Allport’s (1958) theory on generalized prejudice stipulates that individuals who dislike a particular outgroup are more likely to express negative views about other groups. It is indeed a recurrent finding that prejudices towards groups seen as “other” or inferior tend to cluster as parts of a broader phenomenon (e.g., Akrami et al., 2011; Bierly 1985; Bratt 2005; Friehs et al., 2022; Meeusen & Kern, 2016). The concept of GFE, for instance, implies that derogatory attitudes towards different groups—including Jews, Muslims, people experiencing homelessness and sexual minorities—cluster together into a stable structure of interrelated prejudices that is rooted in an ideology of inequality (Heitmeyer, 2002; Zick et al., 2008; Zick et al., 2009). In many of these studies, second-order CFA has been the statistical ‘weapon of choice’ to capture GFE as the common denominator of specific prejudices.

Although research consistently confirms the existence of a generalized prejudice factor, there are good reasons to assume that the specific constellation of prejudices that constitute the GFE-syndrome is context-dependent. Which outgroups become target of derogatory attitudes is contingent on “the options a specific society offers” (Zick et al., 2008, p. 367). Depending on the specific intergroup context, ‘Others’ can be perceived as a threat to the ingroup’s sense of group position in various ways (Blumer, 1958). Integrated threat theory (Stephan et al., 1998; Stephan et al., 1999) distinguishes between realistic threats (to physical and material well-being) and symbolic threats (to cultural traditions, norms and values). Stereotypes that emerge as a reaction against these forms of threat (Fiske et al., 2002) can differ widely across outgroups (Cottrell & Neuberg, 2005; Fiske et al., 2002). Based on the nature of perceived threat, the differentiated threat approach (Meuleman et al., 2019) categorizes outgroups as deviant (high symbolic, low realistic threat), competing (high realistic, low symbolic threat), and dissident groups (high on both realistic and symbolic threat).

The essence of the differentiated threat argument is that GFE cannot be understood in purely abstract terms but should instead be socially situated in concrete intergroup contexts (Meuleman et al., 2019). Although this general premise is widely acknowledged (Choma & Hodson, 2008), few empirical studies focus on context-driven variations in the constellation of GFE. To address this hiatus, this paper compares the structure of prejudices between majority-group Belgians and Belgians of Turkish or Moroccan descent. Given their distinct social positions and intergroup experiences, this majority-minority comparison offers a good vantage point to explore group-based variation in the configuration of GFE.

For the majority-group perspective, we focus on prejudices directed at four outgroups that figure prominently in contemporary Belgian public debate: Muslims, Jews, sexual minorities, and unemployed persons. First, in Belgium, as in other Western countries, Muslims have become a primary target of Othering (Ciftci, 2012; European Union Agency for Fundamental Rights, 2024a; O’Brien, 2016; Zemni, 2011). Negative attitudes towards Muslims and Islam are rooted in the assumption that Islamic values are incompatible with European traditions. Muslims are depicted as a homogenous group that is religiously fundamentalist and therefore inherently sexist, anti-democratic and violent (Ahmed & Matthes, 2017; Ivarsflaten & Sniderman, 2022). Although Muslims are, in the eyes of the prejudiced individual, primarily a symbolic threat, fears related to terrorist attacks have added an element of realistic threat. Second, homonegativity refers to negative attitudes towards sexual minorities, such as gay, lesbian, bisexual, transgender, and queer individuals (Aerts et al., 2014; Slootmaeckers & Lievens, 2014). It is rooted in the beliefs that deviations from cis-gender, heterosexual norms undermine the traditional social-cultural order (thus construing sexual minorities as deviant). Third, antisemitism refers to hostile attitudes toward an imagined and generalized ‘collective Jew’. Even if critical attitudes towards the state of Israel are, both conceptually and empirically, distinct from antisemitism (Schmidt et al., 2011), escalating conflicts in the Middle East have put antisemitism back to the fore (European Union Agency for Fundamental Rights, 2024b; Jacobs et al., 2011). Notwithstanding this geopolitical context, in the Belgian context, perceptions of Jews are often linked to the ultra-orthodox haredim community, which is concentrated in Antwerp and strongly embedded in in its diamond industry. Common stereotypes about Jewish persons in Europe are that they refuse to assimilate culturally while simultaneously having a strong grip on the economy (Meer, 2013; Schiffer & Wagner, 2011). Finally, this analysis also includes a competing outgroup primarily perceived as a realistic threat, namely unemployed persons. Negative imagery often portrays unemployed persons as work-shy and unwilling to take responsibility for their situation (Buffel & Van de Velde, 2019; Furåker & Blomsterberg, 2003; Laenen & Meuleman, 2017; Rossetti et al., 2022).

