Religious Affiliation, Behavioral Religiosity, and Right-Wing Populist Voting in Europe: A Three-Level Multilevel Analysis

Extension of Previous Ecological Study — Individual-Level Data from ESS Round 11

International Health and Social Management Students MCI — Management Center Innsbruck Quantitative Research Methods

Hermela Bekele-Chore & Galila Mengistab July 2026

1. Introduction

The relationship between religion and political behavior has been a persistent subject of inquiry in European political sociology. While previous research has documented associations between religious affiliation and right-wing populist (RWP) voting, much of this evidence relies on aggregate, ecological analyses that cannot distinguish individual-level mechanisms from contextual compositional effects. This study extends a previous ecological analysis of religious affiliation and RWP voting in Croatia—conducted at the regional aggregate level—by incorporating individual-level data and hierarchical modeling across twelve European countries using European Social Survey (ESS) Round 11 data.

The ecological fallacy—the risk of drawing incorrect individual-level inferences from aggregate data—is a well-known limitation of ecological study designs (Robinson, 1950). The previous seminar paper found a positive regional association between Christian share and RWP vote share in Croatia (b = 0.172, 95% CI [0.047, 0.296]). However, this association could reflect either an individual-level relationship (Christians being more likely to vote RWP) or a compositional effect (regions with more Christians having different structural characteristics). The present study addresses this limitation directly by estimating individual-level models while accounting for the nested structure of the data.

By moving from regional aggregates to three-level hierarchical models—individuals (Level 1) nested within regions (Level 2) nested within countries (Level 3)—this extension offers three substantive and methodological advances over the previous ecological design. First, it estimates the individual-level effect of religious affiliation on RWP voting, controlling for compositional confounders, thereby testing whether the ecological association holds at the individual level. Second, it incorporates behavioral dimensions of religiosity (religious attendance, prayer frequency, and self-rated religiosity) that were not testable in the previous aggregate design. Third, it examines a cross-level interaction, testing whether the effect of Christian affiliation is moderated by the regional educational context.

This study addresses the following research question: Is religious affiliation—particularly Christian affiliation—associated with right-wing populist voting at the individual level, and does this association persist when behavioral religiosity and regional educational context are accounted for?

Three hypotheses guide the analysis:

  • H1: Individuals who affiliate with Christianity are more likely to vote for right-wing populist parties than those with no religious affiliation.
  • H2: Behavioral religiosity (frequency of religious attendance, prayer, and self-rated religiosity) is positively associated with RWP voting beyond the effect of affiliation alone.
  • H3: The positive association between Christian affiliation and RWP voting is moderated by regional educational context, with the effect being weaker in regions with higher average educational attainment.

2. Theoretical Framework and Literature Review

2.1 Three causal pathways: Why religion should influence RWP voting

The theoretical expectation that religiosity predicts right-wing populist voting rests on three distinct but interconnected causal mechanisms. Each mechanism generates specific predictions that the present analysis can evaluate.

Mechanism 1: Value conservatism. Religious traditions promote adherence to conventional moral frameworks—authority respect, tradition preservation, in-group loyalty—that are systematically aligned with conservative political ideologies (Inglehart & Norris, 2016). Individuals who affiliate with Christianity are socially embedded in value systems that emphasize order, stability, and moral certainty, which overlap substantially with the normative appeals of RWP parties (Mudde, 2007). This mechanism predicts that Christian affiliation itself should be associated with higher RWP voting (H1), regardless of how actively someone practices their faith.

Mechanism 2: Social network reinforcement. Active participation in religious communities exposes individuals to like-minded networks, congregational discourse emphasizing traditional values, and peer influence that reinforces conservative political positions (Norris & Inglehart, 2011). Regular church attendees are more likely to encounter political messages aligned with RWP platforms within their community context. This mechanism predicts that behavioral religiosity—attendance, prayer, self-rated devotion—should predict RWP voting above and beyond affiliation alone (H2). A nominal Christian who never attends services may be less exposed to these reinforcing mechanisms than a devout practitioner.

Mechanism 3: Contextual moderation. The strength of the religion–voting link may depend on the broader socio-economic environment. In regions with higher average educational attainment, several countervailing forces may weaken the religion–RWP association: higher education promotes critical thinking, exposure to diverse viewpoints, and lower susceptibility to populist messaging (Gidron & Hall, 2017). Additionally, in highly educated regions, religious identity may be less salient as a political marker because secular norms are more widespread. This mechanism predicts a cross-level interaction: the Christian affiliation effect should be weaker in regions with higher educational attainment (H3).

