Does the Muslim penalty in the British labour market dissipate after accounting for so-called “sociocultural attitudes”?

ABSTRACT Using multilevel modelling, this paper investigates ethno-religious penalties in unemployment and inactivity among men and women using the Understanding Society survey. The paper confirms previous findings of a Muslim penalty and a British labour market hierarchized by colour (ethnicity) and religion (culture). However, by including a greater range of ethnic groups the paper provides a corrective to accounts in the sociological literature that being White is not a protection against the Muslim penalty. Rather, while affiliation with the Muslim White British group does not appear to be associated with penalization, Muslim Arabs who traditionally identify as White are found to experience significant disadvantage. This suggests that the Muslim penalty might also be moderated by a person’s country of origin. The paper also finds that considerable penalties remain for Muslims even after adjusting for so-called “sociocultural attitudes”, challenging the assumption that value orientations offer a suitable explanation for the Muslim penalty.


Introduction
Ethnic differences in labour market outcomes have long attracted scholarly and policy interest in Britain. They have been studied from various perspectives, including pay gaps (Longhi and Brynin 2017;Li and Heath 2020), occupational attainment (Cheung 2014), and the probability and duration of unemployment (Longhi 2020). The fact that these variations remain even after accounting for factors that are likely to impact employment (e.g. education, age, region, language proficiency, health) resulted in these opportunities. The insinuation, therefore, is that if Muslims did not hold "isolationists tastes" there would be little variance in their employment outcomes relative to majority group members (Koopmans 2016).
In terms of holding traditional gender norms, the more conservative a person, the more sympathetic they are assumed to be to the "male breadwinner model" (Lewis 2001). The corollary is that women prioritize childrearing and household work, dedicating less time to finding employment. This is posited as another explanation for Muslim women's poor labour market outcomes (Koopmans 2016; see also Khoudja and Fleischmann 2015). It is worth noting, however, that the mechanism could also operate in reverse with women who are unable to find successful employment potentially validating their labour market status retrospectively by holding more traditional views on the division of labour. Khoudja and Platt (2018) capture gender attitudes through participant views on female employment, namely whether they believe it is a husband's role to earn money, and whether they feel a mother working is detrimental to her child's wellbeing. The authors find that "gender attitudes are not related to labour market entries of Indian and Sri Lankan and Pakistani and Bangladeshi women" (Khoudja and Platt 2018, 13). Nevertheless, the authors do find that traditional gender views are associated with labour market exits, but this is found to be the case across all ethnic groups, and not only with Muslims.
In light of how Muslims are often problematized and critically discussed in the public discourse as "segregationists" because of their faith (Field 2007;Joppke 2009), it is notable that many ethnic penalty studies do not account for religiosity. Among the exceptions are Heath, Li, and Woerner-Powell (2018) who capture religiosity through how much difference religion makes to a person's life (see also Connor and Koenig 2015), and frequency of attendance at religious services as a proxy for bonding capital, with those who have high religiosity being more likely to spend time actively engaging with their faith community and therefore developing intra-community ties. However, since there is no religious obligation on Muslim women to attend the mosque, focusing solely on how important religion is to a woman's life is a more accurate measure of religiosity.
Given the claim that these key sociocultural variables "are not often taken into account in ethnic penalty studies" (Koopmans 2016, 198), but that when they are included "there are hardly any statistically significant differences left" (Koopmans 2016, 213), there is a need for novel and updated quantitative research on the Muslim penalty in Britain which also adopts a more heterogenous reading of Muslims. This is particularly needed since, in the UK, research has tended to focus on Muslims with a Pakistani, Bangladeshi and Indian background, yet the population today includes a reasonable number of Muslims with White, Black African and Arab ethnicities (MCB 2015).
Based on the established evidence of a religious (Muslim) and colour (Black) penalty at play in the British labour market (Khattab and Modood 2015), one might assume that any penalty Muslim Arabs face on account of their religion is mitigated by the fact they identify as White (Modood 2005). As such, their penalty should be close to that of White British Muslims. However, recent findings that Muslim male job applicants originating from the Middle East and Africa "experience a double burden: independent of the stigma they face for signalling their closeness to a Muslim association, they are also penalised for the geographic region they originate from" (Di Stasio et al. 2021, 13; emphasis added) suggest that our initial assumption might need to be revisited. Indeed, if we also account for the evidence that "respondents from North Africa and Sub-Saharan Africa report the highest levels of discrimination" in Europe (Fundamental Rights Agency 2017, 24), we can see that a study accounting for the plurality of ethnicities that constitute the Muslim community (i.e. distinguishing between labour markets participants who are Muslim Arabs, Muslim Black Africans and White British Muslims) that also accounts for so-called "sociocultural attitudes" is essential to better understand the potential drivers of the Muslim penalty.