While most research on GFE has focused on the majority perspective, exclusionary attitudes are not confined to dominant groups (Meeusen, Abts & Meuleman, 2019). Belgians of Turkish and Moroccan descent occupy the position of ‘established outsiders’ (Abts et al., 2024). According to recent figures of Statistics Belgium, in 2024 over 200,000 persons of Turkish and 500,000 persons of Moroccan descent reside in Belgium¹1

https://statbel.fgov.be/en/themes/population/structure-population/origin, last visited on 17/07/2025.

Based on the described intergroup context, we hypothesize that a GFE factor underlies the structure of group-specific prejudices for both majority and minority group Belgians. However, we anticipate that the salience and configurations within the latent structure of the GFE construct might differ between groups. Based on previous research (Meuleman et al., 2019), we expect anti-Muslim attitudes to constitute the key component of GFE among majority-group Belgians. It remains an open question whether anti-Western attitudes play a similar structuring role for GFE among Belgians of Turkish and Moroccan descent. Furthermore, we anticipate levels of antisemitism and homonegativity to be higher among the Belgians of Turkish and Moroccan descent than among majority group members.

This analysis integrates two data sets. First, the 2020 Belgian National Election Study (BNES; Meuleman et al., 2021) is used to measure the prejudice dimensions among majority group members. This survey contains 1,659 Belgian adults eligible to vote, randomly selected from the National Register (response rate: 32.8%). The main BNES survey was planned to be conducted face-to-face. However, respondents hesitant to host interviewers in their homes due to the covid19 pandemic were offered an alternative survey mode, including a self-completion online survey (19.5% of sample) or a video interview (4.4%). After the initial face-to-face survey, participants were invited to complete a paper-and-pencil follow-up drop-off questionnaire, which was completed by 1,129 respondents (68.1% of the original sample). This analysis focuses on respondents who completed the drop-off, as items on antisemitism and attitudes toward the unemployed were included in this section.

Second, the 2020 Belgian Ethnic Minorities Election Study (BEMES) is used to gain insight in prejudiced attitudes among minority group members in Belgium. Specifically, BEMES contains a sample of 374 first- and second-generation Belgians of Turkish and Moroccan descent, randomly selected from the population registers of Antwerp (self-completion survey; 70% on paper; 30% online). The reason for only focusing on Antwerp is that (1) ethnic background is not present in the national register, but registered at the municipal level, and (2) this city is the home of sizeable Turkish and Moroccan communities. A consequence of this operational choice is that BEMES cannot be considered as representative of the entire population of Belgians of Turkish and Moroccan descent.

Among both majority and ethnic minority Belgians, we measure prejudice towards four outgroups. Each dimension is measured via the respondents’ agreement with a battery of 2 to 4 statements (measured on a 5-point scale). For homonegativity, antisemitism and attitudes towards the unemployed, the question formulations are identical across BNES and BEMES. The fourth prejudice dimension included is anti-Muslim attitudes (for the majority group) and anti-Western attitudes (for Belgians of Turkish and Moroccan descent, who are mostly Muslims). Both scales are mirrored in question wording but replacing the term ‘Islam(ic)’ by ‘West(ern) or Europe(an’).

The exact question wording of all prejudice items, as well as descriptive statistics, can be found in Table 1. The reliability, validity and comparability of these scales is assessed in detail in the results section.

To study the structure of interrelated prejudices among majority and ethnic minority group members, we make use of second-order confirmatory factor analysis (for similar approaches, see: Bratt 2005; Meeusen, Meuleman, Abts & Bergh, 2019; Meuleman et al., 2019; Zick et al. 2008). These models estimate the specific dimensions of prejudice as first-order latent variables that are each measured using multiple observed items. Based on these first-order dimensions, a more general second-order latent variable that captures the common denominator of the specific prejudices is constructed (see Figure 1).

