WPS 1I91s POLICY RESEARCH WORKING PAPER I 9 12 Pay and Grade Differentials At the World Batk. only about half of salary and grade at the World Bank differentials between men and women cnd between Deon Filiner- staff from high- and, low- income couritries ar-e Margaret Groslh attributabie to d.fferences in Elizabetb King worker charactenstics Neither Dominique van de Walle omitted-variable hia s nor quotas iniposed to ensure diversity are compelling explanations for remaining differentials in salary and grade. Discriminat:on is likely to explain some of the remainder. The World Bank Development Research Group April 1998 POLICY RESEARCH WORKING PAPER 1912 Summary findings Large international organizations such as the World Bank from high- and low-income countries are attributable to pursue many objectives in hiring policies, including differences in worker characteristics. reduced costs, cultural diversity, and the avoidance of They explore a number of alternative explanations for discrimination. the rest of the salary and grade differentials, including There can be sharp tradeoffs between these objectives. omitted-variable bias, quotas imposed to ensure diversity, Diversity is enhanced by recruiting from an international and discrimination in hiring and promoting. labor market, for example, but international They argue that neither omitted-variable bias nor organizations face unusually large differences in quotas are compelling explanations for disparities and reservation wages for staff capable of doing the same that discrimination probably exists, although certainly work. less than would be implied by a cost-minimizing hiring One way to reduce costs would be to pay employees policy. their reservation wages, wvhich implies unequal pay for A shift seems to be occurring in the hiring process at equal work, or discrimination. the Bank, possibly because (1) the application pool, Filmer, Grosh, King, and van de Walle show how including women and Part II nationals (from developing these tradeoffs are resolved in the World Bank's hiring countries) has significantly improved in quality, (2) processes. They estimate disparities in salary and grades information gathering during hiring has intensified, between men and women and by country of origin that decreasing guesswork, (3) there is more incentive to staff cannot be attributed to differences in the productive from minority groups, and (4) the Bank's increasing characteristics of workers. diversity in terms of gender and nationality groups is The results indicate that about half the salary and more conducive to high performance by the people grade differentials between men and women and staff against whom there may previously have been bias. This paper - a product of the Development Research Group - is part of a larger effort in the group to apply economic analysis to policy issues. Copies of the paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Cynthia Bernardo, room MC2-501, telephone 202-473-1148, fax 202-522-1154, Internet address cbernardo@worldbank.org. April 1998. (72 pages) The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should he cited accordingly. The fiudings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries they represent. Produced by the Policy Research Dissemination Center Pay and Grade Differentials at the World Bank Deon Filmer, Margaret Grosh, Elizabeth King and Dominique van de Walle** Development Research Group World Bank ** The authors, listed in alphabetical order, are all of the World Bank's Development Research Group in Development Economics. The paper originated as a request to the Development Research Group to examine the patterns of pay and grade of staff at the Bank, focusing on differences between men and women, and to assess the presence of any gender bias. For their help, the authors thank Berk Ozier, Asli Demirguc-Kunt, Dan Levy, Thomas Hastings, Selina Khan, Susheela Jonnakuty and Rajeev Jain (HRSIS) who kindly assisted in preparing the database. They also thank colleagues and seminar participants for excellent comments on earlier versions. A. Introduction The World Bank has multiple objectives in its hiring policies. On the one hand (like other companies and organizations), it wants to cut costs where possible. But it also pursues a number of other, potentially conflicting, objectives as an employer: 1. A preference for cultural diversity. The World Bank's charter has long championed a composition of employees that broadly reflects the composition of country membership.' In more recent times, thiis preference has expanded to include sensitivity to the gender balance across nationalities. The flip side of cultural diversity goals is the likelihood of systematic differences in the characteristics of workers which may, in turn, explain salary and grade differences. There are far fewer PhDs from developing countries. It is harder to hire women from countries which have not traditionally encouraged women to work outside the home, or to induce them to move to Washington. Language and communication skills vary across individuals for whom English is not the primary language. 2. An international labor market. Most firms hire locally or, at best, nationally. Due in part to its aim for diversity, the Bank faces a much larger labor market in which supply prices vary enormously. The labor pool it confronts is one aggregated across diverse and only weakly integrated national labor markets across the world. Wages for the same labor differ among these national labor markets. So the Bank faces unusually large differences in reservation wages for Assuming diversity can be measured through the mix of nationalities, the Bank has done well. As of May 1997, 129 nationalities were represented at levels 18 and higher (para-professional and professional grades), and 177, over all staff. However, the passport one holds does not necessarily reflect one's experience. I staff capable of performing the same duties.2 Yet the Bank, like other employers, seeks to reduce its labor costs. In principle, one way to do this would be to pay potential employees their reservation wage. This could well result in substantial discrimination, defined as different wages for workers with the same productivity. 3. A desire to avoid discrimination. The World Bank, in common with many employers, wants to avoid discrimination or, at any rate, the appearance of discrimination. This works to relieve the worst effects of facing world labor markets with skewed wage distributions for similar workers. It would not be politically feasible to minimize costs. This problem is common to many local firms not least because of the ramifications of not complying with the law. But in the particular case of the World Bank, this further entails an implicit desire not to simply mirror the (large) differences in wages for the samne labor in the world at large. The Bank's personnel department takes an active role-and one that has become more aggressive in recent years-in trying to ensure that the Bank meet its mandate of achieving staff diversity, while also treating hires equitably. A conscious effort has been made through, for example, greater centralization of recruitment decisions and close monitoring to reduce horizontal inequity. 4. A dual recruitment structure. There are two main ways in which staff are recruited to the World Bank. Around 20 percent of current staff started out as Young Professionals (YP). The program, which sets an age limit and starts all recruits at the same initial level, is designed to 2 There are other reasons besides labor market segmentation associated with restricted labor mobility internationally. Some potential hires are educated elites who are quite desperate to flee bad economic or political situations at home, or simply to improve their work environments and bring up their children with more opportunities, and so are willing, in order to ensure employment, to accept lower pay than counterparts who are already settled in the U.S. Others may have much reduced bargaining power because of the necessity of having visas renewed in order to stay in the U.S. As is often noted, married women's bargaining power is also reduced as they frequently face lower mobility and choose work conditional on a husband's work location (for example, see Hersch and Viscusi 1996). U.S. men (say) who face none of these constraints are likely to have much higher bargaining power. 2 recruit a young cadre of professionals who have a little work experience and recent academic qualifications. In general, YP recruits undergo a more uniform and rigorous interview and selection process than other recruits. The applicant pool is also far larger relative to the number of positions being filled than for other job openings. For this reason, YPs can be expected to be a relatively homogeneous group in terms of characteristics such as technical ability and communications skills. Non-YP recruitments follow a number of less formal routes, which tend also to be less controlled by Personnel or by concerns of the greater good of the institution at large. Some come in through open recruitment, while others are directly identified and hired by managers. The difference between the YP recruitment process and the less regulated process for the rest of the staff hired provides an interesting comparison for tests of discrimination. Without the third feature, and possibly the fourth, it would not be surprising to find large disparities in salaries between men and women, and by country of origin, at the World Bank-disparities which cannot be attributed to differences in the characteristics of workers. The interesting question here is how far the third and fourth features of World Bank hiring processes are able to reduce such disparities. That is the main question addressed by this paper. We classify staff by gender and by whether they come from a Part I or Part II member country of the Bank.3 The data cover all regular and fixed-term staff in the professional (non- support) grades who vvere on active duty as of May 23, 1997.4 The data do not include full pay Data on staff members' pay, grade, and characteristics were prepared by Human Resource Services Department (Personnel) from its staff files. Part I member countries are considered "developed" and are no longer entitled to borrow from the Bank. In general, Part I countries consists of OECD countries. Part II member countries are considered "developing" and are therefore entitled to borrow. 4 Staff in grades 22 to 30 are included. Note that some staff in these grades entered at lower grades. In particular, those who entered the World Bank through the Young Professionals Program enter at grade 21. In addition, some entered at non-professional levels. The total number of regular staff in grades 22-30 was 3,003 as of the date the data were extracted from the records. 