ArticlePDF AvailableLiterature Review

Abstract and Figures

This article introduces the Causal Attitude Network (CAN) model, which conceptualizes attitudes as networks consisting of evaluative reactions and interactions between these reactions. Relevant evaluative reactions include beliefs, feelings, and behaviors toward the attitude object. Interactions between these reactions arise through direct causal influences (e.g., the belief that snakes are dangerous causes fear of snakes) and mechanisms that support evaluative consistency between related contents of evaluative reactions (e.g., people tend to align their belief that snakes are useful with their belief that snakes help maintain ecological balance). In the CAN model, the structure of attitude networks conforms to a small-world structure: evaluative reactions that are similar to each other form tight clusters, which are connected by a sparser set of "shortcuts" between them. We argue that the CAN model provides a realistic formalized measurement model of attitudes and therefore fills a crucial gap in the attitude literature. Furthermore, the CAN model provides testable predictions for the structure of attitudes and how they develop, remain stable, and change over time. Attitude strength is conceptualized in terms of the connectivity of attitude networks and we show that this provides a parsimonious account of the differences between strong and weak attitudes. We discuss the CAN model in relation to possible extensions, implication for the assessment of attitudes, and possibilities for further study. (PsycINFO Database Record
Content may be subject to copyright.
Toward a Formalized Account of Attitudes:
The Causal Attitude Network (CAN) Model
Jonas Dalege
University of Amsterdam and University of Hamburg Denny Borsboom and Frenk van Harreveld
University of Amsterdam
Helma van den Berg
TNO (Netherlands Organization for Applied Scientific
Research), Soesterberg, the Netherlands
Mark Conner
University of Leeds
Han L. J. van der Maas
University of Amsterdam
This article introduces the Causal Attitude Network (CAN) model, which conceptualizes attitudes as
networks consisting of evaluative reactions and interactions between these reactions. Relevant evaluative
reactions include beliefs, feelings, and behaviors toward the attitude object. Interactions between these
reactions arise through direct causal influences (e.g., the belief that snakes are dangerous causes fear of
snakes) and mechanisms that support evaluative consistency between related contents of evaluative
reactions (e.g., people tend to align their belief that snakes are useful with their belief that snakes help
maintain ecological balance). In the CAN model, the structure of attitude networks conforms to a
small-world structure: evaluative reactions that are similar to each other form tight clusters, which are
connected by a sparser set of “shortcuts” between them. We argue that the CAN model provides a
realistic formalized measurement model of attitudes and therefore fills a crucial gap in the attitude
literature. Furthermore, the CAN model provides testable predictions for the structure of attitudes and
how they develop, remain stable, and change over time. Attitude strength is conceptualized in terms of
the connectivity of attitude networks and we show that this provides a parsimonious account of the
differences between strong and weak attitudes. We discuss the CAN model in relation to possible
extensions, implication for the assessment of attitudes, and possibilities for further study.
Keywords: network models, attitudes, tripartite model, connectionism, small-world
The attitude concept continues to occupy a central role in the
social sciences. Not only are attitudes a core topic in social
psychology; they also play an important role in economics and the
political sciences (e.g., Latané & Nowak, 1994). While research on
attitudes has spawned a vast literature, it has been argued that the
theoretical integration of empirical findings is still limited (e.g.,
Monroe & Read, 2008). In particular, a realistic formalized theo-
retical framework is lacking that can be directly related to empir-
ical data through statistical estimation and fitting techniques. In the
current article, we put forward such a formalized measurement
model of attitudes, and argue that this model shows promise in
integrating our understanding of the structural and dynamical
properties of attitudes.
Any formal measurement model of attitudes needs to fulfill
two basic properties. First, it must address how multiple re-
sponses on an attitude questionnaire relate to the attitude con-
struct. Second, it must provide an explanation of the correla-
tions among these multiple responses. Historically the most
influential model of attitudes that fulfills these two basic prop-
erties has been the tripartite model of attitudes. In this model,
attitudes are assumed to consist of cognitive, affective, and
behavioral components (e.g., Bagozzi, Tybout, Craig, & Stern-
thal, 1979;Breckler, 1984;Eagly & Chaiken, 1993;Fishbein &
Ajzen, 1975;Rosenberg, Hovland, McGuire, Abelson, &
Brehm, 1960). Formalized accounts of the tripartite model
assume that attitudes act as latent variables that cause these
three components, which in turn cause specific responses to
attitude questions. Due to a number of problems discussed
below this view of attitudes has fallen out of vogue (e.g., Fazio
& Olson, 2003a;Zanna & Rempel, 1988). However, no domi-
nant alternative formalized measurement model of attitudes has
yet replaced it. To fill this gap, the present article presents a
This article was published Online First October 19, 2015.
Jonas Dalege, Department of Psychology, University of Amsterdam and
Department of Psychology, University of Hamburg; Denny Borsboom and
Frenk van Harreveld, Department of Psychology, University of Amster-
dam; Helma van den Berg, TNO (Netherlands Organization for Applied
Scientific Research), Soesterberg, the Netherlands; Mark Conner, School
of Psychology, University of Leeds; Han L. J. van der Maas, Department
of Psychology, University of Amsterdam.
The authors thank Angelique Cramer, Juliane Degner, Stephen Read,
Joop van der Pligt, and Lourens Waldorp for comments and discussion.
Correspondence concerning this article should be addressed to Jonas
Dalege, University of Amsterdam, Weesperplein 4, 1018 XA Amsterdam,
the Netherlands. E-mail: j.dalege@uva.nl
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Psychological Review © 2015 American Psychological Association
2016, Vol. 123, No. 1, 2–22 0033-295X/16/$12.00 http://dx.doi.org/10.1037/a0039802
2
formalized measurement model of attitudes based around net-
work theoretical principles.
In evaluating the tripartite model’s contributions to attitude
research, it is important to distinguish between how the model
describes and explains attitude structure. In the descriptive sense,
the tripartite model represents the basic correlational structure of
responses to attitude questions: Different responses to the same
attitude object are substantively interrelated. However, in the ex-
planatory sense, the tripartite model is limited because it explains
the correlations among responses to attitude questions in terms of
the shared influence of a small number of latent variables. As
discussed below, it is our view that such latent variable models do
not offer plausible representations of the structure of attitudes in
relation to formation and dynamics.
Recently developed connectionist models of attitudes, on the
other hand, may better capture the dynamics of attitude formation
and change (Monroe & Read, 2008;van Overwalle & Siebler,
2005). In these connectionist models, human cognition is simu-
lated with a network of several interrelated nodes. These models
provide plausible mechanistic explanations of how attitudes form
and change as a result of the interplay between evaluative reactions
that concern the attitude object. However, a weakness of current
connectionist models of attitudes is that they are separated from
the dominant statistical approaches used in the analysis of attitude
data, and their relevance to empirical phenomena is thus indirect.
For instance, while connectionist models are suited to address the
inherent complexity of attitudes, currently implemented models
rely solely on simulated data (e.g., Monroe & Read, 2008). There-
fore, their description of attitudinal processes tends to be at the
metaphorical level. Thus, while the tripartite model can be fitted to
empirical data but relies on a substantively implausible model of
attitudes, current connectionist models of attitudes provide a plau-
sible account of attitude formation and change but cannot be fitted
to empirical data.
Here we propose a new network model of attitudes that can
address both the above concerns: the Causal Attitude Network
(CAN) model. This model conceptualizes attitudes as networks
that consist of evaluative reactions and interactions between these
reactions. Relevant reactions include beliefs, feelings, and behav-
iors toward an attitude object. Interactions between these reactions
arise through direct causal influences (e.g., the belief that snakes
are dangerous causes fear of snakes) and mechanisms that support
evaluative consistency between related contents of evaluative re-
actions (e.g., people tend to align their belief that snakes are useful
with their belief that snakes help maintain ecological balance). In
the CAN model, the structure of attitude networks is held to
conform to a small-world structure (e.g., Watts & Strogatz, 1998):
Evaluative reactions that are similar to each other form tight
clusters, which are connected by a sparser set of “shortcuts”
between them. Importantly, the CAN model allows for the appli-
cation of empirical network models (i.e., models in which ob-
served variables are treated as causally related nodes in a network;
Schmittmann et al., 2013;van der Maas et al., 2006). Such models
can be fitted to actual data, but can also deal with complex systems
(Barabási, 2011) such as attitudes, and thus combine the explan-
atory power of connectionist approaches with empirical analyses
of attitude structure. This combination allows us to derive a real-
istic psychometric conceptualization of attitudes.
The outline of this article is as follows. First, we discuss the
current lack of formalized measurement models in the attitude
literature. Second, we combine the notion of cognitive consistency,
recent connectionist modeling of attitudes (Monroe & Read, 2008)
and recent advancements in applying network theory in psychol-
ogy (e.g., Cramer, Waldorp, van der Maas, & Borsboom, 2010; for
excellent discussions of the relevance of network analysis to the
social sciences in general and psychology in particular see Bor-
gatti, Mehra, Brass, & Labianca, 2009;Westaby, Pfaff, & Red-
ding, 2014) to derive a set of requirements for a realistic formal-
ized measurement model of attitudes. Third, based on these
requirements we develop the CAN model and discuss the proposed
small-world structure of attitudes that underlies it. Fourth, we
discuss the CAN model’s perspective on attitude formation and
structure, attitude stability and change, and attitude strength. Fifth,
we discuss possible extensions of the CAN model, the model’s
implications for the assessment of attitudes, and some possible
avenues for further study of the CAN model.
The Need for a Formalized Measurement Model
of Attitudes
Attitude research was one of the first fields in psychology in
which researchers developed and tested formalized measurement
models (e.g., Bagozzi, 1981;Bagozzi & Burnkrant, 1979;Breck-
ler, 1984). These models rested on the tripartite model of attitudes,
which, as we noted, holds that attitudes consist of affective, be-
havioral, and cognitive components (e.g., Eagly & Chaiken, 1993;
Fishbein & Ajzen, 1975;Rosenberg et al., 1960) and treated the
components as unobservable common causes of observable vari-
ables (i.e., reflective latent variables). This conceptualization of
attitudes, however, has been criticized for not being able to inte-
grate inconsistencies between attitude and behavior, as the model
assumes that behavior is part of the attitude (e.g., Cacioppo, Petty,
& Green, 1989;Fazio & Olson, 2003a;Tesser & Shaffer, 1990;
Zanna & Rempel, 1988). Based principally on this shortcoming,
the formalized account of the tripartite model has largely fallen out
of vogue despite the lack of an acceptable formalized measurement
model to replace it.
Due to this omission, currently there is no satisfactory explana-
tion for a pervasive finding across the attitude field: attitude items
display a positive manifold (i.e., attitude items of the same valence
are substantively positively related and attitude items of different
valence are substantively negatively related; Bagozzi & Burnkrant,
1979;Bagozzi et al., 1979;Breckler, 1984;Conner, Godin,
Sheeran, & Germain, 2013;Haddock, Zanna, & Esses, 1993;
Kothandapani, 1971;Ostrom, 1969;van den Berg, Manstead, van
der Pligt, & Wigboldus, 2005). This positive manifold is reflected
in the findings that: (a) items, which assess the same component,
are highly interrelated (e.g., judging snakes as dangerous, ugly);
and (b) the fitted latent variables representing the components are
substantively interrelated (e.g., beliefs and feelings toward
snakes).
The formalized account of the tripartite model appeared to
provide a parsimonious explanation of the interrelations of items
and variables assessing the same component—such items are
related because they depend on the same underlying component.
However, adding to the former discussed critique of the tripartite
model, this explanation is likely to be incorrect because it rests on
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
3
CAUSAL ATTITUDE NETWORK MODEL
two assumptions of latent variable modeling that are improbable in
the context of attitudes. First, items, or indicators, must be locally
independent and second, items must be exchangeable (e.g., Bollen,
1989;Borsboom, 2008;Borsboom, Mellenbergh, & van Heerden,
2003;Cramer et al., 2010;Jöreskog, 1971).
Local independence refers to the assumption that indicators,
which measure the same latent variable, have no direct causal
influence on each other (e.g., two thermometers are independent
when temperature is held constant; Borsboom, 2008). In the con-
text of attitudes, this assumption is unrealistic, as it would mean
that there are no interactions between evaluative reactions. As
discussed later, this assumption is at odds with the notion of
cognitive consistency and with the basic mechanisms of recent
connectionist models of attitudes.
Exchangeability of indicators refers to the assumption that add-
ing indicators to a questionnaire only increases reliability but does
not add substantial information (e.g., adding thermometers to a
perfect thermometer is superfluous; Bollen & Lennox, 1991). If a
value of a person on any one of the observed indicators were
measured perfectly without any error (i.e., the true score), this
value would exhaustively characterize that person’s position on the
latent variable; thus, as soon as the true score on any one indicator
is known, the other indicators cannot add independent information.
In other words, the relations between true indicator scores are
deterministic: The reflective latent variable model can in fact be
derived from the assumption that all true indicator scores are
perfectly correlated (Jöreskog, 1971). In the context of attitudes,
this would mean that if a given evaluative reaction (e.g., a given
belief toward a presidential candidate) truly changes (i.e., change
caused by the underlying latent variable), the other evaluative
reactions, which belong to the same attitude component (e.g., all
other beliefs toward the presidential candidate), have to change the
exact same amount. All other variation is attributed to random
error. In our view, such an account of attitudes is much too
restrictive and not in accordance with views that emphasize inter-
nal inconsistencies within attitudes (e.g., Newby-Clark, McGregor,
& Zanna, 2002;Thompson, Zanna, & Griffin, 1995;van Harrev-
eld, van der Pligt, & de Liver, 2009).
If one is to reject the idea that beliefs, feelings, and behaviors
regarding an attitude object reflect underlying attitude compo-
nents, the question arises as to how such evaluative reactions relate
to the attitude concept. The currently prevalent view holds that
beliefs, feelings, and behaviors represent different classes of chan-
nels through which attitudes are formed (e.g., Zanna & Rempel,
1988).
1
As Fabrigar, MacDonald, and Wegener (2005, p. 82) state:
. . . the contemporary view holds that an attitude is an entity distin-
guishable from the classes of affect, behavior, and cognition. An
attitude, therefore, does not consist of these elements, but is instead a
general evaluative summary of the information derived from these
bases (Cacioppo et al., 1989;Crites, Fabrigar, & Petty, 1994;Zanna
& Rempel, 1988).
This perspective on attitudes also forms the basis for Fazio’s
(1995,2007) influential account of attitudes, in which attitudes are
defined as associations between an attitude object and a summary
evaluation. According to Fazio (1995,2007), this summary eval-
uation is derived from cognitive, affective, and/or behavioral in-
formation.
Conversely, the view that attitudes are causes of cognition,
affect, and behavior can also be found in current theorizing. This
view is, for example, expressed by Eagly and Chaiken (2007), who
state that “Attitude is...atendency or latent property of the
person that gives rise to judgments as well as to many other types
of responses such as emotions and overt behaviors” (p. 586).
To summarize, current theorizing on the relation between eval-
uative reactions and attitudes holds that attitudes can be formed by
cognition, affect, and behavior and that attitudes in turn also
influence cognition, affect, and behavior. A formal model of
attitudes should thus integrate this bidirectional influence between
the components of attitudes and the attitude itself. Such a model
should furthermore integrate the pervasive finding that evaluative
reactions are generally substantively positively interrelated (e.g.,
Bagozzi & Burnkrant, 1979;Breckler, 1984;Conner et al., 2013;
Haddock et al., 1993;Kothandapani, 1971;Ostrom, 1969;van den
Berg et al., 2005).
From Cognitive Consistency to Network Models
In this section, we discuss the notion of cognitive consistency
and use this notion to derive a formalized measurement model of
attitudes. In contrast to the formalized conceptualization of the
tripartite model of attitudes, which assumes that correlations be-
tween evaluative reactions are spurious, we propose that the cor-
relations between evaluative reactions are meaningful because
they stem from direct interactions between the evaluative reac-
tions. Through these interactions the attitude construct is formed.
The notion of cognitive consistency plays a central role in both
classic and contemporary theorizing in social psychology (Gawron-
ski, 2012;Gawronski & Strack, 2012). Both Heider’s balance
theory (1946,1958) and Festinger’s (1957) cognitive dissonance
theory make the assumption that humans have a basic need for
consistency between their cognitions. Consistency theories also
extend to the notion that feelings and beliefs should also be at least
somewhat consistent (i.e., affective-cognitive consistency; Rosen-
berg et al., 1960) and that people are motivated to reduce incon-
sistencies within their attitudes (e.g., van Harreveld et al., 2009).
Consistency theories thus assume that evaluative reactions have a
tendency to align with each other.
Recently, the mechanism of parallel constraint satisfaction was
used to implement a formalized conceptualization of cognitive
consistency in a connectionist model of attitudes (Monroe & Read,
2008). Parallel constraint satisfaction in the context of attitudes
holds that beliefs impose constraints on other beliefs. For example,
if two beliefs are positively connected and one of the beliefs
becomes activated, the other belief is likely to become activated as
well. On the other hand, if two beliefs are negatively connected
and one belief becomes activated, activation of the other belief is
subsequently inhibited. Whenever the attitude object is activated in
working memory, the associated beliefs self-organize in such a
way that the constraints are more and more satisfied (i.e., become
1
This view on attitudes is more related to formative factor models, in
which the indicators cause the construct (e.g., Bollen & Lennox, 1991). We
refrain from discussing such models in detail, because formative factor
models do not make any assumptions on the correlational pattern between
the measurements (e.g., Schmittmann et al., 2013) and are therefore of
limited use as measurement models.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
4DALEGE ET AL.
more consistent). The system thus strives for an optimized repre-
sentation of the attitude object.
