Ethnic Hiring David Neumark UCI Department of Economics and Center for Economics & Public Policy, NBER, and IZA December 2011 I. Introduction Economic migration brings with it the challenge of racial, ethnic, or national minorities assimilating into their adopted labor markets. Barriers to the employment of these minorities will clearly inhibit their successful assimilation. This chapter focuses on the hiring side of the equation: Are there barriers to the hiring of racial and ethnic minorities? What is the nature of these barriers? And how do workers overcome these barriers? The chapter focuses on three key influences on the hiring of racial, ethnic, or national minorities: discrimination, spatial mismatch, and networks. The barriers posed by discrimination and spatial mismatch are obvious. Networks can also pose barriers to the extent that ethnic minorities have fewer network connections than majority groups, but networks can also be a way for ethnic minorities to overcome barriers to employment. The chapter integrates recent research my co-authors and I have done on these topics. It is not an exhaustive survey of all of the research on these specific topics, nor does it cover other topics that could bear on ethnic hiring. The focus is not on economic migrants per se, but on ethnic and racial minorities, with much evidence coming from research on blacks in the United States. Moreover, the research I discuss is limited to the U.S. setting, but I touch on related recent work on ethnic hiring in European countries. II. Discrimination—Evidence from Field Experiments, and Problems with that Evidence1 Research on discrimination in labor markets has a long history. The “workhorse” of early economics research on labor market discrimination is the “residual discrimination” approach of Oaxaca (1973), in which wage regressions are estimated on individual-level data, and discrimination is estimated from the ethnic differential that remains unexplained after including many proxies for productivity.2 This approach suffers from numerous criticisms. First, the proxies may not adequately capture group differences in productivity, in which case the “unexplained” differences cannot be interpreted as discrimination. Second, mean differences between groups are treated as non-discriminatory, and differences in the equation coefficients as discriminatory. But differences in coefficients can arise for other 1 2 This section draws heavily on Neumark (forthcoming). The approach is sometimes applied to employment (e.g., Fairlie, 2005). 1 reasons – for example, a productivity-enhancing effect of marriage on men but not women (Korenman and Neumark, 1991, 1992) – and differences in means can reflect discrimination (Neumark and McLennan, 1995). Finally, arbitrary assumptions have to be made about the counterfactual – i.e., what the wage structure would look like in the absence of discrimination (Neumark, 1988). Audit or correspondence studies are a response to these criticisms.3 In these studies, fictitious individuals who are identical except for race, sex, or ethnicity apply for jobs. Evidence of group differences in outcomes – such as fewer job offers for blacks – is generally viewed as compelling evidence of discrimination, because there is no reason to expect important differences between, for example, black and white job applicants. As a consequence, this strategy has come to be widely used in testing for discrimination in labor markets (as well as housing markets).4 An advantage of audit or correspondence studies with regard to the present inquiry is that they directly address discrimination in hiring, rather than pay discrimination. Across numerous countries and minority groups, audit or correspondence studies find evidence consistent with discrimination, including discrimination against blacks, Hispanics, and women in the United States (Mincy, 1993; Neumark, 1996; Bertrand and Mullainathan [BM], 2004), Moroccans in Belgium and the Netherlands (Smeeters and Nayer, 1998; Bovenkerk et al., 1995), and lower castes in India (Banerjee et al., 2008). Researchers have, over time, shifted from audit to correspondence studies, in response to critiques of the audit study method (e.g., BM, 2004). For example, Heckman and Siegelman (HS, 1993) noted that in the well-known Urban Institute audit studies (Mincy, 1993), white and minority testers were told, during their training, about “the pervasive problem of discrimination in the United States,” possibly introducing experimenter effects. Correspondence studies address this problem by using applications on paper, cutting out the potential influence of live job applicants. However, a fundamental critique that applies equally well to correspondence studies has not been 3 Another approach is to use worker and firm data to compare productivity and wage differentials between groups (Hellerstein et al., 1999). 4 Thorough reviews are contained in Fix and Struyk (1993), Riach and Rich (2002), and Pager (2007). 2 addressed by researchers. In particular, HS and Heckman (1998) consider what most researchers view as the ideal conditions for an audit or correspondence study – when the observable average differences between groups are eliminated, and the applications are sufficiently detailed that it is safe to assume that employers believe there are no average differences in unobservable characteristics between groups. HS show that, even in this case, these studies can generate evidence of discrimination (in either direction) when there is none, and can also mask evidence of discrimination when it in fact exists. To see this in a simple setting , suppose that productivity depends on two individual characteristics, X’ = (XI,XII). Two testers (or applications) are sent to firms to apply for jobs; R is an indicator equal to 1 for ethnic minority applicants (E) and 0 for non-minorities (NM). Denote by XEj and XNMj the values of XI and XII for the two groups, j = I, II. The study controls only XI in the resumes or interviews, standardizing XI across applicants so XEI = XNMI = XI*. PE* and PNM* denote expected productivity of the two groups. Assume that productivity is P(XI,XII) = βI’XI + XII, and that the parameter γ’ captures possible discrimination against the ethnic minority, resulting in the “discounting” of the productivity of the ethnic minority, as in Becker (1971). Then each individual test (two applicants to a firm) provides an observation equal to (1) PE* + ’ − PNM* = βI’XEI + E(XEII) + ’ (βI’XNMI + E(XNMII)) = ’ + E(XEII) E(XNMII).5 These observations identify ’ under the assumption that E(XEII) = E(XNMII). We cannot, in practice, rule out employers holding different expected means for unobserved productivity, although by including a rich set of resume characteristics this can be mitigated. Moreover, one can always reinterpret {’ + (E(XEII) E(XNMII))} as capturing illegal discrimination, at least in countries like the United States where statistical discrimination is illegal. As HS emphasize, a more troublesome problem arises because in the hiring process firms likely evaluate a job applicant’s productivity relative to a standard, and offer the applicant a job (or an interview) if the standard is met. In this case, even when E(XEII) = E(XNMII) a correspondence study can generate 5 Pk* and E(XkII), k = E, NM, can be interpreted as conditional on XjI. 3 spurious evidence of discrimination in either direction, or of its absence; in other words, discrimination is unidentified.6 In this “best-case” scenario, suppose the correspondence study standardizes on a low value of XI. Because an employer offers a job interview only if it perceives the sum βI’XI + XII to be sufficiently high, the employer has to believe that XII is high (or that the probability that it is high is large) to offer an interview. For example, although the employer does not observe XII, if the employer knows that the variance of XII is higher for non-minorities, the employer correctly concludes that non-minorities are more likely than minorities to have a sufficiently high sum of βI’XI + XII, and will therefore be less likely to offer jobs to non-minorities. The opposite holds if the standardization is at a high value of XI, in which case the employer only needs to avoid very low values of XII, which will be more common for non-minorities. And the results flip if the unobserved variance is higher for minorities.7 To formalize this, suppose that a job offer or interview is given if a worker’s perceived productivity exceeds the threshold c’. The hiring rules are (2) Hire E if I’XI* + XEII + ’ > c’ (2’) Hire NM if I’XI* + XNMII > c’. Assume that the unobservables XEII and XNMII are normally distributed, with equal means (set to zero, without loss of generality), and standard deviations EII and NMII.8 Then the hiring probabilities are (3) Pr[Hire E] = 1 [(c’ I’XI* − ’)/EII] = [( I’XI* + ’ – c’)/EII] (3’) Pr[Hire NM] = 1 [(c’ I’XI*)/NMII] = [( I’XI* − c’)/NMII], where denotes the standard normal distribution function. The difference between equations (3) and (3’) is supposed to be informative about discrimination. But even if ’ = 0, so there is no discrimination, these two expressions need not be equal because EII and 6 Since this problem arises even in correspondence studies, I now refer exclusively to correspondence studies. Differences in the variances of unobservables across groups were introduced in early models of statistical discrimination (Aigner and Cain, 1977). 