WHAT’S THE DIFFERENCE?
A DESCRIPTIVE ANALYSIS OF THE EVOLUTION OF THE FAMILY GAP IN S WEDEN
Submitted by Anna Fornwall
Department of Economics
Supervisor Håkan Selin
Spring term, 2019
ABSTRACT
In this study, I compare men and women with and without children to analyze the effect of children on wages and earnings. By comparing the gender wage gap to the family gap for men and women respectively, I find that there is still a persistent, yet rather small, family gap for women. The constant family gap for women supports the notion that a greater fraction of the gender wage gap can be explained by effects of having children now than previously.
When using yearly earnings instead of hourly wages, the gender wage gap increases whereas the family gap for women decreases.
This implies that although there are several policies with the aim of reducing gender wage differences and creating possibilities for women to combine work and family, there are still concrete effects that arise from taking the responsibility for children. Because the effect of having children is seemingly constant over time for women, the results from this study imply that specific policies are needed to prevent and battle the difference in labor market outcomes that arise because of the differing effects from caring for children.
Keywords: Gender wage gap, Family gap, Wage inequality, Child penalty
C ONTENTS
1 INTRODUCTION 4
2 BACKGROUND ANDLITERATUREREVIEW 6
2.1 THE GENDERWAGEGAP . . . 6
2.2 DIFFERENCES IN THE USE OF PARENTAL LEAVE AND FAMILY RESPONSI- BILITY . . . 6
2.3 EFFECTS OF BEING OUT OF THELABORFORCE . . . 7
2.4 THE FAMILYGAP– ESTIMATING THEWAGEIMPACT OFCHILDREN . . . . 8
3 THEORETICAL FRAMEWORK 11 3.1 LABOR SUPPLY THEORY . . . 11
3.2 HUMAN CAPITAL THEORY . . . 12
3.3 SIGNALING THEORY AND STATISTICAL DISCRIMINATION . . . 14
3.4 SELF-SELECTION THEORY . . . 14
4 DATA 16 4.1 DATA SOURCE . . . 16
4.2 VARIABLE DESCRIPTION . . . 17
5 EMPIRICAL METHOD 21 5.1 ORDINARY LEASTSQUARES . . . 21
5.2 FAMILY AND GENDER GAP INEACH WAVE . . . 21
5.3 INTERPRETATION OF ESTIMATES . . . 23
5.4 LABOREARNINGS AS THEDEPENDENT VARIABLE . . . 23
5.5 PROBABILITY OF BEING IN EMPLOYMENT . . . 24
6 RESULTS 25 6.1 FAMILY AND GENDER GAP . . . 25
6.2 YEARLY EARNINGS . . . 31
6.3 PROBABILITY OF BEING IN EMPLOYMENT . . . 36
6.4 DISCUSSION . . . 37
7 CONCLUSION 39
1 I NTRODUCTION
For the past few decades, there has been a growing body of research examining the wage differential between men and women. Decomposing the gender wage gap into an explained and an unexplained part, the explained part has slowly converged over time and what remains could to some extent be attributed to the gender difference in effects of children [Kleven et al., 2018].
Sweden has one of the most generous parental leave insurances in the world, and was the first country to allow fathers to receive benefit on the same terms as mothers when caring for their children [SCB, 2018]. Despite the possibility to share family responsibilities equally, the outtake of parental leave days is heavily skewed with women still accounting for the vast majority of the parental leave [Försäkringskassan, 2018]. The difference in outtake varies with income and educational level of the parents and the sector in which they work.
There is also a lot of research examining the effects of having children and taking parental leave. These effects differ between men and women, with women experiencing a much more negative effect than men. This negative effect is referred to as the “family penalty”, which is the effect of children on labor market outcomes that cannot be explained statistically
[Staff and Mortimer, 2012]. Because of the longstanding wage differences between men and women, one approach to examine the effect on wage due to children is to compare women with children to women without children. The differences in wage between the two groups, when accounting for observable characteristics, is defined as “the family gap” – the difference in wage that occurs between comparable individuals when one of the two has at least one child and the other one has no children [Waldfogel, 1998]. The family gap has received growing research interest in the past few years, with Kleven et al. [Kleven et al., 2018] accounting for the most recent research. In their paper “Children and Gender Inequality: Evidence from Denmark” they are investigating the family gap in Denmark using an event study approach, and they argue that although the gender wage gap has decreased in the past few decades, the family gap remains constant.
The aim of this thesis is to investigate how the family gap has developed over time in Sweden and how it can be related to the general gender wage gap during the same time period.
The data used for the analysis is the Swedish level of living surveys (SLLS) conducted in six waves. SLLS is unique in that they have collected hourly wage data for individuals in all waves, which is normally not possible to access. In the most similar study conducted by Kleven et al in
Denmark, a measure of wage is constructed by using yearly earnings divided by hours worked, which they themselves point out likely biases the results somewhat.
The research question of this paper is therefore “how has the family gap evolved in Sweden over time and how does this development relate to the general gender wage gap?”. The research question is of relevance in a policy perspective as the wage gap between men and women is still a challenge in the Swedish labor market. An increased understanding of the reasons for its development over time with regards to having children is of importance in order to reduce this difference in the future. Because Sweden has a particularly well-developed parental leave and child benefit system, the results may very well differ from other countries, either in a more positive or more negative direction.
There are two main contributions of this study. Firstly, there has, to the best of my knowl- edge, not been any similar historical analysis of the development of the family gap in Sweden.
Secondly, the study uses gross hourly wages as the dependent variable in the analysis. Data on hourly wages is often difficult to retrieve for earlier years, which is the reason for why most historical analyses depend on monthly or yearly labor income instead. Using hourly wage as the dependent variable makes it possible to separate the effect of children on wages from the effect of children on hours worked. For this reason, I believe that this study will fill a gap in the existing literature.
The outline of the paper is organized as followed: Section 2 provides background and previ- ous literature in the research area; section 3 describes the theoretical framework. Furthermore, the data used for the analysis is presented in section 4, along with some variable descriptions;
section 5 covers the method used to retrieve the results, which are presented in section 6. Lastly, section 7 provides a brief summary and conclusion of the results.