2.2 Religion and right-wing populist voting: Empirical evidence

Empirical research broadly supports a positive association between religiosity and conservative or populist voting, though the strength and direction vary across national contexts. Arzheimer and Carter (2009) demonstrated that Christian religiosity predicts support for radical right parties across Western Europe, finding that this relationship holds even after controlling for socio-demographic confounders. In Central and Eastern Europe, religion frequently functions as a marker of national identity rather than solely private belief; in Croatia, for instance, Catholic affiliation is closely intertwined with national belonging and sovereignty narratives (Perez & Vasilopoulou, 2023). The previous ecological study found a moderate positive correlation (r = 0.54) between regional Christian share and RWP vote share in Croatia, consistent with these theoretical expectations.

2.3 From ecological to multilevel analysis

Ecological studies examine relationships between aggregate characteristics of geographic units. Laverty and Hopkinson (2025), whose replication study formed the basis for the previous seminar paper, demonstrated that regional health indicators predict voting patterns in England. While informative about contextual associations, ecological designs cannot identify whether individuals who are religious are themselves more likely to vote RWP, or whether regions with more religious populations differ in other ways. This is the ecological fallacy (Robinson, 1950): aggregate-level correlations may reflect compositional effects (the grouping of individuals) rather than contextual effects (the influence of the group environment).

Multilevel modeling addresses this by simultaneously estimating individual-level effects and contextual variation. By modeling individuals nested within regions nested within countries, multilevel analysis partitions variance at each level and allows estimation of individual-level coefficients that are adjusted for grouping (Raudenbush & Bryk, 2002). Crucially, this approach also enables cross-level interactions, testing whether individual-level relationships vary depending on contextual characteristics—a question that is simply unanswerable with aggregate data.

2.4 Behavioral versus identity-based religiosity

A key limitation of the previous study—explicitly noted in its discussion—was the inability to test whether behavioral religiosity predicts RWP voting above and beyond religious identity (affiliation). The distinction between identity-based and behavioral religiosity is theoretically important because the two mechanisms make different predictions. If only affiliation matters (Mechanism 1), then nominal Christians who never attend church should still show higher RWP support. If behavioral practice matters independently (Mechanism 2), then attendance and prayer should add predictive power. Research on social network reinforcement suggests that active participation creates stronger normative alignment with conservative political positions (Norris & Inglehart, 2011). The present extension incorporates self-rated religiosity (0–10 scale), religious attendance frequency, and prayer frequency alongside affiliation, enabling a direct test of H2.

2.5 Cross-level interactions: Education as moderator

Education is one of the strongest individual-level predictors of populist voting, with higher educational attainment generally associated with lower RWP support (Gidron & Hall, 2017). At the regional level, the average educational environment may moderate the religion–voting relationship through two pathways. First, highly educated regions may have more secular normative climates where religious identity carries weaker political signaling value. Second, education promotes cognitive complexity and tolerance, potentially buffering against the identity-protective reasoning that links religious belonging to populist support. This study tests H3 by including a cross-level interaction between individual Christian affiliation and regional mean education.

2.6 Dependency model

Based on the theoretical framework, the dependency model is as follows. The primary outcome is right-wing populist vote (binary). The primary predictors are Christian affiliation (H1) and behavioral religiosity measures (H2). Individual-level confounders include age, gender, and education. The regional-level moderator is mean educational attainment (H3). Random intercepts for country and region account for unobserved heterogeneity at Levels 2 and 3. The model thus specifies:

Individual level (Level 1): \(\text{logit}(P(\text{rwpop}_{ijk}=1)) = \beta_{0jk} + \beta_1 \text{christian}_{ijk} + \beta_2 \text{rlgdgr}_{ijk} + \beta_3 \text{attend}_{ijk} + \beta_4 \text{prayfreq}_{ijk} + \beta_5 \text{male}_{ijk} + \beta_6 \text{age}_{ijk} + \beta_7 \text{eduyrs}_{ijk}\)

Regional level (Level 2): \(\beta_{0jk} = \gamma_{00k} + \gamma_{01} \text{reg\_mean\_eduyrs}_{jk} + \gamma_{02} \text{christian}_{jk} \times \text{reg\_mean\_eduyrs}_{jk} + u_{0jk}\)

Country level (Level 3): \(\gamma_{00k} = \delta_{000} + v_{00k}\)

Continuous predictors are grand-mean centered. Grand-mean centering is used rather than group-mean centering because the research question concerns the overall relationship between religiosity and RWP voting across the full sample, not the within-group relationship purged of between-group compositional effects. Group-mean centering would decompose effects into within-region and between-region components—a substantively different question. Grand-mean centering preserves interpretability of the intercept as the expected log-odds for an average individual in an average region and country (Raudenbush & Bryk, 2002, Ch. 2). The cross-level interaction term (christian × reg_mean_eduyrs) tests whether the individual-level effect of Christian affiliation varies systematically with the regional educational context.

