Beyond the far-right umbrella: cultural-hierarchical radicalism and the legitimation of political violence across 64 countries

1 Introduction

The category “far right” gathers under one term two party families that the literature had long kept apart: the extreme right, which rejects democracy as such, and the radical right, which accepts democracy as an ideal while opposing key features of its liberal variant (Mudde, 2007, 2016, 2017). Whether the umbrella should be retained is now openly contested. Pirro (2023) defends it on the ground that the two families share nativism and authoritarianism and have become difficult to separate in practice; Zulianello (2025) replies that it conflates properties which vary independently, and that recovering the distinction is a condition of asking the field’s questions well. The dispute has been conducted almost entirely on the supply side, among parties and organizations, and about the demand side we know much less. It matters beyond classification: contemporary right-wing actors combine authoritarian, religious-conservative, populist and nativist elements in different proportions across contexts, so that what the label gathers in one region it may not gather in another (López Aguilar and Pino Uribe, 2024).

This is the question the article asks: whether the distinction survives the passage from parties to publics. Does the radical–extreme distinction identified among political organizations have an attitudinal counterpart among mass publics, or do cultural-hierarchical radicalism and the legitimation of political violence form a single continuum there? If the two properties are strongly nested, treating the demand-side foundations of radical and extreme politics under a common umbrella costs little. If they are not, the umbrella conceals heterogeneity twice over: among the organizations it classifies, and among the attitudinal reservoirs on which those organizations draw.

Across sixty-four countries and 80,222 respondents of the seventh wave of the World Values Survey, the two properties barely travel together. Their association is 0.039, and 0.076 within countries, so the weak association is not produced by pooling. A strict necessity claim fails: 15 % of the sample grant political violence at least limited justifiability while ranking low on the measured cultural-hierarchical dimension, crossing the violence threshold without crossing the other, and no country reaches the conventional necessity threshold under country-specific cut-points. A further 26 % rank high on that dimension while placing political violence at the rejecting end of its scale, which bears on sufficiency, a condition funnel models do not assert. The two properties coincide in under 8 %.

An assumption of nestedness organizes two literatures at once: explicit and processual in the study of radicalization, in Moghaddam’s (2005) staircase and the cruder conveyor-belt image of policy discourse, and tacit in the study of the far right, embedded in the umbrella itself. McCauley and Moskalenko (2008, 2017) broke the funnel in two, distinguishing a pyramid of opinion from a pyramid of action, but their model locates individuals on two ladders without naming or counting the resulting configurations, and it has not been examined in nationally representative samples.

We therefore treat the separation as a hypothesis and convert it into a property space. Crossing the two dimensions, each dichotomized, generates a symmetric two-by-two partition in the Lazarsfeld–Barton tradition (Lazarsfeld, 1937) that Sartori (1970) identified as the proper instrument of typological analysis. Section 2.4 sets out the four configurations and Section 4.1 the scope of their names. The cuts are analytical, and the central result does not depend on where the boundaries fall.

The article does not claim that individual respondents can be classified as members of the radical or the extreme right, nor does it seek to estimate the true prevalence of support for political violence. It asks whether two properties the umbrella tends to bring together actually covary in mass publics. They do not, and the four configurations they generate have different social correlates. The clearest indication is a coefficient that changes sign: religious importance is associated with a higher probability of cultural-hierarchical radicalism and a lower probability of violence legitimation without measured radicalism.

The article makes three contributions. The first is to the debate the umbrella has opened: it extends the radical–extreme distinction from the classification of parties to the study of mass attitudes, and shows that the properties the label combines are separable on the demand side as well. The second is comparative: it offers, to our knowledge, the first cross-national test of this separation on nationally representative samples, and reports the social correlates of each configuration. The third is methodological and less comfortable. The World Values Survey items on the justifiability of violence share more than a quarter of their variance with a general willingness to justify petty dishonesty, a component we cannot fully partition into response behavior and substantive permissiveness, and the survey’s authoritarianism items do not travel across contexts. Both results changed the measurement model.

2 The umbrella, the distinction, and the demand side

2.1 What the umbrella joins, and what its critics separate

The umbrella reflects a considered position rather than an accident of usage. Its proponents defend it on a specific ground: the extreme and the radical right share nativism and authoritarianism, and distinguishing them empirically has become difficult in practice. Pirro (2023) argues that the label registers a family resemblance across actors that differ in degree rather than in kind, and that the analytical cost of abandoning it exceeds the cost of retaining it. The position is neither isolated nor dated: Rovira Kaltwasser (2026) adopts the broad concept on those grounds for a comparative study of eight Latin American countries, holding that narrower terms travel worse outside the contexts that produced them.

Zulianello (2025) rejects that defense, and the terms of his rejection are the terms this article adopts. The distinction he defends is the one Mudde (2017) had already drawn between the populist radical right and the extreme right, and which implies, at the level of parties, that radical ideological orientations need not entail the legitimation of violence. Our question is whether that implication survives the passage to individuals. His objection is that the umbrella conflates two properties which may vary independently: the ideology of an actor and its status within the party system. Extreme-right parties in contemporary Europe remain, almost without exception, excluded from coalitions and from government; radical-right parties increasingly are not, and experience what he calls negative integration, incorporation into coalitions and governments without moderation of their core ideology. To account for the difference he recovers Sartori’s (1976) notion of a disjointed political space, in which “no coalition” points separate the actors the mainstream treats as legitimate players from those it does not. The upshot is a methodological injunction: use the precise term wherever possible, because the umbrella overlooks differences that matter.

Wondreys and Zulianello (2024) supply the empirical warrant. Across European extreme-right and radical-right parties they show that the two families diverge systematically even in foreign policy, a domain remote from the issues on which the umbrella is built: the extreme right tends to support Palestine and Russia, the radical right Israel and Ukraine. Two families the label treats as one behave differently where the label predicts nothing at all; the distinction does the empirical work.

Nor is the distinction confined to ideology: recent comparative work shows it operating on three levels at once. On the supply side, radical-right parties negotiate their integration through strategies that reach as far as the visual identity of the party (Curini et al., 2025), and Fratelli d’Italia repositions itself between a post-fascist legacy and a radical-right identity (Puleo and Piccolino, 2022). On the demand side, radicalization within the Italian populist right does not spread to center-right voters (Piccolino et al., 2025), and in the digital context the boundary is more permeable still: the followers of Italian radical and extreme-right accounts overlap substantially, and the clusters that emerge do not follow the party-family line (Taddei, 2026). Audience overlap and attitudinal covariation are different quantities, and the second is what this article measures.

The dispute is about parties, and so is the greater part of the evidence: the review literature and the handbooks (Golder, 2016; Rydgren, 2018) organize the demand side around the correlates of voting for radical-right parties, not around the attitudinal properties that distinguish radical from extreme politics. Halikiopoulou and Vlandas (2019) show how much rests on that organizing choice, since what looks like a single nationalist demand resolves into distinct components with distinct social bases. If the properties that distinguish the two families are distinguishable only in the supply of politics, the umbrella is a defensible shorthand for the constituency. If they are separable on the demand side too, the umbrella conceals heterogeneity twice, and the concealment matters for anything the field wants to explain about who supports whom, and why.

