BUS 670 Final Paper Already Started, with outline and sources provided
equestrian14
discrimination_in_the_hiring_process.pdf
Empir Econ (2012) 43:537–563 DOI 10.1007/s00181-011-0496-6
Detecting discrimination in the hiring process: evidence from an Internet-based search channel
Stefan Eriksson · Jonas Lagerström
Received: 6 September 2010 / Accepted: 11 May 2011 / Published online: 12 August 2011 © Springer-Verlag 2011
Abstract This article uses data from an Internet-based CV database to study how job searchers’ ethnicity, employment status, age, and gender affect how often they are contacted by firms. Since we know which types of information that are available to the recruiting firms, we can handle some of the problems with unobserved hetero- geneity better than many existing discrimination studies. We find that searchers who have non-Nordic names, are unemployed or old get significantly fewer firm contacts. Moreover, this matters for the hiring outcome: searchers who get more contacts have a higher probability of getting hired.
Keywords Job search · Unobserved heterogeneity · Discrimination · Ethnicity · Employment status · Age
JEL Classification J64 · J71
1 Introduction
Studies of firms’ hiring behavior indicate that some employers use information about ethnicity, employment status, age and gender to sort applicants, and thus that discrim- ination is a feature in most labor markets. However, a crucial issue in discrimination studies is how to handle the fact that the recruiting firms typically have access to
S. Eriksson (B) Department of Economics, Uppsala University, PO Box 513, 751 20, Uppsala, Sweden e-mail: [email protected]
J. Lagerström School of Business and Economics, Åbo Akademi University, Fänriksgatan 3B, 20500, Åbo, Finland e-mail: [email protected]
123
538 S. Eriksson, J. Lagerström
much more detailed information about the job applicants than the researcher. This unobserved heterogeneity is difficult to take into account in studies using registry data. Therefore, some researchers use field experiments, but this approach also has some weaknesses. The emergence of Internet-based search channels gives research- ers a new way to study discrimination. In an online CV database, the initial contact between the job seeker and the firm is based only on the information in the CV. This means that the researcher will know which types of information that are available to the firms and, given that this information is handled correctly, can obtain accurate esti- mates of discrimination in the early stages of the hiring process. Thus, studies based on data from CV databases, together with decomposition and audit studies, should give us a better understanding of the extent and consequences of labor market discrimination.
The purpose of this article is to study how factors which may be used as a basis for discrimination, such as ethnicity, employment status, age, and gender affect how often job searchers are contacted by firms. We also investigate if searchers who get contacts through this search channel have a higher probability of getting hired, i.e., if the number of firm contacts received matter for the actual hiring outcome.
We use data from ‘My CV’ which is an Internet-based search channel provided by the Swedish Public Employment Service. Anyone who wants to find a job is invited to submit details—personal characteristics and requirements about the jobs—to the database. Recruiting firms are invited to search among the CVs in the database, and can contact searchers for interviews by e-mail within the system. Our dataset covers 18,167 job searchers, and includes the information that the searchers have entered about themselves and the number of firm contacts they have received through this search channel. We combine this data with data on hiring from the Employment Service.
Our dataset has several advantages. First and most importantly, we know which types of information that are available to the recruiting firms. We have access to essentially the same information as the firms (i.e., all information except for the search- ers’ names and short personal statements; see Sect. 2). Thus, even though there are some remaining issues (e.g., the information in the personal statements and interaction effects; see Sect. 3), we have much more control over the information observed by the firms than most studies using registry data. Second, our sample is quite large and includes a diverse pool of searchers looking for work in many occupations. In contrast to studies based on data from field experiments, our data reflect genuine job search and the types of jobs covered have not been chosen by us. Third, even though our focus is on the early stages of the hiring process, where firms are most likely to use easily observable characteristics to sort workers, we can also study if the number of firm contacts received matter for the hiring outcome.
We show that searchers who have non-Nordic names (especially Arabic names), are unemployed or old get fewer firm contacts. The results are both economically and statistically significant, and are supported by an extensive robustness analysis which checks for signs of problems with unobserved heterogeneity. Thus our results indicate that firms use ethnicity, employment status and age as sorting criterions, implying that discrimination is a feature in the Swedish labor market. We also show that searchers who receive contacts through ‘My CV’ have a higher probability of actually getting hired.
123
Detecting discrimination in the hiring process 539
Two related papers are Edin and Lagerström (2006) and Eriksson and Lagerström (2006) which use data from a more limited predecessor to ‘My CV’ to study discrimina- tion. Edin and Lagerström (2006) find that women have a lower contact probability than men, while searchers with foreign names do not have a lower contact probability than searchers with Swedish names. Eriksson and Lagerström (2006) find that unemployed searchers have a lower contact probability than employed searchers. Compared with the data used in the previous papers, our dataset is larger and more representative of the Swedish labor market. Also, our dataset has three major advantages. First, we have access to more detailed data on the job seekers’ ethnicity. This is important since we find that the results differ for searchers from different ethnic groups. To study ethnic discrimination in Sweden, it is crucial to distinguish immigrants from the other Nordic countries from other immigrants. This may explain why we, in contrast to Edin and Lagerström (2006), find evidence of ethnic discrimination. Second, the size of our dataset allows us to use very detailed control variables for occupation. Together with the fact that our dataset is more representative, this may explain why we, in contrast to Edin and Lagerström (2006), find less evidence of gender discrimination. Third, since we can link our data with data on hiring, we can study if contacts received through ‘My CV’ matter for the actual hiring outcome. This was not possible in the previous studies.
Our article is also related to the literature using field experiments to study dis- crimination.1 Typically in such studies fake job applications are designed and sent to recruiting firms, and the responses of the firms are then analyzed. This approach gives the researcher control over which information is observed by the firms, but has also been criticized; it may be difficult to construct good job applications, many studies are limited in size and focus on just a few selected occupations, and it may be argued that it is unethical to subject employers to fake job search (see e.g., Heckman 1998, Riach and Rich 2002, Lahey and Beasley 2009 for a discussion of the pros and cons of field experiments). Three examples of papers using data from field experiments to study ethnic discrimination are Bertrand and Mullainathan (2004), who find that searchers in the US with White-sounding names receive more callbacks for interviews than searchers with African-American sounding names, Carlsson and Rooth (2007), who find a negative effect for searchers with Arabic-sounding names in Sweden, and Oreopoulos (2009), who find a negative effect for searchers with Chinese, Indian or Pakistani names in Canada. Examples of similar studies of gender discrimination are Neumark et al. (1996), Riach and Rich (1997), Weichselbaumer (2004), and Carlsson (2011), which all find evidence of gender discrimination. Also, Lahey (2008) finds evidence of age discrimination in the US. Discrimination based on employment status is studied in, e.g., Belzil (1996) and Blau and Robins (1990).
The rest of this article is organized as follows. Section 2 introduces the dataset, presents descriptive statistics and discusses selection issues. Section 3 contains the analysis of how the job seekers’ characteristics affect the number of firm contacts they get, and includes a discussion of identification and estimation issues, a presentation of the results and a discussion of robustness issues. Section 4 contains the analysis
1 There are also some papers which analyze anonymous application procedures, see e.g., Goldin and Rouse (2000) and Åslund and Nordström Skans (2011).
123
540 S. Eriksson, J. Lagerström
of how the number of firm contacts received affects the probability of getting hired. Section 5 concludes.
2 Data
The database ‘My CV’ is a search channel offered to job seekers by the Swedish Public Employment Service. Anyone who wants to find a job, irrespective of current employment status, is invited to submit a CV to the database over the Internet or at the Employment Service. The searchers submit their information by entering their per- sonal details into a number of standardized forms. In the forms, they are asked to enter information about their education, labor market experience, language and computer skills, other skills, the requirements they have about the jobs they want to find, and are asked to write a short personal statement. The information is only made visible to employers if all forms have been completed. Employers are invited to search online in the database, and can contact searchers for interviews etc. by e-mail within the system. The registration process, the information in the database, and the way employers can use it is described in more detail in the Appendix.
All searchers—both new and previously registered—who logged into the system in December 2004 were asked if they wanted to participate in a research project on the recruitment behavior of firms. Around 40% agreed to participate and were also asked to immediately answer a short online questionnaire.2 All searchers who agreed to participate, answered the questionnaire and had profiles visible to employers are included in our sample, which consists of 18,167 searchers. We have data on every- thing the searchers have registered in ‘My CV’ (except for their names and personal statements), the number of firm contacts they have received and their subsequent employment histories.3 Table 1 presents descriptive statistics.
In Table 1, we see that the searchers are quite diverse: the average age is 34 years, 53% are women, 18% have foreign names, 27% are employed, 53% are unemployed, more than half have a post-secondary education, most workers have some labor market experience, and most workers search for work in the metropolitan areas.
