appear to
be induced due to conditioning on a collider in this case. Conversely, the absence
of an edge in the GGM does not always correspond to the absence of an edge
in the underlying weighted DAG. In half the members of the SE-set there is at
least one directed relationship present which is absent in the GGM: These are
weighted DAGs which are unfaithful, in the sense that two or more conditional
dependencies approximately cancel one another out (as discussed in the previous
section).
We can further explore the equivalence set by seeing in what proportion of
the SE-set edges of a particular direction occur. This is represented as a weighted,
directed network in Figure 2.7(b), where the edges A → B describe how often a
corresponding directed relationship A → B is present in the SE-set.11 We can
see from Figure 2.7b that for most pairs of variables, edges of opposite directions
occur with equal frequency. However, this is not the case for edges relating to
the Resilience variable, which was more often the child (or effect) than the par-
ent (cause). Resilience was the child of the working memory in 71 percent of the
equivalence-set DAGs, and a parent in 29 percent of the cases. A similar split
is present in the connections between Resilience and Resid Depress (71/29), Re-
silience and Adapt ER (69/31) and Resilience and Maladapt ER (69/31).
It is interesting to note the different ideas we get about a possible underlying
directed causal structure from Figure 2.7(b) than the relative importance network
in Figure 2.5(b). From the latter researchers may be tempted to conclude, that,
because Resilience has a greater predictive value for other variables than vice
versa, that Resilience is more likely to cause those variables than vice versa. From
the SE-set analysis we can say that, if we think the data-generating mechanism is
a linear weighted DAG, and each member of the SE-set is equally plausible, then
it is more than twice as likely that Resilience is caused by the other variables in
our graph than vice versa.

2.4.2.3 Controllability Distribution in the SE-set

In general, the SE-set algorithm can be used to examine how any particular char-
acteristic of a weighted DAG varies over different members of the SE-set. Node
predictability, described above, gives us an upper-bound on the variance explained
in a node Xi by its parents in the true underlying DAG, that is, the controllabil-
11 Pairs of directed edges in this network are mutually exclusive: if A → B is present then A ← B
cannot be present. The weights of mutually exclusive pairs add up to the weight of the corresponding
undirected edge in Figure 2.7(a)

29
2. The Challenge of Generating Causal Hypotheses

Resilience

Resid Depress
Variable

Maladapt ER

BRIEF_WM

Adapt ER

0.0 0.2 0.4 0.6 0.8

Controllability value R2

Figure 2.8: Distribution of controllability values for each variable, over DAGs in the SE-set. Control-
lability is here defined as the explained variance R2 of each variable when predicted by its parents in
a given DAG.

ity. However, the true controllability will differ depending on which variables
are causes and effects of others. For example, Resilience will have the highest
controllability in Figure 2.6(d), where it is a common effect of all other variables,
and a controllability of zero in Figure 2.6(c), where it is a common cause of all
others.
Using the SE-set algorithm, we can try to quantify our uncertainty about con-
trollability, by calculating a distribution of this value over the different possible
weighted DAG structures. This is shown in Figure 2.8. From this figure we can
see that, in most elements of the SE-set, the controllability of Resilience is high,
as we would expect given that Resilience is more often an effect than a cause of
other variables. We can also see that the controllability of both maladaptive and
adaptive emotion regulation strategies is quite low across different DAGs, with
parents of both strategies explaining a maximum of 20 percent of the variance
in each. However there is much more uncertainty about the controllability of
residual depressive symptoms and working memory. The distribution of con-
trollability for each is quite wide, ranging from zero to sixty percent, and with
peaks of similar height all across this range. This analysis shows us that, while
predictability in a standard GGM analysis captures the upper bound of these con-
trollability distributions, the SE-set analysis can supplement this information by
showing the variance and peaks of these distributions.

2.5 Discussion
In this paper we have critically evaluated a widespread practice in the current lit-
erature on psychological network analysis: Using estimates of PMRF-based mod-
els to generate hypotheses about an underlying directed causal structure. We

30
 2.5. Discussion

have shown that the utility of PMRF models for causal hypothesis generation is
critically dependent on what target causal structure is specified. When taking a
causally-agnostic approach, hypothesis generation is a fundamentally intractable
problem: The heuristics used by researchers to generate these hypotheses cannot
be assumed to apply to any and all causal structures. Defining a target structure,
however, allows us to generate causal hypotheses in a principled way. We illus-
trate this procedure using the DAG and a linear path model (weighted DAG) as
those target structures. In each case, we have shown that the specification of a
target structure allows for detailed and novel causal hypotheses to be generated,
based on the global structure of the estimated PMRF. In contrast, we have shown
how the isolated application of heuristics, even in combination with other tools
which encode predictive relationships, can lead to at best incomplete and at worst
incorrect inferences regarding the underlying pattern of causal relationships.
Throughout the paper we have made use of DAGs and weighted DAGs, in the
form of linear path models, as candidate target causal structures. D-separation
rules allowed us to derive the moral-equivalent set of DAGs from a PMRF, and the
SE-set algorithm introduced here allowed us to derive the statistically-equivalent
set of weighted DAGs from a GGM. It should be noted that the SE-set is best
considered an illustrative or exploratory tool. The benefit of using the SE-set al-
gorithm in the current paper is that it directly relates GGMs to potential under-
lying causal structures. As a stand-alone DAG estimation method, however, the
SE-set algorithm is limited for that same reason - inferences are entirely reliant
on the GGM estimate, and uncertainty in those estimates is not accounted for. Re-
searchers who wish primarily to estimate a (weighted) DAG structure from data
would be better served in using algorithms specifically designed for that purpose,
such as the PC, LiNGAM and FCI algorithms, amongst others (Spirtes et al., 2000;
Shimizu, Hoyer, Hyvärinen, & Kerminen, 2006; Spirtes, Meek, & Richardson,
1995), many of which are implemented in R packages such as pcalg (M. Kalisch,
Mächler, Colombo, Maathuis, & Bühlmann, 2012) and bnlearn (Scutari, 2010).
Note that typically these algorithms are unable to uniquely identify a DAG from
data, instead estimating a Markov-equivalent class, that is, a set of DAGs with fully
equivalent conditional dependency relationships (cf. Andersson et al., 1997).12
Numerous tutorials on the use and advantages of different DAG estimation algo-
rithms, and the assumptions necessary for each, are available for the interested
reader (e.g., Malinsky & Danks, 2018; Spirtes & Zhang, 2016).
However, the possibility of directly estimating DAG structures from data
should not be taken as a panacea to the problem of inferring multivariate pat-
terns of causal relationships from data. The focus on DAGs in the current paper
is motivated by our focus on the heuristic mapping rules which are applied by
researchers to generate causal hypotheses: If the underlying structure is a DAG,
this represents the best case scenario for this practice, in the sense that we are
guaranteed those heuristics are correct. A focus on DAGs as target causal struc-
tures also confers some additional benefits, as they have been extensively studied
12 Furthermore, since these algorithms take disparate approaches to estimating DAG structures, we
are not guaranteed that the moral graph implied by the estimated Markov-equivalent DAGs will be
equal to the PMRF estimate, even when the underlying structure is a DAG.

31
2. The Challenge of Generating Causal Hypotheses

in the causal inference literature, and can be easily used in conjunction with, for
example, interventionist theories of causality (Spirtes et al., 2000; Pearl, 2009;
Dawid, 2002; Richardson & Robins, 2013). That being said, it is possible that
the DAG is not the optimal formalism for causal structure in psychological set-
tings, and that some alternative causal structure may be more appropriate (for
various discussions on this point, see McNally, 2016; Dawid, 2010; Cartwright,
1999, 2007). The problem however, is that it appears as though no such alter-
native formalism is readily available, or at least well-known in the psychology
literature.
Cyclic causal models (i.e., graphs which allow causal “loops” Xi  Xj ) are
in general underdeveloped, and the interpretation of cyclic causal effects with-
out invoking a notion of time-forward dependency is notoriously difficult (e.g.,
Spirtes, 1995; Hayduk, 2009). As noted by, amongst many others, Borsboom
et al. (2012) and Epskamp, Waldorp, et al. (2018), cyclic effects between dy-
namic processes can easily be represented as acyclic time-forward relationships
Xi,t → Xj,t+1 → Xi,t+1 , either in a DAG or other related structure, as in time
series analysis (e.g., Bringmann et al., 2013; Hamaker & Dolan, 2009). For-
malized approaches to defining these time-forward causal relationships exist
both in discrete-time (Eichler & Didelez, 2010; Dahlhaus & Eichler, 2003) and
continuous-time (Didelez, 2000; Aalen, 1987). Alternatively, undirected graphs
such as the Ising model can potentially be given a dynamic interpretation, reflect-
ing symmetric time-forward directed relationships, by invoking Glauber dynam-
ics (Glauber, 1963; Haslbeck, Epskamp, Marsman, & Waldorp, 2018; Marsman
et al., 2018), as applied in recent theoretical models in psychology (Cramer et
al., 2016; Dalege et al., 2016; Borsboom, 2017). However, if we take these types
of dynamic systems to be the target structure(s), many additional assumptions
must be invoked in order to ensure that PMRF edges estimated from a given data
source reflect causal relationships therein. For example, assumptions about the
frequency of the underlying process are needed to recover the structure of a dy-
namic system from time-series data (Papoulis & Pillai, 2002; Marks, 2012) and
the oft-criticized ergodicity assumption is needed to ensure recovery of dynamic
relationships from cross-sectional data (Molenaar, 2004; Hamaker, 2012; Mole-
naar, 2008). There may be yet other approaches to cyclic causal effects which
are more promising for psychological applications: For example, cyclic effects
which represent dynamic systems in an equilibrium state (Mooij, Janzing, Hes-
kes, & Schölkopf, 2011; Forré & Mooij, 2018, 2019). However, more work is
needed to assess their suitability and applicability in practice, and to establish
any potential links between those types of causal structures and the conditional
dependency patterns estimated by methods such as the PMRF.
Outside of their use for causal hypothesis generation, there are many attrac-
tive reasons to use PMRF-based models in practice. Amongst other things, they
allow for the identification of predictive relationships, sparse descriptions of sta-
tistical dependency relationships in a multivariate density, and may be used as
a variable clustering or latent variable identification method (e.g., Golino & Ep-
skamp, 2017). However, we have shown that, when the aim is to uncover some
aspects of the underlying causal structure, it is generally unclear what conclu-

32
 2.5. Discussion

sions we can draw directly from an estimated network structure alone (that is,
without further specification and exploration of a target structure). Ours is not
the first paper to come to such a conclusion: Dablander and Hinne (2019) showed
that node centrality in a GGM is a poor indicator of causal influence in an un-
derlying DAG; Bos et al. (2017) showed that conclusions about causal structure
based on a cross-sectional GGM do not generalize to those made if a time-forward
causal structure is assumed; Haslbeck and Ryan (2019) showed that it is typically
unclear how to use various statistical methods common in the network approach
literature, including the GGM, to draw conclusions about an underlying dynamic
system from time series data. This pattern of results is concerning, as our read-
ing of the literature suggests that discovering patterns of causal relationships
is the primary motivation behind most of the applications of these methods in
empirical research: These statistical methods appear to hold the promise of di-
rectly uncovering the causal interactions which are a cornerstone of the theoret-
ical framework which motivated their development (Borsboom & Cramer, 2013;
Borsboom, 2017).
In this paper we suggest one route by which the search for causal relationships
in psychological networks can move forward: The investigation and specification
of target causal structures. Specifying a target causal structure establishes the
rules and boundaries through which we can make some inductive or abductive
inference from an estimated network model to the underlying system of causal
relationships. The feasibility of this approach relies on first the availability of
suitable formalized causal structures and second the establishment of how infor-
mation captured by different statistical methods relates to those structures. An
alternative to this route would be to use methods such as the PMRF as purely
deductive or confirmatory tools, to test the implications and predictions of pre-
specified causal theories: This in turn relies on our ability to specify these the-
ories a-priori. In either case, the critical observation is that to use PMRFs, or in
essence, any statistical method, to discover potential causal relationships, it is
necessary to move beyond the causally-agnostic approach. Only when we make
clear what it is we wish to make some inference about, can we hope to use any
tool to make those inferences in a principled way. Only then can the power of this
and other statistical approaches be fully realized: By understanding what types
of inferences we can make, when we can make them, and how to go about doing
this in practice.

33
2. The Challenge of Generating Causal Hypotheses

Appendix 2.A Moral-Equivalent DAGs: Violations

of Sufficiency and Faithfulness
In this appendix we provide additional insight into moral-equivalent DAGs
which violate the assumptions of either sufficiency or faithfulness.

2.A.1 Sufficiency-Violating DAGs

In the main text, Figure 2.4(a), we show a single example of a DAG which gen-
erates the PMRF depicted in the left-hand column of Figure 2.3, in the presence
of an unobserved variable E which acts as a common cause of both A and B. If
this is the data-generating DAG, we can say that variables A − D do not meet the
assumption of sufficiency - to derive the true DAG, the variable E is needed.

(a) (b) (c) (d)

E E E E
A B A B A B A B
C C C C

D D D D
(e) (f) (g) (h)
E E E E
A B A B A B A B
C C C C

D D D D

Figure 2.9: Eight additional DAGs which generate the same PMRF between the variables A − D (as
shown in Figure 2.3). For each DAG, the set of variables A − D violates the sufficiency assumption,
with respect to the variable E, a common cause of both A and B.

If we assume that sufficiency is violated only in that an unobserved variable E is

a common cause of A and B, then there are eight more DAGs which generate the
same PMRF for the variables A − D, and contain one less directed edge between
the variables A−D than any of the DAGs in the moral-equivalent set shown in the
right-hand column of Figure 2.3. These are depicted in Figure 2.9. Many more
moral-equivalent DAGs are possible if the assumption of sufficiency is relaxed
further, that is, if different and/or additional connections between E and other
variables or allowed, or if more than one unobserved common cause is allowed.

2.A.2 Faithfulness-Violating Weighted DAGs

In the main text, Figure 2.4(d), we depict a weighted DAG which generates the
PMRF depicted in the left-hand column of Figure 2.3, but which violates the

34
 2.B. The SE-set Algorithm: A Tool to Aid Causal Hypothesis Generation

faithfulness assumption. The weights matrix of this weighted DAG is given as

0.000 0.200 −0.400 0.000

 
0.000 0.000 0.000 0.000
B =  (2.5)
 
0.000 0.485 0.000 0.152

0.000 0.100 0.000 0.000
 

Although there is a positive direct effect from B to D, denoted by the parameter

b41 = 0.100, the correlation between B and D when C is conditioned on is zero.
This is because conditioning on C induces a slight negative relationship between
B and D, which cancels out the positive directed effect from B to D.
It is well known that if two variables X and Y are independent but share
a collider Z on which both have a positive effect, there will be a negative condi-
tional dependency between X and Y conditional on Z (cf. VanderWeele & Robins,
2007). Suppose X and Y are standard normal variables, and fully determine the
variable Z, both having the same positive effect on Z. Now consider that we con-
dition on Z taking on its average value, Z = 0. If we also know that X takes on
a moderately low value (say, X = −0.5) then it must be the case that Y takes on a
high value (Y = 0.5) – otherwise, it would not be possible for Z to take on a value
of zero. In this way, although both X and Y are causally independent, if we know
the value of their common effect Z, then X and Y contain information which can
be used to predict one from the other.

Appendix 2.B The SE-set Algorithm: A Tool to Aid

Causal Hypothesis Generation
In this appendix we describe a tool which allows us to directly relate a given GGM
to a set of equivalent weighted DAGs (i.e., linear path models), based on SEM
principles. This tool takes as input an (estimated) precision matrix and outputs
a set of weighted DAG structures between the observed variables. While each of
these weighted DAGs may lead to very different substantive interpretations, they
all lead to the same pattern of partial correlations between the observed variables
described by the GGM.
The SE-set algorithm can be used to explore the different weighted DAG
structures which may underlie a given GGM, aiding in the generation of causal
hypotheses. As the weighted DAG structures in question are based on lin-
ear path models, we can say that they are statistically-equivalent in the sense
of a structural equation model, in that they produce the same model-implied
variance-covariance matrix (Bollen, 1989; MacCallum et al., 1993; Raykov & Mar-
coulides, 2001; Tomarken & Waller, 2003). This tool is thus called the statistical-
equivalence-set or SE-set algorithm.
Note that we are limited in the current implementation to deriving DAGs
which meet the sufficiency assumption (i.e., weighted DAGs between observed
variables). However it is not necessary to invoke the faithfulness assumption,
and so the SE-set algorithm may produce DAGs which are unfaithful to the set

35
2. The Challenge of Generating Causal Hypotheses

of conditional (in)dependencies described by the GGM. Furthermore, note that

the SE-set algorithm is intended primarily as an illustrative and exploratory tool:
The limitations of this algorithm as a stand-alone DAG estimation procedure are
addressed in the discussion section of this paper.

2.B.1 Relating the GGM and Weighted DAGs

The basis of the SE-set algorithm is that both the weights matrix of the p−variate
GGM and the weights matrix of a weighted DAG can be related to each other
through the p × p inverse variance-covariance matrix, known as the precision ma-
trix Ω̂ (see Equation (2.2)). In most applications of the GGM, for example those
using regularized estimation methods, the estimated precision matrix will not be
equivalent to the inverse of the observed covariance matrix. Instead, we can de-
scribe the inverse of the estimated precision matrix as a model-implied covariance
matrix, which we will denote Σ̃
−1
Σ̃ = Ω̂ . (2.6)
The weights matrix of a weighted DAG was defined in Equation (2.4) as the ma-
trix of regression parameters B in which child variables are regressed on their
parents. In other words, the weighted DAG can be seen as a SEM model with
uncorrelated residual terms. From the SEM literature, there is a straightforward
expression relating the parameters of such a path model to a model-implied vari-
ance covariance matrix. This relationship is given by

Σ̃ = (I − B)−1 Ψ (I − B)−T (2.7)

where Ψ is a matrix of residual variances of X, diagonal in the case of uncor-

related residuals (Bollen, 1989). As such, the weights matrix of both the GGM
and the weighted DAG can both be seen as (estimated) decompositions of some
variance-covariance matrix Σ̃. A similar expression is used by Epskamp, Wal-
dorp, et al. (2018) to illustrate the relationship between the precision matrix and
an underlying weighted DAG.
Given that we now have expressions relating the weights matrices of a GGM
to the weights matrix of a DAG, through the model-implied covariance matrix
(Equations (2.6) and (2.7)), we can use this relationship to define the SE-set al-
gorithm. If Ψ and B are both known, then they can be combined to find Σ̃ us-
ing Equation (2.7). However, the inverse operation, solving uniquely for B and
Ψ from Σ̃ is not typically possible without additional information or assump-
tions. This is a well-known issue in the SEM literature, which has long identi-
fied that many different path models imply the same variance-covariance matrix
(MacCallum et al., 1993; Raykov & Marcoulides, 2001).
One circumstance in which it is possible to find B directly from the covariance
matrix is when the topological ordering of the DAG is known, and the residual
terms are assumed to be uncorrelated (Levina et al., 2008; Shojaie & Michailidis,
2010). The topological ordering of a DAG is defined as an ordering of nodes
such that every parent comes before every child. The graph in Figure 2.1(c) has
two valid topological orderings: {Support, Pressure, Stress, Worry} and {Pressure,

36
 2.B. The SE-set Algorithm: A Tool to Aid Causal Hypothesis Generation

Support, Stress, Worry}. If the rows and columns of the covariance matrix Σ̃ are
sorted according to the topological ordering, then Equation (2.7) gives a unique
solution. In that case, B will be a lower triangular matrix with zero’s on the
diagonal, and Equation (2.7) will be equivalent to an LDLT matrix decomposition
(Abadir & Magnus, 2005).

2.B.2 From GGM to Statistical Equivalent Set

In the settings of interest in the current paper, the topological ordering of the
underlying DAG is unknown. In a system of p variables, there are p! possible
topological orderings. Thus, every one of the p! possible orderings of the rows
and columns of Σ̃ produces a (possibly distinct) weights matrix B. Typically
the number of distinct B matrices will be less than p! as some DAG structures
will have more than one equivalent topological ordering. For example, in Fig-
ure 2.1(b), the GGM is made up of four variables, Support, Pressure, Stress and
Worry. This means there are 4! = 24 possible topological orderings of an underly-
ing weighted DAG. However at least two of these topological orderings describe
the same weighted DAG structure, as described above.
The SE-set algorithm takes as input a p × p precision matrix and calculates a
corresponding B for every p! possible topological ordering of the observed vari-
ables. It does this by repeatedly re-ordering the variables in Σ̃ and applying the
transformation in Equation (2.7). This gives a set of weighted DAG models of size
p!, the weights matrices of which are collected in the SE-set B. Each weighted
DAG in B leads to the same model-implied variance covariance matrix Σ̃. This
means that, by construction, each weighted DAG in B is statistically equivalent.
This definition of statistical equivalence comes from the structural equation
modeling literature, where statistically equivalent path models are those which
yield the same fit, defined by the distance between the model-implied variance
covariance matrix Σ̃ and the population covariance matrix S, and the restrictive-
ness of the model, typically defined by the number of freely estimated parameters
(Bollen, 1989). Each element of B is statistically equivalent to each other in this
sense. By construction, each matrix B defines a path model with p(p − 1)/2 paths
and p residual variance terms, which all yield the same model-implied covariance
matrix Σ̃. That a single variance-covariance matrix can typically be generated by
a large number of distinct statistically equivalent path models is a well known is-
sue in the SEM literature (MacCallum et al., 1993; Raykov & Marcoulides, 2001;
Tomarken & Waller, 2003).
After deriving all p! possible weighted DAGs, the size of the SE-set B can be
reduced in a second step by rounding or thresholding the values in the weights
matrices, and removing duplicates. In the next subsection we discuss in greater
detail the mathematical details of the algorithm.

2.B.3 Mathematical Basis of the SE-set Algorithm

Take it that we have some p−variate set of variables V , related by a p ×p precision
matrix Ω. Each path model in the SE-set of Ω is defined by a unique ordering of

37
2. The Challenge of Generating Causal Hypotheses

the variables V , of which there are p!. The output of the algorithm is thus a set of
matrices
B = {B1 , B2 , . . . Bp! } (2.8)
of size p!. To generate B we first generate V , the set containing each p! topological
ordering of the variables V . The elements of V are thus ordered sequences of p
variable names

V = {{V1 , V2 , V3 , . . . Vp }, {V2 , V1 , V3 , . . . Vp } . . . } (2.9)

In a second step, the input precision matrix Ω is transformed to a model-implied

variance-covariance matrix Σ̃ by taking its inverse (see Equation (2.6))

Ω−1 = Σ̃

under the restriction that Σ̃ is positive definite. Each element of V defines an

ordering of the rows and columns of this matrix Σ̃. We will denote by Σ̃q the
model-implied variance covariance matrix with rows and columns ordered ac-
cording to Vq .
For each Σ̃q , we can calculate a corresponding weights matrix Bq by equating
the relationship
Σ˜q = (I − Bq )−1 Ψ q (I − Bq )−T (2.10)
where Ψ q is a p- dimensional diagonal matrix of residual covariances, to an LDLT
matrix composition
Σ˜q = LDLT (2.11)
where L = (I −Bq )−1 and D = Ψ . The LDLT decomposition is unique for positive-
definite matrices (Abadir & Magnus, 2005), giving us a unique weights matrix Bq
for a given topologically ordered covariance matrix Σ˜q .
This equivalence is based on the property that a topologically ordered weights
matrix of a DAG B will always be lower-triangular, with zero’s on the diagonal.
As a result, we can define the inverse (I − Bq )−1 as a convergent infinite series

(I − Bq )−1 = I + B + B 2 + B 3 + . . .
X∞
= Bz
z=0

where B z will also be lower-triangular with zero-diagonal for all z > 0. This
means that (I − Bq )−1 will be a lower triangular matrix with diagonal elements
equal to one (Abadir & Magnus, 2005).
In practice the LDLT decomposition of a matrix can be done by first calculat-
ing the cholesky decomposition

Σ˜q = GG T (2.12)

where G is an upper-triangular matrix, and thus

LDLT = (LD 1/2 )(LD 1/2 )T ) = G T G (2.13)

38
 2.B. The SE-set Algorithm: A Tool to Aid Causal Hypothesis Generation

Letting D1/2 = S, we can write

GT = LS (2.14)
and so
GT S−1 = L (2.15)
As S is a diagonal matrix, taking it’s inverse is the same as taking the scalar in-
verse of each diagonal element and collecting these in a diagonal matrix of the
same dimensions. As L = (I − Bq )−1 has ones on the diagonal, it must be the case
that Gii /sii = 1, for all i, which means sii = Gii . To find L = (I − Bq )−1 we simply
divide each column of G by the corresponding diagonal element in that column.
Then we have
Bq = I − L−1 (2.16)
Repeating this operation for all topological orderings in V defines the SE-set algo-
rithm. The output of this algorithm thus yields a set of matrices Bq for 1 ≤ q ≤ p!,
which are collected in the SE-set B.

2.B.4 Details of R implementation

The SE-set algorithm is implemented in the statistical programming language
R. The current implementation of this algorithm is computationally feasible for
systems of p ≤ 12 variables. This is due to vector size limits in R, which has
maximum vector length 231 − 1. The topological orderings V are generated all at
once and stored in a vector using the combinat library in R (Chasalow, 2012). This
means that topological orderings cannot be generated for p > 13 as this would
violate the maximum vector length (13! > 231 − 1).
For p ≤ 12 the current implementation runs extremely quickly on a standard
machine, taking an average of .23 seconds for the empirical analysis in the current
paper, and yields a matrix as output which is easily manipulated as a standard R
object. Limited increases in the maximum number of variables may be possible
using recursive programming, and storing objects outside of the R working envi-
ronment, but this is not implemented currently. The size of the output, and time
taken to apply all of the transformations, grows at a greater than exponential
rate with p, which means that practically applicable versions of this algorithm
may not be feasible for much larger systems.
With the intention of aiding researchers in exploring the SE-set, a number
of options are implemented to simplify and reduce the number of matrices in
the output B, and to mimic the typical context of researchers using GGM based
techniques. First, it is recommended that researchers input a precision matrix
which corresponds to a standardized covariance matrix, or which is calculated
on the basis of standardized variables, as is typically recommended in GGM ap-
plications. Pre-standardizing variables ensures that the sizes of paths are directly
comparable. By default thresholding and rounding of matrix elements is applied,
up to two decimal places, to eliminate small paths of less interest to researchers.
Second, for ease of comparison, the rows and columns of each matrix are
re-ordered to some reference-ordering, typically V1 , the ordering of rows and
columns in the input precision matrix. Third, in practice, many elements of B

39
2. The Challenge of Generating Causal Hypotheses

will describe identical linear path models: for a given weighted DAG, there may
be more than one equivalent topological ordering. The intended use of this tool is
for researchers interested in the different distinct (causal) structures which may
generate Σ̃, where we consider a-priori all distinct structures to be equally plau-
sible. As such, the implementation of this algorithm by default removes any
duplicate matrices in B.
One disadvantage to rounding and thresholding matrix elements is that el-
ements of B may no longer be statistically equivalent, that is, they may imply
different variance-covariance matrices. However these differences are likely to
be relatively small, and this disadvantage is offset by the increased simplification
and interpretability of the different path models.

Appendix 2.C Empirical Illustration Details

In this section we will provide further technical details on the empirical illus-
tration shown in the main text. We will first review the methodology used for
constructing a GGM based on the results reported by Hoorelbeke et al. (2016),
and the necessary differences in methodology used in the current paper. Follow-
ing this we will validate the SE-set analysis used in the current paper by showing
that the weighted DAGs which result from this analysis reproduce the precision
matrix which was estimated in the GGM analysis.

2.C.1 GGM Re-analysis

Hoorelbeke et al. used the parcor package in their GGM analysis, which takes as
input a full dataset, and returns a partial correlation matrix (Kraemer, Schaefer,
& Boulesteix, 2009). As the SE-set algorithm requires a precision matrix, and we
do not have access to the original dataset, it is not possible to perfectly reproduce
their estimated GGM.
Instead, we use the reported sample correlation matrix, shown in the upper
triangle of Table 2.1. We estimate the GGM using the EBICglasso function from
the package qgraph (Epskamp et al., 2012). This function estimates a regularized
precision matrix, requiring only the zero-order (marginal) correlation matrix and
sample size as input, with the selection of the tuning parameter performed au-
tomatically by using EBIC fit indices. Internally, this function first estimates the
precision matrix, before transforming the estimated precision matrix to a par-
tial correlation matrix: The estimated precision matrix can be returned using the
returnAllResults option.
The marginal correlations reported by Hoorelbeke et al. (2016), used as input
to the qgraph package, are shown in the upper-triangle of Table 2.1. The estimated
partial correlations, used to create Figure 2.5(a) are shown in the lower triangle
of Table 2.1.

40
 2.C. Empirical Illustration Details

BRIEF WM PASAT ACC Adapt ER Maladapt ER Resilience Resid Depres

BRIEF WM - .060 -.100 .280 -.660 .500
PASAT ACC 0 - -.260 .03 -.21 .090
Adapt ER 0 0 - .120 .390 -.260
Maladapt ER 0 0 0 - -.330 .300
Resilience -.325 0 .092 -.039 - -.640
Resid Depres .094 0 0 0 -.301 -

Table 2.1: Reported correlations and estimated partial correlations for the empirical example. The
upper triangle shows the marginal correlations reported by Hoorelbeke et al. (2016). The lower tri-
angle shows the partial correlations estimated using the reported correlation matrix as input to the
EBICglasso function from the qgraph package.

2.C.2 SE-set Input Recovery

To verify that the linear DAG weights matrices resulting from the SE-set algo-
rithm Bq ∈ B all approximately reproduce the input precision matrix Ω̂, we re-
calculate the precision matrix given by each of the 720 un-rounded matrices in
B using Equation (2.7). For this purpose, the Ψ matrices corresponding to each
Bq were saved when the SE-set was generated. In Table 2.2 we show the bias of
the re-calculated precision matrices, compared to the input precision matrix. We
can see that the weights matrices B reproduce the correct precision matrix, up to
the 6th decimal place. The RMSE is not reported as the error for each element
of B was very similar, meaning the RMSE is approximately equal to the absolute
value of the bias.
BRIEF WM PASAT ACC Adapt ER Maladapt ER Resilience Resid Depress
BRIEF WM 2.8 ∗ 10−6
PASAT ACC 0 0
Adapt ER −1.2 ∗ 10−8 0 4.3 ∗ 10−8
Maladapt ER 5.3 ∗ 10−9 0 −1.4 ∗ 10−9 7.5 ∗ 10−9
Resilience 1.1 ∗ 10−6 0 −4.4 ∗ 10−8 1.7 ∗ 10−8 8.3 ∗ 10−7
Resid Depress −2.3 ∗ 10−7 0 −1.1 ∗ 10−8 4.8 ∗ 10−9 9.5 ∗ 10−8 2.0 ∗ 10−7

Table 2.2: Average error (“bias”) of elements of B in reproducing the input precision matrix Ω̂. Zero
elements of Ω̂ are highlighted in red.

We can see from Table 2.2 that positive elements of Ω̂, for example the diag-
onal elements as well as ω51 , ω54 and ω65 appear to have a slight positive bias,
and negative elements (ω53 , ω61 ) a slight negative bias. The true zero-elements,
highlighted in red in the table below, have a mix of positive, negative and zero
bias. All parameters relating to PASAT ACC, which is unconnected to all other
variables, are recovered as exactly zero. However other non-zero elements are
recovered as positive or negative non-zero elements, but the size of these ele-
ments is very small. This is not necessarily problematic, as the input precision
matrix makes use of thresholding in any case, where such parameters would be
set exactly to zero.
This performance analysis is conducted on the un-rounded equivalence set,
also including duplicates. Removing duplicates does not make a substantial dif-
ference to the results due to the small variances in individual errors. Rounding

41
2. The Challenge of Generating Causal Hypotheses

makes some difference to the performance, but not substantially: the signs of
some biases change, and the absolute values of the largest biases are now larger,
of the order 10−3 .

42
 Chapter 3

A Continuous-Time Approach
to Intensive Longitudinal
Data: What, Why and How?
Abstract

The aim of this chapter is to: a) provide a broad didactic treatment

of the first-order stochastic differential equation model – also known as
the continuous-time (CT) first-order vector autoregressive (VAR(1)) model;
and b) argue for and illustrate the potential of this model for the study of
psychological processes using intensive longitudinal data. We begin by de-
scribing what the CT-VAR(1) model is, and how it relates to the more com-
monly used discrete-time VAR(1) model. Assuming no prior knowledge
on the part of the reader, we introduce important concepts for the analy-
sis of dynamic systems, such as stability and fixed points. In addition we
examine why applied researchers should take a continuous-time approach
to psychological phenomena, focusing on both the practical and concep-
tual benefits of this approach. Finally, we elucidate how researchers can
interpret CT models, describing the direct interpretation of CT model pa-
rameters as well as tools such as impulse response functions, vector fields,
and lagged-parameter plots. To illustrate this methodology we re-analyze
a single-subject experience-sampling dataset with the R package ctsem; for
didactic purposes, R code for this analysis is included, and the dataset it-
self is publicly available.

This chapter has been adapted from: Ryan, O., Kuiper, R. M. & Hamaker, E. L. (2018). A
Continuous-Time Approach to Intensive Longitudinal Data: What, Why and How? In K. v. Montfort,
J. H. L. Oud, & M. C. Voelkle (Eds.) Continuous Time Modeling in the Behavioral and Related Sciences.
New York, NY: Springer. Author contributions: OR and ELH conceptualized the initial project. OR
and RMK wrote the R code and generated figures. OR wrote the paper. RMK and ELH helped further
develop the ideas in the project, discussed progress and provided textual feedback.

43
3. A Continuous-Time Approach to Intensive Longitudinal Data

3.1 Introduction
The increased availability of intensive longitudinal data – such as obtained with
ambulatory assessments, experience sampling, ecological momentary assess-
ments and electronic diaries – has opened up new opportunities for researchers
to investigate the dynamics of psychological processes, that is, the way psycho-
logical variables evolve, vary and relate to one another over time (cf. Bolger &
Laurenceau, 2013; Hamaker et al., 2005; Chow, Ferrer, & Hsieh, 2011). A useful
concept in this respect is that of people being dynamic systems whose current
state depends on their preceding states. For instance, we may be interested in the
relationship between momentary stress and anxiety. We can think of stress and
anxiety as each defining an axis in a two-dimensional space, and let the values of
stress and anxiety at each moment in time define a position in this space. Over
time, the point that represents a person’s momentary stress and anxiety moves
through this two-dimensional space, and our goal is to understand the lawful-
ness that underlie these movements.
There are two frameworks that can be used to describe such movements: 1)
the discrete-time (DT) framework, in which the passage of time is treated in dis-
crete steps; and 2) the continuous-time (CT) framework, in which time is viewed
as a continuous variable. Most psychological researchers are at least somewhat
familiar with the DT approach, as it is the basis of the vast majority of longi-
tudinal models used in the social sciences. In contrast, CT models have gained
relatively little attention in fields such as psychology: This is despite the fact
that many psychological researchers have been advocating their use for a long
time, claiming that the CT approach overcomes practical and conceptual prob-
lems associated with the DT approach (e.g., Boker, 2002; Chow et al., 2005; Oud
& Delsing, 2010; Voelkle, Oud, Davidov, & Schmidt, 2012). We believe there are
two major hurdles that hamper the adoption of the CT approach in psychological
research. First, the estimation of CT models typically requires the use of special-
ized software (cf. Driver, Oud, & Voelkle, 2017; Chow, Ferrer, & Nesselroade,
2007; Oravecz, Tuerlinckx, & Vandekerckhove, 2016) or unconventional use of
more common software (cf. Boker, Neale, & Rausch, 2004; Boker, Deboeck, Edler,
& Keel, 2010; J. S. Steele & Ferrer, 2011). Second, the results from CT models are
not easily understood, and researchers may not know how to interpret and rep-
resent their findings.
Our goal in this chapter is twofold. First, we introduce readers to the per-
spective of psychological processes as CT processes; we focus on the conceptual
reasons for which the CT perspective is extremely valuable in moving our under-
standing of processes in the right direction. Second, we provide a didactic de-
scription of how to interpret the results of a CT model, based on our analysis of
an empirical dataset. We examine the direct interpretation of model parameters,
examine different ways in which the dynamics described by the parameters can
be understood and visualized, and explain how these are related to one another
throughout. We will restrict our primary focus to the simplest DT and CT mod-
els, that is, first-order (vector) autoregressive models and first-order differential
equations.

44
 3.2. Two Frameworks

The organization of this chapter is as follows. First, we provide an overview of

the DT and CT models under consideration. Second, we discuss the practical and
conceptual reasons researchers should adopt a CT modeling approach. Third,
we illustrate the use and interpretation of the CT model using a bivariate model
estimated from empirical data. Fourth, we conclude with a brief discussion of
more complex models which may be of interest to substantive researchers.

3.2 Two Frameworks

The relationship between the DT and CT frameworks has been discussed exten-
sively by a variety of authors. Here, we briefly reiterate the main issues, as this is
vital to the subsequent discussion. For a more thorough treatment of this topic,
the reader is referred to Voelkle et al. (2012). We begin by presenting the first-
order vector auto-regressive model in DT, followed by the presentation of the
first-order differential equation in CT. Subsequently, we show how these mod-
els are connected, and discuss certain properties which can be inferred from the
parameters of the model. For simplicity, and without loss of generalization, we
describe single-subject DT and CT models, in terms of observed variables. Exten-
sions for multiple-subject data, and extensions for latent variables, in which the
researchers can account for measurement error by additionally specifying a mea-
surement model, are readily available (in the case of CT models, see for example
Boker et al., 2004; Oravecz & Tuerlinckx, 2011; Driver et al., 2017).

3.2.1 The Discrete-Time Framework

DT models are those models for longitudinal data in which the passage of time
is accounted for only with regards to the order of observations. If the true data-
generating model for a process is a DT model, then the process only takes on
values at discrete moments in time (e.g., hours of sleep per day or monthly
salary). Such models are typically applied to data that consist of some set of vari-
ables measured repeatedly over time. These measurements typically show auto-
correlation, that is, serial dependencies between the observed values of these vari-
ables at consecutive measurement occasions. We can model these serial depen-
dencies using (discrete-time) auto-regressive equations, which describe the rela-
tionship between the values of variables observed at consecutive measurement
occasions.
The specific type of DT model that we will focus on in this chapter is the
first-order Vector Auto-Regressive (VAR(1)) model (cf. Hamilton, 1994). Given a
set of V variables of interest measured at N different occasions, the VAR(1) de-
scribes the relationship between yτ , a V × 1 column vector of variables measured
at occasion τ (for τ = 2, . . . , N ) and the values those same variables took on at the
preceding measurement occasion, the vector yτ−1 . This model can be expressed
as

yτ = c + Φyτ−1 + τ , (3.1)

45
3. A Continuous-Time Approach to Intensive Longitudinal Data

where Φ represents a V × V matrix with autoregressive and cross-lagged coef-

ficients that regresses yτ on yτ−1 . The V × 1 column vector τ represents the
variable-specific random shocks or innovations at that occasion, which are nor-
mally distributed with mean zero and a V × V variance-covariance matrix Ψ .
Finally, c represents a V × 1 column vector of intercepts.
In the case of a stationary process, the mean µ, and the variance-covariance
matrix of the variables yτ (generally denoted Σ), do not change over time.1 Then,
the vector µ represents the long-run expected values of the random variables,
E(yτ ), and is a function of the vector of intercepts and the matrix with lagged
regression coefficients, that is, µ = (I − Φ)−1 c, where I is a V × V identity matrix
(cf. Hamilton, 1994). In terms of a V -dimensional dynamical system of interest,
µ represents the equilibrium position of the system. By definition, τ is limited to
positive integers; that is, there is no .1th or 1.5th measurement occasion.
Both the single-subject and multilevel versions of the VAR(1) model have fre-
quently been used to analyze intensive longitudinal data of psychological vari-
ables, including symptoms of psychopathology, such as mood- and affect-based
measures (Browne & Nesselroade, 2005; Moberly & Watkins, 2008; Rovine &
Walls, 2006; Bringmann et al., 2016; Bringmann, Lemmens, Huibers, Borsboom,
& Tuerlinckx, 2015). In these cases, the auto-regressive parameters φii are often
interpreted as reflecting the stability, inertia or carry-over of a particular affect or
behavior (Kuppens, Allen, & Sheeber, 2010; Kuppens et al., 2012; Koval, Kup-
pens, Allen, & Sheeber, 2012). The cross-lagged effects (i.e., the off-diagonal ele-
ments φij for i , j) quantify the lagged relationships, sometimes referred to as the
spill-over, between different variables in the model. These parameters are often
interpreted in substantive terms, either as predictive or Granger-causal relation-
ships between different aspects of affect or behavior (Granger, 1969; Ichii, 1991;
Watkins, Lei, & Canivez, 2007; Gault-Sherman, 2012; Bringmann et al., 2013).
For example, if the standardized cross-lagged effect of y1,τ−1 on y2,τ is larger than
the cross-lagged effect of y2,τ−1 on y1,τ , researchers may draw the conclusion that
y1 is the driving force or dominant variable of that pair (Schuurman, Ferrer, de
Boer-Sonnenschein, & Hamaker, 2016). As such, substantive researchers are typ-
ically interested in the (relative) magnitudes and signs of these parameters.

3.2.2 The Continuous-Time Framework

In contrast to the DT framework, which treats values of processes indexed by
observation τ, the CT framework treats processes as functions of the continuous
variable time t: The processes being modeled are assumed to vary continuously
with respect to time, meaning that these variables may take on values if observed
at any imaginable moment. CT processes can be modeled using a broad class of
differential equations, allowing for a wide degree of diversity in the types of dy-
namics that are being modeled. It is important to note that many DT models have
1 The variance-covariance matrix of the variables Σ is a function of both the lagged parameters and
the variance-covariance matrix of the innovations, vec(Σ) = (I −Φ ⊗Φ)−1 vec(Ψ ), where vec(.) denotes
the operation of putting the elements of an N × N matrix into an N N × 1 column matrix (p.27 Kim &
Nelson, 1999)

46
 3.2. Two Frameworks

a differential equation counterpart. For the VAR(1) model, the CT equivalent is

the first-order Stochastic Differential Equation (SDE), where stochastic refers to
the presence of random innovations or shocks.
The first-order SDE describes how the position of the V -dimensional system
at a certain point in time, y(t), relative to the equilibrium position µ, is related
dy(t)
to the rate of change of the process with respect to time (i.e., dt ) in that same
instant. The latter can also be thought of as a vector of velocities, describing in
what direction and with what magnitude the system will move an instant later
in time (i.e., the ratio of the change in position over some time-interval, to the
length of that time-interval, as the length of the time-interval approaches zero).
The first-order SDE can be expressed as

dy(t) dW (t)
= A(y(t) − µ) + G . (3.2)
dt dt
The model representation in Equation (3.2) is referred to as the differential
dy(t)
form as it includes the derivative dt . The same model can be represented in the
integral form, in which the derivatives are integrated out, sometimes referred to
as the solution of the derivative model. The integral form of this particular first-
order differential equation is known as the CT-VAR(1) or Ornstein-Uhlenbeck
model (Oravecz, Tuerlinckx, & Vandekerckhove, 2011). In this form, we can de-
scribe the same system, but now in terms of the positions of the system (i.e., the
values the variables take on) at different points in time. For notational simplic-
ity, we can represent y(t) − µ as y c (t), denoting the position of the process as a
deviation from its equilibrium.
The CT-VAR(1) model can be written as

y c (t) = eA∆t y c (t − ∆t) + w(∆t) (3.3)

where A has the same meaning as above, the V × 1 vector y c (t − ∆t) represents the
position as a deviation from equilibrium some time-interval ∆t earlier, e repre-
sents the matrix exponential function, and the V × 1 column vector w(∆t) repre-
sents the stochastic innovations, the integral form of the Wiener process in Equa-
tion (3.2). These innovations are normally distributed with a variance-covariance
matrix that is a function of the time-interval between measurements ∆t, the drift
matrix A, and the diffusion matrix Γ (cf. Voelkle et al., 2012).2 As the variables
in the model have been centered around their equilibrium, we omit any intercept
term. The relationship between lagged variables, that is, the relationships be-
tween the positions of the centered variables in the multivariate space, separated
by some time-interval ∆t, is an (exponential) function of the drift matrix A and
the length of that time-interval.
2 Readers should note that there are multiple different possible ways to parameterize the CT
stochastic process in integral form, and also multiple different notations used (e.g., Oravecz et al.,
2011; Voelkle et al., 2012).

47
3. A Continuous-Time Approach to Intensive Longitudinal Data

3.2.3 Relating DT and CT Models

It is clear from the integral form of the first-order SDE given in Equation (3.3) that
the relationship between lagged values of variables is dependent on the length of
the time-interval between these lagged values. As such, if the DT-VAR(1) model
in Equation 3.1 is fitted to data generated by the CT model considered here, then
the auto-regressive and cross lagged effects matrix Φ will be a function of the
time-interval ∆t between the measurements. We denote this dependency by writ-
ing Φ(∆t). This characteristic of the DT model has been referred to as the lag or
time-interval problem (Gollob & Reichardt, 1987; Reichardt, 2011).
The precise relationship between the CT-VAR(1) and DT-VAR(1) effects matri-
ces is given by the well-known equality

Φ(∆t) = eA∆t . (3.4)

Despite this relatively simple relationship, it should be noted that taking the ex-
ponential of a matrix is not equivalent to taking the exponential of each of the
elements of the matrix. That is, any lagged effect parameter φij (∆t), relating vari-
able i and variable j across time-points, is not only dependent on the correspond-
ing CT cross-effect aij , but is a non-linear function of the interval and every other
element of the matrix A. For example, in the bivariate case the DT cross-lagged
effect of y1 (t − ∆t) on y2 (t), denoted φ21 (∆t), is given by
1
√2 2 1
√2 2
a21 (e 2 (a11 +a22 + a11 +4a12 a21 −2a11 a22 +a22 )∆t − e 2 (a11 +a22 − a11 +4a12 a21 −2a11 a22 +a22 )∆t )
q
a211 + 4a12 a21 − 2a11 a22 + a222
(3.5)

where e represents the scalar exponential. In higher dimensional models, these

relationships quickly become intractable. For a derivation of Equation 3.5 we
refer readers to Appendix 3.A.
This complicated non-linear relationship between the elements of Φ and the
time-interval has major implications for applied researchers who wish to inter-
pret the parameters of a DT-VAR(1) model in the substantive terms outlined
above. In the general multivariate case, the size, sign, and relative strengths of
both auto-regressive and cross-lagged effects may differ depending on the value
of the time-interval used in data collection (Oud, 2007; Reichardt, 2011; Dor-
mann & Griffin, 2015; Deboeck & Preacher, 2016). As such, conclusions that
researchers draw regarding the stability of processes, and the nature of how dif-
ferent processes relate to one another may differ greatly depending on the time-
interval used.
While the relationship in Equation 3.4 describes the DT-VAR(1) effects matrix
we would find given data generated by a CT-VAR(1) model, the reader should
note that not all DT-VAR(1) processes have a straightforward equivalent repre-
sentation as a CT-VAR(1). For example, a univariate discrete-time AR(1) process
with a negative auto-regressive parameter cannot be represented as a CT-AR(1)
process; as the exponential function is always positive, there is no A that satisfies

48
 3.2. Two Frameworks

Equation 3.4 for Φ < 0. As such, we can refer to DT-VAR(1) models with a CT-
VAR(1) equivalent as those which exhibit ‘positive autoregression’. We will focus
throughout on the CT-VAR(1) as the data-generating model.3

3.2.4 Types of Dynamics: Eigenvalues, Stability and Equilib-

rium
Both the DT-VAR(1) model and the CT-VAR(1) model can be used to describe a
variety of different types of dynamic behavior. As the dynamic behavior of a sys-
tem is always understood with regards to how the variables in the system move
in relation to the equilibrium position, often dynamic behaviors are described
by differentiating the type of equilibrium position or fixed point in the system
(Strogatz, 2015). In the general multivariate case, we can understand these dif-
ferent types of dynamic behavior or fixed points with respect to the eigenvalues
of the effects matrices A or Φ (see Appendix 3.A for a more detailed explanation
of the relationship between these two matrices and eigenvalues). In this chapter
we will focus on stable processes, in which, given a perturbation, the system of
interest will inevitably return to the equilibrium position. We limit our treat-
ment to these types of processes, because we believe these are most common and
most relevant for applied researchers. A brief description of other types of fixed
points and how they relate to the eigenvalues of the effects matrix A is given in the
discussion section – for a more complete taxonomy we refer readers to Strogatz
(2015).
In DT settings, stable processes are those for which the absolute values of the
eigenvalues of Φ are smaller than one. In DT applications researchers also typi-
cally discuss the need for stationarity, that is, time-invariant mean and variance,
as introduced above. Stability of a process ensures that stationarity in relation
to the mean and variance hold. For CT-VAR(1) processes, stability is ensured if
the real-parts of the eigenvalues of A are negative. It is interesting to note that
the equilibrium position of stable processes can be related to our observed data
in various ways: In some applications µ is constrained to be equal to the mean
of the observed values (e.g., Hamaker & Grasman, 2015; Hamaker et al., 2005),
while in others the equilibrium can be specified a-prior or estimated to be equal
to some (asymptotic) value (e.g., Bisconti, Bergeman, & Boker, 2004).
We can further distinguish between dynamic processes that have real eigen-
values, complex eigenvalues, or in the case of systems with more than two vari-
ables, a mix of both. In the section “Making Sense of CT Models” we will focus
on the interpretation of a CT-VAR(1) model with real, negative, non-equal eigen-
values. We can describe the equilibrium position of this system as a stable node.
In the discussion section we examine another type of system which has been the
focus of psychological research, sometimes described as a damped linear oscilla-
3 In general, there is no straightforward CT-VAR(1) representation of DT-VAR(1) models with real,
negative eigenvalues. However it may be possible to specify more complex continuous-time models
which do not exhibit positive autoregression. Notably, M. Fisher (2001) demonstrates how a DT-AR(1)
model with negative autoregressive parameter can be modeled with the use of two continuous-time
(so-called) Ito processes.

49
3. A Continuous-Time Approach to Intensive Longitudinal Data

tor (e.g., Boker, Montpetit, Hunter, & Bergeman, 2010), in which the eigenvalues
of A are complex, with a negative real part. The fixed point of such a system is
described as a stable spiral. Further detail on the interpretation of these two types
of systems is given in the corresponding sections.

3.3 Why Researchers Should Adopt a CT Perspective

There are both practical and theoretical benefits to CT model estimation over
DT model estimation. Here we will discus three of these practical advantages
which have received notable attention in the literature. We then discuss the fun-
damental conceptual benefits of treating psychological processes as continuous-
time systems.
The first practical benefit to CT model estimation is that the CT model deals
well with observations taken at unequal intervals, often the case in experience
sampling and ecological momentary assessment datasets (Oud & Jansen, 2000;
Voelkle et al., 2012; Voelkle & Oud, 2013). Many studies use random intervals
between measurements, for example to avoid participant-anticipation of mea-
surement occasions, potentially resulting in unequal time-intervals both within
and between participants. The DT model, however, is based on the assumption
of equally spaced measurements, and as such estimating the DT model from un-
equally spaced data will result in an estimated Φ matrix that is a blend of differ-
ent Φ(∆t) matrices for a range of values of ∆t.
The second practical benefit of CT modeling over DT modeling is that, when
measurements are equally spaced, the lagged effects estimated by the DT models
are not generalizable beyond the time-interval used in data collection. Several
different researchers have demonstrated that utilizing different time-intervals of
measurement can lead researchers to reach very different conclusions regarding
the values of parameters in Φ (Reichardt, 2011; Oud & Jansen, 2000; Voelkle
et al., 2012). The CT model has thus been promoted as facilitating better com-
parisons of results between studies, as the CT effects matrix A is independent
of time-interval (assuming a sufficient frequency of measurement to capture the
relevant dynamics).
Third, the application of CT models allows us to explore how cross-lagged
effects are expected to change depending on the time-interval between measure-
ments, using the relationship expressed in Equation (3.4). Some authors have
used this relationship to identify the time-interval at which cross-lagged effects
are expected to reach a maximum (Deboeck & Preacher, 2016; Dormann & Grif-
fin, 2015). Such information could be used to decide upon the ‘optimal’ time-
interval that should be used in gathering data in future research.
While these practical concerns regarding the use of DT models for CT pro-
cesses are legitimate, there may be instances in which alternative practical so-
lutions can be used, without necessitating the estimation of a CT model. For
instance, the problem of unequally spaced measurements in DT modeling can
be addressed by defining a time grid and adding missing data to your observa-
tions, to make the occasions approximately equally spaced in time. Some simu-

50
 3.4. Making Sense of CT Models

lation studies indicate that this largely reduces the bias that results from using
DT estimation of unequally spaced data (De Haan-Rietdijk, Voelkle, Keijsers, &
Hamaker, 2017).
Furthermore, the issue of comparability between studies that use different
time-intervals can be solved, in certain circumstances, by a simple transforma-
tion of the estimated Φ matrix, described in more detail by Kuiper and Ryan
(2018). Given an estimate of Φ(∆t) we can solve for the underlying A using
Equation 3.4. This is known as the “indirect method” of CT model estimation
(Oud, van Leeuwe, & Jansen, 1993). However this approach cannot be applied
in all circumstances, as it involves using the matrix logarithm, the inverse of the
matrix exponential function. As the matrix logarithm function in the general
case does not give a unique solution, this method is only appropriate if both the
estimated Φ(∆t) and true underlying A matrices have real eigenvalues only (for
further discussion of this issue see Hamerle, Nagl, & Singer, 1991).
However, the CT perspective has added value above and beyond the poten-
tial practical benefits discussed above. Multiple authors have argued that psy-
chological phenomena, such as stress, affect and anxiety, do not vary in discrete
steps over time, but likely vary and evolve in a continuous and smooth manner
(Boker, 2002; Gollob & Reichardt, 1987). Viewing psychological processes as CT
dynamic systems has important implications for the way we conceptualize the
influence of psychological variables on each other. Gollob and Reichardt (1987)
give the example of a researcher who is interested in the effect of taking aspirin
on headaches: This effect may be zero shortly after taking the painkiller, substan-
tial an hour or so later, and near zero again after twenty-four hours. All of these
results may be considered as accurately portraying the effect of painkillers on
headaches for a specific time-interval, although each of these intervals considered
separately represent only a snapshot of the process of interest.
It is only through examining the underlying dynamic trajectories, and ex-
ploring how the cross-lagged relationships evolve and vary as a function of the
time-interval, that we can come to a more complete picture of the dynamic sys-
tem of study. We believe that – while the practical benefits of CT modeling are
substantial – the conceptual framework of viewing psychological variables as CT
processes has the potential to transform longitudinal research in this field.

3.4 Making Sense of CT Models

In this section, we illustrate how researchers can evaluate psychological variables
as dynamic CT processes by describing the interpretation of the drift matrix pa-
rameters A. We describe multiple ways in which the dynamic behavior of the
model in general, as well as specific model parameters, can be understood. In
order to aid researchers who are unfamiliar with this type of analysis, we take a
broad approach in which we incorporate the different ways in which researchers
interested in dynamical systems and similar models interpret their results. For
instance, Boker and colleagues (e.g., Boker, Montpetit, et al., 2010) typically inter-
pret the differential form of the model directly; in the econometrics literature it is

51
3. A Continuous-Time Approach to Intensive Longitudinal Data

typical to plot specific trajectories using Impulse Response Functions (Johnston

& DiNardo, 1972); in the physics tradition, the dynamics of the system are in-
spected using Vector Fields (e.g., Boker & McArdle, 1995); the work of Voelkle,
Oud and others (e.g., Voelkle et al., 2012; Deboeck & Preacher, 2016) typically
focuses on the integral form of the equation, and visually inspecting the time-
interval dependency of lagged effects.
We will approach the interpretation of a single CT model from these four
angles, and show how they each represent complimentary ways to understand
the same system. For ease of interpretation we focus here on a bivariate system;
the analysis of larger systems is addressed in the discussion section.

3.4.1 Substantive Example from Empirical Data

To illustrate the diverse ways in which the dynamics described by the CT-VAR(1)
model can be understood, we make use of a substantive example. This exam-
ple is based on our analysis of a publicly-available single-subject ESM dataset
(Kossakowski, Groot, Haslbeck, Borsboom, & Wichers, 2017). The subject in
question is a 57-year old male with a history of major depression. The data
consists of momentary, daily and weekly items relating to affective states. The
assessment period includes a double-blind phase in which the dosage of the par-
ticipants anti-depression medication was reduced. We select only those mea-
surements made in the initial phases of the study, before medication reduction;
it is only during this period that we would expect the system of interest to be
stable. The selected measurements consists of 286 momentary assessments over
a period of 42 consecutive days. The modal time-interval between momentary
assessments was 1.766 hours (inter-quartile range of 1.250 to 3.323).
For our analysis we selected two momentary assessment items, “I feel down”
and “I am tired”, which we will name Down (Do) and Tired (T i), respectively.
Feeling down is broadly related to assessments of negative affect (Meier & Robin-
son, 2004), and numerous cross-sectional analyses have suggested a relationship
between negative affect and feelings of physical tiredness or fatigue (e.g., Denol-
let & De Vries, 2006). This dataset afforded us the opportunity to investigate the
links between these two processes from a dynamic perspective. Each variable was
standardized before the analysis to facilitate ease of interpretation of the param-
eter estimates. Positive values of Do indicate that the participant felt down more
than average, negative values indicate below-average feelings of being down, and
likewise for positive and negative values of T i.
The analysis was conducted using the ctsem package in R (Driver et al., 2017).
Full details of the analysis can be found in Appendix 3.B. Parameter estimates
and standard errors are given in Table 3.1, including estimates of the stochastic
part of the CT model, represented by the diffusion matrix Γ . The negative value
of γ21 indicates that there is a negative co-variance between the stochastic input
to the rates of change of Do and T i; in terms of the CT-VAR(1) representation,
there is a negative covariance between the residuals of Do and T i in the same
measurement occasion. Further interpretation of the diffusion matrix falls be-
yond the scope of the current chapter. As the analysis is meant as an illustrative

52
 3.4. Making Sense of CT Models

Parameter Value Std. Error

a11 -0.995 0.250
a21 0.375 0.441
a12 0.573 0.595
a22 -2.416 1.132
γ11 1.734 0.612
γ21 -0.016 0.650
γ22 4.606 1.374

Table 3.1: Parameter estimates from substantive example

example only, we will throughout interpret the estimated drift matrix parameter
as though they are true population parameters.

3.4.2 Interpreting the Drift Parameters

The drift matrix relating the processes Down (Do(t)) and Tired (T i(t)) is given by
" #
−0.995 0.573
A= . (3.6)
0.375 −2.416

As the variables are standardized, the equilibrium position is µ = [0, 0] (i.e.,

E[Do(t)] = E[T i(t)] = 0). The drift matrix A describes how the position of the
system at any particular time t (i.e., Do(t) and T i(t)) relates to the velocity or rate
of change of the process, that is, how the position of the process is changing. The
system of equations which describe the dynamic system made up of Down and
Tired is given by

 h dDo(t) i  " #" #

 E
 = −0.995 0.573 Do(t)

 h dt i (3.7)
dT i(t) 0.375 −2.416 T i(t)
E dt
 

such that
" #
dDo(t)
E = −0.995Do(t) + 0.573T i(t) (3.8)
dt
" #
dT i(t)
E = 0.375Do(t) − 2.416T i(t) (3.9)
dt

where the rates of change of Down and Tired at any point in time are both de-
pendent on the positions of both Down and Tired at that time.
Before interpreting any particular parameter in the drift matrix, we can deter-
mine the type of dynamic process under consideration by inspecting the eigen-
values of A. The eigenvalues of A are λ1 = −2.554 and λ2 = −0.857; since both
eigenvalues are negative, the process under consideration is stable. This means
that if the system takes on a position away from equilibrium (e.g., due to a ran-
dom shock from the stochastic part of the model on either Down or Tired), the

53
3. A Continuous-Time Approach to Intensive Longitudinal Data

system will inevitably return to its equilibrium position over time. It is for this
reason that the equilibrium position or fixed point in stable systems is also de-
scribed as the attractor point, and stable systems are described as equilibrium-
reverting. As the eigenvalues of the system are real-valued as well as negative,
the system returns to equilibrium with an exponential decay; when the process is
far away from the equilibrium, it takes on a greater velocity, that is, moves faster
towards equilibrium. We can refer to the type of fixed point in this system as a
stable node (Strogatz, 2015).
Typical of such an equilibrium-reverting process, we see negative CT auto-
effects a11 = −0.995 and a22 = −2.416. This reflects that, if either variable in
the system takes on a position away from the equilibrium, they will take on a
velocity of opposite sign to this deviation, that is, a velocity which returns the
process to equilibrium. For higher values of Do(t), the rate of change of Do(t) is
of greater (negative) magnitude, that is, the velocity towards the equilibrium is
higher. In addition, the auto-effect of T i(t) is more than twice as strong (in an
absolute sense) as the auto-effect of Do(t). If there were no cross-effects present,
this would imply that T i(t) returns to equilibrium faster than Do(t); however,
as there are cross-effects present, such statements cannot be made in the general
case from inspecting the auto-effects alone.
In this case the cross-effects of Do(t) and T i(t) on each others rates of change
are positive rather than negative. Moreover, the cross-effect of T i(t) on the rate
of change of Do(t) (a12 = 0.573) is slightly stronger than the corresponding cross-
effect of Do(t) on the rate of change of T i(t) (a21 = 0.375). These cross-effects
quantify the force that each component of the system exerts on the other. How-
ever, depending on what values each variable takes on at a particular point in
time t, that is, the position of the system in each of the Do(t) and T i(t) dimen-
sions, this may translate to Do(t) pushing T i(t) to return faster to its equilibrium
or to deviate away from its equilibrium position, and vice versa. To better un-
derstand both the cross-effects and auto-effects described by A, it is helpful to
visualize the possible trajectories of our two-dimensional system.

3.4.3 Visualizing Trajectories

We will now describe and apply two related tools which allow us to visualize the
trajectories of the variables in our model over time: Impulse Response Functions
and Vector Fields. These tools can help us to understand the dynamic system
we are studying, by exploring the dynamic behavior which results from the drift
matrix parameters.

