THE EFFECT OF WAGE DISCRIMINATION ON PRIVILEGED GROUPS

A Dissertation

Submitted to the Graduate School

of the University of Notre Dame

in Partial Fulfillment of the Requirements

for the Degree of

Doctor of Philosophy

by

Daniel Matthew Custance Lawson, D.E.F., B.A., M.A.

Teresa Ghilarducci, Director

Graduate Program in Economics

Notre Dame, Indiana

July 2005
 UMI Number: 3265640

UMI Microform 3265640

Copyright 2007 by ProQuest Information and Learning Company.
All rights reserved. This microform edition is protected against
unauthorized copying under Title 17, United States Code.

ProQuest Information and Learning Company

300 North Zeeb Road
P.O. Box 1346
Ann Arbor, MI 48106-1346

c Copyright by

Daniel Matthew Custance Lawson

2005

All Rights Reserved

 THE EFFECT OF WAGE DISCRIMINATION ON PRIVILEGED GROUPS

Abstract

by

Daniel Matthew Custance Lawson

Wage discrimination reduces the wages of groups subject to discrimination. Its

effect on the wages of groups not subject to discrimination is less obvious. A powerful

effect on one group in a work relationship could likely affect the other members of

the relationship. This study explores the effect of wage discrimination on the wages

of managers and coworkers not targeted for discrimination. Existing theories of

discrimination predict disparate effects on the wages of the managers and coworkers

of workers targeted for discrimination.

This study develops models in which managers’ tastes cause discrimination. The

models predict that as managers’ taste for discrimination rises, the wage gap rises,

and managers’ pay falls. Discrimination hurts profits, but managers are willing to

pay for it, making up for the losses in order to be allowed to hire fewer workers from

groups they dislike.

I test this theory against other theories using data about sex and race discrimi-

nation. Using 2000 Census data to measure the wage discrimination against female

non-managerial workers by Oaxaca decomposition, I find that for every one percent

discriminatory decrease in the wages of female non-managerial workers, the wages

of managers fall by 0.531%. This is consistent with the model of manager discrimi-

nation. I also find that a one percent discriminatory decrease in female wages leads
 Daniel Matthew Custance Lawson

to a 0.227% decrease in the wages of their male coworkers in the same geographic

area. This is not consistent with the basic model, although it could be explained

by extensions to the model. In the context of the race wage gap, I find the oppo-

site effect for managers: when the unexplained wage gap increases by 1%, manager

wages rise by 0.196% (significant at the 99% confidence level).

These results, taken together, suggest that while a model of manager taste

discrimination may explain some of wage discrimination based on sex, race wage

discrimination needs another explanation, perhaps statistical discrimination. The

findings on race wage discrimination are inconsistent with a model that rogue man-

agers produce the wage gap by gratifying their own tastes in opposition to corporate

interests.
To Lisa and Madeleine

ii
 CONTENTS

FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

CHAPTER 1: MODELS OF WAGE DISCRIMINATION AND ITS IMPACT

ON THE WAGES OF NON-TARGETED GROUPS . . . . . . . . . . . . 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Manager Discrimination . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Taste-based Discrimination . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.1 Owner Taste . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.2 Coworker Taste . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.3 Consumer Taste . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Divide and Conquer . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 Statistical Discrimination . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.6 Comparing the Theories . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6.1 Statistical Discrimination . . . . . . . . . . . . . . . . . . . . 22
1.6.2 Consumer Taste . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.6.3 Divide and Conquer . . . . . . . . . . . . . . . . . . . . . . . 24
1.6.4 Coworker Taste . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.6.5 Owner Taste . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

CHAPTER 2: EVERYONE LOSES: HOW WAGE DISCRIMINATION AGAINST

FEMALE NON-MANAGERIAL WORKERS AFFECTS THE WAGES
OF MANAGERS AND MALE COWORKERS . . . . . . . . . . . . . . . 27
2.1 Who Pays for Sex Discrimination? . . . . . . . . . . . . . . . . . . . . 27
2.2 Discriminating Managers as Dollar Store Merchandise . . . . . . . . . 29
2.3 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.1 Testing the Effect of Discrimination . . . . . . . . . . . . . . . 34
2.3.2 Measuring Discrimination . . . . . . . . . . . . . . . . . . . . 35
2.3.3 Inferring Discrimination from the Unexplained Wage Gap . . . 39
2.3.4 Factors Other than Discrimination Entering the “D” Variable 41
2.4 Data and Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . 43

iii
 2.4.1 Census of Population and Housing, 2000 . . . . . . . . . . . . 43
2.4.2 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.5 Variables in the Wage Regression . . . . . . . . . . . . . . . . . . . . 46
2.6 Results and Interpretation . . . . . . . . . . . . . . . . . . . . . . . . 47
2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

CHAPTER 3: A PRELIMINARY FORMAL MODEL OF DISCRIMINAT-

ING MANAGERS AS DOLLAR STORE MERCHANDISE . . . . . . . . 57
3.1 Labor Supply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2 Labor Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.3 Equlibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4 On Profit-neutrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.5 Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.5.1 Effect on Managers’ Pay . . . . . . . . . . . . . . . . . . . . . 66
3.5.2 Effect on Non-managerial Pay Gap . . . . . . . . . . . . . . . 67
3.5.3 Effect on Type-ω Workers’ Pay . . . . . . . . . . . . . . . . . 67
3.5.4 Effect on Type-o Workers’ Pay . . . . . . . . . . . . . . . . . . 68
3.6 Extended Implications . . . . . . . . . . . . . . . . . . . . . . . . . . 69

CHAPTER 4: A DIFFERENT STORY: HOW WAGE DISCRIMINATION

AGAINST BLACK NON-MANAGERIAL WORKERS AFFECTS THE
WAGES OF MANAGERS AND WHITE COWORKERS . . . . . . . . . . 71
4.1 Do Managers Pay for Race Discrimination? . . . . . . . . . . . . . . . 71
4.2 Testing the Specific Model . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2.1 Predictions about White Non-managers’ Wages . . . . . . . . 72
4.2.2 Predictions about Managers’ Wages . . . . . . . . . . . . . . . 73
4.3 A Generalized Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3.1 White Non-managers . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.2 Managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.4 Data and Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4.1 Census of Population and Housing, 2000 . . . . . . . . . . . . 78
4.4.2 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 Variables in the Wage Regressions . . . . . . . . . . . . . . . . . . . . 80
4.5.1 Controlling for Non-discriminatory Wage Differences . . . . . 80
4.5.2 Specific Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.6.1 Specific Test on Wages of White Non-managers . . . . . . . . 82
4.6.2 Specific Test on Wages of Managers . . . . . . . . . . . . . . . 87
4.6.3 Generic Test on Wages of Managers . . . . . . . . . . . . . . . 91
4.7 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

EPILOGUE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

APPENDIX A: TAYLOR APPROXIMATION . . . . . . . . . . . . . . . . . 100

iv
APPENDIX B: VARIABLES IN THE SEX DISCRIMINATION WAGE RE-
GRESSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

WORKS CITED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

v
 FIGURES

1.1 Unexplained Male—Female and Black—Non-black Wage Gap Among

Full-time Workers, Current Population Survey, United States, 1979-
1998 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2.1 Unexplained Sex Wage Gap in Each POWPUMA in the Continental

U.S., 2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.1 The Unexplained Race Wage Gap by POWPUMA in the Continental

U.S., 2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

vi
 TABLES

1.1 PREDICTED EFFECTS OF DISCRIMINATION MODELS ON GROUPS

NOT TARGETED BY DISCRIMINATION . . . . . . . . . . . . . . 23

2.1 DETERMINANTS OF LOG WAGES OF FULL-TIME WORKERS,

UNITED STATES, 2000 . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.2 MODELS CONSISTENT WITH DISCRIMINATION REDUCING
THE WAGES OF ALL NON-MANAGERIAL WORKERS AND THEIR
PREDICTIONS FOR THE EFFECT OF INCREASED DISCRIMI-
NATION ON MANAGERS’ WAGES . . . . . . . . . . . . . . . . . . 54

4.1 STRUCTURALLY SUGGESTED DETERMINANTS OF LOG WAGES

OF FULL-TIME WHITE NON-MANAGERS, UNITED STATES, 2000 84
4.2 STRUCTURALLY SUGGESTED DETERMINANTS OF LOG WAGES
OF FULL-TIME MANAGERS, UNITED STATES, 2000 . . . . . . . 87
4.3 REDUCED FORM DETERMINANTS OF LOG WAGES OF FULL-
TIME MANAGERS, UNITED STATES, 2000 . . . . . . . . . . . . . 92

B.1 MEAN VALUES OF CHARACTERISTICS, MALE NON-MANAGERIAL

WORKERS, UNITED STATES, 2000 . . . . . . . . . . . . . . . . . . 102
B.2 MEAN VALUES OF CHARACTERISTICS, MANAGERS, UNITED
STATES, 2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
B.3 PAIR-WISE CORRELATION BETWEEN CHARACTERISTICS OF
NON-MANAGERIAL MALE WORKERS AND THEIR WAGES AND
THE UNEXPLAINED SEX WAGE GAP IN THEIR LOCALITY,
UNITED STATES, 2000 . . . . . . . . . . . . . . . . . . . . . . . . . 110
B.4 PAIR-WISE CORRELATION BETWEEN CHARACTERISTICS OF
MANAGERS AND THEIR WAGES AND THE UNEXPLAINED
SEX WAGE GAP IN THEIR LOCALITY, UNITED STATES, 2000 . 114

vii
 ACKNOWLEDGMENTS

Financial support for this dissertation came from a University of Notre Dame

Graduate School Presidential Fellowship. Office space was provided by the Higgins

Center for Labor Research. The computer facilities for algebraic manipulation,

data analysis, writing, and typesetting were funded by a generous gift from Hilda

N. Burns.

Assistance and feedback on the content of this dissertation came from Katharine

Abraham, David Betson, Justin Dubas, Teresa Ghilarducci, Lisa Workman Gloege,

David Hachen, Sharon Hermes, Alicia Knoedler, David Lawson, Lisa Lawson, BJ

Lee, Bruce Meyer, Derek Neal, Kevin Neuman, James Rakowski, Jim Sullivan, Jen-

nifer Warlick, Lance Wescher, Marty Wolfson, and three anonymous grant reviewers

at the National Science Foundation. Their advice and contributions have been in-

valuable. All remaining errors are my own.

Powerful and well-written software contributed to this study; without the tools

provided by the authors of Mac OS X Panther, Mathematica 5.0, Perl 5.8.1, Stata

8.2, TEXShop 1.34, and TextWrangler 2.0, this study would have taken far longer

to complete.

viii
 CHAPTER 1

MODELS OF WAGE DISCRIMINATION AND ITS IMPACT ON THE WAGES

OF NON-TARGETED GROUPS

1.1 Introduction

Wage discrimination reduces the wages of groups subject to discrimination. Its

effect on the wages of other groups is less obvious. This study explores the effect of

wage discrimination on the wages of privileged groups — those not subject to wage

discrimination. This study specifically examines the effect of discrimination against

black workers on the wages of white workers and on the bosses of these workers, and

the effect of discrimination against female workers on the wages of male workers and

the bosses of these workers. Besides the group discriminated against, who is helped

or harmed by discrimination?

Forty years ago, after hearing testimony from people knowledgeable about eco-

nomic theory and considering other evidence, Congress enacted the Equal Pay Act

of 1963, which, among other things, required that employers not discriminate on the

basis or race and sex in paying equally trained and experienced employees for equal

work. The economic theory of the day explained how wage discrimination could

exist, but predicted that in a competitive market, it would and indeed could not

persist over time (Becker 1957). The Equal Pay Act and many non-governmental

anti-discrimination efforts were designed to impede the mechanisms of wage discrim-

ination explained by existing economic theory. Forty years later, wage discrimina-

1
tion is in many cases reduced, but by no means gone. Economic theory has offered

several alternative theoretical explanations for persistent discrimination, many of

which would suggest different policies to combat them. Just as the economic theory

of owner taste-based discrimination available in 1963 informed the Equal Pay Act,

contemporary economic theory must inform current efforts to combat discrimina-

tion.

Wage discrimination on the basis of race and sex appears to be stubbornly per-

sistent. If, after Altonji and Blank (1999), one accepts the gap between the mean log

wages of two groups that is not explained by differences in the groups’ mean levels of

education and potential experience (i.e. a Oaxaca decomposition of the wage gap) as

at least a crude measurement of the wage discrimination against one of the groups1 ,

then the time trend over the past twenty years indicates that wage discrimination is

still a pressing problem. While the unexplained wage gap among full-time workers

due to sex has declined, it has distinctly not gone to zero, and the unexplained

wage gap among full-time workers due to race has actually increased. To the extent

that these unexplained wage gaps provide a crude measurement of discrimination,

it would seem that further investigation of persistent wage discrimination would be

a worthwhile project.

Figure 1.1 on page 3 is calculated by taking a Oaxaca decomposition of the

gap in mean log hourly wage between males and females and then between blacks

and non-blacks controlling for education and potential experience for workers who

worked over 35 hours the previous week from the monthly outgoing rotation groups

of the Current Population Survey from January 1979 to December 19982 .

1
But see, for instance, Neal and Johnson (1996) for why this might not be the case. Workers
select whether to enter the workforce based on both the wage they face and factors that determine
their reservation wage. If the reservation wage varies systematically between two groups, the wage
gap observed between those who choose to work could misrepresent the wage gap actually faced
by the average potential worker.
2
Analysis software code is available from the author upon request.

2
 0.45

0.4

0.35
Unexplained difference in mean log wages

0.3

0.25

0.2

0.15

0.1

0.05

0
Jan-79

Jan-80

Jan-81

Jan-82

Jan-83

Jan-84

Jan-85

Jan-86

Jan-87

Jan-88

Jan-89

Jan-90

Jan-91

Jan-92

Jan-93

Jan-94

Jan-95

Jan-96

Jan-97

Jan-98

Unexplained Wage Gap by Sex Unexplained Wage Gap by Race

Figure 1.1. Unexplained Male—Female and Black—Non-black Wage Gap Among

Full-time Workers, Current Population Survey, United States, 1979-1998

3
 If wage discrimination both exists and persists, its effects become much more

important than if it were but a transitory phenomenon, quickly and asymptotically

approaching zero. One question that has not been thoroughly explored in the litera-

ture to date is the question of what relationship exists between wage discrimination

and the wages of the group not discriminated against. This basic empirical ques-

tion will be of importance for two distinct reasons: as an empirical phenomenon in

its own right and as a means of distinguishing between some competing economic

models of wage discrimination.

On one level, the relationship between discrimination against one group and the

wages of the group not discriminated against is a question of who might benefit and

who might suffer from the removal of the discrimination, and to what extent. On a

completely positive level, these questions deserve accurate answers. Public policy to

combat discrimination should be informed by the effects such policy might have on

the wages of those not subject to discrimination. It is reasonable to expect that a

powerful effect on one group in a work relationship would affect the other members

of the relationship.

Members of groups not discriminated against have historically played many roles

in the struggle for pay equity. In some cases, these actions may arise from a sense

of altruism, justice, or solidarity. Theoretical models exist, however, predicting that

promoting pay equity would not just be the honorable and just course of action, but

also the economically profitable one for the group not discriminated against. Insti-

tutional evidence of union strategies may suggest that these models may motivate

much modern union support for racial equality. In other theoretical models, such

actors’ wages would decline, suggestion that their actions were altruistic, gaining

them psychic utility benefits that offset the wages they would forego by eliminating

the discrimination that worked to their advantage.

4
 A relationship between overall wage level for the baseline group and the effect

of wage discrimination on members of groups facing such discrimination could help

describe the nature of development paths that might be characterized by relatively

more or less discrimination. This could be true even if this relationship is completely

bi-directionally non-causal. In other words, even if the relationship between overall

wage level and the level of discrimination offset indicates nothing about one causing

the other, if economic and social structures that lead to high wages tend to be

accompany either greater or less discrimination, this is noteworthy in and of itself.

If discrimination does not cause low (or high) wages, and low wages do not

cause discrimination, but both are caused by a particular pattern of industrial or-

ganization, or the industrial make-up of an area, or something else, the relationship

between discrimination and wages could be fundamentally interesting in looking

at patterns that relate to the development of higher-wage jobs even though the

regression does not identify a direction of causality.

The second, and perhaps more critical reason that examining the relationship

between wage discrimination and the wages of groups not discriminated against

is important is that its answer can help shed empirical light on the long-standing

question of why wage discrimination persists. Since Gary Becker first broached the

topic of the economics of discrimination at the very beginning of his project to

understand human behavior in terms of economic analysis, many economists have

explained both the existence and the persistence of a wage gap between men and

women and between blacks and whites. Multiple theories of discrimination explain

wage gaps between different groups in the face of the prima facie presumption that

market forces ought to set one’s wage equal to the marginal product of one’s labor.

5
1.2 Manager Discrimination

This study develops a new model of wage discrimination in which managers’

decisions cause hiring discrimination. Managers working for firm owners often act

as their agents in hiring, firing, and wage determination decisions. Many economic

models that explain wage discrimination or its absence assume that a firm’s poten-

tially discriminatory decisions either maximize owner utility (Becker 1957) or profits

(Arrow 1973; Aigner and Cain 1977; Phelps 1972; Reich 1984; Blau, Ferber, and

Winkler 1992, p. 224). The model proposed in this study introduces principal/agent

conflict so that managers make hiring decisions to maximize their own utility rather

than the firm’s profits.

Perhaps unsurprisingly3 , firms facing discrimination lawsuits often claim that

discrimination happens not because of company policy but because of the discrimi-

natory actions of individual managers. In such scenarios, firms claim to be against

discrimination and to have policies to that effect, while individual managers violate

such policies, bringing about discriminatory results that are not the fault of the

firm.

In January 2001, computer software company Microsoft faced a $5 million racial

discrimination lawsuit, claiming systemic deficiencies in the promotion and pay of

black employees relative to white employees. Microsoft’s head of human relations,

Deborah Willingham, replied that “Microsoft has a zero-tolerance policy towards

discrimination in the workplace,” suggesting that discrimination was not due to

company policy, but rogue managers acting contrary to company policy (New York

Times 2001).
3
Demonstrating that discrimination took place despite efforts of the company to prevent it is
not a defense against charges of discrimination, but can have drastic effects in limiting damages
in such cases. See Burlington Industries, Inc. v. Ellerth, Faragher v. Boca Raton, and Kolstad v.
American Dental Ass’n.

6
 In March 2001, a class of 2,600 current and former black employees sued food

and facilities-management company Sodexho, alleging discrimination in promotion.

Sodexho spokesperson Leslie Aun and Sodexho court filings called the cases of dis-

crimintion “isolated”, with Aun claiming,“When you have a company that is the

size of 100,000 people, do you have people who act inappropriate? Yes.. . . We have

the same issues as a town of 100,000. But is that reflective of the culture or the kind

of company you’re trying to build? No. We’ve been very aggressive about taking

a stand on these issues.” (Shin 2004) Representatives of the company claim that

discrimination is perpetuated not by company policy but by individual managers

acting contrary to company interests.

A September 2003 federal anti-discrimination lawsuit accused managers at New

York’s Plaza hotel of discriminating against Muslim staff members after the Septem-

ber 11, 2001 terrorist attack. Plaza spokesperson Amanita Duga-Carroll acknowl-

edged some discriminatory behavior on the part of managers, but claimed that it was

contrary to company policy, and noted that the company terminated or suspended

the offending management staff (Padgett 2003).

A sex-discrimination class action lawsuit against aerospace corporation Boeing

went to trial in May 2004 claimed that Boeing benefitted from paying women less

than men. Boeing’s lead counsel Barbara Berish Brown countered, claiming that

Boeing’s policies specifically prohibit salary discrimination on the base of sex, explic-

itly blaming the claims of discrimination on “a few rogue managers.” (Bowermaster

2004)

Jewelry store chain Friedman’s Inc. faced a class-action suit in 2003 claiming

hiring and promotion discrimination against blacks. Freidman’s Inc attorney Denis

Shanagher identified the chain as “an equal opportunity employer [that] does not

tolerate employment discrimination of any kind.” (Jewlers’ Circular-Keystone 2003)

7
Shanagher noted that outrageously discriminatory action by one single manager was

found, and the company terminated that manager’s employment as a result.

If one takes these claims at face value that firms find discrimination undesirable

(because of either natural inefficiency, reputation damage, or legal costs resulting

from discrimination), but some managers desire to discriminate, and thus persist

in this behavior even in the face of lower expected wages (either probabilistically

because of the risk of firing or deterministically as in the pay schemes linking pay to

diverse hiring), an economic model should be able to reflect this dynamic. Such a

model would come to a conclusion directly opposite the conclusion of a “statistical

discrimination” model, in which discrimination is efficient for the firm. Evidence

that managers are penalized for discriminating is not consistent with profit maxi-

mization if discrimination contributes to profit maximization. Arrow (1998) noted

that a model in which discriminating managers are paid less had never been devel-

oped or tested, but suggested that such effects were unlikely to actually exist.

Several firms have explicitly adopted policies linking managerial pay to hiring

records, suggesting that the theory that managers take a pay cut for discriminating

is not as far-fetched as previously thought, and deserves further study. Some firms

have responded to discrimination lawsuits with policies that explicitly link manage-

rial pay with the diversity of managers’ hiring records. According to Wall Street

Journal reporter Mehta (1996), firms implementing such pay policies include Mar-

riott International Inc., Deloitte & Touche, Tenneco, and Hoechst Celanese Corp.

In June 2004, the United States’ largest employer, Wal-Mart, announced that it

would link the bonuses of its top executives to their ability to meet diversity goals.

