Foreign Policy Attitudes As Networks
Joshua D. Kertzer* and Kathleen E. Powers†
Last revised: August 8, 2014
ABSTRACT: What’s systemic about foreign policy belief systems? We introduce political
scientists to a new — networked — paradigm for political attitudes, and develop a new
experimental design that disentangles the directional nature of attitude constraint while
leveraging tools from network analysis. We find that established models of foreign policy
attitudes make unrealistically strict assumptions about the directionality and uniformity
of attitude structure. Specific policy attitudes like the Iraq war play more central roles
than our existing theories give them credit for, and the topology of attitude networks varies
substantially with individual characteristics like political sophistication.
ACKNOWLEDGMENTS: This research was supported by the Behavioral Decision Making initiative at Ohio State, and benefited from presentations at Ohio State, the 2013 NYU Experiments
Conference, and the 2013 MPSA and APSA meetings. A special thanks to Robin Baidya, Jan BoxSteffensmeier, Greg Caldeira, Chris Gelpi, Brian Greenhill, Matt Hitt, George Ibrahim, Luke Keele,
Samara Klar, Vlad Kogan, Neil Malhotra, Kathleen McGraw, Joanne Miller, William Minozzi, Irfan
Nooruddin, Dave Peterson, Brian Rathbun, Jason Reifler, Jonathan Renshon, Molly Roberts, Chris
Skovron, Ben Valentino, and Ben Wagner for their helpful assistance and feedback on previous versions of this project. Authors are listed in alphabetical order.
*
Assistant Professor of Government, Harvard University. 1737 Cambridge St, Cambridge, MA 02138. Email: jkertzer@gov.harvard.edu. Web: http:/people.fas.harvard.edu/~jkertzer/.
†
PhD Candidate, Department of Political Science, The Ohio State University. 2140 Derby Hall, Columbus, OH
43210. Email: powers.276@osu.edu. Web: www.kepowers.com.
Introduction
How do foreign policy attitudes work together in a system? Ever since Converse (1964) first asked
whether the American public was “innocent of ideology,” public opinion scholars have searched
for signs of constraint in the public’s beliefs, in both domestic (Peffley and Hurwitz, 1985; Feldman,
2003) and foreign policy (Wittkopf, 1986; Hurwitz and Peffley, 1987) domains. By pointing to structural patterns that characterize how attitudes and beliefs relate to one another as part of a broader
system — what psychologists call interattitudinal structure (Judd and Krosnick, 1989; Lavine, Thomsen, and Gonzales, 1997; Dinauer and Fink, 2005) — political scientists have demonstrated that
attitudes are indeed better organized “than a bowl of cornflakes” (McGuire, 1989, 50), and that
the public perhaps deserves more credit than originally thought. Against the Almond-Lippmann
consensus, the past thirty years have produced a plethora of models examining the many different
ways in which the public’s foreign policy attitudes are connected to one another (e.g. Holsti, 1979;
Rathbun, 2007; Liberman, 2013a).
These investigations provide considerable insight into the various component relationships of
Americans’ foreign policy belief systems — demonstrating piece by piece how certain attitudes predict one another. Past work highlights individual relationships between attitudes that are theoretically and logically linked, yet they do so at the expense of researching the holistic properties that
make belief systems systemic in the first place. We seek to build upon these foundations, while developing a theoretical framework for understanding systems of foreign policy attitudes at a higher
level of analysis.
In this paper, we introduce political scientists to a new — networked — paradigm for political
attitudes and a method for investigating them. We begin by reviewing the foreign policy attitude
literature, highlighting the structural assumptions and implicit reductionism that underlie existing
work. Next, we develop a novel experimental design — a directed multigraph experiment — that dis-
1
entangles the directional nature of attitude constraint while leveraging tools from network analysis
to study belief systems systemically. These tools have been used elsewhere to yield insights about interpersonal (Sokhey and Djupe, 2011) and international (Hafner-Burton, Kahler, and Montgomery,
2009) networks, but as far as we are aware, have not been applied to the study of interattitudinal
structure.
As a “plausibility probe” for our theoretical framework, we present our experimental results,
suggesting that studying foreign policy attitudes as networks leads to three important findings. First,
we find evidence that belief systems are often driven from the bottom-up rather than the top-down:
specific policy attitudes play more of a role, and core values less of one, than existing work would
lead us to expect. Second, although IR scholars often model the public as a monolithic entity, we
find that the topology of attitude networks varies dramatically across individuals. Third, we see
that political knowledge and interest play major roles in shaping the structure of foreign policy
attitude networks: knowledge is critical in shaping the degree to which attitudes are structured,
while levels of interest in foreign policy determine whether more specific or abstract attitudes are
the most central elements of belief systems. Finally, we conclude by reviewing implications for IR
scholars, political psychologists, and public opinion scholars more broadly.
Three things we know about foreign policy attitudes
Ever since Converse (1964) claimed that many Americans were innocent of ideology, political scientists have searched for signs of “constraint”, or functional interdependence, among political attitudes.1 Although these questions implicate political attitudes more generally, we focus here specifically on foreign policy attitudes, both for reasons of tractability, and because they were traditionally
1
While political scientists invoke “constraint” in a variety of ways, sometimes referring to response (in)stability over
time (Zaller and Feldman, 1992), and other times to simple intercorrelations between issue attitudes (Nie, Verba, and
Petrocik, 1979), we specifically use constraint here to describe “functional interdependence” among attitudes (Converse,
1964, 3). We follow Eagly and Chaiken (2007, 583) in defining attitudes as “an individual’s propensity to evaluate a
particular entity with some degree of favorability or unfavorability.”
2
believed to be far less structured than their domestic counterparts (Holsti, 2004). The result is a rich
and robust research program, which has taught us at least three things about how foreign policy attitudes are structured.
First, foreign policy attitudes “go together”: evaluations of one foreign policy issue are often
related to attitudes towards other foreign policy issues. We know from Wittkopf’s (1990, 33) early
factor analyses that the same Americans who, for example, favored arms control agreements with the
Soviet Union tended to also support cultural exchanges with them. Similarly, those who prioritize
human rights support a stronger United Nations (Rathbun, 2007), and citizens heavily concerned
about defending US national security during the Cold War were likely to call for the containment
of communism (Chittick, Billingsley, and Travis, 1995). These examples demonstrate that foreign
policy attitudes are not entirely unorganized; instead, evidence suggests that preferences cluster in
predictable ways. Holding one type of foreign policy attitude makes people more likely to hold
another, related, position.
Second, specific foreign policy attitudes depend not only upon each other, but also on more
general dispositions and values: they have abstract antecedents. In some cases scholars arrive at
the antecedents inductively by estimating an exploratory factor analysis and naming the ensuing
dimensions — as in Wittkopf’s (1990) cooperative and militant internationalism. Others specify a
particular abstract attitude and use regression models to demonstrate that the variable of interest
predicts foreign policy attitudes. Moral values, for example, are related to foreign policy stances:
individuals who value authority are more likely to favor the war in Iraq, while those who value
equality support cooperation with the United Nations (Kertzer et al., 2014). Drawing from an array
of potentially influential dispositions, others show that beliefs about retributive justice (Liberman,
2013b) and human nature (Brewer and Steenbergen, 2002), as well as commitments to hierarchy or
community (Rathbun, 2007) predict a host of foreign policy attitudes. Hurwitz and Peffley (1987)
employ structural equation modeling to go further, theoretically mapping a series of relationships
3
from the most general core values to foreign policy “postures” and specific policy attitudes (defense
spending, nuclear disarmament, etc.).
Third, we know that these connections — both between policy attitudes and to more abstract
dispositions and values — are not coincidental. They are believed to tell us something about how
people form and update their political attitudes. Models of social cognition suggest not only that
how we respond to new information in the world depends heavily on information already in our
heads, but also that we form associational links between two ideas when we consider them simultaneously (Judd and Krosnick, 1989; Lodge and Taber, 2013). In other words, attitudes are neither
formed nor updated in a vacuum. We expect that individuals rely on existing ideas to shape their attitudes toward global politics, and that related attitudes will interact with one another. Hurwitz and
Peffley (1987) make this argument explicitly. According to their model, people already carry ideas
about whether or not warfare is moral, which informs their level of militarism. When they determine their position on a specific policy item, they are doing so with those beliefs in the background
shaping and informing their expressed opinion.
A key implication of previous work, then, is that attitudes are systemic: the various component
parts form a whole. We know for example that Wittkopf’s (1990) cooperative internationalism is
associated with a desire to strengthen the United Nations (Holsti and Rosenau, 1990), and that a
moral commitment to avoid harm predicts cooperative internationalism (Kertzer et al., 2014). We
can thus think of the three attitudes as embedded within a broader system — a network in which
attitudes have various, multidirectional connections to one another.
Attitudes as networks
Previous work thus suggests a level of complexity in the structure of foreign policy attitudes that
should encourage us to think about attitudes at a higher level of analysis. While past work tells us
much about specific relationships between particular attitudes, we know little about the system itself.
4
We suggest that one way to study belief systems as systems is to explicitly conceptualize attitudes as
networks and leverage tools from network analysis. This enables us to make claims about systemlevel dynamics, such as how densely connected foreign policy belief systems are, whether abstract
attitudes shape more specific positions in the top-down process described by Hurwitz and Peffley’s
(1987) structural equation model, and which types of attitudes are most likely to spark change in
other attitudes situated throughout the network.
In this respect, although our methods differ, our interests here are similar to those of operational code scholars, who study leaders’ belief systems in a fashion that values the whole as much
as its parts (Holsti, 1970; Schafer and Walker, 2006; Renshon, 2009). A leader’s operational code —
their “philosophical and instrumental beliefs about the nature and use of power in the international
system” — is composed of 10 distinct components (Renshon, 2009, 650). While knowing a leader’s
stance on one component is instructive, it does not provide sufficient insight into their decisionmaking process and offers limited scholarly value on its own. Upon understanding how each of the
10 components combine and interact as part of a system, though, scholars can explain the leader’s
approach to foreign policy events. Individual attitudes and bivariate relationships between them
are useful objects of study, too, but we argue that as in operational code analysis, their dynamic
interactions within the larger network of attitudes enable more powerful predictions.
Public opinion scholars have done much to connect the foreign policy public opinion literature
with psychological research on social cognition, which explores how social information is stored in
memory (Fiske and Taylor, 1984, for a review in political science, see McGraw, 2000). We build on
this work here, arguing that one underappreciated implication of social cognition research is that
attitudes are best conceptualized as networks. Inspired by early investigations into the organization
of human memory (Collins and Quillian, 1969; Loftus and Collins, 1975), psychologists have long
understood attitudes as “networks of associated cognitions” (Monroe and Read, 2008, 734), whether
through associative (Bower, 1981; Anderson, 1983) or connectionist (Feldman and Ballard, 1982;
5
Smith, 2009) models. Network models of attitudes envision attitude objects (Democrats, gun control, the war in Afghanistan) to be stored as a system of nodes connected to one another through
associational links, the precise structure of which has implications for how attitudes are accessed
and formed (Lodge and McGraw, 1995; Lavine, Thomsen, and Gonzales, 1997). Just as models of
semantic memory demonstrate that links form between concepts brought to mind simultaneously
(such as pets and dogs), network models of attitudes posit that links emerge or are strengthened
when multiple attitudes are thought about once (Loftus and Collins, 1975). Indeed, assuming that
attitudes are linked to one another necessarily means that “the representational structure becomes
a network” (Judd and Krosnick, 1989, 109).
Like Sinclair (2012, 14) argues in reference to the influence of social networks on political behavior, little work in political science has explicitly modeled attitudes as networks, but it has been
nevertheless suggestive of the paradigm. Whenever we talk about “spreading activation” in the context of priming effects (Valentino, 1999), the role of memory-based processing in theories of survey
responses (Zaller, 1992), or schemas in political belief systems (Conover and Feldman, 1984), we
are implicitly studying the consequences of networked attitude structure.
Despite invoking networked metaphors in our work, we have not fully embraced the implications of treating attitudes as networks, both for our theories of how attitudes interact within a
system and the methods we use to test them. Treating attitudes as networks enables us to move
beyond investigations of component parts to assess the system as whole, without relying on restrictive assumptions about the nature of the structure. Thus, rather than asking whether one attitude
shapes another, or whether specific policy positions go together, we can ask whether attitudes at different level of abstraction — values, orientations, or policy positions2 — stimulate more change in
the system as a whole, whether the system is tightly organized (implying close associations among
the elements) or relatively sparse (implying that people do not readily connect attitudes in the pro2
We draw here from the language of Hurwitz and Peffley 1987, who specify three levels of abstraction in their
hierarchical model of foreign policy attitude structure.
6
posed network, even those otherwise found to be correlated), and explore, rather than assume, the
directionality of attitude constraint.
Most research testing the impact of one or more attitudes (the IV(s)) on another attitude (the
DV) assumes that attitudes are shaped from the top-down, such that core values like authority or
ethnocentrism predominantly shape attitudes towards the Iraq War via militant internationalism or
directly, but not the other way around. Yet we know that even a leader’s operational code, rooted in
her deep philosophical and instrumental beliefs about the world, is subject to change with shifting
world events: John Foster Dulles’ code changed in response to WWII (Holsti, 1970), as did Lyndon
B. Johnson’s following Operation Rolling Thunder in Vietnam (Walker and Schafer, 2000). In the
domain of public opinion, the link between values and partisanship changes with one’s social network (Lupton, Singh, and Thornton, 2014). Research from cognitive and social psychologists also
argues that the formation and direction of links between attitudes is not necessarily organized by
abstract schemas like ideology, but may instead be based on alternative considerations including
other more concrete positions (Lavine, Thomsen, and Gonzales, 1997). Thus, both political scientists and psychologists raise the possibility of bottom-up processes of attitude constraint — a notion
that we can test directly within the network, where directed ties (which flow unidirectionally from
one attitude element to another) indicate whether the stimulation of one ‘type’ of attitude element
leads to change in another. If the stimulation of core values leads to widespread changes in the system of attitudes, top-down processes are at work. If it is instead the case that an individual’s attitude
toward the Iraq War can stimulate change, we learn that a network of attitudes as a whole can change
in response to shifting world events.
Similarly, while most factor analytic models treat militant and cooperative internationalism as
independent dimensions that do not interact directly with the policy positions they organize (Wittkopf, 1986), studying the system allows us to adjudicate this claim. We can determine whether an
individual’s position on one dimension stimulates changes in the other or on attitudes throughout
7
the system; if so, they are more central sources of constraint than current research suggests.
Previous work also suggests an important degree of uniformity in belief systems across members of the public. While individual attitudes vary, the manner in which they are connected—the
underlying structure—remains the same (e.g. Wittkopf, 1990; Rathbun, 2007). In Hurwitz and Peffley’s (1987) model, for example, individuals can be more or less militaristic, but it is always the case
that militarism shapes attitudes towards defense spending. In contrast, we argue that properties
of the system of attitudes will vary across subsets of the population. In a representational network
in which objects exist cognitively as nodes, links are formed and strengthened over time as individuals “[come] to believe that one object implies, favors, contradicts, or opposes the other object”
(Judd and Krosnick, 1989, 109), or generally through a mental process in which two or more elements are simultaneously represented. Individuals must have formed an association between two
attitude objects in order for one to influence the other in the system (Lavine, Thomsen, and Gonzales, 1997). Given the diversity of experiences and encounters with political information among the
public, the elements that are central to an attitude network and the direction of constraint between
nodes should vary with individual-level characteristics.
