Instruction

The instruction is: 'Wenn Sie an Ihre Zeit in Deutschland denken, wie oft wurden Sie in alltäglichen Situationen diskriminiert?' (English translation, for comprehension only: 'When you think about your time in Germany, how often have you been discriminated in everyday situations?'; English original version: In your day-to-day life, how often do any of the following things happen to you?' (Williams et al., 1997). 

Items

Table 1

Items of the German-Language translation of the Everyday Discrimination Scale and the original English Version

Response specifications

Table 2

Response specifications

Scoring

The instrument assesses perceived discrimination in everyday life. The unidimensional structure is drawn from the theory and supported by factor analysis. An unweighted mean score can be calculated and in addition, a factor score can be used. Pairwise exclusion of missing values is recommended. Lower scores are associated with few experiences of discrimination, higher scores with many.

Application field

The target group of the survey instrument are adults aged 18 years or older speaking German, Arabic, or Dari. The use is restricted to the web-based self-administered mode, as the instrument was validated in this mode. The instrument can be used for research purposes only. Considering the results of the reliability and validity analysis, the instrument can be used in each language group separately and also in the heterogeneous language samples. Comparability of correlations is given between Arabic and German languages and between Arabic and Dari respectively. Mean scores can be compared accordingly when not restricting the intercepts of items 1, 2, and 9 in the register sample and  7 and 10 in the facebook sample to be equal among the language groups (Pokropek et al., 2019). Further research and analysis are needed to enable simultaneous comparisons between the three languages.

The scale used in this study was adapted and translated from the Everyday Discrimination Scale (Williams et al., 1997), designed to measure subjective experiences of disadvantage or unfair treatment in day-to-day life. Williams et al. (1997) highlight this perceived discrimination as stressor contributing to health disparities. They emphasize that discrimination can be both individual and institutional, affecting access to resources and opportunities. It can manifest in various areas of life, including healthcare, the labor market, and social interactions (Bonilla-Silva, 1997; Schmitt et al., 2014; Williams et al., 1997). It is often rooted in cultural biases that promote the notion of certain groups being inherently superior to others. These biases frequently lead to the development of negative perceptions and attitudes about outgroups (prejudice), resulting in unequal treatment of individuals within these groups by individuals as well as social institutions. Discrimination thus represents the behavioral component of stereotypes (Blank et al., 2004). This unequal treatment can manifest in disparities, limiting socio-economic opportunities, social mobility, and life chances for members of outgroups (Williams et al., 1997).

Perceived discrimination, referring to the subjective perception of individuals being disadvantaged on the basis of personal characteristics such as gender, ethnicity or cultural origin, is one of the most frequently researched forms of discrimination. In recent years, the effects of perceived discrimination on health are of growing scientific interest. Numerous studies have demonstrated that experiencing discrimination is closely associated with adverse health effects, such as stress, which can significantly affect psychological well-being (Pascoe & Smart Richman, 2009; Scott et al., 2022; Williams & Mohammed, 2009). Williams et al. (1997) identified perceived discrimination as a critical determinant of disparities in health. While socio-economic status (SES) explains part of the variance in health outcomes, perceived discrimination remains a significant indicator affecting both physical and mental health.

Perceived discrimination is measured by assessing individuals’ perception of experiences of discrimination in everyday situations. Questions determine whether the respondents have the feeling of being discriminated against, to enable a precise assessment of the construct. Refugee migration poses new challenges for survey research, as established measurement instruments can be understood, interpreted and answered differently by people from different linguistic and cultural backgrounds. In order to ensure the comparability of studies across different cultures and languages, equivalent survey instruments are to be developed and their comparability examined and evaluated.

The Everyday Discrimination Scale has been validated in numerous countries, including the U.S., Germany, and South Africa, for both general populations and specific demographic groups, such as minorities (Barnes et al., 2012; Clark et al., 2004; Ulusoy et al., 2023; Williams et al., 1997). Krieger et al. (2005) found an internal consistency (Cronbach´s alpha) of .88. Similarly, Gonzales et al. (2016) reported a Cronbach’s alpha of .92. Regarding validity, the Everyday Discrimination Scale has been shown to correlate with measures of mental and physical health outcomes, including depression, anxiety, and hypertension, providing strong evidence for its criterion validity (Lee & Turney, 2012; Sawyer et al., 2012; Seaton et al., 2008; Williams et al., 2003).

