Policy Research Working Paper 10509 Female-Worker Representation Effect Gender Pay Variation in the Kuwaiti Civil Service Mohamed Ihsan Ajwad Simon Bilo Ha Nguyen Ebtesam AlAnsari Lama AlHumaidan Faleh AlRashidi Social Protection and Jobs Global Practice & Middle East and North Africa Region June 2023 Policy Research Working Paper 10509 Abstract Kuwaiti women working in Kuwait’s civil service earn, on with a higher ratio of women to men tend to have lower average, 18 percent less than Kuwaiti men. Using a unique wages for both genders when compared to workers in occu- data set of all Kuwaiti nationals working in Kuwait’s civil pations with a lower ratio of women to men. This finding service, this paper analyzes the relationship between wages, is especially true for women. Workplaces with a higher gender, and the relative dominance of women in occupa- female-to-male ratio exhibit lower male wages but slightly tions and workplaces. The main finding is that an important higher female wages than workplaces with lower female- portion of the association between gender and wages is to-male workplace ratios. The paper calls this latter novel explained not by human capital but by occupational and finding the female-worker representation effect. workplace segregation of men and women. Occupations This paper is a product of the Social Protection and Jobs Global Practice and the Middle East and North Africa Region. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank. org/prwp. The authors may be contacted at majwad@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Female-Worker Representation Effect: Gender Pay Variation in the Kuwaiti Civil Service Mohamed Ihsan Ajwad (World Bank, Washington, DC, USA) Simon Bilo (World Bank, Washington, DC, USA) Ha Nguyen (International Monetary Fund, Washington, DC, USA) Ebtesam AlAnsari (Public Authority for Applied Education and Training, Kuwait City, Kuwait) Lama AlHumaidan (Public Authority for Applied Education and Training, Kuwait City, Kuwait) Faleh AlRashidi (Public Authority for Applied Education and Training, Kuwait City, Kuwait)* Keywords: gender wage gap, civil servants’ wages, Kuwait, MENA JEL Codes: J15, J24, J31, J42, J45, J61, J71 * The authors may be contacted at majwad@worldbank.org, sbilo@worldbank.org, hnguyen7@imf.org, im.alansari@paaet.edu.kw, lt.alhumaidan@paaet.edu.kw, and fm.alrashidi@paaet.edu.kw. This paper, its analysis, and its conclusions solely represent the views of the authors and not necessarily their respective organizations. The authors would like to acknowledge the financial support received from the Gender and Labor Markets Research Programs under the aegis of the Middle East and North Africa Chief Economist’s Office (MNACE), as well as comments received from MNACE colleagues. The authors are grateful to Dr. Khaled Mahdi, Issam A. Abousleiman, Ghassan N. Alkhoja, and Anush Bezhanyan for their support. In addition, the authors are grateful to S Anukriti, Roberta V. Gatti, Daniel Lederman, Ismail Radwan, Fatma Ahmad Al Ibrahim, and Nelly Elmallakh for helpful comments and valuable suggestions. Any remaining errors are the responsibility of the authors. Female-Worker Representation Effect: Gender Pay Variation in the Kuwaiti Civil Service 1. Introduction Earnings differences between men and women are observed in almost every country. For example, in the US, the women/men earnings ratio was roughly 60 percent for many years, though it began to rise sharply in the 1980s and “by 2014, women full-time workers earned about 79 percent of what men did on an annual basis” (Blau and Kahn, 2017). A vast literature documents the gender pay difference around the world. For example, Altonji and Blank (1999) report that women’s wages have increased relative to men. Defined as the difference between median earnings of men and women relative to median earnings of men, the OECD finds that the gender wage gap is as low as 1.2 percent in Belgium and 2.6 percent in Bulgaria, and as high as 31.1 percent in the Republic of Korea and 24.3 percent in Israel (OECD 2023). Blau and Kahn (2017), using Panel Study of Income Dynamics data, discuss the stylized facts regarding the gender wage difference in the United States. While human capital variables used to explain a significant share of the gender wage difference, more recent estimates show that they do not. An increasingly important determinant of the gender wage difference is the occupation and industry in which an individual works. Importantly, men and women often work in different occupations and sectors, with corresponding variation in gender pay differences (Blau and Kahn 2017; Kunze 2018). The effect of occupational gender segregation on gender pay differences is demonstrated by Kidd and Goninon (2000), Hegewish et al. (2010), Grönlund and Magnusson (2013), and Cortes and Pan (2018). Works using employer–employee matched data sets such as Groshen (1991), Reilly and Wirjanto (1999), Bayard et al. (2003), Card et al. (2016), Cardoso et al. (2016), and Casarico and Lattanzio (forthcoming) show the importance of workplace gender segregation for explaining gender wage differences. It is difficult to empirically determine whether lower representation of women in higher-paying occupations is due to discrimination, subtle barriers women face when seeking work, or selection (specific job characteristics that lead women not to pursue such jobs). Recent research points to multiple mechanisms leading women toward lower- paying occupations, including higher risk aversion (Croson and Gneezy 2009; Eckel and Grossman 2008), worse performance in highly competitive environments (Bertrand 2011; Gneezy et al. 2003), lower confidence (Niederle and Vesterlund 2007), lower salary expectations (Bertrand 2017; Bowles, Babcock, and McGinn 2005; Babcock and Laschever 2006; Roussille 2022), higher involvement in family care (Kunze 2018; Kleven et al. 2019a; Kleven et al. 2019b; Kleven et al. 2021), and men’s desire to maintain their occupational group’s social status under asymmetric information (Goldin 2002). 