Labour Mobility and Returns to Education. Jiayuan Teng. A Thesis presented to The University of Guelph

Transcription

1 Labour Mobility and Returns to Education by Jiayuan Teng A Thesis presented to The University of Guelph In partial fulfilment of requirements for the degree of Doctor of Philosophy in Economics Guelph, Ontario, Canada c Jiayuan Teng, April, 2015

2 ABSTRACT LABOUR MOBILITY AND RETURNS TO EDUCATION Jiayuan Teng University of Guelph, 2015 Advisors: Dr. Miana Plesca, Dr. Louise Grogan This dissertation applies statistical methods to understand labour market issues in China and Canada. The first chapter uses an instrumental variable method to identify the causal effect of migrant networks on the probability of rural-urban labour migration in China. It uncovers a substantial heterogeneity in migrant network effects by gender, age groups, and between people with and without migration experience. Evidence shows that migrant networks affect migration decisions through increasing job tenure and improving work environments of migrants. The second and third chapters answer research questions related to gender wage gap and returns to postgraduate education in Canada. Using a broader set of occupational characteristics than previous studies, the second chapter adopts a quantile decomposition method to reveal that women with different educational levels experience the gender gap for different reasons. DOT-skills used in previous studies are important in explaining the gender gap for most workers in Canada, but not for high-school dropouts and for the top 10% of wage earners among the university-educated workers. For the latter, men working in more competitive jobs and taking more managerial responsibilities are the explanations underlying Canada s glass-ceiling phenomenon. By applying imputation techniques in a novel way, this chapter quantitatively demonstrates that correcting for

3 selection into work makes little difference in estimating the gender gap for individuals with post-secondary education. For individuals without post-secondary education, the use of observed characteristics is sufficient to capture the selection rule. The third chapter documents up-to-date evidence on the decline in returns to postgraduate education relative to four-year university degrees from 1995 to The return has declined in all major fields of study except engineering and computer science in which workers with postgraduate education have experienced a substantial gain over the same period of time. By focusing on the supply side of the labour market, this paper provides an explanation for the decline in returns to postgraduate education by exploring changes over time in the occupational composition of workers with postgraduate education. Keywords: Rural-urban migration, China, migrant networks, gender gap, quantile analysis, workplace competitiveness, DOT-skills, duncan index, returns to a Master s degree, returns to a Doctorate, STEM fields

4 iv ACKNOWLEDGEMENTS First and foremost I want to thank my advisor Miana Plesca for her support and consistent encouragement. She provided me with tremendous help in my research and also taught me how to be a successful economist. The joy and enthusiasm she has for her research was contagious and motivational for me, especially during tough times in the pursuit of my Ph.D. I am also very thankful to my advisor Louise Grogan who has provided extremely useful guidance in my dissertation. I appreciate the contribution of time, ideas, and funding from both Miana Plesca and Louise Grogan to make my Ph.D. a gratifying experience. I am truly grateful to Bram Cadsby who recommended me to the Master s program in Economics at the University of Guelph. As my best friend and mentor, Bram has been a great support to my professional and personal life since we met in My achievement of the Ph.D. degree in Economics would not be possible without him. I also want to thank Fei Song for her continued support during my graduate studies at Guelph. Other professors have contributed immensely to my graduate studies. I am especially grateful to my dissertation committee member Alex Maynard for his advice on my research and for making a graduate level econometrics course very enjoyable. I would like to thank Chris McKenna, David Prescott, and Ana Ferrer for their helpful suggestions on my dissertation. I also want to thank professors Francis Tapon, Thanasis Stengos, Ross McKitrick, Steve Kosempel, René Kirkegaard, Yiguo Sun, Mei Li, Michael Hoy, and Asha Sadanand for their contribution to my understanding of economic theory. I also gratefully acknowledge access to data provided by the Statistics Canada Data Centre Network. Lastly, I would like to thank my family and friends for all their love and encouragement. For my parents who supported me in all my pursuits, for Esmond whose support during the final stages of this Ph.D. is very appreciated, and for the time spent with my friends Diana Alessandrini, Fraser Summerfield, and Joniada Milla at Guelph!

5 v Table of Contents List of Tables List of Figures vii ix 1 Social Networks and Migration Decisions: Evidence from China Introduction Literature Review Data Variable Construction Probit Estimation and Instrumental Variable Method Unexpected Changes in Rainfall and Network Effect Empirical Results Instrumental Variable Results Network Effects for First-time and Repeat Migrants Employment Outcomes and the Size of Migrant Networks Conclusion Appendix Occupational Characteristics and Gender Wage Inequality: A Distributional Analysis Introduction Data Variable Construction Summary Statistics The Gender Gap Across the Wage Distribution Empirical Results Quantile Decomposition Method Explained and Unexplained Proportion of Gender Gap Gender Differences in Work Experience, Union, Sector, Degree Attainment, and Fields of Study

6 vi Gender Differences in Occupational Characteristics Accounting for Selection into Paid Work Conclusion Appendices The Evolution of Returns to Education in the High-End Labor Market in Canada Introduction Literature Review Data Empirical Results Wage Premium on Postgraduate Education Relative to BA Returns to Postgraduate Education by Age Group Returns to Postgraduate Education by Major Field of Study Returns to Professional Degrees by Major Field of Study, Relative to MA Potential Explanations for the Decline in the PG-BA Wage Gap Conclusion References 136

7 vii List of Tables 1.1 Summary Statistics Network Effects on Migration with Alternative IV Migrant Network Effects Migrant Network Effects for First-time Migrants Migrant Network Effects for Repeat Migrants The IV Estimate of Network Effects on Labour Market Outcomes for Migrants The IV Estimate of Network Effects on Labour Market Outcomes by Migration Experience A1 Geographical Information in Each Province A2 The IV Estimate of Network Effects on Labour Market Outcomes for Migrants A3 The IV Estimate of Network Effects on Labour Market Outcomes for First-time Migrants A4 The IV Estimate of Network Effects on Labour Market Outcomes for Repeat Migrants Sample Size by Educational Category and Gender Skill Classifications and Examples of Occupations Summary Statistics: Labour Market Attributes of Full-time Employees Summary Statistics: Occupational Characteristics for Full-time Employees Percentage of Full-time Employees in Each Occupation Gender Differences in Managerial Responsibilities The Proportion of Workers at Different Parts of the Wage Distribution Variables Used in Decomposition Analysis Explained and Unexplained Proportion of Gender Wage Gap The Contribution of Subsets of Covariates for the Low-Educated Workers

8 viii 2.11 The Contribution of Subsets of Covariates for the High-Educated Workers Fraction of Gender Gap Explained by Differences in Occupational Characteristics and Industry (%) for Low-educated Workers Fraction of Gender Gap Explained by Differences in Occupational Characteristics and Industry (%) for High-educated Workers O*Net Characteristics in Service, Trade and Manufacturing Occupations 74 B1 O*Net Characteristics B2 The Contribution of Subsets of Covariates in Model 1 for the Low- Educated Workers B3 The Contribution of Subsets of Covariates in Model 1 for the High- Educated Workers Summary Statistics The Evoluation of PG-BA Wage Gap from 1995 to The PG-BA Wage Gap by Age Group for Men The PG-BA Wage Gap by Age Group for Women The Difference in Average Job Tenure between BA and PG by Age Group The PG-BA Wage Gap by Major Field of Study for Men The PG-BA Wage Gap by Major Field of Study for Women Returns to Professional Degrees by Major Field of Study for Men, Relative to MA Returns to Professional Degrees by Major Field of Study for Women, Relative to MA The PG-BA Wage Gap by Occupation for Men The PG-BA Wage Gap by Occupation for Women The PG-BA Wage Gap by Occupation for Young Workers

9 ix List of Figures 1.1 Response of Self-Migration and Networks to Alternative IV The Gender Gap at Various Points of the Wage Distribution Gender Gap Correcting for Selection B1 Gender Gap across the Wage Distribution by Employment Status B2 Gender Gap across the Wage Distribution by Employment Status with 95% Confidence Interval The Evolution of PG-BA Wage Gap from 1995 to Average Weekly Wage by Birth Year in 1995 and 2010 for Men Average Weekly Wage by Birth Year in 1995 and 2010 for Women Proportion of Men with the Same Education in Different Occupations Relative Wage Gap in 2010 for Men Proportion of Women with the Same Education in Different Occupations Relative Wage Gap in 2010 for Women

10 1 Chapter 1 Social Networks and Migration Decisions: Evidence from China 1.1 Introduction In the China s labour force a tremendous number of workers are from rural areas. They are called migrant workers and they have made a significant contribution to the fast growth of the Chinese economy since Migrant workers in China experience the difficulty of establishing a life in destination cities because of China s rigid rural-urban migration system. 1 They often rely on their migrant contacts for support in cities. While research on 1 China s interregional migration system is based on a person s hukou. The hukou system works as if it were citizenship. It is determined by the person s birth region, rural or urban, and birth place and affects various aspects of people s life. Examples are people s own and children s education, health care, and pension plans.(fan, 2008) When rural residents work in cities, they apply for a temporary residential permit (TRP) and can only stay in cities with a valid TRP. Most migrants cannot obtain a permanent residential permit in their destination cities.(chen et al., 2010)

11 2 international immigration has shown a causal link between migrant contacts and individual immigration decisions, a similar link between migrant contacts and rural-urban migration has not been well established. Using Chinese household survey data in 2002, this chapter contributes to the literature by examining how migrant networks affect rural residents migration decisions and their job market outcomes in cities. In this study migrant networks are defined as interpersonal ties linking migrants to family members, friends, and people in their villages of origin. Theoretically, these networks could affect a person s migration decision either positively or negatively. They could positively affect the probability of migration by reducing the costs of migration because migrant contacts offer information on the urban labour market and provide support to prospective migrants in the workplace. They could negatively affect the probability if migrant contacts share the challenges of living in cities and thereby influence village residents not to migrate. I attempt to identify the causal effect of migrant networks on the migration decision of rural residents using the Chinese Household Income Project Survey (CHIPS) I use the ratio of migrants in the village of origin to the number of village residents to measure the size of a person s migrant network. I examine whether the network effects differ between people with and without previous migration experience, and whether migrant networks affect people s migration decisions through the channel of improving people s labour market outcomes in cities. 2 A difficulty in identifying the causal effect is that some village characteristics, such as 2 The data set for this study does allow me to observe whether rural residents reported living outside their township for more than one year. I use this information to distinguish rural residents with migration experience from rural residents without migration experience.

12 3 the size of arable land in the village, affect the migration decision of village residents. If those characteristics are not observed by researchers, which is often the case, the measure of migrant networks is endogenous. Within the literature examining rural-urban migration in China, a few studies have attempted to identify migrant network effects, but the results are mixed depending on the identification strategies used. In this study the network effect is estimated using an instrumental variable method. I construct the instrumental variable using unexpected changes in rainfall in I find daily rainfall between April and October from 1996 to 2002 in a township and compute the mean of daily rainfall over 7 years in the township. I calculate the average amount of daily rainfall in the township in 1999 and subtract the average daily rainfall from the mean of daily rainfall over 7 years. The instrumental variable is the absolute value of the difference. I argue that this is a better identification strategy than those used in existing studies. 3 Statistical tests show that my instrument is positively and strongly correlated with the size of networks at the 1% level. The F-statistic value on the rainfall instrument is 30, which is 3 times greater than the critical value under which the instrument is weak. The validity of my instrumental variable estimates relies on the assumption that the instrument does not have a direct impact on migration. The use of sudden changes in rainfall in an earlier year reduces the possibility that rainfall affects people s migration decisions. Throughout the analysis I directly include unexpected changes in rainfall at the survey time as a covariate. This accounts for the plausible correlation between the instrument 3 The idea of using average daily rainfall as the instrument comes from Munshi (2003). In the literature review, I provide a detailed discussion on the identification strategy used in the previous studies and why my identification strategy is better than the existing ones.

13 4 and the residual term of the migration determinants function. In comparison to household characteristics used as the instruments in earlier studies, which could be correlated with an individual s migration decision, unexpected changes in rainfall in 1999 is a better instrumental variable in identifying the network effect. My results show that the magnitude of network effects on migration varies by age group, gender, and between people with and without previous migration experience. Overall, an increase of 10 percent in the proportion of migrants in a person s village increases the probability of migration by 8 percent. The network effect is smaller for people without migration experience (first-time migrants) compared to people with migration experience (repeat migrants). I find that migrant networks play a significant role in determining migration decisions only for repeat migrants and first-time migrants younger than 30. For everyone else, the impact is insignificant. To the best of my knowledge, my paper is the first study providing evidence of heterogeneity in network effects between first-time and repeat migrants. Turning to the labour market outcomes of migrants, my results show that migrant networks do not affect migrants annual earnings in cities, but they do help young firsttime migrants and repeat migrants in finding jobs that have longer tenures and better work environments. There is no significant impact of networks on job quality for first-time migrants older than 30. These findings support the conclusion that migrant contacts in the villages affect migration decisions for some migrants, by improving their job quality after they migrate to cities.

14 5 The chapter is organized as follows. Section 1.2 summarizes the existing studies on migration network effects. Section 1.3 describes the data and identification strategy used in this chapter. Section 1.4 presents the results. Section 1.5 concludes. 1.2 Literature Review Microeconomic theory explains the choice of migration as an outcome of an individual s expectations regarding positive net returns from movement.(massey et al., 1993) A number of studies examining international immigration show that migrant networks, which consist of people s friends, family and community members in their countries of origin and destination, affect migrants immigration decisions by reducing the cost of immigration through jobsearch assistance (Munshi, 2003), helping to find welfare and health care programs (Bertrand et al., 2000; Devillanova, 2008), helping to find a destination location(bartel, 1989), and providing relational support (Schwartz, 1973; Berry, 1997). In the context of rural-urban migration in China, previous studies have found a positive relationship between migrant networks and the probability of rural residents migrating to cities. Most of these studies attribute the impact to job-search assistance (Zhao and Li, 2003; Du et al., 2005; Bao et al., 2007; Ioannides and Topa, 2010; Knight et al., 2011). 4 Zhao (2003) finds that the importance of early migrants comes from the guidance and assistance that they offer to new migrants. In line with Zhao (2003), Knight and Song (2005) use a 4 This is also found in other countries such as India (Banerjee, 1983; Iversen et al., 2009) and Uganda (Muto, 2012).

15 6 survey conducted in 8 provinces in 1995 and report that lack of contacts and information about the labour market in cities are primary factors preventing rural labor from migrating. Even though the positive relationship is consistent with social network theory, the causal link of this relationship is subject to considerable debate. Since people living in the same village will all experience any shocks that occur to the local labour market, shocks that change the probability of neighbors migrating will similarly change the probability of survey participants migrating. Empirical analysis, which fails to take account of this fact, would generate a biased estimate of migrant network effects (Manski, 1995). Two methodologies are used to identify the migrant network effect on China s ruralurban migration. The first methodology is to estimate the proportion of migrants in the current year with the proportion of migrants in people s villages of origin in the previous year (Zhao, 2003). However, Bertrand et al. (2000) empirically show that the ratio of migrant workers in the prior year could still be correlated with omitted personal and group characteristics that affect people s migration decision in the current year. The second methodology is to use an instrumental variable method (Lu et al., 2008; Chen et al., 2010). The results in these studies are mixed. Let s denote a village resident as M, Lu et al. (2008) use the political identity of M s father in the Mao era as the instrumental variable for the size of migrant networks for M in the CHIPS Using the two-stage least square (2SLS) estimator, they find that migrant networks do not affect the probability of migration. Using the 2006 China Agricultural Census, Chen et al. (2010) use three instrumental variables to estimate the migrant network effect for M: the percentage of adult

16 7 residents living in M s village whose first birth has two or more children, the percentage of female adults residing in households with a girl firstborn, and the percentage of male adults residing in households with a girl firstborn. Their 2SLS estimates show that a 10% increase in the migration rate raises the probability of migration by 7.3%. As explained in Chen et al. (2010), the migrant network effect is equivalent to the impact of 7-8 years of education on migration. This is not trivial, given that the average years of schooling in their sample is 7 years. A better identification strategy is needed in defining the migrant network effect. Chen et al. (2010) argues that M s father s political identity, which is used in Lu et al. (2008) as the instrumental variable, affects M s migration decision by affecting M s social ties. However, it is not clear why it affects M s neighbours migration decisions. The latter is the measure of the size of M s migrant networks. Therefore, Chen et al. believe that the political identity of M s father should be used as a covariate in the main regression rather than the instrumental variable. However, there are also concerns with the instruments in Chen et al. (2010). The first concern, as discussed by the authors themselves, is that missing information on adults own fertility may affect the accuracy of the analysis. The second concern is related to the validity of the instruments. Other studies have found that in China and other developing countries, the fertility rate of rural-urban migrants is significantly lower than that of non-migrants. 5 Thus, the proportion of first-born girls is directly correlated with migration. 5 See Goldstein et al. (1997) for evidence in China, Lee and Farber (1984) for Korea, Chattopadhyaya et al. (2006) for Ghana, and Lee and Pol (1993) for Korea and Mexico.

