Age discrimination in hiring decisions: Evidence from a field experiment in the labor market

ARTICLE IN PRESS

JID: LABECO

[m5GeSdc;April 4, 2019;20:13]

Labour Economics xxx (xxxx) xxx

Contents lists available at ScienceDirect

Labour Economics journal homepage: www.elsevier.com/locate/labeco

Age discrimination in hiring decisions: Evidence from a field experiment in the labor market Magnus Carlsson a, Stefan Eriksson b,∗ a b

Centre for Discrimination and Integration Studies, Linnaeus University, SE-391 82 Kalmar, Sweden Department of Economics, Uppsala University, PO Box 513, SE-751 20 Uppsala, Sweden

a r t i c l e JEL classification: J23 J71 Keywords: Demographic challenge Age Gender Discrimination Field experiment Hiring

i n f o

a b s t r a c t This paper shows the results of a field experiment in which over 6000 fictitious resumes with randomly assigned information about age (35–70 years) were sent to Swedish employers with vacancies in low- and medium-skilled occupations. We find that the callback rate begins to fall substantially for workers in their early 40s and becomes very low for workers close to the retirement age. The decline in the callback rate by age is steeper for women than for men. Employer stereotypes about the ability to learn new tasks, flexibility, and ambition seem to be an important explanation for age discrimination.

1. Introduction In many countries, older workers have lower employment rates and experience much longer unemployment spells than younger workers (Chan and Stevens, 2001; OECD, 2018). In this paper, we investigate the extent to which this pattern is due to employer discrimination. There are several reasons why employers may discriminate against older workers. They may believe that older workers do not have the same ability to learn new tasks or that they are less able to adapt to changes in the workplace. They may also perceive older workers as less ambitious and less willing to work hard. Such concerns may result in statistical discrimination (Arrow, 1973). Age discrimination may also be taste-based (Becker, 1957), i.e., driven by ageism, a dislike of older workers.1 Age discrimination could also interact with gender, even though the sign of the interaction is difficult to predict (Levy, 1988). Among older workers, employers may prefer women because they have, on average, better health than men (OECD, 2013). Conversely, they may prefer men because women tend to retire earlier (OECD, 2015). In 2015 and 2016, we conducted a large-scale field experiment to explore these issues and investigate the extent to which age discrimination contributes to the observed differences in labor market outcomes. More than 6000 fictitious resumes for female and male applicants aged 35–70

∗

Corresponding author. E-mail address: [email protected] (S. Eriksson). 1 There may also be implicit discrimination (Bertrand et al., 2005) and attention discrimination (Bartoš et al., 2016).

were sent to Swedish employers with a vacancy. The measured outcome was the response from employers in the form of callbacks (e.g., invitations to job interviews). We can identify a causal effect of age, since this characteristic is randomly assigned to the fictitious resumes. This approach is in sharp contrast to studies using survey or administrative data, where it is difficult to separate the effects of age from the effects of other worker characteristics that are observed by the recruiting firms but not by the researcher. If such characteristics exist and are correlated with age, a classical omitted variable bias arises. Field experiments to detect discrimination were developed in response to such concerns and have frequently been used to study discrimination, particularly ethnic discrimination (e.g., Riach and Rich, 2002; Bertrand and Mullainathan, 2004; Carlsson and Rooth, 2007; Rich, 2014; Neumark, 2018). Our main results are consistent with the existence of employer discrimination against older workers. The callback rate begins to fall substantially for workers already in their early 40s and becomes very low for workers close to the retirement age. Ten years of aging leads to a callback rate that is approximately five percentage points lower. These results are robust to several ways of taking into account that older workers have more potential work experience. We also provide evidence that age discrimination is worse for women than for men. At age 35, women have a higher callback rate, but the callback rate for women declines faster with age. We explore the mechanisms behind age discrimination in several ways. First, we investigate the importance of perceived productivity differences by varying signals about employment status and flexibility, but we find no conclusive results. Second, we consider heterogeneity across occupations and firms, and find that the negative age effect is very stable. Third, we conducted a survey on a representative sample of Swedish

https://doi.org/10.1016/j.labeco.2019.03.002 Received 23 October 2018; Received in revised form 17 March 2019; Accepted 21 March 2019 Available online xxx 0927-5371/© 2019 Elsevier B.V. All rights reserved.

Please cite this article as: M. Carlsson and S. Eriksson, Age discrimination in hiring decisions: Evidence from a field experiment in the labor market, Labour Economics, https://doi.org/10.1016/j.labeco.2019.03.002

JID: LABECO

ARTICLE IN PRESS

M. Carlsson and S. Eriksson

[m5GeSdc;April 4, 2019;20:13] Labour Economics xxx (xxxx) xxx

employers, who were asked how they perceive that different skills vary by age. The answers suggest that employers’ worry that workers already in their 40s are starting to lose their ability to learn new tasks, flexibility/adaptability, and ambition. Such stereotypes could result in statistical discrimination. Two of the most recent large-scale field experiments on age discrimination in the labor market are those by Lahey (2008) and Neumark et al. (2018), both conducted in the US. Lahey (2008) studies female job applicants and finds that older applicants (aged 50, 55, or 62) are less likely to be invited to a job interview than are younger applicants (aged 35 or 40). Neumark et al. (2018) find that applicants, especially women, around age 50 and 65 are less likely to be invited to a job interview than are those around age 30, and the difference is larger for those around age 65. Other studies include those by Bendick et al. (1996, 1999), Riach and Rich (2010), Albert et al. (2011), Ahmed et al. (2012), Baert et al. (2016a), and Farber et al. (2017). A related body of literature uses field experiments to study gender discrimination, e.g., studies by Neumark et al. (1996), Riach and Rich (1987, 2006), Weichselbaumer (2004), Petit (2007), Carlsson (2011), and Baert et al. (2016b). Most of the studies of gender discrimination consider relatively young applicants and do not explicitly consider how age interacts with gender. Our study contributes to this literature in several ways. First, we include age as a continuous variable, while most previous studies compare groups of applicants in specific age categories (e.g., Neumark et al., 2018). This approach allows us to study the dynamic pattern of age discrimination, e.g., to measure the effect of one year of aging in a wide age interval and investigate at what age discrimination starts. Second, we explicitly analyze gender differences in age discrimination. The previous large-scale studies of age discrimination have either focused on only one gender (e.g., Lahey, 2008) or have not included both female and male applicants in the same occupations (e.g., Neumark et al., 2018, use resumes from both genders in only one occupation, namely, sales work).2 For Sweden, it is reasonable to randomize gender in all occupations in the experiment. Third, we conducted an employer survey with questions designed to reveal why employers discriminate. This work adds to the survey-based literature on why firms discriminate based on age (Posthuma and Campion, 2009) and helps us interpret the experimental results. In Section 2, we show that there are systematic age differences in key labor market outcomes. In Section 3, we describe the field experiment. In Section 4, we present our results. In Section 5, we explore the mechanisms that may explain the experimental results. Section 6 concludes the paper.