2 In this paper we explore whether men’s and women’s pay relates to female- worker representation, defined as the share of women in an occupation or workplace. Our analysis explores a unique cross-sectional administrative data set containing anonymized individual-level information on all 341,522 civil service employees in Kuwait employed in 2019. In addition to a standard analysis of gender wage differences, we identify functional relationships between wages and the proportion of women in an occupation or workplace. We find, as others before us, that human capital variables do not explain the gender wage difference (Hosni and Al-Qudsi 1988; Shah and Al-Qudsi 1990; and AlAnsari 2018). Instead, we find that occupational and workplace segregation play an important explanatory role in explaining wages in Kuwait (like many other countries). The richness of our data, which identifies people’s occupations and workplaces, allows for novel scrutiny of the possible underlying mechanisms behind labor market segregation and gender wage differences. Men and women tend to make less on average in female- dominated occupations than male dominated occupations. However, women’s wages are even lower in such occupations. We also find that a higher proportion of women in a workplace is associated with higher female wages and lower male wages 1 than workplaces with a lower proportion of women. Others find that in the private sector, a higher proportion of women is associated with lower wages in general (Reilly and Wirjanto 1999; Bayard et al. 2003). We call this relationship the female-worker representation effect. Our work has the most in common with Groshen (1991) and Bayard et al. (2003), which use private sector employer–employee matched data sets from the US to show that occupational and establishment-level gender segregation accounts for a sizable fraction of the gender wage difference. Bayard et al., who use more comprehensive data than Groshen, find that higher proportions of women in occupations and workplaces are associated with larger gender wage gaps. Unlike them, we analyze male and female wages of civil service workers in the Kuwaiti public sector. In addition, rather than focusing on the wage gap, we focus on the relationship between gender representation and men’s and women’s wages. Our findings are thus more nuanced, allowing us to analyze the relationship between each gender’s wage and the proportion of women in occupations and workplaces. Other related research examines the effects of gender concentration and gender diversity on other outcomes, such as job satisfaction. Bender et al. (2005) use US data and find that women report higher job satisfaction in female-dominated workplaces because such workplaces offer more flexibility (flexible hours, managers who are more accommodating, more liberal sick- and family-leave policies). Similarly, Qian and Fan 1 Our finding is consistent with Whaley et al. (2020), who find that female doctors get paid less in practices with an overwhelming proportion of male doctors. 3 (2019) show that women report more unpleasantness and less meaningfulness at work when working in male-dominated workplaces. Women also report facing pervasive stereotypes, such as that of the caring mother (Sarathchandra et al. 2018) or office housekeeper (Berdahl et al. 2018). Finally, women report facing a lack of mentoring and career-development opportunities (Campuzano 2019). In contrast, Clark et al. (2021) show that women perceive job quality to be lower when coworkers are mainly women, while men report the highest perceived job quality when the gender mix is equal. They argue that women might not report lower job quality in male-dominated workplaces, as these jobs may signal their economic and social advancement (Moore 2018). Meanwhile, men working in female-dominated occupations may suffer from negative stereotyping and therefore report lower job evaluations (Lupton 2000; Torre 2018). The remainder of the paper is organized as follows. Section 2 provides background information on the Kuwaiti labor market and highlights some key opportunities and constraints faced by women in the labor market. Section 3 describes the data used in the empirical analysis. Section 4 discusses the results, and Section 5 concludes. 2. Background AlAnsari (2018: 37–45) summarizes key developments of Kuwaiti women’s status in the labor market. Female labor force participation has increased substantially, from 1.8 percent in 1965 to 39.3 percent in 2015, with an increasing share employed in executive positions such as managers and senior government officials. Most Kuwaiti women, like Kuwaiti men, are employed in the public sector. Kuwaiti women face many formal and informal limitations when entering the labor market. Formal limitations include legal restrictions on participation in occupations “that are hard, dangerous, or harmful to health, and from occupations that utilize their femininity, violate their manners or are not in line with public morals” (AlAnsari 2018: 42). Informal limitations for women in the labor market include sociocultural pressures, such as the expectation that women “have the approval of a male relative (husband, father or brother) to work [and] choose a female-appropriate job which is gender-segregated, safe, and has flexible and short working hours” (AlAnsari 2018: 44). Employment opportunities are also limited because women cannot participate in informal male social gatherings. Such gatherings, called diwaniya, are important forums for networking and decision-making, including those related to employment and promotions (Redman 2014: 113; AlAnsari 2018: 45, 74). A female Kuwaiti business owner illustrates the importance of exclusion from male social networks: “I could never penetrate a male network because I am a woman. And a lot of what happens, happens in the diwaniya. A lot of the discussions—you don’t have access to them. You’re really excluded” (quoted in Garrison 2015: 157). 