Social Mobility and Perceived Discrimination: Adding an Intergenerational Perspective*

This article adds an intergenerational perspective to the study of perceived ethnic discrimination. It proposes the conjecture that perceived discrimination tends to increase with parental education, particularly among those children of immigrants who have attained only mediocre levels of education themselves. I discuss that this conjecture may be developed as an argument that comes in two versions: a narrow version about explicit downward (intergenerational) mobility and a wide version about unfulfilled mobility aspirations more generally. Analyses based on the six-country comparative EURISLAM survey support the argument: parental education positively predicts perceived discrimination in general, but among the less educated this relation is most pronounced whereas it is absent among those with tertiary education. A replication and falsification test based on the German IAB-SOEP Migration Sample reconfirms the main finding and provides further original pieces of evidence. The analyses suggest processes associated with unfulfilled mobility aspirations as the more plausible underlying reason.


Introduction
Ethnic inequalities remain stark across classical (European) immigration countries, but altogether they have declined over the last decades in legal (Koopmans, Michalowski, and Waibel 2012), political-representative (Bloemraad and Schönwälder 2013), and socio-economic (Li and Heath 2016) terms. Scholars who expect declining grievances about ethnic penalties as a result of these improved conditions, might be surprised by current research on the 'integration paradox' (Buijs, Demant, and Hamdy 2006;Verkuyten 2016). Under this concept, evidence has amounted according to which perceived discrimination is stronger, the better integrated persons of immigrant origin are. But of course, not all drivers of perceived discrimination are counterintuitive. For instance, members of minority groups that are most despised by the majority also report having made experiences of discrimination the most. The reason for the parallel existence of intuitive and counterintuitive findings is that perceptions of ethnic discrimination are a peculiar product of social exposure to actual discrimination and cognitive susceptibility to frame experiences in terms of discrimination. They are nevertheless consequential as stressors which erode mental and physical health or undermine academic test and work performance. All this renders the explanation of perceived ethnic discrimination a relevant sociological endeavor.
This article attempts to complement established scholarship by adding an intergenerational perspective, according to which social mobility matters to account for perceptions of ethnic discrimination among the children of immigrants. In particular, I propose the conjecture that perceived ethnic discrimination tends to increase with parental education, especially among those children of immigrants who have attained only mediocre levels of education themselves. I discuss that this conjecture may be developed as an argument that comes in two versions: a narrow version about explicit downward (intergenerational) mobility and a wide version about unfulfilled mobility aspirations more generally.
The empirical analyses are primarily based on the six-country comparative EURISLAM survey and confirm the general conjecture of this article. The results speak specifically in favor of the wide version of the social mobility argument, according to which unfulfilled mobility aspirations are a driver of perceived discrimination. A robustness replication based on the IAB-SOEP Migration Sample reconfirms the main finding and provides several further pieces of evidence.

Theoretical Background
Perceptions of ethnic discrimination are a menace to persons of immigrant origin, irrespective of the actual discrimination they face. As a stressor, perceived discrimination diminishes well-being (Safi 2010) along with both mental and physical health (Nandi, Luthra, and Benzeval 2016;Pascoe and Smart Richman 2009). It undermines self-confidence and thereby causes underperformance in school and at the workplace (Spencer, Logel, and Davies 2016). Verwiebe et al. (2016) unearth fear of discrimination as motivation among graduates of immigrant origin to apply for mismatching jobs. For similar reasons, immigrant minorities do not fully exploit opportunities to socialize with natives (Hewstone et al. 2015). Questioning the social significance of perceived discrimination as simply being an invalid indicator of actual discrimination overlooks these insights.
The study of how perceptions of ethnic discrimination emerge is an interesting and relevant sociological endeavor, exactly because they are highly consequential but not simply a reflection of actual discrimination . They are a combination of social exposure to actual discrimination and cognitive susceptibility to frame experiences in terms of discrimination (Doorn, Scheepers, and Dagevos 2013;Maxwell 2015). This explains intuitive findings, for instance, that persons of those origins who tend to be most rejected by natives report higher average levels of perceived discrimination (Habtegiorgis, Paradies, and Dunn 2014;Simonsen 2016). Another example is that immigrants who experience status loss (as compared to their country of origin) tend to suffer from depression over-frequently (Euteneuer and Schäfer 2018; Nicklett and Burgard 2009) and report more experiences of discrimination (Doorn et al. 2013;. Social exposure and cognitive susceptibility also account for counterintuitive findings, which currently accumulate under the concept 'integration paradox' (Buijs et al. 2006;Verkuyten 2016). Perceptions of discrimination among persons of immigrant origin tend to increase with their educational level (Silberman, Alba, and Fournier 2007;Sizemore and Milner 2004;ten Teije, Coenders, and Verkuyten 2013;de Vroome, Martinovic, and Verkuyten 2014; but see Tolsma, Lubbers, and Gijsberts 2012); Diehl and Liebau (2017) even present longitudinal evidence. They are on average higher among the domestically as compared to the foreign born (Jasinskaja-Lahti, Liebkind, and Perhoniemi 2006;Litchmore and Safdar 2015;Skrobanek 2009;Yazdiha 2018), and among the employed as compared to the unemployed (Alanya, Baysu, and Swyngedouw 2015; André and Dronkers 2017). As  argue, increasing integration among persons of immigrant origin goes along with a growing awareness about their marginalized (especially racial (Flores 2015)) status in society. This seems to come about by growing social exposure to natives at school, work, in associations and in their neighborhood (Rollock et al. 2011;Doorn et al. 2013;, as well as by increased local news media consumption that raises exposure to negative portrayals of ethnic minorities (Holtz, Dahinden, and Wagner 2013;. At the same time, investments into better integration may also increase immigrant minorities' cognitive susceptibility to frame experiences in terms of discrimination, as the 'hypothesis of rising expectations' explains (Cooney 2009;Silberman et al. 2007). Rejections from a mainstream are particularly difficult to endure, if persons of immigrant origin aspire to be a regular part of it (Runfors 2016). But even without increased expectations,  theorize that integration in terms of acculturation, education and language skills enables minorities to cognitively comprehend actual experiences of discrimination as such.
Where perceptions of discrimination stem from failure and lack of success (e.g., at school or work), they are again a peculiar product of both social exposure and cognitive susceptibility. On the one hand correspondence studies testify that actual discrimination is a major cause of non-success among persons of immigrant origin (for recent meta-analyses see Auspurg, Schneck, and Hinz 2018;Zschirnt and Ruedin 2016). Yet, lack of success might also increase one's cognitive susceptibility to frame experiences in terms of discrimination. As an ex-post rationalization of failure such framing helps to retain a sense of agency and selfesteem, as the idea of the 'self-protective properties of stigma' proposes (Crocker and Major 1989). Note that this cognitive susceptibility increasing mechanism differs from the above-discussed ones, in that it focuses on the self-esteem incentives to frame experiences in terms of discrimination among the unsuccessful, rather than on the cognitive capacities and rising expectations among the successful.