We predict that political sophistication will shape systemic properties of attitude networks. With
tools from network analysis we can assess whether how densely foreign policy attitudes are connected in individuals with high and low levels of political sophistication. Much of the public opinion literature has sought to show that foreign policy elites and the mass public structure their foreign
policy views in similar ways (e.g. Holsti and Rosenau, 1988; Chittick, Billingsley, and Travis, 1995,
but see Oldendick and Bardes, 1981), but our networked paradigm implies otherwise: individuals
with relatively high levels of political knowledge will more readily connect abstract and specific attitudes (Zaller, 1992). Further, expanding upon Judd and Krosnick’s (1989) finding that political
attitudes are more consistent among people interested in politics, we expect that those who are most
interested in foreign policy will exhibit higher levels of interdependence and connectedness in their
8
network of foreign policy attitudes.
To summarize, thinking about attitudes as networks suggests a new understanding of foreign
policy attitude structure that differs from previous work in three substantive ways. First, we have
shifted the conversation away from isolated relationships toward a holistic discussion of how the
system of attitudes operates. Second, while public opinion scholars have largely sought to exonerate
the public by showing that constraint flows from the top-down (from abstract values to concrete
policy positions), we expect that it can also flow from the bottom-up: if attitudes are networks of
elements connected to one another in various ways, then attitudes at lower levels of abstraction
(e.g. specific policy positions) may stimulate more change in the system than we often assume.
Finally, we may be overestimating the uniformity of attitude structure: if the particular pattern of
connections varies with individual experiences, then the topology of attitude networks should vary
with individual characteristics.
Methods and materials
Observational data, network analysis, and directionality
Thinking about attitudes as networks also raises two important methodological implications, the
first highlighting the importance of causal inference, and the second the benefits of incorporating
network analysis tools. First, if a single attitude can impact multiple others throughout the system simultaneously, we need to employ research designs that give us leverage on this directional
question: which attitudes stimulate change, and which are stimulated? Although our existing models have made considerable contributions towards understanding how people think about world
politics, political scientists have traditionally approached the structure of foreign policy attitudes
as a measurement problem to be resolved by a structural equation modeling approach, in which
the structure of the system is largely a product of prior theoretical assumptions, rather than some-
9
thing to be tested empirically.3 Because structural equation modeling on static observational data
is poorly suited to testing causal claims (Imai et al., 2011), this literature thus tends to adopt the
directional language of causal inference without employing any of the methodological techniques
conducive to carrying it out.4 One potential solution is to forgo claims of directionality altogether,
but whenever we estimate a regression model where we treat one attitude as a dependent variable,
and another as an exogenous independent variable, we are inherently making a directional claim
about how constraint operates. Another is to follow the advice of Judd et al. (1991), who call for an
exploration of the “dynamic implications” of attitude structure (p. 193) and Peffley and Hurwitz
(1993), who close their panel study by noting that they remain “unable to say much about the exact
reasoning processes that give rise to top-down... patterns of attitude constraint, a task better left to
future, experimental studies” (p. 83). We opt for the latter, employing an experimental design that
lets us tackle this directional question head-on.
Second, if attitudes truly are networks, they should be studied as such. Although network analysis is increasingly popular in both political science and psychology, to our knowledge no one has
yet applied these tools to the study of political attitudes. Given that network analysis is used to study
everything from connections between organizations to ties between neurons (Butts, 2009), there is
nothing to preclude its use here, particularly when our theories are already suggestive of a network
structure – and influential works in public opinion have even gone so far as to invoke network properties like degree centrality as ways of thinking about the structure of political belief systems (Luskin,
1987, 859). We therefore develop an experiment that we analyze using methods from network analysis, allowing us to derive attitude structure from the patterns of relationships that emerge in the
data rather than simply imposing them ex-ante. As we note above, our intent here is to suggest a
different paradigm for the study of belief systems, which requires far more than a single experiment
3
For this reason, there has been a spirited debate in the public opinion literature about the specific criteria used to
establish the topology of foreign policy beliefs (e.g. Kegley 1986; Chittick, Billingsley, and Travis 1995).
4
The notable exceptions include uses of panel data (e.g. Peffley and Hurwitz 1993; Murray 1992; Goren 2005), but
these studies stand out in the foreign policy public opinion literature precisely because their dynamic setup is so rare.
10
to test (Lakatos, 1970); in this sense, the empirical tests we outline below should be understood as
a “plausibility probe” showing what sorts of insights political scientists gain by studying attitudes at
a higher level of analysis.
Study sample
We conducted an online experiment in the winter of 2013 to investigate the systemic properties
of dynamic attitude networks. The sample consisted of 1,307 participants recruited using Amazon.com’s Mechanical Turk.5 Participants, 59.6% of whom identified as male, and 78.4% as white/Caucasian,
ranged in age from 18-75 (median: 28). Samples from Mechanical Turk are increasingly widely used
in the social sciences (e.g. Arceneaux, 2012; Grimmer, Messing, and Westwood, 2012; Huber, Hill,
and Lenz, 2012) because of the extent to which they can replicate classic experiments on more diverse samples than traditionally employed in political psychological research (Buhrmester, Kwang,
and Gosling, 2011; Berinsky, Huber, and Lenz, 2012). This sample is especially valuable for our
purposes both because the length of our experimental instrumentation and complexity of our experimental design would render the study infeasible while embedded in a national survey, but also
because of the nature of the theories being tested. First, since most studies presume that attitudes are
structured similarly across individuals, we should not require a nationally representative sample to
find evidence in favor of their tenets; Druckman and Kam’s (2011, 46) simulations make this point
clear. Second, since conceptualizing attitudes as network assumes that attitude structure varies considerably based on individual characteristics, we do not expect there to be a single model of attitude
structure that applies to the population as a whole.
5
Following the advice of Berinsky, Huber, and Lenz (2012), participants were paid $1.00 in exchange for their participation. In order to test our hypotheses in an appropriate sample of American adults, we required participants to be
at least 18 years of age, US citizens, and have completed at least 50 human intelligence tasks with a 94% approval rate
on previous work.
11
Measures, design, and instrumentation
As Sinclair (2012, 27-28) notes, there are at least two ways to experimentally study network structure. One is to directly manipulate the network by altering its structure; the other is to “stimulate
some part of the network directly and then observe to what extent these manipulations spill over
into other parts of the network.” We adopt this latter strategy, employing an experimental design
where we stimulate one foreign policy attitude and trace the network structure by measuring how
these effects spill over onto other attitudes.
Network models of memory in psychology are often tested using simulation methods, which
allow for networks with a relatively large number of nodes (Monroe and Read, 2008). However, in
order to maintain the tractability and statistical power of the experimental design, the foreign policy
attitude network we test here is far smaller, consisting of 11 nodes at three levels of abstraction: four
core values (authority/respect, harm/care, ingroup/loyalty, and fairness/reciprocity), three foreign
policy orientations (militant internationalism, cooperative internationalism, and isolationism), and
four specific policy attitudes (towards the war in Iraq, the NATO-led intervention in Libya, the potential for intervention in Syria, and renewing the Kyoto protocol). For the sake of feasibility, the
elements of the network tested here thus represent a theoretically-motivated sample of the potentially massive number of relevant nodes — an often necessary step in the analysis of large network
populations.
The four core values of authority/respect, harm/care, ingroup/loyalty, and fairness/reciprocity
were chosen for three reasons. First, they are central in the contemporary psychological literature on moral values (Haidt and Graham, 2007). Whereas fairness/reciprocity and harm/care —
reflecting concerns about equality and helping others in need, respectively — are “individualizing
foundations” important for both liberals and conservatives, authority/respect and ingroup/loyalty
– the former emphasizing tradition and deference to authority figures and the latter focusing on
remaining loyal to one’s fellow group members — are “binding foundations” fundamental only to
12
political conservatives (Graham, Haidt, and Nosek, 2009). Including both types ensures variation
in what values individuals are likely to rely on. Second, two of the foundations (authority/respect
and fairness/reciprocity) relate closely to what Rathbun (2007) terms hierarchy and community, the
values used in his model of the foundation of foreign policy views. Third, recent survey research
corroborates our expectations that these moral foundations are elements of foreign policy attitude
networks, demonstrating that each is associated with both militant and cooperative internationalism, and also that they predict stances on specific policy issues (Kertzer et al., 2014).
Militant internationalism (MI) and cooperative internationalism (CI) — referring to an individual’s support for the idea that the US military should play an active role in international affairs,
or that the US should work with other countries to solve global problems, respectively — are selected as the two general foreign policy postures or orientations primarily because of the popularity
of Wittkopf’s (1990) model, the independent structure of which is typically inferred from factor
analysis rather than experimentally investigated. The third orientation, isolationism, is included
in order to assess general preferences for disengagement independent of MI and CI (Holsti, 1979).
Finally, we include four specific policy attitudes: positions on the Iraq War and the intervention in
Libya, as well as attitudes about potential U.S. military involvement in Syria and the U.S. role in
establishing a new treaty to replace the Kyoto protocol. These policy items are selected because they
vary in two key respects. First, attitudes toward the Iraq war are likely to be widely connected due to
their long-term salience, allowing us to determine the extent to which individuals extrapolate from
one battleground to another when forming their views. Second, they represent a range of policies
that are positively associated with MI (Iraq, Syria), CI (Kyoto), or both (Libya), allowing us to test
whether stimulating one follows the logical patterns of constraint proposed in previous research
(Kertzer et al., 2014).
Following terminology from graph theory, we test networks of foreign policy attitudes with
what we call a directed multigraph experiment, because it allows us to simultaneously adjudicate the
13
direction of constraint between multiple attitude nodes as well as make claims about system-level
properties such as the density of the network.6 In the experiment, we employ a persuasive-message
approach (Dinauer and Fink, 2005), in which, after answering a series of demographic questions,
participants are exposed to one of eleven stimulus paragraphs designed to stimulate their position
on the node being manipulated, or a control message unrelated to foreign policy attitudes.
This approach has the effect of activating the targeted attitude object in the participant’s mind,
together with both the positive associations offered in the message and any negative associations
that they already have. Where associational links exist, after one attitude is stimulated, attitudes
toward linked nodes should be open to influence while unconnected attitude objects will not be
implicated. Just as Judd et al. (1991) show that priming one attitude with survey questions produces
more polarized responses to linked attitudes through spreading activation, we expect that attitudes
towards linked nodes will differ across treatment conditions — changing indirectly in response to
the treatment. Importantly, while we target a message at one attitude, we are interested in these
indirect, or “downstream” effects of our stimulus paragraphs, rather than on whether the message
changed the targeted attitude itself. Indeed, we have reason to predict that we would find very
little change on the explicitly targeted position: Blankenship, Wegener, and Murray (2012) show
that changing an attitude with a direct message is challenging — particularly where it concerns
controversial policy positions — as participants guard against the obvious effort. They find evidence
that an “indirect route” (a message targeted at a different part of the attitude network), is “more
effective in changing the policy attitudes” (Blankenship, Wegener, and Murray, 2012, 609). If a
message successfully leads to attitude shifts, we expect to see evidence through changes in linked
attitudes. Thus, this design enables us to get at the process through which some attitudes shape
others and identify the directions in which this constraint occurs.
The stimulus materials – presented in full in the appendix §2.1 – were similar in length, compo6
In graph theory, attitude constraint implies a directed graph; given distinct attitudes A and B, if A constrains B,
we we draw a directed edge (A, B). If both (A, B) and (B, A) exist, we have a multigraph.
14
sition, and the inclusion of an expert quote identifying reasons for an individual to have a positive
attitude toward the element of the network being manipulated; for example, authority/respect is
emphasized because “authority figures represent the wisdom of their position and can help people
have better lives by protecting them.” These short messages were created to only activate the relevant node in the network, deliberately avoiding language that might directly reference important
components of any of the other elements.7 As a result, the manipulation ensured that only the element of interest was directly stimulated, such that changes in participants’ position on other nodes
in the network, relative to participants who received the control message, could be attributed the
association within the network.
Our experimental approach thus differs dramatically from most investigations into attitude
structure in political science, which infers structure by searching for correlations in individuals’
responses to survey items. We contend that our approach has three advantages. First, as we note
above, to suggest that one attitude determines an individual’s position on another posits a dynamic,
directed process, and experiments are better suited to determining whether A constrains B or B
constrains A within the network. Second, it ties into Converse’s (1964) original understanding of
dynamic attitude constraint: the idea that “when new information changes the status of one ideaelement in a belief system, by postulate some other change must occur as well.” (p. 208) Our experimental design mimics this process by targeting new information at each node in the network, and
measuring what changes occur with respect to the other nodes. Third, by mapping changes across
the network rather than from one attitude to another, we are able to investigate how the system
operates holistically.
We first exposed participants to the persuasive message. Next, we measured each participant’s
position on the eleven elements of our proposed attitude network. Questions were presented to par7
Because they relate to salient foreign policy issues, some of the messages may be more likely to trigger affective
reactions while others may be read as more cognitively-laden. Consistent with the psychology literature on automatic
attitude activation, where both affect and cognition are implicated in spreading activation (Fazio, 2001), we expect that
either process will facilitate change in associated attitudes.
15
ticipants in nine blocks in random order to avoid potentially confounding order effects. To measure
participants’ endorsement of the abstract moral values, we relied on the twenty-four items pertaining to these four values that are included on the full version of the moral foundations questionnaire
developed by Haidt and Graham (2007). In these questions, participants record how relevant particular considerations are to their judgments of right and wrong on a 5-point scale ranging from “not
at all relevant” to “extremely relevant” and the extent to which they agree with statements about the
importance of each value on a 6-point scale from “strongly disagree” to “strongly agree.” These two
types of value questions are each presented in a separate block, with items measuring all of the four
values randomized within each block.
While widely used scales exist for the purposes of measuring militant and cooperative internationalism (Wittkopf, 1986), many of their items measure these orientations by gauging support for
specific policy items such as Cold War containment or the importance of intervening in humanitarian crises. Because determining whether attitudes at one level of abstraction spark more widespread
change than others requires keeping general orientations and specific policy attitudes distinct, we
chose subsets of the standard MI and CI items for use in this study. These questions tap participants’
general ideas about the use of force in global politics (MI) and about the need for the U.S. to cooperate with other nations and organizations to solve transnational problems (CI). Participants recorded
the extent to which they agreed with each of the items on a scale from 1 (“strongly disagree”) to 7
(“strongly agree”). Isolationism is measured using a four item scale that taps the extent to which
individuals think that the U.S. would be better served by focusing on national issues as opposed to
becoming involved in the world. The final four blocks of questions measured attitudes toward the
2003 U.S. intervention in Iraq, the 2011 airstrikes conducted by NATO in Libya, the potential for
U.S. military involvement in the Syrian conflict, and the possibility of a new agreement to replace
the expired Kyoto Protocol. Items included in these blocks were drawn from questions used on a
variety of public opinion polls conducted by organizations such as ABC News and the Pew Research
16
Center. Responses were provided on 4-7 point scales, with higher scores reflecting more support
for the policy option.8
Finally, participants reported their political ideology and party identification, indicated their
personal interest in foreign policy issues, and completed a 9-item scale measuring political knowledge. The scale included four questions recommended by Delli Carpini and Keeter (1993), together
with five additional questions, three of which specifically tested knowledge on international political
issues.9
Creating the directed multigraphs
The conventional method of presenting the results would be to simply analyze the magnitude and
statistical significance of each of the treatments on each of the dependent variables in sequence.
