Discrimination in the Rental Market for Apartments*
Magnus Carlsson† and Stefan Eriksson‡
Abstract
Discrimination in the housing market may create large inefficiencies, but is difficult to measure. To circumvent the problems with unobserved heterogeneity, most recent studies use the correspondence testing approach (i.e. sending fictitious applications to landlords). In this study, we extend the existing methodology by (i) randomly assigning all relevant applicant characteristics to the applications, and (ii) carefully taking into account the interactions between applicant, landlord, apartment and regional characteristics. Then, we demonstrate how this approach can be implemented by considering how an applicant’s gender, ethnicity, age and employment status affect the probability of being invited to an apartment viewing in the Swedish housing market. Our results confirm the existence of widespread discrimination, but also show that the degree of this discrimination varies substantially with applicant, landlord, apartment and regional characteristics. This heterogeneity highlights the importance of using of using a broad approach when conducting correspondence studies.
JEL classification: C93, J15, J16, R39
Keywords: Discrimination; Field experiment; Correspondence testing; Gender;
Ethnicity; Age; Employment status; Housing market
* We are grateful for helpful comments from colleagues at Uppsala University and the Linnaeus University Centre for Labor Market and Discrimination Studies. Financial support from the Swedish Research Council and the Jan Wallander and Tom Hedelius Foundation is gratefully acknowledged.
†Linnaeus University Centre for Labor Market and Discrimination Studies, Linnaeus University, SE-391 82 Kalmar, Sweden, Magnus.Carlsson@lnu.se
‡Department of Economics, Uppsala University, PO Box 513, SE-751 20 Uppsala, Sweden, Stefan.Eriksson@nek.uu.se
1. Introduction
Discrimination in the housing market may create large inefficiencies. If people with certain characteristics are unable to find housing in some areas due to discrimination, it becomes much more difficult to reduce high unemployment and to avoid the social problems which may arise in segregated areas. Hence, it is crucial for policymakers to be informed about how factors such as people’s gender, ethnicity, age and employment status affect their probability of finding housing.
Many studies document substantial differences in housing market outcomes between different groups, e.g. between ethnic groups (cf. Dymski, 2006). This may reflect discrimination, but proving it is often very difficult.1 In studies using observational data (i.e. administrative or survey data), it is often impossible to separate the effects of unobserved factors and discrimination. To circumvent the problems with unobserved heterogeneity, it has become increasingly popular to use field experiments to study discrimination. In these studies, fictitious applications are sent to landlords (correspondence studies) or auditors are sent to apartment viewings (audit studies).2 Recently, most studies use the first approach.
In a typical correspondence study, qualitatively identical applications are sent to landlords with a vacant apartment – the only difference being the name of the applicant,
1 By discrimination, we mean any situation when two applicants who are identical in all dimensions except for one – e.g. gender, ethnicity or age – are treated differently. Only some forms of discrimination are illegal. According to the Swedish Discrimination Act, it is illegal to discriminate based on gender, transgender identity or expression, ethnicity, religion or other belief, disability, sexual orientation or age.
Moreover, the discrimination may preference-based or statistical or any combination of these two types.
2 Examples of studies using the correspondence testing methodology are Carpusor and Loges (2006), Ahmed and Hammarstedt (2008), Ahmed et al. (2010), Bosch et al. (2010), Baldini and Federici (2011), Hanson and Hawley (2011), Ewens et al. (2012), and Andersson et al. (2012). Examples of studies using the audit study methodology are Yinger (1986), Page (1995), Ondrich et al. (2000), Zhao (2005), and Zhao et al. (2006). The correspondence testing approach is described in e.g. Riach and Rich (2002). A reason for the popularity of this approach is the emergence of Internet-based search channels which have made it much easier to conduct these studies. Moreover, audit studies have been severely criticized on methodological grounds; c.f. Heckman and Siegelman (1993) and Heckman (1998).
which is chosen to signal e.g. a certain gender or ethnicity. Discrimination is then measured as the difference between the groups in the probability of being invited to an apartment viewing.
This approach has a clear advantage in terms of identifying discrimination, but also suffers from some potentially important weaknesses. First, the measured degree of discrimination may depend on the experimenter’s choice of the constant attributes in the applications, especially the level at which these attributes are held constant. This issue is highlighted by Heckman and Siegelman (1993) and Heckman (1998) who show that these choices may affect estimates of the degree of discrimination.3 Second, this approach may not fully capture the heterogeneity of discrimination. Many studies include only a few types of landlords and apartments located within a limited geographical area or do not take such factors into account in the empirical analysis. This is problematic if the landlords’ preferences are heterogeneous across different types of landlords (e.g.
male/female, ethnic majority/minority, and individual/company), apartments (e.g. size and rent) or regions (e.g. metropolitan and other areas). Hence, there is room for methodological improvements in the way correspondence studies are conducted.
In this study, we describe a more general methodology for conducting a correspondence study in the housing market. Then, we demonstrate how this approach can be implemented in practice by investigating the effects of gender, ethnicity, age and employment status in the Swedish rental market for apartments. In the experiment, we incorporate substantial variation in applicant, landlord, apartment and regional characteristics, and carefully analyze how these factors interact.
3 These studies mainly criticize audit studies, but many of their arguments are valid for correspondence studies as well. These issues are also discussed in Neumark (2012).
For applicant characteristics, the most general design is to randomly assign all relevant characteristics. This has several advantages: First, it makes it possible to analyze to what extent the measured degree of discrimination depends on the experimenter’s choice of the applicant characteristics that are typically held constant in these studies, i.e.
everything except the applicant’s name. Second, it enables a simultaneous analysis of different types of discrimination, and how they interact. Third, it makes it easier to interpret the estimates of discrimination since the effect of different characteristics can be compared. To our knowledge, no previous correspondence study in the housing market has implemented a randomization of a substantial number of distinct applicant characteristics.4
For the landlord, apartment and regional characteristics, we include many different types of landlords and apartments located in many different types of regions, and collect detailed information about the landlords and apartments from the advertisements.