This move came in response to lawsuits alleging wide-spread gender discrimination

in promotion through the Wal-Mart chain. Wal-Mart CEO Lee Scott announced

up to 7.5% cuts in bonuses for all top executives, including himself, if a smaller

8
proportion of women and minorities were promoted than applied for management

positions. (Foster 2004)

Firms explicitly linking pay to a manager’s record of diversity in hiring and

promotion are rare but the group includes some large and influential firms. Repre-

sentatives of the firms interviewed in the article, however, claimed that their firms

believed that diversity in hiring benefited the firms’ bottom lines. If diversity in hir-

ing does indeed benefit profitability, as in the model developed in this study, then

firms explicitly linking diversity in hiring to managerial pay capture only the tip of

the iceberg: many more firms would implicitly link diversity in hiring to managerial

pay. If diversity in hiring is good for profitability either inherently (in the absence of

statistical benefits from discrimination) or because of legal costs of discriminating,

any firm that links managerial pay with performance would penalize managers for

discriminating.

Note that finding support for the claim that discrimination is driven by man-

agers’ preference to hire males does not necessarily imply that discrimination is solely

the fault of rogue managers. In the profit-neutral version of the manager discrimina-

tion model, firms pay managers according to their marginal revenue product for the

owners, giving the firm no incentive to promote discrimination. However, a model

in which owners cultivate a preference for discrimination among their managers so

they can pay them less could also be consistent with these findings, particularly

if managers are hired within internal labor markets. To test for the distinction

between a profit-neutral manager discrimination model and a profit-advantageous

manager discrimination model (and to examine the owners’ incentive for hiring dis-

criminating managers), one would need data on localized profit rates, which I do

not have.

Before developing new models, however, existing models ought to be surveyed to

9
show that they do not explain the same results. The theoretical story of discrimina-

tion and wage gaps has not lacked for attention in any of the decades since Becker

first tackled the issue in 1957 (Becker 1957; Phelps 1972; Arrow 1973; Aigner and

Cain 1977; Reich 1984; Black 1995; Lundberg and Startz 1998). For a summary,

see Altonji and Blank (1999). These multiple theoretical models of discrimination,

however, posit different causes for the wage gap, and thus would imply different pol-

icy solutions. Determining the extent of wage discrimination itself has proven to be

a difficult empirical challenge, although many empiricists have risen to the challenge

(Cain 1986; Blau and Beller 1988; Groshen 1991; Blau and Beller 1992; Blau and

Kahn 1992; Neal and Johnson 1996; Blau and Kahn 1997). A more daunting chal-

lenge has been to determine to what extent wage discrimination is due to different

causes. Economists have several well-developed stories of the sources of persistent

wage differentials not based directly on productivity factors, but determining which

story to tell about which group at which time remains a difficult question.

1.3 Taste-based Discrimination

Formal economic modeling of discrimination began with Gary Becker. In his

introduction to The Economics of Discrimination (1957), Becker claims that with

regard to discrimination, “other social scientists, by their early entrance into this

field may have established it as their property; the economist here, as elsewhere,

has respected the property rights of others.” (Becker 1957, p. 2) Property rights of

other disciplines notwithstanding, Becker “attempted to remedy this neglect” of an

economic theory of discrimination.

In this model, a preference for discrimination (or prejudice, in Becker’s terms)

takes the form of disutility from contact with members of a particular group. Prej-

udice refers to the case where an individual experiences disutility from associating

10
with certain groups of people. Prejudice becomes measurable as discrimination

when that individual pays money or foregoes income in order to associate and not

associate with the people of his or her choice. (Becker 1957, p. 6). The amount

an individual pays for the “good” of association with desired groups measures the

extent to which that individual discriminates. This foregone income reveals that the

discriminating party values the lack of contact with people of the disliked group at

least as much as the amount of money foregone. Becker refers to this as a “shadow

price,” an indication of value not obviously displayed in a market, but revealed

through observed decisions.

For a utility-maximizing employer, then, the hiring decision enters the utility

function through two paths. First, hiring affects the profitability of the firm, which

affects the income available to the firm owner. This path alone leads to the normal

profit maximization story — firms maximize profits because their owners maximize

income in order to maximize utility. Becker, however, introduces an element previ-

ously unconsidered in economic models. There are more inputs to the firm owner’s

utility function than simply profits. The firm owner maximizes utility with respect

not only to profits but also with respect to contact with groups of people the owner

likes or dislikes.

Becker introduces the notion of a discrimination coefficient (DC) to link the

money costs of hiring a worker, buying a product, working at a job, etc., with the

net utility costs of each of these actions. Becker then uses this as a springboard

to introduce the notions of employer discrimination, employee discrimination, and

consumer discrimination.

11
1.3.1 Owner Taste

If an employer faced a wage of π for a particular type of labor i, but the employer

incurred disutility from hiring that type of worker, the employer would behave as if

the wage paid to that worker were π(1 + di ) for workers belonging to group i. The

employer’s discrimination coefficient for type-i workers would thus be di .

Suppose there are two factors of production that are perfect substitutes in the

production process but one factor Lo produces disutility di for the employer, while

the other factor Lω does not. If the market wages for both were equal, there would

be less demand for Lo than for Lω . This would lead to a shortage of Lω and a

surplus of Lo . The theoretical Walrasian auctioneer, however, would notice that the

market does not clear, and adjust prices accordingly. Thus, wages for Lω would

go up, increasing supply and decreasing demand, and wages for Lo would go down,

decreasing supply and increasing demand, until the markets clear. This is commonly

referred to as a crowding model — Lo workers are crowded into a smaller number

of firms that are willing to hire labor of this type.

By this mechanism, employer distaste for Lo drives down the price of Lo until

equilibrium is reached. The employers with the largest discrimination coefficients

would pay higher factor prices, while employers with lower or zero DCs would pay

lower factor prices. In competitive product markets, firms facing higher costs would

be driven out of the industry in the long-run adjustment process. This would reduce

demand for Lω and increase demand for Lo , leading to a new short-run equilibrium

wage regime in which the wage for Lo is close to the wage for Lω than it was in the

previous regime. This process would continue until only firms with a DC of zero

remain in the industry, and Lω equals Lo .

Becker thus concluded that in competitive product markets, owner taste dis-

crimination would be competed away in the long run as discriminating employers,

12
with their higher costs, would be driven out of business. Stiglitz (1973) noted that

even in the presence of competitive product markets, imperfections in factor mar-

kets could allow owner taste discrimination to persist. Minimum wage legislation,

efficiency wages, or other factors that yield non-clearing factor markets can produce

job queues. In the presence of queues, employers could discriminate in whom they

select from the queues without long-run cost pressure.

1.3.2 Coworker Taste

Becker explored the possibility that the distaste for working with members of

another race could come not from the owners of the firm but from the workers.

Becker’s model of employee taste discrimination predicted that in the presence of

this form of discrimination, employers would find it least costly to employ completely

segregated workforces. Stiglitz (1973) argues that this model developed by Becker

(1957) and expanded by Arrow (1972a, 1972b) and Krueger (1963) explains only

segregation and not discrimination in the long run as long as there is the equivalent

of “free trade” between the firms employing the two groups.

Welch (1967) observes that worker distaste for members of other groups need

not lead to a segregated workforce if workers of different groups are complements for

one another rather than substitutes. If workers from two groups are complements,

the increase in marginal productivity for a worker (and, presumably, wage) caused

by working with members of the complementary group could more than offset the

disutility of working with said group. Welch does conclude, however, that “The

incentive to integrate is derived solely from the possibility of obtaining factor com-

binations different from those which exist under segregation” (Welch 1967 p. 232) —

in other words, workers from the two groups must be fundamentally different from

one another in a way that is useful to the production process. Integration takes

13
place only when there are significant differences between two groups that comple-

ment each other in production. If workers are basically the same but dislike one

another, segregation is more efficient.

A manager taste model is basically a subclass of coworker taste discrimination.

One element these models omit are the legal or public relations costs of maintaining

a segregated workforce. The manager discrimination model developed in Chapter 3

will add this element, which helps explain how worker taste could lead to discrimi-

nation and not mere segregation with long-run tendencies toward wage equalization

in the absence of complementary fundamental differences between groups.

1.3.3 Consumer Taste

Becker notes the potential for discrimination based not on the tastes of firm

owners or employees, but rather on the tastes of consumers. Welch (1967) con-

cluded that if some consumers derive no disutility from associating with members of

the disfavored group, these consumers could act as intermediaries for “impersonal”

products, arbitraging away the price difference for goods and services that do not

require contact between the producer and the end consumer. Thus, consumer taste

discrimination should only mater in the case of more direct services.

Hammermesh (Hammermesh and Biddle 1994; Donald and Hammermesh 2004)

suggests that in the presence of customer preferences for goods or services produced

by one group of workers over another where such customers are willing to pay a

premium for goods or services from non-minority workers over minority workers,

the minority status of a worker becomes an actual factor in that worker’s marginal

revenue product for the firm. In this case, it is impossible to separate discrimina-

tion from productivity differences by looking at the inputs because the same input

(minority status) is the source of both potential discrimination by managers or

14
coworkers and at the same time differences in productivity (because customers may

be willing to pay more for goods or services from non-minority workers). If these

theorized effects exist, decomposition cannot separate “productive differences” from

“discriminatory differences” in the pay gap. Hammermesh and Biddle (1994) find

evidence that lawyers that consumers find attractive earn more than ones consumers

find ugly. This means that lawyers whose physical appearance is more desirable to

clients are by definition more productive — clients are willing to pay more for their

services, and clients’ willingness to pay determines the worth of the lawyers’ time.

Race and sex are both linked to physical appearance. It is a small leap to go from

physical appearance as a productivity factor to race or gender as a productivity

factor.

1.4 Divide and Conquer

Becker (1957 p. 13ff) makes reference to what he describes as the “falacious”

notion that capitalists benefit from discrimination and non-minority workers suffer.

Becker, of course, develops a model in which this is not the case, despite what Becker

acknowledges as extensive literature4 to the contrary, much of it outside economics.

The simple act, however, of developing a model in which something does not hap-

pen does not prove that the phenomenon does not occur. Similarly, Becker’s 1957

assertion that capitalists are harmed by discrimination and non-minority workers

benefit from discrimination requires empirical testing. Models can and indeed do ex-

ist in which capitalists benefit from discrimination and workers (even non-minority

workers) are hurt by the same. Reich (1981) finds that where the median black

household income is a larger percentage of the median white household income, in-
4
Becker cites Saenger (1953), Allport (1955), Rose (1951), Cox (1948), Dollard (1937),
McWilliams (1948), and Aptheker (1946) as claiming that capitalists benefit from discrimination
at the expense of all workers.

15
equality among whites goes down, or the greater the inequality between whites and

blacks, the greater the inequality between rich whites and poor whites.

The Becker model, and indeed, many neoclassical models, requires that wage

determination (and, by extension, the buying and selling of labor) takes place in

perfectly competitive markets, with all the attendant prerequisites (full employ-

ment, perfect information, uniform goods, etc.). In the absence of these prerequi-

sites, many institutionalist economists offer the counterproposal that in the context

of wage determination, relative power matters. Even within a purely neoclassical

framework, imperfectly competitive markets behave differently from perfectly com-

petitive markets. If the firms in an industry can effectively collude to fix prices,

profits will go up at the cost of consumer surplus. Similarly, if workers can effec-

tively collude to fix labor prices, wages could go up at the cost of profits. This

implies that profit share could be related to the ability of workers to effectively form

a cohesive group.

Reich’s model (1981 p. 204) sets the bargaining power of workers as a function

of the relative wages and quantities of black and white workers. Reich notes that the

labor market sets a wage floor below which a firm cannot hire workers, but, at least

in the short run, a firm could certainly choose to pay more to workers. In a twist on

Stiglitz’s efficiency wage theory, if a firm selects one group of its workers to receive

a wage premium in exchange for their not joining with another group in collective

bargaining, the firm could potentially buy a lower total wage bill by dividing the

pool of workers that potentially might bargain as a unit. White capitalists thus earn

more in the presence of discrimination, while white workers earn less. Discrimination

thus increases white income inequality.

From this model, Reich concludes that “racial inequality can be stable in a

competitive economy even in the absence of collusion among employers.” The model

16
is consistent with Reich’s finding that capitalists benefit from discrimination, while

white workers receive reduced wages. This model relies on a unionized labor force

being the alternative scenario; if employers do not fear the threat of unionization,

the mechanism outlined in Reich becomes unlikely.

1.5 Statistical Discrimination

Taking the opposite approach to Becker, theories of statistical discrimination

argue that discrimination persists because it improves a firm’s profitability. In this

model, non-discriminating firms would be driven out of a competitive market be-

cause discrimination is not the indulgence of a firm-owner’s tastes but rather the use

of valuable information. These models, developed by Arrow (1972a, 1972b), Stiglitz

(1973), Phelps (1972), Spence (1973) and Aigner and Cain (1977) are united by

the theme that there is a profit motive for discriminating. In the absence of dis-

crimination, then, workers in the group not discriminated against would receive

lower wages and managers not engaging in the best practice of discriminating would

receive lower wages, in direct opposition to the predictions of a manager taste dis-

crimination model. In some ways, a model of this class is the most direct alternative

hypothesis to the manager taste discrimination hypothesis.

Statistical discrimination models grew out of the imperfection and asymmetric

information literature of the 1970’s. Spence (1973), for instance, explicitly treats

the hiring decision as another instance of investing in the presence of uncertainty.

With transition costs associated with worker change and imperfect knowledge of

worker ability, Spence compares hiring a worker to purchasing a lottery. “Signals,”

or costly messages that individuals can choose to invest in, are the focus of Spence’s

analysis, but he also acknowledges the existence of “indices,” which are unalterable

observable characteristics that convey information. Spence classifies race and sex as

17
unalterable indices rather than investable signals.

Stiglitz (1973) extends this framework to explicitly address the impact of indices

in wage determination. If some unobservable productivity factor is correlated with

an index, members of a group with a higher mean productivity will receive a higher

return to the same increase in productivity than members of a group with a lower

mean productivity. Thus, if productivity is something individuals can invest in,

members of the group with the higher mean productivity have a larger incentive to

invest in additional productivity, making the difference in means grow even greater.

Stiglitz further notes that firms that do not discriminate (i.e. that do not use the

information conveyed by indices) will take in lower profits. Information is valuable,

and firms that disregard information conveyed by race or sex earn less than firms

that use all available information. This model is consistent with both individual

utility maximization (without needing to invoke hedonic utility functions at all)

and firm profit maximization.

Aigner and Cain (1977) critique earlier models of statistical discrimination by

noting that if the effect of the imputation of average group characteristic information

to members of that group is linear, as it is in Phelps (1972), then while there are

some winners and some losers in each group, the average wage of workers of each

group will equal the average marginal productivity of each group: thus, no economic

discrimination. Aigner and Cain note that if both the average level of productivity

for a group and the variance matters then there could be systemic differences in

wages, but they find little empirical evidence that such a model explains much if

any actual discrimination in modern labor markets.

In any of these models, information is both costly and valuable. If indices (i.e.

workers’ ascriptive characteristics such as race and sex) convey information, and

information has value, any employer not using such information would be incurring

18
extra costs. In such a case, competitive product markets would drive out non-

discriminating employers. Statistical discrimination models assume no distaste on

the part of the employer for hiring members of any group. Indeed, hedonic utility

functions do not enter the analysis. Similarly, because firms discriminate in the

context of profit maximization, long run competition does not drive out discrimina-

tion5 .

1.6 Comparing the Theories

In 1953, Milton Friedman claimed that “The ultimate goal of a positive science

is the development of a ‘theory’ or ‘hypothesis’ that yields valid and meaningful (i.e.

not truistic) predictions about phenomena not yet observed” (Friedman 1953, p. 5).

He further claimed that “The choice among alternative hypotheses equally consistent

with the available evidence must to some extent be arbitrary...” (Friedman 1953, p.

10).

This Friedman-esque approach to economic modeling seems to have character-

ized most existing tests of discrimination theories. Theories have been proposed

and non-trivial testable predictions have been tested. Friedman’s widely-accepted

methodology claims that choices among the theories, however, to a large extent have

been arbitrary, or at least not explicitly tested.

Some efforts have taken place attempting to pit theories against one another,

and to explain why this has been done as little as it has. Levitt (2003) notes that

distinguishing between competing theories of economic wage discrimination poses

a persistent empirical challenge, and suggests that economists may have to turn to

unusual data sources in order to gather the revealed preference information neces-
5
This assumes that employers are not incorrect about the information conveyed by the index.
Being “not incorrect” means that the assumptions about the index are either objectively correct
or self-perpetuating.

19
sary to help determine whether taste-discrimination or information-discrimination

is prevalent in a given situation. Levitt extensively analyzes data from a television

game show (The Weakest Link ), in which the structure of the game gives contes-

tants an incentive in early rounds to vote against players who do poorly (or whom

they expect to will poorly), and in later rounds to vote against players who do well

(or whom they expect will do well). Because Levitt’s data also includes measure-

ments of the players’ observed “productivity,” he shows how a pattern of more votes

against a particular demographic group than the number that would be would be

consistent with their performance in both early rounds and later rounds would be

indicative of taste-based discrimination. A pattern of voting against a particular

demographic group in early rounds that reverses itself in later rounds, on the other

hand, would be suggestive of information-based discrimination.

In Levitt’s analysis, he found information-based discrimination against Hispanics

and taste-based discrimination against older players; no discrimination was found

in his data set against women or Blacks. Despite the statistical significance of his

findings, he concludes that

Given the highly stylized nature of the interactions on this televi-

sion show, one must use extraordinary caution in trying to draw general
conclusions from these results. Indeed, one could imagine that the ab-
sence of observed discrimination towards Blacks in this artificial context
might arise precisely because of the presence of real-world discrimina-
tion towards Blacks which has sensitized Americans to the importance
of not appearing outwardly racist, regardless of inward beliefs. At some
conscious or unconscious level, contestants may shy away from targeting
Blacks on a nationally televised program. In contrast, players may be
less concerned about appearing to target Hispanics and the elderly. Ide-
ally, one would like to isolate real-world settings in which the strategic
incentives flip as they do on Weakest Link to provide a more readily
generalizable test of competing theories of discrimination (Levitt 2003,
p. 22).

In other words, Levitt’s proposes that identification of competing theories of

20
discrimination ideally requires a data set representative of real-world market be-

havior (rather than stylized game show interactions) that uses changing incentive

structures to reveal the extent and nature of people’s discriminatory tendencies.

Even this, however, does not seem to identify the truly interesting question.

The real world discriminatory tendencies of people in general are not especially

important; the way that discrimination is manifest in actual market interactions

is of much more economic relevance. Thus, while Levitt proposes a strong test of

measuring individual tendencies to discriminate, even if the stylized interactions

of the game show were representative of real-world settings and the contestants

were nationally representative, it would be a non-trivial leap to go from Levitt’s

conclusions to any relevant policy questions about the ways in which discrimination

affects hiring, wages, promotion, and other labor market outcomes. Levitt’s ideal

data could show how representative individuals might discriminate if they had they

had the opportunity. Everyone, however, does not have equal opportunities to

discriminate, and thus Levitt’s ideal data would not explain the discrimination that

does take place, nor who benefits in an economic exchange.

The story of wage gaps between pairs of groups (black-white, male-female, etc)

has been well documented; Altonji and Blank (1999) summarize this quite thor-

oughly. Little previous empirical work explicitly examines the relationship between

discriminatory wage gaps and the pay of workers not discriminated against; one of

the rare examples comes from Reich (1984). Reich examines the relationship be-

tween white income inequality and the wage equality between black and white work-

ers. Using the ratio between median black household and median white household

income as a proxy for discrimination, he finds when the ratio is higher, inequality

among whites goes down, or the greater the inequality between whites and blacks,

the greater the inequality between rich whites and poor whites. Little corresponding

21
work has been published on the issue of the gender pay gap; this project will fill

that gap in the literature.

This study will thus examine the effect of the unexplained wage gap between

men and women on the wages of non-managerial males and on the wages of their

managers. It will also examine the effect of the unexplained wage gap between

blacks and whites on the wages of non-managerial whites and on the wages of their

managers.

Chapter 2 develops a simple model of manager taste discrimination, and tests its

reduced form predictions about the effect of discrimination against female workers on

the wages of managers. Chapter 3 develops the manager taste model in more detail,

and formulates several structural predictions. Chapter 4 tests these hypothesis in

the context of the effect of discrimination against black non-managers on the the

wages of managers. The alternative hypotheses come from the models reviewed in

this chapter.

Directly comparing the predictions of these models on the wages of managers

and white non-managers requires some extrapolation; most of the models do not

introduce managers directly as a distinct entity to be considered. Nonetheless,

some rough predictions can be drawn under a few interpretative scenarios.

1.6.1 Statistical Discrimination

In a statistical discrimination model, group membership conveys information

about the productivity of current or potential employees. Firms not employing all

available information are discarding a valuable resource. Profits should consequently

be lower at such firms. Similarly, managers’ wages should be lower because the

marginal revenue product of a manager is lower at such a firm.

If some firms engage in statistical discrimination and others do not, the wages of

22
 TABLE 1.1

PREDICTED EFFECTS OF DISCRIMINATION MODELS ON GROUPS NOT

TARGETED BY DISCRIMINATION

Effect on Effect on Effect on

owners’ managers’ non-targeted
Model profits wages workers’ wages
Statistical Discrimination ↑ ↑ ↑
Consumer taste ↑ ↑ ↑
Divide and conquer ↑ ↑ ↓
“Advantageous” manager taste ↑ ↓ ↑
Coworker taste — — ↓
“Neutral” manager taste — ↓ ↑
Owner taste ↓ — ↑

individuals from groups with higher expected productivities would be higher at firms

who compute their expected productivity conditional on their group membership

than at firms that do not. If the unexplained wage gap is higher where more firms

assign wages based on expected productivity conditional on group membership, then

where the unexplained wage gap is higher, wages should be higher for members of

groups with high conditional expected productivities.

