Are Women Discriminated Against in Countries with Extensive Family Policies? A Piece of the “Welfare State Paradox” Puzzle from Sweden

Abstract

A common assumption in comparative family policy studies is that employers statistically discriminate against women in countries with dual-earner family policy models. The empirical evidence cited in support of this assumption has exclusively been observational data, which should not be relied on to identify employer discrimination. In contrast, we investigate whether employers discriminate against women in Sweden—frequently viewed as epitomizing the dual-earner family policy model—using field experiment data. We find no evidence supporting the notion that Swedish employers statistically discriminate against women.

Introduction

In this study, we bring together two disconnected strands of research: comparative research on family policies, and research that uses field experiments to test for gender discrimination in the hiring of employees. We evaluate an accepted theory in the existing literature, and we do so using experimental data relating to a case that has been chosen strategically with this specific purpose in mind. Theoretical arguments on how rational—sometimes discriminatory—employers are expected to act depending on the family policy context are extremely common in comparative family policy studies. However, there has to date been a complete absence of studies testing whether employers actually do act in accordance with these theoretical conjectures, and ours is the first study to use adequate data to convincingly test whether employer choices actually conform to these conjectures, which are routinely predicted in the comparative literature on family policies. We do not claim originality, since the specific theoretical and methodological points underlying this article have been made before. However, they have not been brought together in a systematic way, and our innovation therefore lies in unifying the theory with a method adapted to testing it.

Sweden (together with the other Nordic countries) is well-known for its generous family policy model. Parental leave periods at almost full pay are extensive, and subsidized, high-quality, public childcare is easily accessible for parents of small children. These policies, paired with individual taxation, have created strong economic incentives for couples with children to be dual earners. As a result, the female and maternal employment rates are higher in Sweden than in most other OECD (Organisation for Economic Co-operation and Development) countries (OECD 2017; United Nations Development Programme 2016). Because Sweden is often heralded as the most clear-cut example of this family policy model, it is frequently at the center of the scholarly debates about these policies, and Swedish family policies are often used as blueprints for reforms in other countries (Duvander and Ferrarini 2010; Shalev 2008).

There are dissenting voices, however, and a steady stream of researchers have expressed concern that these family policies actually harm women’s careers rather than helping them. Since such effects are very far from the intention behind these policies, this has been labeled a “welfare state paradox” (e.g., Bergmann 2008; Blau and Kahn 1996; Gangl and Ziefle 2009; Hakim 2004; Mandel 2011; Mandel and Semyonov 2006; Mandel and Shalev 2009; Mun and Jung 2018; Pettit and Hook 2009; Ruhm 1998; Shalev 2008). Although there is agreement that Swedish family policy has facilitated female employment, the argument is that it inadvertently obstructs women’s opportunities to have successful careers. Mothers in Sweden stay home with small children to a greater extent, and for longer periods, than fathers. Parents in Sweden are also entitled to reduced working hours during their children’s preschool and elementary school years and, again, this entitlement is mostly used by mothers (Løvslett Danbolt 2016). Thus, if mothers, or mothers-to-be, are hired, they may spend long periods away from work, or work reduced hours. Swedish employers should therefore be more reluctant to hire them or to promote them to important positions. In other words, employers are expected to discriminate against mothers, not primarily based on taste, i.e., because they dislike mothers as employees, but rather because they practice statistical discrimination. That is, they purportedly use aggregate information about women as a group to screen out individual women in recruitment and promotion processes, based on the premise that female employees are on average likely to be less productive or reliable than male employees.

Without doubt, this is a convincing theoretical argument. But do Swedish employers statistically discriminate against women? Unfortunately, discriminatory behaviors are not identifiable in the kind of (quantitative) data employed in comparative family policy studies focused on analyses of gender gaps in labor market outcomes. Assuming the theoretical argument to be correct, however, one unequivocal prediction is that we should find negative hiring discrimination against women of fertile age in Sweden.

In this article, we approach this empirical question in a more direct way than the standard empirical approach employed in comparative family policy studies. We use field experiment data from a large-scale correspondence audit, in which non-authentic job applications were sent to real job openings with the gender of the job applicant randomly assigned, and in which employer responses to these applications were subsequently recorded. Our study was conducted in Sweden, often singled out as the country in which the “Nordic” family policy model has been implemented in the most consistent way. Thus, we are able to provide direct evidence on the question of whether Swedish employers—exemplifying employers in the Nordic family policy context—discriminate against women and mothers in hiring situations, and whether such patterns differ by job qualification level.

The Theory and its Empirical Support

Sweden and the other Nordic countries rank high on the United Nations Gender Development Index and Gender Equality Index respectively, indicating relatively small gender differences in, for example, life expectancy, education, standard of living, reproductive health, empowerment, and labor market outcomes (United Nations Development Programme 2016, tables 4 and 5). In relation to other OECD countries, female labor force participation is high in Sweden (ibid., table 5), the gender employment rate gap is small, even for mothers with young children (OECD 2017, figures 1.6 and 1.7), and so is the gender difference in unpaid working time (OECD 2012, figure 17.1). However, there are still also substantial gender inequalities in the Swedish labor market. Whereas Swedish men most often carry on working fulltime when they become fathers, women often start working (long-term) part-time when they return to work after parental leave (Kennerberg 2007). Moreover, mothers utilize a majority of all paid parental leave days and days spent caring for sick children (Swedish Social Insurance Agency 2017, 20–26). Regarding the gender pay gap, it is almost as large as the OECD average, and for the top decile in the earnings distribution it is larger than the OECD average (OECD 2012, figure 13.1). This gap can partly be explained by gender differences in working hours and job characteristics but a substantial part remains unexplained (ibid., figure 13.5). Female employees in Sweden and the other Nordic countries are underrepresented in senior management, perhaps more so than would be expected given the relative equality in other labor market domains, and more so than in many other countries (ibid., figure 14.1; also see Baxter and Wright 2000; Mandel and Semyonov 2006; Rosenfeld, van Buren, and Kalleberg 1998; Wright, Baxter, and Birkelund 1995; Yaish and Stier 2009). Using panel data, Bygren and Gähler (2012) have shown that this gap occurs in connection with parenthood. As long as women and men are childless, there is no difference in their chance of having a managerial position but when men become fathers their careers take off, whereas women’s stall when they become mothers.