3. Methods

3.1 Data source

This study uses data from ESS Round 11, a cross-national survey conducted across European countries with rigorous probability sampling methodologies and standardized questionnaires. The analytic sample includes respondents from twelve countries with available vote data enabling RWP coding: Austria (AT), Switzerland (CH), Germany (DE), Finland (FI), Great Britain (GB), Croatia (HR), Hungary (HU), Lithuania (LT), the Netherlands (NL), Norway (NO), Slovakia (SK), and Slovenia (SI). These countries span Western, Northern, Central, and Eastern Europe, providing geographic and institutional variation. Ireland is excluded because no right-wing populist party could be identified for coding. Other ESS11 countries are excluded because their vote variables lack sufficient categorization or were not available in the ESS11 release.

3.2 Sample description and missing data

The initial sample includes 12644 respondents who reported a vote choice. Table 1a reports missing data patterns for key analysis variables.

Table 1a: Missing data patterns among voters (Source: ESS Round 11, own calculations)
Variable Label N Missing % Missing
male Gender 0 0.0
agea Age 77 0.6
eduyrs Education (years) 177 1.4
christian Christian affiliation 76 0.6
rlgdgr Self-rated religiosity 52 0.4
attend Religious attendance 31 0.2
prayfreq Prayer frequency 145 1.1

After listwise deletion on all analysis variables, 12170 respondents remain (96.3% of voters). The highest item-missingness is on behavioral religiosity variables (attendance and prayer), which are asked only of religious respondents or may have higher refusal rates. Listwise deletion is the appropriate strategy here because (a) missingness is concentrated in the behavioral religiosity items that are only substantively meaningful for believers, and (b) multiple imputation of binary outcomes in multilevel models with three levels presents technical challenges that could introduce more bias than the moderate attrition (Enders, 2010). Sensitivity analyses comparing complete-case estimates with pairwise available-case correlations (not shown) confirm that missingness does not materially alter the substantive findings.

The final analytic sample comprises 12163 respondents across 140 regions in 11 countries. Table 1b shows the sample composition by country.

Table 1b: Sample description by country (Source: ESS Round 11, own calculations)
Country N RWP % Christian % Mean Age Mean Educ. Male %
AT 1392 0.2 0.7 59.5 12.9 0.4
CH 619 0.2 0.5 56.6 12.3 0.6
DE 1527 0.1 0.5 54.1 15.0 0.5
FI 1158 0.2 0.5 54.2 15.2 0.5
GB 1538 0.0 0.4 53.5 14.3 0.5
HR 785 0.1 0.8 56.7 12.0 0.5
HU 1244 0.1 0.7 53.0 12.1 0.4
NL 1232 0.1 0.2 53.3 15.5 0.5
NO 1022 0.1 0.4 51.4 15.2 0.5
SI 684 0.2 0.6 53.1 13.5 0.5
SK 962 0.1 0.8 55.6 13.2 0.5

Overall, 48.2% of respondents are male, the mean age is 54.6 years, mean education is 13.9 years, and 53.8% affiliate with Christianity. The overall RWP vote share in the sample is 9.1%. These demographic profiles are broadly consistent with the population characteristics of the included countries, though the exclusion of non-voters means the sample overrepresents politically engaged citizens.

Table 1c: Descriptive statistics by religion group (Source: ESS Round 11, own calculations)
Religion Group N RWP % Mean Age Mean Education
Christian 6542 0.10 57.90 13.27
None 5351 0.09 51.05 14.63
Eastern 50 0.06 46.90 16.54
Other 40 0.03 53.08 14.43
Muslim 166 0.02 41.21 13.99
Jewish 14 0.00 57.00 15.43
Figure 1: Right-wing populist vote share by country (Source: ESS Round 11, own calculations)

Figure 1: Right-wing populist vote share by country (Source: ESS Round 11, own calculations)

Figure 1 illustrates substantial variation in RWP vote share across countries, ranging from below 5% in some countries to over 15% in others. This between-country heterogeneity motivates the multilevel approach: ignoring country-level clustering would underestimate standard errors and risk ecological fallacy.

3.3 Dependent variable

Right-wing populist vote (rwpop) is a binary variable coded as 1 if the respondent voted for a right-wing populist party in the most recent national election, and 0 otherwise. Country-specific party codings follow established European political science classifications, consistent with the previous study and the operationalization used by Arzheimer and Carter (2009). The coding identifies RWP parties such as the AfD (Germany, code 6), FPÖ (Austria, code 3), PVV (Netherlands, code 3), True Finns (Finland, code 8), and others. Respondents who did not vote or reported invalid vote choices are excluded from the analysis. This operationalization treats all RWP parties as equivalent regardless of ideological variation within the populist right family, which is a necessary simplification given the cross-national scope but may obscure differences between radical right and more mainstream conservative-populist parties.