4 In Kuwait, most employed nationals—about 83 percent—work for the government (Ajwad et al. 2022). In turn, most public sector employees are hired by the Civil Service Commission, which matches Kuwait’s ministries and other government organizations with new employees. The concentration of nationals in the government is important for our analysis, as our data set covers all civil service employees in 2019 and thereby a significant majority of nationals. The Civil Service Commission hires most government employees, but the government hires through other channels for positions in the security forces and state-owned enterprises. Although we have not come across a systematic analysis of the hiring process at the Civil Service Commission, evidence suggests significant inefficiencies, such as long waiting periods and sizable job-skills mismatch (Garrison 2015: 147–53). The general compensation criteria of people employed through the Civil Service Commission are public and well known. AlAnsari (2018: 65–66) and AlHumaidan (2019: 22) divide the criteria into two sets of components. The first set applies to Kuwaiti nationals irrespective of gender and concerns years of service, education level, and specialization. The second set is gender dependent and designed to increase the average male wage. This set includes a social allowance, which is paid to all nationals but is generally higher for married men, and a child allowance, predominantly also paid to men. The differential treatment by gender follows Personal Status Law no. 51/1984, according to which “the male in the family is responsible for providing the living expenses for the family” (AlAnsari 2018: 65–66). We show below that these gender-specific factors explain much of the gender difference in compensation. The rules for retirement are also gender dependent. In January 2020, women could retire with a full pension after 25 years of service if they were 50 or older. Men could do so after 30 years of service at the age of 55. While the retirement rules have become stricter over time, the differential treatment of the genders has been a constant (Ajwad et al. 2022; Bilo et al. 2021). 3. Data We analyze 2019 payroll data from the Civil Service Commission. This is a unique and rich data set. It comprises a cross section of individual-level anonymized information on civil service employees, including age, years of service, marital status, monthly wages, workplace, and type of employment. There are 341,522 observations in the data set, which we narrow down to 262,771 observations after excluding non-nationals, duplicate records, and people with implausible age values. We exclude approximately 70,000 observations of non-nationals in the civil service because Kuwait has a dual labor market in which compensation follows very different rules for nationals and non-nationals (AlAnsari 2018: 41–45), requiring an analysis beyond the scope of this 5 paper.2 Our data cover about 64 percent of all nationals working in Kuwait and 76 percent of nationals working in the public sector, as shown by Ajwad et al. (2022) and Bilo et al. (2021: 15). To our knowledge, only Alqattan, Stergioulas, and Al-Zayer (2012) use a comparably large and detailed data set when estimating returns to schooling in Kuwait. Five variables in the data are critical for our analysis—wages, occupation type, occupation, public sector organization, and workplace. Wages are monthly wages in Kuwaiti dinars (1 Kuwait dinar was on average USD 3.29 during 2019). These exclude child benefits, which are almost always paid to men for their families, because they are not labor compensation and therefore do not represent wages in the traditional sense of the term. 3 Occupation type has 15 usable categories – a broad occupational classification of the dataset. 4 In contrast, the occupation variable has 6,322 categories and is the detailed occupational classification. Public sector organization, or organization, classifies all workers into one of 36 governmental organizations. 5 Workplace is a more detailed variable classifying each worker into one of 11,166 workplaces. Table 1 presents summary statistics of the data set. It shows that women represent most Kuwaitis among civil service employees. This is mostly because our data do not include the security forces, which are male dominated. The average woman employed in the civil service earns 1,181 Kuwaiti dinars (KD) per month, about 81 percent of the wages of the average man in the civil service, who earns KD 1,457. When including child benefits, the average monthly wage of a man is KD 1,571, while that of women is KD 1,184. Table 1 also explores possible explanations for the difference: male civil servants tend to be older and have more work experience; and are less likely to have a college education but have more master’s degrees and PhDs. We confirm below that education does not explain gender wage differences. 6 The remaining row of Table 1 shows the potential importance of the social allowance. As three-quarters of men are married, they receive a higher marital social allowance. We confirm below that the allowance is important for explaining the observed gender wage difference. 2 However, we use the number of non-Kuwaitis when computing our two derived variables, female occupation ratio and female workplace ratio. 3 It would have been, for the same reason, useful to net the wages of marital social allowance. Unfortunately, we do not have enough information to make the adjustment and the wage variable includes this allowance. We do control for marital status in our analysis, however, to capture the effect of the differences in social benefits between married men and everyone else. 4 Appendix 1, Table A1, gives the full list of occupation types. 5 Appendix 1, Table A2, gives the full list of public sector organizations. 6 See Appendix 2 for the list of education levels. 