2 B ACKGROUND AND L ITERATURE R EVIEW
This section starts by a presentation of research on the general wage differences between men and women. Then, statistics and research regarding the gender differences in factors relevant to labor market outcomes are discussed. Lastly, previous literature on the family gap and its implications are reviewed.
2.1 T
HEG
ENDERW
AGEG
APUnderstanding the wage differentials between men and women has been an important research agenda for decades. The difference can be decomposed into two parts – explained and un- explained. Much of the variation between men and women can be explained by observed differences in educational and career choices, differences in the amount of unpaid work and family responsibilities [Blau and Kahn, 2017]. The explained gap has somewhat decreased over time, but the pace has slowed down and there is still a persistent unexplained gap that remains [Kleven et al., 2018]. This has been explained in different ways; statistical discrimi- nation towards women, differences in preferences and psychological attributes being some of the most common examples from literature [Bertrand, 2011]. The gender wage gap opens up mainly in ages 30-40 (SCB, 2013), when many individuals have young children, with men and women following similar trends before becoming parents. [Kleven et al., 2018]
2.2 D
IFFERENCES IN THEU
SE OFP
ARENTALL
EAVE ANDF
AMILYR
E-
SPONSIBILITY
One reason for why studying Sweden is particularly interesting is because of its generous poli- cies on parental leave benefits and public childcare. The parental leave insurance was first introduced in 1974, replacing the previous maternal leave insurance [SCB, 2018]. The new insurance made it possible for men to receive social benefits on the same terms as women when taking leave from work to care for their children. This encouraged men to be more involved in their children and acknowledged the child’s right to both parents. Although the insurance was changed to induce economic incentives for couples to share family responsibilities more equally, women accounted for 99,5 percent of the outtake in 1974. Women have since grad- ually decreased their share of parental leave and men have increased theirs, but the averages
between the groups still differ significantly, with women being responsible for a vast majority of the parental leave still. [Försäkringskassan, 2018]
Since being introduced in 1974, the scheme of the insurance has underwent a number of changes; the maximum length of the parental leave benefit has been prolonged gradually through several reforms and is presently set at 480 days, with 90 days being reserved for each parent since 2002. (Ibid.) The parental leave benefit can be divided into two levels of replace- ment. For 90 of the total 480 days, the replacement is set at a low, fixed level whereas benefit for the remaining 390 days is calculated based on previous earnings. There is, however, a cap for maximum level of benefits. (Ibid.)
Women on average take more time away from the labor market when becoming parents, thus also taking on a larger share of unpaid work in the home [Försäkringskassan, 2018]. The difference in parental leave may impact long-run patterns within families, with women tak- ing temporary parental leave more than men to care for sick children or choosing to work part-time. Differences in outtake between men and women and the design of the parental leave insurance causes women to fall behind men when they have children. Angelov et al.
[Angelov et al., 2016] show that there are small gender differences in wage prior to having children, but significantly larger differences 15 years after the birth of the first child. The great- est differences occur in couples where the woman would have had a lower income and wage development than the man even without children.
2.3 E
FFECTS OFB
EING OUT OF THEL
ABORF
ORCEHaving children has a negative effect on women’s wages, but not men’s, and the penalty remains after controlling for the length of the leave and other characteristics [Staff and Mortimer, 2012].
The wage penalty could be because of a loss of human capital due to the time spent away from the labor market but if so, men and women would be equally affected by taking the same amount of parental leave, which is not the case. Another explanation of the penalty could be statistical discrimination, with employers promoting men more than women in terms of wages and advancement [Benard and Correll, 2010]. Since mothers more than fathers take the initial parental leave associated with having a child, this also affects the amount of unpaid work the parents perform at home and the pattern of which parents divide temporary parental leave when the child is sick later [Forssell, 2002, Försäkringskassan, 2013]. Knowing this, it may be rational for the employer to discriminate against women, knowing that they on average
will spend less time at work than men. In contrast to the punishment that women experience after becoming mothers, fathers seem to rather be rewarded for having children; with wages increasing after having a child [Benard and Correll, 2010]. This can be compared to the wage premium of married compared to single men, perhaps suggesting that being married and having children signal responsibility [Chun and Lee, 2001].
However, men and women taking out the same amount of parental leave are not equally affected. Assuming that men and women, conditioning on a set of characteristics, have the same potential outcomes, being out of the labor force should have the same impact on all individuals independent on gender. Furthermore, the loss of human capital should be the same for a given time period independent on what the time out is devoted to, i.e. taking parental leave should not cause a different loss in human capital than taking a leave for other non-educational or work-related reasons. Albrecht et al. [Albrecht et al., 1999] studied this in Sweden in the nineties, by comparing the effect of different types of career interruptions on wages to estimate the specific effect of parental leave. The results suggested that women’s labor market outcomes were seemingly unaffected by taking parental leave, whereas men experienced a negative effect.
This either suggests that men taking parental leave sends a negative signal to employers whereas it’s expected of women, or that women are already being statistically discriminated against in terms of expectations of parental leave, so that they have been penalized already before taking the leave.
2.4 T
HEF
AMILYG
AP– E
STIMATING THEW
AGEI
MPACT OFC
HILDREN One of the first and most eminent researchers on the family gap is Jane Waldfogel [Waldfogel, 1998, Waldfogel, 1997]. In a paper from 1997, she examines the evolution of the family gap in the US using the National Longitudinal Survey of Young Women between 1968 and 1988. While high- lighting that the gender wage gap has decreased over time as a result of higher labor force par- ticipation and education among women, Waldfogel discuss the persistent tendency of women having less labor market experience than comparable men. Likewise, the wage differential be- tween women with and without children – the family gap – appears to be rather constant over time. In order to account for the difference in labor market experience, she uses a measure of“potential labor market experience” in her estimates, as well as controlling for both parental and marital status. She finds that even after controlling for actual labor market experience and other characteristics, women with children earn lower wages than women without children.
Additionally, after performing some heterogeneity tests, Waldfogel suggests that the wage dif- ferential cannot be attributed to differences in unobserved characteristics between women with and without children. The results ultimately imply that women suffer a family penalty when having children that persists when controlling for labor market choices and personal attributes.