Item generation and selection

The German version of the survey instrument is based on a translation of the Everyday Discrimination Scale by Williams et al. (1997). The items were retained in the context of the study, but the instruction was adapted to the survey context in Germany. In addition, the response scale was changed from a 6-point scale to a 5-point scale to include a midpoint (Menold & Bogner, 2016). Finally, in order to survey Arabic and Dari speakers, the measurement instrument was translated into the respective languages. The TRAPD approach (Translation, Review, Adjudication, Pretest, Documentation) (Behr et al., 2016) was chosen as the translation method. In this approach, two professional translators for Arabic and Dari independently translated the German version of the survey instrument into Arabic and Dari respectively. Problems as well as inconsistencies were then discussed in moderated sessions to reach consensus on the translations. The two Arabic translations were identical for eight of the nine items, with only item 6 differing. In the discussion, a standardized wording was found for item 6.

Cognitive pretests

Cognitive interviews were conducted with 18 refugees (Syrian N = 6, Iraqi N = 6, Afghan N = 6) of different gender, age, and education (Hadler et al., 2021) between December 2020 and January 2021. The instrument was not revised after interviews, due to minor or absent cognitive problems of respondents. For more details on cognitive interviews see Hadler et al. (2021).

Samples

Register sample

A probabilistic sample of three large Cities in Saxony (Dresden, Leipzig, Chemnitz) was implemented. The target population included German-speaking residents, Arabic-speaking Syrian and Iraq residents (refugees). The addresses of sampled persons, together with information on gender, age and country of origin was provided by municipal registration offices. We only required the selection of addresses for refugees who had been resident in the city since 2014. In total, contact details for 1,000 German speakers, 2,000 Arabic speakers and 1,000 Dari speakers were provided. We used mail contact and re-contacts to recruit the participants to a web-based survey. Response rates RR6 (AAPOR, 2023) were 28% for German-speaking residents, 19% for Arabic-speaking refuges and 16% for Dari-speaking refugees. The sample was randomized, with one group of respondents answering the perceived discrimination questionnaire version before the cognitive pretest, and the other group answering the questionnaire version after the cognitive pretest. For the perceived discrimination scale version administered before the cognitive pretest, relevant to this report, the sample size was N = 134 for Arabic speakers and N = 102 for German speakers.

Facebook sample

The Facebook survey was hosted on the servers of TU Dresden and targeted people from Iraq, Syria and Afghanistan living in Germany. An external service provider was commissioned to advertise the web-based survey. In addition, Bielefeld University supported advertising, but without the possibility to collect socio-demographic information. To ensure the participants are refugees speaking these languages, we conducted a screening and allowed to participate refugees living in Germany only. The questionnaire was available in Arabic and Dari only. The realized sample (N = 3448) consisted of N = 1925 Arabic and N = 1523 Dari speaking persons. A random part of them administered the adapted instrument as developed after the cognitive pretests. For the perceived discrimination scale version administered before the cognitive pretest, relevant to this report, the sample size was N = 443 for Arabic speakers and N = 256 for Dari speakers. See Tables 3 and 4 for a description of both samples.

Table 3

Descriptives Register sample

Note. N refers to the total sample size of respondents for German/Arabic with available demographic data for the specific question. Deviations in N due to Data Missing; n.a.: not applicable.

Table 4

Descriptives Facebook sample

Note. N refers to the sample size of respondents for Arabic/Dari with available demographic data. Deviations in N due to Data Missing; n.a: not applicable.

Item analyses

Tables 5-7 show the item distribution parameters for the total register sample (Table 5) as well as for the Arabic (Table 6), and German (Table 7) subsample of the instrument. Tables 8-10 display item distribution parameters for the total Facebook sample (Table 8) as well as the Arabic (Table 9), and Dari (Table 10) subsample.

Table 5

Register sample (total): N, Means, Standard Deviations, Skewness, Excess/Kurtosis, Item-scale correlation (corrected), and POMP scores of Items

Note. Scale ranging from 1 (never) to 5 (all the time). POMP score = Percentage of Maximum Possible score.

In the total register sample, items 3-6 and 8-9 have rather high item difficulties with mean values < 2.00. This indicates that there are fewer experiences of discrimination for these items (e.g. "You have been harassed" or "You have been threatened"). None of the 9 items are normally distributed (Shapiro-Wilk test p < 0.001). All items show positive significant skewness. All items with significant kurtosis are broad, while item 7 ("People behave as if they are better than you") is narrowly peaked. All items have a sufficiently high discriminatory power (item-scale correlation, corrected) (>0.4). 

In the Arabic speaking subsample (Table 6), the number of items with high item difficulties is reduced (items 3, 6, and 9), whereas it is increased in the German speaking subsample, where all items except item 1 exhibited mean values < 2.00 (Table 7). The number of items with a significant kurtosis is also lower in the Arabic subsample compared to the overall sample, whereas in the German-speaking subsample all items except item 1 have a significant kurtosis and are broadly peaked. All other item statistics in both subsamples are comparable to these found in the total register sample. The items provide a high discriminatory power in each subsample as well.