Fig. 1. Unemployment duration by age. Notes: The graphs are constructed by calculating the average unemployment duration for each age. Data is obtained from the Swedish Public Employment Service, which contains information about all unemployed individuals registered at the Employment Service. The sample we use consists of individuals registered at the Employment Service from January 1, 2011, to February 18, 2012. Right censoring is not an important issue in these graphs, since all individuals are observed at least two years after the start of a period of unemployment because we observe individuals until February 18, 2014. The sample contains 46,201 unemployment periods and 44,258 individuals (a few persons have more than one period of unemployment).

aged, and then drop when the worker gets closer to the retirement age (Willis, 1985). Studies suggest that it is hours worked rather than hourly wages that fall with age (Johnson and Neumark, 1996; Rupert and Zanella, 2015), so the decline should reflect the reduction in hours worked. This pattern is true for Sweden as well (Online Appendix Figure A2). The shapes of the earnings and employment profiles are similar for women and men, but the levels are always higher for men (Blau and Kahn, 2017). In an international comparison (OECD, 2018), labor force participation among older workers is high in Sweden (in 2015, the employment rate was 74.6% for workers aged 55–64, according to Statistics Sweden). If there is discrimination in hiring, unemployment durations should be affected. Fig. 1 shows the unemployment durations for female and male workers registered at the Swedish Public Employment Service. The durations increase almost linearly with age, and the increase is somewhat steeper for women. The finding that durations increase with age is observed in most other Western countries (OECD, 2018). Worker mobility could also be affected by discrimination. If workers expect discrimination, they may not search for a new job or are unable to find a new job. If we consider the probability of changing jobs by age for Swedish workers, this measure of worker mobility declines almost linearly with age (Online Appendix Figure A3). Mobility is very low among older workers and is always lower for women than for men. The finding that worker mobility declines with age and tenure is observed in most Western countries (Farber, 1999; Theodossiou and Zangelidis, 2009). In the field experiment, we investigate whether demand effects play an important role in explaining these patterns.

2. The labor market for women and men at different ages3 Before moving on to our experiment, we provide evidence on the distribution of labor market outcomes across older and younger workers in Sweden, and we briefly discuss how it compares with corresponding findings in other developed countries. If we consider total labor market earnings in Sweden by age 35– 65 for female and male workers, the age-earnings profile has a positive slope until a peak at approximately age 45–50, and after age 55, the profile begins to decline markedly (Online Appendix Figure A1). This hump-shaped relationship between total labor market earnings and age is well documented in the literature. Earnings tend to increase in the early stages of a worker’s career, flatten out when the worker is middle-

3. Experimental design

2

A few studies consider both female and male job searchers in specific age categories (e.g., Baert et al., 2016a). However, to fully analyze the age-gender interaction, we include age as a continuous variable. We also conduct an experiment that is larger in scale, which makes it possible to draw more precise conclusions on how age and gender interact. Lahey and Oxley (2016) consider how age interacts with gender in a laboratory environment. 3 These patterns are very similar if we consider only low educated workers, as in the experiment.

3.1. Age, gender, and other worker characteristics In the experiment, age is included as a continuous variable in the interval of 35–70 years. The lower bound is chosen because we assign all resumes at least ten years of relevant work experience (see below), and at age 35, most workers have finished their education and have worked 2

JID: LABECO

ARTICLE IN PRESS

M. Carlsson and S. Eriksson

[m5GeSdc;April 4, 2019;20:13] Labour Economics xxx (xxxx) xxx

for a number of years. The upper bound is chosen based on the current retirement age in Sweden (which is 61–67), but it is often argued that people should work a few additional years. Age is randomly assigned to each of the resumes and is drawn from a uniform distribution. Gender is signaled through the name of the applicant. In Sweden, there is usually a clear distinction between female and male names. Names are randomly assigned to the resumes, where approximately 50% were assigned a typical female and male name, respectively. We use three of the most common female and male names for people in the age interval of 35–70 years, according to Statistics Sweden’s name register.4 We use two characteristics to signal productivity – employment status and willingness to participate in on-the-job training. Employment status is signaled by randomly assigning the job applicants to a spell of unemployment 0–36 months in duration.5 The idea is to investigate whether older workers are statistically discriminated against because employers worry that they have worse unobserved skills, e.g., insufficient occupational skills. We analyze whether being employed (or shortterm unemployed) rather than (long-term) unemployed raises the callback rate more for older than for younger job applicants.6 If this type of statistical discrimination is important, older employed workers should, to a larger extent, be perceived as having unobserved skills that are comparable with those of younger workers than if we compare unemployed older and younger workers (i.e., the fact that an older worker is employed should be a signal that he or she has sufficient skills). Employment status is randomly assigned so that 1/3 of the applicants are on-the-job searchers and 2/3 of the applicants are uniformly distributed over the interval of 1–36 months of unemployment. We signal that an applicant is interested in participating in on-thejob training by expressing a willingness to take part in such training in the resume. The idea is to investigate whether employers discriminate against older workers because they are not perceived as flexible and adaptable to workplace changes.7 Survey evidence suggests that employers are concerned that older workers do not have these skills (see Section 5.3). We randomly assigned half of the resumes a sentence signaling this characteristic.

part of the employment history. This is the standard approach in field experiments on discrimination on other grounds, e.g., ethnicity, in which the job applications are designed to be qualitatively identical. In the case of age discrimination, this design can make the age effect difficult to interpret (Riach and Rich, 2010). Employers may assume that older workers have more relevant work experience than the resume reveals, which could potentially generate statistical discrimination that favors older workers. In our experiment, this design means that we assign all job applications ten years of relevant work experience, which is realistic for workers in the age interval we study.8 The third design attempts to hold perceived relevant work experience constant by explicitly signaling that older applicants do not have more relevant work experience. This is typically done by including a complete employment history in which it is stated that the older applicant has been engaged in a previous activity that is assumed not to affect the applicant’s productivity in the current occupation. Examples are working in a very different (unqualified) occupation, which is the design we use, taking care of children, and being active in the military (e.g., Ahmed et al., 2012). A potential problem with this approach is that it creates a correlation between being older and having experience in a particular previous activity. Hence, this strategy will only work if this activity is irrelevant for the recruiting employers. The advantage of using all three designs in the experiment is that we can investigate whether the (theoretical) arguments that have been raised against each of the designs are important when conducting a field experiment. Our approach is similar to that of Neumark et al. (2018), who use the first two designs by including both older and younger job applicants who have the same amount of work experience and older job applicants who have a full CV history of work experience. They find that the choice of design does not affect the age results in most occupations (an exception being male janitors).9 3.3. Generating resumes To create realistic resumes, we studied a large number of real job applications available in a database at the Swedish Public Employment Service. We also consulted colleagues and used our previous experience in conducting field experiments. Each generated resume consists of two parts – a cover letter and a CV (Online Appendix Figures A4 and A5 show an example). The cover letter starts with a short presentation that includes the applicant’s name and age, a description of work experience, and some information about personal interests. The CV includes the applicant’s full name, date of birth, contact details, work experience, education, computer skills, driver’s license, and some occupation-specific certificates.10 All resumes were generated before the experiment began. The first step was to create templates that specify the structure, layout, and typeface of the resume and contain some general phrases and information. We use a specific template for each occupation. In the second step, the resume templates were filled with content, which depended on the randomized variables, occupation, and city. The

3.2. Employment histories for younger and older job applicants The construction of employment histories in a field experiment on age discrimination must inevitably consider the fact that older workers have lived longer than younger workers have. Previous studies have used three alternative designs to handle this fact (Neumark et al., 2018; Baert et al., 2016a). We implement all of these alternatives. The first design fills the resumes with a full history of relevant work experience. The argument for this design is that it is the most reasonable, since older workers tend to have more relevant work experience than younger workers do (Riach and Rich, 2010). The argument against this design is that it does not capture the pure age effect, since older workers may be favored by always having more relevant work experience than younger workers do. The second design assigns both younger and older workers a fixed number of years of relevant work experience and then omits the earlier

8 Job applicants in Sweden are often given the advice to include only the most relevant work experience in their resumes. Lahey (2008) argues that including ten years of work experience is what US firms prefer. Neumark et al. (2018) argue that only including ten years of experience could lead to a disadvantage for older applicants. 9 Baert et al. (2016a) also analyze this issue and label it “the post-educational years problem”. They consider job applicants aged 38, 44, and 50. They assign all applicants the same amount of work experience in the relevant occupation immediately after leaving school as well as immediately before applying to the job in the experiment. However, for an intermediate period, they either leave the period empty, add a job that is assumed to be unrelated to the productivity in the job applied for, or add a job in the relevant occupation. 10 All applicants had a driver’s license for a car. In addition, truck drivers had a driver’s license for a truck. Truck drivers were also given a few other occupationrequired certificates.

