Introduction
Today, public services are frequently offered in markets with both public and private providers. These markets are believed to increase cost efficiency and provide citizens with more choice opportunities. Critics, on the other hand, argue that market-based incentives may lead to cream skimming (this argument is common in the public debate; for academic examples, see e.g., Jilke et al., Reference Jilke, Van Dooren and Rys2018; West et al., Reference West, Ingram and Hind2006; Lacireno-Paquet et al., Reference Lacireno-Paquet, Holyoke, Moser and Henig2002; Six, Reference Six2003), i.e., the intentional selection or avoidance of certain clients in order to reach program goals (cf. Lipsky, Reference Lipsky1983). If this criticism is correct, then the ideal of equal treatment – a cornerstone of public sector service – is no longer upheld. Such a state of affairs would indeed be challenging for proponents of market-based reforms.
What, then, is known about the presence of discrimination in marketized welfare services, and what do we know about the potential drivers of discrimination in such markets? In fact, despite the rapid increase in private actors in welfare sectors all over the world, experimental research on the matter is conspicuously limited. Field experiments are perhaps the most reliable method of investigating discrimination (cf. Pager and Shepherd, Reference Pager and Shepherd2008; Quillian, Reference Quillian2006). However, only a small number of published experiments on the subject exist. Interestingly, the findings are mixed, not always indicating that private providers are more likely than public providers to discriminate (Jilke et al., Reference Jilke, Van Dooren and Rys2018; Milkman et al., Reference Milkman, Akinola and Chugh2015; Brown and Hilbig, Reference Brown and Hilbig2021; Oberfield and Incantalupo, Reference Oberfield and Incantalupo2021).
However, prior field experiments have several limitations. First and foremost, the effect of competition levels between providers has not been studied. Second, we know of only one experiment analyzing discrimination among publicly financed for-profit providers, e.g., providers with the strongest incentives to cream skim (Jilke et al., Reference Jilke, Van Dooren and Rys2018). Finally, previous experiments have ignored socioeconomic status (SES) discrimination. This may be of importance since service providers could be expected to cream skim clients based on SES (see more on our contributions in the next section).
Pursuing a twofold purpose, our article contributes on all of these points. First, we study whether discrimination is more present when local competition is stronger. This purpose is new to the literature, focusing on a previously neglected aspect of marketization. Second, we investigate differences in discrimination levels between different types of providers, including for-profit providers. Using an email correspondence experiment, we study the email replies from 3,430 Swedish elementary school principals who were contacted by fictional parents interested in placing their children at each principal’s school. The ethnicities of the fictional parents were randomized (Swedish- or Arabic-sounding names), as were their professions (low or high SES). The study covers all school principals working at a unique school in Sweden. A notable number of the principals are employed at publicly financed for-profit schools, allowing a nuanced picture of the importance of various types of providers. Moreover, high-quality registry data were added to the experiment, allowing the investigation of competition effects, a new approach in the research field.
Discrimination indicators include whether the emails from our fictional parents received responses, how many of the posed questions were answered, and to what extent emails were friendly and promoted the school in question. The latter qualitative dimension we refer to as promotion is new to the literature. Promotion takes place when public officials selectively highlight positive service qualities and provide information about service availability, which could be used to cream skim clients.
We are agnostic about the results of our study. Client discrimination may very well occur unrelated to the competition level or type of provider. For example, as argued by Lipsky (Reference Lipsky1983), public officials tend to cream skim and discriminate in general to manage resource scarcity. Only further empirical investigation can reveal which of these theoretical perspectives is empirically accurate.
Our study focuses on schools in the Scandinavian welfare state of Sweden, i.e., a notably different context than the countries studied in previous experimental research of this kind (Belgium and the US; see references below). Moreover, the elements and levels of marketization in the Swedish case currently resemble how public services are offered in many sectors in the Western world, making more general conclusions possible (see the institutional context section).
The findings of our study may have important societal implications. The Swedish school choice reform has led to increased school segregation in terms of ethnicity and SES in Sweden (Holmlund et al., Reference Holmlund, Sjögren and Öckert2019), and students with parents with higher education are much more likely to attend private schools (cf. Table 17 in Appendix J). While schools are not allowed to discriminate against students, principals may influence the applicant pool via the kinds of messages and information sent to potential clients. For example, being informed about admission processes and quality differences between schools could affect school choice decisions (Hastings and Weinstein, Reference Hastings and Weinstein2008; Hoxby and Turner, Reference Hoxby and Turner2013; Kapor et al., Reference Kapor, Neilson and Zimmerman2020). If school segregation is partly a result of differences between schools in such ‘informal discrimination’, then it could have adverse effects not only for the affected students but for social mobility and social cohesion in society more generally. School segregation prevents important intergroup contact from taking place and may create hostility between groups (cf. Pettigrew and Tropp, Reference Pettigrew and Tropp2006). Furthermore, the concentration of less affluent students in public schools could reinforce existing social inequalities. Students attending private schools have higher grades and national test scores in Sweden than those attending public schools after adjusting for student composition. While it is not clear whether these differences are a result of grading or school quality differences (Holmlund et al., Reference Holmlund, Sjögren and Öckert2019), the higher scores in private schools could potentially lead to better life and career chances among their students.
