Economic Reforms and Gender Inequality in Urban China

Transcription

1 Economic Reforms and Gender Inequality in Urban China Haoming Liu Department of Economics National University of Singapore May 31, 2007 Abstract This paper jointly examines the gender earnings gap and employment rate in urban China using a longitudinal data set that covers the period of Consistent with previous studies, we find that the earnings gap increased between 1989 and The rise is mainly driven by unobserved gender specific factors, such as a decline in women s unobserved skill or an increase in the discrimination against them. After 1993, the gap fell gradually, which is achieved at the cost of less educated women s employment. Consequently, the narrowing gap is not necessarily a good news to all urban women. Since the exit from employment is mostly involuntary, women still lost relative to men even in the later reform period. Keywords: Economic reform, gender inequality, discrimination JEL classification: J3, J7, P25 Corresponding author: Haoming Liu, Department of Economics, National University of Singapore, 1 Arts Link, Singapore ecsliuhm@nus.edu.sg, Phone: , Fax:

2 1 Introduction The impacts of economic reforms on the gender earnings gap in transition economies have been examined by several authors with mixed findings. For example, Brainerd (2000) finds that economic reforms increased the gender earnings gap in Russia and Ukraine but reduced it in most Eastern European countries such as Hungary, Poland, Czech Republic and Slovak Republic. Hunt (2002) shows that while the gender earnings gap fell in East Germany after German unification, almost half of it was attributable to the exit from employment of low skill women. Münich, Svejnar, and Terrell (2005a) suggest that economic reforms narrowed the gap in Czech Republic. Studies on China mostly find that economic reforms raised the gender earnings gap. For instance, Gustafsson and Shi (2000) show that China s gender wage gap in urban areas rose slightly, from 15.6% in 1988 to 17.5% in Similar pattern has also been documented by Knight and Song (2003) and Yang (2005). Comparing with other transitional economies, the gender earnings gap in China is small and little affected by economic reforms. Several factors might be responsible for China s small and relatively stable gender earnings gap. First, although China s economic reforms were commenced in the late 1970s, early reforms mainly affected rural areas. The urban labor market has not been affected utile the mid-1990s (Meng 2004). This is evident by changes in the payroll employment share of State-Owned Units (SOU) and of Collective-Own Enterprises (COE). The payroll employment share of SOUs is 78.4% in 1978, 73.4% in 1988, 73.5% in 1995, and 63.1% in This suggests that the 1995 data is still too early to reveal the impacts of recent fundamental urban reforms. Second, as has been pointed out by Hunt (2002), the small increase in the gender earnings gap might be attributable to the decline in the employment rate of low skilled workers. Because the average skill level of women is lower than that of men, a decline in the employment rate of low skilled workers has a larger effect on women s than on men s 1 The information is extracted from Table 1-14 of the 2005 China Labour Statistical Yearbook. The payroll employment only includes staff-and-workers, is called zhi gong in Chinese. Hereafter, we will use payroll employment and employment interchangeably. 1

3 employment, which narrows the gender skill gap. This paper uses information extracted from the China Health and Nutrition Survey (CHNS) to tackle this issue. Comparing with previous studies that largely used the 1988 and 1995 China Household Income Project (CHIP), the CHNS enables us to document the evolution of the gender earnings gap over a much longer period, between 1989 and By extending the sample coverage to post 1995 reforms, we can compare the impacts of recent reforms that mostly affected urban areas with these of earlier reforms. In addition, we can use CHNS to jointly examine changes in employment probability and changes in the gender earnings gap. Consistent with previous studies, we find that the gender earnings gap in urban China raised from 13% in 1989 to 18% in 1993, and then stabilized around that level for the rest of the sample period. The increase in the earnings gap in the earlier reform period was mainly driven by the widening gap at the top of the earnings distribution. Our quantile regression results indicate that while the gap (9.6%) hardly changed at the 10th percentile between 1989 and 1993, it increased from 15.4% in 1989 to 22.3% in 1993 at the 90th percentile. The stable gender earning gap at the lower percentile in the early reform period reflects the fact that there were no obvious losers during this period. According to Zhao and Li (1999), while the incomes of the top 3% urban households increased by 53% from 1988 to 1995, the incomes of the bottom 20% urban households increased by 20% over the same period. In contrast to the earlier period, the earnings gap at the bottom of the earnings distribution increased considerably while it actually declined slightly at the top in the later reform period. For example, while the gap rose from 8.9% in 1993 to 15.6% in 2004 at the 10th percentile, it fell from 22.3% to 7.7% over the same period at the 90th percentile. As a result, the mean gender earnings gap hardly changed. The evidence documented here suggests that while the gender earnings gap in the early reform period is characterized by men gaining ground on women at the top of the earnings distribution, high wage women regained some ground in the later reform period. 2