Comparing properties of latent variables across groups requires that these constructs are measured invariantly. The importance of measurement invariance has become widely acknowledged and invariance tests are now commonly applied in comparative research, often using multigroup CFA (Davidov et al., 2014; Leitgöb et al., 2023; Meuleman et al., 2023). The specific case of second-order models, however, brings along a couple of additional complexities for invariance testing (Chen et al., 2005; Rudnev et al., 2018). After all, these models contain two layers of latent variables that each have their respective measurement parameters. Depending on the specific purpose of comparison, different types and levels of measurement invariance can be distinguished.

Starting at the level of the first-order factors, a distinction can be made between configural, metric and scalar invariance (similar to conventional CFA models; Davidov et al., 2014). Configural invariance implies that the pattern of salient and non-salient loadings is equal between the groups. Configural invariance does not guarantee that the first-order factors can be compared numerically, however. A more direct numerical comparison assumes the equality of certain sets of measurement parameters. If the factor loadings are equal across groups—metric invariance—covariances and regression parameters involving the first-order factors can be compared in a valid manner across groups. Additionally constraining item intercepts to be equal—scalar invariance—is a prerequisite for comparing the means of the first-order latent variables.

A similar logic can be applied to the second-order factor (Chen et al., 2005; Rudnev et al., 2018). Comparing covariances and regression coefficients involving the second-order latent variable presupposes that the second-order factor loadings are equal across groups. Comparing second-order latent means assumes that also the intercepts of the first-order latent variables are equal across groups. However, given that the first-order factors serve as indicators based on which the second-order factor is constructed, invariance of the higher-order factor requires that invariance of the lower-order factors is established first. If first-order factor loadings differ across groups, the covariances between the first-order factors are incomparable. Under these conditions, it becomes useless to test the equality of second-order factor loadings (that are derived from these covariances between first-order factors). Similarly, if scalar invariance does not hold for first-order factors, first-order latent means lack comparability, which makes it meaningless to even test scalar invariance of the second-order factor (Rudnev et al., 2018, pp. 51–52).

These observations imply that the equality of measurement parameters needs to be tested separately for first- and second-order factors. In the results section, we will subsequently test the following series of increasingly constrained models:²2

Note that the order of tests deviates slightly from what is proposed by Chen et al. (2005) and Rudnev et al. (2018). The slight change in approach stems from the fact that this paper presents an exploratory analysis that is interested in comparing first-order factors (the prejudice dimensions) as well as the second-order factor (GFE). Before turning to constraints on the second-order measurement parameters, we fully explore the parameters of the first-order factors.

All presented models are estimated using Mplus Version 8.4. Item non-response is dealt with efficiently using the Full Information Likelihood (FIML) estimation procedure that takes all available information into account. Despite the ordered-categorical nature of the indicators, we opted for a Maximum Likelihood-based estimator instead of the robust weighted least squares estimator (WLSMV). This simplifies the cross-group comparison of measurement parameters considerably and is justified given that all indicators contain five answer options at least (DiStefano, 2002; Pokropek et al., 2019).

Table 1 displays the percentage of respondents that (strongly) agree with statements measuring homonegativity, anti-Muslim / anti-Western attitudes, antisemitism and negative attitudes towards unemployed persons. At the item-level, several notable differences between majority and minority Belgians emerge. Prejudiced attitudes are prevalent in both samples, but homonegative and antisemitic sentiments are considerably more outspoken among Belgians of Moroccan or Turkish descent compared to majority-group Belgians, especially regarding the statement ‘children should be taught that homosexuality is normal’ (22.3% support vs. 72.2% respectively). The gap in support for equal rights for homosexual persons is less outspoken, but still clear. Almost half of the minority respondents agree that Jews have too much influence or feel superior over others. Among the majority Belgians, the percentage endorsing these statements is considerably lower at 12.3 and 19.8% respectively.

For the items measuring the other two prejudice dimensions, between-group differences are smaller. Unexpectedly, majority and minority group Belgians generally display rather similar levels of prejudice towards unemployed persons. Support for anti-Muslim statements among majority group Belgians is about as strong as support for anti-Western claims among minority group Belgians. A notable exception is the item on value conflicts: while 37.8% of the majority sample see Islamic values as incompatible with European values, this is considerably weaker (20.3%) for the mirror version in the minority sample. In anti-Islam discourse, Islamic values are constructed as a foreign element that is rejected entirely (e.g., Árnadóttir & Meeusen, 2024; Moftizadeh et al., 2021). Conversely, the vast majority of European Muslims do not see a contradiction between Islamic and European values (Abu-Rayya & Sam, 2017; Hillekens et al., 2019).