3 and promotion histories, but do include both entry and current levels, so that average annual growth in salaries and grades over time can be measured. Table 1 presents descriptive statistics for salaries, indexed to be 100 for Part I men. All underlying salaries are net of taxes and expressed in 1990 US dollars.5 The ranking from highest average salary to lowest is Part I men, Part II men, Part I women, then Part II women. Current average salary for Part II men is 95 percent of that for Part I men; for Part I women, it is 87 percent; and for Part II women, 82 percent. The same rankings are revealed for pay at entry, but with even larger quantitative differences. Some catch-up in relative terms is thus indicated. This especially favors Part II men relative to Part I men. Annual salary growth rates are on average a little higher for women (3.8 percent versus 3.2 and 3.3 percent for Part I and Part II men, respectively). Table I also shows mean tenure for the four staff groups. The slightly lower mean years of employment duration of women (10.7 for Part I and 11.2 for Part II women against 11.9 and 12.5 for Part I and II men respectively) probably reflects the Bank's active efforts at female recruitment in recent years. In May 1997, Part I men accounted for 45 percent of all staff in levels 22-30; Part II men, for 29 percent, Part I women, 18 percent, and Part II women, 8 percent. A pattern similar to that for salaries is found for the distribution of staff groups across grades. Table 2 shows the summary statistics, while Figure 1 graphs the distribution of each staff group. Women are in general recruited at lower grades than men, and Part II staff at lower grades than Part I staff. Figure 1 shows clearly that the grade distribution of men is skewed S According to the international tax treaties that pertain to World Bank employment, non-U.S. citizens are exempt from U.S. income taxes, but U.S. citizens are not. To maintain comparability among employee groups, all regular staff are paid salaries net of taxes, with U.S. citizens being given a separate tax allowance. 4 towards the higher grades relative to the grade distribution of women, even though the average duration of employment tends to be similar across groups. What accounts for these disparities? Should they be a source of concern? The disparities in pay and grade documented above are unadjusted for differences in observed individual characteristics, such as length of Bank employment, highest education degree, and area of specialization. Clearly, differing values of such characteristics will be rewarded differently. If individual characteristics that matter to productivity systematically differ across groups, then that could explain the differences in salary and grade. But, perhaps they don't. We may in fact be seeing the effect of underlying differences in the returns to the characteristics of the four groups. The analysis that follows examines the sources of these disparities. These issues are clearly of concern to the World Bank. But the Bank's experience is of wider interest from a number of points of view. It is of general interest to see how the evolving objectives of this large international organization facing a global labor market influence the way in which its internal labor market functions; armned with data covering a long enough period, we are able to throw light on the effects on wage differentials of the Bank's changes in hiring policies, with potential implications for other large organizations. Another aspect of the Bank's experience which is of wider interest is the institution's unusual dual entry system, with differing possibilities for discrimination between the two streams of recruits. The structure of the internal hiring process at the Bank allows us to implement a novel test for omitted variable bias-a common problem in past empirical research on wage differentials.6 6 Recent discussions in the literature include Jones and Makepeace (1996) and Berger (1995). 5 Below we first discuss our empirical models. We then present our findings on the decomposition of salary and grade differentials-including at entry, current and in growth over time-across staff groups. A number of alternative assumptions and samples are tested in order to gauge robustness of the results, as well as to better understand them. A final section interprets the results and discusses their implications for Bank policy and practice on hiring and promotion. B. Models and Decompositions We use an approach standard in the literature on labor economics.' Models The reduced-form model for the log of current salary (Wij ) for the ih staff member in the jth group can be written as: (1) lnWij = Xijaj + eij, where Xij is a vector of individual characteristics-including age, nationality, education, area of specialization, work experience prior to Bank employment and duration of employment in the Bank. The corresponding parameters are aj, while eij is a zero mean error term uncorrelated with Xij. The estimated coefficients are used to calculate predicted salaries for the four groups as well as predicted salaries based on the measured characteristics of one group and the parameters of another and vice versa. These predicted values are next used to decompose the differentials in 7 The approach is originally due to Oaxaca (1973) and Blinder (1973). A previous study of the World Bank by Oaxaca and Ransom (1993) uses the methodology to explore current salary differentials and promotion rates for staff on board in 1988 and 1992. The present study expands the analysis to examine entry level salaries, growth in pay, current and entry grade and the rate of promotions. It has access to somewhat richer data and the more recent data also allows an exploration of the effects of recent hiring policies on gender and nationality differentials. 6 mean salaries across groups into three component parts. For example, the salary differential between Part I men (subscripted by ml) and women (fl) is expressed as: (2) lnW*mi - lnW*fl = amI(X*ml - X*fl) + X*fln( ml - a) [Total difference] [Characteistics] [Structure] where the lnW*s and X*s represent the predicted mean (log) current salaries and the mean characteristics of the respective groups. This is simply an identity derived from the original regression model. The first component is the salary differential attributable to differences in the observed characteristics of the groups. The second is that attributable to between-group differences in the returns to given individual characteristics; it is in effect, the extra income Part I women would expect to get if Part I men's parameters were used to value Part I women's characteristics. The latter is usually referred to as the difference due to "structure". If the underlying model is correctly specified, then structure is interpretable as discrimination relative to what Part I men are currently paid. If the model is incorrectly specified-for example if there are omitted variables correlated with gender-then the structure component could also pick up the effects of misspecification. To the extent that there are important omitted variables, the coefficient estimates will be biased. If the biases are small or about equal between staff groups, this would not be problematic as they would cancel out. In general, as will be evident below, our models predict current salaries and growth in pay quite well, though biases in the underlying parameter estimates of the decompositions due to omitted variables or endogeneity can never be ruled out conclusively. 8This is often referred to as the Oaxaca decomposition. Blinder (1973) proposed essentially the same technique simultaneously to Oaxaca (1973). 7 In the example above, we used Part I men as the reference wage structure against which discrimination is measured. Why Part I men? As has been pointed out by Neumark (1988), the appropriate reference should depend on the underlying nature of discrimination. This is not known or discernible from the data. One argument for choosing Part I men is that, in the absence of discrimination, the Part I male wage structure would prevail. We do not, of course, know what the wage structure would be like in a different, no-discrimination, situation. However, given that Part I men have historically dominated the World Bank staff and that they remain the largest staff group (accounting for 45 percent), it might be argued that this reference is the most defensible. It is the main reference group we will use throughout the analysis. We test the sensitivity of the decomposition results to this assumption by also using the coefficients from a pooled regression model as the reference wage structure, as suggested by Neumark (1988) and Oaxaca and Ransom (1994). The decomposition given in equation (2) then takes the form: (3) lnW*mi - lnW*ti = 'Ž(X*m - X*fn) + [X*mI( aml - X) - X*fl( a- a [Total difference] [Characteristics] [Structure] where the cc 's are the coefficients obtained from the pooled regression and represent the non- discriminatory wage structure. We also examine the differences in initial salary and differences in growth in pay since entry. The equation for annual salary growth is: Alternatively we could use a population weighted combination of the parameters for each sub-group as the reference as in Cotton (1988). As Oaxaca and Ransom (1988) discuss, the pooled model doesn't constrain the wage structure parameters to lie within the range of the existing structures as does the Cotton method. Oaxaca and Ransom see this as an advantage of the pooled model. The pooled model will also give lower standard errors for the weighted coefficients. Finally, a problem arises in implementing the Cotton method when there are non-overlapping characteristics across groups. For example, in the present case we include certain important nationalities -such as US-as controls in 8 (4) rij = XijYj + vj where rij is the average annual rate of growth of salary for individual i in group j (given by (W/ Wo) - 1, where t is tenure, W is current salary and Wo is the initial salary, and all are indexed to the individual and group). As before, X is a vector of characteristics that are thought to affect salary at the time of entry and over time, and v is a zero mean innovation error. Similarly to equation (2) above, we then decompose the differences in average salary growth for pairs of groups. For example, for the decomposition between Part I men and Part I women, we calculate: (5) r*ml - r*fl = Yml(X*ml - X*fl) + X *fl( Yml - rl), [Total difference] [Characteristics] [Structure] where the Y 's are the coefficients from the estimated equations; and (as before) the "*"s denote mean values. The identity giving the growth rate as a function of current wage, entry wage and tenure allows us to back out entry salary predictions and their decompositions. ' For example, the predicted log of entry salary for a given group (such as Part II women if they had the parameters of Part I men) is given by the mean (log) current salary of that group minus the mean for that group of t ln( 1 + re) where t is tenure and r is the predicted growth rate of each individual in that group. Once the predicted log entry salary has been calculated for all groups (including for all combinations of characteristics and parameters), then the decompositions can be derived as before. the regressions. But, clearly, there are no US nationals in the Part II groups. In principle this means that there are unobserved coefficients and it is unclear what to do in taking weighted averages. '0 Alternatively, we could run an initial salary regression. However, as we explain shortly, pre-entry and at time of entry information is sparse. 