Ethnic Discrimination
A study on Swedish Municipalities
Bachelor Thesis
Authors: Måns Boström
Aaron Åberg
Supervisor: Magnus Carlsson
Examiner: Mats Hammarstedt
Term: Spring 2020
Subject: Economics
Level: Bachelor
Abstract
This paper analyzes whether there are differences in treatment for Western-Balkan and Arabic-sounding names when requesting information about the Swedish for Immigrants (SFI) program by Swedish municipalities. To answer this question, a correspondence test was conducted in which all 290 municipalities in Sweden were contacted via emails from two fictitious inquirers. We were able to gather data on six outcome variables from this correspondence test, which were chosen to measure the time and effort spent on replies to each inquirer. The results suggest that government officials provide differential treatment in favor of a Western-Balkan-sounding name, but that the outcome variables are relatively small. Moreover, we found no strong evidence for whether this differential treatment is due to taste-based or statistical discrimination. The findings in our study highlights potential consequences for immigrants with an Arabic background as limited access to the SFI program could have implications for their integration into the labor market.
Acknowledgements
We want to extend our gratitude to our supervisor, Magnus Carlsson, for his enormous support and encouragement. His expertise and helpful guidance have been of great help to us throughout this process. We would also like to thank our discussant, Marcus Hansen, for his constructive feedback during the seminars. Additionally, we would like to thank Jasmin Huskanovic and Amir Alic in assisting us with choosing the name for the Western-Balkan inquirer. Finally, we would like to thank the members of faculty at the School of Business and Economics who have provided us with great insight and knowledge about the topic of economics over these past three years. In particular, we would like to mention Lina Aldén, Hyunjoo Kim Karlsson, Chizheng Miao, and Dingquan Miao.
Keywords
Table of Contents 1. Introduction ... 3 2. Theoretical Framework ... 5 2.1 Taste-Based Discrimination ... 5 2.2 Statistical Discrimination ... 7 2.3 Connection to Municipalities ... 8 3. Literature Review ... 9
3.1 Methodology in Ethnic Discrimination Studies ... 9
3.2 Ethnic Discrimination in the Housing and Labor Markets ... 12
3.3 Ethnic Discrimination by Public Authorities ... 14
4. Methodology ... 15
4.1 The Experiment... 15
4.2 Choice of Ethnicities and Names ... 17
4.3 Outcome Variables ... 18 4.4 Econometric Specification ... 19 4.5 Statistical Power ... 19 4.6 Robustness Check ... 20 4.7 Ethical Considerations ... 20 5. Data ... 20 5.1 Descriptive Statistics ... 20 6. Results ... 21 6.1 Regression Results ... 23 7. Discussion ... 25
7.1 Taste-Based or Statistical Discrimination? ... 26
8. Conclusion ... 28
8.1 Limitations and Future Research... 28
Appendix ... 30
1. Introduction
Sweden is a country with a reputation for being an egalitarian welfare state with ambitious multicultural policies, where equal treatment of all citizens before the state is of utmost importance for a democratic rule to be fulfilled (Adman and Jansson, 2017). One such state entity is the typical Swedish municipality, which constitutes the lowest level of government. Following the Principle of Equality, Swedish municipalities are legally obligated to treat all of their citizens in an equal manner (Swedish Code of Statutes, 2017:725). Furthermore, municipalities are responsible for a substantial number of local services, including educational, emergency, and other social services. One such service is the Swedish for Immigrants (SFI) program, which is intended to provide newly arrived immigrants with basic skills in the Swedish language as well as a basic understanding of the Swedish society. Moreover, according to the Education Act of 2010, Swedish municipalities are required to actively reach out and motivate immigrants to enroll in the SFI program (Swedish Code of Statutes, 2010:800). A vast array of scholars argue that host-country specific skills, such as language proficiency, have a profound positive effect on immigrant labor market outcomes (Berman et al., 2003; Bleakley and Chin, 2004; Dustmann and Fabbri, 2003). This notion has also been demonstrated to exist in Sweden, as Aslund and Rooth (2006) confirm that there is a positive correlation between proficiency in the Swedish language and the labor market success of immigrants.
The SFI program is one possible way to acquire language proficiency for immigrants, which the municipalities are responsible for providing. However, previous studies suggest that differential treatment exists between natives and immigrants in interactions with government officials. Although such discriminatory behavior is unlawful, immigrants, especially with an Arabic background, have been subject to discrimination by government officials.
Only a few select studies have attempted to analyze discriminatory behavior by local government entities in Sweden. These studies have used correspondence tests in which two fictitious applicants with different ethnic backgrounds have expressed interest in the same municipal service. The purpose of these studies has been to determine if differential treatment frequently exists when people with Swedish- or Arabic-sounding names attempt to interact with the municipalities. If differential treatment is found to exist, it is then analyzed by comparing the mean values of a group of outcome variables to see if these are statistically significant. However,
observation per municipality. As a result, we have identified a gap in these studies, which we will attempt to address.
First, this study will attempt to determine if there are signs of differential treatment between immigrants with an Arabic background and other immigrant groups. Discriminatory behavior towards immigrants with an Arabic background is well documented and studied in various marketplaces such as the housing and labor markets (see Ahmed and Hammarstedt, 2008; Ahmed et al., 2010). The other immigrant group of interest will be individuals from the Western-Balkan region. This group is interesting to investigate as it argued that individuals from this region have been more successful in their integration into the Swedish labor market, but also because they represented the fourth-largest immigrant group in Sweden in 2019 (Ruist, 2018 and Statistics Sweden, 2020). Despite this, no major studies are underway investigating whether the Arabic immigrant group is subject to greater discrimination by Swedish government entities than other groups. In order to achieve higher statistical power, this study will also attempt to collect two samples from each municipality, essentially doubling the total number of observations as compared to previous studies.
The main objective of this study is to analyze and determine if there are any statistically significant differences between the treatment of immigrants with Western-Balkan- and Arabic-sounding names when requesting information about the Swedish for Immigrants (SFI) program administered by Swedish municipalities. Since individuals with an Arabic background are treated differently than individuals with Swedish-sounding names by public authorities, the objective is also to investigate if individuals with an Arabic background are treated differently when compared to the Western-Balkan immigrant group.
The rest of this study is structured as follows. Section 2 describes the strengths and weaknesses of four commonly used research methods as well as previous literature on ethnic discrimination. Section 3 discusses the theories of discrimination. Section 4 will cover the methodological framework. Section 5 will describe data gathered. Section 6 will present the results, and section 7 will discuss the ramifications of the results. Finally, section 8 will summarize and conclude the results of this study.
2. Theoretical Framework
In the late 20th century, two main theories regarding discrimination in various markets have been
used in economic literature. These are taste-based discrimination and statistical discrimination and below follows an explanation of the two theories.