4 Our decomposition exercise shows that the increase in the earnings gap in the early period was mainly driven by changes in unobserved characteristics the gap effect, and their prices the unobserved prices effect. The gap effect can be due to either a decrease in women s unobserved skill or an increase in discrimination. The unobserved prices effect reflects the impact of rising earnings inequality. The slight decrease in the earnings gap between 1997 and 2004 was mainly attributable to the rising education level of employed women, which was largely due to the exit from employment of low skilled women. Among individuals who worked in 2000, our regression results show that a one year increase in schooling raises women s 2004 employment probability by 5.5 percentage points, while it can only raise men s 2004 employment probability by 1.9 percentage points. The high exit rate from employment of less educated women is not a concern if they withdraw from the labor market voluntarily. Unfortunately, our regression results suggest the opposite. Therefore, our analysis suggests that the recent decline in earnings gap is not necessarily a good news to urban women. The paper is structured as follows. The next section provides some background knowledge on the process of economic reforms in urban China and a brief discussion of the data source. Section 3 documents the evolution of the gender earnings gap in urban China between 1989 and Section 4 analyzes the determinants of employment rate. A short conclusion is given in Section 5. 2 Institutional Background and the Data Before the economic reform, both employment and compensation in China were controlled by the state. The wage system was centrally regulated into occupational based wage scales: administrative personnel were put into 20 salary grades, technicians into 17 grades and manual employees into 8 grades (Knight and Song 2003). Employers had little control over who they could employ and how much to pay. The first step to increasing labor market flexibility is the reintroduction of bonuses and piece wage in Since then, the share of bonuses in total wages for all enter- 3

5 prizes increased from 2% of wage bill in 1978 to around 16% in 1997 (Brooks and Tao 2003). Although, in theory, employers were free to set bonuses level (subject to a ceiling), they rarely used bonuses as an incentive mechanism. Actually, some researchers (Walder 1987) argue that most employers distributed bonuses equally among their employees so as to minimize contention. A labor contract system was introduced in the mid-1980s, these contract workers were largely treated the same as permanent workers. But then in the mid-1990s, the whole institution setup changed dramatically. The radical reform, known as xia gang, first on trial in 1994 and finally launched in 1997 (Appleton et al. 2002). According to Brooks and Tao (2003), 25 million employees from Sate-Owned Enterprises (SOE) and COEs were laid-off during By the end of 1999, the accumulated laid-off workers accounted for 13.2% of the entire urban labor force (Appleton et al. 2002). The mass layoff has a larger negative impact on women s than on men s employment. According to the statistics documented in Table 1 of Appleton et al. (2002), the incident rate of layoff is 12% for men and 22% for women. Although some of the male-female difference in the probability of being laid-off is due to differences in personal characteristics, females probability of being laid-off is still 5.5 percentage points higher than that of males even after controlling for a series of observed characteristics. To make things worse, females are not only more likely to be retrenched than males, but also more likely to be forced to retire early. In addition, once separated from previous jobs, females have a much lower reemployment probability as well. For example, Giles, Park, and Cai (2006) show that while 44.3% of year old males were reemployed within 12 months of leaving their jobs, the corresponding figure for females is only 22.1%. Facing the harsh reality, many females dropped off the labor market. The female labor force participation rate fell from 74.4% in January 1996 to 63.1% in November 2001 while the male labor force participation rate declined from 93.0% to 86.3% (Giles, Park, and Cai 2 SOE consists a subset of SOU. For example, while workers employed by government institutions, hospitals, education and research institutions are a part of SOU employment, they are not a part of SOE employment. 4

6 2006, p67) during the same period. The male-female difference in the labor force participation rate is the largest among the year old group. The labor force participation rate of the year old declined by 7.9 and 14.5 percentage points for males and females, respectively. Among individuals who are in the labor force, females generally have to face higher unemployment rates. In November 2001, the unemployment rate of the year old is 10.3% for men and 17.1% for women. Actually, among all age groups, only women of the youngest group (16-29 year old) have a lower unemployment rate than men of the same age (Giles, Park, and Cai 2006, p67). As a result of the mass layoff, the employment shares of SOEs and COEs have decreased considerably since Table 1 reports the number of payroll employment by registration type. The data show that both total employment and the number of employees in State-Owned Units (SOUs) increased year by year between The employment share of SOUs stayed at around 73% over this period with little year-to-year variation. As urban economic reforms gained momentum in the mid- 1990s, both total employment and SOU employment started to fall, with the latter outpaced the former. As a result, SOU s employment share decreased from 73.5% in 1995 to only 60.9% in During the same period, the employment share of Other-Ownership Units increased from 5.9% to 31.1%. This evidence suggests that previous studies that largely used the 1988 and 1995 data might not able to capture the impacts of more recent radical urban reforms on gender inequality in the labor market. To examine the impacts of the later radical reforms, we use data extracted from the China Health and Nutrition Survey (CHNS). The CHNS is designed jointly by the Carolina Population Center at the University of North Carolina at Chapel Hill and the Chinese Academy of Preventive Medicine. It is implemented by related city/county anti-epidemic stations under the organization of the Food Inspection Services of the provinces. Although the CHNS is designed to examine the effects of health, nutrition, and family planning policies, it does contain detailed 5

