IS THE DEMOCRATIC PEACE COLORBLIND? A RACIALIZED PEACE

Noting the contestation over the mechanisms that link joint democracy to peace (Bueno de Mesquita et al. Reference Bueno de Mesquita, Morrow, Siverson and Smith1999; Dafoe Reference Dafoe2011; Gartzke Reference Gartzke2007; Rosato Reference Rosato2003), as well as the importance of the mass public in many theories of the democratic peace, Tomz and Weeks (Reference Tomz and Weeks2013) experimentally validated the deference of the U.S. and British mass publics when engaging with democracies as opposed to nondemocracies. In a hypothetical scenario of a country developing nuclear weapons, respondents were more likely to support the use of preventive strikes against a nondemocracy than a democracy. Respondents perceived nondemocracies as more dangerous and striking them to be less immoral, justifying violent action for self-defense. Johns and Davies (Reference Johns and Davies2012), using a similar scenario in a survey fielded at roughly the same time, also found a pacifying effect of democracy, as well as a greater willingness to use force against an Islamic as opposed to a Christian target.

However, the experimental basis of the democratic peace in public opinion has recently faced criticism on the grounds of “information equivalence.” As Dafoe, Zhang, and Caughey (Reference Dafoe, Zhang and Caughey2018, 400–1) explain, “Manipulating whether a country is described as ‘a democracy’ or ‘not a democracy’…is likely to affect subjects’ beliefs about such background features as the country’s geographic location or demographic composition. If it does, then any differences between experimental groups cannot be reliably attributed to the effects of the beliefs of interest.” In other words, in the minds of those surveyed, the term “democracy” is potentially indexed to a number of factors associated with democratic countries that might drive the effect of interest, rather than democratic regime type per se.

One of the most likely elements of “demographic composition” is race. We understand race not as an immutable biological characteristic but rather as the “product of a complex fusion of factors including societal values, skin color, cultural traits, physical attributes, diet, region of ancestry, institutional power relationships and education” (Sen and Wasow Reference Sen and Wasow2016, 506). In the terms of Sen and Wasow (Reference Sen and Wasow2016), perhaps democracy is part of race’s “bundle of sticks.” In their replication of Tomz and Weeks (Reference Tomz and Weeks2013), Dafoe, Zhang, and Caughey (Reference Dafoe, Zhang and Caughey2018) asked respondents to report beliefs about the characteristics of the hypothetical country described in the experiment, characteristics that were not explicitly mentioned in the vignette. Subjects assigned to a hypothetical democracy were more likely to infer that the country is wealthy, Christian, majority white, a military ally of the United States, and a significant trading partner, despite the fact that the vignette did not explicitly mention these traits (Dafoe, Zhang, and Caughey, Reference Dafoe, Zhang and Caughey2018, 400). And, although Tomz and Weeks (Reference Tomz and Weeks2013) and Johns and Davies (Reference Johns and Davies2012) fix or manipulate a number of strategic factors associated with the hypothetical scenario, no previous study, to the best of our knowledge, fixes the racial characteristics of the country in question. This raises the possibility that what appears to be a “democratic” peace effect might instead be one driven by implicit racial inferences about the country’s population; perhaps members of the public read “democracy” and infer “white” despite the vignette’s lack of explicit information about race.

Recent contributions to critical IR theory, remarkably given their strikingly different epistemological foundations, offer a criticism that parallels this concern about information equivalence: the concept of democracy is racialized. Racialization is the “extension of racial meaning to a previously racially unclassified relationship, social practice or group” (Omi and Winant Reference Omi and Winant2014, 64). As Sabaratnam (Reference Sabaratnam2020, 10) explains, “A key point, yet one readily forgotten, is that ‘the West’ is on any plausible reading a racialised category indexed to ‘Whiteness.”’ While the concept has often been applied to the racialization of particular social groups as “white” or “non-white” through historical processes of subjugation and domination (Maass Reference Maass2023), racialization also applies to the institutions that emerge from such social practices—institutions like democracy. If American identity is amenable to implicit racialization (Devos and Banaji Reference Devos and Banaji2005), it is not a stretch to argue that assumptions about democracies are as well.

In a way that methodologically minded criticisms of “information equivalence” do not, critical theorists offer an explanation for this racialization: the self-perceived cultural superiority of white democratic societies. “Though recently established, white men’s countries sought legitimacy through locating themselves in the long tradition of Anglo-Saxon race history that dated back to the mythic glories of Hengist and Horsa. They shared an English-speaking culture and newly ascendant democratic politics, priding themselves, as Anglo-Saxons, on a distinctive capacity, indeed a genius, for self-government,” write Lake and Reynolds (Reference Lake and Reynolds2008, 6). Importantly, much of this sense of civilizational superiority, particularly for Anglo-Saxon countries, rested on beliefs in a unique white capacity for enlightened democratic rule (Bowden Reference Bowden, De Carvalho, Lopez and Leira2021, 164). In a careful study of the origins of the Anglo-American “special relationship,” often attributed to a shared commitment to democracy, Vucetic (Reference Vucetic2011b, 413) finds instead that “Anglo-American elites indeed understood each other as fellow democrats, but their antecedent ontology was always race, not regime type. To the extent that democracy was recognized as a cause of Anglo-American peace and cooperation, this frame was typically deployed in the context of a racial theory of history… [I]t can therefore be said that what caused peace…was Anglo-Saxon democracy, not Anglo-Saxon democracy.”