Table 2 summarizes the procedure testing the invariance of the second-order models. The configural invariance model (M1) has a good fit judging by RMSEA, CFI, TLI and SRMR.³3

In the ethnic minority group, the initial model has an inadmissible estimate, namely a negative error variance for the antisemitism factor. This shows that among the minorities, GFE and antisemitism cannot be distinguished. To overcome this problem, the error variance was fixed to 0.

When also the first-order intercepts are set equal across groups (M3), non-ignorable misfit is introduced. Not only do we see a substantive increase in the chi-square, also the change in RMSEA, CFI and TLI is considerable. The modification indices show that the lion’s share of misfit stems from constraints on two specific items. First, the second item of the anti-Muslim / anti-Western construct functions differently for majority and minority group: Among the majority members, the item ‘Islamic values are incompatible with European values’ has a higher intercept and a stronger factor loading than the statement ‘European values are incompatible with Islamic values’ among ethnic minority Belgians. Among the majority group, this item is a better indicator of the underlying concept and, conditional on the latent mean, agreement with the statement is more outspoken. The section with descriptive findings already discussed that this item provoked a clearly distinct response pattern among both groups. Second, the antisemitism-item ‘Most Jews are only after money’ is found to function differently. This item has a higher intercept among the majority population. If we compare majority and ethnic minority group members who have the same level of antisemitism, agreement with the statement is thus stronger among the majority group. Apparently, the stereotype of the greedy Jew occupies a more central position in the antisemitic imagery among majorities than it does for Turkish and Moroccan minorities. For the latter, the geopolitical context is probably a more central element. Removing these two equality constraints results in partial scalar equivalence for the first-order factors (M4). Given that at least two items per first-order construct have identical loadings and intercepts, partial invariance is established (Byrne et al., 1989) and cross-group comparisons of first-order latent means are possible (Rudnev et al., 2018, p. 51).

Now partial scalar invariance is established for the first-order latent variables, we focus on the comparability of the second-order factor. Constraining second-order factor loadings (M5) leads to mixed results: While RMSEA and SRMR deteriorate somewhat, CFI remains stable and TLI even improves. There are some cross-group differences in second-order loadings, but these differences are not so large for the fit indices to pick them up easily. When first-order intercepts (M6) are additionally constrained, however, model fit deteriorates very clearly (judging by the difference in CFI and RMSEA). Thus, while it is possible to identify a second-order factor representing GFE in both groups, the specific configuration of the GFE factor differs considerably, in terms of which prejudices are the central elements of GFE (loadings) and particularly how prevalent the specific prejudices are (intercepts). As a result, direct numerical comparisons involving the GFE factor (e.g. comparing the GFE mean) cannot be made in a meaningful manner.

Based on the invariance tests, M4 as well as M5 can be suitable for further interpretation. If the purpose is to create a higher-order construct that is as comparable as possible (e.g., to formulate explanation models for GFE), the model with metric invariance of the second-order factor (M5) is preferable. If one is rather interested in mapping the differences in the configuration of GFE, it can be more insightful to interpret M4 with its differences in second-order loadings. Here, we take the latter approach (see Table 3). For the majority group, the strongest loading is found for anti-Muslim attitudes (0.78), implying that anti-Muslim sentiment overlaps largely with the general GFE factor (communality: 61%). In the eye of majority Belgians, the Muslims serve as the master template of Othering onto who general prejudices are projected. The second-order loadings for antisemitism and attitudes towards the unemployed are slightly weaker (around .60), and the weakest loading is found for homonegativity (0.46). Among Belgians of Turkish and Moroccan descent, the configuration of the GFE factor is quite different. Antisemitism stands out as the strongest component of GFE, with a standardized second-order loading equaling 1 (meaning that, empirically, antisemitism cannot be distinguished from GFE⁴4

To avoid inadmissible estimates, the error variance of the antisemitism factor was fixed to 1 for the minority group.

The intercepts of the first-order latent variables (representing the group-level average on the four prejudice dimensions⁵5

This interpretation of the latent intercepts stems from the fact that the latent mean of the second-order factor is fixed at 0 in both groups.