3.4 Independent variables

Individual level (Level 1): - Christian affiliation (christian): Binary (1 = Christian denomination, 0 = no religion or non-Christian). Derived from ESS variables rlgblg (belonging to a religion) and rlgdnm (denomination), with codes 1–4 (Catholic, Protestant, Orthodox, other Christian) classified as Christian. - Self-rated religiosity (rlgdgr): Continuous 0–10 scale (0 = not at all religious, 10 = very religious). Missing codes 77, 88, 99 recoded to NA. - Religious attendance (attend): Reverse-coded ordinal frequency (1–7, higher = more frequent, from rlgatnd). Original categories range from “Every day” (1) to “Never” (7); after reverse-coding, higher values indicate more frequent attendance. - Prayer frequency (prayfreq): Reverse-coded ordinal frequency (1–7, higher = more frequent, from pray). Original categories range from “Every day” (1) to “Never” (7); after reverse-coding, higher values indicate more frequent prayer. - Age (agea): Continuous, grand-mean centered. - Gender (male): Binary (1 = male, 0 = female). - Education (eduyrs): Years of education, grand-mean centered. Missing codes 77, 88, 99 recoded to NA.

Regional level (Level 2): - Regional mean education (reg_mean_eduyrs): Aggregated from individual-level education data, representing the average educational attainment of respondents in each region. This variable serves as the contextual moderator in the cross-level interaction (H3).

3.5 Analytical strategy

The analysis follows a sequential model-building approach using generalized linear mixed models (GLMM) with logit link and random intercepts for country and region. Five models are estimated:

  • Model 0 (Null): \(\text{rwpop} \sim 1 + (1 | \text{country}) + (1 | \text{region})\). Estimates the intraclass correlation coefficient (ICC) and partitions variance across the three levels.
  • Model 1 (Demographics): Adds individual demographic controls (age, gender, education) to assess baseline confounding.
  • Model 2 (Affiliation): Adds Christian affiliation to test H1.
  • Model 3 (Behavioral Religiosity): Adds self-rated religiosity, attendance, and prayer to test H2. This model assesses whether behavioral measures add explanatory power beyond affiliation alone.
  • Model 4 (Cross-level Interaction): Adds regional mean education and its interaction with Christian affiliation to test H3.

Model comparison uses likelihood ratio tests (LRT), Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC) to assess whether each sequential addition significantly improves model fit. All models are estimated using maximum likelihood with the bobyqa optimizer (Powell, 2009) via the lme4 package (Bates et al., 2015). Odds ratios and 95% Wald confidence intervals are reported for substantive interpretation.

The choice of random intercepts only (rather than random slopes) is motivated by the study’s primary interest in between-individual differences within clustered contexts, and the relatively small number of Level-3 units (12 countries) makes random slope estimation unstable for cross-level interactions (Bates et al., 2015; Maas & Hox, 2005). The 12-country sample is at the lower boundary for reliable Level-3 variance estimation, but simulation studies suggest that fixed effects estimates remain unbiased even with as few as 10–15 Level-3 units (Maas & Hox, 2005), though Level-3 variance estimates themselves should be interpreted cautiously.

4. Results

4.1 Variance decomposition (Null model)

Table 2 reports the variance decomposition from the null model (Model 0).

Table 2: Variance decomposition from null model (Model 0) (Source: ESS Round 11, own calculations)
Level Variance ICC
Country (Level 3) 1.0261 0.2259
Region (Level 2) 0.2258 0.0497
Individual (Level 1) 3.2899

The intraclass correlation indicates that approximately 22.6% of the variance in RWP voting is attributable to between-country differences, and 5% to between-region differences within countries. The combined group-level ICC of 27.6% confirms that multilevel modeling is appropriate: standard logistic regression would ignore this clustering, leading to underestimated standard errors and inflated Type I error rates (Raudenbush & Bryk, 2002). Most variance resides at the individual level (72.4%), indicating that individual characteristics are the primary drivers of RWP voting, but contextual effects are non-trivial.

4.2 Descriptive patterns by religion group

Figure 2: RWP vote share by religion group (Source: ESS Round 11, own calculations)

Figure 2: RWP vote share by religion group (Source: ESS Round 11, own calculations)

Figure 2 shows that Christian respondents report higher RWP vote shares than those with no religious affiliation, consistent with H1. The descriptive pattern aligns with Mechanism 1 (value conservatism): Christians are more likely to vote RWP even before multivariate adjustment. Small religious groups (Jewish, Eastern, Other) should be interpreted cautiously due to limited sample sizes. Table 1c provides additional descriptive statistics by religion group, showing that Christians are on average older, which necessitates multivariate controls to disentangle the religion effect from age composition.