17 8 I argue that my instrument is better suited to identify the network effect on China s rural-urban migration. The idea of using rainfall information, so-called distant-past rainfall, is borrowed from Munshi (2003), who uses distant-past rainfall as the instrument for the size of immigrant networks, specially networks made up of Mexican immigrants who are from a new immigrant s community in Mexico and immigrated to the US. 6 He finds that having more established immigrant contacts, immigrants who are located continuously at the destination for three or more years, in the new immigrant s networks increases the probability of the new immigrant finding a nonagricultural job. Munshi (2003) shows that the amount of rainfall in a year only affects the local labour market in the same year. Distant-past rainfall does not affect the migration decision in the year of interest. A similar strategy is used to examine the impact of migration on household consumption growth (Giles and Yoo, 2007) and educational attainment of youth (Brauw and Giles, 2008) in rural China. Empirical evidence suggests that migrant networks are implemented as a dynamic process. Information on migration flows from experienced migrants to new migrants at a point of time. The latter become experienced migrants in a few years and then help new migrants find jobs at their places of destination (Banerjee, 1983; Shah and Menon, 1999; Hooghe et al., 2008). This suggests that it is the new migrants, not the experienced migrants, who benefit from migrant networks. This study will provide evidence on this issue by 6 Rainfall conditions are used in a number of studies to examine internal and international migration (Henry et al., 2004; Deshingkar and Grimm, 2005; Choi and Yang, 2007; Giné et al., 2008). For example, Henry et al. (2004) analyze how the rainfall condition affects internal migration in Burkina Faso villages. Their study suggests that people from the drier regions are more likely than those from wetter regions to move to other areas.

18 9 examining the migrant network effect separately for people who never migrated before 2002 and people who had migrated prior to Data The main source of data in this study comes from the Chinese Household Income Project Survey (CHIPS) Questionnaires are designed separately for urban and rural residents to account for the different geographic and demographic characteristics between rural and urban regions. I use the rural household survey conducted during the period of Chinese New Year in 2003 when most migrants go back home to meet their families. The survey questions are related to various aspects of people s life in In the rest of the paper the year 2002 is referred to as the survey time. I use rural residents whose age is between 16 and 65 in They are from 22 provinces, which are Beijing representing the three metropolitan cities; Jiangsu, Zhejiang, Guangdong, Shandong, Liaoning, and Hebei representing the eastern region; Shanxi, Jilin, Anhui, Jiangxi, Hubei, Henan, and Hunan representing the central region; and Yunnan, Gansu, Guizhou, Sichuan, Chongqing, Shaanxi, Guangxi, and Xinjiang representing the western region. 7 The use of broad geographical regions ensures that the results reported in this study are representative of China. 7 Sample selection frames can be found in Knight and Gunatilaka (2010) and Knight et al. (2011).

19 Variable Construction Using the definition from the National Bureau of Statistics of China, I define a migrant worker as a rural resident who spent at least 30 days working/looking for jobs in cities that are outside of their townships of origin. A person s migrant network is measured using the proportion of migrant workers over the number of residents in his/her village of origin. The main question used to identify migrant workers is how long did you stay out of the household in 2002? A potential measurement error in this exercise is that it classifies people as migrant workers if they live in the township but do not live with their families in In order to mitigate this measurement error, I use the questions asking how much time a person spent on agricultural and non-agricultural activities during the harvest time. 8 People who spent 330 days or more on farm and/or home production in the countryside during the survey time are not considered to be migrants. Similar to previous studies, I cannot directly control for the labour market experience of rural migrants in cities because the CHIPS does not report years of experience for rural migrants. However, the survey does ask have you lived outside of the township for at least one year? and I use this question to identify rural residents with and without migration experience prior to Throughout the paper, repeat migrants refers to rural residents who reported living outside of the township for at least one year and were also migrants in 8 Survey participants were asked to answer during the harvest season, how many days you spent in planting, raising livestock (including in the yard), and on nonproductive activities (schooling, housework, taking care of sick family members and so forth. 9 If a person lived outside of township for at least one year, this person was a migrant before the year of 2002.

20 , while first-time migrants refers to rural residents who had not lived outside of the township for one year but were migrants in A disadvantage of the CHIPS is that I cannot observe which city a person moved to in 2002 or which city the person stayed in before 2002, thus I cannot observe the person s network in the destination cities. Network effects estimated here do not account for the impact of migrant contacts in cities on the migration decision. 10 For each township studied, I collect the size of arable land and the total population in 1989 from the provincial yearbooks that were published in The measure of migrant networks is the proportion of migrants over the number of residents in a village. The use of arable land per person a decade before the survey time is preferred to the arable land per person in 2002 because the latter would cause multicollinearity in the analysis. This paper uses two geographical units: villages and townships. There are a number of villages in one township. A person s migrant networks are measured as the proportion of migrants from the person s village of origin. Unfortunately, the yearbooks in 1990 do not report information on land and population for people s villages of origin. Thus, I construct the arable land per person by dividing the size of arable land by the population in the township where a person was born. Daily rainfall information from 1996 to 2002 in townships is from the China Meteorological Administration Data Center. 11 Data from 1996 is used because most townships in my dataset do not have rainfall information before I first select the weather station that 10 For the same reason, I do not account for the labour market conditions in destination cities. 11 Details of climate data are available at

21 12 is nearest to a township by comparing the distance of all weather stations to the township with their latitude/longitudinal points. Then I collect the amount of daily rainfall that was recorded by the nearest weather station Probit Estimation and Instrumental Variable Method This study uses a Probit model that is presented in equation (1), Y i = α 0 + α 1 M ( i) + α 2 X i + ε 1i (1.1) where Y i is a binary variable that describes a person s migration status in It is 1 if person i migrates to cities, 0 otherwise. M ( i) is the migrant networks in the person s village of origin, which is the proportion of migrants in the village, excluding person i. X i is a set of explanatory variables that represent individual, household, village,and township characteristics. 12 Standard errors are clustered by villages Specifically, individual characteristics are whether a person is female, age, whether a person is married, and whether a person has high school or university education. A household characteristic is the years of schooling for household heads. A village characteristic is the distance to the nearest bus/train/dock station. The township characteristics are arable land per person in 1989 and the absolute value of deviation of average daily rainfall in 2002 from the mean of daily rainfall over the years of Equation (1) is presented in a linear function format in order to help readers understand the functional form. A formal Probit model is displayed as follows. Let y be unobserved, a latent variable, and determined by y i = γ 0 + γ 1 M ( i) + γ 2 X i + e, y i = 1 [y i > 0], where e has the standard normal distribution. If income in countryside is normalized to be 0, an example of y is expected income from working in a city. When expected income from migration is greater than the income in countryside, a person decides to migrate. The response probability of y is P (y = 1 X, M) = P (y i > 0 X, M) = P (e > (γ 0 + γ 1 M ( i) + γ 2 X i ) X, M) = 1 G( (γ 0 + γ 1 M ( i) + γ 2 X i )) = G(γ 0 + γ 1 M ( i) + γ 2 X i ),

22 13 Table 1.1: Summary Statistics Migrants Nonmigrants (1) (2) (3) (4) All First-time Repeat Obs(#) 4,224 2,542 1,682 12,571 Charateristics Average size of migrant networks Age Female Married Completed high school or college Schooling of household head Distance from village to a nearest station(km) Average arable land per person in 1989 (mu) Distance of 2002 s rainfall from the mean over 7 years (mm) I estimate a binary instrumental variable model with maximum likelihood estimator. The proportion of migrants is estimated with the following equation, where Z i is the instrumental variable. M ( i) = β 0 + +β 1 Z i + β 2 X i + ε 2i (1.2) In Table 1.1, I present the summary statistics of explanatory variables that are included in the estimation. 14 In total, 26% of rural residents are migrant workers. 60% of them are first-time migrants. Overall, migrants are 10 years younger than non-migrants. They are more likely to be men, less likely to be married, and are more educated than non-migrant workers. 15 Migrants live in townships that had less arable land per person in 1989 and a where G is a function taking on values between zero and one: 0 < G(X, M) < Appendix Table A1 presents the number of villages, townships, and migrants in each of the 22 provinces. 15 These findings are consistent with Zhao (2003).

23 14 Table 1.2: Network Effects on Migration with Alternative IV Panel A: IV estimate of network effects on migration (1) (2) (3) (4) (5) Migrant network (0.14) (0.27) (0.18) (0.31) (0.35) Panel B: The effect of alternative instrument on the size of network Excluded instrument (0.0005) (0.0005) (0.0005) (0.001) (0.001) F-statistics on the excluded instrument Controls demographical variables abs. deviation from mean over time in 2002 abs. deviation from provincial mean in 2002 Excluded Instrument abs. deviation from mean over time in 1999 abs. deviation from mean over time in 2000 abs. deviation from mean over time in 2001 abs. deviation from mean over time in abs. deviation from provincial mean in Notes: The table reports the IV estimate of network effects and the effect of excluded instrument on the size of network when a different excluded instrument is used. The sample size is 14,211. Results are estimated with a Probit instrumental variable model. The Probit estimate of network effect corresponding to the IV estimates in columns (1) - (5) is Standard errors are clustered at village level and reported in parentheses. Demographical variables are listed in Table 1.1. larger amount of unexpected rainfall than non-migrants. This satisfies the hypothesis that people are more likely to migrate to cities, when they live in townships where land and rainfall conditions are less favorable to farming Unexpected Changes in Rainfall and Network Effect I use unexpected changes in rainfall to identify exogenous variations in the ratio of migrants across townships. To estimate an expected change in the amount of rainfall, I collect data on daily rainfall between April and October and between 1996 and Using this data, I calculate the mean of daily rainfall over the 7-year period in each township. I also calculate the average daily rainfall in each year for each township. Then I calculate the absolute difference between the 7-year mean and the yearly averages. The absolute

24 15 difference tells us how different average daily rainfall is in each year relative to the level of rainfall that agricultural products have been adapted to. A critically large absolute value of deviation would result in a lower quantity of agricultural products, leading to a decrease in a rural person s income and an increase in the person s incentive to migrate. I use the absolute value of the deviation of average daily rainfall in each year as the excluded instrument and report the IV estimate of network effects on migration and the effect of instrument on the size of migrant network in Panel A and B in Table 1.2. I find that F-statistics on the absolute value of deviation between 1996 and 1998 (column (4)) are below 10, suggesting that this variable is weakly correlated with the size of migrant networks. 16 Among the three variables that pass the weak instrument test, the absolute value of deviation in 1999, 2000, and 2001, the absolute value of deviation in 1999 has the largest F-statistics and has a positive and significant impact on the size of network in the first stage. Therefore, I choose the absolute value of the deviation in 1999 as the excluded instrument for the study. 17 Another reason why the year of 1999 is preferred over 2000 and 2001 is related to the validity of the identification strategy that relies on the assumption that the instrument is 16 Stock et al. (2002) suggests that when F-statistics on the excluded instrument are smaller than 10, the excluded instrument is weakly correlated with the endogenous variable and the IV estimate is biased towards the estimate without accounting for endogeneity. The greater the F-statistics value, the better it is in ruling out the weak instrument problem (Angrist and Pischke, 2008). 17 An earlier version of this chapter used average daily rainfall between 1996 and 1998 as the instrument. The idea was to use variation in average daily rainfall across townships to predict the exogenous variation in the proportion of migrants in To test whether this idea provides a good instrument, I used daily rainfall between 1996 and I computed the mean of daily rainfall of all the townships located in the same province and subtracted the average daily rainfall in a township between 1996 and 1998 from the provincial mean. Row (5) shows that when I use the absolute value of the deviation between 1996 and 1998 as the instrument, the instrument cannot pass the weak instrument test.

25 16 Figure 1.1: Response of Self-Migration and Networks to Alternative IV not correlated with the residual term in the migration determinants equation (Equation (1)). Suppose a person lives in a village that flooded in 2001 and decided to migrate in The absolute value of deviation of average daily rainfall in 2001 is not an appropriate instrument, since it directly causes the person to migrate. Since the year of 1999 is furthest away from the survey year, the absolute value of deviation in 1999 is used to minimize the possibility that the instrument has a direct impact on migration decisions in The instrumental variable is constructed with daily rainfall between April and October because many agricultural products (e.g. rice, wheat, and corn) are planted before April and harvested by October. It is crucial to have rainfall conditions that are favorable to

26 17 farming between April and October. Since farming is a source of income for rural residents, unexpected changes in average daily rainfall that occurred between April and October is an ideal choice for the instrumental variable. In Figure 1.1, I plot two graphs using a nonparametric estimator. 18 The graph on the left is the correlation of the self-migration status variable and the size of migrant networks with the deviation of average daily rainfall from the time mean in The y-axis is deviation of average daily rainfall in 1999, with the mean of daily rainfall over the 7-year period in townships normalized to 0. The graph shows that when people live in townships that had very little or a lot of rain in 1999 relative to the mean, they have a greater size of migrant networks and are more likely to migrate. The graph on the right is the correlation of the self-migration status variable and the size of migrant networks with the instrumental variable. Points on the solid curve are comparable to the reduced-form estimates and points on the dashed curve are comparable to the first-stage estimates. It is clear that the probability of self-migration and the size of migrant networks increase as the instrumental variable increases. It is the exogenous changes in average daily rainfall three years prior to the survey that increased the number of migrants at that time. This enables me to estimate the causal effect of network on migration. 18 For the graph on the left, I use the Epanechnikov kernel function and run a local polynomial kernel regression of migration status variable and the size of migrant networks on the deviation of average daily rainfall in For the graph on the right, I run the local polynomial kernel regressions on the absolute value of the deviation.

27 Empirical Results Instrumental Variable Results In Table 1.3, I present the results of the migration network effects analysis of the Probit model, the instrumental variable method (IV), and the reduced-form equation. In order to explore the heterogeneity in migrant network effects for people in different age groups, I report the estimates separately for rural residents in four age groups: 16-30, 31-40, 41-50, and The estimates are reported in columns (2)-(5), respectively. The IV estimate of migrant networks is similar to the Probit estimate for most age groups. Table 1.3 shows that an increase of 10 percent in the proportion of migrant workers increases the probability of migration by 8.2 percent for all residents, 8.7 percent for men, and 7 percent for women. The corresponding figures that are estimated with the Probit model are 8.2, 10, and 6.1 percent. In contrast to the Probit estimate, the IV estimate of network effects for workers in the age group and women in the age group is not significantly different from zero. This suggests that the significant impact of migrant networks for people in these age groups, as estimated with the Probit model, only reflects the positive correlation between people s own migration decisions and the migration decision of their village fellows. As expected, reduced-form results show that when people are making migration decisions, most of them respond strongly to the unexpected changes in rainfall three years before the survey time. Specifically, the absolute value of deviation of average daily rainfall

28 19 Table 1.3: Migrant Network Effects (1) (2) (3) (4) (5) Panel A: All Residents Proportion of Migrant Workers Probit Model Migrant Networks (0.04) (0.05) (0.05) (0.06) (0.04) IV Maximum Likelihood Estimator: 1 st stage Absolute Value of Deviation in (0.0004) (0.0005) (0.0005) (0.0004) (0.0005) IV Maximum Likelihood Estimator: 2 nd stage Migrant Networks (0.13) (0.17) (0.25) (0.22) (0.14) Reduced Form Absolute Value of Deviation in (0.001) (0.001) (0.0008) (0.0006) (0.0004) # of obs 14,211 4,712 3,688 Panel B: Men Proportion of Migrant Workers Probit Model Migrant Networks (0.05) (0.06) (0.08) (0.09) (0.06) IV Maximum Likelihood Estimator: 1 st stage Absolute Value Deviation in (0.0004) (0.0005) (0.0005) (0.0004) (0.0005) IV Maximum Likelihood Estimator: 2 nd stage Migrant Networks (0.19) (0.24) (0.42) (0.31) (0.22) Reduced Form Absolute Value of Deviation in (0.0008) (0.001) (0.001) (0.0009) (0.0006) # of obs 8,044 2,559 2,028 1,815 1,642 Panel C: Women Proportion of Migrant Workers Probit Model Migrant Networks (0.04) (0.07) (0.06) (0.05) (0.04) IV Maximum Likelihood Estimator: 1 st stage Absolute Value of Deviation in (0.0004) (0.0005) (0.0006) (0.0006) (0.0007) IV Maximum Likelihood Estimator: 2 nd stage Migrant Networks (0.14) (0.20) (0.30) (0.20) (0.32) Reduced Form Absolute Value of Deviation in (0.0005) (0.001) (0.0006) (0.0005) (0.0004) # of obs 6,177 2,153 1,660 1, Notes: This table reports the marginal effects that are estimated with the Probit model and an instrumental variable method. Dependent variable is 1 if one is a migrant, 0 otherwise. IV estimates in bold are significantly different the Probit estimates at the 10% level. Standard errors are clustered by villages. The Probit model and the instrumental variable method control for the covariates that are presented in Table 1.1.