Using simulation studies, it was shown that the model built on
the mechanism of constraint satisfaction can account for several
well-known empirical findings on attitudes (e.g., thought-induced
attitude polarization, systematic vs. heuristic processing, implicit
attitude processes; Monroe & Read, 2008). In addition to the
insights from cognitive consistency theories, parallel constraint
satisfaction further establishes that the need for consistency is a
driving factor in attitude formation and change (see also Holyoak
& Simon, 1999;Simon & Holyoak, 2002;Simon, Krawczyk, &
Holyoak, 2004;Shultz & Lepper, 1996;Simon, Snow, & Read,
2004).
To derive a measurement model of attitudes that integrates the
above-discussed notions of consistency and optimization, we use
empirical network models that have recently been applied to
research on clinical disorders and personality (Borsboom & Cra-
mer, 2013;Cramer et al., 2012;Cramer et al., 2010). In empirical
network models, relations between observed variables are not
assumed to reflect an underlying factor as, for example, general
intelligence, major depression, or in the current discussion, an
attitude (Borsboom, 2008;Cramer et al., 2010;van der Maas et al.,
2006). Rather, relations between variables are assumed to stem
from a network of causally related variables. The relation between,
for example, judging snakes as dangerous and as ugly arises
through a direct causal connection between the judgments in order
to make the attitude more consistent.
A network is a system of such interrelated variables. Some of
these variables are directly connected and others only indirectly
connected through other variables. While judging snakes as ugly is
probably directly related to judging snakes as dangerous, judging
snakes as ugly might cause fear of snakes only through also
judging snakes as dangerous. As long as the variables in a network
are all connected at least indirectly, information can flow from one
variable to all other variables in a network. The variables in the
network thus all align to some extent (i.e., they become corre-
lated), so that the variables can be regarded as parts of the same
system.
2
A fundamental difference between latent variable models and
empirical network models concerns the causal power of observable
variables. In latent variable models, observable variables have
virtually no causal power, as all causation flows from the latent
variable to the observable variables. Empirical network models
allow for more causal power of observable variables. From a
network perspective, observable variables are not merely indica-
tors of a psychological construct—the construct is isomorphic to
the observable variables and their causal connections. From this
perspective, the attitude construct is isomorphic to evaluative
reactions toward a given attitude object and the interactions among
those evaluative reactions (i.e., the whole network of the evalua-
tive reactions and the interactions between these reactions repre-
sents the attitude construct). It then follows that observable vari-
ables are part of the construct and therefore can influence the
construct (i.e., the whole system of the causally connected vari-
ables) and the construct can also influence the observable vari-
ables. For example, judging a snake as dangerous influences the
system of the attitude toward snakes through its causal connections
to the other evaluative reactions but is also influenced by the
system of the attitude through these causal connections. Judging
snakes as dangerous can thus both (to some extent) be a cause of
a negative attitude toward snakes and also a result or expression of
a negative attitude.
Network models are furthermore equally well suited to explain
interrelations between observable variables as latent variable mod-
els are. Several simulations that aimed to explain the positive
manifold of intelligence subtests with a network model showed
that networks can shape data that resembles different factor ana-
lytic findings (van der Maas et al., 2006).
By combining insights from cognitive consistency theories,
connectionist models of attitudes and empirical network models,
we derive two main requirements for a measurement model of
attitudes. First, correlations between evaluative reactions stem
from pairwise interactions between the reactions and, second,
these interactions are aimed at optimization of the consistency of
the evaluative reactions. These two requirements are readily ful-
filled by the recently proposed application of the Ising model
(Ising, 1925) to psychological data (Epskamp, Maris, Waldorp, &
Borsboom, in press;van Borkulo et al., 2014). The Ising model
belongs to the class of Markov Random Field models used in
network analysis and constitutes the Markov Random Field for
binary data (Kindermann & Snell, 1980). Originating in statistical
physics, the Ising model has been applied in several research areas.
The interested reader is referred to Epskamp et al. (in press) for a
thorough discussion of the Ising model.
In the Ising model, nodes can be of two states (!1, 1; Kinder-
mann & Snell, 1980). These nodes can represent all kinds of
objects that can be in two states (van Borkulo et al., 2014). For
example, dipoles of a magnet that are either in “spin up” or “spin
down,” neurons that either fire or not fire, psychopathological
symptoms that are either present or absent, or evaluative reactions
that are endorsed or that are not endorsed. Each node in a network
is connected to a given number of neighboring nodes and these
connections can be either positive or negative. If the connection
between two nodes is positive then two neighboring nodes will be
pressured to assume the same state and if the connection is neg-
ative then the two nodes will be pressured to assume the opposite
states. The connections can also vary in strength. Because of the
connections between the nodes, the first requirement of a mea-
surement model of attitudes—pairwise interactions between eval-
uative reactions—is fulfilled.
The second requirement of a measurement model of attitudes—
optimization of consistency—is fulfilled by the axiom of the Ising
model that the system strives to reduce energy expenditure. In a
given Ising model, configurations of the system that cost a lot of
energy to maintain are less likely than configurations that cost less
energy. Inconsistent attitudes therefore cause much energy expen-
diture and the system will strive to reduce this inconsistency.
Given two positively connected nodes, the configurations in which
both nodes are either positive ("1, "1) or negative (!1, !1) are
thus more likely than the configurations in which the nodes are not
2
One might object that alignment of variables suggests some exchange-
ability between the variables. To the extent that they share variance, this
objection is correct. However, variables seldom perfectly align and are
therefore also not completely exchangeable in terms of shared variance.
Furthermore, even if two variables perfectly align for some time, this does
not exclude the possibility that they are influenced by different factors at
other times.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
5
CAUSAL ATTITUDE NETWORK MODEL
aligned [("1, !1), (!1, "1)], as the latter configuration costs
more energy to maintain. The extent to which the unaligned
configurations cause energy expenditure depends on the strength
of the connection between the nodes.
To summarize, the Ising model represents a promising concep-
tualization of attitudes as this conceptualization readily integrates
both interactions between evaluative reactions and the need to
maximize cognitive consistency. In the next section, we present
the specifics of the Causal Attitude Network (CAN) that imple-
ments this conceptualization.
The Causal Attitude Network (CAN) Model
In the CAN model, attitudes are conceptualized as networks of
interacting evaluative reactions (e.g., feelings, beliefs, and behaviors
toward an attitude object) and the dynamics of the networks conform
to the Ising model. Evaluative reactions are represented as nodes in
the networks and causal influences between these reactions are rep-
resented as edges (i.e., links between the nodes). Edges can represent
either excitatory or inhibitory influence and can have varying weights
(i.e., the causal influence between evaluative reactions varies). Atti-
tude networks strive for an optimized consistent representation of the
attitude object to reduce energy expenditure. To acquire a consistent
state, evaluative reactions of the same valence generally have excit-
atory influence between them and evaluative reactions of different
valence generally have inhibitory influence between them.
However, as individuals are also motivated to hold at least some-
what accurate attitudes (e.g., Chaiken, Liberman, & Eagly, 1989;
Petty & Cacioppo, 1986), optimization is bound by the motivation to
have an accurate attitude. While striving only for consistency would
lead to perfectly aligned evaluative reactions, striving only for accu-
racy can, in some instances, lead to completely unaligned evaluative
reactions. To deal with this trade-off between optimization of consis-
tency and accuracy, attitude networks are proposed to show clustering
(i.e., different sets of evaluative reactions are highly interconnected).
Clustering allows for energy reduction within clusters (e.g., all eval-
uative reactions toward a person that pertain to the dimension of
warmth are highly aligned) but also allows for accuracy by having
unaligned or even misaligned clusters that do not cost much energy
(e.g., the evaluative reactions that pertain to the dimension of warmth
are not highly aligned to the evaluative reactions that pertain to the
dimension of competence).
The trade-off between optimization of consistency and accuracy is
somewhat reminiscent of the trade-off between effort and accuracy in
decision-making (e.g., Payne, Bettman, & Johnson, 1988,1993).
Noteworthy from the line of research on the trade-off between effort
and accuracy is that in many scenarios decision-strategies that do not
require much effort (e.g., heuristics) fare equally well or even better
than decision-strategies that require much effort (e.g., Payne et al.,
1988;Gigerenzer, Todd, & the ABC Research Group, 1999). This
might indicate that also in the case of attitudes, a representation that
optimizes the trade-off between consistency and accuracy might be
the most adaptive representation.
Attitude Formation and Structure
From the perspective of the CAN model, attitudes start out with one
or just a few specific evaluative reactions. These first reactions then
serve as a model for the person to predict which other characteristics
the attitude object might have. This is in accordance with the free-
energy principle, which holds that humans need to make inferences
from learned information to derive predictions (Friston, 2009;Friston,
Daunizeau, Kilner, & Kiebel, 2010). Then, also in line with the
free-energy principle, information is sought that conforms to this
prediction (see also Hart et al., 2009 for a meta-analysis that indicates
that individuals generally prefer information that supports their atti-
tudes). Evaluative reactions therefore cause readiness of other evalu-
ative reactions. However, this readiness is not deterministic as it is
also possible that no confirming information is found. Such a situation
is likely to arise when individuals are highly motivated to be accurate,
which lowers their preference for information that supports their
attitudes (Hart et al., 2009).
As individuals are motivated to make correct inferences, it is likely
that the strength of the readiness depends on the similarity of the
reactions because correct inferences are more likely to be made for
similar evaluative reactions. For example, two judgments that pertain
to the dimension of warmth (e.g., friendly and sincere) would cause
more readiness of each other than two judgments that pertain to
different dimensions (e.g., friendly and competent; cf. Fiske, Cuddy,
Glick, & Xu, 2002). Reactions from different attitudinal components
are also generally less similar than reactions from the same compo-
nent and are therefore also generally less closely connected. Another
factor that makes it more likely that an evaluative reaction causes
readiness for another evaluative reaction is that the evaluations share
the same valence because evaluations of differing valence can behave
relatively independent of each other (Cacioppo & Berntson, 1994).
Connecting this line of reasoning to research on the development of
networks, a recent study showed that nodes that make new connec-
tions are more likely to connect to similar nodes than to less similar
nodes (e.g., a U.S. web page is more likely to connect to another U.S.
web page than to a Russian web page; Papadopoulos, Kitsak, Serrano,
Boguña, & Krioukov, 2012).
Another factor of importance in the attachment of new nodes to the
network is the popularity of the nodes (i.e., how many connections a
node already has; Barabási & Albert, 1999). This means that nodes
are more likely to connect to nodes that already have many connec-
tions (a phenomenon known as preferential attachment). In the case of
attitudes, this would mean that evaluative reactions that already have
many connections are more likely to lead to the activation of addi-
tional evaluative reactions. The proposed mechanism behind this
effect is that evaluative reactions that are strongly connected already
have proven to be predictive in the past,whichmakessuchevaluative
reactions more likely to cause readiness of other evaluative reactions
in the present.
Let us illustrate how attitude networks might take shape using the
example of someone, say Bob, developing a positive attitude toward
Barack Obama during the American presidential election campaign of
2008. First, Bob formed the impression of Obama being an honest
person. Quickly after forming this initial simple impression, Bob also
formed similar judgments about Obama, such as judging him as a
moral person who cares about people like Bob. The formed judg-
ments hold each other in check, so that none of these judgments can
readily change without inflicting some change on the other judgments.
This already constitutes a small network, see Figure 1a.
After Bob has thought a bit more about Obama, he also extended
his judgment to other dimensions and came to the conclusion that
Obama is also intelligent and competent and that he will probably
be a good leader. These judgments form a new cluster but are also
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
6DALEGE ET AL.
to some extent linked to the other judgments through the connec-
tion between judging Obama as a good leader and judging him to
care about people like Bob, as Bob thinks that this is a crucial
aspect of being a good leader. The attitude network has thus grown
and now consists of two clusters (i.e., set of nodes that are highly
interconnected), which are connected by a shortcut, see Figure 1b.
The reason that this edge represents a shortcut is that removing this
edge would substantially decrease the global connectivity (i.e.,
average connectivity of each node with all other nodes; West,
1996) of Bob’s attitude network (Watts, 1999). In the current
example, removing the edge between judging Obama as caring and
judging Obama to be a good leader would result in Bob’s attitude
network no longer being fully connected.
At some point, Bob also had evaluative reactions of a more
affective nature toward Obama. Because he judged Obama as
honest and moral, he also started to feel hopeful toward Obama
and this in turn caused him to feel proudness and sympathy toward
him (see Figure 1c). Again, these different affective reactions
cannot change without exerting some pressure to change on the
other affective reactions. Furthermore, Bob’s feeling of hope to-
ward Obama and his judgments that Obama is honest and moral
are closely connected, so that when one of these evaluative reac-
tions increases or decreases, the other reactions will also more
readily increase or decrease.
Due to his already very positive attitude, Bob also started to
convince other people to vote for Obama, he placed a sticker on his
car saying “Vote Obama” and, of course, at the election he voted
for Obama. From the more specific evaluative reactions toward
Obama, more general evaluations subsequently emerged. For ex-
ample, Bob would state that he likes Obama and that he generally
sees him as a good person. These new clusters attached to the
evaluative reactions of judging Obama as caring and feeling hope-
ful toward Obama due to the popularity (i.e., number of connec-
tions) of these evaluative reactions (see Figure 1d).
As the above example illustrates, attitude networks are expected
to show a structure with high clustering, in which these clusters are
connected through shortcuts. Through these shortcuts, attitude
networks have high global connectivity (i.e., all nodes on average
are closely connected to each other). The combination of high
clustering and high connectivity is known as a small-world struc-
ture (Albert & Barabási, 2002;Watts & Strogatz, 1998). The
formalized definition of small-world networks holds that such
networks have a higher clustering than a random graph (i.e.,
network in which the nodes are randomly connected) but about the
cares
moral
honest
a
cares
moral
honest
leader
intelligent
competent
b
cares
moral
honest
leader
intelligent
competent
hope
proud
sympathy
c
cares
moral
honest
leader
intelligent
competent
hope
proud
sympathy
sticker
convince
vote
good
like
d
Figure 1. Hypothetical attitude network at four points in time (a–d). Nodes represent evaluative reactions and
edges represent causal influence between the evaluative reactions. Note that for reasons of simplification, all
edges represent excitatory influence and all edges have the same strength. The layouts of these graphs are based
on the Fruchterman-Rheingold algorithm (Fruchterman & Reingold, 1991), which places closely connected
nodes near each other. As all networks shown in this article, this network was created using the R-package
qgraph (Epskamp, Cramer, Waldorp, Schmittmann, & Borsboom, 2012). See the online article for the color
version of this figure.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
7
CAUSAL ATTITUDE NETWORK MODEL
same connectivity as a random graph. For example, the network
shown in Figure 1d has a clustering index almost three times as
high (C#.59) as the clustering index of a corresponding random
graph (C#.21) but only a slightly higher average shortest path
length (L#2.74) than the corresponding random graph (L#
2.33), implying slightly lower connectivity. For all network-based
calculations in this article, we used either the R-package qgraph
(Epskamp et al., 2012) or the R-package igraph (Csárdi & Nepusz,
2006).
The small-world structure has been observed for several net-
works from distinct areas of research. For example, small-world
structures have been found for the power grid of the western
United States; collaboration between film actors; neural networks
of worms (Watts & Strogatz, 1998), monkeys (e.g., Stephan et al.,
2000), and humans (e.g., Achard, Salvador, Whitcher, Suckling, &
Bullmore, 2006); human language (Ferrer-i-Cancho & Solé,
2001); and symptoms of psychological disorders (Borsboom, Cra-
mer, Schmittmann, Epskamp, & Waldorp, 2011).
To provide a first test of the hypothesis that attitude networks
have a small-world structure, we used the American National
Election Study (ANES) of 1984 (see the Appendix for a detailed
description of the data set). In the ANES of 1984, evaluative
reactions toward the presidential candidates were assessed in a
nationwide random sample of 2,257 participants. Participants were
asked whether or not they attributed several positive characteristics
to each candidate (e.g., whether the candidate is a decent, intelli-
gent or a moral person) and whether they had ever had positive or
negative feelings toward each candidate (e.g., feelings of hope or
anger). We used the participants’ responses toward these evalua-
tive reactions to estimate attitude networks for the attitudes toward
each presidential candidate.
3
To estimate attitude networks, we used the eLasso-procedure,
which is designed to find the optimal Ising model for a set of data
by regressing each variable on all other variables (van Borkulo et
al., 2014). The regression function is subjected to regularization to
control the size of the statistical problem (see Friedman, Hastie, &
Tibshirani, 2008;Tibshirani, 1996). For each node, the set of edges
that displays the best fit to the data is selected based on the fit of
the regression functions according to the Extended Bayesian In-
formation Criterion (Chen & Chen, 2008). Weights of the edges
are then based on the regression parameters in the selected neigh-
borhood functions (van Borkulo et al., 2014).
