A.  Probit Estimates of the Effect of Manager Race on the Race of New Hires

Tables 4–7 show the results from the probit models estimating the effect of manager race on the race of a new hire.13 The dependent variable in table 4 is a dummy variable equal to one if the new hire is white and zero otherwise. In tables 5–7, the dependent variables are dummies indicating that a new hire is black, Hispanic, and Asian. In each regression, the omitted manager race category is white.

Cross‐sectional analysis.—In tables 4–7, column 1 shows the results from the regressions with no controls. These results simply confirm the strong correlations between manager race and the race of a new hire that were seen in table 3. The representation of each race group among new hires is significantly higher when the manager is of that race group.

Columns 2 and 3 show how much of the correlation between manager race and employee race is explained by observable differences across stores and locations. Not surprisingly, a significant part of this correlation is explained by differences in community demographics. In column 2, which controls for the population share of each race group in the community, the manager race coefficients that were significant in column 1 are reduced in magnitude by at least one‐quarter. However, most of them remain significantly different from zero. Column 3 adds more location variables, including population density, median household income, and store location type. Although these variables are often statistically significant, they explain little of the correlation between manager race and employee race.

In column 4, the inclusion of month dummies controls for unobserved differences across time. This also has little effect and suggests that the similarities between managers and new hires are not explained by seasonal and national fluctuations in the labor market.

Store fixed effects estimates.—After controlling for observed differences across stores and locations (cols. 2 and 3) and unobserved differences across time (col. 4), a substantial amount of the correlation between manager race and the race of new hires still remains to be explained. Column 5 of tables 4–7 shows the store fixed effects specifications of the probit models. These estimates control for all fixed differences across stores and locations. A comparison of columns 4 and 5 reveals that, in many cases, unobserved differences across stores and locations account for nearly all of the remaining correlation between manager race and the race of new hires. In the equation predicting that a new hire is white (table 4, col. 5), the Hispanic and Asian manager effects are very small and not significantly different from zero. In the equations predicting that a new hire is Hispanic (table 6, col. 5) or Asian (table 7, col. 5), none of the manager race effects are significantly different from zero at a 5% level.

However, even after controlling for store fixed effects, one salient pattern remains. There is a significant difference between the hiring patterns of black managers and all nonblack managers (i.e., white, Hispanic, and Asian managers). This difference lies mainly in the share of whites and blacks hired by these two groups of managers. First, the manager race coefficients in table 4, column 5, indicate that the probability a new hire is white is between 3.7 and 4.4 percentage points lower under black managers than it is under nonblack managers in the same store. Second, the coefficients in table 5, column 5, indicate that the probability that a new hire is black is between 3.5 and 4.0 percentage points lower under nonblack managers than it is under black managers. In both cases, the estimated differences between black managers and each nonblack group are significant at a 1% level of confidence, and the estimated differences among nonblack managers are statistically insignificant.

We would stress three points about the fixed effects estimates. First, the main finding is that all nonblack managers hire more whites and fewer blacks than do black managers. Second, hiring patterns on average are similar among all nonblack manager groups—white, Hispanic, and Asian managers all hire roughly the same proportions of white, black, Hispanic, and Asian employees. Finally, it is notable that black managers differ very little from the three nonblack groups in the shares of Hispanics and Asians hired.

Two caveats also apply. First, we shall see below that the effect of manager ethnicity on the probability that a new hire is Hispanic becomes significant when we restrict the sample to locations with larges shares of Hispanics in the local population. Second, we note that, in the analysis of Asian hires, the lack of significant manager race effects is due partly to a lack of precision. We cannot rule out moderate effects. For example, the Asian manager coefficient in table 7, column 5, suggests that the probability a new hire is Asian is 1.5 percentage points higher under an Asian manager than it is under a white manager in the same store; however, the 95% confidence interval for this estimate is between −0.1 and 3.1 percentage points.