1.6.2 Consumer Taste

In the consumer taste model, consumers have a preference for services from one

group of workers over another. Profits should be higher at firms that pay wages ac-

cording to marginal productivity. If this model holds, firms that pay workers from

different groups equally even though their effect on revenue is unequal pay their fac-

tors inefficiently. If an increased inter-group pay gap reflects an increased response

to consumer preferences, it follows that firms that pay factors more efficiently will

receive greater profits than firms that pay inefficiently.

23
 Because managers are the ones who implement pay structures, managers who

implement efficient pay structures have a higher marginal revenue product than

managers who implement inefficient pay structures. Thus, managers should earn

more in the presence of a inter-group higher pay gap.

If firms in some areas respond to consumer tastes with their pay structure more

on average than firms in other areas do, workers of the groups with a higher marginal

revenue product (i.e. those favored by consumers) would earn more at firms that

recognize differences in consumer tastes. It follows, then, that as the inter-group

gap increases, the pay for workers of favored groups should also increase.

The consumer taste model could be considered a special case of the statistical

discrimination model. While the two models grew out of different traditions, con-

sumer taste is the most intractable statistical discrimination model. The common

argument about statistical discrimination is that the solution to statistical discrim-

ination is better or cheaper information so that employers do not need to rely on

noisy but cheap sources of information such as race and sex. In consumer taste

models, an employee’s race or sex conveys information about productivity, as in

statistical discrimination models, but it is not a noisy signal. Race or sex itself is

the source of the productivity difference because consumers are willing to pay a

premium for services from someone of a particular race or sex.

1.6.3 Divide and Conquer

In the divide and conquer model, firm owners pit two groups of workers against

each other to enhance profitability. Where the practice is more prevalent, profits

should increase. This profit motive drives the practice; if profits did not increase,

firms would not invest in dividing their workforce.

The effect on managers’ wages is less explicit in the model, but two aspects

24
of Reich (1981) suggest it should be positively signed. First, Reich’s empirical

findings show that controlling for other effects, the income Gini coefficient among

whites increases as the black-white wage gap increases. If this were caused entirely

by increased profits and not by increased manager pay (including pay for upper

management), this increase in the income Gini coefficient would be accompanied

by no corresponding increase in the earnings Gini coefficient. This is not the case,

however. Secondly, if it takes skillful manipulation of the workforce to accomplish

the increased profits resulting from “dividing and conquering” the workers, and

managers are the ones to carry out this practice, it follows that managers who can

carry out this practice would have value to the firm, and be rewarded for it. While

Reich does explicitly discuss the effect of his model of discrimination on the wages

of managers, both empirically and theoretically it appears that the Reich model

predicts that managers should earn more in the presence of discrimination.

Reich’s primary prediction is that white workers earn less in the presence of

discrimination than they would have earned in the absence of discrimination, even

though they earn more than black workers. This is a primary distinguishing feature

of “divide and conquer” models — by dividing the workforce into privileged and

unprivileged groups, all wages are kept lower than they would be in the event that

the workforce were a single collective bargaining unit.

1.6.4 Coworker Taste

Profits and managers’ wages should be unaffected by the level of segregation

workers choose to buy, as long as wages are set correctly. If hiring a more segregated

workforce imposes costs on the firm, the workers who desire this workforce bear

that cost. Managers and owners are unwilling to bear the costs of discrimination; if

workers will not absorb the costs, discrimination will not take place in this model.

25
Thus, the model predicts no link between the wage gap and profits or managers’

wages.

At least in the short run, this model predicts that the wages of members of

majority groups should be lower in the presence of a higher wage gap. If legal, insti-

tutional, or cost structural constraints preclude completely segregated workforces,

the wage gap and its effect on majority wages should persist. Workers from major-

ity groups, however, elect to earn less in exchange for the utility they receive from

working in a more segregated workplace.

1.6.5 Owner Taste

In the owner taste model, owners are willing to forego profits in order to “buy”

a workforce comprised of members of a group they like. In areas where owners are

more willing to do this, profits will be lower. This is a key feature of this model.

Managers’ wages should not be affected by the wage gap. In both the case of dis-

criminatory hiring and non-discriminatory hiring, managers are simply responding

to owners’ whims. Nothing in the model suggests that either tactic calls for more

skill or finesse in implementation; all managers are required to act in accordance

with the desires of firm owners. Thus, does not predict any link between the wage

gap among workers and the wages of managers.

As demonstrated above, through a crowding model, demand for workers from

favored groups increases and demand for workers from non-favored groups decreases.

This model does not call for any secondary mechanisms that would cancel out this

effect. This means that when the wage gap increases, wages of workers from favored

groups should increase.

26
 CHAPTER 2

EVERYONE LOSES: HOW WAGE DISCRIMINATION AGAINST FEMALE

NON-MANAGERIAL WORKERS AFFECTS THE WAGES OF MANAGERS

AND MALE COWORKERS

2.1 Who Pays for Sex Discrimination?

Managers working for firm owners often act as their agents in hiring, firing, and

wage determination decisions. Many economic models of wage discrimination as-

sume that a firm’s potentially discriminatory decisions either maximize owner utility

(Becker 1957) or profits (Arrow 1973; Aigner and Cain 1977; Phelps 1972; Reich

1984; Blau, Ferber, and Winkler 1992, p. 226; Blau, Ferber, and Winkler 1992, p.

224). This study shows that a model in which hiring decisions maximize the man-

agers’ utility rather than the firm’s profits predicts a different effect of discrimination

on managers’ wages than many models that ignore this principal/agent distinction.

This study finds that under a reasonable set of assumptions, for every percent

that discrimination reduces female non-managerial workers’ pay below its level in the

hypothetical absence of discrimination, the wages of male non-managerial workers in

that locale fall by at least 0.227%, and the wages earned by managers in that locale

go down by at least 0.531%. Discrimination lowers women’s pay most directly;

managers and male coworkers also suffer lower pay from the presence of a male-

female wage gap in their locality. These empirical effects are consistent with the

discrimination model based on the tastes of both coworkers and managers developed

27
here.

Becker (1957) argued that discrimination is not stable in the long run because of

a zero-profit condition1 ; an employer who indulges a taste for discrimination incurs

additional costs that non-discriminating employers do not incur. Basic dynamics in

competitive markets will drive out any firm with higher than normal costs.

If discrimination persists, either something must protect discriminating firms

from long-run competition, or discrimination must be efficient for the firm. Indus-

trial organization models that predict discrimination in the absence of competitive

product markets include Hellerstein, Neumark and Troske (1997) and Black and

Strahan (1999). Statistical discrimination models (Arrow 1973; Aigner and Cain

1977; Phelps 1972) claim that discriminating firms take advantage of all available

information, and thus are efficient. Strategic segmentation models (Reich 1984) ar-

gue that firms can lower payroll costs across the board by maintaining two groups

of workers at odds with one another.

Others argue that although discrimination is inefficient, it can persist in compet-

itive product markets despite that inefficiency because someone other than the firm

bears the cost of discrimination. Customer discrimination models (Blau, Ferber,

and Winkler 1992, p. 226) explain the ongoing “inefficiency” by customers who are
1
While economists such as Gunnar Myrdal had written about wage discrimination earlier
(Myrdal 1944), the first widely-noted formal model of wage discrimination (and its apparent viola-
tion of the so-called “law of one price”) came from University of Chicago sociologist and economist
Gary Becker (Becker 1957). Becker’s basic model of employer discrimination accounts for dis-
crimination in the short run, but predicts that in competitive product markets, in the long run
discrimination should disappear due to market forces alone. (Becker 1957; Blau and Kahn 2000).
Prior to Becker’s formulation of his theory, competitive product markets existed; firms had been
formed and had gone out of business, and still, discrimination was, as Kenneth Arrow quotes
Samuel Johnson, “too evident for detection and too gross for aggravation” (Arrow 1998). In other
words, when Becker developed his model predicting that competitive markets eliminate wage dis-
crimination, facts indicated that competitive markets had not eliminated wage discrimination even
at firms that had survived multiple long-run market adjustments. Thus, while Becker’s basic model
opened the door to economic modeling of wage discrimination, it did not explain persistent wage
discrimination in competitive markets. Welch (1967 p. 227) concurs with the assessment that “[i]n
view of the rather long history of discrimination, it seems to me that explanations which are stable
only in the short run are unsatisfactory.”

28
willing to pay more for services rendered by people who are not in a group they

dislike. Coworker discrimination models (Blau, Ferber, and Winkler 1992, p. 224)

project that some workers would forego a portion of wages in order to not have to

work with workers in a group they dislike.

Manager discrimination extends the coworker discrimination model to consider

managers in a firm as agents of the employer who make hiring decisions but seek

to maximize their own utility, not firm profitability. This principal/agent situation

allows for “bosses” to perpetuate discrimination and yet not hurt the firm’s bottom

line, thus negating the push of market forces that would have otherwise driven

discriminating firms out of business.

2.2 Discriminating Managers as Dollar Store Merchandise

In the manager discrimination model, the assumptions of Becker’s original model

hold with one exception. Some people who run a company experience psychic disu-

tility from hiring female employees. If enough people making hiring decisions prefer

to hire males to females, ceteris paribus this reduces demand for female employees,

causing the going wage for female employees to fall below that of male employees,

as in the Becker model.

In this model, as in Becker’s, managers who discriminate (either fully by not

hiring females at all, or partially by hiring fewer females) are less efficient than those

who do not. The difference lies in long-run adjustments. Firm owners reviewing

these managers find they are not maximizing profits: the agent does not represent

the principal well. To actually represent the interest of the principal well (by hiring

in a non-discriminatory manner) would incur psychic disutility for the agent. If firms

compensate managers according to their marginal revenue product, a discriminating

manager would earn less money than a non-discriminating manager because the

29
discriminating manager does not hire efficiently. This model predicts that where

levels of discrimination are higher, manager pay is lower. Discriminating managers

are like low-quality manufacturing equipment or dollar store merchandise: they do

not perform as well as higher-quality managers, but they also cost less.

A manager discrimination model would suggest that managers under-hire or un-

derpay female workers because of their own distaste for female workers, and not

because of pressure from the owners. If managers make hiring and firing decisions

and set pay, then by definition, managers will carry out wage discrimination. The

mere fact that managers are the ones carrying out the discrimination does not dif-

ferentiate a model of “manager discrimination” from Becker’s 1957 model of “owner

discrimination.” If owners do not want to employ women because doing so would

be a source of disutility, and the firm’s managers do not hire women to please the

owners, the agents (managers) simply accommodate the principals (owners), and the

original Becker (1957) model applies. If the owners of a firm want to strategically

divide male and female workers to segmenting the labor force, as in racial divides in

Reich (1984), a manager who systematically underpays or under-promotes women in

this context would simply act as a good agent, and traditional segmentation models

would apply.

This distinction between managers discriminating on behalf of owners and man-

agers discriminating because of their own tastes is not trivial. On June 22, 2004,

Federal Judge Martin J. Jenkins of the United States District Court in San Francisco

certified 1.6 million female current and former Wal-Mart employees for a class-action

discrimination lawsuit against Wal-Mart. The suit claims that Wal-Mart systemati-

cally paid and promoted women less than men. Wal-Mart appealed the class-action

certification, claiming that discrimination was not due to systemic corporate pol-

icy (pressure from the principals), but rather the tastes of the individual managers

30
(Greenhouse and Hays 2004). Even if managers’ tastes drive discrimination, owners

could strategically choose to hire discriminating managers in an attempt to capture

rents.

Note that in a manager discrimination model, managers do not “lose” in an

absolute sense; they lose wages. While their wages are lower than those of non-

discriminating managers, discriminating managers make up in psychic utility gained

from being able to discriminate at least what they lose in wages; if not, they would

act like non-discriminating managers. Arrow (1998) noted that this model had never

been developed or tested, but suggested that such effects were unlikely to actually

exist.

Assume that firms in an industry face two perfectly substitutable factors of

production, male and female labor, and that the production function of any firm is

the same:

Y = F (L) (2.1)

where in Equation (2.1), Y is the output of the firm and L is the input of generic

factor of production Labor, composed as follows:

L = Lm + L f (2.2)

where Lm represents the quantity of male labor and Lf represents the quantity of

female labor.

A profit-maximizing firm would maximize π where

π = Py F (Lm + Lf ) − Lm Wm − Lf Wf (2.3)

where in Equation (2.3), Py represents the price of the good produced, a constant

in a perfectly competitive product market and a function of Y in a imperfectly

competitive market.

31
 Lf
Let r = L
.

π ∗ = max π = max Py F (L) − L(1 − r)Wm − LrWf (2.4)

L,r L,r

If an individual firm maximizes profits and Lf is perfectly substitutable for Lm ,

if Wf < Wm , the profit maximizing firm would hire only female labor (i.e. r = 1)
∗)
at the quantity L∗ where Py dFdL
(L
= Wf .

If managers rather than owners make the hiring decisions, managers do not nec-

essarily maximize π, but rather their own utility, Ub , where r, the fraction of female

employees, and Wb , the manager’s wages both affect managers’ utility. Assume a

separable form of the utility function:

Ub = G(Wb ) − D(r) (2.5)

If a manager makes inefficient decisions, the firm loses Λ, the additional amount

the firm would have earned had the manager maximized profits. Let Λ ≡ π − π ∗

where π is the actual level of profits earned under the manager, and π ∗ is the level

of profits that would have been earned had the manager maximized profits. Thus,

Λ = Λ(r). Let the manager’s pay Wb be a function of lost profits.

Wb = Wb (Λ) (2.6)

Thus, the utility-maximizing condition for a manager is

∂
∂
G Wb (Λ(r)) = D(r) (2.7)
∂r ∂r

which decomposes to
∂G ∂Wb ∂Λ ∂
= D(r) (2.8)
∂Wb ∂Λ ∂r ∂r
∂G
In decomposed form, ∂Wb
represents the marginal utility of money to a manager.
∂Wb
The term ∂Λ
represents how firms penalize managers for lost profits. In the profit-

32
 ∂Wb
neutral case2 , ∂Λ
= 1. If managers had market power and their leadership wanted
∂Wb
to reward discriminatory managers, ∂Λ
could be less than one. If firms have bar-
∂Wb
gaining or information advantages over managers, it could be the case that ∂Λ
> 1,

meaning that firms over-penalize discriminating managers, extracting more profits

∂Λ
from such operations. The term ∂r
can be solved algebraically, and represents the

loss in profits caused by changing r from its optimal value of r∗ = 1. Finally, ∂

∂r
D(r)

represents the marginal disutility to a manager of increasing r.

Firms would thus have depressed demand for female labor. This model alone
∂
would predict that as ∂r
D(r) increases (i.e. the disutility of a higher fraction of

female labor in the workforce increases, or managers have stronger tastes for not
∂Wb
hiring female labor), r will fall, and as long as ∂Λ
> 0 (i.e. firms do not reward

managers for losing money and punish them for increasing the firm’s profitability),

managers would earn less money. The wages for female labor, Wf would be lower

than their non-discriminatory value, and the wages for male labor, Wm would be

higher than their non-discriminatory value.

It would be likely that the tastes of managers are not entirely disconnected from

the tastes of male non-managers in the same geographic area. This model could

combine with a standard employee discrimination model(Becker 1957, p. 46) to

predict that both managers and non-manager male workers would lose wages due

to discrimination. In the simplest form of such a combination, assume an economy

with multiple industries, some of which were governed by the manager discrimination

model and others governed by the employee discrimination model; more complex
2
In the “profit-neutral” version of this model, firms pay managers according to their marginal
revenue product for the owners. A divide-and-conquer model in which owners cultivate a preference
for discrimination among their managers so they can pay them less could also exist, particularly in
a model with limited job transition options among managers. To test for the distinction between a
“profit-neutral” manager discrimination model and a “profit-advantageous” manager discrimina-
tion model (and to examine the owners’ incentive for hiring discriminating managers), one would
need data on localized profit rates, which I do not have.

33
(and realistic) combinations could also produce the same results.

2.3 Empirical Strategy

2.3.1 Testing the Effect of Discrimination

This section lays out an econometric strategy to measure the relationship be-

tween wage discrimination against one group and the wages of other groups. This

study uses individual workers in the non-target groups as the unit of analysis . With

this approach, the measured wage of each worker is the dependent variable, while

individual levels of various forms of human capital, other determinants of pay, and

the level of discrimination in the labor market containing that individual are the

independent variables. Ordinary least-squares (OLS) regression could determine

the relationship between the level of wage discrimination against a target group

in a given labor market and the wages of non-target group members in that labor

market. Using this approach, the standard errors need adjustment because of the

market-by-market clustering of our independent variable of interest. This would off-

set some of the benefit to standard error gained by the large sample size, although

measurements of the control variables would still improve in accuracy.

Wi = α + βHi + γOi + θMj + δDj + i,j (2.9)

In Equation (2.9), Wi is the wage of male worker i, who works in labor market j,

Hi is worker i’s level of human capital, Oi is a measurement of other determinants of

the wages of worker i, Mj are other determinants of the wages of all workers in labor

market j, Dj is the level of wage discrimination against women in labor market j,

i is the residual term for individual i, and uj is the residual term for labor market

j.

As in Rogers (1993), this study uses ordinary least squares regressions to find

34
the effect of the unexplained wage gap on the wages of male non-managerial work-

ers and managers, but adjusts the variances according to a modified Huber-White

technique to account for the clustering in the error term. Repeating the study using

a maximum-likelihood estimation of an error composition modelchanges the mag-

nitude of the coefficients slightly, but the sign and significance of the variables of

interest remain the same, as do the implications of the results.

2.3.2 Measuring Discrimination

As Kenneth Arrow notes (1998), because of the passage of legislation against

employment discrimination, the direct evidence of discrimination that was available

in 1950 by simply noting racial restrictions in “job wanted” classified advertisements

is no longer available3 . Many empiricists have risen to the challenge of determining

the extent of wage discrimination (Cain 1986; Blau and Beller 1988; Groshen 1991;

Blau and Beller 1992; Blau and Kahn 1992; Neal and Johnson 1996; Blau and Kahn

1997, to list but a few).

Most modern attempts to measure the level of wage discrimination employ a

method known as the Oaxaca decomposition, after its developer, econometrician

Ronald Oaxaca (Oaxaca and Ransom 1999; Altonji and Blank 1999). This study

will employ the Oaxaca decomposition in every geographic labor market (“Place of

Work Public Use Microdata Area,” or POWPUMA as defined by the United States

Census Bureau) to measure the size of the pay gap between males and females that

productivity proxies do not explain.

3
While economists can measure unemployment or property rental costs or education simply
by asking a survey question, there is great incentive for discriminating employers to not respond
honestly to a survey question asking them how much they discriminate, particularly on a sur-
vey administered by the federal government. While straightforward methods might have been
usable in a time when neither legal consequence nor social stigma were attached to the practice
of discrimination, now researchers must find signs that discrimination has taken place, given that
employers have an incentive to claim that practices other than discrimination cause apparently
discriminatory outcomes.

35
 In a Oaxaca decomposition of the gap between the mean log wages of two groups,

one performs two wage regressions: one for each group. One can decompose the wage

gap, the difference in mean log wages between the two groups, into the “explained”

portion and the “unexplained” portion.

The explained portion (i.e. the difference not due to discrimination) for each

observable characteristic is the product of the difference between the second group’s

mean value and the first group’s mean value multiplied by the first group’s coefficient

for that characteristic. The sum of this value for all characteristics is the portion

of the wage gap that differences between the two groups’ observable characteristics

explain. The unexplained portion of the gap equals the total gap minus the portion

of the gap explained by differences in observable characteristics.

GU = (W̄M − W̄F ) − (X̄M − X̄F )BM (2.10)

It is an arbitrary normalization which group to select as the baseline group.

Barsky et al. (2001) propose a non-parametric model that does not need this ar-

bitrary normalization. While such a model would be ideal here, data constraints

(obtaining a large enough sample to perform a nonparametric decomposition of

the wage gap for each labor market) preclude the use of the more data-intensive

nonparametric model in this study4 .

In this study, this calculation will take place with males as the first group and

females as the second group. The unexplained gap serves as a proxy for the wage

gaps due to sex discrimination. For each locale, a Oaxaca decomposition split
4
Juhn, Murphy, and Pierce (1991) suggest that the measurements of discrimination would
ordinarily be close, but not necessarily equal depending on which normalization one selects. In
this case, however, there exists a reason to prefer one normalization over the other. If one group is
significantly larger than the other, using the coefficients from the larger group to multiply by the
smaller group’s mean in the decomposition (i.e. group one is the larger group) would result in an
estimation of the difference with a smaller standard error.

36
by sex will determine the unexplained sex wage gap (a proxy for the wage gap

due to discrimination by sex). Using the general method of Oaxaca decomposition

described above, the wage regression for each locale and sex for the male/female

decomposition will be the following:

ln Wi =α + β1 Ai + β2 A2i + β3 A3i + β4 A4i +

β5 E1i + β6 E2i + β7 E3i + β8 E4i + β9 E5i + β10 Bi +

β11 S1i + β12 S2i + i (2.11)

In Equation (2.11), Ai is the age of individual i, E1i through E5i are dummy

variables for education as defined in the Appendix, Bi is a dummy variable for black

workers, and S1i and S2i are dummy variables for extractive5 and manufacturing6

industries, respectively, with the baseline being the service sector.

The variable Ai (or potential experience) serves as a proxy for human capi-

tal due to actual experience (and on-the-job training), and the various education

dummy variables attempt to capture human capital due to education. This will

over represent discrimination to the extent that there are productivity characteris-

tics correlated with race or gender that the human capital proxies (such as quality

of school system) do not capture. Measurements of educational attainment rather

than years of schooling (or dummies for various sheepskin effects) have been demon-

strated in the context of race to be helpful in measuring human capital (Neal and

Johnson 1996). If members of groups facing discriminatory wage offsets tend to sys-

tematically acquire experience-based human capital differently with age than groups

facing no such offsets in a manner not due to discrimination, this would make age
5
Census Industry codes between 0 and 56, inclusive
6
Census Industry codes between 77 and 406, inclusive

37
a particularly poor proxy in that case, and would lead to overrepresentation of dis-

crimination. This might be the case for sex, but it might not; due to lower labor

force attachment, women often have lower rates of acquiring firm-specific human

capital than men do. One still acquires experience, however, in time spent out-

side the formal-sector labor force. The degree to which the human capital acquired

outside the formal sector is transferable to subsequent formal-sector employment is

subject to empirical investigation, but it is far from obvious that lower rates of labor

force attachment should automatically mean that women acquire experience-based

human capital slower than men do.