Building on the previous paper, Waldfogel continued studying the matter in mainly the US and Britain. When the first few studies were published just before the millennium, she argued that although the gender wage gap had decreased in the United States in the previous decades, the family gap had actually widened. One reason for this may, according to Waldfogel, be that several policies have been implemented to encourage and legalize equal pay, but fewer on maternity and childcare. The direction in which maternity leave affects wage is according to most research ambiguous. On the one hand, maternity leave enable new mothers who would otherwise have remained out of the labor force to come back to their previous job without a decrease in wage. On the other hand, some mothers likely stay at home longer than they otherwise would have. However, previous studies suggest that the main impact of maternity leave is that more women choose to return to the labor force after childbirth. In her paper
“Understanding the “Family Gap” in Pay for Women with Children” [Waldfogel, 1998], she compares the US to some Scandinavian countries including Sweden and Denmark. She argues that as opposed to the US, which at the time had a family gap of about 10-15%, neither Denmark nor Sweden has any noteworthy family gap. What is interesting in this context is that Kleven et al. [Kleven et al., 2018], has published a study showing that Denmark now has a family gap.
Whereas the gender gap was lower in Denmark than in the US before the millennium, as shown by Waldfogel, the levels are now much more similar.
Kleven et al. use an event study approach based on the birth of the first child. The identi- fication strategy relies on the assumption that although the decision of having children is not random, the timing of the first child is. Their results show that men and women follow the same trends in labor market outcomes prior to childbirth. However, directly following the birth of the first child, women fall behind men and the trend for women never recovers. In addition to studying the impact of children on wages, Kleven et al. also investigate the incentives among women to switch to “family friendly” jobs after forming a family. Family friendly jobs are described as firms with more flexible characteristics and a larger share of women employed.
Self selection into more family friendly work environments have been discussed as a possible determinant of the gender wage gap in previous studies [Goldin, 2014], but this is the first study
to show how the segregation is associated to having children. Ultimately, the results of Kleven et al. [Kleven et al., 2018] imply that the wage differential caused by having children, i.e. the family gap, has evolved over time in Denmark and is presently responsible for the remainder of the unexplained gender wage gap.
Another recent study on the family gap is one by Lundborg et al. [Lundborg et al., 2017].
In their paper “Can Women Have Children and a Career”, they use an IV approach with IVF treatments. The identification strategy is based on the assumption that the chance of having a successful IVF treatment is random and not correlated with previous labor market outcomes.
Individuals that have all received IVF treatments, but where only some has succeeded, are then compared in terms of labor market outcomes. Thus, it is possible to interpret the difference in labor market outcomes between parents and non-parents as the causal effect of having children.
Lundborg et al. find that having children has large negative impacts on women’s labor market outcomes, both in the short and the long term. The results suggest that women choosing to work in lower paid jobs when they become mothers can explain a large portion of the wage decrease.
They also suggest that although women receiving IVF treatments likely differ to other women, the results are likely generalizable to all women.
Lastly, Angelov et al. [Angelov et al., 2016] investigate the income and wage trajectories for Swedish couples before and after parenthood in their paper “Parenthood and the Gender Gap in Pay”. The approach is based on comparing the within-couple gaps in wage before and after having children. They find that the effects on wages are substantially negative and long lasting for women. 15 years after the birth of the first child, the wage gap between men and women have increased by 32 and 10 percentage points. They provide some evidence of explanations for this in terms of comparative advantages of working at home and in the labor market.
3 T HEORETICAL FRAMEWORK
This section provides a theoretical framework for the following analysis. First, a short overview of labor supply theory is presented. Then an introduction to human capital theory follows, and lastly signaling theory and self-selection theory are presented.
3.1 L
ABOR SUPPLY THEORYSimplified, the labor market consists of firms and workers; firms accounting for the demand and workers accounting for the supply of labor. On a market with free competition, wages are determined by setting supply equal to demand [Borjas, 2016]. In the textbook example, wage represents marginal productivity of the worker and since there is variation in marginal productivity, there is variation in wages. Wages and labor market outcomes are thus determined by several factors affecting the skills and marginal productivity of workers. These consist of observable factors as educational level and working experience, and unobservable factors as effort. [Mankiw and Taylor, 2017]
Workers account for the supply of labor in the economy, offering their time and produc- tivity to firms at a cost. The labor supply varies between demographic groups and over time [Borjas, 2016]. The traditional framework for analyzing labor supply is called the “neoclassi- cal model of labor-leisure choice”, in which any worker has a fixed number of hours a day to allocate between labor and leisure. The allocation depends on the individual’s preferences, as taking time for leisure has an alternative cost in terms of the wage that individual could have earned if the time was instead spent on labor. There is thus a trade-off between labor and leisure (ibid.).
Gary Becker [Becker, 1965] formalized a model for explaining household’s choice of al- location of time. In his unitary model, households are both the consumers and workers and they gain utility from consuming and means of consuming, though disutility, from working.
This model has later been criticized in that it assumed that couples have joint preferences and that there is no “work” in the household. In terms of labor supply, the work performed in the household is then grouped with consumption of labor, although the labor choice may be that of working in the labor force or in the household [Johnson, 2010]. Becker suggested however that as one person’s consumption depends on the others time spent working in the joint model, it is rational that women perform more of the work in the household and men in the labor market.
This argument is based on women’s potential earnings being lower than the potential earnings of men traditionally.
The labor supply of women has gradually increased in the past few decades, but is still below that of men in most countries. Much of this increase has been explained as a result of changing attitudes towards women working, as well as increased wages for women and thus higher incentives to work [Gardiner, 1997, Bertrand, 2011]. As the potential earnings of women increase, Becker’s argument of the rational choice in households no longer holds, but the division of house and labor market work is still determined by gender to a large extent [Gardiner, 1997].