7 income and job related information that can be used to analyze labor market issues. The survey took place over a 3-day period, draw a sample of about 4,400 households with a total of 16,000 individuals in nine provinces, including Liaoning, Heilongjiang, Jiangsu, Shandong, Henan, Hubei, Hunan, Guangxi and Guizhou. Households from Heilongjiang were only sampled after the 1993 survey and households from Liaoning were not surveyed in The CHNS has a high retention rate. For example, among the 15,923 individuals who were surveyed in 1989, 14,028 were also surveyed in Even after 15 years of the initial survey, the CHNS still managed to survey 56.3% of the original sample population whose community still participated in The retention rate is even higher among these who were born between 1929 and 1969 (aged in 1989), with a value of 61.4%. Since 1993, all new households formed from sample households residing in sample areas were added to the survey, resulting in a total of 13,893 individuals. Since 1997, additional households were added to replace those no longer participating. In this paper, we focus on individuals aged between 17 and 60 in the survey year, were not enrolled in any types of schools, and lived in urban sites. Their earnings are measured by self-reported average monthly wages (excluding bonus and subsidies) of the primary job in the year prior to the survey year except for the 1989 survey. Since this information was not collected in the 1989 survey, we have to use the monthly wage constructed by the survey center from two questions: (1) daily wage on the current job for workers who were paid by time; (2) piece rate for workers who were paid by the number of pieces finished. Consequently, our wage information covers the period of The monthly wage is top coded at 999 yuan in 1991 and at 9,999 yuan between 1993 and Since only 27 observations are affected, these observations are excluded from our analysis. In the 2000 survey, all reported monthly earnings are lower than 9,999. However, there are 4 observations whose reported wages are greater than or equal to 9,999 yuan in 2004 and 45 observations whose constructed monthly wages are greater than or equal to 999 yuan in

8 To keep consistency, these observations are excluded as well. 3 Because it is difficult to measure labor income for self-employed or family workers, they are also excluded from our analysis. The nominal wages are deflated by the CPI (1978=100) of all urban areas. The change in survey design prevents us from constructing a consistent measure of the amount of subsidies and bonuses. Therefore, subsidies and bonuses are not included in our earnings measure. The amount of subsidies and bonus were broken into several categories in the period of , but only the total amount was asked in the last two surveys. Curiously, while all urban wage earners reported some types of subsidies between , only around half of the urban earners reported a positive amount of subsidy in the last two surveys. Because the missing information on subsidy could be either the result of not receiving any or refusing to answer the question, we use basic monthly salary as our earnings measure throughout the entire sample period. To gauge the potential bias induced by excluding various subsidies, we calculate the correlation between the amount of total subsidies and basic monthly wages among workers who reported both positive wage and subsidies. The correlation coefficient between these two variables never exceeds 0.14, suggesting that they are weakly correlated. The average amount of subsidies received by women is equal to 91% of that received by men, which is larger than the female to male wage ratio of 82%. The above evidence suggests that our wage regression results should not be sensitive to whether we include subsidies, and including them will slightly narrow the gender earnings gap. Another labor market variable used in our analysis is the type of registration of an employee s work units. Although the survey question on this issue varies over time, two groups of workers can be consistently identified throughout the entire sample period: those worked for SOUs and those worked for COEs. Consequently, we group all employed workers into three categories: SOU, COE, and others. The last group largely consists of employees of private enterprises, foreign owned enterprises, 3 Among these excluded individuals, 63% of them are males. Including these topcoded (or would be topcoded in some years) observations will slightly increase the gender earnings gap. 7

9 overseas Chinese owned enterprises and joint ventures. In the later discussion, we will refer this group as employees in the private sector. Other variables used in the analysis are the province of residence, years of schooling and potential years of experience. Years of schooling are constructed from two variables: years of formal education completed and the highest level of education obtained. The existence of two education related variables enables us to check the consistency of the education information. Presumably, the reported years of schooling are likely to be accurate if these two measures conform with each other. Therefore, the reported years of schooling are used whenever these two variables are consistent. If these two variables are only consistent in some survey years, then the years of schooling in these years are used for the nearest inconsistent years. If these two variables are inconsistent throughout the entire sample period, then the respondent s average years of schooling is used for each period. Average years of schooling over the entire sample period might differ from actual schooling in a particular year for individuals whose formal education consists of several disconnected intervals. In our urban sample, only 64 of them went back to school after leaving school in the previous survey, and only 17 of them whose education variables were not consistent with each other. Therefore, we are confidence that the benefits of using our constructed years of schooling rather than the reported years of schooling outweighs its costs. Having constructed years of schooling, the potential years of working experience are then calculated as age minus years of schooling minus 6. Table 2 reports the basic sample statistics by gender and survey year. The female to male earnings ratio fell from 84.1% in 1989 to 82.6% in 1991, which is consistent with the evidence documented by previous studies. Interestingly, women s relative earnings seem to have gained some grounds since Unfortunately, the narrowing earnings gap was accompanied by a declining employment rate. The most dramatic decline in employment rate happened in the late 1990s and early 2000s. Women s employment rate fell faster than men s, which widened the difference in employment rate between men and women. For instance, as shown in the first row of panel A 8