2.2 Violence legitimation as a boundary, not as a definition

A clarification is owed before the empirical argument begins, because the second dimension of our property space could be misread. The extreme right cannot be reduced to violence. Its defining feature, in the literature that produced the distinction, is the rejection of democracy as such, of which violence against political opponents is one expression among others (Rovira Kaltwasser, 2026), and violent dispositions are neither necessary to that rejection nor exclusive to it: violence is justified by people of many ideological persuasions and of none. We do not propose it as a definition of the extreme right, and nothing in what follows depends on one.

What the legitimation of political violence captures is a willingness to grant abstract political violence some degree of justifiability. That may indicate a relaxation of the norm that political conflict is processed without physical coercion, but the item does not identify the agent, the target or the context, and we do not read it as establishing a rejection of democratic restraint. Kalmoe and Mason (2026) provide a systematic guide to measuring violent political views. Read against it, what we have is a single abstract-justifiability question: it does not record conditional support for violence in specified circumstances, behavioral intention, felt moral obligation or reported action, and we do not treat it as a proxy for any of them. We use it as a partial proxy for one possible antilegalist boundary, and not the decisive one: the rejection of popular sovereignty, of pluralism, and of democracy itself would each be others, and this survey does not measure them well enough to cross them with anything. The other dimension captures a high-intensity cultural-hierarchical orientation that comparative research associates with some forms of right-wing radicalism; it does not directly measure a program for transforming the liberal order. Whether the two are manifestations of a common constituency, as the umbrella may encourage analysts to assume, is the empirical question this article asks.

The conceptual apparatus we use for that boundary is not improvised. Pino Uribe and Cely (2023), analyzing left-wing political extremism in Colombia, distinguish the antilegalist component of extremism, the disposition to break the state’s monopoly on legitimate violence, from its monist component, the rejection of pluralist coexistence with adversaries, which is the anti-pluralism Müller (2014) places at the center of populism. The distinction does not belong to the right, which is why it can be used to ask whether the right’s demand side is organized by it. The violence dimension approximates the antilegalist component; the cultural-hierarchical dimension approximates a content compatible with monist orientations rather than measuring monism. The configuration in which both coincide is the one in which the two measured properties occur together, which is narrower than anything the literature means by the extreme right. The configurations in which they come apart are the ones it has no name for.

The two properties do in fact come apart, for reasons anticipated in the classical sociology of deviance. Some justifications of violence issue from an elaborated ideological frame; others do not. Merton (1938) showed that anti-systemic dispositions can arise from the disjunction between socially prescribed goals and the institutional means available to reach them, in patterns that need contain no ideologically coherent premise. Work on state weakness, criminal violence and post-conflict societies has since documented the legitimation of violence operating without that scaffolding (Kalyvas, 2006; Roy, 2017; McLaughlin and Robitaille, 2024). The Mertonian reading remains a hypothesis these data cannot test.

One feature of the design should be stated whole rather than in parts, because its two halves are usually defended separately and it is their joint effect that bounds the conclusion. The distinction this article takes from the study of parties is defined by the relation an actor holds to democracy: the extreme right rejects it, the radical right accepts it as an ideal while opposing its liberal form. Neither pole of that definition is observed here. In its place stand two substitutions: for the ideological content of right-wing radicalism, a culturally hierarchical, religiously exclusivist orientation; for the extreme pole, the legitimation of political violence, which this section has argued is one observable antilegalist boundary and not a definition. What is tested here is therefore the observable version of the distinction, and the conclusion inherits whatever validity the two proxies possess: if either is a poor stand-in, the finding is a finding about the proxy.

2.3 The funnel, the two pyramids, and what they do not do

The dominant models of radicalization are processual and funnel-shaped, and they supply the assumption of nestedness in its most explicit form. Moghaddam’s (2005) staircase describes a narrowing ascent through five floors, from a perceived injustice to the moral legitimation of violence and finally the readiness to kill. Both the staircase and the cruder conveyor-belt image share a premise that has organized the field for two decades: the population holding radical beliefs contains, as a proper subset, the population prepared to use violence. The premise has not fared well under scrutiny. Lygre et al. (2011) find support for the individual factors the staircase names and none for the transitions between its steps. Hafez and Mullins (2015) describe radicalization as a puzzle whose pieces do not assemble into a sequence, and the meta-analysis of Wolfowicz et al. (2020) reaches the same conclusion from the opposite direction: the risk factors for radical attitudes, intentions and behavior are not the same factors.

The most consequential response is McCauley and Moskalenko’s (2008, 2017) two-pyramids model, which breaks the funnel in two. One pyramid orders opinion, from neutrality through sympathy and justification to the felt moral obligation to use violence; the other orders action, from inactivity through legal activism to terrorism. Movement up the one does not imply movement up the other, and Moskalenko and McCauley (2009) supplied the measurement counterpart in the activism and radicalism intention scales. The model is the closest relative of the argument advanced here, and this article builds directly on it.

It is also, for our purposes, incomplete in three respects. It is a model of individual position on two ladders, not a typology of population configurations: it does not name, count or characterize the resulting groups. Its empirical base is narrow, developed on undergraduates and an American internet panel. And it is silent on ideological content: its opinion pyramid is defined by the intensity of endorsement of violence, not by the worldview that may or may not accompany it. Our dimensions are not its dimensions: both of ours are attitudinal, an ideological orientation and a normative judgment about the legitimacy of violent action. We do not observe behavior, and we do not claim to test the opinion–action separation in its behavioral form.

One recent proposal comes closer to the two-dimensional logic we adopt: the Dynamic Matrix of Extremisms and Terrorism (Risius et al., 2021) crosses ideological engagement with the presence of violence and is explicit that a highly ideological actor without violence and a violent actor without ideology are different objects. It was built to classify online content, not mass attitudes. The measurement debate on public support for political violence, in which Kalmoe and Mason (2022a, 2022b) read the survey evidence as registering a real willingness and Westwood et al. (2022a, 2022b) read much of it as an artefact of the instrument, converges on one point: instruments that conflate hostility with the endorsement of violent action overstate the prevalence of the latter, and Section 5.3 reports a confirmation of that warning which we did not seek.

2.4 The property space

We take the separation the two-pyramids model asserts and the separation Zulianello recovers among parties, and convert both into a cross-sectional property space of mass attitudes. Two dimensions, each dichotomized, generate a symmetric two-by-two space with four configurations, set out in Table 1. The two off-diagonal cells bear on different propositions. Violence legitimation without measured radicalism contradicts a strict necessity claim directly, because it contains respondents who cross the violence threshold without crossing the measured cultural-hierarchical one. Cultural-hierarchical radicalism without violence legitimation bears instead on sufficiency, which funnel models ordinarily allow, since they suppose that only a fraction of the radical population proceeds. Taken together they show that the properties should not be collapsed into a single continuum, but they are not the same test of nestedness and we do not present them as one.

Table 1. Cultural-hierarchical radicalism by the legitimation of political violence.

The cells are analytical constructs, not empirical clusters; the cut-points are specified in Section 3.

Each of the off-diagonal cells has a literature, even if the far-right literature has not gathered them. The radical who rejects violence is the configuration Pickerill (2006) recorded in social-movement research, and that the comparative study of religious-conservative parties has repeatedly found operating inside liberal-democratic frameworks (Kalyvas, 1996): too non-moderate to be counted among the moderates, too unwilling to grant violence justifiability to be counted among the extremists, and invisible to a dichotomous classification. The violence legitimizer without measured radicalism is the Mertonian figure of Section 2.2. Our claim is not that these configurations are unknown, but that the umbrella has no place for them and that they are not marginal.