The searchers in the sample have received 12,994 contacts from firms during their time in the database (a contact is an e-mail sent from a firm to a searcher, and typi- cally contains an invitation to an interview). Table 2 gives descriptive statistics on the contacts received. Also included in the table is data from the Employment Service on the fraction of searchers who have found jobs—using any search channel—after they have registered in ‘My CV’.
In Table 2, we see that 28% of the searchers have been contacted by an employer, and that the average number of contacts is 0.72. The average probability of finding a job at any time after registration (but before May, 2005) is 22%. Looking at the sub-
2 The details of the data collection are described in the Appendix. 3 Due to privacy concerns, the Employment Service did not give us access to the personal statements, except for their lengths. The statements may contain both information which is also registered in the other forms and new information. Of course, the statements may affect the firms’ contact decisions, but even if we had access to these statements it would be very difficult to control for their quality in an objective way. We deal extensively with this issue in the robustness analysis in Sect. 3.4.
123
Detecting discrimination in the hiring process 541
Table 1 Descriptive statistics about the searchers (in fractions)
Ethnicity Foreign name 0.18 Of which Nordic name 0.06 African name 0.003 Arabic name 0.02 Asian name 0.01 Other foreign name 0.09 Age Mean (years) 34.3 Age 20–25 0.27 Age 26–35 0.33 Age 36–50 0.28 Age 51– 0.12 Employment status Employed 0.27 Unemployed 0.53 Labor market program 0.05 University student 0.05 Other adult training 0.02 High school student 0.01 On parental leave 0.01 Other 0.06 Gender Female 0.53 Highest level of completed education Primary 0.13 Secondary 0.33 Post-secondary 0.54 Work experience Less than 1 year 0.15 1–2 years 0.12 2–5 years 0.19 5–10 years 0.14 10–15 years 0.10 15–20 years 0.09 More than 20 years 0.21 Limited relevant experience 0.28 Some relevant experience 0.39 Almost only relevant experience 0.33 Other skills Managerial experience 0.32 Foreign work experience 0.10 Telecommuting experience 0.11 Research experience 0.05 Driving license 0.77 Good language skills—Swedish 0.998 Good language skills—English 0.61 Good language skills—French 0.04 Good language skills—German 0.13 Good language skills—Spanish 0.04 Number of languages 3.5 Number of computer programs 2.3 Other skills 4.7 Region Stockholm 0.27
123
542 S. Eriksson, J. Lagerström
Table 1 continued
Notes: The ethnicity variable is based on a question in the questionnaire. Post-secondary education includes all types of education which is above the secondary level, i.e., university education, training in crafts etc. Relevant experience refers to experience in occupations that the searcher is looking for work in. Region and occupation refer to the regions and occupations where the searcher is looking for work and a searcher may look for work in several regions/occupations. The occupations are classified according to the system used by the Employment Service.
Uppsala 0.08 Södermanland 0.07 Östergötland 0.07 Jönköping 0.05 Kronoberg 0.04 Kalmar 0.04 Gotland 0.01 Blekinge 0.03 Skåne 0.17 Halland 0.07 Västra Götaland 0.19 Värmland 0.04 Örebro 0.06 Västmanland 0.06 Dalarna 0.04 Gävleborg 0.05 Västernorrland 0.05 Jämtland 0.02 Västerbotten 0.04 Norrbotten 0.03 Occupation Legislators, senior officials, and managers 0.04 Professionals 0.32 Technicians and associate professionals 0.31 Clerks 0.31 Service workers and shop sales workers 0.28 Skilled agricultural and fishery workers 0.04 Craft and related trades workers 0.13 Plant and machine operators and assemblers 0.13 Elementary occupations 0.31 Mean number of weeks in ‘My CV’ 50.4 Median number of weeks in ‘My CV’ 36.9
groups, we see that searchers who have Nordic names, are old, employed or male get more contacts than searchers who have non-Nordic names, are young, unemployed or female, respectively.
An important issue is how representative the sample is. In principle, there are three selection issues which we need to consider. First, the searchers who agreed to partici- pate in our study may differ from those who did not. Unfortunately, we do not have any information about the searchers who did not agree to participate. However, this should only affect our results if we fail to include important (observable) variables which are correlated with ethnicity, employment status, age or gender in the regressions, or if the regressions do not fully capture the potentially very complex way the employers use the information. We investigate this issue further in the robustness analysis (see Sect. 3.4). It should also be noted that when the searchers were asked if they wanted to participate in the study, the question did not spell out the exact details of the study.
Second, since both searchers and employers choose whether or not to use ‘My CV’, we may be concerned that those who use it differ from those who do not. To check if searchers in ‘My CV’ differ from other Swedish job seekers, Table A1 in the Appendix compares our searchers with data on other job seekers at the Employment Service. In the table, we see that the searchers are quite similar. The most noteworthy differ- ences are that our sample contains more women and searchers with post-secondary
123
Detecting discrimination in the hiring process 543
Table 2 Descriptive statistics about the contacts received and the probability of finding a job (in fractions)
Group Fraction receiving at least one contact
Average number of contacts received
Probability of finding a job after registration
All 0.28 0.72 0.22 Swedish name 0.28 0.73 0.23 Foreign name 0.27 0.65 0.18 Nordic name 0.30 0.80 0.21 African name 0.25 0.43 0.05 Arabic name 0.23 0.44 0.18 Asian name 0.18 0.38 0.12 Other foreign name 0.27 0.64 0.19 Age 20–25 0.22 0.46 0.21 Age 26–35 0.28 0.69 0.27 Age 36–50 0.32 0.89 0.22 Age 51– 0.34 1.00 0.14 Employed 0.35 0.98 0.40 Unemployed 0.24 0.57 0.18 Labor market program 0.24 0.61 0.12 University student 0.32 0.87 0.20 Other adult training 0.31 0.71 0.11 High school student 0.18 0.42 0.10 On parental leave 0.35 0.84 0.16 Male 0.28 0.84 0.23 Female 0.28 0.61 0.22 Primary education 0.33 0.97 0.17 Secondary education 0.23 0.48 0.21 Post-secondary education 0.30 0.80 0.25
Notes: The ‘probability of finding a job’ is the probability of being deregistered by the Employment Service because the searcher has found a job at any time after he or she registered in ‘My CV’ (but before May 9, 2005). The last column only includes searchers who are registered both in ‘My CV’ and at the Employment Service.
education, and that our searchers are less likely to search for work as service workers. Turning to the employers, our dataset does not include direct information about the employers who use the database; we only have data on the offers received by the searchers. However, to get a rough sense if the vacancies that employers try to fill using ‘My CV’ differ from other vacancies, Table A2 in the Appendix compares the preferred occupation/region of the searchers in ‘My CV’ who have received at least one contact with the inflow of vacancies to the Employment Service. This is obviously not an ideal measure, but it should give us a rough estimate since it is reasonable to expect that most employers contact searchers who have stated that they are interested in the same types of jobs as the firms are trying to fill. The comparison shows that the searchers in ‘My CV’ who have been contacted are more likely to search for jobs as clerks, and less likely to search for jobs as service workers. To see if such differ- ences matter for our results, we run separate regressions for different subgroups in our empirical analysis (see Sect. 3.3).
Third, it may be that employers who use ‘My CV’ have access to less informa- tion than employers using other search channels, and therefore are more likely to use
123
544 S. Eriksson, J. Lagerström
indicators, such as ethnicity, to sort workers. If this is the case, our results may over- estimate the true effects for the whole labor market. However, it is not obvious that employers who use ‘My CV’ have access to less information than employers who, e.g., choose among written applications.
3 The number of contacts received
We want to investigate how the job searchers’ characteristics affect the number of firm contacts they get. In this section, we discuss identification issues, define the variables and the econometric specification, present the results, and discuss robustness issues.
3.1 Identification
Suppose that an employer has chosen to use ‘My CV’ to fill a vacancy. The employer obviously wants to locate the most productive worker. However, a lot of factors will affect the productivity of an applicant in a particular job, and only some of these factors are directly observable in the database. Which characteristics should we expect such an employer to consider relevant? Probably, the employer will consider two types of information. First, all factors that he or she believes directly will affect the pro- ductivity of the applicants, e.g., education and work experience. Second, all factors (e.g., ethnicity, employment status, age and gender) that the employer believes are indicators for other important factors (e.g., ability to cooperate, motivation, and other social skills) that are unobservable when the decision is made. Typically, these indica- tors are not important for productivity by themselves, but instead serve as indicators for unobservable characteristics (c.f. statistical discrimination). Also, employers may have preferences over the indicators directly (c.f. preference-based discrimination).
Now, how can we identify the effect of the searchers’ characteristics on the number of firm contacts received? Here it is important to note that the information about the searchers in ‘My CV’ to a high degree coincide with the information set of the firm which could potentially contact the applicant (we have access to all information about the searchers except for their names and personal statements). The firm acts on the basis of the expected value of the applicants’ ability conditional on observable attri- butes, and the latent variable may, therefore, be viewed as the expected ability of the applicant conditional on his or her observable characteristics. Thus, we can write the econometric model as a regression function with regressors which are orthogonal to the error term.