3.4.3.1 Impulse Response Functions

Impulse Response Functions (IRFs) are typically used in the econometrics litera-
ture to aid in making forecasts based on a DT-VAR model. The idea behind this
is to allow us to explore how an impulse to one variable in the model at occa-
sion τ will affect the values of both itself and the other variables in the model
at occasions τ + 1, τ + 2, τ + 3 and so on. In the stochastic systems we focus on

54
 3.4. Making Sense of CT Models

in this chapter, we can conceptualize these impulses as random perturbations or

innovations, or alternatively as external interventions in the system.4 IRFs thus
represent the trajectories of the variables in the model over time, following a
particular impulse, assuming no further stochastic innovations (see Chapter 9 of
Johnston & DiNardo, 1972, for more detail).
To specify impulses in an IRF, we generally assign a value to a single variable
in the system at some initial occasion, yi,τ . The corresponding values of the other
variables at the initial occasion yj,τ , j , i are usually calculated based on, for
instance, the covariance in the stochastic innovations, Ψ , or the stable covariance
between the processes Σ. Such an approach is beneficial in at least two ways:
first, it allows researchers to specify impulses which are more likely to occur in
an observed dataset; second, it aids researchers in making more accurate future
predictions or forecasts. For a further discussion of this issue in relation to DT-
VAR models, we refer the reader to Johnston and DiNardo (1972) pages 298-
300. Below, we will take a simplified approach and specify bi-variate impulses at
substantively interesting values.
The IRF can easily be extended for use with the CT-VAR(1) model. We can
calculate the impulse response of our system by taking the integral form of the
CT-VAR(1) model in Equation (3.3) and a) plugging in the A matrix for our sys-
tem, b) choosing some substantively interesting set of impulses y(t = 0) and c)
calculating y(t) for increasing values of t > 0. To illustrate this procedure, we will
specify four substantively interesting sets of impulses. The four sets of impulses
shown here include y(0) = [1, 0], reflecting what happens when Do(0) takes on
a positive value 1 standard deviation above the persons mean, while T i(0) is at
equilibrium; y(0) = [0, 1] reflecting when T i(0) takes on a positive value of corre-
sponding size while Do(0) is at equilibrium; y(0) = [1, 1] reflecting what happens
when Do(0) and T i(0) both take on values 1 standard deviation above the mean;
and y(0) = [1, −1] reflecting what happens when Do(0) and T i(0) take on values
of equal magnitude but opposite valence (1SD more and 1SD less than the mean
respectively). Figures 3.1(a) to 3.1(d) contain the IRFs for both processes in each
of these four scenarios.
Examining the IRFs shows us the equilibrium-reverting behavior of the sys-
tem: Given any set of starting values, the process eventually returns, in an expo-
nential fashion, to the bivariate equilibrium position where both processes take
on a value of zero. In Figure 3.1(a), we can see that when T i(t) is at equilib-
rium and Do(0) takes on a value of plus one, then T i(t) is pushed away from
equilibrium in the same (i.e., positive) direction. In substantive terms, when our
participant is feeling down at a particular moment, he begins to feel a little tired.
Eventually, both Do(t) and T i(t) return to equilibrium due to their negative auto-
effects. The feelings of being down and tired have returned to normal around
t = 4, that is, four hours after the initial impulse; stronger impulses (|Do(0)| > 1)
will result in the system taking longer to return to equilibrium, and weaker im-
pulses (|Do(0)| < 1) would dissipate quicker.
Figure 3.1(b) shows the corresponding reaction of Do(t) at equilibrium to a
4 Similar functions can be used for deterministic systems (those without a random innovation part),
however in these cases the term initial value is more typically used.

55
3. A Continuous-Time Approach to Intensive Longitudinal Data

1.0

1.0
Do(t) Do(t)
Ti(t) Ti(t)
0.5

0.5
Variable Values Y(t)

Variable Values Y(t)

0.0

0.0
−0.5

−0.5
−1.0

−1.0
0 1 2 3 4 0 1 2 3 4
Time t Time t

(a) Do(0) = 1, T i(0) = 0 (b) Do(0) = 0, T i(0) = 1

1.0

1.0

Do(t) Do(t)
Ti(t) Ti(t)
0.5

0.5
Variable Values Y(t)

Variable Values Y(t)

0.0

0.0
−0.5

−0.5
−1.0

−1.0

0 1 2 3 4 0 1 2 3 4
Time t Time t

(c) Do(0) = 1, T i(0) = 1 (d) Do(0) = 1, T i(0) = −1

Figure 3.1: Impulse response function for the model in Equation (3.9) for four different sets of im-
pulses; red solid line = Do(t) and blue dashed line = T i(t).

positive value of T i(0). We can further see that the deviation of Do(t) in Fig-
ure 3.1(b) is greater than the deviation of T i(t) in Figure 3.1(a): a positive value
of T i(t) exerts a greater push on Do(t) than vice versa, because of the greater
cross-effect of T i(t) on Do(t). In this case this strong cross-effect, combined with
the relatively weaker auto-effect of Do(t), results in Do(t) taking on a higher value
than T i(t) at around t = 1, one hour after the initial impulse. Substantively, when
our participant is feeling physically tired at a particular moment (Figure 3.1(b)),
he begins to feel a down over the next couple of hours, before eventually these
feelings return to normal (again in this case, around 4 hours later).
Figures 3.1(c) further demonstrates the role of the negative auto-effects and
positive cross-effects in different scenarios. In Figure 3.1(c), both processes take
on positive values at t = 0; the positive cross-effects result in both processes re-
turning to equilibrium at a slower rate than in Figures 3.1(a) and 3.1(b). In sub-
stantive terms this means that, when the participant is feeling very down, and
very tired, it takes longer for the participant to return to feeling normal. Here

56
 3.4. Making Sense of CT Models

also the stronger auto-effect of T i(t) than Do(t) is evident: although both pro-
cesses start at the same value, an hour later T i(1) is much closer to zero than
Do(1), that is, T i(t) decays faster to equilibrium than Do(t). In substantive terms,
this tells us that when the participant is feeling down and physically tired, he re-
covers much quicker from the physical tiredness than he does from feeling down.
In Figure 3.1(d), we see that Do(0) and T i(0) taking on values of opposite signs
results in a speeding-up of the rate at which each variable decays to equilibrium.
The auto-effect of Do(t) is negative, which is added to by the positive cross-effect
of T i(t) multiplied by the negative value of T i(0). This means that Do(0) in Fig-
ure 3.1(d) takes on a stronger negative velocity, in comparison to Figures 3.1(a)
or 3.1(c). A positive value for Do(0) has a corresponding effect of making T i(0)
take on an even stronger positive velocity. Substantively, this means that when
the participant feels down, but feels less tired (i.e. more energetic) than usual,
both of these feelings wear off and return to normal quicker than in the other
scenarios we examined. The stronger auto-effect of T i(t), in combination with
the positive cross-effect of Do(t) on T i(t), actually results in T i(t) shooting past
the equilibrium position in the T i(t) dimension (T i(t) = 0) and taking on positive
values around t = 1.5, before the system as a whole returns to equilibrium. Sub-
stantively, when the participant is feeling initially down but quite energetic, we
expect that he feels a little bit more tired than usual about an hour and half later,
before both feelings return to normal.

3.4.3.2 Vector Field

Vector fields are another technique which can be used to visualize the dynamic
behavior of the system by showing potential trajectories through a bivariate
space. In our case the two axes of this space are Do(t) and T i(t). The advantage of
vector fields over IRFs in this context is that in one plot it shows how, for a range
of possible starting positions, the process is expected to move in the (bivariate)
space a moment later. For this reason, the vector field is particularly useful in
bivariate models with complex dynamics, in which it may be difficult to obtain
the full picture of the dynamic system from a few IRFs alone. Furthermore, by
showing the dynamics for a grid of values, we can identify areas in which the
movement of the process is similar
h dy(t) i or differs.
To create a vector field, E dt is calculated for a grid of possible values for
y1 (t) and y2 (t) covering the full range of the values both variables can take on.
The vector field for Do(t) and T i(t) is shown in Figure 3.2. The base of each
arrow represents a potential position of the process y(t). The head of the arrow
represents where the process will be if we take one small step in time forward,
that is the value of y(t + ∆t) as ∆t approaches zero. In other words, the arrows
in this vector field represent the information of two derivatives, dDo(t)/dt and
dT i(t)/dt. Specifically, the direction the arrow is pointing is a function of the sign
(positive or negative) of the derivatives, while the length of the arrow represents
the magnitude of this movement, and is a function of the absolute values of the
derivative(s).
If an arrow in the vector field is completely vertical, this means that, for that

57
3. A Continuous-Time Approach to Intensive Longitudinal Data

1.0
0.5 b) c)
Ti(t)

0.0

a)
−0.5
−1.0

d)

−1.0 −0.5 0.0 0.5 1.0

Do(t)

Figure 3.2: Vector field for Do(t) and T i(t), including blue and red nullclines.

position, taking one small step forward in time would result in a change in the
system’s position along the T i(t) axis (i.e., a change in the value of Tired), but
not along the Do(t) axis (that is, dDo(t)/dt = 0 and dT i(t)/dt , 0). The converse
is true for a horizontal arrow (that is, dDo(t)/dt , 0 and dT i(t)/dt = 0). The
two lines in Figure 3.2, blue and red, identify at which positions dDo(t)/dt = 0
and dT i(t)/dt = 0, respectively; these are often referred to as nullclines or equiv-
alently, solution lines. If the nullclines are not perfectly perpendicular to one
another, this is due to the presence of at least one cross-effect. The point at which
these nullclines cross represents the equilibrium position in this two-dimensional
space, here located at Do(t) = 0, T i(t) = 0. The crossing of these nullclines splits
the vector field in four quadrants, each of which is characterized by a different
combination of negative and positive values for dDo(t)/dt and dT i(t)/dt. The top
left and bottom right quadrants represent areas in which the derivatives are of
opposite sign, dDo(t)/dt > 0 & dT i(t)/dt < 0 and dDo(t)/dt < 0 & dT i(t)/dt > 0,
respectively. The top right and bottom left quadrants represent areas where the
derivatives are of the same sign, dDo(t)/dt < 0 & dT i(t) < 0 and dDo(t)/dt > 0 &
dT i(t) > 0, respectively.
By tracing a path through the arrows, we can see the trajectory of the sys-
tem of interest from any point in the possible space of values. In Figure 3.2, we

58
 3.4. Making Sense of CT Models

include the same four bi-variate trajectories as we examined with the IRFs. In-
stead of the IRF representation of two variables whose values are changing, the
vector field represents this as the movement of one process in a two-dimensional
space. For instance, the trajectory starting at Do(t) = 0 and T i(t) = 1 begins in
the top-left quadrant, where dDo(t)/dt is positive and dT i(t)/dt is negative; this
implies that the value of Down will increase, and the value of Tired will decrease
(as can be seen in Figure 3.1(b)). Instead of moving directly to the equilibrium
along the T i(t) dimension, the system moves away from equilibrium along the
Do(t) dimension, due to the cross-effect of T i(t) on Do(t), until it moves into
the top-right quadrant. In this quadrant, dDo(t)/dt and dT i(t)/dt are both nega-
tive; once in this quadrant the process moves towards equilibrium, tangent to the
dDo(t)/dt nullcline. The other trajectories in Figure 3.2 analogously describe the
same trajectories as in Figure 3.1(a), 3.1(c) and 3.1(d).
In general, the trajectories in this vector field first decay quickest along the
T i(t) dimension, and slowest along the Do(t) dimension. This can be clearly seen
in trajectories b), c), and d). Each of these trajectories first change steeply in the
dDo(t)
T i(t) dimension, before moving to equilibrium at a tangent to the red ( dt )
nullcline. This general property of the bi-dimensional system is again related to
the much stronger auto-effect of T i(t), and the relatively small cross-effects. In a
technical sense we can say that that Do(t) represents the ‘slowest eigen-direction’
(see Strogatz, 2015, Chapter 5).

3.4.4 Inspecting the Lagged Parameters

Another way to gain insight into the processes of interest is by determining the
relationships between lagged positions of the system, according to our drift ma-
trix. To this end, we can use Equation (3.4) to determine Φ(∆t) for some ∆t.
For instance, we can see that the auto-regressive and cross-lagged relationships
between values of Down and Tired given ∆t = 1 are
" #
0.396 0.117
Φ(∆t = 1) = . (3.10)
0.077 0.106

For this given time-interval, the cross-lagged effect of Down on Tired (φ21 (∆t =
1) = 0.077) is smaller than the cross-lagged effect of Tired on Down (φ12 (∆t = 1) =
0.117). However, as shown in Equation (3.5) the value of each of these lagged ef-
fects changes in a non-linear way depending on the time-interval chosen. To
visualize this, we can calculate Φ(∆t) for a range of ∆t, and represent this infor-
mation graphically in a lagged parameter plot, as in Figure 3.3.
From Figure 3.3, we can see that both cross-lagged effects reach their maxi-
mum (and have their maximum difference) at a time-interval of ∆t = 0.65; fur-
thermore, we can see that the greater cross-effect (a12 ) results in a stronger cross-
lagged effect φ12 (∆t) for a range of ∆t. Moreover, we can visually inspect how
the size of each of the effects of interest, as well as the difference between these
effects, varies according to the time-interval. From a substantive viewpoint, we
could say that the effect of feeling physically tired has the strongest effect on
feelings of being down around 40 minutes later.

59
3. A Continuous-Time Approach to Intensive Longitudinal Data

1.0
φ11(∆t)
φ22(∆t)
φ12(∆t)
φ21(∆t)
0.8
0.6
Φ(∆t)

0.4
0.2
0.0

0 1 2 3 4

Time−Interval ∆t

Figure 3.3: The elements of Φ(∆t) for the bivariate example (i.e., φ11 (∆t), φ12 (∆t), φ21 (∆t), φ22 (∆t))
plotted for a range of values for ∆t.

While the shape of the lagged parameters may appear similar to the shapes
of the trajectories plotted in the IRFs, lagged parameter plots and IRFs represent
substantively different information. IRFs plot the positions of each variable in
the system as they change over time, given some impulse (y(t) vs t given some
y(0)). In contrast, lagged parameter plots show how the lagged relationships
change depending on the length of the time-interval between them, independent
of impulse values (eA∆t vs ∆t). The lagged relationships can be thought of as the
components which go into determining any specific trajectories.

3.4.5 Caution With Interpreting Estimated Parameters

It is important to note that in the above interpretation of CT models, we have
treated the matrix A as known. In practice of course researchers should take ac-
count of the uncertainty in parameter estimates. For example, the ctsem package
also provides lagged parameter plots with credible intervals to account for this
uncertainty.
Furthermore, researchers should be cautious about extrapolating beyond the
data. For instance, when we consider a vector field, we should be careful about

60
 3.5. Discussion

interpreting regions in which there is little or no observed data (cf. Boker &
McArdle, 1995). The same logic applies for the interpretation of IRFs for im-
pulses that do not match observed values. Moreover, we should also be aware that
interpreting lagged parameter plots for time-intervals much shorter than those
we observe data at is a form of extrapolation: It relies on strong model-based
assumptions, such as ruling out the possibility of a high-frequency higher-order
process (Voelkle & Oud, 2013; Voelkle et al., 2012).

3.5 Discussion
In this chapter we have set out to clarify the connection between DT- and CT-
VAR(1) models, and how we can interpret and represent the results from these
models. So far we have focused on single-subject, two-dimensional, first-order
systems with a stable node equilibrium. However, there are many ways in which
these models can be extended, to match more complicated data and/or dynamic
behavior. Below we consider three such extensions: a) systems with more than
two dimensions (i.e., variables); b) different types of fixed points resulting from
non-real eigenvalues of the drift matrix; and c) moving from single-subject to
multi-level datasets.

3.5.1 Beyond Two-Dimensional Systems

In the empirical illustration, we examined the interpretation of a drift matrix in
the context of a bivariate CT-VAR(1) model. Notably, the current trend in ap-
plications of DT-VAR(1) models in psychology has been to focus more and more
on the analysis of large systems of variables, as typified, for example, by the dy-
namic network approach of Bringmann et al. (2013, 2016). The complexity of
these models grows rapidly as the number of variables is added: To estimate a
full drift matrix for a system of three variables, we must estimate nine unique
parameters, in contrast to four drift matrix parameters for a bivariate system. In
addition, we must estimate a three-by-three covariance matrix for the residuals,
rather than a two-by-two matrix.
The relationship between the elements of A and Φ(∆t) becomes even less in-
tuitive once the interest is in a system of three variables, because the lagged pa-
rameter values are dependent on the drift matrix as a whole, as explained earlier.
This means that both the relative sizes, as well as the signs of the cross-lagged
effects may differ depending on the interval: The same lagged parameter may be
negative for some time-intervals and positive for others, and zero-elements of A
can result in corresponding non-zero elements of Φ (cf. Deboeck & Preacher,
2016; Aalen, Røysland, Gran, & Ledergerber, 2012; Aalen et al., 2016; Aalen,
Gran, Røysland, Stensrud, & Strohmaier, 2018). Therefore, although we saw in
our bivariate example that, for instance, negative CT cross-effects resulted in neg-
ative DT cross-lagged effects, this does not necessarily hold in the general case
(Kuiper & Ryan, 2018).
Additionally, substantive interpretation of the lagged parameters in systems

61
3. A Continuous-Time Approach to Intensive Longitudinal Data

with more than two variables also becomes less straightforward. For example,
Deboeck and Preacher (2016), Aalen et al. (2012, 2016) and Aalen et al. (2018) ar-
gue that the interpretation of Φ(∆t) parameters in mediation models (with three
variables and a triangular A matrix) as direct effects may be misleading: Deboeck
and Preacher argue that instead they should be interpreted as total effects. This
has major consequences for the practice of DT analyses and the interpretation of
its results.

3.5.2 Complex and Positive Eigenvalues

The empirical illustration is characterized by a system with negative, real, non-
equal eigenvalues, which implies that the fixed point in the system is a stable node.
In theory, however, there is no reason that psychological processes must adhere
to this type of dynamic behavior. We can apply the tools we have defined al-
ready to understand the types of behavior that might be described by other types
of drift matrices. Notably, some systems may have drift matrices √ with complex
eigenvalues, that is, eigenvalues of the form α ± ωi, where i = −1 is the imagi-
nary number, ω , 0, α is referred to as the real part, and ωi as the imaginary part
of the eigenvalue. If the real component of these eigenvalues is negative (α < 0),
then the system is still stable, and given a deviation it will return eventually to
a resting state at equilibrium. However, unlike the systems we have described
before, these types of systems spiral or oscillate around the equilibrium point,
before eventually coming to rest. Such systems have been described as stable spi-
rals, or alternatively as damped (linear or harmonic) oscillators (Voelkle & Oud,
2013; Boker, Montpetit, et al., 2010).
A vector field for a process which exhibits this type of stable spiral behavior
is shown in Figure 3.4, with accompanying trajectories. The drift matrix corre-
sponding to this vector field is
" #
−0.995 0.573
A= (3.11)
−2.000 −2.416

which is equivalent to our empirical example above, but with the value of a21
altered from 0.375 to −2.000. The eigenvalues of this matrix are λ1 = −1.706 +
0.800i and λ2 = −1.706 − 0.800i. In contrast to our empirical example, we can
see that the trajectories follow a spiral pattern; the trajectory which starts at
y1 (t) = 1, y2 (t) = 1 actually overshoots the equilibrium in the T i(t) dimension
before spiraling back once in the bottom quadrant. There are numerous ex-
amples of psychological systems that are modeled as damped linear oscillators
using second-order differential equations, which include the first- and second-
order derivatives (cf. Boker & Nesselroade, 2002; Bisconti et al., 2004; Boker,
Montpetit, et al., 2010; Horn, Strachan, & Turkheimer, 2015). However, as shown
here, such behavior may also result from a first-order model.
Stable nodes and spirals can be considered the two major types of stable fixed
points, as they occur whenever the real part of the eigenvalues of A are negative,
that is α < 0. Many other types of stable fixed points can be considered as spe-
cial cases: when we have real, negative eigenvalues that are exactly equal, the

62
 3.5. Discussion

b) c)

1.0
0.5
y2(t)

0.0

a)
−0.5
−1.0

d)

−1.0 −0.5 0.0 0.5 1.0

y1(t)

Figure 3.4: Vector field for a stable spiral corresponding to a drift matrix with negative real part
complex eigenvalues.

fixed point is called a stable star node (if the eigenvectors are distinct), or a stable
degenerate node (if the eigenvectors are not distinct). In contrast, if the real-part
of the eigenvalues of A are positive then the system is unstable, also referred to
as non-stationary or a unit-root in the time series literature (Hamilton, 1994).
This implies that, given a deviation, the system will not return to equilibrium; in
contrast to stable systems, in which trajectories are attracted to the fixed point,
the trajectories of unstable systems are repelled by the fixed point. As such we
can also encounter unstable nodes, spirals, star nodes and degenerate nodes. The
estimation and interpretation of unstable systems in psychology may be fruitful
ground for further research.
Two further types of fixed points may be of interest to researchers; in the
special case where the eigenvalues of A have an imaginary part and no real part
(α = 0), the fixed point is called a center. In a system with a center fixed point,
trajectories spiral around the fixed point without ever reaching it. Such systems
exhibit oscillating behavior, but without any damping of oscillations; certain bi-
ological systems, such as the circadian rhythm, can be modeled as a dynamic
system with a center fixed point. Such systems are on the borderline between
stable and unstable systems, sometimes referred to as neutrally stable; trajectories

63
3. A Continuous-Time Approach to Intensive Longitudinal Data

are neither attracted to or repelled by the fixed point. Finally, a saddle point oc-
curs when the eigenvalues of A are real but of opposite sign (one negative, one
positive). Saddle points have one stable and one unstable component; only trajec-
tories which start exactly on the stable axis return to equilibrium, and all others
do not. Together spirals, nodes and saddle points cover the majority of the space
of possible eigenvalues for A. Strogatz (2015) describes the different dynamic be-
havior generated by different combinations of eigenvalues of A in greater detail.

3.5.3 Multilevel Extensions

The time series literature (such as from the field of econometrics) as well as
the dynamic systems literature (such as from the field of physics) tends to be
concerned with a single dynamic system, either because there is only one case
(N = 1), or because all cases are exact replicates (e.g., molecules). In psychol-
ogy however, we typically have data from more than one person, and we also
know that people tend to be highly different. Hence, when we are interested in
modeling their longitudinal data, we should take their differences into account
somehow. The degree to which this can be done, depends on the number of
time points we have per person. In traditional panel data, we typically have
between two and six waves of data. In this case, we should allow for individ-
ual differences in means or intercepts, in order to separate the between-person,
stable differences from the within-person dynamic process, while assuming the
lagged relationships are the same across individuals (cf. Hamaker et al., 2015).
In contrast, experience sampling data and other forms of intensive longitu-
dinal data consist of many repeated measurement per person, such that we can
allow for individual differences in the lagged coefficients. This can be done by
either analyzing the data of each person separately, or by using a dynamic multi-
level model in which the individuals are allowed to have different parameters (cf.
Driver et al., 2017; Boker, Staples, & Hu, 2016). Many recent studies have shown
that there are substantial individual differences in the dynamics of psychological
phenomena, and that these differences can be meaningfully related to other per-
son characteristics, such as personality traits, gender, age, and depressive symp-
tomatology, but also to later health outcomes and psychological well-being (e.g.,
Kuppens et al., 2012, 2010; Bringmann et al., 2013).
While the current chapter has focused on elucidating the interpretation of
a single-subject CT-VAR(1) model, the substantive interpretations and visual-
ization tools we describe here can be applied in a straightforward manner to,
for example, the fixed effects estimated in a multilevel CT-VAR(1) model, or to
individual-specific parameters estimated in a multilevel framework. The latter
would however, lead to an overwhelming amount of visual information. The de-
velopment of new ways of summarizing the individual differences in dynamics,
based on the current tools, is a promising area for future research.

64
 3.5. Discussion

3.5.4 Conclusion
There is no doubt that the development of dynamical systems modeling in the
field of psychology has been hampered by the difficulty in obtaining suitable
data to model such systems. However this is a barrier that recent advances in
technology will shatter in the coming years. Along with this new source of psy-
chological data, new psychological theories are beginning to emerge, based on
the notion of psychological processes as dynamic systems. Although the statis-
tical models needed to investigate these theories may seem exotic or difficult to
interpret at first, they reflect the simple intuitive and empirical notions we have
about psychological processes: Human behavior, emotion and cognition fluctuate
continuously over time, and the models we use should reflect that. We hope that
our treatment of CT-VAR(1) models and their interpretation will help researchers
to overcome the knowledge-barrier to this approach, and can serve as a stepping
stone towards a broader adaptation of the CT dynamical system approach to psy-
chology.

65
3. A Continuous-Time Approach to Intensive Longitudinal Data

Appendix 3.A Matrix Exponential

Similarly to the scalar exponential, the matrix exponential can be defined as an
infinite sum
∞
X 1 k
eA = A
k!
k=0
The exponential of a matrix is not equivalent to taking the scalar exponential of
each element of the matrix, unless that matrix is diagonal. The exponential of a
matrix can be found using an eigenvalue decomposition

A = V DV −1

where V is a matrix of eigenvectors of A and D is a diagonal matrix of the eigen-

values of A (cf. Moler & Van Loan, 2003). The matrix exponential of A is given
by
eA = V eD V −1
where eD is the diagonal matrix whose entries are the scalar exponential of the
eigenvalues of A. When we want to solve for the matrix exponential of a matrix
multiplied by some constant ∆t we get

eA∆t = V eD∆t V −1 (3.12)

Take it that we have a 2 × 2 square matrix given by

" #
a b
A=
c d

and we wish to solve for eA∆t . The eigenvalues of A are given by

1 √
λ1 = (a + d − a2 + 4bc − 2ad + d 2 )
2
1 √
λ2 = (a + d + a2 + 4bc − 2ad + d 2 )
2
where we will from here on denote
√
R = a2 + 4bc − 2ad + d 2

for notational simplicity. The exponential of the diagonal matrix made up of

eigenvalues, multiplied by the constant ∆t is given by
 1 
D∆t  e 2 (a+d−R)∆t 0 
e =  1 (a+d+R)∆t 

0 e2


The matrix of eigenvectors of A is given by

" #
a−d−R a−d+R
V = 2c 2c
1 1

66
 3.B. Empirical Example Data Analysis

assuming c , 0, with inverse

−c a−d+R
" #
−1 R 2R
V = c −a+d+R .
R 2R

Multiplying V eD V −1 gives us
b(−eλ1 ∆t +eλ2 ∆t )
 
 R−a+d eλ1 ∆t + R+a−d eλ2 ∆t 
A∆t
e =  2R c(−eλ1 ∆t +eλ2R2 ∆t )
R  (3.13)
R+a−d λ1 ∆t R−a+d λ2 ∆t

R 2R e + 2R e

Note that we present here only a worked out example for a 2 × 2 square matrix.
For larger square matrices (representing models with more variables), the eigen-
value decomposition remains the same although the terms for the eigenvalues,
eigenvectors and determinants become much less feasible to present.

Appendix 3.B Empirical Example Data Analysis

## Data Preparation
# Load the ctsem package
library(ctsem)

# Data is available form https://osf.io/c6xt4/ #

setwd("./ESMdata")

#Load Data#
rawdata <- read.csv("ESMdata.csv",
header=TRUE,
stringsAsFactors = FALSE)

# Select only measurements which take place in

# the control and initial (no medication reduction) phase
rawdata <- subset(rawdata,rawdata$phase==1|rawdata$phase==2)

# Select only the variables of interest

sel <- c("mood_down","phy_tired")
data <- rawdata[,(names(rawdata) %in% sel)]

# Standardize the selected variables

for(j in 1:dim(data)[2]){
data[,j] <- (data[,j]-mean(data[,j]))/sd(data[,j])
}

# Create a time vector representing hours since first measurement

# Required by ctsem function ctIntervalise
t1 <- as.POSIXct(paste(rawdata$date, rawdata$resptime_s),
format="%d/%m/%y %H:%M:%S")

67
3. A Continuous-Time Approach to Intensive Longitudinal Data

time <- as.numeric(difftime(t1,t1[1], units="hours"))

# Attach this time variable to the selected items

data$time = time

# Create an ID variable
data$id = rep(1,dim(data)[1])

# Rename pat_agitate = Y1 -- event_import = Y2

colnames(data) = c("Y1", "Y2", "time", "id")

# Get data in wide format for ctsem

datawide <- ctLongToWide(datalong = data, id = "id",
time = "time",
manifestNames = c("Y1","Y2"))

# Create time-interval variable

datawide <- ctIntervalise(datawide = datawide,
Tpoints = dim(data)[1],
n.manifest = 2,
manifestNames = c("Y1","Y2"))

## Data analysis

# First specify the bivariate model, with 2 observed variables

model <- ctModel(n.manifest = 2, n.latent = 2,
Tpoints = 286,
LAMBDA = diag(nrow = 2),
MANIFESTMEANS = matrix(data = 0, nrow = 2, ncol = 1),
MANIFESTVAR = matrix(data = 0, nrow = 2, ncol = 2),
DRIFT = "auto",
CINT = matrix(data = 0, nrow = 2, ncol = 1),
DIFFUSION = "auto",
TRAITVAR = NULL,
MANIFESTTRAITVAR = NULL,
startValues = NULL)

# Fit the model to the data

fit <- ctFit(data = datawide,
ctmodelobj = model,
objective = "Kalman",
stationary = c("T0VAR", "T0MEANS"),
iterationSummary = T,
carefulFit = T, showInits = F,
asymptotes = F,

68
 3.B. Empirical Example Data Analysis

meanIntervals = F,
plotOptimization = F,
nofit = F, discreteTime = F,
verbose = 0)

summary(fit)

69
 Chapter 4

Time to Intervene: A
Continuous-Time Approach to
Network Analysis and
Centrality
Abstract

Dynamical network analysis of experience sampling data has become

increasingly popular in clinical psychology. In this approach, discrete-
time (DT) VAR model parameters are interpreted as direct relationships
between psychological processes, and centrality measures are used to iden-
tify which variable should be targeted for an intervention. There are two
methodological problems with this practice. First, VAR models suffer from
time-interval dependency, yielding different conclusions based on an often
arbitrary choice of sampling scheme. Second, the exact link between cen-
trality measures and intervention effects is unclear in this context and re-
lies on intuitive interpretations of model parameters that may not be valid.
In this paper we address both issues by proposing a continuous-time (CT)
approach to network analysis. We make use of CT-VAR models, which
allow one to explicitly model time-interval dependency. To aid interven-
tion targeting we develop new centrality measures based on CT networks.
We formulate these measures within the interventionist causal inference
framework, and show how these measures can be used to identify optimal
targets for either an acute or continuous intervention.

This chapter has been adapted from: Ryan, O. & Hamaker, E. L. (under review). Time to Inter-
vene: A Continuous-Time Approach to Network Analysis and Centrality. Author contributions: OR
and ELH conceptualized the initial project. OR wrote the paper and R code. ELH helped further
develop the ideas in the project, discussed progress and provided textual feedback.

71
4. A Continuous-Time Approach to Network Analysis and Centrality

4.1 Introduction
Dynamical network analysis, based on lagged regression models, has become a
popular approach for the analysis of experience sampling data in psychology
(Bringmann et al., 2013; Borsboom & Cramer, 2013). In clinical psychology in
particular, such analyses have been promoted as an aid in developing personal-
ized treatments for psychopathology. To facilitate this, centrality measures cal-
culated from parameter estimates are often used to identify which variable in
the network to target for an intervention (Bringmann et al., 2013; A. J. Fisher &
Boswell, 2016; Kroeze et al., 2017; Epskamp, van Borkulo, et al., 2018; Rubel,
Fisher, Husen, & Lutz, 2018; Bak, Drukker, Hasmi, & van Os, 2016; Bringmann
et al., 2015; Bastiaansen et al., 2019; A. J. Fisher, Reeves, Lawyer, Medaglia, &
Rubel, 2017; Christian et al., 2019).
Two developments in the methodological literature highlight problems with
this practice. First, the estimation of dynamical network structures is based on
path estimates from a discrete-time (DT) first-order Vector Auto-regressive (VAR)
model. The DT-VAR model suffers from the problem of time-interval dependency,
meaning that the parameters of this model can potentially lead to dramatically
different conclusions based on how the observations are spaced in time (Gollob &
Reichardt, 1987; Kuiper & Ryan, 2018). This is an issue which numerous authors
have argued could be resolved by modeling psychological processes as unfold-
ing continuously over time using continuous-time (CT) models (Boker, 2002; van
Montfort, Oud, & Voelkle, 2018; Ryan, Kuiper, & Hamaker, 2018). Second, the
use of out-of-the-box centrality measures for the identification of intervention
targets has been critiqued from various sides, both in terms of their applicability
to psychological networks (Bringmann et al., 2019) and in their actual ability to
detect optimal intervention targets (Borgatti, 2005; van Elteren & Quax, 2019;
Dablander & Hinne, 2019).
Although the consequences of taking a CT approach have been discussed in
general elsewhere (cf. Boker, 2002; Voelkle et al., 2012; Aalen et al., 2016, 2012;
Deboeck & Preacher, 2016), the specific implications for dynamical network
analysis have not yet been investigated. Taking a CT approach yields practical
benefits, such as an ability to deal with unequally spaced measurements, but also
leads to a new outlook on the meaning and interpretation of lagged regression
parameters. As a consequence, taking a CT perspective not only undermines the
typical interpretation of DT-VAR networks as well as their accompanying central-
ity measures, but it also creates opportunities for the development of new ways
to understand the underlying dynamic process and how best to intervene on it.
In this paper we introduce a CT approach to dynamical network analysis,
and develop new centrality measures which can be used to gain more insight
into what variables should be targeted for interventions, and what kinds of in-
terventions can be expected to lead to what kinds of outcome. In so doing we
tie together diverse strands of the psychological and methodological literature.
Specifically, we show how centrality measures are related to path-specific effects
from the SEM literature, which allows us to connect previous research on path-
specific effects in CT models (Aalen et al., 2012; Deboeck & Preacher, 2016) and

72
 4.2. Current Practice: DT-VAR Networks

intervention-based approaches to causal modeling (Eichler & Didelez, 2010; Van-

derWeele, 2015; Pearl, 2009).
This paper is organized as follows. In the first part of the paper, we provide an
overview of the DT-VAR model, how path-specific effects and centrality measures
are used to identify intervention targets based on this model in practice, and the
problems that arise therein. Second, we present the CT approach to dynamic
network analysis based on the CT-VAR model, introducing a new network repre-
sentation into the psychological canon from the causal inference literature, and
exploring the new insights this approach yields. Third, we introduce new fit-for-
purpose centrality measures which both reflect the CT nature of the underlying
process and have a clear and direct link to interventions and the choice of optimal
intervention targets. Throughout, we use a hypothetical example for illustrative
purposes. Moreover, for simplicity, the developments on this paper focus on sin-
gle subject models, though the critiques and measures developed here generalize
in a straightforward way to within-subjects parameters of multi-level models.

4.2 Current Practice: DT-VAR Networks

In this section we discuss the first-order discrete-time Vector Auto-Regressive
(DT-VAR) model and its representation as a dynamical network. Subsequently,
we describe how path-specific effects and centrality measures based on this
model are used by applied researchers to identify intervention targets, and estab-
lish how these two practices are connected. We end by discussing two problems
with current practice, namely the lack of clarity regarding how centrality mea-
sures relate to the effects of interventions (i.e., the intervention problem) and the
time-interval dependency of parameter estimates (i.e., the time-interval problem).

4.2.1 The DT-VAR model

The DT-VAR model is a single-subject time-series model that describes dynamic
relationships between variables measured repeatedly over time. Lagged regres-
sion parameters encode the effect of a variable on itself (an auto-regressive effect)
or another variable (a cross-lagged effect) at the next measurement occasion (i.e.,
at a lag of one). This model can be written as

Yτ = c + ΦYτ−1 + τ (4.1)

where, given p variables, the p × 1 vector of random variables Y at occasion τ

is regressed on the p × 1 vector of those same variables at the previous occasion,
Yτ−1 . The p × p matrix of lagged regression parameters is denoted Φ, while the
p × 1 vectors c and τ denote the intercepts and random shocks respectively, the
latter being normally distributed with mean zero and variance-covariance ma-
trix Ψ (Hamilton, 1994). Two crucial assumptions of the DT-VAR model are,
first, that the same amount of time (denoted ∆t) elapses between two subsequent
measurement occasions, and second, that the underlying process is stationary,

73
4. A Continuous-Time Approach to Network Analysis and Centrality

í í
... Y1,ɒ Y1,ɒ൅ͳ Y1,ɒ൅ʹ
... 
Y Y
í

... Y2,ɒ Y2,ɒ൅ͳ Y2,ɒ൅ʹ
... í
 í í 

Y3,ɒ Y3,ɒ൅ͳ Y3,ɒ൅ʹ í

... ... í

í
Y4 Y

... Y4,ɒ Y4,ɒ൅ͳ Y4,ɒ൅ʹ ...
í 

(a) DT-VAR path-model and dynamical network representing Φ(1)

í 

... Y1(t0) Y1(1ï»๜) Y1(t1) Y1(2ï»๜) Y1(t2)

... í
Y Y


í
... Y2(t0) Y2(1ï»๜) Y2(t1) Y2(2ï»๜) Y2(t2)
... 
í   

... Y3(t0) Y3(1ï»๜) Y3(t1) Y3(2ï»๜) Y3(t2) 
...
í
Y Y

... Y4(t0) Y4(1ï»๜) Y4(t1) Y4(2ï»๜) Y4(t2)
...
í 

(b) DT-VAR model with latent waves and dynamical network representing Φ( 12 )

Figure 4.1: Path-model (left-hand side) and network (right-hand side) representations of two four-
variable DT-VAR models. In the path models, the presence of an arrow linking two variables denotes
some non-zero dependency between them, conditional on all variables at the previous wave. For the
networks solid arrows denote positive parameters and dashed arrows represent negative parameters.

which entails that the means, variance and covariances, and lagged regression
parameters remain the same over time.
The DT-VAR model describes a relatively simple dynamic system. Each vari-
able has some baseline level or equilibrium position defined by its mean value.
Typically, the variables themselves are centered around this mean value (e.g. As-
parouhov, Hamaker, & Muthén, 2018), and so the intercept term is often omitted
for notational simplicity (c = 0), a convention we will adopt throughout the re-
mainder of the paper. The random shocks τ push the system away from equi-
librium, and the lagged parameters Φ determine how the variables react to these
shocks, eventually returning to equilibrium over time.
The DT-VAR model can be represented as either a path-model, as shown in the
left-hand panel of Figure 4.1(a), or as a dynamical network structure, as shown
in the right-hand panel, where the nodes represent the random variables, and the
edges represent the values of the lagged parameters Φ (Bringmann et al., 2013;
Epskamp, van Borkulo, et al., 2018). The DT-VAR model, and its multi-level
extension, are popular for the analysis of experience sampling data in clinical
psychology (Bringmann et al., 2013; Pe et al., 2015; A. J. Fisher & Boswell, 2016;

74
 4.2. Current Practice: DT-VAR Networks

Kroeze et al., 2017; Rubel et al., 2018; Bak et al., 2016).1 In dynamical network
approaches, each variable is typically used to represent some psychological con-
struct (such as an emotion or symptom of disorder) that varies over time, which
we refer to as a process. The lagged parameters in Φ are typically interpreted as
direct effects of these processes on each other over time.
For example, take it that the four variables in Figure 4.1(a) represent (re-
peated measurements of) Stress (Y1 ), Anxiety (Y2 ), feeling Self-Conscious (Y3 ) and
feelings of Physical Discomfort (Y4 ). We will refer to this throughout as the Stress-
Discomfort system. We can see from the parameter values in the dynamical net-
work that all variables share reciprocal cross-lagged relationships with all other
variables: The network is fully connected, with a mix of positive and negative
cross-lagged and auto-regressive parameters of different sizes. Typically, a cross-
lagged parameter such as φ41 = 0.05 would be interpreted as the direct effect of
current Stress (Y1,τ ) on Physical Discomfort at the next measurement occasion
(Y4,τ+1 ), conditional on (i.e., controlling for) current feelings of Anxiety, Self-
Conscious, and Physical Discomfort (Y2,τ , Y3,τ , Y4,τ ). This parameter is weakly
positive, leading to the interpretation that a high level of current Stress has a
small positive direct effect on feelings of Physical Discomfort at the next occa-
sion.

4.2.2 Intervention Targets from DT-VAR models

To identify which variables should be considered targets for an intervention
based on a DT-VAR model, psychology researchers have mainly used two ap-
proaches: a) path-specific effects, which are inspired by the SEM literature (Bollen,
1987); and b) centrality measures, which come from the broader network analysis
literature (Freeman, 1978; Opsahl, Agneessens, & Skvoretz, 2010).

4.2.2.1 Path-Specific Effects

Path-specific effects have been used to describe the total, direct and indirect effects
of one variable on another, and are calculated using the well-known path-tracing
rules from the SEM literature (Bollen, 1987). The total effect of Stress levels now
(Y1,τ ) on Physical Discomfort two measurement occasions later (Y4,τ+2 ) is the sum
of direct effect pathways (Y1,τ → Y4,τ+1 → Y4,τ+2 and Y1,τ → Y1,τ+1 → Y4,τ+2 )
and indirect effect pathways through the mediating variables Anxiety (Y1,τ →
Y2,τ+1 → Y4,τ+2 ) and Self-Conscious (Y1,τ → Y3,τ+1 → Y4,τ+2 ) (Cole & Maxwell,
2003). Note here that to define a path-tracing effect between variables Y1 and Y4
we must also specify the number of measurement occasions that elapse between
the relevant instances of those variables, that is, the lag of interest.
1 Note that in many applications, the multi-level extension of the VAR(1) model is used, for the pur-
poses of regularizing the parameter estimates and modeling between-person differences in parame-
ters. The primary interest of researchers using these multi-level models is in the individual within-
person parameter estimates (Bringmann et al., 2013; Schuurman et al., 2016; Asparouhov et al., 2018;
Suls, Green, & Hillis, 1998; Lodewyckx, Tuerlinckx, Kuppens, Allen, & Sheeber, 2011; Liu, Kuppens,
& Bringmann, 2019). For the sake of notational simplicity —and without loss of generality— we focus
on the single-subject version of the VAR(1) model throughout the paper.

75
4. A Continuous-Time Approach to Network Analysis and Centrality

If we accept the interpretation of Φ parameters as direct causal effects, we

may suggest that interventions should target variables that have strong direct
or total effects on others, and/or those mediators through which the strongest
indirect effects pass (for examples in a general SEM context, see Bernat, August,
Hektner, & Bloomquist, 2007; Bramsen et al., 2013). For instance, based on the
parameters in Figure 4.1(a), we might suggest Anxiety as an intervention target
due to the relatively strong lag-one direct effect on Physical Discomfort (φ42 = .08),
or because it is a mediator of the largest lag-two indirect effect, from Stress to
Physical Discomfort (Y1,τ → Y2,τ+1 → Y4,τ+2 = −.005).

4.2.2.2 Centrality Measures

Researchers adopting a network approach have used centrality measures to deter-
mine which variable should be considered the optimal intervention target in a
network (e.g., Bringmann et al., 2013; A. J. Fisher & Boswell, 2016; Kroeze et al.,
2017; Epskamp, van Borkulo, et al., 2018; Rubel et al., 2018; Bak et al., 2016;
Bringmann et al., 2015; Bastiaansen et al., 2019). Centrality measures are used
to summarize the relationships a particular variable has with the network as a
whole, typically summing over the individual relationships that variable shares
with all others in the network. To our knowledge, the precise connection between
path-specific effects and centrality measures for DT-VAR models has not yet been
described in the literature. However, a close inspection of the computation and
interpretation of many popular centrality measures reveals that they are very
similar to path-tracing effects: Specifically, many centrality measures are inter-
preted as capturing either total, direct and indirect effects, and in turn these mea-
sures are often closely related to summaries of the corresponding path-specific
effects. Here, we will focus on three such measures (with the exact connection
between these measures and path-tracing effects given in Appendix 4.A.)
(2)
The Two-Step Expected Influence measure (EIi ; Robinaugh et al., 2016; Kaiser
& Laireiter, 2018) is typically interpreted as a summary of total effects: In path-
tracing terms it is the sum of lag one direct effects and lag two total effects (see
Appendix 4.A for details). As such, variables with a high Two-Step Expected In-
fluence could be expected to exert a high overall influence on the system, making
it an attractive intervention target. For the Stress-Discomfort system, this mea-
(2)
sure would suggest Anxiety as an intervention target, EIi = 0.035.
(1)
The One-Step Expected Influence (EIi ;
Robinaugh et al., 2016; Kaiser & Laire-
iter, 2018) and Out-Strength centrality (Opsahl et al., 2010) measures are inter-
preted as summarizing direct effects. They are both sums of lag-one direct effects,
with the latter taking the absolute value. Due to their close connection we will
focus exclusively on the One-Step Expected Influence measure in this paper. For
the Stress-Discomfort System, Anxiety also has the highest One-Step Expected
(1)
Influence (EI2 = φ12 + φ32 + φ42 = 0.038).
Finally, Betweenness is interpreted as indicating the degree to which a variable
“funnels information flow”, similar to how mediating variables funnel indirect ef-
fects (e.g. Bringmann et al., 2013; Opsahl et al., 2010; Freeman, 1977). This mea-

76
 4.2. Current Practice: DT-VAR Networks

sure is conceptually similar to determining which variables are strong mediators,

although paths are calculated by summing, rather than multiplying parameters,
as in path-tracing rules. The Betweenness Centrality of Stress, Anxiety and Physi-
cal Discomfort are all equal (BC1 = BC2 = BC4 = 1), with Self-Conscious attaining
a lower score (BC3 = 0), indicating that the former three are equally attractive in-
tervention targets in terms of the role they play in indirect effects in the network.

4.2.3 Two Problems with Current Practice

From the above we can conclude that there is a clear logical thread that can be fol-
lowed from the interpretation of DT-VAR parameters as direct effects through to
the use of centrality measures for the identification of intervention targets. How-
ever, there are two major problems associated with the approach outlined above,
both of which relate to the intuitive interpretation of lagged model parameters as
direct effects. We detail these problems below, and elaborate on how to overcome
these problems in the remainder of the paper.

4.2.3.1 The Intervention Problem

The first problem with current practice is that it is not clear exactly how or why
particular parameters or centrality measures should be informative about the
effects of interventions. The idea that a cross-lagged parameter φji represents a
direct effect can be defended from two perspectives. First, from a purely path-
tracing point of view, if we assume that the path model in Figure 4.1(a) is the
true model, then φji represents the path that we can trace from Yi,τ to Yj,τ+1
without passing through any values of other variables in the model. Second, we
can use the notion of Granger causality, which states that, if we assume that all
relevant variables have been measured (i.e. no unobserved confounding) then
the parameter φji represents a direct causal effect of Yi,τ on Yj,τ+1 (Granger, 1969).
This is a notion of causality which is often invoked in empirical applications of
the DT-VAR model (cf. Cole & Maxwell, 2003; Hamaker et al., 2015; Bulteel et
al., 2016).
However, what is absent from both conceptualizations of lagged parameters
as direct effects is an explicit link between a particular parameter and the effect
of a particular intervention. Without such a link, it is impossible to give mean-
ingful answers to some of the most basic questions which arise regarding how
path-specific effects and centrality measures should be applied in practice, such
as: Should we prefer to intervene on a variable with high direct, total, or indi-
rect relationships? What kind of intervention should we apply, and what kind of
effect would we expect to see? And under what conditions do these measures in-
form us of this? Establishing an unambiguous link between interventions on the
one hand, and parameters and centrality measures on the other, is a fundamen-
tal prerequisite if we are to use these measures to inform intervention targets
in a principled way. Such a link is necessary as, despite their intuitive appeal,
any given centrality measure cannot be simply assumed to be informative about
intervention targets. This issue has been highlighted by, amongst others, van El-

77
4. A Continuous-Time Approach to Network Analysis and Centrality

teren and Quax (2019) and Dablander and Hinne (2019), who show that central
nodes are often not the optimal target for an intervention, as well as Borgatti
(2005) and more recently Bringmann et al. (2019), who detail at length how the
utility of common centrality measures is highly dependent on the particular type
of network under consideration.
To solve this problem, we will make use of the interventionist causal inference
framework (Hernan & Robins, 2019; Pearl, 2009; Rubin, 1974). In this frame-
work, a causal relationship between X and Y is defined as the change we would
expect to see in Y if we were to intervene and change the value of X. The key
appeals of this approach are first, that it places an emphasis on an intervention-
effect as the target of inference, and second that it allows one to derive the con-
ditions under which the effects of such an intervention can be identified from
relationships between variables in observational data. Crucially, intervention-
based definitions have also been extended to path-specific effects: VanderWeele
and Tchetgen Tchetgen (2017) have shown that—under certain circumstances—
path-tracing total, direct and indirect effects in a lagged regression model (similar
to the DT-VAR model discussed above) can be interpreted as the expected change
in Y , given interventions to change the values of the cause variable X and the
mediating variable(s) M. Although, to our knowledge, intervention-based defi-
nitions have not been extended to centrality measures, the connection between
centrality and path-specific effects established above opens up a potential avenue
by which such a definition could be established. In Section 4.4.1 we will make
use of these connections to establish new centrality measures which are inspired
by this framework, ensuring an unambiguous connection between the value of
those measures and the choice of intervention target.
Of course, the identification of these intervention effects in practice will still
rely on a number of idealized assumptions. We must assume that there are no
unmeasured confounder variables, that we can intervene in the system without
changing how the variables in the system relate to one another (an assumption
known as modularity; Pearl, 2009), and that the first-order time series model we
fit to the data correctly characterizes the process. Assumptions regarding con-
founders and modularity are fundamental to the identification of intervention-
effects from data in any context (e.g. from experiments, cross-sectional data or
time-series). However, even if we are willing to make these strong and optimistic
assumptions, there is still a fundamental issue specific to lagged regression mod-
els which we must account for in order to come to sensible conclusions about
intervention targets. That fundamental issue is the time-interval problem.

4.2.3.2 The Time-Interval Problem

As pointed out by Gollob and Reichardt (1987), and since then by many others
(Oud & Jansen, 2000; Reichardt, 2011; Voelkle et al., 2012; Deboeck & Preacher,
2016; Kuiper & Ryan, 2018), the effect of one variable on another depends crit-
ically on the time-interval between them. This is referred to as the time-interval

78
 4.2. Current Practice: DT-VAR Networks

problem.2 This problem arises when the processes under investigation vary at
a higher frequency than we observe them, which means that the processes take
on values at points in time in-between those particular occasions at which we
measure the process. In path-modeling terms, this means that there are unmea-
sured and un-modeled values of the observed variables Y in-between measure-
ment occasions, depicted as latent variables in Figure 4.1(b). The time-interval
problem has two consequences. First, the value of the regression parameters Φ
are a function of the time-interval between observations (denoted Φ(∆t)), and
so, except under restrictive circumstances, may change their sign, size and rela-
tive ordering if a different time-interval is used (Kuiper & Ryan, 2018). Second,
the lagged regression parameters at any interval should be interpreted as total,
rather than direct effects from a path-tracing perspective. The reason for this is
that cross-lagged parameters constitute both direct pathways and indirect path-
ways through latent values of the other processes in the model (Cole & Maxwell,
2003; Reichardt, 2011; Deboeck & Preacher, 2016).
To illustrate both issues, let us take it that the parameters that we introduced
in Figure 4.1(a) represent the lagged relationships at one-hour intervals; we de-
note these parameters as Φ(∆t = 1). In theory, we could also have obtained ob-
servations of the Stress-Discomfort system at half-hour intervals: The path model
representing this shorter-interval system is depicted in the left panel of Figure
4.1(b), where the potential half-hour measurements are depicted as latent vari-
ables Y (t = 1 12 ) and Y (t = 2 21 ). The effects matrix relating the half hour realiza-
tions of the process is denoted Φ(∆t = 12 ). A well-known result from the time-
series literature allows us to relate the parameters of these two models through
the matrix exponent (Hamilton, 1994). In this case, the matrix of lagged param-
eters of the half-hour system is related to the matrix of lagged parameters at the
longer one-hour time-interval by the expression

Φ( 12 )2 = Φ(1) (4.2)

that is, by squaring the matrix of parameters at the shorter interval, we obtain
the parameters at twice that interval. It is important to note here that squaring
a matrix is not equivalent to squaring the parameters of that matrix: Instead,
any given parameter in Φ(1) is a function of multiple parameters in Φ( 12 ). For
instance, the cross-lagged parameter which regresses Y4,τ+1 on Y1,τ can be re-
written in terms of the shorter-interval parameters as φ41 (1) = φ11 ( 12 )φ41 ( 12 ) +
φ21 ( 12 )φ42 ( 12 )+φ31 ( 12 )φ43 ( 12 )+φ41 ( 12 )φ44 ( 12 ). Clearly, this has major consequences
for the interpretation of lagged regression parameters.
First, we can see that the types of conclusions we would intuitively make
based on the parameter values in the half-hour network are entirely different
than those we made based on the parameters in the one-hour network. To be-
2 This is also sometimes referred to as the lag problem (e.g., Hamaker & Wichers, 2017). The lag of
a model also describes a notion of time-spacing, but more generally refers to the order of a lagged re-
gression model: Lag-one models relate current values to the exactly preceding measurement occasion,
and lag-two models relate current values to the variables two measurement occasions previously, and
so forth. Thus, the lag of a model and is distinct from the time-interval between observations, and so
for clarity we refer to this as the time-interval problem throughout.

79
4. A Continuous-Time Approach to Network Analysis and Centrality

gin with, the magnitude of each individual parameter, as well as their signs and
relative ordering, are different in the two networks. For example, in the one-
hour network, Stress and Anxiety both have positive lagged relationships with
Physical Discomfort (φ41 (1) = 0.047 and φ42 (1) = 0.077), with the effect of Anx-
iety being slightly larger. In the half-hour network, the corresponding lagged
relationships are both strongly negative, with the effect of Stress now the larger
(φ41 ( 12 ) = −0.275 and φ42 ( 12 ) = −0.151). If we calculate centrality measures for the
half-hour network, we may conclude that either Physical Discomfort is a promis-
ing intervention target, as it has the highest scores on the total and direct influ-
(2) (1)
ence measures (EI4 = 0.555 and EI4 = 0.557), or that Anxiety is a promising
target as it has the highest Betweenness Centrality (BC2 = 3; See Appendix 4.B
for a full table of centrality values).
Second, on a more fundamental level, the presence of the time-interval prob-
lem alters how we should interpret a set of lagged regression parameters. Specif-
ically, based on the relationship in Equation (4.2), the lagged parameters of the
one-hour network Φ(1) should be interpreted as total rather than direct effects
(Deboeck & Preacher, 2016; Aalen et al., 2016). Indeed, taking the power of a
matrix of direct effects, as in Equation (4.2), is suggested by Bollen (1987) exactly
as a method of calculating total effects in the SEM literature. This interpretation
also helps clarify how we come to seemingly contradictory conclusions based
on the time-interval used. For example, the parameter φ42 (1) = 0.077 would
typically be interpreted as a direct effect of Anxiety now (Y2,τ ) on Physical Dis-
comfort an hour from now (Y4,τ+1 ). However, when we examine how these vari-
ables are related to one another in Figure 4.1(b), we can see that the relationship
between them is made up of a number of different pathways, including direct
paths (Y2,τ → Y2 (1 12 ) → Y4,τ+1 and Y2,τ → Y4 (1 12 ) → Y4,τ+1 ) as well as paths that
pass through latent values of Stress (Y2,τ → Y1 (1 12 ) → Y4,τ+1 ) and Self-Conscious
(Y2,τ → Y3 (1 12 ) → Y4,τ+1 ). In this instance, a combination of strong positive indi-
rect paths and slightly weaker negative direct paths add up to the weak positive
total effect in the one-hour network: A high feeling of Anxiety has the effect of
lowering Physical Discomfort a half hour later, but also sets off a chain reaction of
Stress and Self-Conscious feelings that ultimately leads to an increase in Physical
Discomfort an hour later.
In summary, the time-interval problem is present whenever there are latent,
unobserved values of our processes of interest in-between measurement occa-
sions. The consequences of the time-interval problem for the identification of in-
tervention targets, using current practice, is twofold. First, any parameter, path-
specific effect, or centrality measure may lead to conclusions which are specific
to the time-interval used in data collection: The effect of X on Y should in fact
be interpreted as the effect at one specific time-interval (Gollob & Reichardt, 1987).
Second, the intuitive interpretation given to path-specific effects and centrality
measures based on a path-tracing logic are misleading: Direct effects are better
conceptualized as total effects (Deboeck & Preacher, 2016; Reichardt, 2011). This
leaves the usual path-tracing based notion of “direct”, “indirect” and “total” ef-
fects, and therefore also of centrality measures which summarize these effects,

80
 4.3. CT Network Analysis: Accounting for Continuity

open to question. However, our presentation here also highlighted one poten-
tial solution to the time-interval problem: Decomposing lagged relationships be-
tween observations into truly direct effects operating over a shorter time-interval.
This decomposition opens up a new perspective on how lagged relationships
should be interpreted, a perspective which we can use to explore time-interval
dependency, and avoid coming to misleading or contradictory choices regarding
intervention targets.

4.2.4 Conclusion
We have now identified two problems with current DT-VAR network approaches,
in particular with the use of these models and associated centrality measures for
the identification of intervention targets. The first is that the link from inter-
ventions to model parameters and centrality measures is unclear. The second
problem is that the parameters and centrality measures themselves suffer from
time-interval dependency. In the next section we will present one potential solu-
tion to the latter problem: the use of continuous-time models. We will then use
this CT approach in the fourth section to address the other problem, describing
how CT models can in principle be used to identify intervention targets.

4.3 CT Network Analysis: Accounting for Continuity

In the example given in the previous section we focused on the problems that
arise when there is a single un-modeled latent wave of measurements when
studying processes that vary over time. In psychological research, there are likely
to be many more of these potentially observable but unmeasured process values
between measurement occasions. While we are limited to observing psycholog-
ical processes such as momentary stress and anxiety at discrete measurement
occasions, they are likely to vary, evolve and exert influence on one another con-
tinuously over time: One’s feelings of anxiety influences one’s feelings of stress a
second, thirty seconds, a minute from now and so forth, not only stress levels
an hour from now. This means that there will be infinitely many latent values of
these variables in-between measurement occasions, and moreover, that there are
different lagged relationships between variables over a range of different time-
intervals.
This perspective is consistent with viewing psychological phenomena as
continuous-time processes, a perspective described in detail by Boker (2002) and
promoted by proponents of CT statistical models in psychology (e.g. Coleman,
1968; Oud & Jansen, 2000; Oravecz et al., 2011; Deboeck & Preacher, 2016; Ryan
et al., 2018; van Montfort et al., 2018; Ou, Hunter, & Chow, 2019; Driver et al.,
2017; Voelkle et al., 2012). In SEM terms, we can represent a CT process as a
path model in which there are infinitely many latent unobserved variable val-
ues in-between any two measurement occasions, spaced an infinitesimally small
time-interval apart, as depicted in the left-hand panel of Figure 4.2 (see also De-
boeck & Preacher, 2016).

81
4. A Continuous-Time Approach to Network Analysis and Centrality

Modeling CT processes is based on breaking down the relationship between

observed measurement waves into their fundamental building blocks: direct
lagged relationships operating over an infinitesimally small, hereby referred to
as moment-to-moment, time-interval. These continuous moment-to-moment dy-
namics are captured by differential equation models, and the parameters of these
models can be obtained from ESM-type data observed at longer intervals by fit-
ting their integral form (Strogatz, 2015; Voelkle et al., 2012). Fitting this type
of CT model allows us to overcome the time-interval problem by explicitly ac-
counting for how lagged effects depend on the time-interval: The same moment-
to-moment dynamics produce different lagged relationships depending on how
far apart in time those variables are spaced.
In this section we will describe how CT models can be applied in the context
of dynamical network analysis using a type of network representation known as
a local dependence graph. First, we describe a simple differential equation and how
its parameters can be interpreted. Second, we describe the CT-VAR model, which
is the integral form of that simple differential equation and is the CT equivalent
of the DT-VAR model. Third, we describe how using the CT-VAR model for net-
work analysis affects our perspective on path-specific effects (based on previous
work by Deboeck & Preacher, 2016), and what the consequences are for existing
centrality measures.

4.3.1 Differential Equations: Moment-to-Moment Dynamics

The continuous-time equivalent of the VAR model is the first-order stochastic
differential equation (SDE), which can be written as
dY (t)
= AY (t) + W (t) (4.3)
dt
dY (t)
where dt on the right is the first derivative or the rate of change of the variables
Y at time t (denoted Y (t)). We can think of this derivative as being equivalent
to a (scaled) change score Y (t + s) − Y (t) for a very short time-interval (lim s → 0).
This derivative is dependent on the current value of the mean-centered variables
Y (t), and the p × p matrix of regression parameters which relates these two is
called the drift matrix A. The W (t) term represents a Wiener process, special kind
of mean-zero white noise residual term (described in greater detail by, among
others, Oud & Jansen, 2000; Voelkle et al., 2012; Voelkle & Oud, 2013). The type
of dynamic system described by this differential equation is the same as that of
the DT-VAR model: Each variable Y (t) has an equilibrium value, here defined
by its mean; the Wiener process pushes the system away from this equilibrium;
and the drift parameters A determine how the variables react to these shocks,
eventually returning the system back to equilibrium over time (Strogatz, 2015;
Ryan et al., 2018).
The drift parameters for the Stress-Discomfort system are plotted in network
form on the right-hand side of Figure 4.2. Here, we use the hexagonal nodes to
denote that the weights of the network are drift matrix parameters rather than
lagged regression parameters, relating the value of one variable to the rate of

82
 4.3. CT Network Analysis: Accounting for Continuity

í í
Y1(t0) Y1(t)Y1(t)Y1(s) Y1(sn) Y1(t1) Y1(t)Y1(t)Y1(s) Y1(sn) Y1(t2)
... ... ... ... 1.25
Y1 Y2
5.5

Y2(t0) Y2(t)Y2(t)Y2(s) Y2(sn) Y2(t1) Y2(t)Y2(t)Y2(s) Y2(sn) Y2(t2)

... ... ...
...
5 í 5.9

Y3(t0) Y3(t)Y3(t)Y3(s) Y3(sn) Y3(t1) Y3(t)Y3(t)Y3(s) Y3(sn) Y3(t2)

... ... ...
...

Y4 2.5 Y
Y4(t0) Y4(t)Y4(t)Y4(s) Y4(sn) Y4(t1) Y4(t)Y4(t)Y4(s) Y4(sn) Y4(t2)
... ... ... ...
í
í

Figure 4.2: CT-VAR path-model (left-hand side) and CT network in the form of a weighted local
independence graph representing A (right-hand side). In the path model, the latent variables and
ellipses represent an infinite number of latent unobserved process values in-between measurement
occasions, spaced an infinitesimally small time-interval apart; the presence of an arrow linking two
variables in the path model denotes some non-zero dependency between them, conditional on all
variables at the previous “wave”, that is, a local dependency. For the network, solid arrows denote
positive parameters and dashed arrows represent negative parameters.

change of the other, instead of the value of the other at the next occasion. This
type of network representation is known as a local dependence graph (Schweder,
1970; Aalen, Borgan, Keiding, & Thormann, 1980; Didelez, 2008), of which the
right-hand side of Figure 4.2 is a weighted variant. We will refer to this as a CT
network, where the edges depict direct moment-to-moment dependencies between
the time-varying processes.
The interpretation of drift matrix parameters is similar to that of a change-
score model from the time-series literature. The diagonal parameters are known
as auto-effects and are typically negative: If Stress takes on a high positive value,
its negative auto-effect (a11 = −6) ensures it will move back towards equilibrium
in response. The higher the absolute value of the auto-effect, the quicker the pro-
cess will move back towards equilibrium (ignoring all other parameters), similar
to an auto-regressive parameter close to zero. As such, the diagonal parameters
are also referred to as reflecting the “centralizing tendency” of a variable, for in-
stance by Oravecz et al. (2011). Off-diagonal elements known as cross-effects have
a similar interpretation to cross-lagged parameters: The negative cross-effect of
Anxiety on the rate of change of Physical Discomfort (a42 = −7.3) means that if
Anxiety takes on a positive value, this will result in a decrease in the value of
Physical Discomfort. The higher the absolute value of the parameter, the greater
the magnitude of the effect (for more details on the interpretation of these pa-
rameters, see also Ryan et al., 2018; Voelkle et al., 2012; Oravecz et al., 2011).
From the CT network, we can see that all of the Stress-Discomfort variables
have a negative auto-effect on themselves (a11 = a33 = a44 = −6); the auto-effect of
Anxiety is closer to zero, which roughly corresponds to a stronger auto-regressive
effect than the other variables (a22 = −2.5). Notably there are much fewer direct
dependencies (i.e. fewer connections in the network) between processes on a
moment-to-moment basis than we saw in the lagged DT-VAR networks in Fig-
ures 4.1(a) and (b): For instance, Stress has no direct moment-to-moment ef-

83
4. A Continuous-Time Approach to Network Analysis and Centrality

fect on Physical Discomfort (a41 = 0). The dependencies which are present are
mostly positive: Stress and Anxiety exert reciprocal positive effects on one an-
other (a12 = 1.25 and a21 = 5.5); Anxiety has a direct positive effect on the rate of
change of Self-Conscious; Anxiety has a strong direct negative effect (a42 = −7.3),
and Self-Conscious a positive direct effect (a43 = 2.5) on the rate of change of
Physical Discomfort; finally, Physical Discomfort has a positive direct effect on
the rate of change of Stress (a14 = 5). We will see why there is a discrepancy be-
tween the moment-to-moment CT network and the lagged relationships in the
DT-VAR network by connecting these two models through the integral form of
the differential equation.