4
We then calculated the small-world index for the unweighted
networks (Humphries & Gurney, 2008). A small-world index
higher than one indicates that the network has a small-world
structure. To test whether the small-world index was significantly
higher than one, we calculated confidence intervals using 1,000
Monte-Carlo simulations of random graphs (Humphries & Gurney,
2008).
The estimated networks are shown in Figure 2. The network of
the attitude toward Ronald Reagan had a small-world index of 1.16
and the upper limit of the 99.9% confidence interval for the
corresponding random graphs was 1.13 (C
attitude network
#
.62, L
attitude network
#1.47, C
random network
#.53, L
random network
#
1.47). The network of the attitude toward Walter Mondale had a
small-world index of 1.25 and the upper limit of the 99.9%
confidence interval for the corresponding random graphs was 1.19
(C
attitude network
#.57, L
attitude network
#1.55, C
random network
#
.46, L
random network
#1.54). Both networks thus had a small-world
structure. Apart from the global structure, it can also be seen that
similar evaluative reactions are more closely connected than dis-
similar evaluative reactions. For example, both positive and neg-
ative feelings form distinct clusters (see Figure 2). Also, judgments
that pertain more to the warmth-dimension (e.g., the candidate is
fair, cares about people and is compassionate) and judgments that
pertain to the competence-dimension (e.g., the candidate is intel-
ligent, knowledgeable, and hardworking) are each closely con-
nected to each other.
To summarize, the CAN model holds that evaluative reactions
cause readiness of related evaluative reactions to the same attitude
object and through this process attitude networks take shape.
Similar evaluative reactions tend to cluster and these clusters are
connected by shortcuts, which give rise to the small-world network
structure of attitudes. Having discussed the formation and structure
of attitude networks we now turn to the implications of the CAN
model for change and stability in formed attitudes.
Attitude Stability and Change
After an attitude network has formed, the CAN model suggests
that different evaluative reactions tend to hold each other in check.
While in the formation phase of an attitude, the influence of
evaluative reactions on other evaluative reactions is unidirectional,
evaluative reactions in formed attitude networks probably have
bidirectional influence on each other. The reason for this is that an
evaluative reaction B that was activated by another evaluative
reaction A will also cause further readiness in evaluative reaction
A. As an example let us return to Bob’s attitude network. After
having formed the judgment that Obama cares for people like Bob
because he earlier judged Obama to be honest, judging Obama as
caring also serves as a model of how honest Obama is. So, if Bob
observes an incident that strengthens his judgment of Obama as
caring, his judgment of Obama being honest will also strengthen to
some extent.
While we have discussed attitude formation and the behavior of
formed attitudes in isolation, the distinction between these two
phases of attitudes will not be that clear-cut in reality. New nodes
3
The reason that we analyzed the ANES of 1984 instead of any other
ANES is that the number of assessed evaluative reactions was the largest
in the ANES of 1984. As the evaluative reactions are represented as nodes,
the size of the network (i.e., number of nodes) depends on the number of
assessed evaluative reactions. As the small-world index linearly increases
with the size of the network in small-world networks (Dunne, Williams, &
Martinez, 2002;Humphries & Gurney, 2008), the power to detect a
small-world structure is higher for the ANES of 1984 than for any other
ANES. Note, however, that the size of the estimated networks based on the
ANES of 1984 is still relatively small.
4
Note that also other methods than the eLasso-procedure can be used to
estimate networks from psychometric data (e.g., Costantini et al., 2015).
Two common practices involve estimating edges by using zero-order
correlations or using partial correlation between any given pair of the nodes
in the network while controlling for all other nodes (e.g., Cramer et al.,
2013;Cramer et al., 2010). The eLasso-procedure, however, has clear
advantages compared with these procedures. First, the eLasso-procedure
provides an estimation of the causal structure underlying the data, which
zero-order correlations do not. Second, the eLasso-procedure produces less
false positives than does using partial correlations (Costantini et al., 2015).
The reason for that is that the eLasso-procedure uses regularization, which
reduces the number of false positives. In fact, simulation studies have
shown that the eLasso-procedure has very high specificity (van Borkulo et
al., 2014).
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
8DALEGE ET AL.
and new connections will probably be formed continuously and the
extent of this likely depends on how often a person interacts in
some way with the attitude object. We discuss this point further in
the section on the relation between attitude strength and network
connectivity.
Apart from implications for attitude stability, conceptualizing
attitudes as networks also has implications for attitude change.
The most straightforward implication of conceptualizing atti-
tudes as networks for attitude change is that attitudes can be
changed via a plethora of different processes as each node in the
attitude network can serve as a gateway to instigate change in
the network. Looking back at Figure 1d, change in the network
could, for example, be instigated by cognitive dissonance (e.g.,
Bob did not vote for Obama because of minor situational
constraints and because of this a more negative evaluation
spread through the network; cf., Festinger, 1957), evaluative
conditioning (e.g., pairing Obama with images related to hope
through which an even more positive evaluation spread through
the network; cf., De Houwer, Thomas, & Baeyens, 2001;Jones,
Olson, & Fazio, 2010) or by presenting arguments (e.g., a friend
convinces Bob that Obama is not so competent after all from
which again a more negative evaluation spread through the
network).
Whether the state of the targeted evaluative reaction changes
(e.g., from "1to!1), however, is a function of not only the
strength of external pressure but also of the state of the neighbor-
ing nodes and the strength of the links between the targeted node
and the neighboring nodes. If a single node changes to a state that
is incongruent to its links with neighboring nodes, the energy
expenditure of the system rises. The amount of energy expenditure
of a given configuration is calculated using the Hamiltonian func-
tion H!!":
H(!)"#
#
i
$i!i##
i,j
%i,j!i!j, (1)
where $iis the threshold of any given evaluative reaction $
i
and
represents the disposition of the given evaluative reaction to be
endorsed (1) or not endorsed (!1). A threshold higher than 0
indicates that the probability of the given evaluative reaction to be
endorsed is higher (given the absence of any influence of neigh-
boring evaluative reactions) than the probability that the evaluative
reaction is not endorsed. Conversely, a threshold lower than 0
indicates that the probability of the given evaluative reaction to be
endorsed is lower than the probability that the evaluative reaction
is not endorsed.
%
i,j
is the weight of the interaction between !iand its neighbor-
ing evaluative reaction !j. A weight higher than 0 indicates that
that the interaction between two evaluative reactions is positive
(e.g., if one evaluative reaction is endorsed, the probability that the
other evaluative reaction is endorsed heightens). Conversely, a
weight lower than 0 indicates that that the interaction between two
evaluative reactions is negative (e.g., if one evaluative reaction is
endorsed, the probability that the other evaluative reaction is
endorsed lowers).
The probability of a given configuration can be calculated with
the following equation:
hard-wor king
decent
compassionate
respect
intelligent
moral
kind
inspiring
knowledgeable
good example
cares
leadership
understands
fair
in touch
angry
hope
afraid
proud
disgusted
sympathetic
uneasy
Ronald Reagan
hard-working
decent
compassionate
respect
intelligent
moral
kind
inspiring
knowledgeable
good example
cares
leadership
understands
fair
in touch
angry
hope
afraid
proud
disgusted
sympathetic
uneasy
Walter Mondale
Figure 2. Estimated attitude networks toward the two main candidates in the U.S. presidential election in 1984.
Red (gray) nodes represent positive judgments, blue (light gray) nodes represent positive feelings, and green
(dark gray) nodes represent negative feelings (see the Appendix for the complete wording of the items). Green
(solid) edges indicate excitatory influence between the nodes and red (dashed) edges indicate inhibitory
influence between the nodes. Thicker edges represent higher weights of the edges. The same algorithm as for
Figure 1 was used for the layout of these graphs. See the online article for the color version of this figure.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
9
CAUSAL ATTITUDE NETWORK MODEL
Pr(&"!)"1
zexp(#H(!)), (2)
where Zis the normalizing constant that guarantees that the prob-
abilities sum to 1 and is given by:
Z"#
!
exp(#H(!)) (3)
To acquire a consistent state, weights between the evaluative
reactions of the same valence (e.g., honest and fair) have to be
positive and weights between evaluative reactions of different
valence have to be negative (e.g., honest and mean). The strength
of the weights is a function of the amount of interaction with the
attitude object (Monroe & Read, 2008) and the similarity of the
evaluative reactions.
The CAN model holds that the threshold depends on several
factors, implying that evaluative reactions differ in their disposition to
be endorsed. While a high threshold indicates that a given evaluative
reaction has a disposition to be endorsed, a low threshold indicates
that a given evaluative reaction has a disposition to be not endorsed
(see Equation 1). First, some evaluative reactions are likely to have
inherently higher thresholds than other evaluative reactions. For ex-
ample, some emotions are experienced more frequently (e.g., joy,
anger) than others (e.g., euphoria, contempt; Schimmack & Diener,
1997) and are therefore likely to have higher thresholds. Second,
thresholds of different persons probably also vary. For example,
persons differ in their disposition to adopt negative or positive
attitudes (Eschleman, Bowling, & Judge, 2015;Hepler & Albar-
racín, 2013). Third, whether one has endorsed (not endorsed) an
evaluative reaction in the past will probably also heighten (lower)
the threshold of this evaluative reaction in the future, implying that
the evaluative reaction is more likely to be endorsed (not endorsed)
in the future. This postulate is based on the finding that rehearsal
of attitudes results in the strengthening of attitudes (e.g., Fazio,
1995). Fourth, external persuasion attempts might also affect the
threshold of a given evaluative reaction. One route to change an
attitude would thus be to change the threshold of evaluative
reactions.
Let us return to Bob to illustrate how an attitude network would
be affected if the threshold of an evaluative reaction were to
change as a result of a persuasion attempt. For simplification, we
will only look at the cluster of the evaluative reactions that Obama
is competent, intelligent, and a good leader. Let us assume that the
thresholds before the persuasion attempt were all equal to .3 and
that the weights between all evaluative reactions were all equal to
.5. These represent moderately high thresholds and weights (using
uniform weights and thresholds is a simplification, as these pa-
rameters are likely to vary in real attitude networks, a point we
address later).
The persuasion attempt on Bob’s attitude was directed at the
evaluative reaction that Obama is competent and the attempt was
quite strong. Because of this, Bob’s threshold of judging Obama as
honest dropped to !.3. We can now calculate the energy that each
of the configurations will cost and how probable each configura-
tion is. As can be seen in Table 1, the configuration in which all
three evaluative reactions are endorsed, is the most likely config-
uration. The most likely scenario for this illustration is thus that,
while the persuasion attempt was quite strong, Bob will still think
of Obama as competent. This illustrates our point that the success
of a persuasion attempt not only depends on the strength of the
persuasion attempt but also on the parameters of the attitude
network. It is also important to note that it is more probable that
evaluative reactions will become uniformly negative after the
persuasion attempt, than that the judgment of Obama as competent
will change in isolation. This leads to a testable prediction of the
CAN model: If one evaluative reaction changes and this change
persists, other evaluative reactions are also likely to change.
Changes in the threshold of a given node can be interpreted as
rather subtle attitude change. Such changes can, for example, arise
if arguments are presented that make one doubt one’s current
evaluation. However, there might also be situations, in which one
not only doubts one’s current evaluation but discards it completely.
Such attitude change could be modeled by fixing one of the nodes
to a given state. This would be the case if, for example, Bob were
to receive information which would make it impossible to remain
viewing Obama as competent. In this event, only the configura-
tions in which judging Obama as competent is not endorsed feed
into Equation 3. As a result, the configuration in which none of the
positive evaluative reactions are endorsed becomes the most likely
scenario.
While up to this point we have discussed how attitude change
takes place in a network, in which all nodes are uniformly con-
nected to each other, nodes in real attitude networks will differ in
their connectivity. This can be seen in both the hypothetical
attitude network of Bob and the estimated networks of the ANES
data. How a given node is connected in the network will influence
whether and how change in this node will spread to other nodes.
Connecting attitude change to the small-world structure of atti-
tudes, the first important implication of network structure for
attitude change is the existence of clusters within the network. As
can be seen in Figure 2, negative feelings toward the presidential
candidates form rather distinct clusters in both attitude networks. If
a node in this cluster were to be changed, this change would mostly
spread to other nodes in this cluster.
If the different clusters in an attitude network were not strongly
connected, configurations of the network, in which clusters do not
align to the same evaluation, would cost rather low energy. Be-
cause of this, even an attitude network consisting of clusters that
are not aligned to each other can be relatively stable. Even mis-
aligned clusters might cause low energy when the weights between
Table 1
All Possible Configurations of the Evaluative Reactions
Competent, Good Leadership, and Intelligent and Their
Associated Energies and Probabilities
Competent Good leader Intelligent H!!"Pr(X #$)
!1!1!1!1.2 .24
1!1!1 1.4 .02
!11!1 .2 .06
11!1 .8 .03
!1!1 1 .2 .06
1!1 1 .8 .03
!111!.4 .11
111!1.8 .44
Note. Competent has a threshold of !.3 and good leader and intelligent
have thresholds of .3, respectively. All weights between the evaluative
reactions are equal to .5.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
10 DALEGE ET AL.
the clusters are lower than the inconsistent thresholds of the
different clusters. For example, if both negative feelings and
positive beliefs in the attitude network toward Ronald Reagan had
highly positive thresholds and the (negative) weights between the
clusters were low, having both negative feelings and positive
beliefs toward Ronald Reagan would not cause much energy
expenditure.
Apart from belonging to different clusters, nodes can also differ
in their centrality. Centrality refers to the structural importance of
a given node in the network and three of the most popular cen-
trality measures are betweenness,degree, and closeness (Freeman,
1978;Opsahl, Agneessens, & Skvoretz, 2010). Betweenness refers
to how often a node lies in the shortest path between two other
nodes. The shortest path between two nodes is defined as the
shortest distance it takes to “travel” between two nodes over the
edges in the network. The distance is a function of both the number
and strength of the edges that lie between two nodes.
Betweenness can be linked to the notion of clustering, as nodes
that connect different clusters will generally have high between-
ness. If change affects one cluster of the attitude network, whether
the change will spread through the whole network depends on
the behavior of the nodes that connect this cluster to other parts
of the network. In the network of the attitude toward Ronald
Reagan, the node with the highest betweenness is the evaluative
reaction of whether he sets a good example. As can be seen in
Figure 2, this node is rather closely connected to the negative
affect cluster. Thus, whether change in the negative affect cluster
would spread through the network would depend on whether you
change your mind that Ronald Reagan sets a good example.
Degree is a function of the number of neighbors a given node
has and of the weights of the edges between the given node and its
neighbors (Freeman, 1978). Degree thus refers to how strongly a
given node is directly connected to all other nodes in the network.
Closeness is the inverse of the shortest path length between a given
node and all other nodes in the network and thus refers to how
strongly a given node is both directly and indirectly connected to
all other nodes in the network.
While a node with high degree and/or high closeness is very
unlikely to change independent of change in the whole network
and vice versa, a node with low degree and low closeness can
change rather independently of the other nodes in the network.
Therefore, it will be more difficult to change nodes with high
degree and/or high closeness than nodes with low degree and low
closeness. If change, however, takes place, it will be more conse-
quential if it takes place in a node with high degree and/or
closeness than when it takes place in a node with low degree and
low closeness. For example, the evaluative reaction with both the
highest degree and highest closeness in the network of the attitude
toward Ronald Reagan is the judgment of whether he cares about
people like oneself. It is thus likely that it would be very difficult
to change this judgment but that change in this judgment would
affect the attitude network to a large extent.
To summarize, different routes of attitude change can be inte-
grated in the CAN model. First, subtle attitude change can be
modeled by decreasing or increasing the threshold of a given node
and drastic attitude change can be modeled by fixing a given node
to a given state. Attitude change will also depend on the position
of the targeted evaluative reaction in the attitude network. Change
of an evaluative reaction will first spread to evaluative reactions
that belong to the same cluster and the extent to which an attitude
network will be affected by change in a single evaluative reaction
will depend on the centrality of the evaluative reaction. While
highly central evaluative reactions will be likely to resist change,
their change will also be more consequential than change in an
evaluative reaction that is not central.
Attitude Strength
Strong attitudes are defined by their stability, resistance to
change, and impact on behavior and information-processing (Kros-
nick & Petty, 1995;Visser, Bizer, & Krosnick, 2006). Apart from
these key features of attitude strength, several other attributes have
been identified that are related to attitude strength, such as extrem-
ity (Abelson, 1995), elaboration (Petty, Haugtvedt, & Smith,
1995), and importance (Boninger, Krosnick, Berent, & Fabrigar,
1995). In this section, we show that the CAN model integrates the
key features of attitude strength into a single framework. Specif-
ically, in the CAN model, the global connectivity (i.e., average
shortest path length; West, 1996) of an attitude network can be
regarded as a mathematically formalized conceptualization of at-
titude strength. We first discuss how the global connectivity of a
network affects the dynamics of a network and link these differing
dynamics to the key features of attitude strength. Then we discuss
some empirical support for the proposition that attitude strength
and connectivity of attitude networks are related, after which we
discuss how connectivity of attitude networks relates to the other
attributes related to attitude strength.
As we already hinted at in our discussion of attitude change,
stability and resistance to change of a given node in a network
depends on how strongly this node is connected to other nodes in
the network. By extending this argumentation to the whole net-
work, it follows that nodes in a highly connected network are more
stable and resistant than nodes in a weakly connected network and
that because of this the whole network can be regarded as more
stable and resistant (Kindermann & Snell, 1980;van Borkulo et al.,
2014). Note that for these dynamics to occur, the assumption is
made that the weights are organized in a way that not much
conflicting influence is present (e.g., connections between evalu-
ative reactions of the same valence are mostly excitatory).
