Towards a Generative Model for Emotion Dynamics

Oisín Ryan∗1 , Fabian Dablander∗2,3 , and Jonas Haslbeck∗4,5

1
Department of Data Science and Biostatistics, University Medical Centre Utrecht
4
Institute for Biodiversity and Ecosystem Dynamics, University of Amsterdam
5
Institute for Advanced Study, University of Amsterdam
4
Department of Psychological Methods, University of Amsterdam
5
Department of Clinical Psychological Science, Maastricht University

Abstract
Most theories of emotion suggest that emotions are reactions to situations we encounter in
daily life. Process theories of emotion further specify a feedback loop between our environment,
attention, emotions, and action that clarifies the adaptive nature of emotions. In principle,
such process theories describe how emotions develop in daily life, and consequently, emotion
measures collected from individuals many times a day in studies using the Experience Sampling
Methodology (ESM) should be highly useful in developing these theories. However, current
emotion theories are predominantly verbal theories and therefore do not make clear predictions
about such data. In this paper, we take a first step towards a generative model of emotion
dynamics by formalizing the link between situations and emotions, which provides us with a
basic generative model of emotions in daily life. We show that this incomplete model already
reproduces nine stable phenomena in emotion time series related to (temporal) associations
between emotions and their distributional form. We then discuss how we can draw on existing
(process) theories of emotion to extend our basic model into a complete generative model
of emotion dynamics. Finally, we discuss how generative models of emotion dynamics can
facilitate theory development and advance measurement, study design, and statistical analysis.

1 Introduction
Emotions play a critical role in how we navigate our environment. They track the extent to which
outcomes align with our goals and motivate us to act on our internal and external environments.
Many theories have been proposed for emotions, such as affect-program, appraisal, and construc-
tivist theories (e.g., Barrett, 2009; Cunningham et al., 2013; Ekman & Cordaro, 2011; Ellsworth
& Scherer, 2003; Izard, 2007; Lazarus, 1991; Panksepp, 2004); for a recent review, see Lange et
al. (2020). While these theories offer different conceptualizations of emotion, focus on different
underlying mechanisms, and aim to explain different phenomena, essentially all emotion theories
are consistent with the idea that emotions are responses to situations we encounter in daily life
(Gross, 2015; Keltner & Gross, 1999; Suri & Gross, 2022). Process theories of emotion go further,
suggesting that emotions are an adaptive part of a feedback loop involving situations we encounter,
our perceptions of those situations, the emotions they evoke, and the actions we take based on
our emotions to change the situation we are in (e.g., Barrett et al., 2007; Carver & Scheier, 1990;
Frijda, 2007; Gross, 1998, 2015; Scherer et al., 2001; Tracy, 2014).
In addition to theoretical work, a large body of empirical work on emotion has been built up
in the psychological literature, including experimental (e.g., Lench et al., 2011), neuroscientific
(e.g., Adolphs & Anderson, 2018; Panksepp, 2004), and physiological (e.g., Behnke et al., 2022)
research on emotion. A recent addition to this vast empirical literature are studies using self-report
measures of emotion collected throughout the day using the Experience Sampling Method (ESM;
Csikszentmihalyi et al., 2014), which has gained great popularity since the advent of smartphone
technology. The increasing availability of such time series data led to a rapid growth of the field
of emotion dynamics (Kuppens & Verduyn, 2017). Here, statistical analyses of emotion time
∗ All authors contributed equally to this work.

1
series have provided new insights into empirical phenomena such as how emotions correlate with
themselves over time (Kuppens, Allen, et al., 2010; Kuppens et al., 2022); how they co-vary with
other emotions (Zelenski & Larsen, 2000); and how these patterns differ across individuals (Trampe
et al., 2015). By providing researchers with the ability to directly study how emotions evolve in
daily life, data from ESM studies seem particularly well suited to advancing our understanding of
the adaptive and dynamic nature of emotions in daily life.
However, there is a growing recognition in many areas of psychological science that statistical
models fit to empirical data are often disconnected from theories about psychological processes
(e.g., Borsboom et al., 2021; Fried, 2020; Haslbeck et al., 2021; Proulx & Morey, 2021; van Rooij
& Baggio, 2021). The central problem is that theories are often not described with sufficient
precision to make clear predictions about what we expect to see in empirical data or in statistical
models fitted to it. In the emotion literature, while emotion theories largely agree that emotions
are caused by situations and that the emotion system is generally adaptive, they typically do not
make concrete predictions about how emotions evolve in daily life. For example, existing theories
do not predict the extent to which we would expect emotions to correlate with themselves and
with other emotions on a given timescale, or how this might vary across individuals. At the same
time, the statistical models fitted to the data typically imply dynamics and behavior that are too
simple to serve as plausible substantive theories, suggesting that these statistical models should
be interpreted as descriptive rather than generative. For example, popular autoregressive models
used to analyze emotion time series, while providing useful descriptions of co-occurence patterns
within and between emotions over time, do not directly model the interaction and feedback between
emotions and the environment, the core component of process models of emotion regulation. The
consequence of this disconnect between statistical models and theories is that theory development
has advanced slowly, despite recent rapid progress in data collection and statistical analysis.
Statistical modeling and theorizing can be integrated by formalizing theories about psychologi-
cal processes as computational or generative models, a practice that has begun to gain considerable
traction in recent years (e.g., Epstein, 2008; Haslbeck et al., 2021; Robinaugh et al., 2021; Smaldino,
2017). This means translating the theory into mathematical equations that can be used to simulate
the behavior of the system over time. By formalizing a theory, we make existing theories more
amenable to development using empirical data and statistical analysis: A formalized theory makes
clear predictions about the patterns we expect to see in a particular type of empirical data, pre-
dictions that can then be tested using statistical models (Haslbeck et al., 2021; van Dongen et al.,
2022). In addition to allowing researchers to clearly state their predictions and then test them,
formalized theories have a number of other significant advantages: They require clearly specifying
the relevant variables and their relationships, which can expose gaps in understanding and reduce
conceptual ambiguity; they allow the specification of complex relationships that go beyond what
(linear) statistical models generally assume; they can be tested on different (types of) data and
hence can integrate results across different studies, which facilitates cumulative instead of frag-
mented science; and they can inform the design of studies and help optimizing resources and time
(Epstein, 2008; Guest & Martin, 2020; Haslbeck et al., 2021; Robinaugh et al., 2021; Smaldino,
2017; van Rooij & Baggio, 2021).
In this paper, we take a first step toward developing a generative model of the dynamics of
emotions in daily life. The core feature of our model is the seemingly simple idea that emotions in
daily life are reactions to situations we encounter. Synthesizing the emotion dynamics literature,
we identify nine stable phenomena about how emotions unfold in daily life, and demonstrate that
this simple and incomplete model already reproduces all of these phenomena. We then outline
concrete steps for using ideas from existing emotion theories to extend our model to a complete
feedback loop between situations, attention, emotion and action, and learning mechanisms. These
extensions would make our model an adaptive system (that is, a system that can change its
behavior in response to its environment; Holland, 1992), increasing its potential to explain other
phenomena related to emotion, such as experimental findings or developmental trajectories. In
addition to creating a basic generative model derived from existing emotion theories that explains
a number of key phenomena of emotion dynamics in daily life, we discuss how our model can guide
measurement, study design, and statistical analyses.

2
2 A Basic Generative Model for Emotion Dynamics
In this section, we take a first step toward developing a generative model of emotion dynamics.
A complete version of such a model would capture the basic adaptive function of emotion, en-
compassing situations, attention, emotions, and actions, and thus would potentially be able to
explain many different phenomena related to emotions in daily life. Here, we present a first step
toward such a complete model by creating a basic model that models how emotions are elicited as
responses to situations. We take this first step for three reasons: First, it is the aspect of emotions
on which essentially all theories and models of emotion agree (Gross, 2015; Keltner & Gross, 1999;
Suri & Gross, 2022); second, it is a natural starting point for the feedback loop that must be part
of any complete model of emotion dynamics; and third, as we will show in the following section,
this basic model is already able to reproduce all stable phenomena about emotion dynamics in
daily life that we are aware of.
We begin by briefly reviewing a number of central emotion theories and identifying a feedback
loop between situations, attention, emotions, and actions which is broadly consistent with most
emotion theories (Section 2.1). On the basis of this feedback loop, we create a generative model
by formalizing the Situation → Emotion link (Section 2.2). In Section 3, we show that the incom-
plete model, consisting only of the Situation → Emotion link, already explains some of the core
phenomena which have been established in the emotion dynamics literature, which studies how
emotions evolve and vary over time within individuals as they go about their daily lives (Kuppens
& Verduyn, 2017).

2.1 Brief Review of Theories of Emotion

Emotions have many aspects that have been theorized extensively. One way to organize existing
theories is to distinguish between three categories, depending on how they conceptualize emotions
(Lange et al., 2020): Affect-program theories propose that there are a limited number of basic
emotions that are activated in important situations (e.g., Ekman & Cordaro, 2011; Izard, 2007;
Tracy, 2014); constructivist theories see emotions as a result of people’s tendency to categorize
affective states that can change in response to the environment (Lindquist et al., 2015); and ap-
praisal theories see emotions as playing a central role in how individuals evaluate the world around
them. Here, situations encountered by an individual are evaluated along a number of appraisal
dimensions, such as novelty, valence, and how conducive they are to achieving a psychologically
relevant goal, and different patterns in these appraisal dimensions then elicit different emotions
(Ellsworth & Scherer, 2003; Frijda, 2007). While these theories differ in how they conceptualize
emotions, which part of the emotion system they focus on, and what phenomena they seek to
explain, they all agree on two core features that characterize emotions: First, that emotions un-
fold over time (typically over seconds to minutes; Gross & Feldman Barrett, 2011; Mauss et al.,
2005); and second, that emotions are a response to events or situations that are meaningful to the
individual (Keltner & Gross, 1999), with different patterns of emotions being elicited by different
situations. Notably, this is also a feature that distinguishes emotions from the broader category
of affective states, which are characterized by valence (e.g., positive or negative), motivation (e.g.,
implying a directional tendency such as approach or avoidance), and arousal (i.e., physiological
arousal), but may not necessarily be reactions to situations we encounter (Schiller et al., 2022), and
moods, which are typically longer in duration, may not be attributable to any particular situation,
and which have may not have any specific motivational functions (Schiller et al., 2022).
Within the broad range of different emotion theories, a number of models emphasize the dy-
namic nature of emotions in daily life, focusing on how emotions play an adaptive role in shaping
how we interact with our environment (Bargh, 2013; Carver & Scheier, 1990; Gross, 1998). For
example, cognitive theories of emotion focus on how emotions are generated by comparing the situ-
ation one is in with a desired goal or goal state (Carver & Scheier, 1990; Eldar et al., 2021; Emanuel
& Eldar, 2022; Erdman & Eldar, 2023; Scherer et al., 2001). Models such as the modal model
of emotion and the closely related process models of emotion regulation (Gross, 1998) emphasize
how both cognitive and behavioral strategies, which can be both adaptive and maladaptive, can
be used to regulate emotions. At a broad conceptual level, these types of emotion theories can be
roughly characterized as describing a feedback loop between situations and emotions, as shown in
panel (a) of Figure 1.

3
 (a) Feedback Loop (b) Time Forward Representation

Situation Situation Situation

Attention Attention
Action Attention

Emotion Emotion

Emotion Action Action

Figure 1: Left: The central feedback loop of the emotion system, depicted as a cycle. Right:
Time-forward version of the feedback loop. The black components and relationships are the ones
we will include in our basic generative model presented in Section 2.2.

The feedback loop begins when an individual is in a psychologically relevant situation. A

situation may refer to some feature of the external environment, such as driving to one’s office,
giving a public talk, or encountering a mugger in an alley on the way home from work. Alternatively,
the situation encountered could be conceptualized as the activation of some internal representation,
such as thinking about having to go to work the next day, considering whether to accept a speaking
engagement, or planning how to get home from work later that night. When the individual
perceives this situation, that is, when they focus their attention on this situation, then this situation
produces an emotion, which we understand as the subjective experience of emotion. Appraisal
theorists would consider this emotion to be an appraisal or evaluation of the situation. Noticing
the knife-wielding mugger in front of us and evaluating this as a threat to our physical safety,
the individual may experience a feeling of fear. In turn, the experience of an emotion will lead
the individual to choose an action, such as giving the mugger one’s wallet and jewelry, or turning
around and running away. This action results in a change in the external environment, and the
feedback loop begins again: for example, by fleeing the mugger, we may reduce our level of fear.
This may lead us to choose a new action, and so on and so forth, or it may be that the current
situation does not evoke any new emotions and no emotion regulation is necessary. As such, the
feedback loop could also be represented as a time-forward process, as shown in Figure 1(b). While
Figure 1 represents a huge simplification and a lot of nuance can be added at all stages (see Section
4), we believe that any model of emotion dynamics that captures the adaptive nature of emotion
must be broadly consistent with this feedback loop.
Emotion theories, as we have described them here, typically take the form of verbal rather than
computational or generative models. As a number of researchers in different areas of psychological
science have argued in recent years (e.g., Borsboom et al., 2021; Fried, 2020; Haslbeck et al., 2021;
Robinaugh et al., 2021; Robinaugh et al., 2019; van Rooij & Baggio, 2020), verbal theories in
psychology are often difficult to evaluate and develop because they are not precise enough to make
predictions about patterns we expect to find in empirical data. To make unambiguous predictions
about phenomena, and thus to better connect theoretical and empirical work on emotion, we
need generative models of emotion dynamics. Broadly speaking, the models used in the emotion
literature either (a) focus on aspects of the emotion system that are somewhat removed from the
dynamics of how emotions evolve in daily life, or (b) are generic statistical models that do not
formalize theories of emotion and thus do not represent plausible generative models.
Examples of the first category are the models formulated by Reisenzein (2009) and Suri and
Gross (2022). The primary purpose of these models is to explain the nature, i.e., the cognitive
underpinnings of emotions themselves, rather than to test whether these basic theories of emotions
can explain the patterns of emotion dynamics we find in empirical research. There are also several

4
computational models which use the framework of reinforcement learning to connect learning with
emotion and mood. For example, Eldar et al. (2016), Eldar and Niv (2015), and Bennett et
al. (2022) conceptualized mood valence as a moving average of reward prediction errors, thereby
providing an account for how mood can be adaptive in learning. Rutledge et al. (2014) propose
a different model based on reinforcement learning which suggests that momentary happiness can
be better explained by the recent reward expectations and the discrepancies between expected
and received outcomes (prediction errors), rather than by immediate earnings from a task. These
models provide critical connections between learning, emotions, and constructs evolving at slower
time-scales such as mood. We discuss how these models could be integrated with our model when
discussing how to include learning in our model in Section 4.2.
Computational process models have also been developed in the computer science literature
(Bosse et al., 2010) in the area of affective computing (Broekens et al., 2013; Marsella & Gratch,
2014; Picard, 2000; Scherer et al., 2010). However, in this field the focus is often more on en-
gineering than on science. The goal typically is to develop artificial systems and devices that
can recognize, interpret, process, and simulate human emotions, often with the goal of facilitat-
ing human-computer interactions (c.f., Bosse et al., 2014). It is therefore not straightforward to
test these relatively complex models against existing empirical emotion measurements in humans.
However, we think that integrating the study of human emotion and the creation of artificial
agents that display human-like emotion would be extremely promising, similar to how cognitive
psychology and computer science collaborate with huge success on artificial intelligence (Lake et
al., 2017). However, such a collaboration requires a shared language, which is facilitated with
formalized models (Reisenzein et al., 2013).
Examples of the second category of models are statistical models that capture statistical associ-
ations between different emotions over time, either by using regression models to capture the lagged
correlations in emotion time series (Bringmann et al., 2016; Hamaker et al., 2015; Kuppens, 2015;
Kuppens, Oravecz, et al., 2010), or by modeling the probability that experiencing one emotion will
lead to the activation of other emotions (as in the Affective Ising Model; Loossens et al., 2020).
These statistical models could also be interpreted as generative models (Vanhasbroeck et al., 2021),
representing the hypothesis that emotions directly cause other emotions at later points in time.
Wirth et al. (2022) proposed another interesting statistical model in which they model the impact
of positive and negative events on emotions measured in ESM-data. However, while these models
formalize specific aspects of emotions in daily life and could be used to specify a generative model
(see Section 4), they are themselves not generative models that formalize emotion theories.
In the remainder of this section, we take a first step toward a generative model of emotion
dynamics. Specifically, we will implement the Situation → Emotion link, as well as transitions
between situations that represent changes in an individual’s external environment. This means
that for our initial model of emotion dynamics, we will ignore the gray components and links in
Figure 1(b). We choose these links as the first modeling step both because they are the natural
starting point for the feedback loop, and because they represent one of the few core features of
emotions (that they are responses to situations) on which essentially all emotion theories agree
(Suri & Gross, 2022). In Section 4 we will discuss how to take aspects from different process
theories, such as cognitive theories of emotion or the process model of emotion regulation, to
create a complete model of emotion dynamics that captures all aspects of the feedback loop shown
in Figure 1.