The move is, at the level of individuals, formally analogous to the move Zulianello makes at the level of parties. He separates the ideology of an actor from its status in the system and shows they vary independently, a separation already at work in his account of anti-system parties, where positive integration requires ideological moderation and negative integration does not (Zulianello, 2019); we separate an ideological orientation from the legitimation of political violence and show the same. The analogy is methodological rather than substantive: his second dimension is a systemic property of parties, ours an attitude of individuals. What the two approaches share is a commitment to separating properties that umbrella labels combine.

If culturally radical and violence-legitimating attitudes belong to partly different constituencies, parties may have to mobilize them through different issues and frames: radical-right parties may draw on culturally hierarchical but non-violent publics, of the kind the grievance-mobilization literature has described (Ivarsflaten, 2008), while actors at the extreme may find support among citizens whose tolerance of coercion is not embedded in any radical-right worldview. Treating them as one attributes to a single public a combination of properties few of its members hold. Section 4.1 takes a first step with an expert classification of parties.

The property space is not proposed as a model of four latent natural kinds, and Section 4.1 reports a latent profile analysis that does not reproduce its cells. Its analytical value lies elsewhere, and can be stated precisely. It makes the necessity and the sufficiency of one property for the other assessable rather than assumed. It renders the two off-diagonal configurations visible and countable instead of leaving them as the residual of a correlation. It allows their social correlates to be compared rather than averaged. And it exposes a predictor, the importance of religion, that operates in opposite directions on the two properties, a contrast that a single continuum would have canceled. Continuous specifications recover the association without loss of information, and can be made to yield most of these quantities; the property space complements them by making the set relations and the theoretically relevant configurations directly interpretable.

2.5 Defining cultural-hierarchical radicalism

The cultural-hierarchical dimension is defined substantively rather than identificationally. We operationalize one culturally hierarchical component of right-wing radicalism through religious exclusivism and gender traditionalism. The measure does not reproduce the ideological morphology of radical-right parties and is not an indicator of affiliation with them; where the shorter form is used below, it denotes that component and nothing wider. Following Bobbio (1996), the analytical core of the right lies in the treatment of hierarchy and inequality as legitimate expressions of natural or cultural difference. We operationalize the dimension through two components. Religious exclusivism captures the claim that one religious tradition is uniquely legitimate and overrides secular knowledge. Gender traditionalism captures the treatment of gender hierarchy as natural: that men make better political leaders and better business executives, that men should have priority in employment when jobs are scarce, and that a pre-school child suffers when its mother works. The second component is the more directly Bobbian, since it names inequality as such, and it is the one that travels: it is the moral conservatism that Rovira Kaltwasser et al. (2026) find in every Latin American case they examine.

The operational strategy follows Pino Uribe and Castillo Fell (2026), who measure right-wing values through devotional religiosity and exclusionary belonging as legitimizing myths of social hierarchy in the Bobbian sense; the two components are also those the comparative literature on values has established as cross-nationally measurable (Inglehart and Welzel, 2005; Norris and Inglehart, 2011, 2019).

One decision requires an explicit defense, because it will otherwise look like an omission. Left–right self-placement is not a constitutive criterion of the index. Self-placement carries well-documented social-desirability and translation problems in cross-national research: it is not fielded in eleven of the countries here, it does not mean the same thing where the axis does not structure competition, and it produces centripetal misreporting in polarized and authoritarian settings. Building it into the measure would import those problems into the definition of the types, so it is preserved as an external criterion, in Section 4.1, where its association with the index is informative precisely because it is imperfect, and in Section 4.3, where its regional heterogeneity becomes a finding in its own right.

Bobbio establishes what makes these orientations right-wing. He does not establish what makes them radical, and the distinction matters. We do not infer radicalism from the substantive content of any single item: a person may hold conservative religious or gender views without being radical in any politically useful sense. We use the term to identify an intense position at the non-moderate end of a culturally hierarchical orientation, and the intensity is relative to the pooled distribution. “Cultural-hierarchical radicalism” is therefore an attitudinal construct denoting relative non-moderation, not a proxy for party affiliation, voting behavior, organizational membership or a complete ideological worldview, and respondents who score highly on it do not constitute a radical-right electorate. Section 4.1 assesses the construct against criteria that are not part of it: left–right self-placement, hostility to immigrants as neighbors, immigration policy preferences, and the party a respondent chooses, matched to an expert classification.

One consequence of the umbrella’s two pillars shaped the measure and deserves recording. The available indicators of nativism do not travel: an anti-immigration agenda is absent from all but one of the seven Latin American cases coded by Rovira Kaltwasser et al. (2026), so a measure anchored in nativism would classify Bolsonaro, Milei and Bukele as something other than far right. The point is not confined to the coding of parties, for the reason given in Section 2.1: nationalist demand resolves into components whose reach differs by region. The available indicators of authoritarianism do not travel either, for measurement reasons reported in Section 5.3. The difficulty is not only ours. Baldner et al. (2026) argue that the instruments the field relies on conflate ideological content with antidemocratic disposition, and report that under perceived threat the desire for an authoritarian leader is not confined to one side of the political spectrum, although it remains more prevalent on the right. An attribute whose available indicators behave in that way cannot anchor a right-wing index estimated across sixty-four countries. The two attributes on which the case for the umbrella rests are the two whose available World Values Survey indicators a globally distributed measurement model cannot sustain. That is a claim about instruments and not about concepts, and it is why the index is built from religious exclusivism and gender traditionalism alone.

3 Data and measurement

3.1 Data: the World Values Survey, wave 7

The seventh wave of the World Values Survey (Haerpfer et al., 2022) is used because it is the only instrument that fields, in a single questionnaire and across a wide range of political regimes and religious traditions, both the cultural-moral items required to measure the cultural-hierarchical dimension and an item on the justifiability of political violence. The pairing is what the argument needs, and it does not exist in the regionally bounded alternatives. The political-violence item was introduced in the seventh wave, so earlier waves cannot support the second dimension; the cost is a cross-sectional design that cannot speak to change over time. Fieldwork within the wave was conducted between 2017 and 2023. No country in the analytical sample was surveyed across more than one calendar year, so survey-year indicators are collinear with the country indicators and add no identification; the models therefore carry country fixed effects only, and survey-year effects are not entered.

The strengths of the sample should not be overstated. The WVS provides nationally representative samples within the countries it includes; it is not a probability sample of countries, and we do not describe it as globally representative. Coverage is uneven: Africa has seven countries, North America and Oceania two each. We therefore report regional figures with the number of countries attached, treat them as descriptive, and do not aggregate regions with fewer than five countries into a mean with an interval. The survey is also known to carry an urban bias in several low- and middle-income countries: if rural respondents are more religious and more gender-traditional, it would understate the prevalence of the cultural-hierarchical configuration there. Whether it does cannot be established with these data, and we record it as a limitation.

Several analytical samples appear in what follows, all drawn from the sixty-six countries of the file. Turkey is dropped for want of the violence item. Listwise deletion on the six index items and the violence item, applied before the index is built and not after, yields the typology sample of 80,222 respondents in 64 countries, with Egypt lost entirely. The multinomial adds eight predictors and yields 72,530 in 63 countries. The measurement model estimating the association between the two axes needs the two secondary violence items and the four permissiveness items, and runs on 74,719 in 60 countries. The criterion samples vary with the criterion, and their sizes are given where used. Appendix A.2 reports the arithmetic item by item, and the replication package the retention by country; the loss is not concentrated in any region.