8 In an audit study the randomness that generates a statistical model arises naturally, as variables unobserved by the econometrician but observed by the firm can generate variation in hiring. In a correspondence study, given XEI and XNMI (assumed equal), the employer hires the higher variance group if the level of standardization is low, and vice versa. One way to introduce unobservables that generate random variation, with different variances across groups, is to assume that there are random productivity differences across firms that are multiplicative in the unobserved productivity of a worker. Alternatively, employers may make expectational errors and rather than assigning a zero expectation to the unobservable assign random draws based on the distributions of the unobservables. 7 4 NMII can be unequal. Thus, even when the means of the unobserved productivity-related variables are the same, and firms use the same hiring standard (’ = 0), correspondence studies can generate evidence consistent with discrimination against ethnic workers, or in their favor. So are correspondence (or audit) studies rendered useless by this criticism? In a recent paper, I show that with the right data from a correspondence study, and using the same framework as in HS, it is possible to recover an unbiased estimate of discrimination, conditional on an identifying assumption. The intuition is as follows. The HS critique rests on differences between groups in the variances of unobserved productivity. The fundamental problem, as equations (3) and (3’) show, is that we cannot separately identify the effect of ethnicity (γ’) and a difference in the relative variance of the unobservables (EII/NMII). However, a higher variance for one group implies a smaller effect of observed characteristics on the probability that an applicant from that group meets the standard for hiring. Consequently, information on how variation in observable qualifications is related to employment outcomes can be informative about the relative variance of the unobservables, and this, in turn, can identify the effect of discrimination. Based on this idea, the identification problem is solved by invoking an identifying assumption – specifically, that there is variation in some applicant characteristics in the study that affect perceived productivity and have effects that are homogeneous across groups. This is an assumption, but it implies overidentifying restrictions that can be tested in the data, because if the effects on hiring of multiple productivity controls differ between two groups only because of the difference in the variance of the unobservables, the ratios of the estimated probit coefficients for the two groups, for each variable, should be equal. To see why this works, the difference in outcomes is (4) [( I’XI* + ’ – c’)/EII] − [( I’XI* − c’)/NMII]. Impose the normalization that NMII = 1. The parameter EII is then the variance of the unobservable for the ethnic group relative to non-minorities. Denoting that parameter RELII, and dropping the prime subscripts to indicate that the coefficients in equation (4) are now ratios relative to NMII, equation (4) becomes 5 (4’) [(IXI* + − c)/RELII] − [IXI* − c]. If there is variation in the level of qualifications used as controls (XI*), and these qualifications affect hiring outcomes, then we can identify c, I/RELII, and I. The ratio of the latter two parameters identifies RELII, and identification of RELII then implies identification of γ. Note that variation in XI* that affects hiring is essential, since otherwise we cannot separately identify RELII, c, and γ. The parameters can be estimated using a heteroskedastic probit model, with the variance of the unobservable varying across groups. As an application, BM’s well-known correspondence study of race discrimination is unusual in that – for reasons unrelated to the concerns of this paper – it used data with applicants at, roughly speaking, two different levels of qualifications. BM report probit models estimated for whites and blacks separately (their Table 5). These estimates reveal higher callbacks for whites, and also substantially stronger effects of measured qualifications for whites than for blacks. Viewed in light of the preceding discussion, and assuming that the true effects of qualifications are the same for blacks and whites, the smaller estimated probit coefficients or marginal effects for blacks implies that blacks have a larger variance of the unobservable. If BM standardized applicants at low levels of qualifications, then the HS analysis would imply that there is a bias towards finding discrimination in favor of blacks, as the high-variance group would be preferred; that is, the evidence of discrimination against blacks would be stronger absent the bias from differences in the variances of the unobservables. But there is no way to assess whether the characteristics of applicants were low, since we do not know the population of applicants. Hence, implementation of the estimation procedure outlined above is likely the only way even to sign the bias, let alone to recover an unbiased estimate of discrimination. Because BM’s data include applicants with different levels of qualifications, and the qualifications predict callbacks, their data can be used to implement the methods described above. Panel A of Table 1 shows their baseline results. Marginal effects are reported for two different specifications. Estimates are shown for males and females combined, and for females only; as the sample sizes indicate, the male sample is considerably smaller. The estimated effects of race and a few of the other controls are shown. Callback 6 rates are much lower (about 33 percent) for blacks than for whites, and the race difference is robust. The table also shows that a number of the resume characteristics have statistically significant effects on the callback probability, as needed. Panel B begins by reporting the estimated overall marginal effects of race from the heteroskedastic probit model. These estimates are slightly smaller than the estimates from the simple probits, but trivially so. They remain statistically significant and indicate callback rates that are lower for blacks by about 2.42.5 percentage points (or about 25 percent). However, the marginal calculation is more complicated in the heteroskedastic probit model, because if the variances of the unobservable differ by race, then when race “changes” both the variance and the level of the latent variable that determines hiring can shift. As long as we use the continuous version of the partial derivative to compute marginal effects from the heteroskedastic probit model, there is a natural decomposition of the effect of race into two pieces: the partial derivative with respect to race affecting the level of the latent variable – corresponding to the counterfactual of race changing the valuation of the worker without changing the variance of the unobservable; and the partial derivative with respect to changes via the variance of the unobservable. In the table, these two separate effects are reported. The effect of race via shifts in the latent variable – or how race affects the employer’s valuation of worker productivity – is of greatest interest. The point of the HS critique is that differential treatment of blacks and whites based only on differences in variances of the unobservable should not be interpreted as discrimination. And the effect of race via the latent variable captures discrimination likely to be manifested in the real economy, whereas its effect through the variance is more of an artifact of the study (Neumark, forthcoming). The marginal effect via the level of the latent variable is larger than the marginal effect from the probit estimation, ranging from −0.054 to −0.086. The effect of race via the variance of the unobservable, in contrast, is positive, ranging from 0.028 to 0.062 (not statistically significant). The implication of the first marginal effect is that race discrimination is more severe than indicated by the analysis that ignores the role of differences in the 7 variances of the unobservables.9 Thus, in the context of the BM study, implementing a method that addresses the HS unobservables critique and recovers an unbiased estimate of discrimination leads to stronger evidence of discrimination. More generally, the method proposed here can be easily implemented in any future correspondence (or audit) study. All that is needed is for the resumes or applicants to include some variation in characteristics that affect the probability of being hired. As it turns out, some past audit and correspondence studies that have helped to establish the consensus that ethnic minorities face discrimination in hiring also use applicants of different quality, along a number of dimensions.10 Results from these studies are described briefly in Table 2. In almost every case, these studies indicate discrimination against an ethnic group. But given the HS critique, is it possible that in some cases these studies overstate discrimination. We cannot determine this without fully implementing the methods described above. However, if a low level of standardization was used, then we would expect stronger rather than weaker evidence of discrimination in cases where qualifications do more to boost hiring among the non-minorities than among the minorities. The evidence appears to be mixed on this point. For example, for Pager et al. (2009), the ratio of callback rates for the more- versus lessqualified applicants is higher for non-minorities – or the higher level of qualifications does more to boost non-minority employment. The same result holds for Ravaud et al. (1992). But Pager (2003) reports the opposite: the non-criminal/criminal ratio is 2.0 for whites but 2.8 for African-Americans. And the same is true in the data for indigenous versus white Australians in Booth et al. (2010). We do not actually know whether the level of standardization was high or low. The implication is that the conclusion we might draw from this body of literature, once we apply methods that can identify discrimination, could give us a different impression from the current one that nearly all the evidence points to discrimination. It is possible that the case for ethnic discrimination is not as overwhelming as it 9 The implication of the second is that, as the preceding argument implies, with low standardization a high variance increases the probability of a callback. 10 This was originally pointed out to me by Judith Rich. 8 appears.11 III. Spatial Mismatch—Is the Problem Where Minorities Live?12 Another potential source of hiring barriers for ethnic minorities is a lack of jobs near where they live, or “spatial mismatch,” driven by exogenous residential segregation and other frictions. As a result of the segregation of minorities in areas with fewer jobs, the net wage (defined as the wage minus commuting costs) is more likely to be below their reservation wage, and fewer will choose to work. This will be truer of lower-skilled minorities for whom commuting costs represent a larger share of earnings. Spatial mismatch requires frictions that prevent labor markets from reaching an equilibrium in which employment rates are largely equalized across neighborhoods. Spatial mismatch has been widely invoked to explain black employment problems in the United States. In that context, the disequilibrium is attributed to numerous factors, including the movement of jobs out of central city areas, discrimination in housing that prevents mobility of blacks to where jobs are located, customer discrimination against blacks that can reduce black employment prospects in white areas, employer discrimination that deters employers from moving to urban black areas where wages are lower, and poor information about jobs in other areas (Ihlanfeldt and Sjoquist, 1998). The role of spatial mismatch in Hispanic employment problems in the United States has been much less studied. Moreover, in the international context, there appear to be more parallels between the Hispanic immigrant population in the United States and major immigrant populations in Western Europe than there are between the situations of blacks in the United States and immigrants in Europe, including: language differences in some cases, such as Turks in Germany (Hillman, 2002) and Asians in Sweden (Zenou et al., 2006); residence in ethnic enclaves (Schönwälder, 2007; Drever and Clark, 2006); continuing economic and political ties with the origin countries of the immigrants; and of course the absence of a history of slavery. Like for blacks and Hispanics in the United States, there is considerable residential segregation of minorities in Europe (Musterd, 2005). 11 12 Analysis of these data using the methods described here is underway (Neumark and Rich, in progress). Some of the discussion in this section and the next is taken from Hellerstein and Neumark (2011). 9 Newer research testing spatial mismatch tries to incorporate direct information on access to jobs that is related to either travel time or the extent of nearby jobs (e.g., Ellwood, 1986; Ihlanfeldt and Sjoquist, 1990; Weinberg, 2000). These studies tend to show that blacks live in places with fewer jobs per person, and that this lower job access implies that blacks face longer commute times to jobs – although the differences may not be large and could conceivably be overcome relatively easily (Ellwood, 1986). However, if blacks with jobs and therefore higher incomes choose to live in areas with less job access (e.g., consuming suburban amenities), this generates a bias toward zero in the estimated relationship between job access and employment (Ihlanfeldt, 1992). Evidence of longer commute times for blacks also does not point to spatial mismatch per se, as simple employment discrimination against blacks can imply fewer job offers and hence on average longer commute times for blacks even if they live in the same places as whites. Overall, two comprehensive reviews argue that there is a good deal of evidence consistent with the spatial mismatch hypothesis (Holzer, 1991; Ihlanfeldt and Sjoquist, 1998), although Jencks and Mayer (1990) provide a more negative assessment of the hypothesis. In recent work, Hellerstein et al. (2008) ask whether other sources of barriers to minority employment could erroneously be interpreted as spatial mismatch. The pure spatial mismatch hypothesis implies that it is only the location of jobs, irrespective of whether they are held by blacks or whites, which affects employment prospects. But if discrimination, or labor market networks in which race matters, play important roles, then the distribution of jobs held by members of one’s own race may be a more important determinant of employment status. Given that urban areas with large concentrations of black residents may also be areas into which whites tend to commute to work, it is possible that the employment problems of low-skilled inner-city blacks may not reflect simply an absence of jobs where they live, even at appropriate skill levels, but rather that the jobs that do exist are more likely to be available to whites. Hellerstein et al. (2008) therefore study whether the relationship between job access and employment of blacks is driven solely by the spatial distribution of jobs, or whether the racial composition of those jobs is also important in explaining black employment. They construct measures of job access at a disaggregated level, using confidential Census information on place of work. In