In recent work, Blau and Kahn [Blau and Kahn, 2013] examine how the labor supply of women is affected by the presence of more family-friendly policies. They build on the exist- ing theory by suggesting that these policies can affect women’s labor supply in two opposite directions: on the one hand by making it possible for women to combine work and family, they facilitate labor market entry for women who would have otherwise remained at home. On the other hand, long, paid parental leave likely causes even career-oriented women to stay at home longer than they otherwise would have. Likewise, women with a strong labor market commit- ment may have incentives to adjust their labor supply to part-time and possible positions of lower levels.
3.2 H
UMAN CAPITAL THEORYGary Becker [Becker, 1962] formalized the first modern human capital theory. Human capital refers to the stock of skills an individual has, and so human capital investments refer primarily to investments made in an individual’s skills and productivity, but it can also include invest- ments in physical and emotional health. These are factors that increase individual productivity and subsequently wage. There are different types of human capital investments, where the main investments are usually seen as education and on-the-job training. On the job training can be either of a general character or firm specific. General training increases the individual’s productivity regardless of workplace, whereas specific training increases productivity on the current job, but cannot be transferred to a different job or sector. There are also other types of human capital investments such as acquiring knowledge of the labor market and economic system and increasing emotional and physical health. Previous research has found that higher levels of human capital are positively related to earnings. Likewise, skills are negatively corre-
lated with unemployment. When trying to measure a causal effect of human capital on wages, there are two main problems. Firstly, human capital can normally not be directly observed, so instead some proxy is used as a measure of human capital, normally years of schooling and labor market experience. Secondly, there is a potential ability bias. It could be argued that this should not cause any trouble, as human capital is the stock of skills and knowledge of an individual. However, there is a positive correlation between being more able and the amount of education and training, which may attribute an excessive amount of credit to investments in human capital.
Mincer and Ofek [Mincer and Ofek, 1982] developed Becker’s model further. By inves- tigating how wages respond to career interruptions, they showed that human capital must be maintained or it depreciates. In their paper from 1982, they find that wage rates are lower at reentry for individuals that have left the labor market and increases with the length of the in- terruption. These results supported the theory of human capital depreciation, as a break from the labor force would cause a decrease in the human capital stock because of a loss of specific on-the-job investments, but it would be constant regardless of the length of the interruption, unless human capital depreciated over time. When “leavers” reenter the labor market, they thus enter with lower wages than when they left. However, Mincer and Ofek found that there was a rapid initial wage growth for the leavers once they came back. Because human capital is less costly to reconstruct once lost than to construct new human capital, they explain this as leavers having decreased their human capital stock, they are less productive and earn lower wages.
The alternative cost for investing in human capital is thus lower for leavers than for those who stayed, and so when reentering the labor market they make large investments in their human capital.
According to this theory, being out of the labor force does not only keep human capital from increasing, but even reduces it – human capital is an investment that must be maintained continuously or it depreciates [Mincer and Ofek, 1982]. This implies that taking leave to care for children would cause a decrease in human capital for the individual taking leave, thus implying that a decrease in earnings following the leave could be a consequence of the decrease in productivity. However, if this is the case, then men and women would be affected equally by the leave and the effect of the leave should be the same as the effect of any other leave, which is not the case [Albrecht et al., 1999].
3.3 S
IGNALING THEORY AND STATISTICAL DISCRIMINATIONAs opposed to the human capital theory, which assumes that investing in schooling increases actual productivity, signaling theory suggests that schooling does not increase productivity per se. Signaling theory instead proposes that education signals productivity to employers and that it is that signal that is reflected in offered wages. [Spence, 1973]
The signaling theory builds on the assumption that the costs of investing in education, sig- naling costs, are negatively correlated with productivity capability. If this assumption does not hold, then all individuals would invest the same amount so that they are not distinguishable in their signal. [Spence, 1973]
Signals are thus characteristics that employers use as an implication for true productivity. In this context, the signaling of being a parent rather than that of education is of interest. This can be either positive or negative and likely differs between men and women. Gender should not be a signal in terms of productivity per se, since the same amount of education should signal the same amount of productivity, regardless of sex [Spence, 1973]. There are however other aspects of the signal of gender, which have been brought up earlier in the literature review.
Given that women on average, conditional on education and previous labor market outcomes, tend to take on more family responsibility and care for children once becoming a parent, simply being a woman sends a signal of a lower average labor market commitment of the group. As the employer cannot distinguish which individuals will be more committed, being a woman may cause employers to statistically discriminate and invest more in men. [Bielby and Baron, 1986]
Previous research indicates that whereas women with children are perceived as less compe- tent and engaged in work, men with children are considered to be loyal and reliable.
[Benard and Correll, 2010]
3.4 S
ELF-
SELECTION THEORYHuman capital assumes that different amounts of human capital stock affect wages, and as- sumes that human capital can be seen as homogeneous. Self-selection theory criticizes this homogeneity assumption that all the variation in wages can be explained by different amounts of human capital. Self-selection theory instead suggests that there are different kinds of human capital, with different pay-offs in terms of wage [Polachek, 1981]. The self-selection theory introduces occupational choice and its importance for wages. Since occupational patterns dif-
fer with demographic groups, the main interest in this context is occupational sorting based on gender. The basic assumption of the model assumes that the goal for all individuals is to maximize lifetime earnings. For different occupations, there are different levels of atrophy.
Atrophy is defined as “the loss in potential lifetime earnings from labor force intermittency”. It is thus the wage decrease caused by skills not being continuously used. The smaller the loss in wage because of an interruption, the lower is the atrophy [Polachek, 1981]. In terms of gender differences, the model assumes that men and women have the same characteristics on average and differ only in expected lifetime labor force participation. It follows that gender differences in occupation can be attributed to the differences in lifetime labor force participation. Because atrophy differs between jobs, it is rational for individuals to choose an occupation that maxi- mizes lifetime earnings given the expected dropout rates from the labor force. Since women have lower labor force participation than men, it may be rational for men and women to choose different occupational paths, and for women to choose occupations with lower atrophy and thus lower wages [Polachek, 1981]. Traditionally, women tend to take more parental leave – both directly following childbirth and later temporary to care for sick children. Knowing that they will likely have career interruptions, it is rational for women to invest in education and choose occupations for which the wage decrease following an interruption is low, even if the initial wage is lower than for other occupations. A related but different possible explanation of occu- pational sex segregation could be because of differing preferences between men and women.