3.2 Variables and estimation
To identify the effects of ethnicity, employment status, age, and gender on the contact rate, we need to control for all other factors which may affect these decisions. In the estimation, we only use data directly visible to the employers, except for ethnicity where we use data from a question in the questionnaire (see below).
123
Detecting discrimination in the hiring process 545
For education, we use variables for the highest level of completed education; primary, secondary and post-secondary. For experience, we use variables for seven lengths of experience; less than 1, 1–2, 2–5, 5–10, 10–15, 15–20, and more than 20 years.4 Also, we use variables for the fraction of the experience which is in the occu- pations where the searcher is looking for work; limited, some or almost all. For skills, we use variables for managerial experience, foreign work experience, telecommut- ing experience, research experience, driving skills, computer skills, language skills in Swedish and other languages, and other skills.
For ethnicity, we use variables for Swedish and foreign names, dividing the second group into searchers with Nordic, African, Arabic, Asian, and other foreign names. Often we combine the last four groups into searchers with non-Nordic names. We get the data on ethnicity from a question in the questionnaire.5 However, since employ- ers have access to the searchers’ names, we expect that very similar information is available to them. For age, we divide the searchers into five groups; less than 26, 26–35, 36–50, and over 50-years old.6 For gender, we use naturally defined variables. For employment status, we divide the searchers into eight groups; employed, unem- ployed, participant in a labor market program, university student, participant in other adult training, high school student, on parental leave, and others.
To capture differences across occupational and regional labor markets, we include variables for occupation and region. This is important since we expect that an em- ployer’s choice usually will be limited to searchers who have stated that they are interested in jobs in a particular occupation at a particular location. For occupation, we use the Employment Service’s classification system to divide the searchers into 114 occupations.7 For location, we divide the searchers into 23 regions (21 counties, all of Sweden and abroad).
We also need to include controls for the length of time searchers have been reg- istered in ‘My CV’, since those who have been in the database longer, on average, have received more contacts. Thus, we include a vector of the variables time and time squared (time is measured in weeks). In Sect. 3.4, we consider several alternatives as a robustness check.
Our data is count data and the conventional model used to analyze such data is the Poisson model. This model requires that events occur randomly over time. In our case, we believe that the ‘memoryless’ feature of the Poisson model is appropriate for several reasons. First, ‘My CV’ contains a large number of job seekers and thus it is reasonable to expect that most employers will not remember individual job seekers from one search to the next. Second, it is likely that most employers recruit workers only occasionally. Third, information on how long a searcher has been registered in the database is not available to the employers who use it. In particular, employers cannot specify search criteria based on the time searchers have been registered in the
4 We use this categorization since we received the data classified into these seven groups. However, we have tried other alternatives and all results remain unchanged. 5 Ethnicity is not registered in ‘My CV’, but in the questionnaire the searchers were asked how they believe employers perceive their name (see the Appendix for details). 6 The results are unchanged if we include age as a continuous variable. 7 The results are unchanged if we use broader occupational categories.
123
546 S. Eriksson, J. Lagerström
database. Also, employers cannot observe whether other employers have looked at a searcher’s CV. To investigate if our results are robust, we consider several alternatives, e.g., the negative binomial model. Also, we estimate the probability of receiving at least one contact using the Probit model. However, we prefer to use the count data models since these models use all the variation in the dependent variable (i.e., that some searchers have received more than one contact).
3.3 Results
Table 3 presents the results for the number of firm contacts received. In column 1 in Table 3, we only include factors which firms may use as indicators
for unobservable characteristics, i.e., we include ethnicity, employment status, age and gender (the time vector is included in all regressions). We see that searchers who have non-Nordic names, are women, unemployed or participate in labor market pro- grams get fewer contacts. This indicates that firms use these factors to sort workers, but we cannot exclude the possibility that these effects are explained by other differ- ences across groups. In columns 2–4, we therefore successively introduce variables for other observable factors. In column 2, we add the skill variables (education, experience, language skills etc.), and see that many of these factors have strong positive effects on the contact rate; all measures of education and experience are highly significant. In columns 3 and 4, we add the variables for the requirements the searchers have about the occupation and region of the desired jobs, and see that these factors also matter. In column 4, we see that the contact rate is lower for searchers with non-Nordic names, older searchers, and searchers who are unemployed or participate in labor market programs, even though the effect from being old or a program participant is signifi- cant only at the 10% level. The negative effect for women disappears when we control for the search region.8 Since it is likely that labor market conditions differ across both occupations and regions, this indicates that women to a larger extent than men search for work in weak labor markets. In addition, the women in our sample, on average, search for work in fewer regions than the men, and this may also partially explain why they get fewer contacts. In column 5, we divide the searchers with non-Nordic names into subgroups based on ethnicity, and find that searchers with Arabic or other foreign names get fewer contacts. For searchers with African or Asian names the negative point estimates are not statistically significant, but this may be because these groups are small.
Since we have data on most of the information firms observe prior to hiring, our results suggest that firms use ethnicity, employment status and age to sort workers and thus that discrimination is a feature in the Swedish labor market. The magnitudes of the effects are substantial. The relative effects are between 13 and 23%. Since the average number of contacts received is 0.72, this implies that a searcher who has a
8 This variable measures in which region(s) the searcher is looking for work, and not the region of resi- dence. If we use the region of residence in the regressions, we find that the negative effect for women is similar to the result in column 3.
123
Detecting discrimination in the hiring process 547
Table 3 Poisson estimates of the number of contacts received
(1) (2) (3) (4) (5)
Non-Nordic name –0.111* –0.223*** –0.211*** –0.228*** – (0.059) (0.059) (0.052) (0.050)
Nordic name – – – – –0.030 (0.063)
African name – – – – –0.110 (0.358)
Arabic name – – – – –0.286** (0.134)
Asian name – – – – –0.184 (0.173)
Other foreign name – – – – –0.228*** (0.054)
Age 26–35 −0.015 −0.295*** −0.116** −0.125** −0.124** (0.049) (0.059) (0.057) (0.054) (0.054)
Age 36–50 0.065 −0.495*** −0.145* −0.134* −0.132* (0.052) (0.083) (0.078) (0.074) (0.074)
Age 51– 0.004 −0.563*** −0.174* −0.166* −0.165* (0.065) (0.106) (0.096) (0.091) (0.091)
Unemployed −0.203*** −0.135*** −0.129*** −0.125*** −0.125*** (0.041) (0.040) (0.037) (0.035) (0.035)
Labor market program −0.280*** −0.160* −0.103 −0.129* −0.128* (0.093) (0.092) (0.081) (0.077) (0.077)
University student 0.075 0.065 0.076 0.071 0.071 (0.082) (0.082) (0.077) (0.075) (0.075)
Other adult training −0.028 0.076 −0.076 0.003 0.003 (0.097) (0.094) (0.099) (0.095) (0.095)
High school student 0.126 0.256 −0.029 0.045 0.045 (0.248) (0.243) (0.193) (0.189) (0.189)
On parental leave 0.043 0.028 0.107 0.161 0.162 (0.147) (0.147) (0.132) (0.128) (0.126)
Other 0.032 0.046 −0.051 −0.067 −0.067 (0.083) (0.075) (0.069) (0.065) (0.065)
Female −0.178*** −0.072** −0.101** 0.019 0.018 (0.034) (0.034) (0.041) (0.038) (0.038)
Secondary education – 0.121** 0.088* 0.063 0.063 (0.056) (0.053) (0.049) (0.049)
Post-secondary education – 0.164*** 0.142*** 0.086* 0.086* (0.050) (0.049) (0.045) (0.045)
1–2 years experience – 0.111 0.096 0.132** 0.130** (0.072) (0.066) (0.065) (0.065)
2–5 years experience – 0.347*** 0.317*** 0.334*** 0.332*** (0.068) (0.063) (0.061) (0.061)
5–10 years experience – 0.387*** 0.321*** 0.336*** 0.333*** (0.075) (0.071) (0.069) (0.069)
10–15 years experience – 0.473*** 0.316*** 0.378*** 0.377*** (0.083) (0.079) (0.078) (0.078)
15–20 years experience – 0.646*** 0.447*** 0.538*** 0.534*** (0.101) (0.088) (0.085) (0.085)
20– years experience – 0.556*** 0.316*** 0.405*** 0.402*** (0.099) (0.089) (0.087) (0.087)
Some relevant experience – 0.178*** 0.104** 0.118*** 0.118*** (0.045) (0.042) (0.041) (0.041)
123
548 S. Eriksson, J. Lagerström
Table 3 continued
(1) (2) (3) (4) (5)
Almost only relevant – 0.127*** 0.107** 0.131*** 0.131*** experience (0.047) (0.043) (0.043) (0.043) Other skills No Yes Yes Yes Yes Dummies for occupation No No Yes Yes Yes Dummies for region No No No Yes Yes Number of observations 18,167 18,167 18,167 18,167 18,167 R2 0.26 0.26 0.39 0.42 0.42
Notes: The regressions also include the time vector (time and time squared in weeks) and a constant. The reference category is an employed man who is less than 26-years old, has a Nordic name (Swedish name in column 5), primary education, and less than one year of non-relevant experience. Robust standard errors are in parentheses. ***, **, and * denote significance at the 1, 5, and 10% level.
non-Nordic name, are unemployed or are over 50-years old get around 0.16, 0.09, and 0.12 fewer contacts, respectively.