Prior Empirical Research and Our Contributions
The last decade has seen a growing number of empirical studies of public officials’ discrimination of clients. Discrimination based on ethnicity has been particularly acknowledged and here refers to unequal treatment based on physical characteristics (e.g., skin/hair) or cultural factors (e.g., name or language). Field experiments have become the preferred design in this literature since they allow a more objective assessment of the extent to which discrimination takes place and do not suffer from unrepresentative samples or social desirability bias (cf. Pager and Shepherd, Reference Pager and Shepherd2008; Quillian, Reference Quillian2006). They also make it possible to capture discrimination in real-world contexts. Most existing studies are based on correspondence experiments, where emails from fictional persons of varying ethnicities are sent to public officials. In the US, clear signs of ethnic discrimination have been found, even in terms of response rates (for a meta-study focusing especially on the US, see Costa, Reference Costa2017). Research in Europe also reports biased client treatment, although discrimination levels seem to be lower and primarily concern the quality and friendliness of responses but not response rates (see, for example, Hemker and Rink, Reference Hemker and Rink2017; Olsen et al., Reference Olsen, Kyhse-Andersen and Moynihan2020; Larsson Taghizadeh, Reference Larsson Taghizadeh2022a).
Clearly, ethnic discrimination against clients occurs among public officials in Western democracies. Thus, a natural next step is to examine whether some officials or organizations are more likely to discriminate than others. In contrast to our knowledge on the prevalence of discrimination, much less is known about the contextual factors that can explain differences in discrimination against clients.
One important contextual factor that has repeatedly been suggested to cause biased client treatment is marketization. Marketization here refers to a process where market mechanisms such as competition, private provision and economic incentives are introduced into the public sector to make it more economically efficient (cf. Pollitt and Bouckhaert, Reference Pollitt and Bouckhaert2011). Critics of marketization argue that market-based incentives may lead to “cream skimming” defined as the intentional selection – or avoidance – of certain groups in an effort to reach program goals or work in a cost-efficient manner (cf. Lipsky Reference Lipsky1983, 107). However, the experimental discrimination literature does not reveal whether discrimination is more common in marketized welfare services or what the potential drivers of biased treatment in such sectors may be. It almost exclusively concerns public services operating as public monopolies and/or within less competitive environments. Private providers are largely missing, and the effects of different levels of competition between providers are not investigated. While there exists a relatively large non-experimental and quasi-experimental literature on marketization and cream-skimming (see e.g., Zimmer and Guarino, Reference Zimmer and Guarino2013; Lacireno-Paquet et al., Reference Lacireno-Paquet, Holyoke, Moser and Henig2002; West et al., Reference West, Ingram and Hind2006), more field experiments are needed, as they generally allow a more objective assessment of the extent to which discrimination takes place.
We identify only four published discrimination experiments on public and private welfare services operating within competitive environments (Jilke et al., Reference Jilke, Van Dooren and Rys2018; Milkman et al., Reference Milkman, Akinola and Chugh2015; Brown and Hilbig, Reference Brown and Hilbig2021; Oberfield and Incantalupo, Reference Oberfield and Incantalupo2021). All but one of these studies (Jilke et al., Reference Jilke, Van Dooren and Rys2018) focus on the US and only two of them cover publicly financed services, reducing their generalizability outside these areas. In regard to ethnicity, the results of these studies are mixed. Both Jilke et al. (Reference Jilke, Van Dooren and Rys2018) and Milkman et al. (Reference Milkman, Akinola and Chugh2015) find more bias among private than public providers when studying Belgian elderly care and US universities, respectively. However, Brown and Hilbig (Reference Brown and Hilbig2021) and Oberfield and Incantalupo (Reference Oberfield and Incantalupo2021) find no such differences when analyzing colleges and elementary schools in the US. In regard to other discrimination forms, Brown and Hilbig (Reference Brown and Hilbig2021) find public-private differences in discrimination against students with criminal records.
Due to the limited number of studies, several aspects of marketization have not been studied. We seek to address these limitations. First and foremost, the effect of competition levels between providers has not been studied. This is a crucial aspect of marketization (cf. Pollitt and Bouckhaert, Reference Pollitt and Bouckhaert2011) but has not been taken into account in the experimental literature on discrimination. Second, publicly financed for-profit providers have largely been neglected and are therefore included in our study. For-profit providers should have stronger incentives to select applicants who are more profitable and avoid those they regard as costly, but we know of only one experiment on discrimination explicitly investigating such providers (Jilke et al., Reference Jilke, Van Dooren and Rys2018).
Third, when evaluating the occurrence of biased treatment, prior research has not included treatments based on a potentially crucial characteristic: SES. Discrimination based on SES here refers to unequal treatment based on an individual’s education, occupation and/or income. SES is included in our study, since public officials are likely to cream skim based on this client characteristic and not only on ethnicity. To give an example from the education sector, principals should have strong incentives to attract students with highly educated parents because such students are probably seen as less resource-demanding. Furthermore, these students seem more likely to raise average grades and to increase the school’s reputation. Hence, SES could send clear signals on future costs, and service providers may be particularly prone to cream skim based on this characteristic.
Fourth, with two exceptions (Jilke et al., Reference Jilke, Van Dooren and Rys2018; Brown and Hilbig, Reference Brown and Hilbig2021), prior studies investigate only to what extent client emails are replied to. However, the larger literature on discrimination among public officials in Europe mainly finds discrimination in the quality and friendliness of replies (e.g., Hemker and Rink, Reference Hemker and Rink2017; Olsen et al., Reference Olsen, Kyhse-Andersen and Moynihan2020). Although hardly investigated before, this tendency may very well occur in competitive markets, too. In addition, when investigating cream-skimming among providers, we argue that it is particularly important to analyze a new qualitative aspect we refer to as promotion. This denotes when public officials highlight positive service qualities and inform potential clients about service availability. This should be a highly relevant communicational aspect, as public officials may cream skim by selectively promoting their own service and thereby increasing certain clients’ willingness to use that service. In our case, principals may highlight positive aspects of the schools, e.g., regarding pedagogic quality or the physical school environment, to certain clients and selectively inform certain parents about free slots. As a consequence, parents’ willingness to apply to the school in question may be affected.