Assuming that some women and men are family-oriented and therefore willing to forego higher earnings in exchange for employment that provide better work-life balance, this would suggest that there are self-selection effects [Gash, 2009]. The self-selection theory thus provides some alternative explanation of the family penalty in wages. Family friendly employment is charac- terized by jobs with a more flexible working environment and, in more recent research, a higher share of women in the workplace. [Gash, 2009]
4 D ATA
This section presents the data used to conduct the analysis. It also contains some variable descriptions in addition to the descriptive statistics.
4.1 D
ATA SOURCEThis paper uses the Swedish level-of-living surveys (SLLS) collected by the Swedish institute for social research at Stockholm University. The survey has been carried out in six different waves, the first one in 1968 and the most recent one in 2010.
The original aim of SLLS, beginning with the first wave in 1968, was to survey a random sample of approximately one per mille of the Swedish population in ages 15-74. In the latest wave however, the sample consisted of about 6000 individuals in ages 18-75, with a response rate of 72%.
Throughout the years, the survey has developed. After the three first waves, further di- mensions were introduced in the 1991 survey. Family events, information on each individuals partner and the composition of labor market interruptions were all included in the question- naire. For the purpose of this study, all available waves are used in order to examine how the family and gender gaps have developed over time. The additional variables for the later waves regarding children and parental leave are of particular interest for this study. Unfortunately, no suitable proxy is available for similar estimation for the earlier waves, and for this reason parental leave outtake is not included in the analysis.
There are however great advantages of using the SLLS. Because many of the dimensions in the survey are the same over time, this presents a rare opportunity of comparability over time.
Likewise, because of the extensive time period for which SLLS has been carried out, it provides historical data for completely comparable variables. Furthermore, SLLS contain hourly wage data for all survey years, which is unusual for data from the time of the first few surveys years.
This is also one of the main advantages of the SLLS data, as previous research has suggested that the use of monthly or yearly labor earnings may bias the results somewhat.
The data is conducted on mainly different population samples at each point in time, thus constituting repeated cross-sectional data on multiple dimensions. About a thousand individu- als have however been surveyed in each wave, thus compiling panel data for these individuals.
Although panel data would be preferable in order to answer the research question, the panel
data sample is not large enough for the purpose of the thesis, so the approach uses the repeated cross sections for all individuals instead. This also means that there is no possibility of estimat- ing causal effects in terms of the research question using the available data. Therefore, the aim of this thesis is to give a descriptive analysis of the associations between cause and effect.
4.2 V
ARIABLE DESCRIPTIONTable 1 shows the variable descriptions of the variables of interest in the following analysis.
Table 1: Variable Description
Variable Description
Gross hourly wage Gross hourly wage, collected for each wave. The wage is expressed in SEK and is based on several wage variables in the SLLS data, including bonuses and an average of weekly hours worked. All wage variables adjusted with CPI, using 2010 as the index year and thereafter logged.
Yearly labor earnings The variable is constructed by using the gross hourly wage and then multiplying it by hours worked in the previous year.
Employed Dummy variable encoded as 1 if the individual had a positive number of hours worked in the year before the survey and 0 otherwise.
Married Dummy variable with levels [ Married, Not married ]. Generated by using the variable
"Civil state according to the interview" and with "Married" as the reference. For the survey years of 2000 and 2010, couples living together but not being married are also indicated as married.
Age Age in years by the survey year. Calculated by using survey year and year of birth.
The sample has been limited to only include observations aged between 25-65 years at the time of the survey.
Education Total number of years of education. Observations with more than 20 years of education has been removed from the sample.
Experience Total number of years employed in the labor force. Individuals with more than 50 years of experience have been removed from the sample.
Children Dummy variable with levels [0,1], where 1 indicates having children and 0 indicates not having children. The variable was generated by first adding the number of children living at home and the number of children not living at home.
Table 2 shows the descriptive statistics for the observations with a positive gross hourly
wage, i.e. the subsample used for the main analysis. Because the dependent variable used in the main analysis is gross hourly wage, all observations with no value in that variable has been removed from the analysis. This is mainly because the Swedish Level of Living Survey (SLLS) cover ages 18-74, which indicates that removing observations lacking wage data likely decreases the age span of the analysis somewhat further than the initial restriction of 25-65 years old at the time of the survey. This is also confirmed in the descriptive statistics, where the mean age is higher than for the full samples.
The descriptive statistics imply that the women in the sample have on average lower educa- tion and experience in the first few waves, but catches up to men in more recent year, and even exceed men in terms of education for the last three waves. This is in line with previous studies.
Tables 3 and 4 give the same descriptive statistics of means but grouped on individuals with and without children.