29 20 in 1999 has a positive impact on the probability of migration and the impact is greatest for people in the age group Network Effects for First-time and Repeat Migrants The previous section shows that migrant networks affect the probability of migration differently for people in different age groups. It is also useful to investigate whether the networks have different impacts for people with different levels of migration experience. For people who never migrated before 2002, the migrant contacts in their villages could be the only source of information about cities. However, if people who migrated before 2002 have contacts in the cities in which they stayed, their migration decision would be less dependent on the migrant contacts who live in their villages. In this section, I explore the network effects for two types of migrants. The first type of migrants is first-time migrants. I investigate their migration decision relative to people without migration experience. The second type of migrants is people who were migrants before I investigate the migration decision for past migrants who also migrate in 2002, relative to past migrants who do not migrate in Overall, I find that people with migration experience in all age groups and people with no migration experience prior 2002 who are younger than 30 are more likely to migrate when they have more migrant contacts. The network effect is slightly greater for young first-time migrants than repeat migrants. I repeat the same analysis shown in Table 1.3 for people without migration experience and report the results in Table 1.4. As only 4% of people in the age group are first-time

30 21 Table 1.4: Migrant Network Effects for First-time Migrants (1) (2) (3) (4) Panel A: All Residents Proportion of First-time Migrants Probit Model Migrant Networks (0.04) (0.07) (0.06) (0.04) IV Maximum Likelihood Estimator: 1 st stage Absolute Value of Deviation in (0.0005) (0.0005) (0.0005) (0.0005) IV Maximum Likelihood Estimator: 2 nd stage Migrant Networks (0.16) (0.23) (0.31) (0.15) Reduced Form Absolute Value of Deviation in (0.0006) (0.001) (0.0008) (0.0004) # of obs 11,704 3,454 3,145 5,105 Panel B: Men Proportion of First-time Migrants Probit Model Migrant Networks (0.06) (0.08) (0.10) (0.06) IV Maximum Likelihood Estimator: 1 st stage Absolute Value of Deviation in (0.0005) (0.0005) (0.0006) (0.0005) IV Maximum Likelihood Estimator: 2 nd stage Migrant Networks (0.27) (0.42) (0.58) (0.26) Reduced Form Absolute Value of Deviation in (0.0008) (0.001) (0.001) (0.0007) # of obs 6,386 1,835 1,653 2,898 Panel C: Women Proportion of First-time Migrants Probit Model Migrant Networks (0.04) (0.08) (0.05) (0.03) IV Maximum Likelihood Estimator: 1 st stage Absolute Value of Deviation in (0.0005) (0.0005) (0.0006) (0.0005) IV Maximum Likelihood Estimator: 2 nd stage Migrant Networks (0.15) (0.22) (0.31) (0.10) Reduced Form Absolute Value of Deviation in (0.0005) (0.001) (0.0005) (0.0003) # of obs 5,318 1,619 1,492 2,130 Notes: This table reports the marginal effects that are estimated with the Probit model and an instrumental variable method. Dependent variable is 1 if one is a first-time migrant in 2002, 0 if one has never migrated. The proportion of first-time migrants is the ratio of firsttime migrants over the number of people who have never migrated. IV estimates in bold are significantly different the Probit estimates at the 10% level. Standard errors are clustered by villages. The Probit model and the instrumental variable method control for the covariates that are presented in Table 1.1.

31 22 Table 1.5: Migrant Network Effects for Repeat Migrants (1) (2) (3) (4) (5) (6) All Residents All Residents All Residents All Residents Men Women Proportion of Repeat Migrants Probit Model Migrant Networks (0.08) (0.09) (0.20) (0.13) (0.09) (0.12) IV Maximum Likelihood Estimator: 1 st stage Absolute Value of Deviation in (0.0006) (0.0007) (0.0007) (0.0006) (0.0006) (0.0006) IV Maximum Likelihood Estimator:2 nd stage Migrant Networks (0.20) (0.30) (0.41) (0.27) (0.20) (0.34) Reduced Form Absolute Value of Deviation in (0.001) (0.001) (0.002) (0.001) (0.001) (0.001) # of obs 2,490 1, , Notes: This table reports the marginal effects that are estimated with the Probit model and an instrumental variable method. Dependent variable is 1 if one is a repeat migrant in 2002, 0 if one has migration experience but is not migrant in The proportion of repeat migrants is the ratio of migrants who are not first-time migrants over the number of people who had migration experience prior to the survey time. IV estimates in bold are significantly different the Probit estimates at the 10% level. Standard errors are clustered by villages. The Probit model and the instrumental variable method control for the covariates that are presented in Table 1.1. migrants, I combine the and age groups and report the migrant network effect for the age group of 16-30, 31-40, and The most important finding in this table is that, compared to the Probit estimate, the IV estimate of migrant network effects is significant only for first-time migrants younger than 30. An increase of 10 percent in the size of migration networks increases the probability by 13 percent for men in the age group and 15.7 percent for women in the same age group. The IV estimate for young women is statistically larger than the Probit estimate, suggesting that the Probit estimate of network effects for young women without migration experience is downward-biased. For people older than 30, the probability of migration is not significantly affected by the size of their migrant networks. On the contrary, the IV estimate of network effects for people with migration experience

32 23 is positive and significant for both genders in all age groups. 19 Table 1.5 shows that an increase of 10 percent in the ratio of migrants in a person s village increases the person s probability of returning to the urban labour market by 13.2 percent. The IV estimate for men is statistically greater than the corresponding Probit estimate, while the IV estimate for women is statistically the same as the Probit estimate. To explain the difference between Probit and IV estimates, we need to consider what these two methods estimate. The instrumental variable method estimates the impact of migrant networks for people at the margin: rural residents who would not have migrated if they did not know the migrant contacts who migrated to cities because of the unexpected rainfall changes in The Probit estimates reflect the positive correlation between an individual and the individual s village cohort who migrate in Migrant contacts who migrated in earlier years have more information related to the labour market in cities. When we compare the influence of migrant contacts who migrated in 1999 with migrant contacts in 2002, Tables 1.3 and 1.4 imply that the former, as estimated with the instrumental variable method, have more influence on the migration decision of rural residents than the latter, as estimated with the Probit method, for men with migration experience and young women without migration experience. Finally, when rural residents have migration experience, they likely have connections in the cities where they worked. Their migration decisions are also affected by these contacts. Table 1.5 shows that the proportion of repeat migrants among people with migration 19 The analysis by gender shows that the IV estimate is positive and significant for each of the three age groups and the magnitude of IV estimates is greater than the magnitude of Probit estimates.

33 24 Table 1.6: The IV Estimate of Network Effects on Labour Market Outcomes for Migrants (1) (2) (3) (4) (5) (6) Earnings Days worked Hours worked Working indoors High temperature Toxics Panel A: All Migrants Mean of DV (1.00) (92.9) (1.23) (0.50) (0.31) (0.28) Network Effect (0.86) (66.91) (1.00) (0.28) (0.27) (0.20) # of obs 3,796 3,985 3,967 4,048 4,048 4,048 Panel B: Male Migrants Mean of DV (1.02) (93.0) (1.52) (0.50) (0.33) (0.32) Network Effect (0.90) (73.5) (1.02) (0.32) (0.30) (0.24) # of obs 2,633 2,754 2,741 2,801 2,801 2,801 Panel B: Female Migrants Mean of DV (0.95) (89.0) (1.55) (0.41) (0.23) (0.17) Network Effect (1.11) (84.06) (1.63) (0.41) (0.26) (0.16) # of obs 1,163 1,231 1,226 1,247 1,247 1,247 Notes: This table presents the impact of migrant networks and other labour market characteristics on various labour market outcomes. The results are estimated with the two-stage least square (2SLS) estimator using migrant workers in Standard errors are clustered by villages. Omitted groups are male migrants and people who are not married and who have education below high school. Columns (1) - (6) present the estimate of impacts on log annul earnings from taking a job that is not related to agricultural activities, the number of days spent in working at the job in 2002, the number of hours spent at the job per day, whether the job requires the person to work indoors, whether the job requires the person to work in a very hot environment, and whether the job involves exposure to toxics, respectively. I present the mean and the standard deviation of dependent variables in the first row of each panel. experience is substantially large relative to the proportion of first-time migrants among people without migration experience. The impact of contacts in cities would decrease the impact of migrant contacts in the countryside. Since the Probit estimator does not control for the size of migrant networks in cities, the Probit estimates of network effects are downward-biased Employment Outcomes and the Size of Migrant Networks One potential reason for the effect of migrant networks on migration decisions is that migrant contacts in the countryside help fellow villagers to find jobs in cities. In this section,

34 25 I focus on the migrants who are employed in 2002, which make up approximately 95% of the migrants in the sample, and investigate the extent to which migrant networks in the countryside affect their labour market outcomes in cities. Overall, I find that knowing more migrant contacts in a person s village of origin does not increase the person s annual earnings, but it does improve the quality of the person s job. The positive impact on job quality is particularly strong for people without migration experience. Table 1.6 presents the second-stage results that are estimated with the 2SLS instrumental variable method for migrants aged between 16 and 65 and separately for each gender. 20 Column (1) shows that the size of migrant networks does not have a significant impact on annual earnings in However, Panel A in Table 1.6 shows that an increase of 10 percent in the proportion of migrants in the home village increases the number of days worked by 24 days and decreases the probability of working outdoors and exposure to toxics by 7.5% and 5.4%, respectively. The impact of migrant networks on job tenure and the probability of exposure to toxics are found for both genders. On top of that, a 10% increase in the size of migrant networks also increases the probability of working indoors by 8%. In Table 1.7 I show that migrant networks affect job market outcomes for first-time migrants younger than 30 and for repeat migrants. For the former, a 10 percent increase in the proportion of migrants in the home village increases the number of days worked by approximately a month and decreases the probability of exposure to toxics at the workplace 20 In Appendix I present the 2SLS estimate of all covariates included in the analysis. The instrument has a positive impact on the size of migrant networks at the 1% level. The first-stage results are available upon request.

35 26 Table 1.7: The IV Estimate of Network Effects on Labour Market Outcomes by Migration Experience (1) (2) (3) (4) (5) (6) Annual Earnings Annual Days Weekly Hours Indoor High temperature Toxics Panel A: First-time Migrants Mean of DV (0.98) (94.3) (1.57) (0.50) (0.31) (0.30) Network Effect (1.14) (95.70) (1.31) (0.47) (0.41) (0.31) # of obs 2,300 2,410 2,400 2,452 2,452 2,452 Panel B: First-time Migrants, age 30 Mean of DV (0.94) (92.1) (1.43) (0.48) (0.29) (0.26) Network Effect (1.50) (110.50) (1.65) (0.55) (0.46) (0.29) # of obs 1,314 1,389 1,384 1,405 1,405 1,405 Panel C: First-time Migrants, age>30 Mean of DV (1.02) (92.2) (1.46) (0.46) (0.34) (0.34) Network Effect (1.33) (131.68) (1.77) (0.58) (0.53) (0.51) # of obs 986 1,021 1,016 1,047 1,047 1,047 Panel D: Repeat Migrants Mean of DV (1.02) (83.64) (1.38) (0.48) (0.30) (0.26) Network Effect (1.02) (66.39) (1.16) (0.34) (0.23) (0.15) # of obs 1,485 1,548 1,540 1,584 1,584 1,584 Notes: This table presents the 2SLS estimates on the labour market outcomes for first-time migrants and the repeat migrants. Standard errors are clustered by villages. Omitted groups are male migrants and people who are not married and who have education below high school. Columns (1) - (6) present the estimate of impacts on log annul earnings from taking a job that is not related to agricultural activities, the number of days spent in working at the job in 2002, the number of hours spent at the job per day, whether the job requires the person to work indoors, whether the job requires the person to work in a very hot environment, and whether the job involves exposure to toxics, respectively. I present the mean and the standard deviation of dependent variables in the first row of each panel.

36 27 by 6.5%. The same pattern is found for repeat migrants, with an increase in days worked of 21 days and decrease in probability of toxic exposure of 4%. In addition, a 10% increase in the size of migrant networks also increases the probability of working indoors for young first-time migrants by 14%. For first-time migrants who are older than 30, migrant networks have little impact on labour market outcomes. These results support the hypothesis that migrant networks in the countryside affect migration through the channel of improving the job quality of rural residents working in cities. 21 Knowing more migrant contacts increases job arrival rate for all migrants. 22 As repeat migrants are likely to have contacts in cities, their labour market outcomes are not only dependent on their networks in the countryside, but also migrant networks in cities. This could explain why migrant networks in the countryside play a smaller role in affecting labour market outcomes for repeat migrants than first-time migrants. 1.5 Conclusion This chapter examines the effect of migrant networks on migration decisions of village residents in 2002 at the village of origin. Using unexpected changes in average daily rainfall in 1999 as an instrumental variable, this study identifies a substantially large network effect 21 In Table A4 I present the network effects on labour market outcomes for young repeat migrants and repeat migrants older than 30. There is a significant impact of migrant networks on annual number of days worked and the probability of exposure to toxics for both groups of repeat migrants. These impacts are smaller for young repeat migrants than young first-time migrants. 22 Social network theory demonstrates that word-of-mouth communication among unemployed individuals and their contacts reduces the search frictions, which in turn increases the number of vacancies learned about by unemployed individuals. For a theoretical explanation, see Calvó-Armengol and Jackson (2004); Calvó-Armengol (2004); Calvó-Armengol and Zenou (2005).

37 28 for male and female first-time migrants younger than 30 and for repeat migrants in all age groups. I explore whether migrant networks in the countryside affect migration decisions through the channel of helping rural residents find a job. My results show that knowing more migrant contacts does not have a significant impact on a person s annual earnings, but it seems to improve a first-time migrant s job quality in terms of job tenure, the probability of working indoors and exposure to toxics. The impact of migrant network on job tenure and the probability of toxic exposure are also found for people with migration experience but the impact is much smaller than that for first-time migrants. A drawback of the chapter is that the CHIPS 2002 does not have information on which city a person migrate to in 2002 and which cities repeat migrants stayed at before This limits the ability to explore the network effect on labour market outcomes in cities, since a migrant often interacts with other migrants residing in the same location. In order to link migrant workers to their contacts in their destination cities, new datasets are needed. My findings suggest that first-time migrants are worse off in terms of wage and work environment relative to people with migration experience. Expanding public services for new migrants, such as services that provide information about local labour market and community supports, would be beneficial.

38 29 Appendix Table A1: Geographical Information in Each Province Province Migrants Average Size of Average Amount of Villages Townships Migrant Networks Unexpected Rainfall in 2002 in 1999 Anhui Beijing Chongqing Hebei Henan Hunan Jiangsu Jiangxi 1, Jilin Liaoning Gansu Guangdong 1, Guizhou Nanning Shaanxi Shandong 1, Shanxi Sichuan Xinjiang Yunnan Zhejiang Wuhan

39 30 Table A2: The IV Estimate of Network Effects on Labour Market Outcomes for Migrants (1) (2) (3) (4) (5) (6) Earnings Days worked Hours worked Working indoors High temperature Toxics Panel A: All Migrants Mean of DV (1.00) (92.9) (1.23) (0.50) (0.31) (0.28) Explanatory Variables: Migrant networks (0.86) (66.91) (1.00) (0.28) (0.27) (0.20) Female (0.04) (3.44) (0.05) (0.02) (0.01) (0.01) Age (0.003) (0.24) (0.004) (0.001) (0.001) (0.001) Married (0.05) (4.40) (0.07) (0.02) (0.02) (0.01) High school (0.06) (4.80) (0.07) (0.02) (0.02) (0.02) # of obs 3,796 3,985 3,967 4,048 4,048 4,048 Panel B: Male Migrants Mean of DV (1.02) (93.0) (1.52) (0.50) (0.33) (0.32) Explanatory Variables: Migrant networks (0.90) (73.5) (1.02) (0.32) (0.30) (0.24) Age (0.003) (0.25) (0.004) (0.001) (0.001) (0.001) Married (0.06) (4.91) (0.07) (0.03) (0.02) (0.02) High school (0.06) (5.60) (0.09) (0.03) (0.02) (0.02) # of obs 2,633 2,754 2,741 2,801 2,801 2,801 Panel B: Female Migrants Mean of DV (0.95) (89.0) (1.55) (0.41) (0.23) (0.17) Explanatory Variables: Migrant networks (1.11) (84.06) (1.63) (0.41) (0.26) (0.16) Age (0.006) (0.57) (0.01) (0.003) (0.002) (0.001) Married (0.11) (9.15) (0.17) (0.04) (0.03) (0.02) High school (0.08) (7.16) (0.13) (0.04) (0.02) (0.02) # of obs 1,163 1,231 1,226 1,247 1,247 1,247 Notes: This table presents the impact of migrant networks and other labour market characteristics on various labour market outcomes. The results are estimated with the two-stage least square (2SLS) estimator using migrant workers in Standard errors are clustered by villages. Omitted groups are male migrants and people who are not married and who have education below high school. Columns (1) - (6) present the estimate of impacts on log annul earnings from taking a job that is not related to agricultural activities, the number of days spent in working at the job in 2002, the number of hours spent at the job per day, whether the job requires the person to work indoors, whether the job requires the person to work in a very hot environment, and whether the job involves exposure to toxics, respectively. I present the mean and the standard deviation of dependent variables in the first row of each panel.