Theoretical Expectations and Mechanisms
New public management (NPM) reforms, with their emphasis on competition and measurable performance targets, have changed the conditions for all types of providers in many societal sectors (cf. Bohte and Meier, Reference Bohte and Meier2000; Bevan and Hood, Reference Bevan and Hood2006; Considine et al., Reference Considine, O’Sullivan, McGann and Nguyen2020). There are several reasons why discrimination and cream skimming tendencies may increase as a result of the introduction of competition. With more competition, providers face greater pressure to decrease costs (cf. Le Grand and Bartlett, Reference Le Grand and Bartlett1993; Savas, Reference Savas2000). Furthermore, the risk of closure due to poor finances or a bad reputation is likely to be higher. While competition over more ‘average clients’ should be intensified, attracting less resource-demanding clients could potentially be even more beneficial when providers are subject to competition. Such clients may lower costs and improve the provider’s reputation and therefore decrease economic risks, often associated with high competition. Finally, the broad use of measurable performance targets causes public officials to focus, sometimes almost exclusively, on measured tasks (cf. Bevan and Hood, Reference Bevan and Hood2006). As a consequence, more discrimination might follow among all sorts of providers, regardless of the type of ownership involved (see, e.g., Epp et al., Reference Epp, Maynard-Moody and Haider-Markel2014). An example from the school sector is the common use of average grades as a quality indicator, which could induce all kinds of schools to engage in cream skimming when competition is substantial.
The incentives to cream skim should be particularly strong in markets where providers are not fully compensated for recruiting costly clients (such as the Swedish school market; more on this is provided below). Incentives should also be strong in markets where clients make their choices based on quality indicators that are not adjusted for client composition (e.g., grades, again, as in Sweden). However, even in systems where such objective quality measurements are available, using the school example, families could still see an intrinsic value in placing their child in a school with a certain student composition (high SES, few immigrants).
On the other hand, there are also reasons why we should not expect to find a relationship between competition and discrimination. First, a high level of competition could result in providers also competing over more costly clients, which should decrease discrimination. Second, as argued by Lipsky (Reference Lipsky1983), public officials in general are likely to use cream skimming. It is one of several coping mechanisms often used to handle a high workload. Discrimination could then follow when staffing and budgetary resources are scarce, regardless of competition levels. Finally, if discrimination is primarily caused by discriminatory attitudes that are not largely affected/activated by the organizational context in which officials operate, then we should not expect to find a causal relationship between competition and any form of discrimination. Biased treatment in general social situations – i.e., not particularly focusing on public officials and clients but on communication between individuals in general – has been related to taste-based motives that are both conscious and unconscious (i.e., motives outside the discriminator’s awareness; Greenwald and Banaji, Reference Greenwald and Banaji1995). Hence, a principal may unknowingly discriminate against, e.g., individuals from the Middle East due to subconscious negative feelings toward individuals belonging to that group, unrelated to the competition levels her/his school is exposed to.
We now move on to our theoretical expectations regarding our second research question. As argued by several scholars, there are reasonable reasons to expect private providers to discriminate more than public ones. Translated to the context of our study, private schools (both for- and non-profit) are generally more vulnerable than public schools. Owners do not necessarily protect them, e.g., if they have economic problems or suffer from a bad reputation. Hence, private schools face a higher risk of closure/bankruptcy. By focusing their recruitment efforts on self-perpetuating students from high SES families, private schools can lower their costs for teachers and hence economic risks in general. Such students also automatically improve school results (in terms of grades) and school reputation, which makes future recruitment of more students easier. Everything being equal, from a rational choice perspective, incentives for cream skimming should therefore be stronger among private providers, regardless of whether they are for-profit or not. For-profit providers should have the strongest incentives to adopt cream skimming, however, as their owners can obtain a direct financial gain from not attracting costly or low-performing clients who may damage the provider’s reputation.
As with competition, however, there are also sound reasons to expect similar levels of discrimination among all sorts of providers. High workload and coping strategies might be just as frequent among public actors as private for-profit ones. Moreover, taste-based discrimination may of course be present among publicly employed officials and not just among private or even for-profit officials. Furthermore, public providers may also have strong incentives to discriminate against less affluent clients due to the overrepresentation of such clients using their services (cf. Table 17 in Appendix J).
Institutional Context
Sweden is known as a well-functioning, egalitarian and tolerant democracy. However, while the country’s economic equality and earnings/education mobility are still high according to OECD standards, they have decreased rapidly in recent decades (OECD, 2015; 2018). Following decades of large-scale immigration, today, the population is rather diverse, ethnically, linguistically and culturally. Approximately 19 per cent is foreign-born, with a substantial share with a background in the Middle East (approximately four per cent of the Swedish population was born in Syria, Iraq, or Somalia alone; Statistics Sweden, 2019). In general, Swedish public opinion shows comparatively positive views on ethnic minorities (World Value Survey, 2014). However, the right-wing populist party the Swedish Democrats has grown to be the third largest, with 17.5 per cent of the votes in the 2018 election. It mobilizes particularly against immigrants with a background in the Middle East, and the general discourse surrounding this group has hardened. Moreover, immigrants from this part of the world have been found to be discriminated against in both the labor and housing markets and, to some extent, in interactions with public officials and politicians (e.g., Bursell, Reference Bursell2014; Larsson Taghizadeh et al., Reference Larsson Taghizadeh, Åström and Adman2022). In sum, from an SES as well as ethnic and political perspective, Sweden might be less divergent than its reputation holds.