Drawing on multilevel modelling, the purpose of this study is therefore to measure inequalities between hitherto included but not disaggregated Muslim ethnic groups, and, in doing so, to contribute to and incentivise research that develops explanations for the differences the data show. To do so, this paper contributes to the literature in two ways. First, it examines differentials in terms of unemployment and inactivity while also accounting for oft-excluded so-called "sociocultural variables" that have been posited as an explanation for Muslims' poor labour market outcomes. Second, by adopting a more heterogenous reading of Muslims and disarticulating between hitherto included but not disaggregated groups (namely Arabs and British Whites), the paper investigates whether there is evidence of a country-of-origin penalty among Muslim men from the Middle East and North Africa.
The specific research questions considered are: 1. Does the Muslim penalty, among men and women, dissipate once socalled "sociocultural attitudes" are accounted for? Specifically, are religiosity, traditionalist views, and lower civic participation associated with a higher risk of unemployment and inactivity? 2. Do both Muslim groups that identify as White -Arabs and White British peopleexhibit a similar risk of being unemployed and inactive relative to White British Christians? Specifically, is there evidence of a country-of-origin effect moderating the Muslim penalty among men?

Data and method
Data This paper uses information from the first ten waves of the adult panel of Understanding Society: the UK Household Longitudinal Study (UKHLS) (University Of Essex 2021). This annual survey started in 2009 and collects information over a 24-month period, primarily through face-to-face interviews, on people's socio-economic situation and events that have occurred between each wave. It offers the most detailed and highest-quality source currently available on the labour market position of British Muslims in their social, religious and cultural contexts. The survey design involves clustering and stratification, meaning it is not a simple random sample, and allows researchers to draw nationally representative statistics after adjusting for the complex survey design. It counts approximately 100,000 individuals from 40,000 households (38,000 households in wave one (2009/11) which includes 4,000 household from an ethnic minority boost sample), making it one of the largest studies of its kind. It also benefits from an Immigrant and Ethnic Minority Boost Sample added in wave six (2014/16), which provides an additional 2,500 households.

Measures
Dependent variable. By pooling data from the first ten waves, I analyse two types of labour market status. Both analyses are restricted to those of working age (16-64). First, I focus on estimating the average probability of unemployment within the active population between 2009 and 2020, distinguishing between those who are unemployed (1) and those who are employed (0). Respondents are considered to be employed if they report being in either employment (full-or part-time) or self-employment. Those on maternity leave, government training schemes, and apprenticeships are also considered employed. Individuals are categorized as unemployed if they self-report as such. All other groups are excluded from the analysis. Second, I assess the likelihood of inactivity. Those in full-time education, retired, working in a family business in an unpaid capacity, focused on "family care or home", "doing something else", or who identify as long-term sick or disabled are classified as inactive (1). Those who are employed and unemployed (as defined above) are coded as (0).
Explanatory variable. Ethno-religious groups: Data on ethnicity and religious affiliation are combined in order to capture the interaction between the two and create a set of ethno-religious categories (Khattab 2009;Khattab and Modood 2015). To identify ethnic membership responses to the question "What is your ethnic group?", which is asked once of participants when they first enter the study, is used. For religious affiliation, I use information from the question that asks, "Which religion do you regard yourself as belonging to?". When information is missing at a particular wave, I fill the gap using information from the closest prior wave. Otherwise, I use information from the closest next wave. I also use information from a question that asks "Do you regard yourself as belonging to any particular religion?" to create No Religion groups based on those who answered "no" to the question.
Based on the ethnicity and religious affiliation questions which have 18 and eight modalities each (once "Other religion" is included, Christian denominations are combined, and a No Religion group is created), there are 144 different possible combinations of ethno-religious groups that can theoretically be created. Only those groups which had at least 100 observations were assigned their own group in the regression analysis. The groups with too few observations or with missing ethno-religious identity information were combined into one "Other" group which also includes all those who specified their ethnicity to be Other/Other Mixed. While this group is not of substantive interest, their inclusion in the model guards against producing biased estimates. Table 1 lists the categories of ethno-religious groups derived.
The models are adjusted for the survey's complex design which involves clustering based on postcode sectors, and stratification within primary sampling units. Weights provided by UKHLS (Knies 2018) are also applied to adjust for over-sampling due to the survey design and non-response. This is to ensure coefficient estimates are unbiased and standard errors are efficient.
To get a clearer understanding of the association between ethno-religious grouping and risk of unemployment/inactivity, the study adjusts for other influences known in the literature to affect these two. These are discussed below and a statistical description of each of these variables, including the number of observations, is available online (Tables A1, A2, A3, and A4).