The results are statistically significant at conventional levels and remain stable across different specifications and estimation methods (e.g., using the negative bino- mial model or the Poisson QML model). Also, the results are similar if we restrict the sample to searchers who have received at least one contact, or if we treat the unemployed and labor market program participants as one category.9
We have also run regressions including interaction effects and separate regressions for different subgroups. Table 4 presents results of regressions with interaction effects included.10
In column 1 in Table 4, we include interaction effects between the searchers’ per- sonal characteristics (i.e., ethnicity, employment status, age, gender and education). From the results, we see that only one interaction effect is statistically significant, and that all our main results remain qualitatively unchanged. The interaction effect between non-Nordic name and female is positive (significant at the 10% level), but this effect disappears when we include more interaction effects (i.e., between non-Nordic name and education/occupation). In column 2, we also include interaction effects between the searchers’ characteristics and their preferred occupation (i.e. white-collar and blue-collar).11 From the results, we see that two interaction effects are statistically significant; there are negative effects for non-Nordic and female searchers who look for white-collar jobs. The estimate for non-Nordic name is still negative, but is no longer statistically significant. These results suggest that it is non-Nordic and female job seekers searching for white-collar jobs who are most likely to face discrimination.
Table 5 presents results of regressions for different subgroups. Comparing the results for the different subgroups the following are worth noting:
(1) Searchers with non-Nordic names get fewer contacts in most groups. The results
9 Since unemployed searchers do not always update their information when they enter a program firms may not be able to distinguish program participants from other unemployed searchers. 10 More results including interaction effects are available from the authors upon request. 11 The results are similar if we use other categorizations of the occupations, e.g., the nine categories used by the Employment Service.
123
Detecting discrimination in the hiring process 549
Table 4 Poisson estimates of the number of contacts received, interaction effects
Baseline (1) (2)
Non-Nordic name −0.228*** −0.216** −0.090 (0.050) (0.101) (0.104)
Age 26–35 −0.125** −0.119** −0.107** (0.054) (0.054) (0.053)
Age 36–50 −0.134* −0.126* −0.114* (0.074) (0.074) (0.074)
Age 51– −0.166* −0.137 −0.183 (0.091) (0.116) (0.138)
Unemployed −0.125*** −0.153** −0.150** (0.035) (0.063) (0.069)
Female 0.019 0.063 0.155** (0.038) (0.061) (0.064)
Secondary education 0.019 0.051 0.047 (0.038) (0.050) (0.050)
Post-secondary education 0.063 0.074 0.068 (0.049) (0.064) (0.082)
Non-Nordic × female – 0.156* 0.134 (0.094) (0.092)
Non-Nordic × age 51– – 0.017 0.031 (0.161) (0.164)
Non-Nordic × unemployed – –0.052 −0.069 (0.095) (0.094)
Non-Nordic × post-secondary – –0.099 −0.000 education (0.094) (0.102)
Female × age 51– – –0.097 −0.010 (0.083) (0.062)
Female × unemployed – 0.004 −0.068 (0.062) (0.087)
Female × post-secondary education – −0.076 −0.025 (0.061) (0.066)
Age 51– × unemployed – −0.079 −0.081 (0.085) (0.086)
Age 51– × post-secondary education − 0.093 0.093 (0.084) (0.086)
Unemployed × post-secondary – 0.073 0.070 education (0.062) (0.065)
White collar × non-Nordic – – −0.252** (0.099)
White collar × female – – −0.171*** (0.058)
White collar × age 51– – – 0.061 (0.096)
White collar × unemployed – – 0.004 (0.060)
White collar × post-secondary – – −0.024 (0.066)
Number of observations 18,167 18,167 18,167 R2 0.42 0.42 0.42
Notes: The regressions also include the time vector (time and time squared in weeks), a constant and all other variables included in Table 3. The reference category is an employed man who is less than 26-years old, has a Nordic name and primary education. ‘White-collar’ is occupation groups 1–3, and ‘blue-collar’ is occupation groups 4–9. Robust standard errors are in parentheses. ***, **, and * denote significance at the 1, 5, and 10% level.
123
550 S. Eriksson, J. Lagerström
T ab
le 5
P oi
ss on
es ti
m at
es of
th e
nu m
be r
of co
nt ac
ts re
ce iv
ed ,s
ub gr
ou ps
B as
el in
e W
om en
M en
A ge
≥ 40
A ge
< 40
U ne
m pl
oy ed
N on
-N or
di c
na m
e N
or di
c na
m e
B or
n in
S w
ed en
N on
-N or
di c
na m
e −0
.2 28
** *
−0 .1
39 **
−0 .2
61 **
* −0
.3 25
** *
−0 .1
29 **
−0 .2
66 **
* –
– −0
.1 59
* (0
.0 50
) (0
.0 66
) (0
.0 71
) (0
.0 80
) (0
.0 60
) (0
.0 71
) (0
.0 92
) A
ge 26
–3 5
−0 .1
25 **
−0 .0
88 −0
.1 10
– –
0. 05
1 −0
.2 42
** −0
.1 26
** *
−0 .1
78 **
* (0
.0 54
) (0
.0 63
) (0
.0 91
) (0
.0 86
) (0
.1 17
) (0
.0 36
) (0
.0 69
) A
ge 36
–5 0
−0 .1
34 *
−0 .1
53 −0
.1 03
– –
0. 12
8 −0
.2 90
** −0
.1 39
** *
−0 .1
67 **
* (0
.0 74
) (0
.0 94
) (0
.1 15
) (0
.1 12
) (0
.1 43
) (0
.0 50
) (0
.1 05
) A
ge 51
– −0
.1 66
* −0
.2 30
* −0
.1 52
– –
0. 03
1 −0
.0 24
−0 .1
61 **
* −0
.1 11
(0 .0
91 )
(0 .1
21 )
(0 .1
36 )
(0 .1
32 )
(0 .1
89 )
(0 .0
60 )
(0 .1
25 )
U ne
m pl
oy ed
−0 .1
25 **
* −0
.1 02
** −0
.1 29
** *
−0 .1
10 **
−0 .1
65 **
* –
−0 .2
65 **
* −0
.1 20
** *
−0 .1
73 **
* (0
.0 35
) (0
.0 47
) (0
.0 50
) (0
.0 51
) (0
.0 45
) (0
.0 78
) (0
.0 22
) (0
.0 48
) F
em al
e 0.
01 9
– –
−0 .0
04 0.