2.2 Specification of Basic Generative Model

We have identified three core theoretical principles of emotion: (a) emotions vary over time, (b)
emotions are reactions to situations, and (c) different types of situations elicit different patterns
of emotion. In this section, we formalize these theoretical principles in a basic model for emotion
generation. This model explains how emotions evolve over time within an individual in response to
changes in that person’s external environment. The model is basic in the sense that it implements
only a small but core aspect of the verbal emotion theories described in the previous section.
We focus primarily on implementing a mechanism by which different types of situations produce
different patterns of emotions, and for now we consider situations to represent features of an
individual’s external environment only. Extensions that would make this model a complete model
for emotion dynamics and emotion regulation are discussed in Section 4.

5
 The three theoretical principles provide us with some of the key ingredients necessary for build-
ing a generative model; for example, our computational model must consist of different situations
S that produce different emotions Y, and both emotions and situations should be indexed by time
t. In order to translate these theoretical principles into a complete generative model, we must
make a number of auxiliary assumptions and modeling choices that are not explicitly stated by
the verbal theories of emotion described above. We encountered one such auxiliary assumption in
the previous section, namely that one’s external environment, i.e., the situations one encounters,
can change over time independently of one’s emotions.
Our generative model is shown schematically in Figure 2. To keep the explanation of our
model simple, we choose five situations S ∈ {s1 , s2 , s3 , s4 , s5 } and six different emotions Y =
{Y1 , Y2 , Y3 , Y4 , Y5 , Y6 }. We further distinguish (Y1 , Y2 , Y3 ) as representing three distinct positively
valenced emotions (e.g., happy, relaxed, and satisfied) and (Y4 , Y5 , Y6 ) as representing three distinct
negatively valenced emotions (e.g., sad, anxious, and angry), a distinction often used in models and
measures of affect and emotion (Russell, 1980; Watson et al., 1988). We chose to treat situations
as distinct “states” of the external environment, such that only one situation is encountered at each
time step, i.e., the variable S can take on five distinct values at time t, denoted si . Emotions are
treated as continuous random variables, so that all six emotion variables Yj take some real value
at each time step. The choice of these numbers is of course somewhat arbitrary, and deciding how
many situations and emotions are good abstractions is an interesting problem beyond the scope of
this paper. However, we will show below that the qualitative behaviors produced by our model,
which are of interest in this paper, do not depend on the specific number of situations and emotions
chosen.
Our basic model has two parts. The first part, shown in Figure 2, formalizes the principle that
emotions are reactions to situations we encounter by imposing that the current situation St causes
the level of currently experienced emotions Yt . This is the core of the model, which represents
a small part of the temporal process model shown in figure 1 (b). In this simple model, there
are no perception or action components, emotions do not directly propagate forward in time, and
emotions do not directly affect the situation at the next time point. The second part of the model
specifies how to transition between situations independent of emotion regulation and is shown in
Figure 2 (b). At each time step, a situation St = si occurs, followed by either the same situation
or a new situation at the next time step. Since we have chosen to treat situations as distinct and
mutually exclusive binary states, we can mathematically model the time evolution of these states
with a Markov model. In a Markov model, the current state depends on the previous state through
a transition matrix P, shown as a network in Figure 2 (b).

6
 (a) (c) Neutral Situation
St -1 St
s1
0 0

Y1,t-1 Y1,t Y1 Y4

Y2 Y3 Y5 Y6

Positive Situations
Y6,t-1 Y6,t s2 s3
+ 0 + 0
(b) α
Y1 Y4 Y1 Y4
s1 Y2 Y3 Y5 Y6 Y2 Y3 Y5 Y6

s3 Negative Situations
s2
s4 s5
β
α α 0 + 0 +

Y1 Y4 Y1 Y4
s4 s5 Y2 Y3 Y5 Y6 Y2 Y3 Y5 Y6

α α

Figure 2: Model illustration. (a) Shows how the situations and emotions evolve over time. Emo-
tions are independent at each time point and across time points conditional on current situation.
(b) Shows the transition matrix between the five situations as a directed graph. (c) Shows the
situations and the emotions they elicit.

The transition matrix P specifies the probability of being in any state si at time t + 1, given
that one is in state sj at time t. It allows us to describe an individual’s environment by specifying
how many (types of) situations there are, which situation typically follows a given situation, and
how long we tend to stay in different situations. We will discuss in Section 4 how this allows
one to specify many abstract characteristics of the environment of a given individual. For now,
however, we want to keep our model as simple as possible, and so we parameterize P with only
two parameters by writing

s1 s2 s3 s4 s5 
s1 α β β β β
s2 
β α β β β 
P = s3 
β β α β β  ,
s4  β β β α β
s5 β β β β α

where α is the probability of remaining in a situation from one time step to the next, and β is
the probability of moving from one type of situation to another, both assumed to be the same
for all situations and transitions for simplicity. To choose appropriate values of α and β, we need
to decide on a timescale on which we want to model situations and emotions. While emotions
typically unfold on a timescale of minutes (Gross & Feldman Barrett, 2011; Houben et al., 2015),
their duration can be highly variable and can even unfold over hours (Verduyn, 2021; Verduyn &
Lavrijsen, 2015). Here we assume that emotionally relevant situations that we encounter in daily
life unfold on an hourly timescale: Most time intervals in daily life that we would call “situations”,
such as having lunch with a colleague, going for a run, enjoying a theater performance, or reading
a paper on emotion regulation, typically last longer than a minute. If we choose a timescale of
hours (with an abstract time step equal to one hour), a value of α = 0.60 would reflect that there
is a 60% chance of being in the same situation from hour to hour. After two hours, there is a
36% (0.60 × 0.60) chance of staying in the same situation, meaning that it is more likely than not
that the situation will change, and so on. Given that α = 0.60, there is a 40% chance of moving

7
to a different situation. Since we assume that the probability of moving to another situation is
the same regardless of what that other situation is, and combined with the fact that we only have
four other situations, this means that the probability of moving to any other situation is 10%, or
β = 0.10. From these parameters, we can derive the parameters corresponding to a timescale of
minutes (see the Appendix A for details).
Finally, Figure 2 (c) shows how the principle that different situations elicit different emotions
is formalized in the model. As a modeling choice, we impose that state s1 is neutral and does
not elicit any emotion. This reflects the fact that individuals do not experience emotions at all
times (e.g., Trampe et al., 2015), which is an auxiliary assumption that does not follow directly
from most emotion theories. Situations s2 and s3 elicit positive emotions, and situations s4 and
s5 elicit negative emotions. We conceptualize the emotions Yj themselves as continuous rather
than discrete. Thus, in the model implementation itself, states “elicit” emotions by determining
the mean of the (Gaussian) distribution from which the emotion is drawn at that time. This is
encoded in the matrix M, which defines the mean of each emotion variable (columns) for each
different situation (rows). For simplicity, we let

Y1 Y2 Y3 Y4 Y5 Y6 
s1 0 0 0 0 0 0
s2 
γ θ λ 0 0 0 
M = s3 
λ γ θ 0 0 0  , (1)
s4  0 0 0 γ θ λ
s5 0 0 0 λ γ θ
where the neutral situation s1 causes all emotions to have a mean of zero, which encodes the aux-
iliary assumption that neutral situations elicit no emotions on average. With only three positive-
valued parameters γ, θ, and λ, we can enforce that states s2 and s3 elicit positive emotions (positive
means of Y1 , Y2 , and Y3 ) and s3 and s4 elicit negative emotions (positive means of Y4 , Y5 , and
Y6 ), but to different degrees. For example, setting γ = 90, θ = 60, λ = 45 could refer to a situation
in which one is surprised with a party (e.g., situation s2 ), which could mainly cause the emotions
Happiness (Y1 ) and Satisfaction (Y2 ), and to a lesser extent the emotion Relaxed (Y3 ). Similarly,
for negative emotions, one could be cut off by a car on the road (e.g. situation s5 ), which could
cause the emotion Anger (Y5 ), but less so the emotions Worry (Y4 ) and Sadness (Y6 ). In our basic
model, we assume that the elicited emotions can be factored as a product of p univariate Gaussian
distributions. In other words, we have

Yi,t | St ∼ N (µi,s , σ 2 ) , (2)

where i = {1, 2, . . . , 5} and Yi,t and µi,s denote a particular emotion and its mean at time point t.
We made the intensity of the emotion stochastic to represent the fact that each of our situations
does not present a specific fixed situation but a type of situation, which elicits slightly different
patterns of emotion intensity depending on the specifics of the situation. To keep things simple,
we assume that the variance σ 2 is the same across emotions and time points. As we did for the
transition matrix parameters above, we can also choose specific values for the parameters γ, θ, λ,
and σ, although in this case the choice is not based on considerations of timescale, but rather of
what scale we consider the emotion variables to be measured on. For instance, choosing γ = 90,
θ = 60, λ = 45 and σ = 16 ensures that most of the emotion values will fall in the range of 0 and
100. An implementation of our model in the R programming language including documentation is
available on the Github repository https://github.com/ryanoisin/GenerativeEmotion.
In summary, we have combined three theoretical principles and a number of auxiliary assump-
tions to create a generative model that implements the Situations → Emotions link of the process
model of emotion regulation. Next, we illustrate the power of such generative models to derive
concrete predictions from empirical data.

2.3 Simulate Data from the Generative Model

We can now use our basic generative model to simulate how situations and emotions evolve together
over time. The left panel of Figure 3 shows the evolution of the five situations (S1-S5; coded in

8
colors at the bottom of the left panel) and one positive (green lines) and one negative (orange
lines) emotion over the period of 20 days, displayed on an hour-to-hour timescale.

Simulated Data Empirical Data

S1 S2 S3 S4 S5 Happy
Anxious
100
Emotion Intensity

75

50

25

0 4 8 12 16 20 0 4 8 12 16 20
Days Days

Figure 3: Left: Simulated data from our simple generative model as explained in the main text.
Right: Empirical data from a single individual collected by Rowland and Wenzel (2020). Note
that we scaled the simulated data slightly to be on the same scale as the empirical data measured
on a 0-100 slider scale.

A simple way to compare the predictions of the generative model with empirical data is to
compare them visually. We do this by plotting the emotions Happy (green) and Sad (orange) over
20 days for a single individual from the dataset of Rowland and Wenzel (2020). For this particular
individual, we see that the predicted and observed time series are similar in some ways, but not
in others. They are similar in that in both cases the individual alternates between high happiness
and low anxiety states and low happiness and high anxiety states. They are different, however, in
that the empirical time series of this particular individual, the co-occurrence of low negative and
low positive affect generated by the neutral state, is very rare. Of course, a proper validation of
our model requires data from many individuals.
Although we believe that evaluating a model’s prediction by visualizing empirical and model-
generated data is a powerful and underused tool, it can be impractical for large numbers of in-
dividuals, and it is neither transparent nor reproducible. To make the comparison between the
model-predicted and the empirical time series more systematic, a general strategy is to choose one
or more summary statistics that capture some important feature of the empirical data (Haslbeck
et al., 2021). For example, we can evaluate means, variances, covariances, autocovariances, or other
measures of variation in time series such as root mean square successive differences (RMSSD). One
could also examine multivariate conditional relationships, for example in lag-0 partial correlations
or the lagged effects in a vector autoregression (VAR) model (Epskamp et al., 2018; Hamilton,
1994). However, it is also possible to go beyond linear models and consider other characteristics
of the data generating distribution such as skewness, nonlinear relationships, and multimodality
(Haslbeck et al., 2022).
Once we have decided which features of the empirical data are of substantive interest, we can
check whether the data generated by the generative model are consistent with the empirical data
by comparing the summary statistics computed from each. If they are very similar, we can take
this as tentative evidence for the adequacy of the generative model. If not, the model needs to
be adjusted, and the way in which the model’s predictions do not match the empirical data can
provide information about the necessary changes to the generative model. In the following section,
we present a more systematic empirical validation of our basic generative model, following exactly
the line of reasoning outlined above.

9
3 Empirical Validation of Basic Generative Model
In the previous section, we introduced a basic generative model of emotion dynamics based on
the process model of emotion regulation. In this section, we systematically examine the empirical
validity of our model. To do this, we identify consistent statistical patterns that are present across a
number of empirical studies of emotion time series. These can be considered empirical phenomena
that our generative model should reproduce and thereby explain.
Ideally, we would evaluate our model using empirical studies that measure both situations and
emotions and their statistical relationships, as this would be the most direct way to validate the
hypothesized Situation → Emotion relationship in our model. However, situation (i.e., context)
information is not measured in many existing empirical studies, and when it is, it is often ana-
lyzed in a variety of different ways (e.g., Castro-Alvarez et al., 2022; Dejonckheere & Mestdagh,
2021; Silk, 2019), making it difficult to distinguish stable empirical phenomena from idiosyncratic
features of a particular dataset. We must therefore resort to a less direct test of our generative
model. Specifically, we will consider only the (temporal) statistical relationships between emotion
measures. This still allows us to test our generative model because our model makes predictions
about these relationships. In order to test our generative model in this way, we must first deter-
mine what stable patterns of this kind exist in the empirical data. In the remainder of this section,
we will briefly review existing studies of emotion dynamics to synthesize stable phenomena about
statistical relationships in VAR models and associated residual (partial) correlations (Section 3.1),
and phenomena about the distributional form of emotion measures (Section 3.2).