To illustrate that highly connected attitude networks are more
stable than weakly connected attitude networks, let us return to the
example of Bob’s judgments that Obama is competent, intelligent,
and a good leader. To compare the dynamics of these judgments in
a highly connected network to the dynamics in a weakly connected
network, we can set the weights between the nodes to 1 (repre-
senting a highly connected network) or .1 (representing a weakly
connected network). To first focus on stability without external
pressure, the thresholds were all set to .3. As can be seen in Table
2, the probability that all judgments are positive is much higher in
the highly connected network than in the weakly connected net-
work. These probabilities can be interpreted as the likelihood that
we would observe a given configuration if we measure a person’s
attitude at a given point in time. So, if we were to measure attitudes
of a group of individuals who have a highly connected attitude
network at two points in time, we would find a higher correlation
between their attitude scores than were we to measure a group of
individuals who have a weakly connected network, which mirrors
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
11
CAUSAL ATTITUDE NETWORK MODEL
how attitude stability is assessed in the attitude strength literature
(e.g., Bassili, 1996;Prislin, 1996).
To illustrate that highly connected attitude networks are also
more resistant than weakly connected attitude networks, we return
to our example, in which persuasion was targeted at Bob’s eval-
uative reaction that Obama is competent. Again, we set the thresh-
old of this evaluative reaction to !.3 and calculate the probability
of what would happen in a highly connected network (all weights
equal to 1) and in a weakly connected network (all weights equal
to .1). Based on the probabilities shown in Table 2, the chance that
the targeted evaluative reaction will change is higher for the
weakly connected network than for the highly connected network.
The chance that the sum score of the three evaluative reactions will
reflect a less positive evaluation is also higher for the weakly
connected network than for the highly connected network.
5
The
probability that the targeted evaluative reaction will become neg-
ative is .62 and the expected mean evaluation of the three evalu-
ative reactions is .12 for the weakly connected network compared
with .38 and .28 for the highly connected network. Highly con-
nected attitude networks are thus more likely to resist persuasion
attempts than weakly connected networks, which fits what is
known about the resistance of strong versus weak attitudes to
persuasion attempts (e.g., Bassili, 1996;Visser & Krosnick, 1998).
Impact of attitude on behavior in the attitude strength literature
is generally assessed by measuring individuals’ attitude at a given
point in time and then predicting a related behavior at a later point
in time (e.g., Fazio & Williams, 1986;Holland, Verplanken, & van
Knippenberg, 2002). There are two reasons to expect attitudes with
highly connected networks to be more predictive of behavior than
weakly connected networks. First, due to the stability and resis-
tance of highly connected networks, it is more likely that the
attitude will be the same at the time it is measured and at the time
the behavior is executed.
Second, as we can see in Table 2, evaluative reactions in highly
connected attitude networks are more likely than evaluative reac-
tions in weakly connected attitude networks to align to each other.
An aligned attitude network is likely to be more informative for a
decision on whether a related behavior should be executed or not.
For example, it would be easier for Bob to decide to vote for
Obama if he thinks that Obama is competent, a good leader, and
intelligent than when he only thinks that Obama is competent and
intelligent but not really a good leader. In the latter case, Bob
would probably base his decision on other factors (e.g., which of
the candidates his friends prefer), which would make his attitude
less predictive of his voting behavior. Furthermore, it is possible
that the salience of evaluative reactions differs between situations
(cf., Sparks, Conner, James, Shepherd, & Povey, 2001). If Bob’s
evaluative reactions differed in their endorsement, it might be that
at the time his attitude is assessed, mostly negative evaluative
reactions are salient. By the time he executes the behavior, other
evaluative reactions that are mostly positive might be salient. In
such a case, his measured attitude would have low predictive
value.
The impact of strong attitudes on information processing refers
to the power of strong attitudes in directing attention and influ-
encing the way in which incoming information is integrated (e.g.,
Fazio & Williams, 1986;Houston & Fazio, 1989;Roskos-
Ewoldsen & Fazio, 1992). The connectivity of an attitude network
influences the way in which incoming information is integrated,
because evaluative reactions that are not aligned to each other cost
more energy in a highly connected attitude network than in a
weakly connected attitude network. There is thus more pressure to
fit incoming information to one’s evaluative reactions in a highly
connected network than in a weakly connected network. This
pressure might also lead to heightened attention to attitude objects
in order to detect “attacks” on the attitude network at an early
stage.
As we have shown, the key features of attitude strength follow
from conceptualizing strong attitudes as highly connected net-
works. From this reasoning, the hypothesis follows that strong
attitudes correspond to highly connected networks. Indeed, this is
exactly what we found in a study focusing on attitudes toward
presidential candidates in the American presidential elections from
1980–2012 (Dalege, Borsboom, van Harreveld, & van der Maas,
5
In the CAN model, the unweighted sum score of the evaluative
reactions can be used as a measure of the overall state of the network. A
possibility for a more sophisticated calculation for the overall state of the
network might be to weigh the evaluative reactions by their closeness
centrality, as evaluative reactions with a high closeness hold more infor-
mation about the other evaluative reactions than evaluative reactions with
low closeness.
Table 2
Configurations of Three Evaluative Reactions (Competent, Good Leader, Intelligent) and
Associated Probabilities With Either Congruent or Incongruent Thresholds and Either Low or
High Weights
Pr!X"!"
&
i
#{.3, .3, .3} &
i
#{!.3, .3, .3}
Configuration %
ij
#{.1, .1, .1} %
ij
#{1, 1, 1} %
ij
#{.1, .1, .1} %
ij
#{1, 1, 1}
(!1, !1, !1) .06 .14 .11 .33
(1, !1, !1) .07 .00 .04 .00
(!1, 1, !1) .07 .00 .13 .01
(1, 1, !1) .13 .01 .07 .01
(!1, !1, 1) .07 .00 .13 .01
(1, !1, 1) .13 .01 .07 .01
(!1, 1, 1) .13 .01 .24 .02
(1, 1, 1) .35 .83 .20 .61
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
12 DALEGE ET AL.
2015). In this study, we first assigned participants into three
strength groups based on whether the participants were interested
in presidential campaigns (interest in an attitude object is known to
be a reliable indicator of attitude strength; Krosnick, Boninger,
Chuang, Berent, & Carnot, 1993). We then checked whether these
groups differed in their attitude’s stability, extremity, and impact
on behavior. All of the attributes differed between the groups in the
expected direction, so that the high attitude strength group had
attitudes with the highest stability, extremity and impact on be-
havior.
We then estimated networks for each attitude strength group for
the attitudes toward the Democratic and Republican candidate at
each election and compared the global connectivity of the attitude
networks, giving us a total of 18 sets of attitude networks of groups
with either low, moderate, or high strength. Confirming the hy-
pothesis that strong attitudes correspond to highly connected net-
works, we found that the groups’ networks differed robustly and
strongly in their global connectivity. In every set of attitude net-
works, the network of the strong attitude group had a higher
connectivity than average and in all but one set, the network of the
weak attitude group had a lower connectivity than average.
Thus, the conceptualization of strong attitudes as highly con-
nected networks directly implies that strong attitudes are stable and
resistant. By adding a few simple assumptions, the other two key
features of attitude strength—impact on behavior and information
processing—can also be integrated in the current framework.
Furthermore, initial empirical results show that attitude strength
and connectivity of attitude networks are indeed strongly and
robustly related. The question then arises of how attributes that are
related to attitude strength relate to the connectivity of attitude
networks.
In their review on attitude strength-related attributes, Visser,
Bizer, and Krosnick (2006) identified elaboration (Petty et al.,
1995), importance (Boninger, Krosnick, Berent, & Fabrigar,
1995), knowledge (Wood, Rhodes, & Biek, 1995), accessibility
(Fazio, Jackson, Dunton, & Williams, 1995), certainty (Gross,
Holtz, & Miller, 1995), ambivalence (Thompson et al., 1995),
structural consistency (Chaiken, Pomerantz, & Giner-Sorolla,
1995), extremity (Abelson, 1995), and intensity (Cantril, 1946) as
the most prominent attributes related to attitude strength. It is our
view that these attributes can be roughly grouped into three clus-
ters in relation to connectivity of attitude networks: attributes that
are determinants of heightened connectivity, attributes that are
consequences of connectivity, and attributes that moderate the
consequences of connectivity.
Elaboration is probably a direct cause of connectivity. Building
on the proposition of connectionist models of attitudes that eval-
uative reactions self-organize when the attitude object is activated
in working memory (Monroe & Read, 2008) and the principle that
humans are motivated to decrease free-energy (e.g., Friston, 2009),
it follows that connections between evaluative reactions will be-
come stronger when a person elaborates on his or her attitude. The
same holds for determinants of attitude importance. Three key
determinants of attitude importance are self-interest, social iden-
tification and value relevance (Boninger, Krosnick, & Berent,
1995). Self-interest refers to the extent that one perceives that a
given attitude has influence on one’s life, social identification
refers to whether an attitude is relevant to groups a person iden-
tifies with, and value relevance refers to how strongly an attitude
is linked to a person’s value. What these three determinants of
attitude importance have in common is that they all make it likely
that an attitude will play a large role in a person’s life. The attitude
object therefore is frequently activated in working memory. This
is, for example, illustrated by the finding that individuals often
think about attitudes that are important to them (Herzog, 1993;
Krosnick et al., 1993). It thus follows that determinants of attitude
importance are also likely to determine connectivity of attitude
networks.
Knowledge on the attitude object can rather be seen as an
amplifier of attitude strength than as being directly related to
attitude strength—the effects of attitude strength are more pro-
nounced when there is a large amount of knowledge (e.g., a strong
attitude in combination with much knowledge is very resistant to
change; Wood et al., 1995). In the CAN model, large amounts of
knowledge about an attitude object could best be modeled by
having a network with many nodes, as it is likely that an individual
with much knowledge about an attitude object will also have many
different evaluative reactions. Knowledge about an attitude object
would thus affect the size of the attitude network but not neces-
sarily the connectivity of the attitude network. The size of a
network was shown to amplify the effect of connectivity on
network dynamics (Cramer et al., 2013) leading to the hypothesis
that attitude networks that are both highly connected and consist of
many different evaluative reactions will correspond to stronger
attitudes. This is in line with the idea that knowledge amplifies the
effects of attitude strength (Wood et al., 1995).
In our view, accessibility, certainty, structural consistency, in-
tensity, and extremity are all likely to be caused by connectivity.
Accessibility refers to how fast a person can judge whether a given
attitude object is positive or negative (e.g., Fazio & Williams,
1986). We would argue that it is easier to judge an attitude object
as either positive or negative if one has an attitude that consists of
aligned evaluative reactions (which is more likely in a highly
connected attitude network) than when the attitude consists of
unaligned evaluative reactions (which is more likely in a weakly
connected attitude network). This reasoning is supported by the
finding that judging an attitude object, to which an individual holds
an ambivalent attitude, takes longer than judging an attitude object,
to which an individual holds an univalent attitude (Bargh, Chaiken,
Govender, & Pratto, 1992;van Harreveld, van der Pligt, de Vries,
Wenneker, & Verhue, 2004).
Attitude certainty can be divided into the constructs attitude
clarity and attitude correctness (Petrocelli, Tormala, & Rucker,
2007). Attitude clarity refers to how certain a person is what his or
her attitude actually is and attitude correctness refers to how
convinced a person is that his or her attitude is valid. It is our view
that attitude clarity rather than attitude correctness can be directly
linked to the connectivity of attitude networks. First, having an
aligned attitude network will probably make it more likely that a
person is certain about his or her attitude. Second, both attitude
clarity and connectivity of attitude networks are likely to be higher
when one has frequently interacted with the attitude object. This
could also explain why some ambivalent attitudes are held with
certainty. This would be the case if, for example, one has received
much conflicting information regarding an attitude object and the
strongly connected attitude network was not able to settle in a
state, in which the evaluative reactions are aligned. We will return
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
13
CAUSAL ATTITUDE NETWORK MODEL
to this point in more detail when we discuss the relation between
ambivalence and connectivity of attitude networks.
Structural consistency refers to attitudes being evaluative-
affective consistent, evaluative-cognitive consistent and affective-
cognitive consistent (Chaiken et al., 1995). Affective-cognitive
consistency directly follows from an attitude network being highly
connected because the affective and cognitive nodes are likely to
align. Also, having an attitude network, in which all (or most) of
the evaluative reactions are aligned, makes it likely that both
feelings and beliefs will be consistent with one’s general evalua-
tion (i.e., evaluating an attitude object as positive or negative
overall).
Intensity refers to how strongly an attitude object elicits emo-
tional reactions (Visser et al., 2006). As emotional reactions rep-
resent nodes in an attitude network, strong emotional reactions are
expected in attitudes with highly connected networks. Returning to
Bob’s attitude toward Obama, Bob will have strong positive emo-
tional reactions toward Obama when his attitude network is highly
connected as these reactions are pressured to align to the overall
positivity of the attitude network. In a weakly connected network,
his emotional reactions can vary more freely, so that weaker
emotional reactions are expected.
Finally, extremity follows directly from high connectivity. As
can be seen in Table 2, in highly connected attitude networks it is
virtually impossible that the sum score of the evaluative reactions
takes a moderate value. This finding also links the CAN model to
the catastrophe model of attitudes (Flay, 1978;Latané & Nowak,
1994;Zeeman, 1976), which holds that attitude importance deter-
mines whether attitudes act like categories (i.e., attitudes are only
stable in discrete states of positive or negative evaluations) or
dimensions (i.e., attitudes can take any place on the dimension
ranging from positive to negative): Important attitudes act more
like categories and unimportant attitudes act more like dimensions.
Studies that focused on political attitudes provided support for this
hypothesis (Latané & Nowak, 1994;Liu & Latané, 1998;van der
Maas, Kolstein, & van der Pligt, 2003). It is our view that impor-
tant attitudes or more generally, strong attitudes, act as categories
because strong attitudes correspond to highly connected networks.
Linking attitude strength to the connectivity of attitude networks
can also shed more light on the consequences of ambivalence. In
relation to attitude strength, ambivalence is a puzzling phenome-
non because, on the one hand, ambivalent attitudes are weak
predictors of behavior (Armitage & Conner, 2000;Conner &
Sparks, 2002), and on the other hand, have high impact on infor-
mation processing (Jonas, Diehl, & Bromer, 1997;Nordgren, van
Harreveld, & van der Pligt, 2006). High impact on information
processing of ambivalent attitudes was mostly found in the context
of felt ambivalence (i.e., ambivalence that causes discomfort; e.g.,
van Harreveld, Rutjens, Rotteveel, Nordgren, & van der Pligt,
2009;van Harreveld, van der Pligt et al., 2009). In our view, felt
ambivalence is likely to arise in highly connected networks with
unaligned evaluative reactions because highly connected networks
are in an unstable state when the nodes are not aligned (Cramer et
al., 2013). This proposition can explain why ambivalent attitudes
are less predictive of behavior while they do influence information
processing. Due to the high connectivity, motivation arises to
process information in a way that resolves the unstable state of
ambivalence (cf., Nordgren et al., 2006;van Harreveld, Nohlen, &
Schneider, 2015;van Harreveld, van der Pligt et al., 2009) but
because the evaluative reactions are not aligned yet, the impact on
behavior is low.
To summarize, conceptualizing strong attitude as highly con-
nected networks provides a framework in which the key features of
attitude strength—stability, resistance, and impact on behavior and
information processing—as well as attributes related to attitude
strength, such as importance, elaboration and extremity can be
integrated. Furthermore, this conceptualization sheds more light on
the underlying process of the catastrophe model of attitudes and
can explain why ambivalent attitudes can be regarded as both weak
and strong.
Discussion
In the present article, we introduced the Causal Attitude Net-
work (CAN) model as a formalized measurement model of atti-
tudes. In this model, attitudes are conceptualized as networks of
interacting evaluative reactions (e.g., beliefs, feelings, and behav-
iors toward an attitude object). Interactions arise through direct
causal connections and mechanisms that support evaluative con-
sistency. Attitude networks are driven by the trade-off between
optimization (i.e., consistency between evaluative reactions) and
accuracy. This trade-off results in a small-world structure, in
which evaluative reactions, that are similar to each other, tend to
cluster. Conceptualizing attitudes as networks provides testable
hypotheses for attitude change (e.g., change in an evaluative re-
action will foremost affect the cluster it belongs to) and a parsi-
monious explanation for the differences between strong and weak
attitudes by conceptualizing attitude strength as connectivity of
attitude networks. Initial empirical tests of the CAN model support
both the proposed small-world structure of attitude networks and
the hypothesis that strong attitudes correspond to highly connected
attitude networks.
Extensions of the CAN Model
The CAN model as presented in this article can only deal with
binary data but items are often assessed on nominal or continuous
scales in the attitude literature. A natural extension of the model
therefore would be to allow for nominal or continuous data. Such
extensions are relatively straightforward as there are both Markov
Random Field models available for nominal data (Potts model;
Wu, 1982) and continuous data (Gaussian Random Field; Laurit-
zen, 1996) and their application to psychometric data is currently
under development (Epskamp et al., in press;van Borkulo et al.,
2014).