Systematic employer discrimination against women, and in particular discrimination against women of fertile age, routinely figures as a plausible underlying mechanism for such gender gaps in countries with extensive family policies. Based on the notion of statistical discrimination, the expectation is that employers favor or disfavor individual job applicants based on group characteristics, e.g., known or expected lower average productivity (Phelps 1972) among women as a result of their greater family responsibilities, and longer and repeated expected periods of part-time work and work absence. Another version of statistical discrimination theory refers to group differences in variance rather than mean differences (Aigner and Cain 1977), but it is the latter that are (tacitly) claimed to lie behind employer discrimination against mothers (to be).

Because maternity leaves imply ex ante costs to employers, they generate effects that disadvantage all working women because employers either directly pass on these costs by lowering wages for women generally (cf. Ruhm 1998), or pass them on via statistical discrimination against women or working mothers in hiring and promotion decisions (e.g., Blau and Kahn 1996). The expectation is that the latter effect is particularly important in the context of lengthy statutory leave periods that strongly affect the expected duration of work interruptions following childbirth (Gangl and Ziefle 2009). That is, this effect should be particularly important in societies such as Sweden that have generous parental-leave programs.

Moreover, the statistical discrimination of women should be higher in highly qualified jobs. In well-developed welfare states in which gender-neutral social rights in practice support women’s absence from work, the exclusion of women from jobs which require costly firm-specific investment will be more acute, and employers are expected to prefer male workers for these positions (Mandel and Semyonov 2006). In a similar vein, Bergmann (2008) has argued that Nordic-style policies have some clear negative effects on female careers, and that these are caused by employer choices in this institutional context:

Employers would have strong incentives and more excuses to resist placing women in any but routine jobs—the kind of jobs where one person can smoothly and easily fill in for another. … Lengthy paid parental leave, as well as more part-time work, would increase discrimination and make women’s chances of getting such jobs considerably lower. (p. 353)

Arguments more or less identical to this can be found in Bygren and Gähler (2012), Gasser and Liechti (2015), Hakim (2004), Mandel (2011), Mandel and Shalev (2009), Pettit and Hook (2009), and Shalev (2008). Highly educated career-oriented women are thus likely to suffer more as a result of these policies, because employers find it riskier to hire or train women for jobs where they are hard to replace. That is, employers will discriminate more against women in highly qualified jobs in the event that family policies give them incentives to go on parental leave for extended periods of time. Among mothers, the number of parental leave days is inversely related to educational level. Nevertheless, mothers with a university education and mothers in leading positions on average spend around eight full-time months away from paid work after the birth of a child (Swedish Social Insurance Agency 2016). Indeed, as has been noted, fewer women have high-authority and elite positions in countries with developed family policy models (e.g., Mandel and Semyonov 2006; Shalev 2008). In line with this argument, Mandel (2012) has found that the gender income gap is particularly large among highly educated, high-income earners in countries with extensive family policies, and has concluded that progressive family policies are a likely cause of this via their adverse impact on women’s chances of entering highly paid positions. To conclude, then, among scholars researching gender gaps in the labor market, and motherhood penalties, there seems to be considerable consensus in the acceptance of the plausibility of the statistical discrimination explanation. A large number of scholars argue that statistical discrimination against women should be higher in societies with extensive family policies.

A rather blatant shortcoming in this field of research is that the mechanism hypothesized to generate the effects is never observed in the data on which these scholars base their conclusions. They all infer demand-side effects from observational supply-side data. To give one example, Shalev (2008) bases his conclusion that Swedish employers practice statistical discrimination against women on a cross-tabulation of gender by class, using individual-level survey data from the United States and Sweden. Many other examples of this type of theorizing about employer choices, using individual-level survey data to illustrate or test the empirical implications of the theory, could be listed. This ‘residual approach’ to the conceptualization and measurement of discrimination has—for good reason—been abandoned in modern approaches to the measurement of discrimination (Gaddis 2018). The disconnect between the theoretical mechanism and the observational data is obvious here. This empirical void is unfortunate, since employers are the crucial actors in the discrimination explanation for perverse effects of Nordic family policies, but the question of whether and in what ways generous family policies affect employer behaviors in the way predicted remains to be tested (see Mun and Jung 2018).

Observing and Quantifying Discrimination

While we have a great deal of descriptive knowledge about the way gender gaps develop over the life course, we know much less about the way employers act in relation to men and women and their (different) family responsibilities, or about the potential role of gender discrimination for these gaps. As has been noted, there has been a great deal of theorizing about the way employer discrimination is supposed to work to produce negative effects on women’s career attainment. Much less is known empirically, however, because observing discrimination in real labor markets is a challenge for research, not least because discrimination appears to be subtle, and to occur at a sometimes unconscious level (Petersen 2006; Van Bavel and Cunningham 2010).