2.1 Taste-Based Discrimination
The Nobel Laureate Gary Becker first introduced the theory of taste-based discrimination in his 1957 doctoral dissertation, “The Economics of Discrimination”. Becker’s theory is considered the birth of modern economic analysis of discrimination, as much of the subsequent academic literature on discrimination is guided through the analytical framework of his dissertation (Borjas, 2020).
Becker’s theory primarily bases the concept of taste-based discrimination on the labor market, where the concept attempts to explain the notion of ethnic prejudice in an economic context. For example, assume that there are two types of workers in the labor market: native and immigrant workers with wage rates wN and wI, respectively. Now, if the employer is prejudiced towards immigrants, the employer in question will receive disutility from hiring an immigrant worker. Put differently; the employer will act as if it costs wI(1+d) to hire an immigrant worker even though
the actual cost is wI. The discrimination coefficient (d) will work as a percentage markup in the hiring cost of an immigrant worker, based on the employer’s prejudice. Thus, the discrimination coefficient essentially monetizes ethnic prejudice. Moreover, the more prejudiced an employer is, the more disutility he or she will receive from hiring an immigrant worker, and the greater the discrimination coefficient will be (Borjas, 2020).
In his dissertation, Becker presents different types of discrimination that can occur in the labor market, e.g., employer, employee, and customer discrimination. For simplicity, an example of ‘employer discrimination’ will be presented. This example assumes that there are two types of workers in the labor market, similar to the example presented above. Furthermore, suppose a competitive firm is going to decide how many of these workers to hire. This model assumes that both workers are perfect substitutes in production. As a result, the production function can be written as:
where (q) represents the output of the firm, EN represents the number of hired native workers, and
EI denotes the number of hired immigrant workers. Further, Equation 3.1 claims that the output of
the firm depends on the total number of hired workers, irrespective of their ethnicity. Therefore, the firm will get the same output if it hires 100 natives and 0 immigrants as it would vice versa. This notion explains that the marginal product of labor (MPE), or the output produced by an
additional worker, will be the same regardless if the worker is a native or an immigrant. Further, since both workers are equally productive, differences in economic status across the two groups cannot be attributed to differences in skill, rather it must arise from discriminatory behavior of labor market participants.
A discriminatory firm will base their hiring decision by comparing wN with wI(1+d) and hire
the worker that has a lower utility-adjusted price. If the employer discriminates against immigrant workers, then he or she will decide based on the following rules:
Hire only immigrants if: wI(1+d) < wN
Hire only natives if: wI(1+d) > wN (3.2)
Becker further argues that there are two types of firms: firms that only hire native workers (native firms) and firms that only hire immigrants (immigrant firms). This depends on the magnitude of the discrimination coefficient; if the coefficient is very high then a firm will be a native firm and if the coefficient is relatively low it will be an immigrant firm.
As illustrated in Figure 1, a native firm will hire native workers up to the point where the native wage and the marginal product value are equal. In Figure 2, it can be seen that an immigrant firm will hire immigrant workers up to the point where the value of marginal product is equal to the utility-adjusted wage. Furthermore, a firm with a discrimination coefficient d0 will act as if the price of immigrant workers is wI(1+d0) and this coefficient is small enough for the firm to still
want to hire an all-immigrant workforce. The firm will hire immigrant workers up to the point where the utility adjusted wage is equal to the value of marginal product at EI0 workers. If the
discrimination coefficient were to be larger (dI), the firm would hire less immigrant workers EI1.
As a result, the number of hired immigrant workers will be smaller for firms with larger discrimination coefficients as employers who do not prefer immigrant workers will minimize their discomfort by hiring fewer immigrant workers.
2.2 Statistical Discrimination
Statistical discrimination remains a well-known topic within the theory of discrimination. The theory was developed in order to appraise the rationality and economic perspectives of a decision-makers process. Rational decisions may contribute to cost efficiency, meanwhile obtaining additional information provides costs for the employer and is more inefficient (Arrow 1971; Phelps 1972). However, rational decisions may not be the optimal choice for the employers, since decisions are based on observable characteristics such as ethnicity or gender and do not determine the individual’s productivity.
To explain this further, a commonly used example defining statistical discrimination will be presented. Assume that a firm is at the initial stage of hiring a worker. The recruiter received applications from two potential workers with equal skills, where the only difference is their country of origin, one native and one foreign-born. Earlier, the recruiter had encountered foreign workers with lower productivity relative to native workers and therefore assumed that the new foreign applicant would also generate lower productivity than the native applicant. As a result, the firm hires the native worker and declines the foreign worker’s application.
As both Phelps (1972) and Arrow (1971) argued, the recruiter is only able to detect observed characteristics (e.g., ethnicity and gender). Besides, the recruiter is more likely to use these observables as a simpler method for presuming the unobservable characteristics of an individual.
A scoring test may be a reasonable method to perform for the firm to determine an individual’s true productivity. Though, the construction of the test is associated with costs for the employer. However, if firms are assumed to maximize profits, the usage of observable characteristics as a proxy for the unobserved does not provide additional costs for the firm. If observable characteristics seem trustworthy to the firm as low-cost information in order to determine the applicant’s productivity level, prejudices towards certain ethnicities or genders may be easier introduced during the hiring process. Therefore, firms rely on the average aggregated group characteristics to evaluate individual skills. As in the mentioned example, the said recruiter had encountered lower productivity among foreign workers than natives. Therefore, ethnicity is one determinant factor of which applicant the firm hires (Arrow, 1973). Statistical discrimination seems more connected to beliefs and prejudices towards certain groups, regardless of whether the assumptions are true. Nygren (2004) stated that decision-makers might not be aware that they are discriminating towards a certain ethnicity or gender.
2.3 Connection to Municipalities
In terms of this study, it is important to stress that there is no actual theory explaining how government officials might discriminate against their citizens based on ethnicity. The theory of statistical discrimination mainly focuses on differences in treatment of natives and immigrants across various markets (see, for example, Heckman 1998; Ahmed et al. 2010). The theory deviates from the subject investigated in this study. However, the fundamentals of discrimination remain the same. Consequently, it is problematic for us to conclude if statistical discrimination exists within the municipalities based on the existing theory.
Since government officials do not “employ” individuals moving to their municipalities, the theory of taste-based discrimination is not directly applicable to our study. Although Becker’s theory of discrimination can help us get a better understanding of why government officials might have an incentive to discriminate against their citizens. In this case, the discrimination coefficient is highly relevant as it can explain why a government official might want to provide different amounts of information to inquirers, which would be based on their taste for certain immigrant groups. By slightly modifying the theory of ‘Employer Discrimination’, we can assume that the employers will be the government officials, and the workers will be inquirers. Thus, a government official will refuse to provide information about the SFI program to inquirers with a higher
marginal cost than their marginal benefit to the municipality in which the inquirer intends to live and attend the SFI program. A government official faced with the cost (ci) will act as if ci(1+di) is the net cost of providing information to an inquirer. Suppose that we have two immigrant groups, Western-Balkans (B) and Arabs (A), who are perfect substitutable citizens in a municipality (i.e., they incur the same cost and benefits to the municipality). Further, the government official has a discrimination coefficient of value (di) towards (A). If it is assumed that the cost of providing
information to (B) is cB and this is less than cA(1+di), then a government official will only provide
information to the Western-Balkan individual as the intensity of tastes is greater than the benefits of helping the individual with an Arabic background. With that said, the concept of taste-based discrimination could also be relevant in situations where government officials and citizens are in contact.