Moreover, we combine the data from the experiment with administrative data from official records on important economic, social and demographic characteristics of the municipalities where the apartments are located.5 This design enables us to carefully analyze to what extent the measured degree of discrimination depends on landlord, apartment and regional characteristics.
To demonstrate how this more general approach to correspondence testing in the housing market can be implemented in practice, we conducted a large-scale field
4 Instead, some studies have used a bundle of characteristics (for example high/low quality applications), and then randomly assigned either all or none of the characteristics in the bundle to the applications; c.f.
Ahmed et al. (2010) and Bosch et al. (2010). In the labor market, there are some correspondence testing studies that vary a number of worker characteristics; e.g. Rooth (2011) and Eriksson and Rooth (2013).
5 There are a few studies that have made attempts to do this. Carpusor and Loges (2006), Ahmed and Hammarstedt (2008) and Bosch et al. (2010) analyze the effects of some landlord and apartment characteristics. Ewens et al. (2012) and Hanson and Hawley (2011) analyze the importance of the neighborhoods’ racial composition.
experiment on the largest Swedish online classified advertisement website, where individual and small company landlords search for tenants.6 During a six-month period in 2010 and 2011, we replied to more than 5,800 advertisements. We quantify the degree of discrimination by estimating the difference in the landlords’ callback rate (i.e. the probability of being invited to an apartment viewing), and analyze how this estimate depends on the various dimensions of the applicant, landlord, apartment and regional characteristics discussed above.
Our results confirm the existence of widespread discrimination, but also show that the degree of this discrimination varies substantially with applicant, landlord, apartment and regional characteristics. This suggests that it may not be possible to summarize the degree of discrimination in the rental market by one representative figure. The landlords in the experiment use ethnicity and employment status, but not gender and age, to sort applicants. That landlords avoid ethnic minority applicants constitutes illegal discrimination, while the fact that they avoid unemployed applicants does not. However, the consequences for labor mobility, unemployment and social problems may be severe in both cases. Moreover, we find that there are important interaction effects between a number of applicant characteristics. This implies that estimates of the degree of discrimination actually may depend on the experimenter’s choice of the constant attributes in the applications. In addition, our results show that the effects vary substantially with landlord/apartment characteristics (in particular, the landlords’
ethnicity) and regional characteristics (in particular, the labor market situation and the number of immigrants in the municipality). This highlights the importance of including
6 This dataset is also used in Carlsson and Eriksson (2012) to analyze the link between public attitudes towards immigrants and ethnic discrimination.
many different types of landlords and apartments located in many different types of regions, and carefully taking this information into account in the empirical analysis.
Overall, our findings highlight the importance of designing correspondence studies as broad as possible.
The rest of the paper is organized as follows. Section 2 describes the field experiment. Section 3 presents descriptive results, and Section 4 contains the empirical analysis. Section 5 concludes the paper.
2. The field experiment
In this section, we discuss the main features of our experimental approach. Then, we describe some remaining details of the experiment.
2.1 The main features of the experiment
The key aspects of our experimental approach are random assignment of all relevant applicant characteristics to the applications, inclusion of many different types of landlords/apartments located in many different types of regions, gathering of extensive information about the landlords/apartments and regions, and a careful analysis of how these factors interact.
The applicant characteristics that are randomized in the experiment can be divided into two distinct categories. The first category contains the characteristics which signal the groups that we study discrimination against, while the second category contains all other characteristics which are typically included in an apartment application.
Concerning the first category of characteristics, we restrict the experiment to study discrimination with respect to gender, ethnicity and age. Also, we consider the effect of employment status. These restrictions are motivated by the fact that, although we want to study discrimination against as many groups as possible, the expected sample size puts a limit on how many groups we can include in the experiment. To study gender discrimination, we include applicants with male and female names.7 Concerning ethnic discrimination, previous studies show that men with Arabic/Muslim names face widespread discrimination in the Swedish housing and labor markets and, therefore, we include applicants with Arabic/Muslim names in the experiment (cf. Ahmed and Hammarstedt, 2008, and Carlsson and Rooth, 2007). Interestingly, ethnic minority women are a group that, so far, has received less attention in correspondence studies.8 Therefore, we include applicants with both male and female Arabic/Muslim names.9 An applicant’s gender and ethnicity is signaled by the name, and all messages to the landlords contained the applicant’s name after the initial greeting, at the end of the message, and in the e-mail address which was used for receiving responses from the landlords.10
Two applicant characteristics that, to our knowledge, have never been considered in a correspondence study in the housing market are age and employment status.11 In this experiment, we include age as a random integer between 25 and 55, and employment status as being unemployed (expressed as ‘looking for work’), working as a shop sales
7 We used the most common names in the name register kept by Statistics Sweden. The names were Erik Johansson and Anna Nilsson.
8 Two exceptions are Bosch et al. (2010) and Andersson et al. (2012).
9 The names were Ali Hassan and Fatima Ahmed.
10 The applicants’ e-mail addresses were registered at a well-known e-mail provider.
11 Andersson et al.(2012) study something they label ‘class’ where they compare how the callback rate to a viewing is affected by being employed as an ‘economist’ as opposed to as a ‘warehouse worker’. However, they do not include unemployed applicants.
assistant or working as a financial manager. Obviously, it is not illegal for landlords to avoid applicants who are unemployed or low skilled. However, the consequences for society in terms of labor mobility, unemployment and social problems may be severe if many landlords sort applicants based on their employment status (cf. Section 4.1).