Social Mobility and Perceived Discrimination
This article sets out to investigate another implication of the idea that perceptions of ethnic discrimination are a peculiar product of social exposure and cognitive susceptibility: I propose the conjecture that perceived discrimination tends to increase with parental education, particularly among those children of immigrants who have attained only mediocre levels of education themselves. Those who attain mediocre education despite a favorable educational background are more likely to have been exposed to actual discrimination (and hence failed), and they have cause to frame their experiences in terms of discrimination as an ex-post rationalization of their lack of success. In the following, I further elaborate the argument why a lack of social mobility matters for perceived discrimination. I distinguish a narrow and a wide version of the argument, although I will ultimately only be able to provide a thorough test of the general implications of the argument that are compatible with both versions.
One can argue that the proposed conjecture narrowly captures a consequence of explicit downward (intergenerational) mobility. That is, perceived discrimination derives from a loss of (educational) standing from one generation to the next.
Indeed, the 'motive of status maintenance' (Breen and Goldthorpe 1997;Stocké 2007) proclaims that people seek to reproduce their parents' social status and regard it as a major failure in life if they are not able to do so. This implies that being at par with one's parents is the cardinal demarcation that distinguishes success from failure. This narrow version of the social mobility argument is closelyrelated to the literature on the consequences of downward international (in contrast to intergenerational) mobility; immigrants who experience status loss due to their international immigration tend to suffer from depression over-frequently (Euteneuer and Schäfer 2018;Nicklett and Burgard 2009) and report more experiences of discrimination (Doorn et al. 2013;. However, there is good reason to widen the social mobility argument and consider that the conjecture may be driven by unfulfilled mobility aspirations more generally. The wider version acknowledges that people might aspire to surpass their parents' level of education, so that failure even lurks in attaining (mediocre) levels of education that lie above the level of education of one's parents. In contrast to the narrow version of the argument, the wide version proposes no fixed demarcation at which failure turns into success, apart from suggesting that those who reached the highest levels of education, by definition, succeeded in attaining their social mobility aspiration. This wider version might particularly apply to persons of immigrant origin. According to so-called 'immigrant optimism' or 'immigrant paradox' research, the children of immigrants over-frequently aspire to surpass their parents' level of education (Ichou 2014;Luthra and Soehl 2015;Salikutluk 2016). It might be in this aspiration to surpass their parents that they fail and which then turns into perceived discrimination--an unanticipated downside of immigrant optimism.
To establish the intergenerational perspective according to which lacking social mobility matters for perceived discrimination, this article lays its main focus on hypotheses that are compatible with both versions of the argument. Initially, the social mobility argument may be used to motivate the overall expectation that on average persons of immigrant origin with better-educated parents report experiences of discrimination more frequently (H1). The reason is that according to both versions of the argument, parental education serves as a central reference point and therefore indicates mobility aspirations. Indeed, research on persons of immigrant origin's status aspirations shows how these strongly increase with parental education (Teney, Devleeshouwer, and Hanquinet 2013; albeit less so than among natives: Relikowski, Yilmaz, and Blossfeld 2012;Salikutluk 2016). 1 Higher parental education thus raises the bar and thereby increases the probability of making experiences of failure, which may be framed in terms of discrimination.
Yet, the social mobility argument also qualifies that Hypothesis 1 conceals important heterogeneity in the strength of the association between parental education and perceived discrimination that is driven by respondent's own level of education. That is, because the intergenerational perspective shifts the focus to deficient social mobility, it implies that higher levels of own education mitigate the degree to which parental education predicts perceptions of discrimination (H2).
1 This may be taken as indication in favor of the wide unfulfilled mobility aspirations version of the argument, because the smaller correlation stems from the fact that immigrant kids over-frequently try to surpass their parents' level of education (see previous paragraph).
The theoretical development and empirical test of Hypothesis 1 and Hypothesis 2 are the main aim of this article. Answering the subsequent question whether they are driven by explicit downward mobility or more general unfulfilled mobility aspirations is largely beyond the article's capacity. Nevertheless, the following sub-hypothesis attempts to provide a first step into that direction. The narrow version's emphasis of explicit downward mobility suggests that the difference between parental and own education (∆ Edu = Edu Parents − Edu Self ) is the most direct measure of the driving mechanism and thus a better predictor of perceived discrimination than the interaction implied in Hypothesis 2. But what specifically defines the narrow version is its claim according to which being at par with one's parents is the demarcation that distinguishes success from failure. It follows that only explicit downward mobility in terms of positive differences ∆ Edu (increasingly less education than one's parents) predicts perceived discrimination (H2a). It is difficult to formulate an accompanying sub-hypothesis for the wide version of the social mobility argument, because it simply makes less explicit and hence testable assumptions. I will instead emphasize findings throughout the results section that are compatible with the wide version but less so with the narrow one.

Data and Methods
The primary analyses of this study are based on the publicly-available EURISLAM data that were collected through computer assisted telephone interviews in Belgium, France, Germany, the Netherlands, Switzerland, and the UK in 2011 (Hoksbergen and Tillie 2012). The survey consists of more than 7,000 interviews with natives, as well as oversamples of persons of ex-Yugoslav, Turkish, Moroccan, and Pakistani origin, which constitute the most important Muslimorigin countries in the six destination countries entailed in the survey. Respondents of immigrant origin were sampled from the latest available electronic phonebooks via an onomastic procedure. At the beginning of the questionnaire, all respondents were screened as being of immigrant origin if they themselves or at least one of their parents was of Muslim and the required ethnic background. The EURISLAM survey allowed for bilingual interviews, so that respondents could choose between the language of the country of residence or their country of origin.
Like most contemporary telephone surveys (particularly among minority populations) EURISLAM has low response rates, with "47,8% for the ex-Yugoslav group, 41,2% for the Moroccan origin group, 30,5% for the Pakistani origin group and 55,7% for the Turkish origin group" (Hoksbergen and Tillie 2012 p. 12). Differentiated by country, the lowest response rate is 20.3% among the target population of Moroccan origin in France. These response rates are adjusted for "unknown eligibility", that is, to incorrect telephone numbers, screen out, but also to non-contact. The latter of the three unknown eligibility reasons is problematic, because it means that response rates are probably even lower. Moreover, EURISLAM does not provide any weights. The analyses presented below may therefore provide compelling tests of the raised hypotheses, but should not be considered as representative of the populations studied.
Persons originating from ex-Yugoslavia, Turkey, Morocco, and Pakistan are on average phenotypically distinguishable from central Europeans and tend to hold distinctively ethnic names. Moreover, their religion, Islam (97% of the sample identify as Muslim), 'has become the key site for demarcating boundaries between majority populations and individuals of immigrant origin across Europe' (Statham and Tillie 2016). In consequence, this analysis focuses on persons of immigrant origin who can plausibly claim to have experienced failure due to discrimination.
Analyzing EURISLAM also means that respondents originate from countries where the majority of the population holds rather conservative gender-role attitudes (Carol 2016). This could indicate gender-specific social mobility aspirations among immigrant parents and their children. Fleischmann and Kristen (2014) discuss, for instance, how reservations against female labor market participation among immigrant parents could affect the educational attainment of girls. Against this background, one could wonder about gender differences with regards to the hypotheses. Online Supplement A discusses and empirically tests this possibility in detail but finds no evidence of any gender differences.
Among the EURISLAM respondents, this paper focuses on children of immigrants who have attended school in their country of residence. This sample definition is important since the argument about perceived ethnic discrimination does not apply to those who have only attended schools abroad, which a falsification test supports (see online Supplement G). Moreover, I exclude all respondents who are currently enrolled in education, because their educational attainment visá-vis their parents' is not yet established. Additional results show that the conclusions of this study are robust to further excluding respondents who are younger than 25 (see online Supplement K). Overall this results in an analysis sample of n = 2,478 respondents, 1,135 of which are 1.5 th (i.e., born abroad but hold a general or vocational degree from the country of residence) and 1,343 of which are second (i.e., born domestically or immigrated before the age of six, and have at least one parent who was born abroad) generation persons of immigrant origin.
Note that this sample definition does not restrict age and hence includes respondents for whom (discriminatory) experiences at school are long past. Online Supplement A therefore tests whether the results are affected by age or time since schooling, but finds no indication of that.