40 31 Table A3: The IV Estimate of Network Effects on Labour Market Outcomes for Firsttime Migrants (1) (2) (3) (4) (5) (6) Earnings Days worked Hours worked Working indoors High temperature Toxics Panel A: First-time Migrants Mean of DV (0.98) (94.3) (1.57) (0.50) (0.31) (0.30) Explanatory Variables: Migrant networks (1.14) (95.70) (1.31) (0.47) (0.41) (0.31) Female (0.05) 5.00 (0.07) (0.02) (0.02) (0.01) Age (0.003) (0.30) (0.005) (0.002) (0.001) (0.001) Married (0.06) (5.87) (0.09) (0.03) (0.02) (0.02) High school (0.07) (6.32) (0.09) (0.03) (0.02) (0.02) # of obs 2,300 2,410 2,400 2,452 2,452 2,452 Panel B: First-time Migrants, age 30 Mean of DV (0.94) (92.1) (1.43) (0.48) (0.29) (0.26) Explanatory Variables: Migrant networks (1.50) (110.50) (1.65) (0.55) (0.46) (0.29) Female (0.06) (5.58) (0.08) (0.02) (0.02) (0.02) Age (0.01) (1.00) (0.01) (0.005) (0.004) (0.003) Married (0.08) (8.14) (0.12) (0.04) (0.03) (0.03) High school (0.09) (8.06) (0.13) (0.04) (0.03) (0.03) # of obs 1,314 1,389 1,384 1,405 1,405 1,405 Panel C: First-time Migrants, age>30 Mean of DV (1.02) (92.2) (1.46) (0.46) (0.34) (0.34) Explanatory Variables: Migrant networks (1.33) (131.68) (1.77) (0.58) (0.53) (0.51) Female (0.1) (10.45) (0.17) (0.05) (0.04) (0.03) Age (0.005) (0.46) (0.007) (0.002) (0.002) (0.002) Married (0.18) (15.43) (0.29) (0.07) (0.08) (0.07) High school (0.08) (7.77) (0.13) (0.04) (0.03) (0.04) # of obs 986 1,021 1,016 1,047 1,047 1,047 Notes: This table presents the 2SLS estimates on the labour market outcomes for first-time migrants. Standard errors are clustered by villages. Omitted groups are male migrants and people who are not married and who have education below high school. Columns (1) - (6) present the estimate of impacts on log annul earnings from taking a job that is not related to agricultural activities, the number of days spent in working at the job in 2002, the number of hours spent at the job per day, whether the job requires the person to work indoors, whether the job requires the person to work in a very hot environment, and whether the job involves exposure to toxics, respectively. I present the mean and the standard deviation of dependent variables in the first row of each panel.

41 32 Table A4: The IV Estimate of Network Effects on Labour Market Outcomes for Repeat Migrants (1) (2) (3) (4) (5) (6) Earnings Days worked Hours worked Working indoors High temperature Toxics Repeat Migrants Mean of DV (1.02) (83.64) (1.38) (0.48) (0.30) (0.26) Explanatory Variables: Migrant networks (1.02) (66.39) (1.16) (0.34) (0.23) (0.15) Female (0.05) (4.32) (0.07) (0.03) (0.02) (0.01) Age (0.005) (0.38) (0.007) (0.002) (0.001) (0.001) Married (0.08) (6.43) (0.10) (0.03) (0.03) (0.02) High school (0.09) (6.2) (0.11) (0.03) (0.02) (0.02) # of obs 1,497 1,560 1,552 1,596 1,596 1,596 Panel B: Repeat Migrants, age 30 Mean of DV (1.01) (78.8) (1.25) (0.45) (0.27) (0.26) Explanatory Variables: Migrant networks (1.18) (71.46) (1.32) (0.40) (0.45) (0.19) Female (0.06) (5.03) (0.08) (0.03) (0.23) (0.02) Age (0.01) (0.94) (0.01) (0.005) (0.02) (0.003) Married (0.10) (7.65) (0.11) (0.04) (0.003) (0.02) High school (0.11) (7.41) (0.14) (0.04) (0.03) (0.03) # of obs 1,061 1,101 1,095 1,130 1,130 1,130 Panel C: Repeat Migrants, age>30 Mean of DV (1.05) (91.54) (1.64) (0.50) (0.34) (0.27) Explanatory Variables: Migrant networks (1.17) (88.75) (1.66) (0.53) (0.38) (0.24) Female (0.12) (11.20) (0.01) (0.06) (0.04) (0.03) Age (0.007) (0.62) (0.01) (0.003) (0.002) (0.002) Married (0.30) (16.90) (0.29) (0.10) (0.07) (0.06) High school (0.13) (10.54) (0.17) (0.05) (0.04) (0.03) # of obs Notes: This table presents the 2SLS estimates on the labour market outcomes for repeat migrants. Standard errors are clustered by villages. Omitted groups are male migrants and people who are not married and who have education below high school. Columns (1) - (6) present the estimate of impacts on log annul earnings from taking a job that is not related to agricultural activities, the number of days spent in working at the job in 2002, the number of hours spent at the job per day, whether the job requires the person to work indoors, whether the job requires the person to work in a very hot environment, and whether the job involves exposure to toxics, respectively. I present the mean and the standard deviation of dependent variables in the first row of each panel.

42 33 Chapter 2 Occupational Characteristics and Gender Wage Inequality: A Distributional Analysis 2.1 Introduction Using a quantile wage decomposition method, this chapter explores the question of why men and women are paid differently when they work in different occupations. Since the 1970s, women have made great improvements in educational achievement and labour market participation, but a male-female wage gap still persists in Canada (Baker and Drolet, 2010). By controlling for occupational dummy variables, several Canadian studies have found that a considerable proportion of the gender gap can be attributed to men and women

43 34 working in different occupations (Fortin and Huberman, 2002; Drolet, 2002a; Boudarbat and Connolly, 2013). 1 A drawback of using the occupational dummies, however, is that this approach does not reveal why gender-specific occupational distribution has an impact on the gender wage gap. For this reason, a few studies replace the occupational dummies with occupation-specific skills, which are extracted from sources such as the Dictionary of Occupational Titles (DOT), and examine how the DOT-skills affect the average gender wage gap (Baker and Fortin, 2001). 2 This study adds to the literature in two respects. First, it constructs a broader set of skill measures and examines how gender differences in these occupation-specific skills affect the gender gap at different points of the wage distribution other than the mean. In addition to the DOT-skills, such as verbal, numerical, and clerical skills, I include workplace competitiveness and the ranking of an individual s managerial position (i.e. non-manager, junior manager, or senior manager.). This analysis shows that more of the gender gap at various points of the wage distribution is explained when occupational dummy variables are replaced with the DOT-skills, workplace competitiveness, and the ranking of managerial positions. Moreover, gender differences relating to workplace competitiveness and the ranking of managerial positions, which were not examined in the existing Canadian literature 1 The extent to which gender differences in occupation contribute to the gender wage gap varies at different point of the wage distribution. Examining the wage gap for young post-secondary graduates, Boudarbat and Connolly (2013) show that the inclusion of the occupational dummies reduces the gender gap by 37% at the mean, 112% at the 10 th percentile of the wage distribution, and 17.7% at the 90 th percentile of the wage distribution. 2 Macpherson and Hirsch (1995), Black and Spitz-Oener (2010), and Bacolod and Blum (2010) conduct a similar analysis for the US. To simplify the explanation, I call the occupation-specific aptitude factors, as used in the previous studies, the DOT-skills.

44 35 on the gender wage gap, explain 30.5% of the gender gap at the 95 th percentile of the wage distribution for university-educated workers, 27 percentage points greater than the 3.5% of the gender gap explained by DOT-skill variables. 3 Workplace competitiveness and the ranking of managerial positions therefore appear to be the principal determinants underlying the glass ceiling phenomenon high-paid women experience a greater wage gap than low-paid women. Previous studies have documented this phenomenon, but reasons as to why women are prevented from obtaining the wage levels of the highest-paid men are missing in the literature (Baker et al., 1995; Drolet, 2002b; Boudarbat and Connolly, 2013). 4 The second contribution of the study is to reveal that women with different educational levels experience the gender wage gap for different reasons. In line with previous studies, I find that gender differences in DOT-skills explain up to 50% of the gender gap for high school and community college graduates, as well as most of the university graduates. The extent to which differences in DOT-skills contribute to the gender gap varies at different 3 A number of studies use laboratory experiments to test the hypothesis that gender differences in attitudes toward competition have a significant impact on the gender gap in productivity (Niederle and Vesterlund, 2011; Cadsby et al., 2013). However, the hypothesis has not been tested using large-scale Canadian micro data yet. The use of micro data has pros and cons in testing the competitiveness hypothesis, compared to laboratory experiments. The main purpose of the study is not to demonstrate whether the use of micro data is more proper in testing the hypothesis than laboratory experiments, but rather to test whether working in jobs with different levels of competitive pressure is an explanation for the gender gap for highly-educated workers. 4 Albrecht et al. (2003) and Christofides et al. (2013) have found the evidence of the glass ceiling phenomenon in European countries, and Blau and Kahn (2006) in the U.S. The glass ceiling phenomenon can be defined in two ways. First, when men and women work in the same occupation, a number of studies have found that high-skilled women are less likely to be promoted than high-skilled men because women experience more career interruption due to child-rearing.(wood et al., 1993; Bertrand and Hallock, 2001; Ginther and Kahn, 2004; Bertrand et al., 2010; Goldin and Katz, 2011; Gicheva, 2013; Goldin, 2014). Second, when the highest-paid women work in occupations different from the highest-paid men, the highest-paid women are paid less than the highest-paid men. A typical example is the pay difference between male top executives and female pharmacists, where pharmacists earn less than top executives, suggesting that the highest paid women lack access to the highest paying jobs of men. Explanations for this type of glass ceiling, which were not subject to comprehensive examination in the literature, are explored in this chapter.

45 36 points of the wage distribution for each educational group. However, there are two groups of workers for whom the DOT-skills are not significantly different between men and women but the gender gap still exists. The first group is university-educated workers above the 90th percentile of the wage distribution. In this group, the analysis shows that men are compensated more because they work in more competitive jobs and take more managerial responsibilities than women. The second group is workers without high school education. In this group, the analysis show that men are compensated more because they experience unpleasant work conditions more often than women. Within previous work that has investigated the impact of occupational characteristics on the gender wage gap for Canada (Baker and Fortin, 1999; Drolet, 2002b), the only Canadian study I am aware of that addresses the relationship of the DOT-skills to pay differences between male- and female-dominated occupations is Baker and Fortin (2001). Using Canadian data from 1987 and 1988, Baker and Fortin (2001) examine whether female-dominated occupations on average pay less than male-dominated and mixed occupations, conditional on occupational characteristics that are extracted from the Canadian Classification and Dictionary of Occupations. They found that men were paid significantly less in femaledominated occupations than in other occupations; however, a significant penalty for women in female-dominated work only exists among women with university education. This study extends their work in two dimensions. First, my study uncovers heterogeneity in the impact of detailed occupational attributes on the gender wage gap at different points of the wage distribution. Second, the analysis for the university-educated workers provides some expla-

46 37 nations as to why university-educated women are paid significantly less in female-dominated occupations than in other occupations, explanations that go beyond the analysis in Baker and Fortin (2001). An important secondary analysis for understanding the gender gaps relationship to education levels is an examination of sample selection induced by non-employment. This exercise is particularly important for examining the gender gap for workers without postsecondary education, because of the relatively low employment rate of women at these educational levels. To account for selection effects, I use alternative imputation techniques to recover the missing wage values of non-working individuals. The first approach that allows for selection through unobserved characteristics is closely related to that of Olivetti and Petrongolo (2008). Previous studies using this imputation approach examined the impact of sample selection on the wage gap at the median of the wage distribution. 5 This is the first study that applies the approach to investigating the impact of selection on the gender wage gap at various points of the wage distribution other than the median. The second approach that allows for selection through observed characteristics is built upon the reweighting method introduced by DiNardo, Fortin and Lemieux (1996). The DiNardo-Fortin-Lemieux method is commonly used in analyzing the wage gap between demographic groups, but it has not previously been applied for the purpose of correcting for sample selection. Different from the first approach, the second approach does not rely on longitudinal data. Therefore, it enables researchers who use cross-sectional data such as Current Population Survey (CPS) 5 See Johnson et al. (2000) and Neal (2004) for an application to the white-black wage gap, Blau and Kahn (2006) and Olivetti and Petrongolo (2008) for an application to the gender wage gap.

47 38 to analyze data with missing wage values. While these alternative imputation techniques reveal different economic channels of selection, results with both imputation approaches confirm that correcting for sample selection makes little difference in estimating the gender gap for individuals with postsecondary education. For individuals without post-secondary education (low-educated), correction for sample selection on observables makes greater changes in the gender gap than selection on unobservables, suggesting that the use of observed characteristics is sufficient to capture the selection rule for low-educated individuals. The rest of the chapter is organized as follows. Sections 2.2 and 2.3 introduce data and variable construction, followed by empirical findings. Section 2.4 concludes. 2.2 Data The data for this study are from the Survey of Labour and Income Dynamics (SLID) for the years 1993 to The choice to use the SLID is motivated by the fact that it contains rich information on individuals work history and educational attainment, including both education levels and major fields of study. More importantly, the SLID provides a longitudinal dimension that offers information on fluctuations in income and on changes in labour market activity over time for up to 6 years. This data is needed to recover missing wage values for individuals who worked in some years but did not work in other years. The SLID has 461,693 people in the age group who were not enrolled

48 39 in school at the time of the survey. I restrict my analysis to full-time employees whose highest education level is known. These restrictions result in a total sample size of 257,937 observations, where 49.6% are women and 50.4% are men Variable Construction The measure of earnings used is the hourly wage. 6 The hourly wage is measured using the total annual earnings (including tips, bonuses and commissions) divided by the annual hours worked. Hourly wage over the period is evaluated in 1993 constant dollars. I focus on the wage gap for full-time employees because it allows a comparison between similar types of workers. Part-time employees tend to have different labour market characteristics than full-time employees. 7 To define educational groups, I use the survey variable highest level of education of a person. There are four educational categories: below high school, high school graduates (low-educated workers), post-secondary education below a four-year university degree, and Bachelor s degree and graduate education (high-educated workers). 8 6 Hourly wage is preferred to other measures such as weekly and yearly earnings because it eliminates the impact of gender differences in working hours on the gender pay gap. 7 The use of full-time workforce data is common in previous studies. See Fortin and Huberman (2002), Baker et al. (1995), and Boudarbat and Connolly (2013) as examples using Canadian data. In Figure B2 in Appendix B, I plot the gender gap at each decile of the wage distribution for full-time employees and for part-time employees with a 95% confidence interval. It is noticeable that the pattern of the gender gap is very different between full-time and part-time employees, which indicates the potential dissimilarities between the two types of employees. 8 The sample of observations with education below high school includes individuals who had zero to 13 years of schooling and did not complete a high school diploma. The category of post-secondary education without completing a four-year university degree includes trade programs, community colleges (with or without certificates), and some university education with no degree. University education includes four-year university degrees, university certificate or diploma above BA but below Master s degree, Master s degree,

49 40 Table 2.1: Sample Size by Educational Category and Gender Total Percentage of Share of Observations Full-time Workers Missing Wage Observations Men Below HS 40,408 58% 37% HS 37,955 69% 26.6% College 108,999 71% 24.1% University 36,840 72% 21.3% Women Below HS 33,059 37% 50% HS 41,694 50% 34.5% College 119,475 56% 26.8% University 43,263 64% 20.3% Notes: Author s calculations using SLID Referenced population: year old, not currently attending school. College includes people who have attended trade programs, community colleges, or universities but did not complete a four-year university degree. Table 2.1 summarizes the number of observations by educational category for men and women. It shows that the employment gap between men and women varies across educational groups. The second column tells us that, in all educational groups, the proportion of full-time employees is larger for men than women, but the difference is reduced by 13% ((58%-37%)-(72%-64%)) when one compares the employment gap for the lowest educational group with that for the highest educational group. 9 The third column shows that the difference in the proportion of non-working individuals between men and women decreases as the level of education increases. Since the employment and non-employment degree in medicine, dentistry, veterinary medicine, optometry or first professional degree in law, and Doctorate. 9 Compared to Boudarbat and Connolly (2013), full-time employees here have a smaller share amongst the college, trade and university observations. This difference is due to the fact that Boudarbat and Connolly (2013) examines the wage gap for employees 2 years and 5 years after graduation, while this study examines the wage gap for all employees. As mature workers are more likely to have career interruptions than young workers, the full-time employment rate for each gender is smaller in this study.