6 Table 1: Averages for key variables among Kuwaiti civil servants Men Women Difference in the mean (men minus women) Number of observations 97,926 164,845 Monthly wages (KD) 1,457 1,181 276*** Age (years) 37.15 36.22 0.92*** Years of service in the civil service 12.10 10.34 1.77*** Bachelor’s degree only 36% 49% −13*** (pctg. points) PhD or master’s degree 6.66% 2.97% 3.69*** (pctg. points) Married (%) 75% 68% 6.83*** (pctg. points) Notes: The currency unit is the Kuwaiti dinar (KD), and monthly wages exclude child benefits of KD 50 per child for up to seven children. The benefits are predominantly paid as part of male wages on behalf of their families. * p < 0.05, ** p < 0.01, *** p < 0.001 Source: Kuwait Civil Service Commission (2019) and authors’ calculations Table 2 presents the distributions of monthly wages of male and female civil service employees. Depending on the wage percentile, women earn between 79 and 90 percent of what men earn. When defined as the difference between median earnings of men and women relative to median earnings of men, Kuwait’s gender wage gap is 14.1 percent, which is above the OECD aggregate gap of 11.9 percent (OECD 2023). Table 2: Monthly wages (KD) of Kuwaiti civil servants by percentiles Men Women Women/Men Mean 1,457 1,181 81% 10th percentile 778 698 90% 25th percentile 972 825 85% Median 1,235 1,061 86% 75th percentile 1,679 1,371 82% 90th percentile 2,339 1,842 79% Notes: The currency unit is the Kuwaiti dinar (KD), and monthly wages exclude child benefits of KD 50 per child for up to seven children. The benefits are predominantly paid as part of male wages on behalf of their families. Source: Kuwait Civil Service Commission (2019) and authors’ calculations As noted, Blau and Kahn (2017) show that occupation- and industry-related differences between men and women play an important role in explaining the gender wage difference. Our data highlight the importance of occupational and sectoral differences in the public sector. In our study, occupation types, such as engineering jobs, administrative support jobs, or medical and health services jobs, allow us to compare occupational differences. 7 Similarly, different public sector organizations, such as the Ministry of Education and Ministry of Health, represent different sectors. Figure 2 and 7 Appendix 1 gives the full list of organizations and occupation types in the data. 7 Figure 1 show that women tend to work in occupations and organizations paying lower average monthly wages. 8 Our analysis below confirms this is an important part of the explanation of the overall wage variation. Figure 1: Average monthly wage against the proportion of employed women by various occupation types Notes: The currency unit is the Kuwaiti dinar (KD), and monthly wages exclude child benefits of KD 50 per child for up to seven children. The benefits are predominantly paid as part of male wages on behalf of their families. See Appendix 1 Table A1 for the list of occupation types. Four occupations are excluded. In three cases, one of the genders had fewer than 25 observations. In the fourth, the occupation had an ambiguous title. The regression line is a simple OLS regression for the two plotted variables. The line is not necessarily the best possible fit for the data; it only illustrates one possible relationship that we exploit rigorously in the analysis. The average compensation is based on both male and female Kuwaiti employees. Source: Kuwait Civil Service Commission (2019) and authors’ calculations 8 The linear regression lines are not necessarily the best fit for the data. It is possible that a nonparametric or a quadratic function would fit better. Our goal now is, however, only to illustrate a possible relationship that we exploit rigorously in the analysis. 8 Figure 2: Average monthly wage against the proportion of employed women at various public sector organizations Notes: The currency unit is the Kuwaiti dinar (KD), and monthly wages exclude child benefits of KD 50 per child for up to seven children. The benefits are predominantly paid as part of male wages on behalf of their families. See Appendix Table A2 for the list of organizations. The regression line is a simple OLS regression for the two plotted variables. The line is not necessarily the best possible fit for the data; it only illustrates a possible relationship that we exploit rigorously in the analysis. The average compensation is based on both male and female Kuwaiti employees. Source: Kuwait Civil Service Commission (2019) and authors’ calculations The figures above suggest that our lists of organizations and occupation types are analytically meaningful. But they are also somewhat limited. Fortunately, our data set also offers more comprehensive variables representing specific occupations and workplaces that we translated to English from Arabic using Google Translate API. While the categorizations prove to have high explanatory power, Table 3 suggests the need for caution when using the two variables because many categories represent very few employees. For example, the table indicates that at least 50 percent of classified occupations have only one observation. We show below that robustness checks—in which we drop occupations and workplaces representing fewer than 10 employees—do not substantively change the results. 9 Table 3: Number of employees (Kuwaiti and non-Kuwaiti) in various occupations and workplaces by percentiles # of employees in occupations # of employees in workplaces 10th percentile 1 1 25th percentile 1 2 50th percentile 1 6 75th percentile 6 18 90th percentile 67 80 Source: Kuwait Civil Service Commission (2019) and authors’ calculations To better understand the occupation and workplace variables, we list the five most populous categories for each in Table 4 and Table 5, respectively. The two detailed variables on occupations and workplaces allow us to pinpoint people who are coworkers and those working the same types of jobs. The detailed information then allows us to construct variables capturing the extent to which one’s occupation and workplace is male- or female-dominated and explore how this relates to the gender wage difference. We define the two derived variables as follows: = + = + The female occupation ratio represents the percentage of women in a given occupation j relative to the total number of workers in that same occupation. The female workplace ratio in workplace w represents the number of women in a workplace divided by the total number of workers in that same workplace. By design, the two variables have a range in the closed interval [0,1], and the underlying numbers of men and women include Kuwaiti and non-Kuwaiti workers. 