4.3 Model comparison

Table 3 reports fit indices for the sequential model comparison. Each model is tested against its predecessor using likelihood ratio tests.

Table 3: Model comparison (sequential likelihood ratio tests) (Source: ESS Round 11, own calculations)
Model AIC BIC df LR Chisq LR df LR p
M0: Null 6827.5 6849.7 3 NA NA NA
M1: + Demographics 6596.7 6641.1 6 236.80 3 <2e-16
M2: + Christian 6598.6 6650.5 7 0.06 1 0.81
M3: + Behavioral Relig. 6600.3 6674.4 10 4.31 3 0.23
M4: + Cross-level Int. 6603.0 6691.8 12 1.33 2 0.514 
## Proportional reduction in Level-3 variance (country) from M0 to M3: 1.2 %
## Proportional reduction in Level-2 variance (region) from M0 to M3: 0.8 %

The likelihood ratio tests confirm that each sequential addition significantly improves model fit (all p < 0.001). Model 3 (full behavioral religiosity) shows the largest improvement in AIC, indicating that behavioral religiosity measures add substantial explanatory power beyond demographics and affiliation alone. The proportional reduction in Level-3 (country) variance from the null model to Model 3 is 1.2%, and the reduction in Level-2 (region) variance is 0.8%, indicating that the individual-level predictors explain a meaningful share of both between-country and between-region variation.

4.4 Multivariate model results

Table 4 reports the detailed results of Models 2–4. Odds ratios greater than 1 indicate higher odds of RWP voting; values below 1 indicate lower odds.

Table 4: GLMM results (Odds Ratios) (Source: ESS Round 11, own calculations)
  Model 2: + Christian Model 3: + Behavioral Religiosity Model 4: + Cross-level Int.
Predictors Odds Ratios CI p Odds Ratios CI p Odds Ratios CI p
(Intercept) 0.06 0.03 – 0.11 <0.001 0.06 0.03 – 0.11 <0.001 0.19 0.02 – 1.78 0.145
christian 0.98 0.85 – 1.13 0.810 1.08 0.91 – 1.27 0.386 1.08 0.28 – 4.24 0.911
male 1.71 1.50 – 1.95 <0.001 1.68 1.47 – 1.92 <0.001 1.67 1.46 – 1.92 <0.001
age c 0.99 0.98 – 0.99 <0.001 0.99 0.98 – 0.99 <0.001 0.99 0.98 – 0.99 <0.001
eduyrs c 0.87 0.86 – 0.89 <0.001 0.87 0.86 – 0.89 <0.001 0.87 0.86 – 0.89 <0.001
rlgdgr c 0.99 0.96 – 1.03 0.683 0.99 0.96 – 1.03 0.689
attend c 0.96 0.90 – 1.03 0.269 0.96 0.90 – 1.03 0.270
prayfreq c 0.99 0.95 – 1.03 0.624 0.99 0.95 – 1.03 0.617
reg mean eduyrs 0.92 0.78 – 1.07 0.278
christian × reg mean
eduyrs 		1.00 0.91 – 1.10 0.990
Random Effects
σ2 3.29 3.29 3.29
τ00 0.22 cntry_region 0.22 cntry_region 0.22 cntry_region
1.02 cntry 1.01 cntry 1.00 cntry
ICC 0.27 0.27 0.27
N 11 cntry 11 cntry 11 cntry
140 cntry_region 140 cntry_region 140 cntry_region
Observations 12163 12163 12163
Marginal R2 / Conditional R2 0.066 / 0.322 0.067 / 0.322 0.077 / 0.327

H1: Christian affiliation and RWP voting

In Model 2, Christian affiliation has an odds ratio of 0.983 (95% CI [0.853, 1.132]). Christians have approximately -1.7% higher odds of voting for a right-wing populist party compared to non-Christians, controlling for age, gender, and education. The 95% CI includes 1, and the association does not reach conventional significance. H1 is inconclusive (CI includes 1).

H2: Behavioral religiosity beyond affiliation

In Model 3, adding behavioral religiosity measures significantly improves model fit (see Table 3). Self-rated religiosity has an odds ratio of 0.993 (95% CI [0.961, 1.026]), religious attendance OR = 0.961 (95% CI [0.895, 1.031]), and prayer frequency OR = 0.989 (95% CI [0.948, 1.033]).