50 41 gaps are much larger for low-educated individuals, it makes sense to examine the impact of sample selection on the gender gap for low-educated observations. The proportion of missing wage observations within each education-gender group is larger than the unemployment rate for that group because missing wage observations result from both unemployment and nonresponse. If individuals who are not employed and individuals who refuse to answer income questions systematically differ from employed individuals who are willing to supply this information, then the estimated gender wage gap based on the the observed wage value would be biased. 10 Therefore, when I account for sample selection, I include missing wage observations due to unemployment and nonresponse. To investigate the impact of gender differences in occupation on the wage gap, I use nine aptitude factors that are extracted from the Career Handbook, which is the Canadian version of Dictionary of Occupational Titles. The aptitude factors are general learning ability, verbal ability, numerical ability, spatial perception, form perception, clerical perception, motor coordination, finger dexterity, and manual dexterity. 11 The level of requirement is between 1 and 5, with 1 denoting the lowest requirement. 12 I match the skills that are required by an occupation to the people who worked in the occupation in a given year. Table 2.2 provides an example of occupations that require a specific aptitude skill. 10 Because middle-income people are more likely to be income respondents than low-income and highincome people, such selection is prone to have a larger impact on the estimates of gender gaps at the tails of the distribution. Thus, correcting for a nonrandom selection amongst the nonrespondents is particularly important in a distributional analysis. 11 Those variables were adopted in Baker and Fortin (2001) for Canada. Macpherson and Hirsch (1995) and Bacolod and Blum (2010) adopted a similar set of variables in studying the gender wage gap in the US. 12 The decomposition results are estimated by treating the skill requirement as continuous variables. I also conduct the analysis by treating the skills as categorical variables. Findings are robust to this change.

51 42 Table 2.2: Skill Classifications and Examples of Occupations Aptitude Skills from the Career Handbook General Learning Ability Verbal Ability Numerical Ability Clerical Perception Spatial Perception Form Perception Motor Co-ordination Finger Dexterity Manual Dexterity Most needed in managerial and natural science occupations. Most needed in occupations such as senior government managers, judges, and university professors. Most needed in occupations such as financial senior managers, professionals in natural science, accountants, and economists. Most needed in administrative work. Most needed in occupations for engineers, computer programmers, and graphic designers. Some blue-collar jobs require a high level of this skill, e.g. machinist and aircraft mechanics. Highest requirement is in landscape architects, physicists, astronomers, and chemists. Highest requirement is in dentists, jewellers, watch repairers, electronics assemblers, and related occupations. Blue-collar jobs, e.g., electrical cable workers, require a level above median. Most needed in occupations such as aircraft technician. Other jobs such as shoe repairers and hair dressers require a level above median. Needed in occupations such as electronic technicians, physicians, cabinetmakers, and craftspersons. O*Net Characteristics Workplace Competitiveness Jobs such as graphic designers, orthodontists, investment fund managers, and top executives require workers to work under the highest level of stress. Jobs such as kindergarten teachers, librarians, cashiers, and general office clerks require workers to work under the lowest level of competition.

52 43 In addition to the aptitude variables, I include the level of workplace competitiveness, as extracted from the O*Net database, 13 and the ranking of an individual s managerial position, as extracted from the SLID. Workplace competitiveness (referred to by O*Net as level of competition at workplace ) measures the extent to which a job requires the worker to compete or to be aware of competitive pressures. It is a continuous variable and is normalized between 0 and 1, with lower values meaning lower competitiveness. Ranking of an individual s managerial position is a categorical variable: 0 if the individual is not a manager, 1 if he/she works at a lower/middle managerial position (junior manager), and 2 if he/she works at an upper/top managerial position (senior manager). When accounting for the impact of occupations on the gender wage gap, most studies control for occupational binary variables, where jobs are grouped into a small number of occupational categories. There are many reasons why aggregated occupational dummy variables are adopted in those studies, e.g. disaggregated occupational categories are not available in the dataset. In applying wage decomposition methods to study the gender wage gap, a necessary condition to make the analysis valid is that men and women within occupations must be comparable in observed labour market characteristics, which is called the common support assumption. 14 If disaggregated occupational dummy variables are applied to the analysis, there are occupations in which 90% of the employees are men. 15 The common support assumption would be violated if the labour market characteristics were 13 The O*Net database is the new version of Dictionary of Occupational Titles. I use the O*Net database because workplace competitiveness is not available in the Career Handbook. 14 See Fortin et al. (2011) for a detailed explanation. 15 An example of such occupations in the SLID is textile machinery mechanics, where only 3 women were employed in the sample period.

53 44 very different between women and men in these occupations. In contrast to those studies, this study controls for nine aptitude factors and the level of competition, which reflect the characteristics of more than 500 occupational categories, and the ranking of a managerial position, which is unique to each individual. Such detailed occupational attributes are preferable to the occupational dummy variables because the comparison in occupational attributes between men and women provides an informative message underlying wage differences between male and female work. Moreover, the use of occupational attributes rather than disaggregated occupational dummy variables yields a sample size for men that is comparable to the sample size for women for each of the eleven occupation-related variables. This enables the computation of meaningful decompositions at deciles of the wage distribution. In the rest of the paper, occupational attributes refers to the nine aptitude variables, workplace competitiveness, and the ranking of an individual s managerial position. This study controls for the proportion of female workers in an occupation and 17 industrial categories. Baker and Fortin (2001) showed that conditional on occupational characteristics, the proportion of female workers plays a significant role in determining wage for men and university-educated women. For this reason, the decomposition analysis accounts for the proportion of female workers in order to capture the unobserved occupational characteristics that are wage determinants. In reality, we see variation, such as office staff in the oil and mining industry being paid very differently than in the finance and business industry. Thus, I control for industry dummy variables in order to account for such pay

54 45 differences Summary Statistics Table 2.3 offers sample mean statistics for men and women by educational group. To keep the interpretation simple, I define the sample of employees with education below high school as the HSD (high school dropouts) group, high school graduates as the HS group, employees with some post-secondary education as the community college group, and employees who have four-year university degrees or above as the university group. The first row presents the proportion of women in the full-time labour force by educational group. Women make up close to half of the full-time employees in three of the educational groups, with the exception of the HSD sample in which only 35% of the fulltime employees are women. This indicates that labour supply behaviour is particularly different between the very low-educated women and women in other educational groups. If selection into full-time employment is not random, it could cause a substantial bias in estimating the gender gap for workers in the HSD group. Job attributes suggest that even though the gender wage gap decreases as women achieve more education, men on average earn more than women in four of the educational groups. Table 2.3 shows that there is a larger proportion of women than men employed in the public sector, but a smaller union coverage rate among women than men for educational levels below university. For the university group, the fraction of women in the public sector and the fraction of union members are both greater than the fractions for men. Previous studies

55 46 Table 2.3: Summary Statistics: Labour Market Attributes of Full-time Employees HSD HS College University M F M F M F M F % of women in FT jobs Job Attributes Hourly Wage (0.39) (0.37) (0.40) (0.40) (0.42) (0.41) (0.47) (0.44) Experience (9.79) (9.76) (9.02) (9.07) (9.1) (8.55) (8.8) (8.16) Union (0.49) (0.44) (0.49) (0.45) (0.48) (0.47) (0.46) (0.50) Public (0.27) (0.30) (0.34) (0.38) (0.39) (0.45) (0.47) (0.50) Highest Education Certificates (0.43) (0.42) BA (0.47) (0.45) Below MA (0.29) (0.29) MA (0.39) (0.35) Professional (0.24) (0.19) Notes: Standard deviations are reported in parentheses. SLID cross-sectional weights are applied to the analysis. Hourly wage is the log hourly wage that is converted to 1993 constant dollars. Experience is measured using the number of years worked at all jobs (part-time and full-time) since the first full-time paid job. Union includes individuals who are union members or covered by collective agreement. Certificates includes individuals with community college degrees that are below four-year university degrees. Below MA includes individuals who attended Master s programs but did not complete the degree, and BA graduates who completed additional courses for jobs such as accountants and teachers. See text for details.

56 47 have found that there is a positive premium for working in the public sector and being a union member. This indicates that university-educated women have a smaller average wage gap than other women partly because a greater proportion of university-educated women work in the public sector and are union members. For the degree/diploma attainment, Table 2.3 shows that women outperform men in achieving college certificates and four-year university degrees; however, they still fall behind men in completing graduate degrees. The finding that women are underrepresented among workers with post-graduate education is particularly important in explaining the gender gap for high-wage earners, since workers with higher education are more likely located on the upper-part of wage distribution. Table 2.4 presents the summary statistics of occupational characteristics. It is clear that women work in jobs that have less competition and that women are more likely to work with female coworkers than men. Furthermore, the percentage of female workers in the occupations of university-educated men is 43%, about 20 percentage points greater than that for other educational groups, suggesting that compared to women in other educational groups, university-educated women are more likely to opt out of female work, e.g., administrative occupations, and to participate in male-dominated/gender-integrated occupations, e.g. occupations in law This finding is consistent with Blau et al. (2013) and Fortin and Huberman (2002), where they documented that since 1970 the proportion of women working in jobs requiring postsecondary education, which used to be male-dominated, has increased at the expense of fewer women working in clerical jobs. But among low-educated workers, there is little change in female employment.

57 48 Table 2.4: Summary Statistics: Occupational Characteristics for Full-time Employees HSD HS College University M F M F M F M F Aptitude Skills General Learning (0.19) (0.20) (0.22) (0.22) (0.24) (0.23) (0.27) (0.22) Verbal (0.15) (0.15) (0.18) (0.17) (0.19) (0.17) (0.18) (0.16) Numerical (0.17) (0.19) (0.20) (0.21) (0.21) (0.20) (0.23) (0.19) Clerical Perception (0.17) (0.21) (0.19) (0.22) (0.19) (0.20) (0.19) (0.15) Spatial Perception (0.20) (0.12) (0.21) (0.14) (0.25) (0.17) (0.30) (0.21) Form Perception (0.18) (0.15) (0.19) (0.16) (0.22) (0.18) (0.24) (0.21) Motor Coordination (0.17) (0.18) (0.19) (0.18) (0.20) (0.18) (0.15) (0.14) Finger Dexterity (0.14) (0.17) (0.16) (0.18) (0.20) (0.18) (0.15) (0.16) Manual Dexterity (0.15) (0.18) (0.18) (0.18) (0.20) (0.17) (0.15) (0.14) O*Net Characteristics Workplace Competitiveness (0.10) (0.10) (0.10) (0.11) (0.11) (0.13) (0.11) (0.13) Ranking of Managerial Positions Not a Manager (0.31) (0.31) (0.40) (0.39) (0.43) (0.40) (0.49) (0.44) Junior Manager (0.26) (0.27) (0.34) (0.34) (0.37) (0.36) (0.43) (0.40) Senior Manager (0.19) (0.18) (0.24) (0.21) (0.26) (0.21) (0.38) (0.26) % of Women (0.23) (0.26) (0.25) (0.25) (0.26) (0.25) (0.24) (0.22) Notes: Standard deviations are reported in parentheses. denotes that the requirement is significantly greater for men (women) than for women (men) at 10% level, at 5% level, and at 1% level. Aptitude Skills are extracted from the Career Handbook. An aptitude skill reflects the requirement for that skill in occupations that are coded with National Occupational Classification for Statistics 2006 (four-digit NOC). Aptitude skills in this table are normalized between 0 and 1. Workplace Competitiveness is extracted from the O*Net database. O*Net occupations are matched to a four-digit NOC in the SLID. When there is more than one O*Net occupation for an SLID occupation, the characteristic is weighted by the fraction of workers in each of the O*Net occupations that comprise a single SLID occupation. Each of the O*Net characteristics has a score between 0 and 1 (inclusive). Ranking of Managerial Positions is extracted from the SLID. Junior Manager is 1 if one takes a middle or lower management position; 0 otherwise. Senior Manager is 1 if one takes a upper or top management position; 0 otherwise. % of Women is the fraction of female employees at each of the four-digit NOC by gender and educational group.

58 49 Occupations Table 2.5: Percentage of Full-time Employees in Each Occupation HSD HS College University Management Occupations M Business, Finance and Administrative F F F Occupations Natural and Applied Sciences and Related M Occupations Health Occupations Occupations in Social Science, Education, F Government Service and Religion Occupations in Art, Culture, Recreation and Sport Sales and Service Occupations F F F 8.89 Trades, Transport and Equipment Operators M M M 1.80 Occupations Unique to Primary Industry Occupations Unique to Processing, M M Manufacturing and Utilities Duncan Segregation Index Notes: Author s calculation of the percentage of full-time employees in each of the ten occupational categories within educational groups. An occupation with a superscript F (M) is considered as a female(male)-dominated occupation, where the proportion of female (male) workers in the occupation for an educational group exceeds 60% of the workforce in the occupation for that educational group. In this context, the Duncan Index is a demographic measure of the evenness with which two genders that belong to the same educational group are distributed across the ten occupations. It is between 0 and 100. When it is 0, the proportion of female workers equal to the proportion of male workers in any of the ten occupations (no segregation). When it is 100, men work in some occupations while women work in other occupations (complete segregation).

59 50 Women are less likely to take managerial positions. This is particularly clear for workers in the university group for whom the proportion of women taking junior and senior managerial positions is 5.6 and 9.5 percentage points, respectively, lower than the 24.6% and the 17% of men taking junior and senior managerial positions, respectively. The corresponding figures are 2 and 2.5 percentage points lower for women in the college group taking junior and senior managerial positions, and 0.2 and 1.6 percentage points lower for women in the high school group taking junior and senior managerial positions. For workers in the HSD group, close to 89% of men and women do not take managerial positions. There is no significant gender difference in the proportion of workers taking managerial positions for the very low-educated workers. Men and women are required to have different aptitude skills. Women are required to have higher levels of verbal, numerical, and clerical abilities, as well as finger dexterity, while men are required to have higher levels of spatial perception, form perception, motor coordination, and manual dexterity. This is found for employees with education below university. In the university-educated group, women are required to have a higher level of finger dexterity, while men are required to have higher levels of general ability, numerical, and spatial perception. Workers in different educational groups are required to have different aptitude skills. Very low-educated workers take jobs that require high levels of motor coordination, finger dexterity, and manual dexterity, while university-educated workers take jobs that require high levels of general learning, verbal, numerical abilities, and clerical perception.

60 51 In Table 2.5, I present the occupational distribution by education. While three quarters of workers in the HSD group work in sales and service, trades, and manufacturing occupations, the same fraction of workers with university education work in management, business and finance, natural science, and social science. Workers with high school or college education are mostly hired in administrative jobs, sales and service, and trades. On top of that, 13% of high school graduates work in manufacturing. Individuals and aptitude skills are linked through individuals occupations. Since people with different levels of education work in different occupations, aptitude skill requirements are different across educational group. The last row in Table 2.5 represents the Duncan segregation index, a measurement of occupational segregation. The index is computed as, S = 1 2 j=1 M j F j, where M j and F j are the proportion of male and female workers in job j, respectively. It measures the proportion of women (men) who would have to change occupations to obtain equal distribution of occupations between men and women. The measure is between 0 and 100, with 0 indicating no segregation and 100 indicating complete segregation. The Duncan index falls over the four educational groups, but the clear drop appears only when one examines the segregation for university-educated workers. For those without university education, the Duncan index is close to 60%. For those with university education, it drops to 40%. This means that university-educated men and women are more likely to work in similar occupations than other workers. This is consistent with the finding in Table 2.4 that university-educated men working in occupations with a greater average

61 52 Table 2.6: Gender Differences in Managerial Responsibilities Upper Level Management Budget Promotion Supervising Highest-paid in University Sample (0.028) (0.03) (0.023) (0.03) FT Employees in University Sample (0.005) (0.008) (0.013) (0.01) Notes: In this table and the following tables, standard errors are reported in parentheses. denotes that the coefficient is significantly different from zero at 10% level, at 5% level, and at 1% level. The SLID provides information on a person s managerial duties. This table reports Probit (marginal) estimates of gender differences in the probability of taking the managerial duties. A negative value for a managerial duty means that compared to men, women are less likely to be responsible for the duty. Probit model controls for ten occupations, work experience, union status, sector of employment, marital status, major fields of study, residential provinces, survey year, whether one is an immigrant, and whether one is handicapped. Highest-paid in University Sample represents the numbers that are estimated using the full-time universityeducated employees whose wage is above the 90 th percentile of the wage distribution. FT employees in University Sample represents the numbers estimated using all full-time employees in university group. A managerial duty is coded as a binary variable, 1 if one takes the duty, 0 otherwise. There are five of such duties: Upper Level Management: Whether one takes a upper/top level management position. Budget: Whether one has an influence on budget or staffing. Promotion: Whether one has an influence on pay raise or promotion. Supervising: whether one s job involves supervising employees. percentage of females than other men. Lastly, I examine the likelihood of university-educated women taking managerial responsibilities, relative to their male counterparts, and compare this with the likelihood of the top 10% of the wage earners of university-educated women, relative to the top 10% of university-educated men. Table 2.6 demonstrates that while compared to men, women on average have a lower probability of being responsible for upper level management, supervising coworkers, determining coworkers promotion or pay raise, and planning a company s budget, these differences are even larger between highest-paid men and women than the differences between average men and women in the university group. In particular, highestpaid men are 12% more likely to work in upper-level managerial positions than highest-paid

62 53 Figure 2.1: The Gender Gap at Various Points of the Wage Distribution Notes: Author s calculations of the logarithm of male-female wage ratio at each decile of the distribution. The curve, FT, connects the wage gap at each decile of the wage distribution for the full-time employees. Other four curves plot the wage gap along the wage distribution by educational groups: HSD for the full-time employees whose education is below high school, HS for the full-time employees who graduated from high school, College for the full-time employees who attended/completed some post-secondary education, and University for the full-time employees who completed university degrees. women. The corresponding figure is 5% when I use the entire university group. This implies that the fact that women are underrepresented in the managerial positions is more important in accounting for the gender gap at the 90 th percentile of the wage distribution than the average gender gap, which is supported by the decomposition results The Gender Gap Across the Wage Distribution Figure 2.1 plots the gender gap at each decile of the distribution, where the solid curve is the gender gap for the entire full-time sample and the other four curves plot the gender gap by educational group. For example, the solid curve tells us that the gender gap at the

63 54 30 th percentile is approximately 22%. This means that at the 30 th percentile of the wage distribution men earn approximately 25 cents more than women for every dollar earned. 17 This figure shows that the pattern of the gender gap is strikingly different across educational groups, in particular between the HSD group and the university group. Contrary to the HSD group for which the wage gap curve displays an inverse U-shape, the gender gap for the university group increases throughout the wage distribution and the increase accelerates above the 80 th percentile. The gender wage gap changes from 12% at the bottom of the wage distribution to 21% at the top of the wage distribution, increasing by nine percentage points. This indicates the existence of a glass ceiling phenomenon: women on the upper-tail of the wage distribution experience larger wage gaps than women on the lower-tail of the wage distribution. For the HS and college groups, the gender gap displays a small variation along the wage distribution. As workers with high school and community college education compose 65% of the sample, the gender gap for the entire sample displays a slightly declining trend along the wage distribution, which hides the glass ceiling phenomenon because it only exists for university-educated women. Another way to observe the existence of glass ceiling phenomenon is to examine the underrepresentation of women among high-paid workers. In Table 2.7, I present the proportion of workers by educational group at different parts of the wage distribution for full-time workers in Panel A, and the proportion of workers by gender in Panel B. Panel A shows that low-educated workers and workers in the college group compose more than 50% of 17 Log-wage differentials reported throughout the paper are used as an approximation to percentage differences. The exact percentages can be obtained as the exponential of the log differential minus 1.

