Are Affirmative Action Hires Less Qualified? Evidence from Employer- Employee Data on New Hires

by David Neumark, Harry Holzer
Citation
Title:
Are Affirmative Action Hires Less Qualified? Evidence from Employer- Employee Data on New Hires
Author:
David Neumark, Harry Holzer
Year: 
1999
Publication: 
Journal of Labor Economics
Volume: 
17
Issue: 
3
Start Page: 
534
End Page: 
569
Publisher: 
Language: 
English
URL: 
Select license: 
Select License
DOI: 
PMID: 
ISSN: 
Abstract:

 

Are Affirmative Action Hires Less Qualified? Evidence from Employer-Employee Data on New Hires

Harry Holzer, Michigan State University

David Neumark, Michigan State University and 
National Bureau of Economic Research 

We use microlevel data on employers and employees from a sample of establishments in four major metropolitan areas in the United States to investigate whether Affirmative Action leads to the hiring of minority or female employees who are less qualified. Our measures of qualifications include the educational attainment of the workers hired and a variety of outcome measures related to worker perfor- mance on the job. We find evidence of lower educational qualifica- tions among women and minorities hired under Affirmative Action. However, we do not find evidence of weaker job performance among most groups of minority and female Affirmative Action hires.

I. Introduction

Affirmative Action have always been controversial, largely because of the allegation that they cause employers to give preference in hiring to less-qualified minorities or females over more-qualified white

We are grateful to Jess Reaser for outstanding research assistance, to the Rockefeller Foundation for financial support, and to Jeff Biddle, Bill Evans, Robert LaLonde, Angelo Melino, Ed Montgomery, and seminar participants at Maryland, Michigan State, the National Bureau of Economic Research, Ohio State, Rochester, Toronto, and the Upjohn Institute for helpful comments.

[Journal of Labor Economics, 1999, vol. 17, no. 31 
O 1999 by The University of Chicago. All rights reserved. 
0734-306X/99/1703-0003$02.50 

males. Survey evidence suggests that, even among whites, there is wide- spread public support for outlawing employment discrimination and also for policies that compensate the past victims of discrimination through targeted education, job training, and recruitment efforts. At the same time, policies that give "preference" in employment (or university admis- sions) to less-qualified members of these groups are strongly opposed (e.g., Lipset and Schneider 1978; Bobo and Smith 1994).'

Similarly, many critics of Affirmative Action policies support strong enforcement of Equal Employment Opportunity (EEO) laws and even compensatory recruitment and training but argue that the alleged prefer- ences for less-qualified minorities and females in most current Affirmative Action practices create labor market inefficiencies and/or inequities.' In contrast, proponents of Affirmative Action policies frequently argue that labor market discrimination continues to be prevalent, despite EEO laws.3 They also argue that gaps between different groups in relative qualifications measured by educational attainment, experience, and so on likely reflect past or current discrimination (Bergmann 1989), and that such qualifications are weak predictors of actual performance on the job, which itself may be difficult to measure, especially in the short term.4 In

In these data, much hinges on the wording of questions describing Affirmative Action; references to "quotas" or "reverse discrimination" generally elicit the most negative responses. There is also some evidence of more tolerance for compensatory policies in education than in employment, though quotas or other forms of preferences in university admissions are still widely opposed. For a view that distinguishes Affirmative Action in universities from that in employment, see Carter (1991).

Critics of current Affirmative Action policies that stress both inefficiencies and inequities (from violating principles of rewards based on individual merit) include Glazer (1975), Sowe11 (1990), and Epstein (1992). Coate and Loury (1993) argue that Affirmative Action can reduce the incentives of "preferred" groups to invest in human capital formation, while Carter (1991) also emphasizes the stigma borne bv aualified minorities because of these volicies.

, i

Evidence of discrimination against minorities or females, even when control- ling for observable credentials, can be found in several recent "audit" studies of employers (e.g., Fix and Struyk 1994; Neumark 1996), in which matched pairs of applicants with comparable credentials but of different races or sexes are sent out to apply for identical jobs.

The difficulty of predicting or observing individual-level productivity has been stressed in the "statistical discrimination" literature (e.g., Cain 1986), and more recently by Altonji and Pierret (1996) and Oettinger (1996). The question of whether required qualifications accurately predict job performance, and whether any resulting "disparate impacts" across demographic group can be considered discrimination, has been addressed in a variety of court cases (e.g., Griggs v. Duke Power in 1971 and Ward Cove v. Atonio in 1989) and in the Civil Rights Act of 1991. The debate over the latter, including the allegations of critics that such

this view, the efficiency costs of preferential policies are considered very small or even negative.

To a large extent the debate over Affirmative Action hinges on whether labor markets in the absence of Affirmative Action are better character- ized by competition or by discrimination. If Affirmative Action is used in establishments that otherwise would discriminate in hiring, there would not necessarily be any shortfall in the qualifications of minorities or females hired there relative to white males. In fact, if more qualified minorities and women were being passed up because of discrimination, Affirmative Action could lead to the hiring of more productive minorities and females. On the other hand, if Affirmative Action is imposed on (or chosen by) firms with no such discriminatory practices, it could lead to the hiring of employees with weaker qualifications or performance, since such firms were presumably already hiring the best-qualified individuals regardless of race or sex.

The likelihood that Affirmative Action leads to the hiring of less- qualified minorities and women may be reinforced because the educa- tional attainment of minorities (especially among Hispanics) continues to lag behind that of whites, as do their average test scores (Mare 1995; Neal and Johnson 1996). These shortfalls in average qualifications might reflect past discrimination in the labor market or current discrimination in the housing market, either of which could lead to lower schooling attainment of young people through their effects on family incomes or neighborhood quality (Yinger 1995; Cutler and Glaeser 1997). Of course, as we empha- size below, education (and test scores) may represent only a narrow set of relevant job qualifications. Regardless, lower distributions of qualifica- tions among some groups of minorities or women would make it more likely that Affirmative Action would lead to less-qualified minority or female hires in both discriminating and nondiscriminating companies. Thus, evidence on the presence or absence of discrimination in labor markets is an important component of the Affirmative Action debate but is not sufficient to fully evaluate the policy.

Given these considerations, the current debate on Affirmative Action turns heavily on the existence and empirical magnitudes of shortfalls in qualifications and job performance of women and minorities hired under Affirmative Action. Yet despite the intensity of the viewpoints held, the evidence to date on this issue remains quite thin. To the best of our knowledge, the only systematic empirical study of the effects of Affir- mative Action on productivity is that of Jonathan Leonard (1984). He estimates production functions using state-by-two-digit industry data in

legislation constituted a "quota bill," indicates the difficulties of distinguishing strong EEO policies from Affirmative Action.

manufacturing, in which the fractions of employment accounted for by minorities/females and by federal contractors (who are typically required to have Affirmative Action plans) appear as independent variables. But the usefulness of this approach is limited by the highly aggregated nature of the data and its focus only on manufacturing.' Also, Badgett (1995) provides more qualitative evidence on the effects of Affirmative Action on the qualifications of minority hires from a case study of a large manufacturing firm and concludes that-at this company-Affirmative Action did not result in lower-quality hires, based in part on evidence that upper-level management promoted Affirmative Action as a means of finding the best employees.

In this article, we provide what we believe to be the first survey-based microlevel empirical evidence on the relative qualifications of workers hired under Affirmative Action. Using data on new hires collected from establishments, we are able to compare the qualifications of women or minorities hired under Affirmative Action to those of white men in comparable jobs and to those of women or minorities or white men hired in establishments that do not use Affirmative Action. Because workers' qualifications and performance are multifaceted and difficult to measure, we take a broad-ranging approach, focusing on a variety of variables related to both qualifications and job performance.

11. The Data The Sample

The data we use are from a survey that was administered to about 800 establishments in each of four metropolitan areas: Atlanta, Boston, De- troit, and Los Angeles (see Holzer 1995,1996, and 1998).~ The survey was administered between June 1992 and May 1994, as the national economy was recovering from the recession of the early 1990s.'

Most of Leonard's articles deal with the effects of Affirmative Action on relative employment, occupational status, and the like. These are reviewed in various survey papers (1989, 1990); see also Smith and Welch (1984). Given the small effects of Affirmative Action on the overall composition of employment that Leonard shows in his other work, it is very unlikely that he would find strong effects on productivity at the two-digit industry level.

The data are drawn from a broader project known as the Multi-City Study of Urban Inequality, which also consists of household surveys and an in-depth, qualitative study of a smaller sample of establishments.

'The survey was administered to establishments in Detroit between June 1992 and February 1993 and the other areas between March 1993 and May 1994. The timing was deliberately chosen in order to coordinate with the surveying of households in each area, as part of the Multi-City Study. Monthly unemployment rates averaged under 6% in Atlanta and Boston during the survey period and between 8% and 10% in Detroit and Los Angeles. Dummy variables for metro-

The survey was administered over the phone and averaged roughly 35 minutes in length. The sample of establishments was drawn from two sources: (I) a listing of establishments and their phone numbers provided by Survey Sampling, Incorporated (SSI), which is drawn primarily from local phone directories and supplemented by other sources and (2) the establishments of employment of respondents in the household surveys that were administered in each of these four metropolitan areas. The latter were drawn in order to generate a sample of "matched pairs" of individ- uals and establishments. For the establishments in the SSI part of the sample, the main respondent to the survey is the person who is respon- sible for hiring noncollege workers. For the sample drawn from the household survey, the respondent is the person responsible for hiring into the occupation of the household respondent.'