9 9 Since the goal is to identify whether an individual works in a female-dominated occupation (or a female- dominated workplace), it seems reasonable to compute corresponding variables for all men and women working in the occupation (or workplace). Below we use analogous variables that include only Kuwaitis as part of our robustness checks. 10 Table 4: Five occupations representing the most employees (Kuwaiti + non-Kuwaiti) Occupation (translated) Number of employees Nurse 10,804 Assistant transaction coordinator 9,506 Transaction executor/officer 6,900 Secretary 6,633 Senior nurse 5,348 Notes: Occupation titles were computer translated from Arabic and are displayed in English after editing. Also note that teachers are not among the five occupations. This is because teachers are not classified as a single group but as several smaller occupation categories, depending, for example, on their subject of instruction. Source: Kuwait Civil Service Commission (2019) and authors’ calculations Table 5: Five workplaces with most employees (Kuwaiti + non-Kuwaiti) Workplace (translated) Number of employees Islamic Studies Department 6,056 Al-Adan Hospital / Al-Ahmadi Health District 4,018 Al-Farwaniyah Hospital / Al-Farwaniyah Health District 3,970 Mubarak Al-Kabeer Hospital / Hawalli Health District 3,557 Jahra Hospital / Jahra Health District 3,537 Notes: Workplace titles were computer translated from Arabic and are displayed in English without editing. Source: Kuwait Civil Service Commission (2019) and authors’ calculations Table 6 provides examples of occupations that are common at various levels of female occupation ratio. To get a sense of perspective, it is useful to note that the median female occupation ratio is 36 percent for men and 76 percent for women. Table 7 then gives examples of workplaces for various levels of the ratio. The median female workplace ratio is 18 percent for men and 91 percent for women. Table 6: Examples of common occupations at different levels of female occupation ratio Female occupation ratio, value range Examples of common occupations 0.15-0.25 Senior accountant Adviser 0.45-0.55 Legal Researcher Civil Junior Engineer 11 0.75-0.85 Teacher (C) Islamic education Bookkeeper Table 7: Examples of common workplaces at different levels of female workplace ratio Female workplace ratio, value Examples of common workplaces range 0.15-0.25 Management of treated water Street lighting management 0.45-0.55 General Secretariat of the Supreme Council for Planning and Development Faculty of Business Studies 0.75-0.85 Al-Jahra Health Zone Al-Farawaniyah health area 4. Results This section presents three sets of results: (1) the size of the average gender wage difference and the extent to which it can be attributed to observed variables; (2) the effect of female dominance in an occupation and at a workplace on wages; (3) robustness checks. 4.1 Attributing the gender wage difference to occupation, workplace, and marital status When estimating the average monthly wages of men and women, we use wages after removing child benefits. We use a version of the Mincerian human capital model (Mincer 1974) to attribute the wage to various variables. We estimate the following: = + + + � + � + � + � + � ℎ ℎ ℎ + � + Wi represents the monthly wages person i was making in 2019; femalei indicates whether the person is female; Hi is a vector of human capital covariates, including age, age squared, years of service, years of service squared, and a set of dummy variables indicating the person’s highest achieved education level; Mid is a dummy variable 12 representing marital status; 10 Oif is a dummy for an occupation type as listed in Table A1 of Appendix 1; Gie is a dummy for a public sector organization; Jih is a dummy representing one of the 6,322 detailed occupation categories; Pim is a dummy for a workplace. β is the main coefficient of interest and it indicates the extent to which we can attribute a person’s wage to their gender rather than other included independent variables. Table 8 shows this attribution progressively: we include additional independent variables until we reach the full specification of our model. Model (1) shows that Kuwaiti women earn approximately 18 percent less on average than Kuwaiti men. This is the unconditional wage gap and is similar in magnitude to the gender wage gap in the US for all workers (Blau and Kahn 2017). Model (2) controls for human capital, proxied by level of education, years of service, and age. Human capital has high explanatory power (measured by adjusted R-squared) for wage. However, β, the coefficient of the gender dummy variable, changes very little between models (1) and (2), increasing from 18 to 19 percent. The implication is that including human capital variables as wage determinants has very little impact on the gender coefficient. Model (3) shows that marital status plays an important role. Married men earn about 8 percent more than unmarried men (the baseline). Married women, however, get paid 9 percent less than unmarried men. This result is expected given the additional benefits married men receive as part of their compensation. After controlling for marital status, the gender coefficient drops to 12 percent from 19 percent. Models (4) and (5) additionally control for occupation type and public sector organization. The gender coefficient implies that once the controls are in place, unmarried female Kuwaitis earn 4.4 percent less than their male counterparts. Models (6) and (7) include detailed controls for occupations and workplaces. Looking at the full specification in model (7), the estimated impact of gender on wages of unmarried people drop to 0.5 percent. The above results show that what one does and where one works are key to explaining the gender wage difference among unmarried Kuwaitis working in the civil service. When all observables are controlled for, the portion of wages we can attribute to gender almost disappears for unmarried Kuwaitis. This finding is confirmed in model (8), which is a robustness check that excludes observations in which people work in occupations and workplaces with fewer than 10 people. 