Store‐specific trends.—If trends in local demographics lead to parallel trends in the demographics of applicant pools for both managers and employees, then the fixed effects model might overestimate the causal effect of manager race on hiring patterns. The specification in column 6 of tables 4–7 addresses this concern by including store‐specific trends as controls. The results of this specification are somewhat surprising. In almost all cases, the magnitudes of the manager race coefficients become larger rather than smaller, suggesting that, if anything, the fixed effects model underestimates the causal effect of manager race. However, the differences between the estimates in columns 5 and 6 are generally quite small.

Figure 1A and B provides additional evidence that the estimates of the relationship between manager and employee race are not driven by trends in store demographics. The figure also provides some insight as to why the column 6 estimates are larger in magnitude than those in column 5. The figure shows trends in the white share of new hires (fig. 1A) and the black share of new hires (fig. 1B) before and after a change in manager.14 Trends are shown for three cases: (1) a nonblack manager is replaced by a black manager, (2) a black manager is replaced by a nonblack manager, and (3) a nonblack manager is replaced by a nonblack manager. In all cases, the introduction of a new manager whose race differs from the outgoing manager results either in a clear intercept shift or in a break with the previous trend rather than in a continuation of the current trend. Hence, figure 1A and B clearly shows that manager race is not simply following the same trends as the race of new hires.

However, the graphs of figure 1A and B do raise another concern. It appears that nonblack managers who hire increasingly more whites (fig. 1A) and fewer blacks (fig. 1B) over a 4‐month period are more likely to be replaced by black managers. One possible explanation of this pattern is that the company tries to maintain a certain level of diversity in their stores and that this goal results in the replacement of nonblack managers with black managers in stores where there is a noticeable decline in the ratio of blacks to whites. But in a robustness test described below (Sec. V.B), we use regression analysis to examine whether the company’s choice of a new manager may be influenced by the hiring patterns of the previous manager. We find that any such balancing effect is small and statistically insignificant.

B.  Fixed Effects Multinomial Logit Estimates

The multinomial logit results are remarkably similar to the binomial probit results. These results are shown in tables 8 and 9, and they facilitate the comparison of hiring patterns across all four manager race groups. First, table 8 shows the coefficients of the model (cols. 1–6) and Wald tests assessing the overall similarity in hiring patterns between managers of different races (col. 7). This table again shows that there are no significant differences in the hiring patterns of the three nonblack groups of managers (white, Hispanic, and Asian). Moreover, it confirms that the hiring choices of each nonblack group differ significantly from those of black managers. Specifically, the differences between nonblack managers and black managers lie in the ratios of black hires to hires of other races and mainly in the ratio of black hires to white hires (col. 1).

Next, table 9 presents the predicted hiring probabilities for each manager race group.15 Differences among the three nonblack groups are very small—the largest being the 1.3 percentage point difference between white and Asian managers in the share of Asians hired.16 Black managers differ very little from the three nonblack groups in the shares of Hispanics and Asians hired. Once again, the differences that stand out are those between nonblack managers and black managers in the shares of blacks and whites hired. The estimates imply that, when any nonblack manager replaces a black manager in a typical store, the share of new hires that is black falls by 3.8–4.8 percentage points. In all cases, this decline is offset mainly by an increase in the share that is white.

C.  Geographic Differences in the Effects of Manager Race and Ethnicity

The estimates presented thus far are national averages, and hence they may obscure important geographic differences in the effects of manager race and ethnicity on hiring patterns. We are able to perform clean tests of whether the effects of manager race vary by geographic location because the workplaces in our sample are all very similar except for their geographic locations. Here we report the results of two such geographic tests. First, we show that differences between nonblack and black hiring patterns are particularly strong in the South. Second, we show that differences between Hispanic and white hiring patterns are large and statistically significant in locations with large Hispanic populations.17