A number of steps were taken to ensure that the data could be used to draw inferences regarding the underlying population. Sample weights were generated to account for any differences in establishment character- istics that might be attributable to these different sampling strategies, so that we can pool data from these two sources. Despite the differences between these two sources, both were designed to generate employee- weighted samples of establishments. For the SSI sample, this was accom- plished by ex ante stratification of the sample based on establishment size, with the distribution of establishments chosen to approximate the distri- bution of employees across size categories in the workf~rce.~

For the household-generated sample, the distribution of establishments should approximate the distribution of employment in the population by con- struction, at least when household respondents are sample weighted.'' Thus, no additional size weighting of establishments is necessary with this sample. When focusing on the characteristics of each establishment's most

politan area and year are included in the multivariate analyses that follow to control for these differences in local labor market conditions.

* Most characteristics of workers and jobs do not differ significantly across the samples of establishments generated by the two data sources.

The stratification scheme was as follows: 25% in establishments with fewer than 20 employees; 50% in establishments with 20-99 employees; and 25% in those with 100 or more employees. These distributions were drawn from a weighted sample of establishments in the Employment Opportunity Pilot Project (EOPP) of 1980 and 1982.

lo Sample weights are applied to the household-generated establishments that adjust for (1) the underrepresentation of jobs requiring college, since the SSI sample focused on noncollege jobs; (2) the oversampling of low-income and minority residents in the household surveys; and (3) the incompleteness of the Boston and Los Angeles samples of households from which establishments were drawn. More information on the construction of these weights is available from the authors.

recently filled job, the sample will provide extra weight to establishments that do a lot of hiring because of their size (but not because of high turnover)." The sample of new jobs should thus approximate the stock of available jobs.

The overall response rate for the survey was 67% for establishments that were successfully screened." This response rate compares favorably with other phone surveys of employers (e.g., U.S. Department of Labor 1993). In addition, because we have some measured characteristics of establishments in the SSI sample that did not respond to the survey (i.e., establishment size, industry, and location), we could check for differences in response rates across these observable categories that might generate sample selection bias. Few significant differences were found in response rates across the categories measured by these variables, and response rates for all categories were within .I0 of the mean.

Comparisons of the industries and sizes of establishments in our sam- ple were also made with the most recently available published data from County Business Patterns for these metropolitan areas, and the two samples appeared to be quite comparable. Finally, we compared the distributions of occupations among our most recently filled jobs with those in the 1990 census of population for these areas and with the distributions of occupations and worker characteristics among all em- ployees in our establishments to see whether or not the sample of "marginal" employees (i.e., new hires) differs greatly from the "average." While we found only minor evidence of such differences, our results speak more directly to qualifications of new hires.

The Affirmative Action Variable

The variables used in this study are primarily drawn from survey questions on the last job filled and the worker hired into that job. In particular, respondents were asked whether or not "Affirmative Action or Equal Employment Opportunity Law play any role in your recruiting activities for this position" and also whether or not these factors "play any role in whom you actually hire."13 The measure we use for Affirma-

11 The lack of extra weighting for high-turnover establishments seems appro- priate since a single job that turns over frequently is only available to a single worker at any time. Unfortunately, there was no easy way to put extra weight on establishments whose rate of hiring is temporarily high due to their net employ- ment growth.

12 Successfully screened establishments were those where the correct establish- ments and the person responsible for new hiring into the relevant types of positions were contacted.

l3 Clearly, a positive response to this latter question does not necessarily mean that the worker was an "Affirmative Action hire" but just that Affirmative Action

tive Actim in this article is the latter because lower qualifications or performance are more likely when Affirmative Action is used in hiring rather than in recruiting. In addition, data on Affirmative Action in hiring are obviously more pertinent to who actually gets hired, which is the focus of the debate.

Given the wording of the question, there is some ambiguity over whether we are picking up the effects of Affirmative Action or EEO law more generally. In addition, the variety of reasons a respondent might indicate use of Affirmative Action raises questions about the exogeneity of the variation in Affirmative Action across establishments. All firms with 15 or more employees are bound by Title VII of the Civil Rights Act, which covers just over 80% of the private sector workforce (Bloch 1994). On the other hand, explicit Affirmative Action plans are manda- tory only for employers with federal contracts and 50 or more employees or a contract worth $50,000 or more. In our sample, 60% of establish- ments report the use of some kind of Affirmative Action in recruiting and 45% in hiring. Although our self-reported measure is not explicitly based on actual contractor status, the fraction of units reporting Affirmative Action and some of their characteristics are similar to what Leonard (1989) found regarding federal contractor status among firms required to file EEO-1 reports. Specifically, 60% of his sample were contractors who were therefore bound by executive orders mandating the use of Affirma- tive Action; and the use of Affirmative Action in Leonard's data was highly correlated with employer size (at the firm level, in his case), as it is in our data. On balance, then, we believe that our measure of Affir- mative Action is likely to be related to contractor status-which is to some extent exogenous-and is therefore much more likely to be captur- ing explicit Affirmative Action in hiring, rather than mere compliance with EEO law.I4

In addition to the role of contractor status, establishments or firms may behave as ifthey have an explicit Affirmative Action plan, even when they do not." This might arise if management feels some pressure to improve

played a role in the decision. This is reflected in the considerable number of observations in which employers provide a positive response to this question but hire a white male (see table 1). An instructive example comes from academic hiring, where Affirmative Action nearly always plays a role in the hiring decision but by no means precludes hiring white males.

l4 Because the survey used in this article was not designed explicitly to study Affirmative Action, it did not elicit information that might further clarify the interpretation of the questions, such as federal contractor status or the size of federal contracts.

15 Conversely, given that there is imperfect compliance with Affirmative Action

guidelines for federal contractors (as emphasized, e.g., by Leonard 1989), employ

the hiring and promotion of protected groups. We presume this is espe- cially likely for establishments belonging to firms with 100 or more employees, which are required to file EEO-1 forms with the government. Although only compliance with EEO laws is required of firms without federal contracts, those who file EEO-1 reports may engage in Affirma- tive Action hiring to avoid litigation or other problems that could be triggered by these reports; as noted by Bloch (1994), employment dis- crimination suits can stem from charges brought by the Equal Employ- ment Opportunity Commission (EEOC) based on review of EEO-1 reports, even without a single complainant. In this case, establishments belonging to such firms would likely report the use of Affirmative Action in hiring, and their use of Affirmative Action would be exogenously determined by their size and the practice of federal monitoring.

Of course, there may be other reasons why employers might report that they are using Affirmative Action besides their status as federal contractors or their size. In addition to requirements for contractors, explicit plans may be implemented by the courts as a remedy for a finding of past discrimination (see the discussion in Epstein 1992, ch. 19), or they may be implemented by companies as a deterrent to claims of discrimi- nation under Title VII (see Badgett 1995) or to increase workplace diversity for other reasons. These latter plans are voluntarily chosen and have been permitted by the courts if they are based on a specific plan, correct a previous imbalance, protect the interests of non-Affirmative Action candidates, and will end when specific goals are met (Gold 1993). We know of no actual count of the number of such plans that exist and suspect that the number may be fairly small; but their existence at least raises the possibility that an employer's decision to use Affirmative Action may be endogenous with respect to other outcomes that we are measuring, an issue we explore further below.

Other Variables Used in the Study

The other job-specific questions include whether or not a college degree is required for the job; whether or not high school, specific previous experience, vocational training, or references are each required;16 whether each of a set of tasks (dealing with customers directly, reading or

ers in principle bound by such guidelines may not in fact be adhering to them. Thus, a self-reported measure of Affirmative Action use may be preferable to a measure based on federal contractor status or explicit Affirmative Action plans.

l6We code most of these requirements as dummy variables which take on a value of one if the requirement is "absolutely necessary" or "strongly preferred" at the time of hiring and zero if it is "mildly preferred" or "doesn't matter." In contrast, the college requirement is based on an explicit "yes" or "no" question in the survey.

writing paragraphs, arithmetic, or computer use) is performed daily on the job; and a set of one-digit occupational dummies. The characteristics of the last worker hired include race/ethnicity, sex, age, and educational attainment as well as information on wages, promotions, and a supervi- sor's performance rating. Establishment-specific characteristics include establishment size, percent of workforce covered by collective bargaining, one-digit industry dummy variables, and dummy variables for location within the central city of the metropolitan statistical area (MSA). We also use variables for the race and sex of the main respondent to the survey and the worker's supervisor, and the racial composition of the establishment's customer pool, to control as much as possible for determinants of racial or gender preferences among survey respondents or other agents.

111. Descriptive Statistics on Differences in Establishment, Job, and Employee Characteristics by Use of Affirmative Action

Descriptive Statistics

We begin in table 1 by providing simple descriptive information on differences between worker-establishment matches in which Affirmative Action is used in hiring and those in which it is not. Looking first at demographic characteristics of recent hires, we see that the largest abso- lute difference between hires in which Affirmative Action plays a role and those in which it does not is between white males and females, where the proportion of recent hires accounted for by these groups is .06 lower and .07 higher, respectively, when Affirmative Action is used in hiring. The proportions of recent hires among the other demographic groups are not very different between Affirmative Action and non-Affirmative Action hiring; the proportions of black females and Hispanic males are actually slightly lower. Of course, these are univariate comparisons, and the influence of Affirmative Action on hiring by demographic group may change once account is taken of establishment, job, and other individual characteristics."