4 We used lists of the most common first names of individuals in the population born in the same years as our fictitious applicants. We also compiled a list of the most common surnames. Then, we randomly combined the first names and surnames. The female names are Anna Eriksson, Eva Olsson, and Lena Persson. The male names are Anders Karlsson, Lars Johansson, and Peter Nilsson. 5 In official unemployment statistics, a duration of unemployment of 12 months or longer is classified as long-term unemployment. We chose 36 months as the upper limit to include long-term unemployed workers with long durations, which is realistic among the older unemployed population. 6 Previous research has shown that employers prefer not to recruit long-term unemployed workers (Kroft et al., 2013; Eriksson and Rooth, 2014). 7 This approach is similar to that of Lahey (2008), who includes the phrase “willing to embrace change” in her experiment.

3

JID: LABECO

ARTICLE IN PRESS

M. Carlsson and S. Eriksson

[m5GeSdc;April 4, 2019;20:13] Labour Economics xxx (xxxx) xxx

first two randomized variables are the age and name of the applicant, which are stated both in the cover letter and the CV. In Sweden, it is very common for job applicants to explicitly state their age in their resumes.11 The third random variable is employment status (duration of unemployment), which is signaled in the CV part of the resume. The CV lists previous jobs and an unemployed worker has a gap in his or her employment history. The fourth randomly assigned variable is the signal of being willing to participate in on-the-job training, which is conveyed by the following sentence in the cover letter: “I enjoy taking courses and participating in on-the-job-training to improve my skills” (as opposed to no information about this). We also randomized the design of the employment history, as described above, in the CV. The education and work experience of the applicant are determined by the occupation. All applicants were given an occupation-relevant high school education and at least ten years of occupation-relevant work experience (two jobs with random tenure summing to ten years). Since we know from our previous field experiments that employers very seldom respond by postal mail, we used a fictitious address in the same city as the job. The addresses are located in similar suburbs (in terms of the socioeconomic characteristics of the residents) that are not too far from the city centers. To enable employers to contact the job applicants, an email address and a telephone number (with an automatic answering service) were included in the applications, which were registered at a large Internet provider and phone company, respectively. Each of the six names we used in the experiment had a separate email address and telephone number. There is an efficiency argument for sending several resumes to each employer, since a given number of observations can then be collected using fewer resources. However, the risk of making employers suspicious also increases when more resumes are sent to the same employer. We had to make a trade-off between these two issues and decided that it was reasonable to send three resumes to each employer. This design entailed the construction of three types of resume templates, which differed in terms of structure, layout, typeface, and general phrases to prevent employers from becoming suspicious. The resume template used for each application was randomly assigned, and hence cannot affect our measure of discrimination.

the resources available for this project, and our previous experience in conducting field experiments. Vacancies were sampled in the three major cities in Sweden – Stockholm, Gothenburg and Malmö – that together encompass a clear majority of all Swedish vacancies. In these cities, we randomly sampled firms in seven occupations that had an advertisement on the website of the Swedish Public Employment Service.13 When choosing the occupations, our aim was to include a sufficient number of common occupations to obtain a representative picture of the Swedish labor market. We compiled a list of the most common occupations in Sweden.14 Starting from the top of this list, we included occupations that fulfilled a number of criteria. The first criterion was that it must be possible to apply by email. Many employers in the private sector accept applications by email. In contrast, most employers in the public sector use web-based recruitment systems, in which applicants must state their social security number. Therefore, we could not include occupations in healthcare and teaching, which in Sweden are dominated by the public sector. This criterion excluded ten occupations on the top-25 list. The second criterion was that the occupation is not high-skilled, i.e., requiring a university education. For such jobs, employers are likely to use the Internet to screen job applicants, e.g., by using LinkedIn. However, the fictitious job applicants cannot be found on the Internet, which could make employers suspicious. High-skilled occupations also often require more elaborate resumes, tailored to a specific advertisement, which are difficult to generate in an automated way. This criterion excluded two more occupations on the top-25 list. The third criterion was that the number of advertisements in an occupation posted on the website we use is large enough to, at least exploratory, study heterogeneity across occupations. We checked for this by counting the number of new advertisements available per day in each occupation. This criterion did not exclude any further occupations. After these three criteria were applied, we had 13 possible occupations. However, we decided that it is reasonable to treat office assistants, accounting assistants, and secretaries as the same occupation, since the job tasks are similar. We gave this occupation the label of administrative assistant. We also treated food and non-food retail salespersons as one occupation, which we label retail sales persons and cashiers. These changes reduced the number of occupations to ten. Finally, to avoid too much heterogeneity (and hence estimates with excessive noise), we did not want to include extremely male- or female-dominated occupations. Four occupations on our list were maledominated, with at least 80% men. We kept the largest of these occupations (truck drivers) and excluded the other three (warehouse workers, janitors, and construction workers). We retained the sole occupation with at least 80% women (administrative assistants). We also checked for extreme deviations in the age distribution of the occupations, but we did not find reasons to exclude any further occupations. In the end, we include the following seven occupations: administrative assistants, chefs, cleaners, food serving and waitresses, retail sales persons and cashiers, sales representatives, and truck drivers. These occupations are common not only in Sweden but also in most other Western countries. Online Appendix Table A1 presents descriptive statistics about the number of workers employed, the age and gender distributions, and the share of reported vacancies in these occupations. Each occupation employs 1–4% of the total number of workers in the labor market, which are relatively large numbers given that there are 429 occupations at the four-digit level. The fractions of vacancies reported to the Employment Service in these occupations seem representative of the

3.4. Sampling of vacancies We planned for a sample of approximately 6000 job applications, which would be sent in groups of three to approximately 2000 employers. The size of the experiment was determined by power calculations,12

11 We have access to a large database of job applications, and in that database, the majority of the job applicants mention their age in both the cover letter and the CV. 12 Before conducting the experiment, we performed a simple simulation exercise to determine the required minimum sample size. Recall that we use a continuous explanatory variable, while most other studies use a discrete variable. Suppose we want to be able to detect an age effect of five percentage points for every ten years of aging with 90% certainty. We use Stata 15 and make 1000 draws from a population characterized by such an age effect, an average callback rate of 10% (which is somewhat higher than that of Bertrand and Mullainathan, 2004, but lower than that of Carlsson and Rooth, 2007), and an error term that affects the probability of a callback. Then, we need a sample size of 400 to obtain p-values of less than 0.05 in more than 90% of the cases for the age estimate, i.e., a statistical power of 90%. Ex post, we have confirmed that we obtain a qualitatively similar number if we instead use the statistical software G∗ Power. None of these methods take into account that the applications are sent in groups of three. Lahey and Beasley (2016) show that with three applications sent to each firm, the number of observations must be inflated by 20–60%. This finding suggests that we need approximately 400 × 1.6 = 640 observations. However, since we had resources to collect data for at least six months, we decided to gather substantially more observations, meaning that we should have a sample size well above the minimum size required to obtain the desired statistical power for the age effect.

13 The Employment Service estimates that 30–40% of all vacancies are reported to them. 14 We define an occupation at the most detailed level in the Swedish Standard Classification of Occupations (SSYK 2012 which is similar to ISCO-08). The total number of categories is 429.