Regarding the school sector, a comprehensive school choice reform in the 1990s radically changed the Swedish school system. Today’s relatively high frequency of Swedish private schools was one consequence (cf. Table 17 in Appendix J). The proportion of students attending public schools in Sweden is still close to the OECD mean, however (OECD, 2017). What is somewhat unusual about this country is the existence of publicly financed private for-profit schools that are allowed to return profits to owners. Besides Sweden, we know of the existence of such schools in the US and Estonia (cf. OECD, 2017). However, in other sectors, publicly financed for-profit providers are internationally common, e.g., healthcare and elderly care (Välfärdsutredningen, 2016; Hoppania et al., Reference Hoppania, Karsio, Näre, Vaittinen and Zechner2022). Hence, findings based on the Swedish school sector are of potential relevance both to other countries and to other sectors, although it is difficult to know more precisely to what extent our results are generalizable in this sense.
In Sweden, municipalities (local governments) have administrative responsibility for organizing and financing schools (therefore, competition is measured at this level, c.f. below). Parents can freely choose between schools in their municipality and are encouraged to actively do so. There are public schools managed by the municipalities, private schools run by for-profit businesses, and non-profit private schools run by associations and foundations. All schools are obliged to follow the same laws and to adhere to the Swedish curriculum for compulsory schools. Forty-three per cent of Swedish municipalities do not compensate for low-SES and immigrant students; in the municipalities that do have compensatory systems, only a relatively small part of the school budget is redistributed (SKL, 2018). A voucher system is used. Hence, none of the students have to pay for their education. If a private school is oversubscribed, it can choose students based on proximity to the school, waiting lists (by date of application), and/or priority for children whose older siblings are already enrolled. The same criteria apply to public schools, although students are always guaranteed a slot in the public school nearest to their home.
Method
In our correspondence experiment, every elementary school principal governing a unique school in Sweden was randomly contacted via emails from fictional parents (N = 3430).Footnote 1 Emails were sent out in several waves between 11 and 14 January 2020.Footnote 2 A factorial design was employed based on ethnicity, SES, and genderFootnote 3 . Factorial designs are a standard approach in the field (cf. Brown and Hilbig, Reference Brown and Hilbig2021; Oberfield and Incantalupo, Reference Oberfield and Incantalupo2021) and are based on the basic idea that experimental units take on all possible combinations of the levels of the factors of interest. Accordingly, the school principals’ email addresses were randomly divided into eight groups corresponding to the aliases and SES signals used (see Table 1). An advantage of factorial designs is their efficiency with respect to using experimental subjects; for a given number of treatments, factorial designs require fewer experimental subjects than alternative experimental designs to maintain the same level of statistical power (cf. Collins et al., Reference Collins, Dziak and Runze2009). As all treatment conditions are randomly distributed in the sample in a balanced way (cf. Appendix C, Table 7), we can estimate the individual independent effects of each treatment in a single model with little loss of efficiency. While it does not make a difference for the results if all three treatments are included in the statistical models or not, again showing that the randomization procedure was successful, we include them in all models to be prudent.
Table 1. Our treatments
Regarding the ethnicity treatment, Swedish-sounding and Arabic-sounding names are easily distinguished from each other. We tried to avoid stereotypical names that might signal SES or religious beliefs. More specifically, we chose the male name Mahmoud and the female name FatimahFootnote 4 . Both were ascribed the surname Hassan, which is rather common in Sweden (Statistics Sweden, 2017). Commonly, names have a certain SES association. Names of individuals belonging to ethnic minorities are more strongly associated with low education and low income levels and disadvantageous employment positions (Aldrin, Reference Aldrin2017; Elchardus and Siongers, Reference Elchardus and Siongers2011). Therefore, in a study such as this one, majority group names should convey SES levels similar to those of minority group names. In a prestudy, we asked upper secondary school principals to assess the education/income levels of school parents with different names (see Appendix D, Table 8). The results are in line with those of previous Swedish studies on names (Aldrin, Reference Aldrin2017). We chose the Swedish names Kevin and Melissa, as they were associated with SES levels similar to those of the Arabic names used in this study. The Swedish surname that we employ, Andersson, is very common and unlikely to be associated with any particular SES level.Footnote 5
SES discrimination is investigated by signaling highly skilled professions (dentist) in half of the e-mails and signaling low-skilled professions in the other half (care assistant). These professions are common among both immigrants and native-born individuals (Socialstyrelsen, 2018). In contrast to care assistants, who require only a high school education, dentists require extensive university education as well as a license to practice. Their average wage (47,400 SEK/month) is almost twice that of care assistants’ (24,800 SEK; Statistics Sweden, 2020). Hence, the chosen professions should clearly signal SES.