3. Literature Review
3.1 Methodology in Ethnic Discrimination Studies
In economic literature, researchers have primarily used four methods to investigate discrimination in various markets: attitude surveys; regression analysis; audit studies; and correspondence studies. A discussion on the four methodologies below highlights the potential strengths and weaknesses associated with each method.
Attitude Surveys: A standard method often employed to investigate discrimination is to use attitude surveys. Such surveys have been used in a large body of work (see, for example, Lange, 1998 and Zussman, 2013). Together, the authors used these surveys to capture attitudes towards certain ethnic groups in society. Lange (1998) conducted a survey to measure whether individuals of Sami ethnicity have been exposed to unequal treatment due to their origin when they contact local services. The survey results concluded that one-fifth of the respondents had been exposed to harassment or insults once or twice at work, and one-third have been subject to verbal insults in their personal life. Charles and Guryan (2008) tested the predictions of a model referred to as Becker’s model of discrimination in the American labor market. They collected data from the General Social Survey to capture prejudiced attitudes towards African-Americans among whites in the United States. Zussman (2013) developed an attitude survey to investigate discrimination and prejudiced behavior on the Israeli online market for used cars. The author found that Jewish
sellers responded differently if the buyer was Arabic and that this behavior was linked to the seller’s attitude, as stated in the attitude survey.
There are several drawbacks to using survey data (Bradburn, 1984; Carlsson and Eriksson, 2017). Carlsson and Eriksson (2017) point to three main drawbacks of using surveys to detect ethnic discrimination. First, respondents might have incentives to report false attitudes in a survey if their true attitude violates that of social norms. Second, respondents might have hidden attitudes (which they are not aware of) that can affect their behavior, but not their stated explicit attitudes. Third, situational factors could affect the magnitude of actual discrimination without affecting the attitudes of an individual. An example of a situational factor would be the overall ease of finding workers or tenants during a particular historical period. The authors argue that this can affect the magnitude of discrimination in the labor and housing markets, but not the respective attitudes of employers and landlords.
Regression Analysis: Another frequently used method for studying discrimination is regression analysis (see, for example, Yinger, 1998; Arai and Skogman Thoursie, 2009; and Bertrand and Duflo, 2017). The main purpose of using regression analysis is to isolate the impact of discrimination on outcome variables without directly observing discriminatory behavior. The test for discrimination is whether the regression produces a significant coefficient for the proper group membership variable. For example, Arai and Skogman-Thoursie (2009) studied how earnings changed when immigrants from Asian, African, or Slavic countries changed their surnames to a Swedish- or neutral-sounding name. To estimate this effect, the authors regressed annual earnings on a dummy variable indicating whether an individual had changed their surname or not. They found that immigrants who changed their surname saw a substantial increase in annual earnings.
However, Yinger (1998) argued that a regression analysis presented three limitations. He also argued that omitted variable bias might arise if relevant control variables correlated with group membership are excluded from the regression. Second, included variable bias arises if some independent variable is influenced by the economic agent whose discriminatory behavior is being investigated. Third, diverting variable bias arises when variables are correlated with ethnicity but are not appropriate control variables included in the regression. The potential presence of said biases could make estimated coefficients challenging to interpret and, in turn, make it more challenging to detect actual effects of discrimination.
Audit Studies: Scholars developed the audit study due to the limitations of the regression analysis. Yinger (1998) defined an audit study as a matched-pair survey technique in which at least two physical auditors from different groups (e.g., gender, ethnicity, or age) are selected, coached, and sent to apply or negotiate for a good or service. The main purpose of conducting an audit study is to observe whether or not physical auditors from one specific group are systematically treated differently than auditors from a different group. The advantage of the audit study is that potentially omitted, included, and diverting variable biases can be eliminated. Researchers can carefully design their study such that auditors do not differ on any characteristic other than one specific trait relevant to their study.
With that said, there has also been criticism directed towards audit studies in several publications (Heckman and Siegelman, 1995; Heckman, 1998; Yinger, 1998; Bertrand and Duflo, 2017). The authors are concerned about whether auditors are required to know the purpose of the study. They are concerned that physical auditors might have their own perception of gender and ethnic issues. They might consciously or sub-consciously create motives to generate reliable or unreliable data. Furthermore, to make a pair of auditors utterly identical in all factors that might affect productivity, researchers would need to employ auditors with similar physical attributes and train them for many days to coordinate interviewing styles.
Correspondence Studies: One method that is highly recommended for detecting discrimination in various markets is the correspondence testing (CT) method (Bertrand and Mullainathan, 2004; Rooth, 2014). This method is similar to the audit study, but rather than using physical auditors, CT uses fictitious applicants. Bertrand and Duflo (2017) describe the CT as researchers responding to housing, job, mortgage, and other advertisements by sending numerous fictitious resumes or letters of interest to the subjects being studied. These fictitious resumes have similar characteristics except for a minority trait relevant to the study in question. Discrimination is estimated by comparing the outcomes (usually call-backs) between the fictitious applicants with the minority trait and without the minority trait. Bertrand and Duflo (2017) further argue that CT is advantageous for several reasons. First, by using fictitious applicants rather than physical applicants, researchers can create strict comparability across different applicants and make the minority trait the only difference between the applicants. Besides, CT circumvents the issue of auditors knowing the purpose of the study. Finally, the marginal cost of conducting a CT is
samples, more precise estimates, and the ability to analyze differential treatment from various angles and connect it to specific theories of discrimination.
There are potentially several disadvantages of using a CT. Carlsson et al. (2013) argue that correspondence methods without strong assumptions are unable to detect whether differences in callbacks across groups are due to taste-based or statistical discrimination. Bertrand and Duflo (2017) argue that one can separate between taste-based and statistical discrimination by comparing differential gaps between minority and non-minority applicants with weaker or stronger productivity signals on respective resumes. By including more productivity relevant information on applications and resumes, the average difference in unobservable characteristics will be reduced, and statistical discrimination. Furthermore, correspondence studies might be subject to ethical concerns. Researchers sending resumes and applications do so without the consent of the affected agents (e.g., employers, landlords, and sellers) and, as a result, might use up their time, which could be a scarce resource. Another ethical concern is that the decision-makers discover that the fictitious applicants declining offers are similar to real-life applicants, thereby making the real-life applicants appear less favorable.