The second category of applicant characteristics contains other attributes which are typically included in apartment applications. To decide which characteristics to include, we registered a few fictitious apartment advertisements on the website we use in the experiment, and analyzed the applications we received. In addition, we examined a number of posted advertisements on the website to learn which information landlords typically request. Based on this exercise, we decided to include three additional applicant attributes: Leisure time interests, smoking habits, and whether the applicant had a reference from a previous landlord or not. Leisure time interests were enjoying evenings at home, being engaged in sport activities, or enjoying restaurant life and nightclubs;
smoking habits was a dichotomous variable indicating that the applicant was a smoker (who never smoked indoors) or a non-smoker; references were also a dichotomous variable which corresponds to either a sentence expressing that the applicant had a reference from a previous landlord or no information at all about references.
To enable an analysis of the importance of landlord/apartment and regional characteristics, it was essential to include many different types of landlords/apartments located in many different types of regions, and – equally important – to collect extensive information about the landlords/apartments and regions. Information about the landlords and apartments was retrieved from the advertisements. In most cases, it was possible to identify the name of the landlord and, thus, to deduce the landlord’s gender and ethnicity
(i.e. native Swede or ethnic minority) as well as whether the landlord was an individual or a company. The information about the apartments which we extracted from the advertisements included the rent, the number of rooms, the size of the apartment in square meters, whether it was a rented or a co-operative apartment, and the municipality where the apartment was located.
Regional data for the municipalities where the apartments were located was collected from Statistics Sweden. This data include information about important economic, social and demographic characteristics at the municipality level. Specifically, we retrieved information about labor market characteristics (the employment rate; the share on income support), characteristics of the general population (the share of highly educated; the average age; the share of men), characteristics of the immigrant population (an ethnic segregation index; the share of immigrants; the share of immigrants from countries outside the EU), and rental market characteristics (the share of vacant apartments).
2.2 The details of the experiment
To collect apartment advertisements, we used the largest Swedish online classified advertisement website, where landlords search for tenants.12 The choice of an online search channel means that the advertisements and all communication with landlords are conducted online. To make a large-scale data collection feasible, we took advantage of this fact by writing a computer program that handled most of the data collection. Another important advantage of using a computer program for the data collection is that the risk of human error is minimized.
12 The site is called Blocket (www.blocket.se).
The computer program started by navigating to the website with the advertisements and created a list of all currently available apartments. Then, the program went through this list, and displayed one advertisement at a time to the research assistant. The assistant checked that the landlord did not state that he or she only accepted responses by telephone, which happened in only a very limited number of cases. In addition, the assistant checked that the same advertisement had not been replied to before (i.e. if the same advertisement was repeated). If the advertisement was approved, two steps were taken by the program. The first step was to retrieve all available information about the landlord and the apartment from the advertisement and temporarily store this information.
The second step was to make consecutive random draws – one draw for each applicant characteristic (e.g. age) – from a uniform distribution where the outcome determined the value of each applicant characteristic (e.g. 35 years old). In principle, each applicant characteristic corresponded to a specific sentence in the application (e.g. ‘I’m a 35 year old’). Then, the program created a message with the correct typeface and layout. This message always started with a polite greeting followed by the sentences obtained from the randomization procedure.13 Next, the program pasted the constructed message into the form on the website intended for responding to an advertisement and pushed the reply button. Finally, – after the research assistant had confirmed that the message had been sent to the landlord – the program permanently stored information about the applicant
13 An example of a message is (translated from Swedish):
Hi,
I’m a 35 year old woman who is very interested in the apartment. To tell you a little about myself, I work as a financial manager. In my spare time, I like to visit restaurants and nightclubs. I should also mention that I smoke (but never indoors). I have good references from my previous landlord. Is it possible to come and take a look at the apartment? Please contact me by e-mail.
Best,
Anna Nilsson
characteristics and the retrieved information about the landlord and the apartment in a spreadsheet. Only one application was sent to each advertised apartment.
During a six month period between October 2010 and March 2011 a random sample of more than 5,800 advertisements was collected. We applied to apartments all over Sweden and, since the competition for apartments – especially in some parts of Sweden – is intense, we always responded quickly to newly posted advertisements.14 Almost all the landlords on the website accepted applications sent by e-mail through the form available on the website. Responses from the landlords where always received by e-mail. For each response, the research assistant recorded the type of response in a spreadsheet and promptly declined all invitations to viewings to minimize any inconvenience to the landlords.
The website we used for collecting the advertisements is widely used by individual and small company landlords to find tenants for vacant apartments.15 Often, people look for temporary tenants. However, public housing companies and large commercial landlords often use other ways to find tenants.16 Hence, this website should be fairly representative of the Swedish market for individual and small company landlords.
3. Descriptive results
The average degree of discrimination measured in the experiment is the starting point for the analysis of to what extent the degree of discrimination varies with applicant, landlord,
14 Usually, we replied to advertisements within 24 hours after they were posted.
15 The site is the third largest website in Sweden and has close to four million unique visitors per week (www.kiaindex.net). It contains around 500,000 advertisements in various categories. According to the polling firm Synovate, 95 percent of the Swedish population has heard about the site and almost 70 percent has at some point in time bought or sold something on the site.
16 Typically, these landlords have their own queuing systems which they use to find tenants. In many cases, an applicant must wait several years to get an apartment from these companies. People who need an apartment quickly are therefore often forced to look for temporary contracts.
apartment and regional characteristics.