10 and B of Table 2, while men s employment rate declined by 45.9 percentage points between 1989 and 2004, women s employment rate decreased by 46.3 percentage points. This finding is consistent with the statistics documented by the National Bureau of Statistics (2001) using the second survey on women s status in China. According to the National Bureau of Statistics (2001), the urban employment rate of year old fell from 90.0% in 1990 to 81.5% in 2000 for men, and from 76.3% to 63.7% for women. The relative high employment rate documented by the National Bureau of Statistics (2001) is likely due to its loss definition of employment. An individual is considered as employed as long as he has worked for income in the last week. Consequently, individuals who are temporarily employed and self-employed are also considered as employed. However, individuals are considered as employed only if they reported a positive value on the question on average, what was your monthly wage/salary last year, excluding subsidies and bonuses?, which excluded self-employed and temporarily employed. The decline in the employment share of SOU started a few years later than the fall in the employment rate. It changed little between 1989 and 1993 for both men and women, but decreased by 6.7 percentage points for men and 4 percentage points for women between 1993 and The timing of the decline in SOU s employment share is consistent with what has been documented in Table 1 at the national level. One potential reason for the decline in the employment rate is that many workers started their own business and become self-employed. 4 Including self-employed workers into our sample indeed raises the employment rate, but it does not change the downward trend. For example, the employment rate (including self-employed workers) decreased from 90% in 1989 to 69.1% in 2004 for men and from 82.5% in 1989 to 52.8% in 2004 for women. In contrast to the decline in the employment rate, monthly earnings of both men and women increased steadily over the entire sample period. Men s average earnings are 3.51 times larger in 2004 than in 1989 while women s average earnings are 3.63 times higher in 2004 than their 1989 level. Years 4 Self-employed workers are excluded from our sample. 9

11 of education of employed workers also increased steadily for both men and women. It raised from 9.58 years in 1989 to years in 2004 for men, and from 9.14 to for women. The average education level of employed workers is always higher than that of the entire adult population, especially for women, suggesting that better educated individuals have a higher employment rate. The positive relationship between education and employment should have a larger impact on women s employment since their average years of education are lower, which in turn narrows the gender earnings gap. Similar conclusion has been drawn by Hunt (2002) using a German data set. The decline in employment rate and the increase in monthly earnings of employed workers suggest that changes in the gender earnings gap need to be jointly examined with changes in employment opportunities. 3 Wage results To depict a general picture of the evolution of the gender earnings gap, we first use the CHNS as 6 cross-section data sets. We run log earnings regressions for each of the 6 waves separately. The control variables include a gender dummy (=1 for male and 0 otherwise), years of education, potential years of experience and its square, the registration type of work unit, and the province of residence. To help us understand the contributions of these observed variables to the earnings gap, we add them into our regression gradually. The regression results are reported in Table 3. In Panel A, gender dummy is used as the only control variable. Similar to Table 2, the largest increase in the gender earnings gap occurred between 1991 and 1993, and has declined since then. The slight difference between Table 2 and Panel A of Table 3 is that while the former reveals the difference in the average earnings, the latter reflects the difference in the average of log earnings. Panel B shows that controlling for education and experience reduces the gap by 1.4% to 5.1%. 5 Including types of work units in Panel C and the province of residence in Panel D 5 Hereafter, the coefficient on gender dummy γ will all be converted to percent using the formula (exp(γ) 1) as suggested by Halvorsen and Palmquist (1980). 10

12 have no obvious impact on the gap, suggesting gender earnings gap does not vary much across different types of work units. After including all control variables, the gender earnings gap measured by the coefficient on gender dummy increased from log points (13.2%) in 1989 to (14.0%) in 1991 and then to (18.4%) in These estimates are very close to Yang (2005) who finds that the coefficient on gender dummy is in 1988 and in Interestingly, the gender earnings gap stabilized around till the end of the sample period. The increase in the rate of returns to education (as reported in Panel D) started much later than the rise in the gender earnings gap. The coefficient on years of education stayed at around 0.02 for the period of , then increased to in 1999 and in Our estimates are comparable to the rate of returns to education in other central planned economies, such as Czech Republic (0.027 in 1989 Münich, Svejnar, and Terrell 2005b), Russia (0.039 in 1990 Gorodnichenko and Peter 2005), and Ukraine (0.039 in 1991 Gorodnichenko and Peter 2005). It is also consistent with previous Chinese studies. For instance, using the 1988 CHIP data, both Liu (1998) and Yang (2005) find that the rate of returns to education is around 0.03 in China. However, our 1997 estimate is lower than what have been found by Yang (2005) using the 1995 CHIP data set. His estimation suggests that the rate of returns to education increased to in 1995 with a considerable variation across cities. Yang s results show that the rate of returns to education is higher in cities with better information infrastructures and more foreign and joint-ventures. We suggest that the lower rate of returns to education in our estimation is mainly driven by the fact that the CHIP contains more developed provinces, such as Beijing, Guangdong and Jiangsu, than the CHNS does. Another interesting result of Table 3 is that the magnitudes of the coefficients on SOU and COE first increased from 1989 to 1993 and then declined. The coefficients on SOU even became statistically insignificant in the last two surveys. Because workers in the private sector are used as the reference group, these changes suggest the wage gap between private and non-private sector widened in the early reform 11