4.3.2 The CT-VAR model

We can estimate the parameters of the differential equation model from ESM data
by making use of its integral form known as the CT-VAR or Ornstein-Uhlenbeck
process (Oud & Jansen, 2000; Oravecz, Tuerlinckx, & Vandekerckhove, 2009;
Voelkle et al., 2012; Driver et al., 2017).3 This allows us to express the rela-
tionship between measurement waves observed with any spacing ∆t in terms of
the moment-to-moment relationships of the first-order SDE:

Y (tτ ) = c + eA∆tτ Y (tτ−1 ) + (∆tτ ) (4.4)

where variables at the current measurement occasion Y (tτ ) are regressed on vari-
ables at the previous measurement occasion Y (tτ−1 ). Note that τ refers to the
measurement occasion, whereas t refers to the actual time when this measure-
ment took place. Hence, ∆tτ indicates the time interval between two consecutive
measurement occasions, which may differ across pairs of observations. The cur-
rent expression of the CT-VAR model is very similar to the DT-VAR model that
was presented in Equation (4.1). The c term represents a vector of intercepts,
and (∆tτ ) represents the residual vector, normally distributed with mean zero
and variance-covariance matrix that is also a function of the time-interval (for
more details see Oud & Jansen, 2000; Voelkle et al., 2012; Voelkle & Oud, 2013).
In place of the Φ matrix in Equation (4.1), these lagged variables are related
by eA∆tτ , the matrix exponential of the drift matrix A, multiplied by the time-
interval between those measurement occasions.
It follows that the CT- and DT-VAR lagged regression matrices are related to
each other by the expression

eA∆t = (eA )∆t = Φ(∆t) (4.5)

3 With very high frequency data, this model can be estimated by first calculating the derivative and
then fitting the first-order SDE directly to data (see for example Boker, Deboeck, et al., 2010; Chow,
2019). However, when observations are collected with a longer time-interval between measurements
(as is typical in ESM settings), we can estimate these same model parameters by fitting the integral
form of the differential equation model. Yet another approach involves numerically integrating the
differential equation during estimation (e.g., Ou et al., 2019). The current approach, using the ana-
lytic integral form allows us to demonstrate the connection between current DT models and their CT
equivalent most clearly.

84
 4.3. CT Network Analysis: Accounting for Continuity

which states that the lagged parameters for any particular time-interval Φ(∆t)
can be found by taking the matrix exponential of the moment-to-moment drift
matrix A to the power of the length of that time-interval ∆t (cf. Oud & Jansen,
2000; Voelkle et al., 2012). Notice the similarity between this relationship, and
the expression used to relate the half-hour and one-hour parameter matrices in
Equation (4.2): To find the lagged relationships at a longer time-interval, we
again take the appropriate matrix exponent of the lagged relationships at the
shorter interval.
Since there is a clear connection between the CT- and DT-VAR models, we
can now consider the two consequences of the time-interval problem from a CT
perspective. First, the conclusions we would intuitively make based on a DT-
VAR model using any one particular interval may differ from those we would
make at any other interval. Figure 4.3 depicts the lagged relationships of the
Stress-Discomfort system over a range of time-intervals, from zero to two hours,
showing clearly that observing this process at say, 15 minute intervals (∆t = 0.25)
instead of half-hour or one-hour intervals leads to a quantitatively and qualita-
tively different set of conclusions about the underlying process. In Figure 4.4 we
show how the DT network centrality measures also change according to the time-
interval, leading to different conclusions regarding optimal intervention targets
when applying current standard practice (an effect which is particularly pro-
nounced for the betweenness measure in panel (c)).
Second, the CT perspective shows that lagged regression parameters Φ(∆t)
at any interval should be interpreted as total rather than direct effects from a
path-tracing perspective (Aalen et al., 2016; Deboeck & Preacher, 2016). This
follows directly from the observation that the matrix exponential relationship in
Equation (4.5) is simply a generalization of the path-tracing operation described
above when relating the half-hour and one-hour networks (Equation (4.2)): The
lagged relationships at a very short time-interval are described by (a function of)
the A matrix, and to find the relationships between variables spaced further (i.e.,
∆t) apart, we apply a path-tracing operation through the latent values of Y (t) in-
between those occasions. See Appendix 4.C for an accessible derivation of this
term as a path-tracing operation.
This interpretation also allows us to explain the changing patterns of lagged
relationships we observe in Figure 4.3. For example, we know that Stress has
no direct moment-to-moment effect on the rate of change of Physical Discom-
fort (a41 = 0), but we see that the corresponding lagged relationship in the DT
model (φ41 (∆t) in Figure 4.3(c)) is strongly negative at short intervals and weakly
positive at longer intervals: This is because it is a total effect, made up of one
negative indirect pathway (through Anxiety) and one positive indirect pathway
(through Anxiety and Self-Conscious). In order to calculate lagged direct and in-
direct effects from the CT perspective, we can use path-tracing methods defined
by Deboeck and Preacher (2016) and Aalen et al. (2016) and described in detail in
Appendix 4.D. The direct lagged effect of Stress on Physical Discomfort is found
by first omitting the indirect pathways from A (in this case, a31 , a42 , a32 and a43 )
before applying the matrix exponential function in Equation (4.5). In so doing we
trace a path from Stress to Physical Discomfort which does not pass through any

85
4. A Continuous-Time Approach to Network Analysis and Centrality

1.0
Parameter Value
0.5
φ11(∆t)
φ22(∆t)
φ33(∆t)
φ44(∆t)
0.0
−0.5

0.0 0.5 1.0 1.5 2.0

Time Interval (∆t)

(a) Auto-regressive parameters φii (∆t)

1.0
Parameter Value

φ12(∆t)
0.5

φ14(∆t)
φ21(∆t)
φ32(∆t)
φ42(∆t)
0.0

φ43(∆t)
−0.5

0.0 0.5 1.0 1.5 2.0

Time Interval (∆t)

(b) Cross-lagged parameters φij (∆t) for which aij , 0

1.0
Parameter Value

φ13(∆t)
0.5

φ23(∆t)
φ24(∆t)
φ31(∆t)
φ34(∆t)
0.0

φ41(∆t)
−0.5

0.0 0.5 1.0 1.5 2.0

Time Interval (∆t)

(c) Cross-lagged parameters φij (∆t) for which aij = 0

Figure 4.3: Lagged regression parameters as a function of the time-interval for the Stress-Discomfort
system. Black dotted lines indicate the parameter values of the half-hour and one-hour networks in
Figure 4.1(a) and (b).

86
 4.3. CT Network Analysis: Accounting for Continuity

1.0
Centrality Value
0.5
Y1( t)
Y2( t)
Y3( t)
0.0
Y4( t)
−0.5

0.0 0.5 1.0 1.5 2.0

Time Interval (∆t)

(a) Two-Step Expected Influence EI (2)

1.0
Centrality Value
0.5

Y1( t)
Y2( t)
Y3( t)
0.0

Y4( t)
−0.5

0.0 0.5 1.0 1.5 2.0

Time Interval (∆t)

(b) One-Step Expected Influence EI (1)

3.0

●
Centrality Value

● ● ● ●
2.0

Y1(t)
Y2(t)
Y3(t)
● ● ● Y4(t)
1.0

● ●

● ● ● ● ● ● ● ● ●
0.0

0.0 0.5 1.0 1.5 2.0

Time Interval (∆t)

(c) Betweenness Centrality BC

Figure 4.4: Centrality measures as a function of the time-interval for the Stress-Discomfort system.
Black dotted lines indicate the centrality values of the half-hour and one-hour networks in Figure
4.1(a) and (b).

87
4. A Continuous-Time Approach to Network Analysis and Centrality

unobserved, intermediate values of Anxiety or Physical Discomfort. To find the

indirect effect, we can either omit the direct dependency between the two target
variables, or alternatively we can take the difference between the direct and total
effects (see Appendix 4.D for details).

4.3.3 Conclusion
To summarize, the CT-VAR model offers a promising alternative to DT-VAR net-
work approaches, allowing us to directly address the time interval problem. First,
it allows for an elegant treatment of unequal time-intervals between observa-
tions, which is a typical characteristic of ESM data. Second, the CT model offers
an alternative conceptualization of the dynamic network structure by breaking
down lagged relationships between measurements into their most fundamental
building blocks, that is, the moment-to-moment relationships encoded by the
drift matrix A. Given an estimate of this drift matrix (which can be obtained with
software packages such as ctsem (Driver et al., 2017) or dynr (Ou et al., 2019)), the
relationship in Equation (4.5) allows us to explore how the dynamic relations de-
pend on the time-interval, as we illustrated with the Stress-Discomfort system.
Third, we can use path-tracing logic suggested by previous authors to calculate
direct, indirect and total lagged relationships in a CT network.
The CT perspective also has consequences for the interpretation and practical
utility of DT centrality measures. Specifically, existing measures fail to account
for the CT nature of the underlying process, and —as a consequence— even their
intuitive interpretation as a summary of direct, indirect and total relationships is
not valid from a CT perspective. This implies that using DT centrality measures
to decide which variable should be intervened on is simply unfounded. Hence,
we need to define new centrality measures that adhere to the CT perspective, but
that also have a clear link with interventions.

4.4 Interventions and Centrality for CT Networks

In this section we address the intervention problem by developing new fit-for-
purpose centrality measures that both account for the CT nature of the underly-
ing process, and that have an explicit link to the effects of interventions as defined
in the interventionist causal inference framework. This allows us to use the CT
approach to decide which variable in the network to intervene on. To achieve this
goal, we need to take a number of steps.
As we have detailed above, centrality measures tend to serve as summaries
of total, direct or indirect effects: For instance, they are intended to reflect the
degree of influence a particular variable has on all others in the network, or they
are supposed to capture the amount of information flow which passes through
a particular variable. As such, to develop new centrality measures that can be
interpreted in such terms, we first need to link total, direct and indirect effects in
a CT model to the effects of variable interventions.

88
 4.4. Interventions and Centrality for CT Networks

Once we know how to interpret path-specific effects in terms of particular in-

terventions on one or more variables in the system, then we can define three new
centrality measures as summaries of these path-specific effects. By doing this, the
new centrality measures we develop can be interpreted in terms of the expected
change over time in the network given a specific type of intervention on a specific
variable. Hence these measures can be used to inform researchers about which
variable to target, and what kind of intervention should be applied to achieve the
desired effect. At the end of this section we apply these CT centrality measures to
our running example and show that they lead to different and novel insights into
the underlying dynamic system than what would be concluded using current DT
measures.
Throughout, we take a simplified approach, based on a number of assump-
tions that were described briefly in Section 4.2.3.1. That is, we assume that in-
terventions can be made in the system without changing how the variables in the
system relate to one another (i.e. modularity), and that all relevant variables are
measured and included in the model (i.e. no unmeasured confounders). In ad-
dition to assuming a first-order process, we also assume that the CT-VAR model
fully describes the dynamics of this first-order process. These simplifying as-
sumptions allow us to provide a straightforward link between CT-VAR model pa-
rameters and the effects of variable interventions under ideal conditions, which
we will illustrate using the Stress-Discomfort example.

4.4.1 Path-Specific Effects and Interventions

Here we generalize intervention-based definitions of path-specific effects in DT
lagged regression models (cf. Eichler & Didelez, 2010; VanderWeele, 2015) to a
CT-VAR setting. In doing so, we provide a clear conceptual link between a path-
tracing effect in the CT model and an actual intervention.

4.4.1.1 Total Effects and Interventions

In terms of the Stress-Discomfort system, let’s imagine that we are able to in-
duce a momentarily high experience of Anxiety in our participant, for instance
by making the participant view a negative photograph, a manipulation which has
been shown to increase state anxiety in lab studies (Richards & Whittaker, 1990;
Richards, French, Johnson, Naparstek, & Williams, 1992). We refer to such an
intervention at one moment in time as an acute intervention (also referred to as an
atomic intervention by Eichler & Didelez, 2010). In the time-series literature this
is sometimes referred to as an impulse, and in the causal inference literature, they
denote such an intervention using the do operator, with do(Y2 (t) = 1) meaning we
intervene to set Anxiety to a value of one at time t.
We can visualize this intervention by plotting the expected trajectories of the
four different variables in our Stress-Discomfort system following this interven-
tion: This is shown in Figure 4.5(a). Following the initial intervention on Anxiety,
the other three variables leave their equilibrium. Eventually, the effect of the in-
tervention fades, and all variables return to their resting state. This shows that

89
4. A Continuous-Time Approach to Network Analysis and Centrality

an intervention on Anxiety changes the value of all variables in our model. If

we want to describe the effect of this intervention on a particular variable, say
Physical Discomfort, in formal terms we would express this as a difference in ex-
pected value of Physical Discomfort when comparing two different interventions
on Anxiety.4 For example, if we instead set Anxiety to its equilibrium value,
do(Y2 (t) = 0), none of the variables would move away from equilibrium. By con-
sidering the contrast between the two expected values of Physical Discomfort
(some time ∆t after these interventions), we get the Total Effect of Anxiety now
on Physical Discomfort. This is written formally as
h i h i
T E24 (∆t) = E Y4 (t + ∆t) do(Y2 (t) = 1) − E Y4 (t + ∆t) do(Y2 (t) = 0) (4.6)
Since we are dealing with a dynamic system, the value of this total effect depends
on the time-interval that elapses after the intervention is applied: In Figure 4.5(a)
we can see that initially the intervention resulted in Physical Discomfort taking
on a negative value, but at longer intervals the effect was to make Physical Dis-
comfort take on a weak positive value.
In order to calculate the value of this total effect, we need to plug in an ex-
pression or model for these expected values. Using the simplifying assumptions
described above, we plug in the CT-VAR model for each expected value, and we
can express this total effect as
T E24 (∆t) = eA∆t [42] (4.7)
where eA∆t [42] denotes the parameter in the fourth row, second column of eA∆t .
Critically, the expression for the causal effect presented above is identical to the
path-tracing definition of a total effect given by Deboeck and Preacher (2016)
and Aalen et al. (2016) and described in Appendix 4.D. We show the equivalence
of these two different definitions of the total effect in Appendix 4.E. This implies
that the CT total effect, under idealized conditions, reflects a very specific type of
variable intervention: An intervention to increase the value of the cause variables
(Anxiety) at a single point in time. The value of the total effect reflects the change
we would expect to see in the effect variable (Physical Discomfort) some time later.
These definitions may seem intuitive, obvious or overly simplistic, matching the
usual interpretation given to regression parameters: If we increase X, how do we
expect Y to change? However, this conceptual link will prove helpful in defining
other path-specific effects which may be less intuitive, and in defining centrality
measures later.

4.4.1.2 Direct Effects and Interventions

The interventionist definition of a direct effect typically consists of two interven-
tions: One to change the cause variable, just as we did for the total effect, and one
4 Here since we are dealing with a single-subject dynamic process, the expectation is defined with
respect to the stochastic input, that is, the normally distributed noise process, as is standard in time-
series approaches (Hamilton, 1994). Given that we do not change the properties of the system (by the
modularity assumption) and that time points t are interchangeable (by the stationarity assumption),
we can interpret this as an expectation for an individual over a population of time points t. This is
analogous to a causal effect for that individual for an unspecified point in time t.

90
 4.4. Interventions and Centrality for CT Networks

1.0
1.0

0.5
0.5

Variable Values
Variable Values

● Y1(t)
Y2(t)

0.0
0.0

●
● Y3(t)
● Y4(t)
−0.5

−0.5
−1.0

−1.0
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Time (t) Time (t)

(a) T E Intervention (b) DE Intervention

Figure 4.5: Illustration of each total and direct effect of Anxiety (Y2 (t)) on Physical Discomfort (Y4 (t +
∆t)) in terms of the interventions they describe in the system over time. Acute interventions are
indicated by empty diamonds, and continuous interventions by filled diamonds.

to keep one or more mediating variables fixed (cf. Robins, 2003; Robins & Richard-
son, 2010; VanderWeele, 2015). For example, suppose that we can intervene in
our system to keep the variables Stress and Self-Conscious fixed to a normal value
at every moment in time over an interval. We will refer to this type of intervention
as a continuous intervention, denoting it do(Y1 (t + ∆t) = 0) and do(Y3 (t + ∆t) = 0)
for Stress and Self-Conscious respectively.5 For example, we can imagine that we
can prompt the participant to engage in a long-lasting mindfulness meditation
focused on reducing stress reactivity (Hoge et al., 2013), or, alternatively, that we
can administer a drug which keeps stress levels fixed to a low or normal level.
Now suppose that we are interested in evaluating the effect of momentarily
increasing Anxiety, on the value of Physical Discomfort, if we intervene to keep
Stress and Self-Conscious fixed. This is visualized by the trajectories in Figure
4.5(b). Just like the total effect intervention, Anxiety starts at a high level and
dissipates back to equilibrium. Here however, we keep Stress and Self-Conscious
fixed at all moments in time, so they stay at their equilibrium value. We can see
that the effect of increasing Anxiety on Physical Discomfort has changed: Phys-
ical Discomfort is pushed even further from equilibrium, taking on a stronger
negative value at short intervals. Since we keep the mediating variables fixed, we
no longer activate the positive feedback loop present in the total effect. Instead,
Physical Discomfort takes longer to return to baseline, still taking on a negative
value at t = 2.
Just as we did for the total effect, we can define the effect of this intervention
formally as the difference between two conditional expectations. Here we change
5 We can consider this to be a special case of an acute intervention defined by Eichler and Didelez
(2010), that is repeated continuously over time. See Appendix 4.E for details. Note that this type of
direct would be referred to as a controlled direct effect rather than a natural direct effect in the causal
inference literature (VanderWeele & Tchetgen Tchetgen, 2017). Since we assume a linear model, the
controlled and natural direct effect formulations lead to equivalent results.

91
4. A Continuous-Time Approach to Network Analysis and Centrality

Anxiety, but keep Stress and Self-Conscious fixed in both cases, and we can define
the effect of this combination of interventions as the direct effect. It can be written
formally as
h i
DE24·13 (∆t) = E Y4 (t + ∆t) do(Y2 (t) = 1), do(Y1 (t + ∆t), Y3 (t + ∆t) = 0)
h i
− E Y4 (t + ∆t) do(Y2 (t) = 0), do(Y1 (t + ∆t), Y3 (t + ∆t) = 0) (4.8)

and by plugging in the CT-VAR parameters for each conditional expectation, we

can express the effect of this intervention as
(D[−1,−3]) ∆t
DE24·13 (∆t) = eA [42] (4.9)

where A(D[−1,−3]) denotes the drift matrix in which the indirect pathway parame-
ters linking Anxiety to Physical Discomfort through the mediating variables (that
is, a12 , a32 , a43 ) are set to zero; hence, in this drift matrix only the direct links be-
tween Anxiety and Physical Discomfort are retained. The proof that the effect of
this intervention can be expressed in this way is given in Appendix 4.E.
Again, we can see that this expression is exactly equivalent to the path-tracing
definition of a direct effect in a CT model given by Deboeck and Preacher (2016).
This means that the CT direct effect under ideal conditions describes the effect
of a very specific set of interventions: An acute intervention to change the value
of the cause variable, combined with a whole set of continuous interventions to
keep each mediating variable fixed in value. Note that we can also define direct
effects in which only one of the mediating variables is kept fixed with a continu-
ous intervention in an analogous way: For example, we could describe the direct
effect of Anxiety on Physical Discomfort relative to Self-Conscious as DE24·3 (∆t),
in which only Self-Conscious is kept fixed by a continuous intervention. Then we
would need to omit only the parameters a32 and a43 from the altered drift matrix
A(D[−3]) .
Since direct effects consist of combinations of at least two or more interven-
tions (depending on the number of mediators considered), they may not be the
most useful measure to inform interventions in practice. However, it is necessary
to define the direct effect in this way in order to define indirect effects, which we
discuss next, and which form the basis of the indirect effect centrality measure
that we present later.

4.4.1.3 Indirect Effects as Intervention Contrasts

Defining the indirect effect as a combination of unique interventions is somewhat
more challenging than doing so for the total and direct effect. This has been
discussed in detail both for general mediation models, and more recently for CT
models (Robins, 2003; Robins & Richardson, 2010; Didelez, 2019). Here we will
define the indirect effect as a contrast between a total effect intervention and a
direct effect intervention.
Consider that the total effect and the direct effect both describe aspects of the
relationship between a cause variable Yi (t) and an effect variable Yj (t + ∆t): They

92
 4.4. Interventions and Centrality for CT Networks

both describe what happens to Yj (t + ∆t) if we intervene in an acute way on Yi (t).

The only difference between the two is that the influence of this acute interven-
tion may differ if we also intervene to keep one or more mediator variables fixed,
that is, the total effect may not be the same as the direct effect. Hence, for each
mediator or set of mediators, we can quantify exactly what difference it makes if
we keep it or them fixed. We do this by looking at the difference between the total
effect and the direct effect.
In other words, the indirect effect IE(∆t) describes how the effect of acutely
intervening on Yi (t) changes when we also intervene to keep the mediator(s) Yk
fixed from t to t + ∆t (i.e., with a continuous intervention) . For instance, we may
be interested in the mediating roles that both Stress and Self-Conscious play in
the relationship between Anxiety on Physical Discomfort. To quantify this, we
would define the indirect effect of Anxiety on Physical discomfort (relative to
Stress and Self-Conscious) as

IE24·13 (∆t) = T E24 (∆t) − DE24·13 (∆t). (4.10)

In terms of the trajectories in Figure 4.5 the indirect effect is the difference in
value of Physical Discomfort at any point in time in panel (a)(representing the
total effect T E(∆t)), and the corresponding point in time in panel (b) (represent-
ing the direct effect DE(∆t)), and so this particular indirect effect describes the
mediating role of the variables Stress and Self-Conscious combined. We can ex-
press this indirect effect in terms of the CT-VAR parameters as
(D[−1,−3]) ∆t
IE24·13 (∆t) = eA∆t [42] − eA [42] (4.11)

that is, as a difference between the total and direct effect calculations given above.
It follows that the effect of this intervention is equivalent to the path-tracing def-
inition of the indirect effect (given in Appendix 4.D).
We may be interested in the indirect effect if we wish to answer questions like:
How does keeping Stress and Self-Conscious fixed alter the effect of a momentary
increase in Anxiety on Physical Discomfort? We can get an idea of the role of
these mediators by comparing the total effect and the direct effect trajectories in
Figure 4.5(a) and (b) respectively. In this instance we see that keeping both me-
diators fixed changes how Stress effects Self-Conscious, meaning Self-Conscious
reacts in a more strongly negative way to changes in Stress than if both mediators
are not kept fixed. We may also be interested in assessing indirect effects one
mediator at a time. For instance, if we want to decrease the effect that experi-
encing an acutely high level of Anxiety has on feelings of Physical Discomfort,
we may choose to either apply a continuous intervention on Stress (IE24·1 (∆t)) or
Self-Conscious (IE24·3 (∆t)), depending on whichever particular indirect effect is
largest. It is this kind of indirect effect that we will use later in defining a cen-
trality measure that describes the flow of information in the network through a
particular variable.

93
4. A Continuous-Time Approach to Network Analysis and Centrality

4.4.2 Centrality Measures to Identify Intervention Targets

Having established the link between CT path-specific effects as described by pre-
vious authors (Deboeck & Preacher, 2016; Aalen et al., 2018) and variable inter-
ventions from the causal inference framework (Eichler & Didelez, 2010; Pearl,
2009; Robins, 2003), we now propose three new centrality measures for CT net-
works. Each centrality measure is explicitly defined as a summary of one of the
intervention-based path-specific effects defined above. This means that these
centrality measures are functions of the time-interval and have a clear link to
a particular type of variable intervention. The three measures we introduce each
capture a summary of the total, direct, and indirect effects of one variable on all
others. The first and last measures are likely to be most interesting and useful
for researchers in practice, as they allow to choose the optimal target for an acute
and continuous intervention respectively. The second measure, summarizing di-
rect effects, is less straightforward to use to inform interventions, but is included
for the sake of completeness.

4.4.2.1 CT Total Effect Centrality

We define our first new centrality measure as the Total Effect Centrality (T EC) of a
variable, which can be calculated by summing the total effect of Yi (t) on all other
variables, at a particular time-interval
p
X
T ECi (∆t) = T Eij (∆t). (4.12)
j,i

hence, we sum over all the total effects of Yi on other variables in the network
(excluding Yi itself). The T EC thus summarizes the effect of a single intervention,
an acute intervention to change Yi (t), on the system as a whole, that is, the cu-
mulative effect on the network, some time-interval ∆t later. Furthermore, since
we explicitly make this centrality measure a function of the time-interval, we can
examine how the cumulative effect of this intervention varies and evolves fol-
lowing the intervention moment. By calculating this measure for each variable,
we can directly inform the choice of optimal intervention target: Which variable
should I change the acute value of to achieve the biggest change in the cumulative
activation levels in the rest of the system later?
Figure 4.6(a) shows the T EC of each variable in the Stress-Discomfort system
over a range of intervals, from ∆t = 0 to ∆t = 1.5. From this we can see that
at short intervals, an acute intervention to increase Physical Discomfort has the
biggest cumulative effect on the network: Overall, this intervention on Physical
Discomfort results in the other variables increasing in value over the next half an
hour or so, before eventually the effect of this intervention fades away. Interven-
tions to increase Stress and Self-Conscious respectively have similar but weaker
effects. Notably, an intervention to increase Anxiety has a weak net negative ef-
fect on the system at shorter intervals, and a weak net positive effect at longer
intervals: We would expect this based on our visualization of that intervention in

94
 4.4. Interventions and Centrality for CT Networks

Figure 4.5(a), where a pulse to Anxiety resulted in Stress and Physical Discomfort
taking on negative values at short intervals.
Based on this, if we want to pick the optimal intervention target for an acute
intervention, the T EC measure allows us to see that Physical Discomfort is the
optimal target for this type of intervention, assuming we can set Physical Dis-
comfort to a low or negative value (e.g. do(Y4 (t)) = −1).

4.4.2.2 Direct Effect Centrality

To summarize the direct influence of one variable on all others, we propose to
take a summary of different direct effects, where the latter is defined in Equation
(4.8). For instance, we defined the direct effect of Anxiety above as the change
in an outcome variable, Physical Discomfort, following the combination of an
acute intervention to change Anxiety, and continuous interventions to keep all
other mediating variables (Stress and Self-Conscious) fixed. The measure we in-
troduce here, Direct Effect Centrality (DEC) takes the sum of all such direct effects
from Yi (t) to all other possible outcome variables, if all remaining variables in the
model are kept fixed
p
X
DECi (∆t) = DEij·k (∆t) (4.13)
i,j

where for each pair Yi (t) and Yj (t + ∆t) all possible mediating variables k ∈ p\(i, j)
are kept fixed. Note that unlike the T EC measure, DEC reflects a summary
measure of different interventions, as for each pair Yi (t) and Yj (t + ∆t), there is
a different set of mediators that must be intervened on to establish the direct
effect. For instance, in Figure 4.5(c) we showed the direct effect of Anxiety on
Physical Discomfort when intervening to keep Stress and Self-Conscious constant
(DE24·13 (∆t)). This represents only one of the three components which make up
the DEC of Anxiety: The other two terms consider interventions where either
Self-Conscious and Physical Discomfort are kept constant (DE21·34 (∆t)) or Phys-
ical Discomfort and Stress are kept constant (DE23·14 (∆t)). Although it is not
straightforward to use this measure to choose a specific intervention target, this
measure does reflect that part of one variables relationship with the network as
a whole which is truly a result of only direct relationships (from a path-tracing
perspective).
Figure 4.6(b) shows the direct effect centrality of each variable over a range
of intervals. From this we see that Stress has the highest positive DEC, Self-
Conscious and Physical Discomfort have weaker overall positive direct influ-
ences, that fade to zero at a quicker rate, and Anxiety has a net direct influence
closest to zero across a range of intervals.6 Although Stress has the strongest truly
direct influence on the network, the practical value of the DEC —or any measure
that is based on direct effects— for choosing intervention targets is limited, as it is
6 The high DEC of Stress is due to the positive feedback loop between Stress and Anxiety in the
underlying drift matrix (a21 = 5.5, a12 = 1.25). Since the direct effects involve interventions to control
the rest of the variables in the system, and no other pair of variables has this direct feedback loop
relationship, all other values of DEC are lower.

95
4. A Continuous-Time Approach to Network Analysis and Centrality

1
Centrality Value Y1( t)
Y2( t)
0
Y3( t)
Y4( t)
−1

0.0 0.5 1.0 1.5 2.0

Time Interval (∆t)

(a) Total Effect Centrality T EC(∆t)

1
Centrality Value

Y1( t)
Y2( t)
0

Y3( t)
Y4( t)
−1

0.0 0.5 1.0 1.5 2.0

Time Interval (∆t)

(b) Direct Effect Centrality DEC(∆t)

1
Centrality Value

Y1( t)
Y2( t)
0

Y3( t)
Y4( t)
−1

0.0 0.5 1.0 1.5 2.0

Time Interval (∆t)

(c) Indirect Effect Centrality IEC(∆t)

Figure 4.6: Illustration of the new total, direct and indirect centrality measures for CT networks,
applied to the Stress-Discomfort system.

96
 4.4. Interventions and Centrality for CT Networks

based on a complex combination of different interventions, that cannot actually

take place all at the same time.

4.4.2.3 Indirect Effect Centrality

A much more appealing way to identify an intervention target is by quantifying
the role that variable plays as a mediator of other relationships between variables
in the network. To do so, we can use the indirect effect measure described in
Equation (4.10) above. Recall that the CT indirect effect captures the change in
the effect of Yj (t) on Yk (t + ∆t), if we intervene to keep the mediator Yi fixed at
every moment in time (T ECjk (∆t) − DECjk·i (∆t)). Hence, we define the Indirect
Effect Centrality (IEC) of a mediator variable Yi as
X
IECi (∆t) = IEjk·i (∆t) (4.14)
(j,k):j,k,i

that is, the sum of all possible indirect effects between other pairs of variables
Yj (t) and Yk (t + ∆t), in which Yi serves as the only mediator. Note here that in
comparison to how we described the indirect effect above, we have switched the i,
j and k notation to reflect that the IEC(∆t) is defined as a property of a mediator,
instead of a property of one particular cause-effect relationship. The summation
denotes that we omit auto-regressive relationships (j , k) and pairs of variables
where the mediator is either the cause or effect variable (j , i and k , i). The IEC
therefore can be understood as quantifying how a continuous intervention on Yi
changes the effects of other variables on each other.
As such, this measure may be especially of interest in the case of networks
of psychopathology variables. For instance, in the Stress-Discomfort system, we
would like to avoid a high value on all four variables as much as possible. The
current measure can be used to determine which of these variables is most impor-
tant in terms of mediating the effects of one variable on another in the system,
such that by intervening on this variable, these indirect paths become blocked
and the flow of activation from one variable to another is interrupted.
Figure 4.6(c) shows the IEC of each variable over a range of intervals. From
this we can see that Physical Discomfort has the strongest indirect effect central-
ity in absolute terms. A strong negative value of IEC means that keeping Physical
Discomfort fixed actually increases the size the effects of other variables on each
other, since the component direct effects are greater than the corresponding total
effects. This happens because Physical Discomfort plays a key role in the only
negative feedback loop in the network: Since an increase in Anxiety actually de-
creases Physical Discomfort (a42 = −7.3), the total effect of Anxiety on Stress is
less strong than its direct effect. If, however, we intervene to keep Physical Dis-
comfort fixed, then this negative compensating effect is not activated, meaning
an increase to Anxiety in fact has a greater effect on the network as a whole. In
contrast, Stress has the largest positive IEC, meaning that keeping Stress fixed
decreases the effects of other variables on one another.
From this we would conclude that we should choose Stress as a target for a
continuous intervention, as it decreases the short-term impact of other variables

97
4. A Continuous-Time Approach to Network Analysis and Centrality

in the network on each other. However, we should avoid at all costs applying a
continuous intervention on Physical Discomfort: Such an intervention would in
fact increase the strength of positive relationships between the other variables.

4.4.2.4 Comparing CT and Existing Centrality Measures

The new CT centrality measures introduced here have a clear link to the choice of
intervention target: The T EC and IEC measures can be easily used to select the
best target for an acute and continuous intervention respectively. Furthermore, it
is clear from comparing the CT centrality measures in Figure 4.6 with the existing
DT measures in Figure 4.4 that the new measures lead to different choices in
intervention targets.
The T EC closely matches the substantive interpretation typically given to the
EI (2) measure (Figure 4.4(a)); however, its value is actually equivalent to calcu-
lating the EI (1) measure across a range of time-intervals (Figure 4.4(b)). This
implies that, unwittingly, the measure that is typically used to summarize direct
influence in a DT network in fact describes the total influence of an acute in-
tervention at a specific time-interval. In contrast to the EI (1) measure, the DEC
provides a rather different picture of which variable has a strong direct influ-
ence on the network. However, we would argue that truly direct influence is less
important when choosing intervention targets in the sense described above.
Finally, we believe that the IEC is a useful way to summarize the mediating
role a variable plays in the network, allowing us to choose an optimal target for
a continuous intervention. Comparing this measure to Betweenness Centrality
in Figure 4.4(c), which is the measure used to quantify indirect influence in DT
networks, we see again that we would reach entirely different conclusions. Specif-
ically, there are few intervals at which calculating Betweenness would correctly
lead us to choose Stress as an intervention target. Perhaps even more crucially,
there are numerous intervals at which we would choose Physical Discomfort as
an intervention target, although we know from the IEC measure that we should
avoid applying a continuous intervention to this variable.

4.5 Discussion
In this paper we have critically appraised current best practice in dynamical
network analysis. We have especially focused on the interpretation of DT-VAR
models as representing direct causal dependencies, and on the use of central-
ity measures predicated on that interpretation to identify optimal intervention
targets. We identified two major problems with current best practice and ad-
dressed each in turn. First, we addressed the time-interval problem by intro-
ducing a continuous-time approach to network analysis. CT models aim to cap-
ture the moment-to-moment dynamics operating between processes, which can
be represented in network form as a local dependence graph. The estimation of
a CT model overcomes both practical and conceptual shortcomings of the DT
approach, providing a new outlook on how parameters should be interpreted,
and the ability to explore how lagged relations vary and evolve as a function of

98
 4.5. Discussion

the time-interval. Second, we address the lack of clarity regarding how current
centrality measures should be used to inform intervention targets. We do this by
introducing new centrality measures for CT models which have a clear link to
interventionist concepts of causal relationships and can be used to identify the
optimal target for either an acute or continuous intervention.
These developments represent a promising new approach to dynamical net-
work analysis. From a practical point of view, recent R packages ctsem (Driver
et al., 2017) and dynr (Ou et al., 2019), have made it relatively easy to apply the
CT modeling approach in empirical research (e.g., Voelkle et al., 2012; Voelkle
& Oud, 2013; Oud, Voelkle, & Driver, 2018; Ryan et al., 2018), and moreover,
in the supplementary materials of this paper we provide extensive R functions
which can be used to calculate CT path-specific effects and centrality measures
introduced in the current paper based on a given CT-VAR drift matrix.7
As this was a first step towards linking CT approaches, network analyses and
interventionist concepts of causal relationships, our approach was necessarily
simplified, both with respect to the interventions we consider, and our treat-
ment of the assumptions necessary for this type of causal inference. Eichler and
Didelez (2010) describe a variety of different interventions that could be applied
to dynamic systems, and a variety of assumptions that must be met for the effects
of these interventions to be identified from data. Driver and Voelkle (2018) de-
scribe how various hypothetical interventions can be simulated from a CT model,
however, they do not do so within an interventionist framework, and so the as-
sumptions needed to identify the effects of those interventions are unclear. More
research is needed to investigate and evaluate identifiability assumptions in a
psychological context, and to operationalize actual psychological treatments as
dynamic systems interventions. For instance, cognitive-behavioral interventions
may be better defined as interventions on moderators of symptom-symptom re-
lationships (i.e., lagged parameters of a network), rather than interventions on
the symptoms themselves.
Furthermore, the dynamic models we focused on here, that is the DT-VAR
and its CT-equivalent, are very simple models, describing a system that fluctu-
ates around a single fixed point. Although this reflects the majority of dynamical
network analyses in empirical research, it may be more beneficial —from a the-
oretical point of view— to investigate dynamic system models which can show a
qualitative change in behavior, for example bi-stable systems which can transi-
tion between different equilibria (Haslbeck & Ryan, 2019). For those systems we
may be interested in identifying which intervention should be applied in order
to move the system from one equilibrium position to another. We hope that the
current paper will serve as the groundwork for such future developments.
Finally, a further simplification we made in the current paper was to ignore
the role of uncertainty in parameter estimates. Previous research has shown that
existing centrality measure estimates can be somewhat unstable as they reflect
the sums of numerous parameter estimates (Epskamp, Borsboom, & Fried, 2018),
an issue which is likely shared by the CT centrality measures estimated here. In
7 https://osf.io/9sgdn/

99
4. A Continuous-Time Approach to Network Analysis and Centrality

practice the uncertainty around those parameters can be obtained in a relatively

straightforward way using standard Bayesian approaches (as implemented in ct-
sem) or by bootstrapping. In line with this concern regarding parameter uncer-
tainty, researchers should exercise caution when extrapolating from drift matrix
parameters to lagged effects at different intervals. In the main text we assumed
a true underlying CT-VAR process, and so could derive how lagged parameters
vary and evolve over a range of time-intervals. In practice, the CT-VAR must
be estimated from data, which likely contains a relatively limited range of time-
intervals between observations. This means that any inference to lagged relation-
ships outside of the observed intervals is a form of model extrapolation and is
prone to inaccuracies if the model is misspecified Although the CT model at least
allows one to explore what these lagged relationships are expected to look like
at any interval, more research is needed to assess the robustness of inferences to
intervals outside the data in practice.
In conclusion, while network analysis is a promising conceptual framework
for psychopathology, we believe that to move forward in this field we need to be-
come more sophisticated in how we think about the dynamics of the underlying
system and how we conceptualize and learn about interventions in those systems.
In the current paper we have established the first steps in using differential equa-
tion models of the underlying dynamic system to conceptualize network struc-
ture and established how centrality and intervention targets can be approached
from an interventionist perspective. The approach in the current paper repre-
sents a first principled step to moving current practice forward, by placing the
emphasis on the dynamic part of dynamic network analysis and placing inter-
ventions at the heart of centrality-based metrics.

100
 4.A. Centrality Measures as Summaries of Path-specific Effects

Appendix 4.A Centrality Measures as Summaries of

Path-specific Effects
In this appendix we show how path-specific effects in DT-VAR models are re-
lated to three popular centrality measures calculated from DT-VAR networks. For
the measures typically interpreted as quantifying the total and direct influence
of a variable (i.e. both Expected Influence measures), this relationship is quite
straightforward, while for the popular indirect influence measure Betweenness
Centrality, the relationship with path-tracing quantities is much farther removed.
In Table 4.1 we provide the formula and description of the three centrality
measures we consider in the main text. These are expressed in terms of lagged
regression parameters φji , which represent the lagged effect from process i to
process j (i.e., it is the element on the jth row and ith column of the matrix Φ).
The-right hand column of Table 4.1 describes how these calculations relate to
path-tracing quantities from the SEM literature. Note that the Expected Influ-
ence measures were originally developed for undirected networks (Robinaugh et
al., 2016), and so, despite the active applications of those measures for directed
networks (e.g. Kaiser & Laireiter, 2018) their precise definition for direct net-
works is left somewhat ambiguous. For instance, the popular packages qgraph
(Epskamp et al., 2012) and networktools (Jones, 2018) differ slightly in how One-
Step Expected Influence is calculated, with the former excluding diagonal ele-
ments (i.e. auto-regressive effects) as is common for DT-VAR centrality measures,
while the latter includes those elements. The definitions we give here to the One-
(1) (2)
Step and Two-step Expected Influence measures (EIi and EIi ) omit relation-
ships a variable has with itself either one or two occasions later respectively. We
believe this is in keeping with the spirit of how these measures are defined for
undirected networks, and allows us to maintain the standard interpretation of
centrality measures as reflecting a type of relationship the target variable shares
with all other variables in the model.
(1)
From Table 4.1 we can see that EIi , which is typically interpreted as a sum-
mary of direct effects, is in fact the sum of lag-one direct effects of Yi,τ on all other
variables at the next occasion (that is, excluding the auto-regressive direct effect
(2)
of Yi on itself at the next occasion). The EIi measure, which is typically inter-
preted as reflecting the total influence of a variable, comprises two separate parts.
The first part is the sum of lag-two total effects, following standard path-tracing
rules, and excluding the total effect of a variable on itself two occasions later. The
(1) (2)
second part is the EIi measure for that variable. As such, EIi measure is a mix
of total and direct effects at both lags.
Finally, the Betweenness Centrality measure BCi , typically interpreted in
terms of indirect effects, is only tenuously related to path-tracing quantities. In
SEM approaches researchers are typically interested in mediators of indirect ef-
fects, where the size of an indirect effects is defined by the product of the com-
ponent pathways (i.e., path-tracing rules). If we have many indirect pathways,
and many potential mediators, we may wish to know which specific indirect ef-

101
4. A Continuous-Time Approach to Network Analysis and Centrality

Network Measure Formula Description

(1) Sum of lag 1 direct effects

EIi Pp
Yi,τ → Yj,τ+1
One-Step j,i φji
Expected Influence ∀j , i

(2) Sum of total effects at lag 2

EIi Pp Pp
Two-Step j φji k,i φkj + EI1i Yi,τ → Yj,τ+1 → Yk,τ+2 ∀k , i
Expected Influence plus direct effects at lag 1

BCi Mjk (i) = 1 iff Yi ∈ d(jk) Counts how often Yi

Betweenness P Pp is a mediator on
k,j,i Mjk (i)
Centrality the shortest network-path
where d(jk) is Yj,τ → . . . Yi,· · · · → Yk,τ+q
max{φhj + · · · + φkh }

Table 4.1: Relationship between different network metrics and path-tracing quantities, in the context
of a VAR(1) model with p variables, regression coefficient matrix Φ, and corresponding dynamical
network with weights matrix Φ T .

fect is strongest, and in turn, how often a specific variable acts as a mediator
of these strongest indirect effects. It seems that this is how psychological re-
searchers using the BCi measure typically interpret it (e.g., Bringmann et al.,
2013, 2015; David, Marshall, Evanovich, & Mumma, 2018). However, the ac-
tual calculation of this measure differs greatly from the mediator-based metric
described above. Specifically, instead of identifying the largest indirect effect,
Betweenness is based on the identification of the shortest network-path between
two variables (d(jk)). The length of this network-path is based on the inverse
of the sum rather than the product of the individual pathways: While large SEM
paths are those where multiplying each individual part leads to a high num-
ber, we say that short network-paths are those where the sum of each individual
part leads to a small number. Similar to standard path-tracing nomenclature,
these network-paths can be either direct (e.g. Yj,τ → Yk,τ+1 = φkj ) or indirect (e.g.
Yj,τ → Yi,τ+1 → Yk,τ+2 = φij + φki ) and each path may span a different number of
measurement occasions. The Betweenness Centrality of Yi is found by first cal-
culating all the shortest paths between all pairs of variables, and then counting
how often Yi lies on that shortest path. It is clear then that, despite how this mea-
sure is interpreted the relationship of Betweenness Centrality with path-specific
effects is much less direct than for the other measures considered above.

102
 4.B. Centrality Values DT Stress-Discomfort System

Appendix 4.B Centrality Values DT Stress-

Discomfort System
In this appendix we present the centrality metrics for the half-hour network
(Φ(∆t = 0.5)) and the one-hour network (Φ(∆t = 1)) as discussed in section 4.2.
These are shown in Table 4.2.

∆t = 0.5 ∆t = 1
EI (2) EI (1) BC EI (2) EI (1) BC
Stress -0.025 -0.029 1 0.245 0.274 0
Anxiety 0.035 0.038 1 -0.071 -0.109 3
Self-Conscious -0.024 -0.027 0 0.125 0.152 0
Physical Discomfort 0.008 -0.002 1 0.555 0.557 0

Table 4.2: Two-Step Expected Influence (EI (2) ), One-Step Expected Influence (EI (1) ) and Betweenness
Centrality (BC) for each of the four variables in the half-hour (Φ(∆t = 0.5)) and one-hour (Φ(∆t = 1))
networks. These measures are based on the lagged parameter matrices displayed in Figure 4.1. In
each column, the largest centrality values are highlighted in bold.

Appendix 4.C The Matrix Exponential as Path-

Tracing
In this appendix we describe in more detail the relationship between the CT-VAR
or Ornstein-Uhlenbeck model, and the notion of path-tracing effects. The CT-
VAR model is the integral form of the first-order stochastic differential equation
(SDE) model, defined as
dY (t)
= AY (t) + W (t) (4.15)
dt
where A is the drift matrix which regresses the derivative on the value of the
process at that moment in time, and W (t) represents the stochastic innovation
part of the system, also referred to as a Wiener process (which is often denoted
dW (t)
G dt , cf. Oud & Jansen, 2000; Voelkle et al., 2012; Voelkle & Oud, 2013). The
elements of the drift matrix encode direct dependencies between time-varying
processes, with aij representing the direct effect of Yj (t) on the rate of change of
Yi (t).
dY (t)
The first-derivative dt is defined as the change in value of Y (t) over the
time-interval t + s, as the value of s approaches zero

dY (t) Y (t + s) − Y (t)
= lim
dt s→0 s
which means that the deterministic part of the first-order differential equation

103
4. A Continuous-Time Approach to Network Analysis and Centrality

(i.e., ignoring the stochastic innovation part) can be re-written as

Y (t + s) − Y (t)
lim = AY (t)
s→0 s
Re-arranging, we can come to an expression for the relationship between Y (t)
and Y (t + s), as s → 0
Y (t + lim s) = lim s × (AY (t)) + Y (t)
s→0 s→0
= (I + A lim s)Y (t)
s→0
that, is, an expression of the differential equation model as an auto-regressive
model of measurements spaced very closely in time. Thus, the auto-regressive
and cross-lagged relationships between waves spaced an infinitesimally small
time-interval apart (i.e. the moment-to-moment lagged relationships) are given
by I + A lims→0 s.
Now take it that we are interested in finding an expression relating two ob-
served waves of variables Y (t) and Y (t + ∆t). We can think of s as a very small
fraction of ∆t, that is,
∆t
s=
n
such that as n → ∞ we get s → 0. This means that we can re-express the relation-
ship between waves spaced an infinitely small time-interval apart as
∆t ∆t
Y (t + lim ) = (I + A lim )Y (t).
n→∞ n n→∞ n
Now, if we conceptualize the CT-VAR as a path model, as depicted in Figure 4.2 in
the main text, then we can find an expression to relate Y (t) and Y (t +∆t) by a sim-
ple application of path-tracing rules (Bollen, 1987) . That is, we can trace through
the limn→∞ n latent waves in-between those two occasion, by taking the appropri-
ate power of the moment-to-moment lagged-effects matrix I + limn→∞ A ∆t n . This
path-tracing operation gives us
n
Y (t + ∆t) = limn→∞ (I + lim A ∆t
n ) Y (t)
n→∞
n o
= limn→∞ (I + n1 A∆t)n Y (t)
By definition, the first term on the right-hand side is exactly the matrix exponen-
tial (cf. Abadir & Magnus, 2005, p.250)
n o
eA∆t = limn→∞ (I + n1 A∆t)n , (4.16)
giving us
Y (t + ∆t) = eA∆t Y (t), (4.17)
which gives us the deterministic part of the CT-VAR(1) model.
This derivation shows that the CT-VAR model can be seen as a path model,
where the lagged relationships are defined as total effects resulting from path-
tracing through an n → ∞ latent waves. Thus, any DT cross-lagged parameter
matrix Φ(∆t) = eA∆t should be interpreted as reflecting total effects relative to the
CT-VAR model.

104
 4.D. Path-Tracing in CT models

Appendix 4.D Path-Tracing in CT models

In this appendix we describe the calculation of path-specific effects for the CT-
VAR model based on path-tracing rules. Both Deboeck and Preacher (2016) and
Aalen et al. (2016) describe a method for calculating direct, indirect and total ef-
fects in a CT-VAR model, which follow the path-tracing rules laid out by, among
others, Bollen (1987). However, these authors only discuss path-tracing with re-
spect to a lower-triangular tri-variate drift matrix, that is, a drift matrix with
only three variables and without reciprocal lagged relationships. Here we gen-
eralize these path-tracing definitions to drift matrices of arbitrary structure and
number, following path-tracing principles. In large part we follow the methods
described by the original authors, except in the case of the indirect effect, where
non-triangular drift matrices must be approached differently than was done in
the simpler scenario of a lower triangular matrix.
To find the path-tracing total effect of Yi (t) on Yj (t + ∆t), which we will here
denote T Eij (∆t), we simply take the element in the jth row, ith column of the
matrix exponential of the drift matrix:

T Eij (∆t) = eA∆t [ji] (4.18)

This follows from the interpretation of the matrix exponential term eA∆t as a
path-tracing operation, relative to the moment-to-moment auto-regressive effects
matrix (I + A limn→∞ ∆t n ) (described in Appendix 4.C). In Figure 4.7(a) we show
a four-variable CT-VAR model with a full A matrix in path-model form, with
n → ∞ latent values of the processes in between measurement occasions, spaced
at intervals of s → 0. From this it is clear that tracing a path from, for instance,
Y1 (t) to Y4 (t + ∆t) includes paths through latent values of Y1 and Y4 , (e.g. Y1 (t) →
Y1 (t+1s) → Y1 (t+2s) → Y4 (t+3s) → · · · → Y4 (t+∆t)) as well as paths through latent
values of Y2 and Y3 (Y1 (t) → Y2 (t + 1s) → Y3 (t + 2s) → Y4 (t + 3s) → · · · → Y4 (t + ∆t)).
As such, we can interpret this total effect as constituted of all possible pathways
linking Y1 (t) and Y4 (t + ∆t), as is the standard interpretation of a total effect.
In order to find the path-tracing direct effect from Yi (t) to Yj (t + ∆t) relative
to some mediator variable(s) Yk , Deboeck and Preacher (2016) state that the drift
matrix should first be altered so that the parameters which make up the indirect
pathways are omitted. We can alter the drift matrix to achieve this by setting the
kth row and column elements of A to zero, yielding a drift matrix containing only
direct relationships between Yi and Yj , which we will denote A(D[−k]) . The path-
tracing direct effect is then found by applying the matrix exponential function to
the altered drift matrix.
(D[−k]) ∆t
DEij·k (∆t) = eA [ji] . (4.19)

For example, for a four-variable system, to define the path-tracing direct effect of
Y1 (t) to Y4 (t + ∆t) relative to the mediators Y2 and Y3 we would need to alter the

105
4. A Continuous-Time Approach to Network Analysis and Centrality

Y1(t) Y1(t+1s) Y1(t+2s) Y1(t+3s) Y1(t+ns) Y1(t+Δt)

... ...

Y2(t) Y2(t+1s) Y2(t+2s) Y2(t+3s) Y2(t+ns) Y2(t+Δt)

... ...

Y3(t) Y3(t+1s) Y3(t+2s) Y3(t+3s) Y3(t+ns) Y3(t+Δt)

... ...

Y4(t) Y4(t+1s) Y4(t+2s) Y4(t+3s) Y4(t+ns) Y4(t+Δt)

... ...

(a) CT-VAR path model representing total effect pathways

Y1(t) Y1(t+1s) Y1(t+2s) Y1(t+3s) Y1(t+ns) Y1(t+Δt)

... ...

Y2(t) Y2(t+1s) Y2(t+2s) Y2(t+3s) Y2(t+ns) Y2(t+Δt)

... ...

Y3(t) Y3(t+1s) Y3(t+2s) Y3(t+3s) Y3(t+ns) Y3(t+Δt)

... ...

Y4(t) Y4(t+1s) Y4(t+2s) Y4(t+3s) Y4(t+ns) Y4(t+Δt)

... ...

(b) CT-VAR path model representing direct effect pathways

Figure 4.7: Path-model representation of the four-variable CT-VAR model with full drift matrix A. In
the top panel the red pathway highlights a path which is included in the difference-method calcula-
tion of the indirect effect but omitted from the path-method calculation. The bottom panel shows the
pathways which make up the direct effect, with indirect paths (removed from the altered drift matrix
A(D[−k]) ) shaded in gray

drift matrix as follows

Y1 Y2 Y3 Y4 Y1 Y2 Y3 Y4
Y1 a a12 a13 a14  Y1 a 0 0 a14 
   
 11  11
Y  a a22 a23 a24  (D[−2,−3]) Y2  0 0 0 0 
A= 2
 21
, A =

Y3  a31 a32 a33 a34  Y3  0 0 0 0 
  
   
Y4 a41 a42 a43 a44 Y4 a41 0 0 a44

This altered drift matrix defines a new path model, absent of any lagged rela-
tionships linking Y1 to Y4 through the mediators Y2 and Y3 . This is displayed in
Figure 4.7(b). Applying the matrix exponential function to this new drift matrix,
it is clear that we only trace through direct pathways linking Y1 (t) to Y4 (t + ∆t)
(e.g. Y1 (t) → Y1 (t + 1s) → Y1 (t + 2s) → Y4 (t + 3s) → · · · → Y4 (t + ∆t)). This process is
exactly equivalent to how Bollen (1987) describes the calculation of a direct effect

106
 4.E. Interventions and Path-Tracing in CT models

using matrix algebra.

To calculate the indirect effect for a lower-triangular drift matrix, both
Deboeck and Preacher (2016) and Aalen et al. (2016) describe an operation by
which the direct links are omitted from the drift matrix (in the four variable ex-
ample, this would be a14 and a41 ) before applying the matrix exponential term.
We will refer to this as the trace-method of calculating an indirect effect. Alterna-
tively, following path-tracing rules in linear models, we could define the indirect
effect as the difference between the total and direct effect, which we will refer to
as the difference-method. For a lower triangular drift matrix, both methods yield
the same indirect effect (Deboeck & Preacher, 2016).
However, for non-triangular drift matrices, these definitions will not be
equivalent. The reason again follows simple path-tracing rules. The difference
method in this scenario quantifies all paths from Yi (t) to Yj (t + ∆t) that pass
through some latent value of Yk . In contrast, the trace-method quantifies fewer
paths, that is, all paths that pass through Yk , but do not pass along any direct
paths linking Yi (t) to Yj (t + s). In Figure 4.7(a) we have highlighted in red a path-
way which is included as part of the difference-method indirect effect, but which
is not included in the trace-method indirect effect.
In order to maintain the property that the total and direct effects sum to one
another, and to allow an easier link to intervention-based definitions of indirect
effects in Section 4.4.1 of the main text, we recommend the use of the difference
method of calculating indirect effects. As such, we define the path-based indirect
effect as
(D[−k]) ∆t
IEPij (∆t) = eA∆t [ji] − eA [ji] , (4.20)
which is equivalent to the difference between the path-tracing total effect and the
path-tracing direct effect described above.

Appendix 4.E Interventions and Path-Tracing in CT

models
In this appendix we prove the equivalence between the intervention-based def-
initions of total, direct and indirect effects, and the path-tracing definitions of
these quantities (described in Appendix 4.D), given the simplifying assumptions
described in the main text.

4.E.1 Total Effect

We define the total effect of Yi (t) on Yj (t + ∆t) as the expected change in value
of Yj (t + ∆t) given an acute intervention to set the value of Yi (t) from a constant,
yi∗ to a new value yi . We denote such a variable-setting operation using the do
operator (Pearl, 2009), and so can express this total effect as
h i h i
T Eij (∆t) = E Yj (t + ∆t) do(Yi (t) = yi ) − E Yj (t + ∆t) do(Yi (t) = yi∗ ) (4.21)

107
4. A Continuous-Time Approach to Network Analysis and Centrality

By assuming that the system is fully observed (i.e., there are no unobserved con-
founders), and by assuming modularity, we can substitute the expected value of
Yj (t + ∆t) following an intervention do(Yi (t) = yi ) with the expected value given
we observe Yi (t) = yi . This yields the expression
h i h i
T Eij (∆t) = E Yj (t + ∆t) Yi (t) = yi − E Yj (t + ∆t) Yi (t) = yi∗

Now we plug in the CT-VAR model for those expected values. Take it that Y (t)
represents a column vector of variable values with ith element Yi (t) = yi . Using
this, we can express the first expected value as
h i
E Yj (t + ∆t) Yi (t) = yi = {eA∆t Y (t)}[j]

that is, the jth element of the column vector obtained by multiplying the square
matrix eA∆t with the column vector Y (t). To obtain the second expectation, we
take it that Y ∗ (t) represents a column vector of variable values with ith element
Yi (t) = yi∗ but which is otherwise identical to Y (t).
h i
E Yj (t + ∆t) Yi (t) = yi∗ = {eA∆t Y ∗ (t)}[j]

Taking the difference between these two expected values we obtain

T Eij (∆t) = {eA∆t Y (t)}[j] − {eA∆t Y ∗ (t)}[j]

 
= eA∆t · Y (t) − Y ∗ (t)
[j]

Since the vectors differ only with respect to their ith element, we obtain

T Eij (∆t) = eA∆t [ji] × (yi − yi∗ ) (4.22)

where eA∆t [ji] is the element in the jth row and ith column of the matrix eA∆t . If
we define the intervention as increasing the value of Yi (t) by one unit (yi − yi∗ = 1),
this yields an expression exactly equivalent to the path-tracing definition of a
total effect given in Appendix 4.D.

4.E.2 Direct Effect

We define the direct effect of Yi (t) on Yj (t + ∆t) as the expected change in value of
Yj (t+∆t) given an acute intervention to set the value of Yi (t) from yi∗ to a new value
yi , while also intervening to keep the value of the mediator(s) Yk fixed to a con-
stant yk at every moment in time in that interval. We denote this latter continuous
intervention using the do operator over an interval of time as do(Yk (t + ∆t) = yk ,
and so express the direct effect as
h i
DEij·k (∆t) = E Yj (t + ∆t) do(Yi (t) = yi ), do(Yk (t + ∆t) = yk )
h i
− E Yj (t + ∆t) do(Yi (t) = yi∗ ), do(Yk (t + ∆t) = yk ) (4.23)

108
 4.E. Interventions and Path-Tracing in CT models

for some mediator(s) k ∈ p. Intuitively, if we want to block the indirect effect that
acts through a mediator, we would need to ensure that either the mediator does
not react to changes in the cause variable, or that it does not transmit information
to the effect variable, or both. If we wish to achieve this by intervening on a
variable, it is straightforward to see that we must do so by intervening to set the
value of the mediator to a constant at every point in time between t and t + ∆t.
As with the total effect derived above, the next step consists of plugging in the
CT-VAR model for the expected values in this expression. However, note that due
to the need to define the continuous intervention on the mediator do(Yk (t + ∆t) =
yk this proof is a little more involved than that of the total effect above (and relies
on the strong assumption that applying such an intervention does not change
the dynamics of the underlying process). To derive an expression for the direct
effect, we first begin with the expression for the expected value of Y (t + ∆t) given
an acute intervention on the cause variable Yi (t), that is,
h i
E Y (t + ∆t) Yi (t) = yi = eA∆t Y (t)

one of the components of the total effect given above. Recall from the derivation
in Appendix 4.C that we can write the CT-VAR model as describing lagged rela-
tionships over an infinitesimally small time interval limn→∞ ∆tn , hereby referred
to as the moment-to-moment relationship. This gives us
h i
E Y (t + lim ∆t ∆t
n ) = (I + A lim n )Y (t).
n→∞ n→∞

From now we will treat this expression as defining a moment-to-moment path

model, as depicted in Figure 4.2 in the main text, and with a slight abuse of
notation we will substitute limn→∞ ∆t n for s, which we will define as a “moment”
in time. We can express the expected value two “moments” after t as
h i
E Y (t + 2s) = (I + As)Y (t + s).
= (I + As)(I + As)Y (t)
= (I + As)2 Y (t)

where
h thei second and third line follow by substituting in the expression for
E Y (t + s) given above.
Now, to define the direct effect we need to express the expected value of Y (t +
2s) given that we have intervened to set the current value of the mediator (Yk (t +
2s)), the value of the mediator one “moment” previously (Yk (t + s)), and the initial
value of the mediator Yk (t) to zero. In order to derive such an expression, we
introduce two simplifications here. First, since we are focusing on a linear model,
and we are interested in the difference between two expected values in which
in both cases the mediator Yk is set to the same value yk , the specific value we
choose for yk is irrelevant. For ease of notation we will therefore consider only
an intervention by which yk is equal to zero (i.e., the equilibrium position of
Yk ). Second, to aid in our derivation, we will express the do operator in matrix
algebraic terms. That is, we will represent the operation do(Yk (t) = 0) using a

109
4. A Continuous-Time Approach to Network Analysis and Centrality

transformation matrix D[−k] , a p × p matrix with zeros as off-diagonal elements, a

zero on the kth diagonal element, and ones as the other diagonal elements. For
instance, a 3 × 3 matrix D[−2] would be given as

1 0 0
 
D[−2] = 0 0 0 .