Perceived Discrimination
The dependent variable of this analysis is perceived discrimination, for which the EURISLAM survey entails a simple yes/no question: This binary variable closely mirrors how perceived personal discrimination is generally measured in surveys. Its generic nature is insofar unfortunate, as the proposed conjecture (i.e., mediocre educational attainment despite favorable parental background) particularly pertains to discrimination experienced at school.
Moreover, it potentially also captures other forms of discrimination, such as age or gender discrimination, due to its generic nature. The IAB-SOEP robustness and falsification test reported later overcomes both limitations.
The EURISLAM survey also considered two items about perceived group discrimination. But these questions were only asked in four of the six countries.
Moreover, in terms of survey methodology, the questions on group discrimination are more generic and vaguer than the question on perceived personal discrimination.
Respondents are asked what they think others who are alike experience; it is guesswork rather than recollection of a significant personal experience. It therefore comes at little surprise that additional analyses show that many of the estimated associations are weaker or even insignificant for perceived group as compared to personal discrimination (see online Supplement B). For these reasons, this study focuses on perceived personal discrimination as dependent variable.

Predictor and Moderator Variables
The EURISLAM data entail generated variables for maternal and paternal edu-  Table H.1 reports percentages for the coarsened indicator version of ∆ Edu , which is based on parental and respondents' education variables that distinguishes between ISCED levels 0 and 1. In Table 1, these two ISCED levels are both contained in the "Low" category.
These claims therefore need to be tested in multivariate models that further take important country, origin and other group differences into account. I thus turn to the remaining aspects of the methodological design now.

Control Variables
The analyses adjust for a rich set of control variables. The base set of control variables are: gender, country of origin, and country of residence. Further control variables follow existing studies on the integration paradox: immigrant generation, age in years (as it is strongly predictive of immigrant generation), employment status, local media consumption 4 , language skills 5 , and the share of natives 4 'When you read newspapers or watch television, in which language is that?': 'Always in [language of CR]', 'Mostly in [language of CR]', 'About equally often in [language of CR] and my parents' mother tongue', 'Mostly my parents' mother tongue', and 'Always in my parents ' mother tongue'. in the neighborhood 6 . Moreover, I go beyond existing studies and also include 'unconcealed' religiosity 7 . Apart from the base set, these controls may be re-

Methods
I use logistic regression to regress perceived discrimination on respondents' own education, parental education, their interaction, and two alternative measures of their difference (∆ Edu ). Because ∆ Edu is a linear function of respondents' own and parental education, considering all three variables in a common regression model leads to collinearity problems. Yet, it is important to adjust for respondents' own and parental education, because the downwardly mobile are, on average, less educated and have parents who hold at least a primary school degree. Since the integration paradox has shown over and over again how education is associated with higher levels of perceived discrimination and since Hypothesis 1 states the same for parental education, these effects could bias any estimate of the ∆ Edu effect. I therefore use logistic diagonal reference models  to control for the categorical versions of parental and respondents' own education.
Online Supplement I reports results of common logistic regression models that do not control for respondents' own and parental education; the results are similar in conclusion.