However, this sequential approach is problematic for two reasons. First, with eleven treatments
and eleven dependent variables, the results are too complex to clearly convey, and the sequential
approach thus obscures more than it reveals. Second, and perhaps more importantly, our quantities
of interest are not individual treatment effects, but the overall configuration; our hypotheses concern
the network-level structure of foreign policy attitudes, rather than whether one particular attitude
influences another.
Thus, we present our results in a different manner, analyzing the experimental data using tools
from network analysis, which to our knowledge is one of the project’s novel contributions. In our
setup, each attitude object (authority, the Iraq war, etc.) constitutes a node or vertex in a network.
If manipulating one node causes a change in another node above a certain threshold, we deem
the two nodes to be connected to one another (or, in network terminology, to share an edge). For
example, if stimulating authority causes a change in attitudes towards the Iraq war, we draw an
8
Measurement scales for each item are presented in full in appendix §2.2; scale reliabilities in appendix §3.
The additional questions were: “Who is the Speaker of the U.S. House of Representatives?”; “Who is the current
majority leader of the U.S. Senate?”; “Who is the Prime Minister of the United Kingdom?”; “What country is led by
Hamid Karzai?” and “Which five countries are permanent members of the U.N. Security Council?”
9
17
edge from authority to Iraq. Because the experimental design lets us calculate treatment effects
bidirectionally — that is, we can tell both whether manipulating node A causes a change in node B,
and whether manipulating node B causes a change in node A — the network is a directed graph. By
using network analytic techniques to analyze the experimental results, we can not only display the
network graphically, but also use simple statistical techniques to analyze the topology of the network,
thereby highlighting the “big picture” of the results even while comparing across subgroups. This
approach has four steps: (i) calculate the treatment effects, (ii) “thin” or “threshold” the results by
retaining only the largest treatment effects, which are then entered into an adjacency matrix, (iii)
plot the adjacency matrix as a network, and (iv) employ permutation and randomization tests to
calculate the probability that the observed network topologies are due to chance. We discuss each
step in turn.
The first step to generating any network is to define its edges. Whereas some types of relations
are easily dichotomized and thus have fairly straightforward criteria for what constitutes an edge
– for example, whether senators i and j cosponsored a bill in a given time period (Fowler, 2006)
– others are more ambiguous (Butts, 2009). The conventional approach would be to estimate each
of our treatment effects using t-tests, drawing an edge from node i to node j if i’s treatment effect
on j is statistically significant according to threshold α. The difficulty with this method, however,
is that the sheer number of hypothesis tests produces an obvious multiple comparisons problem,
for which even relatively powerful empirical Bayes techniques like local fdr will produce overly
sparse networks. Instead, we propose two solutions. First, we could simply abandon thresholding
altogether (Thomas and Blitzstein, 2011), drawing edges between all nodes and weighting them
according to the magnitude of each respective treatment effect. To facilitate the use of simpler binary
network statistics, however, we instead adopt a thresholding approach, in which we threshold edges
based upon the distribution of potential effect sizes in the data, rather than by making reference to
18
a broader population from which our sample is assumed to be drawn.10
We first calculate all of the treatment effects all at once, both for the full sample of participants,
and for a series of subsamples dividing the participants along two measures of political sophistication.11 We represent our treatment effects with t-statistics, since they capture both the difference
in means between the treatment and control groups and also their sample sizes and standard deviations, but a variety of other statistics could be used, ranging from a simple difference in means to
estimates from regression models or matching estimators that also control for pretreatment covariates. In order to produce the analyses reported below, we take the absolute value of the test statistics
and threshold them by declaring the top 10% of effects to merit consideration and discarding the
remaining 90%; in sensitivity analyses in the Appendix we show that our results are robust to the
selection of different cut points.
The effects that meet this 10% threshold are then used to produce an 11 × 11 adjacency matrix
that depicts the relationships between the 11 nodes in a given sample, where entry (i, j) refers to
the presence of an edge from treatment i to dependent variable j. Since the matrix is binarized and
thresholded, those treatment effects that meet the 10% test are represented with a 1, and those that
fail to meet the test are represented with a 0.12 We can then plot this adjacency matrix, and calculate a variety of network statistics to characterize the topology of the network. Moreover, we can
then engage in network inference to produce uncertainty estimates around these network statistics,
comparing our observed networks with randomly generated networks of similar size and density,
to calculate the probability that features of our network are due solely to chance.
To evaluate the likelihood that the distributions of edges amongst core values, orientations, and
policy attitudes are simply due to chance, we perform a series of permutation tests where we ran10
In using the observed data to determine whether a particular effect merits consideration, our analyses are similar
in spirit to other forms of local inference (Efron, 2010; Keele, McConnaughy, and White, 2012).
11
By calculating the distribution of treatment effects for all of our analyses at once, it prevents the analyses from being
ad-hoc.
12
If no thresholding was employed, a weighted adjacency matrix would be used, in which each entry in the matrix
would depict the magnitude of a treatment effect, rescaled to be bounded by 0 and 1.
19
domly generate 10000 networks of an identical size and density as our observed network, to test
the probability that we would observe a given network statistic (the number of edges connected to
a node or category of nodes) due simply to network size and density alone (Butts, 2008).13
Results
We present our results in two stages. First, we analyze the attitude network across our full sample.
Second, because we expect heterogeneous treatment effects, we generate separate attitude networks
for different subgroups to show how foreign policy attitude structure varies as a function of political
knowledge and interest in foreign policy. Throughout, we assess the topologies of the observed
networks with reference to whether the results support the assumptions implicit in extant models
of attitude structure, and what we learn by “zooming out” to the system level.
The full sample
We begin by plotting the 11-node network for the full sample, in Figure 1.14 Core values are depicted with circles, foreign policy orientations with diamonds, and specific policy attitudes with
triangles. The arrows indicate the direction of the attitude constraint, and nodes are scaled by their
degree centrality, which measures each nodes’ connectedness to other nodes in the sample: the
larger the node, the more important a role it plays in the network.15 Additionally, emphasizing the
importance of directional hypotheses in models assuming that the most abstract attitudes shape
more concrete positions (from the top-down), it is also useful to differentiate between outdegree
centrality — that is, how many nodes a particular attitude influences — versus indegree centrality
— that is, how many nodes influence a particular attitude, which we discuss further with reference
13
See appendix §5 for permutation tests to evaluate network density.
All network plots and statistics are generated using the statnet suite of packages in R. (Handcock et al., 2008)
15
We estimated additional measures of centrality, such as Eigenvector centrality, but because of the small number of
nodes in the network, many of these measures failed to converge. As Butts (2008, 22) notes, however, degree centrality is
highly correlated “with most other measures of centrality, making it a powerful summary index,” and thus a satisfactory
measure of a node’s connectedness in attitude networks.
14
20
Figure 1: Foreign policy attitude network: full sample
Auth
Fair
Syria
Ingroup
Iraq
Libya
Iso
Harm
MI
CI
Kyoto
Note: Arrows depict the thresholded treatment effects, and thus, patterns of attitude constraint; nodes are scaled by
their degree centrality, so larger nodes are connected to a higher number of other attitudes. Core values are
represented as circles, foreign policy orientations as diamonds, and policy attitudes as triangles. Thus, core values play
a smaller role in the results from the full sample — and specific policy attitudes play a larger role — than typical
models of foreign policy attitude structure would suggest.
21
to the specific networks. An observable implication of top-down influence is that values will have
only outgoing connections (as abstract values shape orientations and policy positions below them)
and thus will have higher out-degree centrality, but the fact that more specific attitudes do not shape
values should produce very low levels of in-degree centrality. Policy positions should display the
opposite pattern if these assumptions are correct, with higher levels of in-degree centrality implying
that they are constrained by higher-order concepts, and an absence of out-degree ties that would
otherwise indicate that policy attitudes can stimulate change across elements of a belief system.
Contrary to top-down assumptions about attitude constraint, values do not play a large role in
shaping less abstract orientations or policy attitudes: in fact, fairness/reciprocity, ingroup/loyalty,
and harm/care are isolates, disconnected from other nodes in the network. In other words, there is
a general absence of links between moral values and foreign policy attitudes. Indeed, permutation
tests demonstrate that moral values play significantly less of a role (p = 0.024), and policy positions play significantly more of a role (p = 0.046), as measured by degree centrality, than would
be expected by chance.16 Constraint is evident in multiple directions, both from the bottom-up —
positions on the Iraq war shape both isolationism and authority — and within the same level of
specificity, as policy attitudes shape one another. Attitudes toward the war in Iraq are most central in the full network, with more overall and outgoing connections than what we would expect by
chance (p = 0.012 and p = 0.001, respectively). This corroborates our prediction that, consistent
with operational code analyses and psychological theories of how networks in memory are organized, highly salient foreign policy events can link to other activated attitudes, and thus come to
influence more general ideas about how foreign policy should be conducted. Figure 1 thus offers an
initial indication of the benefits of a network approach, displaying patterns of stimulation flowing
in directions that are inconsistent with our current understanding of attitude structure. At the same
16
See the Table 1 for a complete summary of the permuted analyses. The statistics show the number of total edges
(degree centrality), outgoing edges (outdegree centrality), and incoming edges (indegree centrality) by node type. They
demonstrate whether specific node types (values, orientations, or policy positions) display greater levels of centrality
than we would see in a random network of equal size and density.
22
time, however, with psychological research on interattudinal structure as a foundation (Judd and
Krosnick, 1989), we emphasize that properties of attitude networks are likely to vary considerably
with individual-level characteristics like political sophistication, such that the relatively sparse attitude network derived from average treatment effects across the full sample might belie considerable
treatment heterogeneity when the sample is partitioned into subgroups – a hypothesis we explore
below.17
Subgroup analyses
Political sophistication
Political sophistication offers an opportunity to investigate the appropriateness of “one-size fits all”
models of attitude structure. Although much of the foreign policy attitude literature has sought to
show that both foreign policy experts and the general public structure their foreign policy attitudes
similarly (e.g. Holsti and Rosenau, 1988; Rathbun, 2007), political knowledge has been found to
affect both people’s likelihood of receiving political information (Zaller, 1992), and how they process the information they do receive (Fiske, Lau, and Smith, 1990). Given that the formation of
associational links is affected by both the reception and processing of information, we expect that
participants with different levels of political sophistication will exhibit distinct patterns of constraint
at the system level. Since conclusions about political sophistication depend on how it is measured
(Luskin, 1987), we look at it in two different ways: first, as political knowledge (Delli Carpini and
Keeter, 1993), and second as interest in foreign policy issues (Krosnick, 1990).
Figure 2 displays the attitude networks by three levels of political knowledge. First, network
density statistics that measure how connected each network is as a whole reveal a curvilinear effect,
in that moderately knowledgeable participants have more densely connected foreign policy attitude
networks than both their high and low-knowledge counterparts – the middle group’s network is 1.63
17
We also explore gender- and ideology-based structural differences in Appendix §5.
23
Table 1: Permuted Network Statistics
Degree Centrality
Observed
p-value
Out-degree Centrality
In-degree Centrality
Observed
Observed
p-value
p-value
Full Network
Values
1 [2,8]
0.024
0 [0,5]
0.057
1 [0,5]
0.261
Orientations
3 [1,7]
0.531
1 [0,4]
0.450
2 [0,4]
0.549
Policies
8 [2,8]
0.046
5 [0,5]
0.038
3 [0,5]
0.419
Low Interest
Values 11 [6,15]
0.503
5 [2,9]
0.560
6 [2,9]
0.444
Orientations
4 [4,12]
0.054
2 [1,7]
0.180
2 [1,7]
0.185
Policies 17 [6,16]
0.010
9 [2,9]
0.039
8 [2,9]
0.109
Medium Interest
Values
4 [3,10]
0.221
2 [1,6]
0.380
2 [1,6]
0.367
Orientations
4 [1,8]
0.526
1 [0,5]
0.294
3 [0,5]
0.395
Policies
6 [3,10]
0.578
4 [1,6]
0.343
2 [1,6]
0.387
High Interest
Values
0 [4,13]
0.000
0 [1,8]
0.004
0 [1,7]
0.004
Orientations 13 [3,10]
0.002
6 [1,6]
0.064
7 [1,6]
0.016
Policies
9 [5,13]
0.505
5 [1,8]
0.441
4 [1,8]
0.564
Low Knowledge
Values
6 [4,12]
0.255
2 [1,7]
0.172
4 [1,7]
0.609
Orientations
7 [2,10]
0.378
5 [0,6]
0.138
2 [0,6]
0.400
Policies
9 [4,12]
0.392
4 [1,7]
0.611
5 [1,7]
0.357
Medium Knowledge
Values 10 [9,19]
0.106
6 [3,11]
0.415
4 [3,11]
0.103
Orientations 13 [6,15]
0.196
5 [2,9]
0.578
8 [2,9]
0.104
Policies 13 [9,19]
0.452
7 [3,11]
0.582
6 [3,11]
0.416
High Knowledge
Values
9 [5, 15]
0.445
5 [2,8]
0.599
4 [2,8]
0.413
Orientations 11 [3,12]
0.081
5 [1,7]
0.295
6 [1,7]
0.132
Policies 10 [5,15]
0.559
5 [2,8]
0.592
5 [2,8]
0.603
Note: table shows the observed network total, in-degree, or out-degree centrality — the
number of observed connections — for a given node type [along with 95% CIs
derived from 10000 permuted networks], and the corresponding p-values.
24
Figure 2: Foreign policy attitude network by political knowledge
(a) Low-knowledge
(b) Moderate-knowledge
Kyoto
MI
Kyoto
Iso
Ingroup
Syria
Fair
Iraq
Fair
Libya
Auth
Iraq
CI
Harm
Iso
Libya
Auth
Syria
Harm
Ingroup
CI
MI
(c) High-knowledge
Iso
Fair
Syria
CI
Iraq
Ingroup
Harm
Auth
MI
Kyoto
Libya
Note: Arrows depict the thresholded treatment effects, and thus, patterns of attitude constraint; nodes are scaled by
their degree centrality. Core values are represented as circles, foreign policy orientations as diamonds, and policy
attitudes as triangles. Consistent with Zaller (1992), knowledge displays a curvilinear effect here, as the
medium-knowledge network has a greater density than either the low-knowledge or high-knowledge counterparts.
25
times as dense as those low in political knowledge, and 1.2 times the density of the high knowledge
network. This is consistent with Zaller’s (1992) conclusion that the most sophisticated individuals
have stable attitudes that are more resistant to change, and that the least sophisticated lack the ability
to form and store connections between abstract and specific concepts like values to specific attitudes.
Instead, it is the moderate sophisticates, who hold multiple considerations but have the ability to
connect concepts, that are most open to indirect attitude change via the stimulus messages. Second,
each of these three subgroup networks are more densely connected than the aggregated network in
Figure 1, likely due to the fact that aggregating attitude structure for an entire sample of participants
masks important treatment heterogeneity. Finally, of the three levels of political knowledge, only the
network of the most knowledgeable participants includes a role for orientations that is greater than
what a random assignment of edges would predict (p = 0.081). Thus, as Converse would predict, it
appears that only the most sophisticated individuals engage the abstract views about foreign policy
so familiar to academics when forming their attitudes.