Table 2: Sample Means: All Individuals
1968 1974 1981 1991 2000 2010
Men Women Men Women Men Women Men Women Men Women Men Women
Log wage 4.53 4.24 4.71 4.46 4.67 4.50 4.71 4.55 4.88 4.76 5.10 5.0
(0.342) (0.396) (0.279) (0.294) (0.240) (0.258) (0.235) (0.216) (0.227) (0.214) (0.302) (0.237)
Age 42.7 43.2 41.7 42.0 41.8 41.4 41.6 42.2 42.2 43.4 43.9 45.2
(11.2) (10.8) (11.4) (11.2) (11.1) (10.7) (10.8) (10.6) (11.0) (10.5) (11.1) (10.9)
Experience 25.9 17.7 24.1 17.0 23.4 17.4 22.1 18.8 21.9 20.8 22.8 22.2
(12.1) (10.9) (12.6) (10.6) (12.6) (10.3) (12.1) (10.1) (12.4) (10.8) (12.3) (11.7)
Education 8.22 8.35 9.59 9.40 10.3 10.2 11.3 11.4 12.2 12.5 13.3 13.8
(2.57) (2.75) (3.34) (3.01) (3.31) (3.18) (3.08) (2.96) (2.91) (2.91) (2.67) (2.80)
Share with children 74.1% 72.6% 76.1% 79.6% 75.5% 81.2% 70.7% 80.6% 70.3% 78.8% 71.2% 81.2%
No. of observations 1376 879 1390 1035 1385 1347 1350 1438 1211 1288 1074 1062
Table 3: Sample Means: Individuals With Children
1968 1974 1981 1991 2000 2010
Men Women Men Women Men Women Men Women Men Women Men Women
Log wage 4.56 4.22 4.73 4.45 4.69 4.50 4.74 4.55 4.90 4.76 5.14 5.0
(0.297) (0.394) (0.280) (0.310) (0.238) (0.268) (0.233) (0.215) (0.227) (0.212) (0.264) (0.238)
Age 43.7 44.5 42.9 42.8 43.4 42.3 44.2 43.8 44.9 45.1 46.4 46.8
(10.8) (9.98) (11.0) (10.8) (10.6) (10.2) (9.93) (9.95) (10.1) (9.82) (10.1) (10.2)
Experience 27.2 17.0 25.5 16.4 25.4 17.5 25.0 19.7 24.8 22.2 25.4 23.8
(11.5) (10.3) (12.0) (9.90) (12.0) (9.72) (11.3) (9.53) (11.5) (10.1) (11.4) (11.1)
Education 8.10 8.08 9.44 9.18 10.0 10.0 11.1 11.2 12.0 12.3 13.2 13.6
(2.45) (2.67) (3.32) (2.96) (3.30) (3.17) (3.15) (2.90) (2.94) (2.85) (2.70) (2.75)
No. of observations 1019 638 1057 824 1045 1094 955 1158 851 1015 765 868
Table 4: Sample Means: Individuals Without Children
1968 1974 1981 1991 2000 2010
Men Women Men Women Men Women Men Women Men Women Men Women
Log wage 4.43 4.29 4.65 4.50 4.62 4.54 4.64 4.57 4.83 4.77 5.02 5.0
(0.434) (0.399) (0.268) (0.218) (0.239) (0.206) (0.226) (0.220) (0.219) (0.221) (0.368) (0.235)
Age 40.0 39.9 37.9 39.1 36.8 37.5 35.3 35.7 35.7 37.3 37.8 37.7
(11.9) (12.3) (11.9) (12.5) (11.0) (11.6) (10.1) (10.6) (10.3) (10.7) (11.2) (11.0)
Experience 22.3 19.6 19.6 19.0 17.4 17.4 15.2 15.1 14.8 15.7 16.2 15.3
(13.1) (12.2) (13.1) (12.6) (12.6) (12.4) (11.0) (11.4) (11.5) (11.6) (12.0) (11.7)
Education 8.55 9.05 10.1 10.2 11.1 11.1 11.8 12.6 12.6 13.5 13.6 14.9
(2.86) (2.82) (3.37) (3.06) (3.20) (3.08) (2.86) (2.94) (2.80) (2.90) (2.59) (2.79)
No. of observations 357 241 333 211 340 253 395 280 360 273 309 194
Lastly, table 5 shows the difference in means for men and women, with and without chil- dren. The values are calculated by subtracting the means of individuals without children from the means of individuals with children.
Table 5: Difference in Means for Men and Women, With and Without Children
1968 1974 1981 1991 2000 2010
Has Children: Men Women Men Women Men Women Men Women Men Women Men Women
∆ Log wage 0.13 −0.07 0.08 −0.05 0.07 −0.04 0.1 −0.02 0.07 −0.01 0.12 0
∆ Age 3.7 4.6 5.0 3.7 6.6 4.8 8.9 8.1 9.5 7.8 8.6 9.1
∆ Experience 4.9 −2.6 5.9 −2.6 8.0 0.1 9.8 4.6 10.0 6.5 9.2 8.5
∆ Education −0.45 −0.97 −0.66 −1.02 −1.1 −1.1 −0.7 −1.4 −0.6 −1.2 −0.4 −1.3
∆ Observations 662 397 724 613 705 841 560 878 491 742 456 674
A few notes are of interest here. Firstly, when only looking at sample means, men and women with children show the opposite difference in wages compared to men and women without children. Whereas men with children have higher mean wages than men without chil- dren, women with children have lower mean wages than women without children. The indi- viduals with children are also older than individuals without children. Secondly, for the four most recent survey years, both men and women with children have more experience than all individuals without children. Lastly, regarding the fact that women in more recent years exceed men in terms of education when looking at all individuals, i.e. with and without children, it is interesting to note that this difference is mainly driven by women without children.
The opposite directions of the wage differences for men and women with children could possible be a result of the "child premium" for men discussed in previous literature, as opposed to the child penalty for women. The difference in wages are not likely explained by differences in years of education, as women with children have almost the same or slighlty longer edcu- cation than men. There could however be differences in the type or level of education, which I do not have access to for the whole period. The difference in experience could however be a determinant, if the women’s experience is lower because they take on more responsibility for the family and household once they become mothers.
5 E MPIRICAL M ETHOD
This section is outlined as followed: first, there is a description of the method used to retrieve the results for the analysis. Second, a base model is specified for the family and gender gaps.
Third, an alternative measure of wage is introduced to the model as the dependent variable and lastly, a model for investigating differences in employment probability is presented.
5.1 O
RDINARYL
EASTS
QUARESIdeally, we would have had a large longitudinal dataset where fertility was randomly assigned to individuals. That way, we would have comparable individuals where children would cause the only difference. A second best option would have been to have access to the same type of data, but instead using an event study to see how wages respond to children shortly after having a child and a few years after. However, as this requires a large set of panel data, it is unfortunately not feasible for this paper. For this reason, the main method for the analysis is using a simple ordinary least squares.
Although the method in this paper may not be the best for the research question, I believe it is the most suitable one given the prerequisites. There are however other perks of using the Swedish Level of Living Surveys (SLLS). As previously discussed, the SLLS data contain vari- ables that have been the same for all waves, and are thus completely comparable. Furthermore, it has the very unique perk of containing hourly wage data for all samples. This is especially unusual, as many studies in the area do not have access to hourly wage data for earlier years, although it has been argued that it is the most reliable wage measure when analyzing the gender wage gap.