Demographic factors and human capital. Age: I control for age and include a squared age variable to capture any curvilinear effect.
Marital status: Grouped into three categories: (1) single, (2) married, in a same sex civil partnership, or cohabiting, and (3) divorced, separated (including from a civil partnership) and widowers/surviving civil partners.
Health concern: Coded as (1) those with a long-standing physical or mental impairment and (0) those without.
UK born: Coded as born in the UK (1) or not (0). All White British not born in the UK are dropped to get a more accurate picture of the impact of this variable.
Difficulty with English language: Coded as (1) if the participant affirmed having difficulty (i) speaking English in person or over the phone, (ii) reading English, (iii) filling in official forms in English. The relevant information is only collected in waves one, five, six, and ten. Information from wave one is used for the first four waves, wave five data for the fifth, wave six data are used for the subsequent four waves, and wave ten uses its own information. If information is still missing after this, I use information on whether the respondent completed the survey in English. If they didn't, participants are coded as (1). Otherwise, like those whose first language is English, they are recorded as (0). "Sociocultural attitudes": religiosity, civic participation, and traditionalism. Religiosity: Captured through two questions. First, "How much difference would you say religious beliefs make to your life?". Responses are grouped into three categories: (1) "A great difference", (2) "Some difference" or "A little difference" and (3) "No difference". 1 The second question asks, "How often, if at all, do you attend religious services or meetings?". Responses are grouped into three categories: (1) once a week or more, (2) at least once a month, and (3) once a year, never, or only on special occasions. As both questions are only asked in waves one, four and eight, information from wave one is used for the first three waves, data from wave four for the subsequent four waves, and wave eight for the remaining waves. As discussed, since this second question is not a suitable proxy for Muslim female religiosity, it is not included in the women-only models.
Civic participation: Following a similar logic adopted by Heath, Li, and Woerner-Powell (2018) who use information on the number of social organizations a person is a member of or active in as a proxy for bridging capital, here the information is used to proxy for so-called "isolationist tendencies". The rationale is that the lower the number of civic organizations a person is involved with, the more socially isolated they are, and vice versa. Participants are asked about their involvement with 16 different organizations: a political party, trade union, environmental group, parents' or school association, tenants or residents group, religious or church organization, voluntary services group, pensioners organization, scouts or guides organization, professional organization, other community group, social or working men's club, sports club, women's institute or townswomen's guild, women's group or feminist organization, and any other group or organization. The data are only collected in waves three, six and nine. As such, wave three data are used for waves one to five, wave six data for waves six to eight, and wave nine data are used for the last two waves.
Traditionalism: Gender attitudes towards work are a proxy for traditionalism which in turn shapes labour market participation preferences. The more traditional a person, the more likely they subscribe to a gendered view of the division of paid and unpaid labour, with women understood as being responsible for the latter, and therefore more likely to wait to match with a job which fits around family, rendering them less like to be employed. The degree of traditionalism is captured through two questions where respondents are asked whether they (1) strongly disagree, (2) disagree, (3) neither agree nor disagree, (4) agree, or (5) strongly agree with two statements. The first reads, "Husband should earn, wife should stay at home". The second is, "Family life suffers if mother works full-time". As both questions are only asked in waves two, four and ten, information from wave two is used for the first three waves and wave four data are used for all other waves bar wave ten which utilizes its own information.
Finally, wave is included in the model as a categorical variable to account for period effects.

Model
Since the dependent variable is binary I elect to use a logistic regression. 2 Second, as the dataset is hierarchical in nature with observations (level 1) clustered within an individual (level 2), and there are a large number of observations at the higher level (see below), a multilevel (random effect) model is adopted to ensure standard errors are not underestimated. Adopting a random effect model is suitable since people's ethno-religious grouping is relatively stable in the sample. Among the unemployed, out of a total of 11,469 men and 13,941 women, only 1,463 and 1,773, respectively, showed a change in ethno-religious grouping at one point between waves one and ten. Similarly, among the inactive, out of a total of 14,601 men and 21,272 women, only 1,869 and 2,647 respectively, showed a change in ethno-religious grouping at one point between waves one and ten. Importantly, for both groups, the majority were transitions between Christian White British and No Religion White British. This means that a within-subjects design is less useful for understanding the extent to which ethno-religious background is associated with employment outcomes, particularly among ethno-religious minorities for whom there are fewer observations. Adopting a random effect model, which models both within-and between-person effects concurrently, is therefore a suitable approach to adopt (Gayle and Lambert 2018).