This specification of the Oaxaca decomposition will under represent discrimina-

tion because discrimination causes selection bias (wages are only observed for those

actually in the workforce) and reduced levels of education for minority groups. Nev-

ertheless, this study shall interpret the decomposition of the sex wage gap to proxy

for the wage gap due to sex discrimination in locale j; I discuss further interpreta-

tion of this proxy in Section 2.3.4. The method also leaves out other forms of com-

pensation which could bias results. Non-wage compensation behaves dramatically

differently for women and men (Ghilarducci 1985; Ghilarducci and Lee 2004), sug-

gesting that wage might not accurately proxy total hourly compensation in studies

exploring the wage gap between women and men. It would thus be quite desir-

able to expand this study to consider total hourly compensation as an alternative

dependent variable to wage, but available data do not allow this.

Even if the unexplained wage gap were entirely due to discrimination and not

at all due to unmeasured productivity differences, the unexplained wage gap and

the level of discrimination against female workers are not the same thing. The next

subsection shall derive the relationship between the effect of the unexplained wage

gap on the wages of male non-managerial workers and the effect of the level of dis-

38
crimination against female workers on the wages of male non-managerial workers

under the naı̈ve case that the unexplained wage gap is entirely due to discrimina-

tion. The subsequent subsection will then expand this treatment to include the case

where both discrimination and unmeasured productivity differences between men

and women cause the unexplained wage gap.

2.3.3 Inferring Discrimination from the Unexplained Wage Gap

Even if the explained portion of the wage gap perfectly captured productivity

differences between men and women, leaving the unexplained portion of the wage

gap entirely due to discrimination, there would be a difference between the unex-

plained wage gap and the measurement of the effect of discrimination on women’s

wages. The unexplained wage gap would measure the effect of discrimination on

women’s wages and the effect of discrimination on men’s wages. I disentangle these

effects by examining the effect of the wage gap on the men’s wages.

Let D be the level of wage discrimination against women:

D ≡ Wf∗ − Wf (2.12)

where in Equation (2.12) Wf is the actual level of women’s pay and Wf∗ is the wage

women would receive in the absence of discrimination. Similarly for men, let

Wm = Wm∗ + E(D) (2.13)

where in Equation (2.13) E(D) is the effect of discrimination against women on

men’s wages.

The wage gap between male and female workers, Wm − Wf , can thus be decom-

posed

Wm − Wf = Wm∗ + E(D) − (Wf∗ − D) (2.14)

39
or

Wm − Wf = Wm∗ − Wf∗ + E(D) + D (2.15)

where Wm∗ − Wf∗ represents the wage gap not due to discrimination and G, the wage

gap due to discrimination, can be expressed

G = E(D) + D (2.16)

E(D) is some function of D. A first order polynomial approximation of this

function would be E(D) ≈ µD, by Taylor series expansion (See Appendix A).

Recombining the terms,

G ≈ D + µD (2.17)

and
1
D≈ G (2.18)
(1 + µ)
µ
If ψ ≡ (1+µ)
, then

E(D) ≈ ψG (2.19)

Substituting Equation (2.19) into Equation (2.13),

Wm ≈ Wm∗ + ψG (2.20)

By previous definition,
µ
ψ≡ (2.21)
(1 + µ)
Algebraic manipulation reveals

ψ
µ= (2.22)
(1 − ψ)

The variable µ represents the effect of discrimination against women on the wages

of men. Because decomposition reveals the unexplained wage gap and not the

portion of the gap that affects women’s wages, linear regression cannot identify µ.

40
Nonetheless, with a Taylor series approximation, µ can be expressed as a function

of ψ, which linear regression can identify. Thus, a ψ between zero and one indicates

that where there is more discrimination, male coworkers earn more money. A ψ less

than zero indicates that where there is more discrimination, male coworkers earn less

money. A ψ greater than one would indicate the unlikely result that discrimination

against women reduces men’s wages more than it reduces women’s wages.

2.3.4 Factors Other than Discrimination Entering the “D” Variable

In Section 2.3.3, I assumed that Wm∗ and Wf∗ represented the log wage levels

that men and women, respectively, would earn in the absence of discrimination

while E(D) and D represent the effect of discrimination on male log wages and the

reduction in female log wages due to discrimination. More accurately, Wm∗ and Wf∗

represent the log wage levels for men and women predicted by levels of observed non-

discriminatory explanatory variables, while D is the reduction in women’s wages due

to discrimination and differences in unobserved non-discriminatory determinants of

wages.

Dj = δj + ∆j (2.23)

In Equation (2.23), δj represents reduction in the pay of female workers due to

actual discrimination in locale j, while ∆j represents the difference in male and

female pay in locale j due to unobserved skill differences. E(D) is still the effect of

D on male wages, but not all the assumptions made in the previous section continue

to make sense.

Holzer and Neumark’s (2000) argue that while the size of share of the male-female

wage gap due to discrimination is still open to some empirical dispute, discrimina-

tion does exist and accounts for some positive portion of that gap. Holzer and Neu-

41
mark’s argument puts certain bounds on the terms of Equation (2.23). In particular,

∀j, δj > 0. This implies that there may be a gap in unobserved non-discriminatory

determinants of wages, but that gap neither explains the entire male-female pay gap

nor does it over explain it.

Hammermesh (Hammermesh and Biddle 1994; Donald and Hammermesh 2004)

suggests that in the presence of customer preferences for goods or services produced

by one group of workers over another where such customers are willing to pay a

premium for (in this case) goods or services from male workers over female workers,

the sex of a worker becomes an actual factor in that worker’s marginal revenue

product for the firm. In this case, it is impossible to separate discrimination from

productivity differences by looking at the inputs because the same input (sex) is the

source of both potential discrimination by managers or coworkers and at the same

time differences in productivity (because customers may be willing to pay more for

goods or services from male workers). If these theorized effects exist, decomposition

cannot separate “productive differences” from “discriminatory differences” in the

pay gap.

Such a possibility highlights the need for the approach used in this study. If

local differences in customer taste for male workers drive the pay gap, and man-

agers everywhere pay according to marginal revenue product, there should be no

relationship between the pay gap and manager pay. If manager response to cus-

tomer taste for male workers drive the pay gap (the taste exists everywhere, but

some managers have a sense of fairness or fear of anti-discrimination laws and thus

pay more equally even though marginal revenue product is different for male and

female workers), managers should earn more where there is more discrimination.

If managers earn less in the presence of more discrimination, ceteris paribus, this

implies that manager tastes and not customer tastes or other differences in marginal

42
revenue product between men and women drive the gap.

If customer taste or other differences in marginal product combine with manager

taste and coworker taste, the two effects are opposite in direction. The regression

would then indicate which effect is dominant, but the magnitude of the effect would

actually be stronger than its apparent magnitude because the countervailing effect

would mask part of the dominant effect. If the gap cannot be decomposed into

discrimination by the firm and productivity differences, examining whether man-

agers or coworkers reveal a preference for discrimination by taking lower pay in the

presence of more discrimination helps identify the source of the pay gap.

Thus, if Dj contains two components, only one of which is discrimination, under

a reasonable set of assumptions, unmeasured productivity differences masking as

discrimination will make the effect of apparent discrimination on the wages of privi-

leged groups appear to be more positive than it actually is. If the apparent effect of

discrimination on non-target groups is negative, the magnitude of the effects found

under the assumptions of Section 2.3.3 will change, but the sign should stay the

same. The better the unexplained wage gap measures the male-female wage gap

due to discrimination, the better the magnitudes calculated will accurately measure

the magnitudes of the effects of wage discrimination against female non-managerial

workers on the wages of male non-managerial workers and managers.

2.4 Data and Sample Selection

2.4.1 Census of Population and Housing, 2000

This study uses data from the United States Census of Population and Housing,

2000. The decennial census in the United States attempts to question the entire

population of the United States. The long form questionnaire, however, which con-

tains most information of value to economists, goes out to one sixth of United States

43
households, or about 16.7% of the population. The Census Bureau does not release

the data collected in the decennial census in a way that could identify respondents.

Microdata, or individual responses to select questions, is available, however, with

geographic identifiers broad enough that individual respondents cannot be identi-

fied. The microdata release used in this study is the 2000 1% Public Use Microdata

Sample, or 1% PUMS release. This sample includes observations of 2,818,644 indi-

viduals.

The primary virtue of this data set is its size; my empirical strategy calls for sep-

arate regressions for male and female populations in each geographic labor market

in the country to measure local levels of discrimination, a very observation-intensive

task. The 1% Public Use Microdata Sample reports each respondent’s POWPUMA.

For respondents who live outside a Census Bureau-defined Metropolitan Statistical

Area, this POWPUMA corresponds to the super-PUMA (Public Use Microdata

Area) in which the respondent works, the smallest geographical area reported in the

1% Public Use Microdata Sample for place of work. For respondents in Metropoli-

tan Statistical Areas, the POWPUMA corresponds to a common code for all Public

Use Microdata Areas included in the Metropolitan Statistical Area. Each POW-

PUMA contains enough male and female workers to meaningfully conduct a Oaxaca

decomposition of male-female wage inequality for that locality7 .

7
The 5% Public Use Microdata Sample would provide a larger sample size. The 5% sample,
however, lacks some variables that could identify the respondent within that more specific geo-
graphical area. In particular, the 5% sample only reports “place of work” at the state level, and
no finer. Analysis by place of work is key to this study and the primary reason for selecting a
Census data set over, for instance, the Current Population Survey, which collects variables includ-
ing unionization status and some information about non-cash benefits that would be helpful for
this study. The main reason to use a Census data set over a CPS data set is that the Census
has enough observations to meaningfully measure levels of discrimination in localities smaller than
the state level. While the 5% sample includes more than enough observations for such small-scale
geographic analysis, the fact that it does not report place of work at a finer level than the state
negates any reason for selecting that sample.

44
2.4.2 Sample Selection

The study of how wage discrimination against female non-managerial workers

affects the wages of their managers and male coworkers uses four distinct samples:

a sample of non-managerial male workers to estimate the effect of discrimination

on their wages, a sample of managers to estimate the effect of discrimination on

their wages, and samples of male and female non-managerial workers to measure

the unexplained wage gap.

In the case of male non-managerial workers, using the same population to es-

timate the unexplained wage gap and its effect on their wages leads to statistical

bias. If Equation (2.24) determines the wages w for a male worker i,

ln wi = Wi = Xi BM + i (2.24)

and (2.25) represents a Oaxaca decomposition,

GU =(W̄M − W̄F ) − (X̄M − X̄F )BM (2.25)

NM
1 X
= (Wi − Xi BM ) − W̄F + X̄F BM
NM i=1

then (2.26) shows that the unexplained wage gap GU includes the error term i of

each individual male worker.

NM
U 1 X
G = i − W̄F + X̄F BM (2.26)
NM i=1

This means that if in a subsequent regression GU were used as a explanatory term,

this would violate the regression assumption that the explanatory variables and the

error term are uncorrelated.

To avoid this problem, I use a split sample: one sample of male non-managerial

workers in the Oaxaca decomposition to measure the unexplained sex wage gap in

45
each locale, and a separate sample in the final regression to calculate the effect of

the unexplained sex wage gap on the wages of male non-managerial workers.

Lettau (1997) found that similar part and full time workers have large wage

differentials for identical jobs. Examining the effect of part time workers would be

an important extension to this research, but for this study I simply recognize that

part time workers receive different returns for their skill and experience than full

time workers, and thus exclude workers working less than 30 hours a week from the

sample.

In each discrimination model, the profit maximization motive plays an important

role in predicting firm behavior. For this reason, I exclude military and government

workers (those with Census industry codes over 936, corresponding to NAICS code

92) from the sample.

2.5 Variables in the Wage Regression

In an extensive review of empirical discrimination literature, Altonji and Blank

(1999, p. 3157) report that the standard “textbook” decomposition should control

for education, experience8 , region, occupation, industry, and whether the job is in

the public sector or part time. I control for a number of these characteristics via

sample selection: this study considers managerial and non-managerial occupations

in separate samples, and excludes part-time and public sector workers.

In the first specification, the dependent variable is the hourly wages of male

non-managerial workers. In the second specification, the dependent variable is the

hourly wages of managers in another. The unexplained sex wage gap in locale j
8
Because of restrictions imposed by the data set they use, the Current Population Survey
(CPS), Altonji and Blank would have to control for “potential experience” or age, and not actual
work experience. In 1979, one could adjust potential experience for years of schooling, but in
1995, the CPS does not report years of schooling, making the schooling-adjusted years of potential
experience infeasible given the data. Imputation of years of schooling from the schooling data
could allow them to control for imputed education adjusted potential experience.

46
is the independent variable of interest in both specifications; Figure 2.1 shows this

variable in the various POWPUMAs.

This study also controls for a fourth-degree polynomial of the respondent’s age

(see Appendix), educational attainment dummy variables, race, sex (manager re-

gression only), sector of employment, the percentage of workers in the POWPUMA

with varying levels of education, the unemployment rate for people who live within

the geographical territory of the POWPUMA, the average rental rate for individuals

who live in rented housing within the geographical territory of POWPUMA, and

the percentage of workers in each industrial sector in the POWPUMA.

Table B.1 in the Appendix describes the dataset for the sample of male non-

managerial workers. Table B.2 describes the sample of managers.

2.6 Results and Interpretation

OLS regression of the wages of male non-managers with Rogers’ cluster-correlation

standard error adjustment shows that the unexplained male-female log wage gap on

the log wages of male non-managerial workers has a coefficient of -0.292. If the

unexplained male-female wage gap grows by 1%, wages for male non-managerial

workers fall by 0.290%. Subsequent analysis of this result and its theoretical im-

plications will depend primarily on the finding that the effect of the log wage gap

on male workers’ log wages is negative. Using the cluster-adjusted standard errors,

one can be 98% confident that this effect is, in fact, negative. If the log gap’s co-

efficient on male wages is 0.292, Equation (2.22) implies the effect of the proxy for

discrimination on male wages has a coefficient of -0.226.

The unexplained male-female log wage gap has a coefficient of -0.410 on man-

agers’ log wages. If the unexplained male-female wage gap grows by 1%, managers’

wages fall by 0.407%. Subsequent analysis of this result and its theoretical im-

47
plications will also depend primarily on the finding that the effect of the gap on

managers’ log wages is negative. Using the cluster-adjusted standard errors, one

can be 98.6% confident that this effect is negative.

TABLE 2.1

DETERMINANTS OF LOG WAGES OF FULL-TIME WORKERS

UNITED STATES, 2000

Male non-managers Managers

Variable Coeff. (Std. Err.) Coeff. (Std. Err.)

Locale Unexplained
Sex Wage Gap -0.292∗ (0.125) -0.410∗ (0.166)

Age 0.137∗∗ (0.007) 0.129∗∗ (0.017)

Age2 -0.003∗∗ (0.000) -0.002∗∗ (0.001)

Age3 0.000∗∗ (0.000) 0.000† (0.000)

Age4 0.000∗∗ (0.000) 0.000 (0.000)

Some HS 0.124∗∗ (0.009) 0.029 (0.027)

HS Diploma 0.247∗∗ (0.011) 0.111∗∗ (0.026)

Some College 0.338∗∗ (0.013) 0.231∗∗ (0.026)

College Diploma 0.582∗∗ (0.014) 0.473∗∗ (0.027)

Graduate Degree 0.815∗∗ (0.014) 0.624∗∗ (0.026)

Black -0.128∗∗ (0.006) -0.104∗∗ (0.011)

Female -0.240∗∗ (0.005)

Mining 0.422∗∗ (0.019) 0.773∗∗ (0.052)

49
 TABLE 2.1 (CONT.)

Male non-managers Managers

Variable Coeff. (Std. Err.) Coeff. (Std. Err.)

Utilities 0.503∗∗ (0.014) 0.720∗∗ (0.025)

Construction 0.267∗∗ (0.012) 0.568∗∗ (0.017)

Manufacturing 0.312∗∗ (0.013) 0.694∗∗ (0.017)

Wholesale Trade 0.247∗∗ (0.013) 0.605∗∗ (0.019)

Retail trade 0.127∗∗ (0.012) 0.525∗∗ (0.018)

Transportation and
warehousing 0.256∗∗ (0.013) 0.546∗∗ (0.020)

Information 0.365∗∗ (0.016) 0.662∗∗ (0.019)

Finance and Insurance 0.441∗∗ (0.020) 0.703∗∗ (0.019)

Real Estate and Rental

and Leasing 0.185∗∗ (0.016) 0.433∗∗ (0.021)

Professional, Scientific,
and Technical Services 0.406∗∗ (0.013) 0.662∗∗ (0.018)

Management of Companies
and Enterprises 0.497∗∗ (0.063) 0.761∗∗ (0.044)

Administrative and
Support and Waste
Management 0.042∗∗ (0.013) 0.504∗∗ (0.020)

Education 0.121∗∗ (0.013) 0.407∗∗ (0.017)

Health Care and

Social Assistance 0.299∗∗ (0.014) 0.477∗∗ (0.017)

Arts and Entertainment 0.061∗∗ (0.018) 0.401∗∗ (0.026)

50
 TABLE 2.1 (CONT.)

Male non-managers Managers

Variable Coeff. (Std. Err.) Coeff. (Std. Err.)

Accommodation and
Food Services -0.074∗∗ (0.014) 0.235∗∗ (0.016)

Other Services 0.026† (0.013) 0.382∗∗ (0.019)

Locale Some HS 0.540† (0.313) -0.063 (0.392)

Locale HS Diploma 0.655∗∗ (0.179) -0.497† (0.256)

Locale Some College 0.663∗∗ (0.195) -0.530∗ (0.260)

locale College
Diploma 0.618∗∗ (0.232) -0.337 (0.305)

Locale Graduate
Degree 1.176∗∗ (0.397) 0.477 (0.452)

Locale Unemployment
Rate 1.147∗∗ (0.196) 0.741∗∗ (0.260)

Locale Average
Rental Cost 0.000∗∗ (0.000) 0.000∗∗ (0.000)

Locale Mining 0.463† (0.246) 0.371 (0.287)

Locale Utilities 0.809 (0.550) 0.461 (0.757)

Locale Construction 0.555∗ (0.280) 0.664∗ (0.297)

Locale Manufacturing 0.457∗∗ (0.134) 0.700∗∗ (0.176)

Locale Wholesale Trade 0.804† (0.457) 0.983† (0.505)

Locale Retail Trade -0.525∗ (0.258) 0.425 (0.319)

Locale Transportation
and Warehousing 0.515∗ (0.247) 0.902∗∗ (0.248)

51
 TABLE 2.1 (CONT.)

Male non-managers Managers

Variable Coeff. (Std. Err.) Coeff. (Std. Err.)

Locale Information -0.371 (0.383) 0.407 (0.503)

Locale Finance and

Insurance 0.859∗∗ (0.246) 1.342∗∗ (0.228)

Locale Real estate

and Rental and Leasing 0.050 (0.803) 0.718 (0.867)

Locale Professional,
Scientific, and
Technical Services 0.348 (0.337) 0.789∗ (0.316)

Locale Management
of Companies and
Enterprises 9.887 (6.493) 17.447∗ (7.289)

Locale Administrative
and Support and
Waste Management 0.093 (0.465) 1.919∗∗ (0.521)

Locale Education -0.430† (0.260) -0.112 (0.288)

Locale Health
Care and Social
Assistance -0.126 (0.223) 0.538† (0.296)

Locale Arts and

Entertainment 0.960∗∗ (0.359) 1.933∗∗ (0.378)

Locale Accommodation
and Food Services -0.561 (0.364) -0.887∗ (0.379)

Locale Other services 0.216 (0.406) -0.031 (0.503)

Intercept -1.218∗∗ (0.195) -0.612∗ (0.284)

52
 TABLE 2.1 (CONT.)

Male non-managers Managers

Variable Coeff. (Std. Err.) Coeff. (Std. Err.)

N 293323 111989
R2 0.281 0.289
Significance levels : † : 10% ∗ : 5% ∗∗ : 1%

Log wages for female non-managerial workers would have been at Wf∗ in the

absence of discrimination and unmeasured productivity differences and are observed

at Wf ≡ Wf∗ − D in the presence of discrimination and unmeasured productivity

differences. When the reduction in female pay due to discrimination and unmeasured

productivity differences increases by 1%, male wages fall by 0.227%. Discrimination

depresses the wages of both male and female non-managerial workers.

These results illustrate the importance of multivariate regression for this prob-

lem. By pair-wise correlation alone, the unexplained wage gap has a positive corre-

lation with log wages of both male non-managerial workers and managers. Dismiss-

ing the possibility of second-order effects, the temptation exists to conclude that

a larger unexplained sex wage gap caused higher wages for male non-managerial

workers and managers. Multivariate regression reveals that the opposite is true in

both cases (The correlation tables anticipate this; see discussion of Table B.3 and

Table B.4 in the Appendix on page 122).