Some of these methods are also used in conjunction with each other (Yinger, 1986; Ahmed et al., 2010; Zussman, 2013). In his study, Yinger (1986), used a combination of an audit study and regression analysis to detect ethnic discrimination in the US housing market. By gathering data from an audit study conducted in Boston in 1981, the author was able to determine via a regression analysis that the primary source of this discrimination was economic.
3.2 Ethnic Discrimination in the Housing and Labor Markets
As a result of the observations described in the subsection above, the authors were convinced that correspondence testing would probably prove to be the most time- and cost-efficient method for investigating discrimination in various markets. In addition, since the birth of the Internet, new innovative techniques to measure the magnitude of discrimination have emerged. Thus, this section intends to discuss the use of correspondence testing in the housing and labor market involving the new techniques that were not available in prior studies.
Bertrand and Mullainathan (2004) studied ethnic discrimination in the American labor market by using a correspondence test. Between July 2001 and January 2002, the authors faxed and emailed close to 5,000 resumes to approximately 1,300 employment ads. By randomly assigning
African-American and white-sounding names to the resumes, the authors could guarantee that any difference found was single-handedly based on the manipulation of the minority trait, i.e., the applicant’s name. The study presented three main findings. First, white-sounding names received 50 percent more callbacks. Second, the racial gap was uniform across employer size, industry, and occupation. Third, there was little evidence suggesting that employers inferred social class based on names. To conclude, the authors mention that difference in treatment based on ethnicity was evident in the American labor market. However, this study contains some weaknesses. First, it can only explain discrimination in the first stage of the hiring process. Second, the resume does not directly signal ethnicity, but indirectly through names, which raises two main concerns; (i) employers might not notice or recognize the ethnical content, and (ii) the results are not representative of the average African-American, as they might not have an ethnically distinct name. Finally, job advertisements are not the only form of finding work, as social connections could be an essential factor. As a result, it could have affected the results if employers and African-Americans relied more on social networks to obtain employees or jobs, respectively.
Carlsson & Rooth (2007) likewise conducted a correspondence test in order to measure ethnic discrimination in the Swedish labor market. The authors expanded their method by introducing a survey that targeted employers who invited applicants for an interview. The survey provided information such as work characteristics, employee composition, and information about the recruiter. The results revealed that Swedish-sounding names had a higher probability of receiving a callback from the recruiter than Middle Eastern-sounding names. The survey concluded that there was variability in the treatment of the applicants. Female recruiters were less likely to discriminate than male recruiters, and the extent of discrimination seemed to be reliant on the size of the firm. Small firms tended to discriminate more compared to large firms.
Ahmed and Hammarstedt (2008) investigated whether ethnic discrimination existed in the Swedish rental housing market by conducting an internet-based correspondence test. Three fictitious inquirers exposed their ethnicity and gender by their names while showing interest in rental residences. Two of the inquirers had Swedish sounding names, representing both genders, whereas the third had an Arab Muslim sounding male name. To measure discrimination, the authors compared the reply rate, invitation of viewing, or further contact received by the landlords for each fictitious inquirer. The results showed that landlords most preferred Swedish female
Ahmed et al. (2010) extend the method used by Ahmed and Hammarstedt (2008) by investigating to what extent additional information affected the discrimination towards Arab Muslim-sounding names in the Swedish rental housing market. Fictitious letters of interest were sent to landlords in which one group only described their interest in the property and where the other group provided additional information about their personal life and education. Responses from the landlord indicated that when no additional information was provided, Arab/Muslim-sounding names were less likely to receive a callback from the landlord than Swedish-Arab/Muslim-sounding names. When additional information was provided, the probability of receiving a callback improved for both groups, although the unexplained gap in response rates persisted.
3.3 Ethnic Discrimination by Public Authorities
Previous literature on ethnic discrimination by public authorities has been scarce, especially concerning public authorities in Sweden, as few authors have attempted to shed light on this matter. However, it is reasonable to believe that discrimination can occur within the public sector as well. In the following section, we will discuss relevant studies of discrimination by public authorities.
In the US and China, researchers have studied whether public authorities provide equal treatment across certain ethnic groups (see Distelhorst and Hou, 2014; Giulietti et al. 2019). Distelhorst and Hou (2014) investigated whether local municipalities provided equal treatment for Muslim-sounding names and names that did not expose the individual’s ethnicity or origin in China. The authors sent fictitious inquiries for each group to 258 local municipalities asking for assistance and requirements to be eligible for minimum livelihood benefits. From said fictitious inquiries, the obtained results revealed that local municipalities were 33 percentage points less likely to provide information and requirements to those citizens with Muslim-sounding names relative to the unmarked names. Muslim-sounding names also tended to receive shorter replies from the local municipalities relative to the unmarked group. Giulietti et al. (2019) used a similar methodology to examine ethnic discrimination across Whites and African-Americans in the US. The authors extended their study by including additional local public services (e.g., sheriff offices and school districts) to address the extent of ethnic discrimination and whether it varied among the public providers. Roughly 19,000 emails were sent out across six different local public services asking for support, by using different complexity to the questions asked. The emails were randomly assigned to either an African-American or white name. The results revealed that distinct
African-American names tended to receive fewer replies and a more negative cordial tone from the local public services relative to white names. Also, the complex questions had a higher response rate as compared to the simple questions. Three out of six public services were statistically significant at 1%, indicating unequal treatment across the two groups.
To date, only two studies have attempted to investigate ethnic discrimination by public authorities in Sweden (see Adman and Jansson, 2017 and Ahmed and Hammarstedt, 2019). Both studies examined whether individuals with Arabic-sounding names were discriminated against when contacting Swedish public authorities regarding preschool admission. Both studies implemented a correspondence test using fictitious parents with Arabic- and Swedish-sounding names, in which the respective parents asked all 290 municipalities in Sweden about preschool admissions. The methodological construction of their respective correspondence studies differed somewhat, as Adman and Jansson (2017) used two male and two female aliases to allow for a gendered dimension. In contrast, Ahmed and Hammarstedt (2019) used only two male aliases to achieve higher statistical power. There were some similarities as both studies used a randomized correspondence test, i.e., only one inquiry was sent to each municipality from one randomly assigned alias. The results from Adman and Jansson’s study showed that Arabic-sounding names, on average, received fewer replies, informative answers, extra information, and less welcoming answers. Although differences were found, the authors were unable to find any statistically significant difference. Similar results were found in Ahmed and Hammarstedt’s study. However, they found statistical differences when examining how many words each alias received and the extent to which personal salutations were received. From the results, the authors could ascertain a difference in treatment between Swedish and Arabic-sounding names by municipalities in Sweden.