3.1 (Temporal) Relations in Emotion Time Series

Over the past two decades, a large number of empirical research papers have examined the variation
of emotions in daily life using self-report data collected through the Experience Sampling Method
(ESM; e.g., Csikszentmihalyi et al., 2014; Houben et al., 2015). In these studies, individuals are
prompted typically 2-10 times (see Haslbeck & Ryan, 2022) at pseudo-random moments throughout
the day for a couple of weeks. At each measurement occasion they are asked to report on their
subjective experience, which typically includes emotions but sometimes also behaviors, cognitive
measures, and the physical and social context in which the participants find themselves (Kuppens
et al., 2022, for a recent overview see). These papers have found that individuals show variation
over time in their self-reported emotions, typically measured several times a day over several days.
The majority of these studies analyzed the (co-)variation of emotions over time using the same type
of statistical model, the first-order linear Vector Autoregressive (VAR) model and its multilevel
extension (e.g., Bringmann et al., 2013; Elovainio et al., 2020; Fisher et al., 2017; Rowland &
Wenzel, 2020; Snippe et al., 2017). The VAR model estimates direct statistical dependencies
between emotions measured on consecutive measurement occasions and their residuals (Asparouhov
et al., 2018; Epskamp et al., 2018). The fact that the same type of statistical model is used in
many different studies allows us to synthesize results across many studies and to identify stable
phenomena that should be explained by our generative model.
In a brief literature review of 32 empirical papers (detailed in Appendix B), we summarized
the results of individual or multilevel VAR models estimated from experience sampling data of
self-reported emotions. In the case of multilevel models, we considered the fixed effects, that is, the
average autoregressive or cross-lagged relationships over all participants in the study (Epskamp
et al., 2018). The synthesis of these published results allowed us to identify the following six
characteristics of empirical VAR models:

Phenomenon 1: All autoregressive effects are positive.

Phenomenon 2: The autoregressive effects tend to be larger in absolute value than cross-
lagged effects.
Phenomenon 3: Within-valence cross-lagged effects tend to be positive.
Phenomenon 4: Between-valence cross-lagged effects tend to be negative.

Phenomenon 5: In the residuals, within-valence relationships tend to be positive.

10
 Phenomenon 6: In the residuals, between-valence relationships tend to be negative.

The results of this literature review are summarized in Table 1. We found that Phenomena 1
and 2 are present in all but one study in which they could be assessed (97% and 96%). Phenomena
3 and 4 are present in most studies (91% and 75%), while Phenomena 5 and 6 are less consistent
but present in the vast majority of studies considered (82% and 80%).

Source p M/Day N Days Cov DS P1 P2 P3 P4 P5 P5

Elovainio et al. (2020) 7 3 315 7 3 2 3 7 3 3/0 3 3
Pannicke et al. (2021) 15 4 84 14 3 1 3 3 3 3 3 3
Meng et al. (2020) 8 5 211 13 3 1 3 3 3 0 3 3
Rowland and Wenzel (2020) 8 6 125 40 7 1 3 3 3 3 3 3
Oreel et al. (2019) 8 9 30 7 3 1 3 3 3 3 3 3
Bringmann et al. (2013) 6 10 129 12 7 1 3 3 3 3 3 3
Hoorelbeke et al. (2019) 5-6 6 85 7 7 1 3 3 3 3 3 3
De Vos et al. (2017) 14 3 54 30 7 2 3 3 ∼ ∼ 7 7
Fisher et al. (2017) 21 4 40 30 3 3 3 3 7 7 7 7
Greene et al. (2020) 14 2 96 1-16 3 1 3 3 3 3
Snippe et al. (2017) 5 10 50-119 6 7 8 3 3 3 3
Hasmi et al. (2017) 6 10 704 6 7 6 3 3 3 3
Klippel et al. (2018) 13 10 654 5-6 3 3 3 3 3 3
Vrijen et al. (2018) 7 3 25 30 7 2 3 3 3 3
Bak et al. (2016) 5 10 1 356 3 1 3 3 3 3
Bringmann et al. (2016) 6 10 95 7 7 1 3 3 3 3
Hasmi et al. (2018) 6 10 54-390 5 7 6 3 3 3 3
Kuranova et al. (2021) 5 10 78-81 6 7 2 3 3 3 3
Wendt et al. (2020) 3 3.7 102 21 3 3 3 3 3 3
Faelens et al. (2021) 6 6 98 14 3 2 3 3 3 3
Bringmann et al. (2016) 10 10 79 14 7 1 3 3 3 3
Pavani et al. (2017) 7 5 78 14 7 1 3 3 0
Shin et al. (2021) 9 9 119 8 3 2 3 3 3
Lutz et al. (2018) 16 4 35 14 3 2 3 3
Lutz et al. (2018) 16 4 23 14 3 2 3 3
Blanke et al. (2021) 2 6-10 70-200 6-12 3 2 3 3
Fried et al. (2020) 14 4 80 14 3 1 3 3
Pe et al. (2015) 11 8 106 7 7 2 ∼ ∼
Jongeneel et al. (2020) 6 10 95 6 3 1 3 3
van Winkel et al. (2017) 3 10 417 3 2 3
Rowland and Wenzel (2020) 6 6 125 40 3 1 3
Bakker et al. (2017) 3 10 133 7-21 3 1 3

Table 1: p = number of variables; M/Day = ESM measurements per day; N = Number of individ-
uals; Days = Number of days of data collection; Cov = indicates whether additional non-emotion
variables are included in the model; DS = Number of datasets/groups; ranges refer to several data
sets / groups; 3 indicates the presence of phenomena, 7 indicates the absence of phenomena; ∼
indicates that the information is in principle available but could not be obtained from the paper;
0 indicates that all relevant parameters are zero and therefore it could not be determined whether
the phenomenon is present.

The top row of Figure 4 shows the parameters of a typical VAR model estimated from emotion
time series that exhibits each of the six phenomena listed above. The parameters shown are
the fixed effects of a multilevel VAR model estimated from publicly available data. The data
consist of measures of three positive emotions (Happy, Relaxed, Satisfied) and three negative
emotions (Sad, Anxious, Angry) measured six times a day for 40 days on 125 individuals in an
ESM study (for details, see Rowland & Wenzel, 2020). The top left panel of Figure 4 shows
the estimated lagged effects matrix, where we can see that the autoregressive effects (self-loops)
are both positive (Phenomenon 1) and larger in absolute value than the estimated cross-lagged
effects (Phenomenon 2). Within-valence lagged effects, i.e., cross-lagged relationships between
emotions of the same valence (e.g., Happyt−1 → Satisfiedt ), are positive in value (Phenomenon 3).
The majority of between-valence effects, i.e., cross-lagged relations linking emotions of opposite
valence (e.g., Happyt−1 → Sadt ), are negative (Phenomenon 4), with the exception of some positive
relations that are small in absolute value (e.g., |Angryt−1 → Satisfiedt | = 0.01). The top right

11
panel of Figure 4 shows the partial correlations between the residuals of the estimated VAR model,
a way of analyzing the residual structure introduced by Epskamp et al. (2018). Here we see that
all variables of the same valence have positive partial correlations (Phenomenon 5), while variables
of opposite valence have mainly negative partial correlations (Phenomenon 6), again with the
exception of a small positive partial correlation between the residuals of Sad and Relaxed.

Lagged Effects Residual Effects

0.26 0.34

−0.06
happy sad happy −0.16 sad
−0.06

−0.01 −0.03
0.09 0.06 0.01 0.17
0.07 0.11 0.2 −0.02
0 0
Empirical

−0.02
−0.06 0.02 −0.07 0.17
0.24 relaxed 0.14 0.11 0.08 0.05 anxious 0.18 relaxed −0.06 anxious
−0.02 −0.03 0.38 −0.1
−0.03
0 −0.02
0.09 0.06 −0.03 0.12
0.07 0.04 0.27 −0.07
−0.03 −0.01

0.01
satisfied angry satisfied −0.14 angry
−0.01

0.21 0.23

0.33 0.11

−0.01
happy sad happy −0.04 sad
−0.01

−0.01 −0.01
0.05 0.17 0.54
Model−Implied

−0.04
0.05 0.17 0.27 −0.08
−0.01 −0.01
−0.02
−0.01 −0.01 −0.02 0.23
0.27 relaxed 0.06 0.06 0.06 0.06 anxious 0.27 relaxed −0.08 anxious
−0.01 −0.01 0.23 −0.08
−0.02
−0.01 −0.01
0.17 0.05 −0.04 0.27
0.17 0.05 0.54 −0.08
−0.01 −0.01

−0.01
satisfied angry satisfied −0.04 angry
−0.01

0.11 0.33

Figure 4: Left: Parameters of the VAR model. Right: Partial correlations between the residual
of the VAR model. Positive parameters are colored blue, negative parameters are colored red;
the width of the edges is a function of the absolute value of the parameters. Top row depicts
parameters estimated from the empirical data of Rowland and Wenzel (2020). Bottom row depicts
parameters implied by our simple generative model.

Now that we have established the phenomena, we can examine whether our generative model in
fact predicts these phenomena. Of course, the model we have outlined above is capable of simulat-
ing single subject time series data, whereas in the empirical literature the lagged effect estimates
are typically fixed effects from a multilevel model. Since the fixed effects can be interpreted as
the average within-person lagged relationships, we will consider the generative model as represent-
ing an average or typical individual’s emotion dynamics (for a discussion of extensions to model
interindividual differences, see Section 4.3).
A very general approach to investigate whether a model can reproduce a phenomenon would
be to simulate a large amount of data (here a long time series) from the model and then repeat
the statistical analysis on these synthetic data (Haslbeck et al., 2021). In this way, we obtain the
population model-implied statistical model. Here we have analytically derived the VAR models
implied by our generative model, which is possible because our model is still relatively simple

12
(see Appendix C for details). This has the advantage that sensitivity analyses of our model can
be run much faster (see below and Appendix D). The lower panels in Figure 4 show this model-
implied VAR model, derived using the parameter settings described in Section 2.2. We see that
all the autocorrelations are positive (Phenomenon 1). We also see that the autocorrelations are
significantly larger than the cross-lagged effects (Phenomenon 2). Furthermore, the within-valence
cross-lagged effects are positive (Phenomenon 3) and the between-valence cross-lagged effects are
negative (Phenomenon 4). Turning to the residual partial correlation network in Figure 4, we see
that within-valence partial correlations are positive (Phenomenon 5) and between-valence partial
correlations are negative (Phenomenon 6).
Our model thus successfully reproduces all six phenomena that we have been able to identify
in published VAR models: it produces patterns of lagged relationships that are highly consistent
with empirical VAR models. One could argue that this is trivial, since every VAR model can be
represented by a Markov Model (we show this analytically in Appendix C). However, it is important
to note that we did not estimate or freely choose the parameters of our generative model (which in
this case happens to be equivalent to a Markov model), but rather chose plausible values that are
consistent with our theoretical principles and clearly articulated auxiliary hypotheses. In addition,
to ensure that our results are robust to arbitrary modeling choices, we investigated the extent to
which the production of these phenomena is specific to the parameters chosen in Section 2.2 using
a sensitivity analysis (Saltelli, 2004). In Appendix D, we show that these phenomena are all still
produced when we vary the duration of situations and the Situation → Emotion effects in a manner
consistent with our theoretical principles; furthermore, these same phenomena are not produced
when parameter settings inconsistent with our theoretical principles are used. This suggests that
the ability of our models to reproduce these empirical phenomena is robust to theory-consistent
variations in parameter choices.

3.2 Distributional Forms of Emotion Time Series

While VAR models provide a good first summary of multivariate continuous time series, they
capture only a small fraction of possible phenomena in time series data. One phenomenon that is
not captured by the linear relationships in a VAR model concerns the within-person distributional
form of emotion variables, that is, the distribution of an emotion variable within a single individual
across all time points or observations for that individual. In a recent re-analysis of seven empirical
datasets, Haslbeck et al. (2022) showed that many emotion variables follow either a uni-modal,
skewed, or multimodal distribution; these distributional forms cannot be captured by a VAR model,
which can only model symmetric univariate (Gaussian) distributions. They also found a high degree
of variability across items and individuals in their average skew and number of modes. While
Haslbeck et al. (2022) limit themselves to univariate analyses, Loossens et al. (2020) note that a
key phenomenon of emotion time series concerns the bivariate distribution of positive and negative
affect. In particular, these variables often exhibit a V- or wedge-shaped bivariate distribution,
with combinations of high positive and high negative emotions relatively rare compared to high-
low, low-high, or low-low combinations (see also Diener & Iran-Nejad, 1986; Schimmack, 2001).
In most situations, this makes intuitive sense: we may be happy or sad to varying degrees, but
we typically do not experience high levels of both happiness and sadness at the same time. Taken
together, we can define the following three empirical phenomena of emotion time series related to
the distributional form of responses

Phenomenon 7: Some emotion variables exhibit a skewed distribution.

Phenomenon 8: Some emotion variables exhibit a multi-modal distribution.
Phenomenon 9: The bivariate distribution of positive and negative emotions is often V-
shaped.

In the left panel of Figure 5 we illustrate these phenomena using three participants from the data
of Rowland and Wenzel (2020). We plot for each participant the sum of their positive and negative
emotions at each time point. All three participants exhibit skewed and bimodal distributions of
negative and positive emotions separately, as shown by the marginal densities. Each participant has
one mode around relatively low negative emotion scores and a second, less frequent mode around

13
high negative emotion scores. For positive emotions, Participant 47 has a bimodal distribution
with one mode endorsing high positive emotion scores, while the other two participants have a
single mode centered around relatively low positive emotion scores. All three participants also
show some variation of a V-shaped bivariate distribution of positive and negative emotions, with
very few observations in the high-positive and high-negative emotion (upper-right) quadrant.

Empirical Model−Implied
100 Participant: 47 100
Participant: 72
Participant: 97
80 80
Negative Emotion

Negative Emotion
60 60

40 40

20 20

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Positive Emotion Positive Emotion

Figure 5: Left: Bivariate densities of positive and negative emotions in three example participants
drawn from the empirical dataset of Rowland and Wenzel (2020). Right: bivariate densities of
positive and negative emotions generated from our generative model. The margins of each panel
show the univariate densities of positive and negative emotions.

In the right panel of Figure 5, we show that the positive and negative emotions simulated by
our computational model successfully reproduce these phenomena.1 First, we see that positive and
negative emotion scores separately are skewed (with most of the density close to zero) and bimodal,
as shown by the marginal densities in the right panel. Second, the model reproduces the V-shaped
bivariate density of positive and negative emotions. Since we know the model generating the data,
we know that these distributions are determined by the different situations and the emotions they
elicit: positive events elicit high positive emotions and low negative emotions, negative events
elicit low positive emotions and high negative emotions, while neutral events elicit low positive
emotions and low negative emotions. Since no situation activates both positive and negative
valenced emotions simultaneously, we end up with almost no data points in the high positive
emotion and high negative emotion quadrants, resulting in the wedge-shaped bivariate density.
Of course, with a single parameter setting, the computational model is not able to reproduce
the between-person heterogeneity we see in the empirical data; instead, we consider the model to
generate emotion time-series on an individual-specific level, that is, for a single individual governed
by those parameters. We discuss extensions of the model to account for inter-individual differences
in Section 4.1 below.
In summary, we have seen that our basic model is able to reproduce a number of phenomena
observed in the empirical literature. These include six phenomena that have been defined with
respect to empirical VAR models, and another three phenomena related to the form of marginal
distributions of emotion variables. While we do not intend to claim that our model is a realistic
model of emotion dynamics, it is noteworthy that even this basic model, which implements only a
single link in the feedback loop of the emotion system, can already reproduce a variety of phenomena
1 Data were simulated using the same parameter settings as described earlier in the paper, where positive and

negative emotions were computed by summing the scores of the three positively and negatively valenced items,
respectively.