3.2 The measurement model and its cross-regional equivalence

The radicalism index is built from six items lying outside the survey’s “Justifiable:” battery, a decision explained in Section 5.3. Two further components were tested and not retained: democratic devaluation, which correlates at +0.026 with right self-placement within countries against +0.110 and +0.102 for the two retained, and political authoritarianism, for the invariance reasons reported in Section 5.3. Appendix A.8 reports the four-component intercorrelations and the reliability of the retained index by country and region. Religious exclusivism is measured by two items, gender traditionalism by four; wording, scales and directional recodes are in Appendix A.1. The legitimation of political violence is measured by a single item (E290: political violence is never justifiable, 1, to always justifiable, 10). The three-item battery is a robustness check, and the domestic-violence item is excluded from the primary axis for reasons given in Section 4.5.

A categorical confirmatory factor model is estimated on the typology sample (N = 80,222 in 64 countries) by diagonally weighted least squares on the polychoric correlation matrix, with thresholds from the univariate marginals; the six items are treated as ordered categorical, and listwise deletion on the constituent items and the violence item precedes construction of the index. Fit is acceptable but not uniformly strong: incremental indices are high (CFI = 0.976; TLI = 0.955), the standardized root mean square residual is low (SRMR = 0.045), and the RMSEA of 0.086 (90 per cent interval 0.084 to 0.088) indicates residual misfit. McDonald’s (1999) omega is 0.854 and 0.801 for the two components, and the average variance extracted is 0.745 and 0.517, both above the 0.50 threshold of Fornell and Larcker (1981).

Cross-regional equivalence is a separate question and the evidence is partial. Fitted freely in each of six regions, which between them cover the 80,222 respondents of the typology sample, the two-factor pattern is recovered everywhere, with CFI between 0.956 and 0.981: this is configural similarity, and it is a necessary but not a sufficient condition for comparing latent levels across regions. Constraining the loadings to their pooled values, and applying the conventional benchmarks (Davidov et al., 2014), three of the six regions meet the criterion for metric invariance and three do not, the largest deterioration being a CFI loss of 0.020 in the United States and Canada. Full threshold invariance is not achieved at all: the polychoric thresholds of the six items differ across regions by an average of 1.49 standard deviations of the latent continuum, and by as much as 2.40 for the first religious item. The same latent position therefore produces different observed answers in different regions. Appendix A.8 reports the loadings and the constrained-model fit region by region, and the replication package the full threshold table.

The association between the two dimensions requires a word about how it is estimated, because a single item cannot carry a latent factor and the violence axis of the typology is a single item, a deliberate choice defended below. We do not estimate a latent violence factor for this purpose. We report the manifest Pearson correlation, 0.039, as the headline quantity, and an ordinal-latent estimate of 0.058 as a robustness check. The latter is computed analytically from the polychoric matrix of the items rather than by fitting a latent model: for a unit-weighted composite C of items i and a variable v, the correlation implied by that matrix is the sum of the item-by-v entries divided by the square root of the double sum of the item-by-item entries. It is not a correction for measurement error and we do not describe it as one. A three-item violence composite returns a larger value, for a reason that is itself an argument: the domestic-violence item shares obvious content with gender traditionalism.

A further adjustment is available and we report it with a caveat. The political-violence item correlates at 0.52, and at 0.67 as a polychoric correlation, with a composite of four items on the justifiability of petty dishonesty (benefit fraud, fare-dodging, tax evasion, bribery) that lie in the same battery. Partialling that composite out of the violence item drives its association with radicalism to −0.003, that is, to nothing. The shared variance is consistent with a general permissiveness or response-style component, and these data cannot determine how much of it reflects survey response behavior and how much a substantive willingness to justify transgression. We do not make this the headline number, because the partial may over-correct: a general willingness to justify transgression may be a substantive disposition as well as a response style, and removing it eliminates substance along with method. In sum, estimates across the specifications reported in this section range from −0.003 to 0.058, and nothing in the argument turns on which of them is preferred. Two coding rules complete the construction: high on the cultural-hierarchical dimension means at or above the 67th percentile of the pooled distribution, and above the cut on the violence axis means 3 or more on E290. Section 4.5 and Appendix A.3 examine alternatives. Appendix A.10 reports the four side by side.

Accordingly, regional differences in prevalence are reported as descriptive differences in classifications produced by common coding rules, not as comparisons of latent means: the typology is a partition of observed responses, not a claim that the underlying construct is scaled identically in Lagos and in Auckland. This limitation bounds the regional interpretation in Section 4.3.

3.3 Predictors

Each predictor is motivated by the literature. Education is the most established correlate of radical-right support in the European literature (Lubbers et al., 2002; Ford and Goodwin, 2014; Norris and Inglehart, 2019); whether it extends to the legitimation of violence, as opposed to radical belief, is precisely what the typology allows us to ask. Religiosity is included because comparative research identifies moral and religious conservatism as the core of right-wing radicalism outside the Western European frame (Norris and Inglehart, 2011; Rovira Kaltwasser et al., 2026). Digital exposure is included because the claim that internet use drives radicalization is consequential in policy discourse and little tested at the level of mass publics. Age, sex and income enter as controls, and a preference for restricting immigration enters as an eighth predictor because the comparative literature treats immigration attitudes as the most consistently reported correlate of far-right support; because it also serves as an external criterion in Section 4.1, the Wald tests below are repeated without it in the replication package and the configurations remain distinguishable. Left–right self-placement is excluded from the main specification: it is not fielded in eleven of the sixty-four countries, and the standard economic battery is not compatible with this sample. It is retained as an external validation criterion instead, which is the more demanding use of it.

Throughout we report associations, not effects. The design is cross-sectional; the predictors are not exogenous to the configurations with which they are associated; and internet use in particular is plausibly endogenous to education, age, urbanization and political interest. The regression identifies differential predictor profiles across the four regions of the property space, and nothing causal. Country fixed effects absorb every time-invariant national difference, which means that national-level determinants such as regime type cannot be estimated in the same specification; Section 4.4 addresses the country level separately. Predicted probabilities and their intervals are obtained by simulation from the estimated coefficient vector and its cluster-robust covariance matrix, following King et al. (2000).

4 Results

4.1 Assessing the analytical partition: profiles and external criteria

Two labels require caution. “Without measured radicalism” means without a high score on the dimension this article measures, not the absence of any ideology: a respondent may grant political violence justifiability from the radical left, from secular nationalism, from ethnic conflict or from a religious position the index misses. Nor do we describe any respondent as violent. The item records a judgment in the abstract; a person who answers 3 on a ten-point scale has not acted and does not sanction violence in general, but grants it limited justifiability at most. Section 4.5 reports the typology under demanding cut-points.

A property space is a construction, not a discovery of latent kinds, and the warrant for this one is not that respondents cluster into its cells. The latent profile analysis reported in Appendix A.11 does not reproduce the four configurations, and the reason is instructive. At the cut-points that define the space the two properties are statistically independent, as Section 4.2 reports, and under independence they carry no joint information beyond their marginals, so a mixture that fits the observed density has no reason to form components at their intersections: the four-class solution separates respondents along one dimension at a time rather than crossing them. A latent solution that did reproduce the four cells would require the two properties to be associated, which is the proposition this article tests and rejects. The null result is an implication of the central finding rather than evidence against the partition, and the two procedures answer different questions in any case: a mixture optimizes fit to the observed density, while a property space is specified in advance to test necessity, sufficiency and the off-diagonal combinations a nested account says should be rare. The tradition distinguishing constructed from inductively derived typologies makes the point in general terms (Lazarsfeld, 1937; Sartori, 1970; Elman, 2005; Collier et al., 2012): the cells earn their place by carving out regions that behave differently. The partition is therefore assessed against criteria that are not part of the construct: left–right self-placement, the rejection of racial and immigrant neighbors, the preference for restricting immigration, and the party a respondent chooses, none of which enters the index. Correlations are computed within country, which removes national composition.