Part B of table 1provides information on worker education, while part C provides descriptive information on job requirements for the jobs into which the recent hiring occurred. We see that workers hired under Affirmative Action tend to be more educated, and that skill requirements are higher for these jobs for each of the requirements included in the survey. These results suggest that we may have to compare qualifications

"These results also differ somewhat from those of Leonard (1989) for the 1970s, who finds proportionately bigger effects on black males and females than on white females. But during the 1980s, Leonard finds relative employment of black males and females declining at contractor establishments (relative to non- contractors).

Table 1 Descriptive Statistics

A. Worker demographic characteristics: White male White female Black male Black female Hispanic male Hispanic female Asian

B. Worker education: Dro~out GED High school graduate Trade school or some college Associate's degree Bachelor's degree Some graduate school Graduate degree

C. Job requirements: Hi h school degree coTlege degree References Vocational training Specific experience Customer contact Readinghriting Math Computer

D. Occupation: Management/professional Sales Clerical Agricultural Crafts Operative Labor Service

E. Establishment characteristics: Employer size:

Median Central cit NonccntraPcity MSA Percentage of workforce covered by

collective bargaining 
Percentage black customers 
Percentage Hispanic customers 
Construction 
TCPU 
Wholesale trade 
Retail trade 
FIRE 
Services 

Affirmative Action Affirmative Action Used in Hiring Not Used in Hiring

M SE M SE

(1) (2) (3) (4)

Table 1 (Continued )
--
Affirmative Action Affirmative Action
Used in Hiring Not Used in Hiring
M SE M SE
(1) (2) (3) (4)
Public .02 ,004 .01 ,003
Nondurables manufacturing .10 .01 .09 .01
Durables manufacturing .10 .01 .12 .01
F. Respondent and supervisor        
characteristics:        
Respondent's sex different from        
worker's sex .50 .02 .39 .01
Respondent's race different from        
worker's race .37 .02 .36 .01
Supervisor's sex different from        
worker's sex .62 .02 .68 .01
Supervisor's race different from        
worker's race .30 .01 .28 .01

NOTE.-There are 1,033 observations in cols. 1 and 2 and 1,480 observations in cols. 3 and 4. There are fewer observations for age, education, percentage of customers in each race group, and the demographic characteristics of the supervisor and respondent, because of missing data. In the following tables, a dummy variable for missing data on these variables is included, and the variables are set to zero. GED = general equivalency diploma; MSA = metropolitan statistical area; TCPU = transportation, commu- nications, and public utilities; FIRE = finance, insurance, and real estate.

of women and minorities relative to white males within the subset of establishments using Affirmative Action; otherwise, we might incorrectly conclude that Affirmative Action hires are more qualified.

Consistent with the above results, part D of table 1 reveals that a greater proportion of hiring in establishments using Affirmative Action is into management/professional and clerical jobs than into blue-collar or service jobs.'' Of course, these results do not indicate whether Affirma- tive Action hiring is used more for such occupations, or whether Affir- mative Action leads to more hiring into such occupations. Given these questions, it is unclear whether it is always appropriate to control for occupation (or required skills), which generates only within-job estimates of Affirmative Action effects.

Part E of table 1 provides descriptive information on establishment characteristics, broken down by whether or not Affirmative Action was used in recent hiring. Establishments using Affirmative Action are much larger, have a significantly higher proportion of the workforce covered by collective bargaining, and are significantly more likely to be in the services

l8 This evidence is broadly consistent with Leonard's (1989) findings that Affirmative Action has created the most opportunities for white women in white-collar trainee positions and for black females in managerial, sales, clerical, laborer, and white-collar trainee positions.

Table 2

Logit and Multinomial Logit Estimates of Effects of Affirmative Action on

Demographic Group of Hire

Minority or white female ,027 ,036 ... ... (.019) (.019) White female ... ... .077 ,051

Black male ... ... (.023) ,002 (.025) .015
Black female        
Hispanic male Hispanic female ... ... ... ... (.009) -.016 (.007) -.006 (.009) -.003 (.006) -.003
White male -.042 -.051 (.008) -.060 (.007) -.071
Occupation and job (.019) (.019) (.022) (.025)
requirement controls No Yes No Yes

NoTE.-T~~table reports partial derivatives of the probability of each outcome with respect to the Affirmative Action dummv variable, evaluated at the weighted means. Standard errors calculated from a linear approximation to these derivatives treating the means as fixed are reported in parentheses. Cols. 1 and 1' are based on a logit model, and cols. 2 and 2' on a multinomial logit model. There are 2,513 observations. All specifications include dummy variables for citv and year and establishment controls. Asian hires are also included as an outcome, but results are not'reprted. In cols. 1 and l', "minority" refers to blacks and Hispanics. Estimates are weighted.

industry. These establishments also have higher percentages of black and Hispanic customers.

Finally, part F of table 1 reports on the demographic characteristics of supervisors and respondents to the survey. We see that respondents- who are responsible for hiring-are more likely to be the opposite sex from the new hire in establishments that report using Affirmative Action in hiring, although this is not true of supervisors. In the analyses that follow, we control for the possibility that these characteristics of respon- dents or supervisors affect the outcomes over which respondents or supervisors exert some control-such as promotions and performance ratings.

The Effects of Affirmative Action on Hiring of Women and Minorities

Table 2 presents a multivariate descriptive analysis of the relationship between Affirmative Action hiring and the demographic group of the recent hire, based on logit or multinomial logit estimates accounting-in different specifications-for differences in establishment characteristics, in the occupational distribution of hires, and in job requirements.'9 In columns 1 and 1 ', the dependent variable is a dummy variable for whether the hire was a woman or minority, and in columns 2 and 2', the dependent variable is a categorical indicator for the more-detailed demographic

group of the hire. In columns I and 2, controls for city and year of hire, as well as establishment characteristics (listed in table I), are included. The probability that a minority or female was hired is estimated to be signif- icantly higher, by .03, when Affirmative Action is used in hiring." Columns I' and 2' add controls for the occupations and job requirements listed in table 1. The association between Affirmative Action and hiring of minorities or females strengthens slightly when we include these controls. The estimates in column 2' indicate that the probabilities that white females and black males are hired, relative to the probability that a white male is hired, are higher when Affirmative Action is used in hiring (significantly so only for white females); the hiring of black females and Hispanic males and females appears to be unaffected." Comparing effects on probabilities to the means in table 1 for each group, we see that Affirmative Action is associated with increases of about 15% in the probabilities of hiring white women and black men. On the other hand, the last row indicates that the probability that a white male is hired is lower by about 20% under Affirmative Action.*' Presumably, most of these white males are then hired in establishments not using Affirmative Action; these establishments likely pay less, among other reasons because they are smaller.23

19 In this table and all that follow, estimates are sample weighted (as they were in table I) to produce more accurate estimates of "average" effects. We recom- puted the specifications unweighted, and most of the results were very similar. One exception is discussed below. The unweighted results are available from the authors upon request.

20 rr Minority or female" refers to white women, blacks, and Hispanics. All of the models estimated in the article also included categories for Asian men and women. However, we do not focus on (or report) results for Asians, since most of the debate seems to be about the treatment of relatively disadvantaged sub- groups of the population, a categorization which may not apply to Asians. Omitting the Asians had virtually no effect on the reported results.

The results are partly consistent with those reported by Rodgers and Spriggs (1996), who find that federal contractor status is associated with a higher per- centage of black workers (by 12%) and a lower percentage of Hispanic workers (by 0.45%) in 1992.

22 Note that the combined marginal effects reported for minorities or white females are smaller in absolute value than those for white males. This is because Asians are also included in the estimation, and their hiring appears to be boosted b Affirmative Action.

"'Further, if there is some job segregation by demographic characteristics, or if overall hiring at establishments using Affirmative Action falls because of the

These results, coupled with evidence of strengthened enforcement activity in 1989 and especially in 1993 (Anderson 1996), suggest that Affirmative Action has real consequences for hiring behavior. We next turn to the more contentious issue of the relative qualifications and performance of women and minorities hired under Affirmative Action.

IV. Relative Qualifications and Performance of Affirmative Action Hires

The Empirical Approach

Having documented the associations between Affirmative Action and hiring of minorities and women, we now turn to the question of whether Affirmative Action leads to the hiring of less-qualified women or minor- ities. Our approach is to estimate equations for Q, a measure of the qualifications or performance of the last worker hired, of the form:

where AA is a dummy variable indicating that Affirmative Action was used in hiring, WM is a dummy variable for white males, D is a set of dummy variables for each other demographic group considered, X is a set of job characteristics, and Z a set of establishment characteristics; i, j, and k denote the last worker hired, the most recent job filled, and the establishment, respectively. Note that we include separate intercepts for each demographic group distinguished by the Affirmative Action status of the establishment and no common intercept.