particular, they define 10 local labor markets as the zip code in which a person resides, plus all contiguous zip codes, based on evidence that about one-third of people work in these areas. These job access measures are also constructed by skill (jobs at a skill level per resident at that same skill level). The research departs from the spatial mismatch literature by introducing the idea of racial mismatch, constructing measures of job density not only by location and skill, but also by race. The regression models then estimate whether black employment is more sensitive to the spatial distribution of jobs held by blacks than to job density measured without regard to race. Note that if racial mismatch is important but one simply estimates models of the effects of overall (or skill-specific) job density on black employment, one can still find evidence suggesting that job density matters, consistent with the spatial mismatch hypothesis. The evidence is far more consistent with racial mismatch than with simple spatial mismatch. Black job density (the ratio of local jobs held by blacks to black residents) strongly affects black employment, whereas white job density (the ratio of local jobs held by whites to black residents) does not.13 And the own-race relationship is stronger at low skill levels. Thus, for blacks, the spatial distribution of jobs, alone, is not an important determinant of black urban employment, but rather it is the interaction of the spatial distribution of jobs combined with a racial dimension in hiring, or “racial mismatch,” that matters. Some of the main evidence is reported Table 3. Columns (1) and (4) report linear probability estimates for employment using a measure of overall job density broken down by race (for men and women combined and for men only and blacks and whites only). Only job density for blacks is substantively related to the employment of blacks. In each case, the estimated coefficient on the black job density measure is larger than that of the non-black or white job density measure by a factor of about 10. When job density is measured based on lower educational levels – at most a high school degree (columns (2) and (5)), and high school dropouts (columns (3) and (6)) – the main difference is that the estimated effects of black job density are higher. The sharp differences in the estimated coefficients of the black versus the non-black 13 The finding that black employment tends to be higher when black job density is high is not tautological. The job density measure captures jobs located in an area divided by residents of that area, not the employment rate of residents. 11 or white job density measures (differences that are strongly statistically significant) indicate that black job density is a much more important determinant of black employment than is non-black or white job density.14 Together, this evidence is consistent with the notion that the spatial distribution of jobs matters for the employment of less-educated blacks, but it is only the spatial distribution of jobs held by blacks that matters. Thus, the racial mismatch hypothesis is a better characterization of how the spatial distribution of jobs affects black employment. Even if blacks reside in areas that are dense in jobs at their skill level, if these jobs tend to be held by whites, the employment of black residents can be quite low. Moreover, descriptive statistics reported in Hellerstein et al. (2008) show that the density of jobs where blacks live is in fact quite high, even at low skill levels, suggesting that what is more important is which group is more likely to get hired. Table 4 presents another way of making the point that the spatial distribution of jobs, per se, is not very important. The estimated coefficients from an employment model are used to calculate the employment probability that would be implied if blacks lived where the representative white lived. Attention is restricted to high school dropouts, for whom spatial mismatch (whether race-specific or not) is most important.15 Panel A shows an employment rate gap of 0.231. Panel B reports estimates from the simplest model with race-specific job densities. The estimates reflect the same finding as above; the estimated effect of black job density is more than 10 times that of white job density. Panel C reports the means of the job density measures for blacks. There are considerably more white male jobs per black resident than black male jobs per black resident, averaged across blacks. Panel D, instead reports the 14 Hellerstein et al. (2008) also estimate specifications including interactions of the race- and education-specific job density measures with dummy variables for individuals’ education levels. Regardless of the education level for which density is measured, the effect of black job density on black employment is much stronger than the effect of the corresponding non-black job density, but the difference is much larger for less-educated workers. In addition, the effect of black job density for less-educated blacks is stronger when this job density is defined based on less-educated workers and residents, and the relationship is strongest when looking at less-educated blacks, using the density measures defined for lower education levels. 15 This simulation ignores any general equilibrium effects of many people moving, and is therefore best thought of as calculating the change in predicted employment if a small number of black males moved to areas in which they faced the job densities of the representative white male in their MSA. Results are similar for the broader low-skill group with at most a high school education, and for black and non-black males. 12 means of the job density measures that blacks would face if they lived where the representative white in their MSA lived.16 Whites on average live in areas where there are more jobs per black resident, whether held by whites or by blacks, although the difference is far greater for jobs held by whites. Finally, employment probabilities are predicted using the estimated employment model in Panel B, but substituting the job density measures in Panel D for those in Panel C. Because both job density estimates in Panel D are higher, the predicted employment rate for blacks is higher. However, because the effect on black employment of white job density – which is what would increase most sharply if blacks lived where whites lived – is so small (0.002), the simulated change in residential location has very little effect on the predicted probability of employment for blacks. The new predicted black probability is only higher than the actual mean by 0.025, which is a small share (10.8 percent) of the race difference in employment rates for these groups. Thus, the evidence indicates that, in contrast to the spatial mismatch hypothesis, changing the spatial distribution of black residents would do very little to increase black employment. In contrast, the racial mismatch model predicts that simply shifting black residents to areas with high job density (even at the appropriate skill level) is unlikely to do much to increase black employment. More recent research establishes that the results are very similar for Hispanics in the U.S. labor market (Hellerstein et al., 2010). Columns (1) and (3) of the top panel of Table 5 report estimates of the similar equation to that used for blacks, but using a measure of overall job density broken down by Hispanic ethnicity. The estimates indicate that only job density for Hispanics is substantively related to the employment of Hispanics. In columns (2) and (4) job density is based on poor English speakers. The estimated effects of non-Hispanic job density are very small, and insignificant in one case, while the estimated effects of Hispanic job density are much larger. The bottom panel presents similar evidence, but distinguishing workers by immigrant status instead. Again, it is principally Hispanic job density that matters for Hispanic employment. 16 The first step is to compute the average job densities (on a per black resident basis) for white male high school dropouts. These are averaged across whites in the MSA, assigned to each black based on their MSA of residence, and then averaged across blacks. 