13 period but narrowed recently. The gain in the relative position of SOU workers during was largely driven by the country-wide salary increase of government employees (including workers in the government, state services and institutes), in July 1989, July 1997, January 1999, January 2001 and October As government employees account for more than half of the SOU employment, these salary increases reduced the earnings gap between workers in the private and non-private sectors. 7 The increased competition between private and non-private sector in the labor market could be another reason for the narrowing wage gap. Because some workers monthly earnings depend on the number of hours worked, our results might be sensitive to the variation in working hours. Unfortunately, monthly working hours cannot be consistently calculated using the CHNS. It only contains information on daily working hours for the entire sample period. Weekly working hours are only available in the 1989 survey. 8 Days worked per week are only recorded in surveys. To make our estimation results comparable over time, we use daily working hours to account for variation in working hours. Table 4 reports the estimation results. For the sake of brevity, we only report the estimation results with the full set of control variables. Interestingly, controlling for daily working hours only slightly reduces the coefficient on gender dummy. This is because women s daily working hours are only about 0.1 hours shorter than that of men. Moreover, adding daily working hours does not affect the trend of the gender earnings gap. The coefficients on education and experience are not sensitive to controlling for daily working hours either, suggesting that working hours do not vary systematically across education and experience groups. Another interesting observation from Table 4 is that the coefficient on daily working hours is only significant at the 5% level 6 Outlook Weekly, 2002, No According to our own calculation, government employees accounted for 63% of the total SOU employment in Because SOE employees and government employees were grouped together in the earlier surveys, we cannot calculate the employment share of the government employees. However, as the downsizing mainly occurred in SOEs, the employment share of the government is likely to be smaller than the 2004 value. 8 Weekly working hours in the 1991 to 2004 survey refers to the week prior to the survey week rather than to hours per week usually worked. 12

14 in 1989 and This does not necessarily imply that monthly earnings do not depend on working hours. Rather, the weak relationship between working hours and monthly wages reveals that jobs requiring long working hours do not pay higher wages. Several researchers (e.g. Knight and Song 2003; Meng 2004) find that earnings inequality has increased in the 1990s. These studies also suggest that the increase in the rate of returns to human capital is at least partially responsible for the rise in inequality. Because the skill distribution differs between males and females, their impacts on the earnings gap are unlikely to be uniformly distributed across the earnings distribution. Figure 1 plots female/male earnings ratios at selected percentiles. The percentile rankings refer to each gender group s own earnings distribution. Interestingly, although the overall earnings ratio declined from to between , the earnings ratios at the 20th and 40th percentile actually increased while the earnings ratio at the 80th fell. This suggests that the widening earnings gap in the earlier reform period was mainly driven by the deterioration in female s relative earnings at the top half of the earnings distribution. In contrast, it was the females at the bottom of the earnings distribution who lost ground to men during The earnings gap increased at the 20th, 40th and 60th percentile, but decreased at the 80th percentile. To quantitatively compare changes in the earnings gap at various percentiles, we run a series of quantile regressions at the 10th, 25th, 50th, 75th and 90th percentiles with the full set of controls as in Panel D of Table 3. Unlike the OLS estimates that reveal the returns to observed characteristics at the sample mean, the estimates of quantile regressions reveal the returns to observed characteristics at the particular percentiles. 9 plotted in Figures 2 and 3. These two figures show that the evolution of the gender earnings gap at different percentiles takes different paths. In 1989, the gap was almost uniformly distributed across the earnings distribution. It then increased dramatically at the 75th and 90th percentiles in 1991, but changed little 9 The estimation results are reported in Tables A1 to A5. 13