This specification provides us with a number of potential comparisons for estimating the effects of Affirmative Action in hiring. One interesting comparison is between women or minority hires and white male hires in establishments using Affirmative Action. The difference in Q for this comparison, which is given by (P -y), addresses the question of whether Affirmative Action leads to hiring of women or minorities who are less qualified than the white male workers hired in similar establishments. A second interesting comparison is between women or minority hires in establishments using Affirmative Action and women or minority hires in establishments not using Affirmative Action, which is captured by the

policy, the wages of white males in non-Affirmative Action establishments will be lower because of an outward labor supply shift. However, table 7 below shows that these white males still earn higher wages on average than women or minor- ities hired at either type of establishment.

difference (P -6). A third comparison is between women or minority hires in establishments using Affirmative Action and white male hires in establishments not using Affirmative Action, which is measured by (P -a). This comparison might pertain to white male workers who would otherwise have been hired in establishments using Affirmative Action had the policy not been in place.

We suspect that the first difference, (P -y), may be most relevant to the policy debate. However, there is the potential for misleading infer- ences to be drawn because of differences in skills or qualifications be- tween women or minorities and white males that exist independently of Affirmative Action. For example, suppose that the estimate of (P -y) is a large negative number, indicating that women or minorities hired into establishments using Affirmative Action are less qualified than white males hired into similar establishments. However, there may be an econo- mywide shortfall in qualifications of women or minorities relative to white males, in which case workers in non-Affirmative Action establish- ments should serve as a control group for overall differences between minorities or women and white men. For example, Hispanics may be perceived as less qualified or perform less well because of language barriers, regardless of whether the establishments into which they are hired use Affirmative Action. Then to estimate the independent effects of Affirmative Action on differences between white male and Hispanic workers, we subtract off any shortfall in qualifications attributable to language in establishments not using Affirmative Action and ask instead whether the shortfall is relatively larger in establishments using Affirma- tive Action. In general, the shortfall in qualifications (or performance) in establishments not using Affirmative Action is measured by (6 -a), leading to the difference-in-differences estimate of (P -y)-(6 -a).24 Thus, critics of Affirmative Action may make a potential error if they base their criticisms only on observed shortfalls in qualifications at es- tablishments using Affirmative Action rather than on the net difference in these shortfalls between the two sectors.

On the other hand, the assumption underlying the difference-in-dif- ferences estimation is that the policy does not affect the control group-in this case, establishments not using Affirmative Action. Under this as- sumption, (6 -a) measures the difference between women or minorities and white males in the absence of Affirmative Action. However, this

24 Note that the difference-in-differences estimator (P -y) -(6 -a)is equal to the difference between the first and third comparisons, (P -y) -(P -a),plus the second comparison, (p -6). Note also that this estimator does not distinguish between a larger difference in qualifications in establishments using Affirmative Action that arises from lower standards for women or minorities, or higher standards for white men, relative to establishments not using Affirmative Action.

assumption would be invalid if Affirmative Action affects who is hired at all types of establishments, whether or not they use Affirmative Action. In particular, Affirmative Action could create or enlarge a shortfall in qualifications among women and minorities at both types of establish- ments-by drawing the most-qualified women and minorities into estab- lishments using Affirmative Action and pushing the least-qualified white males out of these establishments-in which case the difference-in-dif- ferences estimate could understate or obscure the effect of Affirmative Action. As a simple example, suppose that initially establishment A employs some men with productivity (per unit of time) equal to four and other men with productivity of three, while establishment B employs some women with productivity of two and other women with produc- tivity of one. After implementing Affirmative Action, establishment A hires the more productive women and fires the less productive men, who are hired by establishment B. Thus, establishment A employs men with productivity of four and women with productivity of two, while estab- lishment B employs men with productivity of three and women with productivity of one. In this case, Affirmative Action has led to the hiring of less-productive women, and (P -y), which equals -2, reflects this. But because (6 -a) also equals -2, the difference-in-differences estimator gives no indication that Affirmative Action led establishment A to hire less-productive women over more-productive men." Although this is a contrived example, it illustrates why the simple difference may be pref- erable to the difference-in-differences estimator. Throughout, we report both. Table 3 summarizes the alternative estimates we present.

Table 4 presents simple means for the dependent variables we analyze, by demographic group and whether the firm reports using Affirmative Action. This table is useful when examining the estimates of equation (1) in the tables that follow, since it provides information on the levels of the dependent variables for each of these groups. In addition, the table is useful in providing an illustration of our empirical approach. Consider, for example, educational levels of black men versus white men. If we simply look at firms reporting using Affirmative Action, the educational difference is -1.39 years (13.22 -14.60). This corresponds to (P -y) in equation (I). Note, though, that the educational attainment of black men also falls short in the non-Affirmative Action sector, by 0.89 years, which corresponds to (6 -a) in equation (I). Thus, the difference-in-differences

25 One might argue that in this simple example, Affirmative Action is irrelevant because it simply results in a reshuffling of workers. But if wages are partly attached to jobs and do not just reflect productivity, perhaps because of restric- tions on paying men and women different wages at the same establishment, then the men in this example are likely to suffer wage declines and the women to gain wage increases.

Table 3 Alternative Estimators of Differentials in Qualifications or Performance Associated with Affirmative Action

Equation:

Simple differences: p -7 = woman/minority AA hire vs. white male AA hire p -6 = woman/minority AA hire vs. woman/minority non-AA hire p -a = woman/minority AA hire vs. white male non-AA hire

(P -Y) -(6 -a) = (P -6) -(Y -a) = (woman/minority AA hire vs, white male AA hire) -(woman/minority non-AA hire vs. white male non-AA hire) = (woman/minority AA hire vs. woman/minority non-AA hire) -(white male AA hire vs. white male non-AA hire)

estimate, which measures how much greater the shortfall is among firms using Affirmative Action, is -0.5 [-1.39 -(-0.89)]. Of course, in the estimations that follow, control variables are introduced to remove the influence of confounding factors; consequently, here we do not go through the whole set of estimated differences.

There are three potential econometric problems in estimating equation (1). First, the differentials associated with Affirmative Action may reflect unmeasured job or establishment characteristics that differ across establishments that use Affirmative Action in hiring and hire minorities or women and establishments that use Affirmative Action but do not hire minorities or women-that is, unobserved characteristics that vary within the subset of establishments using Affirmative Action in such a way as to be correlated with minority or female hiring. Without data on multiple hires from different demographic groups for the same job in the same establishment, we must simply make the assumption that unobserved job or establishment characteristics are not correlated with the Affirmative Action-demographic group interactions (e.g., D . AA) as well as with qualifications or performance, in order to identify relative differentials in qualifications or among minorities or women hired under Affirmative Action.

Second, if the decision to use Affirmative Action is partly endogenous-as discussed above-then D AA may be correlated with the error term. For example, if establishments facing smaller skill differentials between minorities and females, on the one hand, and white males, on the

Table 4 Means for Dependent Variables by Demographic Characteristics and Use of

Affirmative Action  
  Jobs FilledUsing Jobs Filled Not Using
  Affirmative Action Affirmative Action
  (1) (2)
A. Educational level:    
White male    
White female    
Black male    
Black female    
Hispanic male    
Hispanic female    
B. Less than reported education    
requirement:    
White male    
White female    
Black male    
Black female    
Hispanic male    
Hispanic female    
C. Log starting wage:    
White male    
White female    
Black male    
Black female    
Hispanic male    
His~anic female    
D. Log current wage:    
White male    
White female    
Black male    
Black female    
Hispanic male    
Hispanic female    
E. Promoted:    
White male    
White female    
Black male    
Black female    
Hispanic male    
Hispanic female    
F. Performance rating-typical    
rating for job:    
White male    
White female    
Black male    
Black female    
Hispanic male    
Hispanic female    

NOTE.-Sample sizes are given in following tables with regression/logit analyses of dependent vari- ables.

other, are most likely to embrace such hiring (since they may face lower costs from using Affirmative Action), our estimates may be biased in the direction of finding no shortfall in qualifications among women or mi- norities hired under Affirmative Action, relative to the estimates we would obtain from an exogenously imposed policy. Alternatively, if establishments that voluntarily adopt Affirmative Action tend to be those whose owners or managers are favorably disposed toward women or minorities, we might expect higher ratings of women or minorities in such establishments, leading to a positive correlation between the error term and the D .AA interactions and, hence, bias against finding that women or minorities hired under Affirmative Action perform worse on the job. However, we argued above that it appears that much of the variation in reported use of Affirmative Action in our data is induced by policy and therefore likely to be largely exogenous. Nonetheless, toward the end of the article we present some separate estimates for firms with less than 50 or less than 100 employees. As discussed above, federal regulations make both of these employment levels points at which the actual use of Affir- mative Action is likely to increase, so the use of Affirmative Action is more likely to be exogenous above these cutoffs than below. Compari- sons of estimates between samples of smaller and larger establishments should thus indicate the relative importance of any such endogeneity.

The estimates we obtain are informative even if one views our estimates as reflecting in part endogenous variation in Affirmative Action. In particular, we can still draw inferences regarding Affirmative Action as it is used in practice, which is an important input into the policy debate because we know little about differences in hiring under Affirmative Action. However, drawing conclusions regarding changes in Affirmative Action from our evidence requires that one interprets the variation in reported use of Affirmative Action as exogenous.