9 We repeat the above analysis for current grade, grade at entry, and rate of promotion. Except for the dependent variable, the model specifications and decompositions are the same and so need not be repeated. An issue we are unable to deal with because of lack of data is sample selection bias." This bias could potentially come from two sources. The universe of staff who are currently active is the result of hiring choices and staff attrition. To assess the existence and extent of bias in pay and grade levels and promotion, it would be necessary to examine time series data. Full histories for the staff who are currently employed at the Bank are not sufficient, because they leave out others who have already left the Bank. Although potentially a problem, we think that it is not a major one: attrition tends to be low. For example, excluding the last three years, attrition at the Bank has been around 3 percent per year. Variables This section summarizes the variables used in the analysis and described more fully in Appendix Table 1. The variables comprising X are dummy variables for the year of entry to Bank employment (this captures tenure as well as any circumstances specific to the year of appointment), age at entry (entered in linear and squared forms), dummies for maximum completed level of education (doctorate, masters and bachelors degrees), dummies that interact the education degree with its country of location and discipline, dummies for the university where economics PhDs wvere obtained, number of years of work experience prior to joining the World Bank, current marital status, dummy variables for current nationality, and current "This bias might affect both the salary and grade analysis. 10 affiliation within the Bank (operations; central vice-presidential units related to research, knowledge management and dissemination; or others, primarily support units such as accounting, personnel, and finance)." Several additional factors that are thought to influence salary and promotion are not available to us. For example, while data on pay, grade, and departmental affiliation are routinely updated in computerized files, updated information on education completed, in-Bank and other training, language skills and so on, are self-reported and so contain gaps and are difficult to verify. Further, qualitative information about performance and annual performance merit ratings are strictly confidential and not available to us. Omitted variables also include pre-entry or time-of-entry information. For example, we are not able to control for nationality and marital status at entry or for the first hiring unit within the Bank. For some individuals these attributes will have remained fixed, for others they will have changed. The data on pre-Bank work experience are also limited. The length and nature of one's previous work experience as well as previous salary are probably good indicators of one's alternative labor market. Although, as mentioned, we include years of previous work experience in our regressions, we do not have data on salary or location for the previous work. Oaxaca and Ransom (1993), who had data on the location of previous work experience, found only a marginally significant salary effect of experience outside the U.S. for Part I men (1.8 percent) and of multi-country experience only for Part II men (2.5 percent). Their finding would suggest that previous experience does not count for much in Bank initial salaries. Hence, it is probably 12 only indication of pre-Bank work experience is in the time elapsed between the completion of a staff member's last degree and his or her start date at the World Bank. Because some staff finalize their degree after joining our variable for pre-Bank work experience is sometimes negative. A dummy 11 true that the farther away the origin of a job applicant is from Washington, DC, the more imperfect the information about the alternative labor market for a potential hire. As will become apparent below, we find that this imperfect information and possibly stereotypical assumptions about the alternative labor markets of particular groups of staff are a crucial part of the story about the patterns of entry salary and grade differentials at the Bank. Salary growth is presumably the result of staff performance over time. Yet the size of the annual increase associated with the performance rating is well-structured, having a predetermined ceiling each year. Variation in performance is presumably predicted in part by characteristics already observed at entry and in part by those unobserved at entry (not closely correlated with observed characteristics) which become apparent to managers with time. Performance is also affected by internal job choices made by staff throughout their employment. The willingness and ability to travel for many months each year, to learn and master a new language quickly, or to manage a difficult or politically sensitive task are examples of these choices. It is also affected by the willingness of managers to give more responsibility and more rewarding assignments to certain staff. Lastly, the proportion of time spent in operational (lending) versus non-operational (e.g., personnel, accounting, research) jobs can affect one's long-term career at the World Bank. We do not have a history on these choices, but we do include variables on current job affiliation within the Bank. variable captures this. We have singled out certain nationalities that consistently appear to be different in their effect and which each make up at least one percent of total staff. 12 C. Results Current Salary The decomposition results of predicted current salaries using the wage structure of Part I men as the reference are given in Table 3. The total difference in salary is given in the first column, followed by the component explained by characteristics and that attributable to structure. Standard errors for each of the two components are given below in parentheses. The underlying regressions are given in Appendix Table 2. Between 43 and 59 percent of the variation in the logs of current salaries is explained by the model, with the regressions for men having higher explanatory power than those for women. An F-test strongly rejects the hypothesis that the regression coefficients are the same across the four staff groups. Among Part I staff, women are paid 14.2 percent less than men. Of this difference, 8.6 percentage points are explained by women having different characteristics than men, but the remaining 5.6 percentage points (or 39 percent) are attributable to structure i.e., the same characteristics being valued differently for men and women. Controlling for Part II origin, women's salaries are still 14.3 percent less than men's, with 6.9 percentage points (48 percent) of this gap being due to structure rather than to characteristics. Controlling for gender, Part II men are paid 5.0 percent less than Part I men, of which 3.9 percentage points (78 percent) are due to structure; and Part II women are paid 5.1 percent less than Part I women, all of which is due to structure. The largest between-group difference is that between Part II women and Part I men -Part II women are paid 19.3 percent less on average than Part I men, with 10.6 percentage points (53 percent) being due to structure, not characteristics. How do these results differ when we use the alternative pooled model coefficients as the reference wage structure? (The pooled regression model results are given in the last column of 13 Appendix Table 2). As can be seen also in Table 3, the portion of the difference in current salaries that is attributed to structure tends to be smaller under the alternative reference, though the differences are not large. For example, the largest change is that for the decomposition between Part I and Part II women. The previous unexplained difference of 5.1 percent declines to 2.5 percent. In one case-that comparing Part II men and Part I women-4he unexplained component rises, from 1.7 to 2.5 percent. Our current salary model controls for more variables than the one estimated by Oaxaca and Ransom (1993) for the World Bank in 1992. To check the consistency of our findings with theirs and to calculate any shift in salary patterns, we also estimated a simpler model that matches theirs closely. The results from this simpler model differ little from those presented here and are thus not reported. They suggest that total differences relative to Part I men have declined somewhat since 1992, and more so for women than for Part II men. But more importantly, the differences that cannot be explained by individual characteristics and thus are attributed to structure, have dropped considerably for women, especially from Part I countries, and slightly for Part II men. The difference due to structure relative to Part I men decreased by 43 percent for Part I women, by 28 percent for Part II women, and by 10 percent for Part II men. Entry Salary and Growtht in Pay Between-group differences in salaries at entry are larger than in the case of current salaries, but follow the same overall pattern (Table 4). The results indicate that 8.7 percentage points of the 17 percent difference between the initial salaries of Part I men and women are attributable not to differences in observed characteristics for women but to structure. Similarly, had individual characteristics been valued in the same way for Part II women as for Part II men, 14 Part II women's starting salaries would have been, on average, 7.7 percent higher. The highest total, and unexplained, differences at 25 and 15 percent, respectively, are again between Part I men and Part II women. Comparing these results to those obtained when we use the pooled sample coefficients to weight the differences, we find again that the differences attributable to structure are generally reduced (Table 4). There is again one exception: the decomposition for Part II men versus Part I women gives a larger unexplained component under the pooled regression coefficienats (3.7 versus 1.8 %). Appendix Table 3 gives the regressions for annual growth in salaries. The explanatory power of the models is also high with from 55 to 69 percent of the variation explained. The variables controlled for are the same as in the current salary