Figure 3 operationalizes political sophistication differently by looking at the extent to which
participants expressed interest in foreign policy issues, since political interest is often held to be
both a cause and consequence of political sophistication (Fiske, Lau, and Smith, 1990). The results
show strikingly different patterns compared to political knowledge: with political knowledge, the
moderate knowledge network was the most dense, whereas with foreign policy interest, we see that
individuals who expressed a moderate amount of interest in foreign policy issues have attitude networks that are less dense than both those who reported low or high levels of interest. Those least
interested in foreign policy have attitude networks that are driven primarily by specific policy positions, most evidently those regarding the war in Iraq — policy attitudes play a larger role in the
low-interest attitude network than we would expect due to chance (p = 0.010). Further analysis reveals that the policy-based connections are primarily out-going, demonstrating more precisely that
the direction of influence flows from specific policy attitudes to other elements of attitude networks
26
Figure 3: Foreign policy attitude network by interest in foreign policy issues
(a) Low-interest
(b) Medium-interest
CI
Harm
Libya
Syria
Libya
Ingroup
MI
MI
Auth
Iraq
Syria
Iraq
Kyoto
Fair
Iso
Kyoto
Auth
Harm
Iso
Ingroup
CI
Fair
(c) High-interest
Harm
Iraq
Kyoto
CI
Auth
Iso
MI
Fair
Libya
Ingroup
Syria
Note: Arrows depict the thresholded treatment effects, and thus, patterns of attitude constraint; nodes are scaled by
their degree centrality. Core values are represented as circles, foreign policy orientations as diamonds, and policy
attitudes as triangles. Here, we see the inverse curvilinear effect to that from Figure 2, in that the medium-interest
network displays less constraint than its low-interest and high-interest counterparts, with specific policy events playing
a greater role influencing attitudes throughout the network for low-interest participants, and the more abstract foreign
policy orientations playing a greater role influencing other attitudes for high-interest participants.
27
for the uninterested public. Those least interested in foreign policy issues therefore have patterns
of associations that could lead to widespread changes in the network of attitudes based on a single
event in U.S. foreign policy — the chief concern of Almond, Lippmann, and the other postwar realists. Thus, when the uninterested engage with a foreign policy event, they are more likely to base
their broader ideas about how the US should conduct itself internationally on that attitude, as they
come to believe that it implicates other pre-existing but newly associated attitude objects. Multiple
nodes in the network update in response to the attitude.
In contrast, general orientations play a greater role in the networks of those most interested
in international affairs, similar to what we found in the high knowledge subgroup (p = 0.002).
Moreover, values play no role at all in the foreign policy networks of the most interested segment of
our sample, which is telling given the large number of edges in this network (p = 0.000). Consistent
with Judd and Krosnick’s (1989) findings that attitudes are more reliably consistent among the group
of individuals who ascribe importance to the political attitudes being investigated, those participants
who report high levels of interest in foreign policy seem to closely connect their general views on
how foreign policy should be conducted (orientations) to specific issues, but leave out the values that
may not be obviously associated with the rest. The high interest network also exhibits the patterns
most closely consistent with top-down directionality assumptions, given that constraint flows from
orientations to policy items, but bottom-up effects such as the influence of Libya on both MI and
Isolationism, bidrectional edges indicating within-level influences between CI and Isolationism, and
the nonexistent role of core values suggests that this assuming effects in only one direction misses
much of the story.
The results of the two political sophistication analyses thus suggest two findings. First, both
general political knowledge and interest in foreign policy affect how foreign policy attitudes are
structured, as well as the degree of structure itself — something that we can only understand by
turning to the properties of the system. The importance of sophistication thus suggests that, in con-
28
trast to many existing models of foreign policy attitude structure (e.g., Holsti and Rosenau, 1988;
Rathbun, 2014), expertise matters. Second, regardless of how one operationalizes political sophistication, we detect violations of the top-down patterns of constraint assumed by past work, finding
signs of bottom-up patterns of influence, bidirectional edges across levels, and constraint among
attitudes within levels of abstraction.
Conclusion
Political scientists have long sought evidence of structure and constraint in foreign policy attitudes,
producing a variety of models based on divergent sets of assumptions about how these attitudes are
organized. In this paper, we build on these important contributions by proposing a new way of
studying political attitudes: networks. We argued that interattitudinal structure is best conceptualized as a flexible network of nodes, interacting with one another as part of a system, and developed a
novel experimental design that illuminates the properties of the whole system rather than relations
between its component parts while studying attitudes as network data. Our intent here is thus to
explore a different ontology for studying belief systems.
The results from our experimental “plausibility probe” suggest four key conclusions. First, conceptualizing attitudes as networks offers a promising theoretical framework. With our design, we
were able to visualize interactions within a broader system of attitudes, directly observing which
types of attitudes stimulated more change throughout the system. We bring together findings from
across the foreign policy attitudes literature to determine how various relationships reported elsewhere work within an attitude network. Additionally, the networked paradigm allows us to study
system-level properties like network density — which revealed important differences in the attitude
structures of political sophisticates at different tiers — that would have been difficult to uncover in
a traditional framework. The experiment thus provides insight into holistic properties of attitude
networks.
29
Second, Converse’s (1964) pessimistic assessment about the paucity of top-down constraint in
the American public seems to be more accurate than revisionist accounts have suggested: while
regression and SEM-based research designs assume that values and abstract dispositions will be
most influential in ultimately determining policy positions, we find that values generally play less
and policy positions more of a role in the system of attitudes than we would expect by chance.
Given the extent to which IR scholars have long complained about the moralistic nature of public
opinion — Hans Morgenthau claimed that “intoxication with moral abstractions” was “one of the
great sources of weakness and failure in American foreign policy” (Morgenthau, 1951, 4) — it is
noteworthy that we find more evidence of specific policy issues constraining moral values than of
moral values constraining specific policy issues. The outsize role played by bottom-up constraint
suggests that the public is more likely to tie leaders’ hands in response to specific policy issues –
particularly involving salient events – rather than abstract principles per se.
Third, when it comes to attitude structure, one size does not fit all: the oft-dramatic differences
revealed in subgroup analyses clearly demonstrate that there is no single attitude structure that can
be readily applied to all members of the population at large. Instead, we find sensible patterns of
divergence based on political sophistication. This ultimately suggests that it is misleading to speak of
such a thing as “the” structure of foreign policy attitudes — but it also suggests that the IR literature,
which frequently views the public as a monolithic unit responding to international events in tandem
(Fearon, 1994), can benefit from taking this heterogeneity into account.
Finally, we show that political knowledge and levels of interest in foreign policy are particularly important in understanding how attitude networks are organized. Those with moderate levels
of knowledge have more densely connected attitude networks than their high and low-knowledge
counterparts, and interest in foreign policy determines whether the direction of constraint flows
mainly from more specific or abstract attitude objects in the network. The central role of expertise in our findings thus raises questions about whether it is realistic to expect the mass public to
30
organize their foreign policy attitudes in the same manner as do foreign policy elites.
Importantly, the results also have implications far beyond the study of foreign policy attitudes.
Although political scientists are well aware of the danger of order effects in survey responses, the
network patterns of interdependence detected here suggest that order effects may be even more
problematic than many political scientists may realize, as activating one attitude can affect evaluations of other attitudes across levels of abstraction. Moreover, directional debates in attitude
structure are relatively common in political science: do policy preferences cause party ID, or does
party ID cause policy preferences? (Jackson, 1975) Does racial prejudice cause attitudes towards
specific policies like welfare and busing? (Sears, 1988) Do evaluations of political parties stem from
economic assessments, or do economic perceptions drive political support? (Evans and Andersen,
2006) Traditionally, these hypotheses have been tested either using path analysis of cross-sectional
data (whereupon assumptions about the direction of causality is governed entirely by theoretical assumptions), and panel surveys, which offer more leverage on causal inference. The directed multigraph experimental design fielded here offers another way to test theories of directional attitude
interdependence and offers yet another path for matching our methods to our theories.
Finally, the current investigation suggests four fruitful avenues for future research. First, although we investigate a number of individual differences in the subgroup analyses, the analyses are
not multivariate. Given the current experimental design and sample size, we are only able to control
for one variable at a time, but developing a factorial design will facilitate multivariate analyses that
may provide additional insight into the structure of foreign policy attitude networks across individuals. Second, although we employ a thresholding approach when drawing edges between nodes,
based on the distribution of effects in the sample, future work could examine not just the presence of
a significant treatment effect, but its substantive impact as well, treating the experimental results as
a weighted network. Third, the between-subjects component of our design limits our ability to investigate changes in an individual’s unique attitude network. Developing a within-subjects version
31
of a directed multigraph experiment will allow us to more clearly discern causal patterns of attitude
constraint. Finally, the current investigation explored whether individual-level attributes affect the
topology of attitude structure, but future research should examine how particular attitude-level attributes (strength, extremity, level of abstraction, and so on) structure the topology of foreign policy
belief systems.
References
Anderson, John R. 1983. The Architecture of Cognition. Cambridge, MA: Harvard University Press.
Arceneaux, Kevin. 2012. “Cognitive Biases and the Strength of Political Arguments.” American Journal of Political Science 56 (2): 271-285.
Berinsky, Adam J., Gregory A. Huber, and Gabriel S. Lenz. 2012. “Evaluating Online Labor Markets
for Experimental Research: Amazon.com’s Mechanical Turk.” Political Analysis 20 (3): 351-368.
Blankenship, Kevin L, Duane T Wegener, and Renee A Murray. 2012. “Circumventing Resistance:
Using Values to Indirectly Change Attitudes.” Journal of Personality and Social Psychology 103 (4):
606.
Bower, Gordon H. 1981. “Mood and Memory.” American Psychologist 36 (2): 129-148.
Brewer, Paul R., and Marco R. Steenbergen. 2002. “All Against All: How Beliefs about Human Nature
Shape Foreign Policy Opinions.” Political Psychology 23 (1): 39-58.
Buhrmester, Michael, Tracy Kwang, and Samuel D. Gosling. 2011. “Amazon’s Mechanical Turk: A
New Source of Inexpensive, Yet High-Quality, Data?” Perspectives on Psychological Science 6 (1):
3-5.
Butts, Carter T. 2008. “Social Network Analysis: A Methodological Introduction.” Asian Journal of
Social Psychology 11: 13-41.
Butts, Carter T. 2009. “Revisiting the Foundations of Network Analysis.” Science 325 (July 24 2009):
414-416.
32
Chittick, William O., Keith R. Billingsley, and Rick Travis. 1995. “A Three-Dimensional Model of
American Foreign Policy Beliefs.” International Studies Quarterly 39 (3): 313-331.
Collins, Allan M., and M. Ross Quillian. 1969. “Retrieval Time from Semantic Memory.” Journal of
Verbal Learning and Verbal Behavior 8 (2): 240-247.
Conover, Pamela J., and Stanley Feldman. 1984. “How people organize the political world: a
schematic model.” American Journal of Political Science 28 (1): 95-126.
Converse, Philip E. 1964. “The nature and origin of belief systems in mass publics.” In Ideology and
Discontent, ed. David E. Apter. New York: Free Press.
Delli Carpini, Michael X., and Scott Keeter. 1993. “Measuring Political Knowledge: Putting First
Things First.” American Journal of Political Science 37 (4): 1179-1206.
Dinauer, Leslie D., and Edward L. Fink. 2005. “Interattitude Structure and Attitude Dynamics: A
Comparison of the Hierarchical and Galileo Spatial-Linkage Models.” Human Communication
Research 31 (1): 1-32.
Druckman, James N., and Cindy D. Kam. 2011. “Students as Experimental Participants.” In Cambridge Handbook of Political Science, ed. James N. Druckman, Donald P. Green, James H. Kuklinski, and Arthur Lupia. Cambridge University Press.
Eagly, Alice H., and Shelly Chaiken. 2007. “The Advantages of an Inclusive Definition of Attitude.”
Social Cognition 25 (5): 582-602.
Efron, Bradley. 2010. Large-Scale Inference: Empirical Bayes Methods for Estimation, Testing, and
Prediction. Cambridge University Press.
Evans, Geoffrey, and Robert Andersen. 2006. “The Political Conditioning of Economic Perceptions.”
Journal of Politics 68 (1): 194-207.
Fazio, Russell H. 2001. “On the Automatic Activation of Associated Evaluations: An Overview.”
Cognition & Emotion 15 (2): 115–141.
Fearon, James D. 1994. “Domestic Political Audiences and the Escalation of International Disputes.”
33
American Political Science Review 88 (3): 577-592.
Feldman, J.A., and D.H. Ballard. 1982. “Connectionist Models and Their Properties.” Cognitive Science 6 (3): 205-254.
Feldman, Stanley. 2003. “Values, Ideology, and the Structure of Political Attitudes.” In Oxford Handbook of Political Psychology, ed. David O. Sears, Leonie Huddy, and Robert Jervis. Oxford: Oxford
University Press.
Fiske, Susan T., Richard R. Lau, and Richard A. Smith. 1990. “On the varieties and utilities of political
expertise.” Social Cognition 8 (1): 31-48.
Fiske, Susan T., and Shelley E. Taylor. 1984. Social Cognition. New York: McGraw-Hill.
Fowler, James H. 2006. “Legislative cosponsorship networks in the US House and Senate.” Social
Networks 28: 454-465.
Goren, Paul. 2005. “Party Identification and Core Political Values.” American Journal of Political
Science 49 (4): 881-896.
Graham, Jesse, Jonathan Haidt, and Brian A. Nosek. 2009. “Liberals and Conservatives Rely on
Different Sets of Moral Foundations.” Journal of Personality and Social Psychology 96 (5): 10291046.
Grimmer, Justin, Solomon Messing, and Sean J. Westwood. 2012. “How Words and Money Cultivate a Personal Vote: The Effect of Legislator Credit Claiming on Constituent Credit Allocation.”
American Political Science Review 106 (4): 703-719.
Hafner-Burton, Emilie M., Miles Kahler, and Alexander H. Montgomery. 2009. “Network Analysis
for International Relations.” International Organization 63: 559-592.
Haidt, Jonathan, and Jesse Graham. 2007. “When Morality Opposes Justice: Conservatives Have
Moral Intuitions that Liberals may not Recognize.” Social Justice Research 20 (1): 98-116.
Handcock, Mark S., David R. Hunter, Carter T. Butts, Steven M. Goodreau, and Martina Morris.
2008. “statnet: Software Tools for the Representation, Visualization, Analysis and Simulation of
34
Network Data.” Journal of Statistical Software 24 (1): 1-11.
Holsti, Ole. 1970. “The ‘Operational Code’ Approach to the Study of Political Leaders: John Foster
Dulles’ Philosophical and Instrumental Beliefs.” Canadian Journal of Political Science 3 (1): 123–
157.
Holsti, Ole R. 1979. “The Three-Headed Eagle: The United States and System Change.” International
Studies Quarterly 23 (3): 339-359.
Holsti, Ole R. 2004. Public Opinion and American Foreign Policy. Revised edition ed. Ann Arbor,
MI: University of Michigan Press.