13 Labor market discrimination at a local level could give rise to a finding that blacks (Hispanics) are much more likely to be employed when they live in areas where many other blacks (Hispanics) hold jobs, but not when they live in areas where many non-blacks (non-Hispanics) are employed. For example, if the distribution of discriminatory employers or employees varies across areas, this kind of variation in minority employment could arise. Yet when similar “racial mismatch” specifications are estimated for white males, only white job density is associated with increases in white employment (Hellerstein et al., 2008). This casts doubt on a discriminatory explanation as the principal explanation of this result. IV. Networks—Is the Problem Who Minorities Know or Don’t Know? The similar “racial mismatch” findings for whites as well as Hispanics or blacks suggest that racially- or ethnically-stratified networks rather than discrimination may explain the results. Stronger evidence of racial mismatch for low-skilled workers and low-skilled job density measures (Hellerstein et al., 2008, 2010) also suggests that networks may be important; networks may operate along many dimensions, but geographical links between workers are likely to be more important for lower-skill jobs and workers for which labor markets are more local. A large body of evidence is consistent with labor market networks, much of it simply survey evidence indicating widespread reliance on friends, relatives, and acquaintances to find jobs (Ioannides and Datcher Loury, 2004).17 The evidence points to little difference between blacks and whites in the use of informal contacts in job search, higher rates of use of informal contacts among low-educated workers, and substantially higher rates of use of informal contacts among Hispanics. Subsequent work has noted that labor market networks may be race- (or ethnic-) based so that, for example, reliance on informal referrals in a predominantly white labor market benefits whites at the expense of other groups (Kmec, 2007). Focusing on the geographic or spatial dimension of networks, Bayer et al. (2008) look for evidence of network effects among neighbors using confidential Census data on Boston-area workers. They find that two individuals living on the same Census block are more likely to work on the same Census block than are two individuals living in the same block group but not on the same block. As long as informal networks 17 Pellizzari (2010) provides recent, similar evidence for many European countries. 14 are stronger within blocks than within block groups, but unobserved differences are similar within blocks and block groups, this evidence suggests that residence-based labor market networks affect hiring. In recent work, Hellerstein et al. (2011) assess evidence on the importance of labor market networks among neighbors. The evidence improves on Bayer et al. by looking explicitly at who works at which establishment. In addition, the approach is used to ask whether networks are racially stratified, which can help explain the evidence of racial mismatch. The study tests for the importance of residencebased labor market networks in determining the establishments at which people work, using matched employer-employee data at the establishment level, based on a large-scale data set covering most of the United States (the 2000 DEED, described in Hellerstein and Neumark, 2003). The measure of labor market networks captures the extent to which employees of a business establishment come disproportionately from the same sets of residential neighborhoods (defined as Census tracts), relative to the residential locations of other employees working in the same Census tract but in different establishments.18 The method first computes the share of an individual’s co-workers who are his or her residential neighbors, relative to the share that would result if the establishment hired workers randomly from the geographic areas where all individuals who work in the Census tract reside. Residence-based networks would predict that the share of neighbors among a worker’s co-workers would be higher than would result from the random hiring process. While random hiring provides a lower bound for the sorting of workers by neighborhoods across establishments, it is also important to construct an upper bound, because if establishments are larger than networks, perfect sorting by residence-based networks across establishments cannot occur. The measure of the importance of residential labor market networks is then the fraction of the difference between the lower bound and upper bounds of the extent to which a worker can work with neighbors that is actually observed in the data. In some of the analyses this is computed conditional on skill measures. Overall, the evidence indicates that residence-based labor market networks play an important role 18 Place of residence is treated as predetermined, potentially influencing place of work. This appears to be a reasonable assumption, because the results reported below are similar when the sample is restricted to people who have lived at the same location for five or more years but have worked at their current employer for fewer than five years. 15 in hiring. For blacks and whites, about 10 percent of the maximum amount to which residential networks could contribute to the sorting of workers by establishment is actually reflected in the sorting of workers into establishments. However, blacks and whites work in very differently-sized establishments, and when looking at much more homogeneous samples by race with respect to establishment size, the effective network isolation index for blacks is nearly double that for whites. In addition, networks appear more important for less-skilled workers, as would be expected for network connections among neighbors because of the more local nature of low-skill markets. For Hispanics, residence-based networks are considerably more important; the grouping of workers from the same neighborhoods in the same business establishments is about 22 percent of the maximum, and as much as twice as high for Hispanic immigrants and those with poor English skills. These results suggest that informal labor market networks may be particularly important for workers who are not as wellintegrated into the labor market, and for whom employers may have less reliable information. Hellerstein et al. (2010) present a different kind of analysis, for Hispanics, intended to ask whether network effects likely underlie the racial mismatch evidence described earlier. Traditional receiving areas for Hispanic immigrants have been metropolitan Los Angeles, South Texas, and South Florida. The persistent spatial distribution of immigration suggests the importance of immigrant enclaves in helping immigrants integrate into the labor market. Strikingly, however, between 1990 and 2000, when the Hispanic U.S. population doubled, Hispanics established sizable communities in cities that traditionally had small Hispanic populations, with the growth of Hispanic communities in these cities driven primarily by changes in the destinations of new migrants to the United States. For example, 1990 and 2000 Census data indicate that the Greensboro-Winston Salem-Highpoint MSA had fewer than 1,000 non-U.S.