To construct a measure of average discrimination, we use the responses from the landlords. These responses can be divided into three categories: An invitation to a viewing, another positive response, and a negative response. An invitation to a viewing is a reply which contains a phrase that explicitly states that it is possible to arrange a viewing of the apartment. Another positive response is a reply which does not contain an explicit invitation to a viewing, but rather asks for further information about the applicant or otherwise indicates a positive response. A negative response is a reply which indicates that it is not possible to arrange a viewing. An example of a negative response is a reply which states that the apartment has already been rented to someone else.
Table 1 reports descriptive results for the landlords’ responses to the 5,827 vacant apartments we applied for. The number of apartments is very similar for all groups that we consider (cf. the last line in Table 1). This indicates that the randomization of the group characteristics has worked as intended. In the first row are, the callback rates to a viewing, i.e. the number of invitations to a viewing divided by the number of apartments we applied for. The callback rates for the gender and ethnic groups are: 26 (Swedish male name), 28 (Swedish female name), 14 (Arabic/Muslim male name), and 19 (Arabic/Muslim female name) percent. For the age groups, the callback rates are: 21 (aged 25-35), 23 (aged 36-45), and 21 (aged 46-55) percent. Finally, the callback rates for the employment status groups are: 15 (unemployed), 23 (shop sales assistant), and 28 (financial manager) percent. In the second row, are the other positive response rates, which vary between nine and eleven percent. In the third row, are the negative response
rates, which vary between five and ten percent. Finally, in between 55 and 70 percent of the cases, the applicants got no response at all (cf. the fourth row).
In the empirical analysis, we focus on differences in the callback rate to a viewing.17 There are several reasons for this choice. First, in most cases a prerequisite to get an apartment is to be invited to a viewing. Therefore, it is natural to focus on this alternative.
Second, the other alternatives are somewhat difficult to interpret. For example, a quite common positive response was that the landlord had already invited a few people to viewings and would contact the applicant again if the apartment was still available after these viewings, but in almost all such cases the landlord did not get back to the applicant.
Also, a common negative response was that the apartment had already been rented to someone else. These answers may of course be true, but may also be used as an excuse to hide discrimination. For similar reasons, a non-response is also difficult to interpret.
Third, the descriptive results presented above show that the most compelling difference between the groups is the difference in the share of applicants who are invited to a viewing.
4. Estimation and results
In this section, we first study the average degree of discrimination in the rental market.
Then, we consider how the degree of discrimination varies with applicant, landlord, apartment and regional characteristics.
4.1 The average degree of discrimination
To estimate the average degree of discrimination, we investigate differences in the
17 The results are similar if we include other positive responses.
callback rate (i.e. the probability of being invited to an apartment viewing) based on gender, ethnicity, age and employment status. For this analysis, it is crucial to note that since all applicant characteristics are randomly assigned to the applications, there is no scope for unobserved heterogeneity and there are no interdependencies among the explanatory variables. This is the key advantage of using data from a field experiment compared to using observational data. Moreover, this means that we should get the same estimates of discrimination irrespectively of the number of applicant characteristics that we include in the regressions. Therefore, we can analyze discrimination against one group at a time without worrying about problems caused by omitted variables.
The estimated average degrees of discrimination are presented in Table 2.18 In column 1, we focus on the effects of gender and ethnicity. Applicants with Swedish male names are the reference category; the constant gives the average callback rate for this group, which is 26 percent. Compared to applicants with Swedish male names, applicants with Swedish female names have a two percentage points higher callback rate, although this difference is not statistically significant.19 In contrast, applicants with Arabic/Muslim male names have a statistically significant twelve percentage points lower callback rate compared to applicants with Swedish male names. The corresponding difference for applicants with Arabic/Muslim female names is minus eight percentage points, which is also a statistically significant difference. T-tests show that applicants with Arabic/Muslim female names have a statistically significant lower callback rate compared to applicants
18 We use the linear probability model in the estimation. All models in this study basically compare average callback rates between cells defined by dummy variables constructed from the different applicant attributes.
There are no empty cells, and the average callback rate in each cell is always between zero and one. Thus, we do not have to worry about predicted callback probabilities outside the unit interval and, therefore, we can use the linear probability model (this case is discussed in detail in Wooldridge, 2012). However, the results are similar if we use the probit model and calculate the marginal effects.
19 This difference is statistically significant at the ten percent level in the regression with all applicant attributes included.
with Swedish female names, but a statistically significant higher callback rate compared to applicants with Arabic/Muslim male names. Thus, our results show clear evidence of ethnic discrimination, but not of gender discrimination.
How do these findings relate to the results in previous correspondence studies? That ethnic minority men face discrimination is in line with the results in several other studies (e.g. Ahmed and Hammarstedt, 2008, and Ahmed et al., 2010). That the degree of ethnic discrimination is lower for women than for men is similar to the results in Bosch et al.
(2011), while Andersson et al. (2012) find no statistically significant difference. That we find essentially no difference in the callback rate between applicants with male and female Swedish names is in contrast to the results in Ahmed and Hammarstedt (2008) and Andersson et al. (2012). Hence, the results in different studies are rather conflicting despite the fact that many of the studies are conducted in Scandinavian rental markets.
In column 2, we focus on the effect of age. The reference category is applicants aged 25-35, who have an average callback rate which equals 21 percent. Applicants aged 36- 45 have a somewhat higher callback rate (two percentage points higher; statistically significant at the ten percent level) than the reference group, while applicants aged 46-55 have a similar callback rate. Thus, we find essentially no evidence of age discrimination.20 Since age discrimination in the rental market has never been analyzed before in a correspondence study, this is a new result.
In column 3, we focus on the effect of employment status. The reference category is unemployed applicants, who have an average callback rate that equals 15 percent. Both applicants who are employed – as shop sales assistants or financial managers – have a statistically significant higher callback rate; plus 8 and 13 percentage points, respectively.