Nonblack vs. black managers in the South.—We ask whether the differences between nonblack and black hiring patterns are larger in the South. This question is important because the South is home to over half the nation’s blacks (Frey 2001) and because there is evidence that racial attitudes remain more divisive in the South than in the rest of the country. For example, the General Social Survey (GSS; 1990–2000) reveals that, in nonsouthern states, 3.0% of blacks and 11.4% of whites favored laws against interracial marriage but, in the South, 5.7% of blacks and 23.0% of whites favored such laws.18 Also, Kuklinski, Cobb, and Gilens (1997) estimate that, while 10% of nonsouthern whites react negatively to the idea of living next door to a black family, 42% of southern whites react negatively. Too, Dee (2005) finds that, in the South, but not in other regions of the United States, a student is more likely to be seen as disruptive or inattentive by a teacher if the teacher does not share the student’s race.

To compare hiring patterns in the South to those in the rest of the country, we run separate fixed effects probit regressions for the two subsamples.19 The regressions estimate the probability (1) that a new hire is white and (2) that a new hire is black. Table 10 shows the results. The tendency of nonblack managers to hire more whites and fewer blacks than black managers is particularly strong in the South, and the differences between southern and nonsouthern states are substantial.

When a nonblack manager replaces a black manager in a nonsouthern store, the predicted probability of a white hire increases from 54.4% to 57.1% and the probability of a black hire falls from 16.9% to 14.4%. This amounts to a 5% increase in white employment and a 15% decline in black employment. In southern stores, the differences are much larger. When a nonblack manager replaces a black manager in a southern store, the probability of a white hire increases from 52.0% to 61.4% and the probability of a black hire falls from 29.3% to 21.2%. This represents an 18% increase in white employment, and a 28% decline in black employment.20

Again, we use a graphical analysis to confirm that these estimates are not driven by demographic trends. Figure 2A and B shows trends in the white share and black share of new hires for the subsample of southern stores and compare these trends for the cases where a nonblack manager is replaced by a black manager to the cases where the order is reversed. Just as in the full‐sample graphs, the introduction of a new manager whose race differs from the outgoing manager always results either in a clear intercept shift or in a break with the previous trend rather than in a continuation of the current trend.

White versus Hispanic managers in locations with large Hispanic populations.—Census data suggests that there is less assimilation of Hispanics and non‐Hispanics in areas with larger Hispanic populations.21 Regional variation in the assimilation of Hispanics and non‐Hispanics might cause the effect of manager ethnicity to differ by geographic region. Hence, we ask whether the effect of manager ethnicity is stronger in locations with larger Hispanic populations. Specifically, we examine the subsample of stores where the population share of Hispanics in the community is at least 30%.22 These stores are located primarily in the southern parts of Texas, Florida, and California.

To maintain a sufficient sample size, we do not restrict the sample to stores with one hire of each new race group; rather we examine all stores that have at least one Hispanic hire and one white hire. Further, because nearly all of the non‐Hispanic managers in this subsample are white, we focus only on the differences between Hispanic and white managers.

The results of this analysis are shown in table 11. In locations where Hispanics are less than 30% of the population, the hiring patterns of Hispanic and white managers are remarkably similar, and Hispanics account for roughly 10% of new hires on average. However, in stores where the local community is at least 30% Hispanic, we find that Hispanic managers hire significantly more Hispanics and fewer whites than do white managers.23 Our estimates imply that, when a Hispanic manager is replaced by a white manager in one of these stores, the share of new hires that is Hispanic falls roughly from 59% to 48% and the share that is white rises from 22% to 32%. This represents a 17% decline in Hispanic employment and a 45% increase in white employment.

Figure 3A and B shows the graphical analysis of trends in hiring patterns for stores where the local community is at least 30% Hispanic. Once again, the patterns suggest that the estimated effect of manager race on hiring patterns cannot be attributed to trends in the demographics of the labor pool.