4

ARTICLE IN PRESS

JID: LABECO

[m5GeSdc;April 4, 2019;20:13]

M. Carlsson and S. Eriksson

Labour Economics xxx (xxxx) xxx

Table 1 Descriptive statistics. Occupation

Number of resumes

Share of resumes

Callback rate

Administrative assistants Chefs Cleaners Food serving and waitresses Retail sales persons and cashiers Sales representatives Truck drivers All

1110 1059 789 1137 918 558 495 6066

0.18 0.17 0.13 0.19 0.15 0.09 0.08 1

0.053 0.158 0.091 0.048 0.040 0.091 0.170 0.087

Table 2 The probability of a callback.

size of the occupations, i.e., the fractions are similar in magnitude to the share of workers employed in each occupation. 3.5. Conducting the experiment and recording the responses

Panel A: Age Age

Between August 2015 and March 2016, we sent 6066 resumes to 2022 employers with an advertisement on the chosen website. For each advertisement, three resumes were randomly selected and sent in random order to the employer, with a one-day delay between each resume sent. The employers mostly replied by email, or in some cases by leaving a voicemail message.15 After we recorded each reply, we promptly declined all invitations to job interviews. From the recorded replies, we constructed a callback dummy variable, which is coded as one in the case of a positive response (i.e., any response that can be interpreted as expressing an interest in the applicant) and zero in the case of a non-response or a negative response. An alternative would be to use an indicator variable that takes the value of one only for explicit invitations to a job interview. However, it is likely that many employers do not invite applicants to an interview via e-mail (or by leaving a message on an answering service), although such contacts may eventually lead to an interview. We consider cases with only explicit invitations to interviews as a robustness check and obtain very similar results. Column 1 in Table 1 shows the number of resumes sent for each of the seven occupations. For each occupation, we sent between approximately 500 and 1100 resumes. Column 2 shows the corresponding shares of sent resumes (8–19%). The last column shows the callback rates for the seven occupations, which vary between 4.0% and 17.0%, with an average of 8.7%. The variation in the number of resumes and the callback rates likely reflects differences in labor demand among the occupations.

Panel B: Age interacted with gender Female Age x female Age x male p-value (test of equal age coefficients) Covariates included Mean callback rate: 0.087

(1)

(2)

−0.0049∗ ∗ ∗ (0.0004)

−0.0048∗ ∗ ∗ (0.0004)

0.0494∗ ∗ ∗ (0.0181) −0.0058∗ ∗ ∗ (0.0005) −0.0040∗ ∗ ∗ (0.0005) 0.0109 No

0.0484∗ ∗ ∗ (0.0178) −0.0057∗ ∗ ∗ (0.0005) −0.0038∗ ∗ ∗ (0.0005) 0.0071 Yes

Notes: Each panel shows results from separate regressions with N = 6066. The regressions in column 1 includes no additional covariates while column 2 includes covariates (dummy variables) for gender, employment status, flexible, type of employment history, application type, application order, occupation and city. Standard errors are clustered by firm. ∗ ∗ ∗ significant at the 1% level, ∗ ∗ significant at the 5% level; ∗ significant at the 10% level.

Hence, the callback rate falls by approximately five percentage points for every ten years of aging, which shows that the magnitude of the age effect is substantial. The age estimate remains essentially unchanged if we include covariates for resume characteristics, occupation, and city, which is expected since age is randomly assigned (column 2).17 The gender indicator in the regression in column 2 (not shown in the table) reveals that women have a callback rate that is, on average, approximately 1.4 percentage points higher than that of men (significant at the 10% level). The callback-rate-age profiles for women and men in Fig. 2 are similar; both show an early decline in the callback rate and a clear negative relationship between the callback rate and age. However, two highly visible differences also emerge. Early in the age interval, women have a higher callback rate than men, and the decline in the callback rate is steeper for women than for men. Panel B in Table 2 shows the separate age coefficients for women and men. In these regressions, the age variable itself is not included; rather, only its interactions with the female and male indicators are included. With no additional covariates in the regression, the size of the age effect is approximately −0.0058 for women and −0.0040 for men (column 1), and the difference is statistically significant (p-value = 0.0109). For applicants aged 35, women have a callback rate that is 4.9 percentage points higher (statistically significant) than that of men. Again, the estimates remain nearly unchanged if we include the other covariates (column 2). These results show that the negative effect of age is stronger for women.

4. Results of the field experiment 4.1. The effect of age and gender on the callback rate Our main results are consistent with the existence of employer discrimination against older workers. For both women and men, the callback rate drops substantially early in the age interval we consider, and there is a clear negative relationship between the callback rate and age, which panels A and B in Fig. 2 clearly show. Additionally, for resumes with ages close to the retirement age, the callback rate is very low, at approximately 2–3%. Table 2 contains the coefficients from estimating a linear probability model with the callback indicator as the dependent variable and with age (entered linearly) as the explanatory variable.16 Panel A reports the magnitude of the age effect. The coefficient of age in column 1, which is from a regression with no additional covariates, is approximately −0.0049 and is statistically significant at the 1% level. 15 We were able to match almost all responses to an application (we failed to match less than ten responses). 16 An alternative is to use the probit model. All main results remain qualitatively unchanged if the probit model is used (Section 4.2). We also consider other specifications, e.g., discrete age intervals, as a robustness check.

17 The use of random assignment also implies that no pair of random variables should show any sign of a substantial correlation, which is confirmed in Online Appendix Table A2.

5

JID: LABECO

ARTICLE IN PRESS

M. Carlsson and S. Eriksson

[m5GeSdc;April 4, 2019;20:13] Labour Economics xxx (xxxx) xxx

Fig. 2. Callback rate by age. Note: The callback rate for each age is, on average, calculated based on (6066/2 × 36) = 85 job applications.

How do our results compare with the findings in previous studies on age discrimination? The fact that we find evidence of age discrimination, which affects even middle-aged workers, is in line with the findings of most previous studies (e.g., Lahey, 2008; Neumark et al., 2018). The results in Neumark et al. (2018) suggest that age discrimination is worse for women, which is confirmed by our results, and we show that this finding remain when we use both female and male applicants in the same occupations. In contrast to the studies using specific age groups, we show the dynamics of the age effect in the whole age interval of 35–70 years.

We also estimate a model with age divided into discrete intervals (instead of age entered linearly).18 We divide the age interval of 35– 70 years into four categories of equal size (36/4 = 9 years).19 For each age cutoff (44, 53, and 62 years), we construct a variable that is equal to zero if the age of the applicant is below the cutoff; otherwise, the variable is equal to one. We then estimate the change in the callback rate as a result of a change in the age category, e.g., from 35–43 to 44– 52 (Online Appendix Table A5). This analysis indicates that, in absolute terms, the age effect is largest in the lower part of the age interval of 35– 70 years.20 In particular, this applies for female applicants. In relative terms, the age effect is rather similar across the age intervals. Heckman (1998) raises a potential concern about how to interpret the treatment effect in this type of experimental study, which highlights the importance of using realistic job applications to obtain external validity. He formulates a model that shows that if employers perceive a group difference in the variance of the job applicants’ unobserved characteristics, the estimated age effect could depend on the quality of the resumes used in the experiment. To test for this possibility, we follow Neumark et al. (2018) using the Stata code they provide to estimate a heteroskedastic probit model.21

4.2. Robustness analysis An important issue is whether the design of the employment histories in the resumes affects the estimated age effect. We investigate this issue by analyzing the interaction effects between the different designs and age. The results show that the age coefficient is similar for the three different designs (Online Appendix Table A3). Using an F-test we cannot reject the hypothesis of equal age-slope coefficients (p-value = 0.149). This finding suggests that the choice between the three employment history designs is not important for the results in a field experiment on age discrimination. This finding is similar to those of most of the occupations in Neumark et al. (2018). An alternative to positive responses is to use an indicator of only explicit invitations to a job interview as the dependent variable. If we use this outcome, the age coefficient is –0.0026 and statistically significant at the 1% level (column 2 in Online Appendix Table A4). The finding of a smaller age coefficient is expected, since the average callback rate is lower for invitations to job interviews. Our main results are also not substantially affected by the inclusion of firm fixed effects (2022 dummy indicators; column 3), by the exclusion of applicants above 65 years old (column 4), or by the use of the probit model (column 5).