Holsti, Ole R., and James N. Rosenau. 1988. “The Domestic and Foreign Policy Beliefs of American
Leaders.” Journal of Conflict Resolution 32 (2): 248-294.
Holsti, Ole R, and James N Rosenau. 1990. “The structure of foreign policy attitudes among American leaders.” The Journal of Politics 52 (01): 94–125.
Huber, Gregory A., Seth J. Hill, and Gabriel S. Lenz. 2012. “Sources of Bias in Retrospective Decision
Making: Experimental Evidence on Voters’ Limitations in Controlling Incumbents.” American
Political Science Review 106 (4): 720-741.
Hurwitz, Jon, and Mark Peffley. 1987. “How Are Foreign Policy Attitudes Structured?” American
Political Science Review 81 (4): 1099-1120.
Imai, Kosuke, Luke Keele, Dustin H. Tingley, and Teppei Yamamoto. 2011. “Unpacking the Black
Box of Causality: Learning about Causal Mechanisms from Experimental and Observational
Studies.” American Political Science Review 105 (November 2011): 765-789.
Jackson, John E. 1975. “Issues, Party Choices, and Presidential Votes.” American Journal of Political
Science 19 (May): 161-185.
Judd, Charles M., and Jon A. Krosnick. 1989. “The structural bases of consistency among political
attitudes: Effects of political expertise and attitude importance.” In Attitude structure and function,
ed. Anthony R. Pratkanis, Steven J. Breckler, and Anthony G. Greenwald. Hillsdale, NJ: Erlbaum.
35
Judd, Charles M., Roger A. Drake, James W. Downing, and Jon A. Krosnick. 1991. “Some Dynamic
Properties of Attitude Structures: Context-Induced Response Facilitation and Polarization.” Journal of Personality and Social Psychology 60 (2): 193-202.
Keele, Luke, Corrine McConnaughy, and Ismail White. 2012. “Strengthening the Experimenter’s
Toolbox: Statistical Estimation of Internal Validity.” American Journal of Political Science 56 (2):
484-499.
Kegley, Charles W., Jr. 1986. “Assumptions and Dilemmas in the Study of American’s Foreign Policy
Beliefs: A Caveat.” International Studies Quarterly 30 (4): 447-471.
Kertzer, Joshua D., Kathleen Powers, Brian C. Rathbun, and Ravi Iyer. 2014. “Do moral values shape
foreign policy preferences?” Unpublished manuscript.
Krosnick, Jon A. 1990. “Government Policy and Citizen Passion: A Study of Issue Publics in Contemporary America.” Political Behavior 12 (1): 59-92.
Lakatos, Imre. 1970. “Falsification and the Methodology of Scientific Research Programmes.” In
Criticism and the Growth of Knowledge, ed. Imre Lakatos and Alan Musgrave. Cambridge: Cambridge University Press.
Lavine, Howard, Cynthia J. Thomsen, and Marti Hope Gonzales. 1997. “The Development of Interattitudinal Consistency: The Shared-Consequences Model.” Journal of Personality and Social
Psychology 72 (4): 735-749.
Liberman, Peter. 2013a. “Retributive Support for International Punishment and Torture.” Journal of
Conflict Resolution 57 (2): 285-306.
Liberman, Peter. 2013b. “Retributive support for international punishment and torture.” Journal of
Conflict Resolution 57 (2): 285–306.
Lodge, Milton, and Charles S Taber. 2013. The Rationalizing Voter. Cambridge University Press.
Lodge, Milton and Kathleen M. McGraw, eds. 1995. Political Judgment: Structure and Process. University of Michigan Press.
36
Loftus, Elizabeth, and Allan M. Collins. 1975. “A Spreading Activation Theory of Semantic Processing.” Psychological Review 82 (6): 407-428.
Lupton, Robert N, Shane P Singh, and Judd R Thornton. 2014. “The Moderating Impact of Social
Networks on the Relationships Among Core Values, Partisanship, and Candidate Evaluations.”
Political Psychology.
Luskin, Robert C. 1987. “Measuring Political Sophistication.” American Journal of Political Science
31 (November 1987): 856-899.
McGraw, Kathleen M. 2000. “Contributions of the Cognitive Approach to Political Psychology.”
Political Psychology 21 (4): 805-832.
McGuire, William J. 1989. “The Structure of Individual Attitudes and Attitude Systems.” In Attitude
structure and function, ed. Anthony R. Pratkanis, Steven J. Breckler, and Anthony G. Greenwald.
Erlbaum.
Monroe, Brian M., and Stephen J. Read. 2008. “A General Connectionist Model of Attitude Structure
and Change: The ACS (Attitudes as Constraint Satisfaction) Model.” Psychological Review 115 (3):
733-759.
Morgenthau, Hans J. 1951. In Defense of the National Interest. New York: Alfred A. Knopf.
Murray, Shoon. 1992. “Turning an Elite Cross-Sectional Survey into a Panel Study while Protecting
Anonymity.” Journal of Conflict Resolution 36 (3): 586-595.
Nie, Norman H, Sidney Verba, and John R. Petrocik. 1979. The Changing American Voter. Cambridge, MA: Harvard University Press.
Oldendick, Robert, and Barbara Ann Bardes. 1981. “Belief Structures and Foreign Policy: Comparing Dimensions of Elite and Mass Opinions.” Social Science Quarterly 62 (3): 432-442.
Peffley, Mark A., and Jon Hurwitz. 1985. “A Hierarchical Model of Attitude Constraint.” American
Journal of Political Science 29 (Nov.): 871-890.
Peffley, Mark, and Jon Hurwitz. 1993. “Models of Attitude Constraint in Foreign Affairs.” Political
37
Behavior 15 (1): 61-90.
Rathbun, Brian C. 2007. “Hierarchy and Community at Home and Abroad: Evidence of a Common
Structure of Domestic and Foreign Policy Beliefs in American Elites.” Journal of Conflict Resolution
51 (3): 379-407.
Rathbun, Brian C. 2014. “Wedges and Widgets: The Ideological Origins of Trade Attitudes.” Foreign
Policy Analysis Forthcoming.
Renshon, Jonathan. 2009. “When Public Statements Reveal Private Beliefs: Assessing Operational
Codes at a Distance.” Political Psychology 30 (4): 649–661.
Schafer, Mark, and Stephen G Walker. 2006. “Democratic Leaders and the Democratic Peace: The
Operational Codes of Tony Blair and Bill Clinton.” International Studies Quarterly 50 (3): 561–
583.
Sears, David O. 1988. “Symbolic Racism.” In Eliminating Racism: Profiles in Controversy, ed. Phyllis A. Katz and Dalmas A. Taylor. Plenum Press.
Sinclair, Betsy. 2012. The Social Citizen: Peer Networks and Political Behavior. Chicago: University
of Chicago Press.
Smith, Eliot R. 2009. “Distributed Connectionist Models in Social Psychology.” Social and Personality Psychology Compass 3 (1): 64-76.
Sokhey, Anand E., and Paul A. Djupe. 2011. “Interpersonal Networks and Democratic Politics.” PS:
Political Science & Politics 44 (1): 55-59.
Thomas, A. C., and J. K. Blitzstein. 2011. “Valued Ties Tell Fewer Lies: Why Not To Dichotomize
Network Edges With Thresholds.” ArXiv e-prints (January).
Valentino, Nicholas A. 1999. “Crime News and the Priming of Racial Attitudes During Evaluations
of the President.” Public Opinion Quarterly 63 (3): 293-320.
Walker, Stephen G, and Mark Schafer. 2000. “The Political Universe of Lyndon B. Johnson and his
Advisors: Diagnostic and Strategic Propensities in their Operational Codes.” Political Psychology
38
21 (3): 529–543.
Wittkopf, Eugene R. 1986. “On the Foreign Policy Beliefs of the American People: A Critique and
Some Evidence.” International Studies Quarterly 30 (4): 425-445.
Wittkopf, Eugene R. 1990. Faces of internationalism: Public Opinion and American foreign policy.
Duke University Press.
Zaller, John R. 1992. The Nature and Origins of Mass Public Opinion. Cambridge: Cambridge University Press.
Zaller, John R., and Stanley Feldman. 1992. “A Simple Theory of the Survey Response: Answering
Questions versus Revealing Preferences.” American Journal of Political Science 36 (3): 579-616.
39
Foreign Policy Attitudes as Networks
Supplementary Appendix
Last revised: August 6, 2014
Contents
1 The network of foreign policy attitude research, 1970-2013
Table 1: Attitude structure models, 1970-2013 . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Latent position cluster model of citation patterns . . . . . . . . . . . . . . . . . . . . . . . .
Figure 1: The citation network of foreign policy attitude structure research, 1970-2013 . . .
2 Materials and Instrumentation
2.1 Text of Persuasive Messages . . . . . . . . . . . .
2.1.1 Values . . . . . . . . . . . . . . . . . . .
2.1.2 Orientations . . . . . . . . . . . . . . . .
2.1.3 Policy Positions . . . . . . . . . . . . . .
2.2 Instrumentation . . . . . . . . . . . . . . . . . .
2.2.1 Values: Moral Foundations Questionnaire
2.2.2 Orientations . . . . . . . . . . . . . . . .
2.2.3 Policy Positions . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
2
4
5
6
6
6
7
8
10
10
12
13
3 Scale Reliabilities
14
4 Thresholding and sensitivity analyses
15
5 Supplementary Subgroup Analyses
5.1 Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Political Ideology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
17
20
6 Additional permutation tests
Figure 7: Permutation tests of network density . . . . . . . . . . . . . . . . . . . . . . . . .
20
22
References
23
1
1
The network of foreign policy attitude research, 1970-2013
One of our claims in the manuscript is that existing research on foreign policy attitudes has focused
primarily on testing only one type of relationship at a time — either the clusters of policy positions
uncovered through factor analysis, for example, or the specific predispositions thought to constrain
them, without looking holistically at the dynamic properties — the systemic features — of belief
systems. Another way of making this point comes from looking at the conversations that have taken
place within the literature itself.
We conducted a search of the literature on foreign policy attitude structure published from 1970
to 2013. To compile the list, we used Google Scholar’s search engine to find articles using the search
terms “foreign policy attitudes,” “public opinion and foreign policy,” and “foreign policy attitude
structure,” as well as searching for works citing each included reference, to compile what we believe
to be a fairly comprehensive list of existing work on the structure of foreign policy attitudes.1 We
then categorized each of the 70 pieces into one of two camps (the results of which are listed in Table
1), based on the underlying assumptions about patterns of attitude constraint in the of the models
they put forward. A piece is listed in the “horizontal” constraint category — to borrow terminology
from Holsti 1992 — if it assumes and looks for clusters of related policy attitudes along one or
several dimensions (typically recovered through factor analysis). We classify a piece as “vertical”
if it assumes that underlying predispositions or values determine individual policy preferences or
other more specific orientations toward foreign policy — these pieces look for evidence of top-down
constraint, often with regression-based or structural equation modeling approaches.
Differentiating between these two approaches not only informs our understanding of the assumptions that we investigate in our directed multigraph experiment, but also allows us to assess
our claim that much of this research occurs in parallel and thus misses the opportunity to study the
system of attitudes as a whole. Thus, once the pieces were categorized, we examined their citation
patterns. If there is little overlap between the citation networks of pieces that employ horizontal or
vertical models, this offers further support for our claim that the insights from each are not being
combined holistically — a problem that we seek to rectify in our article.
1
A piece of research is included in the literature review if it meets several criteria. First, it must have been published
in 1970 or later. Second, the piece of research must be published in an academic journal, as a book, or as part of an
edited volume, thus excluding conference papers. Third, the piece is included only if it makes a specific argument about
the structural basis of foreign policy attitudes at the individual level, positing certain dimensions or attitudes that shape
the foreign policy beliefs of the elites or the public. This criterion means that the list excludes reviews of previously
published work (e.g. Holsti, 1992; Russett, Hartley and Murray, 1994), those investigating the influence of personality
and other individual differences on foreign policy attitudes (e.g. Federico, Golec and Dial, 2005), models of attitude
structure that focus solely or primarily on domestic politics (e.g. Nie and Andersen, 1974; Peffley and Hurwitz, 1985),
and the robust body of work that looks at the aggregate longitudinal dynamics of the public’s foreign policy ‘mood’ (e.g.
Caspary, 1970; Page and Shapiro, 2010).
2
Table 1: Foreign policy attitude structure model pieces, 1970-2013
Citation
H
Bardes and Oldendick (1978)
✓
✓
Bartels (1994)
Bennett (1974)
✓
✓
Binning (2007)
Bjereld and Ekengren (1999)
V
✓
Citation
H
V
Jenkins-Smith, Mitchell and Herron (2004)
✓
Liberman (2006)
✓
Liberman (2007)
✓
Liberman (2013)
✓
Maggiotto and Wittkopf (1981)
✓
Brewer and Steenbergen (2002)
✓
Mandelbaum and Schneider (1978)
✓
Brewer et al. (2004)
✓
Mandelbaum and Schneider (1979)
✓
Brody and Verba (1972)
✓
Mayton, Peters and Owens (1999)
✓
Chanley (1999)
✓
Modigliani (1972)
Chittick and Billingsley (1989)
✓
Murray (1996)
Chittick, Billingsley and Travis (1990)
✓
Murray, Cowden and Russett (1999)
✓
Chittick, Billingsley and Travis (1995)
✓
Oldendick and Bardes (1981)
✓
✓
✓
Cohrs and Moschner (2002)
✓
Oldendick and Bardes (1982)
✓
Cohrs et al. (2005)
✓
Patchen (1970)
✓
Cohrs et al. (2007)
✓
Peffley and Hurwitz (1992)
✓
Fite, Genest and Wilcox (1990)
✓
Peffley and Hurwitz (1993)
✓
Rathbun (2007)
✓
Herrmann and Keller (2004)
✓
✓
Herron and Jenkins-Smith (2002)
Reifler, Scotto and Clarke (2011)
✓
Hinckley (1988)
✓
Richman (1996)
✓
Hinckley (1992)
✓
Richman, Malone and Nolle (1997)
✓
Holsti and Rosenau (1976)
✓
Rosati and Creed (1997)
✓
Holsti (1979)
✓
Rosati, Link and Creed (1998)
✓
Holsti and Rosenau (1979)
✓
Rosenau and Holsti (1983)
✓
Holsti and Rosenau (1981)
✓
Russett and Hanson (1975)
✓
Holsti and Rosenau (1984)
✓
Russett et al. (1990)
✓
Holsti and Rosenau (1986)
✓
Sulfaro (1996)
✓
Holsti and Rosenau (1988)
✓
Vail and Motyl (2010)
✓
Holsti and Rosenau (1990)
✓
Wittkopf (1981)
✓
Holsti and Rosenau (1993)
✓
Wittkopf and Maggiotto (1983a)
✓
Holsti and Rosenau (1996)
✓
Wittkopf and Maggiotto (1983b)
✓
Holsti (2004)
✓
Wittkopf (1986)
✓
Hurwitz and Peffley (1987a)
✓
Wittkopf (1987)
✓
Hurwitz and Peffley (1987b)
✓
Wittkopf (1990)
✓
Hurwitz and Peffley (1990)
✓
Wittkopf (1994)
✓
Hurwitz, Peffley and Seligson (1993)
✓
Ziegler (1987)
✓
Note: H and V refer to horizontal and vertical models of attitude structure, respectively.