A number of different approaches have been tried in order to detect, or approximate, discrimination. Exploring inequality in outcomes between different groups based on the analysis of large-scale observational datasets probably represents the most common technique (Pager and Shepherd 2008). Discrimination has been defined as the residual gap in the outcome between groups, controlling for observed confounders. In discrimination research, this is considered to be the least convincing approach to measuring the extent of discrimination. Although it provides evidence of unexplained inequality, it cannot provide reliable evidence of discrimination as such, since it depends on a rather unrealistic assumption that all other relevant characteristics are observed in the data. These studies are likely to suffer from endogeneity bias, as individuals who appear similar to researchers might in fact exhibit substantial heterogeneity and appear very different to employers (see Baert 2018). However, the residual gap approach nonetheless constitutes the implicit approach taken to the conceptualization of discrimination in comparative family policy studies (see Mandel 2012; Mandel and Semyonov 2006; Mandel and Shalev 2009; Shalev 2008).

Surveys and qualitative interviews represent another approach to studying discrimination. Employers—the potential discriminators—can be asked about their views and attitudes towards discrimination, or jobseekers—the possible victims—can be asked about their discrimination-related experiences. However, research indicates that there is a clear difference between what employers say and how they actually behave in hiring situations (e.g., Firth 1981; Pager and Quillian 2005). Also, the potential victims of discrimination may easily over- or underestimate levels of discrimination, and there is usually no way for applicants to know whether discrimination actually occurs because employers have good reasons for not revealing this to them.

Discrimination may also be measured by analyzing records of actual recruitment or hiring cases (e.g., Giuliano, Levine, and Leonard 2009; Petersen and Togstad 2006) or anonymous application procedures (cf. Goldin and Rouse 2000). However, the external validity of such analyses is limited by selection, because only companies and organizations that choose to allow the research can be studied. Companies that actually discriminate are probably less likely to participate in such studies, and companies that archive application documents and recruitment records are probably not representative. Thus, although this method may yield interesting in-depth information on specific cases, it cannot provide evidence of the extent of discrimination more generally.

As a response to the limitations of these approaches, the popularity of using experimental methods to study discrimination has grown significantly over recent years. Because of problems regarding the generalizability of the results from laboratory experiments, and the weaknesses of the other methods discussed, many researchers have adopted the use of field experiments to study labor market discrimination. Audit studies—introduced in the 1960s in the United Kingdom and the United States to study racial and ethnic discrimination—uses equally qualified actual test persons (actors) who apply for jobs in person, whereas in correspondence audits the equally qualified job applicants exist only on article, and written, non-authentic applications are used. These types of field experiments give the researcher a high degree of control over a treatment in a real-world context, and by comparison with other methods they give the researcher a better basis for calculating unbiased estimates of the level of discrimination (Gaddis 2018). Discrimination is measured based on the success of applicants in the hiring process, and the differential treatment of one group is interpreted as evidence of discrimination.

Empirical Strategy

The ideal dataset to test hypotheses on whether employers discriminate against women because of extensive family policies would be to have comparative correspondence audit data covering a period during which there is variation in the extent of these policies both within and across countries. Since such data do not (yet) exist, we will resort to using data for a strategic case (Sweden) for which there are clear-cut expectations with regard to employer discrimination patterns, and where high-quality data are available to test such expectations. Single-country studies are—perhaps counterintuitively—becoming increasingly important in comparative policy research because they are often in a much better position to uncover the precise causal mechanisms underlying the effects of policies (see Gangl and Ziefle 2015; Mun and Jung 2018), while at the same time concerns over internal validity are much reduced by comparison with cross-national comparisons (Pepinsky 2019). The data set used in this study is tailored to measure the extent of discrimination in the labor market, thereby putting us in a much-improved position to judge whether statistical discrimination is in fact practiced against women in a country with extensive family policies.

Hypotheses

Based on our review of the literature, we formulate the following empirically testable hypotheses on employer choices in a country with extensive family policies. First,

H1. Women of fertile age are discriminated against by employers in hiring situations.

Empirically, we translate this hypothesis into the following simple empirical model:

$$\mathrm{Prob}\left(C_i\right)={\beta }_0+{\beta }_1{W}_i+{\epsilon }_i$$

(1)

where $C_i$ is a dichotomous variable indicating whether applicant i receives a positive response from the employer to whom s/he has sent a job application, $W_i$ is a dummy for woman, and ${\epsilon }_i$ is an error term. $W_i$ is randomly assigned to job applications, and all job applicants signal an age of thirty-one years, i.e., a fertile age for women. Our expectation is that the coefficient ${\beta }_1$ in this equation will indicate a negative relationship between being a female applicant and the probability of receiving a positive response from the employer.

Secondly, based on the notion that employees in highly qualified jobs are not as easily replaced and that absence is costly for the employer (cf. Bergmann 2008; Mandel and Semyonov 2006; Williams 2000), we formulate the following hypothesis:

H2. In hiring situations, employers discriminate more against women of fertile age applying for highly qualified jobs than against women applying for less qualified jobs.