The emails sent to the principals were written as if sent by someone considering moving to the municipality (see Figure 1; translated from Swedish (see Appendix E, Figure 2). In this way, we prevented unusual names from arousing suspicion among principals in small municipalities where “everyone knows everyone.” We wanted the email questions to be straightforward and not particularly time-consuming to answer but also not too easily answered. We chose three relatively simple but important questions regarding school profile, the registration procedure and open slots.
Figure 1. The letter sent to the school principals.
The coding scheme consisted of 5 variables (see Table 2).Footnote 6 Two of these measured formal aspects, i.e., whether any reply was received within two weeks and how many of the three questions were answered. For the former variable, replies were registered from anyone working at the school or the municipality of the school, excluding autoreplies and noninformative emails from principals who had left their position. The remaining variables were of a more qualitative character, with one of them being an index measuring the friendliness of the emails.Footnote 7 The last two variables belonged to our promotion dimension and were based on information that could be used to motivate parents to choose or not choose a particular school. More precisely, we measured (a) information about open school slots and (b) whether there was any additional positive information in the principal’s reply. The latter variable concerned only information not directly related to the three questions asked by the parent. The first two variables as well as parts of our friendliness index (such as whether the principals use the name of the sender when replying) have been used in previous research (cf. Brown and Hilbig, Reference Brown and Hilbig2021; Olsen et al., Reference Olsen, Kyhse-Andersen and Moynihan2020). The promotion variables are new to the literature. The coding of the variables is described more in depth in Appendix I.
Table 2. Measures of discrimination
Of the original 3,430 emails sent, 3,394 were ultimately included in the dataset. Thirty-one emails bounced back and were therefore excluded. Five emails were omitted, as the principal answered that she/he had left the position and did not forward the email further. In the few cases where several responses were received from a principal or a school, the overall “best” response was chosen, i.e., the one that scored highest on the five variables taken together. No signs of spillover or disclosure of the experiment were found. Ethical concerns are discussed in Appendix A.
Results
General Discrimination Effects
Before turning to our research questions, the occurrence of discrimination in general is discussed (i.e., regardless of the type of provider or competition level). Multiple linear regression (OLS with several independent variables) is used in all models to make the results easier to interpret.Footnote 8 Table 3 (row 1) displays the ethnic discrimination effects for all types of schools, i.e., the treatment effect Footnote 9 of signing emails with an Arabic-sounding name versus a Swedish-sounding name. In terms of the response rate and the number of questions answered, we find small and nonsignificant negative effects (models 1 and 2). However, we observe relatively large and statistically significant negative effects for the other variables. Responses to emails signed with Arabic names are rated as less friendly (model 3, -0.192 friendliness points)Footnote 10 , are less likely to indicate that there are open slots in the school (model 4, 3.2 percentage points less likely) and are less likely to contain positive information about the school or municipality (model 5, 3.9 percentage points less likely). For socioeconomic discrimination, tendencies are similar. We find statistically significant evidence of low SES parents being discriminated against more for all qualitative and promoting aspects of the replies (row 2, models 3-5). Hence, both for ethnicity and SES, discrimination is found regarding the informal and promoting aspects of the emails but not regarding formal aspects, a result highly similar to that of prior research on general client discrimination effects in Sweden and Europe in general (Hemker and Rink, Reference Hemker and Rink2017; Olsen et al., Reference Olsen, Kyhse-Andersen and Moynihan2020). The discrimination coefficients remain largely unchanged after controlling for school ownership and competition (see Appendix F, Table 12), suggesting that these contextual factors may not be the most important drivers behind the effects we observe.
Table 3. Treatment effects on five discrimination outcomes of having an Arabic-sounding name, having a low-SES profession and being a woman
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level.
Competition Effects
Having observed the expected general discrimination effects, we now turn to our research question regarding competition effects. The relationship between school competition in the municipality and discrimination is investigated in Table 4. The Herfindahl–Hirschman Index (HHI, Rhoades, Reference Rhoades1993) is our main measure of competition. It is well established and frequently used to examine, e.g., the school sector. We calculate it by squaring the market share (proportion) of each school competing in the local school market (municipality) and then summarizing the resulting numbers (based on high-quality registry data from Statistics Sweden). The result is proportional to the average market share, weighted by market share. As such, it can range from 0 to 1.0, moving from a large number of schools with small market shares to a single monopolistic school. Swedish school markets are on average quite competitive, although there are notable differences between municipalities. The mean HHI in the dataset is 0.084; the lowest value is 0.006, the highest is 1, and the standard deviation is 0.088. The models include a large number of controls. Hence, the results should not be caused by other municipality characteristics, such as population size.
Table 4. Treatment effects on five discrimination outcomes of having an Arabic-sounding name and having a low-SES profession, by level of school competition in the municipality (HHI)
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level. The models also include controls for all treatments and a number of municipal level variables (municipality population, proportion foreign born, population growth, mean income, proportion voted for Swedish Democrats). The choice of municipality controls is discussed in depth in appendix G. Larger HHI values indicate lower competition.
Models 1, 3, 5, 7, and 9 in Table 4 concern ethnic discrimination. The results are mixed. Negative interaction coefficients in models 1, 7 and 9 (row 4) indicate that parents with Arabic-sounding names are treated worse in less-competitive school markets. The reverse is true for models 3 and 5. In the case of SES discrimination, there are positive interaction coefficients in all models, indicating that low SES parents are treated better in less competitive school markets (row 5, models 2, 4, 6, 8 and 10). However, none of the interaction effects are statistically significant and in most of the models, the coefficients are very small.Footnote 11 Furthermore, most of the individual treatment effects (rows 2-3 in models 5-10) of the low SES and Arabic-sounding names remain substantially and statistically significant even when the competition interactions are included. A complementary measure of competition also showed statistically nonsignificant and very weak interaction effects (see Table 11 in Appendix F). In sum, the overall picture is that biased treatment does not appear to be strongly related to competition. The results in Table 4 as well as Tables 5 and 6 are robust to numerous robustness tests, including controls for potential confounders such as student composition and student test scores as well as removing all controls (see Appendices G and H, Tables 13 to 16).