Table 6

Register sample (Arabic): N, Means, Standard Deviations, Skewness, Excess/Kurtosis, Item-scale correlation (corrected), and POMP scores of Items

Note. Scale ranging from 1 (never) to 5 (all the time). POMP score = Percentage of Maximum Possible score.

Table 7

Register sample (German): N, Means, Standard Deviations, Skewness, Excess/Kurtosis, Item-scale correlation (corrected), and POMP scores of Items

Note. Scale ranging from 1 (never) to 5 (all the time). POMP score = Percentage of Maximum Possible score.

In the total Facebook sample (Table 8), items 3, 6 and 9 have rather high item difficulties with mean values < 2.00. This indicates that there are fewer experiences of discrimination for these items (for example, "People accuse you of being dishonest" or "You have been threatened"). None of the items have a normal distribution (Shapiro-Wilk test p < 0.001). All items show positive significant skewness. Items 3, 6 and 9 with significant kurtosis are broad peaked, while items 4 and 7 are narrowly peaked. All items have a sufficiently high discriminatory power (item-scale correlation, corrected > 0.4).

Item statistics do not differ notably in the subsamples of respondents speaking Arabic (Table 9) or Dari (Table 10). An exception is item 8 in the Dari subsample, which also shows a comparatively higher item difficulty and significant positive kurtosis.

Table 8

Facebook sample (total): N, Means, Standard Deviations, Skewness, Excess/Kurtosis, Item-scale correlation (corrected), and POMP scores of Items

Note. Scale ranging from 1 (never) to 5 (all the time). POMP score = Percentage of Maximum Possible score.

Table 9

Facebook sample (Arabic): N, Means, Standard Deviations, Skewness, Excess/Kurtosis, Item-scale correlation (corrected), and POMP scores of Items

Note. Scale ranging from 1 (never) to 5 (all the time). POMP score = Percentage of Maximum Possible score;

Table 10

Facebook sample (Dari): N, Means, Standard Deviations, Skewness, Excess/Kurtosis, Item-scale correlation (corrected), and POMP scores of Items

Note. Scale ranging from 1 (never) to 5 (all the time). POMP score = Percentage of Maximum Possible score.

Reliability

To assess reliability, the factor analysis based composite reliability coefficient Raykov's Rho was calculated on the basis of confirmatory factor analyses (CFAs) reported in the section “Factorial Validity” for each subsample. Multi-group confirmatory factor analyses (MG-CFA) of configural models (section “Measurement Equivalence”) were used to calculate Raykov's Rho in both samples (Overall Sample in Table 11). We used Mplus 8.7 (Muthén, & Muthén, 1998-2021) for all factor analyses. As result, composite reliability of the adapted perceived discrimination scale was found to be high in all samples and subsamples (Table 11).

Table 11

Register and Facebook sample: Composite Reliability (Raykov´s Rho)

Validity

Factorial Validity

Factorial validity was evaluated by means of CFAs (Table 12). The model fit in the CFAs was evaluated using the chi-square test, the Root-Mean-Square Error of Approximation (RMSEA), and the Comparative Fit Index (CFI) (Beauducel & Wittmann, 2005). An CFI of 0.95 or higher and RMSEA of 0.08 or less indicate an acceptable fit (Hu & Bentler, 1999). The Robust Maximum Likelihood estimator (MLR) was used due to the ordinal nature and non-normality of the data (Muthén & Muthén, 2014). As five categories were used we treated indicators as continuous (Beauducel & Herzberg, 2006). The one factor could be supported as a just acceptable or acceptable model fit was reached in each subsample, with factor loadings exceeding the conservative threshold of 0.4 for all items (Ford et al., 1986).

Table 12

Register and Facebook sample: Results of CFAs

Note. Standardized factor loadings reported.

*p < 0.05, **p < 0.01

Convergent Validity

Convergent validity was assessed through Pearson correlations between the perceived discrimination scale and structural discrimination in different areas of the societal lives, whereas moderate to high correlations of .4 (p < 0.001) and above were expected. There was a moderate positive correlation between structural discrimination (SOEP Group, 2019) and perceived discrimination (r = .63, N_Register = 221, r = .50, N_Facebook = 685). Moderate correlations have been also found for perceived discrimination and the subscales of structural discrimination for 1) health system, 2) offices and authorities, 3) police, 4) leisure and everyday life, 5) labour market and 6) Housing market. This means that individuals who experienced structural discrimination also tended to report higher levels of perceived discrimination (see Table 13). 

Table 13

Register and Facebook sample: Convergent Validities of the adapted Perceived Discrimination Scale

Note. Pearson’s correlations.