This hypothesis is tested by including an interaction term between having a female name and the qualification level of the job applied for, together with main effects of these terms:

$$\mathrm{Prob}\left(C_i\right)={\beta }_0+{\beta }_1{W}_i+{\beta }_2{H}_i+{\beta }_3{W}_i{H}_i+{\epsilon }_i$$

(2)

where $W_i$ is again a dummy for woman, and $H_i$ is a dummy with the value 1 if the job applied for is a highly qualified job, otherwise 0. We expect ${\beta }_3$ in this equation to be negative.

Design of the Correspondence Audits

The data were collected in 2013–2019 and are based on applications for 6,755 jobs. An unpaired design was employed, i.e., one application was sent to each vacancy (see Vuolo, Uggen, and Lageson 2018). The gender of the applicant was randomly assigned. For a portion of the applications, foreign-sounding names were used. These were, in contrast to the Swedish-sounding names, typical Yugoslavian and Middle Eastern names. We consulted Statistics Sweden to find the most common first and last names in Sweden. The age was set to thirty-one for all applicants, which is close to the median age at first parity for women in Sweden.

As the gender of the job applicant was randomized to every application, the data may be seen as an experiment in which the subjects are the employers and the treatment is the gender of the job applicant, signaled by his or her name. We selected the occupations with the intent of creating a mix with regard to gender composition, immigrant composition, educational level, and sector (see Bursell 2014). We further selected occupations in which relatively simple job applications would not be considered out of line. Jobs were consequently applied for in fifteen occupations: accountant/auditor, assistant nurse, carpenter, chef, cleaner, computer specialist, driver, engineer (in machine technology, industrial economics, or electronics), financial assistant, teacher (elementary or high school), nurse, preschool teacher, receptionist, salesperson, and store personnel/cashier. This selection of occupations is reasonably typical for the Swedish labor market; of the ten largest occupations in the Swedish labor market, seven are included here. We classified some of these occupations as qualified because as a rule they require tertiary education (accountant/auditor, computer specialist, engineer, nurse, preschool teacher, and school teacher). We further classified the occupations with regard to the proportion of female employees, using Statistics Sweden’s occupational register for 2016 divided into three categories: female dominated (above 60 percent women), gender balanced (between 40 percent and 60 percent women), and male dominated (below 40 percent women).¹

Job openings were found through the Swedish Public Employment Service’s website, which is the largest website for job postings in Sweden. Most of the jobs we applied for were located in, or close to, the three largest Swedish cities: Stockholm, Gothenburg, and Malmo.

The job applications consisted of two parts. First, a short-form CV that included contact information (name, telephone number, e-mail address, and postal address), date of birth, prior work experience, and education. All applicants had an education and work experience relevant to the job. Second, a cover letter including the reason for applying for a new job, some more information about experiences in the applicants’ current and previous jobs, additional qualifications, and some information on leisure activities. E-mail addresses and registered telephone numbers connected to a voice messaging service were set up for all of the non-authentic applicants.

We took the following steps to make the job applications appropriate for the different occupational contexts. First, we examined authentic job applications, available at the Swedish Public Employment Service’s website, and used these as templates. Second, we examined authentic job announcements to find the merits that employers typically demand for these occupations, and incorporated these into our job applications. Third, we consulted professional recruiters and practitioners in the occupations to review drafts of the non-authentic job applications, to ensure these appeared as “real” as possible. After this review process we reached the final versions of the non-authentic job applications that were sent to actual employers.

We define a callback as a non-automatic and non-negative response by the employer via e-mail, text, or voice message. Typical responses involved direct job offers, job interview offers, and requests for more information. Since our intention with the callback variable was to tap into employer intentions to employ, we treated these as “callbacks”, but treated the explicitly negative callbacks, and nonresponses, as “no callbacks”. We report descriptive statistics in table 1. In table 2, we report a test of random treatment assignment. The gender of the job applicant is regressed on the variables in table 1 (except the indicators of proportion of female employees, as they are determined by occupation). If the randomization has been successful, these variables should not predict treatment status. None of the variables are significantly associated with treatment status, and the P-value for the F-test of joint significance is equal to 0.928, allowing the conclusion that the randomization was indeed successful.

Table 1

Descriptive statistics, percentages

Variable	All	Female name	Male name
Non-negative callback	30.7	31.8	29.7
2013	4.5	4.7	4.3
2014	20.0	20.5	19.6
2015	7.2	6.8	7.6
2017	14.2	14.4	14.0
2018	43.7	43.3	44.2
2019	10.3	10.3	10.4
Foreign name	41.9	41.8	41.9
Store personnel	7.4	7.1	7.7
Engineer	3.5	3.7	3.3
Computer specialist	7.2	6.9	7.4
Financial assistant	6.7	6.4	6.9
Driver	8.7	9.0	8.5
Preschool teacher	7.4	7.4	7.5
School teacher	5.1	4.9	5.2
Chef	11.8	12.4	11.3
Cleaner	6.5	6.3	6.6
Receptionist	3.3	3.5	3.2
Accountant/Auditor	7.2	7.1	7.4
Salesperson	12.1	11.8	12.3
Nurse	4.8	4.8	4.8
Assistant Nurse	5.5	5.7	5.3*
Carpenter	2.9	3.1	2.6
Female-dominated occupation	29.1	30.4	27.8
Gender-balanced occupation	29.3	29.2	29.3
Male-dominated occupation	36.1	38.0	34.2
Number of observations	6,755	3,375	3,380