Third, as discussed above, there is some ambiguity in the classification of establishments using Affirmative Action. Though we believe it is small, measurement error from misclassification is likely to bias the results towards finding no effect of Affirmative Action. Since some of our results point to no effect, the results must be interpreted cautiously, pending development of superior data sources to reexamine some of the questions we consider in this article.

Educational Qualifications

The qualification we can most easily identify with workers is their education. Table 5 reports alternative simple differences and difference- in-differences estimates for years of education of the last worker hired. The estimates in part A (establishment controls only) are based on specifications that include only city and year dummy variables and es- tablishment characteristics as control variables. The estimates in column 1 measure (p -y), the differences between women or minorities and white males hired at establishments using Affirmative Action. For all five groups, the educational level of women or minority Affirmative Action

Table 5 OLS Estimates for Educational Level of Last Hire

Difference Relative to:

Same Demographic White Male, Group, White Male, Difference-in-AA Hire Non-AA Hire Non-AA Hire Differences

AA Hire

A. Establishment

controls only: White female

Black male

Black female

Hispanic male

Hispanic female

B. Occupation and job requirement controls added:

White female

Black male

Black female

Hispanic male

Hispanic female

NOTE.-AA = Affirmative Action. There are 2,381 observations. Standard errors are reported in parentheses. All specifications include dummy variables for ci~

and year and establishment controls, as well as dummy variables and interactions for Asian males an females. In art B, educational require- ments are not included. There are fewer observations than in the previous tatles because of missing data on the recent hire's actual schooling. The educational variable is coded as follows: dropout (10 years); general equivalency diploma (11 years); high school graduate (12 years); trade school or some college (13 years); associate's degree (14 vears); bachelor's degree (16 vears); some graduate school (17 years); and graduate degree (18 years). ~stimates are weighted. The 'difference-in-differences estimates in col. 4 correspond to either of the following two equivalent relative comparisons: (woman/minority AA hire vs. white male AA hire) -(woman/minority non-AA hire vs. white male non-AA hire) or (woman/minority AA hire vs. woman/minority non-AA hire) -(white male AA hire vs. white male non-AA hire). A relevant difference-in-differences estimate with respect to the estimates in col. 3 is (woman/minority AA hire vs. white male non-AA hire) -(woman/minority non-AA hire vs. white male non-AA hire). This estimate corresponds to those in col. (2).

hires is significantly lower than that of white males hired into similar establishments; the differential is less than ?hyear for white females, but about 1 ?hyears for blacks, and2 ?hyears for Hispanics. However, the estimates of (p -a)in column 3 indicate that these educational shortfalls also appear-although to a lesser extent- between women and minorities hired in establishments using Affirmative Action and white males hired in establishments not using Affirmative Action. At the same time, the esti- mates in column 2 indicate that, with the exception of Hispanic males, there are negligible differences between women and minorities in estab- lishments using Affirmative Action and in establishments not using Af- firmative Action. The combined evidence implies that the educational shortfalls in column 1 may overstate the shortfalls that can be attributed to Affirmative Action hiring, if the differences in the establishments not using Affirmative Action reflect overall differences that would occur in the absence of Affirmative Action hiring at any establishments. This is reflected in the difference-in-differences estimates in column 4, which indicate considerably smaller educational shortfalls of women and minorities attributable to Affirmative Action hiring. The differen- tials for white females and blacks are small and relatively insignificant, while only those for Hispanics (especially men) remain large and signif- icant.

In part B (occupation and job requirement controls added), we present similar estimates adding controls for occupation and job requirements, looking at differences in qualifications within more narrowly defined jobs. The most relevant policy question probably concerns the relative skills of individuals hired to do the same job. If the educational shortfalls of some groups of minorities or women are much smaller with the occupation and job controls than they are in part A, this would suggest that minorities or women hired under Affirmative Action are matched to less demanding jobs that are appropriate for their lower skill levels. On the other hand, occupation and job requirements likely reflect the char- acteristics of the worker hired as well as the job, in which case we may be overcontrolling by including them as independent variables; that is, Af- firmative Action may lead to the hiring of less-skilled women or minor- ities who are then allocated to less demanding jobs. Given these ambigu- ities regarding the choice of specification, we think it best to present both types of evidence.

In part B, the estimated education differentials between women or minority workers hired by establishments using Affirmative Action and other workers are a bit smaller for most groups, while for Hispanic males they are substantially smaller. This implies that only Hispanic males are allocated to significantly less demanding jobs when hired under Affirma- tive Action. However, the overall conclusion is the same; the simple difference estimates of (P -7) indicate educational shortfalls among women and minorities hired under Affirmative Action, while the differ- ence-in-differences estimates of (P -y) -(6 -a)reflect a shortfall only for Hispanics.

The evidence from the difference-in-differences estimation of lower qual- ifications of Hispanics hired under Affirmative Action may appear incon- sistent with the results in table 2, suggesting that the hiring of Hispanics is not boosted by Affirmative Action. However, we examined data on the percent- age of applicants at the establishment accounted for by each demographic group, as reported by the respondent. These data reveal that, in establish- ments using Affirmative Action, the percentage of applications from blacks is higher (by roughly 2 percentage points) than in establishments not using Affirmative Action, while this percentage is lower for Hispanics (by 0.8 percentage points). Given their higher overall skill needs and lower applica- tion rates from this group, establishments that use Affirmative Action may have to be less selective in hiring Hispanic applicants to hire the same proportion of Hispanics as establishments not using Affirmative Action. One reason for the lower application rate of Hispanics at establishments using Affirmative Action might be that Hispanic immigrants are more likely to work in smaller establishments that are owned or operated by coethnics, especially in Los Angeles. On the other hand, table 2 suggested that hiring of white women and black men is boosted by Affirmative Action, so the results in column 1 of table 5 indicating lower qualifications of these groups are more easily reconciled.

An alternative approach to the issue of educational qualifications is to ask whether minorities or women hired under Affirmative Action are less qualified relative to the educational requirements of the job, rather than simply relative to white males. We refer to this outcome as indicating that the employee is "underqualified" rather than "less qualified." For exam- ple, if black men are sufficiently qualified for the jobs they hold, while white males are either overqualified or allocated to more demanding jobs, then the criticism of Affirmative Action on the grounds of lower educa- tional levels of black men would be blunted. Conversely, it is possible that women or minorities hired under Affirmative Action are more under- qualified relative to the jobs that they hold.26

To examine this question, we use information on the educational requirement for the job reported by the employer, and estimate logit models for whether the individual hired had less than the reported required amount of education. There is some ambiguity in the coding of this dependent variable, since employers were asked whether a college degree was required, without specifying whether this was an associate's or bachelor's degree. We assume that they were referring to the latter."

The results are reported in table 6; the control variables are the same as in table 5. In column 1 of part A, the positive signs of the estimates

26 Note that this still could be relevant in part B of table 5, because educa- tional requirements are excluded from the set of job requirements.

"About 6% of the sample is underqualified using the data this way. If we instead assume that employers were referring to associate's degrees, about 5% of the sample is underqualified. In either case, more than half of the underqualified workers were those without high school degrees in jobs reported to require high school degrees.

Table 6

Logit Estimates for Probability That Last Hire Had Less than Reported

Education Requirement

Difference Relative to:

Same Demographic White Male, Group, White Male, Difference-in-AA Hire Non-AA Hire Non-AA Hire Differences

AA Hire (1) (2) (3) (4)

A. Establishment controls only: White female ,011 -.0002 ,023 -.011 (.014) (.012) (.014) (.020) Black male .029 .045 ,040 ,034 (.019) (.028) (.018) (.032) Black female,051 ,069 ,062 .057 (.018) (.026) (.018) (.031) His~anic male .009 ,009 ,020 -,002 (.023) (.025) (.022) (.030)

Hispanic female .031 ,045 ,042 ,034 (.021) (.029) (.021) (.033)

B. Occupation and job requirement controls added:

White female .005 -.001 .018 -.014 (.014) (.011) (.013) (.018) Black male .029 ,036 ,042 ,023 (.018) (.026) (.017) (.030) Black female ,045 ,058 ,058 ,045 (.017) (.024) (.017) (.028) Hispanic male ,018 .011 .031 -.002 (.021) (.023) (.021) (.027) Hispanic female .027 .035 .040 .022 (.020) (.026) (.019) (.030)

NOTE.-AA = Affirmative Action. There are 2,027 observations in both parts. The table reports partial derivatives of the probability of being underqualified, evaluated at the weighted means. Standard errors calculated from a linear approximation to these derivatives treating the means as fixed are reported in parentheses. All specifications include dummy variables for city and year and establishment controls, as well as dummy variables and interactions for Asians. The dependent variable is coded as one if the job requires a college degree and the hire has less than a college degree or the job requires a high school degree and the hire has less than a high school degree (including a GED). We assume that a required college degree refers to a bachelor's degree. In part B, the educational requirements are not included. The sample is smaller than in the previous table because data on the education requirement for the job are needed. Estimates are weighted.

indicate that for all groups, women or minorities hired under Affirmative Action are relatively more likely to be underqualified than white males hired into establishments using Affirmative Action. The difference-in- differences estimates are quite similar, although because of increased standard errors the only significant evidence of underqualification of women or minorities hired under Affirmative Action is for black females. In part B, where we add the occupation and job controls, the evidence is similar (although slightly weaker); because this specification looks at education relative to educational requirements, it is not surprising that controlling for other job characteristics has relatively little influence on the estimate^.^' The fact that we find no evidence of higher probabilities that Hispanics hired under Affirmative Action are more likely to be underqualified is consistent with the conclusion we drew from table 5 that such Hispanic hires are allocated to less demanding jobs. Conversely, the evidence for black females in tables 5 and 6 suggests that this allocation does not occur for them.