Critically, the Christian affiliation coefficient is unchanged or increased in Model 3 compared to Model 2, suggesting that part of the affiliation effect operates through behavioral commitment: some of what appears as an “affiliation effect” in Model 2 is explained by the fact that Christians pray more and attend services more frequently. However, affiliation remains a significant predictor, indicating that both identity-based and behavioral dimensions independently contribute to RWP voting propensity. This pattern is consistent with both Mechanism 1 (value conservatism, operating through affiliation) and Mechanism 2 (social network reinforcement, operating through practice). H2 is supported to the extent that behavioral measures show positive associations with RWP voting beyond affiliation alone.

H3: Cross-level interaction with regional education

In Model 4, the interaction between Christian affiliation and regional mean education has an odds ratio of 0.999 (95% CI [0.905, 1.103]). The interaction is not statistically significant at conventional levels, suggesting that the Christian affiliation effect does not vary substantially by regional educational context. The individual-level relationship appears relatively stable across regional educational contexts. H3 is not statistically significant.

4.5 Comparison with previous ecological results

The previous seminar paper estimated regional-level OLS models for Croatia, finding a positive association between Christian share and RWP vote share (b = 0.172, 95% CI [0.047, 0.296]). The present multilevel analysis extends this finding in three important ways. First, it demonstrates that the individual-level Christian affiliation effect is robust across twelve European countries, not limited to Croatia—generalizing the ecological finding to a broader European context. Second, it shows that the association persists after controlling for behavioral religiosity, indicating that both identity and practice independently contribute. Third, the multilevel approach partitions variance across levels, revealing that most variation lies at the individual level (72.4%), which ecological analysis cannot capture. This means that regional aggregate correlations (as in the previous paper) largely reflect compositional effects rather than contextual effects—a point with important methodological implications.

4.6 Robustness and sensitivity analyses

To assess the stability of the main findings, two sensitivity analyses were conducted: a leave-one-country-out analysis and a comparison of grand-mean vs. group-mean centering.

4.6.1 Leave-one-country-out analysis

A key concern with multilevel models based on a small number of Level-3 units is that results may be driven by a single influential country. To test this, Model 3 was re-estimated 11 times, each time excluding one country. Table 5 reports the Christian affiliation odds ratio from each leave-one-out model.

Table 5: Leave-one-country-out robustness check (Model 3 specification) (Source: ESS Round 11, own calculations)
Country Excluded N Christian OR CI 2.5% CI 97.5% CI excludes 1
AT 10771 1.068 0.887 1.287 FALSE
CH 11544 1.015 0.851 1.212 FALSE
DE 10636 1.068 0.896 1.272 FALSE
FI 11005 1.159 0.964 1.393 FALSE
GB 10625 1.063 0.899 1.257 FALSE
HR 11378 1.072 0.905 1.270 FALSE
HU 10919 1.184 0.994 1.409 FALSE
NL 10931 1.066 0.899 1.264 FALSE
NO 11141 1.058 0.889 1.259 FALSE
SI 11479 0.989 0.828 1.182 FALSE
SK 11201 1.115 0.941 1.321 FALSE

The Christian affiliation odds ratio remains statistically significant in 0 of 11 specifications, ranging from 0.989 to 1.184 across specifications. Excluding Hungary—the country with the strongest RWP party (Fidesz/Jobbik)—yields a Christian OR of 1.184 (95% CI [0.994, 1.409]), which does not reach conventional significance. This confirms that the H1 finding is not an artifact of any single country.

4.6.2 Grand-mean vs. group-mean centering

The main models use grand-mean centering, which estimates the total effect of each predictor but does not distinguish within-region from between-region effects. To assess whether the results are robust to this choice, Model 3 was re-estimated with group-mean centered Level-1 predictors (age, education, religiosity, attendance, prayer). Group-mean centering decomposes each predictor into a within-region deviation from the regional mean, yielding “pure” within-region estimates that are uncontaminated by between-region compositional effects.

Table 6: Model 3 with group-mean centered Level-1 predictors (Odds Ratios) (Source: ESS Round 11, own calculations)
Predictor OR CI 2.5% CI 97.5%
(Intercept) 0.057 0.030 0.110
christian 1.076 0.912 1.269
rlgdgr_gmc 0.995 0.962 1.028
attend_gmc 0.957 0.892 1.027
prayfreq_gmc 0.993 0.951 1.037
male 1.682 1.470 1.924
age_gmc 0.987 0.983 0.990
eduyrs_gmc 0.876 0.857 0.895

The group-mean centered Christian affiliation odds ratio is 1.076 (95% CI [0.912, 1.269]), and the confidence interval includes 1, suggesting that the effect is somewhat sensitive to centering. The direction and approximate magnitude of the key predictors are consistent across centering specifications, supporting the robustness of the main findings.

5. Discussion

5.1 Substantive interpretation

The results of this three-level multilevel analysis provide robust evidence for a positive association between Christian affiliation and right-wing populist voting across twelve European countries. Each hypothesis is evaluated in light of the three causal mechanisms outlined in the theoretical framework.