10 The categories include widow/widower, divorced, married, unknown, and unmarried, where unmarried is the base category. 13 Table 8: Explaining variation in wages among Kuwaiti civil servants Log (monthly wages) Independent (1) (2) (3) (4) (5) (6) (7) (8) variables -0.18*** -0.19*** -0.12*** -0.072*** -0.044*** -0.019*** -0.0051*** -0.0050*** Female (0.0017) (0.0011) (0.0024) (0.0020) (0.0018) (0.0012) (0.0012) (0.0012) Years of service 0.024*** 0.023*** 0.021*** 0.022*** 0.015*** 0.015*** 0.015*** Years of service -0.000065 - -0.000075*** -0.000068*** -0.00011*** -0.00013*** -0.00012*** squared *** 0.00012*** 0.0027*** Age -0.0023*** -0.0048*** -0.0042*** 0.0063*** 0.0025*** 0.0031*** Age squared 0.000085*** 0.00011*** 0.000013 -0.000011 0.0000026 0.0000019 -0.0000042 Divorced -0.00096 0.0025 0.0041 0.011*** 0.0098*** 0.011*** Married 0.079*** 0.067*** 0.071*** 0.075*** 0.070*** 0.070*** Unknown -0.022*** -0.040*** -0.032*** -0.016*** -0.014*** -0.016*** Marital Status 0.058*** Widow/Widower 0.035 0.045 0.064* 0.065*** 0.056*** Female * -0.012*** -0.012*** -0.012* -0.015*** -0.013*** -0.012*** Divorced Female * -0.073*** -0.073*** -0.092*** -0.079*** -0.075*** -0.077*** Married Female * 0.012*** 0.014*** Unknown 0.0055 0.018*** 0.031*** 0.014*** Martial Status -0.073*** -0.072*** Female * Widow -0.060 -0.076* -0.091** -0.081*** 7.32*** 7.23*** Constant 7.18*** 6.44*** 6.43*** 6.46*** 7.02*** 7.55*** Highest achieved No Yes Yes Yes Yes Yes Yes Yes education level Occupation type No No No Yes Yes Yes Yes Yes Employing No No No No Yes Yes Yes Yes organization Occupation No No No No No Yes Yes Yes Workplace No No No No No No Yes Yes Observations 262771 262771 262771 262771 262771 262771 257405 235429 R-squared adj 0.046 0.70 0.70 0.79 0.83 0.93 0.94 0.95 Notes: Dependent variable is the natural log of monthly wages. Unmarried is the base category for the marital-status set of dummies. Model (7) has fewer observations because we removed singletons. Standard errors are included for the estimates for Female and are available upon request for the remaining variables. Model (8) is identical to model (7) but excludes observations in which people work in occupations or workplaces with fewer than 10 people. Government benefits, sector, and occupation are the key variables. * p < 0.05, ** p < 0.01, *** p < 0.001 (using robust standard errors) The findings in Table 8 are consistent with multiple hypotheses. First, selection into occupations and workplaces might reflect hiring managers’ preference for gender discrimination (cf. Goldin and Rouse 2000). It is plausible that the gender wage difference translates into informal and formal differential treatment in hiring decisions. Second, along the lines of Montgomery (1991) the selection might reflect the gender bias of informal networks, which provide direct access only to men. In addition to women’s higher risk aversion and underperformance in competitive environments as reported in the literature, restricted access to informal networks might decrease women’s chances of working in higher-paying occupations and workplaces. Third, women might self-select 14 into lower-paying occupations and workplaces because they offer nonpecuniary benefits, such as a more flexible work schedule, not reflected in the data (cf. Goldin 2014). In other words, the choice of occupation, workplace and even education level reflects constraints on and sometimes preferences of females, which in turn could be shaped by social norms. Available data do not allow us to reject any of these hypotheses, but most likely each of them explains a portion of the observed gender wage difference. 4.2 Female-worker representation effect: Gender in an occupation and at a workplace This section sheds light on the mechanism through which occupations and workplaces channel the measured association between gender and wages among Kuwaitis working in the civil service. One mechanism concerns women’s greater likelihood of working in occupations and workplaces that pay less. Other things held constant, one would expect that female-dominated occupations and workplaces are associated with lower incomes for both men and women. To test this proposition, we reformulate our regression model as the following specification: = + + + � + � + � + � + + + + + The model uses the previously defined female occupation ratio (FORi) and female workplace ratio (FWRi), and the coefficients of interest are σ, τ, π, and ρ. Of particular interest are coefficients τ and ρ, which indicate gender differences. It is possible that women choose lower-paying occupations and workplaces. If so, we want to know whether men with the same occupation and in the same workplace suffer the same pay reduction. 