64 55 Table 2.7: The Proportion of Workers at Different Parts of the Wage Distribution Below 1 st 5 th 10 th 25 th 50 th 75 th 90 th 99 th Above 99 th Panel A: the proportion of workers by education at each part of the wage distribution HSD HS College University Panel B: the proportion of workers by gender at each part of the wage distribution Men Women Notes: The wage distribution for all full-time workers is divided into nine parts: (1) below the 1 st percentage (inclusive) of the wage distribution, (2) between the 1 st (exclusive) and the 5 th percentile (inclusive) of the wage distribution, (3) between the 5 th (exclusive) and the 10 th percentile (inclusive) of the wage distribution, (4) between the 10 th (exclusive) and the 25 th percentile (inclusive) of the wage distribution, (5) between the 25 th (exclusive) and the 50 th percentile (inclusive) of the wage distribution, (6) between the 50 th (exclusive) and the 75 th percentile (inclusive) of the wage distribution, (7) between the 75 th (exclusive) and the 90 th percentile (inclusive) of the wage distribution, (8) between the 90 th (exclusive) and the 99 th percentile (inclusive) of the wage distribution, (9) and above the 99 th percentile of the wage distribution. Panel A reports the proportion of full-time workers at each part of the wage distribution that belong to one of the four educational groups. Panel B reports the proportion of full-time men and women at each part of the wage distribution. the workers below the 90 th percentile of the wage distribution, whereas university-educated workers compose more than half of the workers above the 90 th percentile of the wage distribution. Panel B shows that while more than half of the workers below the median are women, the proportion of women drops substantially to 22% among the top 1% of wage earners. This suggests that achievement in university education would help women get into the high-paying occupations; however, it does not change the fact that women are underrepresented among the highest-paid wage earners. Furthermore, when I restrict my sample to the workers who are above the 90 th percentile of the wage distribution for full-time workers, I find that the women, who make it to the top 10% of wage earners, earn statistically the same as their male counterparts. 18 This means 18 To do this, I pool men and women in one sample and use the workers whose log hourly wage is above the 90 th percentile of the wage distribution for the pooled sample. Women in this high-wage group earn about 1%

65 56 that Canada s glass ceiling exists not because of the gender gap among the highest-paid individuals of all workers, but because of the highest-paid women of female workers earning considerably less than the highest-paid men of male workers. Therefore, I examine the explanations for the glass ceiling phenomenon by comparing male versus female labour market characteristics and detailed occupational characteristics at top points of the wage distribution. 2.3 Empirical Results Quantile Decomposition Method Using the regression-based decomposition approach developed by Firpo et al. (2009) (RIF-regression-based decomposition, hereafter), I estimate how much of the gender gap at a decile of the wage distribution is explained by gender differences in labour market characteristics (which is called composition effect ) and how much of the gender gap is explained by gender differences in returns to labour market characteristics (which is called wage structure effect ). A challenge of decomposing differences between the wage distribution for men and women is that the average derivative of the distribution of explanatory variables with respect to a covariate at a quantile of the distribution differs from the average derivative of the unconditional wage distribution with respect to the covariate at that quantile. Firpo less than men and the gender gap is statistically no different than zero.

66 57 et al. (2009) resolve this problem by estimating a gender-specific wage function with the recentered influence function (RIF-regression). The coefficients estimated with RIFregression at a quantile of the wage distribution correspond to the marginal effects of the covariates on the unconditional quantile of the wage distribution. Using RIF-regression estimates, the unconditional decomposition method decomposes the wage gap into the composition effect and wage structure effect at various points of the wage distribution as if it were decomposing the wage gap at the mean. Let ˆγ g,v be the vector of the coefficients of the RIF-regression for group g at the v th percentile of wage distribution. As shown in Fortin et al. (2011), the overall wage gap at the v th percentile of wage distribution, ˆδ O, v can be decomposed with the RIF-regression-based decomposition in the same way as for the wage gap at the mean, where the counterfactual wage function is based on men s covariates as the reference covariates and the coefficients in women s wage regression as the reference coefficients. 19 K K ˆδ O v = (ˆγ m0,v ˆγ w0,v ) + X mk (ˆγ mk,v ˆγ wk,v ) + (X mk X wk )ˆγ wk,v k=1 k=1 = ˆδv S + ˆδv X (2.1) 19 In lay terms, the counterfactual wage function estimates what hourly wage women would have earned if they had the observed characteristics of men and their wage function remained unchanged.

67 58 where ˆδ v S is the wage structure effect at the v th percentile, K ˆδ S v = (ˆγ m0,v ˆγ w0,v ) + X mk (ˆγ mk,v ˆγ wk,v ), k=1 and ˆδ v X is the composition effect at the v th percentile, K ˆδ X v = (X mk X wk )ˆγ wk,v. k=1 Take work experience as an example. The composition effect of work experience at the median is estimated by weighting the difference in the average number of years worked between men and women with the coefficient of work experience for women at the median. The wage structure effect is estimated by weighting the gender difference in coefficient at the median of the wage distribution with the average number of years worked for men. If the gender gap is fully explained by different labor market characteristics between men and women, we would conclude that there is no unfair discrimination against women. Put in another way, gender-pay-equity legislation addresses the gender gap that is not explained by the composition effect Explained and Unexplained Proportion of Gender Gap In Table 2.8, I present the independent variables in the decomposition analysis. Model 1 uses demographic characteristics (e.g. immigration status, marital status, etc), work experi-

68 59 Table 2.8: Variables Used in Decomposition Analysis Variables HSD and HS College and University Dependent Variable Logarithm of Hourly Wage Logarithm of Hourly Wage Independent Variables Model 1 (M1) immigrant, with disability, marital status, # of children, province, experience, union, public sector, year immigrant, with disability, marital status, # of children, province, experience, union, public sector, year fields of study, education Model 2 M1 + occupational dummies M1 + occupational dummies + industry + industry Model 3 M1 + occupational attributes Model 1 + occupational attributes + % of women + industry + % of women + industry Notes: An explanation of variable constructions is provided in the section of Data. Model 2 controls for 10 occupational dummy variables. The ten occupational categories are presented in Table 2.5. Occupational attributes includes aptitude skills, workplace competitiveness, and the ranking of an individual s managerial position. For the college group, the variable of education accounts for whether one has completed a certificate from a post-secondary educational institution. For the university group, variables of education account for whether one has completed a four-year university degree, a Master s degree, or a more advanced degree.

69 60 ence, union status, and the sector of employment. On top of that, Model 2 adds occupational dummy variables and industry, while Model 3 controls for occupational attributes, the percentage of female workers, and industry. Using different specifications, I estimate the fraction of gender gap that is explained by the composition effect (fraction explained) and the fraction of gender gap that is explained by the wage structure effect (fraction unexplained) at various points of the wage distribution for each of the four educational groups. The comparison between Models 2 and 3 reveals how much of the gender gap is explained when I replace the occupational dummy variables, as commonly used in the literature, with detailed occupational characteristics. Table 2.9 presents the results. Relative to Model 1, the inclusion of occupation-attributes and the percentage of female workers (Model 3) makes a larger contribution to the wage gap than the use of occupational binary variables (Model 2). More importantly, while more than 50% of the wage gap on the upper-tail of the wage distribution is due to the wage structure effect when I use Model 2, the wage structure effect no longer plays a primary role in explaining the wage gap for most of the workers on the upper-tail of the wage distribution when I use Model 3. For example, for high-school dropouts at the 90 th percentile of the wage distribution, 61.1% of the wage gap is explained by composition effect, 12.3 percentage points greater than the proportion of 48.8% when I use Model 2. This suggests that gender differences in detailed occupational characteristics are important in explaining the wage gap, in particular for workers who earn more than 50% of the people in their gender-education group. It is of interest to see that the glass ceiling phenomenon for university-educated

70 61 Table 2.9: Explained and Unexplained Proportion of Gender Wage Gap 10 th 30 th 50 th 70 th 90 th 95 th High-school Dropouts (HSD) log hourly wage gap Model 1: Fraction explained 11.3% 17.4% 22.3% 23.75% 30.5% 40.0% Fraction unexplained 89.7% 82.6% 77,7% 76.25% 69.5% 60.0% Model 2: Fraction explained 20.4% 27.6% 29.6% 25.8% 48.8% 56.0% Fraction unexplained 79.6% 72.4% 70.4% 74.2% 51.2% 44.0% Model 3: Fraction explained 27.9% 33.5% 42.7% 41.7% 61.1% 64.6% Fraction unexplained 72.1% 66.5% 57.3% 58.3% 38.9% 35.4% High School Graduates (HS) log hourly wage gap Model 1: Fraction explained 17.1% 23.7% 19.2% 17.4% 21.9% 11.0% Fraction unexplained 82.9% 76.3% 80.8% 82.6% 79.1% 89.0% Model 2: Fraction explained 5.8% -5.1% 4.5% 15.3% 28.6% 20.6% Fraction unexplained 94.2% 105.1% 95.5% 84.7% 71.4% 79.4% Model 3: Fraction explained 22.5% 22.9% 16.5% 28.5% 47.7% 36.4% Fraction unexplained 77.5% 77.1% 83.5% 71.5% 52.3% 63.6% Community College Graduates (College) log hourly wage gap Model 1: Fraction explained 14.6% 9.6% 7.7% 9.06% 15% 18.6% Fraction unexplained 85.4% 90.4% 92.3% 90.94% 85% 81.4% Model 2: Fraction explained -18.1% 6.04% 5.4% 15.9% 29.0% 33.0% Fraction unexplained 118.1% 93.96% 94.6% 84.1% 71.0% 67.0% Model 3: Fraction explained 13.0% 9.8% 21.8% 31.0% 55.0% 53.1% Fraction unexplained 87.0% 91.2% 78.2% 69.0% 45.0% 46.9% University Graduates (University) log hourly wage gap Model 1: Fraction explained -32.2% -36.1% 8.1% 24.0% 28.0% 34.8% Fraction unexplained 132.2% 136.1% 91.9% 76.0% 72.0% 65.2% Model 2: Fraction explained -26.5% 23.3% 32.0% 45.9% 43.9% 48.3% Fraction unexplained 126.5% 76.7% 68.0% 54.1% 56.1% 52.7% Model 3: Fraction explained 24.0% 37.0% 42.8% 55.7% 63.8% 72.6% Fraction unexplained 76.0% 63.0% 57.2% 44.3% 36.2% 27.4% Notes: See Table 2.8 for the explanation of three specifications.

71 62 women is mostly due to high-paid university-educated women having different occupational characteristics than their male counterparts. The raw gap at the 95 th percentile of the wage distribution is 23%, 11 percentage points larger than the gender gap at the 10 th percentile of the wage distribution for university-educated workers. After accounting for differences in labor market characteristics in Model 3, I find that the corresponding figures drop to 5.8% at the 95 th percentile and 9.5% at the 10 th percentile of the wage distribution. This means that once gender differences in occupational characteristics are taken into account, high-paid women with university education do not experience a greater wage gap than low-paid women. On the contrary, when I use Model 2, which adopts occupational binary variables, the unexplained gender gap is 11.3% at the 95 th percentile and 12.4% at the 10 th percentile of the wage distribution. This shows that the use of occupational binary variables is not sufficient to explain the glass ceiling for university-educated women. A smaller explained proportion of the gender gap in Model 2 than in Model 3 arises from aggregated occupational categories that are used to construct occupational binary variables. Take two jobs as an example. Financial Managers and Restaurant Managers belong to the same occupational category (Management Occupation), but the former on average are paid significantly more than the latter. Financial managers are required to have more numerical ability and to work in a more competitive environment than restaurant managers. This example demonstrates how the use of occupational characteristics can reveal the impact of such differences on the pay gap between the two jobs, whereas the use of occupational binary variables cannot reveal it. Accounting for different occupational characteristics between

72 63 men and women is important in determining whether women are treated unfairly in the workplace; since such differences are suppressed by aggregated occupational categories, detailed occupational characteristics are needed in the analysis Gender Differences in Work Experience, Union, Sector, Degree Attainment, and Fields of Study While Table 2.9 presents evidence supporting the hypothesis that the inclusion of detailed occupational characteristics is important in explaining the wage gap for all four of the educational groups, it is also useful to know how gender differences in each covariate contribute to the gender gap. For this purpose, I present the contribution of work experience, union, sector, education, and detailed occupational characteristics in Tables 2.10 for HSD and HS groups and Table 2.11 for college and university groups. 20 I will start with the composition effect of work experience, union, sector and education. Gender differences in work experience make a positive and significant contribution to the wage gap in each of the four educational groups, meaning that men work more than women at all educational levels. Gender differences in union coverage are positive for all educational groups except the university group for whom the differences are negative, because for university-educated women, they are more likely to be unionized than men, while for other educational groups, women are less likely to be union members. Women 20 Decomposition results in Tables 2.10 and 2.11 are estimated using Model 3. To save space, I do not report the estimates for survey years, marital status, the number of children, age groups, residential provinces, whether one is an immigrant, and whether one is handicapped. Full results are available upon request.