4. Methodology
4.1 The Experiment
To study if municipalities provide differential treatment across different ethnic groups, we conducted a correspondence test in which we emailed every municipality in Sweden from two fictitious individuals inquiring about the SFI program. The design of this experiment is similar to that of both Adman and Jansson (2017), as well as Ahmed and Hammarstedt (2019). However, rather than gathering one observation per municipality, we attempt to gather two observations from each municipality, essentially doubling the number of observations compared to the studies
mentioned above. The reasoning behind collecting two observations per municipality is to achieve higher statistical power, which has been a problem previous authors have encountered in their studies (Lahey and Beasley, 2016). Due to the time constraint of this study, the length of this field experiment was limited to seven working days, between April 20 and April 29, 2020. However, previous studies suggest that most municipal responses were received back within one week, and therefore, seven working days were deemed sufficient to receive a substantial amount of responses. Answers received after the end of the study period were considered as ‘no-reply’. A complete list of contact information for the municipalities was collected from the Swedish Association of Local Authorities and Regions (SALAR), from which all 290 municipalities in Sweden were contacted via emails from both fictitious inquirers. Since the two fictitious inquirers sent emails to the same municipalities, there had to be some differentiation in how these emails were formulated while still requesting the same type of information, shown in the boxes below:
Further, for the experiment to go undetected by local public officials, the order in which each inquirer contacted each municipality along with the email template used, were randomly assigned. Each inquirer contacted 145 municipalities with both email templates during a two-day period, Monday and Tuesday, to eliminate possible interactions between the two fictitious inquirers. As a result, this structure allowed us to collect a total sample of 580 observations, in which both inquirers contributed to 290 observations.
Contacts with municipalities in Sweden can be taken through various channels, such as by phone, letters, or via the Internet. We decided that email correspondence was the most optimal choice of communication for this study. First, municipalities offer many of their services via their
Box 1. First email template Hello!
My name is [Name] and I recently moved to Sweden and I would like to learn Swedish. I am wondering about the SFI program, how does it work? How do I apply? When can I be expected to start?
Sincerely,
[Name]
Thanks, Name
Box 2. Second email template Hi!
My name is [Name] and I would like to start taking classes in Swedish as I have moved to Sweden. I heard about the SFI program and I am wondering how it works? When can you start and how do you apply for it?
Sincerely,
websites, where much information about their services can be found. Second, since the outbreak of the COVID-19 pandemic, contacts with municipalities have been conducted via email in an attempt to limit the spread of the virus. Third, by using email correspondence, government officials will only detect the ethnicity of the inquirer through their names. This allows us to make the two inquirers completely comparable and signal their respective ethnicities through their names only (Bertrand and Duflo, 2018). Both inquirers were given similar email addresses stating their given name, surname, and two random digits, in line with “givenname.surnameXX@gmail.com” (see Appendix A1). We chose the same email provider for both names to minimize potential differences across the two inquirers. Furthermore, as each inquirer was required to send 290 emails within two days, we wanted to eliminate the risk of either account being blocked due to spam or other mailing guidelines. Gmail offered users to send up to 500 emails within a 24-hour period and allowed us to schedule mailings, which made our experiment more efficient.
4.2 Choice of Ethnicities and Names
As mentioned in earlier subsections, previous literature has expressed that individuals with an Arabic background are especially subject to discrimination in various markets in Sweden. In both the housing and labor markets, individuals with an Arabic-background are at a disadvantage when applying for housing and employment (see Carlsson and Rooth, 2007 and Ahmed et al. 2010). This pattern can also be seen in contacts with government officials in Sweden as Arabic-sounding names have been proven to be treated differently than Swedish-sounding names (Adman and Jansson, 2017 and Ahmed and Hammarstedt, 2019). These studies have analyzed the difference between natives and individuals with an Arabic-sounding name, but few studies have attempted to analyze whether there are differences between immigrant groups. One immigrant group that is interesting to compare to are individuals from former Yugoslavia, in particular the Western-Balkan region. Studies suggest that these individuals have been able to assimilate into the Swedish labor market more successfully than many other immigrant groups since their arrival in the 1990s (Aslund et al., 2017; Ruist, 2018). Further, immigrants from Former Yugoslavia represent an immigrant group of substantial size. According to Statistics Sweden (2020), there were approximately 113 000 immigrants from this region residing in Sweden in 2019, making them the fourth-largest immigrant group behind Syria, Iraq, and Finland. However, to our knowledge, no
studies have attempted to analyze if immigrants from this region are subject to discrimination in Sweden.
Another critical factor in our correspondence test was to select names that indicated ethnicity and the male gender. For the Arabic-sounding inquirer, we selected the name Mahmoud Abdulrahman as it is a common Arabic name, which also clearly indicates that the inquirer is male. The name for the Western-Balkan-sounding inquirer was more challenging as names from Bosnia and Herzegovina, Croatia, and Serbia could be easily distinguishable as they can have a religious association depending on the country. Thus, it was essential to select a name that could be considered neutral for this region. After consultation with individuals with backgrounds from this region, we selected the name Alen Mandžukić. This name was considered fairly neutral and common in at least two out of the three countries, while simultaneously signaling that the inquirer was male.
4.3 Outcome Variables
To measure whether municipalities provided unequal treatment across the two immigrant groups, we had to choose a series of outcome variables that would measure the effort spent answering the respective inquirer. This was done to see if the government officials exhibited any prejudicial behavior. Therefore, we chose the following outcome variables; Reply, Salutation, Wordcount, Number of questions answered, Forwarding, and Assistance. These responses were coded using a coding scheme. ‘Reply’ was a binary variable denoting the value of 1 if the municipality replied and 0 otherwise, excluding automatically generated responses. This variable was chosen to observe if municipalities systematically replied to one inquirer but not the other. ‘Salutation’ was a binary variable taking on the value 1 for including the name of the inquirer and/or welcoming the inquirer to Sweden and 0 for neither. ‘Wordcount’ was a continuous variable measuring the number of words used in each reply. This variable was chosen to measure the amount of effort that was put into the replies. ‘Number of questions answered’ was measured on a scale from 0 to 3, indicating how many questions each municipality answered. ‘Forwarding’ was a binary variable taking on the value one if a local public official forwarded the email to the appropriate person or department and 0 if the inquirer was only given a link or contact information to the appropriate person or department. ‘Assistance’ was a binary variable denoting whether a local public official offered to assist the inquirer with the SFI application personally or not.
4.4 Econometric Specification
In the following section, our chosen regression model will be presented with a description of how our regressions will answer the research question. Since our dataset mainly contains binary variables, we conduct an OLS-regression to measure the systematic differences in treatment across the two inquirers. Furthermore, we divided all municipalities into clusters due to the study’s matching methodology, where each inquirer contacted the same municipality. This allows us to consider the correlation within the error-terms between the two inquirers.
Yi = α + δArabici + εi (4.1)
Where Yi denotes the outcome variable for individual i; α is a constant; Arabic is a dummy variable for individual i, indicating 1 if the inquirer is Arabic and 0 otherwise; εi is the error-term for individual i. Furthermore, 𝛿 captures the differences in treatment by the municipalities for the two fictitious inquirers. A positive δ-coefficient indicates that an inquirer having an Arabic-sounding name receives better treatment compared to the Balkan-sounding name or vice versa if the coefficient is negative.