Job-Related Outcomes

The results in the preceding sections provide some evidence of rela- tively lower educational or skill qualifications of women or minorities hired under Affirmative Action. But there remains the question of whether the shortfalls in these two observable measures of qualifications imply inferior performance on the job. If they do not, then there is perhaps no reason to be concerned with the apparent lower qualifications of women or minorities hired under Affirmative Action.29

Consequently, in this subsection we look at three employment out- comes that should be related to actual job performance: starting wages, current wages, and promotions. Looking at this broader set of job-related outcomes is useful for another important reason-namely, educational levels and job requirements are only a subset of the many dimensions along which a worker's qualifications can be measured. Measures such as wages and promotions should, if they are related to productivity, provide more of a summary or "sufficient" statistic for a worker's qualifications. A potential objection, however, is that the same pressures that may lead to the hiring of less-qualified workers under Affirmative Action may also lead employers to pay and promote women and minorities at a rate that is more than commensurate with their productivity, in which case wages and promotions would not be useful as measures of relative qualifications or job performance.30

This is the one dependent variable for which the results differ for unweighted estimates. In particular, although the signs of the estimates are the same, the results are not statisticallv significant. This occurs because the shortfall in educa-

, "

tion relative to requirements among some minorities arises with respect to college degrees. Since, as noted earlier, the original sample strongly oversampled jobs requiring a high school degree (for other reasons related to the purposes of the survey), in the unweighted sample, jobs requiring college degrees are underrep- resented, and this shortfall is therefore harder to detect if we fail to account for the nonrepresentative sampling.

29 AS an example, Bloch (1994) discusses James v. Stockham Valves and Fittings Company, in which an employer charged with discrimination claimed that formal education requirements for manual laborers led to hiring of more whites than blacks. But the court ruled that education was unrelated to job performance.

30 Nonetheless, the wage and promotion results are still of interest because

To obtain a more independent measure of worker performance in these jobs, we use a rating of the worker's job performance (measured on a scale of 1-100) elicited from the person intervie~ed.~' The survey also asked for the supervisor's rating of the typical new hire into the job, which lets us stan- dardize across establishments and jobs by looking at the deviation of the new hire's performance rating from the usual or typical rating. If the performance ratings were the product of a formal evaluation procedure used to set wages and determine promotions, the ratings might be contaminated in the same way as data on wages and promotions (as employers might feel constrained to manipulate performance ratings to back up their wage and promotion decisions). However, these ratings are informal and not explicitly related to actual pay and promotion decisions, and survey respondents were promised full confidentiality. Therefore, the ratings seem likely to provide an unbiased measure of a worker's true job performance. Although even the standardized performance ratings are likely to be measured with error, if the measurement error is random with respect to true performance and the independent variables, it should lead to larger standard errors of the estimated coefficients but not to bias in the estimated magnitudes or sigris of these coefficients. Thus, if we find positive rather than negative effects of Affirmative Action on these ratings, it would be difficult to interpret such findings as stemming solely from measurement error.

Wages

Regressions for logs of starting and current wages are reported in parts A and B of table 7. In addition to the establishment, occupation, and job requirement controls, we also add standard human capital controls (education and age to both sets of equations, and tenure to the equations for current wages and promotions).32 As a means of assess- ing the reliability of the data, it is worth noting that wage differentials

there is virtually no empirical evidence to date on the relationship of Affirmative Action to these outcomes.

31 A similar variable is used in the EOPP Survey (e.g., Barron, Berger, and Black 1989) and a more recent, similar survey of members of the National Federation of Independent Businesses (Bishop 1993). Since the main survey respondent was the person responsible for hiring, in small- and medium-sized companies the performance rating was typically elicited from this respondent, who was likelv to be a manager or owner. and who should therefore be able to

"

speak knowledgeably about a worker's job performance. In large companies, these functions are more likely to be separated. As a result, in these cases the interviewer generally elicited the performance rating from a supervisor.

32 One might object that by controlling for differences in qualifications (edu-

cation), we bias the results against finding poorer performance, as measured by wages (or promotions). However, the results reported in table 7 turned out to be insensitive to the inclusion of these variables.

Table 7 OLS Estimates of Log Wage Regressions, and Promotion Logits

Difference Relative to:

Same Demographic

White Male, Group, White Male, Difference-in- AA Hire Non-AA Hire Non-AA Hire Differences

AA Hire (1) (2) (3) 14)

A. Log starting wage: White female Black male Black female

Hispanic male Hispanic female

B. Log current wage: White female Black male Black female

Hispanic male Hispanic female

C. Promotions: White female Black male Black female

Hispanic male Hispanic female

NOTE.-AA = Affirmative Action. There are 2,135 observations in parts A and B, and 2,438 observations inlaft C,. In parts A and B, regression coefficients and standard errors are reported. Part C reports partial erlvatlrres of the probability of the outcome, evaluated at the weighted means; standard errors calculated from a linear approximation to these derivatives treating the means as fixed are reported in parentheses. All specifications include dummy rariables for city and year and establishment, occupa- tion, and job requirement controls, and controls for age and years of education, as well as dummy rariables and interactions for Asian males and females. In addition, controls for tenure are included in the current wage and promotion specifications, and dummy variables for different race or sex of respondent or supervisor are included in the promotion specification only. Estimates are weighted.

between women or minorities and white males in establishments not using Affirmative Action, which can be calculated from the estimates in column 3 minus the estimates in column 2, indicate significantly lower wages paid to women and minorities. The estimated differentials

are -.I5 for white women, -.I6 for black men, -.25 for black women,

and -.29 for Hispanic men and women.

Turning to the three alternative comparisons for women or minority workers in establishments using Affirmative Action, column I indicates that the wage differentials between these workers and white males in similar establishments are considerably smaller than the wage differentials in establishments not using Affirmative Action, noted above.33 In fact, for black men the wage differential is erased. Because the wage differentials compared with white males in establishments not using Affirmative Ac- tion are somewhat larger, as reported in column 3, and the differences between women and minorities in the two types of establishments are small, as reported in column 2, the difference-in-differences estimates in column 4 indicate that Affirmative Action raises the relative wages of women and minorities, although most groups still earn less than compa- rable white males in establishments using Affirmative Action; the increase is substantial for black and Hispanic men, on the order of 20%.~~

The results are quite similar for starting and current wages, indicating no differences in wage growth.

Promotions

Part C of table 7 reports logit estimates of equations for whether the newly hired worker was promoted.35 The independent variables are the same as in the other parts, with the exception of the inclusion of controls for whether the race or sex of the supervisor differed from that of the worker. The results in column 1 suggest that the probability of promotion for blacks and Hispanics hired in establishments using Affirmative Action is higher than for white males in similar establishments, although the differential is statistically significant only for Hispanic females. The dif-

33 Leonard (1990) reports similar evidence indicating that relative wages of minorities to white males are higher in cities and industries with high proportions of employment in establishments subject to Affirmative Action.

34 Table 1 indicates that the percentage of the workforce covered by collective bargaining is higher in establishments using Affirmative Action. However, the relatively higher wages paid to women and minorities in such establishments do not appear to be attributable to the well-documented tendency for race and sex differentials in wages to be lower among union workers. We reestimated the models adding a set of interactions between the demographic dummy variables D and two alternative measures of unionization: the percentage covered by collective bargaining and a dummy variable indicating whether this percentage was greater than zero. The relatively higher wages paid to women or minorities in establish- ments using Affirmative Action, based on the difference-in-differences estimates, did not diminish in this augmented specification.

35 A limitation of the promotions variable is that median tenure with the employer is roughly 2-3 months in our sample. The proportion of workers promoted is only .O8.

ference-in-differences estimates in column 4 convey the same result. Finally, again subtracting the estimates in column 2 from those in column 3, we see that in establishments not using Affirmative Action, promotion probabilities are lower for four of the five groups of minorities or women. However, in contrast to wages, the estimates suggest that the negative promotion differentials for Hispanics and blacks in these establishments are reversed in establishments using Affirmative Action. This is so despite the earlier evidence that blacks and Hispanics are relatively less qualified in terms of educational attainment. These higher promotion rates for women and minorities hired under Affirmative Action may reflect per- formance that at least compensates for lower educational qualifications, although they may also reflect Affirmative Action pressures themselves. As a consequence, we next turn to the performance ratings.

Performance Ratings

Given our uncertainty over whether the wage and promotion results reflect better performance of blacks and Hispanics, or preferential treat- ment, we turn in table 8 to the performance rating regressions, in which the dependent variable is the measured rating minus the typical rating. We report results including only the establishment (and city and year) con- trols in part A, and adding in the occupation, job requirement, and other controls in part B. The estimates in column 1 indicate that black females hired under Affirmative Action obtain higher performance ratings than white males hired in similar establishments, although the differential is not significant at the 10% level. The evidence for white females and black males also indicates that their performance is not lower than that of white males in similar establishments. On the other hand, the performance ratings of Hispanic males and females at these establishments are lower, although significantly so only for Hispanic males.