The implication of these critiques is that democracies receive better treatment from democracies because they are thought to be equivalently culturally superior, and this cultural superiority is implicated by race. The racial connotations of democracy are crucial, because it has previously been theorized that democracies favor other democracies based on a sense of common identity that rests on the common institutions themselves, without reference to race (Hayes Reference Hayes2013). In other words, the democratic peace in public opinion might be ethnocentric, a broader concept from political psychology with clear parallels to “civilizationalist” critiques of liberal IR theory. Ethnocentrism as a psychological attribute brings these elite-level critiques down to the level of the mass public.

TIES THAT BLIND: AN ETHNOCENTRIC PEACE

Ethnocentrism broadly refers to preferential divisions of the world into in-groups and out-groups (Kinder and Kam Reference Kinder and Kam2010). Bizumic and Duckitt (Reference Bizumic and Duckitt2012) and Bizumic, Monaghan, and Priest (Reference Bizumic, Monaghan and Priest2021) report numerous facets of this group self-centeredness: devotion to the group, a preference for group cohesion, a preference of one’s group over others, a sense of superiority of one’s group over another, a desire for group purity, and a willingness to exploit those outside the group. Importantly, Bizumic and Duckitt (Reference Bizumic and Duckitt2012, 893) note in their review of existing conceptualizations of ethnocentrism that “superiority appears to be the most widely emphasized facet.”

This superiority is primarily cultural and moral, rather than innate and biological. Ethnocentrism is a broader construct than racism: individuals can feel culturally superior to those of the same race. But the two concepts overlap significantly. Associations between a sense of cultural superiority and race have more in common with “symbolic” than “old-fashioned,” biological racism (Kinder and Sears Reference Kinder and Sears1981). In the U.S. context, this symbolic racism is more about the perceived violation of cherished American values than the belief that Black people are “naturally” inferior. Therefore, we conceive of ethnocentrism as a sense of cultural superiority of particular ethnic groups, which might (but need not) be defined racially. According to Bizumic and Duckitt (Reference Bizumic and Duckitt2012, 897), ethnic groups are “the real-world groups that their members perceive as having a unique system of shared meanings (e.g., distinct customs, mores, norms, language, or dialect), the perception of common historical past and future, and usually belief in a common origin.”

The concept of ethnocentrism is particularly well-suited for our purposes given its emphasis on culture and broad applicability to different social contexts. Culture often drives unconscious biases, including stereotypes (Banaji and Greenwald Reference Banaji and Greenwald2016). The presumption that democracies are white is likely such an implicit bias. Ethnocentrism, therefore, is a good way for us to index cultural assumptions. More ethnocentric individuals are likely to exhibit racialized assumptions if those assumptions reflect the dominant culture. However, ethnocentrism is a concept that transcends parts of the west historically cleaved along domestic racial lines. For instance, perceptions that immigrants pose a threat to “Dutch culture” are the primary source of hostility to newcomers in the Netherlands (Sniderman, Hagendoorn, and Prior Reference Sniderman, Hagendoorn and Prior2004).

Nevertheless, we expect that a sense of cultural superiority in the United States, as well as in other western countries, likely contains a racial element. If assumptions about race defined eligibility for first-class citizenship well into the twentieth century in the United States (Smith Reference Smith1997), the supposed “beacon of democracy,” it only makes sense that those assumptions inform perceptions of democracy itself. As Mills (Reference Mills1997) argues, non-whites were never intended as parties to the western social contract. This paved the way for discrimination to take root in societies that identified as democracies (Hanchard Reference Hanchard2018).

A clear parallel exists between this conceptualization of ethnocentrism and “civilizationist” distinctions found in critiques of liberal IR theory, of which democracy is an integral part. Western societies define themselves against others by their degree of advancement, notably material, technological, and moral maturity. This concept of advancement explicitly and implicitly ties to race (Persaud and Walker Reference Persaud and Walker2001; Vucetic Reference Vucetic2011a). As Henderson (Reference Henderson2013, 72) explains historically, “Uniquely among the races, whites were assumed to possess civilization while nonwhites were assumed to occupy a lower stage of development characterized as either barbarism or savagery…The lesser races were assumed to be not only biologically inferior to whites, but in a state of almost perpetual conflict.” Hobson (Reference Hobson, Anievas, Manchanda and Shilliam2015, 83) argues that “after 1945 it transmogrified into a much more subliminal form…[A]ll the old manifest Eurocentric-institutional (and scientific racist) tropes, civilisation, barbarism, savagery and imperialism, were whitewashed but appeared in terms…that dare not speak their name, such as ‘tradition versus modernity’ or ‘core versus periphery.”’ This Eurocentrism is a special case of ethnocentrism, one specific to the superiority of European culture, values, and so on (Blaut Reference Blaut2000, 4). Racism and Eurocentrism are often intertwined. Civilizationalism draws a stark line between a civilized, white west and an uncivilized, non-white other that had profound historical implications in practices such as colonialism and humanitarian intervention (Vitalis Reference Vitalis2000).