4.5 Statistical Power
When using an experimental design, an important part is to figure out the minimum sample size needed to find statistically significant results given realistic effect size and a set power (Lahey and Beasley, 2018). To find a necessary sample size using power analysis, one would require information about the desired power, significance level, and the effect size. Issues can arise, however, when one would want to estimate the effect size across different groups. Generally, the effect size would be calculated from a pilot study, but it can also be generated from previous field experiments. If no such studies are available, researchers can use the default effect size of small, medium, or large based on the researcher’s beliefs about the effect size. If the effect size is expected to be small between two groups, then one would require a large set of observations and, vice versa, to find a statistically significant difference between the two groups. For our study, the maximum amount of observations is limited to 290 per inquirer, resulting in 580 observations in total. As a result, if differences in our outcome variables end up being relatively small between the
differences. Nevertheless, it is important to stress that a result indicating no statistically significant difference would be considered the desired outcome as the opposite would show that one group does appear to be treated worse than the other.
4.6 Robustness Check
To evaluate our results' credibility, we extended equation (4.1) to equation (4.2) by including control variables. See the following equation below, where Controlidenotes our selected control variables.
Yi = α + δArabici + βControli + εi (4.2)
By considering Controli within the regression equation, it allows us to appraise the robustness of
the experimental design performed in this study. Therefore, we observed how the coefficient ‘𝛿’ changed when control variables were considered. Minor variations of the coefficient are commonly implicated as evidence of structural validity, in other words, a credible experimental design or contrariwise, if the coefficient has major variations between the two regressions (Lu and White, 2014).
4.7 Ethical Considerations
As discussed above, correspondence tests are subject to ethical concerns, and so is this study. In most experimental research studies, participants should be informed of their participation (Adman and Jansson, 2017). However, a vital aspect of a correspondence test is that the participants are unaware that the experiment is taking place. Thus, all municipalities and local public officials will be anonymized as the main objective of this study is to analyze the overall trends in treatment and not the treatment by one single municipality or public official. Furthermore, in order not to use up an excess of a public official’s time, a brief message expressing gratitude and not wishing to take the matter further will be sent to each respondent.
5. Data
5.1 Descriptive Statistics
Table 1 presents the descriptive statistics for the email variables. For the ‘Reply’ variable, we were able to collect 580 observations, from which 77.9 percent of the municipalities provided either
inquirer with a response. Moreover, in 44.9 percent of the cases, which involved a reply, the inquirer also received a salutation. As for the ‘Wordcount’ variable, the table illustrates that the municipal replies included an average of 64.4 words. In said municipal replies, an average of 1.32 questions were answered out of the three questions posed by the inquirers. The number of observations for the variable ‘Forwarding’ is larger than the other variables because some municipalities automatically forwarded the emails to appropriate individuals or departments. However, these did not provide a reply within the time frame of our experiment. Nevertheless, automatic forwarding occurred 24.6 percent on average. Finally, for the variable ‘Assistance’, the table shows that about 22.2 percent of the local public officials personally offered to assist the inquirer with the application process or to gather further information about the SFI program.
Table 1. Descriptive statistics for the email variables
N Mean Standard deviation Minimum Maximum
Reply 580 .779 .415 0 1
Salutation 448 .449 .498 0 1
Wordcount 448 64.36 50.46 1 314
No. of questions answered 448 1.32 1.022 0 3
Forwarding 468 .242 .428 0 1
Assistance 448 .222 .416 0 1
6. Results
Table 2 presents the differences in means for the email variables between the Western-Balkan- and Arabic-sounding names. The results from this table show that the differences in their respective mean values are relatively similar. An important aspect to consider regarding our results is that the number of observations differs between the ‘Reply’ variable and the remaining outcome variables. This is because we did not omit non-replies. Instead, we coded them as a 0 in the reply variable, which allowed us to obtain a mean value for the number of replies for each inquirer. The remaining outcome variables have a lower number of observations because they are dependent on whether the inquirer in question received a reply, which happened in roughly 78 percent of the cases.
Table 2. Differences in means for the email variables, between Western-Balkan- and Arabic-sounding names N Arabic Western-Balkan Difference Reply 580 .7828 .7759 0.007 Salutation 448 .4493 .4480 0.001 Wordcount 448 63.43 65.32 -1.890
No. of questions answered 448 1.225 1.412 -0.187
Forwarding 468 .2143 .2696 -0.055
Assistance 448 .1982 .2455 -0.047
For the ‘Reply’ variable, it can be illustrated that the Arabic-sounding name on average received more municipal replies, although this difference was minimal. The difference in mean values for the variable ‘Salutation’ was even smaller, although the Arabic-sounding name received more salutations than the Western-Balkan-sounding name. As for the variable ‘Wordcount’, again, the differences between the two names were minimal as the table shows that the Western-Balkan name on average received about two words more in the replies from the municipalities. Along with the variable ‘Wordcount’, the remaining variables ‘Number of questions answered’, ‘Forwarding’, and ‘Assistance’ all indicated the differences in means to be at a disadvantage for the Arabic-sounding name. Further, these differences are substantially larger than those observed in the first three outcome variables. For the variable ‘Number of questions answered’, the table illustrates that municipalities answered fewer questions for an Arabic-sounding name on average. As for the variable ‘Forwarding’, the Arabic-sounding name received five percentage points fewer automatic forwards to the appropriate individual or department at the municipality than the Western-Balkan-sounding name. Finally, for the variable ‘Assistance’, the Arabic-Western-Balkan-sounding name received personal assistance from the local public official in 5 percentage points fewer instances than the Western-Balkan-sounding name.
Table 3. Effects of Western-Balkan/Arabic-sounding name on the outcome variables without
control variables
(1) (2) (3) (4) (5) (6)
Reply Salutation Wordcount No. of
questions answered Forwarding Assistance Arabic 0.007 0.001 -1.890 -0.187*** -0.055** -0.047* (0.029) (0.033) (3.164) (0.065) (0.028) (0.027) Constant 0.776*** 0.448*** 65.321*** 1.412*** 0.270*** 0.245*** (0.025) (0.034) (3.356) (0.069) (0.029) (0.029) Observations 580 448 448 448 468 448 R-squared 0.000 0.000 0.000 0.008 0.004 0.003
Clustered standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
6.1 Regression Results
Table 3 presents the effects of Western-Balkan- and Arabic-sounding names on the outcome variables without using any control variables. For first three columns, ‘Reply’, ‘Salutation’, and ‘Wordcount’, it is clear that the difference between the Western-Balkan- and Arabic-sounding names is insignificant. The Arabic-sounding name received 0.007 percentage points more replies, though with a high clustered standard error of 0.029. Moreover, Arabic-sounding names received 0.001 percentage points more salutations and, on average, two words less than the Western-Balkan-sounding name. For the variable ‘Number of questions answered’ it can be illustrated that the Arabic-sounding name on average received roughly 0.2 less answered questions than the Western-Balkan-sounding name, which is statistically significant at the 1-percent level. The final two variables, ‘Forwarding’ and ‘Assistance’, are significant at the 5-percent level and the 10-percent level, respectively. This means that the Arabic-sounding name received 0.055 10-percentage points fewer automatic forwards from government officials to the appropriate person or municipal department. Furthermore, the regression estimates reveal that an Arabic-sounding name received 0.047 percentage points fewer instances in which a local public official offered to assist personally in the application process or gathering of further information.