20 If age is included as a continuous variable, it is not statistically significant.
Thus, we find clear evidence that landlords avoid unemployed applicants, which is also a new result. Again, it should be stressed that it is not illegal to use employment status to sort applicants. In fact, it may be a completely rational strategy if landlords believe that unemployed or low skilled applicants will be unable to pay the rent. However, the consequences in terms of labor mobility, unemployment and social problems may be severe if many landlords use such a strategy.
In column 4, we include all applicant characteristics in the regression. The purpose is to ensure that the randomization of the applicant characteristics has worked as intended, and to show how the other applicant attributes, which will be further analyzed in the next section, affect the callback rate. By comparing the estimates in this column with the estimates in the previous columns, it is evident that all results remain essentially unchanged when all applicant characteristics are included in the regression. This is exactly what we expect since all attributes are randomly assigned to the applications. All the estimated coefficients for the added attributes have the expected signs and are statistically significant different from zero. Enjoying evenings at home and being engaged in sports, as opposed to enjoying restaurant life and nightclubs, increases the callback rate by seven and nine percentage points, respectively. Being a smoker decreases the callback rate by three percentage points. Having a reference from a previous landlord increases the callback rate by two percentage points.
Having established that discrimination exists on average, and that the estimates of the other applicant attributes are reasonable, we now turn to the next issue: The analysis of whether the measured average degree of discrimination varies with the applicant, landlord, apartment and regional characteristics.
4.2 Applicant characteristics
A natural way of investigating if discrimination varies with applicant attributes is to estimate the return to the attributes separately for each group (i.e. the gender/ethnicity, age and employment status groups).21 If there are differences in the return to the different attributes between the groups, then the experimenter’s choice of the constant attributes in the applications is likely to influence the measured degree of discrimination. Table 3 reports the results for the three categories gender/ethnicity, age, and employment status.
In columns 1-4, are the results for the gender/ethnic groups. In general, the estimated coefficients are rather similar. The only statistically significant difference is that male smokers face a negative penalty, while female smokers do not. However, it should be noted that strong positive signals – such as being employed as a financial manager relative to being unemployed – has very similar effects on the callback rate in the different groups. Thus, including strong positive signals in the applications do not seem to decrease the magnitude of ethnic discrimination.22
In columns 5-7, are the results for the age groups. Here, there is a statistically significant difference concerning employment status: Being unemployed implies a higher callback rate for the age group 25-35 than for the older age groups (cf. the constant), but since the positive effect of being employed is lower, this age group has a similar callback rate as the older age groups when employed. These results suggest that landlords view unemployment as a less important negative signal for younger applicants, and a good job as a more important signal for older applicants.
21 An alternative approach is to include interaction effects in the regression for the full sample. We use this approach to verify that the differences between the groups are statistically significant (with t-tests).
However, we believe that it is more illuminating to present the results for each group separately.
22 This seems to confirm the results in Ahmed et al. (2010) and Bosch et al. (2010) that adding more information in the applications does not eliminate the discrimination.
In columns 8-10, are the results for the employment status groups. Here, there is more heterogeneity in the effects of different characteristics. For applicants working as financial managers, as opposed to being unemployed, there is a statistically significant positive effect from being older. Moreover, the positive effects of enjoying evenings at home or being engaged in sport activities, as opposed to enjoying restaurant life and nightclubs, are stronger (and statistically significant) for employed applicants. Finally, the positive (and statistically significant) effect of having a reference is much stronger for unemployed applicants than for employed applicants where it is close to zero.
Overall, the effects of the applicants’ characteristics are rather similar across the different groups. However, there are a number of potentially important differences, such as e.g. the difference between applicants with male and female Arabic/Muslim names (cf.
Section 4.1), the difference in how landlords perceive unemployment in different age groups, and the difference in the value of having a reference for different employment status groups. This highlights the importance of considering such interaction effects in correspondence studies, and this requires randomization of a sufficient number of distinct relevant applicant characteristics.
4.3 Landlord and apartment characteristics
A dimension where there may be important differences in the degree of discrimination is between different types of landlords and/or apartments. In contrast to the applicant characteristics – which are all randomly assigned and thus truly exogenous –, the landlord and apartment characteristics may be correlated both with each other and with other unobserved factors affecting the probability of being invited to a viewing. Therefore, we
choose to not include the landlord and apartment characteristics as explanatory variables in the regressions. Instead, we divide the data with respect to the landlord and apartment characteristics, and investigate whether the causal effects of gender, ethnicity, age and employment status varies between the different subsamples.23 A limitation of this analysis is that we cannot be certain that the differences in the estimated causal effects that emerge are explained by the specific landlord/apartment characteristic used to divide the data or by other factors (other observed landlord/apartment characteristics or unobserved factors) that are correlated with this characteristic. This should be kept in mind when interpreting the results. An alternative approach would be to run a regression on the full sample including all observed landlord and apartment characteristics as well as their interactions with the applicant characteristics as explanatory variables. We have tried this alternative and get similar results. However, even if we use this approach, we cannot exclude the possibility that unobserved factors may affect the estimates.
Therefore, we focus on the results for the subgroups, which are reported in Table 4.
Panel A contains the results for the gender/ethnic differences in the callback rate.