3
1.1
Latent position cluster model of citation patterns
As Table 1 shows, of the 70 pieces, 47 adopt the tenets of the horizontal model, and 23 contain
vertical models. The predominance of horizontal models in Table 1, however, belies the growing
popularity of vertical models over time: while 27 of the 29 pieces published before 1990 adhere to
the assumptions of horizontal models, a full 12 of the 15 works published after 2000 employ vertical
models. Yet although horizontal and vertical models make different assumptions about how attitudes are structured, the tension between these assumptions is rarely discussed, perhaps because
proponents of each type of model are focusing more on rejecting the null hypothesis of no constraint, rather than adjudicating between the competing claims of the directionality of constraint or
conducting research into the system as a whole. However, it is striking that if we map out how these
70 pieces listed in Table 1 cite one another and analyze the clustering in the citation network using
a latent position cluster model (LPCM) (Handcock, Raftery and Tantrum, 2007), the LPCM can
correctly classify which “camp” a piece falls into with 71.4% accuracy. Figure 1 plots the citation
network in the first two dimensions of a three-dimensional Euclidean latent space; the pie chart
corresponding to each article denotes the posterior probability (calculated using MCMC sampling)
of the article being located in a particular cluster (red depicting horizontal models, and green depicting vertical models). Despite occasional misclassifications, the LPCM is nonetheless relatively
accurate, correctly classifying 73.9% of the vertical models, and 70.2% of the horizontal models —
with relative accuracy, then, the model can predict which assumptions a piece makes about how
foreign policy attitudes are structured based solely on citation patterns, even without looking at the
content of the pieces themselves. This supports our expectations that there is insufficient integration across the literature, and that a systemic perspective can profitably advance insights from across
these parallel research traditions.
4
Figure 1: The citation network of foreign policy attitude structure research, 1970-2013
2
Cohrs.2002
Russett.1975
Vail.2010
1
Hurwitz.1987a
Bardes.1978
Modigliani.1972
Binning.2007
0
Brody.1972
-1
Second principal component of Z
3
Patchen.1970
Maggiotto.1981
Wittkopf.1981
Mandelbaum.1978
Oldendick.1982
Oldendick.1981
Bennett.1974
Wittkopf.1983a
Ziegler.1987
Holsti.1979b
+
Chittick.1989
Holsti.1979a Holsti.1977
Holsti.1981
Wittkopf.1990
Chittick.1995
Holsti.1984
Mandelbaum.1979 Rosati.1997
Rosenau.1983
Holsti.1988
Hurwitz.1987bWittkopf.1987
Peffley.1993
Chanley.1999
Cohrs.2005
Hurwitz.1990
Hurwitz.1993
Brewer.2002
Brewer.2004
Cohrs.2007
Liberman.2013
Liberman.2006
Hinckley.1992
+
Holsti.1996
Herron.2002Bartels.1994
Hinckley.1988
+Holsti.1990
Rathbun.2007
Liberman.2007
Russett.1990
Murray.1999
Murray.1996
Holsti.2004.1996.
Wittkopf.1986
Wittkopf.1983b
Holsti.1986
Bjereld.1999
Wittkopf.1994
Jenkins.Smith.2004
Fite.1990
Peffley.1992
Herrmann.2004
Mayton.1999
Reifler.2011
Chittick.1990
Rosati.1998
Holsti.1993
Sulfaro.1996
-2
Richman.1996
Richman.1997
-2
0
2
4
First principal component of Z
By examining the citation patterns between the 70 foreign policy attitude network pieces listed in Table 1, a latent
position cluster model (LPCM) can successfully predict which “camp” each piece falls into with 71.4% accuracy. An
arrow from node i to node j shows that article j cites article i: thus, for example Patchen (1970) is cited by Russett and
Hanson (1975). The pie chart for each article represents the posterior probability (estimated using 5000 MCMC
samples) that each article falls into a particular camp; here, the red represents horizontal models, and the green
represents vertical models. The x and y axes represent the first two principal components in a three-dimensional
Euclidean latent space as measured by the Minimum Kullback-Leibler (MKL) estimates. The LPCM is estimated using
the latentnet package in R (Krivitsky and Handcock, 2008).
5
2
Materials and Instrumentation
2.1
2.1.1
Text of Persuasive Messages
Values
Authority/respect/hierarchy
Philosophers throughout history have stressed that people should show respect for authority figures in a society. Authority figures represent the wisdom of their position and can help people have
better lives by protecting them. They also act as exemplars, such that people can learn from them
about how to achieve success in their own lives. Prominent psychologists state that many cultures
value “virtues related to good leadership, which involve[s] magnanimity, fatherliness, and wisdom.”
This emphasis has been echoed by research in anthropology and sociology, which demonstrates the
importance of hierarchies and traditions to most cultures. Respecting legitimate authority figures is
considered virtuous and fosters a happy and healthy environment for all members of a community.
For these reasons, respecting persons in positions of authority is widely regarded as good for both
individuals and society.
Fairness/reciprocity/community
Philosophers throughout history have stressed that people should treat all other people just as they
wish to be treated themselves. More contemporary evidence suggests that using this ‘golden rule’
as a guide will produce increasing levels of success in business practices and relationships because
it helps to build trust. Prominent psychologists state that “all cultures have developed virtues related to fairness.” This emphasis has been echoed by research in anthropology and sociology, which
demonstrates the importance of reciprocity and justice to the cores of societies all over the world.
Adherence to such virtues helps to foster happy and healthy environments for all members of a community. For these reasons, equal and just treatment of other humans is widely regarded as good for
both individuals and society.
Harm/care
Philosophers throughout history have stressed that people should avoid doing harm to others, and
seek to alleviate suffering where possible. This principle is enshrined in contemporary medical practice, where doctors vow to “first, do no harm.” Prominent psychologists further state that people in
most cultures “feel approval to toward those who prevent or relieve harm.” This emphasis has been
echoed by research in anthropology and sociology, which demonstrates the importance of kindness
as a virtue in societies throughout the world. Adherence to such virtues helps to foster happy and
6
healthy environments for all members of a community. For these reasons, kind and compassionate,
rather than harmful, treatment of other humans is widely regarded as good for both individuals and
society.
Ingroup/loyalty
Philosophers throughout history have stressed that people should stay loyal to members of groups
to which they belong. Other members of the same group can be trusted and are easy to cooperate
with. Thus, it is considered virtuous to stay loyal to one’s group and to make sacrifices to protect and
improve the group as a whole. Prominent psychologists state that “most cultures have constructed
virtues such as loyalty, patriotism, and heroism.” This emphasis has been echoed by research in
anthropology and sociology, which demonstrates the importance of group solidarity to strong societies. Adherence to such virtues fosters a happy and healthy environment for all members of a
community. For these reasons, being a loyal member of one’s group is widely regarded as good for
both individuals and society.
2.1.2
Orientations
Militant Internationalism
Many countries and international leaders construct policy in order to have a positive impact on the
world. Often, this requires military force on the part of one or both states. For example, countries
sometimes need to use their military to stop another country from starting a war, or to intervene
when there is no other way to solve a dispute. Scholars and diplomats alike recognize the benefits
of using military force under certain conditions, with one prominent leader recently stating that “to
say that force may sometimes be necessary... is recognition of history,” and is often essential for
maintaining order in the world. Countries should use force when they are threatened or when it
will produce a positive outcome, and military interventions in history have been successful ways
for countries to resolve their differences. Thus, many agree that the use of military force abroad is a
good foreign policy tool.
Cooperative Internationalism
Many countries and international leaders construct policy in order to have a positive impact on the
world. Often, this involves working diplomatically with other nations and international organizations to provide aid or produce an acceptable solution to a pressing global problem. For example,
states can come together to establish guidelines for the protection of human rights or to aid the eco7
nomic development of other nations. Scholars and diplomats alike recognize the benefits of such
cooperation. Regarding the provision of aid for the building of new police stations abroad to cut
down on international crime, one prominent leader recently remarked that “the solution of the most
important problems require international cooperation, and this project is a testimony of what international cooperation for development can do.” Many others agree that it is good when countries
have positive diplomatic relations with other countries and work with them and through international institutions to solve global problems.
Isolationism
Many countries and international leaders construct policy in order to have the most positive impact. Often, this involves focusing primarily on domestic issues, so that the country can primarily
establish their own national prosperity. For example, a state might focus on improving its own economy rather than becoming entangled in agreements with or the affairs of other countries. Scholars
and diplomats alike recognize the benefits of such an emphasis on domestic affairs. One prominent
leader, for example, recently stressed that the United States should focus first on “our energy, economic, and domestic policies,” as they all advance the national interest. Many others agree that it is
good when countries put their domestic priorities first and avoid entanglements abroad.
2.1.3
Policy Positions
The Iraq War
In 2003, the United States led a coalition in an invasion of Iraq due to fears that the government
possessed weapons of mass destruction. In the time that has passed since the beginning of the war,
many American citizens have come to see it as a positive intervention. The occupation of Iraq and
activities by the U.S. has helped to bring prosperity to an otherwise faltering nation. The country
has also become increasingly more stable over time. One prominent US leader noted this recently,
stating that he “agrees with the invasion of Iraq” because it “was the right thing to do” and that all
Americans should “desire to see signs of hope across Iraq flourish.” With time, then, we have come
to know that this intervention was crucial in stopping the potential military threat from Iraq and
the various human rights abuses under Saddam Hussein. There would have been substantially more
violence and instability in the state without the intervention of the United States and other coalition
forces.
NATO Air strikes on Libya
In early 2011, armed conflict emerged in Libya between supporters of the current leader Muam8
mar Gaddafi and rebels who sought to remove his government from power. In response, several
countries, including the United States, engaged in military action in support of the rebels. One
prominent US leader recently stated that this involvement was necessary, because “you had a potentially significantly destabilizing event taking place in Libya.” With time, we have come to know
that this intervention was crucial in putting an end to the massacre of many Libyans by oppressive
government forces. It helped to bring a swifter end to the violence, and many agree that it was a
good decision. There would have been many more deaths and dangerous regional instability without the involvement of other states.
Intervention in Syria
In early 2011, armed conflict emerged in Syria between supporters of the current government led by
President Bashar al-Assad and protesters who sought to remove his government from power. The
government crackdown against the protestors and ongoing civil war has resulted in some 60,000
deaths. As a consequence, many prominent analysts and world leaders have called for the US to
intervene in the conflict, by directly arming rebel soldiers or even sending their own troops, despite
the disapproval voiced by fellow United Nations Security Council members China and Russia. One
prominent leader and analyst, stresses this when he states that “we must certainly consider direct
military action” by “[using] force ourselves, or see the Assad regime murder its way to victory.” With
time, it has become evident to many that an intervention will be the only way to put an end to the
violence. An intervention could prevent more deaths and dangerous regional instability.
Kyoto Protocol
In 1997, delegates from 194 countries met in Kyoto, Japan and crafted an agreement to reduce the
amount of greenhouse gas emissions throughout the world. This agreement, the Kyoto Protocol,
expired at the end of 2012, and countries have not yet established a new treaty to replace it. Many
analysts and leaders argue that the United States should become heavily involved in negotiating a
new treaty designed to curb greenhouse gas emissions and protect the global environment. One
prominent leader in the United States stresses this point when he states that “we are ready, willing,
and able” to have a discussion about establishing a new climate agreement. US support for and
promotion of a new treaty will be crucial to its future success.
Control: Online Retail
Internet retail has emerged in recent years as a prominent mode of commerce. One of the most
important ways that sellers increase the number of people interested in purchasing their product
9
is by earning positive customer reviews. These are easily accessible by individuals while they are
shopping via the internet, and many people report using them to determine their ultimate purchase.
As a result, the importance of online retail is only expected to increase. Time Magazine recently
announced that online shopping was the holiday season’s top shopping trend, and $1.25 billion was
spent online on Cyber Monday alone.
2.2
Instrumentation
2.2.1
Values: Moral Foundations Questionnaire
We measure participants’ moral values using four of the moral values from the Moral Foundations
Questionnaire (Graham et al., 2011). As is standard with the MFQ, each moral foundation is measured with six items, the first three of which are responses to the question “When you decide whether
something is right or wrong, to what extent are the following considerations relevant to your thinking?” These items have 6 point Likert responses ranging from “not at all relevant” to “extremely
relevant.” The second three items for each foundation are 6 point Likert responses from “Strongly
disagree” to “Strongly agree.” Items are presented to participant in two blocks – one with the first
3 items from each scale and one with the latter 3 items from each scale – with the order of items
randomized within blocks.
Authority/Respect(α = 0.719)
1. Whether or not someone showed a lack of respect for authority
2. Whether or not someone conformed to the traditions of society
3. Whether or not an action caused chaos or disorder
4. Respect for authority is something all children need to learn.
5. Men and women each have different roles to play in society.
6. If I were a soldier and disagreed with my commanding officer’s orders, I would obey anyway
because that is my duty.
Fairness/Reciprocity (α = 0.666)
1. Whether or not some people were treated differently than others
2. Whether or not someone acted unfairly
3. Whether or not someone was denied his or her rights
10
4. When the government makes laws, the number one principle should be ensuring that everyone is treated fairly.
5. Justice is the most important requirement for a society.
6. I think it’s morally wrong that rich children inherit a lot of money while poor children inherit
nothing.
Harm/Care (α = 0.710)
1. Whether or not someone suffered emotionally
2. Whether or not someone cared for someone weak or vulnerable
3. Whether or not someone was cruel
4. Compassion for those who are suffering is the most crucial virtue.
5. One of the worst things a person could do is hurt a defenseless animal.
6. It can never be right to kill a human being.
Ingroup/Loyalty (α = 0.727)
1. Whether or not someone’s action showed love for his or her country
2. Whether or not someone did something to betray his or her group
3. Whether or not someone showed a lack of loyalty
4. I am proud of my country’s history.
5. People should be loyal to their family members, even when they have done something wrong.
6. It is more important to be a team player than to express oneself.
11
2.2.2
Orientations
Each item has a seven-point Likert response scale ranging from “Strongly disagree” to ”Strongly
agree”. Items for each orientation are presented together in a block, with block order randomized.
Militant internationalism (α = 0.783)
1. The United States should take all steps including the use of force to prevent aggression by any
expansionist power
2. Rather than simply countering our opponents’ thrusts, it is necessary to strike at the heart of
an opponent’s power.
3. Going to war is unfortunate but sometimes the only solution to international problems.
4. In dealing with other nations our government should be strong and tough.
Cooperative internationalism (α = 0.789)
1. The United States needs to cooperate more with the United Nations.
2. It is essential for the United States to work with other nations to solve problems such as overpopulation, hunger, and pollution.
3. In dealing with other nations our government should be understanding and flexible.
4. The best way to ensure peace is to sit down with other countries and work out our disagreements.
Isolationism (α = 0.814)
1. The U.S. should mind its own business internationally and let other countries get along the
best they can on their own.