15 at the bottom half of the earnings distribution, implying that our OLS results were mainly driven by changes at the top half of the earnings distribution (conditional on observables). The difference in the degree of risk aversion between men and women (Jianakoplos and Bernasek 1998) is likely to be responsible for the widening gap at the top of the earnings distribution. As men are less risk aversion than women, they have a higher probability of switching from relatively stable jobs to less stable but better paid jobs at the early stage of urban reforms. Moreover, because working for private firms is still not the norm in the early reform period, it is usually the husbands who quit relatively stable SOU or COE jobs to pursue high paying private jobs while wives stay with their non-private employers to enjoy health and housing subsidies that private firms normally do not provide. As Figures 4 7 show, between 1989 and 1993, better-paid jobs in the private sector confer a considerable premium over jobs in non-private sectors while the premium for lower-paid private jobs is much smaller. Men s higher probability of moving to the private sector will widen the earnings gap at the top of the earnings distribution. This claim is supported by the data. Among the 1,744 men who worked at least in two out of the 6 surveys, 14.7% of them switched from non-private to private sector while 6.8% of them moved from private to non-private sector. The corresponding numbers for the 1,633 women who worked at least in two out of the 6 surveys are 13.2% and 7.2%. As urban economic reform accelerated in the late 1990s, even jobs at SOUs are not for lifetime. Consequently, the gender difference in terms of accepting different type of jobs also narrowed. Actually, women were more likely to work in the private sector than men in the 2004 survey. As more women took higher paid jobs, the gender earnings gap at the top of the earnings distribution were reduced. This argument is consistent with the pattern documented in Figure 1. After 1993, the gap still trended up at the 10th, 25th and 50th percentiles, but trended down at the 75th and 90th percentiles. This finding is consistent with Liu, Zhang, and Kung (2004) who show that the gender wage gap increased by more that 12 percentage 14

16 points between 1988 and 1999 at the 10th percentile while it declined by about 5 percentage points over the same period at the 90th percentile. Another interesting observation from these quantile regressions is that the coefficients on SOUs at different percentiles also evolved differently. The earnings gap between SOU employees and private employees was much larger at the top than at the bottom of the earnings distribution, implying that top earners in the private sector earned much more than their counterparts in the non-private sectors while bottom earners in the private sector earned only slightly more than their counterparts in the non-private sectors. This suggests that the earnings distribution was more compressed in SOUs than in the private sector. The cross-sector differences have narrowed since 1993 as economic reforms progressed. Figures 4 and 5 show that both the 90/10 and 75/25 differences have decreased since 1993 while the coefficient on the SOU dummy increased steadily at all percentiles. Figures 6-7 shows the coefficient on COE at different percentiles. Similar to the coefficient on SOU, the 75/25 difference decreased gradually over the entire sample period. However, unlike the coefficient on SOU, the 90/10 difference shows no clear pattern. The coefficient on COE at the 90 percentile is still negative and significantly differs from 0 at the 7% level in the last two surveys. This suggests that earnings grow faster in SOUs than in COEs, particularly at the top of the earnings distribution. Since women are more likely to work in COEs than men, changes in the earnings distribution in COEs should have a larger impact on women s earnings. To spell out the contributions of various factors to the gender earnings gap, we adopt the decomposition method proposed by Blau and Kahn (1997). Assuming that males are paid fairly according to their personal characteristics, then the coefficients from males earnings regression should be able to reflect the actual price of the corresponding characteristics. Thus the gender earnings gap can be modeled as D t = ln w mt ln w ft = X t β t + σ t θ t (1) where w j (j = m, f) represents monthly earnings, and X is a vector of control vari- 15

17 ables including education, experience, registration type of work units, and residence of province, the subscript m means male and f female, x represents the sample average of x, β is the corresponding estimates from male earnings regression, θ is the standardized residual, σ is the standard deviation of the earnings residuals, and denotes the male-female difference in X and θ at their corresponding means. The change in the earnings gap between year t and 0 can then be decomposed as: D t D 0 = ( X t X 0 )β t + X 0 (β t β 0 ) + ( θ t θ 0 )σ t + θ 0 (σ t σ 0 ). (2) The first term reflects the contribution of changes in characteristics to the earnings gap when evaluated at the period t price. The second term reveals the contribution of changes in returns to characteristics evaluated at the period 0 difference in characteristics. These two terms are called observed X s effect and observed prices effect respectively by Blau and Kahn (1997). The third term represents the contribution of changes in the percentile ranking of the female earnings residuals while holding the standard deviation of male earnings residuals at the period t level, called gap effect. Finally, the fourth term of equation (2), the unobserved prices effect, measures the contribution of changes in the standard deviation of male earnings residuals holding the percentile rankings of female earnings residuals at their period 0 levels. The last two terms of equation (2) can be rewritten as ( θ t θ 0 )σ t + θ 0 (σ t σ 0 ) = [(θ tm θ tf ) (θ 0m θ 0f )]σ tm + (θ 0m θ 0f )(σ tm σ 0m ) (3) Because the averages of both θ tm and θ 0m are zero, expression (3) can be further 16

