.

.7 I 1‘ 3.

q. . .21.

(a. ‘ffiiréual

.V .17}; ‘

z 4.1.2
. h an

I.
.w ".24. ...5..,...!.Lun.
an». n... ‘
r .-

. \
c

s .. :31...

i {‘95.

g nan: .
3.

MM
.1. :94. .

2; A: y :
EVER“... .9“; I

, , 9i: 1 :p. .if
. . . Nb"... i...
. «5:51.. ‘ V . f. 3 t v. . «M».
in. . 3:: ‘

t
3.. :

Ivy b-sa.‘ _

A. ‘9: .. H

hm

.2.

:34; .a J . . 1
.. . est. ~n~ ’2
J.- JvaJwin‘ .5!
\tug. 69‘ 2. .4.
I1 . \ .13" 3 1011.
9):. 1: filial
.21.. .i

 

 

 

vf.)..
T169)»
.5! t . r. 53.1....
Thaj
. V!
\. .ithr.

ioA no,
. film?

 

7 {NEWS

I
r—

/~
1%“

 

 

 

llBflhAlfi!
Michigan State
University

This is to certify that the

l
l
dissertation entitled ‘

TWO ESSAYS ON THE CHALLENGES FACING WOMEN AND MINORITIESI
IN THE LABOR MARKET I

presented by
JENNIFER ANNE TRACEY

 

has been accepted towards fulfillment
of the requirements for

Ph.D degree in ECONOMICS

 

 

Major professor

W/zfl
//

Date August 9. 2001

MS U is an Affirmative Action/Equal Opportunity Institution 0-12771

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

 

| DATE DUE DATE DUE DATE DUE
11mm EDNA

 

FlgB‘3 g 2 25105

 

 

 

 

 

 

 

 

 

 

 

 

6/01 chIRC/DateDuepBS—p. 15

 

 

Ill
0 E55

TWO ESSAYS ON THE CHALLENGES FACING WOMEN AND MINORITIES IN
THE LABOR MARKET

By

Jennifer Anne Tracey

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY
Department of Economics

2001

 

L/)
L)

II'I'O ES

hl‘ A

It}: L
'9 ""3? ' i .
“llhyi '35 1H“

1‘

Deer; ox cries}

The :3
Ethan Tor. us;
“omen u iIh }

A

“6 amt d‘af‘cc

4.13...“ L
“55 MiG-1R
A?!“

ABSTRACT

TWO ESSAYS ON THE CHALLENGES FACING WOMEN AND MINORITIES IN
THE LABOR MARKET

By

Jennifer Anne Tracey

This dissertation contributes to the literature on women and minorities in the labor
market by investigating two unique challenges facing these groups of workers that have
been overlooked in the past.

The first Chapter investigates the relationship between fertility and job search
behavior, uncovering a potential reason behind the poorer labor market outcomes for
women with young Children. Specifically, it looks at how the presence of young children
at home affects the amount of time an individual devotes to job search. Although the
effect of fertility on job search intensity is analyzed for both men and women, the
findings show that Children represent a significant constraint on search intensity for
women, while the results for men are somewhat mixed.

These are significant findings given the current political attention to child care as
a national concern and the importance of search behavior in determining labor market
outcomes, such as wages, unemployment duration and the probability of remaining in the
labor force. This is true regardless of whether the negative relationship between Children
and search intensity is primarily due to the cost of Child care while searching, or to the
expectation of a lower effective wage once work is found due to child care costs while

working.

‘3 ”.3
Thi 35‘

' $621120“. 17'.
. ~ _

‘ I I
‘ IM- ‘ 1 I)!
3-. 3m ”UK. c-

“has [‘3‘ vy- . I’d

Adl‘ut“

'9,
)‘gf3giai
.

V
we :0?

“it“ ’ha 'I ,
Ahlukllkc ~L.
..‘

The second chapter takes a new look at the problem of racial and ethnic job
segregation, investigating both its effects and causes. Even afier controlling for
differences in personal human capital and job characteristics, the results confirm that
racial and ethnic job segregation is an important contributor to the lower wages paid to
blacks and Hispanics than to similar whites. This study also explores the potential impact
Of racial and ethnic segregation on the likelihood of receiving various employment
benefits, yet finds job segregation to play a smaller role in explaining differences between
minorities and whites in the number of employment benefits received than it does in
explaining wage differentials. Finally, the potential causes of racial and ethnic job
segregation are explored. The results show that while minorities who reside in more
segregated neighborhoods are significantly more likely to work in segregated jobs, those
who commute longer distances to work are less likely to work in a segregated job, two
findings consistent with Kain’s (1968) “spatial mismatch hypothesis.” The findings also
indicate that blacks and Hispanics who work in larger firms are less likely to be in
segregated jobs, and that English fluency and citizenship status are strongly associated

with the likelihood of j ob segregation for Hispanics.

For my parents.

iv

Thxs \xcri;
liar} Holzcr has
1555:1311. I lean:
Tori inspired 'Dc
ad'~‘isor. he shit:
imslcth i :3:
wilsgak
if Iran. C\ en ;
Gtergcioxm

M) 031‘];

encouraged me

ACKNOWLEDGMENTS

This work could not have been completed without the help of many people.
Harry Holzer has been a tremendous asset. While working for him as a graduate
assistant, I learned more about conducting research than I could have in any class. His
work inspired both my research style and many of the ideas in this dissertation. As an
advisor, he shared both his insight and incredible knowledge of the literature related to
this study. I greatly appreciated and benefited from the quality of his feedback on my
work. I was also very honored and grateful that he continued to work with me through
the years, even after his career took him to the Department of Labor and then on to
Georgetown.

My other committee members also helped greatly, especially Jeff Biddle who
encouraged me to look at my work more critically. He made a considerable contribution
to both the empirical and theory sections of the second chapter, forcing me to think about
what the results really meant. In addition, his teaching skills were a tremendous asset
both in the Classroom and throughout the process of conducting this research. I am also
grateful to Steve Woodbury for graciously agreeing to help out late in the process and for
redirecting some of the methodology and theory I was unclear about. Finally, David
Neumark deserves a word of thanks for his help during the early stages of this work.

I definitely could not have made it through this program without my family. My
husband Paul cheered me on even when my work seemed at odds with our personal
goals, and my son Max, who arrived during all the craziness, helped me keep things in

perspective. Perhaps the most instrumental in encouraging me to finish were my parents.

.‘W
D
l

. c i
Masons.

computer suppt
. ,.l .
Teri. send as
. \ll
Film}.

." I 1 .‘
kimgiiil. L65.

I was often inspired by my mother’s determination and my father’s strong belief in the
value of higher education.

I would also like to sincerely thank everyone at Christensen Associates, especially
Carl Degen and Dianne Christensen, for generously providing me with office space and
computer support under the guise of part-time employment. I can honestly say that this
work would never have been completed without their help.

Finally, I am thankful to many of my fellow graduate students. In particular, Dan,
Kathleen, Leslie, Jess, and my husband Paul have served as sounding boards for both
ideas and the obligatory graduate-student griping. I also thank them for their moral

support and sharing their thoughts on earlier drafts of this work.

vi

LIST OF TAB
LIST OF FIG

CHAPTER I

THE EEEECI
iii-30C.
Emit
Econc
Tim D
Dcscr

Eccnt

Eric:
App:

4-
RAM).-
“51

CHAPTER :
RACIAL A)
CO\3EQIT
[Eliot
Theo
Dar».

u

R‘CSL,

c0“.

lik

Rn C.)
y 1‘

TABLE OF CONTENTS

LIST OF TABLES ................................................................................. viii
LIST OF FIGURES ................................................................................. x
CHAPTER 1
THE EFFECT OF CHILDREN ON JOB SEARCH INTENSITY ............................ 1
Introduction .................................................................................. 1
Existing Literature ........................................................................... 5
Economic Framework: Implications from a Simplified Model of Job Search ...... 9
The Data .................................................................................... 12
Descriptive Information on Child Care Concerns and Search Effort ............... 17
Econometric Specification and Empirical Results ..................................... 21
Empirical Results .................................................................. 25
Potential Econometric Problems ................................................. 28
Extensions and Concluding Remarks ................................................... 33
Appendices ................................................................................... 37
References ................................................................................... 46
CHAPTER 2
RACIAL AND ETHNIC JOB SEGREGATION: ITS CAUSES AND
CONSEQUENCES ................................................................................. 61
Introduction .................................................................................. 61
Theoretical Framework .................................................................... 65
Empirical Implications ........................................................... 77
Data and Empirical Framework .......................................................... 81
The Data Set ....................................................................... 81
Measuring Segregation ........................................................... 83
Empirical Methods ............................................................... 85
Results ....................................................................................... 89
How Job Segregation Affects Wages .......................................... 92
Does Job Segregation Affect Benefits? ........................................ 99
The Determinants of Job Segregation ........................................ 102
Potential Econometric Problems .............................................. 106
Conclusions and Policy Implications .................................................. 115
References ................................................................................. 1 18

vii

 

CHAPTER 1

CHAPTER 1

Table 1:

Table 2:

Table 3:

Table 4:

Table 5:

Table 6:

Table 7:

Table 8:

CHAPTER 2

Table 1:

Table 2:

Table 3:

Table 4:

Table 5:

Table 6:

LIST OF TABLES

How Child Care Affects Employment: Perceived Child Care
Constraints in Atlanta, Boston and Los Angeles by Sex ..................... 51

How Child Care Affects Employment: Perceived Child Care
Constraints in Detroit by Sex .................................................... 52

Mean Hours Searched per Week by Sex and Presence of

Children ............................................................................ 53
Perceived Child Care Constraints in Atlanta, Boston and

Los Angeles by Sex and Presence of Children ................................ 54
Means and Standard Errors of Variables Used for Job Searchers

by Sex and Metropolitan Area ................................................... 56
Search Intensity Equations by Sex, Detroit Sample .......................... 58
Search Intensity Equations by Sex, Atlanta-Boston—Los

Angeles Sample ................................................................... 59
Difference-in-Differences Estimates of the Effects of Children

on Search Intensity ................................................................ 60
Sample Descriptive Statistics by Demographic Group ..................... 122
Hourly and Log Wages by Demographic Group and Job

Segregation ....................................................................... 124
Log Hourly Wage Regressions without Controlling for Job

Segregation ....................................................................... 125
Log Hourly Wage Regressions: the Impact of Job Segregation ........... 127

Log Hourly Wage Regressions: the Impact of Job Segregation
by Demographic Group ......................................................... 128

Ordered Logit Estimates from Job Benefits Models without
Controlling for Job Segregation ............................................... 134

viii

 

 

labia 7:

labia 8:

1320169:

Table 10:

Table 11:

Table 7: Ordered Logit Estimates from Job Benefits Models: the Impact

of Job Segregation ............................................................... 136
Table 8: Probit Estimates of the Effect of Job Segregation on Benefits... .........137
Table 9: Characteristics of Minorities by Job Segregation ............................ 138
Table 10: Determinants of Job Segregation: Probit Estimates of Marginal

Effects ............................................................................. 140
Table 11: Log Hourly Wage Regressions: the Impact of Job Segregation

Controlling for Family Background ........................................... 142
Table 12: Log Hourly Wage Regressions: Impact of Job Segregation by

Demographic Group Controlling for Family Background .................. 143
Table 13: The Impact of Job Segregation and Residential Segregation on

Wages by Demographic Group ................................................ 144
Table 14: The Impact of Job Segregation on Wages by Demographic Group

Controlling for Neighborhood Effects ........................................ 145
Table 15: Measures of Racial Attitudes in Job Segregation Models: Probit

Estimates of Marginal Effects ................................................. 146

ix

 

CENTER 1

Figs: 1:

LIST OF FIGURES

CHAPTER 1

Figure 1: The Relationships between Children and Labor Market
Outcomes of Women and the Predicted Job Search Link ................... 50

\

THE F

Introduction
The labor
martian in the 1::
offenility and CE
anc'pox'erty.I Yc
ciiidnn on an In:

my. 1}.er

 

10': Search behaxi
may adieu the arr.
to impact search ‘1
{kid cue apcns:

Sm ml 111
entered he ES. 1.

I- .
‘~ 3., “.1331 . .
I‘M \ {Ills |

Chapter 1

THE EFFECT OF CHILDREN ON JOB SEARCH INTENSITY

Introduction
The labor market status of women with young children has received much
attention in the literature over the past several years. Economists have studied the effects
of fertility and child care costs on outcomes such as labor force participation, earnings
and poverty.1 Yet surprisingly, studies to date have largely ignored the impact of
children on an important determinant of such labor market outcomes — job search
activity. The purpose of this paper is to investigate the relationship between fertility and
job search behavior. Specifically, I look at how the presence of young Children at home
may affect the amount of time an individual devotes to job search. Children are expected
to impact search intensity, in part, by imposing additional search costs on parents, such as
child care expenses while looking for work.

Several factors make this topic both interesting and important. First, women have
entered the US. labor market in unprecedented numbers over the past few decades. This
is especially true of married women with young Children. Between 1970 and 1990, labor
force participation rates nearly doubled for married women with children under age six,
from 30 percent to 59 percent; by 1997, their participation rate had risen to 64 percent

(US. Bureau of the Census, 1998). Finding adequate and affordable Child care
arrangements while searching may be a significant barrier to finding employment for

theSe new labor market entrants. The fact that unemployment rates are significantly

\

I
See, for example, Ribar, 1992; Klerman and Leibowitz, 1994.

 

higher for some?
. 2

. .91: .

clpc‘tdmlll
Second:

care for presehoe

..,,I , '43,.
earmark. but.“

A

Poor working far

Income on child e

 

aboul 6 percent 1(
blTfiT [0 )0b 533:
hr lm'ey‘

pinculmy Imps

shorts continue I

hi gher for women with young Children than for those without seems to support this
expectation. 2
Second, market child care is a significant expense. The estimated annual cost of

care for preschool Children is over $4,000 (Casper, 1995). Such child care expenses are
especially burdensome for low-income families, even though federal programs like the

1 988 Family Support Act and the enactment of federal child care legislation in October
1 990 reduced many of the potential income-related differences in access to child care.
Poor working families earning less than $15,000 annually spend about 25 percent of their
income on child care, while families with annual incomes of $54,000 or more spend only
about 6 percent (Casper, 1995). Consequently, Child care costs are likely to be a greater
barrier to job search, and self-sufficiency, for low-income workers.

An investigation of the relationship between Children and job search is
Particularly important, and timely, given the current political climate. As welfare reform
efforts continue to emphasize the transition from welfare to work, child care has become
an iIIlportant national policy issue. The Personal Responsibility and Work Opportunity
Reconciliation Act (PRWORA) of 1996 included a $3.5 billion increase in firnding for
Child care subsidies to low-income families over six years (US. Dept. of Health and
Hunlan Services, 1996). However, provisions of this legislation do not provide direct

Child care subsidies or services to welfare recipients engaged in job search.3 Rather, each

\

2

Th: most drastic difference in unemployment is among single women, whose overall rate of
unstrmleyment in 1997 was 8.8 percent, compared to a rate of 18.8 percent for single women with children
for er 81x. The unemployment rate of all married women in 1996 was 3.2 percent, compared to 4.4 percent
diffn‘an‘ied women with children under six (US. Bureau of the Census, 1998). Obviously, these

1 ertraces in unemployment rates may also be influenced by differences in human capital investment,

3 minclal resources and other personal characteristics.

res e _Fami1y and Medical Leave Act (FMLA) of 1993 sought to alleviate the likelihood of job loss

Te “lung from interrupted labor force participation due to Childbirth by allowing a new parent the right to

turn to their previous employer afler maternity or paternity leave of up to 12 weeks. This legislation could

\

 

 

":iif mUSl "dc’W

3...:-

ernnntl. 3179 70%

Dgament Of E
Finally l
beams and m:
”1.311 fest er ho UT
opportunities arts
Robins (1985.) f:
unemploment.
by affecting the 1
issue in the cum
Wi‘h QIEVTOLLS es:
he} result from 1
Biiiely to be $1:
Numerow

LA.

tinidr '
31] fall Shun

it

‘1‘.
s: :Ten

“A“ 1
“5 165m:

 

state must “develop personal responsibility plans for recipients identifying the education,
training, and job placement services needed to move into the workforce” (U .S.
Department of Health and Human Services, 1996).

Finally, but perhaps most importantly, labor market outcomes such as wages,
benefits and unemployment duration are likely to suffer if search effort is restricted.
With fewer hours spent searching, less information is gathered about labor market
opportunities and fewer job offers are likely to be received. For example, Keeley and
Robins (1985) find that lower search intensity results in a higher duration of
unemployment. It is also likely that search effort will impact the duration of employment
by affecting the quality of employer-employee matches. Job retention is an important

issue in the current welfare reform debate. While policy makers may be less concerned
With previous estimates of lower employment and earnings for women with children if
they result from personal time-allocation and human-capital investment decisions, there
is likely to be greater concern if they arise due to restricted job search activity.
Numerous studies have shown that labor market outcomes for women with young
Children fall short of outcomes for women without Children. While part of these
differences results fi‘om the fact that on average women with children invest less in
market human capital, due in part to interrupted labor force participation, restricted job
Search activities may also play an important role. For instance, Waldfogel (1997) finds a
negative effect of Children on women’s wages, even after controlling for such human
capital variables as age, experience and education. For men, in contrast, she finds that

chudl'en have either a positive effect on wages or no effect at all.

\

T‘entielly reduce the importance of this research. However, its coverage is far from universal. Klerman
“(1 L'Blbowitz (1994) estimate that the federal FMLA covers only about a third of all working mothers.

Figure 1 1‘

T “It

i "r > - llbku
effort m3." Pm

Because It

the. men's job se.
heiatior ofmen 3
even ’he tnerexe
becoming int 011 e

CZTELITICTS 0i ’3

in

19,44 tCnper. 19 ,.
0ho men and ts
consent 3 st 2'1 :7.
503611 hat m x Cd

1n the Terri.
m“Esllgation oft}:
if existing bthdlc.

Senior outlines a

'9“: ~ I
§4‘.5Io .§

Mil french.

 

 

 

 

Figure 1 illustrates the empirical relationships that have been found between
chi 1dren and labor market outcomes for women, along with the proposed role that search
effort may play.