In our view, whether evaluative reactions are binary, categori-
cal, or continuous variables is a challenging question for research
in itself and we would urge attitude researchers to use models
developed to test the underlying distribution of evaluative reac-
tions (e.g., De Boeck, Wilson, & Acton, 2005). Connectivity of the
attitude network might provide a tentative answer of whether
evaluative reactions represent dimensions or categories. As we
discussed in the section on attitude strength, highly connected
attitude networks behave more like categories and weakly con-
nected attitude networks behave more like dimensions. It is likely
that the overall behavior of the attitude network also extends to the
behavior of the evaluative reactions, so that evaluative reactions in
a highly connected network behave more like categories and
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
14 DALEGE ET AL.
evaluative reactions in a weakly connected network behave more
like dimensions.
Another possible extension of the CAN model would be to use
time-series data to estimate individual attitude networks. The mea-
surement model that we developed in this article is only suited to
cross-sectional data but related approaches can be used to model
time-series data (Bringmann et al., 2013). While estimating net-
works on cross-sectional data provides useful information on the
general structure of the attitude networks in the population, mod-
eling individual attitude networks on time-series data provides a
powerful tool to detect individual differences in attitude structure.
Several studies indicate that individuals differ in their attitude
structure (e.g., Huskinson & Haddock, 2004;van der Pligt, de
Vries, Manstead, & van Harreveld, 2000;van der Pligt & de Vries,
1998;van Harreveld, van der Pligt, de Vries, & Andreas, 2000) but
statistical tools to model such differences have, as yet, been
lacking. It is our view that network analysis can fill this gap.
A related possibility to extend the CAN model would be to
model the sizes of the attitude networks of different individuals. It
is likely that individuals will differ in how many evaluative reac-
tions they have toward an attitude object and that therefore their
attitude networks consist of different numbers of nodes. As we
have argued here, the size of the network has implications for the
dynamics of the network. To measure, however, how many eval-
uative reactions an individual has toward an attitude object new
approaches to the assessment of attitudes have to be developed.
The Challenge to Assess Attitude Networks
Devising attitude questionnaires that are tailored to the theoret-
ical background of the CAN model in some instances contrasts
markedly with devising attitude questionnaires from a latent vari-
able perspective. For example, developing questionnaires that can
be used to asses virtually all attitudes is reasonable from a latent
variable perspective (cf., Crites et al., 1994), as items in the
questionnaires are simply indicators, but it is not necessarily rea-
sonable from a network perspective. There is no a priori reason to
assume that every attitude network consists of the same nodes. It
is, for example, very likely that some emotions are often experi-
enced toward some attitude objects (e.g., anger toward presidential
candidates), while they are virtually never experienced toward
other attitude objects (e.g., anger toward a detergent brand).
Another implication of the CAN model for the assessment of
attitudes that differs from the implications of latent variable mod-
els is that researchers should strive to measure all relevant evalu-
ative reactions because otherwise the danger arises that one mea-
sures not the whole attitude network but only parts of it. As
evaluative reactions in the CAN model represent autonomous
causal entities, omissions of relevant evaluative reactions would
decrease the validity of the model (in contrast to only decreasing
reliability in latent variable models).
To construct attitude questionnaires from a network perspective,
a theory-driven approach to questionnaire construction instead of
an empirically driven approach should be adopted (see Borsboom,
Mellenbergh, & van Heerden, 2004). A challenge in the construc-
tion of such attitude questionnaires becomes to assess all relevant
evaluative reactions in the attitude network. This approach shifts
the focus of attitude questionnaire construction from striving for
internal consistency to striving for assessing attitudes comprehen-
sively. As nodes in attitude networks represent autonomous enti-
ties, they might in some instances correlate only weakly—internal
consistency and validity of the attitude questionnaire might there-
fore be incompatible in some instances. It is therefore our view that
estimating reliability of the whole attitude questionnaire is of
limited value—it is, however, possible to assess each evaluative
reaction with several indicators. In this case, reliability can be
estimated to investigate whether the different questions tapping
one evaluative reaction are likely to indeed measure a single entity.
One way to go about constructing comprehensive attitude ques-
tionnaires might be to assess, in an open-ended questionnaire,
which evaluative reactions are most common for the attitude object
of interest.
After having constructed a comprehensive attitude question-
naire, the next challenge is to assess whether individuals differ in
the number of evaluative reactions they have. As we already
discussed here, it is likely that individuals with more knowledge of
the attitude object have larger attitude networks. Furthermore,
elaboration is also likely to increase the number of evaluative
reactions one has (cf., Tesser & Leone, 1977;Wood et al., 1995).
Assessing whether an evaluative reaction represents a relevant
node in an individual’s attitude network might be possible by
measuring how long it takes a participant to respond to an item. An
item that reflects a relevant node in an individuals’ attitude net-
work should be judged faster than an item that does not, because
subjectively important attitudinal beliefs are judged faster (van
Harreveld et al., 2000) and accessible attitudes are more easily
retrieved from memory (Fazio & Williams, 1986).
Another possibility to distinguish between irrelevant and rele-
vant nodes might be to assess the salience or importance of a given
belief (e.g., Ajzen, 1991;van der Pligt et al., 2000). The more
salient or important a given belief, the more likely it would be that
this belief represents a relevant node in the attitude network. Using
this technique to distinguish between relevant and irrelevant nodes
would be relatively straightforward, as instruments to assess the
importance of a given belief have already been developed (e.g.,
van der Pligt et al., 2000).
The distinction between irrelevant or nonexistent nodes and
relevant nodes bears some resemblance to the distinction between
attitudes and nonattitudes (Converse, 1970), which was further
developed by Fazio, Sanbonmatsu, Powell, and Kardes (1986).
They argued that the distinction between attitudes and nonattitudes
should be seen as a continuum rather than a dichotomy. From a
network perspective, attitude networks that consist of only few
nodes that are only weakly connected lie more at the nonattitude
end of the continuum and attitude networks that consist of several
nodes that are strongly connected lie more at the attitude end of the
continuum.
Another issue in the assessment of attitude networks concerns
the boundary specification problem in network analysis (Laumann,
Marsden, & Prensky, 1989). The boundary specification problem
refers to the difficulty of deciding which entities form part of the
network. In attitude networks, it is for example, difficult to decide
whether the attitude toward a presidential candidate and the atti-
tude toward the candidate’s party represent two distinct networks
or one large network. There are, however, algorithms available that
detect so-called community structures in the data (e.g., Clauset,
Newman, & Moore, 2004;Girvan & Newman, 2002;Newman,
2004,2006;Newman & Girvan, 2004). A community structure
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
15
CAUSAL ATTITUDE NETWORK MODEL
refers to nodes that are highly interconnected and that only share
weak connections with other community structures. A possible
solution to the former example might thus be to assess both
evaluative reactions toward the presidential candidate and her
party and then apply an algorithm to detect community structures
in the data. Based on this, one could then investigate whether the
two attitudes represent two distinct networks or one large network.
A problematic aspect of applying community algorithms to
detect the boundaries of attitude networks is that researchers have
to assess a large set of items if they want to assess each variable
that might be closely connected to the attitude network. Especially,
if researchers want to focus on individual attitude networks, the set
of items can quickly become too large to assess them frequently.
However, this issue is less problematic for cross-sectional re-
search. It would be possible to administer an extensive question-
naire with the most important variables that might be related to the
attitude network to a group of individuals. Assessing the bound-
aries of attitude networks therefore currently needs to be done at
the group level instead of at the individual level.
Future Study of the CAN Model
We now turn to some possible avenues for future study of the
CAN model. Related to the implications on assessment of attitude
networks, a possibility for future study would be to model re-
sponses on implicit measurements of attitudes with the CAN
model. The currently most prominent reaction time (RT) based
measure of attitudes, the Implicit Association Test (IAT; Green-
wald, McGhee, & Schwartz, 1998), however, is not suited to
investigate attitudes from a network perspective, as it is not pos-
sible to analyze single items in this measure. Adopted variants of
the Affect Misattribution Procedure (AMP; Payne, Cheng, Govo-
run, & Stewart, 2005), on the other hand, can be used to measure
specific evaluative reactions (stereotype misperception task;
Krieglmeyer & Sherman, 2012; emotion misattribution procedure;
Rohr, Degner, & Wentura, 2015). However, different theoretical
viewpoints on responses toward implicit measures of attitudes
have different implications for which nodes should be measured
with implicit measures. Two influential models of implicit mea-
sures of attitudes are the Motivation and Opportunity as Determi-
nants of behavior (MODE) model (Fazio, 1990;Fazio & Olson,
2003b) and the Associative-Propositional Evaluation (APE) model
(Gawronski & Bodenhausen, 2006,2007,2011).
From the perspective of the MODE model, implicit measures of
attitudes are regarded as more “pure” estimates of attitudes be-
cause they limit the opportunity to conceal the attitude (Fazio &
Olson, 2003b). A central postulate of the MODE model regarding
the relation between implicit and explicit measures of attitudes
holds that the different measures show high correspondence when
the attitude object is a socially nonsensitive issue (e.g., attitudes
toward presidential candidates) and low correspondence when the
attitude object is a socially sensitive issue (e.g., prejudice). From
the perspective of the MODE model, evaluative reactions toward
sensitive issues should thus be measured using implicit measure-
ments. Measuring evaluative reactions toward nonsensitive issues,
on the other hand, would not require the use of implicit measure-
ments.
From the perspective of the APE model, responses on implicit
measures of attitudes represent affective ‘gut’ reactions and re-
sponses on explicit measures of attitudes represent deliberative
judgments that are propositional in nature (Gawronski & Boden-
hausen, 2006). What follows from this postulate is that implicit
measures would be better suited to measure evaluative reactions
that represent feelings toward an attitude object and explicit mea-
sures would be better suited to measure beliefs toward an attitude
object. To acquire a complete picture of an attitude network,
researchers thus would have to use combinations of implicit and
explicit measures.
We emphasize that arguments against conceptualizing attitudes
as latent variables also pertain to models focusing on attitudes
assessed with implicit measures. In our view, implicit measures of
attitudes do not acknowledge the inherent complexity of attitudes
as these measures treat attitudes as unidimensional continua, which
is clearly expressed by how attitudes are assessed with implicit
measures: in the four most used implicit procedures to assess
attitudes (Nosek, Hawkins, & Frazier, 2011)—IAT, AMP, Go/
No-Go Association Task (Nosek & Banaji, 2001), and evaluative
priming procedures (Fazio et al., 1995)—attitudes are reduced to a
single score, which presupposes them to be unidimensional. As we
have shown in this article, however, a conceptualization of atti-
tudes that instead treats attitudes as complex systems has many
advantages over treating attitudes as unidimensional continua.
Developing implicit measures of attitudes from a network perspec-
tive is thus likely to further the understanding of attitudes to a
greater extent than imposing the (untested) assumption that atti-
tudes are unidimensional.
As we argued in this article, the CAN model also aids the
progress of integrating connectionist models of attitudes with
empirical research on attitudes. We showed that the basic tenet of
connectionist models of attitudes—attitudes are conceived as a
product of a network of interrelated nodes—is also a realistic
conceptualization of empirical data on attitudes. A next step in
integrating connectionist models with empirical data would be to
use empirically estimated attitude networks as input for data sim-
ulation (see van Borkulo, Borsboom, Nivard, & Cramer, 2011 for
an example of how an empirically estimated network can be used
for data simulation in Netlogo; Wilensky, 1999). The advantage of
using empirically estimated attitude networks as input for data
simulation is that the connection strengths between the network
nodes are no longer based on arbitrary choices, but can be
grounded in empirical data. In addition, while the CAN model can
be used as a tool for data simulation, it can also produce empirical
predictions on the dynamic behavior of estimated attitude net-
works.
We already discussed two central empirical predictions of the
CAN model in the context of persuasion: (a) changes in a given
evaluative reaction will foremost affect closely connected evalua-
tive reactions, and (b) successful persuasion directed at highly
central evaluative nodes is more likely to result in a fundamental
change in the attitude network. A related prediction holds that
evaluative reactions with high closeness have the highest impact
on decisions. This prediction flows from the notion that evaluative
reactions with high closeness are the most influential nodes in a
network. While evaluative reactions with high closeness are not
necessarily the nodes that are directly related to the decision, they
can still have a profound influence on the decision through their
influence on the other evaluative reactions in the network.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
16 DALEGE ET AL.
While persuasion research might benefit from investigating
closeness of nodes, research on ambivalence might benefit from
investigating the connectivity of attitude networks. The connectiv-
ity of attitude networks is particularly relevant to the distinction
between potential and felt ambivalence (e.g., Newby-Clark et al.,
2002). Potential ambivalence refers to ambivalent evaluations and
felt ambivalence refers to psychological discomfort resulting from
potential ambivalence. Some factors that influence whether poten-
tial ambivalence results in felt ambivalence have been identified.
For example, simultaneous accessibility of ambivalent evaluations
heightens discomfort resulting from ambivalence (Newby-Clark et
al., 2002) and ambivalent attitudes that are relevant for decision-
making are likely to cause aversive feelings (van Harreveld,
Rutjens et al., 2009;van Harreveld, van der Pligt et al., 2009). A
general framework to integrate these different findings, however,
has not yet been proposed. The CAN model might provide such a
framework.
To see this, it is important to consider the notion of attractor
states. Attractor states refer to states that a dynamical system is
driven to (Alligood, Sauer, Yorke, & Crawford, 1997). Which
attractor states exist in an attitude network likely depends on the
connectivity of the network. Attitude networks that are highly
connected probably display only two attractor states (positive or
negative) because the strong connections between the nodes force
all nodes to align to the same state. In a weakly connected network,
several attractor states are likely to exist, as only some nodes have
to settle in the same state.
6
The drive of the attitude network to settle in an attractor state
might reflect aversive feelings caused by ambivalence. Highly
connected attitude networks would thus be more likely to cause
aversive feelings in the light of ambivalence than weakly con-
nected attitude networks. This is because highly connected net-
works must settle in one of only two attractor states, while weakly
connected networks can settle in one of several attractor states.
Factors that heighten the probability of felt ambivalence might
thus be integrated by proposing that they cause stronger connec-
tivity in attitude networks.
Another research area, in which network analysis and the CAN
model in particular might provide new insights, is research on the
development of attitudes over time. This area has received some-
what limited attention and the CAN model might aid progress in
this area as network analysis is well suited to model developmental
changes (Bringmann et al., 2013). A central postulate of the CAN
model is that attitude networks grow (i.e., new evaluative reactions
attach to older evaluative reactions). This growth might depend on
how often an individual elaborates on the attitude (Monroe &
Read, 2008). As we already discussed here, the CAN model
predicts that the structure of attitude growth will be driven by the
similarity and popularity of evaluative reactions.
Another issue in research on attitudes that might benefit from
investigating attitudes from the perspective of the CAN model is
the question of to what extent attitudes are formed through envi-
ronmental factors and genetic factors, and how these factors inter-
act in the shaping of attitudes. While studies investigating influ-
ences on attitudes generally have focused on environmental
influences (e.g., persuasion), genetic influences on attitudes have
also been observed (Eaves, Eysenck, & Martin, 1989;Olson,
Vernon, Harris, & Jang, 2001;Tesser, 1993).
Research on personality has shown that different genes influ-
ence different nodes in a network to a different extent (Cramer et
al., 2012) and it is likely that a similar pattern exists for nodes in
attitude networks. For example, fearful reactions to attitude objects
might have a stronger genetic basis than other nodes in attitude
networks because conservative attitudes were shown to be pre-
dicted by individual differences in arousal in response to threat-
ening stimuli (Oxley et al., 2008), which in turn are predicted by
amygdala activation (e.g., Larson et al., 2006), while amygdala
activation was found to have a genetic basis (e.g., Hariri et al.,
2002). A possible mechanism of how a conservative attitude
network might develop is that an individual, who is genetically
predisposed to fearful reactions, reacts fearfully to threats to social
order and therefore is more vulnerable to persuasion attempts that
promote that political decisions foremost must guarantee safety to
social order. Network analysis would aid in identifying such pro-
cesses.
Conclusion
In this article we introduced the CAN model and argued that (a)
the Ising model provides a psychometrically realistic formalization
of attitudes, (b) attitude networks conform to a small-world struc-
ture, and (c) network connectivity provides a mathematically for-
malized conceptualization of attitude strength. The CAN model
shows promise in integrating existing findings, which are hitherto
disparate, and in furthering the development of new insights in
several areas of attitude research. For these reasons, the CAN
model is likely to make significant contributions to the integration
of different areas of attitude research, and to further new insights
into the complex concept of attitude.
6
This point is also illustrated in the section on attitude strength. As can
be seen in Table 2, configurations of the attitude network, in which not all
evaluative reactions are aligned, are more probable in weakly connected
networks than in highly connected networks.
References
Abelson, R. P. (1995). Attitude extremity. In R. E. Petty & J. A. Krosnick
(Eds.), Attitude strength: Antecedents and consequences (pp. 25–41).
Hillsdale, NJ: Lawrence Erlbaum.
Achard, S., Salvador, R., Whitcher, B., Suckling, J., & Bullmore, E.