08 2*
0. 02
6 −0
.0 42
−0 .0
21 −0
.0 10
(0 .0
38 )
(0 .0
59 )
(0 .0
49 )
(0 .0
57 )
(0 .0
82 )
(0 .0
25 )
(0 .0
51 )
N um
be r
of ob
se rv
at io
ns 18
,1 67
9, 61
8 8,
54 9
5, 59
1 12
,5 76
9, 68
3 2,
25 1
15 ,9
16 11
,3 84
R 2
0. 42
0. 37
0. 47
0. 45
0. 40
0. 41
0. 44
0. 40
0. 38
L ow
H ig
h W
hi te
- B
lu e-
F em
al e
M al
e S
w ed
is h
Im m
ig ra
nt N
on -i
m m
. ed
uc at
io n
ed uc
at io
n co
ll ar
co ll
ar oc
cu pa
ti on
oc cu
pa ti
on oc
cu pa
ti on
co un
ty co
un ty
N on
-N or
di c
na m
e −0
.1 64
** −0
.2 59
** *
−0 .2
99 **
* −0
.1 12
** −0
.0 98
−0 .2
24 **
−0 .5
00 **
−0 .1
80 **
* −0
.3 36
** *
(0 .0
75 )
(0 .0
63 )
(0 .0
62 )
(0 .0
56 )
(0 .0
90 )
(0 .0
98 )
(0 .2
47 )
(0 .0
58 )
(0 .0
83 )
A ge
26 –3
5 −0
.0 82
−0 .1
25 *
−0 .1
81 **
−0 .0
86 −0
.0 21
−0 .0
02 −0
.2 42
−0 .1
85 **
* −0
.0 62
(0 .0
82 )
(0 .0
71 )
(0 .0
75 )
(0 .0
58 )
(0 .0
88 )
(0 .1
74 )
(0 .2
11 )
(0 .0
68 )
(0 .0
79 )
A ge
36 –5
0 −0
.1 32
−0 .1
33 −0
.1 72
* −0
.1 79
** −0
.0 04
0. 15
6 −0
.3 61
−0 .1
00 −0
.2 01
* (0
.1 11
) (0
.0 97
) (0
.1 00
) (0
.0 84
) (0
.1 25
) (0
.1 89
) (0
.2 45
) (0
.0 93
) (0
.1 09
) A
ge 51
– −0
.2 63
** −0
.1 23
−0 .1
60 −0
.2 12
** −0
.2 20
0. 13
6 −0
.4 45
−0 .1
48 **
−0 .2
45 *
(0 .1
32 )
(0 .1
18 )
(0 .1
17 )
(0 .1
06 )
(0 .1
65 )
(0 .2
03 )
(0 .2
81 )
(0 .1
13 )
(0 .1
36 )
123
Detecting discrimination in the hiring process 551
T ab
le 5
co nt
in ue
d
L ow
H ig
h W
hi te
- B
lu e-
F em
al e
M al
e S
w ed
is h
Im m
ig ra
nt N
on -i
m m
. ed
uc at
io n
ed uc
at io
n co
ll ar
co ll
ar oc
cu pa
ti on
oc cu
pa ti
on oc
cu pa
ti on
co un
ty co
un ty
U ne
m pl
oy ed
−0 .1
96 **
* −0
.0 70
−0 .1
14 **
* −0
.1 58
** *
−0 .1
33 **
−0 .1
64 **
−0 .2
89 **
* −0
.1 45
** *
−0 .1
17 **
(0 .0
52 )
(0 .0
47 )
(0 .0
44 )
(0 .0
41 )
(0 .0
61 )
(0 .0
73 )
(0 .1
11 )
(0 .0
44 )
(0 .0
55 )
F em
al e
0. 18
6* **
−0 .0
58 −0
.0 43
0. 05
1 0.
05 1
0. 02
4 0.
05 2
0. 01
1 0.
07 0
(0 .0
60 )
(0 .0
47 )
(0 .0
46 )
(0 .0
47 )
(0 .0
87 )
(0 .0
88 )
(0 .1
83 )
(0 .0
47 )
(0 .0
61 )
N um
be r
of ob
se rv
at io
ns 8,
37 7
9, 79
0 9,
33 0
13 ,4
91 5,
55 5
4, 46
6 1,
64 9
10 ,2
34 7,
93 3
R 2
0. 44
0. 41
0. 42
0. 42
0. 36
0. 47
0. 45
0. 41
0. 45
N o
te s:
T he
re gr
es si
on s
al so
in cl
ud e
th e
ti m
e ve
ct or
(t im
e an
d ti
m e
sq ua
re d
in w
ee ks
), a
co ns
ta nt
an d
al l
ot he
r va
ri ab
le s
in cl
ud ed
in T
ab le
3. T
he re
fe re
nc e
ca te
go ry
is an
em pl
oy ed
m an
w ho
is le
ss th
an 26
-y ea
rs ol
d an
d ha
s a
N or
di c
na m
e. ‘B
or n
in S
w ed
en ’
in cl
ud es
se ar
ch er
s w
ho ar
e re
gi st
er ed
at th
e E
m pl
oy m
en t
S er
vi ce
an d
w h
o ar
e bo
rn in
S w
ed en
. ‘L
ow (h
ig h)
ed uc
at io
n’ is
pr im
ar y/
se co
nd ar
y ed
uc at
io n
(p os
t- se
co nd
ar y)
. ‘W
hi te
-c ol
la r’
is oc
cu pa
ti on
gr ou
ps 1–
3, an
d ‘b
lu e-
co ll
ar ’
is oc
cu pa
ti on
gr ou
ps 4–
9. ‘F
em al
e oc
cu pa
ti on
’/ ‘m
al e
oc cu
pa ti
on ’/
‘S w
ed is
h oc
cu pa
ti on
’ in
cl ud
e se
ar ch
er s
w ho
on ly
se ar
ch in
oc cu
pa ti
on s
w he
re at
le as
t 70
% of
th e
se ar
ch er
s be
lo ng
to th
e re
sp ec
ti ve
gr ou
ps .
‘I m
m ig
ra nt
co un
ti es
’ ar
e th
e fi
ve co
un ti
es w
it h
m os
t im
m ig
ra nt
s, an
d ‘n
on -i
m m
ig ra
nt co
un ti
es ’
ar e
al l
ot he
r co
un ti
es .
S ea
rc he
rs ar
e al
lo w
ed to
se ar
ch fo
r w
or k
in m
or e
th an
on e
oc cu
pa ti
on an
d, th
er ef
or e,
th ey
ca n
be in
cl ud
ed in
m or
e th
an on
e su
bg ro
up .R
ob us
t st
an da
rd er
ro rs
ar e
in pa
re nt
he se
s. **
*, **
,a nd
* de
no te
si gn
if ic
an ce
at th
e 1,
5, an
d 10
% le
ve l.
123
552 S. Eriksson, J. Lagerström
indicate that the negative effect is especially strong for searchers who are over 40-years old, are looking for work in high-skill occupations, or are looking for work in occu- pations where most searchers (in ‘My CV’) have Swedish names. Also, the negative effect from having a non-Nordic name is substantial for searchers who are born in Sweden (i.e., second-generation immigrants). We have also analyzed the effects for different ethnic groups separately. Since some ethnic groups are small in the subsam- ples, it is difficult to get statistically significant results, but the results indicate that the negative effect may be especially strong for women with Arabic names.12 (2) In most subgroups, there is a strong negative effect from being unemployed. The results suggest that this effect is especially strong for searchers with non-Nordic names. (3) In most subgroups, there is a negative effect for older searchers. The results indicate that this effect is especially strong for searchers with low education. (4) The results indicate that, compared with men, high-skilled women have a lower contact rate. The result that the gender effect is stronger for women looking for white-collar jobs is statistically significant (c.f. column 2 in Table 4).
Also, we find that most coefficients for the skill variables are similar to the esti- mates in Table 3, except that the effect of education is stronger for workers with non-Nordic names. This may reflect that the average primary education is short in many non-Nordic countries. We have also run separate regressions for occupations which typically involve extensive contacts with customers, but find no noteworthy differences.
To summarize, we find that searchers with non-Nordic names (especially Arabic names), unemployed searchers and older searchers get fewer firm contacts. The effects are both economically and statistically significant; the relative effects are between 13 and 23%.
3.4 Robustness
To evaluate the robustness of the results, we have performed a number of robustness checks. Some of these results are presented in Table 6.
First, we may be concerned that we have not managed to properly control for all the information firms use when they make their contact decisions. As mentioned before, we have access to the same information as the firms, except for the searchers’ names and personal statements. Thus, unobserved heterogeneity may be a problem if the statements matter a lot or if firms use the information in the database in a way that we cannot capture in a regression analysis. One way of testing for the importance of the statements is to control for their length.13 We have estimated the model including a variable measuring the number of words in the statements as well as run separate regressions for searchers with short and long statements. We find that a long statement has a positive, although not statistically significant, effect on the number of contacts
12 The results for the ethnic groups are not displayed in Table 5, but are available from the authors upon request. 13 An alternative would be to control for the number of misspelled words. However, we do not have access to such data. However, Edin and Lagerström (2006) had access to such data and showed that this did not affect their results.
123
Detecting discrimination in the hiring process 553
Table 6 Poisson estimates of the number of contacts received, robustness
Baseline (1) (2) (3) (4)
Non-Nordic name −0.228*** −0.227*** −0.134 −0.224*** −0.170** (0.050) (0.050) (0.089) (0.048) (0.070)
Age 26–35 −0.125** −0.124** −0.229** −0.120** −0.056 (0.054) (0.054) (0.099) (0.052) (0.078)
Age 36–50 0.128 0.134* −0.145 −0.138* 0.034 (0.112) (0.074) (0.123) (0.072) (0.112)
Age 51– −0.166* −0.163* −0.191* −0.187** −0.119 (0.091) (0.091) (0.144) (0.089) (0.144)
Unemployed −0.125*** −0.126*** −0.162*** −0.108*** −0.255*** (0.035) (0.035) (0.061) (0.034) (0.057)
Female 0.019 0.023 −0.025 0.026 0.026 (0.038) (0.038) (0.064) (0.037) (0.060)
Long statement – 0.048 – – – (0.033)
Previous (log) wage – – −0.016 – – (0.018)
Number of observations 18,167 18,167 7,313 18,167 12,102 R2 0.42 0.42 0.41 0.41 0.41
Notes: The regressions also include a constant and all other variables included in Table 3. The reference category is an employed man who is less than 26-years old, and has a Nordic name. In the columns baseline, 1 and 2, we include a time vector (time and time squared in weeks). In column 3, we use dummy variables for the time in the database (10-week periods) instead of the time vector. In column 4, we only include searchers who have been in the base less than 52 weeks. Robust standard errors are in parentheses. ***, **, and * denote significance at the 1, 5, and 10% level.
received, but that the inclusion of this variable does not affect any of the other results (see column 1 in Table 6). A related concern is that searchers with foreign names have poor language skills in Swedish, and that this is reflected in their statements. However, in Table 5, we see that most of the negative effect from having a non-Nordic name remains even if we estimate the model on a sample only including searchers who are born in Sweden. Since it is natural to assume that second-generation immigrants have better Swedish skills than first-generation immigrants, this indicates that the lower con- tact rate is not explained by non-Nordic immigrants writing lower quality statements. Also, we control for the searchers’ self-reported Swedish skills in the regressions.