18 simplified as ( θ t θ 0 )σ t + θ 0 (σ t σ 0 ) = ( θ tf + θ 0f )σ tm θ 0f (σ tm σ 0m ) (4) where θ tf σ tm and θ 0f σ 0m are the averages of female earnings residuals in year t and 0 imputed based on the corresponding year s males earnings regressions. To impute θ 0f σ tm, we first estimate each woman s percentile ranking in year 0 based on her earnings residual in that year s distribution of male earnings residuals. Then, we assign her the residual of the corresponding percentile of the year t distribution of male earnings residuals. For example, if a woman s year 0 earnings residual is ranked at the pth percentile in that year s male earnings residual distribution, then she would be assigned the residual corresponding to the pth percentile of the year t male earnings residual distribution. Table 5 reports the decomposition results. The β s used in the decomposition are estimated using only male earnings with the same control variables as what have been used in Panel D of Table 3. Given the low male employment rate in the 2000s, it is possible that the estimates from the men s earnings regression might also subject to the selection bias. Therefore, we also tried to estimate β s using the full maximum likelihood estimates of the Heckman selection model. Our results suggest that the estimates are not sensitive to the estimation method. This is because the low male employment rate is mainly due to the difficulty of locating a job rather a low labor force participation rate. For simplicity, we will focus our discussions on the results based on the OLS estimates. To make our results comparable with existing studies, we group the entire sample into three periods , , and The impacts of self-selection on female s earnings will be reflected in either the observed X s effect or the gap effect. The first column of Table 5 reports the decomposition of the changes in the 10 For the sake of brevity, the estimation results of the underlying wage regressions are not reported. To check whether our results are sensitive to controlling for sample selection, we also conduct a decomposition exercise using the coefficients from the full maximum likelihood estimates of the Heckman selection model. The results reported in Table A6 show that our conclusions are not sensitive to whether we control for the sample selection in the men s wage regression. 17

19 earnings gap between 1989 and 1993, the period when the earnings gap increased the most. Changes in observed X s widened the earnings gap by 0.6 percentage points. The increase in women s probability of working for SOUs from 1989 to 1993 is the major contributor. The decline in the relative earnings of SOU employees and of COE employees worked in opposite directions. On the one hand, the increase in women s probability of working for SOUs widened the gap by 1.5 percentage points. On the other hand, the decline in women s probability of working for COEs narrowed the gap by 0.3 percentage points. Changes in the prices of observed X s have a much larger impact on the earnings gap than the observed X s effect. If nothing else had changed except for the prices of X s, the gap would have narrowed by 3.8 percentage points. Again, the decline in the relative wage of SOU employees is the major contributor for the observed prices effect. As males are more likely to work for SOUs, their earnings are more sensitive to the fall in the relative wages of SOU employees than women s earnings. The increase in the gap effect and the unobserved prices effect are the two major culprits for the widening genders earning gap, each contributes 2.4 percentage points. This is in contrast to the German evidence documented by Hunt (2002) who finds that the gap effect reduced the gender earnings gap in Germany. The increase in the gap effect may be the result of either a rise in the degree of labor market discrimination against women, or a relative deterioration in women s unobserved skill. The latter explanation would be correct if women with higher level of unobserved human capital have a higher exit rate than other women. Since observed and unobserved human capital are likely to be positively correlated, this argument also implies that women with higher level of observed human have a higher exit rate. However, our decomposition shows that changes in education level and experience, the two commonly used measures for skill, only have minuscule impacts on the gender earnings gap. In addition, our later analysis on employment rate suggests that better educated women have a higher employment rate than less educated women. Therefor, the gap effect is unlike the result of a deterioration 18

20 in women s unobserved skill. Moreover, if the gap effect is driven by changes in the relative level of unobserved skill, it will not affect the gender earnings gap among individuals who worked in both 1989 and Following this rationale, we repeated the decomposition exercise using a sample individuals worked in both years. For brevity, the detailed results are not reported. Interestingly, while the gross earnings gap only increased by 0.15 percentage points, the gap effect still widened the earnings gap of the balanced panel by 0.16 percentage points. This further confirms that the widening gender earnings gap during this period was not driven by the exiting of high skilled women from the labor market. Another factor that can lead to the increase in the gap effect is a decline in women s relative productivity. If this is the case, then the returns to women s observed human capital should be lower as well. To test this argument, we pool women and men together and interact gender dummy with the full set of control variables. The coefficient on the interaction between education and gender dummy is positive (the return is higher for men) in 1989 and negative (the return is lower for men) in The coefficient is not statistically significant even at the 10% level in either 1989 or This evidence does not support the productivity difference interpretation. Therefore, the increase in the gap effect is likely to be the result of an increasing in the degree of gender discrimination. Column (2) of Table 5 shows that the gross earnings gap has declined slightly from 1993 to This decline is mainly driven by gender specific factors, i.e. the observed X s effect and the gap effect. If the prices of observed and unobserved characteristics had not changed over this period, the gender earnings gap would have declined by 2.8 percentage points. The observed X s effect was mainly driven by the fact that while the employment share of COEs increased for men, it decreased for women. If the decline is the result of women leaving COEs for better paid jobs, it improves women s relative earnings. If it is the result of women leaving for unemployment, then it deteriorates women s employment prospect. Our calculation 19