Because women remain the primary caregivers in our society, it may be
reasonable to expect that Children will affect their search behavior to a greater degree
than men’s job search. Yet Child care concerns may potentially impact the search
behavior of men as well as women. Not only is joint decision making often necessary
given the increased prevalence of two-income households, but more and more men are
becoming involved in the rearing of their Children. For instance, fathers were the primary

caregivers of 18.5 percent of preschool children in families with working mothers during
1994 (Casper, 1997). This study looks at the effect of fertility on the job search intensity
of both men and women. Perhaps not surprisingly, while the findings show that Children
rePresent a significant constraint on search intensity for women, the results for men are
Somewhat mixed.

In the remainder of this paper, I provide what I believe to be the first serious
investigation of the relationship between Children and search effort. I begin by reviewing
the eXisting bodies of literature on job search and child care issues. The following
Section outlines a conceptual model of job search, which suggests the appropriate
Statistical framework later used in the empirical analysis. Next, the data are discussed,

descriptive information child care issues and search effort is introduced, and the

empirical results are presented. Finally, the paper concludes with policy implications and

(1' °
11"muons for areas of further research.

\

MOWOVer, it does not help new labor market entrants.

 

[mung LiteratL

There is s.
fezale labor 3??
tlecished from ”.3
sereh behatior '7
liost of the sear;
adiehotomous t ..
sample by sex tit:
ditierently than rr

pretence to pool;

 

women are likel}
and to be lookingl

Much Of :1
implement 1r.

MCh ciion for I

and none deter?

Existing Literature

There is surprisingly little overlap in the bodies of literature on job search and
female labor supply, especially on the empirical side. Past research on job search, which
flourished from the late-19705 to the mid-19805, pays scant attention to the differences in
search behavior between men and women and the potential reasons for those differences.
Most of the search literature either focuses strictly on men or simply controls for sex with
a dichotomous variable. This study includes both men and women, but stratifies the
sample by sex throughout the empirical work. Since it is likely that women behave
differently than men when it comes to job search, separate behavioral functions are
preferred to pooled regression analysis. Due to a weaker attachment to the labor market,

Women are likely to have fewer business contacts, to possess weaker networking skills,
and to be looking for different types of employment.

Much of the existing literature on job search intensity investigates the role of
unelllployment insurance benefits.4 Barron and Mellow (1979) develop a theory of
S(321mb effort for unemployed individuals, in which effort involves the Choice of both time
and money devoted to search. Their empirical work, which focuses exclusively on
fa“01‘s affecting the time decision, finds a negative impact of unemployment benefits

(and the probability of being recalled from layoff) on hours devoted to job search. The
aut1101‘s rtur pooled regressions with no controls for sex, fertility or child care costs, and
consequently much of the variation in search intensity is left unexplained.5 Keeley and

Robins (1985) estimate separate search intensity equations for married men, women and

\

4
iSee Hamermesh and Rees (pg. 223, 1993) for a detailed discussion of how the receipt of unemployment
“malice benefits is theorized to impact search intensity.

7“

nuns. but may .
titers find in:
swch for mam;
Panécula.’
|
2: factors that at?
Amajor focus of
tensity has on ’.
ramming uncut;
in: he more 11?:

pn-bability ofbc.

lane. The authl

——

Women less likci
313 men, eVen .

A more r.
Helm '
. .mor bet“ cc:

.omponems of IE

333' ‘

.- *“TllS to c... I

-...I
EBLLIS Show am.

a. .

 

youths, but they do control for the impact of fertility. Like Barron and Mellow, the
authors find that unemployment benefits lead to a significant decrease in weekly hours of
Search for married men, however their results for women and youths are insignificant.
Particularly relevant to the findings of this study, Barron and Mellow (1981) look
at factors that affect the labor force transition probabilities of an unemployed individual.
A major focus of the research is on the impact that an unemployed individual’s search
intensity has on the probability of becoming employed, dropping out of the labor force, or
remaining unemployed. Highlighting the importance of search intensity, the results show
that the more time an unemployed individual devotes to job search, the greater the
probability of becoming employed and the less likely they are to drop out of the labor
force. The authors control for sex using a dichotomous variable, and find unemployed
Women less likely to become employed and more likely to drop out of the labor force
than men, even after controlling for differences in search intensity.

A more recent paper by Blau and Robins (1990) focuses on differences in search
behavior between employed and unemployed individuals. The authors look at four
components of the job search process: the choice of search methods, the choice of how
many firms to contact, the rate at which offers are received, and the acceptance or
rejection of an offer. They do not look at the choice of hours devoted to search. Their
r eSults show that employed individuals use fewer methods and generate fewer contacts,

but I‘téceive more offers than unemployed searchers.6 Differences are also found between
men and women, with women found to use fewer methods and make fewer contacts per

Week. However, since neither Barron and Mellow (1981) nor Blau and Robins (1990)

s\
The author's reported R—squared values range from .057 to .083.

 

 

unmlbrknfi
solellv‘to sex. 0'
sacs if“ omen
m men's sear;
The bod;-
g-aw'mg. Most c
mmmmmah.
two”- ample
significant bar? .
1993; Ribar, l9‘~
hxepunnpafit_
King from .1: .
These c:
COfisonlaborlt-
up the indirect c:
.11....
‘MJEIC this POT
fimmdmmh;

he
door force a.

 

 

control for fertility in their empirical work, these differences should not be attributed
solely to sex. One should expect differences in measures of search behavior between the
sexes if women’s search behavior is constrained by child care concerns to a greater extent
than men’s search.

The body of research on female labor supply and child care issues is large and
growing. Most of the studies to date focus on the labor force participation decision (but
not on the mechanism through which child care costs may affect the probability of
becoming employed). The major finding of this research is that child care costs may be a
significant barrier to labor force participation among women (see, for example, Connelly,

1992; Ribar, 1992; Kimmel, 1996). The estimated child care cost elasticities of labor
force participation vary widely among the various studies, however, with estimates
1‘imging from as low as 0 to as high as -0.9.

These cross-sectional studies attempt to measure the direct effect of child care
costs on labor force participation.7 But the estimates may be biased since they likely pick
up the indirect effect of restricted search effort on remaining in the labor force. To
illuStl‘ate this potential estimation problem, consider a two-period model in which an
indiVidual must decide to participate in the labor force. Such an individual initially enters
the labor force as unemployed and engages in job search. In the next period, she enters
one Of three states -— she becomes employed, remains unemployed or drops out of the
labor force. Child care concerns may impact the probability of transition between these

States by affecting a parent’s search behavior. If, for example, children lower search

\

6
rHOIZer (1987), using a different data set and focusing on the behavior of youth only, found a higher offer
ate for the unemployed than for the employed.

 

midi“) b) in:
t0? 01“ ”m i‘
the“? from Ba."
|

5631011 who are
labor force bet;
altered their lat- ’
paticipation m:
|
maplonnent s
worker smdron
Oxera‘rl.
on job search be
determining sear
finale labor sup
militant bamx‘
magnitude. vary r
int'estigated the 1
Shencoming bec
tel-fare recipien:

“ASST

51 With the pr

Sit betti'een thes

 

 

intensity by increasing search costs, then an unemployed parent will be more likely to
drop out of the labor force and be less likely to become employed (see the result cited
above from Barron and Mellow, 1981). Thus, a significant fraction of women in a cross
section who are classified as non-participants may be women who dropped out of the
labor force because child care restricted their search behavior, not because it directly
altered their labor force participation decision.8 That is, observed labor force
participation may understate true participation in the same manner as official
unemployment statistics may understate true unemployment due to the “discouraged
worker syndrome.”

Overall, the existing literature has generated some important findings. Research
on job search behavior has found that search intensity is a significant factor in
determining search outcomes, such as the probability of employment. The literature on
female labor supply and child care illustrates that the cost of market care is likely to be a
Significant barrier to labor force participation for women, although estimates of the
magnitude vary considerably. Nevertheless, virtually no study to date has seriously
investigated the impact of children on job search behavior. This is a significant
ShOITcoming because policy makers are not only interested in whether women (especially
welfare recipients) are participating in the labor force, but also in how society might
aSSiSt with the process of finding and maintaining jobs. This paper attempts to bridge the

gap between these two bodies of research. By investigating the impact of children on job

\

7

Chlld care costs directly affect the labor force participation decision by imposing fixed costs on work and
cy 10""firing the potential net wage (the more a parent works, the more money she must expend on child
aare). See Hamermesh and Reese (1993) for more details.
1 b e may also argue that women with children in a cross section are more likely to be viewed as in the

a or force (as unemployed) because lower search intensity should lengthen unemployment duration.
oweVer, studies investigating the impact of child care costs on labor force participation (LF P) usually use

search Whit}. .

gilt in [he 5 3;"

Economic Fran

lob seal
but is necessar}
employment er. g
lntm these in'.
implement ..

Sex eral s
unemployed no:
basic job search
isnbution ot‘pr
about mese alter:

npresenting the }

 

m a glen period

.»-.;'fi‘7"“‘m State“
fee on. L“ a

search activity, I hope to enrich our understanding of the role that child care constraints

play in the employment process for women and men.

Economic Framework: Implications from a Simplified Model of Job Search

Job search is a costly process that involves large investments in time and money,
but is necessary due to imperfect information in the labor market. Individuals looking for
employment engage in search to acquire information about alternative job opportunities.
In turn, these investments in job search are expected to affect both the length of
unemployment and the ultimate wage obtained.

Several search models exist that characterize the optimal strategy of an
unemployed worker by the choice of a reservation wage and search intensity.9 In the
basic job search model, the job possibilities of an individual worker are characterized as a
distribution of possible wage offers, with search efforts serving to generate information
about these alternative wage offers. This information, denoted by 9, is characterized as
representing the probability that a wage offer is sampled from the wage-offer distribution
in a given period. The production function for 9 can be depicted as
(1) 9 = 9(1, B),
where t represents the fraction of time per period an individual searches, and the term B
is a shift parameter to capture factors changing the likelihood of a wage offer for a given

search intensity, such as the individual’s search productivity. ‘0 The standard assumption

 

employment status as a proxy of LFP status because child care costs are often observed for employed
women only.

9 See, for example, Mortensen (1977) and Barron and Mellow (1979).

'0 Barron and Mellow (1979) also model the effects of monetary expenditures on 9. However, since time is
ofien considered the most significant search cost, other monetary expenditures on search (such as the cost
of transportation, stamps, newspapers, etc.) are ignored in this discussion.

‘u

 

is that 990' in“:
the likelihood of
It is n e1.
distribution of it
earlier a‘eepts .
allothers. The t
contained in;
discounted utiI;t
results of the in.
searching until ‘5
search.
Theretl:
lS'llOd-Eied as a j
P235? 1 model ti
{3)
wheres is Star;

01 fmiiin m"? -
. VQ.

'11
see

u .
“north, ..
“miter a fat.
k _
fa.

5105 are discu

. B ,
‘ (Flt anon ant: .\I '
4L I" ‘

('

 

 

is that 91>O, indicating that the more time per period an individual searches, the greater
the likelihood of generating a wage offer (Barron and Mellow, 1981).

It is well established that the optimal decision rule for sampling from a
distribution of wage offers is characterized by the choice of a reservation wage. The
worker accepts any wage obtained that meets or exceeds the reservation wage, and rejects
all others. The decision rule that determines the optimal search time, 'c, is derived from a
constrained maximization problem in which the unemployed worker maximizes expected
discounted utility from income and leisure, subject to budget and time constraints.11 The
results of the maximization problem show that the individual should expand time spent
searching until the marginal cost of search time equals the marginal expected gain of
search.

Therefore, search intensity, or the fraction of time an individual devotes to search,
is modeled as a function of the costs and benefits of job search. Given the focus of this
paper, I model the time spent searching for employment as
(2) S = f(CHILD, X)
where S is search intensity measured by hours searched per week,12 CHILD is a vector
of fertility measures, and X is a vector of variables reflecting other costs and benefits of
search. Variables expected to influence search contained in X include factors that affect
the mean of the wage offer distribution; variables to measure changes in B, the shift
parameter affecting the productivity of search; along with other control variables (these

factors are discussed in more detail below).

 

” See Barron and Mellow (1979) for a formal presentation of the model.

‘2 Obviously, one could obtain the exact form of T, the fraction of time per period an individual searches, by
dividing S by the constant 168 hours per week.

10

While it =
sezekbehax-ior.
of fenility on se.)
rogue search o-
cre costs will r;
can of looking
income etiect, t

other sources 0

cos. of looking :
income and sulx
cost of Search. 5

Liki‘“ i5;

 

While it would be surprising if the presence of children had no impact on job
search behavior, economic theory alone does not allow us to predict the sign of the effect
of fertility on search effort, a priori. On the one hand, children are likely to impose
unique search costs on parents in the form of child care expenses while searching. Child
care costs will reduce a searcher’s means to finance the transportation and other monetary
costs of looking for work, especially for those with very little savings, producing an
income effect. Child care costs may also produce a substitution effect for those with
other sources of income. Assuming an hourly fee per child for day care services, each
additional child in a household will reduce the searcher’s net income, raising the relative
cost of looking for work (although perhaps not by the same amount per child). The
income and substitution effects work in the same direction, both increasing the marginal
cost of search, so that fertility is expected to reduce time spent searching. ‘3

Likewise, time allocation theory (Becker, 1965) suggests that fertility will reduce
search intensity. Because children are thought to increase the value of an individual’s
time spent at home, this view implies that the presence of children will increase the
marginal cost of search by increasing the opportunity cost of searching.14 Therefore,
searchers with children are predicted to spend less time searching than searchers without

children, ceteris paribus.

 

‘3 Arguably, some parents have the less expensive option of using a spouse or adult family member to
supervise their young children while searching, and can utilize “free” child care services for 6-10 year olds
while they are in school. Yet, as labor force participation among women has increased and the existence of
extended families has diminished, there has been a growing trend toward the use of formal child care
arrangements. What is desired to test the above hypothesis empirically is a measure of child care costs
while searching. Because such data are not available, measures of fertility may be thought of as proxies for
child care costs in the following analyses.

'4 The time allocation theory also predicts that this effect should be stronger for women than for men if
women are more productive at rearing children (Killingsworth and Heckrnan, 1986). However, I expect
this effect to be relatively small. The empirical work that follows includes only current searchers - those
who have already decided to participate in the labor force.

11

 

  
   
  
 

On the L“
preside an 1161er
hitters. most or
rises. Here it nt.
firnetion depend.
intensity. indix'it
more resources
thiidren will in
intensity. all el

For \\ c
the mud fife:
Women are 5
other hand, h
Although (In.
decades‘ on

The Data

On the other hand, the financial responsibility associated with children may
provide an additional incentive to search more intensely for work. Individuals with
children must provide extra consumption (for their children) so that the value of a job
rises. Here it may help to think of parents as being altruistic, such that their utility
function depends positively on the well-being of their children.15 By increasing search
intensity, individuals will shorten expected unemployment duration and therefore have
more resources to spend on their children. Consequently, this effect suggests that
children will increase the marginal benefit of search and therefore increase search
intensity, all else equal.

For women, one might expect the first two effects described above to outweigh
the third effect, and the presence of children to be associated with lower search intensity.
Women are still more likely to take on the responsibility of child rearing. Men, on the
other hand, have traditionally been responsible for the financial well—being of the family.
Although this allocation of responsibility has changed a great deal over the past few
decades, on average men with children may feel more pressure to support their family
and thus search more intensely than men without children. In other words, the third

effect may be strong enough to outweigh the first two effects for men.

The Data
The data used in this study are from the Multi-City Study of Urban Inequality
(MCSUI). The MCSUI survey was administered to adult household residents in four

metropolitan areas: Atlanta, Boston, Detroit and Los Angeles. Interviewing was

 

'5 The degree of altruism should determine the discount rate on consumption (Becker, 1993), implying the
more altruistic parents are, the more they may attempt to concentrate their lifetime earnings during the

12

 

completed in the
of 1994 in Bostc

The sun
ettrrcin' and po
to iield roughly
were similarly o
concentrated pO'
ESS‘d a multistag
cirsterh g. This
in Boston. 1.543

ISIlr'Cd respond:

resWilx'ifil‘lls are
3' to ,

dimbmfil) and

 

completed in the summer of 1992 in Detroit, one year later in Atlanta, and in the summer
of 1994 in Boston and Los Angeles.16

The survey consisted of a probability sample of households, stratified by race-
ethnicity and poverty-status composition of the 1990 Census. Blacks were oversampled
to yield roughly equal numbers of whites and blacks in all locations; Latinos and Asians
were similarly oversampled in Los Angeles, as were Asians in Boston. In addition,
concentrated poverty areas were oversampled in all metropolitan areas. The project also
used a multistage sampling procedure, utilizing cluster sampling with three levels of
clustering. This process generated a total of 8,916 observations — 1,528 in Atlanta, 1,820
in Boston, 1,543 in Detroit, and 4,025 in Los Angeles. Restricting the sample to non-
retired respondents reduces the full sample to 7,570 observations - 1,283 in Atlanta,
1,630 in Boston, 1,182 in Detroit, and 3,475 in Los Angeles.”

In the following analyses, measures are taken to ensure that the data can be used
to draw inferences regarding the underlying population. First, analysis weights for
respondents are used, which were set inversely proportional to the household sampling
weight. Analysis weights also reflect nonresponse (if nonresponse is not uniformly

distributed) and the number of persons eligible for interview in the respective household.

 

period in which their children are most financially dependant.

The US. economy was recovering from the recession of the early 1990s when the survey was
administered. Monthly unemployment rates during this period averaged under six percent in Atlanta and
Boston, approximately eight percent in Detroit and under 10 percent in Los Angeles. To control for
differences in local labor market conditions, dummy variables for metropolitan area are included in the
regression analyses that follow.

Unfortunately, the sample of current searchers is significantly smaller, as discussed below. (Please also
note that three observations were dropped due to missing sex information.)