5.2 F
AMILY ANDG
ENDERG
AP INE
ACHW
AVEFor each of the six waves, a base specification is used, including only the variables of interest in order to get a picture of the uncontrolled family and gender wage gaps. The base specification is:
ln(GHWi,t) = β1W omani,t + β2W oman × Childreni,t + β3Childreni,t+ εi,t, (1)
where the dependent variable is logged gross hourly wages. To make the monetary difference comparable across the survey years, the wage variable has been adjusted to the consumer price index (CPI) with 2010 as the index year. The variable of interest is “Children”, which is a dummy variable that equals 1 if the individual has at least one child and 0 otherwise. It is also of interest to examine and identify whether the effect of children on wages change with the number of children. The initial idea was therefore to compare how different numbers of children as well as having children of different ages affect wages differently. Unluckely, in- cluding all of the population parameters decreased the sample size of each subset to greatly.
Therefore, only the binary variable indicating children or no children is included. Furthermore, the base specification includes an interaction term between being female and having children.
The interaction term estimates the additional effect of having children for women.
In a modification of the base specification, a wage equation in an adjusted form of the one proposed by Mincer will be used. The specification including all controls is:
ln(GHWi,t) = β0+ β1W omani,t+ β2W oman × Childreni,t+ β3Childreni,t
+ β4Educationi,t+ β5Experiencei,t+ β6Experience2i,t+ β7Agei,t + β8Age2i,t+ β9M arriedi,t+ εi,t
(2)
The additional variables in this specification are Education, Experience, Age and Marriage status. Education is total number of years in education and Experience is total number of years employed, thus years of experience. The experience variable is included both linearly and squared, since existing literature in the area suggests that wage increases non-linearly with experience. The specification also includes a dummy-variable for being married. The reason for including civil status is because previous research has shown that women’s wages tend to decrease when they get married, whereas an increase have been found for men. This increase in wages for married men has been referred to as a “marriage premium”. Again, the initial aim was to include different types of civil status such as single, divorced or widowed as well as married. However, in order for the sample population to remain of reasonable size, only an indicator of being married or not being married is included. However, for the waves in 2000 and 2010, being in a couple and living together will also be indicated as married in the variable for Married, as it has become a common alternative to marriage in more recent years. Lastly,
the variable Age is included both linearly and squared, as proposed by previous research in the area [Waldfogel, 1997].
The identification strategy of the method is selection on observables. By controlling for the observable characteristics, we make the assumption that all individuals are absolutely com- parable except for the fact that some of them have children and some do not. Thus having children is as if randomly allocated and the only differing effect on wage between individuals will be that of being a parent. This is however a rather strong assumption, as parenthood is for the most part a conscious choice. Individuals making the choice of becoming parents likely differ in unobserved characteristics from those not making the choice of parenthood. Although this problem remains when only comparing women with and without children as well, the as- sumption is that there are less unobserved characteristics differing between women in general than between women and men. Thus the chance of estimating the wage impact of children specifically likely increases when only comparing women that are parents with women who are not. However, this is also the main reason for why the aim of the analysis is to estimate the associationsbetween variables, rather than the causal effects.
Only variables that are available for all of the waves will be included in the specification.
Unfortunately, this means leaving out some of the most important variables in the later survey years, such as total number of months on parental leave and number of parental leave weeks per child.
5.3 I
NTERPRETATION OFE
STIMATESIt is of importance to note that because of the nature of the data and the research question at hand, no causal inference can be drawn from the results. Instead, coefficients of the parame- ters are interpreted as associations − or correlations − between the variables of interest. The potential causes for the effects are discussed further on but without the possibility to draw any conclusions about the causality.
5.4 L
ABORE
ARNINGS AS THED
EPENDENTV
ARIABLEWhen studying the gender wage gap as well as the family gap, it is of interest to examine how wages differ independent of labor supply. That is, to ensure we are looking at the wage gap and not the difference in labor supply, we need hourly wages. However, as this is not feasible in many cases when studying longer time periods or using old register data, many studies have
instead had to rely on monthly wage or yearly labor earnings. I argue that one of the main contributions of this paper is the fact that hourly wage data is available for all waves, i.e. a time period of about 40 years. In order to investigate the implication this might have on the results, I perform a similar analysis to the main one, but using yearly labor earnings instead of hourly wages. The variable is constructed by multiplying the wage variable used in the main analysis with hours worked last year. Although this is a rough measure that likely has some measurement error in the dependent variable, it should provide some evidence of the differing results obtained by different measures of wage. Since the same independent variables are used as in the main analysis, only the potential measurement error in the dependent variable should have a potential of posing a problem. If that is the case, the estimates will simply suffer from attenuation bias, i.e. they will be less precise yet unbiased.
5.5 P
ROBABILITY OFB
EING INE
MPLOYMENTIn addition to the main analysis using logged gross hourly wages as the dependent variable as well as an alternative wage measure, a supplementary test of probability of being employed will also be analyzed. The dependent variable used for these regressions is constructed as a binary variable taking the value of one if the individual has a positive amount of working hours in the year before the survey, and zero otherwise. For the purpose of this part of the analysis, a linear probability model will be used in order to estimate the likelihood of being in the labor force, given the presented control variables.
When estimating the probability for a binary dependent variable, it could be argued that a logistic probability model (logit) should be used instead of a linear probability model (LPM).
This is because the linear probability model allows for probabilities below 0 and above 1, which is obviously not possible theoretically. A second issue with the LPM is that probabilities are linear in the independent variables. However, the estimates of LPM and logit are often similar, and because the independent variables of interest in this case are binary as well, the issue of probabilities being linear in the independent variables does not pose a problem in this context.
Furthermore, the LPM has more intuitive interpretations, and so for the purpose of this analysis it poses a more suitable option.
Because the labor supply affects the wages in terms of labor market outcomes such as advancement, it is of interest to examine if it differs between groups and if so, what it is that determines the difference.
6 R ESULTS
This section starts by a presentation of the results obtained from the analysis of the evolution of the family and gender gaps, using two different model specifications. Furthermore, results on earnings are presented and lastly on the probability of being employed.