The estimates in column 3 suggest that the differentials between each group of women or minorities in establishments that use Affirmative Action and white males in establishments that do not use Affirmative Action are similar to the corresponding differentials within establishments that use Affirmative Action. But the estimates in column2 also indicate that similar differentials exist between women or minorities hired at the two types of establishments. As a result, the difference-in-differences estimates in column 4 are somewhat similar to those in column I; the only changes are that the relative performance advantage of black women hired in establishments using Affirmative Action is now stronger, and the performance shortfall for His- panic women is erased. The evidence is very similar in part B, when the more extensive set of controls is added.36

36 Earlier we noted the possibility that stronger ratings given to women or minorities in establishments using Affirmative Action might reflect relatively

Table 8 OLS Estimates of (Performance Rating-Typical Rating for Job) Regressions

Difference Relative to:

Same

Demographic White Male,
White Male, Group, Non-AA Difference-in-
AA Hire Non-AA Hire Hire Differences
AA Hire      
A. Establishment controls      
only:      
White female -1.16 -.I6 .74  
  (1.50) (1.16) (1.19)  
Black male .49 2.32 .91  
  (2.24) (2.40) (2.07)  
Black female 3.10 5.18 3.52  
  (2.39) (2.53) (2.23)  
Hisuanic male -6.31 -5.93 -5.89  
  (2.33) (2.39) (2.17)  
Hispanic female -3.49 .39 -3.07  
B. Occupation, job (2.56) (2.88) (2.42)  
requirement, education,        
age, tenure, and        
different race or        
sex of respondent        
or su ervisor controls        
addef        
White female -1.94 -.59 -1.51  
  (1.55) (1.18) (1.28)  
Black male 2.91 1.23 3.34  
Black female        
  (2.62) (2.53) (2.49)  
Hispanic male -2.72 -6.08 -2.29  
  (2.66) (2.41) (2.53)  
Hispanic female -1.23 -.97 -.81  
  (2.86) (2.89) (2.75)  

-

NOTE.-AA = Affirmative Action. There are 2,130 observations. Standard errors are reported in parentheses. All specifications include dummy variables for city and year and establishment controls, as well as dummy variables and interactions for Asian males and females. Estimates are weighted.

favorable views of these workers at establishments that have voluntarily chosen to use Affirmative Action. But it is unclear why this would be true for blacks and not Hispanic males. Our controls for the racial composition of customers and for the race/sex of the respondent should also help to control for such factors. We also note that our job requirement controls include information on customer contact, where immigrant Hispanics might be at some disadvantage relative to native-born whites and blacks. Finally, below we consider evidence on the sensitivity of the estimates to endogeneity. The estimated coefficients of the dummy variables for different race or sex of the respondent or supervisor indicated that supervisors or respondents of the opposite sex resulted in performance ratings that were signif- icantly higher, by about 2 points, whereas supervisors or respondents of the opposite race resulted in ratings that were lower by about 1-2 points (significantly so for supervisors of the opposite race).

Thus, the evidence for Hispanics tells a relatively consistent story, especially for men. We generally find that Hispanic men and women hired under Affirmative Action are less qualified in terms of education, and at least the men are matched to less demanding jobs and get worse performance ratings.37 On the other hand, while we find some evidence that blacks are less qualified, we find no evidence that they receive lower performance ratings, which (like their wage and promotion rates) are relatively higher in these establishments than elsewhere. This is the case despite the fact that they are not allocated to less demanding jobs, suggesting that employers using Affirmative Action find ways to hire blacks who are qualified on grounds other than education. For white females, who appear to be the primary beneficiaries of Affirmative Action (in terms of numbers of hires), we find relatively little evidence of weaker (or stronger) qualifications or performance.38

Sources of Variation in the Use of Affirmative Action in Hiring

The results to this point describe average differences in qualifications or performance of women or minorities hired under Affirmative Action relative to other workers, based on self-reported use of Affirmative Action by the owners or managers who responded to the survey. Previ- ously, we noted some of the similarities between our self-reported mea- sure and more exogenous measures based on EEO-1 reporting or regu-

37 Although Hispanics in this sample are heavily concentrated in Los Angeles

(with 70% of the newly hired Hispanics being located there), the finding of low performance ratings for Hispanic men is not unique to Los Angeles; if anything, the negative difference-in-differences estimates for Hispanic males in Los Angeles were somewhat smaller than those in other cities.

38 When we add these performance measures to the current wage equations in table 7, we generally find that they generate positive and significant effects on wages. But the various race/sex differentials presented there are little changed by their inclusion. On one hand, this suggests that the relatively higher wages paid to women and minorities hired under Affirmative Action may reflect preferential treatment. On the other hand, these results are based on using the performance ratings as an independent variable; the presumed measurement error in these ratings inhibits our ability to ask whether performance ratings explain these higher wages. One potential problem with the standardized performance ratings is that when we look at a worker hired by an establishment using Affirmative Action, the typical rating that is subtracted off may also apply to an Affirmative Action hire. In this case, the standardization may erase any relative differences in performance between Affirmative Action and non-Affirmative Action hires, leading the estimate of the effect of Affirmative Action on performance to be zero. To examine this possibility, we reestimated the equations in table 8 using the raw performance ratings. The col. 4 estimates using the raw ratings were not generally further from zero than the corresponding estimates in table 8, suggesting that standardization does not force the estimated effects of Affirmative Action toward zero.

lations for federal contractors. However, we acknowledged that the potential endogeneity of self-reported use of Affirmative Action, and the biases this could generate, limit our ability to infer causal effects and, hence, limit the ability of our evidence to answer questions regarding the likely effects of imposing, eliminating, or otherwise changing Affirmative Action policies.

In this section, we assess the sensitivity of our conclusions to this potential endogeneity problem by estimating the effects of Affirmative Action only for a sample of relatively small establishments, among which the use of Affirmative Action is more likely to be voluntarily chosen and, therefore, more endogenous. We use cutoffs of 50 and 100 to delineate the relevant size categories; the former reflects the point at which federal contractors must have Affirmative Action plans, while the latter reflects the point at which firms must file EEO-1 forms (and thus be subject to monitoring that may induce hiring behavior comparable to that induced by explicit Affirmative Action plans). Since the cutoffs are based on employer rather than establishment size, we focus on the subset of establishments in our sample that constitute the entire firm (which we refer to as "single-unit firms").

Because the sample of single-unit firms includes a bit less than half of the establishments used so far, we adopt a more restrictive version of equation (I), in which D is a single dummy variable for being female or minority. Also, we adopt a linear probability model for the educational underqualification equation. To show that these changes are relatively inconsequential, in table 9 we first report results for the full sample of single-establishment firms using this restricted specification for the three dependent variables we consider most interesting: educational level, less- than-required education, and standardized performance rating. These estimates are qualitatively similar to those for the full sample, with the less restricted specification, that were reported in the earlier tables (as were the results for this restricted specification when we did not restrict the sample to single-establishment firms). Looking at educational level, there is some evidence that women or minorities hired at firms using Affirma- tive Action are less qualified (but more so relative to white males hired at similar firms than in the difference-in-differences estimation). The prob- ability that educational levels are less than required also seems somewhat higher for women or minorities hired under Affirmative Action, though the differences once again are not always significant. Finally, with respect to performance ratings, there is no evidence that women or minorities hired at firms using Affirmative Action perform any worse on the job; if anything, the point estimates for their ratings are higher.

These estimates are followed in parts B and C by results for the sample of single-unit firms with less than 50 and then less than 100 employees. Because Affirmative Action is more likely to be voluntary in smaller

Table 9 Estimates of Performance Rating and Educational Qualification Regressions, Alternative Sources of Variation in use of Affirmative Action, Single-Unit Firms

Difference Relative to:

WomanIMinority, AA Hire White Male, AA Hire (1) Woman/Minority, Non-AA Hire (2) White Male, Non-AA Hire (3) Difference-in Differences(4) No. of Observations(5)
A. OLS estimates of restricted model for full sample of single-unit firms: 
Educational level 
         
Less-than-required education          
Performance rating-typical rating for job          
B. OLS estimates for single-unit firms with <SO employees: Educational level          
Less-than-required education           
Performance rating-typical rating for job           
C. OLS estimates for single-unit firms with <I00 employees: Educational level          
Less-than-required education           
Performance rating-typical rating for job           

NOTF.-AA = Affirmative Action. Specifications are restricted to include a single dummy variable for white women or minorities. Asians are dropped from the sample. For the specification for underqualification in terms of education, linear probability estimates are reported. Standard errors are reported in parentheses. The control variables correspond to those in the corresponding specifications in part B of tables 5, 6, and 8. Estimates are weighted.

firms, any potential endogeneity bias should be worse, and therefore we might expect any evidence of lower qualifications or performance of Affirmative Action hires to be weaker in these subsamples if endogeneity bias is important. However, the estimates in parts B and C are qualita- tively and quantitatively similar to those in part A, suggesting that there is little bias due to endogeneity (either because endogenous choices are not correlated with the outcomes we are studying or because there are in fact few cases of voluntary adoption of Affirmative A~tion).~~

To summarize, the results in size categories in which the use of Affir- mative Action is more plausibly endogenous also indicate that women or minorities hired under Affirmative Action are somewhat less qualified in terms of education but perform as well or better on the job. Because this evidence is very similar to that based on estimates for the full sample, we are more inclined to interpret our results as informative with regard to the causal effects of Affirmative Action.