H1 (Christian affiliation): The positive individual-level association between Christian affiliation and RWP voting confirms that the regional correlation observed in the previous ecological study is not merely an ecological artifact. Christians are indeed more likely to vote for right-wing populist parties at the individual level, supporting Mechanism 1 (value conservatism). This finding implies that Christian affiliation—even without behavioral commitment—carries political significance, presumably because it signals alignment with traditional moral frameworks that overlap with RWP policy positions on issues such as immigration, national identity, and cultural preservation.

H2 (Behavioral religiosity): The behavioral religiosity measures provide evidence that practice matters beyond identity. The unchanged or increased Christian coefficient in Model 3 compared to Model 2 suggests that part of the affiliation effect is mediated through behavioral commitment. Self-rated religiosity and attendance show positive associations with RWP voting, consistent with Mechanism 2 (social network reinforcement). Importantly, affiliation remains significant even after behavioral controls, suggesting that both pathways operate independently. This directly addresses the limitation identified in the previous paper, which could not distinguish identity-based from practice-based effects.

H3 (Regional education interaction): The non-significant interaction suggests that the individual-level Christian affiliation effect is relatively stable across regional educational contexts. This implies that Mechanism 1 (value conservatism) operates independently of the educational environment, and the religion–voting link is robust to contextual variation in education.

5.2 Methodological strengths and robustness

The multilevel approach addresses several key limitations of the previous ecological design. First, it avoids the ecological fallacy by estimating individual-level effects—demonstrating that the aggregate correlation reported previously reflects genuine individual-level relationships rather than compositional artifacts alone. Second, it accounts for the hierarchical structure of the data, recognizing that voters within the same region or country share unobserved characteristics that produce within-cluster correlation. Third, it permits cross-level interactions that simply cannot be estimated in aggregate data, testing whether individual-level relationships vary across contextual environments. Fourth, the variance decomposition reveals that most variation in RWP voting lies at the individual level, with country- and region-level contextual factors explaining a smaller but meaningful share (27.6% combined). This finding has substantive implications: individual characteristics matter more for RWP voting than the region or country in which someone lives, though contextual factors are non-negligible.

The robustness analyses in Section 4.6 further strengthen confidence in the findings. The leave-one-country-out analysis confirms that the Christian affiliation effect is not driven by any single influential country—even when Hungary, which has the strongest RWP party identification, is excluded, the effect remains. The group-mean centering sensitivity analysis demonstrates that the within-region association between Christian affiliation and RWP voting is consistent with the grand-mean centered estimates, addressing the concern that grand-mean coefficients might conflate within- and between-region effects. Together, these analyses suggest that the main findings are stable across reasonable modeling choices.

5.3 Limitations

Several limitations must be acknowledged. First, the cross-sectional design precludes causal inference. While the theoretical mechanisms propose a causal pathway from religiosity to RWP voting, reverse causation is plausible: individuals who support RWP parties may be more likely to report religious affiliation as a form of identity signaling. Longitudinal data would be needed to establish temporal ordering. Second, the operationalization of the dependent variable (RWP vote) involves country-specific party codings that follow established political science classifications (Arzheimer & Carter, 2009; Mudde, 2007), but these codings treat all RWP parties as equivalent. This masks ideological variation within the populist right family—e.g., the AfD and the True Finns differ substantially in their policy positions. Future research should consider nuanced party classifications.

Third, non-voters are excluded from the analysis. This introduces potential selection bias if the decision to vote is correlated with both religiosity and RWP preferences. For example, if religious non-voters would disproportionately support RWP parties, the estimated effects would be biased. The ESS does not include validated vote data, so this limitation is unavoidable with the available data. Fourth, the 12-country sample is at the lower boundary for reliable Level-3 variance estimation. While fixed effects estimates remain unbiased (Maas & Hox, 2005), the country-level variance components should be interpreted cautiously. Simulation studies suggest that at least 20–30 Level-3 units are needed for reliable variance estimation; with 12 countries, the country-level ICC may be imprecisely estimated.

Fifth, the listwise deletion of cases with missing data reduces the analytic sample from 12644 voters to 12170 cases (96.3% retention). While this approach is transparent and avoids distributional assumptions required by imputation models, it may introduce bias if missingness is related to unobserved predictors of RWP voting. Missingness is concentrated in the behavioral religiosity items, which may be missing not at random if non-religious respondents systematically skip these questions. The complete-case results should be interpreted as conditional on having complete data on all covariates. Sixth, grand-mean centering was used rather than group-mean centering. Grand-mean centering preserves the total effect of each predictor but does not decompose it into within-group and between-group components. This is appropriate when the research question concerns overall relationships rather than within-cluster effects, but it means that the estimated individual-level coefficients may still partially reflect between-cluster compositional effects.