D.  How Large Are the Effects of Manager Race and Ethnicity on Hiring Patterns?

To help interpret the magnitude of our results, we consider what happens to the racial composition of a store’s workforce over the course of a year when a black manager is replaced by a nonblack manager in a nonsouthern store and in a southern store. We also consider what happens to the racial composition of a store in a Hispanic neighborhood over the course of a year after a Hispanic manager is replaced by a white manager. Employee turnover is so high at our company that, within a year of a change in management, almost all employees will have been hired by the new manager.

First, the estimates in table 10 suggest that, when a black manager is replaced by a nonblack manager, the employment share of blacks in nonsouthern stores falls roughly from 16.9% to 14.4% within a year and the share of whites rises from 54.4% to 57.1%. In a store of 40 employees, this change amounts to going from roughly seven blacks and 22 whites to six blacks and 23 whites—in other words, replacing one black employee with one white employee. Now, from the viewpoint of someone (such as a district manager) who is observing just a small sample of stores, this change might either go unnoticed or appear insignificant. However, the change may appear more significant from the point of view of job seekers—especially black job seekers. Indeed, the change amounts to a proportional decline of 15% in the number of blacks employed.

Further, we have seen that the effects of manager race are larger in southern stores. In a typical southern store with a black manager and 40 employees, there are 12 black employees and 21 white employees on average. Our estimates suggest that replacing a black manager with a nonblack manager would result in the replacement of three to four black employees with white employees. This amounts to a 28% decline in black employment and a 17% rise in white employment.

Second, in a store of 40 employees in a Hispanic neighborhood, replacing a Hispanic manager with a white manager would result in the replacement of roughly four out of 24 Hispanic employees with white employees. This is equivalent to a 17% decline in the number of Hispanics and a 45% increase in the number of whites in the store.

A.  Linear Probability Model without Sample Restriction

Our estimation sample is restricted to stores that hire at least one new employee from each of the four main race groups. This sample restriction allows us to estimate all probit models using the same sample of stores and is also necessary to estimate the multinomial logit model with store fixed effects. However, this restriction could introduce sample selection bias. By eliminating stores in which there is no change in the hiring probability for at least one race group, we may be dropping many stores in which a change in manager race has zero impact on hiring patterns.

To examine the implications of the sample selection, we use a linear probability model to estimate all of the binomial choice equations on both the restricted and full samples. The results (reported in appendix table A1) are similar for both samples and are substantively similar to the probit results.24 This similarity suggests that our sample restriction does not have a significant effect on our estimates.

B.  Tests for the Exogeneity of Changes in Manager Race

Our analysis has ruled out differences across workplaces and differential trends in population demographics as sources of the observed differences in hiring patterns between black and nonblack managers. However, we have not explicitly ruled out the possibility that some centrally determined personnel policy is somehow responsible. In this section, we consider two ways in which “company policy” could explain our findings.

First, one could imagine that the company periodically reviews store‐level data and decides on a target racial composition for each store’s workforce. For example, to serve an increasingly black customer base, the company might decide to increase the presence of blacks in the store both by appointing a black manager and by directing that manager to hire more black employees. In this case, the relationship between manager race and the race of new hires could be driven by personnel policies specific to our company.

One testable implication of a policy of placing black managers in designated “black stores” is that there should be some permanence to the changes in manager race. In other words, new black managers should tend to be followed by black managers. To test this hypothesis, we compare two sets of stores: (1) those in which the first observed manager is black and is replaced by a white manager and (2) those in which the first manager is white and is replaced by a black manager. In the first set of stores, the probability that a third manager is observed within the sample period is 59%, and the conditional probability that this third manager is black is 13%. In the second set of stores, the probabilities are 57% that there is a third manager and 11% that this manager is black. These latter probabilities are not significantly different from those of the first set (see appendix table A2). Hence stores where a black manager replaces a white manager are no more likely than stores where the reverse is true either to keep that black manager or to receive another black manager.