Another way of testing for unobserved heterogeneity is to include the searchers’ pre- vious wages as a regressor. We can do this for a subsample consisting of unemployed searchers (this data are from a separate register used to determine unemployment benefits). This variable is not observable to the employers who use ‘My CV’, but it can be used as a test of whether we have managed to properly control for all the information which are observable to them since the previous wage should be highly correlated with all important dimensions of ability. If we include the previous wage as a regressor and find that it is not statistically significant, this is an indication that our control variables capture most major differences across searchers, i.e., the previous wage should only be statistically significant, if we have failed to control for important factors observable to the firms. Including the previous wage in our baseline regression, we find that it is statistically insignificant. Including this variable weakens some of
123
554 S. Eriksson, J. Lagerström
the other effects (in particular the effect from having a non-Nordic name), but most results remain qualitatively unchanged (see column 2 in Table 6). This suggests that our regressions capture most of the relevant differences.
Second, an issue that may cause problems is the stock-flow aspect of the sample. The searchers have been in the database for different lengths of time as people enter and leave the database continuously (remember that our sample consists of both new and previously registered users of ‘My CV’). In the regressions, we have included a time vector, consisting of time and time squared, to take into account that a searcher who has been registered in the database longer is more likely to have received more contacts. To test whether the way we control for the time in the database matters, we have tried a number of alternatives to the baseline specification:14 First, we have divided the sample into groups based on the searchers’ time in the database (e.g., 10 week-groups) and then included dummies for the groups as regressors instead of the time vector. Second, we have split the sample into searchers who have been in the database for a long or short time and run separate regressions. Some results of this sensitivity analysis are presented in columns 3 and 4 in Table 6.
In Table 6, we see that most results remain qualitatively unchanged irrespectively of how we control for time. The coefficients on non-Nordic name, unemployed and age over 50 are always negative, and mostly statistically significant. We have also esti- mated the model on samples only including searchers who have been in the database for a very short time (less than 2, 4, 6, 8, and 10 weeks). Using this approach, we can be fairly certain that very few searchers have deleted their CVs during the time period considered. Here, we find that the point estimates that we are interested in always remain negative, but often turn out to be statistically insignificant. However, there is no clear trend in the size of the estimates when we reduce the time that we allow the searchers to have been in the database. Also, we have run the regressions with the number of contacts per week as the dependent variable and get similar results; having a non-Nordic name still has a clear negative effect. Taken together all this indicates that the way we control for time is not crucial for our main results.15
Third, we may worry that the same contact attempt could result in more than one contact, and thus that our dependent variable is biased upwards. To test whether this is a problem, we have run all regressions with the probability of getting at least one contact as the dependent variable using the Probit model. The results are in the Appen- dix, and we see that most results are similar, except that being a woman now has a positive effect on the probability of receiving a contact, while it has no statistically significant effect in our baseline regression. However, as mentioned before, we prefer the count data models since they use all the variation in the dependent variable.
Finally, we may worry that some employers who search for workers in ‘My CV’ contact them in other ways than by using the e-mail system, e.g., by phone if the
14 An alternative would be to run the regressions including only the contacts which the searchers have received during a short period around the time of the study. However, we are unable to do this since we do not have data on the dates of the contacts. 15 Another interesting issue is whether searchers who have been unemployed long are treated differently than other unemployed searchers. However, we are unable to analyze this issue since we do not have data on the dates of the contacts (c.f. Eriksson and Lagerström 2006).
123
Detecting discrimination in the hiring process 555
searchers have written their phone numbers in their personal statements. However, according to the Employment Service, most employers use the system of mailboxes.
To summarize, the robustness analysis indicates that we have managed to control for the differences across searchers, and thus that our main results are qualitatively robust.
4 The probability of getting hired
The results presented above show that searchers who have non-Nordic names, are unemployed or old get fewer firm contacts. However, from these results we cannot determine whether contacts received through ‘My CV’ have a positive effect on the probability of actually getting hired. In this section, we use data on hiring to study this issue.
4.1 Identification and estimation
We have data on the complete employment histories for the relevant period for 12,969 of our searchers who are also registered at the Employment Service (we have this data until May, 2005).16 Thus, we can observe if, and when, these searchers are deregistered by the Employment Service because they have found jobs. We construct a variable measuring if a searcher has found a job at any time after registering in ‘My CV’, and then investigate if this variable is affected by contacts received through ‘My CV’.
A crude way of measuring the effect on the hiring probability is to run a regression with the number of contacts received as the only explanatory variable. Running such a regression using the Probit model, we find that the coefficient is 0.016 (0.007).17
However, the coefficient in this regression may be biased, since it may reflect the fact that factors such as education, which are observable to employers, have a positive effect on the hiring probability.
Thus, a crucial issue is how we can identify the effect from a contact received through ‘My CV’ on the probability of getting hired. Again, it is important to note that we have access to most of the information that the firms observe. This means that we can specify a regression equation where the probability of finding a job is a function of the contacts received and all other observable factors which may affect the contact rate, and then interpret the coefficient for contacts as an unbiased measure of the effect these contacts have on the hiring probability. The fact that there are other variables, such as, e.g., motivation, which may affect the hiring probability, does not bias this estimate as long as they are unobservable to the firms using ‘My CV’.
16 Searchers who register in ‘My CV’ are not required to also register at the Employment Service. How- ever, searchers who want to receive unemployment benefits must register at the Employment Service, and are also strongly encouraged to register in ‘My CV’. Thus most unemployed searchers in our sample are in both registers, while many non-unemployed searchers in our sample are not registered at the Employment Service. 17 The robust standard error is reported within the parenthesis.
123
556 S. Eriksson, J. Lagerström
Table 7 Probit estimates of the probability of getting hired
Number of contacts At least one contact
Contacts 0.016* 0.094*** (0.010) (0.036)
Non-Nordic name −0.179 −0.179 (0.043) (0.043)
Age 26–35 −0.048 −0.048 (0.041) (0.041)
Age 36–50 −0.251*** −0.253*** (0.061) (0.061)
Age 51– −0.571*** −0.572*** (0.083) (0.083)
Unemployed −0.651*** −0.649*** (0.030) (0.030)
Female −0.021 −0.024 (0.033) (0.033)
Number of observations 12,969 12,969 R2 0.09 0.09
Notes: The dependent variable is the probability of finding a job at any time after registration in ‘My CV’ (but before May 9, 2005). All regressions also include the time vector (time and time squared in weeks), a constant and all other variables included in Table 3. The reference category is an employed man who is less than 26-years old, and has a Nordic name. Robust standard errors are in parentheses. ***, **, and * denote significance at the 1, 5, and 10% level.
4.2 Results
Table 7 presents the results for the probability of getting hired. In Table 7, we see that the variables measuring the number of contacts received and
having received at least one contact have positive and statistically significant effects on the probability of finding a job. This may be interpreted as an indication that the searchers who use ‘My CV’ benefit from this in terms of their search success. The magnitudes of the effects are such that having received at least one contact increases the probability of being hired with 0.03. Since the average probability of finding a job is 0.22, the relative effect from this increase is around 13%. Also, we see that most other variables have the expected signs; i.e. it is negative to have a non-Nordic name, be unemployed or old. However, these results should be treated with caution since all estimates, except the estimates for contacts, are affected by factors which we cannot control for (e.g., motivation). The results are similar if we run the regressions including only jobs received after December 2004.
We have also run separate regressions for different subgroups to investigate if the effects differ across groups. Table 8 presents some of these results.
In Table 8, we see that having received contacts has a statistically significant positive effect on the job finding probability for searchers with non-Nordic names, unemployed searchers and women. For men the effect is close to zero, and for searchers with Nordic names it is positive but not statistically significant. These results may be interpreted as an indication that ‘My CV’ matters more for groups which have a difficult time finding jobs using other search channels.