Table 2

Test of randomization of gender to job applications

Variables
Foreign name	0.003
	(0.013)
Store personnel	–0.024
	(0.028)
Engineer	0.029
	(0.036)
Computer specialist	–0.021
	(0.028)
Financial assistant	–0.021
	(0.028)
Driver	0.015
	(0.026)
Preschool teacher	–0.005
	(0.028)
School teacher	–0.017
	(0.032)
Chef	0.022
	(0.024)
Cleaner	–0.014
	(0.029)
Receptionist	0.016
	(0.037)
Accountant/Auditor	–0.014
	(0.028)
Nurse	–0.003
	(0.032)
Assistant Nurse	0.019
	(0.030)
2013	0.036
	(0.031)
2014	0.023
	(0.018)
2015	–0.017
	(0.025)
2017	0.014
	(0.019)
2019	0.004
	(0.021)
Constant	0.492**
	(0.018)
Observations	6,755
Model F-test P-value	0.928

**

P < 0.01;

*

P < 0.05.

Note: The reference category for occupation is Salesperson, and the reference category for year is 2018.

To assess whether the sample size in this study (6,755) is sufficiently large to rule out a Type-II error, we computed the required sample sizes for different hypothetical (negative) effects of a female name, given that men’s positive response would equal 30 percent (rounded from the men’s observed proportion of 29.7 percent, see table 1). We report the result of this sensitivity analysis in figure A1, where required sample sizes by alternative callback proportions for women are reported, given that α error prob. = 0.05 and 1 – β error prob. = 0.8. In these scenarios, we can see that as long as the women’s true callback rate is less than 28.4 percent, i.e., with a negative effect size in excess of 1.6 percentage points, our sample size is sufficiently large. When the negative effect of a female name is 1.6 percentage points or smaller, however, it would be necessary to increase the sample size to rule out the possibility of a “false null” result. An effect size of that magnitude would mean that women would need to apply for up to 6 percent more jobs than men to reach an expected parity with regard to the number of positive callbacks. For a subgroup analysis using one-third of the sample, the corresponding effect size would be 2.7 percentage points. An effect size of that magnitude would mean that women would need to apply for up to 10 percent more jobs than men to reach an expected parity with regard to the number of positive callbacks. As true negative effect sizes in this region by any standards would be judged to be small, however, we are confident that our sample size is sufficient, at least given the specific aim of the study.

Results

In the introduction, we formulated two hypotheses. The first (H1), based on the reasoning that extensive family policies (potentially) imply lengthy periods of work absence among female employees of fertile age, stated that employers are likely to discriminate negatively against these women in hiring situations. Based on our data, we find no indication that women are discriminated against by Swedish employers. In fact, the aggregate callback rate (for all occupations) is slightly higher for women, 31.8 percent, than for men, 29.7 percent (see table 1). This difference is not statistically significant, however (table 3a, column 1). It turns significant when we control for study year and a number of applicant characteristics, i.e., foreign name (the estimate is not reported here but, in line with previous Swedish studies, e.g., Arai, Bursell, and Nekby (2016), Bursell (2014), and Carlsson and Rooth (2007), indicates a powerful negative effect on callback rates) and occupation. These controls increase the efficiency of the estimate, but its direction and size remain unaltered. That is, if anything employers seem to have a weak preference to hire women over men, which is in direct contrast to H1.

Table 3a

Linear probability model regression of callbacks on applicant characteristics

Variables	1	2	3	4
Female name	0.021	0.021*	0.016	0.014
	(0.011)	(0.010)	(0.012)	(0.012)
Tertiary education requirement			0.258**
			(0.017)
Female name × Tertiary education requirement			0.019	0.018
			(0.024)	(0.023)
Controls	No	Yes	No	Yes
Observations	6,755	6,755	6,755	6,755
Adjusted R2	0.000	0.132	0.077	0.132

Notes: Robust standard errors in parentheses. Significance levels indicated by **P < 0.01;

*

P < 0.05. Controls include the “foreignness” of the job applicant’s name, occupation fixed effects and year fixed effects.

The second hypothesis (H2), based on the notion that employees in highly qualified jobs are more difficult to replace and that absence from these positions is more costly for the employer, stated that women of fertile age applying for highly qualified jobs should be more discriminated against by employers in hiring situations than women applying for less qualified jobs. We find no support for this hypothesis. When we turn to highly qualified jobs, i.e., jobs requiring tertiary education, we find that the general callback rate is higher, but this does not alter the picture for gender; there is no support for the hypothesis that female applicants are discriminated against in relation to male applicants in the competition for highly qualified jobs. Again, we instead find some evidence pointing in the other direction, that employers prefer to hire women for qualified jobs, as the interaction term female name by qualified job is positive. However, this estimate is imprecisely estimated and is not statistically significant (table 3a, column 3).

As a sensitivity test, we re-ran the models, restricting the sample to job applicants with “native”-sounding names (table 3b). All gender-related estimates and t-values are then closer to zero, indicating that the weak preference for women reported in Table 3a seems to be driven by a weak preference for female job applicants with foreign-sounding names over male job applicants with foreign-sounding names. This interaction effect between female name and foreign name is sufficient to move the point estimate somewhat but it is not statistically significantly different from zero (see table A1).

Table 3b

Linear probability model regression of callbacks on applicant characteristics, only native-sounding names.