Results
The established literature states that better-integrated persons of immigrant origin feel more discriminated than the less integrated. As a first step, I therefore discuss in how far the current analysis of the six-country comparative EURISLAM survey corroborates core findings of that literature (but see the IAB-SOEP robustness test below). Indeed, Model 1 of Table 2 shows that perceived discrimination significantly increases with respondents' level of education. Another indicator are differences in perceived discrimination of second generation as compared to 1.5th generation persons of immigrant origin. Second generation persons of immigrant origin, who spent their whole life in the country of residence, are arguably on average better integrated. As Model 1 of Table 2 shows, they also report to have experienced discrimination more frequently.
The integration paradox literature further suggests that media consumption and language skills partly drive these results because they enable persons of immigrant origin to comprehend the extent of negative portrayals against them.
A second line of reasoning found in this literature considers exposure to natives at work, in the neighborhood and so on as a driver of increased perceptions of discrimination among the better integrated. But based on the EURISLAM data I find no evidence for these claims; only the neighborhood share of natives is associated with increased reports of discrimination (Model 1 of Table 2). A mediation analysis confirms that education and second-generation status are not indirectly associated with perceived discrimination via these variables (see online Supplement J). Note however, that earlier studies also report rather mixed evidence for these predictors, in part because they also entail positive contact experiences (Doorn et al. 2013;ten Teije et al. 2013). Neither do those who are openly (i.e., unconcealed) religious by abstaining from drinking alcohol, following dietary rules, and refraining from certain activities during religious holidays report experiences of discrimination more frequently. 0.066 0.069 0.070 Note: Logistic regression results; n = 2,478; + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: EURISLAM survey, author's own estimates.
Turning to the intergenerational conjecture proposed in this study and the question whether it identifies a systematic pattern beyond what has already been established in the literature, Model 2 of Table 2 adds parental education to the model. In line with Hypothesis 1, respondents with better-educated parents report to have experienced discrimination significantly more often. In fact, parental education is nearly as strong a predictor of perceived discrimination as respondents' own education (all continuous variables are z-standardized), which the linear probability models corroborate (see online Supplement D).
The reasoning underlying Hypothesis 1 is that for both the downward mobility and the unfulfilled mobility aspirations versions of the argument, parental education is a central reference point that sets the bar for personal mobility aspirations; it thereby increases the probability of making experiences of failure, which may be framed in terms of discrimination. Unfortunately, the EURISLAM data contain no direct assessment of aspirations that would allow me to validate this assumption. To at least partly plausibilize it, I make use of a variable that captures respondents' desire to make their parents proud. If the assumption holds that parental education, on average, raises the bar and thereby increases the probability to perceive experiences of discrimination, it should hold specifically among those who strive to make their parents proud. Vice versa, for those who do not strive to make their parents proud, parental education is, on average, most likely not an important reference. In line with this idea, Model 3 of Table 2 shows a marginally significant (p = 0.0840, two-tailed) positive interaction between parental education and the desire to make one's parents proud.
The left-hand panel of Figure 1 visualizes the pattern implied in these regression coefficients. We see that the more respondents agree that making one's parents proud is an important life goal (x-axis), the stronger does the association between perceived discrimination and parental education (y-axis) become. Thus, parental education is indeed a positive predictor of perceived discrimination among those who regard making one's parents proud an important life goal, but not among those who disagree. The right-hand panel of Figure 1 shows that these results are not artificially driven by the linear specification of the predictors. Additional analyses based on the IAB-SOEP (see below) provide, by and large, a comparable result based on a correlate of the intention to make one's parents proud. Parental education is a positive predictor of perceived discrimination among those whose parents closely monitored their performance at school (see online Supplement F). Again, if the assumption that parental education serves as an important point of reference is valid, it should hold especially for those who have parents that stressed school performance. These robustness tests are certainly imperfect, but their results are nevertheless hard to square without recognizing a systematic link between parental education and mobility aspirations.  Having thus established parental education as predictor of perceived discrimination, we may turn to Hypothesis 2: do higher levels of own education mitigate this relationship? This idea is investigated in Table 3. Model 1 introduces the respective interaction. As expected, it is sizable (in relation to its main effects as reported in Table 2) and significantly negative. Having already stated that the unconditional standardized coefficient of parental education is one of the largest standardized coefficients of the whole model (see particularly the first linear probability of Table D.1), it is notable that the interaction term implies that a standard deviation change in respondents' own education can either nearly diminish or double the importance of parental education in strength (Model 2 of Table D.1).
This suggests that the intergenerational perspective is indeed an important complement to established predictors of perceived discrimination. Model 1 and Model 5 are logistic regression models. Model 2 to Model 4 are logistic diagonal reference models. Results are controlled for: immigrant generation, employment status, media consumption, language skills, share of native residents in the neighborhood, unconcealed religiosity, country of residence and origin differences, gender, and age. n = 2,478; + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: EURISLAM survey, author's own estimates.
The left-hand panel of Figure 2 visualizes the pattern implied in these coefficients (after all, the three consecutive terms of an interaction are hard to puzzle out on their own, particularly the conditional main effects). It shows that the strength of the association between parental education and perceived discrimination (y-axis) is strongest among the least educated, for whom a standard deviation increase in parental education goes along with an 8.3% (ISCED level 0) to 6.8% (ISCED level 1) increase in the probability to report experiences of discrimination. This association declines as respondents' own education increases (xaxis). But it is only among those with tertiary education that parental education does not matter for perceived discrimination anymore. That is, only among those who reached the highest absolute levels of education does the marginal effect turn insignificant and close to zero. This finding somewhat speaks in favor of the wider version of the intergenerational mobility argument, because many respondents at lower levels of education have been de facto upwardly mobile. At each ISCED (sub-)level 2 or 3, between 38 and 72% of the respondents are better educated than their parents.  One might question whether these results are artificially driven by the linear specification of the model. The right-hand panel of Figure 2 shows that the conclusion also holds if parental and respondents' own education are coarsened to categorical variables: among the low (and in tendency also intermediately) educated, it is those with highly in contrast to less educated parents who feel particularly discriminated.
Another question of concern is whether these results conceal systematic differences between populations. With regard to gender, for instance, one could wonder, whether the reported moderation works similarly for men and women, given that parents might have gendered expectations about intergenerational social mobility (i.e., which level of education their children should achieve). Another important question is whether the results generalize across all of the six countries studied. Online supplement A discusses and tests overall ten plausible dimensions by which one could expect these results to differ. But neither the EURISLAM nor the IAB-SOEP (see below) analyses provide much evidence that this central result differs by gender, country of residence, immigrant generation, age or time since schooling, ethnic identification, religion, desire to make one's parents proud or parental monitoring.
respondents are as educated as their parents). In contradiction to sub-Hypothesis 2a, introducing the spline specification does not change the conclusion. Model 4 tests a simple categorical measure, which truncates the depth of the difference, as an alternative specification. But it also does not support sub-Hypothesis 2a.
The final test is provided by (certainly over-specified) Model 5 of Table 3, which puts the interaction and the difference (∆ Edu ) against each other. It remains to be the interaction between parental and own education that emerges as statistically significant at the marginal level (p = 0.0902, two-tailed). Additional analyses that replicate Models 2 to 5 of  The predictor variable is the educational status of the better-educated parent, captured in a variable that differentiates between 'no degree', 'secondary education', and 'tertiary education'. What is missing is the fourth category of 'primary education'. As moderator, I use respondents' own education indicated by their ISCED 97 score. Compared to EURISLAM respondents' own education is reported on a less detailed ISCED scale (subcategories A, B, and C are missing). It encompasses ISCED code 6, which identifies respondents with a tertiary master degree, but lacks ISCED code 0, which identifies persons without primary education. To calculate ∆ Edu , I mapped the ISCED scale to the scale for parental education, just as I did for the EURISLAM analysis, and subtracted the former from the latter. Moreover, the robustness replication and falsification test adjusts for the same set of control variables, as the primary analysis of the EURISLAM survey with three exceptions. First, I cannot operationalize unconcealed religiosity. But the variable would also not make much sense in the very internationally composed IAB-SOEP. Second, instead of the native neighborhood share I control for the (former) classroom share of natives, which is arguable an even better control given the focus of the falsification test. Third, the IAB-SOEP did not capture local media consumption. Apart from these three deviations, the robustness test similarly considers age in years, gender, immigrant generation, country of origin categories (where case numbers allowed, I used the country of origin. But frequently enough I had to coarsen to wider categories such as persons of African origin), employment status, and language skills. The latter consist of a scale of the competencies to write, listen to and speak German. Each competence was captured on the scale: 'not at all', 'rather badly', 'fair', 'good', and 'very good'. 0283 Note: Linear full information maximum likelihood (FIML) estimates with robust standard errors. Results are controlled for: age in years, gender, immigrant generation, country of origin categories, employment status, language skills, the (former) classroom share of natives, and parental monitoring. The full regression model for experienced discrimination at school with all controls is shown in Table F.2 in online Supplement F. t statistics in parentheses; + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: IAB-SOEP Migration sample, author's own estimates. Table 4 displays seven regression models, with one model for each type of discrimination. The table displays hardly any (even marginally) significant associations. Respondents with better-educated parents feel more harassed by the police. But this association is not moderated by their own level of education. In fact, the only model that shows a (actually highly) significant moderation of parental education by respondents' own education is the regression of reported ethnic discrimination experienced at school. Thereby the robustness falsification test lends further credibility to the outlined intergenerational conjecture: parental education predicts perceived ethnic discrimination at school particularly among those who have attained only a mediocre level of education. But the same respondents do not report more frequently that they have been discriminated on the housing market, during job search and so on.
Yet, Table 4 shows no significant main effect of parental education and even a significantly negative main effect of respondents' own education. As online Supplement F discusses in greater detail, indeed neither the rather well-established integration paradox nor Hypothesis 1 replicate based on the IAB-SOEP-although there is evidence for an interaction between parental education and parental monitoring paralleling the robustness test of Figure 1. The IAB-SOEP replication is thus only partly successful.
Because this is disquieting, we should get a better impression of the moderation that does replicate Hypothesis 2. The left-hand panel of Figure     whether the evidence presented really supports the importance of an intergenerational perspective, or whether the results simply identify a population that has actually been discriminated and hence dropped out of school too early. Like most research on the topic, I am not able to distinguish actual exposure from cognitive susceptibility, and believe that probably both play a role in the underlying processes. Moreover, actual discrimination at school comes in multiple variants, such as mobbing by classmates or stereotypical portrayals of minorities in school materials, so that any attempt to completely rule it out as possible mechanism is deemed to fail. That said, according to a recent study that surveyed experiences discrimination in detail, personal experiences of unfair grading by teachers is one of the main sources of perceived ethnic discrimination at school (Beigang et al. While the core argument of this article revolves around intergenerational social mobility, an underlying assumption is that parental education, in one way or another, serves as a reference for mobility aspirations. It seems hard to puzzle out how the reported findings could come about without mobility aspirations as driving mechanism. As such, this article also speaks to the integration paradox literature: unfortunately, high aspirations may be another factor setting the stage for increased social exposure and cognitive susceptibility to experience discrimination. This may also be an unanticipated downside of immigrant optimism. Nevertheless, the lack of a direct measure of aspirations is problematic against the fact that parental education is a less decisive indicator of educational aspirations among children of immigrants as compared to natives. Future research hopefully fills this gap. Lindemann (2017), for instance, shows that persons of immigrant origin who transition from school to a vocational training that does not meet their desired option report to have experienced discrimination more often.
The results presented here would suggest that it is particularly children with better-educated parents who make this experience.
This directly leads me to the second shortfall of this study, which however haunts the vast majority of studies on perceived discrimination: the inability to distinguish social exposure to actual discrimination from cognitive susceptibility to frame experiences in terms of discrimination. It is this shortfall that fuels the highly controversial debate among pundits, policy makers and academics whether claims about xenophobia and discrimination are overblown, or to the contrary just the tip of the iceberg. In a striking parallel to Marxists' claims about 'false consciousness', some argue that the less integrated misconstrue the true extend of discrimination against them. By contrast, others suggest that the better integrated may be overly susceptible. British sociologists invented field-experimental correspondence studies in the 1960s to overcome surveys of perceived discrimination as best indicator of actual discrimination (Daniel 1968). Today field experiments regularly provide convincing testaments to the extent of discrimination. Hopefully the discipline has the capacity to solve the puzzle of the association of perceived and actual discrimination too.
On a final note, I would like to add that the social mobility argument laid out in this article is theoretically not limited to the children of immigrants. First of all, other groups are also exposed to discrimination and thereby hindered in realizing their aspirations. But more importantly, the tendency to attribute personal failures to external causes is all too human and certainly also characterizes natives--for instance when they scapegoat persons of immigrant origin.  (7):1115-34.