6.1 F
AMILY ANDG
ENDERG
APThe results from the base specification with no controls are shown in table 6. This specification includes only gender, having children and an interaction between being female and having children, as these are the variables of interest. The interaction term between being a woman and having children is included in the model in order to estimate both the gender wage gap and the family gap in the same regression. The gender wage gap is thus the coefficient for gender, and the family gap is the difference between the gender coefficient and the two coefficients indicating the effects of children. The effect of being a woman, that is the estimate of the gender wage gap, is negative for all waves but with estimates decreasing in magnitude. This implies that when not including controls in the model, the estimated gender wage gap has decreased over the time period studied. The estimate for 2010 is not significant, but previous estimates suggest that the gender wage gap is emerging when not including controls. The coefficient for the variable indicating having children is positive but decreasing between the first and the last wave. Similarly, the estimate for the additional effect of being female with children is decreasing between the first and last survey year, but with some fluctuations across years.
Table 6: Specification With No Controls
Dependent variable:
log(as.numeric(GRHW))
1968 1974 1981 1991 2000 2010
Woman −0.139∗∗∗ −0.150∗∗∗ −0.085∗∗∗ −0.073∗∗∗ −0.057∗∗∗ −0.028
(0.030) (0.025) (0.021) (0.017) (0.018) (0.025)
Woman × Children −0.202∗∗∗ −0.132∗∗∗ −0.106∗∗∗ −0.118∗∗∗ −0.090∗∗∗ −0.111∗∗∗
(0.035) (0.028) (0.023) (0.020) (0.020) (0.028)
Children 0.126∗∗∗ 0.086∗∗∗ 0.065∗∗∗ 0.097∗∗∗ 0.073∗∗∗ 0.114∗∗∗
(0.022) (0.018) (0.015) (0.013) (0.014) (0.018)
Constant 4.434∗∗∗ 4.648∗∗∗ 4.622∗∗∗ 4.641∗∗∗ 4.831∗∗∗ 5.024∗∗∗
(0.019) (0.016) (0.013) (0.011) (0.012) (0.015)
Observations 2,255 2,425 2,732 2,788 2,499 2,136
R2 0.144 0.169 0.109 0.127 0.082 0.055
Adjusted R2 0.143 0.168 0.108 0.126 0.081 0.053
Residual Std. Error 0.361 (df = 2251) 0.284 (df = 2421) 0.248 (df = 2728) 0.224 (df = 2784) 0.219 (df = 2495) 0.269 (df = 2132) F Statistic 126.393∗∗∗(df = 3; 2251)164.413∗∗∗(df = 3; 2421)110.968∗∗∗(df = 3; 2728)134.485∗∗∗(df = 3; 2784)74.679∗∗∗(df = 3; 2495)41.152∗∗∗(df = 3; 2132)
Note: ∗p<0.1;∗∗p<0.05;∗∗∗p<0.01
26
When instead using estimates from the regressions including controls that can be found in table 7, figure 1 show the gender wage gap in terms of estimated gross hourly wage when in- cluding the effects of children, and controlling for observable covariates. Men are the reference group, plotted as the intercept line at 0.
Figure 1: Gender Wage Gap Accounting for Effects of Children
-0.25 -0.20 -0.15 -0.10 -0.05 0.00
1968 1974 1981 1991 2000 2010
Survey Year
EstimateofHourlyWage(log)
Women
Error Bars Represent 95 % Standard Errors For Estimates
Estimated Wage Difference
The figure shows an overall decrease of the gender wage gap, with women still falling below men but less and less so. This is in line with the findings of previous studies. However, the difference between men and women is larger in magnitude compared to official statistics as well as previous studies; the magnitude is similar to what is usually found for the explained gender wage gap. This could possibly be explained by the fact that when accounting for the effects of children, the wage estimate for men becomes higher and the wage estimate for women becomes lower. This can be seen more clearly in figure 2, where the family gaps for men as well as women are graphed.
Figure 2: Family Wage Gaps
-0.2 -0.1 0.0 0.1
1970 1980 1990 2000 2010
Survey Year
EstimateofHourlyWage(log)
Men with Children Women with Children Women without Children
Error Bars Represent 95 % Standard Errors For Estimates
Estimated Wage Difference
What is of interest in this graph is the family gap for men and women respectively, i.e. the difference between men with and without children as well as the difference between women with and without children. Men with children are plotted as the top, two-dashed line and men without children are plotted as the intercept line. The effect of children for men is thus seemingly positive, with fathers having higher wage estimates than non-fathers for all the years studied. The difference is however decreasing; the positive effect of children on the wages of men is about half the magnitude in 2010 as compared to 1968.
The trend in the family gap for women shows a rather different pattern. Women encounter the opposite effect of having children from that of men, namely a negative one. This is in line with previous research and the notion that men with children benefit from a "child premium" as opposed to the "child penalty" faced by women. Furthermore, whereas the family gap for men has decreased steadily over time, the family gap for women has remained on a level of between 3-5%. Although the difference may not seem severe, it is of great economic significance if women with children have constantly lower wages than women without children.
It is also worth noting that a number of reforms were implemented in the 70’s in order to increase the possibility of combining family and work, yet there is no significant drop in the family gap for the subsequent years.
The full results of the main analysis are found in table 7. The regressions include controls for experience, education, age and marriage status. These findings, graphed above in figures 1 and 2 show that the gender wage gap (excluding the effect of children) has consistently decreased since the first observed wave in 1968 and until the last observed wave in 2010.
When taking all control variables into account, the remaining gender wage gap is estimated to be about 16% in 1968, and "only" about 8% in 2010. The estimates are significantly different from zero and effects are of both statistic and economic significance. The additional effect of children for women is negative for all survey years but overall decreasing. However, the full effect for women with children is retrieved by adding all of the estimates for which [Woman=1]
and [Children=1], and so the decreasing positive effect of children for men and women with children has to be taken into account as well. Still, the complete wage difference for women with children compared to men without children has decreased from about −21% in 1968 till about −11% in 2010. However, the wage difference for women without children compared to men without children has decreased from about −17% in 1968 till about −8% in 2010, implying no significant differences in the family gap − it seemingly persists.