14
that have been observed in the empirical emotion dynamics literature. In the next section, we
discuss how one could extend our basic model into a complete model of emotion dynamics. By
implementing more features of the feedback loop, a generative model may eventually be able to
explain additional phenomena related to how individuals adapt and change over longer timescales,
and how individuals differ from each other.

4 Towards a Complete Model of Emotion Dynamics

In the previous section, we have shown that a relatively simple generative model, implementing
only the generation of emotions from changing situations, already explains nine stable phenomena
established in the empirical emotion dynamics literature. However, the simplicity of this model
also means that it is clearly not yet a realistic model of the emotion system: The model does
not implement a number of aspects of the feedback loop shown in Figure 1, such as interactions
from an agent to their environment, the role of attention/cognition, nor any other theoretically
interesting aspects of emotion, such as regulation, schemata, learning and individual differences.
The strength of our model, however, is that its simplicity allows researchers to use it as a basis,
developing their own computational model by formalizing and incorporating those aspects of the
emotion system that they deem most interesting. Once implemented, researchers can use the same
approaches we have demonstrated here (described in more detail by, e.g. Borsboom et al., 2021;
Haslbeck et al., 2021) to further test and develop those models.
In this section, we sketch a number of paths forward for the development of a generative emotion
model, by discussing three key directions in which our model can be extended. In Section 4.1, we
describe in broad strokes how to extend the current model of a single individual to the full feedback
loop shown in Figure 1. This would result in a model of an individual who can use emotions to
select actions based on what is happening in their environment. That is, the individual regulates
emotions. However, the individual will not change its emotion regulation strategies based on their
experiences. We address this issue in Section 4.2, where we discuss how to incorporate mechanisms
for learning into the model, so that the individual can learn from experience which strategies work
in which situations, or even learn entirely new strategies. Finally, in Section 4.3, we turn to the
question of how to incorporate interindividual differences in emotion regulation and how these
might be formally related to outcomes of interest, such as symptoms of mental disorders.

4.1 Completing the Feedback Loop of Emotion Regulation

Here we discuss how to extend the basic model introduced in Section 2.2 to complete the feedback
loop shown in Figure 1. We do this in three steps: first, by unpacking the Situation → Emotion
link, which includes the introduction of attention; second, by introducing actions and their feedback
into situations; and third, by considering direct feedback between emotions. Each extension also
potentially allows for the implementation of different emotion regulation strategies that act on
different components of the process model in different ways.

4.1.1 Unpacking the Link from Situation to Emotion

In the model presented here, situations produce emotions following a relatively simple probabilistic
mechanism, with different types of situations typically leading to similar emotional responses. The
mechanism by which situations lead to emotions could, however, be extended and modelled in a
variety of different ways, reflecting a variety of different emotion theories. Cognitive theories of
emotion emphasize the role of emotions in signaling whether the current situation is in line with
one’s goals (Joffily & Coricelli, 2013; Scherer et al., 2001). For example, someone who is aiming to
impress a new social circle with their strategic prowess may experience negative emotions when they
lose at a board game, while someone who cares more about maintaining a friendly relationship will
experience different emotions in this situation. To account for such differences, we could extend our
basic model by modeling a “goal state” component: At each moment in time, the present situation
is first compared to the agents goal state, and discrepancies between the situation and goal state,
not only the nature of the situation itself, produce an emotional response. The introduction of

15
goals and associated rewards is also crucial for the extension we discuss in Section 4.2, which allows
the individual to learn.
Another cognitive component of emotion that could be implemented in the model is that of an
emotion schema; cognitive structures that represent, for instance, learned associations that guide
perception and appraisal (Segal, 1988). For instance, an individual might have learned that certain
situations are threatening, leading one individual to react with fear or anger to the same external
situation which elicits no or positive emotions in another. Schema may represent an important
intermediary step between the external environment and emotions, and again may more easily
allow implementation of learning and development processes. See for instance, the computational
model of panic disorder developed by Robinaugh et al. (2019) in which arousal schema plays a
central role.
Another type of theory that provides additional information about the Situation → Emotion
link are theories of emotion regulation. For example, in the process model of emotion regulation,
(Gross, 1998) situations have an effect on emotions indirectly through attention or perception.
This means that in order for a situation to produce an emotion, it is necessary for the individual
to perceive or pay attention to the situation, and more specifically to some particular emotion-
inducing aspect of that situation (Gross, 1998). To implement this in our generative model,
one approach would be to first expand the situation component so that situations have multiple
different “aspects” associated with different emotional responses. Second, the model would need to
include a mechanism by which agents choose which aspects to pay attention to, with this choice
directly influencing the emotions experienced. The act of choosing which aspect of a situation to
pay attention to can be considered an emotion regulation strategy, which Gross (1998) refers to as
attentional deployment. For example, when watching a scary movie in a movie theater with your
partner, the movie itself may elicit a high anxiety response. In order to regulate this emotion, you
might choose to direct your attention elsewhere, looking away from the screen or at your partner
instead, or distract yourself by thinking about something else. In the model, we could incorporate
this type of emotion regulation by equipping agents which a parameter which determines their
preference for engaging in attentional deployment; when a potentially emotion-eliciting situation
is present, this parameter would determine whether the individual pays attention to that aspect of
the situation, or directs their attention to more emotionally neutral aspect. By allowing a direct
link from current attention to future attention, the emotional responses can be regulated even as
the situation remains stable over time.
In addition to expanding the situation component and adding an attention component, re-
searchers may also choose to expand the way in which the emotion component of the model is
implemented. While in the current model we chose to model six different emotions labeled as
positive or negative in nature, reflecting the fact that in practice a number of different emotions
are often queried in emotion time series data, researchers could choose to represent emotions in
many different ways. For example, we could adopt the circumplex model of affect, by allowing
discrete situations to produce values of valence and arousal, rather than values of individual emo-
tions (Russell, 2003), or we could adopt different categorical models of emotion states, for instance
using the 27 distinct emotion categories described by (Cowen & Keltner, 2017). In either case, re-
searchers would then need to specify how these dimensions or states of emotion produce responses
to different self-reported emotion variables if the model is to be evaluated with that type of data
(for a discussion of measurement model specification in generative models, see Haslbeck et al.,
2021; Robinaugh et al., 2021).
Another extension of the model would be with respect to the time scale of the effect of events
on emotions. So far, we assumed that an event at time point t causes — in one way or another
— a set of emotions at the same time point t. However, emotions might linger for longer. For
example, having a friend say something hurtful to you in the morning might ruin your entire day.
The model could therefore be extended by allowing (some) events to affect (some) emotions also
several time points later. Such effects could be informed by empirical results like the ones of Wirth
et al. (2022) who modeled the (cumulative) effects of positive and negative events on momentary
well-being. Alternatively, such slower time scale effects could be modeled with concepts such as
mood, which in itself can be modeled as an aggregation of emotions across a certain time window
(e.g. Bennett et al., 2022; Eldar & Niv, 2015; Eldar et al., 2016).
Finally, in the current paper, we consider situations to represent features of the external en-

16
vironment. Researchers might consider expanding the definition of emotion to include “internal”
situations (Gross, 2015): For example, rather than being confronted by a mugger in an alley, one
might be triggered to worry about being mugged in an alley while sitting safely in the theater.
This distinction may be important for implementing different emotion regulation strategies or for
explaining certain types of emotional phenomena. For example, internal representations and the
emotions they evoke may be triggered to some extent by the external environment, but may persist
even after the external environment has changed; safe at home in bed, one individual may no longer
worry about potential muggers, while another may lie awake worrying about personal safety and
the possibility of encountering a mugger in some future scenario. Such extensions would pave the
way to model phenomena around psychological trauma (Van der Kolk, 2003).

4.1.2 Adding the Link from Emotions to Situations via Actions

In the basic generative model presented in Section 2.2, situations cause emotions, but there is no
feedback from emotions to situations. To complete this link, we could consider actions that change
the nature of the situation being experienced; for example, encountering a mugger may elicit fear,
which in turn activates the behavior “flight,” which hopefully changes the situation so that the
mugger is no longer present. These types of situation-altering actions are referred to by Gross
(1998) as the closely related emotion regulation strategy of situation modification and situation
selection, the latter referring to the more proactive selection or avoidance of various anticipated
situations.
Introducing this feedback makes the modeled individual adaptive, at least in the narrow sense
that it can react to certain situations and thereby change the situation and thus change one’s
(emotional) experience. This type of feedback could be implemented in our model by introducing
a direct causal link from emotions to situations, or by explicitly introducing actions. Since the
availability of different actions or strategies is a key feature of emotion regulation, we think it
would be worthwhile to model actions explicitly. This requires specifying a set of actions, the
probabilities of choosing a particular action for a particular (type of) situation, and the effect of
the action on staying in or moving to another situation. Choosing a set of actions that provides a
useful abstraction of the complicated human reality is an interesting problem, similar to choosing
the set of situations and emotions. The probabilities of choosing a particular action in a given
situation can encode how effective emotion regulation is. For example, we can specify strategies
that choose random or even worse-than-random actions. For such models, we would predict more
negative emotions than for individuals who use effective action selection. The effect of actions on
the situation can be modeled by the transition matrix, for example, by temporarily setting the
probability of leaving the situation for a particular other situation to a very high value.

4.1.3 Direct Links between Emotions

Of course, not all emotion regulation occurs by changing the object of attention or by acting on
one’s external environment. Individuals also engage in emotion regulation through metacognitions
that can serve to regulate their emotions without distracting themselves or changing their external
environment. Gross (2015) describes such strategies as cognitive change or reappraisal (Kross
& Ayduk, 2011; Ochsner & Gross, 2005). For example, if someone feels anxious about a work
presentation, they may use reappraisal to reframe the situation and see it as an opportunity to
showcase their work or develop their presentation skills. Reappraisal can take many forms, such
as reappraising the meaning of an emotionally relevant situation or one’s ability to cope with that
situation, or any number of other cognitive strategies, but typically refers to strategies aimed at
down-regulating the intensity of negative emotional experiences (Gross, 2015; Ochsner & Gross,
2005; Ray et al., 2010). These types of emotion regulation strategies could be implemented in our
generative model in a number of different ways. One option would be to implement reappraisal and
cognitive change as “direct” feedback from current to future emotions: For example, if individuals
tend to engage in reappraisal, then negative emotional responses will tend to become less intense
over time, even if the external situation remains constant.
Extending the model in this way to include several different emotion regulation strategies greatly
increases the potential complexity of the model. For example, we could model different emotion
regulation strategies, the extent to which appropriate strategies are selected in a given situation,

17
and how efficiently they are executed. This would allow one to derive many new predictions. For
example, someone with too few emotion regulation strategies, poor strategy selection, or poor
efficiency in executing them would stay in negative situations longer than individuals with good
emotion regulation. This would likely lead to both higher mean negative emotions and higher
autocorrelations of negative emotions, which has been described as emotional inertia. Alternative
emotion regulation strategies with direct feedback from action to emotion might make similar
predictions: being in a situation that causes negative emotions and being able to quickly reappraise
the situation would lead to lower means and autocorrelations in negative emotions.
Of course, expanding the model in any of the directions outlined here may require researchers
to use different types of empirical data and empirical designs to test and develop these extensions.
For instance, extensions which include cognitive components or actions should first be tested to
ensure that they explain the same basic emotion dynamic phenomena that the simpler model ex-
plains, such as patterns of auto- and cross-correlations between emotion measurements in Section
3. It may be the case that new empirical phenomena need to be established to validate the specific
model extension; for instance, researchers could use self-reports of emotional regulation strate-
gies, cognitive appraisals, or actions, in combination with self-reported emotions. Alternatively,
researchers could consider applying other empirical designs, such as those that place individuals
in controlled environments and that observe reactions to those situations (Rapee & Lim, 1992) to
establish what empirical patterns the model should be able to produce.

4.2 Learning Emotion Regulation Strategies

In the model for emotion dynamics outlined in Section 4.1, the individual reacts to different
situations with different emotion regulation strategies that are more or less successful. However,
the individual is unable to learn which strategies are effective in which situations. Such a learning
mechanism can be added to the model using the framework of reinforcement learning (Sutton &
Barto, 2018). This requires defining a reward for different (types of) emotion regulation strategies
in different (types of) situations. This is similar to cognitive theories of emotion, which evaluate
the extent to which situations are consistent with one’s goals (Scherer et al., 2001). This can be
non-trivial because people often optimize on multiple dimensions simultaneously. For example,
I may feel very tired and would prefer not to help my friend move, but I do not want to feel
guilty about letting them down. If the individual is not already using optimal strategy selection,
they will explore different emotion regulation strategies and remember the payoffs they receive
from different options. Over time, they will learn which strategies are most successful (i.e., have
the highest reward) in which situations, and they will predominantly choose successful strategies.
Adding a learning mechanism can extend our model in various directions. For example, Rutledge et
al. (2014) showed that momentary happiness is a function of the recent reward expectations and the
discrepancies between expected and received outcomes. This provides a fine-grained explanation
of how emotions are elicited, while in our current model the elicitation of emotions is not derived
from an adaptive learning process. Another example is the computational model of Bennett et al.
(2022) which conceptualized mood valence as a moving average of reward prediction errors, which
benefits agents in the types of environments that humans most likely find themselves in. Such
models could be used to integrate additional related constructs evolving at a slower time scale,
such as mood.
Adding learning mechanisms to the system greatly increases the complexity of the system and
thus the types of predicted behaviors. For example, one could initialize the system in a state of
complete ignorance about which strategies are most effective in which situations and observe how
quickly the system learns which strategies are most effective. The time it takes to learn certain
strategies and when they are effective could be compared to known developmental trajectories. For
example, Gullone et al. (2010) examined the use of expressive suppression and cognitive reappraisal
strategies (Gross, 1998) in children between the ages of 9 and 15 and found that older children
used less suppression (see also John & Gross, 2004). It may also be possible to model the impact
of the social context, such as one’s family, on the development of emotion regulation strategies
and compare predictions with relevant empirical phenomena (Morris et al., 2007). Other possible
phenomena might capture how personality traits affect the development of emotion regulation
strategies (Barańczuk, 2019; John & Gross, 2004) or how individuals adapt their emotion regulation
strategies to changing environments (Aldao et al., 2015).

18
 Naturally, expanding the model to include learning processes may again prompt researchers to
look outside of ESM data on emotion dynamics for empirical phenomena with which to develop
their model. In the reinforcement learning literature, researchers often use experimental task-
based designs to test and validate their learning models; these tasks allow the researcher to put the
participant in (artificial) situations of their control, with varying degrees of reward and punishment,
and track how participants behave and learn aspects of their environment (e.g. Guitart-Masip et
al., 2012). Alternatively, researchers may consider empirical phenomena derived from longitudinal
studies conducted on longer timescales, such as cohort studies or measurement-burst designs (e.g.
Shanahan et al., 2022; Sliwinski et al., 2009).