15 Table 9: Explaining variation in wages among Kuwaiti civil servants Log (monthly wages) Independent variables (5) (9) (10) (11) (12) −0.044*** 0.041*** −0.0076* −0.013*** −0.0062 Female (0.0018) (0.0032) (0.0036) (0.0038) (0.0037) Years of service 0.022*** 0.022*** 0.022*** 0.022*** 0.022*** Years of service −0.00011*** −0.00012*** −0.00012*** −0.00011*** −0.00012*** squared Age 0.0063*** 0.0074*** 0.0076*** 0.0093*** 0.0077*** Age squared −0.000011 −0.000025** −0.000026** −0.000051*** −0.000026** Female occupation −0.079*** −0.058*** −0.086*** ratio Female * Female −0.10*** −0.12*** −0.095*** occupation ratio Female workplace −0.12*** −0.10*** ratio Female * Female 0.14*** 0.13*** workplace ratio Kuwait female −0.045*** occupation ratio Female * Kuwait female occupation −0.12*** ratio Kuwait female −0.12*** workplace ratio Female * Kuwait female workplace 0.14*** ratio Constant 7.02*** 7.02*** 7.05*** 7.03*** 7.05*** Highest achieved Yes Yes Yes Yes Yes education level Occupation type Yes Yes Yes Yes Yes Employing Yes Yes Yes Yes Yes organization Marital status Yes Yes Yes Yes Yes Observations 262771 262771 262771 235462 262771 R-squared adj 0.83 0.83 0.83 0.84 0.83 Notes: Dependent variable in regressions is the natural log of monthly wages. Unmarried is the base category for the marital-status set of dummies. Standard errors are included for the estimates for Female and are available upon request for the remaining variables. Model (11) is identical to model (10) but excludes observations in which people work in occupations and workplaces with fewer than 10 people. * p < 0.05, ** p < 0.01, *** p < 0.001 (using robust standard errors) Column (10) of Table 9 presents the results of the fully specified model, while column (5) is for reference and mirrors column (5) in Table 8. The first finding is that both men and women tend to make less on average in more female-dominated occupations (occupations with a higher female occupation ratio). But women suffer three times more than men for a given female occupation ratio. The two statistically significant coefficients, −0.058 and −0.12, in column (10) imply that an occupation composed of almost all women pays men 5.8 percent less and women 17.8 percent less than an occupation composed of nearly no women. Note that we already control for education, occupation type, 16 organization, and marital status.11 Because the regression contains these other controls, we conclude that the gender proportions matter in addition to selection into occupation types. The second finding is that, holding all else equal, women who work in more female-dominated workplaces earn more than women who work in less female- dominated workplaces; and men who work in more female-dominated workplaces earn less than men who work in less female-dominated workplaces. The statistically significant coefficient −0.12 in column (10) implies that a man moving from a workplace with close to no women to a workplace with almost all women, holding everything else constant, could expect a wage reduction of about 12 percent. But given the coefficient of 0.14 for the female interaction, a woman moving from a workplace with close to no women to a workplace with only women could expect a wage increase by 2 percent. The above finding is consistent with the possibility of discrimination or gendered networks, as women in male-dominated workplaces might receive fewer opportunities for career growth. This is a point widely discussed in the literature (for example, Elliott and Smith 2004), but quantitative evidence on earnings is scarce. That men’s wages are more sensitive to gender composition at the workplace than women’s wages are, is a point not examined in depth before. Again, besides potential discrimination, other explanations are possible. For example, being in a female-dominated workplace might be a cause or a consequence of the informal networks that are important for a man’s career success. Last, columns (11) and (12) represent robustness checks. Model (11) is identical to model (10) but excludes observations in which people work in occupations and workplaces with fewer than 10 people. Column (12) tests whether the results change substantially with different measures of the two ratios. It therefore replaces FOR and FWR and focuses only on the ratio of Kuwaiti women to Kuwaiti nationals in an occupation or workplace as indicated by the two formulas below. Columns (11) and (12) both confirm the direction and relative importance of the effects found in column (10). , = , + , 11 Note that the regression in column (10) does not control for occupation fixed effects and workplace fixed effects. Also, as explained before, there is a difference between occupation type and occupation. Occupation type is a broader category, there are 15 of them (see Appendix Table A1). Occupation is a very finely defined category in our dataset, there are thousands of different occupations. 17 , = , + , We now look at the predictive margins to understand better the economic significance of the ratios FOR and FWR. Predictive margins, also known in this case as adjusted predictions at representative values, estimate average outcome for a specific predictor or combination of predictors, while holding other predictors constant at their observed values. 12 Figure 3 presents predictive margins of logs of monthly wages for the two genders and various female occupation ratios and female workplace ratios. It is important to remember that the datasets underlying the figures are partly fictional, allowing us to explore the predicted difference between male and female salaries holding everything else constant. Panel (a) shows that for both men and women, the predicted log of monthly wages falls as the female occupation ratio increases. That is, as female occupation ratios rise, monthly wages are likely to fall for both men and women. Interestingly, predicted logs of monthly wages are higher for women than men when the ratio is very low, but once it exceeds approximately 25 percent, the predicted logs of wages for men exceed those for women. In contrast, panel (b) in Figure 3 points to a large negative relationship between female workplace ratio and male wages; and a small positive relationship between female workplace ratio and female wages. As noted, we call this finding the female-worker representation effect. Men’s predicted wages exceed those of women for most values of the female workplace ratio. However, when female workplace ratios exceed 90 percent, women are likely to earn more than men. 12 For example, to see the relationship between gender and various values of female occupation ratio, we first assume all observations to be male, even when they are, in fact, female. In the second step, we change everyone’s female occupation ratio to a specific value, say 5 percent, keeping all the other observed variables intact. Thirdly, we use model (10) to find everyone’s predicted log of net wages and compute the average of the predictions for each observation. And finally, we repeat the process for various values of female occupation ratio to see the variation in the predicted log of wages. To compare the two genders, one needs to repeat the process for women. 