I pool ten waves of data to analyse unemployment and inactivity for both men and women distinctly, starting with the former. Pooling waves allows me to disaggregate between groups that have traditionally been combined for sample size reasons, such as Muslim Arabs and Muslim White British, and ethnic minority groups who do not subscribe to a religion. For each analysis, I adopt a stepwise approach. In both instances, first, I run a model examining the differentials in the risk of unemployment and inactivity after common human capital and demographic factors are accounted for. The subsequent models assess how the ethno-religious differences change once so-called "sociocultural attitudes" are considered. Model 2 controls for bonding capital proxied through religiosity. Model 3 adjusts for attitudes towards traditional gender norms and so-called "isolationist tendencies" proxied through the degree of civic participation.
The results are presented as log-odds, which display the average risk of a particular ethno-religious group being unemployed or inactive relative to the Christian White British majority, along with information on the 95 per cent confidence intervals. Confidence intervals that include 0 are not statistically-significant at that level since it implies that the odds of being unemployed are possibly equal to that of the reference category, i.e. equal to 1. A summary of the full models is presented here, with the full regression outputs available online (Tables B1 and B2).

Results
Table 2 (men) and Table 3 (women) examine ethno-religious differences in the risk of being unemployed (Models 1-3) and inactive (Models 4-6) by    Notes: 95% confidence interval in parenthesis; * signifies insufficient sample size to form stand-alone group; in addition to religiosity, traditionalists views, and lower civic participation (where applicable), models are also adjusted for age and its curvilinear effect, marital status, education, health, number of children, whether born in the UK, English language proficiency, region, and period effects.    Notes: 95% confidence interval in parenthesis; * signifies insufficient sample size to form stand-alone group; in addition to religiosity, traditionalists views, and lower civic participation (where applicable), models are also adjusted for age and its curvilinear effect, marital status, education, health, number of children, whether born in the UK, English language proficiency, region, and period effects.

ETHNIC AND RACIAL STUDIES
gender. In each case, Models 1 and 4 show the risk of unemployment and inactivity, respectively, while adjusting for human capital and demographic factors. Models 2 and 5 include religiosity. Models 3 and 6 adjust for the remaining "sociocultural" variables, notably, so-called "isolationist tastes" and commitment to traditionalism. This stepwise analysis is supported by improvements in AIC, BIC and McFadden Pseudo-R 2 (Langer 2017) estimates of the unadjusted models for both men and women. McFadden Pseudo-R 2 for Models 3 are 0.34 (men) and 0.31 (women), while the statistic for Models 6 are 0.49 (men) and 0.47 (women) (Table C1, online), suggesting very good model fit.

Men
Model 1 (  17). Importantly, these differences hold even after controlling for human capital and demographic factors, whose coefficients are in line with expectations. Model 2 adjusts for religiosity. Those who aver religion makes "some difference" to their life have a lower likelihood of being unemployed than those who say it makes "no difference". There is no statistically-significant difference (at the 95 per cent level) between those who say religion makes "a great difference" and the latter group. Meanwhile, relative to those who never attend a religious service or who only attend yearly/on a special occasion, those who attend once a week or more display a greater risk of unemployment. There is no statistically-significant difference between those who attend at least once a month and the reference group. Broadly speaking, there is no major change in the magnitude or significance of the coefficients relative to Model 1 after this adjustment. No Religion Arabs (4.55) remain the group with the highest chance of being unemployed relative to the Christian White British group, those who identify as Black or as mixed Black and White continue to display a higher likelihood of being unemployed irrespective of religious affiliation or lack thereof. Meanwhile, four out of six Muslim groups continue to exhibit a significantly higher likelihood of being unemployed than White British Christians. The hierarchy is Bangladeshi (3.19), followed by Pakistani (2.45), Black African (2.39) and Arabs (2.35). There is one main development, however. The coefficient for No Religion White British has dropped by a quarter and is no longer significant.
Model 3 adjusts for civic participation and commitment to traditionalism. The former is insignificant at the 95 per cent level but the coefficient does show that increased civic engagement reduces the likelihood of unemployment. Only men who neither agree nor disagree (0.53) with the statement "husband should earn, wife should stay at home" have a greater risk of being unemployed relative to those who strongly disagree with the statement. Meanwhile, men's attitudes as to whether "family life suffers if mother works full-time" does not seem to be associated with a lower/ higher likelihood of unemployment. Overall, adjusting for so-called "sociocultural variables" does not dissipate the increased risk for any of the ethno-religious groups relative to the simpler model, not least Muslims whose coefficients do not appear to have reduced in any noteworthy way. The coefficients have also remained largely unchanged across all models for Black and Black and White mixed groups irrespective of religious affiliation. Hindu Indians also continue to display a significantly higher likelihood of being unemployed than the charter population, but, along with Sikh Indians, this is the smallest in magnitude of all significant coefficients.