0 0 1
 

Pre-multiplying a column vector by the matrix D[−k] reproduces the original col-
umn vector but with a zero as the kth element. That means that D[−k] Y (t) denotes
the acute intervention do(Yk (t) = 0). Again, for ease of notation we will drop the
[−k] notation and leave it implied, that is, in the proof below, D = D[−k] unless
otherwise specified.
Using this matrix-representation of the do operator, we can express the ex-
pected value of Y (t + 2s) given that we have intervened to set the current value of
the mediator (do(Yk (t + 2s) = 0)), the value of the mediator one “moment” previ-
ously (do(Yk (t+s) = 0)), and the initial value of the mediator (do(Yk (t) = 0)) to zero.
Subsequently, since we repeat this acute intervention at every “moment” in time
in an interval, we can describe it as a continuous intervention do(Yk (t + s) = 0)
that is, an intervention that is present for all possible time points in an interval.
The expected value of Y (t + 2s) given this continuous intervention can be written
as
h i
E Y (t + 2s) do(Yk (t + s)) = 0 = D(I + As)D(I + As)DY (t)
= (DI D + DADs)2 Y (t)

Now, using the same substitutions as described in Appendix 4.C, we can express
the expected value an arbitrary time interval ∆t later, given that we intervene to
set Yk to zero at each of the limn→∞ time points in that interval. This is given by
h i n o
n
E Y (t + ∆t) do(Yk (t + ∆t)) = 0 = limn→∞ (DI D + DAD ∆t n ) Y (t)

Noting that D is an idempotent matrix, and that DI D = DI = I D, we can simplify

this expression to
h i n o
n
E Y (t + ∆t) do(Yk (t + ∆t)) = 0 = limn→∞ (I + DAD ∆t
n ) DY (t)

which, by the definition of the matrix exponential function simplifies to

h i
E Y (t + ∆t) do(Yk (t + ∆t)) = 0 = eDAD∆t DY (t) (4.24)

where DY (t) ensures that the initial value of Yk (t) is set to zero.
Pre- and post-multiplying A by D[−k] has the effect of setting the kth row
and column of A to zero. Hence, the expression eDAD∆t is exactly equivalent
to the path-tracing definition of the direct effect given in Appendix 4.D, that is,

110
 4.E. Interventions and Path-Tracing in CT models

DAD = A(D[−k]) . This implies that by plugging the above expression in for the
expected values in the direct effect definition, we obtain

DEij·k (∆t) = eDAD∆t [ji] × (yi − yi∗ )

(D[−k]) ∆t
= eA [ji] × (yi − yi∗ ) (4.25)

which shows that the effect on Yj (t + ∆t) of an acute intervention to change Yi (t)
combined with a continuous intervention to keep the mediator Yk fixed is identi-
cal to the path-tracing direct effect.

4.E.3 Indirect Effect

It follows from the equivalence between path-tracing and intervention-based di-
rect and total effects that the indirect effect, defined as a contrast between those
two, can be calculated by taking the different in path-tracing definitions of each
component effect, described in Appendix 4.D.

111
 Chapter 5

Recovering Bistable Systems

from Psychological Time
Series
Abstract

Conceptualizing psychopathologies as complex dynamical systems has

become a popular framework to study mental disorders. Especially
bistable dynamical systems have received much attention, because their
properties map well onto many characteristics of mental disorders. While
these models were so far mostly used as stylized toy models, the recent
surge in psychological time series data promises the ability to recover such
models from data. In this paper we investigate how well popular (e.g., the
Vector Autoregressive model) and more advanced (e.g., differential equa-
tion estimation) data analytic tools are suited to recover bistable dynami-
cal systems from time series. Using a simulated high-frequency time series
(measurement every six seconds) as an ideal case we show that while it is
possible to recover global dynamics (e.g., position of fixed points, tran-
sition probabilities) it is difficult to recover the microdynamics (i.e., mo-
ment to moment interactions) of a bistable system. Repeating all analyses
with a sampling frequency typical for Experience Sampling Method stud-
ies (measurement every 90 minutes) showed that the recovery of the global
dynamics was still successful, but no microdynamics could be recovered.
These results raise two fundamental issues involved in studying mental
disorders from a complex systems perspective: first, it is generally unclear
what to conclude from a statistical model about an underlying complex
systems model; and second, if the sampling frequency is too low, it is im-
possible to recover microdynamics. In response to these results we propose
a new modeling strategy based on substantively plausible dynamical sys-
tems models.

This chapter has been adapted from: Haslbeck, J. M. B.* & Ryan, O.* (under review). Recovering
Bistable Systems from Psychological Time Series. Pre-print: https://psyarxiv.com/kcv3s/. Both
JMBH and OR are considered joint first authors, with both contributing equally to this project.

113
5. Recovering Bistable Systems from Psychological Time Series

5.1 Introduction
Conceptualizing psychopathologies as complex dynamical systems has become a
popular framework to study mental disorders (e.g., Wichers, Wigman, & Myin-
Germeys, 2015; Cramer et al., 2016; Borsboom, 2017). This framework is attrac-
tive because it acknowledges the fact that many mental disorders are massively
multifactorial (e.g., Kendler, 2019), and because it allows one to specify powerful
within-person dynamical systems models that capture many of the characteris-
tics hypothesized for mental disorders. The central goal of this framework is to
obtain such models to further our understanding of mental disorders, and allow
us to develop and test more successful interventions.
The class of dynamic systems that has received most attention in this emerg-
ing literature is the class of bistable systems (e.g., Wichers et al., 2015; Cramer et
al., 2016; Borsboom, 2017; Wichers, Schreuder, Goekoop, & Groen, 2019; van de
Leemput et al., 2014; Nelson et al., 2017; R. Kalisch et al., 2019). The reason is
that its behavior maps well on many phenomena observed in mental disorders:
Bistable systems describe variables that have two stable states, which can be in-
terpreted as different psychological states such as “healthy” or “unhealthy” (e.g.,
depressed). The stability landscape reflecting the dynamics of the system deter-
mines how easy it is to transition from one state to the other, and thereby offers
a possible formalization of properties of the mental disorder, such as vulnera-
bility or resilience to developing it (Scheffer et al., 2018). Bistable systems can
also show sudden transitions from one state to another, thereby mapping well
on, for example, bipolar disorder or the phenomenon of sudden gains and losses
in psychotherapy (Stiles et al., 2003; Lutz et al., 2013).
In parallel, the realization that inferences from between-subjects data to
within-person data are only possible under stringent assumptions (Molenaar,
2004; Hamaker, 2012) together with the increasing availability of psychologi-
cal time series collected from mobile devices has led to a surge in studies aiming
at recovering the within-person dynamics associated with mental disorders (e.g.,
Bringmann et al., 2013; Pe et al., 2015; A. J. Fisher et al., 2017; Snippe et al.,
2017; Groen et al., 2019). This is an exciting development, because within-person
time series potentially allow one to recover bistable systems and other dynamical
systems from empirical data. This would be a major step forward for studying
mental disorders as complex systems, because so far these models were only used
as stylized toy models.
However, so far there has been no systematic investigation into to what extent
dynamical systems models can in fact be recovered from psychological time se-
ries. To investigate this question for a given dynamical system and data analytic
method, it has to be broken down into two parts. The first question is whether
the method at hand can recover (some aspect of) a dynamical system in principle,
that is, with “ideal” data (long time series, extremely high sampling frequency).
If this is the case, the second question is whether the method also works with re-
alistic data (shorter time series, much lower sampling frequency). In the present
paper, we investigate both questions for a bistable dynamical system and a selec-
tion of the most popular (e.g., the Vector Autoregressive (VAR) model; Hamilton,

114
 5.1. Introduction

1994) and some more advanced (e.g., differential equation estimation; Boker, De-
boeck, et al., 2010) methods. Specifically, we use a basic bistable dynamical sys-
tem for emotion dynamics to simulate both an ideal time series with extremely
high sampling frequency (measurement every six seconds) and a more realistic
time series with a sampling frequency common for Experience Sampling Method
(ESM) studies (measurement every 90 minutes). Using these time series, we eval-
uate how useful each method is for recovering bistable dynamical systems in
principle, and how useful it can be in practice when analyzing realistic ESM time
series.
We will show that the popular VAR model (and the Gaussian Graphical Model
fitted on its residuals; Epskamp, Waldorp, et al., 2018) is in principle unable
to recover the global dynamics (e.g., location and variance of stable states, fre-
quency of transitions) and succeeds only in recovering some of the microdynam-
ics (moment-to-moment interactions) of the true bistable system. However, de-
scriptive statistics, data visualization and more flexible statistical models are able
to capture the global dynamics. The only method that recovered the complete
bistable system is an iterative model building procedure that directly estimates
the system of differential equations (DEs). Reducing the sampling frequency
from every six seconds to every 90 minutes affects the considered methods differ-
ently. The VAR model and its extensions no longer recover any microdynamics,
and the DE-estimation procedure fails. However, descriptives, data visualization
and appropriate statistical models still recover the global dynamics. These re-
sults raise two fundamental issues involved in studying mental disorders from
a complex systems perspective. First, it is generally unclear what to conclude
from a statistical model about an underlying complex systems model. Second,
if the sampling frequency is too low, it is impossible to recover microdynamics.
In response to these findings, we outline a different research strategy to arrive at
dynamical systems models for mental disorders: Proposing initial formal models
which can subsequently be scrutinized and developed by deriving data implica-
tions that can be tested empirically. We will show that in this process many of the
presented methods are instrumental to testing predictions of the formal model
and thereby triangulating the formal model that captures the true dynamical sys-
tem best.
Our paper is structured as follows. In Section 5.2 we introduce a simple
bistable dynamical system for emotion dynamics, discuss its dynamics and char-
acteristics, and describe how we generate the ideal and the more realistic time
series from it. We use the ideal data (measurement every six seconds) in Section
5.3 to evaluate for each method to which extent it can recover a bistable dynam-
ical system. Next, in Section 5.4 we evaluate the same methods but using the
time series with a sampling frequency that matches typical ESM studies (mea-
surement every 90min). Finally, in Section 5.5 we discuss the implications of our
results for the framework of empirically studying mental disorders from a com-
plex systems perspective, and outline a new research strategy based on formal
modeling, which avoids shortcomings of a purely data analytic approach.

115
5. Recovering Bistable Systems from Psychological Time Series

5.2 Bistable Emotion System as Data-Generating

Model
In this section we present a bistable dynamical system and describe its dynamics
(Section 5.2.1), show how we generate data from this system (Section 5.2.2) and
discuss its qualitative characteristics (Section 5.2.3).

5.2.1 Model Specification

Bistable dynamical systems are typically formalized within the framework of dif-
ferential equations (e.g., Hirsch, Smale, & Devaney, 2012; Strogatz, 2015) and so
we too adopt this framework to describe the data-generating mechanism. Our
goal is to provide an accessible first investigation of how well bistable systems
can be recovered from psychological time series and therefore we use one of the
simplest multivariate bistable systems as a case study. Specifically, we choose a
system with four variables that is a generalization of the classic Lotka Volterra
model for competing species (e.g., H. I. Freedman, 1980) to four variables; a sim-
ilar model was used by van de Leemput et al. (2014) who interpreted the four
variables as positive and negative emotion variables, an interpretation we also
adopt here. This model is capable of exhibiting two stable states: one in which
positive emotions are high and negative emotions are low (the “healthy” state);
and one in which the positive emotions are low and the negative emotions are
high (the “unhealthy” state).
Note that different types of (bistable) dynamical systems will differ in how dif-
ficult they are to recover with a given method and type of time series, and much
research is needed to map out the space of dynamical systems model classes, data
analytic methods and types of time series. However, the intuition we rely on in
the present paper is that if there are fundamental problems in recovering one of
the simplest multivariate bistable systems, then these problems will be at least
equally severe when recovering a more complicated bistable dynamical system.
The bistable system we use throughout this paper consists of two emotions
with positive valence (Cheerful (x1 ) and Content (x2 )) and two emotions with
negative valence (Anxious (x3 ) and Sad (x4 )). The dynamics of the system is de-
fined by the stochastic differential equations
4
dxi X
= ri xi + Cij xj xi + ai + i , (5.1)
dt
j=1

where ri can be thought of as the main effect of an emotion on itself over time, that
is, the effect of xi on its own rate of change. This parameter is set to 1 for positive
emotions, and will be varied between r3 , r4 ∈ [0.9, 1.1] for negative emotions. We
interpret the variations in r3 , r4 as being related to stress: Higher stress means that
the effects of a high degree of negative emotion stays in the system longer. The
matrix C represents the dependencies between emotions in the form of interaction
effects

116
 5.2. Bistable Emotion System as Data-Generating Model

−0.2 0.04 −0.2 −0.2

 
 0.04 −0.2 −0.2 −0.2
C =  .

−0.2 −0.2 −0.2 0.04 
−0.2 −0.2 0.04 −0.2
 

The interactions in the matrix C show that emotions of the same valence re-
inforce each other, while emotions of different valence suppress each other. For
example, C12 = 0.04 indicates that the rate of change of x1 (Cheerful) depends
on the product of x1 and x2 (Content) weighted by 0.04. Similarly, C13 = −0.2
indicates that the rate of change of x1 depends on the product of x1 and x3 (Anx-
ious) weighted by −0.2. The diagonal elements are quadratic effects: For exam-
ple, C11 = −0.2 indicates that the rate of change of x1 depends on the product
x1 x1 = x12 . Note that we choose the matrix C to be symmetric purely for the sake
of simplicity. Since we aim to specify the simplest possible bistable system, we
specify that all within-valence effects (e.g., C12 and C34 ) are equal to 0.04 and all
between-valence interaction effects (e.g., C13 and C24 ) and quadratic effects (Cii )
are equal to −0.2.
We interpret xi = 0 as the absence of positive/negative emotion, and therefore
do no not allow emotions to become negative. We ensure this with high proba-
bility by setting the constant ai = 1.6 for all i. The Gaussian noise term i has a
mean of zero and a fixed standard deviation σ and represents short-term fluctu-
ations in emotions due to the environment the system interacts with. Note that
we used the same parameterization as van de Leemput et al. (2014), except that
in our model we use an additive noise term instead of a multiplicative noise term
for simplicity and set all ai = 1.6.
Due to the symmetries in C, r and a, emotions with the same valence are ex-
changeable. We can therefore describe the dynamics of the 4-dimensional system
using a 2-dimensional system consisting of one dimension for positive emotions
and one dimension for negative emotions (for details see Appendix 5.A). Figure
5.1 illustrates the dynamics of the deterministic part (i.e., with i = 0) of this
model: Panel (a) displays the stable (solid lines) and unstable (dashed lines) fixed
points for positive (green) and negative (red) emotions, as a function of stress. For
example, for a low stress level of 0.9 there is only a single fixed point: the positive
emotions (PE) have the value 5.28 and the negative emotions (NE) have the value
1.15. We therefore also refer to this fixed point as the healthy state. If the stress
level remains unchanged, the system will always end up at this fixed point, no
matter how one chooses the starting values. This dynamic is illustrated in the
corresponding vector field in panel (b). The arrows depict the partial derivatives
with respect to the two emotions and therefore describe the linearized dynamics
at a given point in the 2-d space. The vector field shows us that whichever ini-
tial values we choose, the system will always end up at the fixed point at (PE =
5.28, NE = 1.15). Thus, the system with stress = 0.9 describes a person whose
emotions can be changed by external influences, but eventually always returns to
the healthy state of having strong positive emotions and weak negative emotions.
The solid lines in panel (b) indicate the values of positive and negative emotions
for which the two differential equations are zero. At the intersections of those

117
5. Recovering Bistable Systems from Psychological Time Series

(a) Fixed Points across Stress Levels

6

5
Emotion Strength

Positive Emotions
3
Negative Emotions
2

0.9 1.0 1.1

Stress (r3, r4)

(b) Stress = 0.9 (c) Stress = 1 (d) Stress = 1.1

10 10 10
Negative Emotion

Negative Emotion

Negative Emotion
●

5 5 ● 5

● ●

0 0 0

0 5 10 0 5 10 0 5 10
Positive Emotion Positive Emotion Positive Emotion

Figure 5.1: The dynamics of the bistable system we will use as the data-generating model throughout
the paper. Panel (a) shows the fixed points of the deterministic part of the model as a function of
stress, operationalized by the rate of change of the negative emotions. Solid lines indicate stable fixed
points and dashed lines indicate unstable fixed points. Panels (b), (c) and (d) show the vector fields
of the system for the stress values r3 , r4 = 0.9, 1 and 1.1. Solid points indicate stable fixed points and
empty points indicate unstable fixed points. The solid lines indicate the values at which derivative of
positive emotion (orange) and negative emotion (light blue) is equal to zero. At the points at which
the two lines meet, both derivatives are equal to zero and the system remains in this (stable) state.

lines both differential equations are equal to zero, which means that the system
does not change anymore, which is the definition of a fixed point.
Panel (a) of Figure 5.1 shows that when increasing stress from 0.9 until around
0.95, the stable fixed point changes quantitatively: The value of positive emotion
value decreases, and the value of negative emotion value increases. However,
from around 0.95 on the dynamics of the system change qualitatively: the system
now has three fixed points. For example, at stress = 1, the fixed points are (PE =
4.89, NE = 1.36), (PE = 2.80, NE = 2.80), and (PE = 1.36, NE = 4.89). The first
fixed point is the stable healthy fixed point we also observed for values smaller
than 0.9. The second fixed point is an unstable fixed point. Specifically, it is a
saddle point, because the arrows in the vector field flow towards this fixed point
in one direction, but flow away in the other direction (Strogatz, 2015). The third
fixed point is again stable, however, now negative emotions have a high value and
positive emotions have a low value. We could call this fixed point the unhealthy

118
 5.2. Bistable Emotion System as Data-Generating Model

fixed point.
The presence of these three fixed points means that, if the system is initialized
anywhere except on the diagonal, the system will end up at one of the two stable
fixed points. This behavior is illustrated in panel (c), which shows the vector
field of the system for stress = 1. We see that eventually all arrows point away
from the unstable fixed point (PE = 2.80, NE = 2.80) and towards one of the two
stable fixed points. Thus, the system will never converge to this point except if
it is initialized exactly on the diagonal. For all other starting values, the system
will converge to one of the two stable fixed points. For the particular case of
stress = 1, starting values above the diagonal line will converge to the unhealthy
fixed point (PE = 1.36, NE = 4.89), whereas starting values below the diagonal
line will converge to the healthy fixed point (PE = 4.89, NE = 1.36). This system
describes a person that starts out in the healthy (unhealthy) state, and always
returns to the healthy (unhealthy) state after small outside influences. However,
a large influence can push the person into the unhealthy (healthy) state, and now
the person remains there until a large enough influence pushes her back into the
healthy (unhealthy) state.
When increasing stress further until around 1.06, we observe again a quan-
titative change of the three fixed points: the negative emotions go up, and the
positive emotions go down. However, from around 1.06 on the system changes
again qualitatively. It now again exhibits only one fixed point, which is now the
unhealthy fixed point. Thus, when stress is larger than around 1.06, the system
will always converge to the unhealthy fixed point. This behavior is illustrated in
panel (d), which depicts the vector field for the system with stress = 1.1. We see
that there is only a single fixed point at (PE = 1.03, NE = 5.98) and the arrows
show that the system will always converge to this point. This system describes a
person that will always return to the unhealthy state, no matter how large of an
outside influence is applied.
So far, we only discussed the deterministic part of the model, that is, our
model with noise set to zero (i.e., with i = 0). Introducing noise changes the dy-
namics of the system, and how exactly it changes depends on the stress level. For
low stress (below 0.95), the system will fluctuate around the healthy fixed point.
For high stress (above 1.06), the system will fluctuate around the unhealthy fixed
point. The interesting behavior is observed for stress values between 0.95 and
1.06: then, the system will fluctuate around one of the two fixed points, but oc-
casionally the noise will be large enough to push the system to the other fixed
point. The frequency of switching is a function of the distance between the two
fixed points, the vector field between the two fixed points, and the variance of
the Gaussian noise process i . If the variance is low, the probability of a noise
draw that is large enough to “push” the system to the other fixed point is small,
and consequently the frequency of switching is low. In contrast, if the variance
is high, the probability of a large enough noise draw to switch to the other fixed
point is high, and consequently the switching frequency is high.

119
5. Recovering Bistable Systems from Psychological Time Series

5.2.2 Generating Time Series from Bistable System

In the previous section we have shown that our dynamical system is bistable
for stress values (r3 , r4 ) ∈ [0.96, 1.06]. In the remainder of this manuscript we
keep stress constant at stress = 1, and therefore study the bistable dynamical
system with the dynamics displayed in panel (c) of Figure 5.1 and the fixed points
described above. Apart from stress we chose all parameters as indicated in the
previous section.
To obtain a plausible switching frequency for emotion dynamics we set the
standard deviation of the Gaussian noise term σ = 4.5. Note that a system can be
bistable, but the outside influences (the noise term) are so weak that the system
switches very infrequently or not at all. In such cases the bistable system is more
difficult (infrequent) or impossible (no switches) to recover. Thus, our choice of σ
represents an ideal situation, and all presented methods will perform worse with
a lower switching frequency.
In the remainder of this section we describe how we generated the two time
series that we will use throughout the paper. First, we generate an “ideal” time
series with an extremely high sampling frequency of 1 measurement every six
seconds (Section 5.2.2.1). Second, we generate a more realistic time series with
measurements every 90 minutes, a sampling frequency typical for ESM studies
(Section 5.2.2.2).

5.2.2.1 Ideal Time Series

We generated data by computing the numerical solution to the model in Section
5.2.1 with stress = 1 on the interval [0, 20160], using Euler’s method (e.g., Atkin-
son, 1989) with a step size of 0.01. We interpret a time step of 1 as one minute,
and therefore the time series spans two weeks (60 × 24 × 14 = 20160). We obtain
a time series by sampling the numerical solution obtained via Euler’s method 10
times per minute (or every six seconds). We therefore obtain the ideal time series
with 20160 × 10 = 201600 measurements, which appears to switch between fixed
points around 17 times.1 Figure 5.2 displays this time series.
We choose this unrealistically ideal time series (two weeks, one measurement
every six seconds, continuous response scale, no measurement error or missing
values) with 201600 measurements to be able to study the usefulness of different
data analytic methods in principle. That is, we study the usefulness of all methods
on the population level, which is the situation in which we have infinitely many
observations and sampling variation does not exist. With 201600 observations,
in this setting we approximate “infinitely many” for all practical purposes.
Note that we would not be able to investigate how well different methods
perform in principle, if we made the time series more realistic by choosing a
shorter time interval or sampling it with a lower sampling frequency. In such
a case we would not know whether a method cannot recover (an aspect of) the
bistable system for fundamental reasons, or because the time series is too short
1 The code to generate data and reproduce all analyses and results shown in this paper can be
found at https://github.com/jmbh/RecoveringBistableSystems/.

120
 5.2. Bistable Emotion System as Data-Generating Model

(a) (b)
8 Cheerful
Content
Anxious
Emotion Strength

Sad
6

0 5 10 14
Days

Figure 5.2: Panel (a) shows the ideal time series of the four emotion variables Cheerful, Content,
Anxious and Sad. We see that the system switches 17 times between healthy and unhealthy state.
Panel (b) displays the twelfth switch, which is a transition from the unhealthy to the healthy state,
which occurs on day 9.

or the sampling frequency too low. We therefore first study all methods with the
ideal time series in order to identify their fundamental limitations. In the second
part of the paper, we make the time series more realistic by taking measurements
at a sampling frequency that is typical for ESM studies. This will allow us to
investigate the impact of sampling frequency on all methods. In the following
section we describe how we generate this ESM time series.

5.2.2.2 Experience Sampling (ESM) Time Series

Clearly, the ideal time series is very different from time series data sets obtained
from typical ESM studies. The perhaps two most important differences between
the ideal time series and realistic time series are the measurement scale and the
sampling frequency. With respect to the measurement scale, most ESM studies
do not use a continuous response but, for example, a 7-point Likert scale. Regard-
ing the sampling frequency, ESM studies investigating psychological variables
typically do not measure more frequently than every 90 minutes (e.g., Bring-
mann et al., 2013; Pe et al., 2015; A. J. Fisher et al., 2017; Snippe et al., 2017;
Groen et al., 2019). Thus, ESM time series have a much lower sampling frequency
(every 90 minutes) than the ideal time series used above (every six seconds). To be
able to explain possible drops in performance of certain methods when making
the time series more realistic, we must only change one aspect of the time se-
ries. While the measurement scale can possibly be made near-continuous, there
are certainly hard limits on how many times one can notify a person each day
with an ESM questionnaire. We therefore consider the sampling frequency the
more fundamental constraint in realistic data, and thus make it the focus of our
investigation in Section 5.4.
Taking a measurement every 90 minutes in the two weeks of the original data

121
5. Recovering Bistable Systems from Psychological Time Series

yields 224 measurements. This would mean that we would compare the “ideal”
time series with 201600 measurements which essentially implies the absence of
sampling variation to an ESM time series with 224 measurement which implies
quite a considerable degree of sampling variation. Thus, any comparison would
be confounded by the difference in the number of measurement points (i.e. sam-
ple size). To avoid this confound, we increase the measurement interval of the
ESM time series to 1800 weeks, which ensures that the new ESM time series has
exactly the same sample size as the ideal data ( 2242 × 1800 = 201600). Thereby,
we ensure that any drop in performance is a function of the lowered sampling
frequency and cannot be explained by lower sample size (and higher sampling
variation). Note that studying the performance of methods as a function of sam-
ple size (sampling variation) is of paramount importance to evaluate how useful a
given method is in a realistic application. However, here we study the more fun-
damental question of the impact of reducing the sampling frequency to a level
that is typical for ESM studies. We do this because if we find that a method is
ill-suited to recover (some aspect of) a bistable system with a realistic sampling
frequency on the population level (i.e., with infinite sample size), then it does
not make sense to investigate the performance of the method in the less ideal
scenario with realistic (small) sample sizes.
So far, we only discussed that we sample every 90 minutes. However, to em-
ulate ESM measurements, we also need to formalize how exactly ESM questions
measure the four emotion variables. This is far from trivial: questions in some
ESM studies refer to the very moment of measurement and are phrased along
the lines of “How cheerful do you feel right now?”. Such measurements could
be formalized by defining the measurement as the set of current values of the
system (a “snapshot” of the system) at the measurement time. In contrast, other
ESM studies refer to the time period since the last measurement. A question of
this type could be phrased “How cheerful did you feel in the time since the last
notification?”. Such measurements could be formalized by defining the measure-
ment as the average values of the system since the last measurement. However,
many other measurement functions are also possible. In this paper we analyze
the first type of ESM question, because its measurement function is the simplest.
However, we also performed all analyses with the second kind of ESM question,
and all our main conclusions also hold in this situation.
Figure 5.3 displays the two week long original time series (see also Figure 5.2
panel (a)) next to the ESM time series which was obtained by taking “snapshots”
of the process at 90 min intervals. The ESM time series in panel (b) appears
less dense, which is what we would expect since it contains only 1/900 of the
time points of the ideal time series in panel (a). However, we see that the system
is still bistable and that the location of and variance around the fixed points is
largely the same. In Section 5.4 we will use this emulated “snapshot” ESM time
series to try to recover the true bistable system using the same array of methods
as in Section 5.3, in which we analyze the ideal time series.

122
 5.2. Bistable Emotion System as Data-Generating Model

(a) Original time series (b) ESM time series

8 8
Emotion Strength

Emotion Strength
6 6

4 4

2 2

0 0

0 5 10 14 0 5 10 14
Days Days

Figure 5.3: Panel (a) shows the original time series that was already shown in panel (a) of Figure 5.2.
Panel (b) shows the ESM time series which was obtained by taking snapshots every 90 minutes in the
series. Note that the ESM time series we analyze in Section 5.4 is much longer (1800 weeks) than the
14 day ESM time series shown here.

5.2.3 Qualitative Characteristics of the Model

In this section we discuss the key qualitative characteristics of the bistable system
introduced in the previous section. We list these characteristics because most
considered methods are models that are misspecified (i.e. they do not contain the
true system as a special case). In such a situation one can only hope to recover
some characteristics of the true system, and we therefore evaluate how well a
method recovers the bistable system based on how well it recovers the following
seven characteristics:

Global dynamics
1. Bistability (two stable fixed points)

2. Position of fixed points

3. Variability around fixed points
4. Frequency of transitions
Microdynamics

5. Suppressing effects between valences, reinforcing effects within valences

6. Relative size of suppressing/reinforcing effects
7. All parameters are independent of time and independent of variables out-
side the model

123
5. Recovering Bistable Systems from Psychological Time Series

The first four characteristics describe the global dynamics of the dynamical
system. The first is bistability, which means that the data-generating mechanism
exhibits two stable fixed points. This is the case for the data-generating mecha-
nism with stress set to 1, which we use to generate data from and aim to recover
throughout the paper (see Figure 5.2, panels (a) and (c)). The second character-
istic is the position of the fixed points, which are at (PE = 4.89, NE= 1.36) for
the healthy fixed point, and (PE = 1.36, NE= 4.89) for the unhealthy fixed point.
Third, we consider the variability around the different fixed points. Figure 5.2
shows that, for both fixed points, the variability of the emotions with lower val-
ues is smaller than the variability of the emotions with larger values. The fourth
characteristic is the frequency of transitions between the area around the healthy
fixed point and the area around the unhealthy fixed point. In the time series
shown in Figure 5.2 we see that the system switches around 17 times.
The remaining three characteristics describe the microdynamics of the dy-
namical system. The fifth characteristic is that emotions of the same valence re-
inforce each other, while emotions of different valence suppress each other. The
sixth characteristic is the fact that the size (absolute value) of the reinforcing ef-
fects (0.04) are smaller than the suppressing effects (0.2). The last characteristic
is that all parameters in the system of differential equations are independent of
time and independent of variables outside the model.

5.3 Recovering the Bistable System from Ideal Data

In this section we analyze the ideal time series to evaluate how well different
methods recover the data-generating bistable system. The methods considered
here primarily consist of the most popular models used in analyzing time series
in clinical psychology and psychiatry, and some extensions thereof. In all but
one instance, this entails the estimation of misspecified models, that is, models
which do not contain the true bistable system as a special case. Thus we will
focus our investigation on whether these models allow one to recover some of the
characteristics of the true model, as outlined in Section 5.2.3.
We analyze each method in order of increasing complexity, moving from
methods which may be helpful in recovering global characteristics alone to meth-
ods which are typically used with the aim of characterising the microdynamic
structure. We begin by inspecting the time series using descriptive statistics (Sec-
tion 5.3.1); in Section 5.3.2 we characterize the switching behavior in the system
using a mean-switching Hidden Markov Model (Hosenfeld et al., 2015; Hamaker,
Grasman, & Kamphuis, 2016). Next, in Section 5.3.3 we analyze the multivari-
ate lag-0 (same time point) relationships using correlations and partial correla-
tions with the popular Gaussian Graphical Model (GGM) (Epskamp, Waldorp, et
al., 2018). In Section 5.3.4 we use the most popular approach to modeling mi-
crodynamics in experience sampling settings, the lag-1 Vector Auto-Regressive
(VAR(1)) model (e.g. Bringmann et al., 2013; Pe et al., 2015; A. J. Fisher et
al., 2017; Groen et al., 2019). Next, in Section 5.3.5 we evaluate the Threshold
VAR model, an extension of the VAR model that allows the modeling of state-

124
 5.3. Recovering the Bistable System from Ideal Data

switching behavior using time-varying parameters (Hamaker, Zhang, & van der
Maas, 2009; Hamaker & Grasman, 2012). While all models so far are misspeci-
fied, we include one final method that is capable of recovering the full bistable
system: a two-step model building approach based on direct estimation of differ-
ential equations from data, following the dynamic systems modeling approach
of Boker, Deboeck, et al. (2010) and Chow (2019).

5.3.1 Descriptive Statistics

To get a rough overview of the behavior of the system, we inspect the time series
plot of all four emotion variables shown above in Figure 5.2 panel (a). We see that
at almost every time point the two positive emotion variables have high values
around 5, and the two negative emotion variables have small values around 1,
or the other way around. At the remaining time points, the variables seem to
transition between those two states (see panel (b) in Figure 5.2). In addition, we
see that the variables switch between states 17 times.
We can extract a considerable amount of information from simply inspecting
the time series plot. There seem to be two stable states (fixed points), one in which
positive emotions are high and negative emotions are low, and one in which the
reverse is true. Further, we see that the variance is higher for the emotion with
higher values, and we saw that the system switches around 17 times in the two
week window. To get a more direct picture of the distribution of variables and
possible fixed points we show the histograms for each variable in Figure 5.4:

Cheerful Content Anxious Sad

0.6 0.6 0.6 0.6
0.5 0.5 0.5 0.5
Density

Density

Density

Density

0.4 0.4 0.4 0.4

0.3 0.3 0.3 0.3
0.2 0.2 0.2 0.2
0.1 0.1 0.1 0.1
0.0 0.0 0.0 0.0

0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8

Emotion Intensity Emotion Intensity Emotion Intensity Emotion Intensity

Figure 5.4: The histograms of the emotion intensity of the four modeled emotions Cheerful, Content,
Anxious and Sad, for the ideal data.

We see that at most time points in the time series, each emotion either takes
on values around 1 or around 5. This is what we would expect from inspecting
the time series plot, however, the histograms give a more precise picture of the
distributions and allow one to guess possible fixed points with greater precision.
For instance, we could separate the two distributions (using a fixed threshold, or
clustering algorithm) and take their means as estimates for the fixed points.
While eyeballing the data should be the first step in any time series analysis,
the conclusions are subjective and do not allow us to quantify how certain we are
about bistability and the switching frequency. We can quantify the observation
that there are two states and that the system is switching between them by fitting
a Hidden Markov Model (HMM) (e.g., Rabiner, 1989) to the data, which we will

125
5. Recovering Bistable Systems from Psychological Time Series

do in the following section. Such quantification is especially valuable in more

realistic situations, in which the two states are probably harder to separate than
in our ideal simulated data.

5.3.2 Hidden Markov Model

In this section we fit a mean-switching Hidden Markov Model (HMM) in order to
scrutinize the intuition that the system switches between two states and to quan-
tify the switching frequency. The HMM models the observed data as consisting of
K latent states or components, characterized by K multivariate Gaussian distri-
butions, which may differ in their means µk and variances σk .2 Each observation
over time is drawn from one or other of these distributions, and the switching
between these states is governed by a matrix of transition probabilities A. For
more details about this model see Appendix 5.B.1.
Here we choose K = 2 components, and fit this model to our time series using
the R-package depmixS4 (Visser & Speekenbrink, 2010), obtaining the following
parameter estimates

 1.47   0.41   4.75   0.63 

       
!
 1.46   0.40   4.76   0.62  0.9996 0.0004
µ̂1 =  , σ̂ =
 0.63 , µ̂2 =  1.45 , σ̂2 =  0.40 , Â = .
 4.71  1
     
      0.0004 0.9996
 
4.71 0.62 1.45 0.40

We can see from the estimate µ̂1 that in state 1 the means of positive emotions
are low, and the means of negative emotions are high. We can therefore identify
state 1 as the unhealthy state. We also see that the standard deviations of positive
emotions are lower than for negative emotions in the unhealthy state which is
what we already observed in the time series plot in Figure 5.2. Similarly, state 2
can be identified as the healthy state, with high means and standard deviations
for positive emotions, and low for negative emotions. The transition matrix A
indicates the probabilities of switching between states. We see that there is a very
high probability for remaining in the same state (Â11 = Â22 = 0.9996), and a cor-
respondingly low probability to switch states (Â12 = Â21 = 0.0004). This is what
we would expect, because we take one measurement every six seconds, but the
system changes states only a couple of times within the two week window. Multi-
plying the number of time points of the time series with the switching probability
we obtain 201600 × .0004 ≈ 81 switches, which is in the same order of magnitude
of the eyeballed number of switches (17) reported in Section 5.2.2.
In addition to obtaining estimates of means and standard deviations of the
two fixed points and the transition matrix, the HMM allows to predict the most
likely state for each time point. We show the predicted states for the entire time
series in Figure 5.5:
2 In principle, the distributions may also differ with respective to their covariances, but in this
analysis, we set all covariances to zero due to limitations of the software package used in estimation.

126
 5.3. Recovering the Bistable System from Ideal Data

State 1 State 2

8
Emotion Strength

6
Cheerful
Content
Anxious
4 Sad

0 5 10 14
Days

Figure 5.5: Time series of the four emotion variables, also shown in panel (a) of Figure 5.2, with
background color indicating whether a given time point is assigned to the first or second component
of the mean-switching HMM.

When inspecting the predicted states visually, it seems that the HMM cap-
tured the switches well. Next to the larger blocks in which the system stays at the
same fixed point, it also identifies switches in which the system switches back
and forth within only a few time points. These switches might have been missed
when inspecting the time series visually alone.
Taking all results together, which characteristics of the bistable system did
we recover with the HMM (based on the list in Section 5.2.3)? We obtained an
estimate of the location (characteristic 2) and variance (characteristic 3) around
two fixed points, which are very close to the healthy and unhealthy fixed points
in the true bistable system. We also quantified the frequency of transitions in
the transition matrix A. Since the transition frequency (characteristic 4) is not
explicit in the true bistable model, there is no clear way to evaluate this estimate.
However, the number of predicted transitions (81) is at least of the same order of
magnitude as the number of transitions eyeballed from the entire time series (at
least 17). Note that while a bistable HMM seems to fit the data well, we provided
K = 2 as an input to the model, and therefore bistability (characteristic 1) cannot
be considered a characteristic we recovered with this model. Instead of fixing
a particular K, an optimal K can be obtained via model selection. However, in
Appendix 5.B.2 we show that at least the standard approach to selecting K in
mean-switching HMMs performs poorly since the data was not generated from a
mean-switching HMM.
One additional way to visualize or ascertain how much of the true systems
behavior a given model is able to capture is by generating new data from the
estimated model parameters. In Figure 5.16 in Appendix 5.C.1 we generate a two
week time series from the estimated mean-switching HMM and compare it to the
original time series. We find that the data generated from the HMM is similar
to the original data, except for two features: First, the system tends to switch
between states somewhat more frequently, and second, there are no observations
on the transitions between states.

127
5. Recovering Bistable Systems from Psychological Time Series

The remaining three characteristics (5-7) are about the microdynamics of the
true bistable system, that is, about how the components are related to each other.
Clearly, the mean-switching HMM we used here cannot elucidate these charac-
teristics since it does not model any dependencies between the four emotion vari-
ables. In the following sections we fit models that include such dependencies.

5.3.3 Lag-0 Relationships / Gaussian distribution

In this section we analyze the relationships between variables at the same time
point. Figure 5.6 panel (a) displays the relationship between Content and Cheer-
ful, two emotions of the same valence. We see that the observations cluster
around two points, one close to (1, 1) with smaller variance, and one close to
(5, 5) with larger variance. The red line indicates the best fitting regression line
(correlation 0.98).

8 (a) 8 (b)

6 6
Anxious
Content

4 4

2 ρ = 0.98 2 ρ = −0.97

0 0

0 2 4 6 8 0 2 4 6 8
Cheerful Cheerful

Correlation Network (c) Partial Correlation Network (d)

Cheerful 0.98 Content Cheerful 0.51 Content

−0.97 −0.24

−0.97 −0.97 −0.24 −0.25

−0.97 −0.24
Anxious 0.98 Sad Anxious 0.51 Sad

Figure 5.6: Panel (a) shows the relationship between Content and Cheerful, two emotions with the
same valence, at the same time point. The red line indicates the best fitting regression model. Sim-
ilarly, panel (b) shows the relationship between Anxious and Content, two emotions with different
valence. Panel (c) displays the correlation matrix as a network, and panel (d) displays the partial
correlation matrix as a network.

128
 5.3. Recovering the Bistable System from Ideal Data

Panel (b) displays the relationship between Cheerful and Anxious, two emo-
tions of different valence. We see that that the observations cluster around two
points, one close to (1, 5) and the other one close to (5,1). The red line indicates
the best fitting regression line (correlation ρ = −0.97).
Panel (c) displays the correlation network for all four emotion variables. As
we have already seen in panel (a) and (b) there is a positive correlation (ρ = 0.98)
between Content and Cheerful, and a negative correlation (ρ = −0.97) between
Cheerful and Anxious. Due to the symmetry in the true bistable system, all cor-
relations between emotions with the same valence are equal to ρ = 0.98 and all
correlations between emotions with different valences are equal to ρ = −0.97.
Panel (d) shows the partial correlation network (i.e., GGM). We see that the par-
tial correlations between emotions with the same valence are equal to θ = 0.51,
and the partial correlations between emotions with different valences is equal to
θ = −0.24 or θ = −0.25.
What can we learn from these results about the underlying bistable system?
From inspecting the pairwise relationships of emotions with same and different
valence in panels (a) and (b) one could guess the location and variance of pos-
sible fixed points, similarly to inspecting the histograms in Section 5.3.1. How-
ever, the 2-dimensional representation offers additional information about the
stability landscape, for example the shape around the fixed points and the most
likely paths to transition between them. When interpreting the correlations in
panel (b) as “contemporaneous” relationships, we would conclude that there are
strong positive linear relationships between emotions with the same valence, and
similarly strong negative linear relationships between variables with different va-
lences at a relatively short time scale. The partial correlations in (d) are smaller
than the correlations, which is what one would expect since all correlations are
high.
Using our knowledge about the true bistable system, which characteristics
did we correctly recover? From inspecting the scatter plots in panels (a) and (b)
one sees that most observations fall in one of two clusters indicating bistability
(characteristic 1). Also, one can obtain rough estimates of the position of the
fixed points (characteristic 2) and sees that the variances around the fixed points
is different (characteristic 3). Note that the shape of the scatter plot in panel (b)
is determined by the vector field in Figure 5.1 (c). The two clusters are exactly at
the location of the two fixed points, and the observations between the clusters are
both due to variance around the fixed points and switches between fixed points.
From the correlation and partial correlation networks, we correctly find that
there are reinforcing effects within valences, and suppressing effects between va-
lences (characteristic 5). However, the correlation network suggests that their
relative size is equal, and the partial correlation network suggests that the rein-
forcing effects are stronger. In the true bistable system, however, the suppressing
effects between valences are larger than the reinforcing effects within valences.
Thus, judging the relative size of suppressing/reinforcing effects within/between
valences from (partial) correlation would lead to incorrect conclusions.
In sum, inspecting scatter plots of pairwise relationships indicated bistability,
and allowed us to obtain a rough estimate of the location of and variances around

129
5. Recovering Bistable Systems from Psychological Time Series

the fixed points. The scatter plots also allowed one to get a projection of the
stability landscape on two dimensions and thereby provided more information
than histograms. While inspecting the scatter plots allows one to recover global
dynamics of the true bistable system, one cannot infer the coupling between the
emotion variables in the true bistable system from (partial) correlations. This
is not too surprising since the Gaussian distribution is very restrictive in that
it only models pairwise linear relationships (opposed to e.g., 3-way, 4-way, etc.
interactions). In addition, it does not model any dependencies across time, which
are the types of dependencies that constitute the microdynamics (characteristics
5-7) of the true model. In the next section, we inspect those dependencies across
time and fit a Vector Autoregressive (VAR) model to the data, which captures
temporal linear dependencies.

5.3.4 Lag-1 Relationships / VAR Model

In this section we aim to characterize the microdynamics between the four emo-
tion variables by modeling the lagged relationships between them. Panels (a) and
(b) of Figure 5.7 show the marginal relationship of Content at time t with Cheerful
at the previous time point t − 1, and Anxious at time t with Content at t − 1, re-
spectively. Although the data is generated from a dynamic model, the marginal
lagged relationships look similar to the contemporaneous relationships shown
in Figure 5.6: averaging over all other variables, lagged relationships within-
valence are positive, and between-valence are negative. The reason is that the
system largely stays around the two stable fixed points, and relative to the length
of the time series, switches are infrequent. As such, the marginal relationships
are largely driven by fixed point locations.
We can gain further insight into the dependencies between variables in our
model by examining the conditional lagged relationships between pairs of emo-
tions, that is, when keeping the other emotion variables at the previous time
point(s) fixed. A popular model for such conditional lagged relationships is the
first-order vector auto-regressive (VAR(1)) model. The VAR(1) model is one of
the simplest multivariate dynamic models which can be fit to repeated measure-
ment data, allowing linear relationships between all pairs of variables observed
at consecutive measurement occasions t and t − 1, that is

Xt = b + ΦXt−1 + et (5.2)
where b is a vector of intercepts, Φ is a matrix containing the auto-regressive (φii )
and cross-lagged (φij , i , j) effects, that is, conditional linear dependencies, and et
is a vector of normally distributed residuals et ∼ N (0, Ψ ), which are independent
across time, with residual variance-covariance matrix Ψ .
The VAR(1) model has been used widely to analyze experience sampling data
in psychopathology research, particularly in the form of dynamic network analy-
sis, wherein the Φ and Ψ matrices are used to construct directed and undirected
network structures, respectively (e.g., Bringmann et al., 2013; Pe et al., 2015;
Epskamp, Waldorp, et al., 2018). The VAR(1) model describes a system which
fluctuates around a single stable fixed point: Stochastic input in the form of a

130
 5.3. Recovering the Bistable System from Ideal Data

residual term pushes the system away from this fixed point, and the system re-
turns to the fixed point with an exponential decay (Hamilton, 1994). The location
of the stable fixed point is given by the mean vector µ, a function of the intercepts
and the lagged relationships µ = (I − Φ)−1 b, where I is the identity matrix.

8 (a) 8 (b)

6 6

Anxioust
Contentt

4 4

2 2

0 0

0 2 4 6 8 0 2 4 6 8
Cheerfult 1 Contentt 1

Lag(1) Network (c) Residual Network (d)

0.91 0.91

0.05
Cheerful Content Cheerful 0.04 Content
0.05
−0.04

−0.02 −0.02
−0.02 −0.02 −0.02 −0.02
−0.02 −0.03 −0.03
−0.02

−0.03
0.05
Anxious Sad Anxious 0.04 Sad
0.05

0.91 0.91

Figure 5.7: Panel (a) shows the relationship between Content and Cheerful, two emotions with the
same valence, spaced one time point apart (at a lag of one). The red line indicates the best fitting
regression model. Similarly, panel (b) shows the relationship between Anxious and Content, two
emotions with different valence, at a lag of one. Panel (c) displays the matrix of lagged regression pa-
rameters, estimated from a VAR(1) model, as a network, and panel (d) displays the partial correlation
matrix of the residuals of the VAR(1) model as a network. This latter network is often referred to as
the contemporaneous network.

Panel (c) of Figure 5.7 displays the network of estimated lagged regression co-
efficients (Φ̂) between Cheerful, Content, Anxious and Sad. We can see that the
auto-regressive parameters are large and positive for all four variables (φ̂ii = .91).
Furthermore, there are positive cross-lagged relationships between variables of
the same valence (φ̂12 = φ̂21 = φ̂43 = φ̂34 = .05) and weaker, negative cross-
lagged effects between variables of opposite valence (φ̂13 = φ̂31 = · · · = −.02).
All within-valence effects, and all between-valence effects, are of roughly equal

131
5. Recovering Bistable Systems from Psychological Time Series

magnitude, respectively. In panel (d) we show the partial correlations of the

residuals (i.e., standardized Ψ̂ −1 ), sometimes referred to as the “contemporane-
ous” network or the residual GGM (Epskamp, Waldorp, et al., 2018). Here we
see a similar pattern as above: The residuals have negative conditional relation-
ships between-valence, and slightly greater in magnitude positive conditional
relationships within-valence. Note that the residual variance in this case is quite
low for each variable (Ψ̂ii ≈ 0.0185 with in-sample explained variance of approxi-
mately 99 percent). The estimated fixed point (that is, the mean) is approximately
µ̂1 = µ̂2 = 3.16 for Cheerful and Content, and µ̂3 = µ̂4 = 3.04 for Anxious and Sad.
Which characteristics of the bistable system can we recover based on the
VAR(1) estimates? The strong auto-regressive effects correctly capture the strong
linear auto-effects present in the true system, defined by the r parameters in
Equation 5.2. The lagged regression parameters suggest that there are sup-
pressing effects between valences, and reinforcing effects with valences, captur-
ing characteristic number five of the data-generating mechanism (Section 5.2.3).
However, the relative size of the suppressing and reinforcing effects is flipped in
the VAR(1): The suppressing effects are in fact larger in absolute value than the
reinforcing ones (see Section 5.2.1).
It is unclear what conclusions we can draw from the weak relationships
present in the residual network — as there are no such additional instantaneous
relationships present in the data-generating system. We assume a-priori when
fitting the VAR(1) model that these parameters are independent of (i.e., constant
across) time. Finally, the VAR(1) model describes a uni-stable system, precluding
us from capturing any characteristics related to bistability. The dynamics implied
by the VAR(1) are illustrated by generating new data from the estimated param-
eters, displayed in Appendix 5.C.2. The estimated location of the single fixed
point is not equivalent to either of the two stable fixed points or the unstable
fixed point in the true system.
Importantly, we can use our knowledge of the true bistable system to deter-
mine how we arrived at these observed parameter estimates. First, the estimated
position of the fixed point (given by µ̂) is roughly halfway between the positions
of the two stable fixed points, and is approximately equal to the sample mean
for each variable, reflecting that the system spends roughly the same amount of
time around each of the two stable fixed points. The main counter-intuitive re-
sult from the VAR(1) model is that the order of magnitude of the between- and
within-valence relationships is different than in the true bistable system. In the
true system, these pairwise relationships are non-linear, taking the form of in-
teraction effects, and the VAR(1) model captures the best linear approximation
of these non-linear relationships. As we can see from panel (a) of Figure 5.7, the
linear approximation of the within-valence relationship is largely driven by the
strong positive relationship present when both variables take on a high value, for
instance when Cheerful and Content are both near the healthy fixed point. From
panel (b) we can see that, in contrast, the linear between-valence relationship of
Content on Anxious is in fact a mixture of the strong negative effect near the un-
healthy fixed point (Content is low, Anxious is high) and the weaker effect near
the healthy fixed point (Content is high, Anxious is low). Combined, this results

132
 5.3. Recovering the Bistable System from Ideal Data

in higher linear within-valence relationships and lower linear between-valence

relationships.
Finally, the residual covariances displayed in panel (d) of Figure 5.7 are pro-
duced by a combination of model-misspecification (linear approximation of non-
linear relationships) and paths between each process at a shorter time scale than
observed, due to the Euler steps used in data generation. We stress here that, even
in the current idealized situation, it is not trivial to derive an exact explanation
for the residual covariance structure, and so its utility in drawing conclusions
about the underlying system should be approached with great caution.
In summary, the VAR(1) model gives us rather limited information regarding
the core characteristics of the bistable model. In principle, the VAR(1) model
is unable to capture any features which relate to bistability (characteristics 1-
4), as one would expect from a model that exhibits only a single fixed point.
What is perhaps more surprising is that, while the sign of the lagged relation-
ships (characteristic 5), and their symmetries are captured, their relative order-
ing (characteristic 6) is not. This observation is critical: While we could expect
that the VAR(1) model would not reproduce the global dynamics of the system,
even when we have ideal data, the linear relationships in the VAR(1) model also
fail to appropriately capture the local microdynamics in this instance. Funda-
mentally, this is due to the non-linear relationships which must be present in the
underlying system in order to induce bi-stability. In general we would not expect
that linear approximations of non-linear effects would preserve the same rank
ordering. This observation has potentially major implications for the analysis of
dynamic network structures, because many network metrics such as centrality
metrics are strongly dependent on the relative ordering of effects. In the follow-
ing we will examine if extending the VAR(1) model to allow for bi-stability brings
us closer to recovering a more accurate characterization of the data-generating
bistable system.

5.3.5 Threshold VAR Model

Regime-switching VAR(1) models extend the VAR(1) model to allow for obser-
vations to be drawn from two different conditional distributions for Xt given
Xt−1 , that is, two different regimes, described by two different sets of model pa-
rameters. These extensions in principle allow us to directly capture a notion of
multi-stability, by interpreting the mean vector of each conditional distribution
as a separate fixed point. Different extensions allow for different mechanisms by
which to model the switch between these regimes.
One popular regime-switching VAR(1) model is the Threshold TVAR(1)
model, where the system enters a different regime whenever a threshold value
or values τ of an a-priori specified threshold variable zt is crossed, written

(1)
Xt = b(1) + Φ (1) Xt−1 + et if zt ≤ τ
(2)
Xt = b(2) + Φ (2) Xt−1 + et if zt > τ

133
5. Recovering Bistable Systems from Psychological Time Series

for a two-regime model with a single threshold, where the VAR(1) parameters are
(r)
indexed by regime, with et ∼ N (0, Ψ (r) ), and mean vectors µ(r) = (I − Φ (r) )−1 b(r)
(Tong & Lim, 1980; Hamaker, Grasman, & Kamphuis, 2010). The threshold vari-
able zt may be an exogenous variable, or one of the variables in the VAR model.
Here we choose to use Cheerful (zt = x1,t−1 ) as the threshold variable. The thresh-
old value τ is a hyper-parameter that is estimated. Here, we estimate the TVAR(1)
model using the R-package tsDyn (Fabio Di Narzo, Aznarte, & Stigler, 2009),
which estimates τ using a grid search which selects the model with minimum
summed squared residuals.
Figure 5.8 displays the main results from the estimated TVAR(1) model, in
which the threshold is estimated as τ̂ = 2.811. In panel (a) of Figure 5.8 we
show the time-series with shading indicating which observations are below (grey)
or above (white) the threshold. We can see that the estimated threshold nicely
separates the time series into periods in which the system is in an unhealthy state
(based on Cheerful values below the threshold) and a healthy state (Cheerful
values above the threshold).

(a) State 1 State 2

8
Emotion Strength

0 5 10 14
Days
(b) Healthy Regime (c) Unhealthy Regime
0.88 0.88 0.87 0.88

0.02 0.01
Cheerful Content Cheerful Content
0.02 0.01

−0.02 −0.08 −0.08 −0.03

−0.03 −0.08 −0.03 −0.08 −0.08 −0.03 −0.08 −0.03
−0.02 −0.08 −0.08 −0.03

0.01 0.02
Anxious Sad Anxious Sad
0.02 0.02
0.88 0.87 0.88 0.88

Figure 5.8: Panel (a) shows the two weeks of the time series, with observations shaded in either grey
or white as a function of whether x1,t−1 is above or below the threshold τ̂ = 2.811. Panels (b) and
(c) show the estimated VAR(1) parameters as lagged networks in the healthy (white) and unhealthy
(grey) regimes respectively.

134
 5.3. Recovering the Bistable System from Ideal Data

Inspecting the lagged networks for each regime in panels (b) and (c) of Fig-
ure 5.8 we can see that the auto-regressive effects, and the within-valence cross-
lagged effects are pretty similar across both regimes. However, the cross-lagged
effects between variables of opposite valence are different. In the healthy regime,
negative valence emotions have much stronger cross-lagged effects on positive
(2) (2) (2) (2)
emotions (φ̂13 = φ̂14 = φ̂23 = φ̂24 = −.08), and vice versa for the unhealthy
(1) (1) (1) (1)
regime (φ̂31 = φ̂41 = φ̂32 = φ̂42 = −.08). Residual partial correlation networks
for both regimes are shown in Appendix 5.D, which display a similar pattern to
the regular VAR(1) model of weak positive residual partial correlation within-
valence and weak negative residual partial correlation between-valence. For the
TVAR, however, the residual covariance matrix is not symmetric across regimes:
In the healthy regime there is a slightly higher covariance between positive emo-
tions than negative emotions, and vice versa. The estimated means are given as
µ̂2 = {4.74, 4.75, 1.45, 1.46} for the healthy state and µ̂1 = {1.49, 1.48, 4.69, 4.69} for
the unhealthy state. Data generated by the TVAR(1) model estimates is shown in
Figure 5.18 in Appendix 5.C.3. From this figure we can see that most of the global
dynamics are well reproduced, although the system contains fewer switches be-
tween regimes than we would expect and there are fewer observations on the
switches between states compared to the original time series.
Which characteristics of the bistable system do we recover on the basis of the
TVAR(1) parameter estimates? First, the model picks up a number of character-
istics related to the bistability of the system. The estimated mean vectors capture
approximately the position of the two stable fixed points (characteristic 2), and
the estimated threshold correctly captures the position of the unstable fixed point
in the Cheerful dimension. However, note that bistability (characteristic 1) has
been specified a-priori and therefore cannot be considered to be recovered by the
model. Second, although the simulated data in Figure 5.18 (Appendix 5.C.3) ex-
hibits less frequent switches between states than we would expect, we can see
that the combination of state-dependent lagged parameters and residual vari-
ances does reproduce higher variability of positive emotion in the healthy state
in comparison to the unhealthy state, and vice versa for negative emotions (char-
acteristic 3). Finally, the lagged regression parameters in each regime correctly
capture that there are reinforcing effects within valence, and suppressing effects
between valence (characteristic 5).
The result that stands out in this analysis is the asymmetry in lagged regres-
sion coefficients across both regimes. This asymmetry would appear to indicate
that the parameters relating processes either change over time or are all explic-
itly a (step) function of the Cheerful variable. This last result is striking because
this intuitive interpretation does not correctly characterize the relationship be-
tween variables of different valences in the true bistable system. This is because
we know that the dependencies in C are invariant over time and fully symmetric.
However, the dependencies in C relate to pairwise interaction effects rather than
linear dependencies in the VAR(1) model. For example, the relationship between
Anxious, denoted x3 , and the rate of change of Content, dx 2
dt , depends both on the
value of C23 and on the current value of x2

135
5. Recovering Bistable Systems from Psychological Time Series

dx2
= r2 x2 + (C23 × x2 )x3 + . . . . (5.3)
dt
If we view x2 as a moderator, we can see that, when x2 is high, the effect of x3
on the rate of change, given by C23 × x2 , is relatively greater than when x2 is low.
In our system, separating the time-series into two regimes based on a threshold
of 2.811 for the Cheerful emotion essentially means we condition on high val-
ues of x1 and x2 in the healthy regime, and low values in the unhealthy regime.
This leads to the relatively stronger linear relationship from negative emotions to
positive emotions in the healthy regime, and vice versa in the unhealthy regime.
As such, we can see that the asymmetry in lagged relationships over time picked
up by the TVAR(1) model is a characteristic of the true bistable system. Notably,
however, the mechanism by which this asymmetry occurs is entirely due to non-
linear relationships between the observed variables and the similarity of vari-
ables that share the same valence, while the TVAR(1) modeler might be tempted
to ascribe this entirely to the effect of the level of Cheerful.
To summarize, the TVAR(1) model allows us to recover global dynamics, and
it recovers some aspects of the microdynamics. However, we saw that a naive in-
terpretation of the TVAR(1) parameter estimates may easily lead to the incorrect
conclusion that there is one time-varying variable which moderates the relation-
ships between all variables. In addition, we provided bistability as an input to
the model, and therefore cannot be considered a characteristic recovered from
data. In principle one could perform model selection between TVAR(1) models
with different numbers of components, however compared to the Mean switching
HMM in Section 5.3.2, the run time for such a model comparison was unfeasible
for the large data set used in our paper.
Furthermore, note that the threshold VAR(1) model does remarkably well for
this specific system for the following reason: While TVAR(1) models have fre-
quently been discussed in the literature (e.g. Warren, 2002; Hamaker et al., 2009,
2010; De Haan-Rietdijk, Gottman, Bergeman, & Hamaker, 2016) a major limi-
tation of this method is the difficulty in choosing a threshold variable. In our
data-generating mechanism, we know there to be an unstable fixed point defined
in multivariate space, x1 = x2 = x3 = x4 = 2.8. It just so happens that in this case,
almost always when we pass this position in one dimension (e.g., x1 > 2.8) we also
do so in all other dimensions (e.g., x2 > 2.8, x3 < 2.8, x4 < 2.8). This means that
the true mechanism of state-switching behavior is very well approximated by the
univariate mechanism in the TVAR(1) model, for this choice of parameter val-
ues. In more general situations, the choice of threshold variable(s), and number
of thresholds, is likely to be less trivial. While the TVAR(1) model does capture
that there are suppressing and reinforcing effects between and within valences,
it does not capture the relative size of these effects, and it may easily lead to the
incorrect conclusion that there is a single time-varying variable which moderates
all of the relationships between other variables in the system.
Finally, the TVAR(1) is only one of a variety of different regime-switching
dynamic models which could be fitted to the data at hand. Another alternative
would be the Markov-Switching (MS-)VAR model (Hamilton, 1989; Hamaker et

136
 5.3. Recovering the Bistable System from Ideal Data

al., 2010; Hamaker & Grasman, 2012; Chow et al., 2018), a combination of the
HMM and VAR models, in which the regime-switching behavior is determined by
a random Markov process operating between latent categorical variables. While
this model is more flexible than the Threshold VAR model, we show here the
TVAR results for two reasons. First, in this instance the switching behavior will
be less well approximated by the MS-VAR model, leading to even less straightfor-
ward conclusions about the data-generating process, but otherwise likely highly
similar lagged parameter estimates. Second, while recent advances such as the
dynR package (Ou et al., 2019) have made this model easier to estimate, it is still
prohibitively difficult and time consuming to fit to data.3
Now that we have shown the capabilities and limitations of the TVAR model
in recovering the bistable system, there are a few different avenues we could pur-
sue to further increase our model complexity in the hope of recovering more and
more of the features of underlying system. For example, both the TVAR and
MS-VAR can be considered special cases of time-varying parameter models, that
assume the true time-varying model is a partition between a finite set of com-
ponents. Other types of time-varying VAR models assume the parameters are a
smooth function of time (e.g., Haslbeck, Bringmann, & Waldorp, 2017). How-
ever, we would not expect these models to outperform the threshold VAR in this
instance for two reasons. First, the threshold VAR model is already able to cap-
ture the major source of variation in parameters over time, that is, the step-like
switches between stable states. Second, since these models are still based on fit-
ting locally stationary VAR models, the fundamental limitations of approximat-
ing the dynamics with linear relationships remain. As such, in the next section
we examine an approach which aims to recover the exact system of differential
equations (DEs) from data, by allowing non-linear terms to enter into a step-wise
model building procedure.

5.3.6 Differential Equation Model Building

In the previous sections we have shown that some models have been able to re-
cover some characteristics of the true model, but that it is generally difficult to
make inferences about the characteristics of the true system from these models.
Also, since all of these models were misspecified they were fundamentally unable
to recover the exact true bistable system. In this section, we aim to recover the
exact system of differential equations (DEs) directly from the ideal time series.

5.3.6.1 Model Building Procedure

The structure of the true model is typically unknown in practice, and therefore
has to be learned from the data. Chow (2019) describes a general methodology for
building dynamic systems models which consists of two steps: In the first step,
we approximate the first-order derivatives by taking difference scores between
3 Despite numerous attempts and correspondence with the authors of the package, we were unable
to get the model estimates for the dataset described here to converge.