Importantly, a sense of cultural superiority lessens inhibitions against violence. Scholars of foreign policy opinion increasingly document a strong relationship between ethnocentrism and hawkishness, the willingness to support the use of military force against out-groups. Ethnocentrism maps onto militant postures toward the world in general (Hurwitz and Peffley Reference Hurwitz and Peffley1987), and much recent work has focused on “terrorism” in particular (Hansen and Dinesen Reference Hansen and Dinesen2023; Kam and Kinder Reference Kam and Kinder2007). For example, individuals higher in ethnocentrism displayed stronger support for the United States’ “War on Terror.” What is unclear is whether this support for force against nefarious out-groups extends to regime type, namely preferences for democracies relative to nondemocracies, and what it is about democracies that garners ethnocentric support. Cultural superiority is a likely answer, and the sense of being better than others could lead to a greater willingness to use violence. Ethnocentrism can translate into perceptions that adversaries are more dangerous (or “barbaric”) and possess less moral worth. These perceptions parallel the two faces of racism described by Freeman, Kim, and Lake (Reference Freeman, Kim and Lake2022) and also emerge empirically as the two most prominent mechanisms found by Tomz and Weeks (Reference Tomz and Weeks2013).Footnote ¹

All of this raises the possibility that past survey experimental work on the democratic peace might be unwittingly tapping into ethnocentrism. When “democracy” is invoked in prior instruments, the term might have less to do with checks and balances, a free press, and so on. Rather, democracy becomes a signifier for “ethnoculturally similar to us.” Nondemocracy implies ethnocultural inferiority and increases the willingness to use force. Perhaps a set of assumptions about the cultural superiority of the west, associated with race, exists in the minds of western publics which form the primary experimental microfoundations for the democratic peace to date. The remainder of the article empirically examines this possibility.

RACE ERASES THE EFFECT OF DEMOCRACY: EVIDENCE FROM TWO ORIGINAL SURVEY EXPERIMENTS

To assess the effect of race potentially misattributed to democracy in past survey experimental work, we use a randomized mediator design described by Acharya, Blackwell, and Sen (Reference Acharya, Blackwell and Sen2018). The intuition of the design is straightforward: if racial inferences drive the effect of “democracy” in past survey work—that is, if democracy conveys “whiteness” that decreases support for force—then the treatment effect of democracy should not change when respondents are told that a hypothetical country is also white; whiteness is already associated with democracy. By contrast, when subjects are told that a democratic country has a non-white population, then this racial manipulation should deflate and possibly even eliminate the democratic peace in public opinion. All democracies would not be created equal, and whiteness drove the original effect rather than democratic regime type per se.

Figure 1 displays our design. The natural-mediator arm (the left side of the figure) assigns regime type information without explicit racial information about the hypothetical country, identical to past democratic peace experiments. This arm provides a “natural” baseline to estimate implicit judgments about race conveyed by democracy in the absence of explicit racial information (and serves as the typical average treatment effect [ATE] reported in past work on the democratic peace).

Figure 1. Experimental Design

By contrast, the manipulated-mediator arm (the right side of Figure 1) includes the same democracy versus nondemocracy assignment in addition to explicit racial information about the hypothetical country. This arm allows us to estimate the same effect of democracy (relative to nondemocracy) while fixing the country’s race to either white or non-white (also known as “average controlled direct effects” [ACDEs]). That is, these regime type assignments in the presence of explicit racial information allow us to reestimate the same democratic peace effect for white countries and non-white countries separately. If whiteness is associated with the democratic peace, then the democratic peace effect should be similar with and without explicit information that the country is white, whereas the provision of non-white information should substantially deflate the democratic peace effect. The difference between the effect of democracy with no racial information and the effects of democracy with explicit white or non-white information quantifies the amount of democracy’s effect eliminated by white and non-white racial information. Acharya, Blackwell, and Sen (Reference Acharya, Blackwell and Sen2018) refer to these differences as “eliminated effects.” Large and significant eliminated effects indicate that race plays a role in the mechanism that links democracy to peace.

If racialization explains the democratic peace in public opinion, then the effect size of democracy with and without white information will be similar—whiteness will not eliminate the effect of democracy. In the absence of explicit racial information, respondents are likely to assume that a democratic country is majority white compared to a nondemocratic country, such that any racial information about whiteness in the context of democracy is superfluous. Conversely, we expect that a predominantly non-white population will significantly eliminate the effect of democracy. This would indicate that respondents presume that democracies are white and, if told otherwise, no longer display the same reluctance to use force against the country.

Furthermore, we expect that the effect of democracy eliminated by non-white information will be even larger for respondents higher in self-reported ethnocentrism, measured in multiple ways, as explained below. If the racialization of democracy is part of prominent cultural assumptions in the United States or even the west more broadly, even if implicitly, then more ethnocentric individuals who defend that culture will be more likely to make such presumptions and act on them through support for the use of force. Moderation of this type allows for better confirmation of the ethnocentric and culturally superior nature of any preference shown to democracies based on racial presumptions.Footnote ²

We field this design through a replication of the now-canonical Tomz and Weeks (Reference Tomz and Weeks2013) counterproliferation instrument. Our fundamental interest is to test whether the causal mechanism underlying existing studies of the democratic peace in public opinion is properly specified. Therefore, we utilize the same instrument as the most influential study in this area. Theoretical changes to the scenario or dependent variable would undermine our efforts.

All subjects were first presented with the following scenario and fixed characteristics associated with a hypothetical country:

The above material retains all of the substantive features from Tomz and Weeks (Reference Tomz and Weeks2013) with minor adjustments for parsimony. Note that we fix rather than randomize alliance and trade status, given that we are interested in the democracy treatment in particular. The next screen presented our experimental interventions:

As described above, subjects were assigned to either a democracy or nondemocracy bullet, as well as one of the following racial assignments: no racial information (and therefore no bullet point) or a predominantly white bullet point or a predominantly non-white bullet point. Because the mention of race could induce information equivalence problems of its own, we re-analyzed data from Dafoe, Zhang, and Caughey (Reference Dafoe, Zhang and Caughey2018).Footnote ³ Their study measured respondents’ posttreatment judgments about multiple characteristics of the country in question beyond regime type. We found that religion posed the greatest risk to inference. Therefore, we fix religion to Christian for external validity reasons. There are more plausible real-world candidates for non-white Christian countries than white non-Christian countries.