18 We have also investigated what happens if we include age-squared in addition to the linear age term in the main regression. The result is an age coefficient of −0.0075 (significant at the 1% level) and an age-squared coefficient of 0.0001 (significant at the 5% level). 19 We obtain similar results if we use other age categories of equal size (e.g., 36/3 or 36/6). 20 The change in the callback rate from age 35–43 to age 44–52 is 7.6 percentage points; in contrast, the change from 44–52 to 53–61 is 4.4 percentage points and from 53–61 to 62–70 is 2.1 percentage points. An F-test rejects the hypothesis of a constant age effect. 21 The idea to empirically test for the role of group differences in the variance of unobservables is developed in Neumark (2012). The method requires that

6

ARTICLE IN PRESS

JID: LABECO M. Carlsson and S. Eriksson

[m5GeSdc;April 4, 2019;20:13] Labour Economics xxx (xxxx) xxx

Table 3 Neumark decomposition of the age effect.

Panel A: Linear probability model Old Panel B: Probit model Old

Age < 52 vs. >= 52

Age 35–46 vs. 58–70

−0.0883∗ ∗ ∗ (0.0075)

−0.1150∗ ∗ ∗ (0.0091)

−0.0836∗ ∗ ∗ (0.0072)

−0.109∗ ∗ ∗ (0.0088)

Panel C: Heteroskedastic probit model and decomposition Old −0.0679∗ ∗ ∗ (0.0103) Overidentification test (p-value) 0.590 Ratio of standard deviations (old/young) 1.327 Test of ratios of standard deviations = 1 (p-value) 0.173 Old, through level −0.1207∗ ∗ ∗ (0.0241) Old, through variance 0.0528∗ (0.0319) Number of job applications 6066

−0.0955∗ ∗ ∗ (0.0138) 0.581 1.134 0.593 −0.1187∗ ∗ ∗ (0.0290) 0.0231 (0.0397) 4238

Notes: The regressions in panels A, B and C follow Neumark et al. (2018) using the Stata code they provide and include covariates for resume characteristics, occupation, and city. ∗ ∗ ∗ significant at the 1% level, ∗ ∗ significant at the 5% level; ∗ significant at the 10% level.

Neumark’s method implies the comparison of two groups of candidates, so we construct a dummy indicator of being older (defined as being above the median age, which is 52). Panels A–C in Table 3 show the estimates of this age indicator for the linear probability, the probit, and the heteroskedastic probit model, respectively, which are between seven and nine percentage points. Panel C also shows the result of an overidentification test and the point estimate of the relative standard deviation of the unobserved characteristics for older and younger applicants.22 Finally, panel C shows the estimated parameters of the indicator for being older through the level and through the variance, i.e., the decomposed marginal effects in the heteroskedastic probit model.23 The decomposed effects provide suggestive evidence (notice the large standard errors compared to the linear probability model) that is consistent with a relatively low quality of the resumes used in our experiment and that older applicants have a higher variance of unobserved characteristics that benefits them. Following Neumark et al. (2018), we also examine differences between the youngest and the oldest applicants. The argument is that there may be larger differences in the variance of unobservables between these groups. We split the sample into three parts of equal size and drop the middle group (i.e., we compare workers aged 58–70 with workers aged 35–46). The results are qualitatively similar (column 2). At face value, the estimates imply that the age effect would have been even larger if we had used higher-quality applications in the experiment. This finding raises the question of what is the “right” quality of standardization of the resumes. To obtain external validity, the resumes should be representative of real-world resumes. If they are representative, then the estimates of the age effect should be externally valid, and the effect that operates through perceived differences in the variance

of unobserved characteristics (i.e., statistical discrimination) should be considered to be part of the total reported age effect, since it will affect real-world job seekers. 5. Analyzing the mechanisms 5.1. Do the signals of productivity matter? As mentioned in Section 3.1, we included two signals (employment status and flexibility) in the resumes to test whether uncertainty about older workers’ productivity is an important channel for age discrimination. For employment status, we construct an indicator that equals one if the applicant is long-term unemployed (i.e., unemployed at least twelve months) and zero if the applicant is an on-the-job searcher or short-term unemployed.24 For flexibility, we use an indicator that equals one if the sentence signaling a willingness to participate in on-the-job-training is included in the resume. Column 1 in Table 4 shows that there is a strong negative effect of long-term unemployment (the coefficient is −0.021 and is highly statistically significant), which is consistent with the results of previous studies (Kroft et al., 2013; Eriksson and Rooth, 2014). Importantly, this finding confirms that the employers have noted this signal in the resumes. In column 2, we investigate whether the negative effect of longterm unemployment interacts with age. The regression does not contain the age variable itself but rather the two interactions with the long-term and non-long-term unemployment indicators. The results show that the age coefficients are similar in magnitude for those who are long-term unemployed and for those who are not; an F-test cannot reject the hypothesis that the age coefficients are equal (p-value = 0.378). In column 3, we estimate separate age coefficients for the four subgroups defined by employment status and gender. For men, there is no evidence of a difference in the age coefficient among the employment status subgroups (p-value = 0.553). For women, there is weakly significant evidence that the age effect is smaller for long-term unemployed than for non-longterm unemployed (p-value = 0.087). These results are not consistent with our theoretical hypothesis, and hence, we cannot draw any clear conclusions about the importance of statistical discrimination. There is no evidence that the flexibility signal has an effect on the callback rate (Online Appendix Table A6; the coefficient is –0.0001, and

there are resume characteristics (other than age in this case) that shift the callback rate. Gender, employment status, and flexibility vary in the resumes in our experiment (the first two characteristics significantly affect the callback rate), and we include these characteristics in the analysis. 22 The high p-value in the overidentification test is consistent with the data satisfying the identifying assumption of equal coefficients of observed applicant characteristics for younger and older applicants. The relative standard deviation indicates that older applicants have a higher perceived variance of unobserved characteristics; the ratio is 1.33, and the p-value is 0.17. 23 The sum of the effect through the level (−0.1207) and of the effect through the variance (0.0528) is −0.0679, i.e., the estimate of the heteroskedastic probit model.

24 Studies show that it is long rather than short spells of unemployment that are perceived as a strong negative signal. The results are very similar if we instead use a continuous measure of the duration of unemployment.

7

ARTICLE IN PRESS

JID: LABECO

[m5GeSdc;April 4, 2019;20:13]

M. Carlsson and S. Eriksson

Labour Economics xxx (xxxx) xxx

Table 4 The probability of a callback, employment status. (1)

(2) ∗∗∗

Long-term unempl.

−0.0209 (0.0074)

(3) ∗

−0.0292 (0.0172)

Female Female x long-term unempl.

−0.0067 (0.0234) 0.0707∗ ∗ ∗ (0.0237) −0.0457 (0.0343)

−0.0052∗ ∗ ∗ (0.0005) −0.0046∗ ∗ ∗ (0.0005)

Age x not long-term unempl. Age x long-term unempl. Age x not long-term unempl. x female Age x long-term unempl . x female Age x not long-term unempl. x male Age x long-term unempl. x male p-value (test of equal age coeff.: all; female/male.)