4.3 Modeling Interindividual Differences

Up to this point, we have created a computational model for the emotion system of a single
individual. As has been argued by a number of researchers, the fundamental target of analysis
in many areas of psychological science is a process that evolves over time within an individual
(Hamaker, 2012; Molenaar, 2004; Molenaar & Campbell, 2009), and so this is a natural and
perhaps necessary starting point for building a generative model of emotion generation. However,
one of the primary goals of building a such a model may be to ultimately explain how interindividual
differences in emotional experience (e.g., Calkins, 1994; John & Gross, 2007; Ramzan & Amjad,
2017) arise. Indeed, in outlining possible extensions to the model previously, we have already
suggested that interindividual differences may be present in all aspects of this system, such as
cognitive and appraisal components, emotion schema, emotion regulation strategies and learning
processes.
A very general way to introduce interindividual differences into the model is to essentially
create many copies of the model that differ in one or more aspects of the model. While this
may seem daunting, such differences can be summarized, for example, by probability distributions
over parameters, similar to hierarchical models in a statistical context (Snijders & Bosker, 2011).
The model developed so far has many components and relationships, and interindividual differences
seem plausible in almost all of them. For example, by manipulating the transition matrix, we could
place people in different environments with a different composition of positive or negative situations.
This would be particularly interesting from a developmental perspective, e.g., to examine how
certain environments (e.g., very unfavorable ones with many negative events) would affect the
development of emotion regulation strategies (Thompson, 2011). On the timescale of days and
weeks, one could introduce many interindividual differences, for example, in emotion differentiation
(Smidt & Suvak, 2015), the number of available emotion regulation strategies (Bonanno & Burton,
2013), the ability to select a good strategy in the situation at hand (Sheppes & Levin, 2013), and
the ability to implement it successfully (Aldao & Nolen-Hoeksema, 2012).
Modeling interindividual variation in emotion generation and regulation is fundamental to un-
derstanding important individual outcomes and related interindividual differences. Perhaps the
most important example are symptoms of mental disorders: by explicitly modeling interindividual
differences in the emotion system, we can specify exactly which types of mechanisms are dysfunc-
tional and therefore lead to symptoms. Such a model would be extremely useful because it would
provide a precise explanation for the etiology and maintenance of mental disorders (Erdman &
Eldar, 2023). Due to the fact that mental disorders are massively multifactorial and heterogeneous
(Kendler et al., 2011), we believe that it is nearly impossible to integrate the many components of
the system that gives rise to mental disorders with verbal theorizing (Borsboom et al., 2022). One
reason is that verbal theories are inherently imprecise. But even if they were extremely precise, we
typically cannot intuit the behavior of systems with more than a handful of variables interacting
non-linearly at different timescales (Robinaugh et al., 2021). A model that clearly links (interindi-
vidual variation in) a mechanism to (interindividual variation in) symptoms of mental disorders
could also help resolve existing discussions in the literature, such as the idea that more “dense”
interrelationships between symptoms are related to the development and/or maintenance of mental
disorders (e.g., Borsboom, 2017; Schweren et al., 2018; van Borkulo et al., 2015) or generally how
emotion dynamics are related to psychopathology (Wichers et al., 2021). In addition, a good for-
mal model of (dysfunctional) emotion regulation could potentially lead to better treatments, since
understanding the mechanisms underlying the disorder allows us to simulate which treatments are
most effective (for a demonstration of this for panic disorder, see Ryan et al., 2023). The effective-

19
ness of these treatments can then be confirmed in traditional experiments. This may be a more
efficient way to identify promising new treatments to test in randomized controlled trials (RCTs).
A generative model for within-person dynamics that includes interindividual variation also
makes predictions about phenomena observed in cross-sectional data. This, in turn, implies that
cross-sectional data can also be used to develop the model. For example, if we observe certain
patterns of means and correlations among symptoms of major depression, then as long as these
symptoms can be defined based on the components of the model, the theory can be tested using
these cross-sectional data. Haslbeck et al. (2021) has demonstrated this type of theory development
with the computational model of panic disorder by Robinaugh et al. (2019), which incorporates
interindividual differences in individuals’ propensity to interpret bodily sensations as dangerous
or catastrophic. By simulating data from many individuals, they determined the computational
model-implemented Ising model (e.g., Finnemann et al., 2021) of three key symptoms of panic
disorder. Comparing this implied Ising model with its empirical counterpart showed that the
current computational model was unable to reproduce certain patterns of symptom co-occurrence.
This analysis inspired a revision of the model to include an additional component that captures
the tendency of individuals to escape from fear-provoking situations.

5 Discussion
In this paper, we have introduced a basic generative model for emotion dynamics that implements
the link from situations to emotions. At its core, our model formalizes the notion that emotions
evolve over time within an individual primarily in response to the situations they experience.
Although the model we developed is incomplete in that it does not model the full feedback loop
between situations, attention, emotions, and action, we have shown that it already reproduces nine
stable phenomena observed in emotion time series. This suggests that our model may be a good
starting point for working towards a complete generative model for emotion dynamics. To this
end, we discuss how our generative model can be extended in three key directions: completing the
feedback loop involving situations, attention, emotion, and action; introducing learning so that
the modeled individual can adapt to its environment; and introducing interindividual differences,
which would help explain outcomes such as symptoms of mental disorders. We conclude the paper
by discussing how the adoption of a generative modeling approach, based on the model presented
in this paper, can advance both theoretical and empirical emotion research.
A direct consequence of adopting such a generative model approach is that it may lead re-
searchers to consider new study designs. The key mechanism of our basic model is that emotions
are caused by situations. Extensions of the model would also include emotion regulation strategies
in the form of physical actions (e.g., leaving a situation) or metacognitive strategies (e.g., reap-
praising the situation). Thus, the most direct way to test such a model using time series data
from ESM studies would be to measure information about the situations or contexts an individual
experiences, as well as the actions or emotion regulation strategies they engage in. Such data would
allow researchers to analyze the statistical relationships between situations, actions, and emotions.
Another option would be to combine ESM studies with, for example, lab-based study designs in
which the researcher can exert more direct control over the types of emotionally relevant stimuli
that the participant encounters (see for example Dejonckheere & Mestdagh, 2021; Dejonckheere
et al., 2020, for a discussion on a similar topic in the context of the signal-to-noise ratio of emo-
tion studies). Our basic model clearly predicts a strong statistical relationship between (types of)
situations and emotions. If we do not find this empirically, our model would be refuted. On the
other hand, if we do find such relationships, these quantitative results can be used to fine-tune our
model, for example by directly informing the number and type of different situations that would
be appropriate, or by specifying the magnitude and sign of the situation→ emotion effects.
Studying the relationship between situations, actions, and emotions requires adequate statistical
methodology. When emotions are discretized and both situations and contexts are observed in an
ESM setting, we can fit Markov models (e.g., Spedicato, 2017) to assess the statistical relationships
between situations, emotions, and actions over time. When emotions are treated as continuous
variables, they can be modeled jointly with categorical situations and actions using mixed vector
autoregressive (VAR) models (Haslbeck & Waldorp, 2020). While these statistical analyses may
provide the most direct way to test and refine our generative model, analyses based on emotion

20
variables alone can also be highly informative. For example, in Section 3.1 we showed that our
generative model predicts patterns of statistical dependence consistent with VAR models fitted
to emotion time series alone. However, many other models or descriptive statistics can be used
to capture relevant features of the time series. In this context, an interesting class of models
that has not been widely used in the emotion dynamics literature are Hidden Markov Models
(HMMs), which model time series as generated by unobserved latent time-varying states (for a
recent application in the context of emotion time series, see Mildiner Moraga et al., 2023). The
number and nature of latent states found by fitting HMMs to empirical data could potentially be
used to inform the situation part of our generative model, or to fine-tune the modeling of emotion
variables as consisting of distinct states.
Having a generative model also helps us think more carefully about measurement. For example,
our generative model simulates emotions and situations on a minute-to-minute timescale, and as
we did in Section 3, it is possible to derive the implications of our model for the longer hour-to-hour
timescale. This is crucial because otherwise the predictions of the model cannot be meaningfully
tested in empirical data. Careful consideration of the timescale is critical to avoid comparing
apples and oranges, but is easily overlooked when working without a generative model (Haslbeck
& Ryan, 2022). Generative models also help us think about measurement by allowing us to specify
more precisely what we think happens when individuals answer (experience sampling) items. For
example, using the generative model, we might think of a question phrased as “How sad are you
right now?” as measuring the emotion value at the current time. In contrast, the question “How sad
have you been since the last measurement occasion?” would likely be an average of sadness since
the last measurement occasion, possibly weighted to model memory biases (Averell & Heathcote,
2011; Horwitz et al., 2023; Leertouwer et al., 2021). Obviously, these measurement mechanisms are
vast simplifications, but they illustrate how a generative model helps us describe the measurement
process more concretely. A core simplifying assumption we have made throughout the paper is
to consider that ESM studies querying momentary emotional experience in fact capture emotions,
and not other aspects of affective experience, such as moods and affective states (Schiller et al.,
2022) which may not involve reactions to situations, and which evolve on a different time scale to
emotions. We consider this simplification as a necessary starting point to building a generative
model of emotion dynamics: To overcome this simplification, much more work would need to be
done with regards to model extensions specifying how moods and affect states arise, differ from
and interact with emotions, as well as considerable work designing measurement instruments and
research designs to distinguish these from one another.
In addition to being able to clearly test and develop theories using empirical data (Borsboom
et al., 2021; Haslbeck et al., 2021; Robinaugh et al., 2021), formalized models also allow us to
connect emotion theories across different areas of psychological research. For example, the Affec-
tive Ising Model (Loossens et al., 2020) can also be seen as a computational model of emotion
that, in contrast to the model presented here and inspired by network models of psychopathology,
emphasizes the role of direct causal relations between different emotions over time. Future work
could examine whether these models make competing predictions about empirical phenomena and
test which model better explains them. Researchers in computational psychiatry are increasingly
recognizing the importance of incorporating emotion into their cognitive enhancement models of
psychopathology (Bennett et al., 2022; Eldar & Niv, 2015; Huys & Renz, 2017). However, in these
settings, emotion is typically modeled very abstractly, as unidimensional and as a function of pre-
diction error. Such simplifications may hinder the development of theoretical insights into cognitive
and neurological mechanisms of psychopathology. In clinical psychology, there has been a recent
push to formalize theories of mental disorders at the level of symptoms and more general concepts
in cognitive-behavioral therapy (Borsboom et al., 2021; Haslbeck et al., 2021; Robinaugh et al.,
2019). In these settings, individual differences in emotions such as anxiety and emotion regulation
processes such as rumination play a central role in the development and maintenance of mental
disorders. Thus, any developments in the formal modeling of emotion regulation in everyday life
have the potential to greatly advance efforts to create formal models of mental disorders.
We would like to end by reflecting on how we used statistical models in this paper. It is quite
common in the psychological literature to interpret statistical time series models as generative
models. However, statistical models tend to be relatively simple, which has the advantage that
they are easy enough to estimate from the data, but the disadvantage that they often do not show

21
interesting behavior, making them implausible generative models (Haslbeck et al., 2021). A case in
point is the VAR model, which is the simplest possible model of the relationships across time in a
multivariate time series and a useful way to describe a time series. However, its dynamics consist of
a single equilibrium state to which the individual returns after (external) perturbations (Hamilton,
1994; Haslbeck & Ryan, 2022), which is hardly an interesting theory of psychological functioning.
In our paper, instead of treating simple statistical models as generative models, we used them to
establish statistical phenomena that our generative model should produce. For example, in Section
3.1 we considered six phenomena in VAR models and tested whether they were reproduced by our
generative model. This type of modeling, involving substantively plausible generative models that
produce phenomena captured by statistical models, has been suggested in recent calls for formalized
theories in psychology (Borsboom et al., 2021; Haslbeck et al., 2021; Robinaugh et al., 2021), and
is the standard way of modeling in more formalized disciplines such as complex systems science
(e.g., Boccara & Boccara, 2010).
We proposed a first generative model for emotion dynamics that, although incomplete, repro-
duces a nine stable phenomena in emotion time series. We outlined how our model can be extended
to a complete model of emotion regulation that incorporates mechanisms for learning and interindi-
vidual differences. We also discussed how our model suggests new study designs and associated
statistical methodology, and can advance research on measurement in experience sampling studies.
We hope that the generative model we developed can help advance emotion research both by in-
spiring new empirical research and by serving as a basis for more sophisticated generative models
for emotion dynamics.

Acknowledgements. We would like to thank Tessa Blanken, Jens Lange, Ijsbrand Leertouwer,
Jonas Petter, and Don Robinaugh for helpful comments on an earlier version of this manuscript. OR
was supported by ERC Consolidator grant nr.865468. JMBH was supported by the project ‘New
Science of Mental Disorders’ (www.nsmd.eu), supported by the Dutch Research Council and the
Dutch Ministry of Education, Culture and Science (NWO gravitation grant number 024.004.016).

Materials. The generative model presented in this paper can be obtained as an R package
on the Github repository https://github.com/ryanoisin/GenerativeEmotion. Code to reproduce
all simulations, analyses, and results in this paper is available at https://github.com/ryanoisin/
GenerativeEmotionCodeArchive.

References
Adolphs, R., & Anderson, D. J. (2018). The neuroscience of emotion. The neuroscience of emotion.
Princeton University Press.
Aldao, A., & Nolen-Hoeksema, S. (2012). The influence of context on the implementation of adap-
tive emotion regulation strategies. Behaviour research and therapy, 50 (7-8), 493–501.
Aldao, A., Sheppes, G., & Gross, J. J. (2015). Emotion regulation flexibility. Cognitive Therapy
and Research, 39, 263–278.
Asparouhov, T., Hamaker, E. L., & Muthén, B. (2018). Dynamic structural equation models.
Structural Equation Modeling: A Multidisciplinary Journal, 25 (3), 359–388.
Averell, L., & Heathcote, A. (2011). The form of the forgetting curve and the fate of memories.
Journal of mathematical psychology, 55 (1), 25–35.
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.
Bakker, J. M., Goossens, L., Lange, I., Michielse, S., Schruers, K., Lieverse, R., Marcelis, M., van
Amelsvoort, T., van Os, J., Myin-Germeys, I., et al. (2017). Real-life validation of reduced
reward processing in emerging adults with depressive symptoms. Journal of Abnormal
Psychology, 126 (6), 713.
Barańczuk, U. (2019). The five factor model of personality and emotion regulation: A meta-analysis.
Personality and Individual Differences, 139, 217–227.
Bargh, J. A. (2013). Social psychology and the unconscious: The automaticity of higher mental
processes. Psychology Press.