Following treatment, subjects responded to the same primary dependent variable from Tomz and Weeks (Reference Tomz and Weeks2013):

Responses were gathered on a five-point scale from “oppose strongly” (=1) to “favor strongly” (=5), where lower values in the democracy condition indicate a negative, pacifying effect of democracy on support for strikes. The key finding from Tomz and Weeks (Reference Tomz and Weeks2013) is that respondents were less likely to support strikes against a fellow democracy (ATE $ =-11.4 $ , 95% CI [ $ -17.0,-5.9 $ ]).Footnote ⁴

As found in Tomz and Weeks (Reference Tomz and Weeks2013), we expect that respondents will be less likely to support strikes against a hypothetical democracy (relative to nondemocracy). However, we expect that ethnocentric respondents will primarily drive this treatment effect. Further, beyond the typical ATE of regime type, as described above, we expect that the provision of non-white racial information will eliminate this democratic peace effect. By contrast, the provision of white racial information will not eliminate the democratic peace in public opinion to any appreciable degree.

Survey 1: Qualtrics Sample

Our first survey sample was recruited by Qualtrics and fielded on April 7–13, 2022 ( $ N=1,626 $ U.S. adults). The survey sample included quotas for race, education, and gender to increase representativeness relative to the U.S. population.Footnote ⁵ All respondents completed the survey through Qualtrics’s online platform.

In addition to the instrumentation described above, the survey included self-report measures of ethnocentrism to assess whether ethnocentrism disproportionately explains the racialization of democracy.Footnote ⁶ To measure ethnocentrism, we used three items from the commonly used Neuliep and McCroskey (Reference Neuliep and McCroskey1997) scale:

• • Most other cultures are backward compared with my culture.

• • My culture should be the role model for other cultures.

• • I am not really interested in the customs and values of other countries.

Given that our argument hinges on cultural superiority, this scale is attractive. For example, another popular scale pioneered by Kinder and Kam (Reference Kinder and Kam2010) relies on generalized hostility toward specific ethnic groups, which is too specific in its target and too culture-specific for our purposes (Bizumic, Monaghan, and Priest, Reference Bizumic, Monaghan and Priest2021, 13). We aim to measure a general sense of cultural superiority. Moreover, because our experimental treatment is explicitly racial—indeed it must be in that we hypothesize that democracy is operating as a euphemistic stand-in for race—we wanted to account for possible social desirability concerns (Huddy and Feldman Reference Huddy and Feldman2009). If, as we hypothesize, ethnocentrism moderates the deflation of democracy’s effect for non-white countries, this would suggest that the sense of cultural superiority is racial to some significant degree.Footnote ⁷ We use factor analysis to reduce the responses to single-dimensional factor scores, where higher values indicate greater ethnocentrism.Footnote ⁸ For ease of presentation, we split respondents into two bins: high ethnocentrism (above the ethnocentrism median) versus low ethnocentrism (at or below the median).

If the democratic peace is racialized, then ethnocentrism should drive the pacifying effect of democracy. Figure 2 displays the results estimated with linear regression, including the full sample (left panel) and the sample broken down by subjects above or below the ethnocentrism median (middle and right panels, respectively).Footnote ⁹ The first row displays the well-known ATE of democracy (relative to nondemocracy) for subjects assigned no racial information. As found in Tomz and Weeks (Reference Tomz and Weeks2013) and replicated multiple times since, the treatment effect is substantively and statistically noteworthy in the full sample (coef = $ -0.31 $ , 95% CI [ $ -0.592,-0.034 $ ]). However, the middle and right panels suggest that respondents above the ethnocentrism median drive the treatment effect of democracy (coef = $ -0.52 $ , 95% CI [ $ -0.920,-0.125 $ ]). Subjects below the ethnocentrism median do not discriminate between democracies and nondemocracies at all (coef = $ -0.09 $ , 95% CI [ $ -0.474,0.298 $ ]). This result is noteworthy, because ethnocentrism should play no role in the democratic peace in theory.

Figure 2. Qualtrics Sample—Ethnocentrism Drives the “Democratic” Peace Effect

Note: Estimates and confidence intervals come from linear regressions with robust standard errors at 95% (thin line) and 90% (thick line) levels. Supplementary Tables A4–A6 present these results numerically.

The second row of Figure 2 reports this same treatment effect of democracy (relative to nondemocracy) for white and non-white countries (i.e., the ACDEs for respondents randomly assigned racial information). In the full sample, the democratic peace effect for white countries is substantively and statistically identical to the democratic peace effect without racial information (coef = $ -0.31 $ , 95% CI [ $ -0.607,-0.015 $ ]). Adding white information does nothing to alter democracy’s effect. By contrast, the effect of democracy for non-white countries deflates to a level statistically indistinguishable from zero, such that non-white countries enjoy no democratic peace effect (coef = $ -0.19 $ , 95% CI [ $ -0.482,0.093 $ ]). Respondents assigned to non-white countries are equally likely to strike democracies and nondemocracies in the full sample. However, we do not overinterpret these ACDEs, because eliminated effects provide the most direct test of our argument.