Supplement A: (Interaction-)Effect Heterogeneity
One might wonder whether the results reported in the manuscript conceal systematic differences between populations with respect to the main finding (i.e., how the relation between perceived discrimination and parental education is moderated by respondents' own education).
Immigrant generation might matter, for instance, because the mechanism might work differently for those children of immigrants who were born and partly socialized abroad, as compared to second generation persons of immigrant origin, who spent their whole life in the country of residence. With regard to gender one could wonder, whether the reported moderation works similarly for men and women, given that parents might have gendered expectations about intergenerational mobility. Time since schooling (or age taken as a proxy) could be important because the failure of not reaching one's (parents') educational aspirations is more salient among those who have just finished school. Ethnic identification could matter by making the self-protective properties of stigma a cognitively more available option -an argument which even more applies to differences between Muslim and non-Muslim persons of immigrant origin.
With respect to the difference between EURISLAM and IAB-SOEP one could also wonder whether the IAB-SOEP results would differ if we only focused on persons originating from the countries entailed in EURISLAM, that is, Morocco, Turkey, Pakistan and Ex-Yugoslavia. One could also wonder whether striving to make one's parents proud or parental monitoring should not alter the strength of failing to reach one's mobility goals. Finally, there might not be a theoretical reason to believe in country differences, but it is nevertheless important to know whether the main finding actually generalizes across all six countries of residence entailed in the EURISLAM survey. This supplement discusses whether there is empirical evidence for socalled 'effect heterogeneity' with respect to immigrant generation, time since schooling (or age), ethnic identification, country of residence, and religious denomination. Because gender is a cardinal dimension, there is a separate discussion of potential gender differences at the end of this supplement.
Testing for interaction-effect heterogeneity warrants estimating three-way interactions, that is, testing whether the interaction reported in the main document is significantly moderated by gender, ethnic identification, and so on. Note however, that three-way interactions are extremely demanding in terms of statistical power, because all three constitutive coefficients of the original interaction (the main effect of parental education, the main effect of respondents' own education, and their joint interaction term) should be interacted with the additional moderator (e.g., immigrant generation) (e.g., Brambor, Clark, and Golder 2006;Braumoeller 2004).
This means four additional coefficients are being estimated: one further conditional main effect and three further joint interaction effects, among which one is the joint effect of three variables.
Where country of origin is considered as additional moderator, this logic results in even more additional terms (see below). Interpreting the overall seven constitutive coefficients of a three-way interaction is a highly complex task. Here however, we have no substantial interest in the overall pattern of the three-way interaction, but simply wish to test whether the moderation reported in the main manuscript varies across populations. To answer this specific question of what we might call 'interaction-effect heterogeneity', only one of these four additional terms is by and large decisive: the joint three-way effect, that is, the interaction of the original interaction term with the additional moderator. This term tests whether the moderation reported in the main manuscript is systematically larger or smaller for certain populations that are defined by the additional main effect. I will thus focus on this term.  Table 3 in the main article. They only differ in that they introduce interaction effects by immigrant generation, age (taken as proxy of time since schooling), ethnic identification, and striving to make one's parents proud. Model 1 in Table 3 in the main article reports an interaction effect of != -0.109 and t=-2.098. The third row of Table A.1 reports the same interaction, but now it is conditional on the "Main effect", meaning it is conditional on the additional moderator being zero. Zero either indicates the reference group (e.g., 1.5th generation persons of immigrant origin), or the mean of the main effect if the additional moderator is continuous. It is already pretty telling that the effect sizes of these conditional interaction effects are pretty similar to the unconditional one reported in the main manuscript with !s of -0.135, -0.121, -0.111, and -0.095.
The important question with respect to interaction-effect heterogeneity is: do the three-way interaction terms ("*parental education*respondent's education") suggest that these conditional interaction terms look different at other values of the main effect of the additional moderator? The answer is clearly: no. With absolute t-ratios smaller than one, Table A.1 displays not even tentative evidence that the interaction of interest looks systematically (i.e., significantly) different among 2 nd generation rather than 1.5 th generation persons of immigrant origin, older persons for whom schooling is further away in the past, persons who identify less or stronger with co-ethnics, and finally those who strive to make their parents proud and those who do not. generation, country of origin and of residence categories, employment status, language skills, frequency of media consumption, the neighborhood share of natives, and unconcealed religiosity. + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001 Source: EURISLAM survey, author's own estimates. 0.075 Note: Logistic regression results; n = 2,478. Results are controlled for: age, gender, immigrant generation, country of origin and of residence categories, employment status, language skills, frequency of media consumption, the neighborhood share of natives, and unconcealed religiosity. + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001 Source: EURISLAM survey, author's own estimates.
Another concern is whether the results are generalizable across all six countries involved in the EURISLAM survey. Testing this idea similarly warrants a three-way interaction, although one that entails five additional main effects for the country differences and therefore 5*3=15 additional interaction effects. Again, the additional terms of interest are the thee-way interaction terms. Note however, that I use weighted effects rather than classic dummy coding (for details see: Grotenhuis et al. 2017). This has the advantage that I do not estimate whether the moderation of interest is significantly different in some country as compared to the reference country. Instead, the conditional interaction of respondents' and parental education is estimated at the sample average and thus equals (apart from some rounding error) the result reported in the main manuscript. The advantage of this procedure is that the additional threeway interaction effects now test whether the results in the specific countries (i.e., UK, France etc.) significantly differ from the overall result reported in the main manuscript. As Grotenhuis et al. (2017) explain, this is in fact what motivated the development of (weighted) effects coding. Table A.2 shows the results. At the sample average, the interaction of respondents and parental education is estimated as != -0.102 and t=-1.655. This result equals the one from the main manuscript, apart from the fact that the somewhat lower t-ratio shows how the 15 additional interaction terms come at the cost of statistical precision. With respect to the question of interest, the moderation is, statistically speaking (although the t-ratios for Belgium and the Netherlands are larger than one), just as negative in the five countries and (as additional analyses show) in Germany as it is for the overall pooled sample. By and large, the main finding of the manuscript may thus be regarded as to generally hold across the six countries entailed in the EURISLAM survey. Note: Linear full information maximum likelihood (FIML) estimates with robust standard errors; n = 1,181. Results are controlled for: age, gender, immigrant generation, country of origin differences, employment status, language skills, and the (former) classroom share of natives. t statistics in parentheses, + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001 Source: IAB-SOEP Migration sample, author's own estimates.
Thus far, there is hardly any evidence for interaction-effect heterogeneity. We now need to take a look at the IAB-SOEP replication. Table A.3 displays the evidence for interaction-effect heterogeneity in the IAB-SOEP analysis. It reports results that are based on Model "School" of Table 4 in the main article. The main article reports an interaction effect of != -0.0858 and t=-2.908. The third row of Table A.3 reports the same interaction, but now it is conditional on the "Main effect", meaning the additional moderator is zero. Zero either indicates the reference group (e.g., non-Muslims) or the mean of the main effect if the additional moderator is zero.
As before, the effect sizes of these conditional interaction effects are pretty similar to the un- Overall, neither the EURISLAM nor the IAB-SOEP analysis provide much evidence that the central result of the main manuscript (i.e., the association between parental education and perceived discrimination is moderated by respondents' own education) differs across populations. The only deviation from this overall conclusion is that the moderation does not seem to exist in Belgium.
Yet, to finally settle the question, we need to investigate potential differences by gender.
Gender receives especially careful consideration, because the EURISLAM survey focuses on Muslim persons originating from countries where the majority of the population holds rather conservative gender-role attitudes. One could therefore expect pronounced gender differences in mobility aspirations. Parents might, for instance, hold lower mobility aspirations for their daughters as compared to their sons (Fleischmann and Kristen 2014) because of pronounced reservations against female labor market participation (Khoudja and Fleischmann 2015). This would suggest that the patterns shown in the main article should be less pronounced among women. On the other hand, if women are not expected to work because of traditional gender roles, they might as well continue schooling (Fleischmann and Kristen 2014). Research has accordingly shown that girls of Turkish origin tend to be more aspiring than boys (Dollmann 2017) and that, by and large, women of immigrant origin (including those originating from Muslim countries) tend to perform better at school and have higher educational attainment as compared to men (Fleischmann and Kristen 2014). This rather suggests that the reported relations could even be stronger for women. Note however, that women in general tend to feel less discriminated (see descriptive statistics in online Supplement H and  There is no indication of any systematic differences by gender.
One reason for that might be the measurement of parental education as the education of the better educated parent. Maybe women base their aspirations on their mother's education while men focus on their father's. Column "Model 1 Table 3 gender-specific parental education" shows, however, that we come to the same conclusion based on this alternative measure of parental education: no indication of systematic gender differences. Finally, Table A.5 provides the same conclusions based on the IAB-SOEP sample. Overall, these twelve additional tests produce not a single indication of gender differences for any of the results reported in the main article. This is irrespective of whether we analyze EURISLAM or the IAB-SOEP, or whether we use a gender-specific measure of parental education.  Table 3 gender-specific Model 2 Table 2 Model 3 Table 2  Model 1 Table 3 parental education Model 2 Table 3 Model 4 Table 3  are logistic regression models. Model 5 and Model 6 are logistic diagonal reference models. Results are controlled for: age, gender, immigrant generation, country of origin and of residence categories, employment status, language skills, frequency of media consumption, the neighborhood share of natives, and unconcealed religiosity. n = 2,478; + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: EURISLAM survey, author's own estimates. 104 Note: Model 1 to Model 3 are linear full information maximum likelihood (FIML) estimates with robust standard errors with n = 1,181. Model 4 and Model 5 are linear diagonal reference models with n = 512. Results are controlled for: age, gender, immigrant generation, country of origin differences, employment status, language skills, and the (former) classroom share of natives. t statistics in parentheses, + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001 Source: IAB-SOEP Migration sample, author's own estimates.