Variables	1	2	3	4
Female name	0.005	0.006	0.005	0.004
	(0.015)	(0.014)	(0.017)	(0.017)
Tertiary education requirement			0.304**
			(0.022)
Female name × Tertiary education requirement			–0.004	0.004
			(0.032)	(0.031)
Controls	No	Yes	No	Yes
Observations	3,928	3,928	3,928	3,928
Adjusted R2	0.000	0.132	0.090	0.132

Notes: Robust standard errors in parentheses. Significance levels indicated by **P < 0.01;

*

P < 0.05. Controls include occupation fixed effects and year fixed effects.

To explore patterns further, we used the scaling of occupations based on the proportion of employees comprised of women, and interacted this with the gender of the job applicant. As part-time work is less common in male-dominated occupations, statistical discrimination against female job applicants should be higher in male-dominated jobs, which might be reinforced by stereotypes and expectations favoring the gender-typical job applicant, i.e., favoring women or men depending on the stereotypes associated with the occupation. These expectations were not borne out by our data. We found female job applicants to have a slight advantage over male job applicants in both male-dominated occupations and female-dominated occupations, but none of the gender differences in callbacks are statistically significantly different from zero (see figure 1; for the gender-specific callback rates, see table 1).

Figure 1

Percentage point employer callback effects, with 95 percent confidence intervals, of signaling a female name, across shares of women in the occupation (reference: male name, n  = 6,755)

As noted, we defined callback as a non-negative response by the employer. This includes, for example, employer requests for more information and direct job offers. Thus, there is a potential risk for gender bias in type of employer callback. As a sensitivity analysis, we redefined callbacks to only include invitations to interview or meeting, or direct job offers. Results from this analysis are very similar to results for analyses on the more inclusive definition of employer callback. In other words, restricting employer callbacks only to interview invitations or job offers does not indicate the presence of employer discrimination by gender, neither in general (table A2) nor for native-sounding names only (table A3), and regardless of share of women in the occupation (figure A2).

Discussion

For quite some time now, there has been a steady stream of articles on the extensive family policies typical for, for example, Nordic societies, and their potential impact on gender gaps in the labor market. A recurrent concern is that these policies encourage statistical discrimination against women. In Sweden and the other Nordic countries, parents of small children are entitled to long periods of paid parental leave, they have the right to work reduced hours, and they have the right to stay home from work when children are sick (Løvslett Danbolt 2016). Although these rights are gender-neutral, i.e., fathers have the same rights as mothers, and although the actual use of these rights has become more gender equal over time, they are still mostly used by mothers. Thus, employed mothers of younger children spend longer periods away from work, and they work fewer hours than fathers.

Based on these statistical facts, quite a few scholars have claimed that employers should be reluctant to hire and promote mothers and mothers-to-be, and especially so in countries like Sweden where these rights are more extensive than in most other countries (e.g., Bergmann 2008; Blau and Kahn 1996; Bygren and Gähler 2012; Hakim 2004; Mandel 2011; Mandel and Semyonov 2006; Mandel and Shalev 2009; Ruhm 1998; Shalev 2008). Faced with two equally qualified applicants in, say, their late twenties or early thirties (the age at which most Swedish women have their first child), then, employers would be expected to pick the male applicant over the female, i.e., to statistically discriminate against the woman, based on the knowledge that most mothers are absent from work for longer periods than most fathers, and that mothers are more likely to reduce their work hours. This discrimination should be even stronger when employers run higher risks, i.e., have more to lose, by hiring or promoting the “wrong” applicant, for example for highly qualified positions in which employees are difficult to replace and where absence is costly for the employer. This theoretical idea has had resonance among scholars, as there is indeed a correlation between extensive family policies and relatively large gender gaps in career indicators (Bergmann 2008; Mandel and Semyonov 2006; Mandel and Shalev 2009). However, correlation does not equal causation and since the theoretical mechanism postulated to be in operation—discrimination—has never been observed empirically in these studies, we do not know whether it is actually operating to produce these effects.

Sweden epitomizes the type of family policy model that is hypothesized to generate these kinds of employer responses. So, do Swedish employers discriminate against women of fertile age in hiring situations? To answer this question we used experimental data from a correspondence audit conducted in 2013–2019, in which 6,755 non-authentic job applications were sent in reply to job advertisements for jobs in fifteen different occupations. We applied for jobs in a large number of occupations typical of the Swedish labor market, exhibiting large variation in gender composition and the qualifications required.

Judging from the aggregated employer callback rates, 31.8 percent for women and 29.7 percent for men, and the statistically non-significant difference of 2.1 percentage points, we cannot reject the null of no gender discrimination. This conclusion is strengthened when we estimate the gender difference for job applicants with ‘native’ names, which is equal to 0.5 percentage points to the advantage of women, and not statistically significant. In other words, we find no evidence that Swedish employers practice hiring discrimination to the disadvantage of women. Moreover, we cannot reject the null of no interaction effect between the job applicant’s gender and the job’s qualification level, and we cannot reject the null of no interaction effect between the job applicant’s gender and the proportion of employees comprised of women in a given occupation. Thus, the null of no discrimination holds regardless of whether or not the job requires tertiary level education, and regardless of occupational gender composition.