13

Second in all :r:
clustering and s:
|

One go;
concerning the s'
proride a rich 51.
collection of var
direct measure c

contains detailec

 

l
|
l
|
|

The prir‘.‘

Stin'eys used in '
ntensity n as pl‘
Intensity questi

I
‘

quuestion in l

 

Second, in all analyses robust standard errors are calculated that are also adjusted for the
clustering and stratification of the survey design.18

One goal of the Multi-City Study of Urban Inequality was to test hypotheses
concerning the status of women and minorities in urban labor markets. The data thus
provide a rich source of information on labor market histories, including an extensive
collection of variables related to job search behavior. In particular, MCSUI provides a
direct measure of search intensity — time spent searching for work. The data set also
contains detailed information on family background and child care issues.

The primary shortcoming of these data concerns a lack of Lmiformity between the
surveys used in the four cities. For example, in Detroit the question regarding search
intensity was phrased differently and pertained to a shorter time period than the search
intensity question asked in Atlanta, Boston and Los Angeles. The sample of those asked
the question in Detroit was also different than in the other three cities. Specifically, only
respondents who had looked for work within the past 30 days in Detroit were asked
“How many hours per week have you spent looking for work?” In Atlanta, Boston and
Los Angeles, all respondents who had searched any time within the last year were asked
“In total, about how many hours did you spend looking for work in (the last thirty
days/the last month) of your job search?”

As a consequence of these differences, the search intensity variable for Atlanta,

Boston and Los Angeles may be plagued with two sources of measurement error.19 First,

 

'3 This was accomplished using Stata survey (svy) commands (see StataCorp, 1997, pp. 305-312). It was
necessary to calculate a robust variance estimator because an examination of the model's residuals plotted
against the fitted values indicated the potential for heteroskedasticity, a violation of the least-squares
assumptions.

'9 If the measurement error is just mean-zero "white noise," it can be absorbed in the disturbance of the
regression and ignored. Although such classical measurement error in the dependent variable does not bias
the coefficient estimates, it will lead to less precise estimates since the errors will inflate the standard error

14

 

561331155 IfSl‘or‘Li'

 

mine p.151 (and '

)-

mount tssur

hOL‘IS searched '.

reSpOllLifill last 1

 

using unemplOI‘l
an tnten oi as sh
Second.
respondents in .
doing: the last I
is likel§ bee an:
reponderrts 11‘:
Edit iduals u e
searched 20 he
I‘OT “‘Ork in (:1:
31 331111pr r=
, L

“Quid hare SC‘

“339012.16 an

because respondents were asked to recall events that occurred as much as twelve months
in the past (and the question pertains to a longer time period), recall bias is likely to be a
significant issue. Unless the period of job search was very recent, the exact number of
hours searched is likely to be forgotten — the more time that has elapsed since the
respondent last looked for work the greater the likelihood of forgetting. Past researchers
using unemployment data have found recall errors to be a significant problem even over
an interval as short as one month (Horvath, 1982).

Second, the potential for measurement error is exacerbated by the fact that
respondents in Atlanta, Boston and Los Angeles are asked to report hours searched
during the last 30 days of their job search, as opposed to hours searched per week. This
is likely because the interpretation of the search intensity question may vary among
respondents whose search duration was less than 30 days. To illustrate, say two
individuals were both unemployed for only two weeks, during which time they both
searched 20 hours. When asked “In total, about how many hours did you spend looking
for work in (the last thirty days/the last month) of your job search?” the first respondent
may simply report 20 hours. The second respondent, however, realizing that he probably
would have searched more hours had he been unemployed for the full 30 days, may
extrapolate and report 40 hours.20

To mitigate the potential for recall bias, the sample of searchers in Atlanta,

Boston and Los Angeles was restricted to searchers who had looked for work within the

 

of the regression (see for example Greene, 1990, p. 295, and Angrist and Krueger, 1998, pp. 70-71).
However, in this case, the measurement error is probably not "white noise" since it is likely to depend on
the level of search intensity, since search intensity and search duration are presumably related, although the
exact relationship is likely to be quite complicated (See Barron and Mellow, 1979, p. 398).

2° A comparison of average search intensity by search duration, along with the results of a Kolmogorov-
Smirnov test (see StataCorp, 1997, pp. 301-303), revealed evidence of these measurement-error problems
in the data.

15

last 30 days in the analyses that follow (this also made the samples more uniform across
all four cities). To address the second potential source of measurement error,
misinterpretation of the search intensity variable among those who had not searched a full
month, the sample was further restricted to those whose search duration was at least 30
days. These two restrictions reduced the sample of searchers in Atlanta, Boston and Los
Angeles from 1613 to 729.21
This presents another potential shortcoming of the data: the size of the sample.

Even before making the adjustments discussed above, the sample of current job searchers
was fairly modest. But when we focus our attention on those who have looked for work
within the past 30 days, impose the search-duration restriction discussed above, and are
forced to analyze the Detroit data separately from the other cities due to inconsistencies
in the survey instruments,22 the potential for small sample size issues intensifies. The
consequences of small sample size can be quite serious. In general, precision of
estimation is reduced. Estimates may have large errors, such that investigators will
Sometimes be led to falsely accept null hypotheses of no relationship between two
Variables. Estimates will also be very sensitive to sample data, such that a single
Observation can sometimes produce drastic shifts in the sample mean. Unfortunately,
there are no easy remedies for small sample size problems, short of collecting a larger
Sample from the same population. Since this is not a possibility, in order to partially

mitigate problems associated with the size of the sample, a few overly influential

I The latter restrrctron may lead to sample selectron bras rf mdrvrduals wrth search duration exceeding 30
days are different from searchers in general. Note, however, that there were no observations in the Detroit
data for which search duration was less than one month, so this restriction also makes the two samples

gore uniform.
A Kolmogorov-Smirnov test (see StataCorp, 1997, pp. 301-303) determined that there are still significant

differences in the distribution of the search intensity variable between Detroit and the other cities, even
after the two sample restrictions were imposed. ‘

l6

obsen'ations (_ it
subsequently 0:“.

To ax 01¢
SCE.’C.rCl'S are P‘

inflow. and em:

 

which “as also
erogenous to I}.
assumption. sir.
perceived prodr.
Robins (1990) 3
flag rous var.
productivity. H
and additional 5'.
to the paucity 0:

Despite I
PFDVides a rich 5
it a useful stam:

W10 Spam?

Descriptive [I] f.

A“ “an:

COW-:4 ..
wherdbie E\ .

-:.,‘_
‘QS‘ I-
D‘~ a; “50%
«5%.? L113 2‘

“v...
.. ,.._
@105."

 

observations (for which the model did not fit) were diagnosed, and outliers were
subsequently omitted fiom the sample as discussed in Appendix A.23

To avoid reducing the size of the sample any further, however, employed
searchers are pooled with nonemployed searchers in the search intensity regressions that
follow, and employment status is included as an explanatory variable. This approach,
which was also used by Blau and Robins (1990), assumes that employment status is
exogenous to the choice of search intensity. Some may question the validity of this
assumption, since the employment status of a job searcher could be related to the
perceived productivity of searching while employed versus unemployed. As Blau and
Robins (1990) point out, employment status while searching should be treated as an
endogenous variable if there is heterogeneity in unobserved components of search
productivity. However, in this study a simultaneous system of equations was not used
and additional stratification of the sample by employment status was not entertained due
to the paucity of observations.

Despite these shortcomings, the fact that MCSUI is a relatively new data set that
provides a rich source of information on job search behavior and child care issues makes
it a useful starting point. Moreover, the search intensity results that follow seem quite

robust to specification in spite of the size of the sample.

Descriptive Information on Child Care Concerns and Search Effort
An examination of summary statistics disaggregated by gender provides

considerable evidence that women’s search effort is more likely to be affected by the

 

23 This was accomplished using the fit command in Stata and DFITS, Cook’s Distance and Welsch
Distance diagnostic tests (see StataCorp, 1997, pp. 372-397).

17

presence ofehzlt
rtfoznttion on F
Respondents \\ 1:
diecied set'eral
rows f able

ere concerns it.
example. almost
concerns caused
0m) 3 percent 0.
Wfiprtteipate tr
0mm by contr

The in:

seemingly lag;
19m of their t

We 3 c
ltestzons that n

udfil Eli‘lhr

weEn \

[fauna

ed 3'0 Ur

33733111 01' “On

7 ,
J“ 8%)?
5;?»

87'“ igm‘ '

n r,
' ‘ua
Jaw“...
y .‘_

presence of children at home than men’s search effort. Table 1 provides descriptive
information on perceived child care constraints in Atlanta, Boston and Los Angeles.
Respondents with children under eighteen were asked whether child care concerns had
affected several aspects of their labor market experience in the last year.24 The first four
rows of Table 1 illustrate that women are much more likely than men to feel that child
care concerns have constrained their ability to gain and maintain employment. For
example, almost one-third of women with children under eighteen reported that child
concerns caused them to “not look or apply for wor ” over the past year, compared to
only 8 percent of men. Moreover, almost 20 percent of women reported that they could
not participate in school or training programs due to child care concerns. Only 7 percent
of men, by contrast, felt this way.

The last four rows of Table 1 indicate that, of those parents who had worked
sometime within the past year, women were significantly more likely than men to report
that child care concerns caused them to be late or absent from work, to change their hours
of work, or to lose out on a promotion or raise. Overall, however, Table 1 shows that a
surprisingly large percentage of men felt that child care concerns had affected several
aspects of their own employment over the past year.

Table 2 contains summary statistics generated from similar, but less-pointed
questions that were asked in the Detroit survey. Respondents in Detroit with children
under eighteen were asked: “Has the cost, availability or quality of child care ever
influenced your employment or that of your (spouse/partner) in any way?” Thirty-one

percent of women answered yes to this question, compared to 25 percent of men. It

 

2‘ These questions, which are listed in the first column of Table l, are admittedly subjective and may be
subject to a significant amount of reporting bias. For this reason, the term “constraint” should be

18

at"

.1
. afhaa \v

an???“
refill.

V .
:1 ‘Jl"
,L .

ha 'I

‘:"l""‘."
..l.t. .11.;

. .
y.;.p .
LL45)¥ l:

seems reasonable to believe that the majority of Detroit men who answered yes may have
been referring to constraints on the employment of their spouse or partner, especially
when one compares the results from Table 2 with Table 1. Moreover, the results
presented in rows two and three of Table 2 reveal that women were significantly more
likely than men to mention that the cost and quality of child care were areas of concern.

As a follow-up question, respondents in Detroit who felt that child care had
impacted their (or their spouse’s/partner’s) employment were asked “In what ways did
these issues influence your (or your spouse’s/partner’s) employment?” The most common
responses to this secondary question included mention of constraints on when and the
amount of time they could work and constraints on the ability to enter the labor force or
maintain a job, with men more likely than women to report the former (32 percent v. 25
percent) and women more likely to report the latter (43 percent v. 35 percent). As
illustrated in the fifth row of Table 2, roughly 16 percent of women and men felt that
child care concerns restricted their ability (or the ability of their spouse or partner) to gain
employment or to choose a certain type of employment.

Table 3 reports summary statistics on search intensity by sex and the presence of
young children. In Detroit, there is a dramatic difference in time spent searching between
those with and those without children. The means presented in panel A of Table 3
indicate that parents of young children in Detroit, whether they are mothers or fathers,
search significantly fewer hours per week than those without young children. This result
is particularly strong for women; on average women with children under six search

roughly four hours per week, while those without any children ten or under search almost

 

interpreted quite loosely in the following discussion.

19

 

(I;

l

'b

C

.4

if“ C?

, C
r: "'
35 Ah

.‘
'1.’ W,
tLi .;.\

l0? 'M

mere

‘51»;

eleven hours per week.25 Perhaps surprisingly, those with children aged 6-10 reported
fewer hours searched per week on average than parents of children under six, although
the finding is much stronger for men than for women. This relationship also holds true
for men in Atlanta, Boston and Los Angeles, as illustrated in Panel B of Table 3, but not
for women in these cities. 26
A potential explanation for this seemingly peculiar finding may be that fathers are
more likely to become involved in the rearing of children who are slightly older.
Evidence that men take on more of the child care responsibilities of children aged 6-10,
and that this responsibility may affect their career, can be seen in panel B of Table 4,
which presents the same information contained in Table 1, but stratified by child age
groupings. The results speak for themselves. In virtually every case fathers with children
aged 6-10 are more likely than fathers of children under six to feel that child care
concerns have affected their employment in the past year (although in some cases the
differences are not statistically significant). For example, 15 percent of men with
children 6-10 reported that child care concerns caused them to not look or apply for work
in the past 12 months, compared to only 7 percent of men with preschool children.
Comparatively, panel A of Table 4 illustrates that women with preschool aged children

were slightly more likely to report that child care concerns have affected their work life

than women with 6-10 year olds.

 

25 The relatively low hours of search per week has been noted by other authors (Keeley and Robins, 1985;
Barron and Mellow, 1979). The results in this sample are slightly lower than in prior studies because I
include employed searchers, who search less intensely on average.

26 It is interesting to note that Connelley (1991) finds that cost per hour of care is higher for school-aged
children (although average weekly expenditures are less).

20

Econometric S

Hit in; .
sigrftcant diff:
die central ques

“053611.

 

older children n
lebat'ior. To te|
women of the i1
(3)

Where 5, represg
PiliiGE, is a p

Vettor of indie,

ll

The ex;

BachelOr'S dnm

Econometric Specification and Empirical Results

Having documented the prevalence of perceived child care constraints, along with
significant differences in mean search time by gender and fertility status, I now turn to
the central question of this paper: Whether children effect the search intensity of men and
women.

Overall, Tables 1-4 provide some initial evidence that children, even slightly
older children who are in school for a portion of the day, are likely to impact job search
behavior. To test this hypothesis further, I estimate separate equations for men and
women of the form:

(3) Si= a + BCHILDI + 5PWAGE,~ + yPTIME; + kXi+ 6;,

where S,- represents hours searched per week, CHILD,- is a vector of fertility measures,
PWA GE,- is a predicted wage, PT IMEi is an indicator of part-time employment, X,- is a
vector of individual and household control variables, and e,- is an error term which
includes unobserved characteristics. Each of the independent variables is discussed in
more detail below.

The explanatory variables contained in X,- include education categorical variables
(high school diploma or GED; Associates degree, vocational or trade school certificate;
Bachelor's degree; Graduate degree; with high school dropout the omitted category),
employment status and a measure of nonwage income. Nonwage income is expected to
have a negative impact on search intensity, since higher nonwage income reduces the

gain to locating a wage offer through search.27 Employed searchers are predicted to

 

27 In the model presented by Barron and Mellow (1979), nonwage income leads to an unambiguous
reduction in search time. This is due to the fact that they assume time and money inputs are combined to
produce 9 additively (i.e., the level of market expenditures does not affect the marginal productivity of
search time and vice versa). The model presented by Tannery (1983) assumes a more general production

21

search f6“ 5’ ht
_ . 1 . 'r-
winch lmp‘} a ’
ch.” SinCc
Zl‘ilfi liner-1&0):

states and non-A

The 6d:

 

productixity. alt
cc-eiicient on tl
wit a higher 10
1:2} not need tc

Tao prc
distribution. In.
education. sex a
increase in one'
it 3150‘? sues
masure the u .2

4 ;
0.130
t

in

h “. -p. E
on :Ur H SO a.
it"s“:"c L;

M ‘5 1 A

‘jf-BIU F nkon‘c u
'V TH. LEE 0" D
4 a“) “

A l"’
‘.N“‘jn. H
LDC" C;

.K c.
‘ Oi he

search fewer hours per week than unemployed searchers, both due to time constraints,
which imply a higher marginal cost of search, and the lower expected benefits of
search.28 Since employed searchers are likely to react differently to nonwage income
than unemployed searchers, X,- also includes an interaction term between employment
status and nonwage income.

The education variables are included to measure changes in [3, the shift parameter
affecting the productivity of search. A higher education level may indicate higher search
productivity, although this does not necessarily imply one would predict a positive
coefficient on the included education categories. In other words, because an individual
with a higher level of education may be more efficient at searching for employment, they
may not need to spend as much time searching as a high-school dropout does.29

Two proxy variables were used to reflect changes in the mean of the wage-offer
distribution, m. First, PWA GE,- is a predicted wage based on the individual's age, race,
education, sex and metropolitan area, which should be directly related to m. Because an
increase in one's mean eamings potential represents an increase in the benefits of search,
the theory suggests a positive coefficient on PWA GEi. A second variable used to
measure the wage-offer distribution is PT IMEg, a dummy variable for individuals likely to
be searching for part-time work. Part-time work is typically associated with lower wages
and benefits, all else equal, and therefore is expected to lower the benefits of search and

have a negative coefficient. Unfortunately, respondents were not asked explicitly if they

 

function for 9 so that time and money can be complementary inputs. Under this assumption, an increase in
nonwage income will raise the productivity of time spent searching, and thus increase search intensity.
Therefore, the overall impact of nonwage income may be ambiguous.

28 By definition, employed searchers are only looking to improve their compensation or non-monetary
aspects of their job.

22

 

in??? 5

Cliff:

, -. .
an.‘4r
\A'urhu

9' , '
Lift]

will:

E4

lip, 25

. 7?
pLwa

.5:

were searching for part time work. PT IME,- is only an indicator of whether the searcher's
current or past job was part time.30

Vector CHILD,- includes two binary variables, the first indicating the presence of
children aged 0-5 and the second indicating the presence of children aged 6-10. As
discussed above, economic theory does not allow us to predict the impact of the fertility
measures contained in CHILD,- on search intensity, although the information provided on
child care concerns suggests we are likely to find a negative coefficient on the fertility
indicators, at least for women. The information provided above also indicates that
slightly older children, those aged 6-10, seemed to have a bigger impact on employment
issues for men than preschool-aged children.

Because children are expected to influence search behavior in part due to the costs
they impose on parents in the form of child care expenses while searching, vector X,- also
contains family structure variables, meant to serve as proxies for the availability of
unpaid care. Access to child care is a function of family structure, or the number and
characteristics of other potential providers available to care for children, such as older
children or an elderly parent.31 I include two measures of non-parental adults in the
household - the number of young adults aged 18-34 and the number of older adults aged

35-64 — to control for the availability of unpaid care. Unfortunately, information on

potential care providers outside the home is not observed.