V. Conclusion

We use microlevel data on employers and employees to investigate whether minority or female employees hired under Affirmative Action are less qualified relative to other groups of workers. Our measures of qualifications include educational attainment of the workers hired (in absolute levels and relative to job requirements), skill requirements on the job, and a variety of outcome measures that presumably are linked to worker performance on the job. The analysis is based on data from a new survey of establishments in four major metropolitan areas in the United States.

On average, we find some evidence that minority employees hired under Affirmative Action have lower educational attainment and are somewhat more likely to fall short of formal educational "requirements" on these jobs when they are hired, although we find very little evidence of this for white females hired under Affirmative Action. However, when we consider measures of outcomes for workers in these jobs, we find that minorities and females hired under Affirmative Action do relatively well. On average, their wages are relatively higher, as are their probabilities of promotion. But since these outcomes might themselves be driven by Affirmative Action policies, and not just by the performance of the workers, we also consider employers' ratings of employee performance.

39 We have also estimated some equations using dummy variables for estab- lishment sizes of 50-99 and 100 or more employees as instruments for reported use of Affirmative Action. These results were qualitatively similar to those reported in the article, although the estimates were considerably less precise, and identification is tenuous because employer size may directly affect the outcomes we study.

The results show that ratings of white female or black employees in establishments using Affirmative Action are generally at least as high as those of other comparable workers. These results are reversed only for Hispanic men, who receive significantly lower performance ratings.

Taken together, the results of this article suggest that critics of Affir- mative Action may be right in pointing to some shortfalls in qualifications among women or minorities hired under Affirmative Action. However, these critics may be focusing too narrowly on one or two easily observ- able measures of qualifications that are not the only predictors of what is ultimately the most important measure-job performance. Our results suggest that most women or minorities hired under Affirmative Action make up in some way-presumably through qualifications or skills other than those measures we observe, or perhaps through effort-for the educational and skill shortfalls that we find. Thus, there may be some redistribution of employment away from white males towards minorities and females at establishments using Affirmative Action, but there does not appear to be substitution of less-able women or minority workers for more-able white male workers.

This does not necessarily imply that there are no costs from using Affirmative Action. One possibility is that these establishments hire relatively more less-skilled workers than they would in the absence of Affirmative Action, which might entail some cost in efficiency. Another is that the same number of less-skilled workers are hired as before, but that there is relatively more redistribution of employment away from less-skilled white males within these establishments. Without panel data on establishments both before and after their use of Affirmative Action, it is impossible to distinguish among these interpretations.

We should also note a number of further caveats with respect to these findings. Given the data that we have, we are only able to estimate the effects of Affirmative Action on the last worker hired in each establish- ment, which is not necessarily a representative sample of all employees hired under these procedures. We are also not able to compare different hires into comparable jobs within each establishment. The focus on recent hires also forces us to consider only short-term outcomes for a sample of employees with very low job tenure. Finally, our self-reported measure of use of Affirmative Action may not allow us to define the relevant set of establishments or activities as clearly as in those studies that use more objective measures, such as federal contractor status or EEO-1 filing, although we have argued that our measure appears to mimic more ob- jective measures quite well and could conceivably be preferable.

Despite these caveats, we believe that our data provide useful informa- tion on the effects of Affirmative Action by providing the first microlevel evidence linking Affirmative Action to worker qualifications and perfor- mance. We interpret the overall evidence as indicating that while Affir- mative Action may lead to the hiring of women and minorities with shortfalls in terms of some observable qualifications, most groups of women and minorities hired under Affirmative Action perform their jobs as well as white males.

References

Altonji, Joseph, and Pierret, Charles. "Employer Learning and the Sig- naling Value of Education." Working paper no. 5438. Cambridge, MA: National Bureau of Economic Research, 1996.

Anderson, Bernard E. "The Ebb and Flow of Enforcing Executive Order 11246." American Economic Review Papers and Proceedings 86 (May 1996): 298-301.

Badgett, M. V. Lee. "Affirmative Action in a Changing Legal and Eco- nomic Environment." Industrial Relations 34 (October 1995): 489-506.

Barron, John; Berger, Mark; and Black, Dan. "Job Matching and On-the- Job Training." Journal of Labor Economics 7 (January 1989): 1-19. Bergmann, Barbara R. "Does the Market for Women's Labor Need Fixing?" Journal of Economic Perspectives 3 (Winter 1989): 43-60.

Bishop, John. "Improving Job Matches in the U.S. Labor Market." Brooking Papers on Economic Activity: Microeconomics, no. 1 (1993): 335-90.

Bloch, Ferrel. Anti-discvimination Law and Minority Employment. Chicago: University of Chicago Press, 1994.

Bobo, Lawrence, and Smith, Ryan. "Antipoverty Policy, Affirmative Action, and Racial Attitudes." In Confronting Poverty: Prescriptions for Change, edited by Sheldon Danziger, Gary Sandefur, and Daniel Weinberg, pp. 365-95. Cambridge, MA: Harvard University Press, 1994.

Cain, Glen. "The Economic Analysis of Labor Market Discrimination: A Survey." In The Handbook of Labor Economics, vol. 1, edited by Orley

C. Ashenfelter and Richard Layard, pp. 693-785. Amsterdam: North- Holland, 1986. Carter, Stephen. Reflections of an Affirmative-Action Baby. New York: Basic, 1991.

Coate, Stephen, and Loury, Glenn. "Will Affirmative Action Policies Eliminate Negative Stereotypes?" American Economic Review 83 (De- cember 1993): 1220-40.

Cutler, David, and Glaeser, Edward. "Are Ghettoes Good or Bad?" Quarterly Journal of Economics 112 (August 1997): 827-72.

Epstein, Richard A. Forbidden Grounds: The Case against Employment Discrimination Laws. Cambridge, MA: Harvard University Press, 1992.

Fix, Michael, and Struyk, Raymond. Clear and Convincing Evidence. Washington D.C.: Urban Institute Press, 1994. Glazer, Nathan. Affirmative Discrimination: Ethnic Inequality and Public Policy. New York: Basic, 1975. Gold, Michael Evan. "An Introduction to the Law of Employment

Discrimination." In ILR Bulletin, vol. 68. Ithaca, NY: Industrial Labor Relations (ILR) Press, 1993.

Holzer, Harry J. "Employer Skill Needs and Labor Market Outcomes across Groups." Working paper. New York: Russell Sage Foundation, 1995.

. What Employers Want: Job Prospects for Less-Educated Workers.

New York: Russell Sage Foundation, 1996. . "Employer Hiring Decisions and Antidiscrimination Policy." In

Generating Jobs: How To Increase Demand for Less-Skilled Workers,

edited by Richard Freeman and Peter Gottschalk, pp. 223-57. New York: Russell Sage Foundation, 1998.

Leonard, Jonathan S. "Anti-discrimination or Reverse Discrimination: The Impact of Changing Demographics, Title VII and Affirmative Action on Productivity." Journal of Human Resources 19 (Spring 1984): 145-74.

. "Women and Affirmative Action." Journal of Economic Perspec- tives 3 (Winter 1989): 61-76.

. "The Impact of Affirmative Action Regulation and Equal Op- portunity Law on Black Employment." Journal of Economic Perspec- tives 4 (Fall 1990): 47-64.

Lipset, Seymour Martin, and Schneider, William. "The Bakke Case: How Would It Be Decided at the Bar of Public Opinion?" Public Opinion (March-April 1978): 38 -48.

Mare, Robert. "Changes in Educational Attainment and School Enroll- ment." In State of the Union, vol. 1, edited by Reynolds Farley, pp. 155-214. New York: Russell Sage Foundation, 1995.

Neal, Derek, and Johnson, William. "The Role of Premarket Factors in Black-White Wage Differences." Journal of Political Economy 104 (October 1996): 869-95.

Neumark, David. "Sex Discrimination in Restaurant Hiring: An Audit Study." Quarterly Journal of Economics 11 1 (August 1996): 915-42.

Oettinger, Gerald. "Statistical Discrimination and the Early Career Evo- lution of the Black-White Wage Gap." Journal of Labor Economics 14 (January 1996): 52-78.

Rodgers, William M., and Spriggs, William E. "The Effect of Federal Contractor Status on Racial Differences in Establishment-Level Em- ployment Shares, 1979-1992."American Economic Review 86 (May 1996): 290-93.

Smith, James, and Welch, Finis. "Affirmative Action and Labor Markets." Journal of Labor Economics 2 (April 1984): 269-301.

Sowell, Thomas. Preferential Policies: An International Perspective. New York: W. Morrow, 1990.

U.S. Department of Labor. "High Performance Work Systems and Es- tablishment Performance." Photocopied. Washington, DC: U.S. De- partment of Labor, 1993.

Yinger, John. Closed Doors, Opportunities Lost. New York: Russell Sage Foundation, 1995.

Comments
  • Recommend Us