Examining inactivity, Model 4 reveals that Black and Black and White mixed groups generally have a higher likelihood of inactivity than White British Christians irrespective of religious affiliation (or lack thereof). Contrary to the case of unemployment, White groupssuch as Christian White Irish (1.97), No Religion White Irish (2.51), No religion White Other (2.19)appear to have a higher chance of inactivity than the reference group. No Religion Arabs (3.80), No Religion Chinese (3.92), Christian Indians (2.89), Hindu Indians (2.35) and Sikh Indians (2.14) also appear to have a significantly higher chance of being inactive than the charter population after controlling for human capital and demographic factors. The same is true for all Muslims, bar the White British group; Arabs (2.14), Indians (2.42), Pakistanis (2.58) and Bangladeshis (3.56). The controls operate more or less the same as they do in the unemployment models, but with two key differences. Having three or more children and being educated to any level below degree is associated with a higher chance of being inactive.
Model 5 controls for religiosity. In general, holding that religion makes "some difference" to life is associated with a lower risk of inactivity relative to holding that religion makes "no difference". There is no difference between the latter and those who aver religion makes a "great difference". Meanwhile, attending religious services once a week or more (0.50) or at least once a month (0.60) are associated with a higher chance of inactivity. This does not necessarily mean that those who are more religious are more focused on worship than finding work, it could also indicate that someone who has been unsuccessful in finding employment might turn to more active worship to ease their situation. Broadly speaking, adjusting for religiosity in Model 5 (marginally) lowers the magnitude of the coefficients for some (e.g. Muslims) more than others (e.g. Christians). Nevertheless, Muslim Bangladeshis (3.24), Muslim Pakistanis (2.32), Muslim Black Africans (2.18), and Muslim Indians (2.12), continue to display a significantly higher risk of being inactive than the Christian White British. The confidence interval for Muslim Arabs now includes zero. Moreover, bar the aforementioned exceptions, those who identify with Black groups remain more likely to be inactive be they Christian or aver being of no religious persuasion. In fact, Christian Black and White Africans (4.79) exhibit the highest log odds. They are followed by No Religion Chinese (4.00). Overall, being White British irrespective of religion is not associated with a higher chance of inactivity. However, other groups appear to have a significantly higher chance of being inactive despite being White; Christian White Irish (1.97), No Religion White Irish (2.53), and No Religion White Other (2.27). Finally, Model 6 adjusts for the remaining socalled "sociocultural variables". The results suggest that there is no relationship between a person's commitment to traditionalism and likelihood of being inactive, nor between the latter and the extent of civic participation. There is virtually no change in the significance or magnitude of the regression coefficients of any ethno-religious group relative to Model 5.

Women
Model 1 shows that Muslim groups generally exhibit the greatest risk of unemployment relative to White British Christians. Among Muslims, Pakistanis (3.53) display the highest risk of unemployment, followed by Black Africans (3.21), Indians (2.49), and Bangladeshis (2.19). Among Christians, Indians (3.51) display the greatest risk of unemployment, followed by Black and White Caribbeans (1.80), Black Africans (1.71), Black Caribbeans (1.08), and White Other (1.07). Hindu Indians (1.49) and Hindu Asian Other (3.52) also have a relatively higher likelihood of being unemployed. Jewish, Sikh, Buddhist, and Other Religion groups do not display significant coefficients. Among those with no religious affiliation, only Indians (1.65), Asian Other (2.08), Black Caribbeans (1.19) and British Whites (0.65) display a significantly higher chance of being unemployed than the reference group. The direction and significance of the covariates are the same as it is for men bar a few exceptions. For women, those with a secondary education or who hold an "other higher degree" are more likely to be unemployed. Having three or more children and weak language proficiency are also associated with a higher likelihood of being unemployed, while being born outside the UK does not have a significant relationship with the likelihood of unemployment.
Model 2 shows that religiosity is not significantly related to a higher chance of being unemployed. Moreover, this adjustment does not alter the significance or magnitude of any of the regression coefficients relative to Model 1. Adjusting for so-called "segregationist tendencies" and commitment to traditionalism in Model 3 shows that women who neither agree nor disagree (0.65) with the idea that a husband should earn and wife should stay at home have a higher risk of being unemployed, albeit to a lesser extent than those who agree with the statement (1.42). Conversely, a woman's position as to whether "family life suffers if mother works full-time", does not seem to be related to having a higher risk of unemployment. There does, however, appear to be a significant negative relationship between civic participation and unemployment (log odds=−0.28). Nevertheless, adjusting for all these factors has not altered the coefficient or significance of any ethnoreligious groups by any considerable amount relative to Model 1. In fact, Muslim women (bar Muslim White British) remain among those with the highest risk of unemployment irrespective of ethnic affiliation. Only Christian Indians (3.65) and Hindu Asian Other (3.47) have a higher likelihood of being unemployed relative to the reference group. The same Black, mixed Black and White, and No Religion groups discussed in the previous models also continue to exhibit a higher risk of being unemployed with their coefficients remaining broadly unchanged. Only the estimate for No Religion Black Caribbean is no longer significant.