137
5. Recovering Bistable Systems from Psychological Time Series

consecutive measurement occasions, divided by the length of the time-interval

between those occasions

dxi,t xi,t+1 − xi,t

≈ (5.4)
dt ∆t
where in each case, ∆t = .1, as described in Section 5.2.2 (cf. Boker, Deboeck, et
al., 2010).
In the second step, we use this approximate derivative as the outcome vari-
able, and try to find a regression model that predicts this outcome variable as
well as possible with as few parameters as possible. Here, we use results obtained
from the statistical models in the previous sections as a starting point, and then
follow the standard model-building approach of fitting models with increasing
complexity and evaluating the improvement in out-of-sample fit.
From the descriptive statistics (Section 5.3.1) the marginal lag-0 (Section
5.3.3) and lag-1 (Section 5.3.4) relationships and the mean-switching Hidden
Markov Model (Section 5.3.2), we saw that the system is bistable. From dynam-
ical systems theory we know that bistability is only possible in the presence of
non-linear terms (Strogatz, 2015). Similarly, we saw from the TVAR(1) model
(Section 5.3.5) that the linear relationships between pairs of variables is depen-
dent on where in the state-space other variables are located (i.e., below or above
a given threshold). Both of these pieces of information suggest the presence of
interaction effects between variables: However, we have no information about
what specific interaction terms, or what other linear or non-linear dependencies
should be in the model. Here, we start out with a main effects-only model, and
then add more and more non-linear terms (interactions, quadratic effects, etc.).
We evaluate the fit using the mean out-of-bag proportion of explained variance
(R2 ) obtained from a 10-fold cross-validation scheme (see Appendix 5.E for de-
tails), and we choose the model that maximizes this value. If two models result
in the same fit, we choose the model with less parameters.
The fit of the all models considered here is shown in Table 5.1. First, we test
a baseline model (Model A) where each derivative is a linear function of all other
variables. As we described above, the absence of interaction effects makes it un-
likely that this is a suitable candidate model, but it gives us a baseline explained
variance value of R2 = 0.04664. In Model B, we add to the baseline model all pair-
wise interactions between the outcome process (e.g., x1 when the DV is dx1 /dt)
and the other variables in the model xj (i.e., x1 ×xj , ∀j ∈ p). Adding these pairwise
interactions increases the variance explained to R2 = 0.06874. In Model C, we
further extend this model by adding all possible pairwise interactions between
all variables xi × xj , ∀(i, j) ∈ p. However, we see that adding these parameters
in fact leads to a slight decrease in explained variance, R2 = 0.06870, indicating
overfitting. For brevity, we display only these three models, but adding further
complexity to model in terms of additional interaction terms, quadratic or cubic
terms also fails to increase the out-of-bag R2 (see Appendix 5.E). As such, we can
take Model B to be our final model.

138
 5.3. Recovering the Bistable System from Ideal Data

dxi,t
Model ∼ ai + ri xi + . . .
dt q R2
P
A j,i rj xj 5 0.04464
P Pp
B j,i Rij xj + j Cij xj xi 9 0.06874
P Pp
C j,i Rij xj + (j,k) βj xj xk 15 0.06870

Table 5.1: Model fit results for each of the four models described in text. The second column gives
the model equation for each variable, q denotes the number of parameters estimated per univariate
regression model, and the final column indicates the mean proportion of explained variance R2 ,
calculated on the hold-out sets of a 10-fold cross-validation scheme (for details see Appendix 5.E)

5.3.6.2 Dynamics and Data Generated by Final Model

We can see that the structure of Model B is highly similar to the structure of our
data-generating model, with additional main effects between variables, that is,
the linear effects denoted by the p × 1 vector r in the true model is replaced by a
p×p matrix R in our chosen model. Furthermore, we can see from the left panel of
Figure 5.9 that the parameter estimates are highly similar, but not exactly equal
to the data-generating parameters:

h iT
â = 1.40 1.37 1.25 1.27
h iT
σ̂ = 1.35 1.34 1.34 1.34
10

 −0.88 −0.02 −0.01 −0.01 

 
 −0.03 −0.95 −0.01 −0.00 
R̂ = 

 −0.02 −0.05 −0.96 −0.04 

Negative Emotion

−0.04 −0.01 −0.08 −0.91

 
5 ●

 −0.18 −0.04 −0.17 −0.18

  ●

 −0.03 −0.19 −0.19 −0.19 
Ĉ =  ●
 
 −0.18 −0.19 −0.19 −0.03


0
−0.19 −0.18 −0.02 −0.18
 
0 5 10

Positive Emotion

Figure 5.9: Left panel: the parameters estimated from the ideal data. Right panel: the vector field
defined by the estimated parameters in the left panel. Solid points indicate stable fixed points and
empty points indicate unstable fixed points. The solid lines indicate the values at which derivative of
positive emotion (orange) and negative emotion (light blue) is equal to zero. At the points at which
the two lines meet, both derivatives are equal to zero and the system remains in this (stable) state.

While we would not expect to recover the exact parameters of the true model
with a different functional form, we see that the signs, size and relative orderings
of parameters in the estimated Ĉ matrix are quite accurate. Based on these pa-
rameters, we recover that there are suppressing effects between valences and re-

139
5. Recovering Bistable Systems from Psychological Time Series

inforcing effects within valences (characteristic 5), that the reinforcing effects are
smaller in absolute value than the suppressing effects (characteristic 6), and by
capturing approximately the correct functional form, we capture that the micro-
dynamic parameters are dependent only on variables inside the model (charac-
teristic 7). Furthermore, we can see that false positive (i.e., off-diagonal) elements
of R̂ are of a much smaller size than the true positive diagonal elements. The full
parameter estimates, with standard errors and p-values are shown in Appendix
5.E.
Beyond inspecting the estimated parameters, we can judge how good of an
approximation of the true bistable system our estimated model represents by
comparing the dynamics implied by that model to that of the true system. The
dynamics of a differential equation model are described by its vector field, which
we depict for Model B in the right panel of Figure 5.9. To construct this vec-
tor field we use the same two-dimensional approximation (positive and nega-
tive emotion) as we did in Section 5.2.1 (see Appendix 5.A for details). The or-
ange and light blue lines are solution lines which indicate the locations where
the rate of change in one dimension (orange for no change in positive emotion,
light blue for no change in negative emotions) is zero. The points at which these
solutions line cross determine the fixed points. We can see that our model cor-
rectly identifies three fixed points in this range of values: one stable healthy
(x1 = x2 = 4.91, x3 = x4 = 1.34), one stable unhealthy (x1 = x2 = 1.39, x3 = x4 =
4.84), and one unstable fixed point approximately halfway between those two
(x1 = x2 = 2.79, x3 = x4 = 2.82). If we compare these global dynamics to the global
dynamics of the true bistable system depicted in Figure 5.1 in Section 5.2.1, we
see that Model B very accurately reproduces these dynamics, approximating the
position of the fixed points in the true system closely. From this we can say that
the estimated DE model captures characteristics 1 (bistability) and 2 (location of
the fixed points) in the true system.
An additional way to evaluate whether the dynamics of the estimated model
are similar to the dynamics of the true model is to generate data from the esti-
mated model and compare this data to the original data. Figure 5.10 shows a time
series generated from Model B using a step size of .1. We can see that the data
looks very similar to the original data generated from the true bistable system, in
that the fixed points are at roughly the same location, there is a difference in vari-
ance across the high and low emotion value fixed points (characteristic 3), and
there is a similar number transitions (around 14) between the healthy and un-
healthy state (characteristic 4). Thus, even though we did not exactly recover the
set of true parameters exactly, we seem to have recovered a model that is equal
to the true model in all relevant aspects, capturing all of the seven characteristics
listed in Section 5.2.3.

140
 5.3. Recovering the Bistable System from Ideal Data

Emotion Strength 8

6
Cheerful
Content
Anxious
4 Sad

0 5 10 14
Days

Figure 5.10: Data generated from the estimated DE model, with the same initial values as the ob-
served data

5.3.6.3 Exact Recovery of Model Parameters

While this model building procedure performed extremely well in this scenario,
the findings here should be approached with a note of caution. Observe that, de-
spite negligible sampling error, we do not succeed in recovering the exact param-
eter estimates. The reason for this is that, while data is generated using an Euler
step of ∆t = .01, our ideal time series is created by sub-sampling with ∆t = .1.
While this is an unrealistically high sampling frequency, it still means that we
cannot estimate the derivative perfectly. As the sampling frequency becomes
lower, we would expect the quality of this approximation to degrade.
In theory, to recover the data-generating parameters, we would need to fit
the integral solution form of the differential equation (Strogatz, 2015), as this de-
scribes the relationships between observed variables spaced ∆t apart, as implied
by the differential equation. It is well known that this integral solution may con-
tain a seemingly different set of dependency relationships than the differential
equation. For instance, variables which are independent in the DE form may be
dependent in the integral form, and the signs and relative orderings of these de-
pendencies may change depending on the value of ∆t (Ryan & Hamaker, 2019;
Kuiper & Ryan, 2018; Aalen et al., 2012). Because methods based on approx-
imating integral solutions are expected to suffer from similar problems as the
two-step DE estimation procedure, and because these methods are difficult to
apply in practice, we limit ourselves to the two-step approach in this paper (see
section 5.5.3 for a further discussion).

5.3.7 Summary: Analysis of Ideal Time Series

In this section we aimed to recover characteristics of the true bistable system
from the ideal time series with 10 measurements each minute using a number of
time series analysis tools. Table 5.2 provides a rough summary of which method
recovered which characteristics of the true bistable system.

141
5. Recovering Bistable Systems from Psychological Time Series

5) ) 7)
) .( (6 t(
1) 3) ( 4 n f e a n
y( 2) ( ns ei Siz onst
ili
t n( ce tio ./R tive
sta
b
sit io
ri an
an
si
pp r
la e-c
i o a r u e i m
B P V T S R T
Data Visualization X X X × × × ×
HMM X∗ X X X × × ×
Lag-0 / GGM × × × × X × X∗
Lag-1 / VAR(1) × × × × X × X∗
TVAR(1) X ∗
X X X X × ×
DE-Estimation X X X X X X X∗

Table 5.2: Summary of which method recovered which of the seven qualitative characteristics listed
in Section 5.2.3 from the ideal time series. The first four characteristics are global dynamics, thee
last three are microdynamics. The check marks with asterisk indicate that the method includes the
characteristic as a model assumption, and can therefore not be considered recovered from the time
series.

We showed that data visualization (Histograms and the pairwise marginal re-
lationships in Sections 5.3.3 and 5.3.4) revealed bistability and provided a rough
estimate of the position of and variances around the fixed points. However, when
comparing the eye-balled number of switches with the estimates of the HMM,
we saw that we missed instances in which the system quickly switched back and
forth. The Mean switching HMM recovered all global dynamics, however, we
provided bistability as a model assumption, which is why we mark the check
mark at the first characteristic with an asterisk.
Turning to methods that capture dependencies between variables, the analysis
of lag-0 relationships with the GGM and the analysis of lag-1 relationships with
the VAR(1) model (and a GGM on its residuals) fundamentally cannot recover
any global dynamics of the bistable system, but they recovered some micrody-
namics: the characteristic that within valence effect are reinforcing, and between
valence effects are suppressing; and that the parameters are constant across time,
however this is again an assumption of the model and therefore cannot be consid-
ered recovered from the data. The TVAR(1) model was able to recover all global
dynamics with the same caveat as in the HMM, that bistability is a model as-
sumption and not recovered from data. Similarly to the VAR model, it recovered
the reinforcing/suppressing characteristic. However, a naive interpretation of
the model parameters would lead one to conclude that the parameters are time-
varying. Finally, the DE-estimation method was able to recover all microdynam-
ics reasonably well, which implies that it also recovered all global dynamics.
The purpose of this section was to establish whether or not each method can
recover, in principle, some aspect of the bistable system. To do this we used
a highly idealized dataset, with an unrealistically high sampling frequency. As
such, the performance of each method described above can be considered an up-
per bound on its performance in any more realistic scenario. It remains to be
seen exactly how the performance of each method, and in general our ability to

142
 5.4. Recovering the Bistable Systems from ESM Data

recover global and microdynamic characteristics of the system, changes when a

more realistic sampling frequency is used.

5.4 Recovering the Bistable Systems from ESM Data

In this section we analyze a time series that is similar to the ideal time series in
all aspects, except that the system is sampled every 90 minutes instead of every
six seconds (see Section 5.2.2). This allows us to investigate how the ability of
each method to recover (some characteristic of) the bistable system is affected by
having only a low sampling frequency time series, as it is typical for ESM studies.

5.4.1 Descriptive Statistics, HMM and Lag-0 Relationships

The descriptive statistics, such as histograms, and the lag-0 relationships ob-
tained from the ESM times series (Figures 5.20 and 5.21 respectively in Appendix
5.F) are essentially identical to those obtained from the ideal time series, depicted
in Figures 5.4 and 5.6 in Section 5.3.1. This makes sense: we have exactly the
same amount of data points, sampled from the same system as in the ideal time
series case. The only difference is that in the ESM data set 900 time points are
“missing” between each measurement of the ESM time series. However, because
lag-0 relations do not pick up on any temporal dependence, the lower sampling
frequency does not affect the lag-0 relations. While this suggests that lag-0 rela-
tions are robust against low sampling frequency, it also puts their utility to infer
the dynamics of an underlying dynamical system into question.
The parameter estimates obtained by fitting a two-component mean-
switching Hidden Markov Model on the ESM dataset were

 1.47   0.41   4.71   0.64 

       
!
 1.46   0.41   4.71   0.64  0.915 0.085
µ̂1 =  , σ̂1 =
 0.63 , µ̂2 =
 
 1.47 , σ̂2 =
 
 0.41 , Â =
  .
 4.71  0.090 0.910

     
4.71 0.63 1.47 0.41

and the predicted states for two weeks of the time series are shown in Appendix
5.F Figure 5.22.
We see a very similar pattern of results as obtained from the HMM fit to the
ideal time series in Section 5.3.2, with the means and standard deviations of state
1 and state 2 reflecting the unhealthy and healthy states respectively. However,
the parameters of the estimated transition matrix A for this time series show sub-
stantially higher switching probabilities, Â12 = .085 and Â21 = .090. As we can
see from Figure 5.3, although the sub sampled ESM time series contains only
224 observations for a two-week period, rather than 201600, the sampling fre-
quency is still high enough to capture each of the 17 switches between states in
this period. That means that, although the amount of transitions that occur over a
period of time remains the same, the number of measurement occasions between
any two transitions is lower, which results in a higher transition probability. We

143
5. Recovering Bistable Systems from Psychological Time Series

can see that this higher transition probability captures the number of transitions
over two weeks quite accurately — the model predicts between 224 × .090 ≈ 20
and 224 × .085 ≈ 18 switches on average over a two week period. As such, the
HMM fitted on the ESM time series still allows us to estimate the location of and
variance around the two fixed points (characteristics 2 and 3), and approximate
the frequency of transitions between these two fixed points (characteristic 4). In
fact, the transition probabilities appear to be even more accurate than the ideal
case — most likely this numerical imprecision in the ideal case is because the
number of transitions relative to total time series was so low that slight changes
in the transition probability value lead to very different predictions regarding the
number of transitions over 201600 time points.

5.4.2 Lag-1 Relationships and VAR model

When analyzing the lagged relationships in the ESM time series, we begin to
see some striking differences from the analysis of the ideal time series. Panels
(a) and (b) of Figure 5.11 show the marginal relationship of Content observed at
time t with Cheerful at the previously observed time point t − 1, and Anxious
at time t with Content at time t − 1, respectively. Focusing on panel (a), we see
that the density of the lagged variables takes on a square-like shape, and each
quadrant seems to be filled with a roughly circular density. This is in contrast to
the density of the lagged relationships in the ideal data displayed in Figure 5.6
in Section 5.3.4, which was described by two elliptical shapes at the two fixed
points.
How can we explain this pattern? In the ideal data, the two elliptical shapes
indicate that Contentt−1 and Contentt tend to be near the same fixed point (two
shapes), and that the two variables are positively correlated (elliptical shape).
Now, in the ESM time series, we still have most of the density in the upper-
right and the bottom-left quadrant, indicating that if Contentt−1 is at the healthy
(unhealthy) fixed point, it is very likely that Contentt is also at the healthy (un-
healthy) fixed point (noting that t − 1 now reflects a 90 min instead of 6 second
time interval). However, we now also observe density at the top-left and bottom-
right quadrant. These densities represent the situation in which Contentt−1 is
in the healthy (unhealthy) state, but Contentt is in the unhealthy (healthy) state.
This situation is created when a switch between states falls within the 90min in-
terval between two ESM measurements. Next, we focus on the shape of the den-
sity within each of the quadrants: we see that each of the densities have roughly
a circular shape, which indicates that Contentt−1 and Contentt are uncorrelated
at each fixed point. This makes sense: in the ESM time series 900 time points
are missing between each pair of measurement, which means that the that there
is essentially no temporal dependence anymore between the variables. The re-
lationship between Anxioust−1 and Contentt can be explained in an analogous
way. Before fitting the VAR model below, this already shows us that it is futile to
recover the microdynamics of the bistable system from these data.
Panel (c) of Figure 5.11 displays the network of estimated lagged regression
coefficients from a VAR(1) model fit to the data. If we were to use these to in-

144
 5.4. Recovering the Bistable Systems from ESM Data

8 (a) 8 (b)

6 6

Anxioust
Contentt

4 4

2 2

0 0

0 2 4 6 8 0 2 4 6 8
Cheerfult 1 Contentt 1

Lag(1) Network (c) Residual Network (d)

0.2 0.21

0.21
Cheerful Content Cheerful 0.51 Content
0.2
−0.23

−0.21 −0.19
−0.2 −0.2 −0.21 −0.19
−0.2 −0.23 −0.24
−0.19

−0.23
0.19
Anxious Sad Anxious 0.51 Sad
0.2

0.2 0.19

Figure 5.11: Panel (a) shows the relationship between Content and Cheerful, two emotions with
the same valence, spaced one measurement occasion apart (i.e. at a lag of one but with 90 minutes
between measurements) for the ESM dataset. The red line indicates the best fitting regression model.
Similarly, panel (b) shows the relationship between Anxious and Content, two emotions with different
valence, at a lag of one. Panel (c) displays the matrix of lagged regression parameters, estimated from
a VAR(1) model, as a network, and panel (d) displays the partial correlation matrix of the residuals
of the VAR(1) model as a network. This latter network is often referred to as the “contemporaneous”
network.

fer the microdynamic characteristics of the system, we would manage to recover

the signs of effects between variables: Negative, suppressing effects between-
valence, and positive, reinforcing effects within-valence (characteristic 5, see Sec-
tion 5.2.3). However, we again fail to recover the relative size of the suppressing
and reinforcing effects. In fact, in this case, all of the estimated auto-regressive
and cross lagged effects have approximately equal absolute value |φ̂ij | ≈ .2. This
means that we also fail to recover the strong auto-regressive relationships en-
coded by the r parameters in the data-generating model, and reflected by the
strong auto-regressive effects estimated by the VAR(1) model in the ideal setting.
In panel (d) we can see that, as was the case for the ideal time series, we obtain

145
5. Recovering Bistable Systems from Psychological Time Series

positive residual partial correlations within-valence and negative residual cor-

relations between-valence, although the magnitude of these correlations is now
quite high, θ̂ = 0.51 and θ̂ = −0.23 respectively. In addition, note that the residual
variances of each variable in the model is considerably higher than the ideal case,
and approximately equal for all variables (Ψ̂ii ≈ 1.1, explained variance ≈ 0.62).
As we would expect, taking only every 900th measurement from the ideal time
series means that the predictive power of the VAR model decreases.
How can we reconcile these parameter estimates with what we know of
the underlying bistable system? Although we may be tempted to interpret
the VAR(1) parameters as reflecting the microdynamic structure, we have al-
ready seen in the analysis of the marginal relationships above that the large
time-interval between observed measurements means that such an interpretation
would be incorrect. Instead, the VAR(1) parameter values in the present situation
are fully determined by the global characteristics of the system. Essentially, the
estimated lagged relationships reflect that, at a time-scale of 90 minutes, the dy-
namics of the system from one observation to the next can be boiled down to four
possibilities: Either the entire system stays in the same state (healthy-healthy or
unhealthy-unhealthy) or moves from one state to the other (healthy-unhealthy or
unhealthy-healthy). Since 1) the most likely behavior is that the system stays near
the same fixed point, and 2) those two fixed points are defined as high-positive
and low-negative emotions, or low-positive and high-negative emotions, we end
up with positive within-valence relationships (e.g., if Cheerful now is near the
high fixed point, it is likely that Content later will be high too) and negative
between-valence relationships (e.g., if Anxiety now is high, it is likely that Con-
tent later will be low). All of the auto-regressive and cross-lagged relationships
are of equal value, as essentially all variables have the same value in predicting
what fixed point each other variable will be near at the next measurement occa-
sion: Enough time elapses between measurement occasions that even the auto-
regressive effect is only as predictive as the cross-lagged effects. As noted above,
this interpretation is also reflected in the joint densities in panels (a) and (b) of
Figure 5.11. Here, each density takes the appearance of four quadrants of uncor-
related Gaussian distributions, indicating that the microdynamic dependencies
present in the ideal times series are totally absent from the ESM time series.
In summary, having a realistic sampling frequency results in the VAR(1)
model providing even less information about the characteristics of the bistable
model than in the ideal scenario. The longer time-interval between observations
implies that interpreting VAR(1) parameters as reflecting truly microdynamic be-
havior would be incorrect. Parameters interpreted as reflecting microdynamics
in fact must be interpreted as reflecting the global characteristics of the system.
Although the sign of the microdynamic relationships (characteristic 5) is recov-
ered, in this instance it happens that the pattern of microdynamic relationships
has the same valence as the pattern of relationships at a longer time-scale, that is,
the movement of the process between fixed points. Thus, while in the ideal case
the VAR parameters were a mixture of the microdynamics (around each fixed
point), and global characteristics (i.e., position of the two fixed points), in the
ESM time series these parameters are only reflective of the global characteristics.

146
 5.4. Recovering the Bistable Systems from ESM Data

5.4.3 Threshold VAR Model

We saw in the previous sections that inferring global characteristics using ESM
data was somewhat successful, but that inferring microdynamic characteristics
using a VAR(1) model was impossible. For the ideal time series, we saw that
the threshold VAR(1) model was in principle able to capture some microdynamic
and some global characteristics, and so in this section we examine how well that
performance generalises to our emulated ESM data. We use the same threshold
variable (Cheerful, X1 ) and model specification as described in Section 5.3.5.
Figure 5.12 displays the main results from the TVAR(1) model estimated on
the ESM data: The estimated threshold is τ̂ = 2.796, very close to the estimated
threshold in the ideal case, and from panel (a) of Figure 5.12 we can see that
this threshold value does well in separating the time-series into the healthy and
unhealthy states. Inspecting the lagged networks for each regime in panels (b)
and (c) of Figure 5.12 we see a similar general pattern of results as the lagged net-
works for the ideal time series in Figure 5.8 in section 5.3.5. In the healthy regime,
the negative variables have much stronger cross-lagged effects on the positive
variables, and vice versa for the unhealthy regime. However, we see even more
differences between regimes in this case than we did for the ideal time series.
For instance, in the healthy regime, the within-valence and auto-regressive rela-
tionships for the negative variables is much stronger than for the positive vari-
ables, a pattern which is flipped for the unhealthy regime. In both regimes, the
within-valence cross-lagged parameters are roughly equal to the auto-regressive
effects of the variables involved. The estimated means of each regime are given
as µ̂2 = {4.31, 4.31, 1.87, 1.87} for the healthy state and µ̂1 = {1.74, 1.71, 4.44, 4.62}
for the unhealthy state.
We can see from this that the TVAR(1) model for the ESM data succeeds in
recovering some global characteristics of the system. Specifically, the estimated
mean vectors capture approximately the position of the two stable fixed points
(characteristic 2), and the estimated threshold correctly captures the position of
the unstable fixed point in the Cheerful dimension. However, the recovery of
this characteristic comes with the same caveats as described in Section 5.3.5, in
that the use of a univariate threshold for this particular configuration of the true
system happens to be a good approximation of the unstable fixed point in multi-
dimensional space.
Regarding the microdynamics, the lagged parameters in each regime approx-
imately capture that there are reinforcing effects within valence and suppressing
effects between valence (characteristic 5). Otherwise, however, the recovery of
microdynamic relationships performs worse than for the ideal time series, as ex-
pected. As was the case for the VAR(1) model, the regime-specific lagged param-
eters here again reflect global characteristics at the 90 minute time-scale rather
than microdynamics. Partitioning the joint densities in panels (a) and (b) of Fig-
ure 5.11 using a threshold does not aid us in any way to reproduce microdynamic
dependencies which are absent due to the low sampling frequency. Thus, the
asymmetry in parameter values across regimes has to be a function of the global
characteristics, influenced by both the different variances around the fixed points

147
5. Recovering Bistable Systems from Psychological Time Series

(a) State 1 State 2

8
Emotion Strength

0 5 10 14
Days
(b) Healthy Regime (c) Unhealthy Regime
0.08 0.14
−0.03 0.06
0.06 0.14
Cheerful Content Cheerful Content
−0.03 0.08

−0.06 −0.15 −0.14 −0.04

0.03 −0.16 −0.06 −0.15 −0.08 −0.04 −0.15 −0.03

0.03 −0.16 −0.09 −0.04

0.15 0.03
Anxious Sad Anxious Sad
0.16 0.04

0.16 0.15 0.04 0.04

Figure 5.12: Panel (a) shows the first two weeks of the time series, with observations shaded in either
grey or white as a function of whether x1,t−1 is above or below the threshold τ̂ = 2.796. Panels (b) and
(c) show the estimated VAR(1) parameters as lagged networks in the healthy (white) and unhealthy
(grey) regimes respectively.

in each state (i.e., high variance for positive emotions, low variance for negative
emotions in the healthy state, and vice versa) and those observations which jump
from one fixed point to the other across consecutive measurement occasions, as
discussed in the previous section.
As we did throughout Section 5.3, we could evaluate how well this model de-
scribes the bistable system by generating data from it. Notably, the dynamics
defined by Φ (1) and Φ (2) reflect an unstable system in both regimes: The eigenval-
ues of both contain a value outside the unit circle (i.e., with absolute value greater
than one) (Hamilton, 1994). This means that, if we were to generate data using
these parameters, the time series would always diverge towards infinity. This in-
stability also precludes us from making any statement regarding the variance of
positive and negative emotions in each regime (characteristic 3), as the long run
variances implied by the model are infinite. As such, we can say that overall, the
set of estimated parameters for the TVAR(1) based on the ESM time series are a
poor characterisation of the microdynamics of the model at any time-scale.

148
 5.4. Recovering the Bistable Systems from ESM Data

In summary, the TVAR(1) model fitted on the ESM time series still picks up a
global characteristic of the system, but the recovery of microdynamic character-
istics fails. In fact, the relationship between the estimated lagged parameters and
the characteristics of the system was much more opaque than in the ideal data
case, and our ability to generalize from the estimated parameters to the behavior
of the system at any time scale was considerably worse than in the ideal case.
Again here, we should note that the only difference between the ideal and ESM
time series is the sampling frequency. Fundamentally, the results here indicate
that, if we do not have a sufficiently high sampling frequency, then fitting increas-
ingly complex models, or extensions to simpler models such as the TVAR(1), does
not aid us in recovering the characteristics we are interested in: Even when we
have an arbitrarily large number of observations, we fail to recover basic charac-
teristics of the microdynamics due to the spacing between measurements.

5.4.4 Differential Equation Model Building

In this section we will examine whether the DE model-building procedure de-
scribed in Section 5.3.6 also succeeds in recovering the bistable system when
applied to the emulated ESM dataset. Recall from Section 5.3.6 that for the
ideal time series, this method succeeded in recovering the microdynamics of
the system: The global characteristics were also considered to be recovered as
the global characteristics implied by the estimated model (bistability, position of
fixed points) matched up with the actual global characteristics of the underlying
system.

5.4.4.1 Model Building Procedure

Similarly to Section 5.3.6 we first estimate the derivatives directly from the data
by differencing the time series, and then search for the best fitting model by
fitting a series of regression models with increasing complexity. Table 5.3 dis-
plays the fit of seven increasingly complex regression models, evaluated using
the mean out-of-bag proportion of explained variance R2 .
Model A (R2 = 0.13991), Model B (R2 = .16827) and Model C (R2 = 0.16928)
are the same models as introduced in Section 5.3.6. However, since we did not
observe a clear drop in R2 as we increased model complexity from Model A to
Model C, we also assess the fit of four additional models. Model D adds cu-
bic main effects x13 , . . . x43 as predictors, increasing the model fit to R2 = 0.19455.
Model E adds four three-way interactions (xi × xj × xk , i , j , k) to this, further
increasing the model fit to R2 = 0.19940. Adding yet more three-way interac-
tions (xi × xj × xk , ∀(i, j, k) ∈ p) in Model F still increases model fit (R2 = 0.19940),
as does adding all possible four-way interactions in Model G (R2 = 0.20420). As
it is not possible to specify more unique product interaction terms, we consider
Model G to be our final model.

149
5. Recovering Bistable Systems from Psychological Time Series

dxi,t
Model dt ∼ a + ri xi + . . . q R2
P
A j,i rj xj 5 0.13991
P Pp
B j,i Rij xj + j Cij xj xi 9 0.16827
P Pp
C j,i Rij xj + (j,k) βj xj xk 15 0.16928
P Pp Pp 3
D j,i Rij xj + (j,k) βj xj xk + j γj xj 19 0.19455
P Pp Pp 3 P p
E j,i Rij xj + (j,k) βj xj xk + j γj xj + j,k,l ζj (xj xk xl ) 23 0.19801
P Pp Pp 3 Pp
F j,i Rij xj + (j,k) βj xj xk + j γj xj + (j,k,l) ζj (xj xk xl ) 35 0.19940
P Pp Pp 3 Pp
j,i Rij xj + (j,k) βj xj xk + j γj xj + (j,k,l) ζj (xj xk xl )
G Pp 70 0.20420
+ (j,k,l,m) ηj (xj xk xl xm )

Table 5.3: Model fit results for each of the seven models described in text, for the emulated snapshot
ESM data. The second column gives the model equation for each variable, q denotes the number
of parameters estimated per univariate regression model, and the final column indicates the mean
proportion of explained variance R2 , calculated on the hold-out sets of a 10-fold cross-validation
scheme (for details see Appendix 5.E)

5.4.4.2 Dynamics and Data Generated by Final Model

Clearly, the model-building procedure for the emulated ESM data failed to re-
cover the functional form of the true bistable system. Furthermore, we have ar-
rived at a final model which is so complex (4 × 70 = 280 vs. 4 × 6 = 24 parameters
in the true model) that it is close to uninterpretable. In theory we could con-
tinue adding complexity to the model in the form of non-linear transformations
or spline functions, which we know to be absent from the data-generating mech-
anism, but which may improve model fit. However, this would make the model
even more difficult to interpret.
In the left panel of Figure 5.13 we present the estimated parameters that are
also contained in the true model, with full parameter estimates and standard er-
rors shown in Appendix 5.E.3. We can see that the estimates deviate widely from
the parameters in the true bistable system. In addition, the estimated parameters
in the C matrix fail to capture the sign and relative ordering of all parameters in
the true C matrix, though a full evaluation of whether suppressing and reinforc-
ing effects of different sizes are present (i.e., characteristics 5 and 6) is infeasible
due to the large number of parameters present in the model. Thus, we can say
that this approach fails to recover the microdynamics of the system at least to the
degree that they can be interpreted.
While the system did not recover the microdynamics in the sense that it cap-
tures the qualitative characteristics of the true bistable system, it could still be
the case that this more complex system exhibits global characteristics that are
similar to the true bistable system. Similarly to Section 5.3.6, we can evaluate
these dynamics by inspecting its vector field, shown on the right-hand side of
Figure 5.13. As in the vector field obtained from the ideal data (Figure 5.9), the
intersections of the two solution lines (blue and orange) indicate the position of
the different fixed points in the shown range of the state space. These fixed points

150
 5.4. Recovering the Bistable Systems from ESM Data

h iT
â = −0.07 −0.16 0.31 0.20
h iT 10
σ̂ = 0.01 0.01 0.01 0.01
●

 −0.19 −0.07 −0.06 −0.01

 


Negative Emotion
 −0.16 −0.18 −0.01 −0.04 
R̂ = 
 
 −0.09 −0.29 −0.05 −0.01

 5
●
−0.12 −0.22 −0.07 0.10
 
●

 −0.02 −0.11 −0.02 −0.05

 

 −0.10 ●
−0.02 −0.01 −0.05 
Ĉ = 
 
 −0.00 −0.05 −0.01 −0.01

0

−0.00 −0.02 −0.04 −0.03
 
0 5 10

Positive Emotion

Figure 5.13: Left panel: the parameters estimated from the snapshot ESM time series. Right panel:
the vector field defined by the estimated parameters. Solid points indicate stable fixed points and
empty points indicate unstable fixed points. The solid lines indicate the values at which derivative of
positive emotion (orange) and negative emotion (light blue) is equal to zero. At the points at which
the two lines meet, both derivatives are equal to zero and the system remains in this (stable) state.

are further denoted by dots, with filled dots indicating a stable fixed point, and
empty dots indicating an unstable fixed point.
We can immediately see from Figure 5.13 that the stability landscape is much
more complex than the one of the true bistable system, with high-degree poly-
nomial solution lines, and with four rather than three fixed points. Interest-
ingly, the system correctly identifies that there are two stable fixed points re-
lating to the healthy state (x1 = x2 = 4.51, x3 = x4 = 1.67) and the unhealthy state
(x1 = x2 = 1.71, x3 = x4 = 4.47), and that there is an unstable fixed point ap-
proximately half-way between those two (x1 = x2 = 2.99, x3 = x4 = 3.19). Despite
having an entirely different functional form, the estimated model does capture
two stable fixed points (characteristic 1) and the approximate position of those
fixed points (characteristic 2). This shows that Model G performs well in captur-
ing the characteristics of the system for emotion values that were observed in the
time series, that is, near the two stable fixed points.
Crucially, however, we cannot say that this system recovers the global dy-
namics of the true system, not least because the system contains an additional
unstable fixed point at (x1 = x2 = 5.31, x3 = x4 = 7.94), which is not present in
the true bistable system. The presence of this unstable fixed point means that if,
for instance, both negative and positive emotions take on a high value simulta-
neously, then the system enters an unstable region and diverges to infinity. If we
examine the behavior of the system even further outside the range of observed
values (−∞ > X > 0 and 10 < X < ∞) even more fixed points and regions of stabil-
ity and instability can be found. We can further demonstrate these dynamics by
generating data from Model G. Figure 5.14 shows a time series generated from

151
5. Recovering Bistable Systems from Psychological Time Series

the difference-form of Model G (i.e., with a step size equal to that of the observed
data).4 We see that the process moves between the healthy and unhealthy fixed
point for the first ten days, exhibiting the bistable behavior we see in the true
system. However around the eleventh day, the stochastic input is large enough
to move the system to an unstable region in the vector field and which leads the
system to diverge.

(a) (b)
8 Cheerful
Content
Anxious
Sad
6
Emotion Strength

0 5 10 14
Days

Figure 5.14: Data generated from the estimated DE model, with the same initial values as the “ideal”
data

Note that the complexity of the final model here is not a result of over-fitting
the data, as we performed model selection based on the out-of-bag R2 , an ap-
proximation of the out-of-sample R2 . Rather, the complexity of this model can
be attributed to two factors. First, due to the low sampling frequency, our ap-
proximation of the derivative at each point in time is poor. The second, as we dis-
cussed in Section 5.3.6.3, is that given the spacing between observations, the best
one can hope for is to approximate the integral solution to the data-generating
equation, which is likely of a highly complex functional form. The ability of the
misspecified Model G to reproduce some characteristics in regions where we have
observed data can be attributed to the high flexibility afforded by the many non-
linear terms. In that sense, this behavior is highly comparable to the problem of
using a high-degree polynomial regression model to make predictions outside of
the range of observed values. The vector field in Figure 5.13 is constructed by
obtaining predicted values for the derivatives across a grid of input values and
as such, it is unsurprising that the vector field is accurate where the input values
are close to the observed data, and inaccurate elsewhere.
In summary, we do not at all recover the functional form or parameters of
the system; we do recover some of the global characteristics and behavior of the
4 This is obtained by re-fitting the differential equation using the unscaled difference x
i,t+1 − xi,t as
the outcome variable, leading to equivalent results with parameters approximately scaled by dt = 90.
The residual variance used is the estimated residual variance scaled down to .65 the magnitude, to
account for the non-normal residual distribution. Using the estimated residual standard deviation
results in shocks which immediately move the system into an unstable region.

152
 5.4. Recovering the Bistable Systems from ESM Data

system in the region where we have observations, capturing that there are two
stable fixed points and one unstable fixed point, and their locations. However,
the estimated model also implies the presence of at least one extra unstable fixed
point, which has major implications for the dynamics of the model, implying
divergent behavior. Thus, the estimated model implies fundamentally different
microdynamic and global characteristics. Based on the simulated data in Figure
5.14, it does not seem that we correctly capture the variability around these fixed
points, or the frequency of transitions, as any reasonable simulation of data from
this model eventually leads the system to diverge. Crucially, we fail in recov-
ering an interpretable approximation of the data-generating model. As such, it
is not feasible to assess whether there are truly suppressing effects between va-
lences and reinforcing effects within valences, or the relative size of these effects
(characteristics 5 and 6).

5.4.5 Summary: Analysis of ESM Time Series

In this section, we aimed to investigate to which extent lowering the sampling
frequency affects the the ability of our considered methods to recover the bistable
system. Our findings are summarized in Table 5.4.

(5
) ) 7)
) 4) f. (6 nt(
ty
(1 2) (3
) s( n ein
e
Siz onst
a
b ili i o n( n ce i tio r. /R ive - c
s t
sta sit ria an pp la m
e
Bi Po Va Tr Su Re Ti
Data Visualization X X X × × × ×
HMM X∗ X X X × × ×
Lag-0 / GGM × × × × × × X∗
Lag-1 / VAR(1) × × × × × × X∗
TVAR(1) X∗ X X X × × ×
DE-Estimation × × × × × × ×
Table 5.4: Summary of which method recovered which of the seven qualitative characteristics listed
in Section 5.2.3 from the ESM time series. The first four characteristics are global dynamics, the
last three are microdynamics. The check marks with asterisk indicate that the method includes the
characteristic as a model assumption, and can therefore not be considered recovered from the time
series.

Our main findings are that, in general, we remain able to recover global char-
acteristics of the system using simple methods, but that we are completely unable
to recover any of the microdynamics. We saw that each approach which aimed to
capture microdynamic characteristics either deteriorated dramatically in perfor-
mance (for the VAR and TVAR approaches) or broke down altogether (for the DE
model building approach) as soon as we applied them to a time series obtained
with a more realistic sampling frequency. This is despite the fact that the time
series we used in this section can be considered a highly idealized approxima-
tion to ESM time series, in terms of the number of observations and the quality

153
5. Recovering Bistable Systems from Psychological Time Series

of measurements, suggesting that sampling frequency is a fundamental barrier

to inference which needs further investigation. As a side result, we can say that
lowering the sampling frequency typically made it much more difficult to inter-
pret and understand the results of different methods: In particular for methods
which involved lagged relationships of some kind (i.e. the VAR, TVAR and DE
approaches), it was difficult to ascertain precisely what features of the system the
estimated parameters reflected.
The recovery of global characteristics was more successful. Using data visu-
alization and the Hidden Markov Model it was still possible in principle to learn
about the position, variance around and frequency of transitions between fixed
points, and the threshold estimate form the TVAR model succeeded in capturing
the unstable fixed point. Finally, the predictions made by the best-fitting differ-
ential equation model did allow us to get some tentative indication of bistable
behavior, and the possible location of stable fixed points. However, the resulting
model suffered from a high degree of complexity, limiting both substantive in-
terpretation and our ability to extrapolate the model parameters to predict the
behavior of the system under different conditions.
In summary, the results in this section call into question to what extent it is
possible to investigate moment-to-moment microdynamics using data sampled at
a rate typical of ESM studies. We have showed that interpreting model estimates
from ESM time series as reflecting the microdynamics can be highly misleading,
when the process of interest is varying at a higher frequency than the sampling
frequency. Although the recovery of global characteristics is more promising, we
remind the reader that the time series considered here is still highly idealized,
with essentially infinite sampling size, and so the performance of these methods
should be considered an upper bound on performance in any realistic situation.

5.5 Discussion
In this paper we explored to what extent dynamical systems models can be re-
covered from psychological time series by investigating two successive questions.
First, how well does a set of popular and more advanced methods recover (char-
acteristics of) a basic bistable system with an ideal data set sampled at extremely
high sampling frequency (every six seconds)? And second, how is the perfor-
mance of each method affected when reducing the sampling frequency to one
measurement every 90min, which is typical for ESM studies.
When analyzing the ideal time series we found that the popular VAR model
(and the GGM fitted on its residuals) can in principle not recover the global dy-
namics of the true bistable system, and only recovers some of its microdynamics.
However, we showed that descriptive statistics, data visualization and statistical
models which are based on mixtures (the HMM and threshold VAR) were able
to capture the global dynamics of the bistable system. The only method that re-
covered the full bistable system was a differential equation (DE) model building
procedure. Reducing the sampling frequency from every six seconds to every 90
minutes affected the considered methods differently. The VAR model and its ex-

154
 5.5. Discussion

tensions no longer recover any microdynamics, and the DE-estimation procedure

fails. However, descriptives, data visualization and appropriate statistical mod-
els still recover the global dynamics. Overall, our analysis therefore suggests that
it is neither possible to estimate dynamical systems directly from realistic time
series, nor is it possible to reliably infer its microdynamics from the parameter
estimates of statistical models.

5.5.1 Implications for Complex Systems Approaches to Study-

ing Mental Disorders
Our results raise fundamental questions about how to study mental disorders
from a complex systems perspective. First, they show that it is unclear what
exactly one can in principle conclude from statistical models estimated from
psychological time series about an underlying dynamical system. Clearly, these
models are always misspecified (i.e., do not include the true system as a special
case), so one cannot hope to directly recover the underlying dynamical system.
More surprisingly, however, recovering the qualitative characteristics of the true
system also turned out to be difficult. While it was possible to recover the global
dynamics, no statistical model correctly recovered the microdynamics. For exam-
ple, the VAR model fundamentally cannot capture the global characteristics (e.g.,
location of fixed points) of the true bistable system and only recovered some of its
microdynamics (e.g., reinforcing vs. suppressing effect between two variables).
This is a problem for the emerging framework of studying mental disorders as
complex systems, because one is typically interested in the microdynamics (the
“mechanics”) of a disorder because one hopes to intervene on them. In contrast,
it is usually less clear how interventions can target global dynamics, since they
can be seen as the aggregate behavior implied by the microdynamics. Especially
the failure of the popular VAR model to correctly recover the qualitative nature
of the microdynamics in the true model is concerning, because it calls into ques-
tion whether it allows any reliable conclusions about an underlying dynamical
system. It therefore seems to be an open question how useful VAR models and
other statistical models are to studying mental disorders from a complex systems
perspective.
Second, the analysis of the ESM time series raises the question of which pro-
cess can be recovered with which sampling frequency. While we were still able
to recover the global dynamics of the system, each method that provides some
approximation of the microdynamics was strongly affected by sampling only ev-
ery 90min instead of every six seconds. The qualitative characteristics of the VAR
and TVAR models were even less in agreement with those of the true model, and
the DE-estimation method, which was the only fully successful method in the
ideal data case, returned a model with uninterpretable parameters and incorrect
global- and microdynamics. Thus, our results suggest what also seems intuitive:
It is extremely difficult — perhaps impossible — to recover microdynamics at a
time scale that is much smaller than the sampling frequency. This intuition is
also in line with sampling theorems from the field of signal processing. For ex-
ample, the Nyquist-Shannon sampling theorem states that a sine wave (a process

155
5. Recovering Bistable Systems from Psychological Time Series

much simpler than our bistable system) that completes one cycle within, say, 2
minutes, has to be sampled at least every minute to be recovered (e.g., Marks,
2012; Papoulis & Pillai, 2002). This suggests that it is futile to try use a time series
sampled every 90 minutes to directly recover dynamics of emotions that operate
on a time scale of seconds or minutes (Houben, Van Den Noortgate, & Kuppens,
2015) or even from moment to moment (Wichers et al., 2015). However, this also
means that ESM time series can certainly be used to recover processes that unfold
at a time scale of several hours or days.
To summarize, we identified two fundamental barriers to studying mental
disorders from a complex systems perspective. First, even with extremely high
sampling frequency it is generally unclear how to make inferences from a statis-
tical model to an unspecified dynamical systems model. Second, the sampling
frequency of the data collection constrains the type of processes one can recover.
Specifically, a process can only be recovered if the sampling frequency is suf-
ficiently high. Clearly, these are profound problems every empirical discipline
struggles with and no simple answers can be expected. Indeed, they might imply
that studying some aspects of mental disorders will always remain out of reach.
That said, we believe that much progress can be made by studying mental disor-
ders as complex systems and that acknowledging and studying the above issues
allows one to do so more efficiently. As a way forward, in the following section we
suggest a new research strategy based on proposing substantively plausible dy-
namical systems, which opens up avenues to creatively tackle the two problems
identified in this section.

5.5.2 Moving Forward: Proposing Plausible Dynamical Sys-

tems Models
A more abstract perspective on the first problem identified in the previous sec-
tion is the following: We have parameters of a statistical model which we es-
timated from a time series sampled from some system, and we hope to infer
some characteristics (e.g., global or microdynamics) of the data-generating sys-
tem from them. The problem, however, is that the mapping from parameters of
statistical model to the parameters and structure (and the implied dynamics) of
the true model is unknown. Thus, this inference cannot be made. The main rea-
son this mapping is unknown is the trivial reason that no true dynamical system
model is specified.
We propose that, in order to overcome this fundamental problem, researchers
must begin the research process by proposing a “first guess” model of the dy-
namical system. While this is clearly difficult and the validity of this model
should certainly be questioned, this approach has one major advantage: It is
much clearer how to draw conclusions from descriptive statistics, data visualiza-
tions or statistical models about the underlying dynamical systems model. This
is because one can generate time series from the “first guess” model and fit a sta-
tistical model of choice; that way, one always knows which statistical model is
implied by the dynamical systems model. This implied model can then be com-
pared to the model fitted to corresponding empirical data. If the implied model

156
 5.5. Discussion

and the empirical model are in agreement, we have tentative evidence that the dy-
namical system model is correct; if not, we can use the nature of the disagreement
to improve the dynamical system model. Clearly, this modeling approach, which
is typical to more quantitative disciplines such as physics, chemistry and biology,
is different to the statistical modeling framework most psychological researchers
are familiar with. On the one hand these formal dynamical systems models
are harder to build, since they cannot be estimated directly from the data. On
the other hand, they are powerful enough to be plausible for complex phenom-
ena such as mental disorders, and have additional benefits such as synthesizing
knowledge, revealing unknowns, laying open hidden assumptions and enabling
checking of the internal consistency of a model (Epstein, 2008; Lewandowsky &
Farrell, 2010; Smaldino, 2017).
This modeling approach also allows to tackle the problem of sampling fre-
quencies that are too low to recover the process of interest directly, as one can
generate a time series from the specified dynamical systems model and reduce
the sampling frequency to a level that is also available in empirical data. Then,
similarly to above, one can again compute the statistical model of choice that
is implied by the dynamical systems model with a given sampling frequency,
compare it to the corresponding model fit on empirical data, and in the case of
disagreement adapt the dynamical systems model accordingly. Of course, this
approach is not a panacea. Less information is available when the sampling fre-
quency is low, which makes model identification more difficult. However, spec-
ifying an initial dynamical systems model allows one to gauge how difficult it is
to recover a given type of process on a given time scale with a given sampling
frequency.
In addition, starting out with a dynamical systems model also allows to study
the measurement function that defines the mapping from the variables in the
dynamical systems to the obtained measurements, a topic we only touched on
briefly in this paper. In our emulated ESM time series we took the measurement
function to return the exact values of variables at the time point of measurement.
However, different questions imply different measurement functions. For exam-
ple, if the phrasing of a particular question refers to the entire period since the
last measurement, one could instead define the measurement as a function of the
variable values since the last measurement, such as the average. Next to formaliz-
ing which experiences an ESM question refers to exactly, defining a measurement
function also allows to formalize known response and memory biases, such as the
recency effect (Ebbinghaus, 1913/2013).
Finally, having a plausible dynamical systems model allows one to explicitly
address a behavior that has been largely ignored in the psychological time series
modeling literature, that is, the fact that humans sleep. Sleep interacts with es-
sentially everything physiological and psychological, is part of the definition of
several mental disorders (e.g., Major Depression) and related to many more (e.g.,
Walker, 2017). Thus, for many mental disorders, it seems necessary for a plausi-
ble model to include sleep. This may also allow using existing data in new ways,
because data around the “day-night shift” does not have to be excluded anymore,
but instead can be used to test hypotheses about the sleep-related assumptions

157
5. Recovering Bistable Systems from Psychological Time Series

of the dynamical systems model.

Clearly, this brief outline of the proposed modeling approach leaves many im-
portant questions unanswered: Where should the initial “first guess” dynamical
system come from? How to formalize different substantive aspects in a dynami-
cal systems model? Which statistical models should one choose to test which im-
plications of the dynamical systems model? Given some disagreement between
predicted and observed statistics, how should one adapt the existing dynamical
systems model? These and other questions are difficult ones and answering them
requires the combined creativity of a large research community. Nonetheless, a
more detailed account of our proposed new modeling approach would be desir-
able. However, since such a detailed account is beyond the scope of the present
work, we address it in a forthcoming paper.

5.5.3 Limitations
Several limitations of our work require discussion. First, our goal was to explore
to which extent one can recover (bistable) dynamical systems for mental disor-
ders from psychological time series. However, we only studied a single bistable
system. Therefore, it could be that the fundamental problems identified in the
paper and summarized in Section 5.5.1 are in fact a particularity of the chosen
bistable system. This, however, seems extremely unlikely: First, because we iden-
tify the problems in our paper as examples of well-known issues such as model
misspecification and sampling systems with a sampling frequency that is suffi-
cient for recovery. Second, the bistable system we chose is arguably the simplest
bistable system for four variables one can find. Choosing a different model there-
fore results most likely in choosing a more complex model, and our intuition is
that the methodological difficulties discussed in this paper become more and not
less relevant in such models.
Second, a more specific criticism of our bistable system could be that the time
scale of the process is unrealistically small, and we therefore exaggerated the
problem of recovering dynamics of psychological processes from ESM time se-
ries. We agree that it is possible that some psychological processes are easier
to recover from ESM data than the dynamical system used in this paper. Thus,
strictly speaking, we only showed that it is impossible to recover a system if the
sampling frequency does not appropriately match the time scale of the system. In
principle, it is therefore an open question whether there is a mismatch between
the time scale of the system of interest and the available sampling frequency.
However, intuition — and the sampling theorems such as the one mentioned in
Section 5.5.1 — strongly suggest that it is impossible (or at least very difficult) to
recover a process that operates at a time scale of seconds or minutes from an ESM
time series that is measured every 1.5 hours. Clearly, however, our investigation
is only a first treatment of the important topic of sampling frequency, and much
work on it is required to establish a tight connection between psychological time
series and dynamical systems models.
Third, one could reverse the argument in the previous two paragraphs and
argue that our model is so ideal that many analyses perform better than in most

158
 5.5. Discussion

realistic applications. This is certainly the case for the Threshold VAR model,
which performs well only because of the simple dynamics of the bistable system
as we discussed in Section 5.3.5. Other examples are the descriptive statistics
and data visualization which may not be as insightful if fixed points are closer to
each other and if there is more noise in the system. Also, the two-step approach
to estimating the differential equations in Section 5.3.6 may work less well for a
more complicated model. Thus, we would agree with this assessment, however
chose to use a simple bistable system in order to make the paper more accessible
to applied researchers.
Fourth, we analyzed a bistable system whose structural parameters do not
change over time. However, much of the framework of considering mental dis-
orders as complex systems is based on the idea that pathology is defined with
respect to a structural change in the underlying system, and therefore structural
change is of central interest. We expect that structural change renders the re-
covery of a system more difficult, and we therefore did not include this feature
in order to keep the paper at a reasonable length. However, we believe that fu-
ture methodological research into how to recover such structural changes both in
principle and with realistic time series would be extremely helpful to better un-
derstand phenomena such as early warning signals (Scheffer et al., 2009; van de
Leemput et al., 2014) and more generally structural change in mental disorders.
Fifth, in order to estimate a differential equation from data, we took a rather
simple two-step approach based on local linear approximation of the derivative
(cf. Boker, Deboeck, et al., 2010). This approach involves first estimating the
derivative itself using scaled difference scores, and then using this derivative as
an outcome variable in a regression model. While this method benefits from be-
ing extremely simple to implement, we could expect that it would perform poorly
in the presence of low sampling frequency as the quality of the derivative approx-
imation degrades (as we noted in Section 5.3.6.3 and observed in Section 5.4.4).
There are multiple alternative approaches to estimating DE equations which we
did not consider here. For example, approaches based on numerical integration
of the DE equation during estimation, such as implemented in dynR (Ou et al.,
2019) and stan (Carpenter et al., 2017) (with additional functionality in the ct-
sem package; Driver et al., 2017) may in general perform better than the two-step
procedure when the sampling frequency is low. However, for the analysis shown
in the present paper, neither the ctsem nor dynr package performed better than
the two-step approach. In general, however, more research is needed to map out
which method deals best with the problem of low sampling frequencies.
Lastly, throughout our paper we studied how well certain analysis methods
can recover the true bistable system in principle. We did this by studying the pop-
ulation properties of these methods, that is, the situation in which one has essen-
tially infinite sample size, which we approximated with a huge number (201600)
of measurements. This was necessary in order to study the more fundamental
questions of (1) whether a given method can recover our bistable system in prin-
ciple and (2) whether a given method can recover our bistable system based on
a time series with realistic sampling frequency. We did this because it would be
meaningless to study the performance of a method as a function of sample size,

159
5. Recovering Bistable Systems from Psychological Time Series

if the method alredy fails with infinite sample size. Clearly, however, to apply
any of the methods we studied in practice, one has to know how reliable they are
with which sample size, and much more research is necessary to map our these
sample size requirements (e.g., Dablander, Ryan, & Haslbeck, 2019).

5.5.4 Summary
In the present paper we identified two fundamental problems involved in study-
ing mental disorders from a complex systems perspective: first, it is generally
unclear what to conclude from a statistical model about an unspecified underly-
ing complex systems model. Second, if the sampling frequency of a time series
is not high enough, it is futile to attempt to recover the microdynamics of the
underlying complex system. In response to these problems, we proposed a new
modeling strategy that takes an initial substantively plausible dynamical systems
model as a starting point, and develops the dynamical systems model by testing
its predictions. In this approach it is much clearer what we can learn from data
and statistical models about an underlying dynamical system, and in addition it
provides avenues to move the field forward by formalizing the sampling process,
measurement, response and memory biases, measurement reactivity and the in-
fluence of sleep.

160
 5.A. Determining Fixed Points

Appendix 5.A Determining Fixed Points

In this section we show how to compute the fixed points of the deterministic
part of our model, which we report in Section 5.2.1. The fixed points of a
set of differential equations is found by setting all equations to zero and solv-
ing that system. In our case this means solving the nonlinear system of equations:

4
X
0 = r1 x1 + C1j xj x1 + a1
j=1
4
X
0 = r2 x2 + C2j xj x2 + a2
j=1
4
X
0 = r3 x3 + C3j xj x3 + a3
j=1
4
X
0 = r4 x4 + C4j xj x4 + a4
j=1

Since we have r1 , r2 = 1 and a = [1.6, 1.6, 1.6, 1.6] in all studied situations, we fill
in those values and write out the summation:

0 = x1 + C11 x1 x1 + C12 x1 x2 + C31 x1 x3 + C41 x1 x4 + 1.6

0 = x2 + C21 x2 x1 + C22 x2 x2 + C23 x2 x3 + C24 x2 x4 + 1.6
0 = r3 x3 + C31 x3 x1 + C32 x3 x2 + C33 x3 x3 + C34 x3 x4 + 1.6
0 = r4 x4 + C41 x4 x1 + C42 x4 x2 + C43 x4 x3 + C44 x4 x4 + 1.6

We can exploit the symmetries in r and C to simplify finding the fixed points.
The derivatives of x1 and x2 are actually identical, and the derivatives of x3 and
x4 are identical. Thus also their integrals are identical. Thus, we can substitute
x1 into x2 , and x3 into x4 to arrive at a simpler 2-dimensional system. Making the
substitutions, and filling in the parameter values, the differential equations then
reduce to

0 = 1x1 − 0.2x12 + 0.04x12 − 0.4x1 x2 + 1.6

0 = r3 x3 − 0.2x32 − 0.4x3 x1 + 0.04x32 + 1.6

where r3 is the stress level for which the fixed points should be computed.
We now solve these systems for a number of stress values (r3 ) using Mathe-
matica (Wolfram Research, Inc., 2019). This way, we computed the fixed points
shown in Table 5.5, which are displayed in panel (a) of Figure 5.1 in Section 5.2.1.

161
5. Recovering Bistable Systems from Psychological Time Series

Stress Healthy:PE Healthy:NE Unhealthy:PE Unhealthy:NE Unstable:PE Unstable:NE

0.90 5.28 1.15
0.91 5.26 1.16
0.91 5.24 1.17
0.92 5.22 1.18
0.93 5.19 1.19
0.93 5.17 1.20
0.94 5.15 1.22
0.95 5.12 1.23
0.95 5.10 1.24
0.96 5.08 1.26
0.90 5.28 1.15
0.91 5.26 1.16
0.91 5.24 1.17
0.92 5.22 1.18
0.93 5.19 1.19
0.93 5.17 1.20
0.94 5.15 1.22
0.95 5.12 1.23
0.95 5.10 1.24
0.96 5.07 1.26 1.83 3.96 2.03 3.66
0.97 5.05 1.27 1.66 4.25 2.25 3.30
0.97 5.02 1.29 1.57 4.42 2.39 3.22
0.98 4.99 1.31 1.50 4.56 2.50 3.10
0.99 4.96 1.33 1.45 4.69 2.61 2.99
0.99 4.92 1.34 1.40 4.79 2.71 2.89
1.00 4.89 1.36 1.36 4.89 2.80 2.80
1.01 4.85 1.39 1.33 4.98 2.90 2.72
1.01 4.80 1.41 1.30 5.06 2.99 2.64
1.02 4.76 1.44 1.27 5.15 3.09 2.56
1.03 4.71 1.47 1.24 5.23 3.19 2.48
1.03 4.65 1.50 1.22 5.30 3.29 2.40
1.04 4.59 1.54 1.19 5.38 3.40 2.32
1.05 4.51 1.58 1.17 5.45 3.52 2.23
1.05 4.41 1.65 1.15 5.52 3.67 2.12
1.06 4.24 1.75 1.13 5.59 3.87 1.98
1.06 1.12 5.63
1.07 1.11 5.66
1.07 1.09 5.72
1.08 1.08 5.79
1.09 1.06 5.85
1.09 1.04 5.91
1.10 1.03 5.98

Table 5.5: Fixed points of the emotion model for different values of stress (rows), rounded to two
decimals. The 2nd and 3rd columns refer to the fixed points of the healthy fixed points for positive
and negative emotions; the 4th and 5th columns reger to the unhealthy fixed points; and the last two
columns refer to the unstable fixed point.

162
 5.B. Mean-Switching Hidden Markov Model

Appendix 5.B Mean-Switching Hidden Markov

Model
In this appendix we provide additional details with respect to the specification of
the mean-switching Hidden Markov Model, and using model selection to obtain
the number of components, described in Section 5.3.2

5.B.1 Model Specification

The mean-switching Hidden Markov Model is denoted
T −1
Y
P (X, S|µ, σ ) = πi N (X1 ) Aji N (Xt+1 ),
t=1

where X = {X1 , . . . , XT } is a matrix of p−variate elements Xj , S ∈ {1, . . . , K}T is a

vector of length T indicating the state at each time point, πi is the probability of
being in state i ∈ {1, . . . , K}, Ai,j is the probability of transitioning from state i to
state j, and µ, σ parameterize the multivariate Gaussian distribution N with zero
covariances.
In Section 5.3.2 we chose K = 2 components, and fix the covariances of the
Gaussian distribution to zero. Since we model four variables, this gives us 2 × 4
means and 2 × 4 standard deviations. The transition matrix A has three param-
eters since the last one is determined by the remaining three. Similarly, the
marginal probabilities π1 , π2 are determined by A and therefore do not count
as additional parameters. We therefore fit a model with 19 freely estimated pa-
rameters.

5.B.2 Model Selection for Mean-Switching HMM

In Section 5.3.2 we inserted bistability as an assumption in the model by specify-
ing that the HMM exhibits two states, and therefore the HMM does not provide
us any evidence with respect to which number of states represents the data best.
This can be done by performing model selection between HMMs with different
numbers of states.
A popular way to select between mean-switching HMMs / Gaussian mixtures
is the Bayesian Information Criterion (BIC) (Schwarz et al., 1978), because it has
been shown to consistent in estimating Gaussian mixtures (Leroux, 1992), and
has outperformed other information criteria (including the AIC) in simulations
(R. J. Steele & Raftery, 2010). Here we fit HMMs with K ∈ {1, . . . , 10} and report
the BIC values in Figure 5.15.
We see that the BIC is highest for K = 1 and then decreases for larger K, however
the the change in BIC becomes less and less when adding additional states. Since
we know from the true bistable system that the number of states is K = 2, we see
that the BIC does not select the true number of states. The reason is that the BIC
has been shown to be a consistent estimator of K if the data is generated from

163
5. Recovering Bistable Systems from Psychological Time Series

2500000
●

2000000

1500000
BIC ●

1000000 ●

●
500000 ●
●
●
●
● ●
0

1 2 3 4 5 6 7 8 9 10
Number of States

Figure 5.15: The figure depicts BIC values as a function of the number of states K, for HMMs fitted
to the ideal data.

a Gaussian mixture. However, in the present case the data is generated from
a bistable dynamical system. This failed attempt at model selection based on
statistical models again highlights the problems of using misspecified statistical
models to make inferences about dynamical systems models.

Appendix 5.C Data Generated from Estimated Mod-

els
In this Appendix we show data generated from estimated models for the time
period of two weeks of the original time series.

5.C.1 Mean Switching Hidden Markov Model

Figure 5.16 displays a time series of two weeks generated from the Mean switch-
ing HMM estimated in Section 5.3.2:

10

8
Emotion Strength

Cheerful
6
Content
Anxious
4 Sad

−2

0 5 10 14
Days

Figure 5.16: A time series of two weeks generated from the HMM estimated in Section 5.3.2.

164
 5.C. Data Generated from Estimated Models

The generated time series looks similar to the original data in that it switches
between the two fixed points at around (1,6) and (6,1). However, there are also
differences. In the original data there are less switches that lead to a long-lived
change in fixed point, but more switches that are very short-lived. Second, due to
the form of the Mean-Switching HMM, there are no “intermediate” observations
leading from one fixed point to the other. These observations exist in the original
time series (see panel (b) in Figure 5.2).

5.C.2 First-order Vector Autoregressive (VAR(1)) model

Figure 5.17 displays a time series of two weeks generated from the VAR(1) model
in Section 5.3.4:

15
Emotion Strength

10
Cheerful
Content
Anxious
5 Sad

−5

0 5 10 14
Days

Figure 5.17: A time series of two weeks generated from the VAR(1) model estimated in Section 5.3.4.

The generated data does not show bistability, which is expected because the
VAR(1) model exhibits only a single fixed point. What looks approximately like
oscillating behavior is a result of the high auto-regressive effects present in the
estimated VAR(1) model: given a stochastic input, the high auto-regressive ef-
fects ensure that the system is slow to eventually return to equilibrium. This
oscillating behavior is also evident in the eigenvalues of Φ, which consist of one
complex conjugate pair (Strogatz, 2015).

5.C.3 Threshold VAR(1) Model

Figure 5.18 displays a time series of two weeks generated from the TVAR(1)
model in Section 5.3.5. The data generated from the TVAR(1) model looks similar
to the original time series in that the position of the fixed points and the variance
around them is very similar. However, the system seems to switch less often be-
tween states, and similarly to the data generated from the HMM above, there are
much fewer observations on the transitions between states.