22
Barrett, L. F. (2009). The future of psychology: Connecting mind to brain. Perspectives on psy-
chological science, 4 (4), 326–339.
Barrett, L. F., Mesquita, B., Ochsner, K. N., & Gross, J. J. (2007). The experience of emotion.
Annual review of psychology, 58, 373.
Behnke, M., Kreibig, S. D., Kaczmarek, L. D., Assink, M., & Gross, J. J. (2022). Autonomic nervous
system activity during positive emotions: A meta-analytic review. Emotion Review, 14 (2),
132–160.
Bennett, D., Davidson, G., & Niv, Y. (2022). A model of mood as integrated advantage. Psycho-
logical Review, 129 (3), 513.
Blanke, E. S., Neubauer, A. B., Houben, M., Erbas, Y., & Brose, A. (2021). Why do my thoughts feel
so bad? Getting at the reciprocal effects of rumination and negative affect using dynamic
structural equation modeling. Emotion.
Boccara, N., & Boccara, N. (2010). Modeling complex systems (Vol. 1). Springer.
Bonanno, G. A., & Burton, C. L. (2013). Regulatory flexibility: An individual differences perspec-
tive on coping and emotion regulation. Perspectives on psychological science, 8 (6), 591–
612.
Borsboom, D. (2017). A network theory of mental disorders. World psychiatry, 16 (1), 5–13.
Borsboom, D., Haslbeck, J. M. B., & Robinaugh, D. J. (2022). Systems-based approaches to mental
disorders are the only game in town. World Psychiatry, 21 (3), 420.
Borsboom, D., van der Maas, H. L., Dalege, J., Kievit, R. A., & Haig, B. D. (2021). Theory
construction methodology: A practical framework for building theories in psychology. Per-
spectives on Psychological Science, 16 (4), 756–766.
Bosse, T., Broekens, J., Dias, J., & Van der Zwaan, J. (2014). Emotion modeling: Towards prag-
matic computational models of affective processes (Vol. 8750). Springer.
Bosse, T., Pontier, M., & Treur, J. (2010). A computational model based on gross’ emotion regu-
lation theory. Cognitive systems research, 11 (3), 211–230.
Bringmann, L. F., Pe, M. L., Vissers, N., Ceulemans, E., Borsboom, D., Vanpaemel, W., Tuer-
linckx, F., & 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., Borsboom,
D., & Tuerlinckx, F. (2013). A network approach to psychopathology: New insights into
clinical longitudinal data. PloS One, 8 (4), e60188.
Broekens, J., Bosse, T., & Marsella, S. C. (2013). Challenges in computational modeling of affective
processes. IEEE Transactions on Affective Computing, 4 (3), 242–245.
Calkins, S. D. (1994). Origins and outcomes of individual differences in emotion regulation. Mono-
graphs of the society for research in child development, 53–72.
Carver, C. S., & Scheier, M. F. (1990). Origins and functions of positive and negative affect: A
control-process view. Psychological review, 97 (1), 19.
Castro-Alvarez, S., Tendeiro, J. N., de Jonge, P., Meijer, R. R., & Bringmann, L. F. (2022).
Mixed-effects trait-state-occasion model: Studying the psychometric properties and the
person–situation interactions of psychological dynamics. Structural Equation Modeling: A
Multidisciplinary Journal, 29 (3), 438–451.
Cowen, A. S., & Keltner, D. (2017). Self-report captures 27 distinct categories of emotion bridged
by continuous gradients. Proceedings of the national academy of sciences, 114 (38), E7900–
E7909.
Csikszentmihalyi, M., Csikszentmihalyi, M., & Larson, R. (2014). Validity and reliability of the
experience-sampling method. Flow and the foundations of positive psychology: The col-
lected works of Mihaly Csikszentmihalyi, 35–54.
Cunningham, W. A., Dunfield, K. A., & Stillman, P. E. (2013). Emotional states from affective
dynamics. Emotion Review, 5 (4), 344–355.
De Vos, S., Wardenaar, K. J., Bos, E. H., Wit, E. C., Bouwmans, M. E., & De Jonge, P. (2017).
An investigation of emotion dynamics in major depressive disorder patients and healthy
persons using sparse longitudinal networks. PLoS One, 12 (6), e0178586.
Dejonckheere, E., & Mestdagh, M. (2021). On the signal-to-noise ratio in real-life emotional time
series. Affect dynamics, 131–152.

23
Dejonckheere, E., Mestdagh, M., Kuppens, P., & Tuerlinckx, F. (2020). Reply to: Context matters
for affective chronometry. Nature human behaviour, 4 (7), 690–693.
Diener, E., & Iran-Nejad, A. (1986). The relationship in experience between various types of affect.
Journal of Personality and Social Psychology, 50 (5), 1031.
Ekman, P., & Cordaro, D. (2011). What is meant by calling emotions basic. Emotion Review, 3 (4),
364–370.
Eldar, E., & Niv, Y. (2015). Interaction between emotional state and learning underlies mood
instability. Nature communications, 6 (1), 6149.
Eldar, E., Pessiglione, M., & van Dillen, L. (2021). Positive affect as a computational mechanism.
Current Opinion in Behavioral Sciences, 39, 52–57.
Eldar, E., Rutledge, R. B., Dolan, R. J., & Niv, Y. (2016). Mood as representation of momentum.
Trends in Cognitive Sciences, 20 (1), 15–24.
Ellsworth, P. C., & Scherer, K. R. (2003). Appraisal processes in emotion. Oxford University Press.
Elovainio, M., Kuula, L., Halonen, R., & Pesonen, A.-K. (2020). Dynamic fluctuations of emotional
states in adolescents with delayed sleep phase - A longitudinal network modeling approach.
Journal of Affective Disorders, 276, 467–475.
Emanuel, A., & Eldar, E. (2022). Emotions as computations. Neuroscience & Biobehavioral Re-
views, 104977.
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.
Erdman, A., & Eldar, E. (2023). The computational psychopathology of emotion. Psychopharma-
cology, 1–8.
Faelens, L., Hoorelbeke, K., Soenens, B., Van Gaeveren, K., De Marez, L., De Raedt, R., & Koster,
E. H. (2021). Social media use and well-being: A prospective experience-sampling study.
Computers in Human Behavior, 114, 106510.
Finnemann, A., Borsboom, D., Epskamp, S., & van der Maas, H. L. (2021). The theoretical and
statistical ising model: A practical guide in r. Psych, 3 (04), 593–617.
Fisher, A. J., Reeves, J. W., Lawyer, G., Medaglia, J. D., & Rubel, J. A. (2017). Exploring the
idiographic dynamics of mood and anxiety via network analysis. Journal of Abnormal
Psychology, 126 (8), 1044.
Fried, E. I. (2020). Lack of theory building and testing impedes progress in the factor and network
literature. Psychological Inquiry, 31 (4), 271–288.
Fried, E. I., Papanikolaou, F., & Epskamp, S. (2020). Mental health and social contact during the
COVID-19 pandemic: an ecological momentary assessment study. Clinical Psychological
Science, 21677026211017839.
Frijda, N. H. (2007). What might emotions be? comments on the comments. Social Science Infor-
mation, 46 (3), 433–443.
Greene, T., Gelkopf, M., Fried, E. I., Robinaugh, D. J., & Lapid Pickman, L. (2020). Dynamic
network analysis of negative emotions and DSM-5 posttraumatic stress disorder symptom
clusters during conflict. Journal of Traumatic Stress, 33 (1), 72–83.
Gross, J. J. (1998). The emerging field of emotion regulation: An integrative review. Review of
general psychology, 2 (3), 271–299.
Gross, J. J. (2015). Emotion regulation: Current status and future prospects. Psychological inquiry,
26 (1), 1–26.
Gross, J. J., & Feldman Barrett, L. (2011). Emotion generation and emotion regulation: One or
two depends on your point of view. Emotion review, 3 (1), 8–16.
Guest, O., & Martin, A. E. (2020). How computational modeling can force theory building in
psychological science. https://doi.org/10.31234/osf.io/rybh9
Guitart-Masip, M., Huys, Q. J., Fuentemilla, L., Dayan, P., Duzel, E., & Dolan, R. J. (2012).
Go and no-go learning in reward and punishment: Interactions between affect and effect.
Neuroimage, 62 (1), 154–166.
Gullone, E., Hughes, E. K., King, N. J., & Tonge, B. (2010). The normative development of emotion
regulation strategy use in children and adolescents: A 2-year follow-up study. Journal of
Child Psychology and Psychiatry, 51 (5), 567–574.

24
Hamaker, E., Ceulemans, E., Grasman, R., & Tuerlinckx, F. (2015). Modeling affect dynamics:
State of the art and future challenges. Emotion Review, 7 (4), 316–322.
Hamaker, E. L. (2012). Why researchers should think" within-person": A paradigmatic rationale.
In T. S. C. M. R. Mehl (Ed.), Handbook of research methods for studying daily life (pp. 43–
61). The Guilford Press.
Hamilton, J. D. (1994). Time series analysis (Vol. 2). Princeton, NJ: Princeton University Press.
Haslbeck, J. M. B., Ryan, O., & Dablander, F. (2022). Multimodality and skewness in emotion
time series.
Haslbeck, J. M. B., Ryan, O., Robinaugh, D. J., Waldorp, L. J., & Borsboom, D. (2021). Modeling
psychopathology: From data models to formal theories. Psychological Methods.
Haslbeck, J. M. B., & Waldorp, L. J. (2020). Mgm: Structure estimation for time-varying mixed
graphical models in high-dimensional data. Journal of Statistical Software.
Haslbeck, J. M., & Ryan, O. (2022). Recovering within-person dynamics from psychological time
series. Multivariate Behavioral Research, 57 (5), 735–766.
Hasmi, L., Drukker, M., Guloksuz, S., Menne-Lothmann, C., Decoster, J., van Winkel, R., Col-
lip, D., Delespaul, P., De Hert, M., Derom, C., et al. (2017). Network approach to un-
derstanding emotion dynamics in relation to childhood trauma and genetic liability to
psychopathology: Replication of a prospective experience sampling analysis. Frontiers in
Psychology, 8, 1908.
Hasmi, L., Drukker, M., Guloksuz, S., Viechtbauer, W., Thiery, E., Derom, C., & Van Os, J. (2018).
Genetic and environmental influences on the affective regulation network: A prospective
experience sampling analysis. Frontiers in Psychiatry, 9, 602.
Holland, J. H. (1992). Complex adaptive systems. Daedalus, 121 (1), 17–30.
Hoorelbeke, K., Van den Bergh, N., Wichers, M., & Koster, E. H. (2019). Between vulnerability
and resilience: A network analysis of fluctuations in cognitive risk and protective factors
following remission from depression. Behaviour Research and Therapy, 116, 1–9.
Horwitz, A. G., Zhao, Z., & Sen, S. (2023). Peak-end bias in retrospective recall of depressive
symptoms on the phq-9. Psychological Assessment.
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.
Huys, Q. J., & Renz, D. (2017). A formal valuation framework for emotions and their control.
Biological Psychiatry, 82 (6), 413–420.
Izard, C. E. (2007). Basic emotions, natural kinds, emotion schemas, and a new paradigm. Per-
spectives on psychological science, 2 (3), 260–280.
Joffily, M., & Coricelli, G. (2013). Emotional valence and the free-energy principle. PLoS compu-
tational biology, 9 (6), e1003094.
John, O. P., & Gross, J. J. (2004). Healthy and unhealthy emotion regulation: Personality processes,
individual differences, and life span development. Journal of personality, 72 (6), 1301–1334.
John, O. P., & Gross, J. J. (2007). Individual differences in emotion regulation. Handbook of emotion
regulation, 351–372.
Jongeneel, A., Aalbers, G., Bell, I., Fried, E. I., Delespaul, P., Riper, H., Van Der Gaag, M., & Van
Den Berg, D. (2020). A time-series network approach to auditory verbal hallucinations:
Examining dynamic interactions using experience sampling methodology. Schizophrenia
Research, 215, 148–156.
Keltner, D., & Gross, J. J. (1999). Functional accounts of emotions. Cognition & Emotion, 13 (5),
467–480.
Kendler, K. S., Zachar, P., & Craver, C. (2011). What kinds of things are psychiatric disorders?
Psychological medicine, 41 (6), 1143–1150.
Klippel, A., Viechtbauer, W., Reininghaus, U., Wigman, J., van Borkulo, C., MERGE, Myin-
Germeys, I., & Wichers, M. (2018). The cascade of stress: A network approach to explore
differential dynamics in populations varying in risk for psychosis. Schizophrenia Bulletin,
44 (2), 328–337.
Kross, E., & Ayduk, O. (2011). Making meaning out of negative experiences by self-distancing.
Current directions in psychological science, 20 (3), 187–191.

25
Kuppens, P. (2015). It’s about time: A special section on affect dynamics. Emotion Review, 7 (4),
297–300.
Kuppens, P., Allen, N. B., & Sheeber, L. B. (2010). Emotional inertia and psychological malad-
justment. Psychological science, 21 (7), 984–991.
Kuppens, P., Dejonckheere, E., Kalokerinos, E. K., & Koval, P. (2022). Some recommendations on
the use of daily life methods in affective science. Affective Science, 3 (2), 505–515.
Kuppens, P., Oravecz, Z., & Tuerlinckx, F. (2010). Feelings change: Accounting for individual dif-
ferences in the temporal dynamics of affect. Journal of Personality and Social Psychology,
99 (6), 1042.
Kuppens, P., & Verduyn, P. (2017). Emotion dynamics. Current Opinion in Psychology, 17, 22–26.
Kuranova, A., Wigman, J. T., Menne-Lothmann, C., Decoster, J., van Winkel, R., Delespaul, P.,
Drukker, M., de Hert, M., Derom, C., Thiery, E., et al. (2021). Network dynamics of
momentary affect states and future course of psychopathology in adolescents. PloS One,
16 (3), e0247458.
Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building machines that
learn and think like people. Behavioral and brain sciences, 40, e253.
Lange, J., Dalege, J., Borsboom, D., van Kleef, G. A., & Fischer, A. H. (2020). Toward an in-
tegrative psychometric model of emotions. Perspectives on Psychological Science, 15 (2),
444–468.
Lazarus, R. S. (1991). Emotion and adaptation. Oxford University Press.
Leertouwer, I., Cramer, A. O., Vermunt, J. K., & Schuurman, N. K. (2021). A review of explicit
and implicit assumptions when providing personalized feedback based on self-report ema
data. Frontiers in psychology, 5159.
Lench, H. C., Flores, S. A., & Bench, S. W. (2011). Discrete emotions predict changes in cognition,
judgment, experience, behavior, and physiology: A meta-analysis of experimental emotion
elicitations. Psychological bulletin, 137 (5), 834.
Lindquist, K. A., Satpute, A. B., & Gendron, M. (2015). Does language do more than communicate
emotion? Current Directions in Psychological Science, 24 (2), 99–108.
Loossens, T., Mestdagh, M., Dejonckheere, E., Kuppens, P., Tuerlinckx, F., & Verdonck, S. (2020).
The Affective Ising Model: A computational account of human affect dynamics. PLoS
Computational Biology, 16 (5), e1007860.
Lutz, W., Schwartz, B., Hofmann, S. G., Fisher, A. J., Husen, K., & Rubel, J. A. (2018). Using
network analysis for the prediction of treatment dropout in patients with mood and anxiety
disorders: A methodological proof-of-concept study. Scientific Reports, 8 (1), 1–9.
Marsella, S., & Gratch, J. (2014). Computationally modeling human emotion. Communications of
the ACM, 57 (12), 56–67.
Mauss, I. B., Levenson, R. W., McCarter, L., Wilhelm, F. H., & Gross, J. J. (2005). The tie that
binds? coherence among emotion experience, behavior, and physiology. Emotion, 5 (2),
175.
Meng, J., Wang, X., Wei, D., & Qiu, J. (2020). State loneliness is associated with emotional
hypervigilance in daily life: A network analysis. Personality and Individual Differences,
165, 110154.
Mildiner Moraga, S., Bos, F. M., Doornbos, B., Bruggeman, R., van der Krieke, L., Snippe, E.,
& Aarts, E. (2023). Evidence for severe mood instability in patients with bipolar dis-
order: Applying multilevel hidden markov modelling to intensive longitudinal ecological
momentary assessment data. https://doi.org/10.31234/osf.io/egp82
Molenaar, P. C. (2004). A manifesto on psychology as idiographic science: Bringing the person
back into scientific psychology, this time forever. Measurement, 2 (4), 201–218.
Molenaar, P. C., & Campbell, C. G. (2009). The new person-specific paradigm in psychology.
Current directions in psychological science, 18 (2), 112–117.
Morris, A. S., Silk, J. S., Steinberg, L., Myers, S. S., & Robinson, L. R. (2007). The role of the
family context in the development of emotion regulation. Social development, 16 (2), 361–
388.
Ochsner, K. N., & Gross, J. J. (2005). The cognitive control of emotion. Trends in cognitive sciences,
9 (5), 242–249.