The third and final row of Figure 2 reports the eliminated effects—that is, the difference between the ATE and ACDE quantities—which assess whether race plays a role in the causal mechanism linking democracy to peace. The full sample results suggest that racial information does not significantly eliminate the effect of democracy, and therefore race plays no role in the causal mechanism linking democracy to peace. However, the middle and right panels illuminate this null finding. For respondents above the ethnocentrism median, the provision of non-white information entirely eliminates the democratic peace effect (coef = $ -0.65 $ , 95% CI [ $ -1.198,-0.092 $ ]), strong evidence for the racialization of democracy among the exact subpopulation that drives the average effect of democracy. This is not the case when the target is identified as white. Without measurement of ethnocentrism, this result would be missed at the average level: respondents below the ethnocentrism median respond in the opposite direction, slightly more likely to display democratic peace tendencies when the countries are non-white (coef = $ 0.42 $ , 95% CI [ $ -0.152,0.987 $ ]). This suggests that individuals low in ethnocentrism are perhaps expressing anti-racist views, which work against the overall effect in the sample and mask this subtle, racialized process. In sum, individuals higher in ethnocentrism drive the treatment effect of democracy, and the provision of non-white information entirely eliminates the democratic peace effect within this subpopulation.

Furthermore, while the above results follow past research that bins respondents by ethnocentrism level (e.g., Hainmueller and Hopkins Reference Hainmueller and Hopkins2015; Hansen and Dinesen Reference Hansen and Dinesen2023), we note that a non-white population eliminates the democratic peace effect for respondents at or above the bottom 43% of the ethnocentrism scale. That is, far from a handful of ethnocentric respondents driving our results, approximately 57% of respondents display no democratic peace tendencies toward non-white countries.

Finally, to assess whether ethnocentrism uniquely drives the democratic peace effect in public opinion, Supplementary Section A2.1.3 examines a host of individual differences that could plausibly drive hawkishness toward out-groups. We find that other variables do not moderate democracy’s treatment effect. Further, Dataverse Appendix B3 shows that we find no differences by respondent race, but our survey was not designed with sufficient statistical power to detect those differences. Still, we note that in theory racialized assumptions can affect the views of even non-white Americans.

Survey 2: Prolific Sample

Our first survey indicates that non-white countries incur a racial penalty and that ethnocentrism drives this effect. However, we were concerned about two potential issues with our first survey. First, given that randomized mediator designs can face statistical power constraints (Acharya, Blackwell, and Sen, Reference Acharya, Blackwell and Sen2018, 363), our first survey might be underpowered. For example, it is difficult to determine whether the absence of eliminated effects in the full sample is a substantively important result or simply because we based our power calculations on Tomz and Weeks (Reference Tomz and Weeks2013), which was not explicitly designed to detect this quantity. Second, although we used well-established items from Neuliep and McCroskey (Reference Neuliep and McCroskey1997) to measure ethnocentrism in the first survey, we wanted to validate the above results with a more robust set of ethnocentrism items from a different inventory.

To address these issues, we fielded a replication of our first survey on September 27 to October 10, 2022 with respondents recruited on Prolific ( $ N=2,\hskip-1pt 659 $ U.S. adults).Footnote ¹⁰ Our second survey differed from the first in only two respects. First, to ensure that we have sufficient statistical power to detect non-white eliminated effects, here we do not include the white country randomization. Our first study already suggested that the provision of white information does not change democracy’s overall treatment effect. Second, we used the 12-item “superiority” subscale from Bizumic et al. (Reference Bizumic, Duckitt, Popadic, Dru and Krauss2009), because this subscale best meshes with the theoretical literature that motivates our expectations. Respondents agree or disagree with statements like “The world would be a much better place if all other cultures modeled themselves on my culture.” The items we use do not mention race explicitly, but the scale correlates highly with Kinder and Kam’s (Reference Kinder and Kam2010) ethnocentrism scale, which asks respondents to assess racial in-groups and out-groups on positive and negative traits.Footnote ¹¹ Responses were gathered on five-point scales from “strongly disagree” to “strongly agree.” We again use factor analysis to reduce the responses to unidimensional factor scores (Cronbach’s $ \alpha =0.93 $ , SS loadings = 6.19), where higher values indicate greater ethnocentrism, and we split respondents at the ethnocentrism median to ease presentation.

Figure 3 (left panel) displays the results of our second survey estimated with linear regression.Footnote ¹² The first row shows that we again replicate the well-known ATE of democracy relative to nondemocracy for subjects assigned no racial information (coef = $ -0.56 $ , 95% CI [ $ -0.747,-0.367 $ ]). The only notable difference from our first survey is that respondents below the ethnocentrism median show a significant democratic peace effect (coef = $ -0.40 $ , 95% CI [ $ -0.669,-0.134 $ ]). But this effect is significantly smaller in magnitude than the ATE of democracy for respondents above the ethnocentrism median (coef = $ -0.72 $ , 95% CI [ $ -0.982,-0.451 $ ]). Therefore, we again find evidence that ethnocentrism disproportionately explains the treatment effect of democracy.

Figure 3. Prolific Sample—A Predominantly Non-White Population Deflates the “Democratic” Peace Effect

Note: Estimates and confidence intervals come from linear regressions with robust standard errors at 95% (thin line) and 90% (thick line) levels. Supplementary Tables A11–A13 present these results numerically.