Two factors appear important for the degree of discrimination: The ethnicity of the landlord (i.e. whether the landlord has a native Swedish or an ethnic minority name) and the size of the apartment. Concerning ethnicity, a striking result is that there is a substantial ethnicity gap in the callback rate for landlords with Swedish names, while there is almost no gap for landlords with ethnic minority names.24 In particular, this is the case for applicants with Arabic/Muslim male names where the gap is thirteen percentage points and strongly statistically significant for landlords with Swedish names, while the
23 We use t-tests to determine if the differences are statistically significant.
24 This result is in sharp contrast to the results in Ahmed and Hammarstedt (2008) who did not find any difference between landlords with Swedish and ethnic minority names.
gap is very close to zero for landlords with ethnic minority names (columns 3-4). For landlords with Swedish names, there is also a statistically significant gap in the callback rate between female applicants with Arabic/Muslim and Swedish names. In this case there is, however, no evidence that the gap is smaller for landlords with ethnic minority names. Moreover, there is some tendency – although not statistically significant – that female landlords discriminate less (columns 1-2). Also, there are essentially no differences between individual and company landlords (columns 5-6). All of these results remain essentially unchanged if we include the size of the apartment, the rent and the county where the apartment is located as control variables. Concerning the size of the apartment, applicants with Swedish female names – as opposed to applicants with Swedish male names – have a statistically significant higher callback rate for small apartments, while both applicants with male and female Arabic/Muslim names have a statistically significant lower callback rate for large apartments (columns 7-8).25 A t-test reveals that this effect is stronger for applicants with male than female Arabic/Muslim names. Panel B contains the corresponding results for the effects of the applicants’ age.
In Section 4.1, we found essentially no differences in the callback rate between the three age groups. However, when the data is divided with respect to landlord and apartment characteristics, a number of differences emerge. First, there are differences in the age effects depending on the gender and ethnicity of the landlord. The oldest age group (applicants aged 45-55) face a statistically significant penalty of five percentage points if the landlord has a female name (columns 1-2). In contrast, the youngest age group (applicants aged 25-35) has a statistically significant lower callback rate if the landlord
25 This is similar to the results in Ahmed and Hammarstedt (2008). All of these results, except for the negative effect for applicants with female an Arabic/Muslim name, remain essentially unchanged if we include the gender, ethnicity, company/individual landlord and county indicators.
has an ethnic minority name (columns 3-4). Second, there are some differences between individual and company landlords. In general, company landlords are more likely to respond, but they are also a statistically significant eight and nine percentage points more likely to respond to applicants in the two older age groups (columns 5-6).26
Panel C contains the corresponding results for the effects of the applicants’
employment status. The most noteworthy difference is that the positive effect of being employed is statistically significant larger when the landlord is a company rather than an individual (columns 5-6).
Overall, the results suggest that there are potentially important differences in the degree of discrimination depending on the type of landlord and apartment. This highlights the importance of both including many different types of landlords/apartments in correspondence studies, and taking information about these factors into account in the empirical analysis.
4.4 Regional characteristics
Another dimension where there may be important differences in the degree of discrimination is between different types of regions. To investigate this, we analyze how the measured degree of discrimination depends on the characteristics of the regions. As described above, we collected data on important economic, social and demographic characteristics at the municipality level (in Sweden, there are 290 municipalities).27 We do the analysis in a similar way as we did for the landlord and apartment characteristics, i.e. by dividing the sample into subsamples based on the regional characteristics and then
26 These results remain essentially unchanged if we include the size of the apartment, the rent and the county where the apartment is located as control variables.
27 There is substantial variation in most of these characteristics across the municipalities.
run the regressions on the subsamples. For each characteristic, we divide the municipalities into two groups of equal size based on the median, i.e. municipalities above the median and municipalities below the median. Then, we divide our sample of apartments into two subsamples based on in which group of municipalities (i.e. above or below the median) the apartment was located. As in the previous section, it should be kept in mind that the regional characteristics may be correlated with each other and with other unobserved variables. Table 5 present the results. For each municipality characteristic, the table contains two columns corresponding to below (left column) and above (right column) the median, respectively.
Again, panel A presents the results for the gender/ethnic groups. The gender effect remains small and not statistically significant in all subsamples, and there are no significant differences between municipalities above or below the median of the municipality characteristics. In contrast, the ethnicity effect varies quite a lot over the different subsamples.28 For the labor market characteristics, both applicants with male and female Arabic/Muslim names have a statistically significant lower callback rate in municipalities where the employment rate is below the median (columns 1-2). Also, applicants with male Arabic/Muslim names have a lower callback rate in municipalities where the share on income support is above the median (columns 3-4). For the population characteristics, applicants with male and female Arabic/Muslim names have a much lower callback rate when the share of highly educated is below the median or the average age is above the median (columns 5-8). Also, applicants with male Arabic/Muslim names have a statistically significant lower callback rate in municipalities where the share of
28 These results differ from the results in Hanson and Hawley (2011). They investigate the effects of a number of neighborhood characteristics and find that the only regional characteristic which has a statistically significant effect on the callback rate is being near ‘a tipping point’ in racial composition.
men is above the median (columns 9-10). Finally, there is more ethnic discrimination in municipalities where the share of immigrants – both total and from countries outside the EU – is low (columns 13-16). We will return to the results for the share of vacant apartments (columns 17-18) below.
In panel B, are the results for the age groups. Here, we find no statistically significant differences between municipalities above or below the median for any of the regional characteristics.
In panel C, are the results for the employment status groups. Here, there are some differences. For the labor market characteristics, the positive effect of being employed is statistically significant higher in municipalities where the employment rate is below the median (columns 1-2). For the demographic characteristics, the positive effect of being employed as a financial manager is statistically significant higher in municipalities where the average age is above the median (columns 7-8). Moreover, the positive effect of being employed is stronger in municipalities where the share of immigrants is below the median (columns 13-14).