26
Oreel, T. H., Borsboom, D., Epskamp, S., Hartog, I. D., Netjes, J. E., Nieuwkerk, P. T., Henriques,
J. P., Scherer-Rath, M., van Laarhoven, H. W., & Sprangers, M. A. (2019). The dynamics in
health-related quality of life of patients with stable coronary artery disease were revealed:
A network analysis. Journal of Clinical Epidemiology, 107, 116–123.
Panksepp, J. (2004). Affective neuroscience: The foundations of human and animal emotions.
Oxford university press.
Pannicke, B., Kaiser, T., Reichenberger, J., & Blechert, J. (2021). Networks of stress, affect and
eating behaviour: Anticipated stress coping predicts goal-congruent eating in young adults.
International Journal of Behavioral Nutrition and Physical Activity, 18 (1), 1–14.
Pavani, J.-.-B., Le Vigouroux, S., Kop, J.-.-L., Congard, A., Dauvier, B., & Denissen, J. (2017).
A Network Approach to Affect Regulation Dynamics and Personality Trait–Induced Vari-
ations: Extraversion and Neuroticism Moderate Reciprocal Influences between Affect and
Affect Regulation Strategies. European Journal of Personality, 31 (4), 329–346.
Pe, M. L., Kircanski, K., Thompson, R. J., Bringmann, L. F., Tuerlinckx, F., Mestdagh, M., Mata,
J., Jaeggi, S. M., Buschkuehl, M., Jonides, J., et al. (2015). Emotion-network density in
major depressive disorder. Clinical Psychological Science, 3 (2), 292–300.
Picard, R. W. (2000). Affective computing. MIT press.
Proulx, T., & Morey, R. D. (2021). Beyond statistical ritual: Theory in psychological science.
Perspectives on Psychological Science, 16 (4), 671–681.
Ramzan, N., & Amjad, N. (2017). Cross cultural variation in emotion regulation: A systematic
review. Annals of King Edward Medical University, 23 (1).
Rapee, R. M., & Lim, L. (1992). Discrepancy between self-and observer ratings of performance in
social phobics. Journal of abnormal psychology, 101 (4), 728.
Ray, R. D., McRae, K., Ochsner, K. N., & Gross, J. J. (2010). Cognitive reappraisal of negative
affect: Converging evidence from emg and self-report. Emotion, 10 (4), 587.
Reisenzein, R. (2009). Emotions as metarepresentational states of mind: Naturalizing the belief–
desire theory of emotion. Cognitive Systems Research, 10 (1), 6–20.
Reisenzein, R., Hudlicka, E., Dastani, M., Gratch, J., Hindriks, K., Lorini, E., & Meyer, J.-J. C.
(2013). Computational modeling of emotion: Toward improving the inter-and intradisci-
plinary exchange. IEEE Transactions on Affective Computing, 4 (3), 246–266.
Robinaugh, D. J., Haslbeck, J. M. B., Ryan, O., Fried, E. I., & Waldorp, L. J. (2021). Invisible
hands and fine calipers: A call to use formal theory as a toolkit for theory construction.
Perspectives on Psychological Science, 16 (4), 725–743.
Robinaugh, D. J., Haslbeck, J. M. B., Waldorp, L., Kossakowski, J., Fried, E. I., Millner, A.,
McNally, R. J., van Nes, E. H., Scheffer, M., Kendler, K. S., et al. (2019). Advancing the
network theory of mental disorders: A computational model of panic disorder.
Rowland, Z., & Wenzel, M. (2020). Mindfulness and Affect-Network Density: Does Mindfulness
Facilitate Disengagement from Affective Experiences in Daily Life? Mindfulness, 1–14.
Russell, J. A. (1980). A circumplex model of affect. Journal of personality and social psychology,
39 (6), 1161.
Russell, J. A. (2003). Core affect and the psychological construction of emotion. Psychological
review, 110 (1), 145.
Rutledge, R. B., Skandali, N., Dayan, P., & Dolan, R. J. (2014). A computational and neural model
of momentary subjective well-being. Proceedings of the National Academy of Sciences,
111 (33), 12252–12257.
Ryan, O., Haslbeck, J., & Robinaugh, D. (2023). Improving treatments for mental disorders using
computational models.
Saltelli, A. (2004). Global sensitivity analysis: An introduction. Proc. 4th International Conference
on Sensitivity Analysis of Model Output (SAMO’04), 27, 43.
Scherer, K. R., Bänziger, T., & Roesch, E. (2010). A blueprint for affective computing: A sourcebook
and manual. Oxford university press.
Scherer, K. R., Schorr, A., & Johnstone, T. (2001). Appraisal processes in emotion: Theory, meth-
ods, research. Oxford University Press.
Schiller, D., Alessandra, N., Alia-Klein, N., Becker, S., Cromwell, H. C., Dolcos, F., Eslinger, P. J.,
Frewen, P., Kemp, A. H., Pace-Schott, E. F., et al. (2022). The human affectome.

27
Schimmack, U. (2001). Pleasure, displeasure, and mixed feelings: Are semantic opposites mutually
exclusive? Cognition & Emotion, 15 (1), 81–97.
Schweren, L., Van Borkulo, C. D., Fried, E., & Goodyer, I. M. (2018). Assessment of symptom
network density as a prognostic marker of treatment response in adolescent depression.
JAMA psychiatry, 75 (1), 98–100.
Segal, Z. V. (1988). Appraisal of the self-schema construct in cognitive models of depression.
Psychological bulletin, 103 (2), 147.
Shanahan, L., Steinhoff, A., Bechtiger, L., Murray, A. L., Nivette, A., Hepp, U., Ribeaud, D.,
& Eisner, M. (2022). Emotional distress in young adults during the covid-19 pandemic:
Evidence of risk and resilience from a longitudinal cohort study. Psychological medicine,
52 (5), 824–833.
Sheppes, G., & Levin, Z. (2013). Emotion regulation choice: Selecting between cognitive regulation
strategies to control emotion.
Shin, K. E., Newman, M. G., & Jacobson, N. C. (2021). Emotion network density is a potential
clinical marker for anxiety and depression: Comparison of ecological momentary assessment
and daily diary. British Journal of Clinical Psychology.
Silk, J. S. (2019). Context and dynamics: The new frontier for developmental research on emotion
regulation. Developmental psychology, 55 (9), 2009.
Sliwinski, M. J., Almeida, D. M., Smyth, J., & Stawski, R. S. (2009). Intraindividual change and
variability in daily stress processes: Findings from two measurement-burst diary studies.
Psychology and aging, 24 (4), 828.
Smaldino, P. E. (2017). Models are stupid, and we need more of them. Computational social
psychology (pp. 311–331). Routledge.
Smidt, K. E., & Suvak, M. K. (2015). A brief, but nuanced, review of emotional granularity and
emotion differentiation research. Current Opinion in Psychology, 3, 48–51.
Snijders, T. A., & Bosker, R. J. (2011). Multilevel analysis: An introduction to basic and advanced
multilevel modeling. sage.
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 (1), 1–10.
Spedicato, G. A. (2017). Discrete Time Markov Chains with R. The R Journal, 9 (2), 84–104.
https://doi.org/10.32614/RJ-2017-036
Suri, G., & Gross, J. J. (2022). What is an emotion? a connectionist perspective. Emotion Review,
17540739221082203.
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction. MIT press.
Thompson, R. A. (2011). Emotion and emotion regulation: Two sides of the developing coin.
Emotion Review, 3 (1), 53–61.
Tracy, J. L. (2014). An evolutionary approach to understanding distinct emotions. Emotion Review,
6 (4), 308–312.
Trampe, D., Quoidbach, J., & Taquet, M. (2015). Emotions in everyday life. PloS one, 10 (12),
e0145450.
Van der Kolk, B. A. (2003). Psychological trauma. American Psychiatric Pub.
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 Dongen, N., van Bork, R., Finnemann, A., van der Maas, H., Robinaugh, D., Haslbeck, J. M. B.,
de Ron, J., & Borsboom, D. (2022). Improving psychological explanations.
Vanhasbroeck, N., Ariens, S., Tuerlinckx, F., & Loossens, T. (2021). Computational models for
affect dynamics. Affect dynamics (pp. 213–260). Springer.
van Rooij, I., & Baggio, G. (2020). Theory before the test: How to build high-verisimilitude ex-
planatory theories in psychological science. https://doi.org/10.31234/osf.io/7qbpr
van Rooij, I., & Baggio, G. (2021). Theory before the test: How to build high-verisimilitude ex-
planatory theories in psychological science. Perspectives on Psychological Science, 16 (4),
682–697.

28
van Winkel, M., Wichers, M., Collip, D., Jacobs, N., Derom, C., Thiery, E., Myin-Germeys, I., &
Peeters, F. (2017). Unraveling the role of loneliness in depression: The relationship between
daily life experience and behavior. Psychiatry, 80 (2), 104–117.
Verduyn, P. (2021). Emotion duration. Affect Dynamics, 3–18.
Verduyn, P., & Lavrijsen, S. (2015). Which emotions last longest and why: The role of event
importance and rumination. Motivation and Emotion, 39 (1), 119–127.
Vrijen, C., Hartman, C. A., Van Roekel, E., De Jonge, P., & Oldehinkel, A. J. (2018). Spread
the joy: How high and low bias for happy facial emotions translate into different daily life
affect dynamics. Complexity, 2018.
Watson, D., Clark, L. A., & Tellegen, A. (1988). Development and validation of brief measures of
positive and negative affect: The panas scales. Journal of personality and social psychology,
54 (6), 1063.
Wendt, L. P., Wright, A. G., Pilkonis, P. A., Woods, W. C., Denissen, J. J., Kühnel, A., & Zim-
mermann, J. (2020). Indicators of affect dynamics: Structure, reliability, and personality
correlates. European Journal of Personality, 34 (6), 1060–1072.
Wichers, M., Wigman, J., & Myin-Germeys, I. (2015). Micro-level affect dynamics in psychopathol-
ogy viewed from complex dynamical system theory. Emotion Review, 7 (4), 362–367.
Wichers, M., Riese, H., Hodges, T. M., Snippe, E., & Bos, F. M. (2021). A narrative review of
network studies in depression: What different methodological approaches tell us about
depression. Frontiers in psychiatry, 12, 719490.
Wirth, M., Voss, A., Wirth, S., & Rothermund, K. (2022). Affect dynamics and well-being: Ex-
planatory power of the model of intraindividual variability in affect (miva). Cognition and
Emotion, 36 (2), 188–210.
Zelenski, J. M., & Larsen, R. J. (2000). The distribution of basic emotions in everyday life: A state
and trait perspective from experience sampling data. Journal of Research in Personality,
34 (2), 178–197.

A Timescale of our Generative Model

Our model does not have an inherent timescale because it does not refer to specific processes in
the real world that have a well-defined timescale. For example, if our model were specifically about
how anger evolves, and if we had good time series data on anger, then we could call a variable
in our model anger and adjust its parameter values so that it reproduces the timescale of anger
in the real world. Lacking this “grounding,” our model is more abstract. Still, we want to have
a timescale that makes sense for emotions in daily life. The parameter that defines the timescale
for our model is α (and correspondingly, β), which is the probability that one remains in the same
situation at t + 1 as one was in t. Assuming that one time step represents one hour, the left panel
of Figure 6 shows the probability of staying in the same situation for α = 0.30 (red), α = 0.60
(blue), and α = 0.90 (green). After four hours, the probability of staying in the same state is a
mere 0.81% (0.304 ) for α = 0.30, 12.96% (0.604 ) for α = 0.60, and a substantial 65.61% (0.904 ) for
α = 0.90.

29
 Staying probability (hours) Staying probability (minutes)
1.0
α = 0.30 α = 0.9801
α = 0.60 α = 0.9915
0.8 α = 0.90 α = 0.9982
Probability

0.6

0.4

0.2

0.0

0 2 4 6 8 10 12 0 120 240 360 480 600 720

Hours Minutes

Figure 6: Left: Probability of staying in the same state on a timescale of hours. Right: Probability
of staying in the same state on a timescale of minutes.

If a time step is one minute and one wants to preserve the staying probabilities defined on a
timescale of hours, one needs to increase the value of α accordingly. The left panel of Figure 6
60
illustrates this. Note that the transformation is αHours = αMinutes , which makes sense since one
hour is sixty minutes. We believe that it is easier to reason about emotions in daily life on the
timescale of hours, not least because values for α on this timescale are easier to distinguish than
values on a timescale of minutes. For the main text, we choose a value of α = 0.60. Note that the
Appendix D shows that the exact value of α does not matter for our empirical results, since our
model reproduces the empirical phenomena for a wide range of values of α.

B Literature Review Details

To establish the generality of the six emotion time series phenomena presented in Section 3.1,
we conducted a literature review. We identified relevant work by considering the papers cited by
three seminal papers: Bringmann et al. (2013) is one of the earliest papers to use multilevel VAR
models to analyze multivariate (p > 6) emotion time series, and perhaps the first to present VAR
parameters as a network structure; Wichers et al. (2015) is an important review article linking
moment-to-moment affect dynamics to psychopathology in the context of complex dynamical sys-
tems; Epskamp et al. (2018) introduced a popular method for estimating multilevel VAR models.
Together, these papers were cited 926 times. From these papers, we identified 32 papers that met
the following criteria: (1) Experience Sampling Study with at least two measurements per day; (2)
at least positive or negative emotions were measured; and (3) the authors analyzed the emotion
time series with a VAR model. We considered both single subject models and the fixed effects of
multilevel VAR models, and did not distinguish between the method or software used to estimate
these models, such as the mlVAR package (Epskamp et al., 2018) or Dynamic Structural Equation
Modeling in Mplus (Asparouhov et al., 2018).
For each of the 32 papers, we recorded the number of participants, the measurements per day,
the number of measured days, the number of variables, and whether variables in addition to the
emotion variables were included in the model. We coded each phenomenon as present or absent.
In most cases, this decision was clear: for example, in Vrijen et al. (2018) all cross-lagged effects
within valence are positive and all cross-lagged effects between valence are negative (Phenomenon
4). One of the less clear cases was the first model reported in Klippel et al. (2018): Here there are
four strong negative cross-lagged between-valence effects and three weaker between-valence effects.
In such cases, we code the phenomenon as present if the majority of (unweighted) parameters are
consistent with the phenomenon.
The results of the literature review are shown in Table 1. 3 and 7 indicate that the phenomenon
is present or absent; 0 means that the presence of the phenomenon cannot be determined because

30
all relevant parameters are zero; and ∼ means that the reporting of parameters does not allow
to clearly determine whether the phenomenon is present (e.g., only a network is plotted without
reporting the actual parameter estimates). A missing entry means that the analysis was not
performed in the respective paper. For example, only about a third of the papers computed the
partial correlations between residuals, which is why there are fewer data available for phenomena
5 and 6. Note that the reviewed studies, and thus our results, consist of a mixture of multilevel
fixed-effects networks and single-subject networks.
Phenomenon 1 (positive autoregression) is present in 31 of 32 studies and could not be deter-
mined in the remaining study due to insufficient reporting. Phenomenon 2 (larger autoregressive
than cross-lagged effects) is present in 26 of 27 studies, and could not be detected in the remaining
study due to insufficient reporting. Cross-lagged effects within valence were predominantly positive
(Phenomenon 3) in 21 out of 23 studies, the opposite was true in one study, and in one study report-
ing was insufficient to decide. Cross-lagged effects between valences were predominantly negative
(Phenomenon 4) in 15 of 20 studies, in two studies all cross-lagged effects were zero, in one study
the phenomenon was present in one data set but not in the other, and in one data set the report-
ing was insufficient to decide. In the remaining partial correlation networks, relationships within
valence were predominantly positive ( Phenomenon 5) in 9 out of 11 studies; the opposite was true
in two studies. Finally, relations between valences were predominantly negative (Phenomenon 6)
in 8 of 10 studies; the opposite was true in two studies. In summary, Phenomenon 1 is confirmed
by all studies, and Phenomena 2, 3, and 4 are present in almost all studies. Phenomena 5 and 6
are less consistent, but still present in the vast majority of studies considered. Although there are
some inconsistencies, we can conclude that all phenomena are robust phenomena of emotion time
series.