regression. The relatively high explanatory power of these models indicates that many of the factors that matter to annual increases in salaries are being captured despite the fact that we have been unable to include measures of actual performnance. Interestingly, the year of entry dummies suggest that the salary boom of the 1980s was mainly enjoyed by men, but that the salary repression of the 1990s was shared by all. Table 5 shows the decompositions of the differences in annual raises between groups. Overall, very small differences are observed. The negative signs mean that the raises are smaller for the first named group than for the second. For example, the average rate of annual salary growth is higher for Part II men than for Part I men in the first between-group comparison. These results suggest that some "catching up" occurs for those receiving lower entry salaries. The magnitude of this catch up, however, is quite small; for example, just 0.6 percent annually 15 for Part I women relative to Part I men, of which 0.2 percentage points is due to structure.'3 Because of the meager differences in salary growth across the groups, between-group differences in entry salaries continue to dominate the patterns in current salaries, even in a sample with a mean tenure of 12 years. The same conclusions hold if we use the pooled regression model coefficients as reference instead (Table 5). Current Grade Current grade regressions are presented in Appendix Table 4. The explanatory power of the model varies from 30 percent for Part I women to 43 percent for Part II men. As with the salary regressions, the model controls for age and date of entry, level of education, dummy variables that interact the level of education with its location and discipline, universities where economics PhDs were obtained, years of work experience prior to joining the World Bank, marital status, dummy variables for nationality, and internal group affiliation. Table 6 reports on the between-group grade differential decompositions. Controlling for Part I origin, women are 0.7 of a grade lower than men, on average; 0.2 (28 percent) of this gap is attributable to structure. Controlling for Part II origin, women are 0.7 of a grade lower, with 0.3 (43 percent) of this difference being due to structure. Controlling for gender, Part I and Part II men are 0.2 of a grade apart, while Part I and Part II women are 0.3 of a grade apart, with these gaps being almost entirely unexplained by observed characteristics. Similarly to what was found for salaries, the greatest difference is observed between Part I men and Part II women. The gap is close to a full grade. About half of this difference is attributable to characteristics being valued 13 We note that part or all of this catch-up is built into the structure of annual salary increases which stipulates that of staff who perform at the same level of satisfaction, those in lower salary levels get a 16 differently for men and women, meaning that Part II women with identical observed characteristics to Part I men are on average employed at approximately half a grade lower. Table 6 also presents the results using the pooled sample coefficients to weight the decomposition. The components attributable to structure are found again to be smaller than those which use Part I men as the reference, with the one exception of the Part II men versus Part I women comparison where the unexplained term is larger (0.07 versus 0.01 of a grade). Entry Grade and Promotion In general, women enter at significantly lower grades than do men, as do Part II staff relative to Part I staff (Table 7). Focusing on differences attributable to structure, women enter at between 0.7 (Part I) and 0.9 (Part II) of a grade lower than do men. Part II men enter at 0.4 of a grade lower than do Part I men, and Part II women 0.5 of a grade lower than Part I women. The largest unexplained gap is that between Part I men and Part II women and is over one full grade. Turning next to promotion patterns, the data indicate that, on average, men can expect a promotion of about 0.19 of a grade each year; Part I women, 0.23; and Part II women, 0.25. These have direct implications for the number of years it takes to advance bv one grade. We find considerable variance in the promotion models' explanatory power ranging from 16 to 49 percent of total variation explained, for Part I men and Part I women respectively (Appendix Table 5). The lower explanatory power of the models is somewhat inconsistent with the results we obtained for the salaly growth regressions, since in neither case do we have satisfactory measures of performance. slightly higher merit pay increase. 17 Table 8 shows between-group promotion decompositions. Differences in predicted average promotions between groups are small. The signs are negative with respect to the numeraire, meaning that the promotion rates are higher than would have been obtained by the numeraire group of similar characteristics. Again focusing on the differences due to structure, we see that Part I and Part II women have been catching up at the rate of 0.02 and 0.03 of a grade per year, respectively, relative to the corresponding male groups. Part II women have also been catching up at the rate of 0.02 of a grade per year relative to Part I women, and 0.04 of a grade relative to Part I men. Yet, these slightly faster promotions are not sufficient to compensate for the differences in grades at entry that we estimate to be attributable to different valuations of identical included characteristics. In the case both of entry grade and promotion, the decompositions obtained using the pooled model parameters as reference give results consistent with those discussed above. In general the differences attributable to structure diminish. D. Sources of Bias In sum, the above analysis indicates that average salaries and grades differ significantly between men and women and between Part I and Part II staff. While these differences are partly explained by differing characteristics (such as education, age and tenure), a notable portion cannot be so explained. Although some catching up occurs after entry, salary differences at the time of hiring largely dictate inter-group differences in current pay. And current salary differences are absolutely larger except between Part I and Part II men. The pattern in grades is similar to salaries, with disparities in grade assignment at entry largely explaining variation in current grade. Well over one-half of a grade difference between men and women at entry is not 18 explained by differences in their mean observed characteristics. Qualitatively similar results are obtained when we use the pooled model parameters as the reference weights rather than Part I men's wage structure parameters. Do the differences in salaries and grades across staff groups-between men and women, and between Part I and Part II staff-that are not explained by differences in observable characteristics indicate discrimination in hiring and promotion? Omitted variables might explain these results, provided that there is a reason for the bias to be consistently stronger in one group (say Part I males) than in another. Recall that the difference in salary and grade attributed to structure is given by the difference between the mean characteristics of group j evaluated using j's set of estimated coefficients and their value using group k's coefficients. The effects of education and previous experience (for example) may be influenced by characteristics such as the quality of education or individual ability which are omitted from our model but were observed at time of hire, promotion or salary raise.'4 We are limited by the data that are available to us. We are able to include level of education completed, country and in some cases, university where degree was obtained-all likely to capture aspects of education quality and individual aptitude. But variables such as other aspects of education quality (controlling for country and university where the education degree has been completed), potential earnings in alternative jobs, and communication skills in various languages, are still missing and may well be correlated with the included variables. If for some reason, the distribution of such omitted characteristics is such that Part I men are relatively better endowed 4 For example, in the human capital economics literature, including a measure of education quality in earnings functions has been found to result in a reduction in the return usually attributed to years of education (Behrman and Birdsall, 1983). Similarly, ability results in a downward adjustment on the return to education (Griliches and Mason, 1972). 19 with them and hence have higher expected productivity, then the returns to their education (for example) would appear to be higher than for other staff groups or (equivalently) they will have higher salaries at the same levels of observed characteristics. This would not, however, be discrimination. A relatively recent hiring policy is worth discussing here. As mentioned earlier, the Bank aims for a staff mix representing the different cultural and social backgrounds and languages of member countries. One consequence has been some tightening in the recruitment of U.S. men within the past decade. ' Such a "quota" might be expected to result in more selective hiring of U.S. men, which could in turn imply an increase in the quality of recent U.S. male hires, assuming no change in the applicant pool. However, binding quotas alone cannot explain our results. Quotas change the attributes of recruits at the time of hiring and this will be reflected in salaries. Some of those attributes are observed by the researcher, others are not but may have been observed bv the hiring manager or the personnel department. Explaining differences due to structure in the face of a quota still rests on there being omitted characteristics in our model which systematically apply to those hired under quotas, but not to others. There is one test. If the U.S. male quota entailed a higher quality in terms of variables that are unobserved by us, but were observed at hiring, then entry salary regressions for Part I men would show a positive coefficient for the dummy variable on U.S. nationality, holding all other included characteristics constant. We ran entry level regressions to test this hypothesis, including many of the characteristics that are included in the general model minus variables that were clearly not present at entry. We found the coefficient on '5 U.S. men comprise 16 percent of all staff in grades 22-30. 