0.378

−0.0065∗ ∗ ∗ (0.0007) −0.0048∗ ∗ ∗ (0.0007) −0.0037∗ ∗ ∗ (0.0007) −0.0042∗ ∗ ∗ (0.0007) 0.087/0.553

Notes: N = 6066. The regressions include no additional covariates. Standard errors are clustered by firm. ∗ ∗ ∗ significant at the 1% level, ∗ ∗ significant at the 5% level; ∗ significant at the 10% level.

insignificant). This finding is rather surprising, since many studies find that employers consider flexibility and adaptability to be important, but there are several potential explanations for this result. First, a willingness to participate in on-the-job training is only one dimension of flexibility, and it may be that other dimensions (e.g., being open to alternative solutions or accepting new tasks and roles) are more important. Second, some employers may interpret the sentence in the cover letter as “cheap talk” and hence ignore it. Third, the employers may not have seen the signal in the cover letters. An analysis of the subgroups shows that the age coefficients are similar in all groups.

Regarding firm heterogeneity, we have information about the firms’ gender composition and the gender of the contact person26 in the advertisement for a subsample of 4607 and 3204 resumes, respectively.27 The age effect is similar in the subsamples and in the full sample, which suggests that the subsamples are representative (Online Appendix Table A8). When we estimate separate age coefficients for male-dominated (0–50% females) and female-dominated (50–100% females) firms, we find that the negative age effect is very similar in both groups.28 The age effect is also very similar for firms with a female and male recruiter, respectively. 5.3. What do the employers in the survey say?

5.2. Does the age effect differ between occupations and firms?

To learn more about why firms may discriminate, we conducted a survey in which employers were asked a number of questions about how they perceive workers of different ages.29 In the spring of 2016, a representative sample of Swedish employers was contacted by phone to find a person involved in recruitment who could answer a web-based survey. For 3937 employers, a contact was established, and 1336 (34%) responded to the questions we use.30

To investigate differences in the age effect across occupations, we estimate a regression in which occupation is interacted with age. We include all interactions in a single regression without the covariate for age; thus, the coefficients for the interactions should be interpreted as the age effect for each occupation. The results show a statistically significant negative age effect in all occupations (panel A in Online Appendix Table A7). This finding confirms that age discrimination is a widespread phenomenon that is not limited to a few occupations. The results also show that there is variation in the magnitude of the age effect across the occupations; an F-test strongly rejects the hypothesis that the age coefficients are equal (p-value = 0.000). The age effect is largest for truck drivers, chefs, and sales representatives. We also estimate separate age coefficients for female and male applicants in each occupation (panel B). Since the regression does not include the covariate for age, the coefficients for the three-way interactions should be interpreted as the age effect for females and males, respectively, for each occupation. The results show a statistically significant steeper age-slope coefficient for women than for men for administrative assistants and cleaners (see the F-tests in the last row). In these occupations, the advantage of younger female applicants gives women a higher average callback rate.25

26 This contact person is likely to be the recruiter or someone who is at least involved in the recruitment process. We coded the gender based on this person’s name. 27 Unfortunately, we have no information about the firms’ age composition or the age of the recruiter. We have also tested for differences between firms located in Stockholm, which constitute approximately 70% of the sample, and those outside Stockholm (i.e., Gothenburg or Malmö) and find no statistically significant differences for the age coefficient. Finally, we have information on the firms’ size from Statistics Sweden. We divided the sample of firms based on the median number of employees (0–19 vs. >19). Again, we find no statistically significant differences for the age coefficient. 28 In this analysis, it may be argued that control variables for occupations and cities should be included, since there may exist a correlation between these variables and the gender composition of the firm or the gender of the recruiter. All results remain qualitatively unchanged if these controls are included. 29 Some of the questions in the survey were constructed specifically for this study, but the survey also contained other questions about aging and the labor market that are used in a larger project on aging. Carlsson and Eriksson (2017) present descriptive results for some additional questions in the survey. 30 A stratified sampling strategy was used, where the strata are based on the sector and the number of employees. The sectors are the government sector

25 An analysis of the effect of gender across the occupations shows that for administrative assistants and cleaners, women are, on average, preferred over men. Among truck drivers, chefs, and sales representatives, women have a slight disadvantage, on average, but this result is only statistically significant for chefs.

8

JID: LABECO

ARTICLE IN PRESS

M. Carlsson and S. Eriksson

[m5GeSdc;April 4, 2019;20:13] Labour Economics xxx (xxxx) xxx

Fig. 3. Difference in employers’ perceptions about worker characteristics relative to a worker aged 20. Notes: The graphs are constructed from the survey answers to the following question: “Suppose that you are recruiting a new employee to a typical position in your workplace. To what extent, do you think that an average employee at age 20, 30, 40, 50, and 60 would have the following characteristics”. Employers’ perceptions are measured in scores on a scale from 0 (to a very small extent) to 10 (to a very high extent). The graphs show the average difference in employers’ perceptions compared to a worker aged 20 (the horizontal line at zero). The confidence intervals are at the 95% level.

One question in the survey asked about a number of worker characteristics at different ages. The question was “Suppose that you are recruiting a new employee to a typical position in your workplace. To what extent do you think that an average employee at age 20, 30, 40, 50, and 60 would have the following characteristics?” The characteristics were as follows: (i) being self-sufficient, (ii) being able to learn new tasks, (iii) being flexible/adaptable, (iv) being ambitious, (v) being structured, (vi) having technical occupational skills, (vii) having communication skills, (viii) being reliable/loyal, (ix) having cooperation skills, and (x) having leadership skills.31 The scale used ranges from 0 (to a very small extent) to 10 (to a very large extent). The responses show that there are three characteristics that employers believe deteriorate with age: the ability to learn new tasks, flexibility/adaptability, and ambition. The results for these questions are shown in Fig. 3, which plots the average differences in the scores of workers aged 30, 40, 50, and 60 compared to the scores of workers aged 20 (i.e., relative to the horizontal line). The first two

characteristics peak at age 30 and then decline, while the last characteristic peaks at age 40 and then declines. The 95% confidence intervals show that all three declines are statistically significant. The fact that the perceived deterioration starts at age 30–40 suggests that at least part of the explanation for why employers discriminate against relatively young applicants could be that they statistically discriminate based on age and their perceptions about these characteristics. For the other seven characteristics (including occupational skills), the answers to the survey suggest that employers do not expect that these factors deteriorate markedly with age. These characteristics either stay rather constant in the older age categories or decline only in the oldest age category (i.e., at age 60; Online Appendix Figure A6). A general concern with surveys is the risk of social desirability bias. However, even if the responses are affected by a social desirability bias, it should be possible to draw conclusions about the relative importance of the characteristics if the bias is similar for all characteristics. There is no obvious reason why the employers in our survey would answer questions about some of these characteristics more honestly than questions about others. The results of other surveys on this topic are consistent with our results (Posthuma and Campion, 2009). Most other surveys ask employers (or students acting as employers) about whether older workers have fewer skills/abilities than younger workers have (Taylor and Walker, 1998; AARP, 2000; Gray and McGregor, 2003; Henkens, 2005; Swedish Pensions Agency, 2012). The results often show that employers worry that older workers are less able to learn new tasks, are less flexible/adaptable and are less ambitious. Additionally, there is often no clear evidence that employers believe that older workers have fewer oc-

(10%), the municipal sector (30%), and the private sector (60%). In each sector, all workplaces in the largest size categories and a random sample in the smallest size category were included. The firm administering the survey contacted 6063 workplaces by phone to find a person working with recruitment who could answer a web-based survey. A total of 3937 workplaces either agreed (3672) or refused (265) to receive the survey. A link to the survey was sent by e-mail, including several reminders to non-responders. 31 Another question in the survey asked about how important, in general, these characteristics were for the employers. The scale ranged from 0 (unimportant) to 10 (crucial). The average score for the ten characteristics varied between 6.45 and 8.87. This finding confirms that employers consider these characteristics to be important. 9

JID: LABECO

ARTICLE IN PRESS

M. Carlsson and S. Eriksson

[m5GeSdc;April 4, 2019;20:13] Labour Economics xxx (xxxx) xxx

cupational skills or are less productive. Some surveys also find that employers worry about older workers’ physical strength and health. These surveys often define older workers quite loosely, e.g., workers above age 50. This approach makes it somewhat difficult to compare these results to those in our experiment, which is very specific regarding the age of the applicants.

expect (firm-specific) training costs to be important, especially since all job applicants have at least ten years of occupation-relevant work experience. Additionally, worker mobility is lower for older workers (Online Appendix Figure A3); thus, it is not obvious that the expected tenure is shorter for older workers. Of course, we cannot exclude the possibility that some employers avoid older workers because they perceive them as costlier to hire, but if these perceptions are false, this is just an example of incorrect negative stereotypes with a similar effect on hiring as statistical discrimination. We conclude that it is unlikely that these other demand-related considerations are the main explanation of the negative age effect we find.