Models 4-6 in Table 3 display the results for when inactivity is the dependent variable. Model 4 shows that after controlling for human capital and demographic factors Muslims display the highest risk of being inactive across the board relative to the Christian White British group. Specifically, Muslims Arabs (5.90) are the group with the highest risk followed by Muslim Pakistanis (4.18), Muslim Bangladeshi (3.20), Muslim Black Africans (2.66) and Muslim Indians (2.27). Only the Asian Other group -No Religion (3.00), Buddhist (2.87), Hindu (2.84)display a similarly high significant coefficient. Among Muslims, only White British and Asian Other do not display a significant coefficient. Among Christians, only Indians (1.78) and Black and White Caribbeans (1.76) display a significantly higher risk of being inactive. Sikh Indians (1.82), Hindu Indians (1.14), and Other Religion White British (0.67) all display a significantly higher likelihood of being inactive relative to White British Christians. The findings for the controls are in line with expectations and similar to those in the unemployment-only models with one notable exception; having children, irrespective of the number, is associated with a significantly higher likelihood of being inactive.
Model 5 adjusts for religiosity and shows that there does not appear to be a significant relationship between how important religion is to a person's life and their risk of being inactive. There is virtually no change to the magnitude of the ethno-religious coefficient estimates or their significance. The final model, Model 6, adjusts for gender attitudes and civic participation. Relative to those who strongly disagree with the statement "husband should earn, wife should stay at home" those who disagree (0.26), those who neither agree/disagree (0.52), and those agree (1.00) all face a relatively higher risk of being inactive and at an increasing rate. Conversely, much like when examining unemployment, views on whether "family life suffers if mother works full-time" do not seem to be associated with the risk of being inactive. The same is true for civic participation whose coefficient is also close to zero. While controlling for these factors reduces the coefficient for all Muslim groups, it only does so marginally. Muslim Arabs (5.57) remain the group with the highest likelihood of being inactive relative to White British Christians, followed by Muslim Pakistanis (3.91) and Muslim Bangladeshis (2.98). Only No Religion Asian Other (2.99) display a similarly high estimate as the latter. Muslim Black Africans (2.31) and Muslim Indians (2.07) also still display a higher risk of being inactive. The Asian Other group -Buddhist (2.73) and Hindu (2.71)also have a higher relative risk compared to the charter population.

Discussion and conclusion
In this paper, I analysed ethno-religious inequalities in exposure to unemployment and inactivity among men and women in Britain using the first ten waves of UKHLS. The large sample size and data on cultural and religious practices allowed me to assess whether certain so-called "sociocultural attitudes" are plausible mediators for the Muslim penalty. The large dataset also enabled me to distinguish between groups that have not typically been disaggregated in similar studies, such as Arabs, British Whites, Black and White Africans and Caribbeans, and between ethnic minorities with no religious affiliation.
Overall, the evidence indicates support for the thesis that there is both a religious (Muslim) and colour (Black) penalty at play in the British labour market (Khattab and Modood 2015;Khattab 2009;Heath and Cheung 2006). Confirming previous research (Li and Heath 2020;Berthoud and Blekesaune 2007), religion is a much better predictor of unemployment and inactivity for women, whereas for men both colour and religion are important. Adjusting for religiosity, so-called "tastes for isolation" and commitment to traditionalism as potential mediators does not dissipate the Muslim penalty in unemployment or inactivity for either men or women, despite the claim that "[a]fter their inclusion in the explanatory model, there are hardly any statistically significant differences left" (Koopmans 2016, 213). In fact, adjusting for so-called "sociocultural variables" had only a minor effect in reducing the size of the estimates relative to the model that only controlled for human capital and demographic factors, with Muslim men and women consistently among those with the highest risk of being unemployed/inactive. Moreover, the risk of a penalty, particularly in terms of unemployment, remained considerably high for Black African and Black Caribbean men irrespective of whether they subscribed to a faith tradition, providing strong evidence in support of previous research (Khattab and Modood 2015) which established that the British labour market is hierarchized based on skin colour.