20 U.S. nationality to be: negative (but insignificant) relative to Part I men from non-English- speaking countries.'6 Another feature of recent hiring policy has been more aggressive recruitment of women, especially from Part [I countries. Following the same logic, this policy might imply a less selective hiring process for Part II women and thus lower quality of hires than before, holding the applicant pool constant. The actual quality of recent female and U.S. male recruits also depends on the effect of these hiring policies on the quality of the applicant pool itself for either group.'7 If the quota on U.S. men is widely known inside and outside the World Bank and discourages many from applying, it could reduce the overall quality of U.S. male applicants. Hence, the more selective process could result iLn lower, not higher, quality of recent U.S. male hires relative to earlier recruits. This seems unlikely, though it is theoretically possible. Similarly, and more plausible, if the more aggressive recruitment of women results in a much larger pool of applicants from many more countries, it could lead to higher, not lower, quality of recent women recruits relative to earlier cohorts. While we cannot prove that any of these factors are actually at work, we let the data inform us in a later section by comparing two cohort groups-those hired between 1980- 1986 and those between 1990-1997-to assess the implication of these recent hiring strategies on salary and grade differentials. Another possible explanation for differences attributable to structure is that discussed in the introduction: controlling for observed individual characteristics, the alternative labor markets 16 One factor that makes it difficult to draw conclusions from this is that other Part I male recruits who are not "local" hires might be paid a premium to induce them to move to Washington. Note that the same dummy variable has a larger and more significant negative coefficient for U.S. women. 21 for women and Part II staff relative to Part I men are, on average, lower-paying, and this may be partially reflected in the World Bank's remuneration offers. The ability of the employer to pay these lower wages is reinforced by a lack of information on the part of new hires-particularly ones from very different labor markets-and cultural factors. For example, there are many countries where salaries are fixed by government, and few countries where wage bargaining is as culturally accepted, and expected, as in the U.S. However, if the World Bank is paying some employees less because they are willing to work at a lower rate even though they are equally productive, that is discrimination if the stated institutional goal is equal pay for equal work. A third explanation is bias on the part of individuals. This is likely to be reinforced by imperfect information. As noted in the introduction, there are two major paths to getting ajob at the World Bank. One is through the Young Professionals (YP) program which is considered further below. The less formal process includes open recruitment-over which Personnel presumably maintains substantial power to enforce institutional objectives-and another which allows line managers to identify, put forward and champion individuals of their choice. Here, discrimination may be reintroduced as managers follow their preferences for working with and better rewarding individuals they may already know well or for whom they have an affinity through being, for example, from the same part of the world and/or the same gender. So far, the discussion has focused primarily on entry salaries and grades. This is because the patterns in current salaries and grades are dominated by those established at entry. In the next sections, we take different cuts of our sample in the hope of disentangling the different explanations for the pay and grade differentials that we attribute to structure. We first compare 7 Developments in alternative labor markets will clearly also affect the applicant pool. 22 YPs and non-YPs in order to learn more about the impact of the hiring process on pay and grade. We then compare recent staff cohorts in order to assess structural changes in the patterns of pay and grade differentials. We also examine how the recruitment quota on U.S. men might have affected these patterns. E. Extensions YPs Versus Non- YPs Table 9 presents descriptive statistics on salaries (indexed to be 100 for Part I men) for YPs and non-YPs, analogously to Table 1. Non-YP current salaries follow a pattern similar to that found for the full sample: Part I men are paid more than Part II men who are paid more than Part I women who, in turn, get more than Part II women. YP mean current salaries also fit this pattern, though with a larger spread. But not salaries at entry. Interestingly, YP entry salaries are highest for Part II women, followed by Part I women, Part I men and lastly, Part II men. But here the differences are small-the widest differential (between Part II women and men) is of 7.2 percentage points. YPs do not all begin with the same salary. Although the starting range is narrower than for others, entry salaries are also determined based on education, work experience and other factors. In contrast, non-YP staff start salaries reveal much greater absolute differences-the largest, between Part I men and Part II women, is of 27.3 percentage points-and follow the same pattern as for current salaries. Mean annual pay growth is higher for YP men than for YP women, while the reverse is found for the non-YP staff where women's annual salary raises exceed those of men. 23 Table 10, analogously to Table 2, provides summary statistics on grades for YPs. Both Part I and Part II men who were recruited as YPs tend to be concentrated in grades 24 through 26. In contrast, the greatest concentrations of women who were YPs are at grades 22 through 24. This difference is probably due, at least in part, to the lower mean tenure of YP women (1 1 rather than 16 years) resulting from the relatively recent efforts to hire at least an equal share of women in each YP cohort (Table 9). Using the same empirical models as above, we now explore the reasons for the noted differences in salary and grade in these two subsamples. Note that the remainder of the decomposition analysis uses Part I men as the reference wage structure. Table 11 gives the decompositions of mean current salaries of YPs and non-YPs. The underlying regressions are given in Appendix Tables 6 and 7. The model's explanatory power is uniformly high across the YP regressions, though less so for the non-YP groups, particularly the women. There is more unobserved variance for non-YPs. Part I and Part II YP women are predicted to be paid 17 and 20 percent less, respectively, than Part I YP men. Around 13 to 15 percentage points of this difference is explained by differences in characteristics. This leaves a 3.8 percentage point (29 percent) difference that is due to structure for Part I women and 5.1 percentage points (34 percent) for Part II women. There is little total difference, and no difference due to structure, between Part I and Part II YP men. We note much higher differences that are not due to characteristics in mean current salary predictions for the non-YP sample. For Part I women, 5.7 (42 percent) of 13.7 percent total difference relative to Part I men is attributable to structure, not characteristics; for Part II men, it is 5.2 (82 percent) of 6.3 percent; and for Part II women, 10.7 (55 percent) of 19.4 percent total difference. 24 For both the YP and non-YP samples, our salary growth regressions explain from 53 to 72 percent (Appendix Tables 8 and 9). Among YPs, starting salary differentials across the four groups are negligible (Table 12). Identical characteristics are essentially rewarded in the same way. There is very little difference that is not explained by observed characteristics. In contrast, for non-YPs (Table 12), we find even larger differences in entry level salaries than were found for the entire sample, with generally larger unexplained components as well. The patterns across gender and nationality groups are very similar to the ones we have previously seen. For example, of the total difference in mean entry salaries of 22 percent for Part I women relative to Part I men, 9 percentage points are due to structure. For Part II men, this portion is 9.2 of 10 percent; and for Part II women, 17 of 32 percent. Turning to growth in pay over time, there is evidence of a catching-up process or convergence for non-YPs over time (Table 13). On average, those with initially low salaries receive higher increases. There are still differences in growth rates controlling for characteristics but there is some catching up. For example, non-YP Part II women have a mean annual salary raise of 3.9 percent, on average. This would have been slightly lower at 3.7 percent, if evaluated using the coefficients of Part I men. Among YPs, the exact opposite is found (Table 13). While entry level salaries showed negligible differences across staff groups, growth rates favor Part I and Part II men. The model suggests divergence in salaries over time. To take the same example as above, Part II YP women are found to have mean annual pay raises of 3.4 percent. But with the same characteristics, yet based on the coefficients for Part I men, the rate of salary growth would have been 3.9 percent. This indicates a complete reversal of the non-YP case. 25 Turning to grade and promotions, the norm is that all YPs enter at grade 21. After entry, however, the grades of YPs diverge. The difference in mean current grade of Part I YP men and women is about 0.8 of a grade. Once we have controlled for year of entry and individual characteristics, the portion attributable to structure is relatively small at 0.09 of a grade (10 percent) (Table 14). Among Part II staff, men are 1.1 grades higher, on average, than women, of which 0.4 of a grade (36 percent) is due to structure. Part II men are 0.1 of a grade higher than Part I men, with most of this difference being due to structure. Among women, Part I YP graduates are 0.2 of a grade ahead of Part II graduates for reasons other than controlled for characteristics. The difference attributable to structure between Part I men and Part II women YPs is only 0.2 of a grade (22 percent of the total difference), compared with over one-half grade in the full and non-YP samples. In the non-YP group, without the "equalizing" influence of the YP recruitment process, generally larger unexplained between-group differences emerge in current grades (Table 14). However, these differences are significantly smaller than for grades at the time of entry (Table 15), suggesting a readjustment over time as more information is revealed about staff's ability to perform (Table 16). For example, the difference due to structure between Part II men and women at entry is cut by more than one-half of a grade at mean tenure of about twelve years. To sum up our results on the YP and non-YP samples, the model suggests that the YP recruitment process largely eliminates inter-group differences in starting salaries for given values of the observed individual characteristics. Disparities in salaries at entry are roughly explained by differences in the attributes of new YPs. This is probably because the process of making offers to YPs is a more transparent and regulated one than for the non-YP hiring process. 