-born Hispanic adult males in 1990, but a decade later had over 20,000 (Hellerstein et al., 2010). Given the high transaction costs of migration, net migration of over 2,000 percent in a decade suggests that these new migrants had information that the returns to moving to the Greensboro area were high, or more specific information that would make the returns high for them – exactly the kind of information that labor market networks might supply. Moreover, network contacts in 16 these new communities may have been especially important in securing employment for new migrants, given that the local economies did not have long histories of Hispanic employment and employers in these areas did not have much experience with Hispanic workers, especially poor English speakers. As a consequence, if the relationship between density of jobs for Hispanics and employment of poor-Englishspeaking Hispanic residents is particularly strong in the cities that experienced rapid recent growth of Hispanic immigrants, it is likely that this relationship is driven by network effects. Regressions similar to those reported earlier, but for the top 50, 30, and 10 metropolitan areas in terms of the rate of growth between 1990 and 2000 of the non-native Hispanic working-age male population, are consistent with this prediction. Table 6 reports estimates of the specification from column (1) of Table 5. The estimation uses the aggregate job density measures defined for either all Hispanic or all non-Hispanic workers rather than the measure defined for poor English speakers alone because networks may well cross skill boundaries when workers are recruited or induced to move to new locales to find employment. Nonetheless, the expectation is still that the effect of Hispanic job density would be particularly pronounced for the poor English speakers for whom networks are likely to be most important. Relative to the baseline estimates in column (1), which repeat the earlier estimates for the full sample, the effects of Hispanic job density are quite a bit larger for the metropolitan areas with the highest Hispanic immigrant growth – especially the narrowest set of such MSAs (the top 10). There is also evidence (not in the table) that the effects of Hispanic job density for those who speak poor English are much stronger in the MSAs with high Hispanic immigrant growth, again most markedly for the top 10 cities. Perhaps more relevant with regard to hiring challenges faced by minorities is the question of whether labor market networks are racially stratified. The simple fact that networks based on neighborhood of residence are important points to racially-stratified networks. After all, given pervasive racial residential segregation in the United States, networks among neighbors have to be partially racebased. However, the methods used in Hellerstein et al. (2011) can also be used to see whether there is racial stratification of networks even within neighborhoods, with labor market information less likely to flow between black and white co-residents than between co-residents of the same race. 17 To examine this, the analysis of blacks is modified, treating the relevant set of a black worker’s neighbors and co-workers to include either blacks or whites, and hence measuring the extent to which black workers are clustered in establishments with black or white co-workers who are their neighbors – not just with black co-workers who are their neighbors.19 If networks among co-residents are racially stratified, then the likelihood that a black works with a neighbor regardless of race should be smaller than the likelihood that a black works with a black neighbor. The evidence points to weaker network connections between black and white neighbors than among black neighbors. Specifically, the empirical importance of networks disregarding the race of neighbors and co-workers falls by more than 40 percent.20 The two findings from this research – that labor market networks are important, and that these networks are racially stratified – can potentially explain the evidence of racial mismatch, i.e., that higher local job density for one’s own race or ethnic group affects employment probabilities, but higher job density for other race or ethnic groups does not. An area rich in jobs held by members of a group that is not networked strongly with residents may do little to boost employment among that group. Moreover, the existence of labor market networks that are stratified along racial or ethnic lines is consistent with evidence of establishment-level segregation by race and ethnicity, documented in Hellerstein and Neumark (2008) for the United States, and Åslund and Skans (2010) for Sweden. V. Conclusions and Discussion This chapter has discussed research on three aspects of potential barriers to ethnic hiring: discrimination, spatial mismatch, and networks. This research presents challenges to or at least questions about the importance of discrimination and spatial mismatch. With regard to discrimination, the essentially unanimous conclusion from audit or correspondence is called into question by the possibility that these studies do not actually identify discrimination. A new method of dealing with this identification problem, applied to Bertrand and Mullainathan’s (2004) study of race discrimination, ends up reinforcing the finding of discrimination. But this method – or other 19 This analysis was only done for blacks and whites. There is some other evidence consistent with racially- or ethnically-stratified networks in both the United States and Europe (Kasinitz and Rosenberg, 1996; Semyonov and Glikman, 2009). 20 18 approaches to the identification problem that may be developed in the future – needs to be applied to data from both new and existing studies to see how robust the evidence of ethnic discrimination actually is. Spatial mismatch is widely viewed as a partial contributor to the employment problems of minorities in the United States – especially blacks for whom the topic has been studied most extensively – and is often invoked in Europe as well (e.g., Gobillon and Selod, 2007; Pattacchini and Zenou, 2005). But new evidence for the United States suggests that the spatial distribution of jobs and workers disadvantages minorities for reasons to do more with the hiring side, with the density of jobs held by one’s own race or ethnic group in an area the only type of local job market density that matters. Networks provide a potential explanation of this evidence. Areas rich in jobs for unskilled blacks, for example, may provide more network connections to the labor market for other blacks, and hence increase their employment. In contrast, given some evidence that networks are racially stratified, areas rich in jobs for groups other than blacks – even at the same skill level – may not afford many opportunities for blacks. At the same time, networks may be a two-edged sword, as ethnic and racial minority groups may take advantage of labor market networks to find jobs when connections to the broader labor market are weak or unavailable. This can give rise to the well-known phenomenon of ethnic enclaves, which almost surely offers short-term labor market gains, although there is debate about whether in the longer-run these enclaves inhibit or encourage economic assimilation (Edin et al., 2003).21 Questions about the role of discrimination, spatial mismatch, and networks have a long history (e.g., Becker, 1971; Granovetter, 1974; Kain, 1968). With respect to all three of these, the economics literature is still – and in some cases newly – fertile and active, especially regarding the reinvigoration of research on discrimination spurred by the increasing popularity of field experiments and the new hypothesis of implicit discrimination (e.g., Bertrand et al., 2005; Rooth, 2010), and the seeming flood of labor economics research on labor market networks. The rising flows of economic migration into the developed economics (OECD, 2010) emphasize the continuing importance of understanding barriers to 21 This echoes a more general question as to whether jobs found through network contacts are better or worse than jobs found in other ways (Pellizzari, 2010). For evidence of positive effects once one accounts for firm-level heterogeneity, see Dustmann et al. (2011). 19 ethnic hiring, with the eventual goal of shaping policies to help eliminate these barriers. 20 References Aigner, Dennis J., and Glen Cain. 1977. “Statistical Theories of Discrimination in Labor Markets.” Industrial and Labor Relations Review, Vol. 30, No. 2, January, pp. 175-87. Åslund, Olof, and Oskar Nordstrom Skans. 2010. “Will I See You at Work? Ethnic Workplace Segregation in Sweden, 1985-2002.” Industrial and Labor Relations Review, Vol. 63, No. 3, April, pp. 471-93. Banerjee, Abhijit, Marianne Bertrand, Saugato Datta, and Sendhil Mullainathan. 2008. “Labor Market Discrimination in Delhi: Evidence from a Field Experiment.” Journal of Comparative Economics, Vol. 38, No. 1, March, pp. 14-27. Bayer, Patrick, Stephen Ross, and Giorgio Topa. 2008. “Place of Work and Place of Residence: Informal Hiring Networks and Labor Market Outcomes.” Journal of Political Economy, Vol. 116, No. 6, December, pp. 1150-96. Becker, Gary S. 1971. The Economics of Discrimination, Second Edition. Chicago: University of Chicago Press. Bertrand, Marianne, Dolly Chugh, and Sendhil Mullainathan. 