Table 5. Treatment effects on five discrimination outcomes of having an Arabic-sounding name, by school type (private/for-profit school vs. a public school)
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. School ownership reference category: public school. Standard errors in parentheses are clustered at the municipal level. The models also include controls for the SES and gender treatments.
Table 6. Treatment effects on five discrimination outcomes of having a low SES profession, by school type (private/for-profit school vs. a public school)
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. School ownership reference category: public school. Standard errors in parentheses are clustered at the municipal level. The models also include controls for the ethnicity and gender treatments.
Provider Effects
We now turn to our second research question regarding provider effects. For ethnic discrimination, the interaction coefficients in rows 4-5 in Table 5 show private/public and for-profit/public differences in discrimination (see also Table 9). Public schools refer to all schools owned by Swedish municipalities (78.2 per cent of all observations). Private schools refer to all schools not owned by Swedish municipalities, including corporations, cooperatives and non-profit associations and foundations (21.8 per cent). For-profit schools are a subset among private schools and refer to schools run as corporations (13.1 per cent).Footnote 12 We analyze the latter schools in separate models, as cream skimming and discrimination may be even more likely among this private school subset. By testing the interaction effects between the type of ownership and the Arabic name treatment, we investigate whether private and for-profit schools are more likely to discriminate than public schools. If ethnic discrimination is more common among private schools, the interaction coefficients in row 4 should be large, negative and statistically significant. However, the table shows only one statistically significant interaction effect. In addition, it is positive, suggesting that private schools more frequently inform Arabic aliases about open school slots (model 7: OLS: P=0.022, Logit: P=0.052). The interaction effect is not robust to multiple hypothesis testing and is not statistically significant above the 5-percent level when using logistic regression. In addition, the other interaction coefficients for private schools (row 4, models 1, 3, 5, and 9) generally show substantially very small interaction effects. The same goes for for-profit schools (row 5).Footnote 13 In sum, the results do not indicate clear differences in ethnic discrimination between private schools and public schools or between for-profit and public schools.
Similarly, the models in Table 6 show private/public and for-profit/public differences in SES discrimination. Only the interaction effect in model 10 is statistically significant (row 5, OLS: P=0.035, Logit: 0.007). This shows that for-profit schools are less likely than public schools to inform low SES parents about positive aspects of the school or municipality. The direction of the other interaction effects in rows 4-5 shows a similar pattern (8 of 10 coefficients are negative). Some of these could be considered to be of substantial size.Footnote 14 However, the parameters carry much uncertainty, as they are not statistically significant. Hence, even though there is a slight tendency of private schools discriminating more based on SES, the safest conclusion is that the hypothesis does not gain support.
Conclusion
It is often argued that market-based incentives may lead providers of public services to discriminate clients. However, previous experimental research is limited, and the findings are mixed, not always indicating that private providers are more likely to discriminate.
In this study, we addressed several limitations in prior scholarly work on discrimination in marketized welfare services. First, we focused not only on ownership and acknowledged the important marketization aspect of the competition level between providers. Second, we included publicly financed private for-profit providers, which should have the strongest incentives to cream skim. Third, we acknowledged discrimination based on SES, a client characteristic based on which service providers could be expected to cream skim clients. Fourth, we analyzed the quality and friendliness of replies and not only whether emails were replied to or not. All of these points should facilitate the discovery of discrimination effects potentially related to specific aspects of the marketization reform such as the introduction of private providers and competition.
Nevertheless, our results do not indicate that discrimination in the Swedish school market varies depending on these factors. We do not find clear evidence of principals at private schools discriminating along ethnic lines more than principals at public schools, and the same is true when we compare discrimination between for-profit schools and public schools. This goes both for ethnic and socioeconomic discrimination, although we – especially when looking at the substantial sizes of the coefficients – find some traces of private/for-profit schools treating low-SES parents less well than high-SES parents, although almost all of these effects are not statistically significant. We also do not find any robust impact of the level of market competition among schools on discrimination. For ownership and competition to qualify as the main drivers of discrimination, more consistent evidence would be needed.
What we do find is a strong general discrimination effect. In line with the more conventional and extensive research on public officials’ treatment of clients, which tends to not focus on marketized welfare services, principals working at both public and private schools were found to discriminate based on ethnicity as well as SES from the qualitative aspects of the emails. Here, our results highlight the importance of a new qualitative dimension of email replies (promotion) that could potentially be used to cream skim clients. Our Arabic and low SES parents were less likely to be informed about positive aspects of the schools and were also less likely to be informed about open school slots. This finding is of substantial importance, as discrimination of this kind could be used to deter certain parents from asking further questions or visiting the school and ultimately affect their choice of school. The occurrence of such biased treatment conflicts with fundamental principles of public sector service such as equal treatment.