 

29 Barron and Mellow (1979) include a variable measuring years of education, as opposed to the categorical
variables used here, and predict and find a positive correlation between years of education and hours
searched per week.

3° See Appendix A for an exact definition of this variable.

3‘ Using data from the 1985 Survey of Income and Program Participation (SIPP), Connelly (1992) finds
that the presence of a teenager, an adult woman other than the mother, or an adult male other than the
father, lowers the probability of a family paying for child care.

23

“1.34

11L .1

to. is;
'vul :
lilLiLl
15:33:

fafla"
i¥‘¥¢

lens:

011 35

her:

\‘ij‘w

“N,
V

: “ ‘

Q's

For the Atlanta-Boston-Los Angeles sample, X,- also includes binary indicators of
whether the searcher received Unemployment Insurance (U1) benefits and/or welfare
payments (at the time of the surveys, welfare payments were administrated under the Aid
to Families with Dependant Children program, or AFDC). These variables were not
included in the Detroit model because survey questions in Detroit only asked if anyone in
the household received welfare payments (the survey did not ask if the respondent herself
received AFDC, nor did it ask about UI benefits for either the respondent or the
household). Empirically, most studies have found UI benefits to have a negative effect
on search intensity. However, one cannot predict, a priori, the impact of UI on weekly
hours searched. Although UI benefits provide a means of support that reduces the
incentives to look for work, they may also provide unemployed workers the means to
finance the transportation and other monetary costs of looking for work. Moreover, one
of the conditions to receive benefits is an active search for work, as most states require
beneficiaries to register with their state employment service and demonstrate that they
have contacted employers (see Hamermesh and Rees, 1993, for a more detailed
discussion). The dummy indicator for AFDC receipt is likely have a negative impact on
search intensity, since such nonwage income reduces the incentives to search for work.
However many states also required AFDC recipients to register with state employment
agencies and demonstrate that they are actively engaged in job search, so the expected
sign of the AFDC coefficient is also ambiguous.

Table 5 presents means of the analysis variables for the final sample of searchers,
stratified by sex and metropolitan area (Detroit versus Atlanta, Boston and Los Angeles).

The final sample consists of 201 observations on Detroit households and 727

24

   

obsen'ations or
more detailed
along truth a di
Pll'AGE, the p
d‘s’ribution. de
and thus is pre:
QEZEELflgg

Tables
the Mo 53mph
PIOVlClfi almOS
551ml} bchaVi‘
'he Detroir 53}
negative 6,31,:
are mp‘iOYCd
income has a
:he preSenCe ‘
“‘th an Older
as MW e
[mi foi u on

expected si e:

&%m1°SUnn
W95 plediqe

observations on households in Atlanta, Boston and Los Angeles. Appendix A contains a
more detailed description of how the analysis variables discussed above were created,
along with a discussion of the omission of outlier values. The method used to construct
PWAGE, the proxy variable used to reflect changes in the mean of the wage offer
distribution, deserves additional attention due to the potential for sample selection bias,
and thus is presented separately in Appendix B.
Empirical Results

Tables 6-7 report OLS regression estimates of the search intensity equation (3) for
the two samples. The results seem to support the theory fairly well; moreover, they
provide ahnost indisputable evidence that young children have a significant impact on job
search behavior. As shown in columns (1) and (2) of Table 6, regression analysis using
the Detroit sample indicates that children aged 0-5 and children 6-10 have a significant
negative effect on hours searched per week for women. Also as predicted, women who
are employed while searching spend significantly fewer hours looking, and nonwage
income has a significant negative effect on search intensity. It is interesting to note that
the presence of an older adult in the household may have a mitigating effect. Women
with an older adult present search more intensely, perhaps because older adults may serve
as in-house child care providers or simply take over normal household duties, freeing up
time for women to search. As shown in column (2), PWA GE also enters with the
expected sign, although its coefficient is insignificant. The only variable that does not
seem to support the theory presented is PT IME, the indicator of part time work, which
was predicted to have a negative impact on search intensity due to the lower benefits

associated with part time jobs. However, the coefficient is insignificant, and the variable

25

inr-
‘I‘I ‘
.1..-

Hen-

CL‘;
~

is not necessarily a true indication of whether the individual is seeking part time work,
only that they were previously employed part time.32

For men in Detroit, the results are equally striking. While slightly older children
aged 6-10 have a significant negative impact on men's search intensity, very young
children 0-5 actually are estimated to have positive effect on weekly hours searched.

This seems to support what we saw in Tables 3 and 4: that slightly older children are
perceived to represent a greater constraint on the search behavior of men than very young
children do. Again, this may be due to the fact that fathers take on more of the child
rearing responsibilities once their children get older. It is also possible that men with
very young children feel a stronger sense of financial responsibility, since they are
probably not yet acclimated to the idea of "an extra mouth to feed."

Employed men are estimated to search significantly fewer hours than unemployed
men, and nonwage income has a negative effect on search intensity (although the latter
effect loses its significance when the predicted wage variable is added). Among men
who are employed, however, nonwage income is estimated to have a significant positive
impact on hours searched. The interaction term may be picking up individual
heterogeneity, indicating that employed searchers who have accumulated more nonwage
income are more ambitious on average, and thus have a higher propensity to search. The
coefficient on PTIME also enters with the correct sign and is significant at the .01 level.
However, since this variable is not an exact measure of whether the individual is

currently searching for a part time job, but only an indicator of whether one's current or

 

32 The education dummies should be thought of as merely controls for [3, since several of these categories
contain very few individuals. For example, among current searchers in Detroit, there were only four
women and four men with a graduate degree. There is also the possibility of multicollinearity with the
predicted wage variable.

26

.‘\

most recent jo':
heterogeneity.
nine work is n
as opposed to '
azth searching

Tabie '

Angeles. Pote:

problems disc
children have

For \M
regent e imp:
ozl} other Va:
emploment s
fewer hours.
q-si;

r“ MutCIEd S: ..,..

most recent job was part time, we may again be measuring some type of individual
heterogeneity. Men who have worked or are working part time (particularly since part
time work is much less likely among men) may simply have a lower propensity to search,
as opposed to the dummy variable only measuring the lower expected returns associated
with searching for part time work.

Table 7 presents results for the sample of searchers in Atlanta, Boston and Los
Angeles. Potentially due to the way the survey question regarding search intensity was
phrased in these cities, the data do not seem to fit the model as well as it did for the
Detroit sample, despite the steps taken to address the potential measurement error
problems discussed above. Nevertheless, the results support the theory that young
children have a significant impact on job search behavior.

For women, the presence of very young children aged 0-5 has a significant
negative impact on search intensity. In the specification presented in column (1), the
only other variable estimated to have a substantial impact on search for women is
employment status, again showing that those who are employed search significantly
fewer hours. Nonwage income and the part time work dummy also enter with the
predicted sign, yet they are insignificant. When the predicted wage is included, however,
the part time dummy becomes significant at the 10 percent level.

Columns (2) and (4) of Table 7 also include binary indicators of whether the
searcher received Unemployment Insurance benefits and/or welfare payments (AFDC).
It is interesting to note that women who collect UI benefits are estimated to search more
intensely than those who do not receive benefits. This finding is contrary to most

previous empirical studies on the issue (as indicated above, most previous studies found

27

F-‘l.

: _4,
ahj

p
LII

o—o

I

UI to weaken search incentives), but may be due to the search requirements associated
with many state unemployment insurance pro grams.3 3
The results for men in Atlanta, Boston and Los Angeles are similar to what we

saw among the Detroit sample of men in that very young children aged 0-5 are associated

with greater search intensity, while children 6-10 are estimated to reduce search intensity

 

p
(although the latter impact is only marginally significant in the model presented in 2'
column (3) and insignificant in column (4)). Again, employed searchers look less j
intensely (with employment status losing significance when the U1 and AFDC dummies !

are added, possibly due to multicolliniarity), and part time work is estimated to
significantly reduce search intensity. Adults present in the household and the predicted
wage also seem to have a positive impact on hours searched as the model predicts,
although the coefficient on PWAGE is insignificant.34
Potential Econometric Problems

In addition to the measurement error problems discussed previously, the present
analysis is potentially affected by four separate, but related, econometric problems. The
first pertains to whether fertility measures can be safely treated as exogenous within a
search intensity equation. Some may argue that fertility and job search decisions are

made simultaneously, and thus would be more properly modeled as contemporaneous

 

33 The empirical work by Tannery (1983) also finds UI benefits to increase search intensity, theorizing that
UI benefits might encourage unemployed workers to allocate greater market expenditures on search
activities. The author's data set, however, differs from most studies in that it contains observations only on
those who have successfully found work (not on unsuccessful job hunters and those on layoff waiting to be
recalled). Blau and Robins (1990) find that those who receive either UI benefits or AFDC use more
methods of search (which they attribute to the requirement to register with a state employment service), but
make fewer contacts with potential employers and receive fewer job offers.

3’ Introducing three additional demographic variables does not significantly change the findings presented
in Tables 6 and 7. Interestingly, the age of an individual is not correlated with search time. Marital status
does not impact the search intensity of women or men in Detroit, but negatively impacts hours searched for

28

etdogenous \ n
whereas job sc
sigiticant iss..
EgteSSion res:

The sec
audible dgm

 

112316 ll‘eon S

srbstantially i
care while sea
With young cl
directly resin.
once they rec:
Parent's effec

Became Pine

lQb search ab

endogenous variables. However, since fertility is a long-term (in fact, life-long) decision,
whereas job search is a relatively short-term prospect, this problem is not likely to be a
significant issue.35 Nevertheless, one should always be careful when interpreting
regression results to imply causation.

The second potential econometric problem centers on the limitations of the
available data. Without a direct measure of child care costs while searching, it is difficult
to discern what the fertility variables in equation (3) are really measuring. As discussed
in the theory section, there are many reasons why one would expect children to impact
job search intensity. I suspect the strongest reason for the negative relationship between
children and search intensity among women centers around the difficulty of finding
affordable child care while searching (i.e., the unique search costs children impose on
parents). Kisker et a1. (1991) find that family day care providers "tend to charge
substantially higher hourly rates for part-time than full-time care," so the costs of child
care while searching may indeed be significant. It is also possible, however, that women
with young children search less intensely, not only because current child care issues
directly restrict their search behavior, but also because their effective wage will be lower
once they receive a job, due to child care costs while working. That is, we can think of a
parent's effective wage as their hourly wage minus the cost per hour worked of child care.
Because parents have lower expected effective wages than nonparents do, the benefits of

job search are lower.36 This is something that the predicted wage variable does not

 

men in the Atlanta-Boston-Los Angeles sample. Race does not generally impact search intensity, but white
men in the Atlanta-Boston-Los Angeles sample spend more time engaged in job search.

35 Fertility, on the other hand, is likely to be endogenous with respect to labor supply because both are long-
term issues.

36 Child care costs will reduce the effective wage of the parent who is the “designated caregiver,”
traditionally the mother.

29

I
i
t
!

 

.qn‘r

LL“).

11;;931
bbkbe

R1 .‘i'

\A

AL.‘

cm.

4’!
('1)
‘5)

control for. It is important to keep in mind, however, that searchers have already
determined that their expected effective wage is high enough to participate (i.e., it
exceeds their reservation wage), and the above analyses condition on such labor force
decisions. Thus, I expect child care costs while searching to have a stronger impact on
search intensity than the expectation of a lower effective wage once a job is found.

The third estimation issue concerns whether the included fertility controls are
picking up longer-term labor supply decisions, rather than short-term search constraints.
In other words, decisions regarding search intensity for women with children may be
driven by decisions about labor supply. For example, women with children who have
decided to find work may be restricting their search to certain types of employment (such
as part-time or flex-time jobs) because of their children. And since these types of jobs
ofien are associated with lower wages and benefits, the marginal benefit of search time
will be lower. Therefore, the negative effects of young children on hours searched for
women seen in Tables 6 and 7 may not be due solely to the search constraints imposed by
children, but may be partly due to decisions regarding labor supply. To the extent that
the dummy variable for working part time in the past controls for such labor supply
decisions, the estimates of the effects of children on search intensity are indeed picking
up true search constraints for women.

A more generalized version of this econometric problem is endogeneity due to
unobserved heterogeneity, or omitted variable bias. Women with children are likely to
differ from those without children in many ways, including perhaps in their propensity to
search for jobs (or propensity to work hard in the short term). In a reduced-form hours

searched equation like (3), one may expect correlation between the fertility measures and

30

3 h‘ti ‘u-l

i» ..
it 5‘

at;

the error term, 6,, which includes unobservables such as the propensity to search.
Consequently, omitting controls for variables like “propensity to search” is likely to
result in overestimation of the effect of fertility on search intensity if such variables are
negatively correlated with the included fertility measures. Of course there may be other

unobservables, such as responsibility, that are positively correlated with fertility,

E
implying the impact of children on search intensity was underestimated. One approach to f

I
address this potential heterogeneity problem is to use a proxy variable for unmeasurable '
characteristics.37 MCSUI does not contain many compelling choices for appropriate L

proxy variables. However, since the above analyses condition on variables that are likely
to be closely related to propensity to search and/or responsibility, such as educational
attainment and labor supply decisions (since all searchers have already made the decision
to participate), the potential for such a heterogeneity problem is likely mitigated.

Another approach for dealing with unobserved heterogeneity, which I attempt
here, is to use a “difference-in-differences” ODD) estimate.38 I include interactions
between the fertility measures and a dummy indicator representing the sex of the parent
in the search intensity equation:

(4) S.- = a + BCHILD; + (pCHILDi-FEM; + nFEM- + 5PWA GE.- + yPTIME; + XX,- + 5,,
where FEM,- indicates the parent is female.

Equation (4) allows identification of the impact of children on search intensity
under certain assumptions on the form of unobserved heterogeneity. That is, assuming
the effects of children on search intensity for fathers reflect only unobserved

heterogeneity and not a true search constraint, then the difference in the estimated effects

 

37 Alternatively, one could use panel data and difference out time-invariant factors such as propensity to
search.

31

I
,4
..

 

 

bcnreen motl'l

mtersity. Tr. |
correlated “it
Table
regressions or
the child dun:

I

Results are pr

Boston-Los

{Appendix D

interaction te:

COfftparison 11

 

indicating m,
intensity. No

Weed lllee

between mothers and fathers yields an estimate of the effect of children on search
intensity. This DD estimate is unbiased so long as no gender-specific unobservables are
correlated with the fertility measures.

Table 8 presents DD estimates of the effects of children on search intensity from
regressions of equation (4) above, which include interactions between sex of parent and
the child dummies, a dummy for sex of parent, and pool men and women together.
Results are presented from two estimated equations for each sample (Detroit v. Atlanta-
Boston-Los Angeles), corresponding to the specifications used in Tables 6 and 7 E
(Appendix D contains coefficients on the control variables). The coefficients on the
interaction terms indicate the difference in the effect children have on mothers in
comparison to fathers, and thus equal the DD estimates. If one assumes that the father
effect reflects unobserved heterogeneity, then we can interpret the results from Table 8 as
indicating that only very young children aged 0-5 have a negative impact on job search
intensity. Note that the magnitudes of the DD estimates for younger children actually
exceed those presented in Tables 6 and 7.

The results for men presented in Tables 1 and 2 and Tables 6 and 7 may call into
question the assumption that the effects of children on search intensity for fathers reflect
only unobserved heterogeneity and not a true search constraint. Nevertheless, if the
father effect is picking up some heterogeneity, we can conclude that the results for
women presented in Tables 6 and 7 represent a lower-bound estimate of the impact that

young children 0-5 have on search intensity.

 

38 See Gruber (1994) and Holzer and Ihlanfeldt (1998).

32

Extensions 3

This s

 

 

 

concerns asscl
These are si;
labor market

remaining in

national conc
children and :
to the expect.
while workin

should influet

 

Politi
135M sever;
Ranting pare]
mempls to a‘.
participation .
Previous strip

the federal 1:)

Extensions and Concluding Remarks

This study provides considerable evidence that young children, and the child care
concerns associated with them, significantly restrict the job search intensity of women.39
These are significant findings given the importance of search behavior in determining
labor market outcomes, such as wages, unemployment duration and the probability of
remaining in the labor force, coupled with the current political attention to child care as a
national concern. This is true regardless of whether the negative relationship between
children and search intensity is primarily due to the cost of child care while searching or
to the expectation of a lower effective wage once work is found due to child care costs
while working. Determining the relative importance of these two factors, however,
should influence the proper policy prescriptions.

Politicians have recently begun to understand the importance of child care issues.
In fact, several important steps have already been taken to meet the challenges that
working parents face. For example, the Family and Medical Leave Act (FMLA) of 1993
attempts to alleviate the likelihood of job loss resulting from interrupted labor force
participation due to childbirth by allowing a recent parent the right to return to their
previous employer afler maternity or paternity leave of up to 12 weeks. Unfortunately,
the federal F MLA covers only about a third of all working mothers (Klerman and
Leibowitz, 1994) and does nothing to help new labor market entrants with children.
Politicians have also significantly increased the amount of money targeted directly

toward child care in recent years. The welfare reform law of 1996, for example, included

$14 billion in funding for child care subsidies to low-income families over six years — a

33

 

$35 billion in
Nevertheless.
that only 10 p
1993 received
cemented oi
esidence that
getfing the cr:
Morel.
welfare into :1
lIlCTCESlllg 1T0
33d litmtan S
atmou‘ledgei
illitl [hose “ill

 

mum}- the;
l

(113311011 of U1

331d Robins. l

$3.5 billion increase (U .S. Department of Health and Human Services, 1996).
Nevertheless, a recent Department of Health and Human Services (DI-IHS) report found
that only 10 percent of low-income families eligible for federal child care assistance in
1998 received it (U .S. DHHS, 1999a). F orrner DHHS Secretary Donna Shalala
commented on the report, noting "This timely and important report provides conclusive
evidence that millions of low-income parents eligible for child care assistance are not
getting the critical support they need to stay employed" (U .S. DHHS, 1999a).