2. We should not think so much in international terms but concentrate more on our own national problems.
3. The U.S. needs to play an active role in solving conflicts around the world. (Reverse-coded)
4. America’s conception of its leadership role in the world must be scaled down.
12
2.2.3
Policy Positions
The Iraq War (α = 0.788)
1. All things considered, do you approve or disapprove of the U.S. decision to intervene in Iraq?
(7-point scale from “Strongly disapprove” to “Strongly approve”)
2. Considering the costs to the United States versus the benefits to the United States, do you think
the war in Iraq was worth fighting, or not? (5-point scale from “Not worth it” to “Worth it”)
NATO Air Strikes on Libya (α = 0.764)
1. All things considered, do you approve or disapprove of the decision of the U.S. and its allies to
conduct military air strikes in Libya? (7-point scale from “Strongly disapprove” to “Strongly
approve”)
2. Do you think the U.S. did the right thing by using military force in Libya now, or should
the US not have been involved in Libya? (4-point scale from “Definitely the wrong thing” to
“Definitely the right thing”)
Intervention in Syria (α = 0.742)
1. Would you favor or oppose the U.S. and its allies sending arms and military supplies to antigovernment groups in Syria? (7-point scale from “Strongly oppose” to “Strongly favor”)
2. Would you approve or disapprove of the use of U.S. forces to protect the anti-government
protesters in Syria? (7-point scale from “Strongly disapprove” to “Strongly approve”)
Kyoto Protocol (α = 0.827)
1. The Kyoto Protocol, the global treaty designed to curb greenhouse gas emissions, expired in
2012. Do you approve or disapprove of the United States working closely with other nations
to create a new international treaty to fight global warming? (7-point scale from “Strongly
disapprove” to “Strongly approve”)
2. How much do you think the U.S. government should do about global warming? (5-point scale
from “Nothing” to “A great deal”)
13
3
Scale Reliabilities
Table 2: Scale reliabilities
Scale
Cronbach’s α
Fairness
α = 0.666
Authority
α = 0.719
Harm
α = 0.710
In-group
α = 0.727
Militant internationalism
α = 0.783
Cooperative internationalism
α = 0.789
Isolationism
α = 0.814
Iraq war
α = 0.788
Libya
α = 0.764
Syria
α = 0.742
Kyoto
α = 0.827
Foreign policy interest
α = 0.907
Political knowledge
α = 0.684
14
4
Thresholding and sensitivity analyses
A central question in network analysis is how to define the network’s edges, particularly when some
types of relations are less easily dichotomized than others, and thus have more ambiguous criteria
as to what constitutes an edge. Typically network analysts make two choices: either include all
potential edges and simply weight them by their strength (Thomas and Blitzstein, 2011), or threshold
based a particular cut point. As illustrated in Figure 2 below, in the main analyses in the text, we
employ a cut point of 10%, taking the top decile of treatment effects in our experimental results
(shaded in the darker grey) and using them to generate our networks. We represent our treatment
effects with t-statistics, since they capture both the difference in means between the treatment and
control groups and also their sample sizes and standard deviations, but a variety of other statistics
could be used — the choice of which depends on the nature of the analyses.2 In order to ensure that
the patterns of results we present above are not simply artifacts of this 10% cut point, we conduct a
series of sensitivity analyses below probing the robustness of our two main findings (the prevalence
of bottom-up constraint, and the network density’s curvilinear relationship with political knowledge
and interest) to our thresholding criteria.
In the sensitivity analyses, we generated six different versions of the network, defining the edging
threshold as the top 1%, 5%, 10%, 15%, 20%, and top 25% of the treatment effects, respectively. As
one would expect, the more restrictive the threshold, the fewer edges are included in the adjacency
matrix, and the less dense the networks, while the less restrictive the threshold, the more edges
are included in the adjacency matrix, and the more dense the networks. Figure 3 plots the total
degree centrality, indegree centrality, and outdegree centrality scores for core values (depicted as
circles) and policy attitudes (depicted as triangles) across all of the attitude networks generated at
each effect threshold.3 If our networks are characterized by bottom-up constraint, we should see
that policy attitudes consistently play a larger role in our networks than core values — that is, the
triangles should be higher than the circles. If, however, our networks are characterized by top-down
constraint, we should see that core values consistently play a larger role in our networks than core
values — that is, the circles should be higher than the triangles. As the three panels of Figure 3 make
clear, although the overall centrality scores for both core values and policy attitudes increase as less
restrictive effect thresholds are employed, the centrality scores for policy attitudes are always higher
2
W statistics from nonparametric rank sum tests, for example, are a function of sample size, so simply thresholding the network based on the effects with the largest W statistics would be inappropriate when the effects are drawn
from samples of different sizes. Similarly, if analysts were interested in employing local fdr or FDR to control the false
discovery rate, the statistics in the distribution should be z statistics and p values from t-tests, respectively.
3
The analyses add across all of the different networks we generate in the piece (the network for the full sample, the
liberal and conservative networks, the low, medium and high knowledge networks, etc.), but the results also hold when
the subsample networks are dropped and only the full network is considered.
15
0.2
0.0
0.1
Density
0.3
0.4
Figure 2: Distribution of effect sizes
-4
-3
-2
-1
0
1
2
3
Effect size (t statistics)
The histogram plots the distribution of effect sizes, represented here using t-statistics. The largest 10% of the treatment
effects are shaded, indicating those effects for which edges are drawn in the attitude networks. Other thresholds can
also be applied, but the key point is that significance is inferred not in reference to particular parametric assumptions
or assumed populations, but rather in relation to the observed data itself.
than for core values – and in this sense, we find strong support for bottom-up constraint regardless
of the effect threshold used in the analyses.
Figure 4 probes the robustness of the other main finding from the piece: the curvilinear relationship that knowledge and political interest have with attitude network density. In the main
analyses in the text that use a 10% effect threshold, we find that medium knowledge respondents
displayed the densest attitude networks compared to their low- and high-knowledge peers, while
medium-interest respondents displayed less dense attitude networks compared to their low- and
high-interest counterparts. As in the analyses reported in Figure 3, we generated six different versions of each network (using a 1%, 5%, 10%, 15%, 20% and 25% effect threshold), this time focusing
on the low, medium, and high political knowledge networks and the low, medium and high political
interest networks. As Figure 4(a) shows, as long as the effect threshold is greater than 1% (whereupon all three knowledge networks have an equally low density), we replicate the same findings
16
detected with a 10% threshold, with the medium knowledge network displaying a greater number
of edge than either the low or the high knowledge networks. Similarly, Figure 4(b) shows that as
long as the effect threshold is greater than 1%, we see an inverse-U shaped relationship between
foreign policy interest and network density. Together, Figures 3 and 4 show that even as the individual edges included in these networks change as the edge threshold is tightened or relaxed, the
network-level properties (the prevalence of bottom-up constraint, and the curvilinear relationships
between network density and political knowledge and interest) persist, mitigating concern about
the effect of measurement error.
1%
5%
10%
15%
Effect threshold
20%
25%
100
80
60
Outdegree Centrality
0
20
40
100
80
60
0
0
20
40
Indegree centrality
200
150
100
50
Degree centrality
120
Figure 3: Sensitivity analyses: relative centrality of core values versus policy orientations
1%
5%
10%
15%
Effect threshold
20%
25%
1%
5%
10%
15%
20%
25%
Effect threshold
The sensitivity analyses show robust evidence for bottom-up constraint: core values (depicted by circles) consistently
have lower centrality scores than specific policy attitudes (depicted by triangles).
5
5.1
Supplementary Subgroup Analyses
Gender
The main text presents an analysis of the full network along with networks separated by political
ideology and sophistication. However, these are not the only dimensions of individual difference
upon which we expect structural variance. An individual’s gender, too, might impact the type and
direction of foreign policy attitude constraint. Some previous scholarship has shown that the foreign policy attitudes of men and women occasionally diverge in limited ways (such as with regard
to defense spending) but can be captured by the same dimensions (Holsti and Rosenau, 1981) or
are determined by the same hierarchically constraining elements (Fite, Genest and Wilcox, 1990).
17
Figure 4: Sensitivity analyses: curvilinear effect of knowledge and interest on network density
(b) Interest in politics
0.25
(a) Political knowledge
25%
0.20
0.3
25%
20%
0.15
15%
10%
0.10
Density
15%
0.2
Density
20%
10%
0.05
0.1
5%
5%
0.0
Low
Medium
1%
0.00
1%
High
Knowledge
Low
Medium
Foreign Policy Interest
The sensitivity analyses show consistent evidence for quadratic effects of political knowledge (panel a) and interest
(panel b) on network density once the effect threshold is above 1%; effect thresholds are labeled on the right-hand side
of each panel.
Other work in psychology, however, suggests that women and men develop and rely on different
sets of values to make judgments (Gilligan and Attanucci, 1988; Jaffee and Shibley Hyde, 2009).
Thus, Figure 5 disaggregates the foreign policy networks for women and men, respectively. We
find that the Female network is driven largely by attitudes about the airstrikes in Libya, and more
generally that policy positions serve as the most important sources of constraint. Attitudes toward
Libya influence two core values as well as positions on Syria, and the nodes representing attitudes toward the war in Iraq, Syria, and the Kyoto Protocol each have outgoing connections. The permuted
analyses confirm the importance of specific policy positions to the female network, demonstrating
that there are more outgoing edges than we would otherwise expect (p = 0.014). The Male network,
in contrast, has more evenly distributed connections — values, policies, and orientations are about
equally central, suggesting that specific policy events may be more important sources of constraint
for women than for men. These distinctions may help to explain why the gender gap in foreign
policy attitudes seems to fluctuate based on the issue and time, with men and women coming to
their states attitudes via different pathways (Holsti, 2004). Additionally, we see patterns that are
only evident from a systemic perspective, and that challenge the assumptions implicit in previous
work: attitudes at the same level of analysis influence each other (as with harm and authority for
18
High
men, or the reciprocal connection between Libya and Syria for women), and we have already noted
that constraint in the female network works primarily from the bottom up.
Figure 5: Foreign policy attitude network by gender
(a) Female
(b) Male
MI
Fair
Kyoto
Iraq
Fair
Syria
Ingroup
Libya
Auth
Harm
Harm
Libya
Iso
CI
Ingroup
Iraq
Auth
Kyoto
Syria
Iso
MI
CI
Note: Arrows represent the patterns of attitude constraint; nodes are scaled by their degree centrality, so larger nodes
are connected to a higher number of other attitudes. Core values are represented as circles, foreign policy orientations
as diamonds, and policy attitudes as triangles. Neither network conforms with patterns expected by either the
horizontal or the vertical models: attitudes at the same level of abstraction influence one another (e.g. harm and
authority, for men), and we see signs of bottom-up constraint (e.g. Libya and ingroup, for women).
19
5.2
Political Ideology
The psychological literature on political ideology tells us that liberals and conservatives believe in
fundamentally different things (e.g. they rely on different moral foundations — Graham, Haidt and
Nosek 2009) for fundamentally different reasons (e.g. they maintain their belief systems in response
to strikingly different social-cognitive motivations — Jost et al. 2003). In contrast to “one-size fits
all” models of attitude structure, we thus expect that liberals and conservatives would organize their
attitudes differently. As shown by Figure 6, liberals and conservatives maintain distinct foreign policy attitude networks, with CI central to the conservative network, and isolationism situated most
centrally in the liberal network. Whereas attitudes toward the airstrikes in Libya, renewing the Kyoto Protocol, and intervening in Syria constrain CI for conservatives, they influence isolationism
for liberals. One thing that the liberal and conservative networks share, then, is the prominence of
bottom-up patterns of constraint going from policy positions to general orientations: for conservatives, for example, values shape fewer attitudes (p = 0.097) and policies shape more (p = 0.062)
than expected by chance. The directions of the observed links, along with the existence of links
between orientations such as CI and isolationism for liberals, are inconsistent with the directional
assumptions from past work. Interestingly, while the structural patterns of constraint clearly differ
in these networks, their overall topology is relatively similar, and Hamming distance scores suggest
that these networks are the least distinct from one another of all of the subgroup networks presented
here — indicating perhaps that other individual level characteristics, such as political sophistication,
are more important for understanding the underlying variation in foreign policy attitude networks.
6
Additional permutation tests
The main analyses employ permutation tests that examine the probability of particular node-level
attributes by randomly generating 10000 networks of an identical size and density as our observed
network, thereby observing the probability that, for example, core values would display a particular level of constraint due to chance alone.4 While conditioning the permutation test on observed
network densities allows us to determine the probability of given node-level statistics, it does not
let us make inferences about the probability of observing given densities themselves. Thus, Figure
7 displays the results from a permutation test to evaluate the density of each network, randomly
generating 99000 networks of an identical size of each observed network, in which the probability
of each edge forming is varied from 0.01 to 0.99 every 1000 permutations. This technique is used
4
As Butts (2008, fn. 5) notes, this type of permutation test is popular amongst network analysts, although the
terminology used to describe the test often varies: Butts refers to them as Conditional Uniform Graph Tests.
20
Figure 6: Foreign policy attitude network by political ideology
(a) Liberals
(b) Conservatives
Iraq
Ingroup
Iso
Libya
Syria
Kyoto
Kyoto
MI
Harm
Iso
Libya
CI
Fair
CI
Syria
Iraq
Fair
Harm
MI
Auth
Auth
Ingroup
Note: Arrows depict the thresholded treatment effects, and thus, patterns of attitude constraint; nodes are scaled by
their degree centrality. Core values are represented as circles, foreign policy orientations as diamonds, and policy
attitudes as triangles. Thus, we see that liberals and conservatives display different patterns of attitude constraint — for
example, CI shapes isolationism for liberals, but isolationism shapes CI for conservatives — but once again, both
attitude networks show considerable bottom-up patterns of constraint, as specific policy attitudes shape foreign policy
orientations and core values.
to test two different null hypotheses: a random null hypothesis in which each edge has an identical probability of forming, and an informative null hypothesis in which the edge probabilities are
constrained to 0 according to the assumptions of the hierarchical model – that is, all constraint is
top-down rather than bottom-up or within-levels. Thus, Panel (a) of Figure 7 shows that against
a random null hypothesis, we are unlikely to see the observed density for our full network if the
probability of any given edge forming is greater than 0.1. Similarly, Panel (b) shows that against a
hierarchical null hypothesis, we are unlikely to see the observed density for our full network if the
probability of any given edge forming is less than 0.1 or greater than 0.3.
21
22
Network
11. H Knowledge
10. M Knowledge
09. L Knowledge
08. H Interest
07. M Interest
06. L Interest
05. Female
04. Male
03. Conservative
02. Liberal
01. Full
11. H Knowledge
10. M Knowledge
09. L Knowledge
08. H Interest
07. M Interest
06. L Interest
05. Female
04. Male
03. Conservative
02. Liberal
01. Full
0.0
0.0
0.1
0.1
0.2
0.2
0.3
0.3
0.5
Edge probability
0.6
0.4
0.5
Edge probability
0.6
(b) Informative (hierarchical) null
0.4
(a) Random null
0.7
0.7
Figure 7: Permutation tests of network density
0.8
0.8
0.9
0.9
1.0
1.0
0.1
0.05
0.5
p value
0.1
0.05
0.5
p value
the expected probability that we would see networks of the observed density for each edge probability.
commensurate with top-down patterns of constraint (e.g. all edges from policy attitudes to core values, for example, are constrained at 0). These heatmaps thus plot
corresponding to the predictions from a hierarchical model, in which each potential edge is drawn with the probability denoted on the x-axis only if the edge is also
each potential edge is drawn with the probability denoted on the x-axis, while the bottom panel shows the results from an informative null hypothesis
the probability of each edge forming varies from 0.01 to 0.99 every 1000 permutations. The top panel shows the results from a random null hypothesis in which
The figures plot the results of permutation tests that compare each observed network density with 99000 randomly generated networks of identical size, in which
Network
References
Bardes, Barbara and Robert Oldendick. 1978. “Beyond Internationalism: A Case for Multiple Dimensions in the Structure of Foreign Policy Attitudes.” Social Science Quarterly 59(3):496–508.