Our best guess is that other discriminatory mechanisms – also present among public schools – are most likely at work behind the discrimination effects we found. For example, discrimination may primarily be a result of widespread conscious or unconscious discriminatory attitudes (cf. Larsson Taghizadeh, Reference Larsson Taghizadeh2022b). What speaks in favor of this view is that similar discrimination effects as those found here have been found in other public services operating under different conditions (e.g., public monopolies) in other countries (cf. Costa, Reference Costa2017). In addition, our results do not exclude the possibility that marketization/NPM reforms introduce incentives that result in some degree of discrimination among all providers (including public ones), regardless of ownership and levels of competition. Only a longitudinal study wherein discrimination levels are captured before and after such reforms would be able to safely determine whether this is the case. Another avenue for future research would be to utilize other methods to capture discrimination. Private providers may have (exclusive) opportunities to discriminate that correspondence experiments such as this cannot capture. For example, in the Swedish case, such providers may manipulate queues (which they administer themselves), that may be studied using a more case-based method. While our study provides valuable findings regarding unequal treatment of clients following a controversial marketization reform, more studies covering other countries and/or utilizing other methods are needed before we can draw safe conclusions regarding the effects of marketization on discrimination.
Funding and data availability
An anonymized version of the data underlying this article will be shared on reasonable request to the corresponding author. The design was approved by the Swedish ethical review board (see 2017/234 and 2018/371). This work was supported by the Swedish Research Council for Health, Working Life and Welfare (2019-00504).
Conflict of interest disclosure
The authors declare none.
Appendix
A. Research Ethics
To reduce social desirability bias and capture real-life discrimination, the research subjects were not made aware that they were part of a field experiment. Discrimination is potentially sensitive, and the results would not have been accurate if the elementary school principals had been informed and asked to participate beforehand. However, we strived to decrease the potential negative effects of the experiment as much as possible. First, the principals were anonymized, and only aggregate-level tendencies are shown and not specific answers. The emails and the discrimination effects are presented in a way that prevents identification of the schools or the municipalities from which they were sent. Second, we minimized the time the principals spent on emails by keeping the questions simple. Some principals asked questions in their replies, but these questions were not answered to prevent them from working additional hours. To the best of our knowledge, no similar experiment has been conducted on the research subjects, and the authors will not implement any future experiments on them (without seeking consent) to avoid affecting future encounters between them and citizens. The design was approved by the Swedish ethical review board.
B. Generating the List of Principals
To minimize the risk of spillover effects and detection, the list of principal email addresses was generated based on registry data following these rules: (1) Only one unique email address per principal was allowed, (2) Principals working at several school units (based on school codes) were contacted only once, (3) If there were several principals working at a school (e.g., vice-principals), the email address to the main principal was used, (4) In the cases where the email addresses for the principals in the registry data contradicted those on the homepages (according to web-scraping), the addresses on the homepages were used, (5) When no email for directly contacting the principal was available for a school and we had no reason to believe that the school’s principal was already included in our dataset, we contacted the school directly via its common address, and (6) 36 principals that participated in a pilot study were not included in the list of email addresses. These rules resulted in 543 unique email addresses to schools that were not used, and some of these email addresses led to principals being excluded from the final sample. Approximately 5 percent of the emails (180/3430) were answered by nonprincipals (e.g., an administrator at the school or municipality) according to the responses.
C. Randomization Checks
Most of the studies cited in this article did not test for balance across the treatment groups. Hence, they provided no direct evidence that the randomization was successful. Thanks to the availability of register data, balance can be tested in the present study. As shown in Table 7, the randomization procedure resulted in an overall high level of balance between the different treatments in terms of both school and municipality characteristics. There was only one case where significant differences in the means were found between the treatment groups (tested with t tests). For-profit schools were slightly more likely to receive emails from low-SES aliases than from high-SES aliases. However, as the private school variable was generally well balanced between the treatment groups and often resulted in similar effects as the public variable, it is unlikely that the imbalance for for-profit schools was driving the results. To be on the safe side, entropy balancing and coarsened exact matching methods were used to determine whether increasing the balance in this variable would change the results. However, these additional analyses resulted in highly similar results.
Table 7. Balance between the different treatment groups in the dataset, for school and municipal level variables (means)
Notes: + p <.10, * p <.05, ** p <.01, *** p <.001; t tests are for significant differences in means. Standard deviations are in parentheses. Units of analysis = schools.
D. Choice of Swedish Names
Before conducting the experiment, a smaller study was conducted where upper secondary school administrators (i.e., not from an elementary school) had to rate names on a 100-point scale based on socioeconomic status (0=low income/education, 100=high income/education). Of the 715 schools that received the invitation, 252 (35.2 percent) answered all survey questions.
Table 8. Results from our study where upper secondary school administrators rated names on a 100-point scale on income/education
E. Original Swedish Letters Sent to the Principals
Figure 2. The original letters in Swedish.
F. Additional Analyses
Table 9. Treatment effects on five discrimination outcomes of having an Arabic/Swedish-sounding name, a low/high-SES profession and being a woman, depending on school type (public or private)
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level.
Table 10. Treatment effects on the four friendliness measures included in the friendliness index of having a low/high-SES profession, an Arabic/Swedish-sounding name and being a woman
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level.
The HHI measure may overestimate competition in larger municipalities where there are more schools (and schools with more than 1,000 students are rare). Therefore, we test another measure of competition that is frequently used in the education literature: the private school market share (the proportion of students in the market attending private schools); see Table 11 below. In municipalities where this share is low, we expect most schools to operate as public monopolies (no competition). In municipalities where the share is higher, the diversity of schools is probably larger, and both public and private schools presumably have to use market principles to be competitive. Unlike the HHI measure, however, this alternative measure may ignore competition between public schools and overestimate competition in municipalities with few but large private schools. As both approaches provide similar results, our conclusions regarding competition effects appear to be rather reliable.