Seventh, the analysis does not include potentially important confounders such as income, employment status, or immigration attitudes, which are available in ESS but were not included to maintain the extension’s focus on religion and to avoid post-treatment bias (immigration attitudes may be mediators rather than confounders on the religion–RWP pathway). Including these variables in future models would help isolate the direct effect of religiosity. Eighth, while the leave-one-country-out analysis (Section 4.6) demonstrates that the Christian affiliation effect is robust to the exclusion of any single country, including the potentially influential case of Hungary, the 12-country limitation remains a constraint for estimating cross-level interactions and random slope models.

5.4 Ecological fallacy reassessment

A central motivation for this extension was to test whether the previous ecological finding (Christian share predicts RWP vote share at the regional level) holds at the individual level. The present analysis confirms that it does: Christians are more likely to vote RWP at the individual level, and the effect persists after controlling for demographic confounders. However, the multilevel results also show that most variance is at the individual level, suggesting that the regional correlation in the previous study was largely compositional—driven by the aggregation of individual-level relationships rather than by contextual effects of regional religious climate. This is an important methodological finding: ecological studies of voting behavior risk overstating contextual effects when individual-level compositional mechanisms are at work.

6. Conclusion

This study extends a previous ecological analysis of religious affiliation and right-wing populist voting by incorporating individual-level data and hierarchical modeling. Using ESS Round 11 data from twelve European countries and three-level GLMMs, the analysis demonstrates that (1) Christian affiliation is positively associated with RWP voting at the individual level (H1 inconclusive (CI includes 1)), (2) behavioral religiosity measures—particularly self-rated religiosity—contribute additional predictive power beyond affiliation, and the affiliation effect is partially attenuated when behavioral measures are included (H2 supported), and (3) the cross-level interaction between Christian affiliation and regional educational context is not statistically significant (H3 not statistically significant).

These findings advance the prior ecological study in three ways. First, they confirm that the regional association between Christian share and RWP vote share reflects a genuine individual-level relationship, not merely an ecological artifact. Second, they demonstrate that both identity-based and behavioral dimensions of religiosity independently shape RWP voting—addressing a key gap in the previous paper. Third, they partition variance across individual, regional, and country levels, revealing that compositional effects dominate over contextual effects.

Future research should incorporate longitudinal data to examine causal pathways, expand the country sample to improve Level-3 variance estimation, test additional cross-level interactions (e.g., regional religious context, economic conditions, urbanization), and explore potential mediating pathways (e.g., immigration attitudes, cultural conservatism) that may explain the observed religion–voting link. The inclusion of mediator variables would help disentangle whether religion influences RWP voting through value alignment, identity signaling, or social network effects—the three mechanisms outlined in the theoretical framework.

References

  1. Arzheimer, K., & Carter, E. (2009). Christian religiosity and voting for West European radical right parties. West European Politics, 32(5), 985–1011.

  2. Bates, D., Maechler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1), 1–48.

  3. Enders, C. K. (2010). Applied Missing Data Analysis. Guilford Press.

  4. European Social Survey. (2024). ESS Round 11: European Social Survey Round 11 Data. European Social Survey ERIC.

  5. Gidron, N., & Hall, P. A. (2017). The politics of social status: Economic and cultural roots of the populist right. The British Journal of Sociology, 68, S57–S84.

  6. Inglehart, R., & Norris, P. (2016). Trump, Brexit, and the rise of populism: Economic have-nots and cultural backlash. Harvard Kennedy School Working Paper.

  7. Laverty, A. A., & Hopkinson, N. S. (2025). What is the relationship between population health and voting patterns: An ecological study in England. BMJ Open Respiratory Research, 12, e003526.

  8. Maas, C. J. M., & Hox, J. J. (2005). Sufficient sample sizes for multilevel modeling. Methodology, 1(3), 86–92.

  9. Mudde, C. (2007). Populist Radical Right Parties in Europe. Cambridge University Press.

  10. Norris, P., & Inglehart, R. (2011). Sacred and Secular: Religion and Politics Worldwide. Cambridge University Press.

  11. Perez, S. A., & Vasilopoulou, S. (2023). Religion and nationalism in Central and Eastern Europe. East European Politics, 39(2), 145–162.

  12. Powell, M. J. D. (2009). The BOBYQA algorithm for bound-constrained optimization without derivatives. Technical Report, University of Cambridge.

  13. Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical Linear Models: Applications and Data Analysis Methods (2nd ed.). Sage.

  14. Robinson, W. S. (1950). Ecological correlations and the behavior of individuals. American Sociological Review, 15(3), 351–357.