Bartels, Larry M. 1994. “The American public’s defense spending preferences in the post-cold war
era.” Public Opinion Quarterly 58(4):479–508.
Bennett, Stephen Earl. 1974. “Attitude structures and foreign policy opinions.” Social Science Quarterly 55(3):732–742.
Binning, Kevin R. 2007. ““It’s Us Against the World”: How Distrust in Americans versus People-InGeneral Shapes Competitive Foreign Policy Preferences.” Political Psychology 28(6):777–799.
Bjereld, Ulf and Ann-Marie Ekengren. 1999. “Foreign policy dimensions: A comparison between
the United States and Sweden.” International Studies Quarterly 43(3):503–518.
Brewer, Paul R, Kimberly Gross, Sean Aday and Lars Willnat. 2004. “International trust and public
opinion about world affairs.” American Journal of Political Science 48(1):93–109.
Brewer, Paul R and Marco R Steenbergen. 2002. “All against all: How beliefs about human nature
shape foreign policy opinions.” Political Psychology 23(1):39–58.
Brody, Richard A and Sidney Verba. 1972. “Hawk and dove: the search for an explanation of Vietnam policy preferences.” Acta Politica 7:285–322.
Butts, Carter T. 2008. “Social Network Analysis: A Methodological Introduction.” Asian Journal of
Social Psychology 11:13–41.
Caspary, William R. 1970. “The “mood theory”: A study of public opinion and foreign policy.” The
American Political Science Review 64(2):536–547.
Chanley, Virginia A. 1999. “US Public Views of International Involvement from 1964 to 1993 TimeSeries Analyses of General and Militant Internationalism.” Journal of Conflict Resolution 43(1):23–
44.
Chittick, William O. and Keith R. Billingsley. 1989. “The Structure of Elite Foreign Policy Beliefs.”
The Western Political Quarterly 42(2):201–224.
Chittick, William O, Keith R Billingsley and Rick Travis. 1990. “Persistence and change in elite and
mass attitudes toward US foreign policy.” Political psychology pp. 385–401.
Chittick, William O, Keith R Billingsley and Rick Travis. 1995. “A three-dimensional model of
American foreign policy beliefs.” International Studies Quarterly pp. 313–331.
Cohrs, J Christopher and Barbara Moschner. 2002. “Antiwar knowledge and generalized political
attitudes as determinants of attitude toward the Kosovo War.” Peace and Conflict: Journal of Peace
Psychology 8(2):139–155.
Cohrs, J Christopher, Barbara Moschner, Jurgen Maes and Sven Kielmann. 2005. “Personal values
and attitudes toward war.” Peace and Conflict 11(3):293–312.
Cohrs, J Christopher, Jürgen Maes, Barbara Moschner and Sven Kielmann. 2007. “Determinants
of human rights attitudes and behavior: A comparison and integration of psychological perspectives.” Political Psychology 28(4):441–469.
Federico, Christopher M, Agnieszka Golec and Jessica L Dial. 2005. “The relationship between
the need for closure and support for military action against Iraq: Moderating effects of national
attachment.” Personality and Social Psychology Bulletin 31(5):621–632.
Fite, David, Marc Genest and Clyde Wilcox. 1990. “Gender Differences in Foreign Policy Attitudes
A Longitudinal Analysis.” American Politics Research 18(4):492–513.
Gilligan, C. and J. Attanucci. 1988. “Two Moral Orientations: Gender Differences and Similarities.”
23
Merrill-Palmer Quarterly 33:223–237.
Graham, Jesse, Brian A. Nosek, Jonathan Haidt, Ravi Iyer, Spassena Koleva and Peter H. Ditto. 2011.
“Mapping the Moral Domain.” Journal of Personality and Social Psychology 101(2):366–385.
Graham, Jesse, Jonathan Haidt and Brian A. Nosek. 2009. “Liberals and Conservatives Rely on
Different Sets of Moral Foundations.” Journal of Personality and Social Psychology 96(5):1029–
1046.
Handcock, Mark S., Adrian E. Raftery and Jeremy M. Tantrum. 2007. “Model-based clustering for
social networks.” Journal of the Royal Statistical Society: Series A 170(2):301–354.
Herrmann, Richard K and Jonathan W Keller. 2004. “Beliefs, values, and strategic choice: US
leaders’ decisions to engage, contain, and use force in an era of globalization.” Journal of Politics 66(2):557–580.
Herron, Kerry G and Hank C Jenkins-Smith. 2002. “US perceptions of nuclear security in the
wake of the Cold War: Comparing public and elite belief systems.” International Studies Quarterly
46(4):451–479.
Hinckley, Ronald H. 1988. “Public attitudes toward key foreign policy events.” Journal of Conflict
Resolution 32(2):295–318.
Hinckley, Ronald H. 1992. People, polls, and policymakers: American public opinion and national
security. Free Press.
Holsti, Ole R. 1979. “The three-headed eagle: The United States and system change.” International
Studies Quarterly pp. 339–359.
Holsti, Ole R. 1992. “Public opinion and foreign policy: Challenges to the Almond-Lippmann Consensus Mershon Series: Research programs and debates.” International Studies Quarterly pp. 439–
466.
Holsti, Ole R and James N Rosenau. 1976. “Meaning of Vietnam: Belief Systems of American
Leaders, The.” International Journal 32:452.
Holsti, Ole R and James N Rosenau. 1979. “Vietnam, consensus, and the belief systems of American
leaders.” World Politics 32(01):1–56.
Holsti, Ole R and James N Rosenau. 1981. “The foreign policy beliefs of women in leadership
positions.” The Journal of Politics 43(02):326–347.
Holsti, Ole R and James N Rosenau. 1984. American leadership in world affairs: Vietnam and the
breakdown of consensus. Allen & Unwin Boston.
Holsti, Ole R and James N Rosenau. 1986. “Consensus Lost. Consensus Regained?: Foreign Policy
Beliefs of American Leaders, 1976-1980.” International Studies Quarterly pp. 375–409.
Holsti, Ole R and James N Rosenau. 1988. “The domestic and foreign policy beliefs of American
leaders.” Journal of Conflict Resolution 32(2):248–294.
Holsti, Ole R and James N Rosenau. 1990. “The structure of foreign policy attitudes among American leaders.” The Journal of Politics 52(01):94–125.
Holsti, Ole R and James N Rosenau. 1993. “The structure of foreign policy beliefs among American
opinion leaders-after the cold war.” Millennium-Journal of International Studies 22(2):235–278.
Holsti, Ole R and James N Rosenau. 1996. “Liberals, populists, libertarians, and conservatives: The
link between domestic and international affairs.” International Political Science Review 17(1):29–
54.
Holsti, Ole Rudolf. 2004. Public opinion and American foreign policy. Ann Arbor: University of
Michigan Press.
Hurwitz, Jon and Mark Peffley. 1987a. “How are foreign policy attitudes structured? A hierarchical
24
model.” The American Political Science Review pp. 1099–1120.
Hurwitz, Jon and Mark Peffley. 1987b. “The means and ends of foreign policy as determinants of
presidential support.” American Journal of Political Science pp. 236–258.
Hurwitz, Jon and Mark Peffley. 1990. “Public images of the Soviet Union: The impact on foreign
policy attitudes.” The Journal of Politics 52(01):3–28.
Hurwitz, Jon, Mark Peffley and Mitchell A Seligson. 1993. “Foreign policy belief systems in comparative perspective: The United States and Costa Rica.” International Studies Quarterly pp. 245–270.
Jaffee, Sara and Janet Shibley Hyde. 2009. “Gender Differences in Moral Orientation: A MetaAnalysis.” Psychological Bulletin 126:703–726.
Jenkins-Smith, Hank C., Neil J. Mitchell and Kerry G. Herron. 2004. “Foreign and Domestic Policy
Belief Structures in the U.S. and British Publics.” Journal of Conflict Resolution 48(3):287–309.
Jost, John T., Jack Glaser, Arie W. Kruglanski and Frank Sulloway. 2003. “Political Conservatism as
Motivated Social Cognition.” Psychological Bulletin 129(3):339–375.
Krivitsky, Pavel N. and Mark S. Handcock. 2008. “Fitting Position Latent Cluster Models for Social
Networks with latentnet.” Journal of Statistical Software 24(5).
Liberman, Peter. 2006. “An eye for an eye: Public support for war against evildoers.” International
Organization 60(3):687.
Liberman, Peter. 2007. “Punitiveness and US elite support for the 1991 Persian Gulf War.” Journal
of Conflict Resolution 51(1):3–32.
Liberman, Peter. 2013. “Retributive support for international punishment and torture.” Journal of
Conflict Resolution 57(2):285–306.
Maggiotto, Michael A and Eugene R Wittkopf. 1981. “American public attitudes toward foreign
policy.” International Studies Quarterly pp. 601–631.
Mandelbaum, M. and W. Schneider. 1979. The New Internationalisms. In Eagle Entangled: U.S.
Foreign Policy in a Complex World, ed. K.A. Oye. New York: Longman.
Mandelbaum, Michael and William Schneider. 1978. “The new internationalisms.” International
Security pp. 81–98.
Mayton, Daniel M. II, Danya J Peters and Rocky W Owens. 1999. “Values, militarism, and nonviolent predispositions.” Peace and Conflict 5(1):69–77.
Modigliani, Andre. 1972. “Hawks and doves, isolationism and political distrust: An analysis of
public opinion on military policy.” The American Political Science Review 66(3):960–978.
Murray, Shoon Kathleen. 1996. Anchors against change: American opinion leaders’ beliefs after the
cold war. Ann Arbor: University of Michigan Press.
Murray, Shoon Kathleen, Jonathan A Cowden and Bruce M Russett. 1999. “The convergence
of American elites’ domestic beliefs with their foreign policy beliefs.” International Interactions
25(2):153–180.
Nie, Norman H and Kristi Andersen. 1974. “Mass belief systems revisited: Political change and
attitude structure.” The Journal of Politics 36(3):540–591.
Oldendick, Robert and Barbara Ann Bardes. 1981. “Belief Structures and Foreign Policy: Comparing Dimensions of Elite and Mass Opinions.” Social Science Quarterly 62(3):432–442.
Oldendick, Robert W and Barbara Ann Bardes. 1982. “Mass and elite foreign policy opinions.”
Public Opinion Quarterly 46(3):368–382.
Page, Benjamin I and Robert Y Shapiro. 2010. The rational public: Fifty years of trends in Americans’
policy preferences. Chicago: University of Chicago Press.
Patchen, Martin. 1970. “Social class and dimensions of foreign policy attitudes.” Social Science Quar25
terly 51(4):649–667.
Peffley, Mark A. and Jon Hurwitz. 1985. “A Hierarchical Model of Attitude Constraint.” American
Journal of Political Science 29(4):871–890.
Peffley, Mark and Jon Hurwitz. 1992. “International Events and Foreign Policy Beliefs: Public Response to Changing Soviet-U.S. Relations.” American Journal of Political Science 36(2):431–461.
Peffley, Mark and Jon Hurwitz. 1993. “Models of attitude constraint in foreign affairs.” Political
Behavior 15(1):61–90.
Rathbun, Brian C. 2007. “Hierarchy and Community at Home and Abroad Evidence of a Common Structure of Domestic and Foreign Policy Beliefs in American Elites.” Journal of Conflict
Resolution 51(3):379–407.
Reifler, Jason, Thomas J Scotto and Harold D Clarke. 2011. “Foreign Policy Beliefs in Contemporary
Britain: Structure and Relevance1.” International Studies Quarterly 55(1):245–266.
Richman, Alvin. 1996. “Trends: American support for international involvement: General and
specific components of post-Cold War changes.” The Public Opinion Quarterly 60(2):305–321.
Richman, Alvin, Eloise Malone and David B Nolle. 1997. “Testing foreign policy belief structures of
the American public in the post-cold war period: Gross validations from two national surveys.”
Political Research Quarterly 50(4):939–955.
Rosati, Jerel A, Michael W Link and John Creed. 1998. “A New Perspective on the Foreign Policy
Views of American Opinion Leaders in the Cold War and Post-Cold War Eras.” Political Research
Quarterly 51(2):461–479.
Rosati, Jerel and John Creed. 1997. “Extending the Three-and Four-Headed Eagles: The Foreign
Policy Orientations of American Elites During the 80s and 90s.” Political Psychology 18(3):583–
623.
Rosenau, James N and Ole R Holsti. 1983. “US leadership in a shrinking world: The breakdown of
consensuses and the emergence of conflicting belief systems.” World Politics 35(3):368–392.
Russett, Bruce M and Elizabeth C Hanson. 1975. Interest and ideology: The foreign policy beliefs of
American businessmen. WH Freeman San Francisco.
Russett, Bruce M et al. 1990. Controlling the sword: The democratic governance of national security.
Harvard University Press Cambridge, MA.
Russett, Bruce, Thomas Hartley and Shoon Murray. 1994. “The end of the Cold War, attitude change,
and the politics of defense spending.” PS: Political Science and Politics 27(1):17–21.
Sulfaro, Valerie A. 1996. “The role of ideology and political sophistication in the structure of foreign
policy attitudes.” American Politics Research 24(3):303–337.
Thomas, A. C. and J. K. Blitzstein. 2011. “Valued Ties Tell Fewer Lies: Why Not To Dichotomize
Network Edges With Thresholds.” ArXiv e-prints .
URL: http://adsabs.harvard.edu/abs/2011arXiv1101.0788T
Vail, Kenneth E., III and Matt Motyl. 2010. “Support for Diplomacy: Peacemaking and Militarism as
a Unidimensional Correlate of Social, Environmental, and Political Attitudes.” Peace and Conflict
16(1):29–57.
Wittkopf, Eugene R. 1981. “The Structure of Foreign Policy Attitudes: An Alternate View.” Social
Science Quarterly 62(1):108–23.
Wittkopf, Eugene R. 1986. “On the foreign policy beliefs of the American people: A critique and
some evidence.” International Studies Quarterly pp. 425–445.
Wittkopf, Eugene R. 1987. “Elites and masses: Another look at attitudes toward America’s world
role.” International Studies Quarterly pp. 131–159.
26
Wittkopf, Eugene R. 1990. Faces of internationalism: Public Opinion and American foreign policy.
Duke University Press.
Wittkopf, Eugene R. 1994. “Faces of internationalism in a transitional environment.” Journal of
Conflict Resolution 38(3):376–401.
Wittkopf, Eugene R and Michael A Maggiotto. 1983a. “Elites and masses: A comparative analysis
of attitudes toward America’s world role.” Journal of Politics 45(2):303–334.
Wittkopf, Eugene R and Michael A Maggiotto. 1983b. “The Two Faces of Internationalism: Public
Attitudes toward American Foreign Policy in the 1970s–and Beyond?.” Social Science Quarterly
64(2):288–304.
Ziegler, Andrew H. 1987. “The structure of western European attitudes towards Atlantic cooperation: Implications for the western alliance.” British Journal of Political Science 17(04):457–
477.
27