A second way in which company policy could explain our results is that the company’s choice of a new manager may be influenced by the hiring patterns of the previous manager. As mentioned earlier, the patterns in figure 1A and B raise this concern because they suggest that nonblack managers who hire increasingly more whites (fig. 1A) and fewer blacks (fig. 1B) over a 4‐month period are more likely to be replaced by black managers. To be sure, the existence of a policy whereby black managers are employed to help reverse declines in black employment would not undermine our conclusion that manager race matters; indeed, such a policy would suggest that the company exploits the “manager race effect” to shape employment demographics. Still, such a policy would mean that the company deliberately fosters the changes in hiring patterns that we observe.

To investigate this possibility more formally, we estimate probit models predicting the probability that a new manager is a given race. In particular, we examine the effects of three covariates on the race of the new manager: (1) changes in hiring patterns over the preceding 6‐month period, (2) changes in employment demographics over the same 6‐month period, and (3) the ratio of employment share to local population share of each race group during the 6‐month period.25 We control for the race of the departing manager, and we interact all of the variables that measure hiring and employment trends with dummy variables indicating the departing manager’s race.

The results of these regressions are shown in appendix table A3. In all regressions, preceding hiring and employment patterns have small effects on the race of the new manager, and none of the effects is statistically different from zero.26 These results support the conclusion that, in stores where manager changes are observed, the race of the new manager is reasonably independent of any recent patterns in hiring or employment demographics.27 In sum, we conclude that it is very unlikely that our findings result from deliberate company policy to manipulate the demographic composition of their workplaces.

A.  Segregated Hiring Networks

If managers use their own social networks to recruit employees and if social networks tend to be segregated by race and ethnicity, then these networks may lead managers to hire employees of the same racial or ethnic group.28 To test for managerial hiring networks, we assume that there is some congruence between social networks and residential proximity, and we ask whether managers tend to hire employees who live near them.29 We calculate the approximate distance between manager and employee residences using data on residential zip codes.

We perform this test using stores that have at least one change in management and that have two managers (call them Manager 1 and Manager 2) who reside in different zip codes.30 For all employees in these stores that were hired either by Manager 1 or by Manager 2, we construct the following variable: Relative distance from Manager 1 = Distance between residential zip codes of employee and Manager 1 / Distance between residential zip codes of employee and Manager 2.31 We then regress the log of this variable on a dummy variable that is equal to one if the employee was hired by Manager 1 and equal to zero if the employee was hired by Manager 2. The results (table 12, col. 1) indicate that the relative distance to Manager 1 is 1.6% smaller if the employee was hired by Manager 1. In other words, employees tend to live slightly closer to the managers that hired them. The effect is small, but the estimate is statistically significant.

Of course, the tendency of managers to hire individuals who live closer by could result from a combination of racial preferences and residential segregation. For example, black managers may tend to hire their neighbors because they prefer hiring blacks (who tend also to be their neighbors) rather than because networking leads black managers to hire their neighbors (who tend also to be black). To address this issue, we estimate regressions that control for an employee’s racial similarity to Manager 1 and to Manager 2. The results (table 12, col. 2) do in fact indicate that employees tend to live closer to managers who share their race or ethnicity. However, even after taking racial similarity into account, a significantly greater proximity to one’s hiring manager remains. This suggests that the greater proximity is caused by the use of networks and not solely by racial biases in hiring.

B.  Discrimination by Managers: Evidence from Dismissals and Promotions

Two types of managerial discrimination could explain why managers hire more employees from their own race group or fewer from certain other race groups. One is a form of “statistical” discrimination—managers may rationally expect to reap efficiency gains from having same‐race employees (Phelps 1972; Arrow 1973; Aigner and Cain 1977).32 The other is “taste‐based” discrimination—managers may select own‐race employees to indulge their personal preferences (Becker 1971). We do not have data on applicants and cannot directly test whether hiring outcomes are affected by either type of managerial discrimination. But we do have data on dismissals and promotions, and we can use these posthire outcomes to gain insight into whether managerial discrimination may affect hiring outcomes.