5.4. How should the results be interpreted? In this section, we discuss the possible mechanisms that could explain our main results, namely, strong discrimination against older workers in virtually all types of occupations. One potential mechanism is statistical discrimination. Our survey indicates that there are three characteristics that employers report as important and that they worry that workers over age 40 have started to lose: the ability to learn new tasks, being flexible/adaptable, and being ambitious. Hence, a potential explanation of age discrimination is that employers statistically discriminate because they believe that workers are beginning to lose these abilities in their 40s. An advantage of this explanation is that these abilities are general enough to be important in most occupations, albeit at different degrees. In contrast, the early onset of the age effect makes it less plausible that stereotypes about occupational skills, physical strength, and health are the main explanations of the age discrimination we find. It is unlikely that employers believe that workers in their early 40s lack important occupational skills, have low physical strength, or have poor health. Our survey contained a question about occupational skills, and the answers suggest that employers believe that these skills actually improve with age, at least until age 50. This finding is also consistent with the absence of a significant interaction effect between age and long-term unemployment, which is likely to capture employer stereotypes about how occupational skills vary by age. However, for the oldest workers (say, those aged 55–70), we cannot rule out that these factors have some relevance. Another potential mechanism is taste-based discrimination (i.e., ageism). However, it seems unlikely that ageism is the main explanation of the age discrimination we find, since it is unreasonable that such sentiments should affect workers in their early 40s. In addition, the responses to other survey questions reveal no evidence of ageism (Carlsson and Eriksson, 2017). Of course, we cannot entirely rule out this explanation, especially for the oldest workers (say, those aged 55– 70). A relevant question is whether there could be other demand-related explanations that explain the negative age effect. One such candidate explanation is that it may be costlier to hire older workers than younger workers. This could be the case if wage setting is based on seniority, such that employers must pay older workers a higher wage irrespective of their productivity. Seniority-based wages are common in some European countries, but in Sweden, age is not a relevant factor in the collective agreements that largely determine wages for workers in lowand medium-skilled occupations. Moreover, most of the theoretical arguments for seniority-based wages are built on long-term implicit contracts for existing employees (Lazear, 1979). Therefore, it is not obvious that these considerations should be important for the entry wages of new hires. Another reason why it may be costlier to hire older workers is that there may be other wage-related costs, e.g., pension contributions, that are higher for older workers. In Sweden, pension contributions for bluecollar workers are not age-dependent. Hence, for most of our occupations, this is not a relevant issue. Moreover, employer costs for vacations and sick pay are not linked to age in blue-collar occupations. In addition, employment protection legislation in Sweden is mostly linked to firm tenure rather than age, so it should not be an important factor for new hires. A third reason why it may be costlier to hire older workers is if employers pay substantial training costs for their newly hired workers. In this case, it may be argued that employers have incentives to avoid older workers, since such workers may be expected to remain employed for a shorter time. However, for the occupations we study, we do not

6. Conclusions In this study, we analyze the extent to which employers use information about a job applicant’s age in their hiring decisions. We conducted a field experiment in which over 6000 fictitious resumes with randomly assigned information about age (35–70 years) were sent to Swedish employers with a vacancy, and their responses (callbacks) were recorded. We find a strong negative age effect in all occupations that we investigate. The callback rate starts to decline substantially early in the age interval we consider, and closer to the retirement age, the callback rate is very low. The decline in the callback rate by age is steeper for women than for men. The early decline in the callback rate suggests that the main reason for age discrimination in the labor market is not about being older, e.g., above age 55, but rather about not being younger, e.g., below age 40–45. Employer stereotypes about the ability to learn new tasks, being flexible/adaptable, and being ambitious appear to be plausible explanations for age discrimination, while ageism and expectations about occupational skill loss due to aging are less plausible explanations. An important issue is the external validity of our results, i.e., the extent to which they are valid beyond the specific search channel, occupations, and cities we consider. In both Sweden and other countries, many vacancies are filled through informal search channels. However, there are no obvious reasons to expect that an applicant’s age should matter less in informal recruitments. Rather, it may be argued that discrimination could be more important when hiring is less transparent. The occupations we include are rather representative of the low- and medium-skilled parts of the labor market in both Sweden and most other Western countries. It is less certain whether the results are valid for highskilled occupations, but those occupations are difficult to study with this method. The fact that experience may be more important in high-skilled occupations suggests less age discrimination.32 However, the similarity of the age profiles for the duration of unemployment in both the entire labor market and in the labor market for low-skilled workers suggests that our results may be representative. Another aspect of external validity is whether our results are specific to Sweden or valid for other countries as well. Labor force participation among older people, especially women, is high in Sweden (OECD, 2017). Therefore, our results may be more representative of countries with high labor force participation in these groups. For countries with lower labor force participation among older people, the negative effects may be even larger, since employers in these countries may have less experience in hiring such workers. An interesting question is whether the strong negative demand effect of age that we find is visible in real-world economic data. In Fig. 1, we show that there is an almost linear increase in unemployment durations beginning at age 35. This pattern is consistent with our experimental results. The callback-rate-age profiles in the experiment are almost the inverses of the profiles for unemployment durations. Of course, labor supply could also matter, especially for workers above age 60, for

32 Kuhn and Shen (2013) and Delgado Helleseter et al. (2016) report evidence that advertisements for jobs with higher skill requirements are less likely to request workers of a specific age or gender.

10

JID: LABECO

ARTICLE IN PRESS

M. Carlsson and S. Eriksson

[m5GeSdc;April 4, 2019;20:13] Labour Economics xxx (xxxx) xxx

whom retirement is a realistic option. The increase in unemployment durations by age is somewhat steeper for women than for men, which is also consistent with the steeper decline in the callback-rate-age profile for women in the experiment. Job mobility also declines by age in a way that is consistent with the callback-rate profiles in the experiment. Although the main reason why job mobility is lower among older workers is most likely because many of them are well matched, the lower mobility may also reflect that some workers expect discrimination and therefore do not search for a new job or are unable to find a new job. Our results suggest that policymakers working with pension reforms face a twofold challenge. The negative demand effect is strongest for the oldest workers, and in addition, there is likely to be a negative supply effect among older workers. Therefore, if policymakers are to succeed in increasing employment among older workers, they must design measures that both increase labor supply and combat age discrimination.