All non-managers earn less in the presence of discrimination. This is consistent

with coworker taste discrimination models, divide-and-conquer models, and manager

taste discrimination models. In a “neutral” coworker taste model, the presence or

absence of discrimination would not affect managers’ pay. In a “neutral” model,

any loss in productivity due to discrimination would be made up by the reduction

53
 TABLE 2.2

MODELS CONSISTENT WITH DISCRIMINATION REDUCING THE WAGES

OF ALL NON-MANAGERIAL WORKERS AND THEIR PREDICTIONS FOR
THE EFFECT OF INCREASED DISCRIMINATION ON MANAGERS’ WAGES

Model Predicted effect of discrimination

“Neutral” coworker taste No effect on managers’ wages
Divide and conquer Increase managers’ wages
Managerial taste Decrease managers’ wages

of payroll costs. A manager who hires discriminating workers (and pays them less)

would be no more or less efficient than a manager who pays a premium for non-

discriminating workers who are more productive. Thus, a neutral model would

predict no significant effect of discrimination against female workers on managers’

wages.

In one form of divide-and-conquer model, the reductions in payroll costs due to

discrimination more than offset the reduction in productivity due to discrimination.

Firms would thus find it desirable to pursue a policy of hiring discriminating workers

as a form of cost minimization. Managers who hire discriminating workers (or culti-

vate a taste for discrimination among their workforce) would thus be more produc-

tive managers (cost-cutting) than managers who hire non-discriminating workers.

Thus, these models would predict managers in localities with higher discrimination

would make more money than managers in localities with less discrimination. If,

however, managers are just another kind of worker, a divide-and-conquer model that

includes dividing and conquering the managers would predict results identical to a

manager discrimination model (indeed, owners would cultivate manager discrimina-

tion in such a model). The only way to distinguish between a shareholder-indifferent

54
manager discrimination model and a shareholder-supported manager discrimination

model would be to have data on rates of return on investment in each of these lo-

calities, which is not available in this dataset.

A manager taste discrimination model would predict that the desire of managers

to not hire female workers drives discrimination at least in part. Managers are willing

to take a pay cut to be able to hire fewer female workers. Thus, a manager dis-

crimination model would predict that discrimination against female non-managerial

workers would have a negative effect on managers’ wages.

The estimation of the manager model revealed that the effect of the unexplained

gap in log wages by sex on the log wages of managers can be represented linearly

with a coefficient of -0.401. The transformation from Equation (2.18) implies that

when wage discrimination against female non-managerial workers increases by 1%,

under the assumptions outlined in section 2.3.4, managers’ wages fall by more than

0.531%.

2.7 Conclusions

Sex discrimination depresses the wages of managers and male coworkers. These

findings imply three real-world conclusions.

First, the results indicate that manager taste for hiring male workers at least

partly drives the sex wage gap. Managers thus appear to pay for wage discrimination

against female workers with foregone wages. This is consistent with the hypothesis

that manager tastes cause at least some part of wage discrimination.

Secondly, the fact that the data support a model of discrimination that can per-

sist in competitive product markets supports a theory that competes with Becker’s

claim (1957) that competitive markets drive out discrimination in the long run. This

means that even in the long run, eliminating discrimination could require interven-

55
tion and cannot be left to competitive markets.

Finally, this study does not have the data to conclude whether firms benefit

from discriminating managers’ foregone wages or if the productivity losses from the

discrimination offset the managers’ wage reduction. If firms can reduce the pay of

discriminating managers by more than their lost profits from discriminatory hiring,

firms have an incentive to promote tastes for sex discrimination, as in the divide-

and-conquer models. If, on the other hand, as in the competitive model, firms

simply recoup their losses due to managers’ tastes for discrimination, firms have no

incentive to promote discrimination. Future research should attempt to answer this

question to help determine whether anti-discrimination campaigns should target

individual managers’ tastes for hiring female workers or need to target firms.

56
 CHAPTER 3

A PRELIMINARY FORMAL MODEL OF DISCRIMINATING MANAGERS AS

DOLLAR STORE MERCHANDISE

This chapter develops a preliminary model in which discrimination is caused

by managers’ tastes. In this model, which Arrow (1998) proposed but did not

develop because he believed the phenomenon has “not been tested but certainly

seems dubious,” managers’ tastes alone drive discrimination. This model predicts

that as managers’ taste for discrimination rises, the wage gap rises, managers’ pay

falls, and non-targeted workers’ pay rises. Simple modifications to model find that

as managers’ taste for discrimination rises, the wage gap rises, managers’ pay falls,

and non-targeted workers’ pay falls also.

Managers working for firm owners often act as their agents in hiring, firing,

and wage determination decisions. Many economic models of wage discrimination

assume that a firm’s potentially discriminatory decisions either maximize owner

utility (Becker 1957) or profits (Arrow 1973; Aigner and Cain 1977; Phelps 1972;

Reich 1984; Blau, Ferber, and Winkler 1992, p. 224). This study shows that a

model in which hiring decisions maximize the managers’ utility rather than the

firm’s profits predicts a different effect of discrimination on managers’ wages than

many models that ignore this principal/agent distinction.

Manager discrimination extends the coworker discrimination model to consider

managers in a firm as agents of the employer who make hiring decisions but seek

57
to maximize their own utility, not firm profitability. This principal/agent situation

allows for people in positions of authority at the firm to perpetuate discrimination

and yet not hurt the firm’s bottom line, thus negating the push of market forces

that would have otherwise driven discriminating firms out of business.

In the manager discrimination model, the assumptions of Becker’s original model

hold with one exception. Some people who run a company experience psychic disu-

tility from hiring employees from some groups. If enough people making hiring

decisions prefer to hire workers from one group to another, ceteris paribus this re-

duces demand for employees from the disfavored group, causing the going wage for

such employees to fall below that of employees from the favored group, as in the

Becker model.

In this model, as in Becker’s, managers who discriminate (either fully by not

hiring from the disfavored group at all, or partially by hiring fewer employees from

the disfavored group) are less efficient than those who do not. The difference lies

in long-run adjustments. Firm owners reviewing these managers find they do not

maximize profits: the agent does not represent the principal well. To faithfully rep-

resent the interest of the principal (by hiring in a non-discriminatory manner) would

incur psychic disutility for the agent. If firms compensate managers according to

their marginal revenue product, a discriminating manager would earn less money

than a non-discriminating manager because the discriminating manager does not

hire efficiently. This model thus predicts that where levels of discrimination are

higher, manager pay is lower. Discriminating managers are like low-quality man-

ufacturing equipment or dollar store merchandise: they do not perform as well as

higher-quality managers, but they also cost less.

A manager discrimination model would suggest that managers under-hire or un-

derpay female workers because of their own distaste for female workers, and not

58
because of pressure from the owners. If managers make hiring and firing decisions

and set pay, then by definition, managers will carry out wage discrimination. The

mere fact that managers are the ones carrying out the discrimination does not dif-

ferentiate a model of “manager discrimination” from Becker’s 1957 model of “owner

discrimination.” If owners do not want to employ women because doing so would

be a source of disutility, and the firm’s managers do not hire women to please the

owners, the agents (managers) simply accommodate the principals (owners), and the

original Becker (1957) model applies. If the owners of a firm want to strategically

divide male and female workers to segmenting the labor force, as in racial divides in

Reich (1984), a manager who systematically underpays or under-promotes women in

this context would simply act as a good agent, and traditional segmentation models

would apply.

This distinction between managers discriminating on behalf of owners and man-

agers discriminating because of their own tastes is crucial and relevant. On June

22, 2004, Federal Judge Martin J. Jenkins of the United States District Court in

San Francisco certified 1.6 million female current and former Wal-Mart employ-

ees for a class-action discrimination lawsuit against Wal-Mart. The suit claims

that Wal-Mart systematically paid and promoted women less than men. Wal-Mart

appealed the class-action certification, claiming that discrimination was not due to

systemic corporate policy (pressure from the principals), but rather the tastes of the

individual managers (Greenhouse and Hays 2004). Even if managers’ tastes drive

discrimination, owners could strategically choose to hire discriminating managers in

an attempt to capture rents.

Note that in a manager discrimination model, discriminating managers do not

“lose” in an absolute sense; they lose pay. While their pay is lower than that of

non-discriminating managers, discriminating managers make up in psychic utility

59
gained from being able to discriminate at least what they lose in wages; if not, they

would act like non-discriminating managers. Arrow (1998) noted that this model

had never been developed or tested, but suggested that such effects were unlikely

to actually exist.

3.1 Labor Supply

In this model, workers maximize utility, given by Equation (3.1).

Lh2
U = wh − (3.1)
2
In Equation 3.1, U is utility, w is a worker’s wage, h is the percentage of normal

working hours that a worker works (which could be more than 100% but could not

be less than zero), and L is a parameter related to the value of non-market activity.

Any utility-maximizing worker would choose to supply labor according to the

following equation: 
 0 w
if <0

L
h= (3.2)
w w
if ≥0


L L

Assume that there exist two perfectly substitutable factors of production, type-o

labor and type-ω labor. These factors of production are identical except in their

impact on the utility of managers.

If in the population of n potential workers, a share of r̃ belong to group o and a

share of 1 − r̃ belong to group ω, the labor supplies for type-o and type-ω labor are

given by Equation (3.3) and Equation (3.4), respectively.

nr̃wo
LSo = (3.3)
L
n(1 − r̃)wω
LSω = (3.4)
L

60
3.2 Labor Demand

Firms hire managers, and call on them to minimize costs for their unit of pro-

duction. Managers are charged with the task of minimizing the cost of production.

Because managers have a pre-determined scale of production, the choice variable of

interest in this cost minimization is r, the fraction of the total number of workers ν

hired by each manager that belong to group o instead of group ω.

Assume the cost function managers face is as follows:

C = νρ(r − r̃)2 + νrwo + ν(1 − r)wω (3.5)

In Equation (3.5), C represents costs, ν represents the number of workers a

manager supervises, and wo and wω represent the wages of workers from group

o and group ω, respectively. The term ρ(r − r̃)2 is a cost the firm incurs if the

group composition of its workforce does not proportionally match r̃, the group

composition of the workforce as a whole. This could be for public relations reasons

or due to regulatory penalties. The parameter ρ represents the strength of legal

anti-discrimination penalties and public relations costs of discriminating.

From the firm’s perspective, the optimal value for a manager to choose would

be rπ∗ , the profit maximizing value of r. This can be determined as follows:


wω −wo



 0 if r̃ + 2ρ
<0

rπ∗ = −wo
r̃ + wω2ρ if 0 ≤ r̃ + wω −wo
≤1 (3.6)
 2ρ


 1 wω −wo
if 1 < r̃ +

2ρ

A manager, however, want to maximize his or her own utility. Assume the

manager’s utility takes the following functional form:

U = λr + (S)γ (3.7)

61
 In Equation (3.7), U represents the manager’s utility, λ is a parameter measuring

how much the manager likes managing members of group o (for managers who dislike

managing members of group o, λ < 0). The variable S is the manager’s pay, and γ

is a parameter related to the manager’s marginal utility of money. Further assume

for the sake of simplicity that managers’ marginal utility of money is constant. This

means that γ = 1.

To enforce cost minimization on the manager, the firm must link the manager’s

pay, S, with the economic consequences of the manager’s decisions on the firm’s

costs. For any value of r a manager chooses other than rπ∗ , the firm does not

minimize costs, and consequently loses money from the manager’s decision. The

losses Λ can be specified:

Λ = ν(ρ((r − r̃)2 − (rπ∗ − r̃)2 ) + (rπ∗ − r̃)(wω − wo )) (3.8)

The firm thus penalizes the manager by a function of the losses the manager

incurs for the firm κ(Λ). Thus the manager’s pay is a base pay level minus a penalty

function based on the losses incurred by the firm as a result of the manager’s choice

of hiring ratios:

S = s − κ(Λ(r)) (3.9)

If the firm penalizes the manager dollar for dollar according to the costs of the

manager’s decision, then κ(Λ) = Λ. This reduces the utility function reduces to:

U = λr + s − Λ(r) (3.10)

In this case, given a manager’s like λ for members of group o, the manager would

62
choose to hire at ratio ru∗ , that manager’s utility-optimizing choice of ratio r:

λ wω −wo



 0 if r̃ + 2νρ
+ 2ρ
<0

ru∗ (λ) = λ
r̃ + 2νρ −wo
+ wω2ρ if 0 ≤ r̃ + λ
+ wω −wo
≤1 (3.11)
 2νρ 2ρ


 1 λ wω −wo
if 1 < r̃ + +

2νρ 2ρ

For values of λ for which the choice of ru∗ is unconstrained, the pay of the manager

can be expressed as a function of λ:

λ2
S(λ) = s − (3.12)
4νρ
Suppose there exists a continuum of managers in a particular labor market whose

taste λ for managing workers of type o is distributed uniformly centered around λ̄

and extending from λ̄− 2δ , the utility parameter for those managers who most dislike1

workers of type o, to λ̄ + 2δ , the utility parameter for those managers who most like

(or least dislike) workers of type o. Since λ is the parameter measuring how much

a particular manager likes working with type-o workers, this will be negative for

managers who do not like managing workers from group o. Because of this uniform

distribution, the formula for λ at any quantile q can be determined:

λ(q) = λ̄ + δ(q − 12 ) (3.13)

Lo
For the entire labor market, the ratio r̄ of Lω +Lo
can be defined:
Z 1
r̄ = ru∗ (λ(q))dq (3.14)
0

δ
If the parameters are such that for both of the extreme cases λ = λ̄ − 2
and

λ = λ̄ + 2δ , the optimal ru∗ (λ) is actually an optimized value and not one of the two

corner solutions, this reduces to:

1
This implies that δ > 0.

63
 1  
λ(q) wω − wo λ̄ + ν(+2rρ + wω − wo )
Z
r̄ = r̃ + + d(q) = (3.15)
0 2νρ 2ρ 2νρ

3.3 Equlibrium

At equilibrium, the labor supplies of each type of labor must match the labor

demand for each type of labor. Equation (3.16) represents the condition needed for

the market for type-o labor to clear, while Equation (3.17) represents the condition

needed for the market for type-ω labor to clear. In these equations, m represents the

number of managers. Total labor demand for type-o workers is the product of νr,

the number of type-o workers hired by the average manager, and m, the number of

managers. Similarly, demand for type-ω workers is equal to the average manager’s

demand for type-ω workers times the number of managers.

nr̃wo
mν r̄ = (3.16)
L
m(1 − r̃)wω
mν(1 − r̄) = (3.17)
L
For unconstrained values of r∗ , the following solution satisfies these equilibrium

conditions.

Lm(Lmν 2 + n(1 − r̃)(2r̃νρ + λ̄))

wo = (3.18)
n(Lmν + 2n(1 − r̃)r̃ρ)

Lm(Lmν 2 + n(r̃)(2(1 − r̃)νρ − λ̄))

wω = (3.19)
n(Lmν + 2n(1 − r̃)r̃ρ)
Let the wage gap be represented G ≡ wω − wo .

Lmλ̄
G=− (3.20)
Lmν + 2n(1 − r̃)r̃ρ)

64
 ∂S
Solving for ∂λ
= 0 yields the intuitively obvious conclusion that an individual

manager earns the maximum salary when λ = 0 and the manager has no incentive

to deviate from the profit-maximizing optimum.

The average manager’s salary can be determined by taking an average of the

managers’ salaries expressed in Equation (3.12).

1
δ2 λ̄2
Z
S̄ = [s − Λ(r[λ(q)])]dq = s − − (3.21)
0 48νρ 4νρ

3.4 On Profit-neutrality

The function κ(Λ) represents how firms penalize managers for lost profits. In the

profit-neutral case, firms extract from manager pay the profits lost by the managers’

discriminatory tendencies. In this case, κ(Λ) = Λ.

This profit-neutral model is necessary if the market for managers is perfectly

competitive. A firm would have no incentive to hire discriminating managers if

κ(Λ) < Λ. In this case, the market for managers would not clear. The excess supply

of discriminatory managers would suppress wages for discriminatory managers until

the wage penalty they faced were the same as their damage to profitability. At that

point, κ(Λ) = Λ.

Similarly, if κ(Λ) > Λ, discriminatory managers would be a source of excess

profits for a firm. Firms without discriminatory managers would have an incentive

to seek them out. The poaching firm would have an incentive to hire for any κ(Λ)

greater than Λ; the manager would have an incentive to switch firms for any κ(Λ)

less than that of the discriminating manager’s current employer. This poaching

process would drive κ(Λ) down to Λ.

If firms have bargaining or information advantages over managers, it could be

the case that κ(Λ) > Λ, meaning that firms over-penalize discriminating managers,

65
extracting more profits from such operations.

If managers had market power and the leadership of these collectively bargaining

managers wanted to reward discriminatory managers, κ(Λ) could be less than Λ.

This would mean that profits would be lower in the presence of discrimination.

If this were the case, the model would not be sustainable in competitive product

markets due to the zero-profit condition.

3.5 Implications
3.5.1 Effect on Managers’ Pay

The central prediction of this model is the predicted effect changes in the make-

up of the discriminatory preferences of managers have on managers’ average pay.

Equation (3.21) gives managers average pay as a function of λ̄, the average dis-

criminatory tendency, and δ, the spread of discriminatory tendencies (as discussed

previously, discriminatory tendencies are distributed uniformly in the range δ wide

centered at λ̄).

∂ S̄ λ̄
=− (3.22)
∂ λ̄ 2νρ
Equation (3.22) shows that as discrimination increases (and λ̄ falls), S̄ falls.

Note that this assumes λ̄ < 0. If this condition did not hold, managers would, on

average, prefer group o over group ω, which contradicts the designations of the two

groups.

∂ S̄ δ
=− (3.23)
∂δ 24νρ
Equation (3.23) demonstrates that as the dispersion around the mean increases,

wages fall. A negative value for δ would not be meaningful.

66
3.5.2 Effect on Non-managerial Pay Gap

The change in the wage gap with respect to λ̄, derived from Equation (3.20),

shows the model’s prediction about the effect of increased discrimination on the

wage gap.

∂G Lm
=− (3.24)
∂ λ̄ Lmν + 2n(1 − r̃)r̃ρ)
Given that L, m, n, ν, and ρ are only meaningful in this model if non-negative,

and r̃ must be between zero and one, exclusive2 , Equation (3.24) demonstrates

that G is monotonically decreasing in λ̄. This mathematically predicts the intuitive

results that the more managers prefer type-ω workers over type-o workers, the larger

the wage premium for type-ω workers. If some managers’ choices of ru∗ (λ) fall into

the constrained corner solution at either end, this result will still hold so long as not

all managers are constrained into a corner solution.

After considering the effect on G, the effect of changing levels of discrimination

on the wages of each group should also be considered.

3.5.3 Effect on Type-ω Workers’ Pay

∂wω Lmr̃
=− (3.25)
∂ λ̄ Lmν + 2n(1 − r̃)
Equation (3.25) demonstrates that as discrimination increases (and λ̄ falls), wω

rises. Thus, manager distaste for type-o workers raises the wages of type-ω workers.

The effect of discrimination on wages of type-ω workers could be expressed as

the level of pay these workers receive in the presence of discrimination minus the

level of pay they would receive in its absence. Equation (3.19) gives the wage for

type-ω workers as a function of λ̄. To find EωD , the effect of discrimination on the
2
The existence of two groups precludes a value of zero or one for r̃.

67
wage of type-ω workers, subtract wω (λ̄) − wω (0) :

Lmr̃
EωD = −λ̄ (3.26)
Lmν + 2n(1 − r̃)r̃ρ

3.5.4 Effect on Type-o Workers’ Pay

∂wo Lm(1 − r̃)

= (3.27)
∂ λ̄ Lmν + 2n(1 − r̃)
Equation (3.27) demonstrates that as discrimination increases (and λ̄ falls), wo

falls also. Thus, manager distaste for type-o workers unsurprisingly reduces the

wages of type-o workers.

The effect of discrimination on wages of type-o workers could be expressed as

the level of pay these workers receive in the presence of discrimination minus the

level of pay they would receive in its absence. Equation (3.18) gives the wage for

type-o workers as a function of λ̄. To find EoD , the effect of discrimination on the

wage of type-o workers, subtract wo (λ̄) − wo (0):

Lm(1 − r̃)
EoD = λ̄ (3.28)
Lmν + 2n(1 − r̃)r̃ρ
Finally, Equation (3.26) and Equation (3.28) reveal that EωD is a function of EoD

and r̃.
r̃
EωD = − EoD (3.29)
1 − r̃
Because the wage gap due to discrimination is made up of the effect of discrimination

on type-o workers and the effect of discrimination on type-ω workers, one can find

the effect of discrimination on either group as a function of the wage gap.

EωD = r̃G (3.30)

68
 EoD = −(1 − r̃)G (3.31)

Note that these findings are dependent on the assumption in Equation (3.1) that

both types of worker have identical utility functions, and in particular, identical

marginal utilities of non-market activity3 . If that assumption does not hold, the

ratio of those parameters would also be a factor in the distribution of the gap.

3.6 Extended Implications

The basic form of this model predicts that as discriminatory tastes strengthen,

managers earn less, type-o workers earn less, and type-ω workers earn more. Minor

modifications to this model could change the sign on the predicted effect on type-

ω worker wages. If type-ω workers also have a taste for working with their own

kind, but they express that taste through their productivity, then as the level of

discrimination rises, type-ω productivity falls, along with type-ω wages. This could

be a “shirking” model, in which type-ω workers, resentful of having to work with

type-o workers, choose to work less hard. It could also be an actual productivity

model — if, despite claims of the benefits of diversity, workers actually are more

productive with co-workers like them because of language, communication style,

cultural commonalities, etc., this would produce the same effect.

In the current model, the crowding effect boosts type-ω wages by a factor of
Lmr̃
Lmν+2n(1−r̃)
. A productivity effect would have to at least offset this effect to reverse

the sign of the prediction. A localized aggregate demand model could also reverse

the wage increase for type-ω workers. If local managers are the primary consumers

of products in localized product markets, and depressed managers’ wages depresses

demand for labor, then type-ω workers could suffer wage decreases where managers

have reduced wages. Finally, wage-spread models in which institutional constraints

3
This is often incorrectly called “marginal utility of leisure.”

69
somehow link workers’ wages and managers’ wages (Lazear 1999) could also apply a

downward pressure on type-ω wages where manager pay falls. Again, such an effect

would have to outweigh the crowding effect in order to reverse the predicted sign of

the effect on the wages of type-ω workers. Future work should explore these models

in detail.