If either type of managerial discrimination explains why black managers hire more blacks and fewer whites, then we might expect that the motives for such discrimination would also affect posthire personnel decisions made by managers and would be reflected in outcomes such as dismissals and promotions. In short, if either type of managerial discrimination explains why black managers hire more blacks, then we might see that black managers dismiss blacks relatively less often and promote them more often. And if either type of managerial discrimination explains why black managers hire fewer whites, then we might see that black managers dismiss whites more often and promote them less often.33 Our analysis, therefore, looks at the effect of having a black manager on the relative dismissal and promotion rates of black and white employees.

We estimate Cox proportional hazard models that express dismissal and promotion rates as a function of employee race indicators, manager race indicators, and the interactions of the black manager indicators with the black employee and white employee indicators.34 The results (table 13) indicate that black employees have significantly better relative outcomes under black managers than under nonblack managers. The dismissal rate of blacks relative to Hispanics and Asians is 20% lower under black managers, and their relative promotion rate is 88% higher. However, we find that the outcomes of whites relative to Hispanics and Asians are no worse under black managers. In fact, the estimated effect on the white promotion rate (though not significant) suggests that whites do relatively well with black managers.

In sum, the finding that black employees have relatively favorable posthire outcomes under black managers suggests that managerial discrimination may help explain why black managers hire relatively more blacks. However, the finding that white employees do no worse under black managers does not support managerial discrimination as an explanation for why black managers hire fewer whites.

Two further points are important for clarity. First, while our analysis of posthire outcomes suggests that managerial discrimination may help explain why black managers hire more blacks, it cannot tell us whether such discrimination is statistical or taste based. Black managers might treat black employees relatively favorably either because of differences in performance (i.e., blacks perform better relative to nonblacks under black managers) or because of taste‐based bias (i.e., blacks are treated more favorably by black managers despite similar performance under all managers). Hence, the favorable posthire outcomes of blacks with black managers would be consistent with the managerial motives for either statistical or taste‐based discrimination in hiring.35

Second, if taste‐based discrimination explains the differences in the posthire outcomes of blacks under black and nonblack managers, then again there are two interpretations: (1) black managers prefer black employees over nonblack employees or (2) nonblack managers prefer nonblack employees over black employees.36 These two taste‐based possibilities are both plausible and are not mutually exclusive. To be sure, the latter interpretation is consistent with the findings of several audit studies and other experiments that have found evidence of discrimination against black job seekers.37 But either possibility could explain why black managers hire more blacks.

C.  Manager‐Employee Racial Matches and Store‐Level Performance Measures

The preceding analysis suggests that managerial discrimination may help explain our hiring results, but it cannot not tell us what type of discrimination may be at work—statistical and/or taste based. In this section, we investigate the statistical discrimination hypothesis further by looking for evidence that racial similarity or dissimilarity affects performance. We examine two store‐level measures of performance—monthly sales and employee turnover. While our findings are consistent with motives for statistical discrimination, they are less than compelling.

Table 14 reports the results of linear fixed effects regressions of our two variables—sales and turnover—on manager race, employment shares of each race group, and the interactions of the black manager indicator with the employment shares of blacks and whites. In the sales regression (col. 1), a comparison of the two interaction coefficients indicates slightly higher sales under black managers (compared to nonblack managers) when there are more black employees and fewer white employees. While this could indicate an efficiency motive for black managers to hire more blacks and fewer whites or for nonblacks to do the opposite,38 the effect is small and not statistically significant.39 In the turnover regression (col. 2), the coefficient on the black manager‐white employment share interaction is weakly significant $$( p=.08) $$ and indicates slightly higher turnover under black managers when there are more white employees. While costs associated with higher turnover could help explain why black managers hire relatively fewer whites, the effect is again small.40

D.  Discrimination by Job Seekers: Evidence from Employee Quit Rates

If job seekers prefer to work for own‐race managers or are biased against managers of certain other race groups, they may discriminate on the basis of these preferences when choosing where to work. While we cannot test directly for discriminatory sorting by job seekers, we can construct a test for such sorting by analyzing employee quit rates.