18 Figure 3: Predictive margins of the natural log of montly wages by gender for (a) female occupation ratio and (b) female workplace ratio (a) Female occupation ratio (b) Female workplace ratio Notes: Figures are based on regression results of model (10). 95% confidence intervals are included. 19 Figure 4 illustrates the predictive margins analysis in terms of respective wages. Panel (a) shows the average predicted effect of a male Kuwaiti civil servant shifting from a job with a low female occupation ratio into a job with a high female occupation ratio. The predicted monthly wage loss when moving from a job with a 10 percent FOR to a job with a 90 percent FOR is about KD 56. Figure 4’s panel (b) shows that the predicted loss of wages is more significant for women. The corresponding predicted wage loss for a female Kuwaiti civil servant is about KD 164. Panel (c) in Figure 4 shows the average predicted drop in wages when a male Kuwaiti civil servant switches from a job with a low female workplace ratio to one with a high ratio. His predicted monthly wage loss is about KD 115 when switching from a job with a FWR of 10 percent to a job with a FWR of 90 percent. The respective job change in case of a woman then comes with a predicted increase in monthly wages of KD 20. 20 Figure 4: Predictive margins of select monthly wages by gender, female occupation ratio, and female workplace ratio (a) Male wages, female occupation ratio (b) Female wages, female occupation ratio (c) Male wages, female workplace ratio (d) Female wages, female workplace ratio Notes: Figures are based on regression results of model (10). Table 10 provides an indication of the gender-based nature of the relationship between wage variation and FOR and FWR. For example, a woman in a more male- dominated occupation and a more female-dominated workplace is expected to have a higher wage than her otherwise identical female counterfactual. The same cannot be said about men. When a man works in a more male-dominated occupation and also in a more female-dominated workplace, it is unclear whether his wage is higher or lower compared to the other man. 21 Table 10: Summarizing predictive margins for female occupation and workplace ratios by gender Female occupation ratio Increases Decreases Male wages: Decrease Male wages: Ambiguous Female Increases Female wages: Ambiguous Female wages: Increase workplace Male wages: Ambiguous Male wages: Increase ratio Decreases Female wages: Decrease Female wages: Ambiguous 4.3 Robustness checks We now look at subsets of the data to test the robustness of the findings. Table 11 shows that the results are consistent for subsets of observations when grouped by years of service. As before, a higher female occupation ratio comes with lower wages for both men and women, but especially women. A higher female workplace ratio correlates with lower male wages but slightly higher female wages. The effects are strongest for younger workers (i.e., those with less than 5 years of service). 22 Table 11: Explaining variation in wages among Kuwaiti civil servants, by years of service (10) (13) (14) (15) Independent variables All observations <5 years of service 5–15 years of service 15+ years of service −0.0076* −0.021*** −0.016** 0.032*** Female (0.0036) (0.0050) (0.0055) (0.0086) Years of service 0.022*** 0.040*** 0.021*** 0.044*** Years of service squared −0.00012*** −0.0030*** −0.00027*** −0.00053*** Age 0.0076*** 0.0068*** 0.0074*** −0.018*** Age squared −0.000026** −0.000028** −0.000041* 0.00023*** Female occupation ratio −0.058*** −0.076*** −0.048*** −0.043*** Female * Female −0.12*** −0.055*** −0.094*** −0.19*** occupation ratio Female workplace ratio −0.12*** −0.054*** −0.11*** −0.15*** Female * Female 0.14*** 0.063*** 0.11*** 0.17*** workplace ratio Constant 7.05*** 6.95*** 7.12*** 7.58*** Highest achieved Yes Yes Yes Yes education level Occupation type Yes Yes Yes Yes Employing organization Yes Yes Yes Yes Marital status Yes Yes Yes Yes Observations 262771 82549 103050 77172 R-squared adj 0.83 0.80 0.79 0.80 Notes: Regressions (13), (14), and (15) apply the model from column (10) to subsets of the data by the three indicated years-of-service intervals. Dependent variable in regressions is the natural log of monthly wages. Unmarried is the base category for the marital-status set of dummies. Standard errors are included for the estimates for Female and are available upon request for the remaining variables. * p < 0.05, ** p < 0.01, *** p < 0.001 (using robust standard errors) Table 12 breaks the analysis down by education subgroups, showing overall consistency with the analysis of the full data set. A higher female occupation ratio generally comes with lower wages, except for men with less than high school education. The wages, again, tend to be lower for women than men when the female occupation ratio is higher. As before, a higher female workplace ratio means lower wages for men, but not necessarily for women. The positive effect of female workplace ratio on female wages is strongest for women with college degrees (column 18). Women with graduate school education are the only exception, as the effect is not statistically significant. 23 Table 12: Explaining variation in wages among Kuwaiti civil servants, by highest achieved education level (10) (16) (17) (18) (19) Independent variables All observations Less than high High school and College and above Graduate school school above (