Figure 1 reports the mean position of each configuration on the two components of the radicalism index, on the violence item, and on nativism, with intervals from a country-clustered bootstrap. The pattern is a crossing. Cultural-hierarchical radicals stand highest on religious exclusivism and gender traditionalism and lowest, with the moderates, on violence; violence legitimizers without measured radicalism stand below the population mean on both radicalism components and at the top, with the radical violence legitimizers, on violence; and radical violence legitimizers are high on everything.

The configurations differ where they are defined to differ and differ far less where they are not, which is the minimum a property space must deliver. Nativism, which is not part of the index and is reported as a cross-cutting attribute, tracks radicalism rather than violence: 29.2 per cent of radical violence legitimizers and 27.5 per cent of cultural-hierarchical radicals reject two or more cultural outgroups as neighbors, against 17.8 of violence legitimizers without measured radicalism and 11.6 of moderates. Exclusionary preferences track radical belief far more closely than the willingness to grant violence justifiability.

The associations run in the expected direction. The within-country correlation of the index with right self-placement is 0.134 (N = 58,977, the item not being fielded in eleven countries), with the rejection of racial and immigrant neighbors 0.123 (N = 79,525), and with the preference for restricting immigration 0.064 (N = 76,128). The violence item behaves differently on every one of them: 0.035, 0.046 and −0.030, respectively. The cultural-hierarchical dimension displays the expected, though modest, associations with right self-placement and with exclusionary criteria; the violence dimension does not, and the divergence is itself part of the argument: the two are not two grades of the same thing.

An obvious objection, that the index merely captures religiosity, can be tested directly. In a country-demeaned regression of right self-placement on the index and on general religiosity jointly (N = 58,688 in 53 countries), the index retains a standardized coefficient of 0.187 (t = 27.5) against 0.069 for religiosity (t = 13.0), and 0.112 (t = 22.5) when restricted to gender traditionalism alone, which contains no religious content.

A first validation against behavior, in Appendix A.6, estimates each party’s position as the mean self-placement of its own voters. That criterion is not external, since it is built from the voters themselves, and it answers a different question: whether a configuration chooses parties further to the right, not parties experts classify as far right.

We merged The PopuList 4.0 (Rooduijn et al., 2024), an expert-informed classification of far-right parties in Europe, and asked whether the configurations differ in the probability of choosing a party that independent experts, not we, classify as far right. Nine countries of our sample have both a party-choice item and PopuList coverage, and parties the PopuList does not classify are not counted as far right. Because nine clusters are few, Appendix A.12 reports the matching party by party, and the replication package the country-by-country estimates and a leave-one-country-out analysis showing that no single case drives the contrast.

What that test can establish has to be stated before its result, because read without the scope of the index in view the result looks like a failure of construct validity and is not one. The index measures a culturally hierarchical, religiously exclusivist orientation and deliberately excludes nativism and authoritarianism, for the measurement reasons set out in Section 2.5; those are the attributes on which the European far right is built. What the test can therefore deliver is a bound on the construct, not a confirmation of it.

The result is not the one we anticipated. In the nine European cases both violence-legitimating configurations choose PopuList far-right parties more often than their countries’ average, by 4.8 percentage points for radical violence legitimizers (country-clustered bootstrap interval +0.6 to +16.6) and 3.6 for the others (+2.1 to +9.2), while the moderates sit 0.8 points below it. The cultural-hierarchical radicals do not: their deviation is −0.2, with an interval from −2.4 to +4.4 containing zero.

Two things follow. The first is external scope validation: the two dimensions have different associations with far-right party choice, on an expert-coded criterion that enters neither the index nor the typology. The second is that the cultural-hierarchical configuration is not the European far-right electorate and should not be described as such. The null is the expected result rather than an anomaly, since the components of nationalist demand that carry far-right support in Europe are the ones this index omits (Halikiopoulou and Vlandas, 2019). The finding also revises the expectation stated in Section 2.4, that radical-right parties might draw on culturally hierarchical but non-violent publics. In the nine cases where it can be tested they do not: the configurations associated with the far-right vote are the two defined by the legitimation of violence, which is the article’s most direct evidence for the double concealment we attribute to the umbrella, on the nine cases where it can be tested. A last limit belongs to the classification: the PopuList identifies far-right parties but does not differentiate their ideological varieties, the gap the PopulisTree was built to fill (Zulianello, 2026).

4.2 The distribution of the four configurations

Of the 80,222 respondents in the sixty-four countries of the analytical sample, 51.6 per cent occupy the moderate cell, low on both dimensions. Cultural-hierarchical radicals, who rank in the upper third of the radicalism distribution while placing political violence at the rejecting end of its scale, are 26.0 per cent, and are the largest non-moderate configuration by a wide margin. Violence legitimizers without measured radicalism, who justify political violence to a certain degree while ranking in the lower two-thirds of the radicalism distribution, are 14.9 per cent. Radical violence legitimizers, who converge on the two properties, are 7.6 per cent. Figure 2 populates the property space of Table 1 with these shares.

The claim under test is about set relations and contains three propositions the literature does not assert equally. The first is a continuum claim: if the two properties were grades of one disposition they would covary substantially. The second is the necessity claim the funnel models assert: if radical opinion were the antechamber of violence, those who grant political violence some justifiability would be contained within those who score high on cultural-hierarchical radicalism. The third is a sufficiency claim, that scoring high on the cultural-hierarchical dimension implies legitimating violence; no funnel model asserts it, since these models suppose that only a fraction of the radical population proceeds; we test it for completeness. The necessity claim is the one the data contradict. Of the respondents above the violence cut, 33.8 per cent score high on cultural-hierarchical radicalism, against the 0.90 consistency that the analysis of necessary conditions conventionally requires, and the configuration that contradicts it directly is 14.9 per cent of the sample. Sufficiency fails as well, at 22.6 per cent, but its failure is not evidence against the funnel. The continuum claim fails too: among respondents high on the cultural-hierarchical dimension the probability of legitimating violence is 22.6 per cent, and among the rest 22.4, an odds ratio of 1.01.

Nor is the size of the off-diagonal cells evidence in its own right. With 33.56 per cent of the sample above the radicalism cut and 22.45 per cent above the violence cut, independence alone would place 26.0 per cent in the radical non-legitimating cell, 14.9 in the violence-legitimating non-radical cell and 7.5 where the two coincide; the observed shares are 26.0, 14.9 and 7.6, and a chi-square test on the dichotomized cross-tabulation returns 0.54 on one degree of freedom (p = 0.46). That the dichotomized association is weaker than the continuous one is no contradiction: the relation is not monotonic, and the probability of legitimating violence rises across the middle of the radicalism distribution and falls in its highest fifth. The correspondence is not a second finding but the same one expressed as a partition: at these cut-points the joint distribution is the product of the marginals. The two off-diagonal configurations are 40.8 per cent of the sample between them, and an account of non-moderation with no place for either describes a minority of it.