Moreover, while welfare reform has shown some success in moving people from
welfare into the labor market, with the percentage of welfare recipients working
increasing from seven percent in 1992 to 27 percent in 1998 (U .S. Department of Health
and Human Services, 1999b), there is clearly room for improvement. It is widely
acknowledged that one of the biggest remaining challenges of welfare reform is ensuring
that those who leave welfare for jobs remain employed. But what is not widely known is
the important role job search can potentially play in meeting this challenge. Search
intensity, measured by weekly hours searched, has been shown to impact both the
duration of unemployment and the probability of dropping out of the labor force (Keeley
and Robins, 1985; Barron and Mellow, 1981). Job seekers spend time searching to
gather information about prospective employers, such as the wages and benefits offered,
the work environment provided, and the expected satisfaction associated with a potential
job. Thus it is also likely that hours searched will affect ultimate wages received and the
success of employee-employer matches, which in turn should impact the probability of

remaining in the labor force.

 

39 The results presented in Tables 6 and 7 also indicate that for men, while children aged 6-10 may
negatively impact search behavior, very young children, those aged 0-5, are associated with an increase in

34

mule
intensity. fur.
rzt'icted sear
employment.
and other an».
dimension of
neciticaliy at
impact this as

senices whip

 

meets of seal

Altho~
maker 0Utco:
01 linen-1p} 0)T
mote PTOduCr
@hcations \
hm” Oliot
156-1 ll 15 p0
methods of 3‘

mPlOB'ers f..\.

While this paper establishes the significant impact that children have on search
intensity, further research should be directed at a thorough investigation of the impact of
restricted search effort due to child care issues on received wages and the duration of
employment. Future research should also investigate the relationships between children
and other aspects of search behavior. Due to data limitations I explore just one
dimension of the search process — the time devoted to job search. This paper looked
specifically at time spent searching under the assumption that children are most likely to
impact this aspect of search effort due to the costs associated with employing child care
services while looking for work. It is certainly possible that children may affect other
aspects of search effort as well, such as the number of employers contacted.

Although time devoted to job search has been shown to significantly affect labor
market outcomes, Keeley and Robins (1985) found that in terms of reducing the duration
of unemployment, forms of job search associated with direct employer contacts are even
more productive. That is, weekly rates of employer contacts, employer visits and job
applications were more strongly related to unemployment duration than more indirect
aspects of job search, such as weekly hours of search and the number of search methods
used. It is possible that individuals with young children may employ less time-intensive
methods of search, such as sending out a standard cover letter and resume to several
employers from their home, as opposed to "pounding the pavement" while paying for

child care services.40

 

hours searched.

’0 Unfortunately, MCSUI is not the best data set to address these issues. Although respondents were asked
about employers contacted, in Detroit the question was open ended, while in the other cities the question
pertained to the last month of job search.

35

 

q 't V
an! 11".)
Luau.» .

11116715111

51.31.}, «a:

tiliiu L

impact c

r .
”u. ~L'l
“ i 51.;
'i-wy:

sv 1*
‘3‘“ \L

mployr

It is clear that more work is needed to fully understand the relationship between
children and job search behavior. By investigating the impact of children on search
intensity, however, I hope to have at least enriched our understanding of the role that
child care constraints play in the employment process for women and men. The potential
impact of search constraints on mothers trying to make the welfare-to-work transition and
the obvious policy-making implications make this paper is an important first step. The
findings presented here provide significant evidence that policy makers need to be aware
that child care concerns not only affect parents once employed, but that they may also
significantly impede the process of finding a job in the first place and maintaining

employment in the future.

36

APPENDICES

37

1'» L411
by hours .
week. aft.
Atlanta a
hours dic‘
job searc‘;

mention-2t

P3130 d,

 

 

"€55 tic

APPENDIX A
Construction of Variables and Discussion of Outliers

1) Search Intensity (S) — For Detroit, this variable represents search intensity measured

2)

by hours devoted to job search per week. The variable ranges from 0 to 48 hours per
week, after omitting two outlier responses of 72 and 50 (discussed below). In
Atlanta, Boston and Los Angeles, respondents were asked “In total, about how many

hours did you spend looking for work in (the last thirty days/the last month) of your

job search?” I divided the resulting variable by four to get a weekly measure. As
mentioned above, I restricted the sample to those who had looked for work within the
past 30 days and to those with search duration of at least one month. The variable
ranges from 0 to 75 hours per week after omitting 3 observations ranging from 90 to
200 hours per week. In addition, three diagnostic tests, Cook's distance, Welsch
distance and DFITS, were performed to detect potential outliers (see Stata Corp. pp.
372-396). As a result, six observations that failed all three of these tests (based on the
suggested cutoffs) were orrritted: specifically, case ids 1375, 354, 2335, 2747, 4901
and 4570. These observations were quite influential and the model did not fit them
well. Direct inspection of the characteristics of the individuals detected by the three
formal tests was used as a secondary test to determine omission. For example, the
Detroit respondent who reported 50 hours had previously reported that she had not
searched for work within the past 30 days, and thus should not have been asked the
search intensity question.

Nonwage Income — Nonwage income is a categorical variable ranging from 1 to 20

meant to measure income sources not derived from labor in order to give an

38

9a)

3)

indication of one’s means of support while searching for a job. It was created by
subtracting employment earnings from family income. Since family income was
reported as a categorical variable ranging from 1 to 20 (the actual income categories
corresponding to these numbers are available upon request), earnings fiom
employment was transformed into a similarly-defined categorical variable before it
was subtracted from family income. Both family income and earnings fiom
employment pertain to the full year prior to the date of the survey. However, if prior-
year earnings from employment were not reported, annual earnings on current/last
main job were used in its place. (In Detroit, respondents were only asked to report
prior-year earnings if they were not currently working at the time of the survey, but
had worked sometime in the previous two years.)

Hourly Wage (HWAGE) — The hourly wage variable was constructed from an
earnings variable that pertains to the respondent’s current or last main job. If the
respondent did not report earnings as an hourly figure, but used another unit of
measure such as annual earnings, earnings were transformed into an hourly measure
based upon the individual’s usual hours worked per week, if available. In cases for
which weekly hours worked were not reported, the earnings variable was transformed
into an hourly rate assuming the individual worked full time, or 40 hours per week.
Wages less than $2 per hour (78 cases) or exceeding $500 per hour (6 cases) were
classified as outliers. In these cases the hourly wage variable was replaced with a
missing value. In defense of this trimming technique, Angrist and Krueger (1998)
found that "extreme wage values are likely to be mistakes," when they investigated

the impact of trimming outliers using CPS data.

39

4)

5)

The potential for recall bias arises since the earnings variable is based upon a
retrospective question. In Atlanta and Detroit all those who had worked sometime
within the past 5 years where asked the earnings question (referring to their current or
last main job), while in Boston and Los Angeles everyone who had worked sometime

within the past 6 years was asked the earnings question.

“I

Current Wage (WAGE) — WAGE is an hourly wage variable created from the log of

 

HWAGE, using observations on those currently employed only. This conversion was

F" .

meant to mitigate the potential recall bias described above. To control for any
remaining differences in local labor market conditions, dummy variables for
metropolitan area are included in all of the regression analyses involving this
variable.

_P_art time emploment (PTIME) - This variable was constructed from a question that
asked respondents how many hours per week they usually worked at their current or
most recent job (respondents who had not worked within the past five years in
Detroit, or since 1987 in Atlanta, Boston and Los Angeles, were not asked this
question). A binary variable was created equaling one if the respondent reported that

they worked less than 35 hours a week.

40

,7

than
:pmni
15!

shed

hand

hmhn

(I)

aim:
m

“uni

N
balm"1 a

tr“

APPENDIX B
Predicting Wages and Sample Selection Bias

In order to predict a wage for the sample of searchers, a model was first specified
to describe the wages of respondents in the full sample. The log wage equations, run
separately for men and women, took the following form:

(5) WAGEi = Bo + BIZ“ + Uli,

where the vector Z“ includes controls for age, education, race and metropolitan area.
Typically, the estimated coefficients from (5) run on the full sample would be applied to
the characteristics of the sample of searchers to calculate their predicted wages.

It is important to note, however, that WAGE is only observed for a nonrandom
sample of individuals who were employed at the time of the survey, which potentially
gives rise to the classic sample selection problem. If individuals who are employed are
not representative of the general population, failure to control for the differences will lead
to biased estimates of the coefficients in (5).41

Since current wages are observed only if the individual is employed, the
employment selection equation will take the form
(6) EMP, = 1 [ZS + U2, > 0],
where Um has a standard normal distribution, and is potentially correlated with U”.
Sample selection bias will be a problem -- that is, OLS applied to equation (5) will lead to

inconsistent results -- if U1, and U2; are correlated.

 

 

4| - . . . . -
Prior research has found that sample selectron bras rs usually only an issue for women in such models.

41

Toc

 

 

 

correction p
such models
affect fifllp‘a'
this analysis
ot'children ;
Appendix C
specified by
alternative 1;

sensitive to t

 

results of 1h.
(11 and (2) 0
men and “-0
Hinds} (5 l pr

2.2:: '9?
\» ..,“~
{31.3 c4\35l1:1
‘\‘ -“ .
11 .

To correct for potential sample selection bias, I employ the Heckman selection
correction procedure“, which provides consistent, asymptotically efficient estimates for
such models. In order to identify the parameters of the model, variables that strongly
affect employment status but not the wage offer must be included in Z, but not in Z“. In
this analysis, Z includes a dummy variable for marital status and measures of the number
of children in the household aged 0-5, 6-10 and 11-17 for identification.43 Column (3) of F

Appendix C presents the results from the maximum likelihood estimation of the model

 

specified by equations (5) and (6) utilizing the Heckman correction procedure. As an . .
alternative test for sample selection bias (since the Heckman procedure is known to be
sensitive to misspecification), I also employed a two-step least squares procedure; the
results of the second step are found in column (4).44 For comparison purposes, columns
(1) and (2) of Appendix C present the results from OLS estimation of equation (5) for
men and women, uncorrected for sample selection bias. The estimated coefficients from
model (5) presented in columns (1) and (3) of Appendix C were used to construct the
predicted wage, PWAGE, for the sample of searchers. However, the results of Appendix
C indicate that there is no strong evidence sample selection bias; the two correction

procedures resulted in little change in the size and significance of estimated coefficients

 

’2 See Heckman (1979) for a more detailed discussion of the problem, and StataCorp (1997, pp. 187-195)
for a description of the correction procedure used in this analysis.
‘3 There is much controversy in the literature regarding the appropriateness of variables typically used for
identification. Obviously, some may argue that the variables included in (5) for identification should also
be included in the wage equation (5). Ideally, a measure of nonwage income would be used for
identification, however, this variable is missing for approximately 17 percent of the full sarrrple (and is
potentially plagued with a similar self-selection problem).

In the first stage, a probit model is specified for employment, which includes the complete set of
regressors used in (5), along with a marriage dummy and indicators of the number of children in the
household for identification. This employment probit was run on the full sample, and used to compute the

inverse Mills ratio. In the second stage, the inverse Mills ratio was included as a Z,- regressor in the log
wage regression (5), which was run on the sub-sample of workers. Several different specifications were
tested, and in virtually every case there was no evidence of selectivity bias.

42

compared t

the estimatr

 

 

 

 

compared to the model that does not control for sample selection bias, and in both cases

the estimated coefficient on the inverse Mills ratio is insignificant.

43

APPENDIX C

Log-of-Wages Equations by Gender, Full Sample

 

 

 

 

 

Heckman Two-Step
Log Wage Model Correction Model Correction Model
Men Women Women Women
(1) (2) (3) (4)
Age .057 .048 .048 .049
(.011) (.014) (.005) (.013)
Age squared -.0005 -.0005 -.0005 -.0005
(.0001) (.0002) (.0001) (.0002)
Atlanta -.150 -.129 -.129 -.120
(.063) (.045) (.033) (.049)
Boston -.081 .043 .047 .057
(.067) (.061) (.028) (.064)
Detroit -.023 -.165 -.l62 -.154
(.052) (.049) (.028) (.051)
Black -.157 -.056 -.058 -.057
(.050) (.040) (.029) (.040)
Asian -. 174 .061 .054 .057
(.075) (.091) (.055) (.092)
Hispanic -.317 -.237 -.237 -.235
(.061) (.059) (.032) (.060)
High school/GED .250 .377 .383 .396
(.049) (.058) (.037) (.070)
Associates/
Vocational/Trade .365 .523 534 .552
(.060) (.063) (.044) (.086)
College degree .536 .722 .723 .752
(.065) (.060) (.043) (.087)
Graduate degree .644 .772 .786 .809
(.101) (.096) (.053) (.126)
Inverse Mills - - .055 .064
(.099) (.118)
Intercept 1.011 .949 .918 .865
(.214) (.252) (.124) (.281)
R2 .346 .282 .282 .282

 

Note: Sample consists of non-retired respondents. Sample sizes are 2,027 and 1,919 for women and men, respectively.
All estimates are sample weighted. Robust standard errors appear in parentheses and are adjusted for survey design.
Omitted categories for region, race and education, are Los Angeles, white and high school drop out, respectively.

44

APPENDIX D
Coefficients (Standard Errors) on Control Variables for Table 8

 

 

 

Detroit Atlanta-Boston-Los Angeles
(1) (2) (3) (4)
Presence of
children 0-5 3.41 "' 4.30 3.32 4.31"
(1.78) (2.16) (2.57) (2.22)
Presence of
children 6-10 -3.81* -4.42** -3.38 -3.15*
(1.96) (2.04) (2.10) (1.86)
Nonwage income -.712** -.680** .302 .226
(.318) (.337) (.332) (.323)
Employed -7.79*** -7.66*** -4.50*** -4.04**
(1.80) (1.82) (1.49) (1.73)
(Nonwage income) x .935*** .921 *** -.271 -.223
(employed) (.354) (.345) (.413) (.398)
Adults 36-55 1.01 1.65 1.52 1.71
(1.34) (1.40) (1.39) (1.41)
Young adults 19-35 2.23" 223*" -.613 -.275
(1.05) (1.03) (.600) (.603)
Part timejob -1.33 -l.27 -3.84*** -3.71***
(1.58) (1.53) (1.47) (1.45)
AFDC - - - -2.9l **
(1.45)
UI benefits - - - 3.35"
(1.80)
Predicted (log) wage - 4.86 - 2.14
(4.70) (3.84)
Controls:
Education Y Y Y Y
Metropolitan area - - N Y
Constant 13.32 2.53 11.46 4.82
R2 .238 .242 .210 .254

 

Note: Sample sizes are 201, 200, 720, 714 for specifications 1-4, respectively. Regressions include a missing value
dummy variable for part time. All estimates are sample weighted. Robust standard errors appear in parentheses and are
adjusted for survey design. Omitted categories are high school drop out for education and Los Angeles for metropolitan
area (‘ significant at the 10% level; ” significant at the 5% level; "* significant at the 1% level.)

45

 

REFERENCES

46

 

AllngSl. .11;
Economics
lune.

Barton. 10‘:
oiHuer. .

Journal Of

  
 

Becker. G.-

01.
‘
.11.

[Pdl'ersj[\'

BEL Dasi.

and L'nern;

Casper, Ly
ot the C en

B'\
M63111 of

Cornelly
FOrce Pan

. Liking."

Foundation

REFERENCES

Angrist, Joshua D. and Alan B. Krueger (1998), "Empirical Strategies in Labor
Economics," Working Paper #401, Princeton University, Industrial Relations Section,
June.

Barron, John, and Wesley Mellow (1979), “Search Effort in the Labor Market,” Journal
of Human Resources, 14(3): 389-404.

(1981), “Changes in the Labor Force Status among the Unemployed,”
Journal of Human Resources, 16(3): 427-44],

 

Becker, Gary (1965), “A Theory of the Allocation of Time,” Economic Journal, 75: 493-
517.

(1991), A Treatise on the Family, Cambridge, Massachusetts: Harvard
University Press.

 

Blau, David M., and Philip K. Robins (1990), “Job Search Outcomes for the Employed
and Unemployed,” Journal of Political Economy, 98(3): 637-655.

Casper, Lynne M. (1995), “What Does it Cost to Mind Our Preschoolers?” US. Bureau
of the Census, Current Population Reports, P-70, No. 52, Washington, DC.

(1997), “Who’s Minding Our Preschoolers? Fall 1994 (Update),” US.
Bureau of the Census, Current Population Reports, P-70, No. 62, Washington, DC.

 

Connelly, Rachel (1992), “The Effect of Child Care Costs on Married Women’s Labor
Force Participation,” Review of Economics and Statistics, 74(1): 83-90.

(1991), "The Importance of Child Care Costs to Women's Decision
Making," in David M. Blau, ed. The Economics of Child Care, New York: Russell Sage
Foundation.

 

Greene, William (1990), Econometric Analysis, New York: Macmillan Publishing
Company.

Gruber, Jonathan, “The Incidence of Mandated Matemity Benefits,” American Economic
Review, 84(3):622-641.

Hamermesh, Daniel S., and Albert Rees (1993), The Economics of Work and Pay, (5th
edition), New York: Harper Collins Publishers.

Heckman, James J. (1979), “Sample Selection Bias as a Specification Error,”
Econometrica, 47(1): 153-161.

47

‘ [v—w—is ”—1

    

Hoilerth. 5.

Sentinel L
Holzer. He?
Labor Ecru

Holzer. Hz;
Employme:
August: 83

Hort 3dr. f I
Data" Mon

Keeley. .\l;

Requiremc:
1‘. q

3: ~36].

Kiliingsyyol
Ashenfelter
Elseyier Sc

 

lisker. Elle
A Profile or
35“ Jersey

1K1mull. l;
Millie”. 1V
Moninsert

P9371109,”
Fm 5‘ ide

”Won: (
“Downer
ofHuman A

Hofferth, Sandra L., April Brayfield, Sharon Deich, and Pamela Holcomb (1990),
National Child Care Survey, 1990, Washington, DC: The Urban Institute Press.

Holzer, Harry (1987), “Job Search by Employed and Unemployed Youth,” Journal of
Labor Economics, 6(1): 1-20.

Holzer, Harry J. and Keith R. Ihlanfeldt (1998), “Customer Discrimination and
Employment Outcomes for Minority Workers,” The Quarterly Journal of Economics,
August: 835-867.

Horvath, Francis W. (1982), “Forgotten Unemployment: Recall Bias in Retrospective
Data,” Monthly Labor Review, 105(3): 40-44.