Supplement B: Perceived Group Discrimination
The EURISLAM survey also entails questions on perceived group rather than personal discrimination (with answer categories 'never', 'rarely', 'occasionally', and 'frequently'). These two items correlate by r = 0.603: In general, how often do you think the following groups experience hostility or unfair treatment because of their ethnic origin or religious faith in [CR]?

Muslims living in [CR]
In light of Quillan's (2006) discussion of how to measure discrimination accurately, one might wonder, whether personal or group discrimination is the comparatively more accurate indicator of discrimination experiences. In this supplement, I first discuss this question first from a survey item perspective and then summarize how the question has been discussed among psychologists. Overall, my discussion concludes that whether personal or group discrimination is a more accurate measure remains unclear. Yet, it is rather evident that personal and group discrimination measure distinct aspects and should therefore not be compared against each other simply in terms of measurement accuracy. Only accounts of personal discrimination are systematically linked to negative outcomes (such as poor health and well-being) and they are more closely related to well-established predictors of the integration paradox. For this reason, perceived personal discrimination is the appropriate dependent variable for this study.
Starting from the angle of survey item quality, the questions on perceived group discrimination are vaguer than the question on personal experiences of unfair treatment; they ask respondents what they think that others who are alike tend to experience. By contrast, the dependent variable analyzed in the main article asks respondents about their own experiences.
So, whereas the questions on group discrimination involve recollection of events and guesswork, the question on personal experiences only involves recollection of personal experiences.
The results of the robustness analysis of the IAB-SOEP also underline that questions about specific personal experiences make a difference; the falsification test shows that discrimination experienced at work, in everyday life, and so on cannot be predicted by the outlined explanation.
To support this argument against analyzing group discrimination, Table B.1 reports OLS regression results of a model that mirrors Model 1 displayed in Table 3 of the main article, but takes perceived group discrimination as dependent variable. The key findings reported in the main article do not emerge; only the direction of the interaction term is similar. In defense of my claims one should note, however, that the relations between perceived group discrimination and nearly all predictors are less precisely estimated as a comparison of the t-ratios reported in Model 2 of Table 2 versus Table B.1 indicates. This holds for the ethnic group and country differences, age, gender, and the generation differences. But it also holds for the integration paradox predictors: the conditional main effect of respondents' own education, the neighborhood share of natives, and being openly (i.e., unconcealed) religious. In conclusion, the results displayed in Table B.1 may somewhat question the findings reported in the main article. But they mostly question the reliability of perceived group discrimination as dependent variable in studies on the integration paradox and related topics. 0.073 Note: OLS regression results; n = 2,478; + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001 Source: EURISLAM survey, author's own estimates.
Under the term "personal-group discrimination discrepancy", there is also a well-established debate among Psychologists about the difference between personal and group discrimination.
This debate was sparked when scholars noted that personal accounts of experienced discrimination are systematically lower than accounts of group discrimination (e.g., Taylor et al. 1990).
The question arose, whether one was more accurate than the other. A prominent perspective holds that personal accounts are less accurate. One reason put forward is that people are seen as discounting personal experiences of discrimination to uphold their self-esteem and an illusion of control (Ruggiero and Taylor 1995). In addition, some argue that single personal experiences are complex events that offer many alternative explanations, whereas societal discrimination against minority groups is more obvious and easy to perceive (Crosby et al. 1986).
However, others note that perceived group discrimination is a function of auto-stereotypes that are generally held in the in-group community (Taylor, Ruggiero, and Louis 1996), or that the higher figures for group discrimination may stem from a cognitive process in which persons add experiences of others to their personal experiences of discrimination (Taylor et al. 1990).
In summary, psychology has established cognitive mechanisms (and evidence for their actual working) according to which personal experiences of discrimination are potentially discounted or minimized, while accounts of group discrimination are potentially inflated. The existence and investigation of the personal-group discrimination discrepancy is therefore inconclusive on whether one is a more accurate than the other.
What seems more important than measurement accuracy is that psychologists come to emphasize how personal and group discrimination are two separate concepts with distinct consequences for minorities. Most importantly, they have noted that whereas perceived personal discrimination is associated with negative outcomes for minorities, the opposite holds for perceived group discrimination. That is, studies report positive associations between perceived group discrimination and higher well-being or better health, but negative ones for perceived personal discrimination (Bourguignon et al. 2006;Hagiwara, Alderson, and Mezuk 2016). The reason seems to be that perceived personal discrimination is a direct stressor, whereas perceived group discrimination triggers cognitive processes of relative advantage (as compared to other in-group members). In summary, it is unclear whether personal or group discrimination is the more accurate measure. But in this analysis, perceived personal discrimination responds closer to all well-established predictors, and as a detrimental outcome it is simply more relevant.