A major concern about the results in Table 5 (in particular, the results regarding ethnic discrimination) is that they may simply reflect differences in supply and demand conditions across the municipalities. Such factors may have a strong impact on the measured degree of discrimination since it should be much easier for a landlord to discriminate if there are few vacant apartments available and/or many applicants to each vacant apartment. Such supply and demand factors are also likely to be correlated with many of the regional characteristics. To investigate the importance of this issue, we include the local share of vacant apartments in the regressions. This measure is likely to
be a good proxy variable for the supply and demand conditions in the rental market.
Therefore, if we include this factor as a control variable and find that the other results do not change, this should be a strong indication that the results are robust to this concern.
We start by running two regressions in a similar way as before where the sample is divided around the median of the share of vacant apartments (columns 17-18). The results show that the degree of ethnic discrimination is actually higher in municipalities where the share of vacant apartments is above the median. In the next step, we repeat the other regressions in Table 5, but with a control variable included for the share of vacant apartments in each municipality. Somewhat surprisingly, we find that the results in Table 5 are actually reinforced. Thus, supply and demand differences across the municipalities do not appear to explain our findings.
An alternative way of analyzing if there are regional differences in the measured degree of discrimination is to divide the sample based on the geographical location of the apartments. In Table 6, we estimate the degree of discrimination separately for the metropolitan areas (i.e. the three municipalities Stockholm, Gothenburg and Malmö), Stockholm County (i.e. the municipality of Stockholm and the surrounding suburban municipalities), the southern parts of Sweden, the central parts of Sweden (except Stockholm County), and the northern parts of Sweden.29
In panel A, are the results for the gender/ethnic groups. A striking result for the gender difference is that the callback rate in the central parts of Sweden is statistically significant higher for applicants with female Swedish names compared to applicants with male Swedish names (column 5). Otherwise, the gender results are very similar in all the
29 The last three groups correspond to a common division of Sweden where the counties the municipalities are located in are divided into three distinct regional groups: Götaland, Svealand and Norrland.
regional groups, i.e. mostly a small, but not statistically significant advantage for women (except in the northern parts of Sweden where the estimate is larger in absolute terms, but still not statistically significant). For the ethnic groups, the differences are much more striking. From the results, it is evident that ethnic discrimination mainly is a phenomenon outside the metropolitan areas in general and outside Stockholm County in particular (columns 2-6).30
In panel B, are the results for the age groups. Here, we essentially did not find any discrimination in the full sample, and this holds for the regional subsamples too.
In panel C are the results for the employment status groups. These results suggest that the positive effect of being employed is statistically significant higher outside the metropolitan areas and Stockholm County.
Overall, the results suggest that there are substantial differences in the degree of discrimination between different regions, and that these differences vary with the economic, social and demographic characteristics of the municipalities. This highlights the importance of both including apartments located in many different types of regions – e.g. metropolitan and other areas –, and taking information about these factors into account in the empirical analysis.
4. Concluding remarks
The emergence of field experiments – especially correspondence studies – represents a major methodological advancement for studying discrimination in the housing market since this approach makes it possible to circumvent the problems with unobserved heterogeneity and, hence, estimate causal effects. However, there is still room for
30 This is similar to the results in Ahmed and Hammarstedt (2008).
methodological improvements in the way these experiments are conducted since most existing studies tend to be very specific in nature; e.g. varying only one (or a few) applicant characteristics and considering only a limited number of landlords and apartments located within a limited geographical area.
In this study, we use an approach where all relevant applicant characteristics are randomly assigned to the applications, many different types of landlords/apartments located in many different types of regions are included, extensive information about the landlords/apartments and regions are collected, and the interactions between all these characteristics are carefully analyzed. To demonstrate how this approach can be implemented in practice, we conducted a large-scale field experiment on the largest Swedish classified advertisement website where individual and small company landlords search for tenants. In total, we applied for more than 5,800 vacant apartments located all over Sweden. Then, we analyzed how the callback rate (i.e. the probability of being invited to an apartment viewing) varied with the applicant, landlord, apartment and regional characteristics.
Our results show that the landlords in the experiment use ethnicity and employment status, but not gender and age, to sort applicants, but that these effects are very heterogeneous. This suggests that it may not be possible to summarize the degree of discrimination by one representative figure. First, there are several potentially important interaction effects between the applicant characteristics. This highlights the importance of varying a substantial number of distinct relevant applicant characteristics. Second, the degree of discrimination varies considerably with the characteristics of the landlords, apartments and regions. This highlights the importance of including many different types
of landlords and apartments located in many different types of regions, and carefully taking this information into account in the empirical analysis.
The correspondence testing methodology clearly has the potential to tell us something about the degree of discrimination in many markets. However, to reach its full potential, it is important that future studies are designed in a way which makes it possible to draw conclusions beyond the sample considered in each particular study. Today, when the Internet have made it easy to collect information about a large number of market transactions, it is difficult to find arguments for not conducting correspondence studies in the most broad and general way possible. For the rental market for apartments, this implies that future studies should use an approach which randomly assign all relevant applicant characteristics, include many different types of landlords/apartments located in many different types of regions, and carefully take into account the possible interactions between applicant, landlord/apartment and regional characteristics.
References
Ahmed, A., Andersson, L. and Hammarstedt, M. (2010), Can Discrimination in the Housing Market be Reduced by Increasing the Information about the Applicants?, Land Economics, 86: 79-90.
Ahmed, A. and Hammarstedt, M. (2008), Discrimination in the Rental Housing Market:
A Field Experiment on the Internet, Journal of Urban Economics, 64: 362-372.
Andersson, L., Jakobsson, N. and Kotsadam, A. (2012), A Field Experiment of Discrimination in the Norwegian Housing Market: Gender, Class, and Ethnicity, Land Economics, 88: 233-240.
Baldini, M. and Federici, M. (2011), Ethnic Discrimination in the Italian Housing Market, Journal of Housing Economics, 20: 1-14.