Table 11. Treatment effects on five discrimination outcomes of having an Arabic/Swedish-sounding name and a low/high-SES profession, by private school market share in the municipality
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level. The models also include controls for all treatments and the same control variables included in the HHI models.
Table 12. Treatment effects on five discrimination outcomes of having an Arabic-sounding name, having a low-SES profession and being a woman, after controlling for school type and competition
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level.
G. Choice of Controls
In Table 4 in the manuscript, we include a number of controls since the level of competition between schools in a municipality could potentially correlate with numerous municipality-level factors that in turn could correlate with discrimination. For example, the HHI of a municipality correlates with the size of the municipality (-0.427***), population growth (-0.532***) and municipality mean income (-0.331***), which could affect both the economic resources the principals have at hand per student and the number of parents contacting them. Similarly, HHI correlates with the proportion of the population with a foreign background (0.358***) and the proportion voting for the far-right Sweden Democrats (0.156***), factors that more directly could affect how ethnic minorities are treated.
Tables 13 to 15 in the Appendix contain a number of robustness tests where we present results when controlling for student composition as well as student test scores. Controlling for these variables could potentially be important, as there may exist systematic differences between different schools and between different school markets in these regards that could also correlate with discriminatory behavior. For example, principals in underperforming schools with few students with parents with higher education may be more likely to discriminate as they are more likely to be under stress and since avoiding more demanding clients is a way to quickly raise the average grades. On the other hand, overperforming schools with more resourceful students may also discriminate demanding/costly students as their way of teaching and their available resources are adapted to a certain type of student composition. As there are differences – for example, in student composition (see Table 17 below) – between public and private schools, it may be important to control for these potential confounders to avoid omitted variable bias. However, as we see in Appendix H, the results remain largely the same when adding these controls.
H. Robustness Tests
Table 13. Treatment effects on five discrimination outcomes of having an Arabic/Swedish-sounding name by school type (private/for-profit school vs. a public school), controlling for student composition and test scores
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level. The models also include controls for all treatments as well as private schools (models 1, 3, 5, 7, 9) and for-profit schools (models 2, 4, 6, 8, 10), as well as the proportion of students with parents with higher education and the proportion of students passing the national tests in Swedish and Mathematics (not shown in the table).
Table 14. Treatment effects on five discrimination outcomes of having an a low/high-SES profession by school type (private/for-profit school vs. a public school), controlling for student composition and test scores
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level. The models also include controls for all treatments as well as private schools (models 1, 3, 5, 7, 9) and for-profit schools (models 2, 4, 6, 8, 10), as well as the proportion of students with parents with higher education and the proportion of students passing the national tests in Swedish and Mathematics (not shown in the table).
Table 15. Treatment effects on five discrimination outcomes of having an Arabic-sounding name and having a low-SES profession by level of school competition in the municipality (HHI), controlling for student composition and test scores
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level. The models also include controls for all treatments, HHI and a number of municipal level variables (municipality population, proportion foreign born, population growth, mean income, proportion voted for Swedish Democrats), as well as the proportion of students with parents with higher education and the proportion of students passing the national tests in Swedish and Mathematics (not shown in the table)
Table 16. Treatment effects on five discrimination outcomes of having an Arabic-sounding name and having a low-SES profession by level of school competition in the municipality (HHI), without controls
Note: + p <.10, * p <.05, ** p <.01, *** p <.001. Standard errors in parentheses are clustered at the municipal level.
I. Coding
Emails were independently coded by two assistants following instructions and a coding form written by the authors of this article. After coding all responses independently, the assistants paid special attention to emails where their coding differed to reconcile the final coding. However, regarding the simplest variables (reply and the variables forming the friendliness index), differing cases were reviewed by a third assistant. The names of the fictitious emailers were removed before coding started. Moreover, nonresponses were coded as zero for all variables, a standard approach in contemporary correspondence studies. The reason is to avoid selection bias induced by posttreatment variable conditioning (cf. Coppock, Reference Coppock2019). To illustrate the coding procedure for the three variables, two complete email answers are presented in Figure 3.
Figure 3. Two examples of email replies from the school principals.
To illustrate the coding procedure, looking at the first email, all three questions are answered (regarding the school profile, open slots and how to apply). Hence, the variable for the number of questions answered is assigned the value 3 (cf. Table 2). Moreover, the sender is welcomed in a friendly way, her or his name is used, and she or he is invited to make future contact and visit the school. Hence, the friendliness index variable is given the maximum value of 4. The open school slots variable is assigned a value of 0, as the sender is not informed that there are open slots available. No additional positive information regarding the school or municipality is provided, so the positive information variable is therefore given the value 0. For email 2, the principal scores only 2 on the number of questions answered (only the questions on how to apply to the school and on the school profile are answered). The email also includes additional positive information regarding the school (information about average grades). Concerning all the other variables, it receives values of 0. Hence, the first email is formally correct and friendly but does not involve any promotion. The second one is somewhat less formally correct and less friendly but includes some aspects of promotion.
J. Student Composition and School Results for Different School Types
Table 17. School composition and test results depending on school ownership
*data not available for 19 percent of all schools
** Differences in test scores between private/for-profit schools and public schools are not statistically significant after controlling for student composition
Open access