Supplement C: Results Adjusted for Base Set of Controls Only
Controls are problematic, if they condition on those mechanisms that we believe to bring about the effect of interest (Morgan and Winship 2007;Pearl 2009). The central result reported in the main manuscript is adjusted for such variables and thus subject to this critique. To show that the main result also holds, if the analysis is not controlled for mediating variables, Table C.1 presents results that are only adjusted for the base set of control variables. Given the crossnational design of the EURISLAM survey, it remains important to control for country of residence and origin differences. Apart from these, only gender is also considered. Table C.1 reconfirms the core result of the main article. 0.061 Note: Logistic regression results; n = 2,478; + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001 Source: EURISLAM survey, author's own estimates.

Supplement D: Linear Probability Models
Non-linear models may be biased with respect to the comparison of groups, interpretation of interaction effects, and comparisons of coefficients across models (e.g., Breen, Karlson, and Holm 2018;Mood 2010). The main article hence visualizes marginal effects. Another remedy against these problems is to estimate linear probability models with robust standard errors. Table C.2 reports results of such models for the same specification of Model 2 of Table 2 and Model 1 and Model 3 of Table 3 in the main article. We see that the reported relations and moderations remain and that parental education is a quite stronger predictor of perceived discrimination than respondents' own education.  Table 2 Model 1 of Table 3 Model 5 of Results are based on linear probability models with robust standard errors; n = 2,478; + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001 Source: EURISLAM survey, author's own estimates.

Supplement E: Categorical Predictors
Another objection to the main model specification reported in the article is the assumption that parental education and respondents' own education are linear predictors of perceived discrimination. The main article visualizes results of models that do not make this assumption (the right-hand panels of Figure 1 and Figure 2). The visualization is based on results of regression models that are displayed in Table E.1. 0.073 0.073 Note: Logistic regression results; n = 2,478. Results are controlled for: age, gender, immigrant generation, country of origin and of residence categories, employment status, language skills, frequency of media consumption, the neighborhood share of natives, and unconcealed religiosity. + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: EURISLAM survey, author's own estimates.

Supplement F: Additional IAB-SOEP Results
This supplement reports several additional results of the IAB-SOEP replication. Most importantly, Model 1 of Table F.2 displays the full regression model underlying Model "School" of Table 4 in the main article. Model 2 and Model 3 of Table F.2 add controls for actual school grades. Model 2 adds a factor score that summarizes the information of the math, German and first foreign language grade. The factor analysis suggests a one factor solution with all three factor scores higher than 0.6. Figure 3 of the main article is based on Model 2 of Table F.2.
Model 3 of Table F.2 shows that the results do not change, if we control for all three grades received separately, although this strategy runs into the risk of producing multicollinearity. The school grades were transformed so that higher scores imply better grades, which also holds for the factor score. Yet, there is no evidence that those with better grades report experiences of discrimination at school more often.  Table 2 of the main article. We see that in contrast to the main analysis and hypothesis 1, there is no direct average association between parental education and perceived discrimination at school. The coefficient shows the expected direction, but is statistically insignificant (p = 0.34, two-tailed). This is insofar astonishing, as Hypothesis 2 does replicate--as does, by and large, the robustness test which plausibilizes parental education as indicator of mobility aspirations (see Figure F.1). A closer inspection of Figure 3 of the main article seems to hint at a potential answer. The non-interacted coefficient of parental education can be understood as the average coefficient across the distribution of respondents' own education. In contrast to the EURISLAM analysis, a comparatively large share of respondents with ISCED level 6 (i.e., a tertiary Master degree) pull the association into (not statistically significant, but nevertheless) negative terrain. Thereby they counterbalance the influence of the less educated on the average coefficient. This contrasts with Figure   2, which is based on EURISLAM. Moreover, about 82% of these respondents are from European or other countries that would not have been considered by EURISLAM. Table F.1 therefore contains a model that excludes respondents with ISCED level 6. We see that the parental education coefficient indeed doubles in strength, but it still does not reach conventional levels of statistical significance (p = 0.16, two-tailed).
It is also noteworthy that the integration paradox does not replicate in the IAB-SOEP. Respondents' own education is either negatively or not associated to perceived discrimination at school, and there are also no generational differences. Only the native classroom share reflects the finding about the share of natives in the neighborhood; both go along with more perceived discrimination. But those with better language skills even report less discrimination experienced at school. This might have to do with the fact that the IAB-SOEP intended to sample on recent immigrants and contains large shares of phenotypically indistinct Christian immigrants. That is, the IAB-SOEP sampled among populations for whom the awareness and cognitive susceptibility increasing mechanisms proposed by the integration paradox hardly matter, as Flores (2015) argues for instance. Nevertheless, these non-findings are disquieting.
Yet, against the fact that even the well-established integration paradox does not replicate in the IAB-SOEP, it is quite reassuring that the main finding of the article does.