We exploit the fact that many employees in our sample receive new managers, and we compare the effect of manager‐employee racial matches on the quit rates of employees who retain their hiring managers and employees who receive new managers. If job seekers who prefer own‐race managers indulge these preferences when deciding where to work, then the racial match between employee and manager should have little effect on quit rates as long as employees keep the managers who hired them (i.e,. the managers they chose to work for). In contrast, when employees get new managers, the discriminatory preferences of biased employees should lead to relatively large effects of racial matches on quit rates.

Table 15 shows the estimation results from a hazard model of quits.41 We focus on the differential effects of having a black manager for black and white employees, and the black manager effects are estimated separately for employees who still have their hiring managers and for those who have new managers. We find evidence of discriminatory sorting by white job seekers but not by black job seekers. For employees with their hiring managers, we find small and statistically insignificant effects of manager race on the quit rates of both black and white employees. Among employees with new managers, the black quit rate is again unaffected by whether or not the manager is black. However, the quit rate of whites is 15% higher under new black managers than under new nonblack managers.

We interpret this increase in the white quit rate as evidence of discriminatory sorting by white job seekers. It implies that whites who dislike working for black managers often avoid working for black managers in the first place and that when such whites involuntarily find themselves working for a black manager, their quit rates increase substantially.42 Hence, we conclude that the reluctance of some white job seekers to take jobs with black managers may be at least partly responsible for our finding that black managers hire relatively low proportions of whites.43

E.  Discussion

We have focused on why black managers hire relatively more blacks and fewer whites. Our strongest finding helps explain why blacks hire fewer whites. An analysis of quit rates suggests that some whites prefer not to work with black managers and hence that discrimination by white job seekers may cause black managers to hire lower proportions of whites.

But why do black managers hire blacks at especially high rates? Our findings suggest two reasons. An analysis of residential zip codes suggests a role for managerial hiring networks. Moreover, an analysis of dismissals and promotions suggests a role for managerial discrimination. With regard to managerial discrimination, two cautions apply. First, we cannot tell whether such discrimination may be statistical or taste based. Second, if the tastes of managers play a role, there are two possibilities: (1) black managers prefer black employees over nonblack employees and/or (2) nonblack managers prefer nonblack employees over black employees.

More generally, we would like to make two points. First, we considered three reasons why a manager’s race may affect the racial composition of new hires: managers may use racially segregated hiring networks, managers may hire racially similar employees because of expected productivity gains from manager‐employee similarity, and both managers and employees may engage in taste‐based discrimination. Our analysis produced evidence consistent with each of these explanations, and we would stress that these explanations may all play a role in hiring.

Indeed, these “explanations” are not mutually exclusive, and they may even reinforce one another. For example, if black managers have difficulty recruiting due to discrimination from a mostly white applicant pool, this could lead black managers to rely more heavily on hiring networks. Or if, for example, taste‐based discrimination results in biased evaluations of black employees by nonblack managers, this could cause or reinforce lower productivity among blacks (MacLeod 2003). Further, when nonblack managers observe this pattern of lower productivity, they might statistically discriminate against black applicants.

Second, we should acknowledge that, even though the studied employer is in many ways typical of the U.S. retail sector (which accounts for roughly 18% of all U.S. jobs), this is nevertheless a case study. Several characteristics of our sample could affect the relative importance of either managerial or employee discrimination. Consider two examples. First, the sample period (1996–98) was a time of historically low unemployment in the United States. Because it was difficult to find qualified workers during these years, manager discrimination may well have been less prevalent and employee discrimination more prevalent than during periods of higher unemployment. Second, racial attitudes could be affected by the relatively high share of women and the relative youth of both management and the workforce. Other studies have suggested that females and youth are less likely to discriminate on the basis of race.44