165
5. Recovering Bistable Systems from Psychological Time Series

10

8
Emotion Strength

Cheerful
6
Content
Anxious
4 Sad

−2

0 5 10 14
Days

Figure 5.18: A time series of two weeks generated from the TVAR(1) model estimated in Section 5.3.5.

Appendix 5.D Residual Partial Correlations

TVAR(1)

Residual Healthy (a) Residual Unhealthy (b)

Cheerful 0.02 Content Cheerful 0.01 Content

−0.05 −0.05

−0.05 −0.05 −0.05 −0.05

−0.05 −0.05
Anxious Sad Anxious 0.02 Sad

Figure 5.19: Residual partial correlation networks for both regimes in the TVAR model described in
Section 5.3.5 in the main text.

Appendix 5.E Differential Equation Model Building

In this appendix we present additional information relating to the two-step DE
model building procedure utilized in Sections 5.3 and 5.4. This includes details
on how model fit is computed, as well as full model fit results and parameter
estimates for each of the models described in the main text.

5.E.1 Evaluating Model Fit

The fit of each model is evaluated with the mean out-of-bag explained variance,
referred to throughout as R2 . This metric is calculated using 10-fold cross-

166
 5.E. Differential Equation Model Building

validation. First, the given dataset is randomly partitioned into ten mutually
exclusive training and test sets. Second, for each partitioned dataset, regression
models A through G, (defined by the expression in the second column of Table
5.6) are fit to the training set four times, once each of the four outcome variables
dxiˆ/dt, ∀i ∈ {1, 2, 3, 4}. Third, the resulting parameters are then used to predict the
values of the outcome variable in the test set dxiˆ/dt. The variance of the resulting
residuals V AR(dxi /dt − dxiˆ/dt) is then divided by the variance of the outcome
variable in the test set, V AR(dxi /dt) yielding an out-of-bag variance explained
for variable i based on model m in partition k, R2i,k,m . Averaging the explained
variance across each of the partitions yields an average explained variance for
variable i in model m, R2i,m , and averaging this number across all four outcome
variables yields the average out-of-bag explained variance for model m.

5.E.2 Ideal Data

In Table 5.6 we show the fit of models A through G for the ideal dataset analysis
in Section 5.3.6. In Table 5.7 we show the full parameter estimates, standard
errors and p−values for the selected model, Model C.

dxi,t
Model dt ∼ a + ri xi + . . . q R2
P
A j,i rj xj 5 .04464
P Pp
B Rij j + j Cij xj xi
x 9 .06874
P j,i Pp
C j,i Rij xj + (j,k) βj xj xk 15 .06870
P Pp Pp 3
D j,i Rij xj + (j,k) βj xj xk + j γj xj 19 .06871
P Pp Pp 3 P p
E j,i Rij xj + (j,k) βj xj xk + j γj xj + j,k,l ζj (xj xk xl ) 23 .06870
P Pp Pp 3 Pp
F j,i Rij xj + (j,k) βj xj xk + j γj xj + (j,k,l) ζj (xj xk xl ) 35 .06860
P Pp Pp 3 Pp
j,i Rij xj + (j,k) βj xj xk + j γj xj + (j,k,l) ζj (xj xk xl )
G Pp 70 .06846
+ (j,k,l,m) ηj (xj xk xl xm )

Table 5.6: Model fit results for each of the seven models described in text in Section 5.3.6 for the ideal
dataset. The second column gives the model equation for each variable, q denotes the number of pa-
rameters estimated per univariate regression model, and the final column indicates R2 , the explained
variance, as calculated based on the prediction error on a hold-out set, using 10-fold cross-validation.

167
5. Recovering Bistable Systems from Psychological Time Series

dx1 /dt dx2 /dt dx3 /dt dx4 /dt

Est SE p Est SE p Est SE p Est SE p
a 1.40 0.13 <.01 1.37 0.12 <.01 1.25 0.12 <.01 1.27 0.12 <.01
x1 0.88 0.05 <.01 0.03 0.02 0.19 -0.02 0.02 0.33 0.04 0.02 0.05
x2 0.02 0.02 0.34 0.95 0.05 <.01 0.05 0.02 0.02 -0.01 0.02 0.57
x3 -0.01 0.02 0.72 0.01 0.02 0.68 0.96 0.05 <.01 0.08 0.02 <.01
x4 0.01 0.02 0.51 <.01 0.02 0.80 0.04 0.02 0.10 0.91 0.05 <.01
x1 × x1 -0.18 0.01 <.01 - - - - - - - - -
x1 × x2 0.04 0.01 <.01 0.03 0.01 <.01 - - - - - -
x1 × x3 -0.17 0.01 <.01 - - - -0.18 0.01 <.01 - - -
x1 × x4 -0.18 0.01 <.01 - - - - - - -0.19 0.01 <.01
x2 × x2 - - - -0.19 0.01 <.01 - - - - - -
x2 × x3 - - - -0.19 0.01 <.01 -0.19 0.01 <.01 - - -
x2 × x4 - - - -0.19 0.01 <.01 - - - -0.18 0.01 <.01
x3 × x3 - - - - - - -0.19 0.01 <.01 - - -
x3 × x4 - - - - - - 0.03 0.01 <.01 0.02 0.01 <.01
x4 × x4 - - - - - - - - - -0.18 0.01 <.01

Table 5.7: Full parameter estimates, standard errors and p-values for Model B in Section 5.3.6, for
the DE model fit to ideal data.

5.E.3 ESM Data

In Table 5.8 we show the fit of models A through G for the emulated ESM dataset
analysis, from Section 5.4.4. In Table 5.9 we show the full parameter estimates,
standard errors and p−values for the selected model, Model G.

dxi,t
Model ∼ a + ri xi + . . .
dt q R2
P
A j,i rj xj 5 0.13991
P Pp
B j,i Rij xj + j Cij xj xi 9 0.16827
P Pp
C j,i Rij xj + (j,k) βj xj xk 15 0.16928
P Pp Pp 3
D j,i Rij xj + (j,k) βj xj xk + j γj xj 19 0.19455
P Pp Pp 3 P p
E j,i Rij xj + (j,k) βj xj xk + j γj xj + j,k,l ζj (xj xk xl ) 23 0.19801
P Pp Pp 3 Pp
F j,i Rij xj + (j,k) βj xj xk + j γj xj + (j,k,l) ζj (xj xk xl ) 35 0.19940
P Pp Pp 3 Pp
j,i Rij xj + (j,k) βj xj xk + j γj xj + (j,k,l) ζj (xj xk xl )
G Pp 70 0.20420
+ (j,k,l,m) ηj (xj xk xl xm )

Table 5.8: Model fit results for each of the seven models described in text, for the ESM time series,
described in Section 5.4. The second column gives the model equation for each variable, q denotes the
number of parameters estimated per univariate regression model. The final two columns indicate R2 ,
the explained variance, as calculated based on the prediction error on a hold-out set, using 10-fold
cross-validation, for the snapshot ESM data and the mean-aggregated ESM data, respectively. R2 for
Model G in the mean-aggregated ESM data case was not available due to multicollinearity problems
encountered when fitting the model.

168
 5.E. Differential Equation Model Building

dx1 /dt dx2 /dt dx3 /dt dx4 /dt

Est SE p Est SE p Est SE p Est SE p
(Intercept) -0.07 0.73 0.92 -0.16 0.73 0.82 0.31 0.74 0.68 0.20 0.74 0.78
x1 0.19 0.31 0.53 0.16 0.31 0.61 -0.09 0.31 0.78 -0.12 0.31 0.71
x2 0.07 0.30 0.80 0.18 0.30 0.54 -0.29 0.30 0.34 -0.21 0.30 0.48
x3 -0.06 0.31 0.86 0.01 0.31 0.96 -0.05 0.31 0.88 -0.07 0.31 0.82
x4 -0.01 0.31 0.97 -0.04 0.31 0.89 0.01 0.31 0.98 0.10 0.31 0.74
x1 × x1 -0.02 0.06 0.78 -0.01 0.06 0.87 -0.03 0.06 0.65 -0.01 0.06 0.81
x1 × x2 -0.11 0.09 0.21 -0.10 0.09 0.22 0.14 0.09 0.10 0.13 0.09 0.13
x1 × x3 -0.02 0.10 0.86 -0.04 0.10 0.66 <.01 0.10 0.99 0.02 0.10 0.80
x1 × x4 -0.05 0.10 0.59 <.01 0.10 0.99 0.01 0.10 0.94 <.01 0.10 0.98
x2 × x2 0.01 0.06 0.83 -0.02 0.06 0.76 0.02 0.06 0.72 0.01 0.06 0.82
x2 × x3 <.01 0.10 0.98 -0.01 0.10 0.94 0.05 0.10 0.58 0.06 0.10 0.57
x2 × x4 -0.01 0.10 0.95 -0.05 0.10 0.60 0.06 0.10 0.57 0.02 0.10 0.86
x3 × x3 0.02 0.06 0.68 0.01 0.06 0.87 -0.01 0.06 0.87 0.02 0.06 0.73
x3 × x4 0.01 0.09 0.92 0.01 0.09 0.92 0.01 0.09 0.94 -0.04 0.09 0.62
x4 × x4 0.02 0.06 0.80 0.03 0.06 0.68 -0.03 0.06 0.69 -0.03 0.06 0.68
x1 × x1 × x1 <.01 0.01 0.71 <.01 0.01 0.92 0.01 0.01 0.38 0.01 0.01 0.33
x1 × x1 × x2 0.02 0.01 0.19 0.01 0.01 0.48 -0.01 0.01 0.33 -0.02 0.01 0.15
x1 × x1 × x3 <.01 0.01 0.95 0.01 0.01 0.61 0.01 0.01 0.64 <.01 0.01 0.82
x1 × x1 × x4 <.01 0.01 0.78 <.01 0.01 0.71 <.01 0.01 0.73 <.01 0.01 0.81
x1 × x2 × x2 <.01 0.01 0.87 0.01 0.01 0.49 -0.01 0.01 0.39 <.01 0.01 0.76
x1 × x2 × x3 0.01 0.02 0.58 0.01 0.02 0.69 -0.02 0.02 0.32 -0.02 0.02 0.27
x1 × x2 × x4 0.02 0.02 0.25 0.02 0.02 0.22 -0.03 0.02 0.21 -0.02 0.02 0.34
x1 × x3 × x3 <.01 0.01 0.79 <.01 0.01 0.96 <.01 0.01 0.81 <.01 0.01 0.98
x1 × x3 × x4 0.01 0.02 0.73 <.01 0.02 0.80 <.01 0.02 0.99 <.01 0.02 0.99
x1 × x4 × x4 <.01 0.01 0.98 -0.01 0.01 0.62 <.01 0.01 0.80 <.01 0.01 0.75
x2 × x2 × x2 <.01 0.01 0.92 <.01 0.01 0.96 <.01 0.01 0.98 <.01 0.01 0.82
x2 × x2 × x3 <.01 0.01 1.00 <.01 0.01 0.87 <.01 0.01 0.79 <.01 0.01 0.79
x2 × x2 × x4 -0.01 0.01 0.70 <.01 0.01 0.90 <.01 0.01 0.89 <.01 0.01 0.96
x2 × x3 × x3 <.01 0.01 0.92 <.01 0.01 0.93 <.01 0.01 0.89 -0.01 0.01 0.50
x2 × x3 × x4 -0.01 0.02 0.70 <.01 0.02 0.79 <.01 0.02 0.83 0.01 0.02 0.59
x2 × x4 × x4 <.01 0.01 0.91 0.01 0.01 0.63 <.01 0.01 0.83 <.01 0.01 0.78
x3 × x3 × x3 <.01 0.01 0.95 <.01 0.01 0.93 <.01 0.01 0.97 <.01 0.01 0.51
x3 × x3 × x4 -0.01 0.01 0.40 -0.01 0.01 0.68 0.01 0.01 0.63 0.01 0.01 0.39
x3 × x4 × x4 0.01 0.01 0.44 <.01 0.01 0.76 -0.01 0.01 0.56 <.01 0.01 0.73
x4 × x4 × x4 -0.01 0.01 0.36 -0.01 0.01 0.42 0.01 0.01 0.33 0.01 0.01 0.40
x1 × x1 × x1 × x1 <.01 <.01 0.94 <.01 <.01 0.71 <.01 <.01 0.91 <.01 <.01 0.48
x1 × x1 × x1 × x2 <.01 <.01 0.82 <.01 <.01 0.55 <.01 <.01 0.52 <.01 <.01 0.92
x1 × x1 × x1 × x3 <.01 <.01 0.99 <.01 <.01 0.74 <.01 <.01 0.46 <.01 <.01 0.40
x1 × x1 × x1 × x4 <.01 <.01 0.80 <.01 <.01 0.78 <.01 <.01 0.83 <.01 <.01 0.88
x1 × x1 × x2 × x2 <.01 <.01 0.25 <.01 <.01 0.23 <.01 <.01 0.15 <.01 <.01 0.36
x1 × x1 × x2 × x3 <.01 <.01 0.70 <.01 <.01 0.66 <.01 <.01 0.43 <.01 <.01 0.22
x1 × x1 × x2 × x4 <.01 <.01 0.33 <.01 <.01 0.76 <.01 <.01 0.84 <.01 <.01 0.81
x1 × x1 × x3 × x3 <.01 <.01 0.37 <.01 <.01 0.44 <.01 <.01 0.19 <.01 <.01 0.17
x1 × x1 × x3 × x4 <.01 <.01 0.27 <.01 <.01 0.16 <.01 <.01 0.36 <.01 <.01 0.19
x1 × x1 × x4 × x4 <.01 <.01 0.47 <.01 <.01 0.12 <.01 <.01 0.23 <.01 <.01 0.17
x1 × x2 × x2 × x2 <.01 <.01 0.35 <.01 <.01 0.51 <.01 <.01 0.43 <.01 <.01 0.52
x1 × x2 × x2 × x3 <.01 <.01 0.96 <.01 <.01 0.92 <.01 <.01 0.97 <.01 <.01 0.80
x1 × x2 × x2 × x4 <.01 <.01 0.82 <.01 <.01 0.35 <.01 <.01 0.28 <.01 <.01 0.44
x1 × x2 × x3 × x3 <.01 <.01 0.28 <.01 <.01 0.18 <.01 <.01 0.10 <.01 <.01 0.04
x1 × x2 × x3 × x4 <.01 <.01 0.39 <.01 <.01 0.16 <.01 <.01 0.27 <.01 <.01 0.12
x1 × x2 × x4 × x4 <.01 <.01 0.13 <.01 <.01 0.05 <.01 <.01 0.07 <.01 <.01 0.06
x1 × x3 × x3 × x3 <.01 <.01 0.76 <.01 <.01 0.90 <.01 <.01 0.99 <.01 <.01 0.98
x1 × x3 × x3 × x4 <.01 <.01 0.96 <.01 <.01 0.99 <.01 <.01 0.66 <.01 <.01 0.80
x1 × x3 × x4 × x4 <.01 <.01 0.82 <.01 <.01 0.94 <.01 <.01 0.74 <.01 <.01 0.88
x1 × x4 × x4 × x4 <.01 <.01 0.75 <.01 <.01 0.53 <.01 <.01 0.52 <.01 <.01 0.59
x2 × x2 × x2 × x2 <.01 <.01 0.52 <.01 <.01 0.57 <.01 <.01 0.56 <.01 <.01 0.50
x2 × x2 × x2 × x3 <.01 <.01 0.76 <.01 <.01 0.60 <.01 <.01 0.64 <.01 <.01 0.49
x2 × x2 × x2 × x4 <.01 <.01 0.71 <.01 <.01 0.68 <.01 <.01 0.63 <.01 <.01 0.69
x2 × x2 × x3 × x3 <.01 <.01 0.47 <.01 <.01 0.32 <.01 <.01 0.30 <.01 <.01 0.44
x2 × x2 × x3 × x4 <.01 <.01 0.57 <.01 <.01 0.26 <.01 <.01 0.18 <.01 <.01 0.29
x2 × x2 × x4 × x4 <.01 <.01 0.41 <.01 <.01 0.43 <.01 <.01 0.37 <.01 <.01 0.36
x2 × x3 × x3 × x3 <.01 <.01 0.64 <.01 <.01 0.80 <.01 <.01 0.89 <.01 <.01 0.31
x2 × x3 × x3 × x4 <.01 <.01 0.28 <.01 <.01 0.48 <.01 <.01 0.76 <.01 <.01 0.36
x2 × x3 × x4 × x4 <.01 <.01 0.55 <.01 <.01 0.77 <.01 <.01 0.74 <.01 <.01 0.77
x2 × x4 × x4 × x4 <.01 <.01 0.68 <.01 <.01 0.95 <.01 <.01 0.86 <.01 <.01 0.91
x3 × x3 × x3 × x3 <.01 <.01 0.70 <.01 <.01 0.99 <.01 <.01 0.95 <.01 <.01 0.64
x3 × x3 × x3 × x4 <.01 <.01 0.44 <.01 <.01 0.83 <.01 <.01 0.96 <.01 <.01 0.83
x3 × x3 × x4 × x4 <.01 <.01 0.79 <.01 <.01 0.96 <.01 <.01 0.77 <.01 <.01 0.72
x3 × x4 × x4 × x4 <.01 <.01 0.75 <.01 <.01 0.76 <.01 <.01 0.48 <.01 <.01 0.54
x4 × x4 × x4 × x4 <.01 <.01 0.30 <.01 <.01 0.33 <.01 <.01 0.18 <.01 <.01 0.23

Table 5.9: Full parameter estimates, standard errors and p-values for Model G in Section 5.4.4, for
the DE model fit to the emulated ESM data.

169
5. Recovering Bistable Systems from Psychological Time Series

Appendix 5.F Additional Results ESM Time Series

In this appendix, we provide additional figures to visualize the results of the
statistical models fit to the ESM time series in Section 5.4.

Cheerful Content Anxious Sad

0.6 0.6 0.6 0.6
0.5 0.5 0.5 0.5
0.4 0.4 0.4 0.4
Density

Density

Density

Density
0.3 0.3 0.3 0.3
0.2 0.2 0.2 0.2
0.1 0.1 0.1 0.1
0.0 0.0 0.0 0.0

0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8

Emotion Intensity Emotion Intensity Emotion Intensity Emotion Intensity

Figure 5.20: The histograms of the emotion intensity of the four modeled emotions Cheerful, Content,
Anxious and Sad, for the ESM data

8 (a) 8 (b)

6 6
ρ = −0.97
Anxious
Content

4 4

2 ρ = 0.98 2

0 0

0 2 4 6 8 0 2 4 6 8
Cheerful Cheerful

Correlation Network (c) Partial Correlation Network (d)

Cheerful 0.98 Content Cheerful 0.51 Content

−0.97 −0.23

−0.97 −0.97 −0.24 −0.24

−0.97 −0.24
Anxious 0.98 Sad Anxious 0.52 Sad

Figure 5.21: Panel (a) shows the relationship between Content and Cheerful, two emotions with the
same valence, at the same time point The red line indicates the best fitting regression model, for ESM
time series. Similarly, panel (b) shows the relationship between Anxious and Content, two emotions
with different valence. Panel (c) displays the correlation matrix as a network, and panel (d) displays
the partial correlation matrix as a network.

170
 5.F. Additional Results ESM Time Series

State 1 State 2

8
Emotion Strength

6
Cheerful
Content
Anxious
4 Sad

0 5 10 14
Days

Figure 5.22: Time series of the four emotion variables, also shown in panel (a) of Figure 5.2, with
background color indicating whether a given time point is assigned to the first or second component
of the mean-switching HMM estimated from the ESM dataset.

171
 Chapter 6

Modeling Psychopathology:
From Data Models to Formal
Theories
Abstract

Over the past decade there has been a surge of empirical research inves-
tigating mental disorders as complex systems. In this paper, we investigate
how to best make use of this growing body of empirical research and move
the field toward its fundamental aims of explaining, predicting, and con-
trolling psychopathology. We first review the contemporary philosophy
of science literature on scientific theories and argue that fully achieving
the aims of explanation, prediction, and control requires that we construct
formal theories of mental disorders: theories expressed in the language of
mathematics or a computational programming language. We then investi-
gate three routes by which one can use empirical findings (i.e. data models)
to construct formal theories: (a) using data models themselves as formal
theories, (b) using data models to infer formal theories, and (c) comparing
empirical data models to theory-implied data models in order to evaluate
and refine an existing formal theory. We argue that the third approach is
the most promising path forward and conclude by expanding on this ap-
proach, proposing a framework for theory construction that details how to
best use empirical research to generate, develop, and test formal theories
of mental disorders.

This chapter has been adapted from: Haslbeck, J. M. B.*, Ryan, O.*, Robinaugh, D.J.*, Waldorp,
L.J. and Borsboom, D. (under review). Modeling Psychopathology: From Data Models to Formal
Theories. Pre-print: https://psyarxiv.com/jgm7f/. Author contributions: JMBH, OR and DJR are
considered joint first authors and contributed equally to this project. LJW and DB helped develop the
ideas in the project, discussed progress and provided textual feedback.

173
6. Modeling Psychopathology: From Data Models to Formal Theories

6.1 Introduction
Mental disorders are complex phenomena: highly heterogeneous and massively
multifactorial (e.g., Kendler, 2019). Confronted with this complex etiological and
ontological picture, researchers have increasingly called for approaches to psy-
chiatric research that embrace this complexity (Gardner & Kleinman, 2019). The
“network approach” to psychopathology addresses these calls, conceptualizing
mental disorders as complex systems of interacting symptoms (e.g., Borsboom &
Cramer, 2013; Schmittmann et al., 2013; Borsboom, 2017). From this perspec-
tive, symptoms are not caused by an underlying disorder, rather the symptoms
themselves and the causal relations among them constitute the disorder.
In recent years, empirical research within the network approach literature
has rapidly grown (for reviews see e.g., Robinaugh, Hoekstra, et al., 2019; Con-
treras et al., 2019). Most of this work employs statistical models that allow re-
searchers to study the multivariate dependencies among symptoms, thereby pro-
viding rich information about the relationships among those symptoms. How-
ever, this quickly expanding empirical literature has raised a critical question:
how can we best make use of this growing number of empirical findings to ad-
vance the fundamental aims of psychiatric science? This problem is not unique to
the network approach. Psychiatry has produced countless empirical findings, yet
genuine progress in our efforts to explain, predict, and control mental disorders
has remained stubbornly out of reach.
In this paper, we will argue that empirical research can best advance these
aims by supporting the development of scientific theories. We will begin in Sec-
tion 6.2 by discussing the nature of scientific theories and how they achieve the
explanation, prediction and control sought by psychiatric science. We will argue
that to fully achieve these aims, psychiatry requires theories formalized as math-
ematical or computational models. In Section 6.3, we will explore how models
estimated from data can best be used to develop formal theories. We examine
three possible routes from data model to formal theory: first, treating data mod-
els themselves as formal theories; second, drawing inferences from data models
to generate a formal theory; and third, using data models to develop formal the-
ories with an abductive approach. We will argue that the third approach is the
most promising path forward. In Section 6.4, we will expand on this approach
and propose a framework for theory construction, detailing how best to use em-
pirical research to advance the generation, development, and testing of scientific
theories of mental disorders.

6.2 The Nature and Importance of Formal Theories

In this section we will examine the nature of scientific theories and how they
support explanation, prediction, and control. We will begin by introducing four
key concepts that we will use throughout the remainder of the paper: theory,
target system, data, and data models. We will illustrate each of these concepts
using the example of panic disorder.

174
 6.2. The Nature and Importance of Formal Theories

Theory Data Model

C5
C1 V3
C3 V1

C2 C4
V2 V4

Target System Data

V1 V2 V3 V4
1.58 3.00 2.47 4.01
2.83 6.13 4.89 2.33
4.82 3.46 6.73 5.44
0.64 5.72 3.91 2.54
5.11 4.49 2.27 4.03

Figure 6.1: The figure illustrates the concepts target system, theory, data and data model. The target
system is the system consisting of interacting components that gives rise to phenomena. Phenomena
are robust features of the world captured by data models. Theories represent the structure of the
target system, proposing a set of components C and the relations among them and positing that they
give rise to the phenomena. Data for variables V are obtained by probing the target system.

6.2.1 Theories and Target Systems

Theories seek to explain phenomena: stable, recurrent, and general features of the
world (Bogen & Woodward, 1988; Haig, 2008, 2014) such as the melting point of
lead, the orbit of planets, and the tendency for some individuals to experience
recurrent panic attacks. Well developed theories can predict these phenomena
and show how they can be controlled. Although the precise nature of theories
remains a subject of ongoing debate among philosophers of science, the past half
century has seen a growing consensus that theories are best understood as mod-
els.1 Specifically, theories are models that aim to represent target systems: the
particular parts of the real world that give rise to the phenomena of interest. We
use the word “system” here because we assume that the part of the real world giv-
ing rise to any psychiatric phenomena can be partitioned into components and
the relations among them. We use the term “target”, because it is this system that
a theory aspires to represent (cf. Elliott-Graves, 2014).
In psychiatry, the most common phenomena to be explained are symptoms
1 The precise relationship between theories and models is muddied by inconsistent and often con-
flicting use of these terms across time, disciplines, and scientists (for a brief history of models and
their relation to theory, see Bailer-Jones, 2009). In this paper, we will adopt the perspective that the-
ories are models (Suárez & Pero, 2019). However, the core arguments presented in this paper do not
require this precise conceptualization of theories and would similarly hold for pragmatic accounts
that regard models as an intermediary between theory and the real world (e.g., Bailer-Jones, 2009;
Cartwright, 1983).

175
6. Modeling Psychopathology: From Data Models to Formal Theories

and syndromes. For example, researchers seek to explain the tendency for some
individuals to experience panic attacks and the tendency for recurrent panic at-
tacks to be accompanied by persistent worry about those attacks and avoidance
of situations in which they may occur (Spitzer, Kroenke, & Williams, 1980). The
target system in psychiatric research comprises the components of the real world
that give rise to these symptoms and syndromes, and may include genetic, neuro-
biological, physiological, emotional, cognitive, behavioral or social components.
Psychiatric theories aim to represent these target systems, positing a specific set
of components and relationships among them that give rise to the phenomena
of interest. For example, researchers have generated numerous theories of panic
disorder, specifying a set of components that they believe interact to give rise to
panic attacks and panic disorder. Among these, perhaps the most influential is
Clark’s cognitive model of panic attacks, which posits that “if [stimuli] are per-
ceived as a threat, a state of mild apprehension results. This state is accompanied
by a wide range of body sensations. If these anxiety-produced sensations are in-
terpreted in a catastrophic fashion, a further increase in apprehension occurs.
This produces a further increase in body sensations and so on round in a vicious
circle which culminates in a panic attack” (Clark, 1986). This cognitive theory of
panic attacks specifies components (e.g., bodily sensations and a state of appre-
hension) and the relations among them (e.g., the “vicious cycle” of positive causal
effects), positing that this is the target system that gives rise to panic attacks.
Because theories represent the target system, we can reason from theory in
order to draw conclusions about the target system. It is this capacity for surroga-
tive reasoning (Swoyer, 1991) that allows theories to explain, predict, and control.
For example, we can explain the rise and fall of predator and prey populations in
the real world by appealing to the relationships between components specified
in mathematical models representing these populations (H. I. Freedman, 1980;
Nguyen & Frigg, 2017). We can predict what will occur when two atoms collide
by deriving the expected outcome from models of particle physics (Higgs, 1964).
We can determine how to intervene to prevent panic attacks by appealing to the
relationships posited in the cognitive model of panic attacks, determining that
an intervention modifying a patient’s “catastrophic misinterpretations” should
prevent the “vicious cycle” between arousal and perceived threat, thereby cir-
cumventing panic attacks (Clark, 1986). It is this ability to support surrogative
reasoning that makes theories such powerful tools.

6.2.2 The Importance of Formal Theories

Surrogative reasoning relies on a theory’s structure: its components and the re-
lations among them (Pero, 2015; Suárez & Pero, 2019). This structure can be
expressed in a written or spoken language (i.e. verbal theory) or in the lan-
guage of mathematics or computation (i.e. formal theory). For example, a ver-
bal theory would state that the rate of change in an object’s temperature is
proportional to the difference between its temperature and the temperature
of its environment. A formal theory would instead express this relationship
as a mathematical equation, such as dT dT
dt = −k(T − E), where dt is the rate of

176
 6.2. The Nature and Importance of Formal Theories

change in temperature, T is the object’s temperature, and E is the tempera-

ture of the environment; or in a computational programming language, such as:
for(t in 1:end) { T[t+1] = T[t]-k*(T[t]-E) }.
Expressing a theory in a mathematical or computational programming lan-
guage gives formal theories many advantages over verbal theories (e.g., Smith &
Conrey, 2007; Epstein, 2008; Lewandowsky & Farrell, 2010; Smaldino, 2017).
There is one advantage especially relevant to the current paper: Formalization
enables precise deduction of the behavior implied by the theory. Verbal theories
can, of course, also be used to deduce theory-implied behavior. However, due
to the vagaries of language, verbal theories are typically imprecise, thereby pre-
cluding their ability to make exact predictions. For example, the verbal theory of
temperature cooling described in the previous paragraph allows for some general
sense of how the object’s temperature will evolve over time, but cannot be used
to make precise predictions about how it will change or where temperature will
be at any given point in time. Indeed, because of the imprecision of verbal the-
ories, there are often multiple ways in which those theories could be interpreted
and implemented, each with a potentially divergent prediction about how the
target system will evolve over time. Consider the interpersonal theory of suicide,
which posits that suicide arises from the simultaneous experience of perceived
burdensomeness and thwarted belongingness (Van Orden et al., 2010). This the-
ory fails to specify many aspects of this causal structure, such as the strength of
these effects or the duration for which they must overlap before suicidal behav-
ior arises (Hjelmeland & Loa Knizek, 2018). As a result, there are many possible
implementations of that verbal theory, each of which could potentially lead to
a different prediction about when suicidal behavior should be expected to arise.
This imprecision thus substantially limits the theories ability to support surrog-
ative reasoning and the degree to which we can empirically test the theory.
In contrast to most verbal theories, formal theories are precise in their im-
plementation as the mathematical notation or code in a computer programming
language forces one to be specific about the structure of the theory (e.g., speci-
fying the precise effect of one component on another). The precision of formal
theories allows for the provision of singular and precise predictions about how
the target system will behave. These predictions can either be obtained analyti-
cally from the mathematical equation or computed by implementing the formula
in a programming language. For example, we can use the formal theory of cool-
ing to predict the exact temperature of our object at any given point in time.
Similarly, a formal implementation of the interpersonal theory of suicide would
make precise predictions that could inform the prediction of suicide attempts. In
other words, formal theories substantially strengthen surrogative reasoning, the
very characteristic of scientific theories upon which we wish to capitalize.2
2 It is, of course, possible to express verbal theories with the same level of precision as is provided
by a mathematical equation (e.g., there are very few equations in the Principia, yet the laws Newton
describes are not lacking in precision). Nonetheless, the specificity required by mathematics or com-
putational programming makes them more amenable to expressing theories precisely and has the
considerable practical advantage of supporting the derivation of predictions from the theory.

177
6. Modeling Psychopathology: From Data Models to Formal Theories

6.2.2.1 A Formal Theory of Panic Disorder

The cognitive model of panic attacks posited by Clark is a verbal theory and is
limited by the imprecision characteristic of most verbal theories. Indeed, in two
recent papers, Fukano and Gunji (Fukano & Gunji, 2012) and Robinaugh and col-
leagues (Robinaugh, Haslbeck, et al., 2019) independently proposed two distinct
formal implementations of this theory, taking the verbal theory and expressing
it in differential equations. Notably, these distinct implementations of the same
verbal theory make divergent predictions about when panic attacks should occur,
illustrating the limitations of failing to precisely specify the theory (for further
detail, see Robinaugh, Haslbeck, et al., 2019).
In this paper, we will make extensive use of the formal theory proposed by
Robinaugh and colleagues. A complete description of the generation of this the-
ory can be found in Robinaugh, Haslbeck, et al. (2019). For the purposes of this
paper, it is sufficient to note that the aim in developing this model was to take
extant verbal theories, especially cognitive behavioral theories, and express them
in the language of mathematics. For example, Clark’s verbal theory posits that
a perception of threat can lead to arousal-related bodily sensations. However,
the actual form and strength of this effect remain unspecified. In our math-
ematical model, we used a differential equation to precisely define this effect:
dA
dt = α(νT −A). In this equation, there is a linear effect of Perceived Threat (T ) on
the rate of change of Arousal (A), with the strength of this effect specified by the
parameter ν. The product of ν and T is the value Arousal is pulled toward: if νT
is smaller than the current level of Arousal, dAdt will be negative and Arousal will
decrease toward νT ; if νT is greater than Arousal, dA dt is positive and Arousal in-
creases toward νT . Each model component was defined as a differential equation
in this way (see middle panel in Figure 6.2).

Causal Formal Simulated

Diagram Theory Data
1.00 Arousal
Perceived Threat
Arousal Panic Attack Escape
0.75
Schema
0.50

0.25
Arousal Perceived Escape
(A) Threat (E) 0.00
(T)
0 3 6 9

Figure 6.2: The left panel displays the key components of the theory proposed by Robinaugh,
Haslbeck, et al. (2019) at play during panic attacks: Arousal, Perceived Threat, Escape Behavior and
arousal schema. The arrows indicate the direct causal relationships which are posited to operate
between these components in the formal theory. The middle panel displays the formal theory that
specifies the precise nature of the relations among these components. The right panel depicts the
simulated behavior implied by the theory.

By specifying the structure of the theory in this way, we are able to solve
the system numerically, thereby deducing the theory’s predictions about how the

178
 6.2. The Nature and Importance of Formal Theories

target system will behave. For example, the theory shows that when the effect of
Arousal on Perceived Threat is sufficiently strong, the positive feedback between
these components is sufficient to send the system into runaway positive feed-
back, producing the characteristic surge of arousal, perceived threat, and escape
behavior that we refer to as a panic attack (see right panel in Figure 6.2). As this
example illustrates, specifying the theory as a computational model substantially
strengthens our ability to deduce the behavior implied by the theory. A full re-
alization of a theory’s usefulness thus all but requires that theory be formalized.
For that reason, we believe the ultimate goal of psychiatric research should not
only be the production of theories, but the production of formal theories.

6.2.3 Data and Data Models

Our brief overview of the philosophy of science literature on theories suggests
that if our aim is the explanation, prediction, and control of mental disorders,
what we are after are well-developed formal theories: mathematical or computa-
tional models that represent the target system. The key question then becomes:
how can we best determine such a formal theory?
The answer to this question will, of course, involve the collection and analysis
of data. Empirical data plays at least two key roles in the development of formal
theories. First, data gathered about the target system are key to establishing what
our theories must explain. Yet, theories typically do not aim to explain data di-
rectly. Data are sensitive to the context in which they are acquired and subject
to myriad causal influences that are not of core interest (Woodward, 2011). For
example, panic disorder researchers collect data from diagnostic interviews, self-
report symptom inventories, assessments of physiological arousal during panic
attacks, time-series data, and a host of other methods. Data about panic attacks
gathered using these methods will be influenced not only by the experienced at-
tacks, but also by recall biases, response biases, sensor errors, and simple human
error. Accordingly, theories do not aim to account for specific “raw” data. Rather,
theories explain phenomena identified through robust patterns in the data that
cannot be attributed to the particular manner in which the data were collected
(e.g., researcher biases, measurement error, methodological artifacts, etc.). To
identify these empirical regularities in data, researchers use data models, which
are representations of the data (Suppes, 1962; Kellen, 2019). Data models can
take many forms. These can range from the most basic canonical descriptive
tools, such as a mean score, a correlation, or a fitted curve, to more complex sta-
tistical tools which are common in different areas of psychology and beyond; such
as structural equation models, time-series models, hierarchical models, network
models, mixture models, loglinear models and so forth. Essentially, we can con-
sider a data model to be any (statistical) model that summarizes the data in some
way. Thus, data models, particularly robust and replicable data models, play a
key role in determining what a theory must explain.
Second, data models also inform our understanding of the components and
the relations among them that are posited to give rise to a phenomenon (i.e. the
theory’s structure). It is this role which we will focus on in the next part of the

179
6. Modeling Psychopathology: From Data Models to Formal Theories

paper. This role is especially noteworthy in the context of the data models most
commonly used in the network approach literature: the Ising model, the Gaus-
sian Graphical Model, and the Vector Autoregressive model. In Section 6.3 we
will describe each of these models in more detail, but here it is sufficient to note
their most salient feature: these analyses estimate the structure of relationships
among a set of variables; specifically, the structure of conditional dependence re-
lationships (see Figure 6.1; Top Right). There is a strong intuitive appeal to these
analyses as they seem to hold the promise of directly informing the very thing
we are after: the structure of relations among components of the mental disorder
(see Figure 6.1; Top Left). In Section 6.3, our overarching aim will be to critically
evaluate that promise and determine how best to use (network) data models to
guide the development of theories about specific mental disorders.

6.3 Identifying Formal Theories from Data

In this section we will explore how data models can best contribute to the devel-
opment of formal theories. We will do so within the broader theoretical frame-
work of conceptualizing mental disorders as complex systems and will focus on
three data models that have become popular among researchers adopting this
framework: the Ising model, the Gaussian Graphical Model (GGM), and the Vec-
tor Autoregressive (VAR) model. Specifically, we evaluate three routes that make
use of data models in different ways to obtain a formal theory. We believe that
the first two routes describe how data models are currently used in the literature,
and the third route is an alternative that addresses some of the shortcomings of
the first two approaches.
The first route arrives at formal theories directly by treating these data mod-
els as formal theories. In this case, the transition from data model to formal
theory is largely an act of interpretation. Instead of interpreting a data model
as a representation of the data, we interpret it as a representation of the target
system (see Figure 6.3, Left Panel). Specifically, the variables of the data model
are treated as the components of the target system, and the statistical relation-
ships are treated as the structural relationships among the components. From
this perspective, research is carried out by conducting an empirical study, es-
timating a data model, and treating the data model as a theory. If viable, this
approach would be extremely powerful, because a well-developed theory would
be just one well-designed study away. We evaluate this route in Section 6.3.1.
The second possible route arrives at formal theories by drawing inferences
from data models (Figure 6.3, Middle Panel). That is, the data model is not di-
rectly treated as a theory, but rather is used to inform the theory. From this
perspective, research is carried out by conducting an empirical study, estimat-
ing a data model, and using the data model to infer characteristics of the target
system, thereby informing the development of a theory. For example, one could
observe a conditional dependence relationship between two variables and infer
the presence of a causal relationship between the corresponding components in
the target system. To evaluate this approach we need to know the data-generating

180
 6.3. Identifying Formal Theories from Data

target system. We do so in Section 6.3.2 by treating the Panic Model introduced

in the previous section as the target system of interest, simulating data from that
target system, and examining how well inferences drawn from data models can
be used to inform our understanding of the target system.

Use Data Models Use Data Models Use Data Models

as Formal Theories to Infer Formal Theories to Develop Formal Theories

Formal Data Formal Data Implied Data

Data
Theory Model Theory Model Model Model

Target Target Simulated

Data Data Data
System System Data

Production Inference Formal Target

Representation Comparison Theory System

Figure 6.3: The figure provides an overview of three routes to developing formal theories using data
models. In the left panel, data models are treated as formal theories. In the middle panel, data models
are used to draw inferences about the target system and, thereby, to generate formal theories of that
system. In the right panel, data models used to develop formal theories by deducing implied data
models and comparing them with empirical data models.

The third possible route puts formal theories at the heart of theory develop-
ment. From this perspective, research is carried out by first generating an initial
formal theory. From this formal theory we simulate data which we use to ob-
tain the theory-implied data model. We subsequently compare the implied data
model with the empirical data model, and adapt the formal theory based on the
discrepancy between the two. This route thus leverages the “immense deduc-
tive fertility” of formal theories to make precise predictions that clarify how the
model must be revised to be brought in line with empirical data (Meehl, 1978).
From this perspective, formal theory is not only the ultimate goal of the research
process, but also plays an active role in theory development. We evaluate this
route in Section 6.3.3 by deriving predicted data models from a formal theory of
Panic disorder, and showing how the model can be improved by comparing the
predicted data models to empirical data models.

6.3.1 Using Data Models as Formal Theories

If data models are to serve as formal theories of a target system, the properties of
those data models must be able to represent the properties we expect in the target
system. Accordingly, in this section, we discuss the properties we expect in the
target systems of mental disorders from the complex systems perspective (Section
6.3.1.1) and evaluate whether these properties are captured by the properties

181
6. Modeling Psychopathology: From Data Models to Formal Theories

of three data models: the VAR model, the GGM, and the Ising model (Section
6.3.1.2).

6.3.1.1 Properties of Mental Disorder Target Systems

Target systems consist of components and the relations among them. From the
network perspective there are a number of properties we would expect to be
present in the target systems of mental disorders. First, feedback loops among
components are likely present. Researchers have frequently posited “vicious cy-
cles”, where the initial activation of one component (e.g., arousal) elicits activa-
tion of other components (e.g., perceived threat) and, in turn, is reinforced by
the activation of those components. Second, causal effects between components
are likely to be asymmetrical. That is, the effect of component A on component
B may differ from the effect of component B on component A. For example, it is
unlikely that concentration has the same effect on sleep as sleep has on concen-
tration or that compulsions have the same effect on obsessions that obsessions
have on compulsions.
Third, interactions among components are likely to occur at different time
scales. For example, the effect of intrusive memories on physiological reactivity
in Post-traumatic Stress Disorder is likely to occur on a time scale of seconds to
minutes, whereas an effect of energy on depressed mood may play out over the
course of hours to days, and the effect of appetite on weight gain may occur on
a time scale of days to weeks. Fourth, it is likely that there are higher order
interactions among components. For example, the presence of sleep difficulties
may strengthen the effect of feelings of worthlessness on depressed mood or the
effect of intrusive trauma memories on physiological reactivity. If data models
are to serve as formal theories of the target system, they must be able to represent
these types of causal structures.
We would further suggest that most, perhaps all, mental disorder target sys-
tems are likely to have multiple stable states, that is, states into which the system
settles and will remain in the absence of external perturbation. In the simplest
case, the system will be characterized by the presence of two stables states: an
unhealthy state (i.e. a state of elevated symptom activation, such as a depres-
sive episode), and a healthy state (e.g., a state without elevated symptom activa-
tion). In other cases, there may be multiple stable states (e.g., healthy, depressed,
and manic states in Bipolar Disorder). The presence of multiple stable states
is, in turn, accompanied by other behavior often observed in mental disorders,
including spontaneous recovery and sudden shifts into or out of a state of psy-
chopathology, further suggesting that a model of any given mental disorder will
almost certainly need to able to produce alternative stable states.

6.3.1.2 Comparing Target System Properties with Data Model Properties

The first model we will consider is the VAR model. The VAR model for multivari-
ate continuous time series data linearly relates each variable at time point t to all
other variables and itself at previous time points (Hamilton, 1994), typically the
time point immediately prior t − 1 (i.e. a first order VAR, or VAR(1), model; e.g.,

182
 6.3. Identifying Formal Theories from Data

Bringmann et al., 2013; Pe et al., 2015; A. J. Fisher et al., 2017; Snippe et al., 2017;
Groen et al., 2019). The estimated lagged effects of the VAR models indicate con-
ditional dependence relationships among variables over time. The dynamic of
the VAR model is such that the variables are perturbed by random input (typ-
ically Gaussian noise) and the variables return to their means, which represent
the single stable state of the system.
As depicted in Figure 6.4, the VAR model is able to represent some key
characteristics likely to be present in mental disorder target systems. Most
notably, it allows for feedback loops. Variables can affect themselves both di-
rectly (e.g., Xt → Xt+1 ), or via their effects on other variables in the system (e.g.,
Xt → Yt+1 → Xt+2 ). The VAR model also allows for asymmetric relationships,
since the effect Xt → Yt+1 does not have to be the same effect as Yt → Xt+1 in
direction or magnitude. However, because the lag-size (i.e. the distance between
time points) is fixed and consistent across all relationships, the VAR model does
not allow for different time scales. Moreover, because the VAR model only in-
cludes relations between pairs of variables, it is unable to represent higher-order
interactions involving more than two variables. Finally, the VAR model has a
single stable state defined by its mean vector and thus cannot represent multiple
stable states of a system, such as a healthy state and unhealthy state.

        

            





*




Figure 6.4: The figure shows whether the five properties of mental disorders discussed above can
be represented by the three most popular network data models, the VAR model, the GGM, and the
Ising model with Glauber dynamics. Note that there is a check mark at feedback loops for GGMs
because one could in principle endow the GGM with a dynamic similar to the Ising model, which
would essentially lead to a restricted VAR model but with symmetric relations. The asterisk is present
because this endowment of dynamics is not done in practice.

183
6. Modeling Psychopathology: From Data Models to Formal Theories

The second model we will consider is the Gaussian Graphical Model (GGM).
The GGM linearly relates pairs of variables in either cross-sectional (Haslbeck &
Fried, 2017) or time series data (Epskamp, Waldorp, et al., 2018). In the case of
time series data the GGM models the relationships between variables at the same
time point. Because it does not model any dependency across time, it is typically
not considered a dynamic model and, thus, could not be used to represent the
behavior of a mental disorder target system as it evolves over time. In principle
the GGM could be augmented by a dynamic rule similar to one commonly used
with the Ising model (i.e. “Glauber dynamics”, see below). However, in that
case, the GGM would become a model similar to, but more limited than, the
VAR model described above (e.g., it would be limited to symmetric relationships).
Accordingly, the GGM is similarly unable to represent key features we expect to
observe in a mental disorder target system.
The final model we will consider is the Ising model. The Ising model again
represents pairwise conditional dependence relations between variables (Ising,
1925), however, it is a model for multivariate binary data. While the original Ising
model does not model dependencies over time, it can be turned into a dynamic
model by augmenting it with Glauber dynamics (Glauber, 1963).3 Like the VAR
model, the Ising model is able to represent feedback loops. Moreover, due to its
non-linear form it is able to exhibit multiple stable states (and the behavior that
accompanies such stable states, such as hysteresis and sudden shifts in levels of
symptom activation, see e.g., Cramer et al., 2016; Lunansky, van Borkulo, & Bors-
boom, 2019; Dalege et al., 2016). It is perhaps not surprising then, that the Ising
model is used as a theoretical model across many sciences (Stutz & Williams,
1999), and to our knowledge, is the only of the three data models examined here
that has been used as a formal theory of a mental disorder target system (Cramer
et al., 2016). Unfortunately, the Ising model falls short in its ability to repre-
sent the remaining characteristics likely to be present in mental disorders. The
relationships in the Ising model are exclusively symmetric; with the standard
Glauber dynamics, there is only a single time scale; and the Ising model includes
exclusively pairwise relationships, precluding any representation of higher-order
interactions.

6.3.1.3 Data Models as Formal Theories?

The analysis in this section shows that the VAR, GGM, and Ising models are un-
able to represent most key properties we would expect in the target systems giv-
ing rise to mental disorders, and therefore cannot serve as formal theories for
those disorders. Of course, more complex models would be able to produce more
of the characteristics likely to be present in mental disorders. For example, one
could extend the VAR model with higher-order interactions or a latent state (Tong
& Lim, 1980; Hamaker et al., 2010), thereby allowing it to represent multiple
3 Glauber dynamics work as follows: After specifying an initial value for each variable, one ran-
domly picks a variable Xi at t = 1 and takes a draw from the distribution of Xi conditioned on the
values of all other variables. This value (either 0 or 1) is set to be the new value of Xi and then the
same process is repeated, thereby allowing the model to evolve over time.

184
 6.3. Identifying Formal Theories from Data

stable states. However, estimating data models is subject to fundamental con-

straints. More complex models require more data, and larger sample sizes which
are often unavailable in psychiatric research. For example, around 90 observa-
tions (about 2.5 weeks of a typical ESM study) are needed for a VAR model to out-
perform the much simpler AR model (Dablander et al., 2019). Models more com-
plex than the VAR model would require even more data to be estimated reliably.
In addition, the sampling frequency (e.g., measurement every 2 hours) might be
too low to capture the structure of the target system of interest (Haslbeck & Ryan,
2019). In this situation a data model still contains some information about the
target system, but cannot capture the structure of the target system to the extent
that it can serve as a formal theory. Even if large amounts of high frequency data
were widely available, it is unclear how to estimate many complex models. For
example, one could extend the Ising model with a second time scale (e.g., Lu-
nansky et al., 2019), but it would be unclear how to estimate such a model from
data. Finally, even if such models could be estimated, more complex models are
often uninterpretable. For example, nonparametric models (e.g., splines; Fried-
man, Hastie, & Tibshirani, 2001, p. 139), which can capture extremely complex
behavior, typically consist of thousands of parameters, none of which can be in-
terpreted individually. Accordingly, it is unlikely that any data model estimated
from the type of data typically available in psychiatric research will be both inter-
pretable and capable of capturing the characteristics of psychopathology in such
a way that would allow it to serve as a formal theory of a mental disorder.

6.3.2 Using Data Models to Infer Formal Theories

An alternative route from data models to formal theories is to use data models
to draw inferences about a target system, inferences that we can use to construct
a formal theory. There is good reason to think that this approach could work.
Because the data are generated by the target system, and data models summarize
these data, the parameters of any data model certainly somehow reflect charac-
teristics of the target system. This means that it should be possible, in principle,
to infer something about the target system and its characteristics from data and
data models. Although we have seen already that the GGM, Ising and VAR mod-
els cannot directly reproduce the key characteristics of the target system, their
parameters could potentially still yield insights into the structure or patterns of
relationships between components. In line with this intuition, it has frequently
been suggested that the GGM, the Ising model, and the VAR models can serve as
“hypothesis-generating tools” for the causal structure of the target system (e.g.,
Borsboom & Cramer, 2013; van Rooijen et al., 2017; Fried & Cramer, 2017; Ep-
skamp, van Borkulo, et al., 2018; Epskamp, Waldorp, et al., 2018; Jones, Mair,
Riemann, Mugno, & McNally, 2018).
Although this approach seems intuitive, in practice it is unclear how this in-
ference from data model to target system should work. For example, if we observe
a strong negative cross-lagged effect of Xt on Yt+1 in a VAR model, what does that
imply for the causal relationship between the corresponding components in the
target system? A precise answer to this question would require a rule that con-

185
6. Modeling Psychopathology: From Data Models to Formal Theories

nects parameters in particular data models to the structure of the target system.
For some simple systems, such a rule is available. For example, if the target sys-
tem can be represented as a Directed Acyclic Graph (DAG), then under certain
circumstances its structure can be inferred from conditional (in)dependence re-
lations between its components: Conditional independence implies causal inde-
pendence, and conditional dependence implies either direct causal dependence
or a common effect (Pearl, 2009; Ryan, Bringmann, & Schuurman, 2019). How-
ever, it is generally unclear how we can use the parameters of typical data models
to make inferences about the types of non-linear dynamic systems we expect in
a psychiatric context (although Mooij, Janzing, & Schölkopf, 2013 and Forré &
Mooij, 2018 have established some links in this regard). The consequences of
this are twofold. First, any inference from data model to target system must rely
instead on some simplified heuristic(s) in an attempt to approximate the link be-
tween the two. Second, it is unclear how well the combination of common data
models and simple heuristics perform in allowing us to make inferences about
the target system.
In this section, we evaluate whether the three data models introduced above
can be used to make inferences about mental disorder target systems. To do this,
we treat the Panic Model discussed in Section 6.2 as the data-generating target
system and compare the causal structure inferred from the data models to the
true causal structure. To yield these inferences we use a very simple and intu-
itive set of heuristics: a) if two variables are conditionally dependent in the data
model, we will infer that the corresponding components in the target system are
directly causally dependent; b) if there is a positive linear relationship, we will
infer that the causal relation between the corresponding components is positive
(i.e. reinforcing); c) if there is a negative linear relationship, we will infer that the
causal relationship among components is negative (i.e. suppressing).

6.3.2.1 Inferring the Panic System from Network Data Models

To be able to evaluate the success of the simple heuristics described above, we
must first represent the structure of the Panic Model (see Section 6.2) in the struc-
ture of a square matrix, that is, in the same form as the parameters of the VAR,
GGM, and Ising models. Since the relationships between components are formal-
ized through differential equations, a natural choice is to represent the Panic Model
as a network of moment-to-moment dependencies, drawing an arrow X → Y if
the rate of change of Y is directly dependent on the value of X (known as a lo-
cal dependence graph; Didelez, 2007). Figure 6.5 (a) displays these moment-to-
moment dependencies. Note that this structure cannot capture many aspects of
the true model, such as the presence of two time scales or the moderating effect of
Arousal Schema (AS) (see Section 6.2 for details). It is, thus, already clear that the
models cannot recover the exact causal structure of the Panic Model. Nonetheless,
we can still investigate whether applying the simple heuristics to these three data
models allows us to infer this less detailed pattern of direct causal dependencies.
We next compare this true causal structure to the causal structure inferred
based on the three data models. To obtain the three data models, we first gen-

186
 6.3. Identifying Formal Theories from Data

erate data from the target system (See Appendix 6.A). Specifically, we use four
weeks of minute-to-minute time-series data for 1000 individuals. These individ-
uals differ in their initial value of Arousal Schema, with the distribution chosen
so that the proportion of individuals for whom a panic attack is possible was
equivalent to the lifetime history prevalence of panic attacks in the general popu-
lation (R. R. Freedman, Ianni, Ettedgui, & Puthezhath, 1985). For the VAR model
analysis, we create a single-subject experience-sampling-type dataset by choos-
ing the individual who experiences the most (16) panic attacks in the four-week
period. To emulate ESM measurements, we divide the four week period into 90-
minute intervals, taking the average of each component in that interval, yielding
448 measurements. For the GGM analysis, we create a continuous cross-sectional
dataset by taking the mean of each component for each individual over the four
weeks. For the Ising model analysis, we obtain cross-sectional binary data by tak-
ing a median split of those same variables. The resulting VAR, GGM and Ising
model networks are displayed in Figure 6.5 panels (b), (c) and (d), respectively.4

(a) True Local Dependence

PT Arousal

AS Context

Avoidance Escape

(b) VAR Model (c) Gaussian Graphical Model (d) Ising Model

PT Arousal PT Arousal PT Arousal

AS Context AS Context AS Context

Avoidance Escape Avoidance Escape Avoidance Escape

Figure 6.5: Panel (a) shows the true model in terms of local dependencies between components; panel
(b) shows the VAR model estimated from ESM data sampled from the true model; panel (c) shows
the GGM estimated from the cross-sectional data of 1000 individuals, generated from the true model;
panel (d) shows the Ising model estimated on the same data after being binarized with a median split.
Solid edges indicate positive relationships, dotted indicate negative relationships. For panels (b) to
(d), the widths of edges are proportional to the absolute value of the corresponding parameter. Note
that in panel (b) we do not depict the estimated auto-regressive parameters as the primary interest is
in inferring relationships between variables.

4 Note that in the Ising model the parameter estimates are somewhat unstable due to near-
deterministic relationships between some binarized variables.

187
6. Modeling Psychopathology: From Data Models to Formal Theories

We will focus our evaluation on two important causal dependencies in the

target system: the positive (i.e. reinforcing) moment-to-moment feedback loop
between Perceived Threat and Arousal, and the positive effect of Arousal Schema
(i.e. beliefs that arousal-related bodily sensations are dangerous) on Avoidance
(i.e. efforts to avoid situations or stimuli that may elicit panic attacks). In the VAR
model (panel (b) in Figure 6.5) we see a lagged positive relationship of Arousal
to Perceived Threat, a strong negative lagged relationship from Perceived Threat
to Arousal, and a weak positive effect of Arousal Schema on Avoidance. Ap-
plying the heuristics, we would infer a reinforcing relationship from Arousal to
Perceived Threat, a suppressing relationship from Perceived Threat to Arousal,
and a reinforcing effect of Arousal Schema on Avoidance. In the GGM (panel
(c) in Figure 6.5) we see a positive conditional dependency between mean values
of Arousal and Perceived Threat, but we also see a weak negative dependency
between mean values of Arousal Schema and Avoidance. Applying the heuris-
tics to the GGM, we would infer a reinforcing relationship between Arousal and
Perceived Threat, and a suppressing relationship between Arousal Schema and
Avoidance. Finally, in the Ising model (panel (d) in Figure 6.5), we see a strong
positive dependency between Arousal Schema and Avoidance, and a very weak
positive relationship between PT and Arousal. This leads us to infer two reinforc-
ing relationships, between Arousal and Perceived Threat, and Arousal Schema
and Avoidance.
For the VAR model, the heuristics yield one correct and one incorrect infer-
ence. For the GGM, we make exactly the opposite inferences, with again one
correct and one incorrect. In the Ising Model, we yield two correct inferences.
However, inspecting the rest of the Ising Model edges we can see a variety of
incorrect inferences about other relationships, with independent components in
the target system connected by strong edges in the Ising model, and the valences
of various true dependencies flipped. At best, we can say that in each of the three
network models, some dependencies do reflect the presence and/or direction of
direct causal relationships, and some do not. Unfortunately, it is not possible to
distinguish which inferences are trustworthy and which are not without know-
ing the target system, and in any real research context, the target system will be
unknown. Consequently, these data models cannot be used to confidently and
reliably draw inferences about the target system using these simple heuristics.

6.3.2.2 The Mapping between Data Model and Target System

Importantly, our inability to draw accurate inferences from these data models
is not a shortcoming of the data models themselves. Each data model correctly
captures some form of statistical dependency between the components in a par-
ticular domain (e.g., lagged 90 minute windows). Moreover, the statistical de-
pendencies in the data models are produced by causal dependencies in the target
system, so we know there is some mapping from the causal dependencies in the
target system to statistical dependencies in the data model. The fundamental bar-
rier to inference is that the form of this mapping is unknown and considerably
more complex than the simple heuristics we have used to draw inferences here.

188
 6.3. Identifying Formal Theories from Data

For example, consider the relationships between Perceived Threat and Arousal.
The VAR model (panel (b) in Figure 6.5), identifies a negative lagged relationship
from Perceived Threat to Arousal in the data generated by the target system. Yet
in the target system, this effect is positive. This “discrepancy” occurs because of
a very specific dynamic between these components: After a panic attack (i.e. a
brief surge of Perceived Threat and Arousal) there is a “recovery” period in which
arousal dips below its mean level for a period of time. As a result, when we av-
erage observations over a 90 minute window, a high average level of Perceived
Threat is followed by a low average level of Arousal whenever a panic attack oc-
curs. That same property of the system produces the observed findings for the
GGM and Ising Model through yet another mapping (for details, see Appendix
6.B).
As this example illustrates, the mapping between target system and data
model is intricate, and it is unlikely that any simple heuristics can be used suc-
cessfully to work backwards from the data model to the exact relationships in
the target system. We can expect this problem to arise whenever we use rela-
tively simple statistical models to directly infer characteristics or properties of a
complex system (c.f. the problem of under-determination or indistinguishabil-
ity; Eberhardt, 2013; Spirtes, 2010). Indeed, the same problem arises even for
simpler dynamical systems when analyzed with more advanced statistical meth-
ods (e.g., Haslbeck & Ryan, 2019). Of course, in principle, it must be possible to
make valid inferences from data and data models to some properties of a target
system using a more principled notion of how one maps to the other. For exam-
ple, under a variety of assumptions, it has been shown that certain conditional
dependency relationships can potentially be used to infer patterns of local causal
dependencies in certain types of dynamic system (Mooij et al., 2013; Bongers &
Mooij, 2018; Forré & Mooij, 2018). However, the applicability of these methods
to the type of target system we expect to give rise to psychopathology (see Sec-
tion 6.3.1) is as yet unclear and even under the strict assumptions under which
they have been examined, these methods still do not recover the full structure of
the target system.5 This means that the intricacy of the mapping between target
system and data model currently precludes us from making reliable inferences
about the target system. Accordingly, we cannot use those inferences to build
formal theories.
5 Specifically, Mooij et al. (2013) and Bongers and Mooij (2018) have shown that cyclic causal mod-
els can be conceptualized as encoding causal dependencies between the equilibrium positions of de-
terministic differential equations and differential equations with random initial values. Forré and
Mooij (2018) formally link the conditional dependencies between equilibrium position values to the
causal dependencies in these cyclic causal models using a considerably more complex mapping rule
than that which holds for DAGs. Their applicability to the current context is limited in the sense
that 1) to our knowledge these rules have not been extended to dynamic systems with time-varying
stochastic terms (SDEs) as we would expect to see in complex psychological systems (and on which
the Panic Model is based), and 2) the use of these methods is reliant on data that reflects equilibrium
positions. Future developments in this area may prove to yield useful tools for psychological theory
development however, and we consider this area to be ripe for future research beyond the current
paper.

189
6. Modeling Psychopathology: From Data Models to Formal Theories

6.3.3 Using Data Models to Develop Formal Theories

In Section 6.3.2, we saw that the mapping between target system and data model
is intricate and would be nearly impossible to discern when the target system is
unknown. However, we also saw that when the target system is known, we can
determine exactly which data models the target system will produce. Indeed,
this is precisely what we did when we simulated data and fit data models to it in
the previous section. In this section we consider a third route to formal theories,
which makes use of this ability to determine which data models are implied by a
given target system (or formal theory).
This third route works as follows. First and foremost, we must propose some
initial formal theory which we take as a representation of the target system. The
quality or accuracy of this representation may be good or bad, but crucially the
theory must be formalized in such a way as to yield unambiguous predictions (see
Section 6.2.2). Second, we can use this initial theory to deduce a theory-implied
data model. This can be done by simulating data from the formal theory and
fitting the data model of interest. Third, we can learn about and adapt the for-
mal theory by comparing implied data models with their empirical counterparts.
This approach is represented in schematic form in the right-hand panel of Figure
6.3. It can be seen as a form of inference, but it is abductive inference: inference to
the best explanation (Haig, 2005). We first infer the best explanation for the core
phenomena to generate an initial theory. We then infer the best explanation for
any discrepancies between empirical and theory-implied data models, inferences
which inform subsequent theory development. Given the importance of abduc-
tive inference to this approach, we adopt this term to refer to this third route.
In this sub-section, we will illustrate this approach using the example of panic
disorder (for an overview, see Figure 6.6).