A pooled association can conceal a different relation inside national publics, and with half the variance of the cultural-hierarchical dimension lying between countries the objection has to be met rather than deflected. Between the sixty-four country means the correlation is −0.020, unweighted, so it is not the between component of the pooled variance decomposition; within countries, with both axes demeaned by country, it is +0.076, slightly larger than the pooled 0.039 rather than smaller. Estimated separately in each country and combined by random-effects meta-analysis on Fisher’s z, with restricted maximum likelihood, the association is 0.059, with a 95 per cent confidence interval from 0.032 to 0.086 and from 0.032 to 0.087 under Hartung-Knapp. Because that interval locates the mean rather than the spread, we also report the 95 per cent prediction interval, from −0.154 to +0.267. The country correlations run from −0.284 in Pakistan to +0.374 in Canada and I-squared is 94 per cent, so the average describes no particular case well. The set relations survive the same treatment more sharply still. Cutting the cultural-hierarchical dimension at each country’s own upper tertile, which removes national composition entirely, the consistency of the necessity relation has a median of 0.380 and a maximum of 0.644 across the sixty-four countries, and not one attains the 0.90 conventionally required (Schneider and Wagemann, 2012); the consistency of sufficiency has a median of 0.195 and a maximum of 0.589, and none attains the 0.80 usually asked. Within national publics the association therefore remains weak and the necessity relation fails everywhere, although its magnitude and direction vary substantially across countries.

An ordinal-latent estimate of the same correlation is 0.058, and partialling out the shared response style drives it to zero, which we report but do not lean on, for the reason given in Section 3.2. We do not claim orthogonality as a fact about the world, only that an association of this magnitude cannot sustain the premise that radical opinion is the antechamber of violent action.

4.3 Regional distribution

The distribution across regions is close to inverted. Figure 3 gives every national sample and Figure 4 the regional means. Among the seven included African cases, cultural-hierarchical radicals are 50.7 per cent of respondents and violence legitimizers without measured radicalism 7.0, the highest share of the first and the lowest of the second; the figures are 36.9 and 14.0 across the twenty-eight Asian cases, 12.0 and 17.0 across the thirteen Latin American ones, 8.5 and 14.0 across the twelve European ones, 3.4 and 28.1 in the two North American countries of the wave, and 2.7 and 16.6 in Australia and New Zealand.

Three cautions bound the reading. Regional figures rest on small numbers of countries, and the two-country regions are not aggregated, since a bootstrap over two clusters cannot estimate uncertainty. Precision is low: the interval for Africa runs from roughly a third to two thirds, so the ordering is robust and the point estimates are not. The third caution carries the most weight: because full threshold invariance is not achieved, these are differences in the classification of observed responses under a common coding rule, not comparisons of latent means. The ordering survives a direct check on that point. Estimating the polychoric thresholds region by region and rebuilding the dimension from the two items whose thresholds vary least, the ranking is unchanged among the four regions large enough to be aggregated and the rank correlation with the six-item ordering is 0.94; the prevalences compress, as a two-item index must, and Appendix A.3 reports the comparison.

The external criterion adds something further. The index correlates with right self-placement at 0.44 in the United States and Canada and 0.03 in the African cases: the orientation it measures is present everywhere, its coding as “right” in the terms of domestic competition is not. We therefore do not claim that Africa and Asia contain more of the radical right, but that they contain more people in a configuration of hierarchical, exclusivist non-moderation that does not legitimate violence, and that whether it is coded as right-wing depends on the party system it inhabits. Section 5.2 develops the point.

With those cautions in place, the substantive point stands. The regions in which the cultural-hierarchical dimension is most widespread are close to those in which political violence is least justified. A progression from radical belief to violent disposition predicts the opposite.

4.4 Predictors of membership

The multinomial is estimated on 72,530 respondents in sixty-three countries, with country fixed effects and standard errors clustered by country, and Appendix A.15 prints it in full. Coefficients are log-odds against the moderate baseline, with average marginal effects, and are associations. Respondents whose sex is not recorded are dropped; those 25 separate the 72,555 a permissive coding would return from the 72,530 used here.

Religion is the pivot and runs in opposite directions. The importance of religion is the strongest positive correlate of the cultural-hierarchical configuration (b = 0.743, p < 0.001; a predicted probability higher by 7.95 percentage points per standard deviation) and a significant negative correlate of the violence-legitimating configuration without measured radicalism (b = −0.094, p < 0.001; −3.14 points). It is positively associated with radical violence legitimation as well (b = 0.482, p < 0.001). In a multinomial model the marginal effect on one category absorbs redistribution across all of them, so the two quantities attached to the second configuration are not commensurable. The contrast does not depend on that mechanism. Estimated separately on each continuous axis, with country fixed effects, religious importance carries +0.222 on the index in a linear model and −0.097 on the violence item in an ordered logit, both p < 0.001: the opposite signs are a property of the two relations (Appendix A.14). Religion remains the only predictor distinguishable from zero in opposite directions across the two off-diagonal configurations, and that is the clearest predictor-level differentiation between them.

That result invites a charge of circularity, since religious exclusivism is a component of the index. The charge can be answered empirically. Rebuilding the radicalism dimension from gender traditionalism alone, with no religious content, and re-deriving the typology at the same percentile, the opposite-signed contrast survives: religious importance corresponds to a probability of cultural-hierarchical radicalism higher by 3.86 percentage points per standard deviation and a probability of violence legitimation without measured radicalism lower by 1.84, both significant at p < 0.001. The contrast is weaker without the religious component, as it must be, and is not an artefact of the index’s composition.

The remaining predictors differentiate the configurations. Younger age is associated with both violence-based configurations (b = −0.156 and −0.215, both p < 0.001) and unrelated to the cultural-hierarchical one (b = 0.022, n.s.). Education is negatively associated with all three but very unequally: strongly with the two that involve radicalism, barely at all with violence legitimation without measured radicalism. Women are less likely than men to occupy any non-moderate configuration.

These differences are what the partition is meant to capture. Wald tests of the null that two configurations share a coefficient vector are rejected in every pairwise comparison (chi-square = 278.3 for radical violence legitimizers against violence legitimizers without measured radicalism, 118.7 against cultural-hierarchical radicals, and 329.3 for the last pair against each other; each on 8 degrees of freedom, p < 0.001). The regions of the property space are not interchangeable (Figure 5).

One predictor deserves separate treatment. The claim that internet use drives radicalization is consequential in policy discourse and poorly tested in mass publics, and on these indicators it finds no support. Frequent use of the internet as an information source is negatively associated with all three non-moderate configurations, while mobile-phone use runs the other way, positively associated with violence legitimation without measured radicalism (b = 0.092, p < 0.001) and, more weakly, with radical violence legitimation (b = 0.090, n.s.), and unrelated to cultural-hierarchical radicalism: the two are not interchangeable proxies for digital engagement (Settle, 2018; Bail, 2021). Adding a religion-by-internet term, that negative association is concentrated among the least religious: the interaction is positive (b = 0.157, p < 0.001), so the negative relation weakens as religiosity rises and disappears among the most religious (Figure 6). The indicators measure how often a respondent uses a medium for information, not exposure to radical content or peer networks, so they cannot adjudicate a causal thesis, and the sign they show is the opposite of the one the popular narrative predicts (Tucker et al., 2018; Munger and Phillips, 2022). Appendix A.7 reports the models.