73 64 are more likely to work in the public sector than men, but this difference is very small for low-educated workers compared to that for high-educated workers. Men and women are equally likely to complete a college/trade program, but the proportion of women completing university degrees is lower than men above the 30 th percentile of the wage distribution. One explanation is that women fall behind men in the completion of Master s and professional degrees (degrees in medicine and Doctorate). Since workers with more education are more likely located on the upper-tail of the wage distribution, fewer women having post-graduate education than men is particularly important for the high-wage earners. Table 2.11 shows that the wage gap would be diminished by 8.7% (0.02/0.23) at the 95 th percentile of the university-educated workers if the proportion of women completing graduate degrees were the same as that of men. Gender differences in the fields of study play a small role in accounting for the gender gap. 21 The small impact is due to the finding that there is a considerable variation in fields of study across genders. Men are more likely to graduate from architecture, engineering and applied sciences while women are more likely to graduate from health and education. Larger gender differences in architecture, engineering and applied sciences offset smaller differences in health and education. Thus, differences in the fields of study, which are the weighted sum of the difference in each field of study, are small This study is not the only study that finds that major fields of study do not appear as important as jobrelated attributes.drolet (2002a) uses the SLID 1997 and finds that while gender differences in actual work experience explain up to 50% of the gender gap, only 5% of the gender gap at the mean is explained by gender differences in major fields of study. 22 In Appendix B, I report the decomposition results when only covariates in Model 1 are included in the

74 Table 2.10: The Contribution of Subsets of Covariates for the Low-Educated Workers 10 th 30 th 50 th 70 th 90 th 95 th HSD Composition Wage Composition Wage Composition Wage Composition Wage Composition Wage Composition Wage Effects Structure Effects Structure Effects Structure Effects Structure Effects Structure Effects Structure Effects Effects Effects Effects Effects Effects Effects Total (0.12) (0.014) (0.015) (0.017) (0.2) (0.2) (0.025) (0.027) (0.038) (0.044) (0.04) (0.05) Accounted for by: Union (0.002) (0.007) (0.003) (0.010) (0.005) (0.01) (0.005) (0.01) (0.006) (0.02) (0.007) (0.02) Public (0.0005) (0.004) (0.0006) (0.003) (0.001) (0.004) (0.002) (0.006) (0.002) (0.008) (0.001) (0.008) Experience (0.004) (0.058) (0.005) (0.05) (0.005) (0.04) (0.007) (0.04) (0.01) (0.05) (0.012) (0.05) Aptitude skills (0.007) (0.07) (0.01) (0.07) (0.01) (0.08) (0.02) (0.11) (0.03) (0.20) (0.02) (0.23) Competitiveness (0.005) (0.06) (0.005) (0.06) (0.008) (0.07) (0.01) (0.12) (0.008) (0.15) (0.02) (0.15) Managerial ranking (0.0003) (0.02) (0.0006) (0.02) (0.008) (0.07) (0.001) (0.03) (0.002) (0.04) (0.003) (0.06) % of women (0.012) (0.013) (0.016) (0.013) (0.02) (0.02) (0.02) (0.015) (0.04) (0.02) (0.08) (0.03) Industry (0.008) (0.02) (0.01) (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.03) (0.02) (0.04) 65 HS Composition Wage Composition Wage Composition Wage Composition Wage Composition Wage Composition Wage Effects Structure Effects Structure Effects Structure Effects Structure Effects Structure Effects Structure Effects Effects Effects Effects Effects Effects Effects Total (0.014) (0.02) (0.02) (0.02) (0.015) (0.02) (0.017) (0.018) (0.02) (0.02) (0.03) (0.03) Accounted for by: Union (0.003) (0.009) (0.004) (0.01) (0.003) (0.01) (0.003) (0.01) (0.003) (0.02) (0.004) (0.016) Public (0.0006) (0.004) (0.001) (0.005) (0.002) (0.005) (0.002) (0.006) (0.002) (0.008) (0.003) (0.01) Experience (0.003) (0.08) (0.004) (0.06) (0.003) (0.04) (0.004) (0.04) (0.006) (0.046) (0.008) (0.06) Aptitude skills (0.01) (0.11) (0.01) (0.09) (0.01) (0.07) (0.01) (0.09) (0.02) (0.12) (0.02) (0.13) Competitiveness (0.003) (0.06) (0.003) (0.06) (0.003) (0.05) (0.003) (0.06) (0.005) (0.08) (0.006) (0.10) Managerial ranking (0.0005) (0.015) (0.001) (0.14) (0.0008) (0.01) (0.001) (0.01) (0.002) (0.02) (0.003) (0.04) % of women (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) (0.03) (0.02) (0.03) (0.02) Industry (0.006) (0.01) (0.008) (0.01) (0.008) (0.01) (0.01) (0.01) (0.01) (0.02) (0.014) (0.024) Notes: Estimation is conducted with the RIF-regression-based decomposition method and uses full-time salaried employees aged who did not attend post-secondary education. Numbers in bold are significantly different from 0 at 10% level. Covariates in Model 3 are included in the estimation. Variables that are not reported are survey years, marital status, the number of children, age groups, residential provinces, whether one is an immigrant, and whether one is handicapped.

75 Table 2.11: The Contribution of Subsets of Covariates for the High-Educated Workers 10 th 30 th 50 th 70 th 90 th 95 th College Composition Wage Composition Wage Composition Wage Composition Wage Composition Wage Composition Wage Effects Structure Effects Structure Effects Structure Effects Structure Effects Structure Effects Structure Effects Effects Effects Effects Effects Effects Effects Total (0.012) (0.014) (0.012) (0.012) (0.01) (0.01) (0.01) (0.13) (0.02) (0.02) (0.02) (0.03) Accounted for by: Education (0.0003) (0.02) (0.0003) (0.014) (0.0002) (0.011) (0.0002) (0.01) (0.0004) (0.017) (0.0004) (0.02) Major (0.008) (0.02) (0.007) (0.016) (0.007) (0.01) (0.007) (0.01) (0.01) (0.02) (0.02) (0.02) Union (0.001) (0.007) (0.001) (0.006) (0.001) (0.005) (0.0007) (0.005) (0.0006) (0.007) (0.0007) (0.008) Public (0.001) (0.004) (0.001) (0.004) (0.001) (0.004) (0.002) (0.005) (0.002) (0.006) (0.003) (0.007) Experience (0.002) (0.056) (0.002) (0.03) (0.002) (0.024) (0.002) (0.02) (0.003) (0.03) (0.003) (0.03) Aptitude Skills (0.01) (0.06) (0.007) (0.05) (0.006) (0.04) (0.008) (0.06) (0.01) (0.07) (0.01) (0.07) Competitiveness (0.003) (0.04) (0.003) (0.035) (0.002) (0.03) (0.004) (0.05) (0.004) (0.05) (0.004) (0.06) Managerial ranking (0.0007) (0.001) (0.0007) (0.007) (0.001) (0.01) (0.0008) (0.008) (0.001) (0.008) (0.002) (0.02) % of women (0.02) (0.02) (0.01) (0.01) (0.01) (0.009) (0.01) (0.01) (0.02) (0.01) (0.02) (0.01) Industry (0.01) (0.006) (0.005) (0.006) (0.004) (0.005) (0.005) (0.006) (0.008) (0.009) (0.01) (0.07) 66 University Composition Wage Composition Wage Composition Wage Composition Wage Composition Wage Composition Wage Effects Structure Effects Structure Effects Structure Effects Structure Effects Structure Effects Structure Effects Effects Effects Effects Effects Effects Effects Total (0.02) (0.02) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.02) (0.02) (0.03) Accounted for by: Education (0.003) (0.02) (0.001) (0.01) (0.001) (0.01) (0.002) (0.01) (0.003) (0.02) (0.004) (0.02) Major (0.01) (0.03) (0.007) (0.01) (0.006) (0.01) (0.006) (0.01) (0.007) (0.01) (0.01) (0.01) Union (0.006) (0.01) (0.004) (0.008) (0.003) (0.006) (0.002) (0.007) (0.003) (0.008) (0.004) (0.01) Public (0.006) (0.02) (0.005) (0.01) (0.004) (0.01) (0.004) (0.01) (0.004) (0.02) (0.006) (0.02) Experience (0.006) (0.1) (0.004) (0.05) (0.004) (0.03) (0.004) (0.03) (0.004) (0.03) (0.005) (0.04) Aptitude Skills (0.01) (0.18) (0.008) (0.10) (0.007) (0.076) (0.007) (0.08) (0.009) (0.1) (0.01) (0.13) Competitiveness (0.009) (0.11) (0.005) (0.07) (0.004) (0.05) (0.004) (0.05) (0.004) (0.07) (0.006) (0.09) Managerial ranking (0.005) (0.01) (0.003) (0.006) (0.003) (0.005) (0.003) (0.006) (0.005) (0.01) (0.008) (0.015) % of women (0.02) (0.06) (0.01) (0.03) (0.01) (0.02) (0.01) (0.03) (0.01) (0.03) (0.015) (0.04) Industry (0.01) (0.03) (0.007) (0.01) (0.006) (0.01) (0.006) (0.01) (0.007) (0.02) (0.01) (0.02) Notes: Estimation is conducted with the RIF-regression-based decomposition method and uses full-time salaried employees aged who completed post-secondary education. Numbers in bold are significantly different from 0 at 10% level. Covariates in Model 3 are included in the estimation. Variables that are not reported are survey years, marital status, the number of children, age groups, residential provinces, whether one is an immigrant, and whether one is handicapped.

76 Gender Differences in Occupational Characteristics Even though the inclusion of occupational characteristics and industry increases the explained fraction substantially for high school dropouts (Table 2.9), Table 2.10 tells us that it is mostly because very low-educated men and women work in different occupations. The composition effect of aptitude skills, workplace competitiveness, and the ranking of managerial positions is not statistically different from 0 along the wage distribution. 23 In contrast, Tables 2.10 and 2.11 show that gender differences in aptitude skills play a significant role for high school and community college graduates and for most of the university graduates, and gender differences in workplace competitiveness and the ranking of managerial positions are particularly important for high-educated workers on the upper-tail of the wage distribution. I present the fraction of the gender gap that is explained by the composition effect of aptitude skills, workplace competitiveness, the ranking of managerial positions, the percentage of women in the workplace, and industry in Tables 2.12 and 2.13 for loweducated and high-educated employees, respectively. I find that for high school dropouts, the fraction of the gender gap explained by differences in aptitude skills is very small, ranging between -5% and 2% below the 90 th percentile of the wage distribution. The composition effect of aptitude skills makes a slightly larger contribution above the 90 th percentile of the RIF-regression decomposition method. While composition effects of work experience, union coverage, sector of employment, achievement of post-secondary education degrees, and major fields of study are greater when I do not include detailed occupational characteristics and industry in the estimation, major findings are consistent with the findings in Table 2.10 and The only exception is at the 10 th percentile of the wage distribution, where the gender difference in aptitude skills is significantly different from 0 at 10% level.

77 68 wage distribution; however, this contribution is modest, relative to the faction explained by the percentage of women in the workplace. 24 This finding suggests that differences between male and female work at the low-end of educational distribution cannot be captured by gender differences in aptitude skill requirements. The extent to which the gender gap is attributable to gender differences in aptitude skills varies at different points of the wage distribution. Examples are the contribution of aptitude skills for community college graduates, which is -52.6% at the 10 th percent of the wage distribution and 46% at the 90 th percentile, and the contribution for high school graduates, which is -29% at the 10 th percentile and 8% at the 90 th percentile. A negative contribution on the lower-tail of the wage distribution means that women are required to have more aptitude skills than men. This finding arises from the fact that more than 60% of administrative positions (e.g. Administrative Clerks) are taken by women with high school or community college education and therefore women are required to have significantly more clerical skills than men. A positive contribution on the upper-tail of the wage distribution comes from men taking occupations that require more clerical and spatial perceptions. For the former, men are more likely to take upper-level administrative positions such as Executive Assistants, which require more clerical skills than office staff positions. For the latter, men are more likely to work as mechanics, computer programmers and engineers, which require a high 24 In the HSD group, the percentage of women in the workplace on average is significantly lower for men than for women. RIF regression results show that the percentage of women in the workplace is negatively correlated with the wage values along the wage distribution for women in the HSD group. Thus, the composition effect, which is the difference in the percentage of women at workplace between men and women (a negative value) multiplied by the (negative) coefficient on the percentage of women at the τ th percentile of the wage distribution, makes a positive contribution to the gender gap at the τ th percentile of the wage distribution.

78 69 level of spatial perception. As workers with different education levels are concentrated in different occupations, Tables 2.12 and 2.13 show that gender differences in specific types of skill play different roles, depending on the educational group of interest. Specifically, clerical perception is important in accounting for the wage gap along the wage distribution for workers in the high school and community college groups. Spatial perception, which is most needed in engineering and computer programming, is important in explaining the wage gap for community college and university graduates. General learning ability, which is most needed in managerial positions, is essential in explaining the wage gap for university-educated workers. Gender differences in aptitude skills explain a substantial proportion of the gender gap for university-educated workers below the 90 th percentile of the wage distribution, but not for the top 10% of wage earners. Table 2.13 shows that the contribution of aptitude skills is 9% of the gender gap at the 90 th percentile of the wage distribution for university-educated workers and 3.5% at the 95 th percentile of the wage distribution. This is trivial relative to the contribution for the university-educated workers below the 90 th percentile of the wage distribution, which ranges between 26% and 31%. Moreover, the contribution of almost all the aptitude skills above the 90 th percentile is negligible, with the exception of general learning ability, which explains about 8% of the gender gap. This finding implies that gender differences in managerial positions play an important role in explaining the wage gap for high-paid workers. However, not all levels of managerial ranking matter in explaining

79 70 the wage gap. It is gender differences in upper/top managerial positions that explain why high-paid women earn much less than high-paid men. Tables 2.12 and 2.13 tell us that gender differences in working at upper/top management levels have little impact in explaining the gender gap for workers without university education; however, for workers in the university group, these differences explain an increasing fraction of the gender gap as one moves from the bottom to the top of the wage distribution. There is 4% of the gender gap explained by the composition effect of senior managerial positions at the 10 th percentile of the wage distribution for university-educated workers, about 20 percentage points smaller than the fraction explained by aptitude skills. The fraction explained by working at senior managerial positions increases to 14% at the 95 th percentile of the wage distribution, 10 percentage points greater than the fraction explained by aptitude skills at the same point of the wage distribution. Turning to gender differences in working middle/lower management levels, I find that they play little role in explaining the gender gap for all of the four educational groups. Another factor that explains the gender gap for the highest-paid university-educated workers is workplace competitiveness. It explains 7% of the gender gap at the 90 th percentile and 5% at the 95 th percentile of the wage distribution, greater than the contribution of nine aptitude skills at the corresponding percentiles of the wage distribution. Thus, men taking (or being given) upper/top levels of managerial responsibility and working in a highly competitive environment are the major reasons that university-educated women are underrepresented among the top 10% of university-educated wage earners.

80 71 Similar to the composition effect of senior managerial positions, the composition effect of workplace competitiveness explains less than 1% of the gender gap for low-educated workers and 1% - 3% of the gender gap for workers in the community college group. Workplace competitiveness and working at senior managerial positions are high-wage determinants. Thus, these two occupation attributes are critical in explaining the gender gap at the high-end of the wage (and education) distribution, but are not important for other workers. We have seen that gender differences in occupational attributes contribute to the wage gap for workers with high school education and post-secondary education, but they play a very small role for workers with education below high school. An explanation is that although men and women in the HSD group work in different occupations men working in blue-collar jobs and women working in sales and service jobs they are not required to have high levels of skills regardless of their gender and occupation. In Table 2.14, I present the average wage for lowest-educated workers in the occupations that hire 80% of the HSD sample. Service and sales occupation are female-dominated, with 60% of the workforce with education below high school being women. Occupations in trade and transport and occupations in manufacturing and utilities are male-dominated: female workers make up 37% of the lowest-educated workforce in trade and transport, and 7% in manufacturing and utilities. Occupations in trade and transport and occupations in manufacturing and utilities on average pay $14 (in 1993 constant dollars) and $12.50 and $3

81 72 Table 2.12: Fraction of Gender Gap Explained by Differences in Occupational Characteristics and Industry (%) for Low-educated Workers 10 th 30 th 50 th 70 th 90 th 95 th High-school Dropouts(HSD) Aptitude Skills General Learning Verbal Ability Numerical Ability Clerical Perception Spatial Perception Form Perception Motor Coordination Finger Dexterity Manual Dexterity Workplace Competitiveness Managerial Ranking Not a Manager Junior Manager Senior Manager % of Women Industry High School Graduates (HS) Aptitude Skills General Learning Verbal Ability Numerical Ability Clerical Perception Spatial Perception Form Perception Motor Coordination Finger Dexterity Manual Dexterity Workplace Competitiveness Managerial Ranking Not a Manager Junior Manager Senior Manager % of Women Industry Notes: Estimation is conducted with the RIF-regression-based decomposition method and uses full-time salaried employees aged who did not attend post-secondary education. Covariates in Model 3 are included in the estimation.

82 73 Table 2.13: Fraction of Gender Gap Explained by Differences in Occupational Characteristics and Industry (%) for High-educated Workers 10 th 30 th 50 th 70 th 90 th 95 th Community College Graduates (College) Aptitude Skills General Learning Verbal Ability Numerical Ability Clerical Perception Spatial Perception Form Perception Motor Coordination Finger Dexterity Manual Dexterity Workplace Competitiveness Managerial Ranking Not a Manager Junior Manager Senior Manager % of Women Industry University Graduates (University) Aptitude Skills General Learning Verbal Ability Numerical Ability Clerical Perception Spatial Perception Form Perception Motor Coordination Finger Dexterity Manual Dexterity Workplace Competitiveness Managerial Ranking Not a Manager Junior Manager Senior Manager % of Women Industry Notes: Estimation is conducted with RIF-regression-based decomposition method and uses full-time salaried employees aged who completed post-secondary education.

83 74 Table 2.14: O*Net Characteristics in Service, Trade and Manufacturing Occupations (1) (2) (3) Sales and Service Trades and Transport Manufacturing and Utilities Physical Activities Noise Contaminants Hazardous Equipment Indoors Without Environmental Control Very Hot or Cold Temperatures Wear Safety Equipment Attributes of Workers with Education Below High school % of Women Average Hourly Wage Notes: Three categories of occupations are Sales and Service occupations, Trades, Transport and Equipment Operators and Related occupations, and Occupations Unique to Processing, Manufacturing and Utilities. O*Net occupations are matched to National occupational code in the SLID. When there is more than one O*Net occupation for an SLID occupation, the characteristic is weighted by the fraction of workers in each of the O*Net occupations that comprise a single SLID occupation. Each of the O*Net characteristics has a score between 0 and 1 (inclusive). The questions of O*Net characteristics are listed in Table B1.