Carlsson, M., 2011. Does hiring discrimination cause gender segregation in the Swedish labor market? Feminist Econ. 7 (3), 71–102. Carlsson, M., Eriksson, S., 2017. In: The Effect of Age And Gender on Labor Demand: Evidence from a Field Experiment, 2017, p. 8. IFAU working paper, IFAU. Carlsson, M., Rooth, D., 2007. Evidence of ethnic discrimination in the Swedish labor market using experimental data. Labour Econ. 14, 716–729. Chan, S., Huff Stevens, A., 2001. Job loss and employment patterns of older workers. J. Labor Econ. 19, 484–521. Delgado Helleseter, M., Kuhn, P., Shen, K., 2016. The Age Twist in Employers’ Gender Requests: Evidence from Four Job Boards NBER WP no. w22187. Eriksson, S., Rooth, D., 2014. Do employers use unemployment as a sorting criterion when hiring? Evidence from a field experiment. Am. Econ. Rev. 104 (3), 1014–1039. Farber, H.S., 1999. Mobility and stability: the dynamics of job change in labor markets. In: Handbook of Labor Economics, 3, pp. 2439–2483. Farber, H., Silverman, D., von Wachter, T., 2017. Factors determining callbacks to job applications by the unemployed: an audit study. RSF J. Soc. Sci. 3, 168–201. Gray, L., McGregor, J., 2003. Human resource development and older workers: stereotypes in New Zealand. Asia Pacific J. Hum. Resour. 41, 338–353. Heckman, J.J., 1998. Detecting discrimination. J. Econ. Perspect. 12, 101–116. Henkens, K., 2005. Stereotyping older workers and retirement: the manager’s point of view. Can. J. Aging 24, 353–366. Johnson, R.W., Neumark, D., 1996. Wage declines among old men. Rev. Econ. Stat. 78, 740–748. Kroft, K., Lange, F., Notowidigdo, M.J., 2013. Duration dependence and labor market conditions: theory and evidence from a field experiment. Q. J. Econ. 128, 1123–1167. Kuhn, P., Shen, K., 2013. Gender discrimination in job ads: evidence from China. Q. J. Econ. 128, 287–336. Lahey, J.N., 2008. Age, women, and hiring an experimental study. J. Hum. Resour. 43, 30–56. Lahey, J.N., Beasley, R., 2016. Technical Aspects of Correspondence Studies. National Bureau of Economic Research Working paper no 22818. Lahey, J.N., Oxley, D.R., 2016. Discrimination at the Intersection of Age, Race, And Gender: Evidence from a Lab-In-The-Field Experiment Working paper. Levy, J.A., 1988. Intersections of gender and aging. Sociol. Q. 29, 479–486. Lazear, E.P., 1979. Why is there mandatory retirement? J. Polit. Econ. 87 (6), 1261–1284. Neumark, D., 2012. Detecting discrimination in audit and correspondence studies. J. Hum. Resour. 47, 1128–1157. Neumark, D., 2018. Experimental research on labor market discrimination. J. Econ. Lit. 56, 799–866. Neumark, D., Bank, R., Van Nort, K., 1996. Sex discrimination in restaurant hiring: an audit study. Q. J. Econ. 111, 915–941. Neumark, D., Burn, I., Button, P., 2018. Is it harder for older workers to find jobs? New and improved evidence from a field experiment. J. Polit. Econ. forthcoming. OECD, 2013. Health At a Glance 2013: OECD Indicators. OECD Publishing, Paris. OECD, 2015. Effective Age of Labour Market exit, in Pensions at a Glance 2015: OECD and G20 Indicators. OECD Publishing, Paris. OECD, 2018. Employment Outlook. OECD Publishing, Paris. Petit, P., 2007. The effects of age and family constraints on gender hiring discrimination: A field experiment in the French financial sector. Labour Econ. 14, 371–391. Posthuma, R.A., Campion, M.A., 2009. Age stereotypes in the workplace: common stereotypes, moderators, and future research. J. Manag. 35, 158–188. Riach, P., Rich, J., 1987. Testing for sexual discrimination in the labour market. Australian Econ. Pap. 26, 165–178. Riach, P., Rich, J., 2002. Field experiments of discrimination in the market place. Econ. J. 112, F480–F518. Riach, P., Rich, J., 2006. An experimental investigation of sexual discrimination in hiring in the English labor market. B.E. J. Econ. Anal. Policy 6 (2) Art. 1. Riach, P., Rich, J., 2010. An experimental investigation of age discrimination in the English labor market. Ann. Econ. Statistics/Annales d’Économie et de Statistique 99/100, 169–185. Rich, J., 2014. What do Field Experiments of Discrimination in Markets Tell Us? A Meta Analysis of Studies Conducted Since 2000 IZA Discussion Papers No. 8584, IZA. Rupert, P., Zanella, G., 2015. Revisiting wage, earnings, and hours profiles. J. Monet. Econ. 72, 114–130. Swedish Pensions Agency, 2012. Arbetsgivares attityder till äldre arbetskraft. Pensionsmyndigheten analyserar 2012, 2 Pensionsmyndigheten. Taylor, P., Walker, A., 1998. Employers and older workers: attitudes and employment practices. Ageing Soc. 18, 641–658. Theodossiou, I., Zangelidis, A., 2009. Should I stay or should I go? The effect of gender, education and unemployment on labour market transitions. Labour Econ. 16, 566–577. Weichselbaumer, D., 2004. Is it sex or personality? The impact of sex-stereotypes on discrimination in applicant selection. Eastern Econ. J. 30, 159–186. Willis, R.J., 1985. Wage determinants: a survey and reinterpretation of human capital earnings functions. In: Ashenfelter, O., Card, D. (Eds.), Handbook of Labor Economics. North Holland, Amsterdam.

Acknowledgments We are grateful for helpful comments from Per Johansson and the seminar participants at the EALE 2018 Conference in Lyon, the conference the Labour Market with an Ageing Population 2018 in Uppsala, the conference Gender Economics of Labor Markets 2017 in Uppsala, the ESPE 2017 Conference in Glasgow, the Conference on Labour Market Policies 2017 in Stockholm, the Institute for Social Research in Oslo, the UCLS Members Meeting, Linnaeus University, and the Members Meetings for the project “The Effects of an Ageing Population: A Life-time Perspective on Work, Retirement, Housing and Health”. We would like to thank Josefine Andersson for her work with the survey. This work was supported by the Swedish Research Council for Health, Working Life and Welfare (FORTE grant number dnr 2013-2482) . Supplementary materials Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.labeco.2019.03.002. References AARP, 2000. American Business and Older workers: A summary of Findings. AARP, Washington. Ahmed, A.M., Andersson, L., Hammarstedt, M., 2012. Does age matter for employability? A field experiment on ageism in the Swedish labour market. Appl. Econ. Lett. 19, 403–406. Albert, R., Escot, L., Fernandez-Cornejo, J.A., 2011. A field experiment to study sex and age discrimination in the Madrid labour market. Int. J. Hum. Resour. Manag. 22 (2), 351–375. Arrow, K.J., 1973. The theory of discrimination. In: Ashenfelter, O., Rees, A. (Eds.), Discrimination in Labor Markets. Princeton University Press. Baert, S., Norga, J., Thuy, Y., Van Hecke, M., 2016a. Getting grey hairs in the labour market. A realistic experiment on age discrimination. J. Econ. Psychol. 57, 86–101. Baert, S., De Pauw, A.-S., Deschacht, N., 2016b. Do employer preferences contribute to sticky floors? Ind. Labor Relat. Rev. 69, 714–736. Bartoš, V., Bauer, M., Chytilová, J., Matějka, F., 2016. Attention discrimination: theory and field experiments with monitoring information acquisition. Am. Econ. Rev. 106, 1437–1475. Becker, G., 1957. The Economics of Discrimination. University of Chicago Press. Bendick M., Jackson, C.W., Romero, J.H., 1996. Employment discrimination against older workers: an experimental study of hiring practices. J. Aging . Soc. Policy 8, 25–46. Bendick M., Brown, L.E., Wall, K., 1999. No foot in the door: an experimental study of employment discrimination against older workers. J. Aging . Soc. Policy 10, 5–23. 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. Bertrand, M., Chugh, D., Mullainathan, S., 2005. Implicit discrimination. Am. Econ. Rev. 95 (2), 94–98. Blau, F.D., Kahn, L.M., 2017. The gender wage gap: Extent, trends, and explanations. J. Econ. Lit. 55, 789–863.

11