Keeley, Michael C. and Philip K. Robins (1985), "Government Programs, Job Search
Requirements, and the Duration of Unemployment," Journal of Labor Economics, 3(3):
337-362.

Killingsworth, Mark R. and James J. Heckman (1986), “Female Labor Supply,” in
Ashenfelter, O. and Richard Layard, eds., Handbook of labor Economics, New York:
Elsevier Science Publishers.

Kimmel, Jean (1996), “Child Care Costs as a Barrier to Employment for Single and
Married Mothers,” typescript, W.E. Upjohn Institute for Employment Research
(forthcoming in Review of Economics and Statistics).

Kisker, Ellen Eliason, Sandra L. Hofferth, Debra A. Phillips, Elizabeth Farquhar (1991),
A Profile of Child Care Settings: Early Education and Care in 1990, Vol. I, Princeton,
New Jersey: Mathematica Policy Research, Inc.

Klerman, Jacob Alex and Arleen Leibowitz (1994), Employment Continuity Among New
Mothers, typescript, RAND.

Mortensen, Dale T. (1977), "Unemployment Insurance and Job Search Decisions,"
Industrial and Labor Relations Review, 30(3): 505-517.

Ribar, David (1992), “Child Care and the Labor Supply of Married Women: Reduced
Form Evidence,” Journal of Human Resources, 27(1): 134-165.

StataCorp (1997), Stata Statistical Software: Release 5.0, College Station, TX: Stata
Corporation.

Tannery, Frederick J. (1983), “Search Effort and Unemployment Reconsidered,” Journal

of Human Resources, 18(3): 432-440.

48

.J u i— m.
I 7 r '
.. .

175, Bure.
olition). V‘

TS D61);
, I, 4'.
110134.11». I
<hnp: we.

LS. Dfpsi
Families 5
Percent." ti
laccessed .'

17.8. Dep .1:
Announce
the ll'elfar-

mitts;

k.
\r

Foundation

 

US. Bureau of the Census (1998), Statistical Abstract of the United States: 1998, (118th
edition), Washington, DC.

US. Department of Health and Human Services (1996), "Fact Sheet: The Personal
Responsibility and Work Opportunity Reconciliation Act of 1996," September 1996
<http://www.acf.dhhs.gov/programs/opa/facts/prwora96.htm> (accessed 2/17/99).

US. Department of Health and Human Services (1999a), "Only 10 Percent of Eligible
Families get Child Care Help, New Report Shows: State Rates Vary From 4 to 24
Percent," October 19, 1999 <http://www.acf.dhhs.gov/news/press/1999/cc98.htm>
(accessed 3/14/00).

US. Department of Health and Human Services (1999b), "President Clinton Will
Announce Record Numbers of People on Welfare are Working as Businesses Hire From
the Welfare Rolls," August 3, 1999 <http://www.acf.dhhs. gov/news/press/ 1999/
w2w0803.htm> (accessed 3/14/00).

Waldfogel, Jane (1997), “Working Mothers Then and Now: A Cross-Cohort Analysis of
the Effects of Maternity Leave on Women’s Pay,” in Blau, Francine D. and Ronald G.
Ehrenberg, eds., Gender and Family Issues in the Workplace, New York: Russell Sage
Foundation.

49

‘ . . . va
! ‘7 m
i. a-
.4

The R:

Childrc

 

Figure l
The Relationships between Children and Labor Market Outcomes of Women
and the Predicted Job Search Link

 

 

( — ) Labor Force

Participation
&

Employment

 

 

 

(+)

 

 

 

 

 

 

 

Search Effort I + Unemployment
( _) ~ (_ ) Duration

 

 

 

(+)

 

( _ ) Wages and
Benefits

 

 

 

 

50

 

 

 

Not look OTI

Turn down

Not partici;
school trair

 

Q1411 or be 1

Be late for

BC absent .

(”“1783 he

Table 1
How Child Care Affects Employment:
Perceived Child Care Constraints in Atlanta, Boston and Los Angeles by Sex

 

“In past 12 months, have child

 

m concerns caused you to”: Women Men
Not look or apply for work .320 .083
(.024) (.017)
Turn down a job offer .127 .038
(.016) (.010)
Not participate in
school/training . 194 .069
(.020) (.015)
Quit or be fired .063 .014
(-011) (.007)
Be late for work .352 .268
(.032) (.028)
Be absent fiom work .394 .259
(.032) (.028)
Change hours of work .340 .217
(.031) (.027)
Lose out on promotion/raise .052 .030
(.014) (.012)

 

Note: Respondents with a child under 18 at home were asked questions generating the first four rows of
data. Sample sizes are 2,197 for women and 1,000 men, with some cells containing fewer observations
due to missing values. Respondents with a child less than 18 at home who had worked in the last 12
months were asked questions generating the last four rows of data. Sample sizes for these rows are

1,213 for women and 880 for men, with some cells containing fewer observations due to missing values.

Standard errors appear in parentheses. All results are sample-weighted and standard errors are adjusted
for stratification and clustering of survey design.

51

A - .“A‘J-i

Emu—2‘:

Perceptior.

501151131111:

Cost rr.

Quality I
Amour;
Getting

Ability
force or

 

_ 11
5:15:30, for nl
Sum Some-

 

Table 2
How Child Care Affects Employment:
Perceived Child Care Constraints in Detroit by Sex

 

 

 

Women Men
Perception of child care .310 .250
constraints on employment (.023) (.029)
Cost mentioned .615 .390
(.046) (.066)
Quality mentioned .204 .104
(.03 8) (.039)
Perceived constraints on:
Amount & timing of work .254 .316
Getting job or choice of job .165 .156
Ability to enter the labor
force or maintain job .432 .347
Other .149 .181

 

Note: Survey question asked of respondents with children under 18 was the following: “Has the cost,
availability or quality of child care ever influenced your employment or that of your (spouse/partner) in
any way?” Data in the last four rows contain responses to the following question: “In what ways did
these issues influence you or your (spouse’s/partner’s) employment?” Sample sizes are 590 for women
and 305 for men in rows one through three, and 143 for women and 52 for men in rows four through
seven. Some cells contain fewer observations due to missing values. Other category includes responses
discussing the difficulty of raising children and working and other non-specific responses. Standard
errors appear in parentheses. All results are sample-weighted and standard errors are adjusted for
stratification and clustering of survey design.

52

Table 3
Hours Searched per Week by Sex and Presence of Children

 

 

 

 

 

 

A. Detroit Child under 6 Child 6-10 No children <10
Women 3.98 3.40 10.59
(.795) (.533) (1.94)
Men 6.67 4.60 8.97
(1.52) (1.53) (1.41) i
B. Atlanta, Boston
and Los Angeles Child under 6 Child 6-10 No children <10
Women 3.36 5.98 6.00
(.552) (1.58) (1.09)
Men 11.10 6.85 9.81
(2.36) (1.94) (1.91)

 

Note: Sample sizes are 111 and 92 for women and men in Detroit, respectively, and 401 and 319 for
women and men in the other cities, respectively. Standard errors appear in parentheses. All results are
sample-weighted and standard errors are adjusted for stratification and clustering of survey design.

53

 

 

A, “one

“in past 1

car cone ~.

Not look I.
Turn dost:
Not part:
school L'sg'
Quit or be

361316 for

BE absent

 

 

 

Change ho

LOSE 0m 0.

Table 4

Perceived Child Care Constraints in Atlanta, Boston and Los Angeles

by Sex and Presence of Children

 

 

A. Women
Child under 6 Child 6-10 Child 11-17
“In past 12 months, have child at home at home at home
care concerns caused you to”:
Not look or apply for work .454 .411 .236
(.037) (.041) (.029)
Turn down a job offer .153 .149 .090
(.023) (.028) (.021)
Not participate in
school/training .269 .247 .170
(.028) (.035) (.033)
Quit or be fired .090 .046 .042
(.020) (.015) (.014)
Be late for work .462 .444 .236
(.050) (.054) (.035)
Be absent from work .526 .479 .268
(.051) (.052) (.038)
Change hours of work .428 .476 .282
(.047) (.051) (.041)
Lose out on promotion/raise .083 .065 .041
(.025) (.027) (.020)

 

Note: Respondents with a child under 18 at home were asked questions generating the first our rows of
data. Sample size is 2,197 for women, with some rows containing fewer observations due to missing
values. Respondents with a child less than 18 at home who had worked in the last 12 months were asked
questions generating the last four rows of data. Sample size for these rows is 1,213, with some rows
containing fewer observations due to missing values. Standard errors appear in parentheses. All results
are sample-weighted and standard errors are adjusted for stratification and clustering of survey design.

54

Table 4 (cont’d)

 

 

B. Men
Child under 6
“In past 12 months, have child at home
care concerns caused you to”:
Not look or apply for work . .073
(.022)
Turn down a job offer .038
(.013)
Not participate in
school/training .091
(.024)
Quit or be fired .006
(.003)
Be late for work .325
(023)
Be absent from work .293
(.022)
Change hours of work .255
(.021)
Lose out on promotion/raise .031
(.008)

Child 6-10
suduuns

.151
(041)

.057
(026)

.070
(026)

.037
(023)

.332
(028)

.307
(027)

.287
(026)

.066
(015)

Child 11-17
at home

.075
(.031)

.024
(.015)

.036
(017)

.013
(011)

.224
(023)

.225
(.023)

.163
(020)

.001
(001)

 

Note: Respondents with a child under 18 at home were asked questions generating the first four rows of
data. Sarrrple size is 1,000 men, with some rows containing fewer observations due to missing values.
Respondents with a child less than 18 at home who had worked in the last 12 months were asked
questions generating the last four rows of data. Sample size for these rows is 880, with some rows
containing fewer observations due to missing values. Standard errors appear in parentheses. All results
are sample-weighted and standard errors are adjusted for stratification and clustering of survey design.

55

Table 5

Means and Standard Errors of Variables Used for Job Searchers
by Sex and Metropolitan Area

 

Hours searched
per week

Presence of
children 0-5

Presence of
children 6-10

Employed

Nonwage income

# of adults 35-64

# of young adults
18-34

High school drop
out

High school/GED
Associates/

Vocational/Trade

Bachelor’s degree

Advanced degree

Log of predicted
Wage

Atlanta-Boston-Los Angeles

 

Women

75H
(L42)

.243
(055)

.232
(.053)

.614
(066)

2198
(546)

.236
(.097)

.364
(105)

.218
(060)

.448
(065)

.128
(037)

.139
(054)

.067
(034)

2.17
(.047)

Men

833
(L25)

.175
(042)

.157
(039)

.549
(.062)

416
(686)

.490
(123)

.353
(061)

.065
(024)

.599
(068)

.144
(052)

.133
(.046)

.059
(029)

256
(045)

56

Women

558
(816)

.171
(038)

.210
(054)

.398
(070)

357
(396)

.341
(133)

.530
(106)

.145
(042)

.324
(.053)

.249
(071)

.198
(046)

.083
(027)

236
(031)

Men

8.75
(1.48)

.133
(036)

.174
(044)

.508
(054)

352
(533)

.371
(107)

.595
(151)

.175
(033)

.437
(059)

.099
(.030)

.190
(.044)

.099
(031)

2.53
(.044)

‘ l
lienpoj

T153311?

AFDC

.lzfat'a

Boston

los hug

.\
Note 5;
Ellis 1

«4,. .. l
‘- .5,
“at“.

Table 5 (cont’d)

 

Unemployment
insurance benefits
AFDC

Atlanta

Boston

Los Angeles

- .248
(046)

- .143
(.039)

- .136
(027)

- .360
(063)

- .504
(.057)

.324
(050)

.042
(.019)

.080
(017)

.330
(060)

.590
(.057)

 

Note: Sample sizes are 111 and 92 for women and men in Detroit, respectively, and 401 and 319 for

women and men in the other cities, respectively. Some cells contain fewer observations due to missing
values. Standard errors appear in parentheses. All results are sample-weighted and standard errors are
adjusted for stratification and clustering of survey design.

57

F'

heserce

1 11 ‘
In. y
(ninth

P’)’ .‘J
.5. bid.

1 l ‘
,“1 1'” |
buUtW¥ll

IV 1‘ “
v

I
vq I‘--
1....~,',l,\§
I o

lhorn

' a
1
U

'V‘ V m
lei-3.01
' 0

rats 2

ion.)

Q.
».
p...
4.
““L-IT

(Olitg

4'. y
.' ‘ an
“-1.4"

Table 6
Search Intensity Equations by Sex, Detroit Sample

 

 

 

Women Men
(1) (2) (3) (4)
Presence of -4.84*** -4.67*** 573*" 669*"
children 0-5 (1.32) (1.53) (1.33) (1.53)
Presence of -5.32*** -5.39*** -7.28*** -7.99***
children 6-10 (1.57) (1.55) (1.35) (1.42) g-
Nonwage income -.960*** -.976*** -.550** -.362 l
(.285) (.339) (.219) (.255) 1
Employed -7.40** -7.57** -10.29*** -9.50*** 1
(3.10) (3.54) (1.87) (1.75) l.
(Nonwage income) x .705* .722 125*" 1.05***
(employed) (.424) (.453) (.355) (.361)
Adults 36-55 4.94" 5.02“ -.856 -.l93
(1.96) (2.13) (1.19) (1.38)
Young adults 19-35 .838 .852 334* 3.37*
(.800) (.795) (1.77) (1.75)
Part time job 2.30 2.30 -5.83*** —5.71***
(1.73) (1.73) (2.05) (2.04)
High school/GED .480 -.123 -2.75 -4.41*
(2.22) (3.52) (2.41) (2.68)
Associates degree/ .433 -.415 2.20 -.340
Vocational/Trade (2.25) (4.85) (3.39) (3.02)
College degree 9.09 8.05 -4.40 -9.05*"'
(5.74) (9.76) (2.87) (3.85)
Advanced degree -4.66* -6. 16 .974 -3.66
(2.53) (8.32) (3.49) (4.51)
Predicted (log) wage - 1.56 - 4.04
(9.09) (4.08)
Constant 13.58 10.92 14.50 5.28
R2 .381 .381 .471 .487

 

Note: Sample sizes are 109 and 91 for women and men, respectively. Regressions include a missing
value dummy variable for the part time. All estimates are sample weighted. Robust standard errors

appear in parentheses and are adjusted for survey design. Omitted category for education is high school
drop out. (" significant at the 10% level; ** significant at the 5% level; **"‘ significant at the 1% level.)

58

 

Table 6
Search Intensity Equations by Sex, Detroit Sample

 

 

 

 

Women Men
(1) (2) (3) (4)
Presence of -4.84*** -4.67*** 573*“ 669*"
children 0-5 (1.32) (1 .53) (1.33) (1.53)
Presence of -5.32*** -5.39*** -7.28*** -7.99***
children 6-10 (1.57) (1.55) (1.35) (1.42)
Nonwage income -.960*** -.976*** -.550** -.362
(.285) (.339) (.219) (.255)
Employed -7.40** -7.57** -10.29*** -9.50***
(3.10) (3.54) (1.87) (1.75) 'e.‘
(Nonwage income) x .705* .722 1.25*** 1.05“”
(employed) (.424) (.453) (.355) (.361)
Adults 36-55 4.94" 5.02“ -.856 -.193
(1.96) (2.13) (1.19) (1.38)
Young adults 19-35 .838 .852 3.34* 3.37“
(.800) (.795) (1.77) (1.75)
Part timejob 2.30 2.30 -5.83*** -5.7l***
(1.73) (1 .73) (2.05) (2.04)
High school/GED .480 -.123 -2.75 -4.4l"'
(2.22) (3.52) (2.41) (2.68)
Associates degree/ .433 -.415 2.20 -.340
Vocational/Trade (2.25) (4.85) (3.39) (3.02)
College degree 9.09 8.05 -4.40 -9.05**
(5.74) (9.76) (2.87) (3.85)
Advanced degree -4.66* -6. l 6 .974 -3.66
(2.53) (8.32) (3.49) (4.51)
Predicted (log) wage - 1.56 - 4.04
(9.09) (4.08)
Constant 13.58 10.92 14.50 5.28
R2 .381 .381 .471 .487

 

Note: Sample sizes are 109 and 91 for women and men, respectively. Regressions include a missing
value dummy variable for the part time. All estimates are sample weighted. Robust standard errors
appear in parentheses and are adjusted for survey design. Omitted category for education is high school
drop out. (* significant at the 10% level; *"' significant at the 5% level; *** significant at the 1% level.)

58

Table 7
Search Intensity Equations by Sex, Atlanta-Boston-Los Angeles Sample

 

 

 

 

Women
(1) (2) (3) (4)
Presence of
children 0-5 -4.15*** -3.86"* 4.65* 5.77""
(1.13) (.993) (2.40) (2.19)
Presence of
children 6-10 1.40 1.72 -2.91 -2.42
(1.70) (1.63) (2.03) (2.08)
Nonwage income -. 145 -. l 73 .782 .659
(.199) (.223) (.461) (.458)
Employed -5.47"* -5.02”* -3.38* -3.60
(1.80) (1.80) (1.85) (2.32)
(Nonwage income) x
(employed) .021 -.O45 -. 822 -.834
(.330) (.357) (.541) (.541)
Adults 36-55 -.214 .069 3.18‘ 3.92"
(.925) (1.06) (1.68) (1.78)
Young adults 19-35 -.490 -.196 -.667 .195
(.571) (.623) (.996) (.884)
Part time job -2.29 -2.43* 599*“ -4.90“
(1.51) (1 .35) (2.05) (2.08)
AFDC - -1.51 - -2.36
(1.48) (3.12)
UI benefits - 3.65" - 2.69
(1.77) (2.45)
Predicted (log) wage - -l .88 - 2.20
(3.25) (4.81)
Controls:
Education Y Y Y
Metropolitan area N Y N Y
Constant 7.97 10.41 11.08 3.72
R2 .242 .298 .314 .371

 

Note: Sample sizes are 401 and 319 for women and men, respectively. Regressions include a missing value dummy
variable for part time. All estimates are sample weighted. Robust standard errors appear in parentheses and are
adjusted for survey design. Omitted categories are high school drop out for education and Los Angeles for
metropolitan area. (‘I significant at the 10% level; " significant at the 5% level; ”‘" significant at the 1% level.)