6.3.3.1 Obtaining Theory-Implied and Empirical Data Models

In this section, we will treat the Panic Model introduced in Section 6.2 as our
initial formal theory, which should represent the target system that gives rise to
panic disorder (Figure 6.6, bottom row). The Panic Model can be used to simu-
late data and, in turn, to derive predictions made by the theory in the form of
theory-implied data models (left-hand column of Figure 6.6). While in principle
many data models can (and should) be used to perform the abductive inference
described above, here we will examine the implied cross-sectional Ising model
of the three core panic disorder symptoms: 1) Recurrent Panic Attacks (PA), 2)
Persistent Concern (PC) following a panic attack and 3) Avoidance (Av) behavior
following a panic attack (American Psychiatric Association, 2013). If our formal
theory of panic disorder is an accurate representation of the target system that
gives rise to panic disorder, the implied Ising model derived from this theory
should be in agreement with a corresponding Ising model derived from empiri-
cal data.
Critically, obtaining an implied data model requires not only a formal the-
ory from which we can simulate data, but also a formalized process by which
variables are “measured” from those data. The Panic Model is used to generate

190
 6.3. Identifying Formal Theories from Data

Implied Ising Model Empirical Ising Model

PA PA α = -6.14

α = −170.08

85.98 85.98 6.12 1.63

α = −88.38 α = −88.38 α = -3.63 α = -26.14

Av 625.85 PC Av 23.72 PC

Estimate Estimate

Av PA PC Av PA PC
Simulated 0 0 0 0 0 0
0 0 0 CPES Data 0 1 1
Data 1 1 1 1 1 1
0 0 0 0 0 0
1 1 1 1 1 0

Emulated Measurement
Measurement

Arousal
Schema
Panic Disorder
Target System

Represents
Avoid Arousal Perceived Escape
Threat

Context

Panic Attacks

Formal Theory

Figure 6.6: Illustration of the third route to formal theories. We take the Panic Model discussed in
Section 6.2 as our formal theory, representing the unknown target system that gives rise to panic
disorder. To obtain an implied data model from this theory, we first formalize how the components
of the theory produce the data of interest, emulating the measurement process. With this in place,
we can simulate data from the model in the form of cross-sectional binary symptom variables. We
obtain the theory-implied Ising Model by estimating it from these simulated data (top-left corner).
To estimate the empirical Ising Model (top-right corner) we make use of empirical measurements of
binary symptom variables from the CPES dataset.

intra-individual time series data for multiple individuals (as described in Ap-
pendix 6.A) and so we need to define how cross-sectional symptom variables can
be extracted from those time series. We specify that Recurrent Panic Attacks (PA
= 1) are present for an individual in our simulated data if there are more than
three panic attacks in the one month observation period. Persistent Concern is

191
6. Modeling Psychopathology: From Data Models to Formal Theories

determined using the average levels of jointly experienced arousal and perceived
threat (i.e. anxiety) following a panic attack. If an individual has a panic attack,
and their average anxiety following a panic attack exceeds a threshold determined
by “healthy” simulations (i.e. those without panic attacks), they are classified as
having Persistent Concern (PC = 1). Avoidance is defined similarly, with this
symptom present if an individual has a panic attack, and their average levels of
avoidant behavior following that attack are higher than we would expect to see
in the healthy sample. A more detailed account of how we generated these data
can be found in Appendix 6.C. This simulated cross-sectional data was used to
estimate the implied Ising Model (top left-hand corner, Figure 6.6).6
We obtained the corresponding empirical Ising model (right-hand column of
Figure 6.6) using the publicly available Collaborative Psychiatric Epidemiology
Surveys (CPES) 2001-2003 (Alegria, Jackson, Kessler, & Takeuchi, 2007). The
CPES is a nationally representative survey of mental disorders and correlates
in the United States, with a total sample size of over twenty thousand partici-
pants (of which n = 11367 are used in the current analysis; for details see Ap-
pendix 6.C). The CPES combines more than 140 items relating to panic attacks
and panic disorder, with a diagnostic manual describing how these items can be
re-coded into binary symptom variables reflecting Recurrent Panic Attacks, Per-
sistent Concern and Avoidance. PA is present if the participant reported more
than three lifetime panic attacks. PC is present if, following an attack, the partic-
ipant experienced a month or more of persistent concern or worry. Av is present
if the participant reports either a month of avoidance behavior following an at-
tack, or a general avoidance of activating situations in the past year.

6.3.3.2 Theory Development: Comparing Model-Implied and Empirical Data

Models
As seen in Figure 6.6, there is a similar pattern of conditional dependencies in
the implied and empirical data models. In both, all pairwise dependencies are
positive, and all thresholds are negative. There is also a similar ordering of con-
ditional dependencies in terms of their magnitude. Within each model, the con-
ditional relationships of PA with Av and PA with PC are of the same order of mag-
nitude, and the conditional relationship between Av and PC is an order of mag-
nitude greater. However, we also see some differences between the models. First,
the absolute value of pairwise dependencies and thresholds are much greater in
the implied Ising Model (Figure 6.6 (a)) than the empirical Ising Model (Figure
6.6 (b)). Second, we see that the relationships in the implied model are perfectly
symmetric, with exactly the same thresholds for Av and PC, and precisely the
same weights relating PA to both. In the empirical network, these weights and
thresholds are much smaller.
The bivariate contingency tables of all symptom-symptom relationships clar-
ify the nature of these relationships (see Figure 6.7). In both the implied and
empirical data models only a small proportion of individuals experience Recur-
rent Panic Attacks (Empirical 4.3%, Simulated 3.72%). Crucially, in the simulated
6 Again the parameter estimates are somewhat unstable due to near-deterministic relationships.

192
 6.3. Identifying Formal Theories from Data

a) Panic Attacks & b) Panic Attacks & c) Persistent Concern &

Persistent Concern Avoidance Avoidance

PC Av Av
0 1 0 1 0 1
Empirical

0 96.06 0.22 0 93.59 2.69 0 93.79 4.89

PC
PA
PA

1 2.62 1.10 1 0.20 3.52 1 0 1.32

PC Av Av
0 1 0 1 0 1
Simulated

0 94.70 0.70 0 94.70 0.70 0 94.70 0

PC
PA
PA

1 0 4.60 1 0 4.60 1 0 5.30

Figure 6.7: Contingency tables showing percentages for each pair of symptom variables (one per col-
umn) for the empirical data (top row) and simulated data (bottom row). The CPES contingency tables
are based on nCP ES = 11367 observations. The simulated dataset contains nsim = 1000 observations.

dataset, the symptom relationships are almost deterministic: If one symptom is

present, so too are all others, and vice versa for the absence of symptoms (apart
from three individuals who experience less than three panic attacks in the time
window). This is because there is a deterministic relationship between the com-
ponents underlying these symptoms in the Panic Model: All participants who
experience one panic attack have Persistent Concern and Avoidance behavior af-
ter those attacks. In contrast, there are non-deterministic relationships in the
empirical data. For example, it is actually more common to have Recurrent Panic
Attacks without Persistent Concern than with Persistent Concern (column (a)).
Similarly, more individuals experience Avoidance without Persistent Concern,
than with Persistent Concern (column (c)) Conversely, there are no individuals
who experience Persistent Concern but not Avoidance.
The discrepancies between the implied and the empirical data model could
arise at any step in the process from formal theory/target system to implied and
empirical data model illustrated in Figure 6.6. It could be the case that a dis-
crepancy is due to inaccuracies in how we emulate the measurement process. For
example, perhaps Persistent Concern and Avoidance co-occur equally, but the
former suffers from a greater degree of recall bias than the latter (for an example
of differential symptom recall bias in depressed patients, see Ben-Zeev & Young,
2010). There are also different time scales at which the simulated and empiri-
cal symptoms are defined. The simulated symptoms are defined over a month
period whereas the CPES items are defined over lifetime prevalence. Due to the
deterministic nature of the Panic Model, we believe a month period is a good ap-

193
6. Modeling Psychopathology: From Data Models to Formal Theories

proximation for lifetime experience of panic symptoms in this case. Nonetheless,

it is a discrepancy in measurement that could lead to discrepancies between the
implied and empirical data models. It could also be that the discrepancy is due
to estimation issues. However, due to the large sample sizes and simple models
used, we suspect it is unlikely that sampling variance is a problem in this in-
stance. For present purposes, we will assume here that discrepancies are to due
to inaccuracies in how the theory represents the target system, and as such we
can use these discrepancies to directly evaluate our theory.7
On a global level there is a good match between the empirical and implied
models: The theory implies positive symptom-symptom dependencies, which we
also observe in the empirical data. However, the implied model over-estimates
the strength of these relationships. This is largely explained by the deterministic
causal effects in the theory. In the simulated data, everybody who experiences
panic attacks also develops Persistent Concern and, in turn, Avoidance. As seen
in Figure 6.7, this is inconsistent with empirical data, identifying a serious short-
coming in the theory. To improve the model, we must include some mechanism
by which individuals can experience a panic attack without developing the re-
maining symptoms of panic disorder. In the empirical data, there is a near deter-
ministic effect of Persistent Concern on Avoidance, suggesting that once Persis-
tent Concern develops, Avoidance will follow; an observation that is consistent
with the formal theory. However, inconsistent with the formal theory, the em-
pirical data suggests a relatively low probability of Persistent Concern following
panic attacks, suggesting that this is where the theory must be revised if it is
to better account for the observation that some individuals experience recurrent
panic attacks without developing the full panic disorder syndrome.
This is just one discrepancy in these data that can inform model development
and more insights may be gained by focusing on others. Many more insights can
be gained by considering different data models based on different data. For ex-
ample, experimental data on the relation between Arousal and Perceived Threat
may allow us to refine the specification of the feedback between those two vari-
ables. In general this route offers a great deal of flexibility in theory development.
Although the theory is likely to be complex, dynamic and non-linear, the form of
the data models used to learn about that theory need not be. Instead, by starting
with an initial theory, the researcher can use any data about the phenomena of
interest to further develop that theory. In the following section we will provide
a full account of this third route by discussing the full process of theory devel-
opment from establishing the phenomenon, to making novel predictions with a
well-developed formal theory.
7 In practice, inaccurate conceptualizations of how measurements represent the target system will
be problematic for any approach to theory development or indeed any scientific endeavor, as evi-
denced by the growing attention on measurement in psychopathology literature (e.g., Flake & Fried,
2019). Our proposed approach to formal theory development forces us to be specific about the mea-
surement process, just as we are forced to explicate the theory itself. Although we focus on the formal-
ized theory itself here, we consider the formalization of measurement to be a significant advantage of
this approach, and one that warrants further development.

194
 6.4. An Abductive Approach to Formal Theory Construction

6.4 An Abductive Approach to Formal Theory

Construction
In Section 6.3.3 we illustrated a clear approach to use empirical data to develop
an existing formal theory. However, our description of this approach so far has
omitted several critical steps, including how to generate an initial formal theory
and how to test that theory. In this section, we propose a three-stage process
of formal theory construction, with an emphasis on the role data models play
at each stage (see Figure 6.8). First, in the theory generation stage, we establish
the phenomenon to be explained, generate an initial verbal theory, and formalize
that theory. Second, in the theory development stage, the theory is developed
beyond this initial proposal by adapting it such that it is consistent with as many
empirical findings as possible. Finally, in the theory evaluation stage, the the-
ory is subjected to strong tests within a hypothetico-deductive framework. The
approach to theory construction proposed here places considerable emphasis on
the theory’s ability to explain phenomena, especially during the generation and
development of the theory. Accordingly, the framework we have proposed is a
largely abductive approach (Haig, 2005).

Stage of Theory Use of

Construction Data Models

Generate Abduction Establish

Theory Phenomenon

Abduction
Compare
Develop
Implied &
Theory
Empirical
Deduction

Test Evaluate Risky

Theory Deduction
Prediction

Figure 6.8: Flowchart depicting the process of developing a formal theory with the abductive ap-
proach put forward in this section. In the theory generation step we first establish the phenomenon
(Section 6.4.1.1) and then generate an initial verbal theory (Section 6.4.1.2) which is subsequently for-
malized (Section 6.4.1.3). In the second step (Section 6.4.2) the theory is validated by testing whether
it is consistent with existing empirical findings that are not part of the core phenomenon. If the for-
mal theory is not consistent with some findings, it is adapted accordingly. If these adaptations lead
to a “degenerative” theory (Meehl, 1990) we return to the first step; otherwise we continue to the
final step, in which we test the formal theory using risky predictions (Section 6.4.3). If many tests are
successful, we tentatively accept the theory. If not, the theory must either be adapted (step two) or a
new theory generated (step one).

195
6. Modeling Psychopathology: From Data Models to Formal Theories

6.4.1 Generating Theory

6.4.1.1 Establishing the Phenomenon
The goal of a formal theory is to explain phenomena. Accordingly, the first step
of theory development is to specify the set of phenomena to be explained. Estab-
lishing phenomena is a core aim of science and a full treatment of how best to
achieve this aim is beyond the scope of this paper (for a possible way to organize
this process see Haig, 2005). However, we suspect that the most appropriate phe-
nomena for initial theory development will often include things that researchers
would not think to subject to empirical analysis, as the most robust phenomena
may simply be taken for granted as features of the real world. For example, in the
case of panic disorder, the core phenomena to be explained are simply the obser-
vations that some people experience panic attacks and recurrent attacks tend to
co-occur with persistent worry or concern about those attacks and avoidance of
situations where such attacks may occur. These are empirical phenomena so ro-
bust that they are typically not empirically examined but instead simply assumed
to be a feature of the real world.

6.4.1.2 Generate Initial Verbal Theory

Once the phenomena to be explained have been established, how do we go about
generating an initial theory to explain them? A brief survey of well-known sci-
entific theories reveals that this initial step into theory is often unstructured and
highly creative. For example, in the 19th century August Kekulè dreamt of a
snake seizing its own tail, leading him to generate the theory of the benzene ring,
a major breakthrough in chemistry (Read, 1995). In the early 20th century, Al-
fred Wegener noticed that the coastlines of continents fit together similar to puz-
zle pieces, and consequently developed the theory of continental drift (Wegener,
1966), which formed the basis for the modern theory of plate tectonics (Mauger,
Tarbuck, & Lutgens, 1996). In the late 20th century, Howard Gardner explained
that he developed his theory of multiple intelligences in the 1980s using “sub-
jective factor analysis” (Walters & Gardner, 1986, p. 176). Although more codi-
fied approaches to theory development exist (e.g., Grounded Theory; Strauss &
Corbin, 1994), we are unaware of any evidence to suggest that any one approach
to theory generation is superior to any other.
Nonetheless, the nature of theories does provide some guidance for how they
might initially be generated. Theories achieve their aim of explaining phenom-
ena by representing a target system. Accordingly, generating an initial theory
will require that we specify the components thought to compose the target sys-
tem. This process entails dividing the domain of interest into its constituent
components (i.e. “partitioning”) and selecting those components one believes
must be included in the theory (i.e. “abstraction”), thereby producing a system
of components that will be the object of the theory (cf. Elliott-Graves, 2014).
For researchers adopting a “network perspective”, the target system is typically
presumed to comprise cognitive, emotional, behavioral, or physiological com-
ponents, especially those identified in diagnostic criteria for mental disorders

196
 6.4. An Abductive Approach to Formal Theory Construction

(Borsboom, 2017). For example, as we have seen in Section 6.2, the target sys-
tem could consist of Arousal, Perceived Threat, Avoidance and other symptoms
of panic attacks or panic disorder. Having identified the relevant components we
next specify the posited relations among them. For example, specifying that Per-
ceived Threat leads to Arousal. Within the domain of the network approach, this
second step will typically entail specifying causal relations among symptoms or
momentary experiences (e.g., thoughts, emotions, and behavior).
Notably, in psychiatry, we do not necessarily need to rely on creative insight
about the components and relations among them in order to generate an initial
theory. There are already a plethora of verbal theories about mental disorders. If
the initial verbal theory is well supported and specific, it will lend itself well to
formalization and subsequent theory development. However, even poor verbal
theories can be a useful starting point to developing a successful formal theory
(Wimsatt, 1987; Smaldino, 2017).

6.4.1.3 Formalize Initial Theory

Once a verbal theory has been specified, the next step is to formalize the theory.
To do so, we first need to choose a formal framework. A common formal frame-
work is the use of difference or differential equations, which model how vari-
ables change across discrete time steps and continuous time, respectively (e.g.,
Strogatz, 2015). Specifically, the relations between components is specified by
defining the rate of change of each component as a function of all other compo-
nents and itself. The Panic Model, which we used as an example throughout the
paper, uses this formal framework. Another common framework is Agent based
Modeling (ABM), in which autonomous agents interact with each other using a
set of specified rules (e.g., Grimm & Railsback, 2005). Here, each agent has local
rules on how to interact with other agents. Both frameworks can be implemented
in essentially any computer programming language and both are likely to be rel-
evant to psychiatric and psychological research as a whole.
Having chosen a formal framework, the next step is to specify the relations
between each component in the language of that framework. This process of
formalizing relations is an exercise in being specific. Mathematics and compu-
tational programming languages require theorists to specify the precise nature
of the relationship between variables. Requiring this level of specificity is one
advantage of computational modeling, as it has the effect of immediately clari-
fying what remains unknown about the target system of interest, thereby guid-
ing future research. However, this also means that theorists will often be in the
position of needing to explicate relationships when the precise nature of those
relationships is uncertain. We believe that, even in the face of this uncertainty,
it is better to specify a precise relationship and be wrong than to leave the rela-
tionship ambiguously defined, as it is in a verbal theory. Nonetheless, we suspect
that theorists will be on firmer foundation for subsequent theory development
the more that they are able to draw on empirical data and other resources to in-
form this initial formal theory. There are several sources of information that can
guide the formalization process.

197
6. Modeling Psychopathology: From Data Models to Formal Theories

First, empirical research can inform specification of components and the re-
lations among them. For example, one could use the finding that sleep quality
predicts next-day affect, but daytime affect does not predict next-night sleep (de
Wild-Hartmann et al., 2013) to constrain the set of plausible relationships be-
tween those two variables in the formal theory. There could also be empirical data
on the rate of change of variables, for example, Siegle, Steinhauer, Thase, Stenger,
and Carter (2002) and Siegle, Steinhauer, Carter, Ramel, and Thase (2003) have
shown that depressed individuals exhibit longer sustained physiological reac-
tions to negative stimuli than healthy individuals, a finding which is echoed in
self-report measures of negative affect (Houben et al., 2015).
Second, we can possibly derive reasonable scales for variables and relation-
ships between variables from basic psychological science. For example, classical
results from psychophysics show that increasing the intensity of stimuli in almost
all cases leads to a nonlinear response in perception (e.g., Fechner, Howes, & Bor-
ing, 1966): When increasing the volume of music to a very high level, individuals
cannot hear an additional increase.
Third, in many cases we can use definitions, basic logic, or common sense to
choose formalizations. For example, by definition emotions should change at a
time scale of minutes (Houben et al., 2015), while mood should only change at
a time scale of hours or days (Larsen, 2000). And we can choose scales of some
variables using common sense, for example one cannot sleep less than 0 and more
than 24 hours a day, and heart rate should be somewhere between 50 and 180.
Fourth, we could use an existing formal model of another target system,
which we expect to have a similar structure as the target system giving rise to
the phenomenon of interest. This approach is called “analogical modeling”. For
example, Cramer et al. (2016) formulated a model for interactions between symp-
toms of Major Depression using the Ising model, which was originally formu-
lated to model magnetism on an atomic level (Ising, 1925). Similarly, Fukano
and Gunji formulated a model for interactions among core components of panic
attacks using a Lotka-Volterra model originally formulated to represent predator-
prey relationships (Fukano & Gunji, 2012).
Fifth, it is also important to note that there are methods by which we can
potentially estimate the parameters for a formal theory from empirical data.8
These approaches require considerable development of the formal theory (e.g.,
the form of a differential equation), suitable data (typically intensive longitudi-
nal data), and a clear measurement model relating observed variables to theory
components (as we did in Section 6.3.3). Accordingly, this approach already re-
quires considerable progress in generating a formal theory and may be limited
by practical considerations. Nonetheless, it remains a valuable resource that, if
successfully carried out, would likely strengthen subsequent efforts at theory de-
velopment.
8 For example, if the theory is formalized in a system of differential equations, the parameters of
such equations can in principle be estimated from time series data using, amongst others, Kalman
filter techniques and state-space approaches (e.g., Einicke, 2019; Kulikov & Kulikova, 2013; Durbin
& Koopman, 2012). For implementations of these estimation methods see Ou et al. (2019); Carpenter
et al. (2017); King, Nguyen, and Ionides (2015)

198
 6.4. An Abductive Approach to Formal Theory Construction

The aim of this initial stage is to generate a formal theory able to explain a
set of core phenomena. As we have emphasized throughout this paper, formal
theories precisely determine the behavior implied by their theory. Accordingly,
explanation in this context means that the theory has demonstrated its ability to
produce the behavior of interest. For example, a theory of panic attacks must be
able to produce sudden surges of arousal and perceived threat; a theory of de-
pression must be able to produce sustained periods of low mood; and a theory of
borderline personality disorder must be able to produce affective instability. We
would note that there are very few theories in psychiatry that have reached this
stage of not merely positing, but demonstrating, that the theory can explain the
phenomena of interest. Accordingly, completing this stage of theory construction
would constitute a significant advance in psychiatric theories. Once a theory has
reached this stage, it is ready for the next stage of theory construction.

6.4.2 Developing Theory

The formal theory produced in the first stage of theory construction will have
demonstrated its ability to explain the core phenomena of interest. However,
the fact that the formal theory provides some explanation does not mean it is the
correct explanation. In other words, demonstrating an ability to explain the phe-
nomena of interest is a critical first step, but does not guarantee that the formal
theory is a good representation of the target system. To achieve this aim, we pro-
pose a stage of theory development in which the theory is refined by deducing
implied data models and comparing them to empirical data models. If the two
data models align we take this as evidence that the current formal theory is ad-
equately representing the target system; if there are discrepancies between the
two data models, we analyze the nature of those discrepancies, consider the best
explanation for how they arose, and adapt the formal theory to be able to account
for these discrepancies.
This process of further developing a theory through comparisons between
implied and empirical data models is exactly the process that we have illustrated
already in Section 6.3.3. In that illustration, we derived the implied Ising model
for the three variables Panic Attacks, Persistent Concern, and Avoidance, and we
compared it to the empirical Ising model for the same variables. We discussed
possible explanations for this discrepancy and corresponding adaptations of the
Panic Model, for example including a mechanism by which individuals can ex-
perience Panic Attacks without experiencing Persistent Concern and Avoidance.
If we were to continue in this stage of theory development, we would iterate this
process, adapting the Panic Model to include such a mechanism, deriving the
implied Ising model from this adapted Panic Model, and determining whether
it better accounts for the empirical Ising model (i.e. the discrepancy between
implied and empirical Ising models is smaller).
This stage of theory development can make use of many different types of
data such as physiological, psychological and behavioral measurements from in-
dividuals, cross-sectional data such as clinical interviews and questionnaire data,
or experimental data. Depending on the empirical phenomenon we would like to

199
6. Modeling Psychopathology: From Data Models to Formal Theories

account for, different kinds of data and data models will be appropriate. In gen-
eral, however, more complex data models tend be more powerful tools to tease
apart competing theories. For example, a great many formal theories might be
consistent with a set of means, but it is likely that fewer are consistent with the
means and the conditional relationships between the variables, captured, for ex-
ample, by a GGM or Ising model. In other words, there are a more constrained
number of possible formal theories that may account for more complex data mod-
els, thereby doing more to guide theory development.
There are two important considerations when working through this stage.
First, it is important to consider how much trust to place in the empirical data
at hand. Discrepancies between the empirical data model and theory-implied
model may be due to shortcomings of the formal theory, but they may also be
due to poor measurement, insufficient samples, or poor estimation of the param-
eters of the data model. How do we know when to adapt the formal theory in the
face of some discrepancy? For some guidance on this question, we can draw on
the large literature on model evaluation and model comparison. A straightfor-
ward way to decide whether to adapt the theory would be to derive an implied
model of the adapted theory, and then compare the likelihood of the empirical
data given the initial theory and the adapted theory. In order to decide whether
to accept the adaptation we can use, for instance, a likelihood ratio test or a Bayes
factor. This procedure ensures that we only make adaptations to our theory if we
are certain enough that they actually lead to a better representation of the tar-
get system, and not only the idiosyncratic features of the empirical data at hand.
That is, whether we accept an adaptation of the formal theory depends both on
how large the improvement is, and how certain we are about it (i.e. how large the
sample size is).
Second, it is important to consider the danger of making too many ad hoc re-
visions to the model that account only for idiosyncratic features of a given data
model or, worse, yield new implications that are inconsistent with other empir-
ical findings. For some guidance on this question, we can draw on the litera-
ture on theory evaluation from the philosophy of science literature (Meehl, 1990;
Lakatos, 1976), which would suggest that the theory development phase has two
possible outcomes. If the theory is adapted almost every time it is tested against
empirical data, if those adaptations are making the theory increasingly unwieldy,
and if additional changes are increasingly difficult to make without causing the
theory to be inconsistent with earlier tested empirical findings, the theory can be
considered to be “degenerative” (Meehl, 1990; Lakatos, 1976). In such a situation,
the initial theory was inappropriate and we return to the first step to generate a
different initial formal theory. On the other hand if modifications to a theory ex-
pand, rather than contract, its ability to account for other empirical data beyond
those it was originally introduced to explain, then we can have greater confidence
in those modifications and, in turn, the formal theory. Ultimately, theorists must
strive for a balance between the simplicity of the model and its consistency with
empirical data models.
The aim of the theory development stage is a formal theory that not only
explains the core phenomena of interest, but is also consistent with a range of

200
 6.4. An Abductive Approach to Formal Theory Construction

empirical data models. Our confidence in such a theory will grow the more data
models and the more complex the data models that are consistent with theory,
especially if the theory is able to achieve this consistency with minimal ad hoc
adjustments. In other words, if a theory has achieved these aims, we can be in-
creasingly confident that it is a good representation of the target system and can
prepare to subject the theory to more rigorous testing.

6.4.3 Testing Theory

Well developed formal theories allow one to derive unambiguous hypotheses (i.e.
predictions), which in turn provide the opportunity to conduct strong tests of the
theory. Notably, much of psychiatric research begins at this stage of hypothesis
testing, typically through null hypothesis testing. However, the theories from
which these hypotheses are derived are often unclear and, as we have argued in
this paper, the process by which hypotheses are derived from these theories is
opaque and likely prone to error. As a result, it is often unclear whether these
hypothesis tests are an appropriate test of the theory and difficult to know how
the results of such tests can further inform theory construction. This is perhaps
not surprising as the hypothetico-deductive framework in which much of this re-
search is conducted has very little to say about where these theories come from
or how they should be developed (Haig, 2005). In contrast, in the framework
proposed here, we have detailed a process by which formal theories can be gener-
ated and developed, equipping us to generate precise hypotheses which, in turn,
allows us to better test the theory. Accordingly, while this framework is primarily
abductive in nature with a focus on a theory’s ability to explain phenomena, we
also believe this framework substantially strengthens hypothesis testing as a tool
for evaluating theories.
Importantly, the theory testing stage calls for strong tests of a theory: risky
predictions (Meehl, 1990) that render the theory vulnerable to refutation. Strong
tests have at least two key features. First, strong tests entail the prediction of ob-
servations that, absent the theory, we would not otherwise expect. For example,
the panic disorder model we have discussed throughout this paper predicts that
the time to recover from an induction of arousal-related bodily sensations should
indicate vulnerability to panic attacks and, thus, should prospectively predict
the onset of panic attacks (for details, see Robinaugh, Haslbeck, et al., 2019).
To our knowledge, this is not a prediction that has arisen in the context of any
other theory and has never been tested. A study testing and finding support for
this prediction would lend more credence to this theory than a study testing a
prediction we would otherwise expect (e.g., that recurrent panic attacks will be
correlated with avoidance behavior).
Second, strong tests entail precise predictions. That is, a prediction that goes
beyond merely positing a refutation of the null hypothesis (e.g., a statistically
significant association) or even a directional prediction (e.g., a positive associa-
tion) to instead make precise point predictions about what should be observed.
For example, a very well-developed theory of panic disorder would be able to
predict the precise value of perceived threat (or interval of perceived threat val-

201
6. Modeling Psychopathology: From Data Models to Formal Theories

ues) which are likely to result from a particular arousal-inducing manipulation

(e.g., by breathing CO2 enriched air; Roberson-Nay et al., 2017). In other words,
just as in the theory development stage, the theory testing stage calls for us to
deduce the precise data models implied by our theory and to compare those im-
plied data models to empirical data models. There is, thus, a fine distinction
between theory testing and theory development in this framework. In the theory
development stage, these comparisons are carried out in the spirit of improving
upon and refining the theory. In the theory testing stage, these comparisons are
carried out with the aim of subjecting the theory to refutation. A discrepancy be-
tween theory-implied and empirical data models in the development stage calls
for refinement of the theory. A discrepancy between the theory-implied and em-
pirical data models in the testing stage calls for the theorist to deeply consider
the appropriateness of the theory.
Importantly, we are not proposing a “naive falsificationism” approach to the-
ory testing in which a failed test requires abandonment of the theory (Meehl,
1990). A discrepancy between theory implied and empirical data models can
provide an opportunity to improve upon a theory, returning us to the stage of
theory development. Nonetheless, a risky test should entail risk and repeated
failures at this stage should push the theorist toward the generation of a new
competing theory. For that reason, we believe that these risky tests should be
engaged in only when the theorist is sufficiently confident in the theory that they
would be willing to stake its survival on the outcome of the test. To that end,
we would make several recommendations. First, as we have stressed throughout
this paper, formal theories will strengthen confidence in the predictions being
tested by ensuring that they have been correctly deduced. Indeed, the level of
specificity required for predictions to constitute a strong test of the theory all but
requires that the theory be formalized. Second, the stage of theory development
should be used to not only improve upon the theory, but also the assumptions
about the instruments, measurements, and analyses that may also be responsi-
ble for any discrepancies between the theory-implied and empirical data models.
The approach we have argued for in producing implied data models is helpful
in this regard, as it forces the theory to formalize not only the theory, but also
the measurement of variables; what we have termed “emulated measurement”
(see Section 6.3.3). Strengthening confidence in these “emulated measurements”
will strengthen the test of the theory, as such issues cannot so readily be blamed
should the test fail. Third, research at this stage must be confirmatory in the
strictest sense of the term (Wagenmakers, Wetzels, Borsboom, van der Maas, &
Kievit, 2012). These studies should be preregistered, ideally with model simula-
tions showing the precise theory, measurement, and analysis that will be used in
the study.
If a theory fails a strong test, the decision of how to proceed depends upon
what Meehl referred to as the “money in the bank” principle (Meehl, 1990): If a
theory has a track record of success, it would be unwise to discard the theory in
the face of a single, or even several, failed tests. For Meehl, money in the bank
was accumulated by passing risky tests. A theory that has passed many such tests
should be retained more readily than a theory with no such record. We would

202
 6.5. Conclusions

argue for a broader conceptualization that draws on a wider range of criteria for
theory appraisal, with particular emphasis on explanatory breadth. A formal
theory that can explain a range of phenomena should be retained more readily
than a theory that accounts for only a narrow set of phenomena. Nonetheless,
we believe that any failure of a strong test should be taken as a serious challenge
to the theory that, at a minimum, warrants careful consideration about how to
proceed.
If a theory passes a strong test, it is corroborated, with the strength of corrob-
oration proportional to the strength of the test. Notably, because strong tests all
but require the evaluation of predictions made by the theory, a theory that has
passed several such tests will have demonstrated a strong capacity for support-
ing prediction. Accordingly, a theory that has moved from generation, through
development, and testing will emerge well equipped to support not only the ex-
planation, but also the prediction and control of mental disorders.

6.5 Conclusions
In this paper, we have argued that psychiatry needs formal theories and we have
examined how data models can best inform the development of such theories.
We focused especially on the network approach to psychopathology and con-
sidered three possible routes by which conditional dependence networks may
inform formal theories about how mental disorders operate as complex systems.
We found that these data models were not themselves capable of representing the
structure we presume will be needed for a theory of mental disorders. Perhaps
more surprisingly, we also found that we were unable to draw clear and reli-
able inferences from data models about the underlying system. Together, these
findings suggest that merely gathering data models alone is unlikely to readily
inform a well-developed formal theory. Instead, we found that the most promis-
ing use of empirical data models for theory development was to compare them
to “implied data models” derived from an initial formal theory. In this approach,
formal theories play an active role in their own development, with initial formal-
ized theories being refined over time through ongoing comparison of implied and
empirical data models.
Importantly, our analysis is not a critique of the specific data models we ex-
amined here, nor is it a dismissal of their value. Quite the opposite. We believe
these data models provide rich and valuable information about the relationships
among components of a system. However, our analysis strongly suggests that
the network approach to psychopathology cannot survive on these data models
alone. Formal theory is needed if the network approach is to move toward the ex-
planation, prediction, and control of mental disorders. Indeed, there is growing
recognition that formal theories are needed if we are to avoid problems associated
with conflicting empirical results (i.e. the “Replication Crisis”; Collaboration et
al., 2015) and move toward an accumulation of knowledge in scientific research
(e.g., Muthukrishna & Henrich, 2019; Szollosi & Donkin, 2019; Yarkoni, 2019;
Ioannidis, 2014). Accordingly, as a field psychiatry must grapple not only with

203
6. Modeling Psychopathology: From Data Models to Formal Theories

methods for the collection and analysis of data, but also methods for the genera-
tion and development of formal theories
The research framework we proposed in Section 6.4 is intended to be a first
step toward such a method of theory construction. At the heart of this approach
is the use of formalized initial theories to start a cycle of theory development in
which empirical data informs ongoing theory development and these improved
theories inform subsequent empirical research. Critically, this research frame-
work is not intended to suggest that all researchers must develop expertise in
computational modeling. Data and the detection of robust empirical phenom-
ena are central to psychiatric research in our proposed framework. However, our
framework does suggest that, as a field, psychiatry must do more to develop ex-
pertise in computational modeling within its ranks. We suspect that it will only
be through ongoing collaboration among theorists and empirical researchers that
we will be able to leverage the empirical literature to produce genuine advances
in our ability to explain, predict, and control psychopathology.

204
 6.A. Simulated Data from the Panic Model

Appendix 6.A Simulated Data from the Panic Model

In this appendix we describe in more detail how the simulated datasets, pre-
sented in both Sections 6.3.2 and 6.3.3 are obtained.
Data is simulated from the Panic Model, the full specification of which is
given by (Robinaugh, Haslbeck, et al., 2019), using the statistical programming
language R. We use the Panic Model to generate time-series data of 1000 individ-
uals, on a single minute time scale, for 12 weeks, using Euler’s method with a step
size of .001. This yields a total of nt = 12, 0960 repeated measurements per per-
son. Each individual starts with a different initial value of arousal schema, drawn
from a normal distribution with µ = 0.25 and σ = 0.0225. The parameters of this
distribution were chosen to roughly generate a representative number of panic
disorder sufferers (for more details see Robinaugh, Haslbeck, et al., 2019). Oth-
erwise each individual obtains the same parameter values and the same starting
values on all processes, with the stochastic noise terms drawn using a different
random seed for each individual. The mapping from this raw data to the vari-
ables used in the network models of Section 6.3.2 is described in the main text.
Code to reproduce this data-generation scheme can be found in the reproducibil-
ity archive of this paper.9

Appendix 6.B Additional Details: The Panic Model

and Statistical Dependencies
In this appendix we describe in more detail the patterns of statistical dependen-
cies produced by the three data models fitted to data simulated from the Panic
Model in Section 6.3.2. While in the main text we discuss the statistical depen-
dencies between between Arousal and Perceived Threat, and Arousal Schema
and Avoidance, here we focus only on the former. Key to the Arousal-Perceived
Threat dependencies is the positive feedback loop between Arousal and Perceived
Threat in the Panic Model (as described in Section 6.2). If Arousal and Perceived
Threat become sufficiently elevated, this “vicious cycle” leads to runaway pos-
itive feedback, with a pronounced spike in both Arousal and Perceived Threat
(i.e. a panic attack). This spike initiates a process of homeostatic feedback that
brings Arousal down and suppresses Arousal below its baseline for a period of
time after this panic attack, a period which we will refer to as a recovery period.
The panic attack itself lasts about 30 minutes. However, the recovery period lasts
for 2-3 hours (see Figure 6.9 panel (a)).
9 https://osf.io/bnteg/

205
6. Modeling Psychopathology: From Data Models to Formal Theories

Figure 6.9: Panel (a) depicts Arousal during a panic attack, showing the short sharp peak of arousal
levels, followed by a longer recovery period of low arousal, before the system returns to the usual
resting state. The dotted line indicates the mean level of arousal over the observation window (0 -
10 hrs). Panel (b) depicts the state-space plot of Perceived Threat and Arousal at the next measure-
ment occasion, as captured by the emulated ESM study and VAR model. Grey points indicate an
observation window of 90 min in which either part of a panic attack or the following recovery pe-
riod is captured. The solid grey line reflects the marginal lagged relationship. Panel (c) depicts the
cross-sectional marginal relationship between the mean of Arousal and mean of Perceived Threat, as
analyzed in the GGM model. Grey dots indicate individuals who suffer from panic attacks, and black
dots represent “healthy” individuals. The solid grey line shows the negative marginal relationship.
The dotted grey lines indicate the median of both variables, by which the binarized values used in
the Ising model analysis are defined.

In the VAR model in Figure 6.5 (b) in the main text, we observed a strong neg-
ative conditional relationship between Perceived Threat at time t and Arousal
at time t + 1, conditioning on all other variables at time t. The distribution of
these lagged variables is shown in Figure 6.9 (b), with the grey line representing
the also negative marginal relationship. This strong negative cross-lagged rela-
tionship is a direct consequence of the recovery period of Arousal: High values
of Perceived Threat are closely followed by a long period of low Arousal values.
This can be seen in Figure 6.9 (b), where observations over windows in which a
panic attack and recovery period occur are shaded in red. By averaging arousal
values over a window of 90 minutes, the strong positive causal effects operating
locally in time (i.e. over a very short time-interval) are not directly captured, but
instead the VAR(1) model describes correctly describes the negative relationship
between the means of each variable over this window.
In the GGM in Figure 6.5 (c) in the main text, we saw a positive linear re-
lationship between Arousal and Perceived Threat in the estimated GGM. This
dependency indicates that high mean levels of Arousal are associated with high
mean levels of Perceived Threat, conditional on all other variables. We stress the
conditional nature of this relationship, because the marginal relationship between
the two variables is in fact negative as can be seen in Figure 6.9 (c). This nega-
tive marginal relationship comes about by combining two groups of individuals
that have different mean values on both variables. Individuals who experience
panic attacks (grey points) have high average Perceived Threat, but low average
Arousal, due to the long recovery period of Arousal after a panic attack. On the
other hand, individuals who do not experience panic attacks have higher average
values of Arousal, and lower average values of Perceived Threat. When inspecting

206
 6.C. Details Empirical vs Simulated Ising Model

the two groups separately, we see that there is a positive linear relationship be-
tween mean Arousal and Perceived Threat in the group without panic attacks; the
group with panic attacks is too small to determine a relationship. Since Escape
and Avoidance behavior only occur after Panic attacks, conditioning on those two
variables amounts to conditioning on whether an individual had panic attacks.
This conditional relationship is then driven mostly by the positive relationship
in the (much larger) group of individuals who have no panic attacks, indicated
by the black dots in Figure 6.9 (c).
Finally, we can explain the weak positive relationship between Arousal and
Perceived Threat in the Ising model (Figure 6.5 (d) in the main text): The levels
of these variables are defined by a median split of their mean values, depicted as
dotted lines in Figure 6.9 (c). Unlike in the GGM, there is a positive marginal rela-
tionship between these binarized variables, as the majority of individuals without
panic attacks (denoted by the black points) end up in the low Perceived Threat
and low Arousal groups (lower left quadrant Figure 6.9 (c)) or high Perceived
Threat and high Arousal groups (upper right quadrant). How then do we end up
with a weakly positive conditional relationship between these two binary vari-
ables? Similarly to the GGM above, it turns out that conditioning on variables
such as Escape behaviour and Avoidance almost entirely separates individuals
into either the low Arousal and low Perceived Threat category (for low Escape
values) or the high Arousal and high Perceived Threat category (for high Escape
value). This means that, once we have conditioned on other variables which have
direct and indirect causal connections to Arousal and Perceived Threat, there is
very little additional information which Arousal can add to predicting Perceived
Threat levels (and vice versa). This produces the weak positive conditional rela-
tionship between Arousal and Perceived Threat, as well as the stronger positive
connections between Avoidance and Perceived Threat.

Appendix 6.C Details Empirical vs Simulated Ising

Model
In this appendix we describe in more detail how the theory-implied and empiri-
cal Ising Models presented in Section 6.3.3 are obtained.

6.C.1 Simulated Data and Implied Ising Model

To obtain the theory-implied Ising Model we use the raw time series data gener-
ated from the Panic Model and described in Appendix 6.A
To create the binary symptom variables in Section 6.3.3 we transformed the
raw time-series data of each individual as follows. First, we define Anxiety at
a given time point as the geometric mean of the Arousal and Perceived Threat
components at that point in time. Second, we define a panic attack as short, sharp
peak of Arousal and Perceived Threat. We code a panic attack to be present in
the time series data if Anxiety takes on a value greater than 0.5. The duration of
a panic attack is the length of time Anxiety variable stays above this threshold,

207
6. Modeling Psychopathology: From Data Models to Formal Theories

and so we define a single panic attack as a sequence of consecutive time points

in which Anxiety stays over this threshold. This allows to define our first binary
symptom variable, Recurrent Panic Attacks:
1. Recurrent Panic Attacks (PA) : PA is present if the individual experience
more than three panic attacks over the observation window.
We define recurrent as more than three over the observation window for consis-
tency with how this symptom is defined in the CPES dataset, detailed below.
Next, we can define the symptom Persistent Concern (PC), again using the
time series of Anxiety. This symptom is typically described as experiencing a
heightened level of anxiety following a panic attack (American Psychiatric As-
sociation, 2013). To define this, for each individual who experiences a panic at-
tack, we calculate the mean level of Anxiety in a window of 1000 minutes (16.67
hours) following the end of each panic attack. If another panic attack occurs in
that window, we instead take the mean level of Anxiety between the end of one
panic attack and before the beginning of the next. This gives us a vector of mean
Anxiety levels per person, one for each panic attack experienced. Next, we must
define what we consider to be a “heightened” level of anxiety. We do this by ob-
taining the distribution of mean Anxiety levels for healthy individuals, that is,
those members of our sample who never experience a panic attack. We consider
mean Anxiety levels following an attack to be “heightened” if they are greater
than the 90th percentile of mean Anxiety levels in the healthy population. This
gives us our second binary symptom variable.

2 Persistent Concern (PC): PC is present if, following at least one panic attack,
higher average levels of Anxiety are present than in the healthy population,
as defined by the 90th percentile of average Anxiety in the healthy popula-
tion.

Finally we take a similar approach to defining the symptom Avoidance (Av),

typically described as engaging in a heightened level of avoidance behaviour fol-
lowing a panic attack. For this symptom, we use the time series of the Avoid
component. For each individual who experiences a panic attack, we calculate the
mean level of Avoid in a window of 1000 minutes (16.67 hours) following the
end of each panic attack, or before the beginning of the next attack, whichever
is shorter. Heightened avoidance behaviour is defined relative to the 90th per-
centile of Avoid levels in the healthy population. This gives us our third binary
symptom variable.

3 Avoidance (Av): Av is present if, following at least one panic attack, higher
average levels of Avoid are present than in the healthy population, as de-
fined by the 90th percentile of average Avoid in the healthy population.

The Ising model of these three symptom variables is fit using the EstimateIsing
function from the IsingSampler package (Epskamp, 2015), that is, using a non-
regularized pseudolikelihood method.

208
 6.C. Details Empirical vs Simulated Ising Model

6.C.2 Empirical Symptom Data

To test the empirical predictions of the Panic Model, we made use of the pub-
licly available Collaborative Psychiatric Epidemiology Surveys (CPES) 2001-2003
(Alegria et al., 2007). The CPES is a nationally representative survey of mental
disorders and correlates in the United States. The CPES is attractive to use for
testing the Panic Model, first because of the large sample size (20,013 partic-
ipants) ensuring reliable estimates of empirical dependencies, and second, be-
cause approximately 146 items in the survey assess either panic attack or panic
disorder experiences, characteristics, and diagnoses, typically in terms of lifetime
prevalence.
To define our three panic disorder symptoms, we first use the diagnostic man-
ual of the CPES to define whether individuals have ever experienced a panic at-
tack based on responses to 18 items. There are three criteria which must be met
for the individual to be classed as having experienced at least one lifetime panic
attack. These are shown in Table 6.1. In coding the presence or absence of a panic
attack, individuals must positively report at least four out of the thirteen charac-
tersistics of a panic attack, according to the second criteria in Table 6.1. Missing
values were taken as a failure to report that characteristic.

Criterion Description Item number(s)

A A discrete period of intense fear or discomfort SC20 or SC20a
B (four or more) Palplitations, pounding heart PD1a
Sweating PD1e
Trembling or shaking PD1f
Sensation of shortness of breath or smothering PD1b
Feeling of choking PD1h
Chest pain or discomfort PD1i
Nausea or abdominal distress PD1c
Feeling dizzy, unsteady, lightheaded or faint PD1d or PD1m
Derealization or depersonalization PD1k or PD1l
Fear of losing control or going crazy PD1j
Fear of dying PD1n
Paresthesias (numbing or tingling sensations) PD1p
Chills or hot flushes PD1o
C Symptoms developed abruptly and reached a PD3
peak within 10 minutes

Table 6.1: Description of the three criteria (A, B and C) necessary to code an individual as having one
lifetime panic attack based on items from the CPES survey, based on the CPES diagnostic manual

With this definition of a panic attack in place, we define the three binary
symptoms of panic disorder, following the definitions laid out in the diagonistic
manual for Panic Disorder.
1. Recurrent Panic Attacks (PA). PA is present if participant reports more than
three lifetime occurrences of an unexpected, short, sharp attack of fear or
panic (item PD4 and all three criteria in Table 6.1), more than one of which
is out of the blue (PD17a)

209
6. Modeling Psychopathology: From Data Models to Formal Theories

2. Persistent Concern (PC). PC is present if participants reports that following

an attack, they experienced a month or more of at least one of: a) persis-
tent concern about having another attack (PD13a), or b) worry about the
implications or consequences of having an attack (PD13b)

3. Avoidance (Av). Av is present if participant reports at least one of a) follow-

ing an attack, changing everyday activities for a month or more (PD13c), b)
following an attack, avoiding situations due to fear of having an attack for
a month or more (PD13d), or c) in the past 12 months, avoiding situations
that might cause physical sensation (PD42).

In coding this, if two out of three PA criteria were present, and the third was
missing, we assigned a positive value to the PA item. The empirical Ising model
was fit using the same procedure as the theory-implied Ising model.

210
 References
Aalen, O. O. (1987). Dynamic modelling and causality. Scandinavian Actuarial
Journal, 1987(3-4), 177–190.
Aalen, O. O., Borgan, Ø., Keiding, N., & Thormann, J. (1980). Interaction be-
tween life history events. nonparametric analysis for prospective and retrospec-
tive data in the presence of censoring. Scandinavian Journal of Statistics, 161–
171.
Aalen, O. O., Gran, J., Røysland, K., Stensrud, M., & Strohmaier, S. (2018).
Feedback and mediation in causal inference illustrated by stochastic process
models. Scandinavian Journal of Statistics, 45, 62–86.
Aalen, O. O., Røysland, K., Gran, J., Kouyos, R., & Lange, T. (2016). Can we
believe the DAGs? A comment on the relationship between causal DAGs and
mechanisms. Statistical methods in medical research, 25(5), 2294–2314.
Aalen, O. O., Røysland, K., Gran, J., & Ledergerber, B. (2012). Causality, media-
tion and time: a dynamic viewpoint. Journal of the Royal Statistical Society: Series
A (Statistics in Society), 175(4), 831–861.
Abadir, K. M., & Magnus, J. R. (2005). Matrix Algebra (Vol. 1). Cambridge
University Press.

Achen, C. H. (2005). Let’s put garbage-can regressions and garbage-can probits

where they belong. Conflict Management and Peace Science, 22(4), 327–339.
Alegria, M., Jackson, J. S., Kessler, R. C., & Takeuchi, D. (2007). Collaborative Psy-
chiatric Epidemiology Surveys (CPES), 2001-2003 [United States]. Inter-university
Consortium for Political and Social Research.

American Psychiatric Association. (2013). Diagnostic and statistical manual of

mental disorders: DSM-5 (5th ed.). Washington, DC.
Andersson, S. A., Madigan, D., Perlman, M. D., et al. (1997). A characteriza-
tion of Markov equivalence classes for acyclic digraphs. The Annals of Statistics,
25(2), 505–541.

Angrist, J. D., Imbens, G. W., & Rubin, D. B. (1996). Identification of causal ef-
fects using instrumental variables. Journal of the American Statistical Association,
91(434), 444–455.

213
REFERENCES

Armour, C., Fried, E. I., Deserno, M. K., Tsai, J., & Pietrzak, R. H. (2017). A net-
work analysis of DSM-5 posttraumatic stress disorder symptoms and correlates
in us military veterans. Journal of Anxiety Disorders, 45, 49–59.
Asparouhov, T., Hamaker, E. L., & Muthén, B. (2018). Dynamic structural equa-
tion models. Structural Equation Modeling: A Multidisciplinary Journal, 25(3),
359–388.
Atkinson, K. E. (1989). An introduction to numerical analysis. New York, NY:
John Wiley & Sons.

Bailer-Jones, D. M. (2009). Scientific models in philosophy of science. Pittsburgh,

PA: University of Pittsburgh Press.
Bak, M., Drukker, M., Hasmi, L., & van Os, J. (2016). An n=1 clinical network
analysis of symptoms and treatment in psychosis. PloS one, 11(9), e0162811.
Bastiaansen, J. A., Kunkels, Y. K., Blaauw, F., Boker, S. M., Ceulemans, E., Chen,
M., . . . et al. (2019, Mar). Time to get personal? the impact of researchers’ choices
on the selection of treatment targets using the experience sampling methodology.
PsyArXiv. Retrieved from psyarxiv.com/c8vp7 doi: 10.31234/osf.io/c8vp7
Ben-Zeev, D., & Young, M. A. (2010). Accuracy of hospitalized depressed pa-
tients’ and healthy controls’ retrospective symptom reports: An experience sam-
pling study. The Journal of Nervous and Mental Disease, 198(4), 280–285.
Bernat, D. H., August, G. J., Hektner, J. M., & Bloomquist, M. L. (2007). The early
risers preventive intervention: Testing for six-year outcomes and mediational
processes. Journal of abnormal child psychology, 35(4), 605–617.

Bisconti, T., Bergeman, C. S., & Boker, S. M. (2004). Emotional well-being in re-
cently bereaved widows: A dynamical system approach. Journal of Gerontology,
Series B: Psychological Sciences and Social Sciences, 59, 158-167.
Bogen, J., & Woodward, J. (1988). Saving the phenomena. The Philosophical
Review, 97(3), 303–352.

Boker, S. M. (2002). Consequences of continuity: The hunt for intrinsic prop-

erties within parameters of dynamics in psychological processes. Multivariate
Behavioral Research, 37(3), 405–422.
Boker, S. M., Deboeck, P., Edler, C., & Keel, P. (2010). Generalized local lin-
ear approximation of derivatives from time series. In S.-M. Chow & E. Ferrar
(Eds.), Statistical methods for modeling human dynamics: An interdisciplinary dia-
logue (p. 179-212). Boca Raton, FL: Taylor & Francis.
Boker, S. M., & McArdle, J. J. (1995). Statistical vector field analysis applied to
mixed crosssectional and longitudinal data. Experimental Aging Research, 21(1),
77-93.

214
 REFERENCES

Boker, S. M., Montpetit, M. A., Hunter, M. D., & Bergeman, C. S. (2010). Mod-
eling resilience with differential equations. In P. Molenaar & K. Newell (Eds.),
Learning and development: Individual pathways of change (p. 183-206). Washing-
ton, DC: American Psychological Association.
Boker, S. M., Neale, M., & Rausch, J. (2004). Latent differential equation model-
ing with multivariate multi-occasion indicators. In K. van Montfort, J. H. Oud,
& A. Satorra (Eds.), Recent developments on structural equation models (pp. 151–
174). Dordrecht, the Netherlands: Kluwer Academic.
Boker, S. M., & Nesselroade, J. R. (2002). A method for modeling the intrin-
sic dynamics of intraindividual variability: Recovering parameters of simulated
oscillators in multi-wave panel data. Multivariate Behavioral Research, 37, 127–
160.
Boker, S. M., Staples, A. D., & Hu, Y. (2016). Dynamics of change and change in
dynamics. Journal for person-oriented research, 2(1-2), 34–55.
Bolger, N., & Laurenceau, J.-P. (2013). Intensive longitudinal methods: An intro-
duction to diary and experience sampling research. NY: New York: The Guilford
Press.
Bollen, K. A. (1987). Total, direct, and indirect effects in structural equation
models. Sociological methodology, 37–69.
Bollen, K. A. (1989). Structural Equations with Latent Variables. Oxford, England,
John Wiley & Sons.
Bongers, S., & Mooij, J. M. (2018). From random differential equations to struc-
tural causal models: The stochastic case. arXiv preprint arXiv:1803.08784.
Borgatti, S. P. (2005). Centrality and network flow. Social networks, 27(1), 55–71.
Borsboom, D. (2017). A network theory of mental disorders. World Psychiatry,
16(1), 5–13.
Borsboom, D., & Cramer, A. O. (2013). Network analysis: an integrative ap-
proach to the structure of psychopathology. Annual Review of Clinical Psychol-
ogy, 9, 91–121.
Borsboom, D., Cramer, A. O., Schmittmann, V. D., Epskamp, S., & Waldorp, L. J.
(2011). The small world of psychopathology. PloS one, 6(11), e27407.
Borsboom, D., van der Sluis, S., Noordhof, A., Wichers, M., Geschwind, N.,
Aggen, S., . . . others (2012). What kind of causal modelling approach does
personality research need? European Journal of Personality, 26, 372–390.
Bos, F. M., Snippe, E., de Vos, S., Hartmann, J. A., Simons, C. J., van der Krieke,
L., . . . Wichers, M. (2017). Can we jump from cross-sectional to dynamic inter-
pretations of networks implications for the network perspective in psychiatry.
Psychotherapy and Psychosomatics, 86(3), 175–177.

215
REFERENCES

Boschloo, L., Schoevers, R. A., van Borkulo, C. D., Borsboom, D., & Oldehinkel,
A. J. (2016). The network structure of psychopathology in a community sample
of preadolescents. Journal of Abnormal Psychology, 125(4), 599–606.
Bramsen, R. H., Lasgaard, M., Koss, M. P., Shevlin, M., Elklit, A., & Banner,
J. (2013). Testing a multiple mediator model of the effect of childhood sexual
abuse on adolescent sexual victimization. American Journal of Orthopsychiatry,
83(1), 47–54.
Bringmann, L. F., Elmer, T., Epskamp, S., Krause, R. W., Schoch, D., Wichers, M.,
. . . Snippe, E. (2019). What do centrality measures measure in psychological
networks? The journal of abnormal psychology.
Bringmann, L. F., Lemmens, L., Huibers, M., Borsboom, D., & Tuerlinckx,
F. (2015). Revealing the dynamic network structure of the beck depression
inventory-ii. Psychological medicine, 45(4), 747–757.
Bringmann, L. F., Pe, M. L., Vissers, N., Ceulemans, E., Borsboom, D., Van-
paemel, W., . . . Kuppens, P. (2016). Assessing temporal emotion dynamics using
networks. Assessment, 23(4), 425–435.
Bringmann, L. F., Vissers, N., Wichers, M., Geschwind, N., Kuppens, P., Peeters,
F., . . . Tuerlinckx, F. (2013). A Network Approach to Psychopathology: New
Insights into Clinical Longitudinal Data. PloS One, 8(4), e60188. doi: 10.1371/
journal.pone.0060188
Browne, M. W., & Nesselroade, J. R. (2005). Representing psychological pro-
cesses with dynamic factor models: Some promising uses and extensions of
ARMA time series models. In A. Maydue-Olivares & J. J. McArdle (Eds.), Psy-
chometrics: A festschrift to Roderick P. McDonald (p. 415-452). Mahwah, NJ:
Lawrence Erlbaum Associates.
Bulteel, K., Tuerlinckx, F., Brose, A., & Ceulemans, E. (2016). Using raw var
regression coefficients to build networks can be misleading. Multivariate behav-
ioral research, 51(2-3), 330–344.
Carpenter, B., Gelman, A., Hoffman, M. D., Lee, D., Goodrich, B., Betancourt,
M., . . . Riddell, A. (2017). Stan: A probabilistic programming language. Journal
of Statistical Software, 76(1).
Cartwright, N. (1983). How the laws of physics lie. Oxford University Press.
Cartwright, N. (1999). Causal diversity and the markov condition. Synthese,
121(1-2), 3–27.
Cartwright, N. (2007). Hunting causes and using them: Approaches in philosophy
and economics. Cambridge University Press.
Chasalow, S. (2012). combinat: combinatorics utilities [Computer software
manual]. Retrieved from https://CRAN.R-project.org/package=combinat
(R package version 0.0-8)

216
 REFERENCES

Chow, S.-M. (2019). Practical tools and guidelines for exploring and fitting lin-
ear and nonlinear dynamical systems models. Multivariate Behavioral Research,
1–29.
Chow, S.-M., Ferrer, E., & Hsieh, F. (2011). Statistical methods for modeling human
dynamics: An interdisciplinary dialogue. New York: Routledge.
Chow, S.-M., Ferrer, E., & Nesselroade, J. R. (2007). An unscented kalman filter
approach to the estimation of nonlinear dynamical systems models. Multivariate
Behavioral Research, 42(2), 283–321.
Chow, S.-M., Ou, L., Ciptadi, A., Prince, E. B., You, D., Hunter, M. D., . . .
Messinger, D. S. (2018). Representing sudden shifts in intensive dyadic in-
teraction data using differential equation models with regime switching. Psy-
chometrika, 83(2), 476–510.
Chow, S.-M., Ram, N., Boker, S., Fujita, F., Clore, G., & Nesselroade, J. (2005).
Capturing weekly fluctuation in emotion using a latent differential structural
approach. Emotion, 5(2), 208–225.
Christian, C., Perko, V., Vanzhula, I., Tregarthen, J., Forbush, K., & Levinson, C.
(2019). Eating disorder core symptoms and symptom pathways across develop-
mental stages: A network analysis. Journal of abnormal psychology.
Clark, D. M. (1986). A cognitive approach to panic. Behaviour Research and
Therapy, 24(4), 461–470.
Cole, D. A., & Maxwell, S. E. (2003). Testing mediational models with longitudi-
nal data: questions and tips in the use of structural equation modeling. Journal
of Abnormal Psychology, 112(4), 558–557.
Coleman, J. S. (1968). The mathematical study of change. Methodology in social
research, 428–478.
Collaboration, O. S., et al. (2015). Estimating the reproducibility of psychologi-
cal science. Science, 349(6251), aac4716.
Contreras, A., Nieto, I., Valiente, C., Espinosa, R., & Vazquez, C. (2019). The
study of psychopathology from the network analysis perspective: A systematic
review. Psychotherapy and Psychosomatics, 1–13.
Costantini, G., Epskamp, S., Borsboom, D., Perugini, M., Mõttus, R., Waldorp,
L. J., & Cramer, A. O. (2015). State of the aRt personality research: A tutorial
on network analysis of personality data in R. Journal of Research in Personality,
54, 13–29.
Cox, D. R., & Wermuth, N. (1996). Multivariate dependencies: Models, analysis
and interpretation. Chapman and Hall/CRC.
Cramer, A. O., van Borkulo, C. D., Giltay, E. J., van der Maas, H. L., Kendler,
K. S., Scheffer, M., & Borsboom, D. (2016). Major depression as a complex
dynamic system. PLoS One, 11(12), e0167490.

217
REFERENCES

Cramer, A. O., Waldorp, L. J., van der Maas, H. L. J., & Borsboom, D. (2010).
Comorbidity: A network perspective. Behavioral and Brain Sciences, 33(2-3),
137–150. doi: 10.1017/S0140525X09991567
Dablander, F., & Hinne, M. (2019). Node centrality measures are a poor substi-
tute for causal inference. Scientific reports, 9(1), 6846.
Dablander, F., Ryan, O., & Haslbeck, J. M. (2019, Jan). Choosing between AR(1)
and VAR(1) models in typical psychological applications. PsyArXiv. Retrieved from
psyarxiv.com/qgewy doi: 10.31234/osf.io/qgewy

Dahlhaus, R., & Eichler, M. (2003). Causality and graphical models in time
series analysis. Oxford Statistical Science Series, 115–137.
Dalege, J., Borsboom, D., van Harreveld, F., van den Berg, H., Conner, M., &
van der Maas, H. L. (2016). Toward a formalized account of attitudes: The
Causal Attitude Network (CAN) model. Psychological Review, 123(1), 2–22.

David, S. J., Marshall, A. J., Evanovich, E. K., & Mumma, G. H. (2018). Intraindi-
vidual dynamic network analysis–implications for clinical assessment. Journal
of psychopathology and behavioral assessment, 40(2), 235–248.
Dawid, A. P. (1979). Conditional independence in statistical theory. Journal of
the Royal Statistical Society. Series B (Methodological), 1–31.

Dawid, A. P. (2002). Influence diagrams for causal modelling and inference.

International Statistical Review, 70(2), 161–189.
Dawid, A. P. (2010). Beware of the DAG! In Causality: Objectives and Assessment
(pp. 59–86).

Deboeck, P. R., & Preacher, K. J. (2016). No need to be discrete: A method for

continuous time mediation analysis. Structural Equation Modeling: A Multidisci-
plinary Journal, 23(1), 61–75.
De Haan-Rietdijk, S., Gottman, J. M., Bergeman, C. S., & Hamaker, E. L. (2016).
Get over it! a multilevel threshold autoregressive model for state-dependent
affect regulation. Psychometrika, 81(1), 217–241.
De Haan-Rietdijk, S., Voelkle, M. C., Keijsers, L., & Hamaker, E. (2017).
Discrete- versus continuous-time modeling of unequally spaced ESM data.
Frontiers in Psychology, 8, 1849.

Dempster, A. P. (1972). Covariance selection. Biometrics, 157–175.

Denollet, J., & De Vries, J. (2006). Positive and negative affect within the realm
of depression, stress and fatigue: The two-factor distress model of the global
mood scale (gms). Journal of affective disorders, 91(2), 171–180.

218
 REFERENCES

Deserno, M. K., Borsboom, D., Begeer, S., & Geurts, H. M. (2017). Multicausal
systems ask for multicausal approaches: A network perspective on subjective
well-being in individuals with autism spectrum disorder. Autism, 21(8), 960–
971.

de Wild-Hartmann, J. A., Wichers, M., van Bemmel, A. L., Derom, C., Thiery,
E., Jacobs, N., . . . Simons, C. J. (2013). Day-to-day associations between subjec-
tive sleep and affect in regard to future depressionin a female population-based
sample. The British Journal of Psychiatry, 202(6), 407–412.
Didelez, V. (2000). Graphical models for event history analysis based on local inde-
pendence. Logos Berlin.
Didelez, V. (2007). Graphical models for composable finite Markov processes.
Scandinavian Journal of Statistics, 34(1), 169–185.
Didelez, V. (2008). Graphical models for marked point processes based on
local independence. Journal of the Royal Statistical Society: Series B (Statistical
Methodology), 70(1), 245–264.
Didelez, V. (2019). Defining causal mediation with a longitudinal mediator and
a survival outcome. Lifetime data analysis, 25(4), 593–610.
Dormann, C., & Griffin, M. A. (2015). Optimal time lags in panel studies.
Psychological methods, 20(4), 489.
Driver, C. C., Oud, J. H., & Voelkle, M. C. (2017). Continuous time structural
equation modeling with R package ctsem. Journal of Statistical Software, 77(5),
1–35. doi: 10.18637/jss.v077.i05

Driver, C. C., & Voelkle, M. C. (2018). Understanding the time course of inter-
ventions with continuous time dynamic models. In K. L. Montfort, J. H. Oud,
& M. C. Voelkle (Eds.), Continuous time modeling in the behavioral and related
sciences (p. 179-203). New York: Springer.
Durbin, J., & Koopman, S. J. (2012). Time series analysis by state space methods.
Oxford university press.