The method used here only allows for the study of employer discrimination at the hiring stage. Women might be discriminated against within workplaces with regard to, for example, promotion, wage-setting or lay-offs. This is a possibility that certainly deserves further study but we would claim that such reasoning, in the case at hand, has three weaknesses: (i) If employers statistically discriminate against women, why wait until the interview phase, or even later in the hiring phase? Statistical discrimination is by its very essence a type of discrimination that is practiced in situations in which there is only surface-level information available, and the first-stage screening of job applicants typifies such a situation. (ii) If employers have a tendency to discriminate against women, why would this not be visible in the initial stage of the hiring process? In fact, studies show that up to 90 percent of discrimination based on ethnicity takes place at this initial stage rather than later on (Allasino et al. 2004; Attström 2007; Cediey and Foroni 2008). We find it reasonable to believe that the pattern, i.e., that most hiring discrimination in the event that it does occur, occurs in the first stage of the process, is similar, or at least not entirely different, regardless of the ground for discrimination. (iii) Discrimination based on gender may still exist among certain employers and against specific individuals. The general pattern found here does not imply that there are no exceptions to this rule of equal treatment. If so, however, then for recruitment decisions in which there has been negative discrimination against female applicants, there should be a roughly equal number of recruitment decisions in which women have benefited from positive discrimination. This is the logical consequence of the fact that we find no evidence of systematic employer gender discrimination at the aggregate level.

To conclude, then, (i) we do not find evidence that women of child-bearing age are discriminated against in hiring situations, and (ii) this result applies regardless of job qualification level and occupational gender composition. Our results rather suggest that women are weakly preferred over men under some conditions: when they signal that they have an immigrant background or when they apply for a job in occupations that are not gender balanced. Our results align rather well with three other Swedish studies based on correspondence audits. In line with our own findings, Ahmed, Andersson, and Hammarstedt (2013), Bygren, Erlandsson, and Gähler (2017), and Carlsson (2011) have all reported very small gender differences in callback rates. Ahmed, Andersson, and Hammarstedt (2013), however, focused on discrimination based on the sexual orientation of childless applicants, the non-authentic applicants in the study by Carlsson (2011) were all relatively young, with expected parenthood still lying years ahead, and the study by Bygren, Erlandsson, and Gähler (2017) was based on an underpowered sample of job applications by comparison with the one used in this study (see the Design section for a discussion on the current study’s statistical power). Thus, the argument made by a rather lengthy list of scholars (e.g., Bergmann 2008; Blau and Kahn 1996; Gangl and Ziefle 2009; Mandel and Semyonov 2006; Ruhm 1998; Shalev 2008), that extensive family policies should instill particularly discriminatory behaviors in employers against women of fertile age, does not seem to hold when it is subject to stringent empirical testing.

Our results are also in line with the most common finding reported in the rapidly expanding literature that is using correspondence audits to test for gender hiring discrimination in different institutional contexts (see Baert 2018). The rule seems to be an absence of statistically significant gender differences in employer callbacks (Baert 2015; Baert, Pauw, and Deschacht 2016; Capéau et al. 2012 [Belgium]; Zhou, Zhang, and Song 2013 [China]; Albert, Escot, and Fernández-Cornejo 2011 [Spain]). In some studies, female applicants have been found to be subject to positive discrimination (Booth and Leigh 2010 [Australia]; Berson 2012 [France]; Jackson 2009 [the United Kingdom]), whereas in some studies, female applicants have been found to be negatively discriminated against (Petit 2007 [France]; Riach and Rich 2006 [the United Kingdom]).

Two facts are notable here: (i) Sweden seems to be a typical rather than deviant case with regard to the degree of gender hiring discrimination practiced by employers. (ii) Family policies in all the aforementioned countries are less extensive than Swedish family policies. This suggests that explanations for gender inequality on the Swedish labor market should be sought elsewhere. Evertsson (2013) has reported that Swedish women’s work commitment decreases after parenthood. This gendered pattern may be a consequence of implicit normative messages in Nordic-style parental leave rights whereby, once you are given these rights, opting out may be stigmatizing for many mothers given the strong cultural and social forces that shape men’s and women’s choices as parents also in this gender egalitarian context (Løvslett Danbolt 2016, also see Gangl and Ziefle 2015). Thus, a possible explanation for the fact that women’s careers stall in countries with relatively ambitious family policies is that these policies—when they are implemented in a context where there is still some rather strong norms on the gendered division of unpaid and paid work—unintentionally cause “career damaging” changes in women’s attitudes and labor market related choices, rather than causing changes in employers’ propensities to hire or promote women. Researchers interested in the mechanisms behind the potential damage that family policies do to women’s careers would therefore probably be well advised to focus on the supply-side effects these policies may produce, in particular the disincentives they provide to full-time paid work and the potential normative messages they carry.

Appendix

Table A1

Linear probability model regression of callbacks on applicant characteristics

Variables	1	2
Female name	0.005	0.006
	(0.015)	(0.014)
Foreign name	–0.134**	–0.119**
	(0.015)	(0.015)
Female name × Foreign name	0.036	0.037
	(0.022)	(0.021)
Controls	No	Yes
Observations	6,755	6,755
Adjusted R2	0.016	0.132

Notes: Robust standard errors in parentheses. Significance levels indicated by

**

P < 0.01;

*

P < 0.05.

Controls include occupation fixed effects and year fixed effects.

Table A2

Linear probability model regression of employer callbacks on applicant characteristics, callbacks defined as interview invitations

Variables	1	2	3	4
Female name	0.011	0.012	0.009	0.008
	(0.009)	(0.009)	(0.009)	(0.009)
Tertiary education requirement			0.177**
			(0.014)
Female name × Tertiary education requirement			0.012	0.012
			(0.021)	(0.020)
Controls	No	Yes	No	Yes
Observations	6,755	6,755	6,755	6,755
Adjusted R2	0.000	0.098	0.056	0.098

Notes: Robust standard errors in parentheses. Significance levels indicated by

**

P < 0.01;

*

P < 0.05.