***p < 0.001

Divergent Validity

For divergent validity, correlations between perceived discrimination and unrelated concepts were anticipated to be non-significant (p > 0.05), with correlation coefficients below |.3|, indicating weak associations. For both samples a weak significant correlation has been found between perceived discrimination and different dimensions of authoritarianism (KSA-3 (Beierlein et al., 2015)). The correlations between perceived discrimination and the Submissiveness subscale of authoritarianism have been found to be the highest for both samples (r = -.23, N_Register = 227, r = -.22, N_Facebook = 626). In contrast, the correlations between perceived discrimination and the Aggression subscale of authoritarianism have been found to be the lowest for both samples (r = -.04, N_Register = 224, r = -.14, N_Facebook = 624). Despite the significant correlations, the magnitude of these associations remained consistently low, suggesting a weak relationship between perceived discrimination and authoritarianism (see Table 14).

Table 14

Register and Facebook sample: Divergent Validities of the adapted Perceived Discrimination Scale

Criterion Validity

The criterion validity has been established through Pearson correlations between perceived discrimination and health outcomes, assumed to be significantly correlated, with scores of r = |.3| (p<0.001) and above. There were moderate to strong correlations with possible external criteria, both in the register sample and in the Facebook sample. There were moderate negative correlations between perceived discrimination and life satisfaction (Richter et al., 2013) (r = -.41, N_Register = 110, r = -.30, N_Facebook = 358) indicating that individuals with high levels of perceived discrimination had lower life satisfaction. Furthermore, perceived discrimination was significantly positively correlated with depression and anxiety (Loewe et al., 2010), which shows that people who report higher levels of perceived discrimination also have higher levels of depression and anxiety (register sample: r_Depression = -.29, r_Anxiety = .42, N = 112; Facebook sample:  r_Depression = .33, r_Anxiety = .35, N = 352 resp. 351). There was a moderate positive relationship between perceived discrimination and loneliness (Richter & Schupp, 2015) for both samples (r_Register =  .53, N_Register = 127, r_Facebook = .35, N_Facebook = 348) indicating higher scores of loneliness with rising scores of perceived discrimination. Furthermore, there was a low to moderate negative correlation between perceived discrimination and general trust to other people (OECD, 2017) for both samples (r_Register =  -.16, N_Register = 230, r_Facebook = -.29, N_Facebook = 627), indicating lower scores of trust with rising scores of perceived discrimination (see Table 15).

Table 15

Register and Facebook sample: Criterion Validities of the Perceived Discrimination Scale

Note. L-1 = General Life Satisfaction Short Scale (Beierlein et al., 2015), PHQ-4 = Short Form Patient Health Questionnaire (Loewe et al., 2010), UCLA = University of California, Los Angeles (Russell et al., 1978), OECD = Organisation for Economic Co-operation and Development (OECD, 2017), Pearson’s correlations.

*p < 0.05, ** p < 0.01, ***p < 0.001

Further quality criteria

Measurement Equivalence

Measurement invariance analyses were conducted by means of MG-CFAs. The analyses were conducted for the register sample comparing Arabic and German and for the Facebook sample comparing Arabic and Dari. A significant change of chi-square (Meredith, 1993) or a change of ΔCFI > .010 and ΔRMSEA > .015 indicated significant differences in model fit (Chen, 2007) and thus lack of exact measurement invariance. See Tables 16 and 17 for the results of the measurement invariance analyses. For both samples it can be seen that the configural and metric measurement invariance are given, but scalar measurement invariance is violated. A few of the items contributed significantly (modification indexes larger than .10) to scalar non-invariance. Partial scalar invariance holds in the register sample when freeing intercepts of the items 1, 2 and 9. It also holds in the Facebook sample when freeing intercepts of items 6 and 9. Therefore, items 1, 2, 6 and 9 violate scalar measurement invariance and should be not restricted to be equal when comparing factor means. Correlations can be compared for German with Arabic and Arabic with Dari respectively. Mean differences for scale scores can be obtained between the respective language on the basis of partial invariance models (Pokropek et al., 2019). For both samples, there were significant factor mean differences showing a lower discrimination for German residents than for Arabic speaking refugees and for Arabic than Dari speaking refugees respectively.

Table 16

Register sample: Results of measurement invariance analysis

Note. ^a Factor mean differences included, unstandardized mean difference of ΔM = -0.84 (SE=.16,
p = .001) showing lower values for German speaking residents.

*p < 0.05, ** p < 0.01, ***p < 0.001

Table 17

Facebook sample: Results of measurement invariance analysis

Note. ^a Factor mean differences included, unstandardized mean difference of ΔM = -0.25 (SE = .09,
p = .01) showing lower values for Arabic speaking refugees.

*p < 0.05, ** p < 0.01, ***p < 0.001

Acknowledgement

This work was founded by German Science Foundation (DFG) in the scope of the ENSURE project as part of the PH-LENS Research Unit (FOR 2928/GZ: ME 3538/10-1).