26 Equally important, YPs are drawn from a very large pool of applicants each year and are likely to be quite homogeneous with respect to omitted characteristics as well, relative to non-YPs. After YPs have entered the common staff pool, however, a divergence in the salaries and grades of men and women emerges. This divergence could indicate that individual characteristics observed at entry by the program administrators, and on the basis of which YPs are chosen, (presumably a richer set of information than our model can account for) are not, after all, good predictors of subsequent performance. It is, however, implausible that this would only apply to women. Alternatively, women as a group may be making different internal assignment choices than men, for example to control levels of travel and stress in their jobs. However, this could presumably have been anticipated at time of hiring. Another possible explanation is that biases slowly emerge which render the work environment less conducive to the progress of women. Men may tend to receive more mentoring, benefit from greater numbers of role models and be favored in task assignments that have higher potential for promotion. In contrast to that for YPs, annual pay raises for non-YP staff appear to gradually compensate over time for the large unexplained differences in entry level salaries. Indeed, while unexplained differences for YPs are greater for current salaries, for non-YPs they are greater at entry level. The World Bank's YP hiring process can help test our model. As discussed above, our results for the total sample could be due simply to the influence of gender or nationality correlated variables that have been omitted from the model yet were observed at the hiring stage. In this case, we could be attributing to discrimination what are in actual fact returns to omitted variables. On the assumption that the YP process is non-discriminatory-in that it follows a 27 systematic and controlled vetting and checking process which successfully eliminates bias across a set of variables that can be observed by the YP program administrators but which we do not know fully-estimating our model on the YP subsample provides a natural test of whether we have included the right set of variables. Large differences attributable to structure rather than to characteristics would then indicate that key explanatory variables are missing in the underlying model. In fact, the YP model results indicate only small differences in mean salaries between groups that are not accounted for by characteristics. These differences can be interpreted as the effect of omitted characteristics on our estimates of the effect of structure. For example, under this interpretation, the return to the omitted characteristics of Part I men compared to Part I women would be 2.9 percent of entry level salary differentials (Table 12). If the omitted variables are similar for the YP and non-YP samples, this means that about 3 percent of the 22 percent premium paid at entry to non-YP Part I men compared to Part I women may be attributable to unobserved variables. The rest of the entry salary difference due to structure (about 6.4 percent) can then be attributed to discrimination. Tale of Two Cohorts We turn now to the analysis of two cohorts of Bank entrants, that is, individuals who entered between 1980 and 1986, and between 1990 and 1997. We focus on entry salaries and grades. Comparing these two groups is of interest for two reasons. It can help explore what impacts the Bank's recent more active efforts to increase diversity while treating staff fairly, may 28 have had and it allows us to further examine the possibility that implicit "quotas" explain our results. Between the two periods, we find an increase in average entry salaries for all groups: the increase is 31 percent for Part I men, 29 percent for Part II men, 47 percent for Part I women, and 46 percent for Part II women. Since the underlying salaries are in 1990 US dollars, increases reflect increases in the "quality" of new employees as well as changes in the structure of their labor market alternatives, not cost of living adjustments. Relative to Part I men, Part I women started at a 24 percent lower salary in the 1980-86 cohort (Table 17). (The current salary and growth in pay regressions from which these decompositions are derived are presented in Appendix Tables 14 and 15.) Of this difference, 8.1 percentage points (34 percent) are not explained by differences in characteristics. In the 1990-97 cohort, the salary differential is much reduced at 10 percent, of which only 4 percentage points (40 percent) are unexplained by differences in individual attributes. On average, Part II men started at 9.5 percent less than Part I men in the 1980-86 cohort, and this difference remains relatively constant at just under 9 percent in the 1990-97 cohort. However, while the difference for the earlier cohort is entirely attributable to structure, 3.1 percentage points are explained by differences in attributes for the latter, leaving 5.6 percentage points (62 percent) unexplained. Relative to Part II men, Part II women started at 21 percent lower salaries in the 1980-86 cohort, a difference largely attributable to differences in characteristics. In the 1990-97 cohort, the difference declines to less than 7 percent, with most of it still attributable to differences in characteristics. Relative to Part I women, entry salaries for Part II women were 7 percent lower in the 1980-86 cohort, with a difference of 2.6 percentage points (37 percent) that is attributable 29 to structure. In the more recent cohort, the difference is reduced to 5 percent, with 2.3 (46 percent) of that still due to structure. On average, Part II women started at 31 percent lower salary than Part I men in the 1980- 86 cohort. One-third of the difference was attributable to differences in structure. For the 1990- 97 cohort, the differential is reduced to about 15 percent, with just over one-third again not attributable to characteristics. As with salaries, mean entry grades for the later cohort are higher than for the earlier cohort. The difference, shown in Table 18, is close to a full grade for Part I men, 1.2 grades for Part II men, 1.8 for Part I women, and 2.1 for Part II women. The current grade and promotion regressions from which the decompositions are derived can be found in Appendix Tables 18 through 21. Table 18 shows the decompositions of the between-group differences in starting grades. A substantial reduction in the differences in entry grades of Part I women relative to Part I men has occurred between the two cohorts -from 1.5 grades to 0.5 of a grade. There is a similarly notable drop in the difference that is attributable to structure--from 0.5 of a grade to 0.1. The difference between Part II men and Part I men falls slightly from 0.6 grade to 0.4 grade, but where characteristics would have predicted a negative difference in the earlier cohort, the difference is evenly spread between that which can be explained by characteristics and that which cannot in the later cohort. Relative to Part II men, the deficit for Part II women falls from the earlier to the later cohort, from 1.4 grades to 0.5 grades, most of which is explained by differences in characteristics. Relative to Part I women, Part II women entered at 0.5 grades lower in the earlier 30 cohort, a difference lunaccounted for by differences in characteristics. By the 1990-97 cohort, the differential was reduced slightly to 0.3 of a grade, with 0.2 attributable to structure. Finally, Part II women tended to start at two grades lower than Part I men on average in the 1980-86 cohort, and 0.9 grade lower in the 1990-97 cohort. Again, the share that is explained by differences in characteristics rises, in this case from 50 percent to 62 percent. Finally, we omit U.S. men from the recent cohort of 1990-97 in order to take account of the tightening of the Bank's hiring policy with respect to U.S. men. If this squeeze is resulting in a significantly higher quality of recent hires of U.S. men which in turn accounts for larger returms to their observable characteristics relative to others, then we would expect to see a decrease in the unexplained difference in entry salaries relative to Part I men when we leave U.S. men out of the sample. We now compare Table 19 which presents the results for the 1990-97 cohort with and without U.S. men. We do find a slight decrease in the difference due to structure for Part I women, Part II men and Part II women relative to Part I men-O.2, 1.2 and 0.9 percent, respectively. One implication of this is that the tightening of recruitment for U.S. men has not really had much of an impact on the average quality of the recent U.S. male hires. Another is that the limitation on the hiring of U.S. men is occurring at a time when other structural changes are being made in hiring, salary and promotion policies that are meant to close some of the gap in entry salaries across staff groups that is not explained by observed characteristics. In sum, the cornparison of staff who were hired between 1980 and 1986 with those who entered the World Bank between 1990 and 1997 provides some evidence that Bank policies attempting to eliminate unwarranted differences are having the intended effects. The comparison 31 does not appear to support the view that quotas are driving our results. According to that hypothesis, quotas limiting the intake of Part I men and, in particular US men, together with those favoring higher Part II women hires, have led on average to higher quality Part I men hires and lower quality Part II women hires, and explain the large differences in salary and grade at entry not explained by our models. We have already shown that a substantial amount of the variation in rewards is explained by our model when it is run on the YP sample. We have also argued that for binding quotas to be responsible for the results, the omitted quality variables must favor Part I men most and Part II women least. We believe that for both these reasons, the quota story is not a compelling explanation for our results. The comparison of pre- and post-quota staff cohorts only reinforces this view. While the quotas came into force in the 90s, over time, between-group differences in entry salary and grades have diminished as has the component of that difference attributable to structure. G. Conclusions Our study has found differences in salaries and grades between men and women and between Part I and Part II staff in the Bank which are not readily accountable to differences in the attributes that one would expect to influence productivity. The difference in mean current salaries between men and women (holding nationality group constant) is 14 percent, almost half of which is structural in that it cannot be attributed to differences in observed characteristics such as age, education, and previous work experience. The differences in mean salaries between Part I and Part II staff (independently of gender) are much smaller-5 percent-and are largely attributable to structure rather than to differences in characteristics. 