The second row of Figure 3 reports this same treatment effect for respondents explicitly told that the country’s population is predominately non-white (the ACDE). Fixing the country’s population to non-white substantially deflates democracy’s ATE in the full sample (coef = $ -0.31 $ , 95% CI [ $ -0.495,-0.116 $ ]). As such, non-white countries do enjoy a democratic peace effect, but this effect is 45.1% smaller than countries without race specified. Respondents higher in ethnocentrism primarily drive this deflation of democracy’s effect when the hypothetical country is non-white. However, it is notable that, even among respondents low in ethnocentrism, the estimate for democracy’s effect deflates by 16.2% for non-white countries (albeit not to the extent that we can conclude a significant deflation).

The final row of Figure 3 shows that, at least at the $ \alpha =0.10 $ level, race plays a role in the mechanism that links joint democracy to pacifism in the full sample (coef = $ -0.25 $ , 95% CI [ $ -0.519,0.017 $ ]). The middle and right panels confirm again that respondents higher in ethnocentrism drive this effect of democracy eliminated by race. Furthermore, although we median split ethnocentrism for presentation, it is worth noting that non-white information significantly eliminates the effect of democracy for a full 76% of this sample; only respondents in the bottom quarter of ethnocentrism hesitate to discriminate on the basis of race. This provides strong evidence that these results do not reduce to a handful of ethnoculturally chauvinistic apples. Racialization of the democratic peace in public opinion is likely more widespread than our first survey revealed, which was possibly limited by statistical power.

Finally, we note that our Prolific survey included posttreatment measures of threat perception and moral inferences based on Tomz and Weeks (Reference Tomz and Weeks2013), with the expectation that subjects would view non-white countries as more threatening and believe that striking them is less immoral, thus easing the psychological brakes on doing those countries harm. Supplementary Section A2.2.3 and Dataverse Appendix B4 show that the beneficial effect of democracy on threat perception and morality substantially declines for non-white countries. These results align with the two faces of racism outlined by Freeman, Kim, and Lake (Reference Freeman, Kim and Lake2022): one face that approaches non-white countries in a paternalistic and legalistic manner, centered in part on moral deficiency, and a second face that views non-white countries as hostile and threatening. This tendency to divide the world into “virtuous in-groups and nefarious out-groups” is central to the concept of ethnocentrism (Bizumic, Monaghan, and Priest Reference Bizumic, Monaghan and Priest2021; Kam and Kinder Reference Kam and Kinder2007), further suggesting that ethnocentrism explains the democratic peace in public opinion. Importantly, these results imply that racialization likely influences other important variables in foreign policy opinion research, beyond our focus on support for preventive strikes.

Taken together, our results across two surveys suggest that racialization explains much of the democratic peace in public opinion. Respondents higher in ethnocentrism drive the overall treatment effect of democracy, and the provision of non-white information eliminates the democratic peace effect within this subpopulation. This implies that the term “democracy” (relative to “nondemocracy”) conveys whiteness, and respondents with ethnoculturally superior beliefs support the use of force when race is specified otherwise.

HIDING IN PLAIN SIGHT: ETHNOCENTRISM IN PREVIOUS FINDINGS

Our original surveys indicate that racial presumptions about democracy help to explain the democratic peace in public opinion, particularly for respondents higher in ethnocentrism. Do these results shed light on past findings? Here, we reanalyze Tomz and Weeks (Reference Tomz and Weeks2013). If democracy implies whiteness, then ethnocentrism will be a major force behind the preference shown to democracies.

Tomz and Weeks (Reference Tomz and Weeks2013) measured ethnocentrism with items that center on anti-immigration attitudes. These items are a useful proxy for ethnocentrism, because recent reviews suggest that hostility to immigration is a function of perceived cultural threat (Hainmueller and Hopkins Reference Hainmueller and Hopkins2014). As described in Supplementary Section A1.2, we reduce the items to unidimensional factors scores, where higher values indicate greater ethnocentrism.

If ethnocentrism explains the democratic peace effect in past work, then ethnocentrism should moderate the treatment effect of democracy. Figure 4 displays the relationship between support for a strike and the interaction between ethnocentrism and regime type assignment. Although ethnocentrism correlates positively with aggressive responses in general, individuals higher in ethnocentrism are significantly less likely to support the use of force against democracies in particular (coef $ =-0.32,p=0.032 $ ). At lower levels of ethnocentrism, respondents do not differentiate at all between democracies and nondemocracies. At higher levels of ethnocentrism, there is a substantively large, 18.1 point difference in the likelihood of endorsing the use of force. Further, splitting the sample at the ethnocentrism median reveals that the ATE for subjects below the median is statistically indistinguishable from zero (coef= $ -0.19 $ , $ p=0.28 $ ). Respondents above the median entirely drive democracy’s effect (interaction coef = $ -0.53 $ , $ p=0.033 $ ).

Figure 4. Tomz and Weeks (Reference Tomz and Weeks2013) Reanalysis: Ethnocentrism Moderates the Democratic Peace Effect

Note: Moderation effect estimated with linear regression. Column 2 of Supplementary Table A2 presents these results numerically.

Supplementary Section A1 reports the same substantive results for the longitudinal data presented in Tomz and Weeks (Reference Tomz and Weeks2013) and for the survey of the British public conducted by Johns and Davies (Reference Johns and Davies2012). The latter also gives initial indication that our argument travels beyond the U.S. public.

PUTTING THE DEMOCRATIC PEACE INTO CONTEXT: RACIAL ASSOCIATIONS IN THE ENGLISH LANGUAGE

Our survey experiments and reanalyses suggest that racialization explains the pacifying effect of democracy in previous public opinion work on the democratic peace. This final empirical section moves beyond the survey experimental world to assess whether these same associations exist in the English language writ large, a demonstration of external validity. This analysis allows us to validate the above results with an entirely different method—word embeddings—and generalize the results far beyond a handful of survey samples.