70
 CHAPTER 4

A DIFFERENT STORY: HOW WAGE DISCRIMINATION AGAINST BLACK

NON-MANAGERIAL WORKERS AFFECTS THE WAGES OF MANAGERS

AND WHITE COWORKERS

4.1 Do Managers Pay for Race Discrimination?

As discussed in Chapter 3, if managers are willing to take lower pay as a com-

pensating wage differential in exchange for not having to manage with as many

black workers, then in places where demand for black employees was reduced by

these discriminatory forces, managers would receive less pay. This study tests this

hypothesis to see if a manager discrimination model is consistent with observable

data about the unexplained race wage gap.

This chapter conducts tests of the model developed in Chapter 3. Equations

derived in Chaper 3 that provide falsifiable, testable hypotheses will be restated as

appropriate in this chapter (with their original numbering), but not re-derived.

4.2 Testing the Specific Model

Examining the signs of the derivatives in Equation (3.24), Equation (3.25), and

Equation (3.27) yields testable, falsifiable predictions that the model makes about

measurable data. These equations will be repeated in the section discussing them.

The model proposed in Chapter 3 predicts that as λ̄, the mean level of discrimi-

natory preferences among managers, increases (i.e. approaches zero from beneath),

71
wages of black non-managers would rise, wages of white non-managers would fall,

and manager pay would increase. Section 4.3 tests these reduced-form predictions.

The test, however, need not be limited to these generalized predictions. The ac-

tual magnitudes of relationships among the various measurable data and estimable

parameters predicted by the functional form of this model can be tested, and not

merely the signs.

In general, any test of these models will depend on separating the effect of

discrimination from other determinants of pay for a particular population. All of

these tests will begin with a regression of the basic form of Equation (4.1).

ln wi = BX + i, j + E D + i,j (4.1)

4.2.1 Predictions about White Non-managers’ Wages

Equation (3.30) predicts the effect of the wage gap due to discrimination, G, on

the wages of type-ω workers (white workers, in this context).

EωD = r̃G (3.30)

This means that if a variable x0 were to be constructed such that x0 ≡ r̃G, it

could be included in a wage regression with the other determinants of the wages of

white workers.

ln wi = BXi,j + θ0 x0j + i,j (4.2)

Because EωD is the predicted effect of how discrimination against blacks affects

white workers’ wages, it follows that θ0 , the coefficient in the wage regression for

x0 should be 1. The model predicts that the effect of discrimination on the wages

of white workers should be the fraction of workers who are black times the total

72
black-white wage gap due to discrimination.

Hypothesis 4.1
θ0 = 1

4.2.2 Predictions about Managers’ Wages

Equation (3.28) defines the effect of discrimination on the wages of type-o workers

(black workers, in this context).

Lm(1 − r̃)
EoD = λ̄ (3.28)
Lmν + 2n(1 − r̃)r̃ρ
Equation (3.31) expresses this same effect in terms of G, the wage gap due to

discrimination.

EoD = −(1 − r̃)G (3.31)

These equations, taken together, yield λ̄, the average loss of profitability incurred

by a manager, as a function of the discriminatory wage gap, G, the number of

workers supervised by each manager, ν, the fraction of the population that is black,
n
r̃, a parameter representing the strength of anti-discrimination penalties,ρ, and Lm
,

a fraction involving n, the number of potential workers, L, a parameter related to

workers’ marginal utility of non-market time, and m, the number of managers.

G(Lmν + 2nρr̃(1 − r̃))

λ̄ = − (4.3)
Lm

Equation (3.21) provides S̄, the average manager’s salary, as a function of λ̄.

δ2 λ̄2
S̄ = s − − (3.21)
48νρ 4νρ
Taking S̄(λ̄) − S̄(0), one can find the effect of average discriminatory preference

λ on the average wage of managers, Ebλ̄ , in a particular locale.

73
 λ̄2 G2 ν G2 r̃(1 − r̃) G2 n2 r̃2 (1 − r̃)2 ρ
Ebλ̄ = − =− − − (4.4)
4νρ 4ρ Lm L2 m 2 ν
Some of these variables are observable from data; νj , r̃, n, and m can all be mea-

sured for each locale j. Similarly, Oaxaca decomposition yields an approximation

of Gj for each locale. If one assumes that ρ and L are fixed across locales, then the

model reduces to a linear regression, with several testable predictions. These terms

can be separated into estimable parameters θ1 . . . θ3 and observable data x1 . . . x3 :

1
θ1 ≡ − (4.5)
4ρ
1
θ2 ≡ − (4.6)
L
ρ
θ3 ≡ − 2 (4.7)
L
x1 ≡ G2 ν (4.8)
G2 r̃(1 − r̃)n
x2 ≡ (4.9)
m
G r̃ (1 − r̃)2 n2
2 2
x3 ≡ (4.10)
νm2
The variables x1 , x2 , and x3 and the parameters θ1 , θ2 , and θ3 do not have

obvious intuitive meaning; they are useful as ways to make hypotheses about their

constituent variables from the model, which do correspond to economically useful

concepts.

Thus, Ebλ̄ can be expressed as the sum of θ1 x1 , θ2 x2 , and θ2 x2 .

ln wi = BXi,j + θ1 x1j + θ2 x2j + θ3 x3j + i,j (4.11)

The model predicts the following hypotheses about the effect of x1 , x2 , and x3

on manager wages:
Hypothesis 4.2
θ1 < 0

74
The model predicts that the effect of the x1 variable on managers’ wages should

be negative. If θ1 > 0, this would imply that ρ were negative. This would mean

that there would be a regulatory or public relations reward for discriminating. In-

stitutional evidence (see Chapter 1) clearly indicates the first possibility is false —

discriminating leads to fines, not subsidies. The second could be true, but if it were,

a model of consumer discrimination would be in order. If the public were more likely

to support a firm that was known to discriminate, this could cause ρ to be negative.

If this happened, the mathematical model might survive, but the intent behind a

model of manager discrimination would not, and it would be more appropriate to

try to capture the effect with a model of consumer discrimination.

Hypothesis 4.3
θ2 < 0

The model predicts that the effect of the x2 variable on managers’ wages should

be negative. If θ2 > 0, this would imply that L were negative. This would mean

that workers have a distaste for time spent in non-market activity. Work would be a

source of utility, and workers would be willing to pay to work, not demanding a wage.

This is inconsistent with the assumptions behind models of wage determination.

Hypothesis 4.4
θ22
θ3 =
4θ1
The model predicts that the coefficient on the x1 and x2 variables should be related

to the coefficient on the x3 variable. The algebra of the model requires that θ3 =
θ22
4θ1
. If this does not hold then the model is misspecified. This has no intuitive

interpretation, but serves as a check on the functional form specified for the model.

4.3 A Generalized Test

If the model fails some of the tests of Hypotheses 4.1 — 4.4, this indicates that

the functional form specified in Chapter 3 does not hold. If, however, an increase in

75
the discriminatory wage gap were to accompany an increase in white workers’ wages

and a decrease in managers’ wages, but some of the tests of Hypotheses 4.1 — 4.4

were to fail, this would suggest that while the specific model proposed in Chapter 3

does not hold, perhaps a different model of manager taste discrimination would.

Accordingly, this section proposes a set of reduced-form hypotheses that would

indicate that a model like this would explain race discrimination.

All of these tests measure the effect of G. The discriminatory wage gap serves as

a proxy for λ̄, the average level of discriminatory taste among managers, because of

the monotonic relationship between them. Specifically, Equation (3.24) shows the

inverse relationship between G and λ̄:

∂G Lm
=− (3.24)
∂ λ̄ Lmν + 2n(1 − r̃)r̃ρ)

4.3.1 White Non-managers

Let Equation (4.12) be a wage regression in which wi is the wage of white worker

i, Gj is the discriminatory wage gap in locale j, Xi,j are the individual and locale-

specific determinants of wages other than the discriminatory wage gap, and i,j is

an error term.

ln wi = BXi,j + θ4 Gj + i,j (4.12)

Equation (3.25) predicts that as G increases due to an decrease in λ̄ (i.e. λ̄

becomes more negative, and its absolute value, |λ̄|, increases), wages increase for

white non-managers.

∂wω Lmr̃
=− (3.25)
∂ λ̄ Lmν + 2n(1 − r̃)
Thus, the testable hypothesis is that the effect of a discriminatory wage gap on

wages is negative:

76
Hypothesis 4.5
θ4 > 0

4.3.2 Managers

Similarly for managers, let Equation (4.13) be a wage regression in which wi is the

wage of white manager i, Gj is the discriminatory wage gap in locale j, Xi,j are the

individual and locale-specific determinants of wages other than the discriminatory

wage gap, and i,j is an error term1 .

ln w = BXi,j + θ5 Gj + i,j (4.13)

Equation (3.22) predicts that as G increases due to a decrease in λ̄ (i.e. λ̄

becomes more negative, and its absolute value, |λ̄|, increases), wages decrease for

managers.

∂ S̄ λ̄
=− (3.22)
∂ λ̄ 2νρ

Hypothesis 4.6
θ5 < 0

If Hypothesis 4.5 were rejected but Hypothesis 4.6 were upheld, this would sug-

gest that a model like the one proposed in the Section 3.6 might be in order. If

Hypothesis 4.6 were rejected, that would suggest that a model of manager discrim-

ination might not be the best explanation of wage discrimination on the basis of

race.
1
Note that this test shows the effect of λ̄j on the wages of a manager in locale j, not the effect of
that individual manager’s λi , which this study does not measure. Firm-level or even more refined
data would be necessary to separate the effects of λi and λ̄j .

77
4.4 Data and Sample Selection
4.4.1 Census of Population and Housing, 2000

This study uses data from the United States Census of Population and Housing,

2000. The decennial census in the United States attempts to question the entire

population of the United States. The long form questionnaire, however, which con-

tains most information of value to economists, goes out to one sixth of United

States households, or about 16.7% of the population. The Census Bureau does

not release the data collected in the decennial census in a way that could identify

respondents; individual responses to select questions are available with geographic

identifiers broad enough that individual respondents cannot be identified. The mi-

crodata release used in this study is the 2000 1% Public Use Microdata Sample,

or 1% PUMS release. This sample includes observations of 2,818,644 individuals in

403 “Place of Work Public Use Microdata Areas”, or POWPUMAs.

The primary virtue of this data set is its size; my empirical strategy calls for sep-

arate regressions for black and white populations in each geographic labor market

in the country to measure local levels of discrimination, a very observation-intensive

task. The 1% Public Use Microdata Sample reports each respondent’s POWPUMA.

For respondents who live outside a Census Bureau-defined Metropolitan Statistical

Area, this POWPUMA corresponds to the super-PUMA (Public Use Microdata

Area) in which the respondent works, the smallest geographical area reported in the

1% Public Use Microdata Sample for place of work. For respondents in Metropoli-

tan Statistical Areas, the POWPUMA corresponds to a common code for all Public

Use Microdata Areas included in the Metropolitan Statistical Area. Not all POW-

PUMAs, however, contain enough black and white workers to meaningfully conduct

a Oaxaca decomposition of black-white wage inequality for that locality2 . Thus,

2
The 5% Public Use Microdata Sample would provide a larger sample size. The 5% sample,

78
POWPUMA’s with fewer than 30 full-time black workers were omitted from the

study.

4.4.2 Sample Selection

The study of how wage discrimination against black non-managerial workers

affects the wages of their managers and white coworkers uses four distinct samples:

a sample of non-managerial white workers to estimate the effect of discrimination

on their wages, a sample of managers to estimate the effect of discrimination on

their wages, and samples of black and white non-managerial workers to measure the

unexplained wage gap.

In the case of white non-managerial workers, using the same population to es-

timate the unexplained wage gap and its effect on their wages leads to statistical

bias. If Equation (4.14) determines the wages w for a white worker i,

ln wi = Wi = Xi BW + i (4.14)

and (4.15) represents a Oaxaca decomposition,

GU =(W̄W − W̄B ) − (X̄W − X̄B )BW (4.15)

NW
1 X
= (Wi − Xi BW ) − W̄B + X̄B BW
NW i=1

then (4.16) shows that the unexplained wage gap GU includes the error term i of
however, lacks some variables that could identify the respondent within that more specific geo-
graphical area. In particular, the 5% sample only reports “place of work” at the state level, and
no finer. Analysis by place of work is key to this study and the primary reason for selecting a
Census data set over, for instance, the Current Population Survey, which collects variables includ-
ing unionization status and some information about non-cash benefits that would be helpful for
this study. The main reason to use a Census data set over a CPS data set is that the Census
has enough observations to meaningfully measure levels of discrimination in localities smaller than
the state level. While the 5% sample includes more than enough observations for such small-scale
geographic analysis, the fact that it does not report place of work at a finer level than the state
negates any reason for selecting that sample.

79
each individual white worker.
NW
U 1 X
G = i − W̄B + X̄B BW (4.16)
NW i=1

This means that if in a subsequent regression GU were used as a explanatory term,

this would violate the regression assumption that the explanatory variables and the

error term are uncorrelated.

To avoid this problem, I use a split sample: one sample of white non-managerial

workers in the Oaxaca decomposition to measure the unexplained race wage gap

in each locale, and a separate sample in the final regression to calculate the effect

of the estimated unexplained race wage gap on the wages of white non-managerial

workers.

Lettau (1997) found that similar part and full time workers have large wage

differentials for identical jobs. Examining the effect of part time workers would be

an important extension to this research, but this study excludes workers working

less than 30 hours a week from the sample.

In each discrimination model, the profit maximization motive plays an important

role in predicting firm behavior. For this reason, I exclude military and government

workers (those with Census industry codes over 936, corresponding to NAICS code

92).

4.5 Variables in the Wage Regressions

4.5.1 Controlling for Non-discriminatory Wage Differences

In both the generic and specific tests, on both the wages of managers and white

non-managers, the regression will need to separate out the effect of “discrimination”

from the effect of “everything else,” as discussed in Section 4.2. In an extensive

review of empirical discrimination literature, Altonji and Blank (1999, p. 3157)

80
report that the standard “textbook” decomposition should control for education,

experience, region, occupation, industry, and whether the job is in the public sector

or part time. I control for a number of these characteristics via sample selection: this

study considers managerial and non-managerial occupations in separate samples,

and excludes part-time and public sector workers.

This study also controls for a fourth-degree polynomial of the respondent’s age,

educational attainment dummy variables, race (manager regression only), sex, sec-

tor of employment, the percentage of workers in the POWPUMA with varying levels

of education, the unemployment rate for people who live within the geographical

territory of the POWPUMA, the average rental rate for individuals who live in

rented housing within the geographical territory of POWPUMA, and the percent-

age of workers in each industrial sector in the POWPUMA. See the Appendix to

Chapter 3 for a more thorough discussion of the variables used to control for non-

discriminatory wage differences.

4.5.2 Specific Test

The two data called for by the model of the impact of race discrimination on the

wages of white workers are r̃j , the fraction of the potential non-managerial workforce

that is black, and Gj , the wage gap between blacks and whites due to discrimination.

For the case of r̃j , this study measures this by counting the number of individuals

in a POWPUMA who are not employed as managers and report that they are black

and divide this by the number of individuals in a POWPUMA who are not employed

as managers and identify themselves as either black or white (for this measurement,

individuals who are neither black nor white are excluded from the population).

As a proxy for the gap between white and black wages due to race discrimination,

this study uses the gap between white and black wages unexplained by observable

81
factors hypothesized to be separate from current labor market discrimination. This

procedure is explained in detail in Section 2.3.2. Figure 4.1 shows this variable in

the various POWPUMAs.

The model of the impact of race discrimination on the wages of managers calls

for several data. As was the case for white non-managers, it requires a proxy for

the discriminatory race wage gap, Gj and the fraction of black workers, r̃j . It also

calls for mj and nj , the raw numbers of managers and non-managers, respectively,

in a POWPUMA, and νj , the number of non-managers managed by each manager.

nj
The variables mj and nj are simple counts, while νj = mj
.

Finally, the generic reduced-form tests only examine the impact of Gj , the dis-

criminatory race wage gap, which employs the same proxy as above.

4.6 Results
4.6.1 Specific Test on Wages of White Non-managers

Regressing the wages of white non-managers on the determinants suggested

above by the structure of the model from Chapter 3 and other standard deter-

minants of wages (see Chapter 2, but include a dummy variable for sex instead of

race, as the worker sample in Chapter 2 was all male but of multiple races, but

this sample includes both males and females, but all members are white) yields the

results shown in Table 4.1.

Hypothesis 4.1 predicts that the coefficient for x0 should be 1; it was estimated

as 1.089. The hypothesis that x0 = 1 cannot be rejected. This bodes well for

the model. The model’s predictions about the wages of white non-managers are

consistent with these results.

82
 TABLE 4.1

STRUCTURALLY SUGGESTED DETERMINANTS

OF LOG WAGES

OF FULL-TIME WHITE NON-MANAGERS,

UNITED STATES, 2000

Variable Coefficient (Std. Err.)

x0 1.089∗∗ (0.219)

Age 0.145∗∗ (0.007)

Age2 -0.003∗∗ (0.000)

Age3 0.000∗∗ (0.000)

Age4 0.000∗∗ (0.000)

Some HS 0.122∗∗ (0.012)

HS Diploma 0.249∗∗ (0.014)

Some College 0.365∗∗ (0.016)

College Diploma 0.635∗∗ (0.016)

Graduate Degree 0.859∗∗ (0.015)

Female -0.249∗∗ (0.004)

Mining 0.415∗∗ (0.019)

Utilities 0.483∗∗ (0.016)

Construction 0.246∗∗ (0.014)

Manufacturing 0.302∗∗ (0.016)

Wholesale Trade 0.243∗∗ (0.016)

84
 TABLE 4.1 (CONT.)

Variable Coefficient (Std. Err.)

Retail Trade 0.101∗∗ (0.014)

Transportation and Warehousing 0.246∗∗ (0.014)

Information 0.329∗∗ (0.017)

Finance and Insurance 0.370∗∗ (0.016)

Real Estate and Rental and Leasing 0.202∗∗ (0.018)

Professional, Scientific,
and Technical Services 0.335∗∗ (0.015)

Management of Companies
and Enterprises 0.450∗∗ (0.037)

Administrative and Support

and Waste Management 0.062∗∗ (0.015)

Education 0.140∗∗ (0.015)

Health Care and Social Assistance 0.228∗∗ (0.014)

Arts and Entertainment 0.040† (0.021)

Accommodation and Food Services -0.067∗∗ (0.015)

Other Services 0.004 (0.015)

Locale Some HS -0.574 (0.352)

Locale HS Diploma -0.079 (0.193)

Locale Some College -0.110 (0.216)

Locale College Diploma -0.156 (0.251)

Locale Graduate Degree 0.170 (0.392)

85
 TABLE 4.1 (CONT.)

Variable Coefficient (Std. Err.)

Locale Unemployment Rate 1.234∗∗ (0.203)

Locale Average Rental Cost 0.000∗∗ (0.000)

Locale Mining 0.342 (0.334)

Locale Utilities 0.476 (0.529)

Locale Construction 0.764∗ (0.312)

Locale Manufacturing 0.746∗∗ (0.187)

Locale Wholesale Trade 0.963∗ (0.467)

Locale Retail Trade -0.174 (0.297)

Locale Transportation and Warehousing 0.703∗∗ (0.234)

Locale Information 0.077 (0.370)

Locale Finance and Insurance 1.367∗∗ (0.251)

Locale Real Estate and

Rental and Leasing 0.166 (0.763)

Locale Professional, Scientific,

and Technical Services 0.792∗ (0.340)

Locale Management of Companies and

Enterprises 7.390 (5.481)

Locale Administrative and Support

and Waste Management -0.021 (0.455)

Locale Education -0.095 (0.310)

Locale Health Care and Social Assistance 0.376 (0.257)

Locale Arts and Entertainment 1.578∗∗ (0.335)

86
 TABLE 4.1 (CONT.)

Variable Coefficient (Std. Err.)

Locale Accommodation and Food Services -0.242 (0.378)

Locale Other Services 1.052∗∗ (0.402)

Locale Government 0.398† (0.215)

Intercept -0.903∗∗ (0.210)

N 361353
R2 0.305
Significance levels : † : 10% ∗ : 5% ∗∗ : 1%

4.6.2 Specific Test on Wages of Managers

Finding initial signs of encouragement, the study continues to the next set of

hypotheses. Regressing the wages of managers on the determinants suggested above

by the the structure of the model from Chapter 3 yields the results shown in Table

4.2; the results are consistent with some of the hypotheses but reject others.

TABLE 4.2

STRUCTURALLY SUGGESTED DETERMINANTS

OF LOG WAGES OF FULL-TIME MANAGERS,

UNITED STATES, 2000

Variable Coefficient (Std. Err.)

x1 -0.044 (0.027)

x2 1.383∗ (0.586)

87
 TABLE 4.2 (CONT.)

Variable Coefficient (Std. Err.)

x3 -3.341∗ (1.669)

Age 0.131∗∗ (0.018)

Age2 -0.002∗∗ (0.001)

Age3 0.000 (0.000)

Age4 0.000 (0.000)

Some HS 0.041 (0.029)

HS Diploma 0.134∗∗ (0.028)

Some College 0.260∗∗ (0.029)

College Diploma 0.510∗∗ (0.029)

Graduate Degree 0.654∗∗ (0.029)

Female -0.234∗∗ (0.005)

Black -0.107∗∗ (0.011)

Mining 0.757∗∗ (0.058)

Utilities 0.697∗∗ (0.029)

Construction 0.544∗∗ (0.021)

Manufacturing 0.669∗∗ (0.022)

Wholesale Trade 0.584∗∗ (0.023)

Retail Trade 0.505∗∗ (0.022)

Transportation and Warehousing 0.517∗∗ (0.024)

Information 0.641∗∗ (0.023)

88
 TABLE 4.2 (CONT.)

Variable Coefficient (Std. Err.)