59

 

Table 8
Difference-in-Differences Estimates of
the Effects of Children on Search Intensity

 

 

 

A. Detroit Sample Difference-in-differences (DD) estimates:
(1) (2)

(Children 0-5) x (female parent) -9.05*** -9.14***
(2.26) (233)

(Children 6-10) x (female parent) -1.28 -.821
(2.81) (3.00)

R2 .238 .242

B. Atlanta-Boston-Los Angeles

Sample

(Children 0-5) x (female parent) -7.88*** -7.96*"‘*
(2.87) (2.59)

(Children 6-10) x (female parent) 5.06* 4.58
(2.81) (2.80)

R2 .210 .254

 

Note: Sample sizes are 109 and 91 for women and men, respectively. Included controls correspond to

those used in Tables 6 and 7, and the estimated coefficients on these controls are reported in Appendix D.

All estimates are sample weighted. Robust standard errors appear in parentheses and are adjusted for
survey design. Omitted category for education is high school drop out. (* Significant at the 10% level;
** significant at the 5% level; "* significant at the 1% level.)

60

' AM...” . _

Chapter 2
RACIAL AND ETHNIC JOB SEGREGATION:
ITS CAUSES AND CONSEQUENCES
Introduction

Although considerable progress has been made in reducing racial inequality since
passage of the Civil Rights Act of 1964, the employment and earnings of blacks and other
minority groups continue to fall short of their white counterparts. Most empirical studies
find a residual wage gap between whites and minorities even after controlling for
standard productivity proxies such as education and experience. Several of the dominant
theories put forth to explain these persistent wage differentials — including Becker’s
(1971) “taste” models of discrimination and the “spatial mismatch hypothesis” — imply
that minorities are likely to be segregated from whites in the labor market. Just as
women are thought to be crowded into lower-paying “female occupations,” minority
groups may be segregated into jobs with lower wages and benefits for a variety of
reasons. This study investigates the importance of racial and ethnic job segregation on
hourly earnings and the likelihood of receiving various employment benefits. The study
also explores the factors likely to contribute to the job segregation of blacks and
Hispanics in the labor market.

While there is a large body of literature on the effect of sex occupational
segregation on the wage gap between women and men (e. g., Macpherson and Hirsh,
1995), evidence on the effects of racial and ethnic segregation on labor market outcomes
for minorities is fairly limited. A potential reason for this shortfall is that prior studies

find that occupational segregation between races and ethnic groups is significantly less

61

 

pronounced than occupational segregation between the sexes (Watts, 1995; Bayard et al.,
1999). For example, Sorensen (1989) investigates the effect of occupational segregation
by sex and race on eamings differentials between demographic groups. She finds that
occupational segregation by sex significantly affects earnings, accounting for roughly 20
percent of the wage gap between men and women (both white and minority), and 3
percent of the wage gap between white and minority men. However, occupational
segregation by race is not found to be a significant factor influencing earnings.

Bayard et a1. (1999) study the wage effects of racial and ethnic segregation by
industry, occupation, establishment, and job, in an attempt to explain the larger wage
gaps found for men than for women.1 While they do not find racial and ethnic
segregation to be pervasive along occupation or industry lines, they find a great deal of
segregation by establishment and at the job level. They conclude that the negative
segregation effect on minority wages stems primarily from j ob-level segregation.
Carrington and Troske (1998) use a measure of racial composition at the establishment
level to decompose the black-white wage gap in the manufacturing industry into within-
and between-plant components. The authors find that most of the racial wage gap among
men is accounted for by within-plant differences in pay. A shortcoming of these two
studies, however, is that the data used provide only limited productivity-related controls;

neither Carrington and Troske (1998) nor Bayard et a1. (1999) control for potentially

 

' Bayard et al. (1999) use matched employer-errrployee data created from the Sample Detail File (SDF) of
the 1990 Decennial Census One-in Six Long Form and the 1990 Standard Statistical Establishment List
(SSEL). Segregation is measured as the percentages black and Hispanic in an individual’s industry,
occupation, establishment, and job cell. The percentages in the occupation and industry are estimated from
the full SDF sample, so measurement error is unlikely. However, as the authors explain, measurement
error in the estimates of establishment and job segregation “could be sizable,” since they are based on the
matched data (see Bayard et al., 1998, for a detailed description of the matching process). Only 19.4
workers are matched to an establishment on average, so their job-level segregation estimates, in particular,

62

important wage determinants, such as tenure, job tasks, job-specific experience, and
whether one is covered by a collective bargaining agreement.2

Utilizing slightly more extensive controls, including union membership and
tenure, Hirsch and Schumacher (1992) investigate the effect of racial segregation on
wage rates and the black-white wage gap using a measure of racial density within
industry-occupation-region cells (measured by the ratio of black workers to the sum of
white and black workers in these cells). They find that wages of both white and black
workers decrease with respect to their measure of racial density, but that the black-white
wage gap does not vary systematically with respect to racial density by industry-
occupation-region group.

Hirsch and Macpherson (1994) illustrate the importance of including adequate
controls for skill differences. Using a national estimate of the racial composition of
occupations, they find that the negative effect of occupational race segregation on wages
is sharply reduced when measures of occupational skill level (e.g., required years of
training, job tenure, computer usage) are included. Moreover, the authors are able to
virtually eliminate the negative effect of racial segregation at the occupation level by
estimating wage-change equations that difference out unmeasured individual skill
differences fixed over time.

The study here contributes to the current body of research in several ways. In

contrast to most prior work (with the exception of Bayard et al., 1999), I use racial and

 

 

are often based on a small number of observations. The authors thus use highly aggregated occupations
(based on 13 Census occupations in most specifications) to measure job-level segregation.

2 Experience and tenure, in particular, are important characteristics that should be included in wage
analyses investigating discrimination. Bratsberg and Terrel (1998) shed light on the persistence of black-
white wage differentials by investigating the differing returns to tenure and general experience. They find
that returns to general experience for black workers trail those for whites, but black workers earn equal if
not higher returns to tenure than white workers earn.

63

ethnic composition information measured at the job level within a firm. Given the results
of past studies that find firm- and job-level segregation by race and ethnicity affects
earnings differentials to a much greater extent than occupational segregation, job-level
segregation information is preferred over occupation-level information. Furthermore,
some of the processes affecting racial and ethnic segregation are likely to occur within a
firm, rather than simply at the point of sorting into an occupation (as discussed below).
The second contribution of this study relates to the breadth of the data used, which were
drawn from a relatively new survey of households administered between 1992 and 1994
in three major metropolitan areas. These data allow wage differentials by demographic
group to be estimated using an extensive collection of productivity-related measures not
available to most previous researchers who have studied this issue, such as job tasks,
supervisory authority, and job-specific experience. Even after controlling for differences
in personal human capital and job characteristics, I find racial and ethnic job segregation
to be an important contributor to the lower wages paid to blacks and Hispanics than to
similar whites. .\

This study also examines the potential impact of racial and ethnic segregation on
the likelihood of receiving various employment benefits. The data used here contain
information on whether jobs provide a retirement plan, paid sick leave, and individual or
family health insurance. Job segregation is found to play a smaller role in explaining
differences between minorities and whites in the number of employment benefits

received than it does in explaining wage differentials. The results indicate that men

working in jobs with mostly Hispanics are less likely to receive retirement benefits and

64

health insurance for themselves and their family, while women working in minority
dominated jobs are actually more likely to receive health insurance.

Finally, this is the first study to my knowledge to explore the potential causes of
racial and ethnic job segregation. The results show that while minorities who reside in
more segregated neighborhoods are significantly more likely to work in segregated jobs,
those who commute longer distances to work are less likely to work in a segregated job,
two findings consistent with Kain’s (1968) “spatial mismatch hypothesis.” I also find
that blacks and Hispanics who work in larger firms are less likely to be in segregated
jobs, and that English fluency and citizenship status are strongly associated with the
likelihood of job segregation for Hispanics.

The remainder of this paper proceeds as follows. I begin by discussing the
leading explanations for wage differentials between minorities and whites, and then
provide a loose theoretical framework for the empirical work that follows. Next, I
describe the data drawn from the Multi-City Study of Urban Inequality and discuss the
approach used to estimate the impact of j ob segregation on labor market outcomes and to
investigate the likely sources of racial and ethnic segregation. Finally, I report the
empirical results, and then conclude with some discussion of the implications of the

results.

Theoretical Framework
Most labor economists agree on the existence of wage gaps by race and ethnicity,
yet often disagree on the source of the gaps. While relatively little attention has been

devoted to how and why wages vary with the racial and ethnic composition of jobs,

65

 

numerous studies have put forth theories in an attempt to explain the persistence of wage
differentials between whites and minorities with similarrmeasured characteristics.
Although I do not attempt to provide an exhaustive review of all these theories, several of
the alternative models are relevant to this study because they imply that discrimination
will lead to segregation of minorities fiom whites in the labor market.3 A few of these
models also shed light on the relationship between wages and labor market segregation.
Perhaps the most widely recognized theories of labor market discrimination used
to explain wage gaps are Becker’s (1971) taste models, which identify the prejudices of
employers, consumers, and employees as sources of discrimination. The employer
discrimination model predicts that discriminatory employers will offer to pay minorities
less than equally productive whites because they incur a psychic cost associated with
hiring minorities. Segregation is likely to result as discriminating employers hire mostly
or only whites, depending on their taste for discrimination, and nondiscriminatory
employers hire otherwise identical minorities to maximize profits. Of course, an
employer’s taste for discrimination may differ depending on the particular job in
question. A discriminating employer, for example, may have a larger distaste for hiring a

minority as a manager than for hiring a minority as a blue-collar laborer. This may

 

3 The employee taste discrimination model is not discussed here because it is not supported in the data.
Widespread employee discrimination implies a wage premium for white workers who work alongside
minority workers. Likewise, statistical discrimination models, which emphasize the role of imperfect
information about individual worker productivity, are not addressed in this paper since they differ in their
implications about the effects of racial composition (see Hirsch and Macpherson, 1994). Finally, a
discussion of the theory of “occupational crowding "is omitted here since it is thought to explain the
segregation of women from men in the labor market, and past studies do not find empirical support for the
model in the context of racial segregation (see Hirsch and Schumacher, 1992; Hirsch and Macpherson,
1994). Moreover, an underlying assumption of the model is that all employers are discriminators. Cain
(1986) provides a complete discussion of these models and their implications for racial and ethnic

segregation.

66

 

explain why we often see a significant amount of racial and ethnic segregation at the job
level.

The resulting wage differential between minorities and whites will depend on the
shape of the distribution of employers, which reflects the extent to which they
discriminate, and the size of the minority group (i.e., the supply of minority labor to a
particular labor market). This theory thus explains the tendency for discriminatory pay
differentials to be largest where the minority group is a greater fraction of the total
workforce. The basic model (which does not allow discriminatory tastes to vary by job)
also predicts that the higher and wider the employer taste distribution, the greater the
extent of segregation in the labor market and the larger the wage gap between minorities
and whites. At the job level, however, the prediction is less clear-cut. It is likely that the
stronger an individual employer’s taste for discrimination, the greater the extent of job
segregation and the lower minority wages within the firm. However, a discriminatory
employer may prefer minorities for some jobs (perhaps those involving servile tasks), and
may actually be willing to pay minorities more than otherwise identical whites for such
jobs. This implies that whites may also incur wage penalties if they work in certain jobs
dominated by minorities.

The employer discrimination model is ofien faulted for its inability to explain
wage differentials between minorities and whites that persist over time if all firms are in
one competitive product market. The common argument is that nondiscriminatory firms
will eventually drive the discriminators out of business due to their cost advantage. Yet
this obviously does not happen since persistent wage differentials and job segregation

along racial and ethnic lines are well documented (e.g., Cain, 1986). This is often

67

 

reconciled by the contention that discriminating firms may choose not to maximize
money profits, but to maximize utility instead, which is affected by both profits and their
taste for discrimination. It is also possible that either noncompetitive forces are at work,
the forces of competition work very slowly, or there are an insufficient number of
nondiscriminatory firms diamermesh and Rees, 1993).4

The theory of consumer discrimination may also help explain the pervasive
segregation seen across the US. labor market and is consistent with the persistent wage
differentials between minorities and whites. In this model, if the price of the labor
service of a white worker is p, then the tastes for discrimination of a customer are
indicated by an offer price of p - d for the same service of a minority, where d measures
the consumer’s taste for discrimination. Employers with discriminatory customers (with
d > 0) will only be willing to hire minorities at a lower wage, since minorities will be
effectively less productive. Again, customers’ tastes for discrimination are likely to
differ by the particular job in question —- a discriminating consumer, for example, may
have little objection to a minority assembling a manufactured product behind the scenes,
but be less likely to buy a car from a minority salesperson than from a white. The

consumer discrimination theory predicts that minorities may segregate themselves into

 

‘ Whether employer discrimination is currently widespread may not be the relevant issue. Some argue that
the adverse effects of the long history of discrimination in our country continue to impact earnings
disparities and job segregation for minorities (e.g., Darity and Mason, 1998). Prior to passage of the Civil
Rights Act of 1964, racial employment exclusion was blatant. Cordero-Guzman (1990) notes that “up until
the early 19603, and particularly in the south, most blacks were systematically denied equal access to
opportunities [and] in many instances, individuals with adequate credentials or skills were not, legally,
allowed to apply to certain positions in firms.” Although such exclusion is now illegal, some residual
amount of employer discrimination may still exist. Furthermore, minorities may have learned over time
that certain jobs are effectively out of their reach, and may base their human capital accumulation and job
search decisions upon this expectation. Thus employer discrimination, present or past, may still play a role
in perpetuating a significant amount of racial and ethnic job segregation.

68

jobs without direct customer contact or into firms that sell to nondiscriminatory
customers only (Cain 1986).

Although the theory of customer discrimination suggests that in competitive labor
markets such segregation could prevent the long-run maintenance of wage discrimination
against minority groups, there are several reasons why wage gaps may persist. Chiswick
(1973) argues that inequality of wages is likely to persist if some white workers have
skills complementary to the skills of black workers, citing the example of white
“foremen” working with black “laborers.” Holzer and Ihlanfeldt (1998) also point out
that the extent to which minorities can avoid wage penalties by segregating themselves in
the workplace depends on several factors, including the relative sizes of the minority and
white workforces, the relative sizes of sectors in which they do and do not face
discrimination, and the production technologies of each sector.

In its purest form, however, the consumer discrimination model implies that the
wage gap should be negatively correlated with the percent minority in a job. In other
words, all else equal, the more effective minorities are at segregating themselves into
firms with nondiscriminatory customers or into jobs without direct customer contact, the
closer minority wages should be to that of whites. Yet there are many reasons why this
prediction may not hold. Holzer and Ihlanfeldt (1998) discuss a number of scenarios in
which the job segregation resulting from customer discrimination may lead to poorer
outcomes for blacks. These explanations may apply to Hispanics as well.

For example, if there is a shortage of jobs for which minorities do not face
customer discrimination and the proportion minority in a job is acting as a proxy for the

supply of minority labor to a particular labor market, then wages might be negatively

69

correlated with percent minority. Moreover, a relative “crowding” of minorities into jobs
in the nondiscriminatory sector may result in lower wages for whites as well as minorities
(unless of course whites require a compensating differential to work in minority
dominated jobs). Holzer and Ihlanfeldt (1998) also discuss other characteristics of
product and labor markets that might cause wages of minorities and whites to be lower in
minority dominated jobs. For instance, they point to Bates (1993), who finds that
establishments in predominately black neighborhoods are likely to have less advanced
technologies and lower capital-labor ratios than firms in white neighborhoods. They note
that it is also possible that firms located in minority dominated neighborhoods will pay
lower wages because their customers are likely to have lower incomes, which may lead to
lower prices and product market rents.

Taking all of these factors into account, customer discrimination may result in
lower wages for minorities and whites working in minority dominated jobs. This seems
particularly reasonable if the nondiscriminatory sector is dominated by minority-owned
businesses selling to minority customers.

Another dominant explanation for estimated wage differentials is that minorities
come to the labor market with productivity shortfalls (possibly due to pre-market
discrimination), and that empirical studies to date do not fully control for such skill
differences. Proponents of this theory argue that omitted or unobservable differences in
productivity may explain residual wage differences between races, differences that most

researchers characterize as labor market discrimination.5

 

5 Although this skills-based explanation does not directly imply segregation, the racial and ethnic
composition of jobs may be correlated with unmeasured skill differences among workers, as discussed

below.

70

In support of this skills-deficiency hypothesis, some studies find that including the
Armed Forces Qualifying Test (AF QT) score in the National Longitutindal Survey of
Youth will significantly reduce racial differences in wages. For example, Neal and
Johnson (1996) find that AF QT scores explain nearly three-quarters of the black-white
wage gap for men (reducing the differential from -24.4 percent to -7.2 percent), and the
entire black-white wage gap for women in their sample. They conclude that the earnings

disadvantages young black workers face in the labor market arise mostly from obstacles

 

they faced as children in acquiring productive human capital. In contrast, AF QT scores
have been found to be much less effective in explaining differences in employment
between whites and minorities.6

The skills-deficiency hypothesis may be particularly relevant to the job
segregation issues investigated here, since the racial and ethnic composition of jobs may
be interpreted as a proxy for unmeasured skill differences among workers. Hirsch and
Schumacher (1992) and Hirsch and Macpherson (1994), for example, present what they
call a “quality sorting” model in which the racial composition of occupations serves as a
skill index for labor quality. They argue that discrimination is likely to lead to a sorting
equilibrium in which higher-skilled black and white workers are sorted into higher—
productivity occupations with a low proportion of blacks, and lower-skilled blacks and
whites are sorted into occupations with relatively lower productivity and higher
concentrations of blacks. This theory does not, however, explain the mechanisms

through which such segregation occurs.

 

6 Evidence based on the AF QT has been criticized by Darity and Mason (1998), who argue that AF QT
scores are not easily interpreted. They note that questions remain concerning what AF QT scores are
actually measuring, and that wage differentials often reappear when additional controls (such as “self-
esteem”) are added.