(2006). A resilient, low-frequency, small-world human brain functional
network with highly connected association cortical hubs. The Journal of
Neuroscience, 26, 63–72. http://dx.doi.org/10.1523/JNEUROSCI.3874-
05.2006
Ajzen, I. (1991). The theory of planned behavior. Organizational Behavior
and Human Decision Processes, 50, 179–211. http://dx.doi.org/10.1016/
0749-5978(91)90020-T
Albert, R., & Barabási, A.-L. (2002). Statistical mechanics of complex
networks. Reviews of Modern Physics, 74, 47–97. http://dx.doi.org/10
.1103/RevModPhys.74.47
Alligood, K. T., Sauer, T. D., Yorke, J. A., & Crawford, J. D. (1997).
Chaos: An introduction to dynamical systems. New York, NY: Springer.
http://dx.doi.org/10.1007/978-3-642-59281-2
Armitage, C. J., & Conner, M. (2000). Attitudinal ambivalence: A test of
three key hypotheses. Personality and Social Psychology Bulletin, 26,
1421–1432. http://dx.doi.org/10.1177/0146167200263009
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
17
CAUSAL ATTITUDE NETWORK MODEL
Bagozzi, R. P. (1981). Attitudes, intentions, and behavior: A test of some
key hypotheses. Journal of Personality and Social Psychology, 41,
607–627. http://dx.doi.org/10.1037/0022-3514.41.4.607
Bagozzi, R. P., & Burnkrant, R. E. (1979). Attitude organization and the
attitude-behavior relationship. Journal of Personality and Social Psy-
chology, 37, 913–929. http://dx.doi.org/10.1037/0022-3514.37.6.913
Bagozzi, R. P., Tybout, A. M., Craig, C. S., & Sternthal, B. (1979). The
construct validity of the tripartite classification of attitudes. Journal of
Marketing Research, 16, 88–95. http://dx.doi.org/10.2307/3150879
Barabási, A.-L. (2011). The network takeover. Nature Physics, 8, 14–16.
http://dx.doi.org/10.1038/nphys2188
Barabási, A.-L., & Albert, R. (1999). Emergence of scaling in random
networks. Science, 286, 509–512. http://dx.doi.org/10.1126/science.286
.5439.509
Bargh, J. A., Chaiken, S., Govender, R., & Pratto, F. (1992). The generality
of the automatic attitude activation effect. Journal of Personality and
Social Psychology, 62, 893–912. http://dx.doi.org/10.1037/0022-3514
.62.6.893
Bassili, J. N. (1996). Meta-judgmental versus operative indexes of psycho-
logical attributes: The case of measures of attitude strength. Journal of
Personality and Social Psychology, 71, 637–653. http://dx.doi.org/10
.1037/0022-3514.71.4.637
Bollen, K. (1989). Structural equations with latent variables. New York,
NY: Wiley. http://dx.doi.org/10.1002/9781118619179
Bollen, K., & Lennox, R. (1991). Conventional wisdom on measurement:
A structural equation perspective. Psychological Bulletin, 110, 305–314.
http://dx.doi.org/10.1037/0033-2909.110.2.305
Boninger, D. S., Krosnick, J. A., & Berent, M. K. (1995). Origins of
attitude importance: Self-interest, social identification, and value rele-
vance. Journal of Personality and Social Psychology, 68, 61–80. http://
dx.doi.org/10.1037/0022-3514.68.1.61
Boninger, D. S., Krosnick, J. A., Berent, M. K., & Fabrigar, L. R. (1995).
The causes and consequences of attitude importance. In R. E. Petty &
J. A. Krosnick (Eds.), Attitude strength: Antecedents and consequences
(pp. 159–189). Hillsdale, NJ: Lawrence Erlbaum.
Borgatti, S. P., Mehra, A., Brass, D. J., & Labianca, G. (2009). Network
analysis in the social sciences. Science, 323, 892–895. http://dx.doi.org/
10.1126/science.1165821
Borsboom, D. (2008). Psychometric perspectives on diagnostic systems.
Journal of Clinical Psychology, 64, 1089–1108. http://dx.doi.org/10
.1002/jclp.20503
Borsboom, D., & Cramer, A. O. J. (2013). Network analysis: An integra-
tive approach to the structure of psychopathology. Annual Review of
Clinical Psychology, 9, 91–121. http://dx.doi.org/10.1146/annurev-
clinpsy-050212-185608
Borsboom, D., Cramer, A. O. J., Schmittmann, V. D., Epskamp, S., &
Waldorp, L. J. (2011). The small world of psychopathology. PLoS ONE,
6, e27407. http://dx.doi.org/10.1371/journal.pone.0027407
Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2003). The theo-
retical status of latent variables. Psychological Review, 110, 203–219.
http://dx.doi.org/10.1037/0033-295X.110.2.203
Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2004). The concept
of validity. Psychological Review, 111, 1061–1071. http://dx.doi.org/10
.1037/0033-295X.111.4.1061
Breckler, S. J. (1984). Empirical validation of affect, behavior, and cog-
nition as distinct components of attitude. Journal of Personality and
Social Psychology, 47, 1191–1205. http://dx.doi.org/10.1037/0022-3514
.47.6.1191
Bringmann, L. F., Vissers, N., Wichers, M., Geschwind, N., Kuppens, P.,
Peeters, F.,...Tuerlinckx, F. (2013). A network approach to psycho-
pathology: New insights into clinical longitudinal data. PLoS ONE, 8,
e60188.
Cacioppo, J. T., & Berntson, G. G. (1994). Relationship between attitudes
and evaluative space: A critical review, with emphasis on the separa-
bility of positive and negative substrates. Psychological Bulletin, 115,
401–423. http://dx.doi.org/10.1037/0033-2909.115.3.401
Cacioppo, J. T., Petty, R. E., & Green, T. R. (1989). From the tripartite
model to the homeostasis model of attitudes. In A. R. Pratkanis, S. J.
Breckler, & A. G. Greenwald (Eds.), Attitude structure and function (pp.
275–310). Hillsdale, NJ: Lawrence Erlbaum.
Cantril, H. (1946). The intensity of an attitude. The Journal of Abnormal
and Social Psychology, 41, 129–135. http://dx.doi.org/10.1037/
h0061991
Chaiken, S., Liberman, A., & Eagly, A. H. (1989). Heuristic and systematic
processing within and beyond the persuasion context. In J. S. Uleman &
J. A. Bargh (Eds.), Unintended thought (pp. 212–252). New York, NY:
Guilford Press.
Chaiken, S., Pomerantz, E. M., & Giner-Sorolla, R. (1995). Structural
consistency and attitude strength. In R. E. Petty & J. A. Krosnick (Eds.),
Attitude strength: Antecedents and consequences (pp. 387–412). Hills-
dale, NJ: Lawrence Erlbaum.
Chen, J., & Chen, Z. (2008). Extended Bayesian information criteria for
model selection with large model spaces. Biometrika, 95, 759–771.
http://dx.doi.org/10.1093/biomet/asn034
Clauset, A., Newman, M. E. J., & Moore, C. (2004). Finding community
structure in very large networks. Physical Review E: Statistical, Non-
linear, and Soft Matter Physics, 70, 066111. http://dx.doi.org/10.1103/
PhysRevE.70.066111
Conner, M., Godin, G., Sheeran, P., & Germain, M. (2013). Some feelings
are more important: Cognitive attitudes, affective attitudes, anticipated
affect, and blood donation. Health Psychology, 32, 264–272. http://dx
.doi.org/10.1037/a0028500
Conner, M., & Sparks, P. (2002). Ambivalence and attitudes. European
Review of Social Psychology, 12, 37–70. http://dx.doi.org/10.1080/
14792772143000012
Converse, P. E. (1970). Attitudes and non-attitudes: Continuation of a
dialogue. In E. R. Tufte (Ed.), The quantative analysis of social prob-
lems (pp. 168–189). Reading, MA: Addison Wesley.
Costantini, G., Epskamp, S., Borsboom, D., Perugini, M., Mõttus, R., Wal-
dorp, L. J., & Cramer, A. O. J. (2015). State of the aRt personality research:
AtutorialonnetworkanalysisofpersonalitydatainR.Journal of Research
in Personality, 54, 13–29. http://dx.doi.org/10.1016/j.jrp.2014.07.003
Cramer, A. O. J., Giltay, E. J., van Borkulo, C. D., Kendler, K. S., van der
Maas, H. L. J., Scheffer, M., & Borsboom, D. (2013). I feel sad therefore
I don’t sleep: Major depression as a complex dynamical system. Man-
uscript submitted for publication.
Cramer, A. O. J., van der Sluis, S., Noordhof, A., Wichers, M., Geschwind,
N., Aggen, S. H.,...Borsboom, D. (2012). Dimensions of normal
personality as networks in search of equilibrium: You can’t like parties
if you don’t like people. European Journal of Personality, 26, 414431.
Cramer, A. O. J., Waldorp, L. J., van der Maas, H. L. J., & Borsboom, D.
(2010). Comorbidity: A network perspective. Behavioral and Brain
Sciences, 33, 137–150. http://dx.doi.org/10.1017/S0140525X09991567
Crites, S. L., Fabrigar, L. R., & Petty, R. E. (1994). Measuring the affective
and cognitive properties of attitudes: Conceptual and methodological
issues. Personality and Social Psychology Bulletin, 20, 619634. http://
dx.doi.org/10.1177/0146167294206001
Csárdi, G., & Nepusz, T. (2006). The igraph software package for complex
network research. Inter Journal Complex Systems, 1695. Retrieved from
http://www.interjournal.org/
Dalege, J., Borsboom, D., van Harreveld, F., & van der Maas, H. L. J.
(2015). A network perspective on attitude strength: Testing the connec-
tivity hypothesis. Manuscript in preparation.
De Boeck, P., Wilson, M., & Acton, G. S. (2005). A conceptual and
psychometric framework for distinguishing categories and dimensions.
Psychological Review, 112, 129–158. http://dx.doi.org/10.1037/0033-
295X.112.1.129
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
18 DALEGE ET AL.
De Houwer, J., Thomas, S., & Baeyens, F. (2001). Associative learning of
likes and dislikes: A review of 25 years of research on human evaluative
conditioning. Psychological Bulletin, 127, 853–869. http://dx.doi.org/10
.1037/0033-2909.127.6.853
Dunne, J. A., Williams, R. J., & Martinez, N. D. (2002). Food-web
structure and network theory: The role of connectance and size. Pro-
ceedings of the National Academy of Sciences of the United States of
America, 99, 12917–12922. http://dx.doi.org/10.1073/pnas.192407699
Eagly, A. H., & Chaiken, S. (1993). The psychology of attitudes. Orlando,
FL: Harcourt Brace Jovanovich.
Eagly, A. H., & Chaiken, S. (2007). The advantages of an inclusive
definition of attitude. Social Cognition, 25, 582–602. http://dx.doi.org/
10.1521/soco.2007.25.5.582
Eaves, L. J., Eysenck, H. J., & Martin, N. G. (1989). Genes, culture and
personality: An empirical approach. San Diego, CA: Academic Press.
Epskamp, S., Cramer, A. O. J., Waldorp, L. J., Schmittmann, V. D., &
Borsboom, D. (2012). qgraph: Network visualizations of relationships in
psychometric data. Journal of Statistical Software, 48, 1–18.
Epskamp, S., Maris, G., Waldorp, L. J., & Borsboom, D. (in press).
Network psychometrics. In P. Irwing, D. Hughes, & T. Booth (Eds.),
Handbook of psychometrics. New York, NY: Wiley.
Eschleman, K. J., Bowling, N. A., & Judge, T. A. (2015). The dispositional
basis of attitudes: A replication and extension of Hepler and Albarracín
(2013). Journal of Personality and Social Psychology, 108, e1–e15.
http://dx.doi.org/10.1037/pspp0000017
Fabrigar, L. R., MacDonald, T. K., & Wegener, D. T. (2005). The structure
of attitudes. In D. Albarracín, B. T. Johnson, & M. P. Zanna (Eds.), The
handbook of attitudes (pp. 79–125). Mahwah, NJ: Erlbaum.
Fazio, R. H. (1990). Multiple processes by which attitudes guide behavior:
The MODE model as an integrative framework. Advances in Experi-
mental Social Psychology, 23, 75–109. http://dx.doi.org/10.1016/S0065-
2601(08)60318-4
Fazio, R. H. (1995). Attitudes as object-evaluation associations: Determi-
nants, consequences, and correlates of attitude accessibility. In R. E.
Petty & J. A. Krosnick (Eds.), Attitude strength: Antecedents and con-
sequences (pp. 247–282). Hillsdale, NJ.
Fazio, R. H. (2007). Attitudes as object-evaluation associations of varying
strength. Social Cognition, 25, 603–637. http://dx.doi.org/10.1521/soco
.2007.25.5.603
Fazio, R. H., Jackson, J. R., Dunton, B. C., & Williams, C. J. (1995).
Variability in automatic activation as an unobtrusive measure of racial
attitudes: A bona fide pipeline? Journal of Personality and Social
Psychology, 69, 1013–1027. http://dx.doi.org/10.1037/0022-3514.69.6
.1013
Fazio, R. H., & Olson, M. A. (2003a). Attitudes: Foundations, functions,
and consequences. In M. A. Hogg & J. Cooper (Eds.), The handbook of
social psychology (pp. 139–160). London, UK: Sage.
Fazio, R. H., & Olson, M. A. (2003b). Implicit measures in social cogni-
tion. research: Their meaning and use. Annual Review of Psychology, 54,
297–327. http://dx.doi.org/10.1146/annurev.psych.54.101601.145225
Fazio, R. H., Sanbonmatsu, D. M., Powell, M. C., & Kardes, F. R. (1986).
On the automatic activation of attitudes. Journal of Personality and
Social Psychology, 50, 229–238. http://dx.doi.org/10.1037/0022-3514
.50.2.229
Fazio, R. H., & Williams, C. J. (1986). Attitude accessibility as a moder-
ator of the attitude-perception and attitude-behavior relations: An inves-
tigation of the 1984 presidential election. Journal of Personality and
Social Psychology, 51, 505–514. http://dx.doi.org/10.1037/0022-3514
.51.3.505
Ferrer-i-Cancho, R., & Solé, R. V. (2001). The small world of human
language. Proceedings of the Royal Society of London, Series B: Bio-
logical Sciences, 268, 2261–2265. http://dx.doi.org/10.1098/rspb.2001
.1800
Festinger, L. (1957). A theory of cognitive dissonance. Stanford, CA:
Stanford University Press.
Fishbein, M., & Ajzen, I. (1975). Belief, attitude, intention and behavior:
An introduction to theory and research. Reading, MA: Addison Wesley.
Fiske, S. T., Cuddy, A. J. C., Glick, P., & Xu, J. (2002). A model of (often
mixed) stereotype content: Competence and warmth respectively follow
from perceived status and competition. Journal of Personality and
Social Psychology, 82, 878–902. http://dx.doi.org/10.1037/0022-3514
.82.6.878
Flay, B. R. (1978). Catastrophe theory in social psychology: Some appli-
cations to attitudes and social behavior. Behavioral Science, 23, 335–
350. http://dx.doi.org/10.1002/bs.3830230404
Freeman, L. C. (1978). Centrality in social networks: Conceptual clarifi-
cation. Social Networks, 1, 215–239. http://dx.doi.org/10.1016/0378-
8733(78)90021-7
Friedman, J., Hastie, T., & Tibshirani, R. (2008). Sparse inverse covariance
estimation with the graphical lasso. Biostatistics, 9, 432–441. http://dx
.doi.org/10.1093/biostatistics/kxm045
Friston, K. (2009). The free-energy principle: A rough guide to the brain?
Trends in Cognitive Sciences, 13, 293–301. http://dx.doi.org/10.1016/j
.tics.2009.04.005
Friston, K. J., Daunizeau, J., Kilner, J., & Kiebel, S. J. (2010). Action and
behavior: A free-energy formulation. Biological Cybernetics, 102, 227–
260. http://dx.doi.org/10.1007/s00422-010-0364-z
Fruchterman, T., & Reingold, E. (1991). Graph drawing by force directed
placement. Software, Practice & Experience, 21, 1129–1164. http://dx
.doi.org/10.1002/spe.4380211102
Gawronski, B. (2012). Back to the future of dissonance theory: Cognitive
consistency as a core motive. Social Cognition, 30, 652–668. http://dx
.doi.org/10.1521/soco.2012.30.6.652
Gawronski, B., & Bodenhausen, G. V. (2006). Associative and proposi-
tional processes in evaluation: An integrative review of implicit and
explicit attitude change. Psychological Bulletin, 132, 692–731. http://dx
.doi.org/10.1037/0033-2909.132.5.692
Gawronski, B., & Bodenhausen, G. V. (2007). Unraveling the processes
underlying evaluation: Attitudes from the perspective of the APE model.
Social Cognition, 25, 687–717. http://dx.doi.org/10.1521/soco.2007.25
.5.687
Gawronski, B., & Bodenhausen, G. V. (2011). The associative-
propositional evaluation model: Theory, evidence, and open questions.
Advances in Experimental Social Psychology, 44, 59–127. http://dx.doi
.org/10.1016/B978-0-12-385522-0.00002-0
Gawronski, B., & Strack, F. (2012). Cognitive consistency as a basic
principle of social information processing. In B. Gawronski & F. Strack
(Eds.), Cognitive consistency: A fundamental principle in social cogni-
tion (pp. 1–16). New York, NY: Guilford Press.
Gigerenzer, G., Todd, P. M., & the ABC Research Group. (1999). Simple
heuristics that make us smart. New York, NY: Oxford University Press.