Finance and Insurance 0.683∗∗ (0.023)

Real Estate and Rental and Leasing 0.418∗∗ (0.026)

Professional, scientific,
and Technical Services 0.636∗∗ (0.023)

Management of Companies
and Enterprises 0.724∗∗ (0.048)

Administrative and Support

and Waste Management 0.490∗∗ (0.024)

Education 0.376∗∗ (0.022)

Health Care and Social Assistance 0.447∗∗ (0.022)

Arts and Entertainment 0.371∗∗ (0.030)

Accommodation and Food Services 0.215∗∗ (0.020)

Other Services 0.355∗∗ (0.023)

Locale Some HS -0.106 (0.452)

Locale HS Diploma -0.147 (0.267)

Locale Some College -0.197 (0.291)

Locale College Diploma -0.061 (0.337)

Locale Graduate Degree 0.741† (0.448)

Locale Unemployment Rate 1.120∗∗ (0.316)

Locale Average Rental Cost 0.000∗∗ (0.000)

Locale Mining 0.149 (0.386)

Locale Utilities -1.007 (0.861)

89
 TABLE 4.2 (CONT.)

Variable Coefficient (Std. Err.)

Locale Construction 0.675† (0.395)

Locale Manufacturing 0.453† (0.250)

Locale Wholesale Trade 0.486 (0.534)

Locale Retail Trade 0.002 (0.421)

Locale Transportation and Warehousing 0.627∗ (0.278)

Locale Information 0.124 (0.533)

Locale Finance and Insurance 1.043∗∗ (0.285)

Locale Real Estate and Rental

and Leasing 0.340 (0.934)

Locale Professional, Scientific,

and Technical Services 0.440 (0.389)

Locale Management of Companies

and Enterprises 17.213∗ (7.779)

Locale Administrative and Support

and Waste Management 2.007∗∗ (0.558)

Locale Education -0.404 (0.344)

Locale Health Care and Social Assistance -0.024 (0.341)

Locale Arts and Entertainment 1.675∗∗ (0.424)

Locale Accommodation and Food Services -1.104∗ (0.444)

Locale Other Services 0.438 (0.552)

Locale Government 0.140 (0.270)

Intercept -0.789∗∗ (0.304)

90
 TABLE 4.2 (CONT.)

Variable Coefficient (Std. Err.)

N 97889
R2 0.28
F (58,337) 534.935
Significance levels : † : 10% ∗ : 5% ∗∗ : 1%

Hypothesis 4.2 predicts that the coefficient for x1 is negative. The coefficient

does appear to be negatively signed, but the data do not allow a claim with 90%

confidence or better that this is the case.

Hypothesis 4.3 predicts that the coefficient for x2 is also negative. It appears

to be the case that one can say with 95% confidence that the opposite is true: the

coefficient for x2 is positive.

Finally, the test of non-linear Hypothesis 4.4 is clearly rejected. Thus, the model

does not fit.

4.6.3 Generic Test on Wages of Managers

If the specific model proposed in Chapter 3 does not fit. The question remains,

however, whether the problem is specific to the functional form imposed in Chapter

2, or whether the data rule out the basic premise of the model that manager wages

fall in the presence of greater discrimination. Regressing the wages of managers

on the reduced-form determinants discussed above yields the opposite result from

the model’s prediction: race discrimination seems to raise managers’ wages. This is

completely inconsistent with a model in which managers pay for discrimination.

91
 TABLE 4.3

REDUCED FORM DETERMINANTS OF LOG WAGES

OF FULL-TIME MANAGERS,

UNITED STATES, 2000

Variable Coefficient (Std. Err.)

G 0.197∗∗ (0.051)

Age 0.131∗∗ (0.018)

Age2 -0.002∗∗ (0.001)

Age3 0.000 (0.000)

Age4 0.000 (0.000)

Some HS 0.041 (0.029)

HS Diploma 0.134∗∗ (0.028)

Some College 0.260∗∗ (0.029)

College Diploma 0.510∗∗ (0.029)

Graduate Degree 0.655∗∗ (0.028)

Female -0.234∗∗ (0.005)

Black -0.106∗∗ (0.011)

Mining 0.758∗∗ (0.058)

Utilities 0.696∗∗ (0.029)

Construction 0.544∗∗ (0.021)

Manufacturing 0.669∗∗ (0.022)

92
 TABLE 4.3 (CONT.)

Variable Coefficient (Std. Err.)

Wholesale Trade 0.585∗∗ (0.023)

Retail Trade 0.505∗∗ (0.022)

Transportation and Warehousing 0.517∗∗ (0.024)

Information 0.641∗∗ (0.023)

Finance and Insurance 0.683∗∗ (0.023)

Real Estate and Rental and Leasing 0.418∗∗ (0.026)

Professional, Scientific, and

Technical Services 0.636∗∗ (0.023)

Management of Companies and

Enterprises 0.724∗∗ (0.048)

Administrative and Support

and Waste Management 0.490∗∗ (0.024)

Education 0.377∗∗ (0.022)

Health Care and Social Assistance 0.448∗∗ (0.022)

Arts and Entertainment 0.372∗∗ (0.030)

Accommodation and Food Services 0.216∗∗ (0.020)

Other Services 0.355∗∗ (0.023)

Locale Some HS -0.027 (0.421)

Locale HS Diploma -0.201 (0.264)

Locale Some College -0.213 (0.287)

Locale College Diploma -0.131 (0.336)

93
 TABLE 4.3 (CONT.)

Variable Coefficient (Std. Err.)

Locale Graduate Degree 0.748† (0.440)

Locale Unemployment Rate 1.106∗∗ (0.296)

Locale Average Rental Cost 0.000∗∗ (0.000)

Locale Mining 0.102 (0.410)

Locale Utilities -0.881 (0.853)

Locale Construction 0.566 (0.404)

Locale Manufacturing 0.420 (0.263)

Locale Wholesale Trade 0.392 (0.542)

Locale Retail Trade 0.076 (0.431)

Locale Transportation and Warehousing 0.623∗ (0.290)

Locale Information 0.061 (0.555)

Locale Finance and Insurance 1.071∗∗ (0.299)

Locale Real Estate and Rental

and Leasing 0.206 (0.929)

Locale Professional, Scientific, and

Technical Services 0.407 (0.403)

Locale Management of Companies and

Enterprises 19.311∗ (8.149)

Locale Administrative and Support

and Waste Management 1.896∗∗ (0.564)

Locale Education -0.415 (0.365)

94
 TABLE 4.3 (CONT.)

Variable Coefficient (Std. Err.)

Locale Health Care and Social Assistance -0.035 (0.354)

Locale Arts and Entertainment 1.734∗∗ (0.436)

Locale Accommodation and Food Services -1.198∗∗ (0.457)

Locale Other Services 0.353 (0.559)

Locale Government 0.135 (0.283)

Intercept -0.744∗ (0.305)

N 97889
R2 0.28
F (56,337) 546.648
Significance levels : † : 10% ∗ : 5% ∗∗ : 1%

Hypothesis 4.6 predicts that the coefficient on G should be negative. These

results (see Table 4.3) do not indicate that this is the case; rather, they suggest

that one can be 99% confident that the effect of G on managers’ wages is actually

positive.

4.7 Interpretation

The results of the analysis of race discrimination support neither the specific

nor the generic form of the model. Rogue managers acting in violation of company

policy that effectively penalizes them for the imposition of their tastes in hiring

decisions do not seem to cause persistent race discrimination. The results are more

consistent with a model in which managers are rewarded for discriminating, not one

95
in which they are penalized. The empirical evidence does not support the claims

cited in Section 1.2 by corporate spokespeople about race discrimination. It could

be the case that the few firms in question who made these claims do, in fact, penalize

managers for discriminating, but many other firms do not, and the average effect

masks the actions taken by the firms claiming to enforce anti-discrimination policies,

but in the absence of firm-level data, this cannot be tested.

96
 EPILOGUE

Manager pay falls where the unexplained wage gap between males and females

increases. This is a remarkable finding that flies in the face of existing discrimination

theory. The theory developed in Chapter 3 begins to offer an explanation. Even

it, however, is not fully consistent with the empirical findings: the wages of male

non-managers, for instance, were predicted to rise where the unexplained wage gap

rises, and instead they fell. The results observed do not confirm the model, but do

suggest that further exploration of this model might be in order. Future work should

introduce effects of integration on productivity and determinants of the propensity

for market/non-market work to help explain the behavior of wages of non-managers.

While the specific model of rogue manager discrimination in from Chapter 3 might

not fit, the data indicate that it may well be on the right track toward explaining

the negative correlation between manager pay and the male-female wage gap.

While a model of manager taste discrimination may explain some of wage dis-

crimination based on sex, another theory, perhaps statistical discrimination, can

explain race wage discrimination. These findings on race wage discrimination are

inconsistent with a model that rogue managers produce the wage gap by gratifying

their own tastes in opposition to corporate interests. Corporate claims that racial

diversity in hiring is rewarded and discrimination punished do not seem consistent

with the aggregated data from the Census. Further research with firm-level data or,

better still, firm-level data with hierarchy information (who reports to whom, who

hires whom, etc.) would be ideal to examine these claims further.

97
 Future work also needs to develop a coherent explanation for why race and

sex discrimination behave so dramatically differently with regard to their effect on

the pay of managers. As the unexplained race wage gap increases among non-

managers, manager pay rises, but as the unexplained sex wage gap increases among

non-managers, manager pay falls. If race but not sex conveys valuable productivity

information, this could help explain the divergence. Further research needs to look

at the value of the information contained in the race index and the sex index.

The unexplained race and sex wage gaps have significant impacts on the wages

of mangers. This study developed an initial model in an attempt to explain some

of this impact; further work needs to continue on the subject to find a model that

explains the fall in male wages in areas where the wage gap is larger.

The data in this study do not examine any particular company. Thus, any gen-

eral findings about race or sex discrimination are not at this point directly relevant

to the ongoing legal cases discussed in Section 1.1. Any particular company could

be an exception to the trends observed. Nonetheless, the findings of this study do

have some relevance to the believability of the story the companies tell in response

to allegations of discrimination.

In the race discrimination cases of Microsoft, Sodexho, and Friedman’s, the data

surveyed do not support their scenario that discriminating managers are punished

financially. Manager pay is higher where the unexplained black-white wage gap is

higher, textitceteris paribus. It could be the case that a particular company does

link manager pay to racial diversity in hiring, but this study finds that this does

not happen on average. A model in which shareholders find racial discrimination

bad for profits, and rogue managers who perpetuate this practice pay for it does

not match the aggregated findings. If firm-specific data were available, it could be

used to make the case that a particular firm does or does not fit this model, but the

98
present data give no indication that the model holds.

In the sex discrimination cases of Boeing and Wal-Mart, the impact of these

findings is murkier. Manager wages are significantly lower where the male-female

pay gap is larger. This is consistent with a model in which rogue managers pay for

discrimination. This study does not have the data to conclude whether firms benefit

from the wage reductions to discriminating managers or if the increased costs from

discrimination offset the managers’ wage reduction. Firms may or may not have an

economic incentive to encourage a taste for discrimination among their managers;

future research with sub-firm-level data and localized profit rates should attempt

to answer this question to help determine whether anti-discrimination campaigns

should target individual managers’ tastes for hiring female workers or need to target

firms.

99
 APPENDIX A

TAYLOR APPROXIMATION

If there exists no discrimination against women, there would be no corresponding

impact of that discrimination on men.

E(0) = 0 (A.1)

Without knowing the exact form of the function E(x), one can approximate it

using a Taylor series expansion around this known point.

In such an expansion the function E(x) is expressed as a series of derivatives

calculated at a known point x0 and distances from that point.

∞
X
E(x) = E i (x0 )(x − x0 )i (A.2)
i=0

In Equation (A.2), E (x0 ) represents the ith derivative of the function E(x)
i

calculated at the point x0 . Furthermore, for some finite n,

n
X
E(x) ≈ E i (x0 )(x − x0 )i (A.3)
i=0

and

n
X
lim E i (x0 )(x − x0 )i = E(x) (A.4)
n→∞
i=0

100
 A two-term Taylor expansion around known point 0 would yield the approxima-

tion

∞
X
E(D) = E i (0)(D − 0)i ≈ E(0)(D − 0)0 + E 0 (0)(D − 0) = µD (A.5)
i=0

where µ is a scalar multiple equal to the value of E 0 (0).

101
 APPENDIX B

VARIABLES IN THE SEX DISCRIMINATION WAGE REGRESSIONS

TABLE B.1

MEAN VALUES OF CHARACTERISTICS,

MALE NON-MANAGERIAL WORKERS,

UNITED STATES, 2000

Variable Mean Std. Dev. Min. Max.

Locale 27881 15815 1100 56100

Log Wage 2.637 0.739 0.004 8.515

Age 39.804 12.89 16 93

Age2 1750.526 1102.878 256 8649

Age3 83648 78886 4096 804357

Age4 4280678 5519589 65536 74805201

Black 0.093 0.29 0 1

Some HS 0.132 0.338 0 1

HS Diploma 0.312 0.463 0 1

102
 TABLE B.1 (CONT.)

Variable Mean Std. Dev. Min. Max.

Some College 0.279 0.449 0 1

College Diploma 0.139 0.346 0 1

Graduate Degree 0.081 0.273 0 1

Agriculture, Forestry,
Fishing and Hunting 0.021 0.142 0 1

Mining 0.009 0.093 0 1

Utilities 0.015 0.123 0 1

Construction 0.138 0.345 0 1

Manufacturing 0.211 0.408 0 1

Wholesale Trade 0.051 0.22 0 1

Retail Trade 0.113 0.317 0 1

Transportation
and Warehousing 0.068 0.251 0 1

Information 0.028 0.166 0 1

Finance and
Insurance 0.031 0.174 0 1

Real Estate and

Rental and Leasing 0.016 0.125 0 1

Professional,
Scientific, and
Technical Services 0.056 0.23 0 1

Management of Companies
and Enterprises 0 0.018 0 1

103
 TABLE B.1 (CONT.)

Variable Mean Std. Dev. Min. Max.

Administrative and
Support and Waste
Management 0.04 0.196 0 1

Education 0.049 0.215 0 1

Health Care and

Social Assistance 0.043 0.203 0 1

Arts and Entertainment 0.019 0.137 0 1

Accommodation and
Food Services 0.043 0.203 0 1

Other Services 0.049 0.215 0 1

Locale Some HS 0.109 0.03 0.045 0.211

Locale HS Diploma 0.292 0.076 0.126 0.516

Locale Some College 0.305 0.039 0.197 0.419

Locale College Diploma 0.162 0.052 0.068 0.327

Locale Graduate Degree 0.09 0.037 0.038 0.259

Locale Agriculture,
Forestry, Fishing
and Hunting 0.021 0.03 0 0.266

Locale Mining 0.005 0.012 0 0.106

Locale Utilities 0.011 0.005 0.001 0.056

Locale Construction 0.083 0.018 0.036 0.136

Locale Manufacturing 0.172 0.076 0.02 0.513

Locale Wholesale Trade 0.04 0.011 0.008 0.083

104
 TABLE B.1 (CONT.)

Variable Mean Std. Dev. Min. Max.

Locale Retail Trade 0.113 0.017 0.046 0.177

Locale Transportation
and Warehousing 0.047 0.017 0.016 0.139

Locale Information 0.03 0.017 0.007 0.092

Locale Finance and

Insurance 0.051 0.026 0.016 0.17

Locale Real Estate

and Rental and Leasing 0.018 0.007 0.004 0.04

Locale Professional,
Scientific, and
Technical Services 0.057 0.033 0.01 0.219

Locale Management
of Companies and
Enterprises 0.001 0.001 0 0.006

Locale Administrative
and Support and Waste
Management 0.036 0.01 0.015 0.065

Locale Education 0.084 0.02 0.043 0.177

Locale Health Care

and Social Assistance 0.111 0.023 0.055 0.259

Locale Arts and

Entertainment 0.018 0.015 0.004 0.142

Locale Accommodation
and Food Services 0.056 0.017 0.031 0.183

Locale Other Services 0.047 0.009 0.025 0.12

105
 TABLE B.1 (CONT.)

Variable Mean Std. Dev. Min. Max.

Locale Unemployment
Rate 0.058 0.021 0.023 0.195

Locale Average
Rental Cost 542.732 176.625 249.715 1178.959

Locale Unexplained
Sex Wage Gap 0.233 0.022 0.153 0.298

N = 293323

TABLE B.2

MEAN VALUES OF CHARACTERISTICS,

MANAGERS, UNITED STATES, 2000

Variable Mean Std. Dev. Min. Max.

Locale 27550 15792 1100 56100

Log Wage 2.975 0.824 0.002 9.018

Age 43.53 11.689 16 93

Age2 2031.528 1080.761 256 8649

Age3 100873 82083 4096 804357

Age4 5295138 6022451 65536 74805201

Black 0.054 0.226 0 1

106
 TABLE B.2 (CONT.)

Variable Mean Std. Dev. Min. Max.

Some HS 0.04 0.197 0 1

HS Diploma 0.183 0.387 0 1

Some College 0.308 0.462 0 1

College Diploma 0.296 0.456 0 1

Graduate Degree 0.159 0.365 0 1

Female 0.35 0.477 0 1

Agriculture, Forestry,
Fishing and Hunting 0.08 0.271 0 1

Mining 0.004 0.066 0 1

Utilities 0.01 0.101 0 1

Construction 0.082 0.274 0 1

Manufacturing 0.155 0.362 0 1

Wholesale Trade 0.038 0.19 0 1

Retail Trade 0.053 0.225 0 1

Transportation and
Warehousing 0.031 0.173 0 1

Information 0.049 0.216 0 1

Finance and Insurance 0.081 0.274 0 1

Real Estate and Rental

and Leasing 0.041 0.198 0 1

Professional, Scientific,
and Technical Services 0.071 0.257 0 1

107
 TABLE B.2 (CONT.)

Variable Mean Std. Dev. Min. Max.

Management of Companies
and Enterprises 0.002 0.039 0 1

Administrative and Support

and Waste Management 0.026 0.159 0 1

Education 0.069 0.254 0 1

Health Care and Social

Assistance 0.074 0.262 0 1

Arts and Entertainment 0.014 0.116 0 1

Accommodation and
Food Services 0.083 0.276 0 1

Other Services 0.037 0.189 0 1

Locale Some HS 0.105 0.029 0.045 0.211

Locale HS Diploma 0.281 0.075 0.126 0.516

Locale Some College 0.305 0.039 0.197 0.419

Locale College Diploma 0.171 0.055 0.068 0.327

Locale Graduate Degree 0.095 0.041 0.038 0.259

Locale Agriculture,
Forestry, Fishing and Hunting 0.02 0.031 0 0.266

Locale Mining 0.005 0.011 0 0.106

Locale Utilities 0.01 0.005 0.001 0.056

Locale Construction 0.082 0.018 0.036 0.136

Locale Manufacturing 0.165 0.074 0.02 0.513

108
 TABLE B.2 (CONT.)

Variable Mean Std. Dev. Min. Max.

Locale Wholesale Trade 0.04 0.011 0.008 0.083

Locale Retail Trade 0.111 0.018 0.046 0.177

Locale Transportation
and Warehousing 0.047 0.017 0.016 0.139

Locale Information 0.033 0.018 0.007 0.092

Locale Finance and Insurance 0.054 0.028 0.016 0.17

Locale Real Estate

and Rental and Leasing 0.019 0.007 0.004 0.04

Locale Professional,
Scientific, and Technical
Services 0.063 0.037 0.01 0.219

Locale Management of
Companies and Enterprises 0.001 0.001 0 0.006

Locale Administrative
and Support and Waste
Management 0.037 0.01 0.015 0.065

Locale Education 0.083 0.02 0.043 0.177

Locale Health Care and

Social Assistance 0.109 0.023 0.055 0.259

Locale Arts and

Entertainment 0.018 0.014 0.004 0.142

Locale Accommodation and

Food Services 0.055 0.016 0.031 0.183

Locale Other Services 0.047 0.009 0.025 0.12

109
 TABLE B.2 (CONT.)

Variable Mean Std. Dev. Min. Max.