In sum, contrary to Koopmans (2016), this study shows that "sociocultural variables" such as gender attitudes, language proficiency, and the extent of inter-and intra-ethnic social ties are not a convincing source of the unexplained ethno-religious differences in labour market participation and unemployment among Muslim men and women. Instead, this study found that "Muslim religiosity and value orientations (…) which sometimes are cited as major individual-level factors hindering socio-economic assimilation turned out to be less consequential" (Connor and Koenig 2015, 199; see also Khoudja and Platt 2018). How can we understand these seemingly opposing findings?
I argue that the divergence could be explained by the variables Koopmans (2016) utilizes to form his "sociocultural assimilation scale". Specifically, measuring the degree of assimilation based on "host-country neighbourhood acquaintances", "host-country friendships", and "host-country family members" ignores (i) the discriminatory housing policies and redlining practices that regulated immigrant neighbourhood settlement (Daniel 1968); (ii) the role racism plays in creating ethnically segregated neighbourhoods (Harrison, Law, and Phillips 2005); (iii) the evidence that White members of the majority culture actively migrate out of, and are less likely to migrate into, neighbourhoods with increased cultural diversity (Bråmå 2006); (iv) that institutional bias directs ethnic minorities towards specific universities less attended by Whites (Shiner and Modood 2002); (v) that interpersonal racism plays an important role in precluding Muslims from establishing multi-ethnic family ties (Pew Research Centre 2018); and (vi) that, in 2011, 46 per cent of the UK Muslim population lived in the 10 per cent most deprived local authority districts in England (MCB 2015) and are, therefore, more likely to live among co-ethnics/co-religionists. Importantly, there is no clear evidence to suggest that the negative impact of residential segregation on social capital accumulation is due more to ethnic concentration than it is to material disadvantage (Laurence 2011).
In other words, the variables used by Koopmans (2016) are likely to be obfuscating important "causes of causes" (Marmot 2018), such as discrimination. This reasoning is compelling in light of the evidence of prevalent racist and prejudicial societal attitudes towards Muslims and ethnic minorities in Britain (Jones and Unsworth 2022;Pew Research Centre 2019;Kelley, Khan, and Sharrock 2017), and the enduring discrimination towards Black and South Asian people which has not subsided since the late 1960s (Heath and Di Stasio 2019). Field experiments that find evidence of discrimination in the British labour market (Thijssen et al. 2021), particularly towards Muslims (Di Stasio et al. 2021), and research showing that ethnic penalties are highly likely to be a reflection of hiring discrimination (Zwysen, Di Stasio, and Heath 2020) lend further support to this argument. In sum, it is not surprising that Koopmans (2016) finds a reduction in ethnic and religious penalties after including as controls variables which are influenced by discrimination if the latter, as the overwhelming research suggests, is an important driver of such penalties.
Positing that discrimination is likely to be playing only a "distant role by affecting sociocultural determinants" (Koopmans 2016, 214) does not acknowledge the complexities of how racism works and how it manifests itself in a intradisciplinary way (Essed 1991). 3 In not recognizing this reality, there is a risk of taking "Muslim behaviour" as "an analytical concept" and therefore "what needs to be represented as a social process and explained is reconstructed as a social fact that can be used to explain other social facts" (Miles and Brown 2003, 91). In such a case "adaptations [to anticipated or experienced discrimination such as where to live or with whom friendships to forge] can easily be coded as choices rather than constraints, as characteristics to be controlled for in estimates of discrimination rather than included as one part of that estimate" (Pager and Shepherd 2008, 199-200). As Virdee articulates it, [t]he focus on religion, culture and the degree of assimilation amongst other factors when evaluating the relative weight of the ethnic penalty shifts the sociological gaze away from racism and external, constraining forces and towards what will inevitably be interpreted as 'problematic' norms, values and patterns of behaviour amongst the minorities themselves. This merely serves to reify the problem and results in a gross underestimation of the multifarious ways in which racism shapes the employment patterns of ethnic minorities. (2010,74) The analyses here also distinguish Arab Muslims and White British Muslims and No Religion Arab and No Religion White British, illustrating notable differences between them and other Muslims on the one hand, and the unreligious on the other. Specifically, previous research (Khattab 2009) analysing White Muslims combined British and non-British Whites which includes Turks and Arabs. However, when distinguishing here between Muslim Arabs and Muslim White British, the results indicate that contrary to Khattab and Modood (2015) and Khattab (2009), being Whitespecifically White British does appear to offer protection against the Muslim penalty. That said, there appears to be a point at which the historically socially constructed privileges of being racialised as White are lost as the association with Islam becomes stronger, possibly resulting in being seen as more culturallydistant and racialised as "more Muslim than White". This might explain why Muslim Arab men, despite identifying as White, have a significantly higher risk of unemployment relative to White British Christians (it is closer to that of Muslim Black Africans than it is to White British Muslims). Among women, Muslim Arabs also display the highest risk of being inactive relative to the charter population, far above the risk associated with being Muslim White British. In fact, among men and women, White British Muslims do not display a significantly different risk of unemployment and inactivity from the charter population in any of the models. If White British Muslims are not being racialised as Muslim but continue to be perceived as White, this might explain why they appear to evade penalization. It might also be the case that White British Muslims are penalized but have better resources with which to counter their adversity (Zwysen, Di Stasio, and Heath 2020).