C Analytic Derivation of Model-implied VAR Models

In the main text, we showed that the data simulated by our basic model exhibit patterns of
autoregressive and cross-lagged effects between the observed variables that are consistent with what
we see in empirical VAR(1) models. In this appendix, we derive how our basic model reproduces
these patterns as a function of the mixture components M and the transition probability matrix
P.

C.1 Model Parameterization

Let P represent an s × s matrix of transition probabilities with Pij representing the probability
of switching from state i at time t to state j at time t + 1. We can define π as a 1 × s row vector
representing the long-run probabilities of being in each state, that is, the so-called stationary state
probabilities (unconditional/marginal probabilities). These are defined by the expression π = πP
and so can be expressed in closed form as a function of P , using the system of equations
(P − I)T π T = 0 ,
Ps
where I is the identity matrix and we have that i πi = 1. Our basic model is equivalent to a
Markov model where the observed variables are assumed to follow a Gaussian distribution, that is,
Yt | St = s ∼ N (µs , Θs ) ,
where µs and Θs represent the conditional mean and variance covariance matrix of the observed
variables in state s. To simplify matters, let’s assume that the variance covariance matrix is
invariant across states (Θs = Θ ∀s). We can parameterize the mapping of states onto Gaussian
emissions by introducing dummy indicator variables Ii,t at each time point, where Ii,t = 1 if St = i
and Ii,t = 0 otherwise. This allows us to represent the relationship between states St and observed
variables Yt through the standard SEM measurement-model expression
Yt = ΛIt + ζt ,
where ζt ∼ N (0, Θ), represents the exogenous shock terms, independent across time, and where Λ
is a p × s matrix of factor loadings with λij representing the mean of Xi,t in state Sj,t . That is,
we have reformulated the Gaussian emission part of the Markov model as a standard confirmatory
factor analysis measurement model at each time point.

31
C.2 Model-Implied Covariance Matrix
Re-formulating the our Markov model in this way allows us to make use of standard SEM expres-
sions for how model matrices produce model-implied variance and covariances between observed
variables. Let Σ = var(Yt ), the so-called lag-0 or same-time-point variance-covariance matrix of
our observed variables. The SEM literature tells us that the variance-covariance matrix can be
expressed as
Σ = ΛΨΛT + Θ ,
where Ψ is the variance-covariance matrix of the latent variables, in this case the s indicator
variables It .
To get an expression for Ψ, we note that the expected values of each indicator variable are
determined by the stationary probabilities of the state to which that indicator belongs, that is,
E[Ii,t ] = πi . This implies that the variance of each indicator variable can be expressed as

var(Ii,t ) = πi (1 − πi ) .

The covariance between any two indicator variables at the same time point t can be expressed as

cov(Ii,t , Ij,t ) = E[Ii,t Ij,t ] − E[Ii,t ]E[Ij,t ] .

Since the indicator variables are by definition mutually exclusive, P (Ii,t = 1, Ij,t = 1) = 0 and so
the first term on the right hand side vanishes. This leaves us with

cov(Ii,t , Ij,t ) = −πi πj ,

that is, the negative product of the stationary probabilities of states i and j.
In matrix form we can express these variances and covariances as Ψ = Dπ − ππ t , where Dπ is
a diagonal matrix with entries Dii = πi . This leaves us with an expression for the lag-0 covariance
matrix
Σ = Λ(Dπ − ππ t )ΛT + Θ .
This expression for the lag-0 covariance matrix is necessary to compute the lag-1 autoregressive
and cross-lagged relationships. Before we can do that however, we need to find an expression for
the covariance of our observed variables across different time points.

C.3 Model-Implied Autocovariance matrix

A similar approach can be taken to derive the lagged relationships between Yt and Yt+1 . The
simplest representation of these lagged relationships is in the so-called autocovariance matrix
Γ = cov(Yt , Yt+1 ) where γij represents the covariance between Xi,t and Xj,t+1 . From a SEM
perspective, following path tracing rules, we know that the expression for Γ will involve the matrix
of factor loadings and the covariances of the indicator variables across time

Γ = Λ cov(It , It+1 ) ΛT .

To find an expression for cov(It , It+1 ), we note that

cov(Ii,t , Ij,t+1 ) = E[Ii,t Ij,t+1 ] − E[Ii,t ]E[Ij,t+1 ] .

We can evaluate the joint expectation E[Ii,t Ij,t+1 ] using

XX
E[Ii,t Ij,t+1 ] = Ii,t Ij,t+1 P (Ii,t , Ij,t+1 ) ,
Ii Ij

and since both indicator variables are binary, this reduces to

E[Ii,t Ij,t+1 ] = P (Ii,t = 1, Ij,t+1 = 1)

= P (Ij,t+1 = 1 | Ii,t = 1)P (Ii,t = 1)
= P (St+1 = j | St = i)P (St = i)
= pij πi .

32
Recalling that E[Ii,t ] = πi , this allows us to express the covariance between any two lagged indicator
variables as
cov(Ii,t , Ij,t+1 ) = pij πi − πi πj .
In matrix form, we can express this covariance as cov(It , It+1 ) = Dπ P − ππ t . This leaves us with
an expression for the lag-1 autocovariance matrix
Γ = Λ(Dπ P − ππ t )ΛT .
With this expression in place we can now derive simple expressions for the autoregressive and
cross-lagged relationships.

C.4 Lagged Relationships and Residual Covariances

The VAR(1) model is defined as
Yt = ΦYt−1 + et ,
where Φ represents a p × p matrix of autoregressive and cross-lagged regression coefficients, and
et ∼ N (0, Ω) represents a p × 1 vector of Gaussian residuals with variance-covariance matrix Ω.
If data is generated by our basic model, then both Φ and Ω will be determined by the expressions
for the model-implied lag-0 and lag-1 covariance matrices derived above. Specifically, we know from
the time series literature (e.g., Hamilton, 1994) that the matrix of lagged regression coefficients
can be found using the standard expression
Φ = ΓT Σ−1 ,
where Γ is transposed to reflect the fact that, unlike the transition matrix P, the lagged coefficients
Φ are read “from column to row”, with φij representing the lagged relationship between Yj,t to
Yi,t+1 .
Furthermore, the residual covariance matrix Ω can also be found by re-arranging the standard
time series expression(Hamilton, 1994)
Ω = Σ − ΦΣΦT .
Typically in network analysis settings the partial correlations of the residuals are visualized as a
residual network. The partial correlations can be found by taking the inverse of this matrix Ω−1
and re-scaling appropriately.
Taken together, these expressions allow us to analytically derive all VAR(1) model parameters
implied by our basic model described in the main text. These expressions are difficult to interpret
on an element-by-element basis, however they do allow us to perform sensitivity analysis in a
straightforward way.

D Sensitivity Analysis
In the main text, we outlined how a particular configuration of our basic model reproduces VAR
model phenomena (lagged parameters and residual partial correlations) found in empirical data.
Here, we perform a sensitivity analysis to assess how robust our findings are to changes in our
model specification. To do this, we vary one part of our model specification at a time: the tran-
sition matrix; the means of the mixture components; and the (residual) variances of the mixture
components. Unless otherwise specified, we vary a particular matrix while keeping all other pa-
rameter values fixed to the values presented in the main text. We explore the parameter space by
sampling new parameter values for each matrix in turn, and assessing whether the model-implied-
VAR model in each parameter setting reproduces the six phenomena observed in empirical data.
This analysis utilizes the analytic derivations presented in Appendix C.
Note that the six empirical phenomena refer to properties of individual parameters in either
the temporal or residual networks. Therefore, we use a proportion metric to indicate the degree
to which a phenomena is endorsed. For instance, if four out of six autoregressive parameters
are positive, but the remaining two are negative, we would record that the degree to which that
phenomena is present is given by 2/3. Using the majority rule, if the endorsement proportion is
greater than 1/2 we would classify a phenomenon as being present.

33
D.1 Varying the Transition Matrix
A core feature of our basic model as specified in the main text is that situations are relatively
stable over time, represented in our model by enforcing that the staying probabilities α (diagonal
elements Pii ) are greater than any transition probability β (off-diagonal elements Pij ). In Figure
7, we show the results of varying α. As we can see, when α is less than or equal to β (α ≤ 0.20),
the model fails to reproduce most of the empirical phenomena. However, once α > β, all of the
phenomena are endorsed. The phenomena relating to the residual structure are endorsed across
all settings.

Sensitivity Analysis Results

1.0
Phenomena Present (Proportion)
0.8
0.6
0.4

Phenomena 1
Phenomena 2
0.2

Phenomena 3
Phenomena 4
Phenomena 5
Phenomena 6
Decision Threshold
0.0

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Staying Probability α

Figure 7: The effect of varying the staying probability α on the model’s ability to reproduce the
empirical VAR-based phenomena. The staying probabilities are varied from α = 0 to α = 0.99
while keeping the structure of the model otherwise fixed. The dashed lines indicate the decision
threshold of 0.50, such that models that exceed the threshold are taken to reproduce the phenomena
(see main text).

Changing α while enforcing that the rest of the model structure remains constant changes
the absolute value of the elements in the Φ matrix, but does so differently for the autoregressive
and cross-lagged effects. In Appendix C, we showed that the lag-1 auto-covariance is determined
by both the factor loadings and the lagged covariance between the latent states, expressed as
pij πi − πi πj , where pij is the transition probability from state i to state j, and πi is the stationary
probability of state i. When we enforce that the rows of the transition matrix have the same
structure (i.e. that the staying probability pii is equal for all i and the switching probabilities pij
are equal for all combinations of i and j), as a consequence all stationary probabilities are equal. For
a five state system, this yields π = 0.20. When pii = pij = 0.20, the latent state covariance is equal
to zero. As a result, there is no auto-covariance present in the system, and so none of the first four
phenomena are present. When α < 0.20, we obtain negative auto-covariances between the latent
states and so non-positive autoregressive effects. However, the absolute values of the autoregressive
effects are still larger than the cross-lagged effects, since the former are driven by auto-covariances
of the latent states and the latter by cross-covariances of the latent states. For instance, when
α = 0, we obtain απi − πi πj = −0.40, but the switching probabilities of β = 0.25 produce a latent

34
 Transition Matrix Mixture Means Mixture Means
Theory-Consistent Theory-Consistent Theory-Inconsistent
Mean % Absent Mean % Absent Mean %Absent
Phenomenon 1 0.989 0.0 1.000 0 1.000 0.0
Phenomenon 2 0.764 0.5 0.739 0 0.840 0.0
Phenomenon 3 0.775 0.3 0.990 0 0.498 41.9
Phenomenon 4 0.532 34.1 1.000 0 0.499 48.4
Phenomenon 5 1.000 0.0 0.989 0 0.498 41.9
Phenomenon 6 1.000 0.0 1.000 0 0.498 48.6

Table 2: Degree to which the Phenomena 1-6 are present (see main text for definition) and per-
centage of runs in which the phenomena are absent, depending on whether the transition matrix
and mixture means are consistent or inconsistent with our theoretical principles.

covariance of βπi − πi πj = 0.01. For this reason, the phenomenon that the autoregressive effects
are larger in absolute value than the cross-lagged effects is present for all settings except when
α = 0.20, while the other phenomena relating to the sign of lagged relationships are present only
when α > 0.20
To further explore the effects of transition matrix values on phenomena recovery, we consider
the model specification where we freely draw the transition matrix parameter from a uniform
distribution U (0, 1), and re-normalize the matrix to ensure that the staying probability for each
state is greater than any switching probability for that state. This allows both staying and switching
probabilities to differ from one another in contrast to the simpler model considered above, while
staying consistent with our theoretical expectation that α > β. We draw nsim = 1000 of these
models and assess in each case the degree to which each phenomena is present. The results are
shown in the first two columns of Table 2. Phenomena 1, 5, and 6 (positive AR parameters,
positive within-valence residual relationships, and negative between valence residual relationships)
are fully reproduced in all 1000 cases. Phenomena 2 and 3 are reproduced in the vast majority of
cases as they only fail to reproduce in 0.5% and 0.3% of cases, respectively. Only Phenomenon 4
— referring to negative cross-lagged effects between positively and negatively valenced emotions
— fails to be reproduced in 34.1% of cases.
From this we can state that the ability of the proposed model to reproduce VAR-based phenom-
ena is dependent on a specification which enforces that the probability of staying in the same state
is greater than that of any one state switch. However, keeping the rest of the model specification
fixed, it appears that most phenomena are robustly produced regardless of ones choice of actual
transition matrix values with the only exception being Phenomenon 4, which is only reproduced
in 65.9% of cases.

D.2 Varying the Mixture Means / Factor Loadings

A second feature of our basic model is the mean structure of the emotion variables and how they
are determined by the latent states. In the model presented in the main text, we chose a fixed
mean structure whereby neutral events elicit emotions with a mean of zero, positive events elicit
positive emotions with a mean value of γ, θ, and λ, and negative emotions with a mean of zero, and
vice versa for negative situations. In Appendix C, we represented this as a factor loading matrix.
To investigate model sensitivity, we here simulate nsim = 1000 new factor loading matrices with
the same basic structure, but we allow the non-zero mixture means to be drawn randomly from a
uniform distribution U (0.10, 5). Keeping the rest of the model fixed as described in the main text,
we assess whether varying the value of the factor loadings has an effect on the degree to which
our model reproduces the empirical phenomena. The results are presented in the third and fourth
column of Table 2. As we can see, the phenomena are always present. As such, we can say that
the models ability to reproduce the phenomena are robust against the specific choice of positive
factor loading values.
Of course, not all specifications of our basic model will produce all phenomena of interest. To
investigate this, we repeated the above procedure, but allowed the factor loadings to be drawn

35
from positive and negative values, that is U (−5, 5). We can consider this to be a version of the
model which is inconsistent with our theory: Positive situations can now produce negative values
of positive emotions, and vice versa. The results of this simulation are shown in the last two
columns of Table 2. As we can see, Phenomena 1 and 2 are reproduced consistently. We would
expect this from the results presented earlier, which showed that these phenomena were critically
dependent on the transition matrix values. However, the remaining phenomena are not consistently
reproduced, being present in only around 50% of cases. This reinforces the notion that it is not all
specifications of our basic model will reproduce these phenomena, but that these phenomena are
present in versions of the model consistent with our theoretical principles.

36