32 A comparison of Oaxaca and Ransom's (1993) earlier results with ours, and our own analysis of staff who entered between 1980 and 1986, show that overall salary and grade differences have declined, as have the components unexplained by differences in characteristics. Recent Bank policies appear to have successfully reduced differences across staff groups. However, significant unexplained differentials remain. The paper has examined a number of explanations. One is omitted variable bias, whereby there are characteristics omitted from our model but observed by those making salary and promotion decisions which are correlated with observed variables. For omitted variables to be contributing to the component of unexplained differences in salary and grade between staff groups, they must be distributed across groups such that the bias is strongest for Part I men, followed by Part II men and so on. Our analysis of the YP sample, inarguably a more homogeneous group hired through a non-discriminatory process, indicates that only negligible inter-group differences in mean salaries are left unexplained by the variables included in our model. Based on these results, we would argue that omitted variable bias is not a compelling explanation for differences attributable to structure. The YP sample also suggests that horizontal inequities can be eliminated through a different hiring process. A second possible explanation for salary and grade differentials which we are unable to explain is discrimination. We have identified two ways in which this could arise. In its attempt to reduce labor costs, the World Bank may be taking advantage of the fact that the four staff groups face different reservation wages. A third explanation puts more onus on individual managers and their preferences for discrimination coupled with imperfect information. We find that although YPs all begin at the same level and at salaries that differ only to reflect differences in characteristics, once they enter the general staff pool, their salary increases and promotions 33 begin to mirror the patterns found for the rest of the staff. This suggests that bias from both of these sources is present. Under a recently introduced system of dollar budgeting at the World Bank, the incentives to minimize labor costs on the part of managers, and thus to discriminate against certain groups, are likely to increase. Nonetheless, a dramatic shift does appear to be occurring in the hiring process at the Bank. The possible sources of this shift include: (a) on the supply side, that the applicant pool, especially of women and Part II candidates, has significantly improved in quality; (b) that information gathering during hiring has intensified, thus decreasing room for guess work and the opportunities for stereotypes to bias grade assignment and wage offers; (c) that a renewed effort to balance the composition of Bank staff has meant providing more incentives to staff from minority groups; and (d) that as the institution has become more diversified-in terms of gender and nationality groups-it has become more conducive to high performance by women and Part II nationals, as well as less biased. The result of the shift has been to reduce salary and grade differences that cannot be attributed to observed differences in attributes. 34 References Behrman, Jere and Nancy Birdsall. 1983. "The Quality of Schooling: Quantity Alone is Misleading, " American Economic Review 73(5): 928-946. Berger, Jacqueline. 1995. "Were you Referred by a Man or a Woman? Gender of Contacts and Labor Market Outcomes," Working Paper 353, Industrial Relations Section, Princeton University. Blinder, Alan S. 1973. "Wage Discrimination: Reduced Form and Structural Estimates," The Journal of Human Resources 8: 436-455. Cotton, Jeremiah. 1988. "On the Decomposition of Wage Differentials," Review of Economics and Statistics 70: 236-243. Griliches Zvi and William Mason. 1972. "Education, Income and Ability," Journal of Political Economy 80(3): S74-S 103. Hersch, Joni and W. Kip Viscusi. 1996. "Gender Differences in Promotions and Wages," Industrial Relations 35(4): 461-472 Jones, David and Gerald Makepeace. 1996. "Equal Opportunities: Pay and Promotion in an Internal Labour Market," The Economic Journal 106: 401-409. Lazear, Edward P. and Sherwin Rosen. 1990. "Male-Female Wage Differentials in Job Ladders," Journal of Labor Economics 8(1): S106-S123. Neumark, David. 1988. "Employers' Discriminatory Behavior and the Estimation of Wage Discrimination," The Journal of Human Resources 23(3): 279-295. Oaxaca, Ronald L. 1973. "Male-female Wage Differentials in Urban Labor Markets,' International Economic Review 9: 693-709. Oaxaca, Ronald L. and Michael Ransom. 1994. "On Discrimination and the Decomposition of Wage Differentials," Journal of Econometrics 61: 5-21. Oaxaca, Ronald L. and Michael Ransom. 1993. "Salary and Promotion Differences Between Gender and Nationality Groups for Specific Grade Levels," prepared for the World Bank, April. 35 Figure 1 Distribution of staff across grades by staff group 40, 35 30 25 ~20 C~~~~~~~~~~~~~~~~~Crd 15 10 5 0 22 23 24 25 26 27 28 29 30 Grade E pan I ren * Partllirmn [IPart Iwornen El Partllwonen 3 6 Table 1: Descriptive statistics on pay and tenure of professional staff, May 1997 Part I men Part II men Part I women Part II women Current salaries 100 95.3 86.6 82.5 Salaries at entry 100 91.4 86.4 80.6 Mean annual growth (%) 3.2 3.3 3.8 3.8 Mean tenure (years) 11.9 12.5 10.7 11.2 Number of staff 1356 857 537 250 (% of total staff) (45%) (29%) (18%) (8%) Note: Data on World Bank (IBRD) professional staff pertain to those in grades 22-30. Table 2: Descriptive statistics on grade of professional staff, May 1997 Grade 22 23 24 25 26 27 28 29 30 Total Part I Men % in grade currently 7 16 37 19 15 3 3 1 0 100 Mean years in grade 2.5 3.9 6.0 4.3 6.7 4.0 5.0 3.0 -- 5.09 N 98 214 506 254 197 36 36 15 0 1359 Part II Men % in grade currently 8 21 38 16 11 2 2 1 0 100 Mean years in grade 3.3 3.6 5.7 3.7 6.7 3.8 4.4 3.1 7.9 4.7 N 69 182 328 141 91 17 21 7 1 857 Part I Women % in grade currently 17 29 32 10 8 2 1 1 0 100 Mean years in grade 2.7 3.1 4.6 2.9 3.8 2.3 4.6 2.1 -- 3.6 N 93 158 174 51 42 11 5 3 0 537 Part II Women % in grade currently 22 33 31 7 5 0 1 0 0 100 Mean years in grade 2.4 2.9 4.7 2.9 4.5 1.5 2.4 -- -- 3.4 N 55 83 77 18 13 1 3 0 0 250 Total % in grade currently 11 21 36 15 11 2 2 1 0 100 Mean years in grade 2.7 3.5 5.6 3.9 6.3 3.6 4.7 3.3 7.9 4.6 N 315 637 1085 464 343 65 65 25 1 3000 37 Table 3: Between-group differences in current salaries Part I men reference Pooled parameter reference Total Characteristics Structure Characteristics' Structure' Differencea b M1-M2 5.0 1.06 3.93 2.54 2.45 (.688) (.883) (.401) (.401) MI-Fl 14.23 8.64 5.59 9.23 4.99 (.439) (.873) (.285) .644 MI-F2 19.33 8.75 10.58 11.87 7.46 (.722) (1.16) (.481) (.684) M2-FI 9.23 7.57 1.66 6.69 2.54 (.851) (1.10) (5.06) (.428) M2-F2 14.33 7.68 6.65 9.32 5.01 (.407) (1.14) (.315) (.838) F1-F2 5.10 0.11 4.99 2.63 2.47 (.745) (1.76) (.481) (.690) Notes: a - Percentage difference in mean annual salaries between indicated groups. b - Differences in mean salary attributable to differences in included characteristics. c - Differences in mean salary attributable to structure. b and c may not add up to a due to rounding off. Standard errors are given in parentheses. Table 4: Between-group differences in entry level salaries Part I men reference Pooled parameter reference Total Characteristics Structure Characteristics Structure Difference M1-M2 9.21 2.35 6.86 5.24 3.97 MI-Fl 17.33 8.66 8.66 9.67 7.66 M1-F2 24.64 10.04 14.61 14.81 9.83 M2-F1 8.12 6.32 1.80 4.43 3.69 M2-F2 15.43 7.69 7.74 9.57 5.86 FI-F2 7.32 1.37 5.94 5.15 2.17 38 Table 5: Between-group differences in mean annual salary growth Part I men reference Pooled parameter reference Total Characteristics Structure Characteristics Structure Difference MI-M2 -0.03 0.14 -0.16 0.01 -0.03 -______ ______(.001) (.001) (.000) (.000) MI-Fl -0.63 -0.42 -0.21 -0.43 -0.20 (.001) (.001) (.000) (.001) MI-F2 -0.52 -0.45 -0.07 -0.56 0.04 _________ (.001) (.001) (.001) (.001) M2-FI -0.60 -0.55 -0.05 -0.44 -0.17 (.001) (.001) (.001) (.001) M2-F2 -0.49 -0.58 0.09 -0.57 0.07 (.001) (.001) (.000) (.001) FI-F2 0.11 -0.03 0.14 -0.13 0.24 _ (.001) .(003) (.001) (.001) Note: Standard errors are given in parentheses Table 6: Between-group differences in current grade Part I men reference Pooled parameter reference Total Characteristics Structure Characteristics Structure Difference MI-M2 0.21 .01 0.20 0.09 0.12 (.053) (.068) (.030) (.031) MI-Fl 0.66 0.45 0.22 0.46 0.20 (.035) (.066) (.021) (.050) Ml-F2 0.93 0.39 0.54 0.57 0.36 (0.57) (.089) (.036) (.052) M2-Fl 0.45 0.44 0.01 .038 .07 (.067) (.081) (.037) (.031) M2-F2 0.72 0.38 0.34 0.48 .24 (.032) (.084) (.023) (.062) F1-F2 0.27 -0.06 0.33 0.11 0.16 (.059) (.129) (.036) (.050) Note: Standard errors are given in parentheses 39 Table 7: Between-group differences in entry grade Part I men reference Pooled parameter reference Total Characteristics Structure Characteristics Structure Difference MI-M2 0.50 0.15 0.36 0.26 0.24 MI-Fl 1.36 0.66 0.70 0.75 0.61 Ml-F2 1.96 0.72 1.24 1.00 0.96 M2-F1 0.86 0.51 0.35 0.49 0.37 M2-F2 1.46 0.57 0.89 0.74 0.71 FI-F2 0.60 0.06 0.54 0.25 0.34 Table 8: Between-group differences in annual rate of promotion Part I men reference Pooled parameter reference Total Characteristics Structure Characteristics Structure Difference M1-M2 -0.01 -0.01 -0.01 -0.01 -0.00 (.009) (.011) (.005) (.005) MI-Fl -0.06 -0.04 -0.02 -0.04 -0.02 (.006) (.010) (.003) (.007) MI-F2 -0.08 -0.04 -0.04 -0.05 -0.03 (.010) (.015) (.005) (.009) M2-F1 -0.04 -0.03 -0.02 -0.03 -0.01 (.01 1) (.011) (.006) (.004) M2-F2 -0.06 -0.03 -0.03 -0.03 -0.03 (.005) (.013) (.004) (.010) Fl-F2 -0.02 -0.00 -0.02 -0.01 -0.01 (.010) (.018) (.005) (.007) Note: Standard errors are given in parentheses 40 Table 9: Descriptive statistics on pay and tenure for YPs and non-YPs , May 1997 YPs Part I men Part II men Part I women Part II women Current salaries 100 99.5 83.3 81.2 Salaries at entry 100 97.8 103.6 105.0 Mean annual growth(%) 3.83 3.94 3.49 3.38 Mean tenure (years) 16.1 16.5 11.1 10.7 Non-YPs Part I men Part II men Part I women Part II women Current salaries 100 92.6 87.2 82.4 Salaries at entry 100 90.5 80.1 72.7 Meanannualgrowth(%) 3.15 3.19 3.90 3.92 Mean tenure (years) 11.1 11.4 10.6 11.3 Table 10: Descriptive statistics on current grades of YP sample 22 23 24 25 26 27 28 29 30 Total % at grade currently 14 6 23 23 23 4 7 3 -- 100 Mean years in grade 1.8 1.9 7.9 4.7 8.3 4.2 6.4 4.6 -- 5.7 N 30 14 50 50 49 9 15 6 -- 217 .....:TT .LEE:ii E2.................... ;..... . I ~.' .: .i , - . : i . :: : : : : : : : : :: : : : : i : :!