Ebbinghaus, H. (1913/2013). Memory: A contribution to experimental psychol-

ogy. Annals of Neurosciences, 20(4), 155.
Eberhardt, F. (2013). Experimental indistinguishability of causal structures.
Philosophy of Science, 80(5), 684–696.

Eichler, M., & Didelez, V. (2010). On Granger causality and the effect of inter-
ventions in time series. Lifetime Data Analysis, 16(1), 3–32.
Einicke, G. A. (2019). Smoothing, filtering and prediction: Estimating the past,
present and future. New York, NY: Prime Publishing.

219
REFERENCES

Elliott-Graves, A. (2014). The role of target systems in scientific practice (Unpub-

lished doctoral dissertation). University of Pennsylvania, Philadelphia, Penn-
sylvania.
Epskamp, S. (2015). IsingSampler: Sampling methods and distribution
functions for the Ising model [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=IsingSampler (R package version
0.2)
Epskamp, S., Borsboom, D., & Fried, E. I. (2018). Estimating psychological
networks and their accuracy: A tutorial paper. Behavior Research Methods, 50(1),
195–212.
Epskamp, S., Cramer, A. O. J., Waldorp, L. J., Schmittmann, V. D., & Borsboom,
D. (2012). qgraph: Network visualizations of relationships in psychometric
data. Journal of Statistical Software, 48(4), 1–18. Retrieved from http://www
.jstatsoft.org/v48/i04/

Epskamp, S., & Fried, E. I. (2018). A tutorial on regularized partial correlation

networks. Psychological Methods.
Epskamp, S., Rhemtulla, M., & Borsboom, D. (2017). Generalized network
pschometrics: Combining network and latent variable models. Psychometrika,
82(4), 904–927.

Epskamp, S., van Borkulo, C. D., van der Veen, D. C., Servaas, M. N., Isvoranu,
A.-M., Riese, H., & Cramer, A. O. (2018). Personalized network modeling in psy-
chopathology: The importance of contemporaneous and temporal connections.
Clinical Psychological Science, 6(3), 416–427.

Epskamp, S., Waldorp, L. J., Mõttus, R., & Borsboom, D. (2018). The Gaussian
graphical model in cross-sectional and time-series data. Multivariate Behavioral
Research, 53(4), 453–480.
Epstein, J. M. (2008). Why model? Journal of Artificial Societies and Social
Simulation, 11(4), 12.

Fabio Di Narzo, A., Aznarte, J. L., & Stigler, M. (2009). tsDyn:

Time series analysis based on dynamical systems theory [Computer soft-
ware manual]. Retrieved from https://cran.r-project.org/package=tsDyn/
vignettes/tsDyn.pdf (R package version 0.7)
Fechner, G. T., Howes, D. H., & Boring, E. G. (1966). Elements of Psychophysics
(Vol. 1). Holt, Rinehart and Winston New York.
Fisher, A. J., & Boswell, J. F. (2016). Enhancing the personalization of psy-
chotherapy with dynamic assessment and modeling. Assessment, 23(4), 496–
506.

220
 REFERENCES

Fisher, A. J., Reeves, J. W., Lawyer, G., Medaglia, J. D., & Rubel, J. A. (2017). Ex-
ploring the Idiographic Dynamics of Mood and Anxiety via Network Analysis.
Journal of Abnormal Psychology, 126(8), 1044–1056. doi: 0.1037/abn0000311
Fisher, M. (2001). Modeling negative autoregression in continuous time.
(http://www.markfisher.net/ mefisher/papers/continuous ar.pdf)
Flake, J. K., & Fried, E. I. (2019, Jan). Measurement schmeasurement: Questionable
measurement practices and how to avoid them. PsyArXiv. Retrieved from psyarxiv
.com/hs7wm doi: 10.31234/osf.io/hs7wm
Fonseca-Pedrero, E., Ortuño, J., Debbané, M., Chan, R. C., Cicero, D., Zhang,
L. C., . . . Fried, E. I. (2018). The network structure of schizotypal personality
traits. Schizophrenia Bulletin, 44(2), S468–S479.
Forré, P., & Mooij, J. M. (2018). Constraint-based causal discovery for non-linear
structural causal models with cycles and latent confounders. arXiv preprint
arXiv:1807.03024.
Forré, P., & Mooij, J. M. (2019). Causal calculus in the presence of cycles, latent
confounders and selection bias. arXiv preprint arXiv:1901.00433.
Freedman, H. I. (1980). Deterministic mathematical models in population ecology
(Vol. 57). Marcel Dekker Incorporated.
Freedman, R. R., Ianni, P., Ettedgui, E., & Puthezhath, N. (1985). Ambulatory
monitoring of panic disorder. Archives of General Psychiatry, 42(3), 244–248.
Freeman, L. C. (1977). A set of measures of centrality based on betweenness.
Sociometry, 35–41.
Freeman, L. C. (1978). Centrality in social networks conceptual clarification.
Social networks, 1(3), 215–239.
Fried, E. I., Bockting, C., Arjadi, R., Borsboom, D., Amshoff, M., Cramer, A. O.,
. . . Stroebe, M. (2015). From loss to loneliness: The relationship between be-
reavement and depressive symptoms. Journal of Abnormal Psychology, 124(2),
256–265.
Fried, E. I., Boschloo, L., van Borkulo, C. D., Schoevers, R. A., Romeijn, J.-W.,
Wichers, M., . . . Borsboom, D. (2015). Commentary:“consistent superiority
of selective serotonin reuptake inhibitors over placebo in reducing depressed
mood in patients with major depression”. Frontiers in Psychiatry, 6, 117.
Fried, E. I., & Cramer, A. O. (2017). Moving forward: challenges and direc-
tions for psychopathological network theory and methodology. Perspectives on
Psychological Science, 12(6), 999–1020.
Fried, E. I., Epskamp, S., Nesse, R. M., Tuerlinckx, F., & Borsboom, D. (2016).
What are ’good’ depression symptoms? comparing the centrality of DSM and
non-DSM symptoms of depression in a network analysis. Journal of Affective
Disorders, 189, 314–320.

221
REFERENCES

Friedman, J., Hastie, T., & Tibshirani, R. (2001). The elements of statistical learn-
ing (Vol. 1) (No. 10). New York, NY: Springer Series in Statistics.
Friedman, J., Hastie, T., & Tibshirani, R. (2008). Sparse inverse covariance
estimation with the graphical lasso. Biostatistics, 9(3), 432–441.
Fukano, T., & Gunji, Y.-P. (2012). Mathematical models of panic disorder. Non-
linear Dynamics, Psychology, and Life Sciences, 16(4), 457–470.
Galles, D., & Pearl, J. (1995). Testing identifiability of causal effects. In Pro-
ceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence (pp.
185–195).
Gardner, C., & Kleinman, A. (2019). Medicine and the mind — the consequences
of psychiatry’s identity crisis. New England Journal of Medicine, 381(18), 1697-
1699. Retrieved from https://doi.org/10.1056/NEJMp1910603 doi: 10.1056/
NEJMp1910603
Gault-Sherman, M. (2012). It’s a two-way street: The bidirectional relationship
between parenting and delinquency. Journal of Youth and Adolescence, 41, 121-
145.
Glauber, R. J. (1963). Time-dependent statistics of the Ising model. Journal of
Mathematical Physics, 4(2), 294–307.
Glymour, M. M. (2006). Using causal diagrams to understand common prob-
lems in social epidemiology. Methods in Social Epidemiology, 393–428.
Golino, H. F., & Epskamp, S. (2017). Exploratory graph analysis: A new ap-
proach for estimating the number of dimensions in psychological research. PloS
One, 12(6), e0174035.
Gollob, H. F., & Reichardt, C. S. (1987). Taking account of time lags in causal
models. Child Development, 58, 80–92.
Granger, C. W. (1969). Investigating causal relations by econometric models
and cross-spectral methods. Econometrica: Journal of the Econometric Society,
424–438.
Greenland, S., & Robins, J. M. (1986). Identifiability, exchangeability, and epi-
demiological confounding. International journal of epidemiology, 15(3), 413–419.
Grimm, V., & Railsback, S. F. (2005). Individual-based modeling and e ecology.
Princeton University press.
Groen, R. N., Snippe, E., Bringmann, L. F., Simons, C. J., Hartmann, J. A., Bos,
E. H., & Wichers, M. (2019). Capturing the risk of persisting depressive symp-
toms: A dynamic network investigation of patients’ daily symptom experiences.
Psychiatry Research, 271, 640–648. doi: 10.1016/j.psychres.2018.12.054
Grömping, U. (2006). Relative importance for linear regression in R: The pack-
age relaimpo. Journal of Statistical Software, 17(1), 1–27.

222
 REFERENCES

Haig, B. D. (2005). An abductive theory of scientific method. Psychological

Methods, 10(4), 371–388.
Haig, B. D. (2008). Precis of ‘an abductive theory of scientific method’. Journal
of Clinical Psychology, 64(9), 1019–1022.
Haig, B. D. (2014). Investigating the psychological world: Scientific method in the
behavioral sciences. Cambridge, MA: MIT press.
Hamaker, E. L. (2012). Why researchers should think “within-person”: A
paradigmatic rationale. In M. R. Meehl & T. S. Conner (Eds.), (pp. 43–61). New
York, NY: Guilford.
Hamaker, E. L., & Dolan, C. V. (2009). Idiographic data analysis: Quantitative
methods—from simple to advanced. In Dynamic process methodology in the social
and developmental sciences (pp. 191–216). Springer.
Hamaker, E. L., Dolan, C. V., & Molenaar, P. C. M. (2005). Statistical mod-
eling of the individual: Rationale and application of multivariate time se-
ries analysis. Multivariate Behavioral Research, 40(2), 207-233. doi: 10.1207/
s15327906mbr4002\ 3
Hamaker, E. L., & Grasman, R. (2012). Regime switching state-space models ap-
plied to psychological processes: Handling missing data and making inferences.
Psychometrika, 77(2), 400–422.
Hamaker, E. L., Grasman, R. P., & Kamphuis, J. H. (2010). Regime-switching
models to study psychological process. In P. C. M. Molenaar & K. M. Newell
(Eds.), Individual pathways of change: Statistical models for analyzing learning and
development (p. 155-168). Washington, DC: American Psychological Associa-
tion.
Hamaker, E. L., Grasman, R. P., & Kamphuis, J. H. (2016). Modeling bas dys-
regulation in bipolar disorder: Illustrating the potential of time series analysis.
Assessment, 23(4), 436–446.
Hamaker, E. L., & Grasman, R. P. P. P. (2015). To center or not to center? inves-
tigating inertia with a multilevel autoregressive model. Frontiers in Psychology,
5, 1492. doi: 10.3389/fpsyg.2014.01492
Hamaker, E. L., Kuiper, R., & Grasman, R. P. P. P. (2015). A critique of the
cross-lagged panel model. Psychological Methods, 20(1), 102-116. doi: 10.1037/
a0038889
Hamaker, E. L., & Wichers, M. (2017). No time like the present: Discovering
the hidden dynamics in intensive longitudinal data. Current Directions in Psy-
chological Science, 26(1), 10–15.
Hamaker, E. L., Zhang, Z., & van der Maas, H. L. (2009). Using threshold
autoregressive models to study dyadic interactions. Psychometrika, 74(4), 727–
745.

223
REFERENCES

Hamerle, A., Nagl, W., & Singer, H. (1991). Problems with the estimation of
stochastic differential equations using structural equations models. Journal of
Mathematical Sociology, 16(3), 201–220.
Hamilton, J. D. (1989). A new approach to the economic analysis of nonstation-
ary time series and the business cycle. Econometrica: Journal of the Econometric
Society, 357–384.
Hamilton, J. D. (1994). Time series analysis (Vol. 2). Princeton, NJ: Princeton
University Press.

Haslbeck, J. M., Bringmann, L. F., & Waldorp, L. J. (2017). How to estimate

time-varying vector autoregressive models? a comparison of two methods. arXiv
preprint arXiv:1711.05204.
Haslbeck, J. M., Epskamp, S., Marsman, M., & Waldorp, L. J. (2018). Interpreting
the Ising model: The input matters. arXiv preprint arXiv:1811.02916.

Haslbeck, J. M., & Fried, E. I. (2017). How predictable are symptoms in psy-
chopathological networks? A reanalysis of 18 published datasets. Psychological
Medicine, 47(16), 2767–2776.
Haslbeck, J. M., & Ryan, O. (2019, Sep). Recovering bistable systems from psy-
chological time series. PsyArXiv. Retrieved from psyarxiv.com/kcv3s doi:
10.31234/osf.io/kcv3s
Haslbeck, J. M., & Waldorp, L. J. (2018). How well do network models predict
observations? on the importance of predictability in network models. Behavior
Research Methods, 50(2), 853–861.

Hayduk, L. A. (2009). Finite feedback cycling in structural equation models.

Structural Equation Modeling, 16(4), 658–675.
Heeren, A., & McNally, R. J. (2016). An integrative network approach to social
anxiety disorder: The complex dynamic interplay among attentional bias for
threat, attentional control, and symptoms. Journal of Anxiety Disorders, 42, 95–
104.

Hernan, M. A., & Robins, J. M. (2019). Causal inference. CRC Boca Raton, FL.
Higgs, P. W. (1964). Broken symmetries and the masses of gauge bosons. Phys-
ical Review Letters, 13(16), 508–509.
Hirsch, M. W., Smale, S., & Devaney, R. L. (2012). Differential equations, dynam-
ical systems, and an introduction to chaos. Academic press.
Hjelmeland, H., & Loa Knizek, B. (2018). The emperor’s new clothes? A critical
look at the interpersonal theory of suicide. Death Studies, 1–11. doi: 10.1080/
07481187.2018.1527796

224
 REFERENCES

Ho, D. D., Neumann, A. U., Perelson, A. S., Chen, W., Leonard, J. M., &
Markowitz, M. (1995). Rapid turnover of plasma virions and cd4 lymphocytes
in hiv-1 infection. Nature, 373(6510), 123–126.
Hoge, E. A., Bui, E., Marques, L., Metcalf, C. A., Morris, L. K., Robinaugh, D. J.,
. . . Simon, N. M. (2013). Randomized controlled trial of mindfulness meditation
for generalized anxiety disorder: effects on anxiety and stress reactivity. The
Journal of clinical psychiatry, 74(8), 786–792.
Hoorelbeke, K., Marchetti, I., De Schryver, M., & Koster, E. H. (2016). The
interplay between cognitive risk and resilience factors in remitted depression: a
network analysis. Journal of Affective Disorders, 195, 96–104.
Horn, E. E., Strachan, E., & Turkheimer, E. (2015). Psychological distress and
recurrent herpetic disease: A dynamic study of lesion recurrence and viral shed-
ding episodes in adults. Multivariate behavioral research, 50(1), 134–135.

Hosenfeld, B., Bos, E. H., Wardenaar, K. J., Conradi, H. J., van der Maas, H. L.,
Visser, I., & de Jonge, P. (2015). Major depressive disorder as a nonlinear dy-
namic system: bimodality in the frequency distribution of depressive symptoms
over time. BMC Psychiatry, 15(222).
Houben, M., Van Den Noortgate, W., & Kuppens, P. (2015). The relation between
short-term emotion dynamics and psychological well-being: A meta-analysis.
Psychological Bulletin, 141(4), 901–930.
Ichii, K. (1991). Measuring mutual causation: Effects of suicide news on suicides
in Japan. Social Science Research, 20, 188-195.
Ioannidis, J. P. (2014). How to make more published research true. PLOS
Medicine, 11(10). doi: e1001747
Ising, E. (1925). Beitrag zur theorie des ferromagnetismus. Zeitschrift für Physik
A Hadrons and Nuclei, 31(1), 253–258.
Isvoranu, A.-M., van Borkulo, C. D., Boyette, L.-L., Wigman, J. T., Vinkers, C. H.,
Borsboom, D., & Investigators, G. (2016). A network approach to psychosis:
pathways between childhood trauma and psychotic symptoms. Schizophrenia
Bulletin, 43(1), 187–196.
Johnston, J., & DiNardo, J. (1972). Econometric methods. New York, 19(7), 22.
Jones, P. J. (2018). networktools: Tools for identifying important nodes in net-
works [Computer software manual]. Retrieved from https://CRAN.R-project
.org/package=networktools (R package version 1.2.0)
Jones, P. J., Mair, P., Riemann, B. C., Mugno, B. L., & McNally, R. J. (2018).
A network perspective on comorbid depression in adolescents with obsessive-
compulsive disorder. Journal of Anxiety Disorders, 53, 1–8.

225
REFERENCES

Kaiser, T., & Laireiter, A.-R. (2018). Daily dynamic assessment and modelling of
intersession processes in ambulatory psychotherapy: A proof of concept study.
Psychotherapy Research, 1–12.
Kalisch, M., & Bühlmann, P. (2007). Estimating high-dimensional directed
acyclic graphs with the PC-algorithm. Journal of Machine Learning Research,
8(Mar), 613–636.
Kalisch, M., Mächler, M., Colombo, D., Maathuis, M. H., & Bühlmann, P. (2012).
Causal inference using graphical models with the R package pcalg. Journal of
Statistical Software, 47(11), 1–26. Retrieved from http://www.jstatsoft.org/
v47/i11/
Kalisch, R., Cramer, A. O., Binder, H., Fritz, J., Leertouwer, I., Lunansky, G.,
. . . Van Harmelen, A.-L. (2019). Deconstructing and reconstructing resilience:
a dynamic network approach. Perspectives on Psychological Science, 14(5), 765–
777.

Kellen, D. (2019). A model hierarchy for psychological science. Computational

Brain & Behavior, 2(3-4), 160–165.
Kendler, K. S. (2019). From many to one to many—the search for causes of
psychiatric illness. JAMA Psychiatry. doi: 10.1001/jamapsychiatry.2019.1200

Kim, C.-J., & Nelson, C. R. (1999). State-space models with regime switching:
Classical and Gibbs-sampling approaches with applications. Cambridge, MA: The
MIT Press. doi: 10.2307/2669796
King, A. A., Nguyen, D., & Ionides, E. L. (2015). Statistical inference for
partially observed Markov processes via the R package pomp. arXiv preprint
arXiv:1509.00503.
Knefel, M., Tran, U. S., & Lueger-Schuster, B. (2016). The association of post-
traumatic stress disorder, complex posttraumatic stress disorder, and borderline
personality disorder from a network analytical perspective. Journal of Anxiety
Disorders, 43, 70–78.

Kossakowski, J., Groot, P., Haslbeck, J. M., Borsboom, D., & Wichers, M. (2017).
Data from ‘critical slowing down as a personalized early warning signal for de-
pression’. Journal of Open Psychology Data, 5(1).
Koval, P., Kuppens, P., Allen, N. B., & Sheeber, L. (2012). Getting stuck in de-
pression: The roles of rumination and emotional inertia. Cognition and Emotion,
26, 1412-1427.
Kraemer, N., Schaefer, J., & Boulesteix, A.-L. (2009). Regularized estimation
of large-scale gene regulatory networks using gaussian graphical models. BMC
Bioinformatics, 10(384).

226
 REFERENCES

Kroeze, R., van der Veen, D. C., Servaas, M. N., Bastiaansen, J. A., Oude Voshaar,
R., Borsboom, D., & Riese, H. (2017). Personalized feedback on symptom
dynamics of psychopathology: A proof-of-principle study. Journal for Person-
Oriented Research, 3(1), 1–11.

Kuiper, R. M., & Ryan, O. (2018). Drawing conclusions from cross-lagged re-
lationships: Re-considering the role of the time-interval. Structural Equation
Modeling: A Multidisciplinary Journal, 25(5), 809–823.
Kulikov, G. Y., & Kulikova, M. V. (2013). Accurate numerical implementation of
the continuous-discrete extended Kalman filter. IEEE Transactions on Automatic
Control, 59(1), 273–279.
Kuppens, P., Allen, N. B., & Sheeber, L. B. (2010). Emotional inertia and psy-
chological maladjustment. Psychological Science, 21(7), 984–991.
Kuppens, P., Sheeber, L. B., Yap, M. B. H., Whittle, S., Simmons, J., & Allen,
N. B. (2012). Emotional inertia prospectively predicts the onset of depression
in adolescence. Emotion, 12, 283-289.
Lakatos, I. (1976). Falsification and the methodology of scientific research pro-
grammes. In S. Harding (Ed.), Can theories be refuted? (pp. 205–259). Dordrecht:
Springer.

Larsen, R. J. (2000). Toward a science of mood regulation. Psychological Inquiry,

11(3), 129–141.
Lauritzen, S. L. (1996). Graphical Models (Vol. 17). Clarendon Press.
Leroux, B. G. (1992). Consistent estimation of a mixing distribution. The Annals
of Statistics, 1350–1360.
Levina, E., Rothman, A., Zhu, J., et al. (2008). Sparse estimation of large covari-
ance matrices via a nested lasso penalty. The Annals of Applied Statistics, 2(1),
245–263.
Lewandowsky, S., & Farrell, S. (2010). Computational modeling in cognition:
Principles and practice. SAGE publications.
Liu, S., Kuppens, P., & Bringmann, L. F. (2019). On the use of empirical bayes
estimates as measures of individual traits. Assessment, 1073191119885019.
Lodewyckx, T., Tuerlinckx, F., Kuppens, P., Allen, N. B., & Sheeber, L. (2011). A
hierarchical state space approach to affective dynamics. Journal of mathematical
psychology, 55(1), 68–83.
Lunansky, G., van Borkulo, C. D., & Borsboom, D. (2019, May). Personal-
ity, resilience, and psychopathology: A model for the interaction between slow and
fast network processes in the context of mental health. PsyArXiv. Retrieved from
psyarxiv.com/mznbw doi: 10.31234/osf.io/mznbw

227
REFERENCES

Lutz, W., Ehrlich, T., Rubel, J., Hallwachs, N., Röttger, M.-A., Jorasz, C., . . .
Tschitsaz-Stucki, A. (2013). The ups and downs of psychotherapy: Sudden
gains and sudden losses identified with session reports. Psychotherapy Research,
23(1), 14–24.
MacCallum, R. C., Wegener, D. T., Uchino, B. N., & Fabrigar, L. R. (1993). The
problem of equivalent models in applications of covariance structure analysis.
Psychological Bulletin, 114(1), 185–199.
Malinsky, D., & Danks, D. (2018). Causal discovery algorithms: A practical
guide. Philosophy Compass, 13(1), e12470.
Marks, R. J. I. (2012). Introduction to Shannon Sampling and Interpolation Theory.
Springer Science & Business Media.
Marsman, M., Borsboom, D., Kruis, J., Epskamp, S., van Bork, R., Waldorp, L., . . .
Maris, G. (2018). An introduction to network psychometrics: Relating ising net-
work models to item response theory models. Multivariate Behavioral Research,
53(1), 15–35.
Mauger, R., Tarbuck, E. J., & Lutgens, F. K. (1996). Earth: An introduction to
physical geology. Prentice-Hall.
Maxwell, S. E., & Cole, D. A. (2007). Bias in cross-sectional analyses of longitu-
dinal mediation. Psychological Methods, 12(1), 23–44.
McNally, R. J. (2016). Can network analysis transform psychopathology? Be-
haviour Research and Therapy, 86, 95–104.
McNally, R. J., Robinaugh, D. J., Wu, G. W., Wang, L., Deserno, M. K., & Bors-
boom, D. (2015). Mental disorders as causal systems: A network approach to
posttraumatic stress disorder. Clinical Psychological Science, 3(6), 836–849.
Meehl, P. E. (1978). Theoretical risks and tabular asterisks: Sir Karl, sir Ronald,
and the slow progress of soft psychology. Journal of Consulting and Clinical Psy-
chology, 46(4), 806–834.
Meehl, P. E. (1990). Appraising and amending theories: The strategy of
Lakatosian defense and two principles that warrant it. Psychological Inquiry,
1(2), 108–141.
Meier, B. P., & Robinson, M. D. (2004). Why the sunny side is up: Associations
between affect and vertical position. Psychological science, 15(4), 243–247.
Moberly, N. J., & Watkins, E. R. (2008). Ruminative self-focus and negative
affect: An experience sampling study. Journal of Abnormal Psychology, 117, 314-
323.
Molenaar, P. C. (2004). A manifesto on psychology as idiographic science:
Bringing the person back into scientific psychology, this time forever. Measure-
ment, 2(4), 201–218.

228
 REFERENCES

Molenaar, P. C. (2008). Consequences of the ergodic theorems for classical test

theory, factor analysis, and the analysis of developmental processes. Handbook
of Cognitive Aging, 90–104.
Molenaar, P. C., & Campbell, C. G. (2009). The new person-specific paradigm
in psychology. Current directions in psychological science, 18(2), 112–117.
Moler, C., & Van Loan, C. (2003). Nineteen dubious ways to compute the expo-
nential of a matrix, twenty-five years later. SIAM review, 45(1), 3–49.
Mooij, J. M., Janzing, D., Heskes, T., & Schölkopf, B. (2011). On causal discovery
with cyclic additive noise models. In Advances in Neural Information Processing
Systems (pp. 639–647).
Mooij, J. M., Janzing, D., & Schölkopf, B. (2013). From ordinary differential
equations to structural causal models: the deterministic case. arXiv preprint
arXiv:1304.7920.

Murphy, K. P. (2012). Machine Learning: a Probabilistic Perspective. MIT press.

Muthukrishna, M., & Henrich, J. (2019). A problem in theory. Nature Human
Behaviour, 3(3), 221–229.
Nelson, B., McGorry, P. D., Wichers, M., Wigman, J. T., & Hartmann, J. A. (2017).
Moving from static to dynamic models of the onset of mental disorder: a review.
JAMA Psychiatry, 74(5), 528–534.
Newman, M. (2018). Networks. Oxford university press.
Newton, I. (1687). Philosophiæ naturalis principia mathematica (mathematical
principles of natural philosophy). London.

Nguyen, J., & Frigg, R. (2017). Mathematics is not the only language in the book
of nature. Synthese, 1–22.
Opsahl, T., Agneessens, F., & Skvoretz, J. (2010). Node centrality in weighted
networks: Generalizing degree and shortest paths. Social networks, 32(3), 245–
251.
Oravecz, Z., & Tuerlinckx, F. (2011). The linear mixed model and the hierar-
chical ornstein–uhlenbeck model: Some equivalences and differences. British
Journal of Mathematical and Statistical Psychology, 64(1), 134–160.
Oravecz, Z., Tuerlinckx, F., & Vandekerckhove, J. (2009). A hierarchical
Ornstein-Uhlenbeck model for continuous repeated measurement data. Psy-
chometrika, 74, 395–418.
Oravecz, Z., Tuerlinckx, F., & Vandekerckhove, J. (2011). A hierarchical latent
stochastic difference equation model for affective dynamics. Psychological Meth-
ods, 16, 468–490.

229
REFERENCES

Oravecz, Z., Tuerlinckx, F., & Vandekerckhove, J. (2016). Bayesian data analysis
with the bivariate hierarchical ornstein-uhlenbeck process model. Multivariate
behavioral research, 51(1), 106–119.
Ou, L., Hunter, M. D., & Chow, S.-M. (2019). dynr: Dynamic modeling in R
[Computer software manual]. Retrieved from https://CRAN.R-project.org/
package=dynr (R package version 0.1.14-9)
Oud, J. H. (2007). Continuous time modeling of reciprocal relationships in the
cross-lagged panel design. In S. M. Boker & M. J. Wenger (Eds.), Data analytic
techniques for dynamic systems in the social and behavioral sciences (p. 87-129).
Mahwah, NJ: Lawrence Erlbaum Associates.
Oud, J. H., & Delsing, M. J. M. H. (2010). Continuous time modeling of panel
data by means of SEM. In K. van Montfort, J. H. Oud, & A. Satorra (Eds.), Lon-
gitudinal research with latent variables (pp. 201–244). New York, NY: Springer.
Oud, J. H., & Jansen, R. A. (2000). Continuous time state space modeling of
panel data by means of SEM. Psychometrika, 65(2), 199–215.
Oud, J. H., van Leeuwe, J., & Jansen, R. (1993). Kalman filtering in discrete and
continuous time based on longitudinal lisrel models. Advances in longitudinal
and multivariate analysis in the behavioral sciences, ITS, Nijmegen, Netherlands,
3–26.
Oud, J. H., Voelkle, M. C., & Driver, C. C. (2018). Sem based carma time series
modeling for arbitrary n. Multivariate behavioral research, 53(1), 36–56.
Papoulis, A., & Pillai, S. U. (2002). Probability, Random Variables, and Stochastic
Processes. Tata McGraw-Hill Education.
Pe, M. L., Kircanski, K., Thompson, R. J., Bringmann, L. F., Tuerlinckx, F.,
Mestdagh, M., . . . others (2015). Emotion-Network Density in Major De-
pressive Disorder. Clinical Psychological Science, 3(2), 292–300. doi: 10.1177/
2167702614540645
Pearl, J. (2009). Causality. Cambridge university press.
Pearl, J., & Verma, T. (1991). A theory of inferred causation. In J. Allen, R. Fikes,
& E. Sandewall (Eds.), Principles of knowledge representation and reasoning: Pro-
ceedings of the second international conference (pp. 441–452). San Mateo, CA.
Pero, F. (2015). Whither structuralism for scientific representation? (Unpublished
doctoral dissertation). Univeristy of Florence.
Rabiner, L. R. (1989). A tutorial on hidden markov models and selected appli-
cations in speech recognition. Proceedings of the IEEE, 77(2), 257–286.
Raykov, T., & Marcoulides, G. A. (2001). Can there be infinitely many models
equivalent to a given covariance structure model? Structural Equation Modeling,
8(1), 142–149.

230
 REFERENCES

Read, J. (1995). From alchemy to chemistry. Courier Corporation.

Reichardt, C. S. (2011). Commentary: Are three waves of data sufficient for
assessing mediation? Multivariate Behavioral Research, 46(5), 842–851.

Richards, A., French, C. C., Johnson, W., Naparstek, J., & Williams, J. (1992). Ef-
fects of mood manipulation and anxiety on performance of an emotional stroop
task. British Journal of Psychology, 83(4), 479–491.
Richards, A., & Whittaker, T. M. (1990). Effects of anxiety and mood manipu-
lation in autobiographical memory. British Journal of Clinical Psychology, 29(2),
145–153.
Richardson, T. S., & Robins, J. M. (2013). Single world intervention graphs
(swigs): A unification of the counterfactual and graphical approaches to causal-
ity. Center for the Statistics and the Social Sciences, University of Washington Series.
Working Paper, 128(30), 2013.

Roberson-Nay, R., Gorlin, E. I., Beadel, J. R., Cash, T., Vrana, S., & Teachman,
B. A. (2017). Temporal stability of multiple response systems to 7.5% carbon
dioxide challenge. Biological psychology, 124, 111–118.
Robinaugh, D. J., Haslbeck, J. M., Waldorp, L. J., Kossakowski, J. J., Fried, E. I.,
Millner, A. J., . . . Borsboom, D. (2019). Advancing the network theory of men-
tal disorders:a computational model of panic disorder. PsyArXiv. Retrieved from
psyarxiv.com/km37w doi: 10.31234/osf.io/km37w
Robinaugh, D. J., Hoekstra, R. H., Toner, E. R., & Borsboom, D. (2019). The
network approach to psychopathology: a review of the literature 2008–2018
and an agenda for future research. Psychological Medicine, 1–14. doi: 10.1017/
S0033291719003404
Robinaugh, D. J., LeBlanc, N. J., Vuletich, H. A., & McNally, R. J. (2014).
Network analysis of persistent complex bereavement disorder in conjugally be-
reaved adults. Journal of abnormal psychology, 123(3), 510–522.
Robinaugh, D. J., Millner, A. J., & McNally, R. J. (2016). Identifying highly influ-
ential nodes in the complicated grief network. Journal of Abnormal Psychology,
125(6), 747–757.
Robins, J. M. (1999). Association, causation, and marginal structural models.
Synthese, 121(1-2), 151–179.

Robins, J. M. (2003). Semantics of causal dag models and the identification of

direct and indirect effects. In P. Green, N. Hjort, & S. Richardson (Eds.), Highly
structured stochastic systems (p. 70-81). New York, NY: Oxford University Press.
Robins, J. M., & Richardson, T. S. (2010). Alternative graphical causal models
and the identification of direct effects. Causality and psychopathology: Finding
the determinants of disorders and their cures, 103–158.

231
REFERENCES

Rohrer, J. M. (2018). Thinking clearly about correlations and causation: Graph-

ical causal models for observational data. Advances in Methods and Practices in
Psychological Science, 1(1), 27–42.
Rovine, M. J., & Walls, T. A. (2006). Multilevel autoregressive modeling
of interindividual differences in the stability of a process. In T. A. Walls &
J. L. Schafer (Eds.), Models for intensive longitudinal data (pp. 124–147). New
York, NY: Oxford University Press.
Rubel, J. A., Fisher, A. J., Husen, K., & Lutz, W. (2018). Translating
person-specific network models into personalized treatments: Development
and demonstration of the dynamic assessment treatment algorithm for individ-
ual networks (data-in). Psychotherapy and psychosomatics, 87(4), 249–252.
Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and
nonrandomized studies. Journal of educational Psychology, 66(5), 688–701.

Ryan, O., Bringmann, L. F., & Schuurman, N. (2019, Oct). The challenge of
generating causal hypotheses using network models. PsyArXiv. Retrieved from
psyarxiv.com/ryg69 doi: 10.31234/osf.io/ryg69
Ryan, O., & Hamaker, E. L. (2019). Time to intervene: A continuous-time approach
to network analysis and centrality. (Manuscript in Preparation)

Ryan, O., Kuiper, R. M., & Hamaker, E. L. (2018). A continuous time approach to
intensive longitudinal data: What, why and how? In K. L. Montfort, J. H. Oud,
& M. C. Voelkle (Eds.), Continuous time modeling in the behavioral and related
sciences (pp. 29–57). New York: Springer.
Scheffer, M., Bascompte, J., Brock, W. A., Brovkin, V., Carpenter, S. R., Dakos,
V., . . . Sugihara, G. (2009). Early-warning signals for critical transitions. Nature,
461(7260), 53.
Scheffer, M., Bolhuis, J. E., Borsboom, D., Buchman, T. G., Gijzel, S. M., Goulson,
D., . . . others (2018). Quantifying resilience of humans and other animals.
Proceedings of the National Academy of Sciences, 115(47), 11883–11890.

Schmittmann, V. D., Cramer, A. O., Waldorp, L. J., Epskamp, S., Kievit, R. A., &
Borsboom, D. (2013). Deconstructing the construct: A network perspective on
psychological phenomena. New ideas in psychology, 31(1), 43–53.
Schuurman, N. K., Ferrer, E., de Boer-Sonnenschein, M., & Hamaker, E. L.
(2016). How to compare cross-lagged associations in a multilevel autoregres-
sive model. Psychological Methods, 21(2), 206-221.
Schwarz, G., et al. (1978). Estimating the dimension of a model. The Annals of
Statistics, 6(2), 461–464.
Schweder, T. (1970). Composable Markov processes. Journal of Applied Proba-
bility, 7(2), 400–410.

232
 REFERENCES

Scutari, M. (2010). Learning bayesian networks with the bnlearn R package.

Journal of Statistical Software, 35(3), 1–22. doi: 10.18637/jss.v035.i03
Shimizu, S., Hoyer, P. O., Hyvärinen, A., & Kerminen, A. (2006). A linear
non-Gaussian acyclic model for causal discovery. Journal of Machine Learning
Research, 7(Oct), 2003–2030.
Shojaie, A., & Michailidis, G. (2010). Penalized likelihood methods for esti-
mation of sparse high-dimensional directed acyclic graphs. Biometrika, 97(3),
519–538.
Siegle, G. J., Steinhauer, S. R., Carter, C. S., Ramel, W., & Thase, M. E. (2003).
Do the seconds turn into hours? relationships between sustained pupil dilation
in response to emotional information and self-reported rumination. Cognitive
Therapy and Research, 27(3), 365–382.
Siegle, G. J., Steinhauer, S. R., Thase, M. E., Stenger, V. A., & Carter, C. S. (2002).
Can’t shake that feeling: event-related fMRI assessment of sustained amygdala
activity in response to emotional information in depressed individuals. Biologi-
cal Psychiatry, 51(9), 693–707.
Smaldino, P. E. (2017). Models are stupid, and we need more of them. In
R. Vallacher, S. Read, & A. Nowak (Eds.), Computational social psychology (pp.
311–331). New York, NY: Routledge.
Smith, E. R., & Conrey, F. R. (2007). Agent-based modeling: A new approach for
theory building in social psychology. Personality and Social Psychology Review,
11(1), 87–104.
Snippe, E., Viechtbauer, W., Geschwind, N., Klippel, A., De Jonge, P., & Wichers,
M. (2017). The impact of treatments for depression on the dynamic network
structure of mental states: Two randomized controlled trials. Scientific Reports,
7, 46523.
Spector, P. E., & Brannick, M. T. (2011). Methodological urban legends: The
misuse of statistical control variables. Organizational Research Methods, 14(2),
287–305.
Spirtes, P. (1995). Directed cyclic graphical representations of feedback models.
In Proceedings of the eleventh conference on uncertainty in artificial intelligence (pp.
491–498).
Spirtes, P. (2010). Introduction to causal inference. Journal of Machine Learning
Research, 11(May), 1643–1662.
Spirtes, P., Glymour, C. N., Scheines, R., Heckerman, D., Meek, C., Cooper, G.,
& Richardson, T. (2000). Causation, Prediction, and Search. MIT press.
Spirtes, P., Meek, C., & Richardson, T. (1995). Causal inference in the presence
of latent variables and selection bias. In Proceedings of the Eleventh Conference
on Uncertainty in Artificial Intelligence (pp. 499–506).

233
REFERENCES

Spirtes, P., & Zhang, K. (2016). Causal discovery and inference: concepts and
recent methodological advances. Applied Informatics, 3.
Spitzer, R. L., Kroenke, K., & Williams, J. B. (1980). Diagnostic and statistical
manual of mental disorders. Washington, DC: American Psychiatric Association.
Steele, J. S., & Ferrer, E. (2011). Latent differential equation modeling of
self-regulatory and coregulatory affective processes. Multivariate Behavioral Re-
search, 46(6), 956–984.
Steele, R. J., & Raftery, A. E. (2010). Performance of bayesian model selection
criteria for gaussian mixture models. Frontiers of Statistical Decision Making and
Bayesian Analysis, 2, 113–130.
Stiles, W. B., Leach, C., Barkham, M., Lucock, M., Iveson, S., Shapiro, D. A.,
. . . Hardy, G. E. (2003). Early sudden gains in psychotherapy under routine
clinic conditions: Practice-based evidence. Journal of Consulting and Clinical
Psychology, 71(1), 14–21.
Strauss, A., & Corbin, J. (1994). Grounded theory methodology. Handbook of
qualitative research, 17, 273–85.
Strogatz, S. H. (2015). Nonlinear dynamics and chaos: with applications to physics,
biology, chemistry, and engineering. Colorado, USA: Westview press.
Stutz, C., & Williams, B. (1999). Obituary: Ernst Ising. Physics Today, 52, 106–
108.
Suárez, M., & Pero, F. (2019). The representational semantic conception. Philos-
ophy of Science, 86(2), 344–365.
Suls, J., Green, P., & Hillis, S. (1998). Emotional reactivity to everyday prob-
lems, affective inertia, and neuroticism. Personality and Social Psychology Bul-
letin, 24(2), 127–136.
Suppes, P. (1962). Models of data. In E. Nagel, P. Suppes, & A. Tarski (Eds.),
Logic, methodology and the philosophy of science: Proceedings of the 1960 interna-
tional congress. CA, Stanford University Press.
Swoyer, C. (1991). Structural representation and surrogative reasoning. Syn-
these, 87(3), 449–508.
Szollosi, A., & Donkin, C. (2019). Neglected sources of flexibility in psycholog-
ical theories: From replicability to good explanations. Computational Brain &
Behavior, 1–3.
Tomarken, A. J., & Waller, N. G. (2003). Potential problems with “well fitting”
models. Journal of Abnormal Psychology, 112(4), 578–598.
Tong, H., & Lim, K. (1980). Threshold autoregression, limit cycles and cyclical
data. Journal of the Royal Statistical Society: Series B (Methodological), 42(3), 245–
268.

234
 REFERENCES

Tzelgov, J., & Henik, A. (1991). Suppression situations in psychological

research: Definitions, implications, and applications. Psychological Bulletin,
109(3), 524–536.
Uhler, C., Raskutti, G., Bühlmann, P., & Yu, B. (2013). Geometry of the faithful-
ness assumption in causal inference. The Annals of Statistics, 436–463.
van Borkulo, C., Boschloo, L., Borsboom, D., Penninx, B. W., Waldorp, L. J., &
Schoevers, R. A. (2015). Association of symptom network structure with the
course of depression. JAMA Psychiatry, 72(12), 1219–1226.

Van Borkulo, C. D., Borsboom, D., Epskamp, S., Blanken, T. F., Boschloo, L.,
Schoevers, R. A., & Waldorp, L. J. (2014). A new method for constructing net-
works from binary data. Scientific Reports, 4, 5918.
van de Leemput, I. A., Wichers, M., Cramer, A. O., Borsboom, D., Tuerlinckx,
F., Kuppens, P., . . . others (2014). Critical slowing down as early warning for
the onset and termination of depression. Proceedings of the National Academy of
Sciences, 111(1), 87–92.
Van Der Maas, H. L., Dolan, C. V., Grasman, R. P., Wicherts, J. M., Huizenga,
H. M., & Raijmakers, M. E. (2006). A dynamical model of general intelligence:
the positive manifold of intelligence by mutualism. Psychological Review, 113(4),
842–861.

VanderWeele, T. J. (2015). Explanation in causal inference: methods for mediation

and interaction. Oxford University Press.
VanderWeele, T. J., & Robins, J. M. (2007). Directed acyclic graphs, sufficient
causes, and the properties of conditioning on a common effect. American Journal
of Epidemiology, 166(9), 1096–1104.
VanderWeele, T. J., & Tchetgen Tchetgen, E. J. (2017). Mediation analysis with
time varying exposures and mediators. Journal of the Royal Statistical Society:
Series B (Statistical Methodology), 79(3), 917–938.
van Elteren, C., & Quax, R. (2019). The dynamic importance of nodes is poorly
predicted by static topological features. arXiv preprint arXiv:1904.06654.
van Montfort, K., Oud, J. H., & Voelkle, M. C. (2018). Continuous time modeling
in the behavioral and related sciences. Cham: Springer.
Van Orden, K. A., Witte, T. K., Cukrowicz, K. C., Braithwaite, S. R., Selby, E. A.,
& Joiner Jr, T. E. (2010). The interpersonal theory of suicide. Psychological
Review, 117(2), 575.
van Rooijen, G., Isvoranu, A.-M., Meijer, C. J., van Borkulo, C. D., Ruhé, H. G., de
Haan, L., et al. (2017). A symptom network structure of the psychosis spectrum.
Schizophrenia research, 189, 75–83.

235
REFERENCES

Visser, I., & Speekenbrink, M. (2010). depmixS4: An R package for hidden

markov models. Journal of Statistical Software, 36(7), 1–21. Retrieved from
http://www.jstatsoft.org/v36/i07/
Voelkle, M. C., & Oud, J. H. (2013). Continuous time modelling with individu-
ally varying time intervals for oscillating and non-oscillating processes. British
Journal of Mathematical and Statistical Psychology, 66(1), 103–126.
Voelkle, M. C., Oud, J. H., Davidov, E., & Schmidt, P. (2012). An SEM ap-
proach to continuous time modeling of panel data: relating authoritarianism
and anomia. Psychological Methods, 17, 176-192.
Volterra, V. (1931). Variations and fluctuations of the number of individuals
in animal species living together. In R. N. Chapman (Ed.), Animal ecology (pp.
409–448). New York, NY: McGraw-Hill.
Wagenmakers, E.-J., Wetzels, R., Borsboom, D., van der Maas, H. L., & Kievit,
R. A. (2012). An agenda for purely confirmatory research. Perspectives on Psy-
chological Science, 7(6), 632–638.
Walker, M. (2017). Why we sleep: Unlocking the power of sleep and dreams. New
York: Simon and Schuster.
Walters, J. M., & Gardner, H. (1986). The theory of multiple intelligences: Some
issues and answers. Practical intelligence: Nature and origins of competence in the
everyday world, 163–182.
Warren, K. (2002). Thresholds and the abstinence violation effect: A nonlin-
ear dynamical model of the behaviors of intellectually disabled sex offenders.
Journal of Interpersonal Violence, 17(11), 1198–1217.
Watkins, M. W., Lei, P.-W., & Canivez, G. L. (2007). Psychometric intelligence
and achievement: A cross-lagged panel analysis. Intelligence, 35, 59-68.
Wegener, A. (1966). The origin of continents and oceans. New York, NY: Dover
Publications.
Wermuth, N., & Lauritzen, S. L. (1983). Graphical and recursive models for
contingency tables. Biometrika, 537–552.
Wichers, M. (2014). The dynamic nature of depression: a new micro-level
perspective of mental disorder that meets current challenges. Psychological
Medicine, 44(7), 1349–1360.
Wichers, M., Schreuder, M. J., Goekoop, R., & Groen, R. N. (2019). Can we pre-
dict the direction of sudden shifts in symptoms? transdiagnostic implications
from a complex systems perspective on psychopathology. Psychological medicine,
49(3), 380–387.
Wichers, M., Wigman, J., & Myin-Germeys, I. (2015). Micro-level affect dynam-
ics in psychopathology viewed from complex dynamical system theory. Emotion
Review, 7(4), 362–367.

236
 REFERENCES

Wimsatt, W. C. (1987). False models as means to truer theories. Neutral models

in biology, 23–55.
Wolfram Research, Inc. (2019). Mathematica, Version 12.0. (Champaign, IL)

Woodward, J. F. (2011). Data and phenomena: a restatement and defense. Syn-

these, 182(1), 165–179.
Yarkoni, T. (2019, Nov). The generalizability crisis. PsyArXiv. Retrieved from
psyarxiv.com/jqw35 doi: 10.31234/osf.io/jqw35

237
 Nederlandse Samenvatting

Psychologische fenomenen zijn het best te begrijpen als complexe, dynamische

processen. Dit perspectief heeft de laatste jaren enorm aan populariteit gewon-
nen, bijvoorbeeld in het onderzoek naar psychiatrische stoornissen. Onderzoek-
ers in de klinische psychologie bestuderen deze processen momenteel meestal
met simpele statistische modellen die gebruik maken van transversale (cross-
sectional) data of intensieve longitudinale data. Deze modellen zijn slechts deels
– of zelfs helemaal niet – in staat om belangrijke aspecten van de complexe dy-
namiek in kaart te brengen. Dit proefschrift onderzoekt hoe de kaders van dy-
namische systeemtheorie en interventionele causale inferentie gebruikt kunnen wor-
den om deze huidige praktijk te verbeteren.
In hoofdstuk 2 wordt het nut van het gebruik van statistische netwerkmod-
ellen om causale hypotheses te genereren onderzocht. Er wordt beargumenteerd
dat het genereren van hypotheses enorm afhankelijk is van de specificaties van
de ‘beoogde causale structuur’. In dit hoofdstuk wordt aangetoond dat het gener-
eren van hypotheses op basis van netwerkmodellen zelfs in de meest ideale situ-
atie zeer lastig is. Om onderzoekers te helpen bij het exploratief genereren van
hypotheses wordt hier ook een software applicatie geı̈ntroduceerd. Deze appli-
catie wordt verder toegelicht met een empirisch voorbeeld. De discussie van dit
hoofdstuk benadrukt de noodzaak om de rol van tijd in acht te nemen bij het
leren over dynamische processen. Derhalve hebben de overige hoofdstukken
voornamelijk een focus op statistische modellen voor intensieve longitudinale
data.
Hoofdstuk 3 is gericht op de analyse van intensieve longitudinale data met
differentiaalvergelijkingsmodellen. Deze continuous-time (CT) modellen vermi-
jden de praktische en conceptuele beperkingen van gebruikelijke discrete tij-
dreeksmodellen met betrekking tot het tijdsinterval probleem. Dit hoofdstuk
is toegespitst op een hele simpele differentiaalvergelijking, de CT versie van het
VAR(1) model (een veelvoorkomend discreet tijdreeksmodel). Aan de hand van
dit model wordt een brede introductie over de concepten van dynamische sys-
teemtheorie gegeven, en de interpretatie van CT modellen wordt aan de hand
van een empirisch voorbeeld uitgewerkt.
In hoofdstuk 4 wordt de CT benadering van dynamische netwerkanalyse
geı̈ntroduceerd, gebaseerd op het CT-VAR(1) model. Er wordt aangetoond dat
huidige netwerkbenaderingen kunnen leiden tot misleidende conclusies met be-
trekking tot causale relaties tussen variabelen en incorrecte suggesties voor in-
terventies, omdat deze afhankelijk zijn van het geanalyseerde tijdsinterval. Het
CT-VAR(1) model heeft deze afhankelijkheid niet, doordat het een netwerk van
moment-tot-moment relaties schat. Dit resulteert in betere eigenschappen met

239
Nederlandse Samenvatting

betrekking tot conclusies over causale relaties. Hiernaast worden nieuwe cen-
trality statistieken specifiek voor CT netwerken ontwikkeld. Geı̈nspireerd door
de interventionele causale inferentie literatuur, stellen deze statistieken onder-
zoekers in staat om variabelen voor verschillende interventies (acuut of continu)
optimaal te identificeren.
De focus van hoofdstuk 5 is bistabiele systemen, een specifiek type dynamisch
systeem dat veel aandacht heeft gekregen in de psychologie literatuur. De ca-
paciteit van verschillende gangbare statistische modellen om de eigenschappen
van zulke systemen te achterhalen wordt onderzocht, waarbij wordt uitgegaan
van twee scenario’s: ideale data en meer realistische (met langere tijdsintervallen
tussen metingen) data. Met de ideale data wordt aangetoond dat sommige statis-
tische modellen gebruikt kunnen worden om de globale dynamiek (de meer sta-
biele relaties tussen variabelen) te achterhalen maar dat het moeilijk is om dit
correct te doen voor de microdynamiek (moment-tot-moment relaties). Voor de
realistische data geldt dat de globale dynamiek nog steeds gevonden kan worden
maar dat dit helemaal niet mogelijk is voor microdynamiek. Deze resultaten be-
nadrukken a) hoe moeilijk het is om inferenties maken op basis van statistische
modellen zonder een sterke theorie, en b) de fundamentele rol van de frequentie
van de dataverzameling bij het statistisch modelleren van tijdreeksen.
In hoofdstuk 6 wordt betoogd dat formele theorieën van kritiek belang zijn
om onderzoek te doen naar psychiatrische stoornissen. Formele theorieën –
geoperationaliseerd als een set differentiaalvergelijkingen – zijn gebruikelijk in
disciplines die dynamische systeemtheorie toepassen, maar ontbreken bijna al-
tijd in de klinische psychologie. Eerst wordt een kort overzicht gegeven van de
wetenschapsfilosofische literatuur om het belang van formele theorieën te be-
nadrukken. Daarna worden drie manieren onderzocht om formele theorieën
te construeren op basis van statische modellen: a) het gebruik van statistische
modellen als formele theorieën, b) het gebruik van statistische modellen om
formele theorieën van af te leiden, en c) het gebruik van statistische modellen
om bestaande formele theorieën te verbeteren. De derde benadering blijkt het
meeste veelbelovende te zijn maar staat ook het verst af van de huidige praktijk.
Tenslotte wordt een kader voorgesteld dat beschrijft hoe empirisch onderzoek het
beste kan worden ingezet om formele theorieën over psychiatrische stoornissen
te genereren, testen, en verbeteren.
Dit proefschrift eindigt met een duidelijke boodschap voor de klinische psy-
chologie, de psychiatrie en de methodologen die binnen deze disciplines werken:
als we hopen een beter begrip te krijgen over de complexe dynamische processen
die ten grondslag liggen aan psychopathologie, moeten we de huidige onder-
zoekspraktijken met betrekking tot de ontwikkeling van formele theorie radicaal
heroriënteren.

240
 About the Author
Oisı́n Ryan was born on November 23rd 1991 in Kilkenny, Ireland. In 2013
he obtained his BSc. in Psychology from the University of Limerick with first-
class honors. During his bachelor program he worked as a research assistant for
Dr. Timothy D. Ritchie and visited Utrecht University in the Netherlands for
six months as part of an Erasmus program. In September 2013 he returned to
Utrecht University and enrolled in the research masters Methodology and Statis-
tics for the Behavioral, Biomedical and Social Sciences, graduating cum laude in
2015.
Awarded a talent grant from the Netherlands Organization for Scientific Re-
search (NWO), he began his PhD project in September 2015 under the supervi-
sion of Prof. dr. Ellen Hamaker at the Department of Methodology and Statistics
in Utrecht. He has given a variety of invited presentations and workshops to
different groups, and has presented his research at several international confer-
ences, including the International Meeting of the Psychometrics Society (IMPS),
the Association for Psychological Science (APS) annual convention, and the Con-
ference on Complex Systems (CCS). In 2019 he spent one month as a visiting
scholar with Dr. Donald J. Robinaugh at the Department of Psychiatry, Mas-
sachusetts General Hospital, Harvard Medical School, Boston.
As of January 2020, Oisı́n holds a post-doctoral position at Utrecht University,
allowing him to continue his research on dynamical systems modeling in clinical
psychology and psychiatry.

241
Publications & Working Papers

Note: (*) denotes joint first authorship

Ryan, O. & Hamaker, E. L. (under review). Time to intervene: A Continuous-

Time approach to network analysis and centrality.

Ryan, O., Bringmann, L. F., & Schuurman, N. K. (under review). The chal-
lenge of generating causal hypotheses using network models. Pre-print DOI:
10.31234/osf.io/ryg69

Haslbeck, J. M. B.*, Ryan, O.*, Robinaugh, D.*, Waldorp, L. J., & Borsboom,
D. (under review). Modeling psychopathology: From data models to formal
theories. Pre-print DOI: 10.31234/osf.io/jgm7f

Haslbeck, J. M. B.* & Ryan, O.* (under review). Recovering within-person dy-
namics from psychological time series. Pre-print DOI: 10.31234/osf.io/dymhw

Dablander, F.*, Ryan, O.*, & Haslbeck, J. M. B.* (under review). Choosing be-
tween AR(1) and VAR(1) models in typical psychological applications. Pre-print
DOI: 10.31234/osf.io/qgewy

Haslbeck, J. M. B.*, Ryan, O.*, & Dablander, F.* (under review). The sum of
all fears: Comparing networks based on symptom sum-scores. Pre-print DOI:
10.31234/osf.io/3nxu9

Robinaugh, D. J., Haslbeck, J. M. B., Ryan, O., Fried, E. I., & Waldorp, L. J. (under
review) Invisible hands and fine calipers: A call to use formal theory as a toolkit
for theory construction. Perspectives on Psychological Science. Pre-print available
from https://psyarxiv.com/ugz7y.

Groen, R. N., Ryan, O., Wigman, J. T. W., Riese, H., Penninx, B. W. J. H., Giltay,
E. J., Wichers, M., & Hartman, C. A. (in press) Comorbidity between depression
and anxiety: Assessing the role of bridge mental states in dynamic psychological
networks. BMC Medicine.

Bastiaansen, J. A., Kunkels, Y. K., Blaauw, F., Boker, S. M., Ceulemans, E.,
Chen, M., Chow, S. M, de Jonge, P., Emerencia, A. C., Epskamp, S., Fisher, A.
J., Hamaker, E.L., Kuppens, P., Lutz, W., Meyer, M. J., Moulder, R., Oravecz, Z.,
Riese, H., Rubel, J., Ryan, O., Servaas, M. N., Sjobeck, G., Snipper, E., Trull,

243
Publications & Working Papers

T. J., Tschacher, W., van der Veen, D. C., Wichers, M., Wood, P. K., Woods,
W. C., Wright, A. G. C., Albers, C. J. & Bringmann, L. F. (2020). Time to get
personal? The impact of researchers’ choices on the selection of treatment targets
using the experience sampling methodology. Journal of Psychosomatic Research.
https://doi.org/10.1016/j.jpsychores.2020.110211

Kuiper, R. M. & Ryan, O. (2019). Meta-analysis of lagged regression models:

A continuous-time approach. Structural Equation Modeling: A Multidisciplinary
Journal, 27(3), 396–413. DOI: 10.1080/10705511.2019.1652613

Hamaker, E. L. & Ryan, O. (2019). A squared standard error is not a measure of

individual differences. Proceedings of the National Academy of Sciences, 116(14),
6544-6545. DOI: 10.1073/pnas.1818033116

Ryan, O., Kuiper, R. M., & Hamaker, E. L. (2018). A continuous-time approach

to intensive longitudinal data: what, why and how? In K. v. Montfort, J. H.
L. Oud, & M. C. Voelkle (Eds.), Continuous time modeling in the behavioral and
related sciences (pp. 27–54). New York, NY: Springer.

Kuiper, R. M., & Ryan, O. (2018). Drawing conclusions from cross-

lagged relationships: Re-considering the role of the time-interval. Struc-
tural Equation Modeling: A Multidisciplinary Journal, 25(5), 809–823. DOI:
10.1080/10705511.2018.1431046

van de Schoot, R., Winter, S. D., Ryan, O., Zondervan-Zwijnenburg, M., &
Depaoli, S. (2017). A systematic review of Bayesian articles in psychology: The
last 25 years. Psychological Methods, 22(2), 217–239.

Hepper, E. G., Wildschut, T., Sedikides, C., Ritchie, T. D., Yung, Y. F., Hansen, N.,
. . . Ryan, O., & Stephan, E., & Vingerhoets, A . J. J. (2014). Pancultural nostalgia:
prototypical conceptions across cultures. Emotion, 14(4), 733–747.

244
 Acknowledgements

Ellen. Since most people only ever do one PhD, I guess everyone feels that their
experience was unique, whether it was or not. But I’m pretty sure mine was, and
a large part of that was down to you. Thank you for sharing your knowledge, for
your support, your openness, your honesty, your humor and your endless supply
of (in)appropriate idioms. Thanks for giving me plenty of rope, and for stepping
in every now and then to ensure I didn’t hang myself with it. I’m not sure how
this all would have turned out if I didn’t have you to act as my mentor, confidant,
“work mom”, and critic. I’m glad we don’t have to deal with that counterfactual.
Jonas. If I were to make a highlight reel of the last few years of my PhD, you
would feature in a striking amount of it. In Utrecht and Amsterdam, Boston,
Greece and Bordeaux; In our apartments, offices and Airbnbs, spas, bars, and too
many restaurants. Our dinners, debates, discussions and arguments all helped
keep me motivated, focused, and having fun. I’m proud of the work we did to-
gether, and I feel tremendously lucky to have you as a friend. My hope is to enjoy
many more years of friendship and collaboration: May we never tire of ranting at
each other, never agree on how to use semicolons, and never stop surreptitiously
editing each other’s punctuation.
Noémi and Laura, if I were to extend the uncomfortable “work family”
metaphor, you two would be my work big-sisters. Laura, thanks for your will-
ingness to discuss any topic at any time, and do it all with good humor. Noémi,
thanks for your willingness to argue with me about any topic at any time. I
learned a lot from both your empathy and your general intolerance of nonsense.
Don, the month I spent working with you and Jonas in Boston was probably
the best experience I’ve had in academia. Thanks for putting up with us demand-
ing your time, shouting so loudly that your boss installed a white noise machine
outside your office, and for the marathon skype sessions that have followed every
few weeks since we left the US. They have been a true inspiration.
Fabian, I’m very happy that we ended up working with each other so much
in the last couple of years. I hope to write many more lengthy, informative, and
ultimately unnecessary appendices with you. I consider myself a fan of your
work, and I’m delighted that I can count you as a good friend too.
Thanks to everyone at the M&S department, especially Fayette, Thomas, Kees,
Jolien, Anne, Erik-Jan and Ayoub, for all the lunches, coffees and walks around
the uithof. All of these were essential distractions and stress-release valves which
kept me sane in the office. Thanks especially to Erik-Jan and Ayoub in the last
couple of years for sharing the office with me, feeding my caffeine addiction,
dealing with my grumpy behaviour, and engaging me in general tomfoolery when
it was needed most. Thanks to Rebecca for always being willing to help with the

245
Acknowledgements

most obscure mathematical questions that came to mind, and thanks to Gerko
for all the random pieces of advice over the years. Thanks to Flip and Chantal
for the cookies and impromptu Dutch lessons, and to Kevin for putting up with
my constant requests for office equipment. Thanks to Irene Klugkist, who has
somehow conspired to play an important role in many big moments in my life
in recent years: You gave me my first job in the Netherlands, helping me be able
to move here in the first place, you introduced me to Roline for the first time
(although it was not the smoothest first meeting), and now you’ve been a member
of my reading committee. I’m grateful for all of your help.
Thanks to everyone outside of work for distracting me, giving me many great
memories and experiences, and making me feel at home in the past years. Thanks
to Thomas, Millitza and Fayette for all the trips, the marathon board game ses-
sions, and the willingness to engage in all sorts of random children’s pastimes.
Thanks to Joris for all the coffees, visits to the uithof sheep, and your boundless
enthusiasm for new adventures. Thanks to EJ and Lara, Marloes and Vincent, for
all the dinners, games, bottles of wine, late nights and shared recipes. Especially
thanks to all the Kamphuises for making me feel part of the family.
Thanks to all my friends in Ireland, especially Ali, Eoin, Declan, Luke, the
Steves, Kate, Niamh, Cillian, Úna, Zoe and Sally. Thanks for all the meet-ups,
whatsapp calls, facebook messages and occasional reminders of the most em-
barrassing moments of my past. Thanks in particular to Luke for designing the
cover of this book, and Ali for your never-ending enthusiasm for planning shared
holidays and visits to London or Utrecht.
Thanks to all my family for putting up with my too-infrequent visits home
and for ferrying me all over the place when I do arrive back. Thanks to my par-
ents for all the support in moving and setting up a new life in the Netherlands;
Colin, Pia and the kids for having myself and Roline stay for as long as we want
every Christmas; thanks to Niamh for “checking in to see if I’m still alive” on a
frequent basis, and thanks to Paudie for always reminding me how similar we
are, and giving me the peace of mind that everything is taken care of back home.
Finally, thanks to Roline. I have threatened many ways of messing with you
in my acknowledgements, all of which you have successfully vetoed: This para-
graph is longer than one word, it won’t feature a proposal of any kind, and there’s
no need to search for cryptic messages spelled out across the first letter of each
preceding line (turns out that’s much more difficult to pull off than you’d think).
You were there for everything the last few years: The joy, the despair, frustrations
and elations, the bouts of self-doubt, stress, mania and lethargy. You’re in the
top two people who suffered most in my bad moments during this PhD, and I’m
genuinely not sure if I’m number one on that list or you. Through all that you
built us an amazing life together in Utrecht. I promise not to do another PhD, I
promise not to use the doctor title during petty disagreements, and I promise not
to argue with you about what constitutes a “real job” anymore. I couldn’t have
done this without you. You’re my rock, my best friend, and the love of my life.
Thanks.

246