2005. “New Approaches to Discrimination: Implicit Discrimination.” American Economic Review Papers and Proceedings, Vol. 95, No. 2, May, pp. 94–98. Bertrand, Marianne, and Sendhil Mullainathan. 2004. “Are Emily and Greg More Employable than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination.” American Economic Review, Vol. 94, No. 4, September, pp. 991-1013. Booth, Alison L., Andrew Leigh, and Elena Varganova. 2010. “Does Racial and Ethnic Discrimination Vary Across Minority Groups? Evidence from a Field Experiment.” IZA Discussion Paper No. 4947. Bovenkerk, F., M. Gras, and D. Ramsoedh. 1995. “Discrimination Against Migrant Workers and Ethnic Minorities in Access to Employment in the Netherlands.” International Migration Papers, No. 4, International Labour Office, Geneva, Switzerland. Drever and Clark. 2006. “Mixed Neighborhoods, Parallel Lives? Residential Proximity and Inter-Ethnic Group Contact in German Neighborhoods.” Unpublished paper, University of Tennessee. Dustmann, Christian, Albrecht Glitz, and Uta Schönberg. 2011. “Referral-Based Job Search Networks.” IZA Discussion Paper No. 5777. Edin, Per-Anders, Peter Fredriksson, and Olof Åslund. 2003. “Ethnic Enclaves and the Economic Success of Immigrants – Evidence from a Natural Experiment.” Quarterly Journal of Economics, Vol. 118, No. 1, February, pp. 329-57. Ellwood, David. 1986. “The Spatial Mismatch Hypothesis: Are There Jobs Missing in the Ghetto?” In R. Freeman and H. Holzer, eds., The Black Youth Employment Crisis. Chicago: University of Chicago Press, pp. 147-85. Fairlie, Robert W. 2005. “An Extension of the Blinder-Oaxaca Decomposition Technique to Logit and Probit Models.” Journal of Economic and Social Measurement, Vol. 30, No. 4, pp. 305-16. Fix, Michael, and Raymond Struyk. 1993. Clear and Convincing Evidence: Measurement of Discrimination in America. Washington, DC: The Urban Institute Press. Granovetter, Mark S. 1974. Getting a Job: A Study of Contacts and Careers. Cambridge, MA: Harvard University Press. Gobillon, Laurent, and Harris Selod. 2007. “The Effect of Segregation and Spatial Mismatch on Unemployment: Evidence from France.” CEPR Discussion Paper No. DP6198. Heckman, James J. 1998. “Detecting Discrimination.” Journal of Economic Perspectives, Vol. 12, No. 2, Spring, pp. 101-16. Heckman, James, and Peter Siegelman. 1993. “The Urban Institute Audit Studies: Their Methods and Findings.” In Fix and Struyk, eds., Clear and Convincing Evidence: Measurement of Discrimination in America. Washington, D.C.: The Urban Institute Press, pp. 187-258. Hellerstein, Judith, Melissa McInerney, and David Neumark. 2011. “Neighbors and Co-Workers: The Importance of Residential Labor Market Networks.” Journal of Labor Economics, Vol. 29, No. 4, October, pp. 659-95. Hellerstein, Judith K., Melissa McInerney, and David Neumark. 2010. “Spatial Mismatch, Immigrant Networks, and Hispanic Employment in the United States.” Annales d’Economie et de Statistique, Vols. 99/100, July/December, pp. 141-67. Hellerstein, Judith K., and David Neumark. 2011. “Employment in Black Urban Labor Markets: Problems and Solutions.” NBER Working Paper No. 16986. Hellerstein, Judith K., and David Neumark. 2008. “Workplace Segregation in the United States: Race, Ethnicity, and Skill.” Review of Economics and Statistics, Vol. 90, No. 3, August, pp. 459-77. Hellerstein, Judith K., and David Neumark. 2003. “Ethnicity, Language, and Workplace Segregation: Evidence from a New Matched Employer-Employee Data Set.” Annales d’Economie et de Statistique, Vol. 71-72, July-December, pp. 19-78. Hellerstein, Judith K., David Neumark, and Melissa McInerney. 2008. “Spatial Mismatch vs. Racial Mismatch?” Journal of Urban Economics, Vol. 64, No. 2, September, pp. 467-79. Hellerstein, Judith K., David Neumark, and Kenneth Troske. 1999. “Wages, Productivity, and Worker Characteristics: Evidence from Plant-Level Production Functions and Wage Equations.” Journal of Labor Economics, Vol. 17, No. 3, July, pp. 409-446. Hillman, Felicitas. 2002. “A Look at the ‘Hidden Side’: Turkish Women in Berlin’s Ethnic Labour Market.” International Journal of Urban and Regional Research, Vol. 23, No. 2, pp. 267-82. Holzer, Harry J. 1991: “The Spatial Mismatch Hypothesis: What Has the Evidence Shown?” Urban Studies, Vol. 28, No. 1, February, pp. 105-22. Ihlanfeldt, Keith R. 1992. Job Accessibility and the Employment and School Enrollment of Teenagers. Kalamazoo, MI: W. E. Upjohn Institute for Employment Research. Ihlanfeldt, Keith R., and David Sjoquist. 1998. “The Spatial Mismatch Hypothesis: A Review of Recent Studies and Their Implications for Welfare Reform.” Housing Policy Debate, Vol. 8, No. 4, pp. 84992. Ihlanfeldt, Keith R., and David Sjoquist. 1990. “Job Accessibility and Racial Differences in Youth Employment Rates,” American Economic Review, Vol. 80, No. 1, March, pp. 267-76. Ioannides, Yannis M., and Linda Datcher Loury. 2004. “Job Information, Networks, Neighborhood Effects, and Inequality.” Journal of Economic Literature, Vol. 42, No. 4, December, pp. 1056-93. Jencks, Christopher, and Susan E. Mayer. 1990. “Residential Segregation, Job Proximity and Black Job Opportunities.” In L. Lynn and M. McGeary, eds. Inner-City Poverty in the United States. Washington, DC: National Academy Press, pp. 187-222. Kain, John. 1968. “Housing Segregation, Negro Employment, and Metropolitan Decentralization.” Quarterly Journal of Economics, Vol. 82, No. 2, May, pp. 175-97. Kasinitz, Philip, and Jan Rosenberg. 1996. “Missing the Connection: Social Isolation and Employment on the Brooklyn Waterfront.” Social Forces, Vol. 43, No. 2, May, pp. 180-96. Kmec, Julie A. 2007. “Ties that Bind? Race and Networks in Job Turnover.” Social Problems, Vol. 54, No. 4, November, pp. 483-503. Korenman, Sanders, and David Neumark. 1992. “Marriage, Motherhood, and Wages.” Journal of Human Resources, Vol. 27, No. 2, Spring, pp. 233-55. Korenman, Sanders, and David Neumark. 1991. “Does Marriage Really Make Men More Productive.” Journal of Human Resources, Vol. 26, No. 2, Spring, pp. 282-307. Mincy, Ronald. 1993. “The Urban Institute Audit Studies: Their Research and Policy Context.” In Fix and Struyk, eds., Clear and Convincing Evidence: Measurement of Discrimination in America. Washington, DC: The Urban Institute Press, pp. 165-86. Musterd, Sako. 2005. “Social and Ethnic Segregation in Europe: Levels, Causes, and Effects.” Journal of Urban Affairs, Vol. 27, No. 3, August, pp. 331-48. Neumark, David. “Detecting Discrimination in Audit and Correspondence Studies.” Journal of Human Resources, forthcoming. Neumark, David. 1996. “Sex Discrimination in Restaurant Hiring: An Audit Study.” Quarterly Journal of Economics, Vol. 111, No. 3, August, pp. 915-41. Neumark, David. 1988. “Employers’ Discriminatory Behavior and the Estimation of Wage Discrimination.” Journal of Human Resources, Vol. 23, No. 3, Fall, pp. 279-95. Neumark, David, and Michele McLennan. 1995. “Sex Discrimination and Women’s Labor Market Outcomes.” Journal of Human Resources, Vol. 30, No. 4, Fall, pp. 713-40. Neumark, David, and Judith Rich. “Do Field Experiments of Markets Overestimate Discrimination?” In progress. Oaxaca, Ronald. 1973. “Male-Female Wage Differentials in Urban Labor Markets.” International Economic Review, Vol. 14, No. 3, October, pp. 693-709. OECD. 2010. “International Migration Database.” OECD International Migration Statistics. http://stats.oecd.org/BrandedView.aspx?oecd_bv_id=mig-data-en&doi=data-00342-en (viewed December 6, 2011). Pager, Devah, Bruce Western, and Bart Bonikowski. 2009. “Discrimination in a Low-Wage Labor Market: A Field Experiment.” American Sociological Review, Vol. 74, No. 5, October, pp. 777-99. Pager, Devah. 2007. “The Use of Field Experiments for Studies of Employment Discrimination: Contributions, Critiques, and Directions for the Future.” The Annals of the American Academy of Political and Social Science, Vol. 609, No. 1, pp. 104-33. Pager, Devah. 2003. “The Mark of a Criminal Record.” American Journal of Sociology, Vol. 108, No. 5, March, pp. 937-75. Pattacchini, Eleonora, and Yves Zenou. 2006. “Spatial Mismatch, Transport Mode and Search Decisions in England.” Journal of Urban Economics, Vol. 58, No. 1, July, pp. 62-90. Pellizzari, Michele. 2010. “Do Friends and Relatives Really Help in Getting a Good Job?” Industrial and Labor Relations Review, Vol. 63, No. 3, April, pp. 494-510. Ravaud, Jean-François, Béatrice Madiot, and Isabelle Ville. 