123
Detecting discrimination in the hiring process 557
Table 8 Probit estimates of the probability of getting hired, subgroups
Baseline Non-Nordic name Nordic name Unemployed Employed Women Men
Contacts 0.016* 0.086** 0.012 0.027* 0.013 0.071*** −0.001 (0.010) (0.038) (0.010) (0.015) (0.013) (0.018) (0.012)
Number of obs. 12, 969 1, 524 11, 361 7, 658 5, 268 6, 802 6, 075 R2 0.09 0.18 0.09 0.06 0.13 0.10 0.11
Notes: The dependent variable is the probability of finding a job at any time after registration in ‘My CV’ (but before May 9, 2005). ‘Contacts’ is the number of contacts received. All regressions also include the time vector (time and time squared in weeks), a constant and all other variables included in Table 3. Robust standard errors are in parentheses. ***, **, and * denote significance at the 1, 5, and 10% level.
To summarize, we find that searchers who have received contacts during their time in ‘My CV’, all else equal, have a somewhat higher probability of actually getting a job. The relative effect is 13%. Thus the lower contact rate for, e.g., searchers with non-Nordic names seems to matter for the actual hiring outcome by decreasing their chance of getting hired. This is problematic since our results indicate that searchers who belong to these groups actually have a higher probability than other searchers of getting hired if they are contacted.
5 Concluding remarks
The emergence of Internet-based search channels introduces new opportunities for research on discrimination. Since in principle the same information is observed by the recruiting firms and the researcher, it becomes easier to identify discrimination in the early stages of the hiring process. Thus the use of such data is an attractive addition to the existing approaches. Together, studies based on registry data, CV databases, natural experiments and field experiments should increase our understanding of the extent and consequences of labor market discrimination.
In this article, we use data from a Swedish Internet-based CV database to inves- tigate how the number of firm contacts received is affected by factors such as ethnicity, employment status, age, and gender. We find that searchers who have non-Nordic names (especially Arabic names), are unemployed or old get fewer firm contacts. Some of these differences are explained by differences in education, expe- rience and the requirements searchers have about the jobs, but even when we con- trol for all such differences substantial negative effects remain. The results are both economically and statistically significant. The relative effects are between 13 and 23%. Moreover, these differences matter for the hiring outcome, since we find that searchers who receive more contacts have a higher probability of actually getting jobs.
Our results indicate that the employers who use ‘My CV’ sort their applicants based on easily observable characteristics such as ethnicity, employment status and age. An important issue is if this should be labeled ‘discrimination’. As mentioned above, discrimination can be of two types: statistical discrimination, where firms use easily observable factors as indicators for important unobservable factors, and
123
558 S. Eriksson, J. Lagerström
preference-based discrimination, where employers have direct preferences over these characteristics. In the first case, it may be argued that it is completely rational for firms to sort workers based on characteristics functioning as indictors and thus that no actual ‘discrimination’ occurs, whereas in the second case most people would agree that actual ‘discrimination’ occurs. We cannot separate these two types of behavior. Statistical discrimination is a more likely explanation for the negative effect for unem- ployed searchers, while the other negative effects (e.g., for non-Nordic searchers) may reflect both preference-based and statistical discrimination. However, irrespec- tive of why firms use ethnicity, employment status, and age to sort their applicants, the implication of this behavior is that searchers who belong to these groups get fewer firm contacts and, as a consequence of this, have a lower probability of finding jobs.
Acknowledgments We are grateful for helpful comments from two anonymous referees, Per-Anders Edin, Erik Grönqvist, Tuomas Pekkarinen, Peter Skogman-Thoursie, and seminar participants at the annual meeting of the Society of Labor Economists in New York, the European Meeting of the Econometric Society in Milan, the annual conference of the European Association of Labour Economists in Oslo, the annual congress of the European Society of Population Economics in Chicago, the Nordic Summer Institute in Labour Economics in Helsinki, the Institute for Labour Market Policy Evaluation and Växjö University. Thanks also to Eva Granath, Anders Wellman, and the Swedish Public Employment Service for providing the data. Financial support from the Institute for Labour Market Policy Evaluation, the Swedish Council for Working Life and Social Research, The Swedish Research Council, and the Jan Wallander Foundation is gratefully acknowledged.
Appendix
Description of ‘My CV’
When registering, the searchers are required to enter information into the follow- ing eleven forms: (1) Personal information, (2) Employment history, (3) Education, (4) Other merits, (5) Computer skills, (6) Language skills, (7) Driving license, (8) Personal statement, (9) Occupations where the searcher is looking for work, (10) Counties/cities where the searcher is looking for work, and (11) Other require- ments. Mostly, information is entered by choosing from lists of alternatives. Searchers are allowed to change or delete their CVs at any time.
The CVs are only visible to employers if all the forms are completed. To stay visible, searchers are required to log into their accounts at least every 12 weeks.
All employers registered at the Employment Service are allowed to search in the database. Employers have access to a search tool where they can specify search crite- ria such as occupation, region, etc., but they cannot specify requirements on, e.g., the time the searcher has been registered in the database, gender, ethnicity, employment status or age. Also employers cannot observe if other employers have looked at a searcher’s CV.
All searchers (both new and previously registered users) who logged into ‘My CV’ during a few of weeks in December 2004 were asked if they wanted to participate in a research project on ‘the recruitment behavior of firms’. The offer appeared as a pop-up box on the computer screen during login, and a short text explained that to
123
Detecting discrimination in the hiring process 559
participate meant that they agreed that their CVs could be used for anonymous re- search. Those who agreed were also asked to immediately answer a short online questionnaire which also appeared as a pop-up box on the computer screen. The ques- tionnaire was designed by us in collaboration with the Employment Service for a larger research project. The questions were: (1) So far, have you received a contact with an employer from using ‘My CV’ [yes, no], (2) What is your main employ- ment status at present [employed, unemployed, in a labor market program, university student, participate in adult training, high school student, on parental leave, none of the above], (3) How long is your total labor market experience? [less than one year, 1–2 years, 2–5 years, 5–10 years, 10–15 years, 15–20 years, more than 20 years], (4) How much of your total labor market experience is in the occupations where you are looking for work? [nothing or almost nothing, some, all or almost all], (5) Are you registered as unemployed at the Employment Service? [yes, no], (6) Do you think that an employer in general perceives your name as Swedish? [yes, no], and (7) If you answered no to the previous question: How do you think that employers generally perceive your name? [Nordic, Asian, African, Arabic, none of the above].
Our dataset includes all the information that the searchers have registered in their CVs (except their names and short personal statements), the answers to the ques- tionnaire and how many firm contacts they have received. Our identification strategy requires that we take into account all the information observable to the employers who use ‘My CV’. Thus, we construct our variables using the information directly available in the CVs. However, since our data does not include the searchers’ names (or any other information on ethnicity), we use the answer to the ethnicity questions in the questionnaire to create the ethnicity variable. This should be similar to the information firms can deduce directly from the searchers’ names. Otherwise, the answers to the questionnaire were used only as a guide in the classification of data already available in the CVs.
Comparison of the characteristics of the searchers
Table A1 Comparison of the characteristics of the searchers in ‘My CV’ and at the Swedish Public Employment Service (in fractions)
Unemployed searchers All searchers
‘My CV’ Empl. service ‘My CV’ Empl. service
Ethnicity Foreign name 0.18 0.19 0.18 0.19 Of which African name 0.004 0.002 0.003 0.001 Arabic name 0.02 0.04 0.02 0.05 Asian name 0.01 0.05 0.01 0.05 Other foreign name 0.09 0.09 0.09 0.09
123
560 S. Eriksson, J. Lagerström
Table A1 continued
Unemployed searchers All searchers
‘My CV’ Empl. service ‘My CV’ Empl. service
Age Mean (years) 35.5 35.5 34.3 39.2 Age 20–25 0.25 0.26 0.27 0.18 Age 26–35 0.30 0.31 0.32 0.26 Age 36–50 0.29 0.28 0.28 0.33 Age 51– 0.15 0.15 0.12 0.23 Gender Female 0.50 0.40 0.53 0.50 Highest level of completed education Primary 0.14 0.20 0.13 0.20 Secondary or post-secondary 0.86 0.80 0.87 0.80 Work experience Limited relevant experience 0.27 0.25 0.26 0.22 Some or almost only relevant experience 0.73 0.75 0.74 0.78 Region of residence Stockholm 0.21 0.19 0.21 0.14 Uppsala 0.03 0.03 0.03 0.03 Södermanland 0.03 0.03 0.03 0.03 Östergötland 0.06 0.05 0.05 0.05 Jönköping 0.02 0.03 0.02 0.03 Kronoberg 0.02 0.02 0.02 0.02 Kalmar 0.02 0.03 0.02 0.03 Gotland 0.01 0.01 0.01 0.01 Blekinge 0.01 0.02 0.01 0.02 Skåne 0.16 0.13 0.15 0.14 Halland 0.03 0.03 0.03 0.03 Västra Götaland 0.15 0.16 0.15 0.17 Värmland 0.02 0.03 0.02 0.04 Örebro 0.03 0.03 0.03 0.04 Västmanland 0.03 0.03 0.03 0.03 Dalarna 0.03 0.04 0.03 0.03 Gävleborg 0.03 0.04 0.04 0.04 Västernorrland 0.04 0.03 0.04 0.03 Jämtland 0.01 0.02 0.01 0.02 Västerbotten 0.03 0.03 0.03 0.03 Norrbotten 0.03 0.03 0.03 0.04 Occupation Legislators, senior officials and managers 0.04(0.02) 0.02(0.01) 0.04(0.02) 0.02(0.01) Professionals 0.30(0.17) 0.23(0.15) 0.32(0.17) 0.18(0.11) Technicians and associate professionals 0.30(0.16) 0.17(0.11) 0.31(0.16) 0.17(0.10) Clerks 0.31(0.17) 0.23(0.15) 0.31(0.17) 0.25(0.16) Service workers and shop sales workers 0.27(0.15) 0.39(0.25) 0.28(0.15) 0.40(0.25) Skilled agricultural and fishery workers 0.04(0.02) 0.03(0.02) 0.04(0.02) 0.04(0.03) Craft and related trades workers 0.13(0.07) 0.13(0.08) 0.13(0.07) 0.16(0.10) Plant and machine operators and assemblers 0.13(0.07) 0.19(0.12) 0.13(0.07) 0.19(0.12) Elementary occupations 0.31(0.17) 0.18(0.11) 0.31(0.17) 0.19(0.12)
Notes: The ethnicity variables in ‘My CV’ are compared with data on the country of birth from the Employ- ment Service. Occupation refers to the occupation(s) where the searchers are looking for work and is reported both as the total fraction of the searchers who searches for a job in a particular occupation and (in parenthesis) the fraction of all searchers who searches for a job in a particular occupation (i.e., the second fractions are required to sum to one). The occupations are classified according to the system used by the Employment Service.