The results given by our robustness check can be found in Table A2 (see Appendix). When control variables were considered within the regression, we obtained minor impacts on our
changed from positive to negative, indicating that having a Western-Balkan sounding name received 0.007 percentage points more replies and 0.025 percentage points more salutation than the Arabic-sounding name. However, the estimates of the standard error preciseness changed, were ‘Reply’ decreases from 0.029 to 0.007, whereas ‘Salutation’ arises from 0.033 to 0.036. Though, both outcome variables remained insignificant at any given significance level. Moreover, ‘Wordcount’ slightly increased by 0,112 words with a slightly decreased standard error from 3.164 to 3.086. These results do not affect the earlier interpretation of the variable’s outcome, Arabic- names received, on average, roughly two words less than Western-Balkan sounding names. Interestingly, the outcome variable ‘No. of questions answered’ dropped out of the 1-percent significance level, instead of being significant at the 5-percent level due to the increasing standard errors from 0.065 to 0.072. However, the interpretation of the variable outcome remained, as when control variables were not included in the regression. Both ‘Forwarding’ and ‘Assistance’ became insignificant at any of the given significance levels.
In Table 4, we included the control variable ‘Same Respondent’ to Equation 4.1. This variable was coded during the experiment as it was coded as a one if the responder was the same and 0 if there were different municipal responders. The table shows that by including control for both inquirers having the same responder decreased the differences in the outcome variables.
Table 4. Effects of Western-Balkan/Arabic-sounding name on the outcome variables with the
control variable ‘Same Respondent’
(1) (2) (3) (4) (5) (6)
Reply Salutation Wordcount No. of questions
answered Forwarding Assistance Arabic 0.006 -0.026 -1.754 -0.170** -0.013 -0.026 (0.010) (0.033) (2.782) (0.065) (0.029) (0.027) Same respondent 0.027* (0.015) -0.085 (0.070) 2.441 (7.291) 0.346** (0.141) -0.011 (0.052) 0.116** (0.050) Constant 0.970*** 0.509*** 61.628*** 1.196*** 0.190*** 0.151*** (0.017) (0.057) (6.490) (0.117) (0.043) (0.036) Observations 346 343 343 343 341 342 R-squared 0.019 0.007 0.001 0.031 0.000 0.018
Clustered standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
Column 1 shows the results for the ‘Reply’ variable, which is about the same (0.006 compared to 0.007). Worth mentioning, however, is that the clustered standard errors become smaller in Table A3, which indicates that the estimates are more precise, but this estimate is not statistically significant. Column 2 shows in contrast to Table 3, that the estimate for our dummy variable ‘Arabic’ now is smaller but that the standard errors now remain precisely the same. Although somewhat of a drop in the estimate can be seen, this estimate is not statistically significant. From Table 3, it was illustrated that the variables ‘Number of questions answered’, ‘Forwarding’, and ‘Assistance’ all were statistically significant at different confidence levels. However, after controlling for the same respondent, these estimates became smaller, and only the variable ‘Number of questions answered’ remained statistically significant.
7. Discussion
The main objective of this study was to identify if Swedish municipalities provided different treatment for Western-Balkan and Arabic-sounding names when said immigrant groups inquire about information regarding the SFI program.
The main results of this study suggest that government officials provide differential treatment in favor of a Western-Balkan-sounding name but that differences in the outcome variables are relatively small. In this case, differences in three out of six outcome variables showed statistically significant results in which the Arabic-sounding name received less answered questions, automatic forwards, and assistance from the government official. The findings of this study are in line with both Adman and Jansson (2017) as well as Ahmed and Hammarstedt (2019). They found that inquirers with Arabic-sounding names received poorer treatment from government officials compared to those with Swedish-sounding names. This study tends to support and expand the results of previous findings, as our findings suggest that people with Arabic-sounding names do appear to receive even poorer treatment from government officials than Western-Balkans immigrants. However, it is worth mentioning that the differences in treatment found in this study are relatively small. The ‘Reply’ and ‘Salutation’ variables clearly showed that the differences were close to zero since no significant difference was found. Furthermore, these estimates were not statistically significant even after adding in control variables. Even though we managed to collect double the amount of observations for the ‘Reply’ variable, this did not yield a statistically significant result. Lahey and Beasley (2016) mentioned that small differences between two groups
need a large number of observations to find a statistically significant result, and we believe that this study suffered from this phenomenon. As the number of observations was limited to 580 (i.e., 290 per inquirer), the small difference we found in the ‘Reply’ variable was not going to be statistically significant and may also explain the high standard errors of the coefficients. Furthermore, an explanation for the relatively small difference could be attributed to the length of the experiment, which might have presented two possible problems. First, since the inquirers sent out emails to the same municipality one day apart, the municipalities might have understood that an experiment was taking place and thus answered with the same email to both inquirers. Second, if the case is that the municipalities did not detect the experiment, it would seem that a template could have been used to answer any questions regarding the SFI program. In addition, since the experiment was only conducted over seven working days, any response after that point was considered a “no-reply”, which resulted in fewer observations than expected. This most likely influenced the statistical power of our results, as the length of the experiment was probably too short to find any actual differences in how the officials responded to the inquirers.
When adding control variables, we could see that the estimates for our dummy variable changed for all outcome variables, however, these changes were relatively small. As argued by Lu and White (2014), minor changes to coefficient estimates between a regression with and without control variables are seen as a sign of structural validity. This notion explains that our experimental design was robust from the start, and that control variables would not have affected our results very much. However, when adding control variables, only the variable ‘Number of questions answered’ remained statistically significant as both ‘Forwarding’ and ‘Assistance’ became non-significant. Notably, however, the standard errors in Table A2, columns 5 and 6, rises from 0.028 and 0.027 to 0.031 and 0.031, respectively, making these estimates less precise than those presented in Table 3. Moreover, when adding control variables, the coefficient estimates for the dummy variables become negative, indicating that the Arabic-sounding name received inferior treatment compared to the Western-Balkan-sounding name even though only one variable proved to be statistically significant.