Girvan, M., & Newman, M. E. (2002). Community structure in social and
biological networks. Proceedings of the National Academy of the United
States of America, 99, 7821–7826. http://dx.doi.org/10.1073/pnas
.122653799
Greenwald, A. G., McGhee, D. E., & Schwartz, J. L. (1998). Measuring
individual differences in implicit cognition: The implicit association test.
Journal of Personality and Social Psychology, 74, 1464–1480. http://
dx.doi.org/10.1037/0022-3514.74.6.1464
Gross, S. R., Holtz, R., & Miller, N. (1995). Attitude certainty. In R. E.
Petty & J. A. Krosnick (Eds.), Attitude strength: Antecedents and con-
sequences (pp. 215–245). Hillsdale, NJ: Lawrence Erlbaum.
Haddock, G., Zanna, M. P., & Esses, V. M. (1993). Assessing the structure
of prejudicial attitudes: The case of attitudes toward homosexuals.
Journal of Personality and Social Psychology, 65, 1105–1118. http://dx
.doi.org/10.1037/0022-3514.65.6.1105
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
19
CAUSAL ATTITUDE NETWORK MODEL
Hariri, A. R., Mattay, V. S., Tessitore, A., Kolachana, B., Fera, F.,
Goldman, D.,...Weinberger, D. R. (2002). Serotonin transporter
genetic variation and the response of the human amygdala. Science, 297,
400403. http://dx.doi.org/10.1126/science.1071829
Hart, W., Albarracín, D., Eagly, A. H., Brechan, I., Lindberg, M. J., &
Merrill, L. (2009). Feeling validated versus being correct: A meta-
analysis of selective exposure to information. Psychological Bulletin,
135, 555–588. http://dx.doi.org/10.1037/a0015701
Heider, F. (1946). Attitudes and cognitive organization. The Journal of
Psychology, 21, 107–112. http://dx.doi.org/10.1080/00223980.1946
.9917275
Heider, F. (1958). The psychology of cognitive organization. Hoboken, NJ:
Wiley.
Hepler, J., & Albarracín, D. (2013). Attitudes without objects: Evidence for
a dispositional attitude, its measurement, and its consequences. Journal
of Personality and Social Psychology, 104, 1060–1076. http://dx.doi
.org/10.1037/a0032282
Herzog, H. A., Jr. (1993). “The movement is my life:” The psychology of
animal rights activism. Journal of Social Issues, 49, 103–119. http://dx
.doi.org/10.1111/j.1540-4560.1993.tb00911.x
Holland, R. W., Verplanken, B., & van Knippenberg, A. (2002). On the
nature of attitude-behavior relations: The strong guide, the weak follow.
European Journal of Social Psychology, 32, 869876. http://dx.doi.org/
10.1002/ejsp.135
Holyoak, K. J., & Simon, D. (1999). Bidirectional reasoning in decision
making by constraint satisfaction. Journal of Experimental Psychology:
General, 128, 3–31. http://dx.doi.org/10.1037/0096-3445.128.1.3
Houston, D. A., & Fazio, R. H. (1989). Biased processing as a function of
attitude accessibility: Making objective judgments subjectively. Social
Cognition, 7, 51–66. http://dx.doi.org/10.1521/soco.1989.7.1.51
Humphries, M. D., & Gurney, K. (2008). Network ‘small-world-ness’: A
quantitative method for determining canonical network equivalence.
PLoS ONE, 3, e0002051. http://dx.doi.org/10.1371/journal.pone
.0002051
Huskinson, T. L. H., & Haddock, G. (2004). Individual differences in
attitude structure: Variance in the chronic reliance on affective and
cognitive information. Journal of Experimental Social Psychology, 40,
82–90. http://dx.doi.org/10.1016/S0022-1031(03)00060-X
Ising, E. (1925). Beitrag zur Theorie des Ferromagnetismus [Contribution
to the theory of ferromagnetism.]. Zeitschrift fur Physik, 31, 253–258.
http://dx.doi.org/10.1007/BF02980577
Jonas, K., Diehl, M., & Bromer, P. (1997). Effects of attitudinal ambiva-
lence on information processing and attitude-intention consistency.
Journal of Experimental Social Psychology, 33, 190–210. http://dx.doi
.org/10.1006/jesp.1996.1317
Jones, C. R., Olson, M. A., & Fazio, R. H. (2010). Evaluative conditioning:
The “how” question. Advances in Experimental Social Psychology, 43,
205–255. http://dx.doi.org/10.1016/S0065-2601(10)43005-1
Jöreskog, K. G. (1971). Statistical analysis of sets of congeneric tests.
Psychometrika, 36, 109–133. http://dx.doi.org/10.1007/BF02291393
Kindermann, R., & Snell, J. L. (1980). Markov random fields and their
applications. Providence, RI: American Mathematics Society. http://dx
.doi.org/10.1090/conm/001
Kothandapani, V. (1971). Validation of feeling, belief, and intention to act
as three components of attitude and their contribution to prediction of
contraceptive behavior. Journal of Personality and Social Psychology,
19, 321–333. http://dx.doi.org/10.1037/h0031448
Krieglmeyer, R., & Sherman, J. W. (2012). Disentangling stereotype
activation and stereotype application in the stereotype misperception
task. Journal of Personality and Social Psychology, 103, 205–224.
http://dx.doi.org/10.1037/a0028764
Krosnick, J. A., Boninger, D. S., Chuang, Y. C., Berent, M. K., & Carnot,
C. G. (1993). Attitude strength: One construct or many related con-
structs? Journal of Personality and Social Psychology, 65, 1132–1151.
http://dx.doi.org/10.1037/0022-3514.65.6.1132
Krosnick, J. A., & Petty, R. E. (1995). Attitude strength: An overview. In
R. E. Petty & J. A. Krosnick (Eds.), Attitude strength: Antecedents and
consequences (pp. 1–24). Hillsdale, NJ: Erlbaum.
Larson, C. L., Schaefer, H. S., Siegle, G. J., Jackson, C. A. B., Anderle,
M. J., & Davidson, R. J. (2006). Fear is fast in phobic individuals:
Amygdala activation in response to fear-relevant stimuli. Biological
Psychiatry, 60, 410417. http://dx.doi.org/10.1016/j.biopsych.2006.03
.079
Latané, B., & Nowak, A. (1994). Attitudes as catastrophes: From dimen-
sions to categories with increasing involvement. In R. R. Vallacher & A.
Nowak (Eds.), Dynamical systems in social psychology (pp. 219–249).
San Diego, CA: Academic Press.
Laumann, E. O., Marsden, P. V., & Prensky, D. (1989). The boundary
specification problem in network analysis. In L. C. Freeman, D. R.
White, & A. K. Romney (Eds.), Methods in social network analysis (pp.
61–88). New Brunswick, NJ: Transaction Publishers.
Lauritzen, S. L. (1996). Graphical models. Oxford, UK: Clarendon Press.
Liu, J. H., & Latané, B. (1998). The catastrophic link between the impor-
tance and extremity of political attitudes. Political Behavior, 20, 105–
126. http://dx.doi.org/10.1023/A:1024828729174
Monroe, B. M., & Read, S. J. (2008). A general connectionist model of
attitude structure and change: The ACS (Attitudes as Constraint Satis-
faction) model. Psychological Review, 115, 733–759. http://dx.doi.org/
10.1037/0033-295X.115.3.733
Newby-Clark, I. R., McGregor, I., & Zanna, M. P. (2002). Thinking and
caring about cognitive inconsistency: When and for whom does attitu-
dinal ambivalence feel uncomfortable? Journal of Personality and So-
cial Psychology, 82, 157–166.
Newman, M. E. J. (2004). Fast algorithm for detecting community struc-
ture in networks. Physical Review E: Statistical, Nonlinear, and Soft
Matter Physics, 69, 066133. http://dx.doi.org/10.1103/PhysRevE.69
.066133
Newman, M. E. J. (2006). Modularity and community structure in net-
works. Proceedings of the National Academy of the United States of
America, 103, 8577–8582. http://dx.doi.org/10.1073/pnas.0601602103
Newman, M. E. J., & Girvan, M. (2004). Finding and evaluating commu-
nity structure in networks. Physical Review E: Statistical, Nonlinear,
and Soft Matter Physics, 69, 026113. http://dx.doi.org/10.1103/
PhysRevE.69.026113
Nordgren, L. F., van Harreveld, F., & van der Pligt, J. (2006). Ambiva-
lence, discomfort, and motivated information processing. Journal of
Experimental Social Psychology, 42, 252–258. http://dx.doi.org/10
.1016/j.jesp.2005.04.004
Nosek, B. A., & Banaji, M. R. (2001). The go/no-go association task.
Social Cognition, 19, 625–666. http://dx.doi.org/10.1521/soco.19.6.625
.20886
Nosek, B. A., Hawkins, C. B., & Frazier, R. S. (2011). Implicit social
cognition: From measures to mechanisms. Trends in Cognitive Sciences,
15, 152–159. http://dx.doi.org/10.1016/j.tics.2011.01.005
Olson, J. M., Vernon, P. A., Harris, J. A., & Jang, K. L. (2001). The
heritability of attitudes: A study of twins. Journal of Personality and
Social Psychology, 80, 845–860.
Opsahl, T., Agneessens, F., & Skvoretz, J. (2010). Node centrality in
weighted networks: Generalizing degree and shortest paths. Social Net-
works, 32, 245–251. http://dx.doi.org/10.1016/j.socnet.2010.03.006
Ostrom, T. M. (1969). The relationship between the affective, behavioral,
and cognitive components of attitude. Journal of Experimental Social
Psychology, 5, 12–30. http://dx.doi.org/10.1016/0022-1031(69)90003-1
Oxley, D. R., Smith, K. B., Alford, J. R., Hibbing, M. V., Miller, J. L.,
Scalora, M.,...Hibbing, J. R. (2008). Political attitudes vary with
physiological traits. Science, 321, 1667–1670. http://dx.doi.org/10.1126/
science.1157627
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
20 DALEGE ET AL.
Papadopoulos, F., Kitsak, M., Serrano, M. Á., Boguñá, M., & Krioukov, D.
(2012). Popularity versus similarity in growing networks. Nature, 489,
537–540. http://dx.doi.org/10.1038/nature11459
Payne, B. K., Cheng, C. M., Govorun, O., & Stewart, B. D. (2005). An
inkblot for attitudes: Affect misattribution as implicit measurement.
Journal of Personality and Social Psychology, 89, 277–293. http://dx
.doi.org/10.1037/0022-3514.89.3.277
Payne, J. W., Bettman, J. R., & Johnson, E. J. (1988). Adaptive strategy
selection in decision making. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 14, 534–552. http://dx.doi.org/10
.1037/0278-7393.14.3.534
Payne, J. W., Bettman, J. R., & Johnson, E. J. (1993). The adaptive
decision maker. New York, NY: Cambridge University Press. http://dx
.doi.org/10.1017/CBO9781139173933
Petrocelli, J. V., Tormala, Z. L., & Rucker, D. D. (2007). Unpacking
attitude certainty: Attitude clarity and attitude correctness. Journal of
Personality and Social Psychology, 92, 3041. http://dx.doi.org/10
.1037/0022-3514.92.1.30
Petty, R. E., & Cacioppo, J. T. (1986). The elaboration likelihood model of
persuasion. Advances in Experimental Social Psychology, 19, 123–205.
http://dx.doi.org/10.1016/S0065-2601(08)60214-2
Petty, R. E., Haugtvedt, C. P., & Smith, S. M. (1995). Elaboration as a
determinant of attitude strength: Creating attitudes that are persistent,
resistant, and predictive of behavior. In R. E. Petty & J. A. Krosnick
(Eds.), Attitude strength: Antecedents and consequences (pp. 93–130).
Hillsdale, NJ: Lawrence Erlbaum.
Prislin, R. (1996). Attitude stability and attitude strength: One is enough to
make it stable. European Journal of Social Psychology, 26, 447–477.
http://dx.doi.org/10.1002/(SICI)1099-0992(199605)26:3'447::AID-
EJSP768(3.0.CO;2-I
Rohr, M., Degner, J., & Wentura, D. (2015). The “emotion misattribution”
procedure: Processing beyond good and bad under masked and un-
masked presentation conditions. Cognition and Emotion, 29, 196–219.
http://dx.doi.org/10.1080/02699931.2014.898613
Rosenberg, M. J., Hovland, C. I., McGuire, W. J., Abelson, R. P., &
Brehm, J. W. (1960). Attitude organization and change: An analysis of
consistency among attitude components. New Haven, CT: Yale Univer-
sity Press.
Roskos-Ewoldsen, D. R., & Fazio, R. H. (1992). On the orienting value of
attitudes: Attitude accessibility as a determinant of an object’s attraction
of visual attention. Journal of Personality and Social Psychology, 63,
198–211. http://dx.doi.org/10.1037/0022-3514.63.2.198
Schimmack, U., & Diener, E. (1997). Affect intensity: Separating intensity
and frequency in repeatedly measured affect. Journal of Personality and
Social Psychology, 73, 1313–1329. http://dx.doi.org/10.1037/0022-3514
.73.6.1313
Schmittmann, V. D., Cramer, A. O. J., Waldorp, L. J., Epskamp, S., Kievit,
R. A., & Borsboom, D. (2013). Deconstructing the construct: A network
perspective on psychological phenomena. New Ideas in Psychology, 31,
43–53. http://dx.doi.org/10.1016/j.newideapsych.2011.02.007
Shultz, T. R., & Lepper, M. R. (1996). Cognitive dissonance reduction as
constraint satisfaction. Psychological Review, 103, 219–240. http://dx
.doi.org/10.1037/0033-295X.103.2.219
Simon, D., & Holyoak, K. J. (2002). Structural dynamics of cognition:
From consistency theories to constraint satisfaction. Personality and
Social Psychology Review, 6, 283–294. http://dx.doi.org/10.1207/
S15327957PSPR0604_03
Simon, D., Krawczyk, D. C., & Holyoak, K. J. (2004). Construction of
preferences by constraint satisfaction. Psychological Science, 15, 331–
336. http://dx.doi.org/10.1111/j.0956-7976.2004.00678.x
Simon, D., Snow, C. J., & Read, S. J. (2004). The redux of cognitive
consistency theories: Evidence judgments by constraint satisfaction.
Journal of Personality and Social Psychology, 86, 814837. http://dx
.doi.org/10.1037/0022-3514.86.6.814
Sparks, P., Conner, M., James, R., Shepherd, R., & Povey, R. (2001).
Ambivalence about health-related behaviours: An exploration in the
domain of food choice. British Journal of Health Psychology, 6, 53–68.
http://dx.doi.org/10.1348/135910701169052
Stephan, K. E., Hilgetag, C. C., Burns, G. A., O’Neill, M. A., Young,
M. P., & Kötter, R. (2000). Computational analysis of functional con-
nectivity between areas of primate cerebral cortex. Biological Sciences,
355, 111–126. http://dx.doi.org/10.1098/rstb.2000.0552
Tesser, A. (1993). The importance of heritability in psychological research:
The case of attitudes. Psychological Review, 100, 129–142. http://dx.doi
.org/10.1037/0033-295X.100.1.129
Tesser, A., & Leone, C. (1977). Cognitive schemas and thought as deter-
minants of attitude change. Journal of Experimental Social Psychology,
13, 340–356. http://dx.doi.org/10.1016/0022-1031(77)90004-X
Tesser, A., & Shaffer, D. R. (1990). Attitudes and attitude change. Annual
Review of Psychology, 41, 479–523. http://dx.doi.org/10.1146/annurev
.ps.41.020190.002403
Thompson, M. M., Zanna, M. P., & Griffin, D. W. (1995). Let’s not be
indifferent about (attitudinal) ambivalence. In R. E. Petty & J. A.
Krosnick (Eds.), Attitude strength: Antecedents and consequences (pp.
361–386). Hillsdale, NJ: Lawrence Erlbaum.
Tibshirani, R. (1996). Regression selection and shrinkage via the Lasso.
Journal of the Royal Statistical Society Series B. Methodological, 58,
267–288.
van Borkulo, C. D., Borsboom, D., Epskamp, S., Blanken, T. F., Boschloo,
L., Schoevers, R. A., & Waldorp, L. J. (2014). A new method for
constructing networks from binary data. Scientific Reports, 4, 5918.
http://dx.doi.org/10.1038/srep05918
van Borkulo, C. D., Borsboom, D., Nivard, M. G., & Cramer, A. O. J.
(2011). NetLogo Symptom Spread model. Evanston, IL: Center for
Connected Learning and Computer-Based Modeling, Northwestern Uni-
versity. Retrieved from http://ccl.northwestern.edu/netlogo/models/
SymptomSpread
van den Berg, H., Manstead, A. S. R., van der Pligt, J., & Wigboldus,
D. H. J. (2005). The role of affect in attitudes toward organ donation and
donor-relevant decisions. Psychology & Health, 20, 789802. http://dx
.doi.org/10.1080/14768320500161786
van der Maas, H. L. J., Dolan, C. V., Grasman, R. P. P. P., Wicherts, J. M.,
Huizenga, H. M., & Raijmakers, M. E. J. (2006). A dynamical model of
general intelligence: The positive manifold of intelligence by mutualism.