Over the past decade, word embeddings, or vector space models of text, have quickly become the gold standard for the analysis of prejudice in human language (e.g., Caliskan, Bryson, and Narayanan Reference Caliskan, Bryson and Narayanan2017; Garg et al. Reference Garg, Schiebinger, Jurafsky and Zou2018; Kozlowski, Taddy, and Evans Reference Kozlowski, Taddy and Evans2019). In contrast to traditional bag-of-words analyses, which rely on comparisons of simple word frequencies in a given document, word embeddings operationalize the notion that we can know a word by the company it keeps (Rodriguez and Spirling Reference Rodriguez and Spirling2022). Embedding models use large inputs of digitized text to estimate the coordinate positions of each unique word relative to all other unique words in some N-dimensional geometric space. The estimated coordinates of each word reveal that words that share (lack) many linguistic contexts occupy similar (dissimilar) locations in this vector space, and these positions tell us something about the semantic similarity of the terms. For example, “if the distance between ‘immigrants’ and ‘hardworking’ is smaller for liberals than for conservatives, we learn something about their relative worldviews” (Rodriguez and Spirling Reference Rodriguez and Spirling2022, 101).

Word embeddings provide particularly strong leverage on the analysis of prejudice. Although explicit bias certainly exists in English documents, embedding models are able to detect far more subtle instances. Inspired by the “general idea that text corpora capture semantics, including cultural stereotypes and empirical associations,” Caliskan, Bryson, and Narayanan (Reference Caliskan, Bryson and Narayanan2017, 1) replicate findings from the implicit association paradigm, such as the finding that European-American names associate with “pleasant” terms and African-American names associate with “unpleasant” terms in the English language. Similarly, Kozlowski, Taddy, and Evans (Reference Kozlowski, Taddy and Evans2019) quantify gendered stereotypes associated with occupation, noting, for example, that the vector for “male” falls closer to “engineer,” whereas “female” falls closer to “nurse.” Garg et al. (Reference Garg, Schiebinger, Jurafsky and Zou2018) go so far as to quantify well-known gender and ethnic stereotypes across a century’s worth of the English language. As Garg et al. (Reference Garg, Schiebinger, Jurafsky and Zou2018, E3635) explain, “word embeddings…capture common stereotypes because these stereotypes are likely to be present, even if subtly, in the large corpora of training texts.” In other words, quotidian texts provide a window into culture.

This emerging work suggests that word embeddings quantify prejudice at massive linguistic scales. Do these same prejudices—notably racial and ethnocentric prejudices—associate with the democratic peace in the English language? We believe that word embeddings are useful for research on the mass public, because textual data (and the embeddings estimated from these data) reflect the same sorts of biases and implicit associations that we would typically measure via survey methods. That is, members of the public generate these texts, and these texts therefore provide useful information about implicit biases and prejudices. For example, if we asked members of the public to describe democracies versus nondemocracies, we would expect terms like white or non-white to, respectively, coappear, despite the fact that we did not explicitly ask respondents about racial characteristics. Embeddings allow us to make similar average-level assessments on a societal scale.

Recall our argument’s intuition: if racialization explains the democratic peace in public opinion, then the relationship between democracy and peace should substantially decrease for non-white countries. In a word embeddings context, we can similarly assess whether a positive correlation exists between “democratic” terms and “peace” terms, followed by a comparison of that same correlation when the “democratic” terms are averaged with “white” or “non-white” terms. To make these assessments, we use well-established pretrained word embeddings from Stanford’s NLP group (Pennington, Socher, and Manning Reference Pennington, Socher and Manning2014), the most commonly used embeddings in past work on stereotypical associations in the English language. The embeddings were estimated using the global vectors for word representation (GloVe) model, and the underlying corpus derives from massive web crawls of digitized, contemporary English materials (like newswire texts and the entirety of Wikipedia). We use the vectors trained in two hundred dimensions, with four hundred thousand unique English terms.

With these English language vectors in hand, we use the following terms to estimate our theoretical constructs of interest:

• • Democracy terms: democracy, democratic, democratically, elect, elections, elected.

• • Peace terms: peace, harmony, agreement, diplomacy.

• • White terms: white, western, caucasian, european.

• • Non-white terms: non-white, non-western, non-caucasian, non-european.

The average word coordinates for each set of dictionary terms defines the construct of interest, which is the most common method of combining vectors in embedding research on prejudice. For example, the average coordinates for the terms “white,” “western,” “caucasian,” and “european” provide a single averaged vector that represents the estimate of our white construct in the vector space. Mirroring the verbiage of our experiments, we use the prefix “non-” to form contrasts to the white terms.

Before estimating our primary quantities of interest, we note interesting descriptive results from these word embeddings. In line with our expectation that democracy is racialized, we find that the democracy terms show a much larger similarity to the white/western terms (cos sim $ =0.389 $ ) than non-white/nonwestern terms (cos sim $ =-0.003 $ ), which suggests that democracy implicitly associates with whiteness in this corpus of English language texts. This association also parallels Dafoe, Zhang, and Caughey’s (Reference Dafoe, Zhang and Caughey2018) finding that respondents presume that democracies are predominately white.