7.1 Taste-Based or Statistical Discrimination?
Based on the existing theory of discrimination, our results hardly address whether our findings of unequal treatment are related to taste-based or statistical discrimination. As mentioned in the
theoretical framework, we believe that statistical discrimination is less applicable to this study’s subject. This is because we cannot determine whether the respondents in each municipality provide more or less information by using the inquirers’ observed characteristics as a proxy for the unobserved. By introducing the control variable ‘Same Respondent’, we were able to directly observe whether the same respondent provided different answers for each inquirer. However, according to Table 4, our results suggested that when the same respondent is considered, the differences in treatment across the two inquirers actually decreased. Therefore, these results may indicate that our findings are not related to taste-based discrimination from the municipalities since we were unable to observe its existence.
However, we encountered a high degree of identical answers when both inquirers had the same respondent, which reduced the experiment’s probability of finding evidence that points towards taste-based discrimination. As Becker (1957) explains in his theory of taste-based discrimination, an individual with prejudiced beliefs about a certain ethnic group should act as if there is a direct cost of associating with that group. Hence, in our case, this should have been signaled through the differences in outcome variables rising if it existed, however, we found no evidence of this phenomenon. The theory of taste-based discrimination does not seem to be the main cause of differential treatment. Moreover, the unexplained differences across the two inquirers may have been affected by other variables that were not considered in this experiment.
Carlsson et al. (2013) mentioned that correspondence methods with weak assumptions would be unable to detect if differences in callbacks across groups are due to taste-based or statistical discrimination. However, Bertrand and Duflo (2017) argued that one could separate between taste-based and statistical discrimination by adding weaker and stronger productivity signals on resumes, which has been conducted in studies of ethnic discrimination in Sweden (Ahmed et al., 2010). What differs in our study is that we were limited in the number of observations that we could gather in a limited amount of time. Adding weaker and stronger productivity signals in the emails would have resulted in our having to use four different inquirers, meaning fewer observations per inquirer would be obtained. Therefore, we would have suffered from low statistical power and fewer statistically significant results, similar to that of Adman and Jansson (2017). It might also explain why we faced difficulties in determining if the differences found in our results were due to taste-based or statistical discrimination.
8. Conclusion
This is the first correspondence study that attempted to examine the extent to which the Swedish municipalities might be providing differential treatment across two ethnic groups. Our results conclude that there appear to be statistically significant differences in how the municipalities treated our two fictitious inquirers. When confronted with an Arabic-sounding name, the municipalities provided fewer answers to the questions asked and less additional assistance as compared to a person with a Western-Balkan name. These findings suggest that Arabic names are more disfavored relative to Western-Balkan names when contacting local public authorities. However, we found no statistically significant difference when comparing the reply rates nor the municipalities’ kindliness towards each inquirer.
The results of this study suggest that there are possible practical implications for immigrants with an Arabic background, as studies indicate that language proficiency correlates with the labor market success of immigrants in Sweden (Aslund and Rooth, 2006). As the SFI program is one possible way to learn about the Swedish language and culture, limited access to the said program could have implications for the integration of immigrants with an Arabic background. Moreover, since the municipalities act according to the Principle of Equality and the Education Act (see, Swedish Code of Statutes, 2017:725; Swedish Code of Statutes, 2010:800), our statistically significant differences are of even more interest.
8.1 Limitations and Future Research
It should be discussed that the details of our experimental design might have affected the results of this study. First, the length of our experiment was rather short compared to similar empirical research studies within the same field. As discussed above, this might have affected the statistical power of our results. This suggests that further research in this area should be conducted over a more extended period. Further, since we waited only one day between the mailings, the government officials might have detected the experiment. Therefore, another suggestion to future research studies is to consider a longer time-period between mailings to reduce the possibility that officials would detect the experiment. Second, distinguishing between taste-based discrimination and statistical discrimination proved to be a difficult task in this study. More sophisticated methods for detecting either type of discrimination by municipalities would be needed. For example, more specific data on how many Arabs and Western-Balkans end up dropping out of SFI may have
allowed us to reach more concrete conclusions. They would have provided us with further evidence as to whether or not statistical discrimination occurs within some of the municipalities. The respondent may provide less information to one inquirer due to the high drop-out rate among the ethnic group. Third, the outcome variable ‘Number of questions answered’ might have been difficult to interpret. It would be interesting to examine, in future studies, to what degree each question is answered, and the quality of responses provided by the officials.
Presently, numerous authors have already provided empirical evidence of discrimination in various markets. Moreover, continuing research within the field of discrimination may allow researchers to plot the fundamental factors of its existence and why certain ethnicities are more exposed to discrimination as compared to other ethnic groups. Until today, existing literature has mainly targeted experiments that measure the differences between natives and immigrant groups within a certain field and, therefore, neglect differences between minorities. Therefore, future research within the area of discrimination should allow for a broader perspective of the situation by comparing the utmost exposed minorities with other minorities in various markets. By introducing new perspectives on the possible existence of discrimination by municipal officials, we believe that our limited research has contributed to existing knowledge on this subject. It may further play an essential role in the understanding of its existence and how we may proceed to prevent it from occurring.
Appendix
Table A1: Descriptive statistics for the control variables
N Mean Standard deviation Minimum Maximum
Foreign-born employees in
the municipality (%) 288 15.69 6.466 6.46 42.84
Foreign born individuals in
Municipality (%) 290 19.01 7.56 8.18 53.85
Did not finish SFI within two
years (%) 254 15.94 9.62 0 72.22
Same respondent 346 0.68 0.47 0 1
Table A2: Effects of Western-Balkan/Arabic-sounding name on the outcome variables with
control variables
(1) (2) (3) (4) (5) (6)
Reply Salutation Wordcount No. of questions
answered Forwarding Assistance
Arabic -0.007 -0.025 -2.002 -0.180** -0.027 -0.034 (0.007) (0.036) (3.086) (0.072) (0.031) (0.031) Foreign-born municipal employees -0.001 -0.027** 1.506 0.014 0.007 -0.002 (0.001) (0.011) (1.155) (0.024) (0.009) (0.009) Foreign-born individuals in the municipality 0.001 (0.001) 0.028*** (0.009) -0.955 (0.949) -0.014 (0.018) 0.001 (0.008) 0.007 (0.008) Did not finish SFI
within two years
-0.001 (0.001) -0.005 (0.004) -0.249 (0.430) -0.000 (0.009) 0.001 (0.003) -0.008** (0.003) Same respondent 0.008 -0.086 0.546 0.332** 0.037 0.091* (0.008) (0.077) (8.428) (0.153) (0.061) (0.053) Constant 1.000*** 0.476*** 61.721*** 1.335*** 0.020 0.197** (0.003) (0.125) (13.677) (0.233) (0.104) (0.084) Observations 298 297 297 297 294 297 R-squared 0.024 0.054 0.009 0.035 0.024 0.041
Clustered standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
Box A1. Email addresses
Mahmoud Abdulrahman mahmoud.abdulrahman67@gmail.com