Controls include the ‘foreignness’ of the job applicant’s name, occupation fixed effects and year fixed effects.

Table A3

Linear probability model regression of employer callbacks on applicant characteristics, callbacks defined as interview invitations, only native-sounding names.

Variables	1	2	3	4
Female name	0.001	0.002	0.006	0.004
	(0.013)	(0.012)	(0.013)	(0.013)
Tertiary education requirement			0.219**
			(0.020)
Female name × Tertiary education requirement			–0.017	–0.005
			(0.028)	(0.028)
Controls	No	Yes	No	Yes
Observations	3,928	3,928	3,928	3,928
Adjusted R2	0.000	0.107	0.065	0.107

Notes: Robust standard errors in parentheses. Significance levels indicated by

**

P < 0.01;

*

P < 0.05.

Controls include occupation fixed effects and year fixed effects.

Figure A1

Required minimal sample sizes for negative effects of a female name to be detectable, by levels of alternative positive response for women.

Note: α error prob. = 0.05, 1 – β error prob. = 0.8, and true contrast proportion positive response = 0.3.

Figure A2

Percentage point differences in employer interview invitations, with 95 percent confidence intervals, of signaling a female name, across shares of women in the occupation (reference: male name, n  = 6,755) Example of job application (nurse).

CV

Name: [First name, Last name]^a

Birthdate: [YYYY-MM-DD]

Address: [Street name and postal code]

Phone: [Phone number]

E-mail: [E-mail address]

Work Experience

2007–to present Nurse, Stockholm South General Hospital, Emergency Department

2006–2007 School nurse, Hässelby High School

2004–2006 Nurse, Danderyd Hospital, Rheumatology

2003–2004 Cashier, ICA

Education

2001–2004 Bachelor of Nursing, Karolinska Institutet

1998–2001 Health Care Programme, St. Göran Upper Secondary School

Language

Swedish (mother tongue) and English

Computer Skills

Office suite, Melior, Take Care

^aThe names used in the study are: Malin Andersson, Sara Eriksson, Maria Johansson, Anna Karlsson, Elin Nilsson (Swedish-sounding female names), Gustav Andersson, Daniel Eriksson, Erik Johansson, Johan Karlsson, Fredrik Nilsson (Swedish-sounding male names), Amina Abdullah, Zahra Ali, Fatima Ahmed, Jelena Jovanovic, Jelena Nikolic, Jovana Petrovic, Milena Popović, Samira Said (foreign-sounding female names), Mohammed Abdullah, Omar Ali, Hassan Ahmed, Aleksandar Jovanovic, Aleksandar Nikolic, Bojan Petrovic, Dragan Popović, and Hassan Said (foreign-sounding male names).

Name: [First name, Last name]^b

Birthdate: [YYYY-MM-DD]

Address: [Street name and postal code]

Phone: [Phone number]

E-mail: [E-mail address]

Application

My name is [First name, Last name] and I am 31 years old. I am very interested in the announced vacancy.

I am a registered nurse and have worked in healthcare since 2004. I am really happy with my current job but I am now looking for new challenges and am therefore applying for the position.

As a person, I am flexible, responsible and find it easy to collaborate with others. After all these years as a nurse in emergency care, I am used to working at a fast pace and I am good at coping with multiple tasks simultaneously. [I always perform my duties in a way that ensures patient safety. In my previous work, I have developed a strong ability to see and respect the patients’ different, individual needs.]^c

I live in Stockholm [with my husband/wife/cohabitant]^d [who is a primary school teacher/engineer]^e [and our two children/with my two children].^f As a person I am sociable, and in my leisure time I spend time with [my husband/wife/partner/family and] my friends. I like to exercise and enjoy cooking. [According to my colleagues and friends, I am an empathetic and caring person who shows consideration and respect for other people’s needs. I like to help people who need support and have worked as a volunteer for the Red Cross.]^g

[Work is an important part of my life and my ambition is always to carry out my duties in the best possible way. My references will confirm that I do that little bit extra to produce results that exceed expectations.]^h

I look forward to meeting you in person. References are available upon request.

Sincerely,

[First name, Last name]

^bp = 0.29 Swedish-sounding female name; p = 0.29 Swedish sounding male name; p = 0.21 foreign sounding female name; p = 0.21 foreign sounding male name.

^c Signals competence: p = 0.17 included, p = 0.83 blank.

^d p = 0.33 married, p = 0.33 cohabitant, p = 0.33 blank.

^e Partner’s occupation if partner is present: p = 0.25 engineer, p = 0.25 primary school teacher, p = 0.5 blank.

^f p = 0.5 included, p = 0.5 blank.

^g Signals warmth: p = 0.17 included, p = 0.83 blank.

^h Signals commitment: p = 0.17 included, p = 0.83 blank.

Funding

This work was supported by the Swedish Research Council for Health, Working Life and Welfare (Forte grant #2018-00594).

Endnotes

1

We classified the following occupations as female dominated: Store personnel, Financial assistant, Preschool teacher, School teacher, Cleaner, Receptionist, Nurse, Assistant Nurse. We classified the following occupations as gender balanced: Chef, Accountant/Auditor, Salesperson. We classified the following occupations as male dominated: Engineer, Computer specialist, Driver, Carpenter.