Locale Unemployment
Rate 0.055 0.02 0.023 0.195

Locale Average Rental Cost 568.596 186.282 249.715 1178.959

Locale Unexplained Sex

Wage Gap 0.234 0.023 0.153 0.298

N = 111989

TABLE B.3

PAIR-WISE CORRELATION BETWEEN

CHARACTERISTICS OF NON-MANAGERIAL MALE WORKERS

AND THEIR WAGES AND THE UNEXPLAINED SEX WAGE GAP

IN THEIR LOCALITY, UNITED STATES, 2000

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

Log Wage 1 0.0405

Age 0.2726 -0.011

Black -0.0797 -0.0251

Some HS -0.1815 -0.0148

110
 TABLE B.3 (CONT.)

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

HS Diploma -0.1026 -0.0518

Some College 0.0014 -0.0121

College Diploma 0.1918 0.0426

Graduate Degree 0.2603 0.0324

Agriculture, Forestry,
Fishing and Hunting -0.0943 -0.0179

Mining 0.0154 -0.013

Utilities 0.0547 -0.0035

Construction -0.0383 -0.0162

Manufacturing 0.0399 -0.0366

Wholesale Trade 0.0052 0.0073

Retail Trade -0.0782 0.0008

Transportation
and Warehousing 0.0097 -0.0032

Information 0.0565 0.0347

Finance and
Insurance 0.1065 0.0311

Real Estate And Rental

And Leasing 0.0025 0.0047

Professional,
Scientific, and
Technical Services 0.1609 0.0397

111
 TABLE B.3 (CONT.)

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

Management of Companies
and Enterprises 0.0131 0.0002

Administrative and
Support and Waste -0.0818 0.0037

Education 0.0432 0.0048

Health Care And

Social Assistance 0.0787 0.0131

Arts And Entertainment -0.0396 -0.0041

Food Services -0.1551 -0.0033

Other Services -0.0669 -0.002

Locale Some HS -0.1322 -0.1142

Locale HS Diploma -0.1134 -0.3019

Locale Some College -0.0207 -0.1777

Locale College Diploma 0.1654 0.2949

Locale Graduate Degree 0.1602 0.2626

Locale Agriculture,
Forestry, Fishing
and Hunting -0.1117 -0.0779

Locale Mining -0.0657 -0.051

Locale Utilities -0.0762 -0.0628

Locale Construction -0.0907 -0.182

112
 TABLE B.3 (CONT.)

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

Locale Manufacturing -0.025 -0.0969

Locale Wholesale Trade 0.0342 0.1532

Locale Retail Trade -0.0867 -0.1714

Locale Transportation
and Warehousing 0.0221 -0.0509

Locale Information 0.1317 0.3304

Locale Finance and

Insurance 0.1196 0.2067

Locale Real Estate

And Rental And Leasing 0.079 0.0908

Locale Professional,
Scientific, and
Technical Services 0.1574 0.2688

Locale Management
of Companies and
Enterprises 0.0725 0.062

Locale Administrative
and Support and Waste
Management 0.08 0.0651

Locale Education -0.0778 -0.0053

Locale Health Care

And Social Assistance -0.0437 -0.034

Locale Arts And

Entertainment 0.0037 -0.0339

113
 TABLE B.3 (CONT.)

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

Locale Food Services -0.0414 -0.2293

Locale Other Services -0.0129 -0.012

Locale Unemployment Rate -0.0828 -0.0338

Locale Average
Rental Cost 0.1579 0.2968

TABLE B.4

PAIR-WISE CORRELATION BETWEEN

CHARACTERISTICS OF MANAGERS AND THEIR WAGES

AND THE UNEXPLAINED SEX WAGE GAP

IN THEIR LOCALITY, UNITED STATES, 2000

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

Log Wage 1 0.0797

Age 0.1643 -0.0232

Black -0.0383 -0.0053

Female -0.1528 0.0085

114
 TABLE B.4 (CONT.)

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

Some HS -0.118 -0.0092

HS Diploma -0.1948 -0.0513

Some College -0.1069 -0.0225

College Diploma 0.1635 0.0369

Graduate Degree 0.2275 0.0423

Agriculture, Forestry,
Fishing and Hunting -0.251 -0.056

Mining 0.0248 -0.005

Utilities 0.0397 -0.0081

Construction 0.02 -0.0167

Manufacturing 0.1523 0.0182

Wholesale Trade 0.0371 0.0003

Retail Trade -0.0122 0.002

Transportation
and Warehousing 0.0077 -0.0098

Information 0.0637 0.0442

Finance and
Insurance 0.08 0.0251

Real Estate And Rental

And Leasing -0.0281 0.0002

115
 TABLE B.4 (CONT.)

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

Professional,
Scientific, and
Technical Services 0.1022 0.0554

Management of Companies
and Enterprises 0.0191 0.0017

Administrative and
Support and Waste -0.0074 0.0011

Education 0.0112 -0.0087

Health Care And Social

Assistance -0.0247 -0.0129

Arts And Entertainment -0.0228 -0.0052

Food Services -0.1737 -0.0262

Other Services -0.0298 -0.0109

Locale Some HS -0.1114 -0.1503

Locale HS Diploma -0.2189 -0.3063

Locale Some College -0.0906 -0.2146

Locale College Diploma 0.2333 0.3119

Locale Graduate Degree 0.2292 0.2837

Locale Agriculture,
Forestry, Fishing
and Hunting -0.2047 -0.0959

Locale Mining -0.099 -0.0515

116
 TABLE B.4 (CONT.)

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

Locale Utilities -0.1316 -0.0819

Locale Construction -0.1288 -0.1918

Locale Manufacturing -0.0641 -0.0774

Locale Wholesale Trade 0.0776 0.1054

Locale Retail Trade -0.1084 -0.1915

Locale Transportation
and Warehousing 0.0313 -0.0991

Locale Information 0.2115 0.3396

Locale Finance and

Insurance 0.1679 0.2272

Locale Real Estate And

Rental And Leasing 0.1742 0.0789

Locale Professional,
Scientific, and
Technical Services 0.2313 0.2854

Locale Management
of Companies and
Enterprises 0.1134 0.0624

Locale Administrative
and Support and Waste
Management 0.188 0.0333

Locale Education -0.1282 -0.0273

Locale Health Care And

Social Assistance -0.0852 -0.0535

117
 TABLE B.4 (CONT.)

Correlation with
Correlation with Unexplained Sex
Variable log wage wage gap

Locale Arts And

Entertainment 0.0112 -0.0507

Locale Food Services -0.037 -0.2579

Locale Other Services 0.0092 -0.053

Locale Unemployment Rate -0.0787 -0.0833

Locale Average
Rental Cost 0.243 0.3174

Age

The use of age as a proxy for experience-related human capital is so common that

even highly-regarded studies sometimes engage in the practice without comment

or even disclosure. Finer distinctions about human capital can be developed in

detail, distinguishing between firm-specific human capital and general human capital

(Becker 1975). Other researchers have noted that time on the job does not transform

into experience-related human capital in the same linear (or polynomial) fashion for

all workers. Nevertheless, it is still common, especially when using data sets that

do not contain information about workers’ actual work experience, to use age as a

proxy for experience-related human capital, if for no other reason than the absence

of a better proxy in most data sets.

While age may not be a good proxy for firm-specific human capital, it may not

be as bad a proxy for general experience-based human capital as some might believe.

118
Different individuals and different activities likely have different rates at which indi-

viduals convert time spent at various activities into experience-based human capital.

In general, however, whether or not a person was in the labor force is far from the

only indicator of whether a person acquired experience-based human capital.

This study includes a fourth-degree polynomial of potential experience (i.e. age).

Census 2000 does not report full-time workers under age 16; any full-time worker

under age 16 would be truncated from the dataset. At age 93, ages are top coded

for anonymity. Thus, full-time workers over age 93 would be coded as age 93. Thus,

both the sample of managers and male non-managerial workers have maximum

reported values for age of 93 and minimum reported values of 16. The average age

of male non-managerial workers is 39.8, while the average age of mangers is 43.5.

This fits with most economic models in which workers tend to rise to fill managerial

roles as they age, making the average age of managerial workers higher than non-

managerial workers. As expected in a case where the variance is non-zero, the mean

of the square (and cube and quart) age terms is greater than the square (and cube

and quart) of the means of the age terms.

Education

The education variables edu1 through edu5, are dummies where edu1i is equal

to one if individual i attended some high school but did not graduate or attend col-

lege and zero otherwise (no high school attendance or high school graduate); edu2i

corresponds to graduation from high school but no college attendance; edu3i corre-

sponds to some college attendance but no bachelor’s level or higher degree; edu4i

corresponds to a bachelor’s degree but no higher degree; finally for the education

variables, edu5i corresponds to a degree higher than a bachelor’s degree.

I use a series of dummy variables to capture education in the wage regression

119
for both data and theoretical reasons. On the data side, a continuous education

variable is not available in the Census data set (nor has it been available in the

Current Population Survey since 1988). On the theoretical side, the wage effects of

education have been shown to be discontinuous at certain points, especially at points

at which degrees have been awarded (Gullason 1999). This empirically supported

“sheepskin effect” hypothesis suggests that the effects of education on wages could

be better captured by well-placed dummy variables than by a continuous time of

education variable. This is a fortunate conclusion to reach, given the constraints of

the data.

The means of the two samples behave as expected, with the sample of managers

having higher percentages of observations with higher education levels than the

sample of male non-managers. For example, 15.9% of managers hold graduate

degrees, but only 9% of male non-managers. Similarly, only 5.4% of managers

have less than a high school diploma, but 18.9% of male non-managers attained less

than a high school diploma.

Industry

The variables miningi , utiliti , constri , manufai , wholesi , retaili , transpi , informi ,

financi , realesi , profesi , managei , adminii , educati , healthi , artseni , foodsei , and

othersi are dummy variables respectively corresponding to employment in the indus-

tries of “Mining,” “Utilities,” “Construction,” “Manufacturing,” “Wholesale trade,”

“Retail trade,” “Transportation and warehousing,” “Information,” “Finance and

insurance,” “Real estate and rental and leasing,” “Professional, scientific, and tech-

nical services,” “Management of companies and enterprises,” “Administrative and

support and waste management services,” “Educational services,” “Health care and

social assistance,” “Arts, entertainment, and recreation,” “Accommodation and

120
food services,” and “Other services (except public administration)” as defined in

the NAICS. The baseline case is that individual i works in the agricultural sector.

These groupings correspond to the two-digit North American Industrial Classi-

fication System (NAICS) codes. For the 2000 Census, the Census Bureau switched

to using the NAICS from the Standard Industrial Classification (SIC) classification

system for industries that it had used from 1940 through 1990. The Office of Man-

agement and Budget of the Executive Office of the President developed the NAICS

in 1997. The NAICS system classifies firms into one of twenty industries (as opposed

to 11 divisions under the SIC codes) based on their primary activities (Bureau of

the Census, 2003). This study includes dummy variables for industry according to

these twenty sectors (The two wage regressions actually use eighteen dummy vari-

ables, as I exclude NAICS sector 92 (government and military) from the sample (see

Section 2.4.2), and the agricultural sector is the baseline group in the regression.).

The means reported in tables B.1 and B.2 for the industry variables represent the

percentage of workers in that sample working in that industry.

Local characteristics

In the absence of regional (or more finely-tuned) locality dummy variables, this

study substitutes various characteristics of the POWPUMA (see Section 2.4.1 on

page 44). The reason for this substitution is similar to the choice of a random-effects

panel data model over a fixed-effects model (and bears with it all the attendant chal-

lenges). A region, state, or PUMA dummy variable would capture all the effect of

one’s locality on wages. The purpose of these regressions, however, is to capture the

effect of the unexplained wage gap on the various left-hand variables. The discrimi-

nation variable necessarily must be a variable not unique to individuals, but rather

to some aggregated group of people. Thus, a locality dummy variable would capture

121
the effect of local discrimination along with the effects of other characteristics of the

locality. This combined effect would fail to identify the very reason for performing

the study; this is completely analogous to using a fixed effects panel data model

when the variable of interest does not change across time for an individual.

Thus, with a fixed-effect model failing to answer the primary question at hand,

this study must use a model that allows the inclusion of variables that do not change

across locales with appropriate adjustment to the error terms of the regression. It

would be a singularly bad idea to use the level of the unexplained sex wage gap

alone as the only term that varied across localities. Looking at the tables of corre-

lations (Table B.3 on page 110 and Table B.4 on page 114), one can note that many

characteristics of a locality correlate with the size of the unexplained sex wage gap.

Using pair-wise correlations, the unexplained sex wage gap is positively correlated

with the wages of managerial and male non-managerial workers. The second column

of the correlation charts, however, provide reason to think that the pair-wise cor-

relation may be misleading. Comparing correlations between other characteristics

of a POWPUMA and the unexplained sex wage gap in that POWPUMA, one finds

that almost to a characteristic, those attributes of a POWPUMA that correlate with

higher wages also correlate with a larger unexplained sex wage gap. Thus, the pair-

wise correlation could easily capture second-order effects and not the direct effect

of the unexplained sex wage gap on wages of managers and male non-managerial

workers. This fact both motivates multivariate regression to understand this sys-

tem, and highlights the need to include other characteristics of a POWPUMA in

addition to the unexplained sex wage gap.

Several theoretical models predict different effects of the local unemployment

rate on wages (and untangling the role unemployment rates play in models of sex

discrimination is not the subject of this research but should be the subject of future

122
extensions). Regardless of the predicted effect, the unemployment rate in a POW-

PUMA could have an effect on wages in that POWPUMA, so I include it in the

regression.

While the Tiebout hypothesis (Tiebout 1956) could be extended to predict that

average housing rental rates in a POWPUMA might not affect wages at all1 , both

Marxian models of the cost of reproducing labor and neoclassical models of compen-

sating wage differentials in job migration suggest that cost of living might play a role

in wage determination in a particular locale. This study includes average housing

rental rates, likely important on their own, but also as a proxy for cost of living as

a whole. The interpretation of this coefficient is not important to the study, but its

inclusion allows me to capture this aspect of a POWPUMA.

It could be the case that POWPUMAs with higher percentages of workers who

have higher levels of education might have spillover effects for all workers in that

POWPUMA. Development literature certainly suggests this is the case internation-

ally; if “better” or more desirable jobs tend to locate in areas with higher-skilled

workers, one might expect to find the percentage of workers in a POWPUMA who

have higher levels of education relates to higher wages for workers in that POW-

PUMA regardless of their individual level of education.

Finally, industrial controls for the POWPUMA also matter. The industry of

an individual is already present, but the industrial composition of a POWPUMA

should also matter. If in general, a janitor in a retail establishment earns less than

a janitor in a financial institution, it still might be the case that in a locality where

most of the jobs are in the financial sector, the retail sector might have to pay higher
1
If the Tiebout hypothesis were to hold such that cost of home rental or ownership in a com-
munity relative to other communities reflected the level of public goods relative to costs in other
communities, workers would not need any incentive from a firm to relocate to an area with higher
housing costs; greater housing costs would correspond to greater levels of public goods at price
levels such that workers would be indifferent to the exchange.

123
wages to its janitors because they have better-paying alternatives. Thus, this study

includes variables measuring what percentage of the jobs in a POWPUMA fall into

each sector, with the expectation that POWPUMAs with more jobs in higher-paying

sectors might have higher pay even for individuals not in those sectors.

Variables ledu1j through ledu5j correspond to the percentage of all full-time

workers in POWPUMA j that have the education level corresponding to edu1

through edu5.

The variables lminingj , lutilitj , lconstrj , lmanufaj , lwholesj , lretailj , ltranspj ,

linformj , lfinancj , lrealesj , lprofesj , lmanagej , ladminij , leducatj , lhealthj , lartsenj ,

lfoodsej , and lothersj correspond to the percentage of full-time private sector workers

in POWPUMA j in each of these respective industrial sectors. For any POWPUMA,

one minus the sum of these variables should equal the percentage of workers in the

agricultural sector.

124
 WORKS CITED

Aigner, Dennis J. and Glen G. Cain (1977) “Statistical theories of discrimination

in labor markets”, Industrial and Labor Relations Review, 30: 175-187.

Altonji, Joseph G. and Rebecca M. Blank (1999) “Race and Gender in the Labor
Market,” Handbook of Labor Economics, Vol. 3C.

Allport (1955) The Nature of Prejudice.

Aptheker, H. (1946)The Negro Problem in America.

Arrow, Kenneth (1973) “The Theory of Discrimination,” Discrimination in Labor

Markets, Ashenfelter and Rees, eds.

Arrow, Kenneth (1998) “What Has Economics to Say about Racial Discrimina-
tion?” The Journal of Economic Perspectives 12(2): 91-100.

Barsky, Robert B., Bound, John, Charles, Kerwin K. and Lupton, Joseph P. (2001)
“Accounting for the Black-White Wealth Gap: A Nonparametric Approach,”
National Bureau of Economic Research Working Paper 8466.

Becker, Gary S. (1957) The Economics of Discrimination.

Becker, Gary S., (1975) Human Capital: A Theoretical and Empirical Analysis
with Special Reference to Education, Second Edition.

Black, Sandra E. and Philip E. Strahan (1999) “Rent Sharing and Discrimination:
The Effects of Deregulation on the Labor Market,” Working Paper.

Blau, Francine D. and Andrea H. Beller (1988) “Trends in Earnings Differentials

by Gender (1971-81),” Industrial and Labor Relations Review 41(4): 513-529.

Blau, Francine D. and Andrea H. Beller (1992) “Black-White Earnings over the
1970s and 1980s: Gender Differences in Trends,” Review of Economics and
Statistics 74(2): 276-286.

Blau, Francine D. and Lawrence M. Kahn (1992) “Race and Gender Pay Differen-
tials,” Research Frontiers in Industrial Relations and Human Resources, David
Lewin, Olivia S. Mitchell and Peter D. Sherer, eds.

125
Blau, Francine D. and Lawrence M. Kahn (1997) “Swimming Upstream: Trends
in the Gender Wage Differential in the 1980s,” Journal of Labor Economics
15 (1 part 1): 1-42.
Blau, Francine D. and Lawrence M. Kahn (2000) “Do Cognitive Test Scores Explain
Higher U.S. Wage Inequality?” Working Paper, Cornell University.
Blau, Francine D., Marianne A. Ferber, and Anne. E. Winkler (1992) The Eco-
nomics of Women, Men, and Work.
Bowermaster, David (2004) “Boeing Sex-Bias Case to Go to Trial,” Seattle Times,
April 10.
Cain, Glen G. (1986) “The Economic Analysis of Labor Market Discrimination: A
Survey,” Handbook of Labor Economics, Vol. 1, O. Ashenfelter and R. Layard,
eds.
Cox, O. C. (1948)Caste, Class, and Race.
Dollard, J. (1937)Caste and Class in a Southern Town.
Donald, Steven and Daniel S. Hammermesh (2004) “What is Discrimination? Gen-
der in the American Economic Association,” National Bureau of Economic
Research Working Paper 10684.
Foster, Lauren (2004) “Wal-Mart ties bonuses to its image,” The Financial Times,
June 5.
Ghilarducci, Teresa (1985) “Pensions and Collective Bargaining: Toward a Na-
tional Policy on Retire Income Support,” Proceedings of 38th Annual Meetings.
Industrial Relations Research Association (IRRA): 257-259.
Ghilarducci, Teresa, and Mary Lee (2004) “Revisiting Dual Labor Markets: To-
tal Compensation and Female Labor Markets,” Working Paper, University of
Notre Dame.
Greenhouse, Steven, and Constance L. Hays (2004) “Wal-Mart Sex-Bias Suit Given
Class-Action Status,” The New York Times.
Groshen, Erica L. (1991) “The Structure of the Female/Male Wage Differential:
Is it Who You Are, What You Do, or Where You Work?” Journal of Human
Resources 26(3): 457-472.
Gullason, Edward T. (1999) “The Stability Pattern of Sheepskin Effects and Its
Implications for the Human Capital Theory-Screening Hypothesis Debate,”
Eastern Economic Journal 25(2): 141-49.
Hammermesh, Daniel S. and Jeff E. Biddle (1994) “Beauty and the Labor Market,”
American Economic Review 84(5): 1174.

126
Hellerstein, Judith K., David Neumark and Kenneth Troske (1997) “Market Forces
and Sex Discrimination,” National Bureau of Economic Research Working
Paper 6321.

Holzer, Harry and David Neumark (2000) “Assessing Affirmative Action.” Journal
of Economic Literature 38(3).

Jewlers’ Circular-Keystone (2003) “Friedman’s faces discrimination lawsuit, EEOC

investigation,” May, 174(5) p. 34.

Juhn, C., K. Murphy, and B. Pierce. (1991) “Accounting for the Slowdown in
Black-White Wage Convergence,” Workers and Their Wages, M. Kosters, ed.:
107-143.

Lazear, Edward P. (1999) “Personnel Economics: Past Lessons and Future Direc-
tions,” Journal of Labor Economics 17(2): 199-236.

Lettau, Michael K. (1997) “Compensation in Part-time Jobs Versus Full-time Jobs:

What if the Job is the Same?” Economic Letters 56(1997): 101-106.

McWilliams, C. (1948)A Mask for Privilege: Anti-Semetism in America.

Mehta, Stephanie N. (1996) “Diversity Pays: Some Companies Link Managers’

Pay to Minority Hiring,” The Wall Street Journal, April 11.

Myrdal, Gunnar (1944) The American Dilemma.

New York Times (2001) “New Bias Lawsuit Is Filed Against Microsoft,” January
4, 2001, Section C; p. 4; Col. 3.

Neal, Derek A. and William R. Johnson (1996) “The Role of Premarket Factors
in Black-White Wage Differences,” Journal of Political Economy 104(5): 869-
895.

Oaxaca, Ronald L. and Michael L. Ransom (1999) “Identification in Detailed Wage

Decompositions,” Review of Economic Statistics 81(1):154-157.

Padgett, Tania (2003) “Class Action Lawsuit Filed against Plaza Hotel Managers
in New York,”Newsday, October 1.

Phelps, Edmund S. (1972) “The Statistical Theory of Racism and Sexism,” Amer-
ican Economic Review 62: 659-661.

Reich, Michael (1984) Racial Inequality.

Rogers, W. H. (1993) “Regression Standard Errors in Clustered Samples,” Stata

Technical Bulletin 13: 1923.

Rose, A. (1951) The Costs of Prejudice.

127
Saenger (1953) The Social Psychology of Prejudice.

Shin, Annys (2004) “Foundation Helps Sodexho Counter Discrimination Suit,”Washington

Post, June 10.

Tiebout, Charles (1956) “A Pure Theory of Local Expenditures,” Journal of Po-

litical Economy 64: 416-424.

U.S. Dept. of Commerce, Bureau of the Census. CENSUS OF POPULATION

AND HOUSING, 2000 [UNITED STATES]: PUBLIC USE MICRODATA
SAMPLE: 1-PERCENT SAMPLE [Computer file]. ICPSR release. Wash-
ington, DC: U.S. Dept. of Commerce, Bureau of the Census [producer], 2003.
Ann Arbor, MI: Inter-university Consortium for Political and Social Research
[distributor], 2003.

U.S. Dept. of Commerce, Bureau of the Census. CURRENT POPULATION SUR-

VEY, 1979-1989 [Computer files]. Washington, DC: U.S. Dept. of Commerce,
Bureau of the Census [producer], 1990. Ann Arbor, MI: Inter-university Con-
sortium for Political and Social Research [distributor], 1990.

128