Psychological Review, 113, 842–861. http://dx.doi.org/10.1037/0033-
295X.113.4.842
van der Maas, H. L. J., Kolstein, R., & van der Pligt, J. (2003). Sudden
transitions in attitudes. Sociological Methods & Research, 32, 125–152.
http://dx.doi.org/10.1177/0049124103253773
van der Pligt, J., & de Vries, N. K. (1998). Expectancy-value models of
health behaviour: The role of salience and anticipated affect. Psychology
& Health, 13, 289–305. http://dx.doi.org/10.1080/08870449808406752
van der Pligt, J., de Vries, N. K., Manstead, A. S. R., & van Harreveld,
F. (2000). The importance of being selective: Weighing the role of
attribute importance in attitudinal judgment. Advances in Experimen-
tal Social Psychology, 32, 135–200. http://dx.doi.org/10.1016/
S0065-2601(00)80005-2
van Harreveld, F., Nohlen, H. U., & Schneider, I. K. (2015). The ABC of
ambivalence: Affective, Behavioral and Cognitive consequences of at-
titudinal conflict. Advances in Experimental Social Psychology, 52,
285–324. http://dx.doi.org/10.1016/bs.aesp.2015.01.002
van Harreveld, F., Rutjens, B. T., Rotteveel, M., Nordgren, L. F., & van der
Pligt, J. (2009). Ambivalence and decisional conflict as a cause of
psychological discomfort: Feeling tense before jumping off the fence.
Journal of Experimental Social Psychology, 45, 167–173. http://dx.doi
.org/10.1016/j.jesp.2008.08.015
van Harreveld, F., van der Pligt, J., & de Liver, Y. N. (2009). The agony
of ambivalence and ways to resolve it: Introducing the MAID model.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
21
CAUSAL ATTITUDE NETWORK MODEL
Personality and Social Psychology Review, 13, 45–61. http://dx.doi.org/
10.1177/1088868308324518
van Harreveld, F., van der Pligt, J., de Vries, N. K., & Andreas, S. (2000).
The structure of attitudes: Attribute importance, accessibility and judg-
ment. British Journal of Social Psychology, 39, 363–380. http://dx.doi
.org/10.1348/014466600164543
van Harreveld, F., van der Pligt, J., de Vries, N. K., Wenneker, C., &
Verhue, D. (2004). Ambivalence and information integration in attitu-
dinal judgment. British Journal of Social Psychology, 43, 431–447.
http://dx.doi.org/10.1348/0144666042037971
Van Overwalle, F., & Siebler, F. (2005). A connectionist model of attitude
formation and change. Personality and Social Psychology Review, 9,
231–274. http://dx.doi.org/10.1207/s15327957pspr0903_3
Visser, P. S., Bizer, G. Y., & Krosnick, J. A. (2006). Exploring the latent
structure of strength-related attitude attributes. Advances in Experimen-
tal Social Psychology, 38, 1–67. http://dx.doi.org/10.1016/S0065-
2601(06)38001-X
Visser, P. S., & Krosnick, J. A. (1998). Development of attitude strength over
the life cycle: Surge and decline. Journal of Personality and Social Psy-
chology, 75, 1389–1410. http://dx.doi.org/10.1037/0022-3514.75.6.1389
Watts, D. J. (1999). Networks, dynamics, and the small-world phenomen.
American Journal of Sociology, 105, 493–527. http://dx.doi.org/10
.1086/210318
Watts, D. J., & Strogatz, S. H. (1998). Collective dynamics of ‘small-
world’ networks. Nature, 393, 440442. http://dx.doi.org/10.1038/
30918
West, D. B. (1996). Introduction to graph theory. Upper Saddle River, NJ:
Prentice Hall.
Westaby, J. D., Pfaff, D. L., & Redding, N. (2014). Psychology and social
networks: A dynamic network theory perspective. American Psycholo-
gist, 69, 269–284. http://dx.doi.org/10.1037/a0036106
Wilensky, U. (1999). NetLogo. Evanston, IL: Center for Connected Learn-
ing and Computer-Based Modeling, Northwestern University. Retrieved
from http://ccl.northwestern.edu/netlogo/
Wood, W., Rhodes, N., & Biek, M. (1995). Working knowledge and
attitude strength: An information-processing analysis. In R. E. Petty &
J. A. Krosnick (Eds.), Attitude strength: Antecedents and consequences
(pp. 283–314). Hillsdale, NJ: Lawrence Erlbaum.
Wu, F. Y. (1982). The Potts model. Reviews of Modern Physics, 54,
235–268. http://dx.doi.org/10.1103/RevModPhys.54.235
Zanna, M. P., & Rempel, J. K. (1988). Attitudes: A new look at an old
concept. In D. Bar-Tal & A. W. Kruglanski (Eds.), The psychology of
knowledge (pp. 315–334). Cambridge, UK: Cambridge University Press.
Zeeman, E. C. (1976). Catastrophe theory. Scientific American, 234, 65–
83. http://dx.doi.org/10.1038/scientificamerican0476-65
Appendix
Description of the ANES 1984 Data Set
Evaluative Reactions
Sixteen evaluative reactions toward the presidential candidates
tapping beliefs and seven evaluative reactions tapping feelings
were assessed. For items tapping beliefs, participants were asked:
“In your opinion, does the phrase ‘he . . .’ describe the candidate?”
The beliefs that completed the items were “is moral,” “is knowl-
edgeable,” “is inspiring,” “would provide strong leadership,” “is
hard-working,” “is decent,” “is compassionate,” “commands re-
spect,” “is intelligent,” “is kind,” “sets a good example,” “really
cares about people like you,” “understands people like you,” “is
fair,” “is in touch with ordinary people,” and “is religious.” We
excluded the item focusing on whether the participants believed
that the candidates were religious because this belief is ambiguous
in relation to valence. Items were assessed on a 4-point scale, with
answer options 4 #extremely well,3#quite well,2#not too
well,1#not well at all. As the eLasso-procedure only allows
dichotomous data (van Borkulo et al., 2014), we scored Options 3
and 4 as 1 and Option 1 and 2 as 0.
For items tapping feelings, participants were asked: “Has the
candidatebecause of the kind of person he is or because of
something he has done, ever made you feel: . . .?” The feelings that
completed the items were “angry,” “hopeful,” afraid of him,”
proud,” “disgusted,” “sympathetic toward him,” and “uneasy.”
These items were assessed dichotomously, with the answer options
1#“Yes” and 0 #“No”.
Missing Data
Some participants had missing data on at least one of the
evaluative reactions because they responded “Don’t know.” We
deleted missing data casewise, which led to the final sample of
1,877 participants for the analysis on the attitude toward Ronald
Reagan and 1,628 participants for the analysis on the attitude
toward Walter Mondale. To check the robustness of our results, we
also imputed missing values with randomly assigning 0 or 1, when
a participant had missing values. The results of this analysis
mirrored the results reported here.
Received March 13, 2014
Revision received July 31, 2015
Accepted August 3, 2015 !
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
22 DALEGE ET AL.
... the CAN model assumes that attitude networks do show a structure with high clustering (Dalege et al., 2016). Clusters are thereby defined as groups of similar evaluative reactions that exert a stronger influence on each other compared to dissimilar evaluative reactions . ...
... However, conceptualising public attitudes towards science as a psychological network may be especially promising because psychological networks allow the visualisation and interpretation of even very complex attitude structures (Dalege et al., 2016). Due to the inherent complexity of science, it seems indeed likely that the attitude towards science is highly complex and consists of various evaluative reactions, including aspects such as scientific knowledge, scientific methodology, and the scientific community itself (e.g., Altenmüller et al., 2021;Bromme et al., 2010;Wingen et al., 2020). ...
... As can be seen in Figure 2, three of the four most central nodes were related to whether science produces valuable knowledge to tackle the Covid-19 pandemic ("Policy", "Knowledge important"), or whether people should rather use common sense instead of relying on science ("Common sense"). Changing these nodes likely has a strong impact on the overall attitude network (Dalege et al., 2016). Further, the two trust-related nodes, in particular "Trust Statement", also had a relatively high centrality. ...
Article
Full-text available
A better understanding of the public attitude towards science could be crucial to tackle the spread of mis- and disinformation related to the Covid-19 pandemic and beyond. We here contribute to this understanding by conceptualising and analysing the attitude toward science as a psychological network. For this analysis, we utilised cross-sectional data from a German probability sample (N = 1,009), the “Science Barometer”, collected during the first wave of the Covid-19 pandemic. Overall, our network analysis revealed that especially the perceived value of science for curbing the pandemic is central to the attitude towards science. Beliefs about this value are related to trust in science and trust in scientific information and to positive and negative evaluations of scientific controversy and complexity. Further, valuing common sense over science was related to seeking less scientific information on official websites, suggesting that this belief, in particular, may drive mis- and disinformation and could be a promising target for interventions. Finally, we found no evidence that seeking scientific information on social media had detrimental consequences for the attitude towards science. Implications for health communication and science communication, limitations, and future directions are discussed.
... We emphasized that the utility of any causal discovery method depends on the validity of assumptions we make about the underlying causal structure. Ironically, the more specific knowledge we have about the nature of the causal system, the less likely it becomes that the PMRF or GGM is the optimal causal discovery method to use, unless of course that system is made up of symmetric and undirected causal relationships (Cramer et al., 2016;Dalege et al., 2016). Moreover, it is important to note that alternative causal discovery tools which are built for purpose are generally likely to outperform statistical network models. ...
... Outside of their use for causal discovery, there are many attractive reasons to use statistical network models. The Ising model, for example, has been used as a theoretical toy model, describing a causal system with fully symmetric causal relationships (Cramer et al., 2016;Dalege et al., 2016). Statistical network models are also useful as descriptive tools, to explore and visualize multivariate statistical relationships; they allow for the identification of predictive relationships (Epskamp, Waldorp, et al., 2018); provide sparse descriptions of statistical dependency relationships in a multivariate density; and may be used as a variable clustering or latent variable identification method (e.g. ...
... Therefore, different from a simple combination of symptoms, it can be that elderly adults trap in a depressive mood for spasmodic waking up earlier and during the night, being unable to fall asleep deepens their depressive mood. Moreover, network analysis provides us with a new perspective to evaluate how symptoms (nodes) function and how symptoms interweave with each other (edges) (22). Hence, we can identify consequential symptoms and relationships to guide more effective treatment and interventions. ...
Article
Full-text available
Background: Besides physical changes, elderly adults are prone to have mental disorders such as anxiety, depression, and sleep disturbance, and the pandemic of COVID-19 worsened the situation. However, internal relationships and co-occurrence of psychopathologies were scarcely examined. Therefore, in the current study, through network analysis, we inspected relationships among symptoms of depression, anxiety, and sleep disturbance and identified key symptoms that espoused the disease. Methods: We asked 1302 elderly adults to fill in Patient Health Questionnaire-2 (depressive symptoms), the Generalized Anxiety Disorder-2 (anxiety symptoms), and the Youth Self-rating Insomnia Scale (sleep disturbance) and then constructed three networks for elderly adults, male elderly, and female elderly. Via network analysis, we accomplished four goals. First, we identified symptom with the highest centrality (i.e., strength) index for each network; then, we found the strongest correlation (i.e., edges) in each network; thirdly, we confirmed specific nodes that could bridge anxiety, depression, and sleep disturbance; the last was to compare networks based on genders. Network stability and accuracy tests were performed. Results. Networks of elderly adults, male elderly, and female elderly were stable, accurate, and intelligible. Among all networks, "Nervousness"- "Excessive worry" (GAD-1- GAD-2) had the strongest correlation, and "Nervousness" (GAD-1) had the highest strength and bridge strength value. When we made a comparison between female elderly's and male elderly's networks, except for the significant difference in the mean value of "Difficulty initiating sleep" (YSIS-3), the findings showed that the two networks were similar. Network stability and accuracy proved to be reliable. Conclusions: In networks of anxiety, depression, and sleep disturbance, anxiety played a conspicuous role in comorbidity, which could be a target for practical intervention and prevention.
... Alternatively to the factor approach, the network approach has recently emerged Borsboom, 2017) and is applied in various psychological domains, such as personality (Beck & Jackson, 2021), mental disorders (Fried et al., 2017), well-being (Govorova et al., 2020), emotions (Lange et al., 2020), empathy (Briganti et al., 2018), and decent work (Svicher et al., 2022). According to this approach, meaningful work and meaningful study did not necessarily stem from latent factors ; however, they may derive from the reciprocal interaction among its observable indicators (scale items) (Dalege et al., 2016). From this perspective, network analysis can provide an estimate of the centrality of single items of WAMI or WAMI-U (nodes) and determine the association between these elements assumed to be causally related (edges) . ...
Article
The present study examined the Italian version of the Work as Meaning Inventory and Work as Meaning Inventory for University students through a network perspective. Network analysis was applied to 505 Italian workers assessed via the Work as Meaning Inventory and 214 Italian university students assessed via the Work as Meaning Inventory for University students. The comparison between the factor and network model of both Work as Meaning Inventory and Work as Meaning Inventory for University students showed a higher fit for network models. Network analysis revealed that both Work as Meaning Inventory and Work as Meaning Inventory for University students items enclosed in the Positive meaning dimension had the highest centrality; the most connected item with the overall network was the item 2 (work/study that contributes to personal growth) able to link together the three dimensions (positive meaning, meaning-making through work, greater good motivations). Findings suggested that a network approach can provide a promising realm to expand the understanding of meaning at work and meaning at study.
... This would involve developing a theoretical network model of the psychological construct of interest: identifying the nodes in the network, as well the dynamic associations between nodes (e.g., how changes in one node might propagate throughout the network), and thereby developing testable (and importantly, falsifiable) hypotheses (Bringmann et al., 2022). Bringmann et al. (2022) note that there are few of these theoretical network models; examples include a network theory of how attitudes change (Dalege et al., 2016) and a computational model of panic disorder (Robinaugh et al., 2019). Testable hypotheses derived from theoretical network models can be examined through fitting statistical network models to data (potentially through "confirmatory" network models which test model structure and fit; Epskamp, Rhemtulla, & Borsboom, 2017), although also through other designs (e.g., experimental). ...
Article
Full-text available
Network analyses have become increasingly common within the field of psychology, and temporal network analyses in particular are quickly gaining traction, with many of the initial articles earning substantial interest. However, substantial heterogeneity exists within the study designs and methodology, rendering it difficult to form a comprehensive view of its application in psychology research. Since the field is quickly growing and since there have been many study-to-study variations in terms of choices made by researchers when collecting, processing, and analyzing data, we saw the need to audit this field and formulate a comprehensive view of current temporal network analyses. To systematically chart researchers’ practices when conducting temporal network analyses, we reviewed articles conducting temporal network analyses on psychological variables (published until March 2021) in the framework of a scoping review. We identified 43 articles and present the detailed results of how researchers are currently conducting temporal network analyses. A commonality across results concerns the wide variety of data collection and analytical practices, along with a lack of consistency between articles about what is reported. We use these results, along with relevant literature from the fields of ecological momentary assessment and network analysis, to formulate recommendations on what type of data is suited for temporal network analyses as well as optimal methods to preprocess and analyze data. As the field is new, we also discuss key future steps to help usher the field’s progress forward and offer a reporting checklist to help researchers navigate conducting and reporting temporal network analyses.
Article
This study aimed to analyze the relationship among different evaluative reactions of the intergroup attitudes and contact in Spanish adolescents evaluating different ethnic minorities and in immigrant-background adolescents evaluating Spanish youth. This study was based on psychosocial models of great impact in the study of intergroup relations such as the Stereotype Content Model and the Behaviors from Intergroup Affect and Stereotypes Map, and incorporated a new approach to the study of attitudes: psychological networks. In total, 1122 Spanish adolescents and 683 adolescents with an immigrant background (Moroccan, Romanian or Ecuadorian origin) participated in the study, aged from 12 to 19 years. They answered a questionnaire with measures of stereotype dimensions (morality, immorality, sociability and competence), emotions (positives and negative), behavioral tendencies (facilitation and harm) and contact (quantity and quality). The results show similar structural patterns in the six studied groups, with emotions acting as links between stereotypes and behavioral tendencies. Moreover, positive and negative stereotype dimensions appeared as independent dimensions that were part of different processes: sociability and morality, and competence to a lesser extent, were related to facilitation behaviors through positive emotions, while immorality was related to harm behaviors through negative emotions. This could indicate that, to achieve successful intergroup relations involving cooperation and the development of friendly relationships, it would be appropriate to intervene in parallel in these two pathways. Due to the centrality of positive emotions (and sociability and immorality) and, therefore, their capacity to affect the entire network, focusing interventions on these variables could be an appropriate strategy to achieve overall positive attitudes.
Article
Full-text available
While characteristics of psychosocial work environment have traditionally been studied separately, we propose an alternative approach that treats psychosocial factors as interacting elements in networks where they all potentially affect each other. In this network analysis, we used data from a prospective occupational cohort including 10,892 participants (85% women; mean age 47 years) and repeated measurements of seven psychosocial work characteristics (job demands, job control, job uncertainty, team climate, effort-reward imbalance, procedural justice and interactional justice) assessed in 2000, 2004, 2008 and 2012. Results from multilevel longitudinal vector autoregressive models indicated that job demands as well as interactional and procedural justice were most broadly associated with the subsequent perceptions of the work-related psychosocial factors (high out-Strength), suggesting these factors might be potentially efficient targets of workplace interventions. The results also suggest that modifying almost any of the studied psychosocial factors might be relevant to subsequent perceptions of effort-reward imbalance and interactional justice at the workplace.