21 supports the latter interpretation. Among the 118 men who were employed by COEs in 1993, 71% of them were still employed in However, among the 141 women who were employed by COEs in 1993, only 54.6% of them were employed in In contrast to these gender specific factors, the observed prices effect and the unobserved prices effect widened the gap by 1.3 percentage points and 0.6 percentage points, respectively. The largest contributor is the increase in the relative earnings of SOU workers. As shown in figures 4-3, the earnings of SOU workers increased at a faster rate than COE workers. Because statistics reported in Table 2 show that men were more likely to work for SOUs than women, they should benefit more from the increase in SOU workers earnings. Column (3) of Table 5 shows that the gross earnings gap declined by another 0.8 percentage points from 1997 to Again, changes in observed characteristics, which narrowed the gap by 4.6 percentage points, is the single largest contributor to the narrowing gap. This observed X s effect was mainly driven by the rise in female workers relative education level, which was because of the higher exit rate of less educated women. For example, while 61.9% of male workers with at most lower secondary education in 1997 were still employed in 2004, the corresponding number was only 37.1% for females. The observed X s effect was almost counterbalanced by the observed prices effect, which widened the gap by 4.1 percentage points. Half of the observed prices effect was attributable to the rising rate of returns to education. Another contributor is the increase in SOU workers earnings. Unlike observed variables, unobserved factors have little impact on the earnings gap during this period. The gap effect widened the earnings gap by only 1 percentage point, and the unobserved prices effect reduced it by 1.3 percentage points. Because changes in the observed X s suggest that unskilled female workers have a higher exit rate from employment, the increase in their relative unobserved skill is also likely to be the result of the higher exit rate of women with less unobserved skills. Column (4) reports the decomposition results for the entire sample period. The 20

22 gross earnings gap in 2004 is almost identical to its 1989 level. However, this does not suggests that nothing has happened over these 15 years. Changes in observed characteristics, mainly driven by the increase in the education level of employed women, narrowed the earnings gap by 4.3 percentage points, while changes in the prices of observed characteristics widened the gap by 1.6 percentage points. If the relative education level had been the same as it was in 1989, then the earnings gap would have increased by 1.7 percentage points. The gap effect contributed another 1.5 percentage points to the increase and the unobserved prices effect contributed 2.6 percentage points. Because the difference in education level between employed women and the entire female population increased from years in 1997 to years in 2004, it is reasonable to suspect that less educated women have a higher exit rate from employment than well educated women. This suggests that the selective exit from employment might mitigate the gap effect over the entire sample period. 4 Employment Results In the previous section, we show that the increase in the relative education level of women helped to close the earnings gap by 4.3 percentage points. If the increase in the relative education level of women is the result of a faster growth in women s education level, then it suggests that women s relative status has improved since However, if the increase is due to a higher exit rate of less educated women from employment, then it implies that women s status has deteriorated if the exit is involuntary. To examine whether the probability of exit from employment differs across gender and skill level, we create 5 panel data sets from the total 6 surveys. Each panel consists of two adjacent surveys, , , , , and The sample of individuals are restricted to those who reported a positive earnings in the first survey and were interviewed in the second survey. 11 Figure 8 11 To have a complete picture on the impact of changes in employment rate on the gender earnings gap, we should compare both the exit rate and the entrance rate. However, because only a very small number of individuals who did not work in the first period were employed in the second period, the 21

23 plots the employment rate in the second survey by workers earnings quintiles in the first survey. In Panels (a) to (e), the quintiles are separately calculated for the male and female earnings distributions. In Panels (f) to (j), the quintiles are calculated using the pooled earnings distribution. During the period of , employment rate was negatively correlated with earnings. On average, women had a lower employment rate than men, and the differences were larger at the top of the earnings distribution. The higher exit rate of better-paid female workers widened the gross gender earnings gap. Because wages largely reflected workers seniority in the early reform period, the negative correlation between earnings and employment might be the result of the higher probability of withdrawing from employment of senior workers. During the period of , the relationship between employment and earnings is not very clear. Nevertheless, as in the earlier period, women still had a lower employment rate than men, and the differences at the top of the earnings distribution were larger than at the bottom of the earnings distribution. The correlation between employment rate and earnings reverted from negative in to positive in for both male and females. Moreover, the gender differences in employment rate were larger at the lower half of the earnings distribution in the period , which is also in contrast to the data. The higher exit rate of lower-paid female workers helped to narrow the gross gender earnings gap. As the narrowing earnings gap was achieved at the cost of the declining employment opportunity of lower-paid women, it did not signal a reduction in discrimination against women. Although the graphic presentation is informative, it cannot reveal whether the high exit rate of lower-paid workers is due to their lower level of human capital or other factors that are positively correlated with their wages. For example, workers in SOUs do not only earn less than workers in the private sector, but might also face a higher probability of being laid off due to the large scale of downsizing in difference in the entrance rate between men and women is unlikely to have any significant impact on the earnings gap. For the sake of concise, we focus our discussion on the exit rate. 22