~ ;iA: ; :S E E ...... i.. S4i0ihE ii,50EiSg .......... . . ... .ESi Pa I Wvomen::g;| ;:-.;;:- % at grade currently 23 21 25 12 13 3 5 -- 100 Meanyearsingrade 1.8 2.3 5.4 1.8 3.3 0.7 4.6 -- -- 3.1 N 25 23 28 13 14 3 5 -- - 111 .... ... . .. .. . .... ..... .... .... . .. ........ ... . ...................................... . Pr II M -- ;:iXi d::;enE::S;;tVS f ;i;:!;;giE E : % at grade currently 5 12 28 21 22 4 8 2 -- 100 Mean years in grade 1.7 2.8 6.8 3.6 6.1 5.1 4.7 4.0 -- 5.0 N 9 21 48 35 37 6 14 4 -- 170 11 .. iE.._.._iE........- % at grade currently 22 18 35 13 8 0 3 -- -- 100 Meanyears in grade 2.1 2.3 5.7 3.6 3.4 0 1.8 -- 3.7 N 13 11 21 8 5 0 2 -- -- 60 ~~~~~~~~~~~... . - i- a d .......... .. . . ... . .- . i l}?. l ... ................ R iE. LE. ... iiE. . ..... % at grade currently 14 12 26 19 19 3 6 1 - 100 Mean years in grade 1.8 2.4 6.8 3.9 6.6 3.9 5.2 4.4 -- 4.8 N 77 69 147 106 105 18 36 6 - 564 41 Table 11: Between-group differences in current salaries: YP and non-YP samples YP Non-YP Total Characteristics Structure Total Characteristics Structure Difference Difference Ml-M2 0.07 1.75 -1.69 6.33 1.11 5.22 (1.73) (2.06) (0.741) (0.965) MI-Fl 17.08 13.24 3.85 13.7 8.02 5.68 (1.23) (1.77) (0.478) (0.974) MI-F2 20.35 15.21 5.13 19.41 8.67 10.75 (1.81) (2.22) (0.846) (1.34) M2-FI 17.02 11.48 5.53 7.37 6.91 0.46 (2.08) (1.90) (0.940) (1.237) M2-F2 20.28 13.46 6.82 13.08 7.56 5.53 (1.43) (1.95) (0.518) (1.34) FI-F2 3.27 1.98 1.29 5.72 0.64 5.07 (1.88) (3.21) (0.834) (2.12) Notes: Standard errors are given in parentheses. Table 12: Between-group differences in entry salaries: YP and non-YP samples YP Non-YP Total Characteristics Structure Total Characteristics Structure Difference Difference M1-M2 0.20 4.35 -4.15 10.18 0.97 9.21 MI-Fl -5.21 -8.11 2.90 21.89 12.54 9.35 MI-F2 -6.43 -5.23 -1.20 31.99 15.16 16.82 M2-FI -5.41 -12.46 7.06 11.71 11.58 0.14 M2-F2 -6.63 -9.58 2.95 21.81 14.20 7.61 FI-F2 -1.22 2.88 -4.10 10.10 2.62 7.47 42 Table 13: Between-group differences in annual salary growth: YP and non-YP sample YP Non-YP Total Characteristics Structure Total Characteristics Structure Difference Difference MI-M2 -0.13 -0.31 0.18 0.03 0.27 -0.24 ________ ____ (0.001) (0.001) (0.001) (0.001) Mi-Fl 0.33 0.28 0.05 -0.86 -0.58 -0.29 _______ __ (0.001) (0.001) (0.001) (0.002) MI-F2 0.49 0.03 0.46 -0.77 -0.58 -0.20 (0.001) (0.002) (0.001) (0.002) M2-FI 0.46 0.59 -0.13 -0.89 -0.85 -0.04 (0.001) (0.001) (0.001) (0.002) M2-F2 0.62 0.34 0.28 -0.80 -0.85 0.05 (0.001) (0.002) (0.001) (0.001) FI-F2 0.16 -0.25 0.41 0.09 0.00 0.09 (0.001) (0.003) (0.001) (0.004) Notes: Standard errors are given in parentheses Table 14: Between-group differences in current grade: YP and non-YP sample YP Non-YP Total Characteristics Structure Total Characteristics Structure Difference Difference MI-M2 -0.12 0.05 -0.17 0.31 0.02 0.29 (0.152) (0.178) (0.059) (0.074) MI-Fl 0.78 0.69 0.09 0.66 0.43 0.23 (0.108) (0.157) (0.038) (0.071) MI-F2 0.98 0.75 0.23 0.96 0.43 0.53 (0.159) (0.203) (0.067) (0.101) M2-FI 0.90 0.64 0.26 0.35 0.41 -0.06 (0.182) (0.160) (0.074) (0.087) M2-F2 1.10 0.70 0.40 0.65 0.41 0.24 (0.125) (0.173) (0.041) (0.096) Fl-F2 0.20 0.06 0.14 0.30 0.003 0.30 _____ __ (0.165) (0.296) (0.066) (0.146) Notes: Standard errors are given in parentheses 43 Table 15: Between-group differences in entry grade: YP and non-YP sample YP Non-YP Total Characteristics Structure Total Characteristics Structure Difference Difference M1-M2 n/a n/a n/a 0.58 0.05 0.54 MI-Fl n/a n/a n/a 1.68 0.91 0.77 MI-F2 n/a n/a n/a 2.42 1.02 1.39 M2-FI n/a n/a n/a 1.10 0.86 0.24 M2-F2 n/a n/a n/a 1.83 0.98 0.85 Fl-F2 n/a n/a n/a 0.74 0.12 0.62 Note: Since all YPs enter at grade 21 there are no differences to decompose. Table 16: Between-group differences in annual promotion rate: YP and non-YP sample YP Non-YP Total Characteristics Structure Total Characteristics Structure Difference Difference MI-M2 0.03 0.03 0.000 -0.02 -0.003 -0.01 (0.016) (0.018) (0.008) (0.011) MI-Fl -0.01 0.002 -0.01 -0.06 -0.03 -0.02 (0.011) (0.016) (0.005) (0.010) M1-F2 0.002 0.005 -0.003 -0.08 -0.03 -0.05 (0.017) (0.023) (0.010) (0.017) M2-F1 -0.04 -0.03 -0.01 -0.04 -0.03 -0.01 (0.019) (0.011) (0.011) (0.013) M2-F2 -0.03 -0.03 -0.003 -0.06 -0.03 -0.03 (0.013) (0.016) (0.006) (0.016) F1-F2 0.01 0.003 0.01 -0.02 0.002 -0.02 (0.017) (0.026) (0.009) (0.022) Notes: Standard errors are given in parentheses 44 Table 17. Between-group differences in entry salary: 1980-86 and 1990-97 cohorts 1980-86 Cohort 1990-97 Cohort Total Characteristics Structure Total Characteristics Structure Difference Difference MI-M2 9.54 -1.86 11.40 8.75 3.13 5.62 MI-Fl 23.82 15.70 8.12 10.29 6.27 4.02 Ml-F2 30.86 20.12 10.74 15.34 9.02 6.32 M2-FI 14.28 17.56 -3.28 1.54 3.14 -1.60 M2-F2 21.32 21.98 -0.66 6.59 5.89 0.70 FI-F2 7.04 4.42 2.62 5.05 2.75 2.29 Table 18. Between-group differences in entry grade: 1980-86 and 1990-97 cohorts 1980-86 Cohort 1990-97 Cohort Total Characteristics Structure Total Characteristics Structure Difference Difference MI-M2 0.56 -0.27 0.83 0.42 0.21 0.21 Mi-Fl 1.52 0.97 0.55 0.55 0.42 0.13 M1-F2 2.00 1.02 0.97 0.88 0.55 0.33 M2-FI 0.96 1.24 -0.28 0.13 0.22 -0.09 M2-F2 1.44 1.29 0.15 0.46 0.34 0.11 FI-F2 0.48 0.05 0.43 0.33 0.13- 0.20 45 Table 19. Between-group differences in entry salary: 1990-97 cohort, with and without U.S. men 1990-97 with U.S. Men 1990-97 without U.S. Men Total Characteristics Structure Total Characteristics Stiucture Difference Difference MI-M2 8.75 3.13 5.62 8.93 4.63 4.30 Ml-Fl 10.29 6.27 4.02 10.46 6.67 3.80 M1-F2 15.34 9.02 6.32 15.51 10.10 5.41 M2-Fl 1.54 3.14 -1.60 1.54 2.04 -0.50 M2-F2 6.59 5.89 0.70 6.59 5.48 1.12 F1-F2 5.05 2.75 2.29 5.05 3.43 1.61 46 Appendix Table 1: Variable definitions and summary statistics Variable Description Part 1 Men Part 2 Men Part 1 Women Part 2 Women Pooled Sample Mean Std.Dev. Mean Std.Dev. Mean Std.Dev. Mean Std.Dev. Mean Std.Dev. dyOe8O l=entered in 1980 0.0435 0.204 0.0443 0.206 0.0391 0.194 0.0320 0.176 0.0420 0.201 dyOe8l 1=entered in 1981 0.0258 0.159 0.0467 0.211 0.0261 0.159 0.0360 0.187 0.0327 0.178 dyOe82 1=entered in 1982 0.0310 0.173 0.0222 0.147 0.0242 0.154 0.0600 0.238 0.0297 0.170 dyOe83 1=entered in 1983 0.0288 0.167 0.0362 0.187 0.0279 0.165 0.0320 0.176 0.0310 0.173 dyOe84 1=entered in 1984 0.0361 0.187 0.0350 0.184 0.0279 0.165 0.0320 0.176 0.0340 0.181 dyOe85 1=entered in 1985 0.0435 0.204 0.0513 0.221 0.0428 0.203 0.0160 0.126 0.0433 0.204 dyOe86 1=entered in 1986 0.0339 0.181 0.0455 0.209 0.0317 0.175 0.0200 0.140 0.0357 0.185 dyOe87 1=entered in 1987 0.0273 0.163 0.0233 0.151 0.0168 0.128 0.0280 0.165 0.0243 0.154 dyOe88 1=entered in 1988 0.0642 0.245 0.0665 0.249 0.0596 0.237 0.0440 0.206 0.0623 0.242 dyOeB9 1=entered in 1989 0.0428 0.202 0.0513 0.221 0.0261 0.159 0.0440 0.206 0.0423 0.201 dyOe9O 1=entered in 1990 0.0347 0.183 0.0443 0.206 0.0484 0.215 0.0400 0.196 0.0403 0.197 dyOe91 1=entered in 1991 0.0450 0.207 0.0618 0.241 0.0540 0.226 0.0400 0.196 0.0510 0.220 dyOe92 1=entered in 1992 0.0524 0.223 0.0688 0.253 0.0521 0.223 0.0720 0.259 0.0587 0.235 dyOe93 1=entered in 1993 0.0605 0.238 0.0420 0.201 0.0950 0.293 0.0960 0.295 0.0643 0.245 dyOe94 1=entered in 1994 0.0575 0.233 0.0338 0.181 0.0875 0.283 0.0680 0.252 0.0570 0.232 dyOe95 1=entered in 1995 0.0354 0.185 0.0163 0.127 0.0298 0.170 0.0440 0.206 0.0297 0.170 dyOe96 1=entered in 1996 0.0494 0.217 0.0362 0.187 0.0484 0.215 0.0280 0.165 0.0437 0.204 dyOe97 1=entered in 1997 0.0170 0.129 0.0035 0.059 0.0261 0.159 0.0280 0.165 0.0157 0.124 ageent Age at entry to Bank 37.0301 7.666 35.5835 7.270 34.3540 7.682 32.8540 7.093 35.7898 7.626 ageen2 Age at entry to Bank squared 1429.9502 587.572 1318.9769 547.100 1239.1031 563.488 1129.4961 520.050 1339.0494 574.390 phd 1=PhD level 0.4146 0.493 0.4508 0.498 0.3148 0.465 0.3498 0.478 0.4018 0.490 mast 1=Masters level 0.4375 0.496 0.4508 0.498 0.5509 0.498 0.5021 0.501 0.4669 0.499 bach 1=Bachelors level 0.1141 0.318 0.0719 0.258 0.0965 0.296 0.0905 0.288 0.0968 0.296 dl_can 1=PhD from Canada 0.0088 0.094 0.0093 0.096 0.0019 0.043 0.0040 0.063 0.0073 0.085 dl_fra 1=PhD from France 0.0339 0.181 0.0210 0.143 0.0186 0.135 0.0080 0.089 0.0253 0.157 dl_grm 1=PhD from Germany 0.0147 0.121 0.0012 0.034 0.0112 0.105 0.0000 0.000 0.0090 0.094 dl ukg 1=PhD from UK 0.0406 0.197 0.0443 0.206 0.0242 0.154 0.0400 0.196 0.0387 0.193 dl_usa 1=PhD from USA 0.2448 0.430 0.3174 0.466 0.2197 0.414 0.2320 0.423 0.2600 0.439 ml_can 1=Masters from Canada 0.0273 0.163 0.0093 0.096 0.0093 0.096 0.0080 0.089 0.0173 0.131 ml fra 1=Masters from France 0.0280 0.165 0.0128 0.113 0.0261 0.159 0.0120 0.109 0.0220 0.147 ml_grm 1=Masters from Germany 0.0088 0.094 0.0012 0.034 0.0037 0.061 0.0000 0.000 0.0050 0.071 mltukg 1=Masters from UK 0.0745 0.263 0.0502 0.218 0.0447 0.207 0.0600 0.238 0.0610 0.239 ml_usa 1=Masters from USA 0.1858 0.389 0.2404 0.428 0.4041 0.491 0.2920 0.456 0.2493 0.433 bI_can 1=Bach from Canada 0.0052 0.072 0.0023 0.048 0.0074 0.086 0.0040 0.063 0.0047 0.068 bI_fra 1=Bach from France 0.0044 0.066 0.0000 0.000 0.0019 0.043 0.0000 0.000 0.0023 0.048 bl grm 1=Bach from Germany 0.0022 0.047 0.0000 0.000 0.0000 0.000 0.0000 0.000 0.0010 0.032 bl_ukg 1=Bach from UK 0.0243 0.154 0.0105 0.102 0.0093 0.096 0.0040 0.063 0.0160 0.125 bIlusa 1=Bach from USA 0.0428 0.202 0.0105 0.102 0.0670 0.250 0.0200 0.140 0.0360 0.186 dl_eco 1=PhD Ivi Economics 0.2345 0.424 0.2754 0.447 0.1639 0.371 0.1840 0.388 0.2293 0.420 dl bus 1=PhD Ivt Bus/Actng/Comrce 0.0044 0.066 0.0012 0.034 0.0000 0.000 0.0000 0.000 0.0023 0.048 dl eng 1=PhD Ivt Engineering 0.0236 0.152 0.0257 0.158 0.0019 0.043 0.0120 0.109 0.0193 0.138 dl fin l=PhD IvI Finance 0.0037 0.061 0.0140 0.118 0.0019 0.043 0.0040 0.063 0.0063 0.079 ml_eco 1I=Masters lvl Economics 0.1829 0.387 0.1424 0.350 0.2011 0.401 0.1720 0.378 0.1737 0.379 ml_bus 1=Masters Ivl Bus/Actng/Comrce 0.0088 0.094 0.0233 0.151 0.0223 0.148 0.0160 0.126 0.0160 0.125 ml_eng 1=Masters Ivt Finance 0.0310 0.173 0.0373 0.190 0.0056 0.075 0.0080 0.089 0.0263 0.160 ml fin 1=Masters Ivl Engineering 0.0413 0.199 0.0828 0.276 0.0577 0.233 0.0680 0.252 0.0583 0.234 bi_eco 1=Bach tvt Economics 0.0302 0.171 0.0152 0.122 0.0130 0.114 0.0120 0.109 0.0213 0.145 bl bus 1=Bach lvt Bus/Actng/Comrce 0.0052 0.072 0.0058 0.076 0.0074 0.086 0.0200 0.140 0.0070 0.083 bl eng 1=Bach Ivl Finance 0.0184 0.135 0.0187 0.135 0.0019 0.043 0.0000 0.000 0.0140 0.118 bt_fin l=Bach lvl Engineering 0.0015 0.038 0.0012 0.034 0.0000 0.000 0.0000 0.000 0.0010 0.032 continued 47 Appendix Table 1 continued: Variable definitions and summary statistics Variable Description Part 1 Men Part 2 Men Part 1 Women Part 2 Women Pooled Sampte Mean Std.Dev. Mean Std.Dev. Mean Std.Dev. Mean Std.Dev. Mean Std.Dev. wrkex Yrs Work prior Bank 10.1924 8.037 8.0121 7.349 7.4018 6.990 6.1023 6.213 8.7236 7.649 d_exp 1=workexp