Bosch M., Carnero, A., and Farre, L. (2010), Information and Discrimination in the Rental Housing Market: Evidence from a Field Experiment.” Regional Science and Urban Economics, 40: 11-19.
Carlsson, M. and Eriksson, S. (2012), Do Reported Attitudes towards Immigrants Predict Ethnic Discrimination?, Working Paper 2012:6, Department of Economics, Uppsala University.
Carpusor, A. G. and Loges, W. E. (2006), Rental Discrimination and Ethnicity in Names, Journal of Applied Social Psychology, 36: 934-952.
Dymski, G. (2006), ‘Discrimination in the Credit and Housing Markets: Findings and Challenges’, in W. Rogers et al. (ed.), Handbook on the Economics of Discrimination, Edward Elgar Publishing, 320 pp.
Eriksson, S. and Rooth, D.-O. (2013), Do Employers Use Unemployment as a Sorting Criterion When Hiring? Evidence from a Field Experiment, American Economic Review, forthcoming.
Ewens, M., Tomlin, B. and Wang, C. (2012), Statistical Discrimination or Predjudice? A Large Sample Field Experiment, Review of Economics and Statistics, forthcoming.
Hanson, A. and Hawley, Z. (2011), Do Landlords Discriminate in the Rental Housing Market? Evidence from an Internet Field Experiment in U.S. Cities, Journal of Urban Economics, 70: 99-114.
Heckman, J. (1998), Detecting Discrimination, Journal of Economics Perspectives, 12: 101-116.
Heckman, J. and Siegelman, P. (1993), The Urban Institute Audit Studies: Their Methods and Findings, in Fix, M. E. and Struyk, R. J. (eds) Clear and Convincing Evidence:
Measurement of Discrimination in America, 187-258, Urban Institute Press .
Neumark, D. (2012), Detecting Discrimination in Audit and Correspondence Studies, Journal of Human Resources, 47: 1128-1157.
Ondrich, J., Ross, S. and Yinger, J (2000), How Common is Housing Discrimination?
Improving Traditional Measures, Journal of Urban Economics, 47: 470-500.
Page, M. (1995), Racial and Ethnic Discrimination in the Urban Housing Markets:
Evidence from a Recent Audit Study, Journal of Urban Economics, 38: 183-206.
Riach, P. and Rich J. (2002), ‘Field Experiments of Discrimination in the Market Place’, The Economic Journal, 112: 480–518.
Rooth, D.-O. (2011), Work Out or Out of Work: The Labor Market Return to Physical Fitness and Leisure Sports Activities, Labour Economics, 18, 399-409.
Wooldridge, J. (2012), Econometric Analysis of Cross Section and Panel Data, MIT Press.
Yinger, J. (1986), Measuring Racal Discrimination with Fair Housing Audits: Caught in the Act, American Economic Review, 76: 881-893.
Zhao, B. (2005), Does the Number of Houses a Broker Shows Depend on the Homeseeker’s Race?, Journal of Urban Economics, 57: 128-147.
Zhao, B., Ondrich, J. and Yinger, J. (2006), Why Do Real Estate Brokers Continue to Discriminate? Evidence from the 2000 Housing Discrimination Study, Journal of Urban Economics, 59: 394-419.
Table 1. Descriptive statistics about the average callback rates (in percent) Swedish
male name Swedish
female name Arabic/Muslim
male name Arabic/Muslim
female name Aged
25-35 Aged
36-45 Aged
46-55 Unemployed Shop sales
assistant Financial manager
Invitation to a viewing 26.1 28.1 14.2 18.5 20.9 23.3 21.33 14.75 22.6 28.1
Other positive response 11.5 11.6 9.0 10.4 10.6 10.9 10.4 9.6 10.7 11.7
Negative response 6.6 5.9 7.9 8.3 7.1 7.1 7.1 9.9 6.5 4.9
No response 55.8 54.5 69.1 62.8 61.4 58.7 61.1 65.8 60.2 55.3
All 100 100 100 100 100 100 100 100 100 100
Number of observations 1,464 1,485 1,413 1,465 1,913 1,842 2,072 1,959 1,913 1,955
Table 2. The callback rate
(1) (2) (3) (4)
Female Swedish name 0.02
[0.02] - - 0.03*
[0.02]
Male Arabic/Muslim name -0.12***
[0.01]
- - -0.11***
[0.01]
Female Arabic/Muslim name -0.08***
[0.02] - - -0.07***
[0.02]
Aged 36-45 - 0.02*
[0.01] - 0.02
[0.01]
Aged 45-55 - 0.00
[0.01] - 0.00
[0.01]
Shop sales assistant - - 0.08***
[0.01] 0.08***
[0.01]
Financial manager - - 0.13***
[0.01] 0.13***
[0.01]
Enjoying evenings at home - - - 0.07***
[0.01]
Engaged in sports - - - 0.09***
[0.01]
Smoker - - - -0.03***
[0.01]
Has reference - - - 0.02*
[0.01]
Constant
Number of observations
0.26***
[0.01]
5,827
0.21***
[0.01]
5,827
0.15***
[0.01]
5,827
0.13***
[0.02]
5,827 Notes: The dependent variable is the probability of being invited to an apartment viewing (i.e. the callback rate). The specifications are estimated with the linear probability model. The reference category for gender/ethnicity is an applicant with a male Swedish name, for age aged 25-35, for employment status unemployed, for leisure time interests enjoying restaurant life and nightclubs, for smoking habits non- smoker, and for reference no mention of a reference, respectively. ***, **, and * denote the 1, 5 and 10 percent significance levels, respectively. The reported standard errors (in parentheses) are robust.