123
Detecting discrimination in the hiring process 561
Table A2 Comparison of the searchers in ‘My CV’ who have been contacted and the vacancies reported to the Swedish Public Employment Service (in fractions)
Variable ‘My CV’ Employment service
Region Stockholm 0.37 (0.23) 0.25 Uppsala 0.10 (0.06) 0.03 Södermanland 0.09 (0.05) 0.03 Östergötland 0.08 (0.05) 0.04 Jönköping 0.05 (0.03) 0.04 Kronoberg 0.04 (0.03) 0.02 Kalmar 0.04 (0.03) 0.02 Gotland 0.02 (0.01) 0.01 Blekinge 0.04 (0.02) 0.02 Skåne 0.14 (0.09) 0.11 Halland 0.08 (0.05) 0.02 Västra Götaland 0.18 (0.11) 0.17 Värmland 0.04 (0.03) 0.04 Örebro 0.06 (0.04) 0.02 Västmanland 0.07 (0.04) 0.03 Dalarna 0.04 (0.02) 0.03 Gävleborg 0.05 (0.03) 0.02 Västernorrland 0.04 (0.02) 0.03 Jämtland 0.03 (0.02) 0.02 Västerbotten 0.04 (0.02) 0.02 Norrbotten 0.04 (0.02) 0.03 Occupation Legislators, senior officials and managers 0.07 (0.03) 0.02 Professionals 0.37 (0.17) 0.15 Technicians and associate professionals 0.41 (0.19) 0.18 Clerks 0.36 (0.17) 0.07 Service workers and shop sales workers 0.32 (0.16) 0.35 Skilled agricultural and fishery workers 0.04 (0.02) 0.01 Craft and related trades workers 0.13 (0.06) 0.05 Plant and machine operators and assemblers 0.13 (0.06) 0.08 Elementary occupations 0.30 (0.14) 0.09
Notes: Region and occupation refer to the regions/occupations where the searchers are looking for work. The data from the Employment Service is the inflow of new vacancies in 2004. The data for ‘My CV’ includes only searchers who have received at least one contact, and is reported both as the total fraction of the searchers who searches for a job in a particular region/occupation and (in parenthesis) the fraction of all searchers that searches for a job in a particular region/occupation (i.e., the second fractions are required to sum to one). The occupations are classified according to the system used by the Employment Service.
The probability of receiving a contact
Table A3 Probit estimates of the probability of receiving at least one contact
(4) (5)
Non-Nordic name −0.110*** – (0.038)
Nordic name – −0.048 (0.051)
123
562 S. Eriksson, J. Lagerström
Table A3 continued
(4) (5)
African name – 0.071 (0.200)
Arabic name – −0.138 (0.089)
Asian name – −0.296** (0.125)
Other foreign name – −0.097** (0.044)
Age 26–35 −0.051 −0.048 (0.039) (0.039)
Age 36–50 −0.007 −0.003 (0.056) (0.056)
Age 51– −0.127* −0.124* (0.072) (0.072)
Unemployed −0.120*** −0.120*** (0.028) (0.028)
Labor market program −0.173*** −0.171*** (0.061) (0.062)
University student −0.023 −0.021 (0.059) (0.059)
Other adult training −0.082 −0.079 (0.071) (0.072)
High school student −0.250 −0.245 (0.154) (0.154)
On parental leave 0.009 0.006 (0.113) (0.113)
Other −0.071 −0.070 (0.054) (0.054)
Female 0.146*** 0.146*** (0.030) (0.030)
Secondary education 0.098** 0.097** (0.042) (0.042)
Post-secondary education 0.116*** 0.115*** (0.041) (0.041)
1–2 years experience 0.071 0.070 (0.048) (0.048)
2–5 years experience 0.186*** 0.185*** (0.044) (0.044)
5–10 years experience 0.225*** 0.223*** (0.052) (0.052)
10–15 years experience 0.223*** 0.222*** (0.060) (0.060)
15–20 years experience 0.172** 0.168** (0.068) (0.068)
20-years experience 0.204*** 0.199*** (0.067) (0.067)
Some relevant experience 0.086*** 0.086*** (0.031) (0.031)
Almost only relevant experience 0.129*** 0.129*** (0.033) (0.033)
Other skills Yes Yes Dummies for occupation Yes Yes
123
Detecting discrimination in the hiring process 563
Table A3 continued
(4) (5)
Dummies for region Yes Yes Number of observations 18,167 18,167 R2 0.32 0.32
Notes: The time vector (time and time squared in weeks) and a constant are always included. The reference category is an employed man who is less than 26-years old, has a Nordic name, primary education, and less than one year of non-relevant experience. The column numbers correspond to the column numbers in Table 3. Robust standard errors are in parentheses. ***, **, and * denote significance at the 1, 5, and 10% level.
References
Åslund O, Nordström Skans O (2011) Do anonymous job application procedures level the playing field? Ind Lab Rel Rev (forthcoming)
Belzil C (1996) Relative efficiencies and comparative advantages in job search. J Lab Econ 14:154–173 Bertrand M, Mullainathan S (2004) Are Emily and Greg more employable than Lakisha and Jamal? A field
experiment on labor market discrimination. Am Econ Rev 94:991–1013 Blau DM, Robins PK (1990) Job search outcomes for the employed and unemployed. J Polit Econ 98:
637–655 Carlsson M (2011) Does hiring discrimination cause gender segregation in the swedish labor market? Fem
Econ 17(3):71–102 Carlsson M, Rooth DO (2007) An experimental study of sex segregation in the Swedish labor market – Is
discrimination the explanation. Labour Econ 14:716–729 Edin PA, Lagerström J (2006) Blind dates: quasi-experimental evidence on discrimination. Working paper
2006:4. IFAU, Uppsala Eriksson S, Lagerström J (2006) Competition between employed and unemployed job applicants: Swedish
evidence. Scand J Econ 108:373–396 Goldin C, Rouse C (2000) Orchestrating impartiality: the impact of “blind” auditions on female musicians.
Am Econ Rev 90:715–741 Heckman J (1998) Detecting discrimination. J Econ Perspect 12:101–116 Lahey JN (2008) Age, women, and hiring—an experimental study. J Hum Resour 43:30–56 Lahey JN, Beasley RA (2009) Computerizing audit studies. J Econ Behav Organ 70:508–514 Neumark D with the assistance of Bank RJ, van Nort KD (1996) Sex discrimination in restaurant hiring:
an audit study. Q J Econ 111:915–941 Oreopoulos P (2009) Why do skilled immigrants struggle in the labor market? A field experiment with six
thousand résumés. University of British Columbia, Vancouver Riach PA, Rich J (1997) Testing for sexual discrimination in the labor market. Aust Econ Pap 26:165–178 Riach PA, Rich J (2002) Field experiments of discrimination in the market place. Econ J 112:F480–F518 Weichselbaumer D (2004) Is it sex or personality? The impact of sex-stereotypes on discrimination in
applicant selection. East Econ J 30:159–186
123
Copyright of Empirical Economics is the property of Springer Science & Business Media B.V. and its content
may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express
written permission. However, users may print, download, or email articles for individual use.