Supplement G: Falsification Test Among Respondents Who Attended School Abroad
The main manuscript clarifies that the argument put forward applies only to children of immigrants who have actually attended school in their country of residence. There is no reason to assume that those who have experienced downward mobility in their country of origin before emigrating should more often feel discriminated by natives in their current country of residence.
The flipside of this is that we should, as a falsification test, not find a significant interaction of respondents' own education and parental education among immigrants who have attended school in their country of origin. Note however, that there could be positive associations of own education and parental education with perceived discrimination. After all,  argue that higher levels of education help people to comprehend the true extent of discrimination against them and increase the probability of international downward mobility (Doorn, Scheepers, and Dagevos 2013;.  0.023 Note: Logistic regression results; n = 2,428. Results are controlled for: age, gender, country of origin and of residence categories, employment status, language skills, frequency of media consumption, the neighborhood share of natives, and unconcealed religiosity. + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: EURISLAM survey, author's own estimates.

Supplement H: Descriptive Sample Statistics
This online Supplement shows descriptive sample statistics for the variables used in the analyses. Note that all tables are based on un-imputed data.

Supplement I: Explicit Downward Educational Mobility, Sensitivity Analyses
This online Supplement presents additional results on the sensitivity of the results regarding explicit downward educational mobility. The measurement of explicit downward educational mobility as the difference between parental and respondents' own education has the disadvantage that it is a direct linear function of the two involved variables. As such all three could not be used in the same model because of perfect collinearity (i.e., full rank). Yet, considering respondents' own and parental education is arguably important, because the integration paradox literature, which was reviewed in the theory section, has repeatedly shown how bettereducated persons of immigrant origin report experiences of discrimination more frequentlyand they tend to be the upwardly mobile. For this reason, the main results reported in the article control for respondents' own and parental education by using diagonal reference models .
For those who are not familiar with diagonal reference models, Table I.1 shows common logistic regression estimates that are not adjusted for respondents' own and parental education. The findings are largely similar to the results reported in Table 3 in the main article, and thus further suggest that the interaction between respondents' own and parental education is simply a more consistent predictor of perceived discrimination than any (continuous or categorical) measure of explicit downward mobility.  Table   3 of the main article. The interaction between parental education and respondents' own education again turns out as a better and more systematic predictor of perceived discrimination than their difference ∆ Edu or a categorical variable that indicates downward, upward of no educational mobility. Models 2b, 3b and 4b replicate the results reported in Table I.1, meaning these models do not control for respondents' own and parental education. The results based on the IAB-SOEP sample also provide no support for sub-hypothesis 2a.  Table 3 of Table 3 of 0.062 0.063 0.063 Note: Logistic regression results. Results are controlled for: immigrant generation, employment status, media consumption, language skills, share of native residents in the neighborhood, unconcealed religiosity, country of residence and origin differences, gender, and age. n = 2,478; + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: EURISLAM survey, author's own estimates. 133 Note: Model 2a, 3a and 4a are linear diagonal reference models with robust standard errors with n = 512. Models 2b, 3b, 4b and 5 are linear full information maximum likelihood (FIML) estimates with robust standard errors with n = 1,181. Results are controlled for: age in years, gender, immigrant generation, country of origin categories, employment status, language skills, and the (former) classroom share of natives. t statistics in parentheses, + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: IAB-SOEP Migration sample, author's own estimates.

Supplement J: Mediation Analysis
The integration paradox literature suggests that language skills, media consumption and the share of natives in the neighborhood drive the higher levels of perceived discrimination among the better educated and among the second generation. The main article claims that based on the EURISLAM data, there is little support for these ideas. However, Table 2 of the main article presents only direct "effects". Given the largely insignificant associations between perceived discrimination and language skills, media consumption and the native neighborhood share, this conclusion is probably correct. This supplement discusses the results of a mediation analysis to properly test these claims.
Mediation analysis based on non-linear models is not straight forward (Breen, Karlson, and Holm 2013), particularly if one intends to test several indirect effects simultaneously. For this reason, I estimate a linear probability path model using R's lavaan package (Roseel 2012). I use full maximum likelihood estimation, because it seems the most direct way to deal with missing values. To deal with the heteroskedasticity resulting from predicting a binary outcome, I use bootstrapped standard errors based on 10,000 resamples. In all other respects the model imitates Model 1 of Table 1 in the main article and contains the same control variables. because it is not a continuous predictor. But the results are controlled for this variable. In line with the literature on immigrant integration, we see that education and second-generation status systematically predict better language skills, media consumption, and living in a neighborhood with higher shares of natives. The one exception is that second generation respondents seem to be significantly less likely than 1.5 th generation respondents to live in neighborhoods with larger shares of natives. However, just as in Table 1's multiply imputed logit model, better language skills and media consumption do not predict perceived discrimination. Only the native neighborhood share shows a significant positive association to perceived discrimination.
Education and second-generation status remain to predict perceived discrimination directly.
So far, these results merely replicate Model 1 of Table 1 of the main article. Exploiting the capacities of path analysis, however, I can now also report indirect coefficients. The indirect "effects" of education on perceived discrimination running via language skills (! = -0.000, p = 0.871), media consumption (! = 0.002, p = 0.355), and the native neighborhood share (! = 0.001, p = 0.100) are all insignificant (although the indirect effect via the native neighborhood share just misses the threshold for marginal significance). Even combined, the three indirect effects do not reach the verge of statistical significance (! = 0.003, p = 0.272). The same holds for second generation status. There are no significant indirect "effects" running via language skills (! = -0.001, p = 0.871), media consumption (! = 0.003, p = 0.353), or the native neighborhood share (! = -0.002, p = 0.131). Unsurprisingly, the combination of these indirect paths is not significant (! = 0.000, p = 0.969). In conclusion, a proper mediation analysis supports

Supplement K: Respondents Who Are at Least 25 Years Old
One might be concerned that excluding respondents who are currently enrolled in education is not enough to ensure that all respondents have completed educational attainment. This seems important, however, because otherwise respondents could still attempt to realize their mobility aspirations. As a further robustness test, this online Supplement therefore presents results that exclude all respondents who are younger than 25. Table K.1 shows that both the EURISLAM and the IAB-SOEP results are robust to this additional sample restriction. 0.126 Note: Logistic regression and linear full information maximum likelihood (FIML) estimates with robust standard errors. Results are controlled for: (EURISLAM) immigrant generation, employment status, media consumption, language skills, share of native residents in the neighborhood, unconcealed religiosity, country of residence and origin differences, gender, and age, or (IAB-SOEP) age in years, gender, immigrant generation, country of origin categories, employment status, language skills, and the (former) classroom share of natives. + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: EURISLAM survey and IAB-SOEP Migration Sample, author's own estimates.

Supplement L: Results for ISCED Without Sub-Levels
It is not obvious whether the ISCED scale that distinguishes all sub-levels (e.g., 3A, 3B, etc.) is a better (linear) predictor than the ISCED scale that simply distinguishes between the important educational levels 0 to 5. The numeric metric is less obvious for the detailed scale, for instance. The results reported in the main article use the detailed ISCED scale, so as to use as much (potentially informative) variation as possible. This online supplement provides additional analyses showing that the main results of the article are robust to using the less detailed ISCED scale.   McFadden's Pseudo-R2 0.067 0.069 0.072 Note: Logistic regression results, controlled for: immigrant generation, employment status, media consumption, language skills, share of native residents in the neighborhood, unconcealed religiosity, country of residence and origin differences, gender, and age. n = 2,478; + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001. Source: EURISLAM survey, author's own estimates.