1992. “Discrimination Towards Disabled People Seeking Employment.” Social Science & Medicine, Vol. 35, No. 8, October, pp. 951-8. Riach, Peter A., and Judith Rich. 2002. “Field Experiments of Discrimination in the Market Place.” The Economic Journal, Vol. 112, No. 483, November, pp. F480-518. Rooth, Dan-Olof. 2010. “Automatic Associations and Discrimination in Hiring: Real World Evidence.” Labour Economics, Vol. 17, No. 3, June, pp. 523-34. Schönwälder, Karen. 2007. “Residential Segregation and the Integration of Immigrants: Britain, the Netherlands and Sweden.” WZB Discussion Paper No. SP IV 2007-602. Semyonov, Moshe, and Anya Glikman. 2009. “Ethnic Residential Segregation, Social Contacts, and AntiMinority Attitudes in European Societies.” European Sociological Review, Vol. 25, No. 6, December, pp. 693-708. Smeeters, B., and A. Nayer. 1998. “La Discrimination a l’Acces a l’Emploi en Raison de l’Origine Etrangere: le Cas de le Belgique.” International Migration Papers, No. 23, International Labour Office, Geneva, Switzerland. Weinberg, Bruce. 2000. “Black Residential Centralization and the Spatial Mismatch Hypothesis.” Journal of Urban Economics, Vol. 48, No. 1, July, pp. 110-34. Zenou, Yves, Olof Åslund, and John Östh. 2010. “How Important is Access to Jobs? Old Question – Improved Answer.” Journal of Economic Geography, Vol. 10, No. 3, May, pp. 389-422. Table 1: Heteroskedastic Probit Estimates for Callbacks: Full Specifications Males and females (1) (2) A. Estimates from basic probit Black -.030 -.030 (.006) (.006) B. Heteroskedastic probit model Black (unbiased estimates) Females (3) (4) -.030 (.007) -.030 (.007) -.024 (.007) -.026 (.007) -.026 (.008) -.027 (.008) -.086 (.038) .062 (.042) -.070 (.040) .045 (.043) -.072 (.040) .046 (.045) -.054 (.040) .028 (.044) Standard deviation of unobservables, black/white 1.37 1.26 1.26 1.15 Wald test statistic, null hypothesis that ratio of standard deviations = 1 (p-value) .22 .37 .37 .56 X X X X X X 4,784 4,784 3,670 3,670 Marginal effect of race through level Marginal effect of race through variance Other controls: Individual resume characteristics Neighborhood characteristics N Notes: Standard probit marginal effects are reported in Panel A. Panel B uses equations (16)-(16’’) from Neumark (forthcoming). Marginal effects are evaluated at sample means. Standard errors are computed clustering on the ad to which the applicants responded, and are reported in parentheses; the delta method is used to compute standard errors for the marginal effects. Individual resume characteristics include bachelor’s degree, experience and its square, volunteer activities, military service, having an email address, gaps in employment history, work during school, academic honors, computer skills, and other special skills. Neighborhood characteristics include the fraction high school dropout, college graduate, black, and white, as well as log median household income, in the applicant’s zip code. Table 2: Results for Callback Rates for Labor Market Studies of Discrimination with Different Qualifications Study/ Year of test Australia Booth, Leigh, and Varganova (2010) Type of C.V. Number of C.V.’s sent Ethnic or minority group Callback rate % High quality High quality High quality High quality High quality Low quality Low quality Low quality Low quality Low quality 394 454 388 432 447 451 381 460 413 390 Chinese Italian Indigenous Middle Eastern White Australian Chinese Italian Indigenous Middle Eastern White Australian 24.0 38.0 33.0 22.0 42.0 18.0 24.0 21.0 22.0 28.0 Highly qualified Highly qualified Modestly qualified Modestly qualified 559 567 556 556 Disabled Able-bodied Disabled Able-bodied 2.3 5.1 1.1 4.5 Pager (2003) No criminal record No criminal record Criminal record Criminal record 200 150 200 150 African-American White African-American White 14.0 34.0 5.0 17.0 Pager, Western, and Bonikowski (2009) No criminal record No criminal record No criminal record Criminal record Criminal record Criminal record 171 171 171 169 169 169 African-American Latino White African-American Latino White 15.2 25.2 31.0 13.0 15.4 17.2 France Ravaud, Madiot, and Ville (1992) Source: Neumark and Rich (in progress). Table 3: Employment Regressions for Black Men, Alternative Race-Specific Density Measures (1) (2) (3) (4) (5) (6) Job density Non-black jobs or black jobs/ Male white jobs or male black jobs/ measure: black resident black male resident Density defined for: All LTHS+HSD LTHS All LTHS+HSD LTHS Non-black or white .001 .001 .0005 .0009 .0006 .0005 job density (.0001) (.0001) (.0001) (.0001) (.0001) (.0002) Black job density R2 .008 (.002) .013 (.002) .016 (.003) .010 (.002) .014 (.002) .018 (.002) .140 .140 .140 .140 .140 .140 Source: Hellerstein et al. (2008). There are 533,198 observations on black men, and 4,030,425 on white men. “LTHS” refers to those without a high school diploma and “HSD” represents high school graduates. Regression estimates are from linear probability models, with standard errors (robust to non-independence of observations within zip code areas) in parentheses. All specifications include controls for age (linear and quadratic terms), marital status (a dummy variable for currently married), highest education (six categories including less than high school, high school degree, some college, Associate’s degree, Bachelor’s degree, and advanced degree), residence in the central city, non-central city, and suburban residence, and MSA fixed effects. Table 4: Calculation of Effects of Space on Black-White Employment Differential, Black Male High School Dropouts A. Mean employment rates Black male employment 0.459 White male employment 0.690 B. Regression estimates of job density coefficients White male jobs/black male resident .002 (.0004) Black male jobs/black male resident .028 (.004) C. Mean job densities for black males White male jobs/black male resident 1.985 Black male jobs/black male resident 0.432 D. Mean job densities for representative white males in same MSA as black males White male jobs/black male resident 7.868 Black male jobs/black male resident 0.886 E. Predicted black male employment rate if black males faced job densities of average white male in MSA (substituting job densities from Panel D into employment model) 0.484 Source: Hellerstein et al. (2008). The estimates are from the same specification is the same in Table 3, column (6), including only high school dropouts in the sample. The sample size is 129,348. The estimates in Panel D come from computing the average job densities (on a per black resident basis) for white male high school dropouts, taking the mean across whites in the MSA, assigning these to each black based on their MSA of residence, and then averaging across blacks. Table 5: Employment Regressions for Hispanic Men, Alternative EthnicitySpecific Density Measures (1) (2) (3) (4) Male white jobs or Non-Hispanic jobs or male Hispanic Hispanic jobs/Hispanic jobs/Hispanic male Job density measure: resident resident Density defined for: All Poor All Poor English English Non-Hispanic or white .001 .0003 .001 .003 job density (.0003) (.0007) (.0003) (.002) Hispanic job density .022 (.006) .016 (.003) .018 (.005) .007 (.001) R2 Density defined for: Non-Hispanic or white job density .058 All .00 (.0003) .058 Immigrant -.0001 (.0013) .058 All .001 (.0003) .058 Immigrant .001 (.001) Hispanic job density .022 (.006) .028 (.005) .01 (.005) .017 (.004) .058 .058 .058 .058 R2 Source: Hellerstein et al. (2010). There are 625,523 observations on Hispanics. See notes to Table 5. Additional controls include dummy variables for the four Census categories of English proficiency, and for immigrant status. “Poor English” refers to the bottom two Census categories, and “Good English” to the other two. Table 6: Employment Regressions for Hispanic Men, Ethnicity-Specific Job Density Measures, With and Without English Proficiency Interactions, Cities with High Growth Rates of Non-U.S. Born Hispanics (1990-2000) (1) (2) (3) (4) Growth rate of non-U.S. born Hispanics in MSA/PMSA: All Top 50 Top 30 Top 10 Non-Hispanic .001 -.0001 -.0002 -.001 job density (.0003) (.0003) (.0002) (.0008) Hispanic job density R2 .022 (.006) .040 (.009) .037 (.012) .088 (.028) .058 .045 .044 .033 Source: Hellerstein et al. (2010). The specification corresponds to column (1) of Table 5.