A second observation, at the country level, points the same way. The share of the variance that lies between countries is 50.3 per cent for the radicalism dimension and 9.1 per cent for the legitimation of violence: radical opinion is organized by national context, the disposition to grant violence justifiability very largely is not. And across the sixty-four countries the prevalence of cultural-hierarchical radicals correlates at 0.75 with aggregate religiosity while the prevalence of violence legitimizers without measured radicalism correlates at −0.51 with the same quantity.

The geography of the configurations invites a country-level question, and regime type is the obvious candidate. Country fixed effects absorb it, so we answer at the country level. Merging V-Dem’s Regimes of the World classification (Lührmann et al., 2018; V-Dem, 2026), taken as the modal category over 2017–2022, the years the classification covers, with the 2023 fieldwork countries taking the 2022 category, the prevalence of cultural-hierarchical radicals falls monotonically across regime types: 41.5 per cent in closed autocracies, 36.8 in electoral autocracies, 30.3 in electoral democracies and 5.3 in liberal democracies. The violence legitimizers without measured radicalism run the other way, from 8.3 to 19.4 per cent. In a country-level regression on 59 countries, regime type alone explains 27 per cent of the cross-national variance in the cultural-hierarchical share (b = −14.0, p < 0.001).

The association does not survive a control for aggregate religiosity. Once religiosity enters the same regression the regime coefficient collapses to −4.1 and is no longer distinguishable from zero (p = 0.15), while religiosity carries a coefficient of +25.2 (p < 0.001) and the explained variance rises to 58 per cent. Regime type and religiosity are deeply confounded in this sample: the closed autocracies of wave 7 are concentrated among the most religious societies, and fifty-nine countries cannot separate them. We therefore report the association and decline to interpret it as a regime effect. Appendix A.13 reports the full models.

4.5 Robustness

4.5.1 Cut-points on the violence item

A score of 3 on a ten-point scale is a weak endorsement, and the typology should not rest on it. In the pooled sample 67.7 per cent of respondents answer 1, never justifiable, and 83.8 per cent answer 3 or below. Raising the cut-point to 5 leaves 4.3 per cent of respondents in the radical violence-legitimating configuration and 8.1 per cent in the one without measured radicalism; raising it to 6 leaves 2.6 and 4.3 per cent. The religion contrast is unaffected: the marginal effect on cultural-hierarchical radicalism is 7.95 points at a cut of 3, 8.10 at 5 and 8.33 at 6, and the coefficient on violence legitimation without measured radicalism grows slightly in magnitude, from −0.094 to −0.101, although it is estimated less precisely as the category shrinks. The pattern does not depend on the weak end of the scale. Figure 7 reports the grid of percentile and violence cut-points, and Appendix A.9 the demanding ones.

4.5.2 Cut-points on the radicalism dimension

The upper-tertile cut on the pooled distribution is relative, and Appendix A.3 compares it with three alternatives. A country-specific cut cannot support a comparative claim: if high radicalism is the upper third of each country’s own distribution, every country contains close to a third of them by construction, 35.2 per cent of the pooled sample once ties on a six-item mean are placed above the cut, and cross-national variation disappears by definition rather than by evidence. Absolute cut-points on the original scales are the meaningful alternative, and they preserve and amplify the ordering: at a mean of 2.5 on the recoded items, cultural-hierarchical radicals are 67.6 per cent among the African cases, 49.6 across the Asian ones, 29.1 across the Latin American ones, 18.3 across the European ones, 9.6 in the United States and Canada and 6.9 in Australia and New Zealand, against 50.7, 36.9, 12.0, 8.5, 3.4 and 2.7 under the percentile rule. Prevalence changes with the rule; the non-nesting result does not. Appendix A.3 reports the comparison.

4.5.3 The violence axis

Appendix A.4 reports the three violence items by configuration, and the result strengthens the argument: cultural-hierarchical radicals are indistinguishable from moderates on interpersonal violence and sit only slightly above them on the domestic item, and rebuilding the typology with the interpersonal item retains 85.2 per cent of respondents in the same configuration. The domestic-violence item is excluded from the primary axis because it is the item most strongly related to the radicalism dimension (r = 0.131, against 0.039 for political violence): gender traditionalism and the justification of wife-beating share obvious content.

4.5.4 Composition and weights

Removing either component of the index in turn leaves the typology broadly intact and degrades the external validation, which is why both are retained (Appendix A.5). The analyzes are unweighted: the WVS country-normalized weight changes the prevalence of every configuration by less than 0.15 percentage points and leaves the coefficients that carry the argument unchanged in sign and significance (Appendix A.2).

5 Discussion

5.1 A partial counterpart on the demand side

The question was whether the radical–extreme distinction, developed to classify parties, has a counterpart in mass attitudes. It does, but only in part. The two properties the umbrella brings together covary only weakly; the configuration that contradicts a strict necessity claim is 14.9 per cent of the sample and the one that bears on sufficiency a further 26.0 per cent; and the four regions of the property space respond to systematically different social correlates. What the evidence does not deliver is a transposition of the party-level distinction in full: the organizational distinction also turns on the rejection of democracy, on popular sovereignty and on systemic integration, and none of those is observed here. The demand-side foundations of radical and of extreme politics are not one constituency observed at two intensities.

The decomposition has a precedent worth naming. Read through the antilegalist and monist components that Pino Uribe and Cely (2023) distinguish in the study of left-wing extremism, the violence dimension is a partial attitudinal approximation to antilegalism and the cultural-hierarchical dimension captures content compatible with some monist orientations without measuring political monism directly, and at the level of mass publics the two do not travel together. A vocabulary built on the configuration in which both coincide has no name for the two in which they come apart, and those two are five times as large. Because the decomposition was developed to analyze the extreme left, it can be used to test the right rather than assuming it.

This is not a refutation of the two-pyramids model but evidence compatible with it, on data its authors did not have: they separated opinion from action among undergraduates and an American panel, and we find an analogous separation between an ideological orientation and the normative legitimation of violent action across regimes, religions and levels of development. Both of our dimensions are attitudes, so the analogy should not be overstated.

What the evidence cannot bear is a claim about process. In cross-sectional mass attitudes, radical belief is neither a necessary nor a sufficient correlate of the legitimation of political violence. Whether, for some minority, radicalization precedes the acceptance of violence, and whether the passage depends on events or organizations, a snapshot of sixty-four publics cannot answer.

The asymmetry between the two sides of the relation is itself a result. On the supply side the taxonomy of actors has been refined rather than abandoned: a disciplined classification sorts sixty-six parties across thirty-three European countries (Zulianello, 2020), and the PopulisTree sorts populist actors into five types and finer sub-types, among them the populist radical right, applied to the parties The PopuList identifies (Zulianello, 2026). No equivalent classification exists for the publics those actors mobilize, and the partition proposed here is one candidate for that place.

5.2 Religion divides where the umbrella unites

The sharpest empirical result is a single coefficient that changes sign. Religious importance is strongly and positively associated with cultural-hierarchical radicalism and negatively associated with violence legitimation without measured radicalism. It holds within countries, net of national composition; it survives the removal of every religious item from the index; and it reappears between countries, where the prevalence of the two configurations correlates with aggregate religiosity at +0.75 and −0.51.