The evidence that No Religion Arab men are among those with the highest likelihood of unemployment/inactivity (even above that of Muslim Arabs, which could be indicative of the importance of religious bonding capital for labour market inclusion), might suggest that perceived Muslimness is more important for predicting religious disadvantage among men than actual attachment to the faith. This reasoning is particularly compelling given evidence (Di Stasio et al. 2021) that, in Britain, where a Muslim is considered to originate from has a bigger effect on that person's labour market positioning than their actual "religious closeness" and that "country-of-origin effects in combination with anti-Muslim discrimination produce severe double penalties for minority [male] applicants" from the Middle East and Africa (Di Stasio et al. 2021, 16; see also Fundamental Rights Agency 2017). This suggests that the ethnic penalty might not, as currently understood, only be masking a colour and a religious penalty (Khattab and Modood 2015;Modood 2005) but rather, for men, their penalty might also be moderated by a third country-of-origin penalty. This means that an understanding that Islamophobia is multidimensional, and relates to colour, religion, culture and country of origin, with any one dimension of difference being "enough" for someone inclined to be prejudiced, is essential to any strategy seeking to attenuate these inequalities.
There are a number of limitations to this study. Some ethno-religious groups have a small sample size, and it would therefore be beneficial to repeat a similar study with a larger dataset, such as the Census or Labour Force Survey, and compare findings. Exploiting larger datasets might also offer the possibility of creating a Christian Arab group, which would allow testing of whether the country-of-origin effect advanced here is particular to Muslims or those who might be raised Muslim despite no longer subscribing to the faith. That said, these datasets do not offer similarly rich information on social and religious attitudes and practices as UKHLS so there is a trade-off. The paper would also benefit from further theory-building around the mechanisms through which the country-of-origin effect advanced here impacts labour market outcomes.
Even so, the analyses summarized in this paper provide important new findings regarding labour market stratification by ethno-religious background. The paper questions the contention that, amongst men, the ethnic penalty is best understood as resulting primarily from two penalties (colour and religion), and suggests that a country-of-origin penalty may also be at play. This highlights important heterogeneity in the causal mechanisms driving the Muslim penalty, showing that this complexity needs to be understood to clarify how the penalty operates differently for men and women from diverse ethnic backgrounds. This paper also challenges the hypothesis that the Muslim penalty is a result of so-called "sociocultural attitudes" such as religiosity, "tastes for isolation" and a commitment to traditionalism on the part of Muslims. Hence, rather than a focus on alleged "oppositional" norms and behaviours that problematize the faith and essentialize an ethnically heterogenous group of people, attenuating ethnic and religious inequalities will requirein large part but not exclusivelyaddressing both systemic anti-Black and anti-Muslim racism, of which country-of-origin prejudice is likely an important dimension. Notes 1. It is worth noting that while the majority of those who identify as having no religious affiliation aver that religion makes "no difference" to their life, not all do. Specifically, out of a total of 94,400 observations who identify as having no religious affiliation, 24,691 aver that religion makes at least "some difference" to their life. This is not surprising, as people might not identify with a religion but still consider themselves to be spiritual. 2. Results were also estimated using a multilevel mixed-effects multinomial logistic model which was fit using generalized structural equation modelling. However, the binary logistic regression is preferred and is presented here. This is because the assumption of Independence of Irrelevant Alternatives (IIA)integral to multinomial regression, and assumed to hold at the person level in a hierarchal logit modelis not considered to be a realistic assumption for this study. This means that a multinomial model would be incorrectly specified resulting in inconsistent and biased estimates. Importantly, formal tests for IIA are found not to be reliable (Cheng and Long 2007). Crucially, IIA is not an assumption of binary logistic regression, a method used by many other researchers investigating the Muslim penalty (e.g., Khattab and Modood 2015;Khattab, Miaari, and Mohamed-Ali 2020). For transparency, the results of the multinomial generalized structural equation model are available online (Tables B3 and B4). 3. Koopmans (2016) controls for discrimination using perceived discrimination which seems to be capturing its more blatant forms, yet evidence indicates that discrimination operates less obviously (Essed 1991;Pager and Shepherd 2008;Rooth 2010).