84 75 more than the average hourly wage in sales and service occupations, respectively. The pay differences between male-dominated and female-dominated occupations partly explain the wage gap for the lowest-educated workers. As explained in hedonic wage theory, one potential reason for higher average wages in male-dominated occupations is an unfavorable work environment in these occupations. Table 2.14 presents the average score of O*Net attributes that reflect the physical work conditions in the occupations. A higher score means worse work conditions. Male-dominated occupations have more physical activities and are more likely to involve exposure to contaminants and hazardous equipment. Workers in these occupations are required to have higher endurance for unpleasant environments, e.g., noise and very hot/cold temperatures in the workplace. To prevent injuries, trade and manufacturing occupations require workers to wear safety equipment much more often than sales and services occupations. Overall, this section presents evidence supporting the hypothesis that male-dominated and female-dominated occupations pay differently largely because they require different levels of aptitude skills and have different work environments such as physical work conditions and workplace competitiveness. Whether the pay difference arising from different occupational characteristics should be a public concern depends on the reasons as to why women do not work in male-dominated occupations (e.g. oil drilling or surgery) knowing that male jobs pay more than female jobs. If women experience barriers that prevent them from entering male-dominated occupations, the pay gap between male and female jobs concerns the pay-equity legislations. However, if women dislike working in unpleas-

85 76 ant work conditions or women prefer less demanding jobs so that they can devote more time to their family, then their occupational choices should be respected. Unfortunately, I cannot observe the characteristics such as innate ability and preferences that determine a person s occupation. Thus, I cannot conclude whether observed gaps in pay are due to unfair discrimination against women in Canada. Finally, evidence in this section suggests that the use of the unexplained gender gap as evidence for unfair discrimination is problematic. Table 2.9 shows that for high school and college graduates, the explained fraction of the gender gap on the lower-tail of the wage distribution is very small. One reason is that gender differences in aptitude skills make a negative contribution to the gender gap on the lower-tail of the wage distribution, while gender differences in the percentage of women in the workplace make a positive contribution to the gender gap at the same position of the distribution. They cancel each other out, which results in a small proportion of gender gap explained by gender differences in labor market characteristics. In fact, the negative contribution of aptitude skills is substantial, ranking between -20% and -50%. This suggests that for high school and community college graduates, the gender gap on the lower-tail of the wage distribution largely arises from men and women taking different jobs and their jobs requiring different aptitude skills. This example shows that the presence of an unexplained gender gap is not conclusive evidence for the presence of unfair discrimination against women in the labour market.

86 Accounting for Selection into Paid Work This section investigates how sample selection that is induced by non-employment affects the gender gap by educational group. This study uses alternative imputation techniques to recover missing wages along the wage distribution. Using the quantile regression method by Koenker and Bassett (1978), I estimate the wage gap by regressing observed and imputed wage values on a gender dummy at various points of the wage distribution. An advantage of the quantile estimator is that as long as the imputed wage value for an individual is at the same side of the v th percentile of the wage distribution as the actual wage if he/she were employed, the estimate of the wage gap at the v th percentile is unbiased. A proof is provided in Appendix A. I first exploit the panel nature of the SLID. For those not in work in a given year, t, the imputation procedure searches backward and forward to recover wage observations from the nearest wave, t, in the sample. In practice, it imputes y it for I it = 0 with w it when I it = 1. This imputation implicitly assumes that for an individual i, his/her latent wage position with respect to the v th percentile of the potential (gender-education-specific) wage distribution in the year t can be predicted by his/her wage in the nearest wave t when he/she was employed. 25 This approach is called imputation on unobservables. 26 Wage information is imputed with wage values from another wave, regardless of the reasons a person did not work in year 25 This assumption is addressed formally with the equation, F (w v D g,i, I it = 0) = F (w v D g,i, I it = 1). This equation is reasonable if one s wage position with respect to the v th percentile does not change when one s employment status changes between t and t. 26 Olivetti and Petrongolo (2008) used this approach to estimate the median gender gap.

87 78 t but worked in t. Hence, selection into work is based on the persons characteristics that are not observable to researchers. This imputation procedure can recover wage values for individuals who worked at least once during the 6-year sample period. The estimate of the gender gap at the v th percentile is unaffected by imputation whenever movements of one s wage position from t to t happen within either side of the v th percentile of the distributions. In order to recover wage observations for those who are never observed in work during the 6 years of longitudinal sample period, I develop an alternative approach, which is built upon the DFL procedure and reveals a slightly different economic mechanism of selection than the first approach. Specifically, for individuals within gender-education groups, I construct a hypothetical wage distribution for the non-employed by reweighting the wage distribution for the employed workers with labour market characteristics of the individuals who never worked during the survey period. As the imputation is based on the observed characteristics of the non-employed and the wage structure of the employed in a given year t, I call it imputation on observables. Mathematical explanations are provided in Appendix A. This imputation is implemented in two steps. In the first step, I split observations within gender-education groups into two samples: the employed sample and the non-employed sample. The employed sample includes both the full-time and the part-time employees. It weights the wage distribution of employed observations with the characteristics of the non-employed to construct a hypothetical wage distribution for each gender-education

88 79 group. 27 In the second step, I construct an imputed sample in which the employed have their observed wage and the non-employed have wage values that are drawn randomly from their gender-education hypothetical wage distributions. The statistic of interest is the gender wage gap, which is estimated with the imputed sample at each decile of the wage distribution for an educational group. The wage gap that corrects for sample selection is presented in Figure 2.2. It has four diagrams, each diagram representing one educational group. In each diagram, I plot the gender gap with a 95% confidence interval for the full-time and part-time employees combined, the gender gap that accounts for selection on the unobservables, and the gender gap that accounts for selection on the observables. The difference between the actual gender gap (the gender gap for full-time and part-time workers combined) and the potential gender gap (the gender gap that accounts for selection) at each decile of the wage distribution measures the impact of sample selection in estimating the gender gap for that educational group at that point on the wage distribution. 28 The gender wage gap responds more strongly to the adjustment of selection for loweducated groups because low-educated workers have higher employment gaps between men and women than high-educated workers. In particular, Figure 2.2 tells us that the wage gap along the wage distribution is largely unaffected for workers with college or university education; however, the potential wage gap is substantially different than the actual wage gap 27 The observed characteristics are work experience prior to being unemployed, marital status, the number of pre-school and school-aged children, parental education, age group, and immigration status. 28 Because sample selection is induced by individuals who were not employed, the comparison group is the wage gap for the entire sample of wage earners who are either full-time or part-time workers.

89 80 Figure 2.2: Gender Gap Correcting for Selection Notes: The sample consists of survey participants aged between and not enrolled in school. for low-educated workers. This quantitatively demonstrates that the inclusion of individuals who did not work would affect the gender gap significantly for individuals without postsecondary education, but not for other individuals. Among the low-educated individuals, correcting for selection on observables makes greater changes in the estimate of gender gaps than selection on unobservables. When imputing missing wage values using individuals observed characteristics, I include individuals who never worked during the longitudinal period. They had weaker labour market attachment than employed individuals and non-employed individuals who worked in some

90 81 years over a 6-year period. Thus, the gender wage gap is more affected when I correct for selection on observables. The largest adjustment occurs at the median of the wage distribution for the HSD group: if non-employed individuals with education below high school worked during the sample period, women at the median would have earned only 57 cents for every dollar paid to men. An interesting finding appears when I examine the selection-adjusted wage gaps for the HS group. The inclusion of the individuals who never worked reduces the wage gap substantially on the lower-tail of the wage distribution. This finding is not contrary to the assumption in studies that examine the impacts of sample selection: women are assumed to opt out of the labour market when they have low-wage characteristics, relative to the characteristics related to home production. Therefore, the imputed wage values would be expected to be lower than observed wage values for women and selection-adjusted gender gaps would be greater than observed gender gaps. I find that the non-employed men are less likely to be married and have fewer children than women. Since marital status and the number of children are used in constructing hypothetical wage distribution for non-working individuals, the non-employed men have lower productivity characteristics than the nonemployed women. Therefore, the selection-adjusted gender gap is smaller than the actual gender gap. Overall, Figure 2.2 suggests that there is heterogeneity in supply behaviour across educational groups. While alternative imputation approaches reveal different economic channels of selection, results with both imputation approaches confirm that correcting for

91 82 sample selection makes little difference in estimating the gender gap for high-educated workers. For low-educated workers, correction for sample selection on unobservables makes a lesser difference in the gender gap than selection on observables, suggesting that the use of unobserved characteristics is insufficient to capture the selection rule for individuals without post-secondary education. 2.4 Conclusion Previous studies have found that male jobs pay more than female jobs partly because female jobs require different DOT-skills than male jobs. This chapter contributes to the literature by demonstrating that the gender gap is explained in part by required DOT-skills, workplace competitiveness, and degree of managerial responsibility. For workers in the high school and community college groups, and for university-educated workers below the 90 th percentile of the wage distribution, men and women work in occupations that require different levels of clerical and spatial perceptions and general learning ability, which accounts for a substantial proportion of the gender gap. This is in line with the conclusion that male-dominated occupations such as auto mechanics pay more than female-dominated occupations such as secretarial work because these different occupations require different aptitudes/ skills (DOT-skills). However, skill requirements do not account for all of the gender gap, and additional factors at the two ends of the education distribution are very different. In the low-paid and

92 83 very low-educated worker groups, men are compensated more than women for working in unpleasant work conditions. Among the highest-paid university-educated workers, men are compensated more than women for taking upper-level managerial duties and working in a more competitive environment. For the latter group, we have seen that the wage structure effect (adjusted gender gap) decreases as one moves to the top of the wage distribution. This suggests that the glass ceiling phenomenon is explained by the finding that highest-paid men work in more demanding jobs than highest-paid women. A limitation of the study is that I cannot observe the characteristics that determine a person s occupation. Thus, I do not know whether women working in different occupations than men are doing so because of gender differences in abilities (e.g. men are better managers), outside options (i.e. spouse s wages), worker preferences, or unfair discrimination based on stereotypes. I examine how the selection into work affects the estimates of gender gap at various points of the wage distribution. Compared to existing methodologies, my study accounts for sample selection with less restrictive assumptions. It empirically demonstrates that correcting for sample selection increases the gender gap substantially for low-educated workers, but not for high-educated workers. This means that low-paid women who work are very different than those who do not. For better-educated women, for whom participation rates are much higher, selectivity into work is much less important. As the main analysis in the study does not account for sample selection, caution should be taken when interpreting the decomposition results for the low-educated workers.

93 84 In this chapter I discussed the potential reasons for the fact that men and women working in different occupations affects the gender wage gap. Future research could try to understand what accounts for the gender wage gap within occupations. While a number of studies have contributed to understanding the gender wage gap within high-skilled occupations, the gender wage gap within low-skilled occupations has not been subject to such comprehensive investigation. Further exploration of the gender gap for low-educated workers is of fundamental importance for policy makers to ensure gender pay equity in blue-collar occupations. Appendix Appendix A In what follows, I explain the idea underlying the imputation techniques used to recover missing wages at various points of the wage distribution. The variable of interest is the difference between (log) male and female wage at each decile of the distributions: δ v = v(w D m ) v(w D w ) (2.2) where v(.) is the wage function at the v th percentile of the distributions, v = (10, 20, 30, 40, 50, 60, 70, 80, 90). The (log) wage distribution for each gender is defined by

94 85 F (w D g ) =F (w D g, I = 1)P r(i = 1 D g )+ F (w D g, I = 0)[1 (P r(i = 0 D g ))], (2.3) where I is an indicator function that equals 1 if an individual is employed and zero otherwise. Wage distributions are estimated on the basis of the F (w D g, I = 1)P r(i = 1 D g ) term alone. It would be misleading if F (w D g, I = 1)P r(i = 1 D g ) and F (w D g, I = 1)[1 P r(i = 1 D g )] were systematically different. This problem typically affects the estimate of female wage offer distributions in the low-educated labour market, as the unemployment rate of women, 1 P r(i = 1 D w ), declines over the levels of education. The goal is to retrieve the gender gap in (potential) wages at the v th percentile of the distribution. The log wage at the v th percentile, w v, for each gender is defined in equation (4) θ =F (w v D g, I = 1)P r(i = 1 D g )+ F (w v D g, I = 0)[1 (P r(i = 0 D g ))], (2.4) where θ = v/100. To identify F (w v D g ), it needs to retrieve the information on F (w v D g, I = 0)[1 (P r(i = 0 D g )] that represents the probability that non-employed observations have

95 86 potential wage below the v th percentile of the distribution. The approach of this study is based on some form of wage imputation for non-employed individual, but it simply requires assumptions on the position of the imputed wage observations with respect to the v th percentile of the wage distribution, and not on their level of potential wage offer. To see it formally, the explanation below uses the gender wage gap at the median as an example. 29 It estimates the median wage gap in potential wage offers using median wage regressions. Let s consider the linear wage equation w i = β 0 + β 1 D m,i + ɛ i, (2.5) where w i denotes the log hourly wage, D m,i = 1 denotes a man, D m,i = 0 denotes a woman, β 0 is a constant term, and β 1 is the parameter of interest. The conditional median of ɛ given D m,i is assumed to be zero. Denote ˆβ as the hypothetical least absolute deviation (LAD) estimator for a median regression. It is based on the potential wage offers, w i, where ˆβ = ( ˆβ 0, ˆβ 1 ). ˆβ = argmin β N w i β 0 β 1 D m,i i=1 29 The explanation for wage imputation at the median is a summary of methodology section in Olivetti and Petrongolo (2008).

96 87 The wage offers are not observed for those who do not work, I i = 0. Suppose that the potential wage offers of the non-employed are categorized into two groups, L and U, such that w i < ŵ i = ˆβ 0 + ˆβ 1 D m,i for i L, and w i > ŵ i for i U. The imputation procedure can construct a dependent variable y i that is equal to w i for I i = 1 and to some arbitrary wage offer w imputed,i for I i = 0 such that w imputed,i < ŵ i for i L and w imputed,i > ŵ i for i U, and then the following condition holds: ˆβ imputed = argmin β N i=1 y i β 0 β 1 D m,i = ˆβ N = argmin β w i β 0 β 1 D m,i (2.6) i=1 Condition (11) states that the LAD estimator is not affected by imputation when the missing wage observations are imputed on the correct side of the median of the potential wage offers. 30 That is to say, the LAD estimation using y i yields the same estimate of the median gender gap as it would yield if potential wage offers, w i, were available for the whole population. The LAD estimator is the solution to the quantile regression by Koenker and Bassett (1978) when θ = 0.5 (v = 50). 31 When v equals to values other than 0.5, one can prove that the quantile regression estimate of the gender gap based on y i at the v th percentile is valid whenever the imputed wage values are on the correct side of the v th 30 See Bloomfield and Steiger (1983), Chapter 2, for formal proof. 31 Koenker and Bassett (1978) ( show that the θ th regression quantile,0 < θ < 1, is defined ) as any solution to the minimization problem: min θ y t x t b + (1 θ) y t x t b. The LAD estimator b R k t t:y t x tb t t:y t x tb is the regression median,i.e., the regression quantile for θ = 1/2.

97 88 percentile of the potential wage offers. 32 Imputation on observables These are the mathematic notes for estimating the wage distribution of people who were never employed during a 6-year window. For each gender-education group, it takes the form of equation (7), FW h :X=x D g,i=0 = = F W X,Dg,I=1(w X = x)df X Dg,I=0(x) F W X,Dg,I=1(w X = x)τ(x)df X Dg,I=1(x), (2.7) where τ(x) = df X D g,i=0(x) df = P (I = 0 D g, X) P (I = 1 D g ) X Dg,I=1(x) P (I = 1 D g, X) P (I = 0 D g ) where P (I = 1 X, D g ) and P (I = 0 X, D g ) are the probability of one belonging to group I = 1 and I = 0 conditional on X, respectively. P (I = 0 D g ) and P (I = 1 D g ) are the sample proportions in group I = 0 and I = 1, respectively. In this case, the imputation rule does not require an assumption of the identical rank throughout the whole wage distribution between the matched pairs of the non-employed and the employed but only with respect to the v th percentile. Formally, it takes the following form 32 See Koenker and Bassett (1978) for formal proof of Theorem 3.5.

98 89 F (w v D g, I = 0) = F h W v:x=x D g,i=0 (2.8) Equation (13) states that if a non-employed individual were employed, his wage position with respect to the v th percentile would have been the same as the wage position of an employed worker who has the same labour market characteristics as the non-employed one The assumption underlying equation (8) is that the labour market characteristics of the non-employed would have been rewarded the same as the employed. In other words, it does not account for the possibility that the non-employed may have been paid lower than the equally productive employed because of his/her unemployment duration.

99 90 Appendix B Figure B1: Gender Gap across the Wage Distribution by Employment Status Figure B2: Gender Gap across the Wage Distribution by Employment Status with 95% Confidence Interval