The implication for the comparative study of the far right is direct. Where religiosity is high the prevalent form of non-moderation is an intense, morally articulated orientation that does not sanction violence; where religiosity is low, it is more likely to be a willingness to grant political violence justifiability unaccompanied by that orientation. The two North American countries in the wave have the highest prevalence of violence legitimation without measured radicalism and one of the lowest of the cultural-hierarchical configuration, a combination the standard radical–extreme vocabulary, formed on European party systems, does not readily classify at the level of individual attitudes. It has described anti-liberal preferences that are ideologically thin and socially embedded (Galston, 2018; Stefanelli et al., 2025) without giving them a place in the same space as the articulated radicalism it takes as its object.

A further implication follows, and it is the one that most challenges the umbrella. The orientation the index measures is present everywhere, but its political coding is not: it correlates with right self-placement at 0.44 in North America and at 0.03 in the African cases. Where the left–right axis structures competition, hierarchical exclusivism is legible as a right-wing position; where it does not, the same orientation exists without that coding. The same asymmetry appears on the supply side. Coding seven Latin American cases on seven ideological attributes, Rovira Kaltwasser et al. (2026, Table 10.1) find an anti-immigration agenda in one of them, Chile, and moral conservatism in all seven, and they read this as a reversal of the European pattern in which nativism is often the paramount issue. The volume adopts the broad concept while its own table shows the attribute that anchors it in Europe present in one case, and its authors are explicit that the evidence they assemble is supply-side and that the bases of support remain to be established, which is the side measured here. That it appears with a different instrument suggests it is a property of the political contexts rather than of the survey items. The far-right umbrella is a category of the supply side, and it presumes a demand side organized on an axis that, outside the contexts in which the category was formed, may organize very little. The PopuList validation makes the point sharper than the correlation does: in the nine European countries where an expert classification can be merged, it is the violence-legitimating configurations, not the cultural-hierarchical one, that vote for far-right parties.

5.3 Measurement implications of studying radical and extreme attitudes comparatively

The measurement work produced two results that bear on any comparative study using these instruments. The first concerns the survey’s items on the justifiability of violence, which share more than a quarter of their variance with a general willingness to justify petty dishonesty fielded in the same battery (r = 0.52). Measures of moral traditionalism drawn from that battery may therefore acquire a negative association with them once reverse-scored, and any estimate of the relation between moral conservatism and the legitimation of violence that draws items from the battery on both sides may be partly attributable to a shared battery component. Westwood et al. (2022a) argued, on American data, that surveys overstate support for political violence; the problem generalizes, and its likely mechanism here is a component shared within a battery that most users treat as a set of independent items. Controlling for the permissiveness marker leaves the religion result untouched (0.743 to 0.742) and roughly halves the age coefficients for the two violence-based configurations, so part of the apparent youth gradient is shared with that marker.

The second concerns authoritarianism, central to the definition of the radical right and which we had intended to include. The items lack cross-context equivalence: fitted separately by region, the standardized loading of the strong-leader item ranges from 0.37 in Latin America to 0.87 in Oceania, the median country-level reliability of the scale is 0.37, and it falls below 0.50 in forty-eight of sixty-four countries. The claim is about instruments, not ideas: we have not shown that nativism and authoritarianism fail as comparative concepts, only that the available World Values Survey indicators for them do not measure the same thing in Caracas and in Auckland. That reading is consistent with a wider reappraisal. Baldner et al. (2026) argue that existing authoritarianism scales confound ideological content with antidemocratic predisposition, and that an ideologically neutral measure locates the disposition on both sides of the spectrum once threat is taken into account, though it remains more prevalent on the right.

6 Conclusion

This article asked whether the radical–extreme distinction, developed for the classification of parties, has a counterpart in mass attitudes. The evidence reveals a discernible but partial one. Cultural-hierarchical radicalism and the legitimation of political violence are not one dimension: their association is 0.039 pooled and 0.059 when country-specific estimates are combined by meta-analysis; a strict necessity claim fails, since violence legitimation without measured radicalism is 14.9 per cent of the sample and no country reaches the conventional necessity threshold under country-specific cut-points; and the four regions of the space have different social correlates and predictor structures that statistical tests reject as equivalent.

The result speaks to three literatures. It supports Zulianello’s (2025) argument for a reason he does not give, that the umbrella conflates properties varying independently not only among parties but among the publics that sustain them, and it meets Pirro’s (2023) defense not by refuting it but by showing that the available World Values Survey indicators of nativism and authoritarianism, on which that defense rests, do not permit a globally comparable operationalization. To the study of radicalization it says that these findings are consistent with the non-nesting intuition of the two-pyramids model in a domain its authors did not examine, and that what Wolfowicz et al. (2020) found for risk factors, and Lygre et al. (2011) for transitions between steps, appears here as the absence of a consistent necessity relation. And to the comparative study of the far right it says that, in many included African and Asian cases, the largest non-moderate category consists of respondents who score high on the measured cultural-hierarchical dimension while rejecting political violence, a configuration a framework developed around European party competition does not readily capture.

The typology is not a discovery of natural kinds but an analytical partition, and it earns its place by carving out regions that behave differently.

Four limitations bound the conclusions. The design is cross-sectional and identifies associations, not causes; whether either configuration drifts or can be recruited under mobilization, as the American ethnographic and affective-polarization literatures suggest (Hochschild, 2016; Mason, 2018), needs panel data we do not have. The measure rests on religious exclusivism and gender traditionalism because the World Values Survey was not designed to study the radical right, and core components of that ideology, among them nativism, welfare chauvinism and democratic rejection, are absent or inconsistently measured. The cultural-hierarchical dimension is therefore one measurable component of right-wing radicalism, not a reconstruction of radical-right ideology, and the PopuList validation turns that limitation into a result: it does not select the European far-right electorate.

The third limitation is interpretative. Regime type does not survive a control for aggregate religiosity, and fifty-nine countries cannot separate the two; likewise the Mertonian reading of violence legitimation without measured radicalism remains a hypothesis the survey cannot test. The fourth is conceptual and, in our view, the most important. The index travels more reliably as a cultural-hierarchical orientation than as a universally recognized position on the political right, correlating with right self-placement at 0.44 in the North American cases and 0.03 in the African ones. That borders on a finding: the demand side is not everywhere pre-organized in the categories European party systems supply.

Several agendas follow. Longitudinal data would establish whether the configurations are stable types or transitional states; behavioral outcomes, from voting to protest, would test whether they differ where it matters; and comparative work should ask whether the same space emerges when the ideological dimension is measured through nativism or authoritarianism where those indicators travel.

None of this is an argument for abandoning the far-right label. It is an argument for not letting the label dissolve the radical–extreme distinction, among organizations or among the attitudes that sustain them. The umbrella may conceal heterogeneity twice: among the organizations it classifies, and among the attitudinal constituencies on which they draw. Recovering the distinction on the demand side changes what researchers count, whom they compare and how they measure. The relationship between cultural-hierarchical radicalism and the legitimation of political violence should not be assumed as a defining feature of political extremism, but investigated as an empirical question whose answer may vary across societies and political traditions.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Generative AI (Claude, Anthropic) was used as an editorial assistant during manuscript revision. The authors designed the study, conducted all analyzes, made all interpretive decisions, and take full responsibility for the content of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpos.2026.1908561/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Mattia Zulianello, University of Trieste, Italy

Reviewed by: Adam Biggs, United States Department of the Navy, United States

Stefano Sangiovanni, University of Milan, Italy