Next, guided by the intuition of our experiments above, we seek to estimate two relationships in particular. First, akin to the ATE of democracy on pacifism, we calculate the cosine similarity between democracy and peace:

(1) $$ \begin{array}{rl}\mathrm{Total}\hskip0.3em \mathrm{similarity}=\mathrm{cosine}({v}_{democracy},\hskip0.3em {w}_{peace}),& \end{array} $$

where $ {v}_{democracy} $ and $ {w}_{peace} $ represent the average of dictionary term vectors for the democracy and peace constructs, respectively. Large, significant similarities would indicate that democracy correlates positively with peace in the English language.

Second, to assess whether the racialization of democracy helps to explain this association between democracy and peace, we estimate the same similarity between democracy and peace for “white democracies” versus “non-white democracies” separately (akin to the ACDEs above). To do so, we average the democracy terms with white versus non-white terms and simply recalculate the similarity between white democracies and peace, as well as non-white democracies and peace. More formally,

(2) $$ \begin{array}{l}\mathrm{Similarity}\hskip0.3em \mathrm{fixing}\hskip0.3em \mathrm{white}=\\ {}\hskip2em \mathrm{cosine}[\mathrm{mean}({v}_{democracy},\hskip0.3em {u}_{white}),\hskip0.3em {w}_{peace}],\end{array} $$

(3) $$ \begin{array}{l}\mathrm{Similarity}\hskip0.3em \mathrm{fixing}\hskip0.3em \mathrm{nonwhite}=\\ {}\hskip1em \mathrm{cosine}[\mathrm{mean}({v}_{democracy},\hskip0.3em {u}_{nonwhite}),\hskip0.3em {w}_{peace}],& \end{array} $$

where $ {u}_{white} $ and $ {u}_{nonwhite} $ represent the average of dictionary term vectors for the white and non-white constructs, respectively. According to our expectations, democracy should display the same similarity to peace with or without white terms. By contrast, averaging the democracy terms with non-white terms will likely deflate the similarity between democracy and peace, such that non-white terms eliminate the positive association between democracy and peace in the English language.

If past survey work on the democratic peace generalizes to the English language writ large, then democracy terms should correlate positively and significantly with the peace terms (akin to the ATE of democracy on pacifism in the above surveys). Figure 5 displays the results of our cosine similarity estimates, with 95% and 90% confidence intervals calculated using one thousand dictionary permutations. The first row displays the total cosine similarity between democracy and peace, the standard cosine similarity estimate reported in almost all social science applications of word embeddings. It is positive and substantively noteworthy (cos sim = 0.43, 95% CI [0.323, 0.524]). This similarity indicates that we can successfully replicate associations between democracy and peace at the mass public level in a corpus that essentially represents the entire contemporary English language, a notable finding in its own right.

Figure 5. Racialization of the Democratic Peace in the English Language

Note: Word embedding cosine estimates following the intuition of Acharya, Blackwell, and Sen (Reference Acharya, Blackwell and Sen2018). Dataverse Appendix Table B14 presents these results numerically.

Furthermore, if racialization underwrites the democratic peace in public opinion, then whiteness should do little to alter the similarity between democracy and peace, whereas non-whiteness should substantially impinge on the relationship between democracy and peace. The second row of Figure 5 shows that the similarity between democracy and peace substantially declines when we average the democracy terms with the non-white terms (cos sim = 0.26, 95% CI [0.152, 0.351]). That is, just as we found in our original experiments above, the pacifying effect of democracy substantially declines in the presence of non-white information. By contrast, averaging the democracy terms with white terms yields no statistical change in the similarity between democracy and peace (cos sim = 0.49, 95% CI [0.369, 0.573]). Similar to our experimental results, whiteness is baked into conceptions of democracy and therefore the democratic peace in public opinion.

The third and final row of Figure 5 simply takes the difference between the aforementioned quantities to estimate the amount of the similarity between democracy and peace eliminated by the presence of racial information, akin to the eliminated effect estimate in Acharya, Blackwell, and Sen (Reference Acharya, Blackwell and Sen2018). The eliminated similarity shows that the decline in association between democracy and peace in the presence of non-white terms is substantively and statistically noteworthy: fixing democracy to non-white (i.e., averaging the democracy and non-white terms) eliminates 40% of the total similarity between democracy and peace. This estimate closely parallels the proportion eliminated in our well-powered experiment fielded on Prolific (45.1%). Indeed, Caliskan, Bryson, and Narayanan (Reference Caliskan, Bryson and Narayanan2017) find that word embeddings yield effect sizes that parallel those found in implicit association tests, despite the methodological difference. By contrast, averaging the democracy terms with the white terms does nothing to statistically or substantively eliminate the similarity between democracy and peace, as we found in our first experiment.Footnote ¹³

Finally, we note that in the context of this English language corpus, we largely expect the results in Figure 5 to capture the dyadic democratic peace phenomenon captured in our survey experiments. However, Supplementary Section A3.1 empirically verifies this expectation, showing that the results are substantively the same when we add “us” and “them” terms to the democracy dictionary, thus explicitly designating a dyadic democratic peace phenomenon. These results provide further evidence for the racialization of the democratic peace in public opinion.

Together, this analysis helps to validate the results found in our original experiments but generalizes these results far beyond two survey samples. The standard interpretation in the embeddings literature is that this underlying corpus represents the contemporary English language writ large. Even if non-white democracies enjoy a democratic peace effect in public opinion, in the English language, this effect is some 40% less pacific than the peace enjoyed by white democracies, which presents a serious challenge to the theoretical microfoundations traditionally thought to underlie the democratic peace.