71

Empirically, Hirsch and Macpherson (1994) find the negative effect of racial
occupational segregation on wages is sharply reduced when controls for occupational
skills (e. g., required years of training, job tenure, computer usage) are included in a cross-
sectional analysis. Moreover, when panel data are used to difference out the impact of
unmeasured individual skill differences, little if any relationship is found between the
racial composition of occupations and wages. The authors conclude that racial
occupational segregation provides an important control for what are typically unmeasured
worker quality and occupational skill differences, but should not be interpreted as a
causal determinant of wages.

Since prior research finds firm- and job-level segregation to be much more
significant than occupational-level segregation for minorities, it remains to be seen how
important unmeasured skill differences are in explaining the impact of racial and ethnic
job segregation on wage outcomes. If the “quality sorting” model accurately describes
the job sorting by race and ethnicity that takes place at the firm level, then any negative
effect of job segregation on wages should decline with the addition of skill controls in
wage regressions. Moreover, if “quality sorting” is the only reason for segregation in the
labor market, it implies that there should be no correlation between job segregation and
wages if differences in skill level are fully controlled for.

The “spatial mismatch hypothesis” is yet another leading theory advanced to
explain persistent wage gaps and employment differentials between blacks and whites,
and may also help explain the significant job-level segregation found in metropolitan
areas. This theory suggests that the movement of employers out of the inner-city areas

toward the suburbs during the 197 Os and 19803 represents a spatial shift in labor demand.

72

Consequently, a “mismatch” is thought to result between the locations of employers and
those who continue to live in the central city, particularly minorities.

There are several likely reasons for such a mismatch between jobs and workers.
The decline in highway transportation costs over the past several decades has caused
many employers to choose suburban locations over the central city. Manufacturing
employers, who have traditionally provided relatively high-wage jobs for low-skilled

workers, have relocated to the suburbs at particularly high rates since their production

I ' L-_‘ L?’
A ‘_
l.‘

4 A

technology uses a relatively high ratio of land to capital. Although residential
suburbanization has also occurred, proponents of the spatial mismatch hypothesis purport
that some people face more barriers in choosing their residential location. For example,
discrimination in the housing market perpetuates residential segregation, and may prevent
minorities from following employers out to these suburban areas. Minorities residing in
the highly segregated inner-city areas may also lack the necessary transportation or
information networks to obtain suburban jobs.

Yinger (1998) argues that the effects housing market discrimination are far
reaching, claiming “Housing discrimination restricts the options of many black and
Hispanic households and contributes to the continuing intergroup disparities in income,
home ownership, wealth, education and employment (pg. 23).” Audit studies of
discrimination in the housing market (both national-level studies and most smaller
studies) find statistically significant levels of discrimination that are large in magnitude
(Turner, et al., 1991; Turner and Mickelsons, 1992; Yinger, 1995). For example, Yinger
(1995) investigates national data from the 1989 Housing Discrimination Study, and finds

that black home buyers learn about 24 percent fewer houses than whites, black renters

73

learn about 25 percent fewer apartments, Hispanics learn about 26 percent fewer houses,
and Hispanic renters learn about 11 percent fewer apartments. Overall this research finds
that the incidence of housing discrimination does not appear to be abating, but that blacks
and Hispanics continue to encounter discrimination in many aspects of a housing
transaction: “they are told about fewer available units and must put forth considerably
more effort to obtain information and to complete a transaction” (Y inger, 1998, pg. 32).7
Nevertheless, there are reasons why minorities may choose to reside in segregated
neighborhoods, particularly foreign-bom Hispanics. Chiswick and Miller (2001) argue
that the propensity of non-English speaking immigrants to cluster in communities formed
on the basis of language and ethnicity is due to the value of “ethnic goods,” which they
broadly define to include conventional foods and services and social networks. They
suggest that there is likely to be a compensating wage differential associated with such
“ethnic goods,” and find that immigrants living in “linguistically concentrated areas”
have lower earnings, even after controlling for their own language skills.8 Borjas (1998)
explicitly studies the potential determinants of racial and ethnic residential segregation.
He finds a strong negative correlation between residential segregation and both
educational attainment and wages, especially among the least-skilled groups. There also

seems to be a significant amount of intergenerational persistence in racial and ethnic

 

7 A hypothesis for such discrimination, which is supported in the literature, is that housing agents
discriminate to protect their actual and potential business with prejudiced white households (Yinger, 1998).
8 Specifically, Chiswick and Miller (2001) find that earnings of foreign-bom men from non-English
speaking countries are lower in states with higher minority language concentrations, especially for those
with greater English language fluency. They conclude that this result appears to be due to an “ethnic
goods” effect, rather than to labor market crowding. That is, they argue that immigrants sort themselves
across the country to equalize real incomes, and that “ethnic goods” will have a lower cost the greater the
concentration of those speaking the same language. They claim that regional wage differentials may
sirrrply reflect ethnic-group specific cost-of-living differentials, rather than a “crowding effect.” However,
the fact that linguistic concentrations are measured at the state level (not at the neighborhood level), and

74

 

residential segregation. Borj as (1998) concludes that “persons in the least skilled groups
wish to move to neighborhoods where they can benefit from contact with highly skilled
groups, while persons in the most skilled groups want to segregate themselves into
wealthier enclaves” (p. 229).

Empirically, the spatial mismatch literature focuses on the effects of ‘mismatch’
on employment, wage, and earnings outcomes for blacks. Holzer (1991) provides a good
overview of this literature, and concludes that blacks in inner-city areas have less access
to employment than blacks or whites in the suburbs, and that unlike most other groups of
workers, less-educated blacks face higher wages in the suburbs than in the central city.
Providing more recent evidence of spatial mismatch, Stoll et a1. (1999) find that while
less—educated people tend to reside in areas with high minority populations, low-skill jobs
are scarce in these areas. On the other hand, the availability of such jobs relative to less-
educated people in predominantly white suburban areas is high. Cutler and Glaeser
(1997) examine the impact of residential segregation (a common measure of ‘spatial
mismatch’) on several outcomes for blacks, and find that blacks in more segregated cities
have lower high school graduation rates, lower income, are more likely to be “idle”
(neither in school or working) and more likely to become single mothers.

In terms of wage effects, the spatial mismatch hypothesis can be thought of as a
variant of the “crowding” hypothesis. It suggests that since minorities tend to be
residentially segregated in the central cities, and face greater challenges securing jobs in

the suburbs, they are more likely to be crowded into jobs within the central city as well.9

 

several important controls (including industry and occupation) are omitted from the earnings regressions,
makes these results less relevant to the job segregation and neighborhood issues addressed in this study.
9The job segregation effects of mismatch problems may be even more apparent if we think of spatial
mismatch coupled with certain types of discrimination. For example, a discriminatory employer may

75

 

Spatial mismatch implies that minorities will receive lower wages than whites not due to
employment discrimination per se (unless, of course, employers are relocating to the
suburbs to gain access to a “whiter” workforce), but due to its impact on local labor
supply and demand. Minorities living and working in the central city face lower labor
demand as a result of firm relocation, and a relatively high supply of labor due to barriers
they face in following jobs out to the suburbs. Spatial mismatch also suggests that whites
competing directly against minorities for the fewer inner-city jobs should receive lower
wages as well. The theory thus predicts a negative correlation between racial and ethnic
job segregation and wages for both minorities and whites.

The theories I have presented thus far to explain wage differentials and job
segregation are all related to some form of discrimination (labor market, housing market
or pre-market discrimination). It is also possible, however, that personal choice plays a
role. Blacks and Hispanics may choose to apply for jobs in which they are likely to work
with others of their own race/ethnicity, perhaps under the expectation that they will have
more in common with coethnics. Minorities may also choose to work in segregated jobs
due to language barriers, a factor that is likely to be particularly important for Hispanics
with little or no English. Moreover, minorities that self-select into segregated jobs may

be willing to pay a compensating differential to do 50.10 Such self-selection is likely to

 

prefer to hire whites, but if the firm is located in the central city, the majority of job applicants are likely to
be minorities. Such an employer may be thus forced to hire more minorities than desired, but may
compensate by hiring them into the most menial positions within the firm.

'0 Since whites “dominate” the US. labor market (and society, for that matter), this idea of self selection is
not thought to apply to them (i.e., they will not accept lower wages to work with mostly whites, since that
is the likely outcome). I am also assuming that whites do not discriminate against minorities by demanding
higher wages to work in minority dominated jobs (as the “employee discrimination” model assumes).
Thus, the self-selection theory suggested here implies that whites that work with mostly minorities should
earn more than the minority group in question, but not necessarily more than whites that work with mostly
whites (Table 2 below supports these predictions). This is a slightly different result than that implied by the

76

result in the same outcomes as the discrimination models predict — racial and ethnic wage
differentials and job segregation — but the policy implications of self-selection are vastly
different from those implied by discrimination.

Overall, it is likely that no one dominant force is solely responsible for
perpetuating wage differentials and job segregation, but that several of the forces
described by the above theories play a limited, and possibly related role. For example,
employer discrimination may exacerbate problems with spatial mismatch, particularly if
firms are relocating to the suburbs in part to gain access to a ‘yvhiter” workforce. Holzer
(1996) finds evidence that employer discrimination is greater in the suburbs than in the
central city, since he finds that the ratio of new hires to job applications from blacks is
lower in suburban firms than central-city firms. This is particularly troubling since he
also argues that the skill needs in suburban firms are generally lower, and the relative
skills of black applicants in the suburbs are likely higher. Consumer discrimination may
also be related to spatial mismatch, since, as Kain (1968) argues, customer discrimination
may contribute to the failure of inner-city blacks to follow jobs out to the suburbs.
Empirical Implications

Although I do not attempt to formally test any of these leading theories, they do
provide a loose theoretical framework for the following empirical work that investigates
the determinants of job segregation. 1' For example, if employers have a taste for

discrimination against minorities as Becker’s model suggests, we may expect to find

 

employee discrimination model, which predicts that whites require a compensating differential to work
with minorities.

'1 Previous studies that have attempted to test the importance of the above explanations for rninority-white
wage gaps have resulted in few, if any, firm conclusions. Cain (1986), who provides a review of several
studies that test the discrimination hypotheses, concludes that these mixed results are due to the fact that the
theories often yield ambiguous predictions. He goes on to describe several of the difficulties in testing the
hypotheses of discrimination theories.

77

certain relationships between the likelihood of job segregation and other variables, such
as region and firm size. It is well known that racial prejudice has historically been more
pervasive in the South, while ethnic tensions and discrimination against Hispanics may
be more pervasive in the Los Angeles area due to the large influx of Latino immigrants
over the past few decades (see, for example, Ortiz, 1996). Thus a strong relationship
between metropolitan area (a proxy for region) and the likelihood of j ob segregation for

minorities may be indicative of employer discrimination (such a relationship would also

 

support the consumer discrimination model if customers and employers share regional a
tastes). Similarly, we may expect to find a negative relationship between firm size and
job segregation if employer discrimination is a factor. Larger firms are more likely to
have a formal human resource department (or even a legal department) and be more
aware of the legal consequences of employment discrimination. Moreover, Affirmative
Action enforcement is likely to be stronger for larger firms.12 Controls for firm size and
metropolitan area are therefore included in the following analysis of job segregation to
shed light on whether employer discrimination is a likely determinant of job segregation
for minorities.

The consumer discrimination model suggests an empirical approach in which to
explore the likelihood that buyers of labor services may help perpetuate labor market
disparities between whites and minorities. I explore the impact of potential customer
discrimination on wages and the likelihood that minorities work in segregated jobs using
an indicator of face-to-face contact with consumers as a proxy for the potential of

customer discrimination. The theory predicts that minorities who face customer

 

'2 Firms with fewer than fifty employees are not subject to Affirmative Action laws. However, the effect of
Affirmative Action on segregation should be present whether discrimination originates from employer or

78

discrimination will receive lower wages than otherwise identical whites.13 We may also
expect to see more job segregation among minority groups for whom customer
discrimination is more pervasive, as they attempt to avoid the associated wage penalties.
If a sufficient amount of job mobility is possible, minorities working in jobs with direct
customer contact may be interpreted as those who do not face discriminatory customers
and therefore we should expect to find a positive relationship between the indicator of
customer contact and segregation into jobs with other minorities.

Finally, I investigate the possible impact of “spatial mismatch” on job
segregation. It is well known that there is a significant amount of residential segregation
between whites and minorities, with minorities often concentrated in the inner-city areas.
Over the past few decades there has also been considerable movement of firms and jobs
from central-city areas to the suburbs. Evidence suggests that inner-city job access has
declined over time for low-skilled workers (Ihlanfeldt and Sjoquist, 1990). Past studies
also find residential segregation (a proxy for spatial mismatch) to be negatively correlated
with employment and income for blacks. Spatial mismatch problems further imply that
minorities living in more segregated neighborhoods should be more likely to work in
segregated jobs. However, if minorities can commute longer distances to follow the jobs
out to the suburbs, they might be able to avoid the wage penalties associated with job

segregation.

 

consumer prejudice.

'3 Unfortunately, the data used here do not contain a measure of customers’ racial and ethnic composition,
which may provide a more direct test of the potential for customer discrimination since one would expect
less discrimination by customers of one’s same race or ethnicity. Holzer and Ihlanfeldt (1998) provide
convincing evidence of customer discrimination in metropolitan areas using the employer-side counterpart
to the household data used here. They find that the larger the fraction of minority customers, the higher the
probability that workers fi'om the same minority group will be hired. Their results are strongest for jobs
with significant contact with customers.

79

I therefore examine the impact of residential segregation and average commute
times on the probability that minorities work in a segregated job to assess the significance
of spatial mismatch. If spatial mismatch plays a role in perpetuating racial and ethnic
segregation in the labor market, residential segregation will be positively related to the
likelihood of job segregation for minorities. On the other hand, holding constant
residential segregation, commute time and job segregation will be negatively related.
Finally, the effect of residential segregation on job segregation should decline with
commute time.14

A few qualifying comments are in order here. First, it is important to note that
neither residential segregation nor commute time provides a perfect measure of the
degree of spatial mismatch. Residential segregation really only tells one side of the story
— that the residences of minorities tend to be concentrated in certain areas within a
metropolis, usually the central city. It tells us nothing about firm location (although the
fact that firms have relocated to the suburbs in large numbers is well established).
Furthermore, residential segregation may have negative effects on minority outcomes
above and beyond the wage and employment effects of spatial mismatch. For example,
Cutler and Glaeser (1997) find that residential segregation negatively affects schooling
and family structure outcomes for blacks, even after controlling for average relative
commute time. Commute time, on the other hand, should give an indication of the
distance between residence and job location. But commute time is only observed for

those who actually find jobs, and thus a measure of average commute for an area is likely

 

" The theory also predicts wages to be negatively related to residential segregation (which is verified below
for all demographic groups but Hispanic women), but positively correlated with commute time, since most
workers require a compensating differential to commute longer distances. Although not reported, commute

80

to understate the true distance between workers and jobs (i.e., the degree of spatial
mismatch), particularly for minorities. Despite these shortcomings, however, measures
of residential segregation and commute time are reasonable proxies for spatial mismatch
and are likely to be related to job segregation as discussed above if mismatch is a
contributing factor to the segregation of minorities from whites in the labor market.

The remaining two explanations for wage differentials and job segregation that do
not necessarily suggest labor market discrimination -— the self-selection explanation and
the skills-deficiency hypothesis — are difficult to test either formally or informally given
the data used here. Ideally, one would use longitudinal data to difference out any
unmeasured skill and taste differences among workers that may be correlated with job
segregation. Difficulties associated with possible bias resulting from such unobserved
heterogeneity, which may affect the following empirical results, are discussed and
addressed below in the subsection of the Results section entitled “Potential Econometric

Problems.”

Data and Empirical Framework

The Data Set

I explore these issues using a relatively new data set drawn from the Multi-City
Study of Urban Inequality (MCSUI). The MCSUI survey was administered to adult
household residents in Atlanta, Boston and Los Angeles.15 Interviewing was completed

in the summer of 1993 in Atlanta, and in the summer of 1994 in Boston and Los

 

time was found to be largely insignificant in wage equations using these data. In equations for white
women, however, commute time was verified to be positively related to wages.

'5 The survey was also administered in Detroit, however the Detroit data is not analyzed here because it
does not contain information on job segregation.

81

 

Angeles.16 Data from the 1990 Census at the block and group level were merged into the
MCSUI data set in order to control for residential segregation.

The MCSUI survey consisted of a probability sample of households, stratified by
race-ethnicity and poverty-status composition of the 1990 Census. Blacks were
oversampled to yield roughly equal numbers of whites and blacks in all locations; Latinos
and Asians were similarly oversampled in Los Angeles, as were Asians in Boston. In
addition, concentrated poverty areas were oversampled in all metropolitan areas. The
project also used a multistage sampling procedure, utilizing cluster sampling with three
levels of clustering. This process generated a total of 7,373 observations — 1,528 in
Atlanta, 1,820 in Boston, and 4,025 in Los Angeles. Restricting the sample to non-retired
respondents reduces the sample to 6,388 observations. The sample was further restricted
to include only whites, blacks and Hispanics (omitting Asians and other respondents due
to limited observations on these individuals) with usable wage information.17 These
additional sample restrictions reduced the full sample used in the analyses that follow to
3,895 observations — 1,260 on whites, 1,392 on blacks, and 1,243 on Hispanics (some
estimates are based on fewer observations due to missing values).

One goal of the Multi-City Study of Urban Inequality was to test hypotheses

concerning the status of women and minorities in urban labor markets, making this data

\
16
a T,h°_ US. economy was recovering from the recession of the early 1990s when the survey was
Bd’oslmtollStered. Monthly unemployment rates during this period averaged under six percent in Atlanta and
dummn, and uIlder 10 percent in Los Angeles. To control for differences in local labor market conditions,
17 l y varlables for metropolitan area are included in the regression analyses that follow.
$2 £386 note that three observations were dropped due to missing sex information. Hourly wages below

e exeeedmg $200 were classified as outliers. In defense of this trimming technique, Angrist and
impafter £1998) found that "extreme wage values are likely to be mistakes," when they investigated the
report“? hmmming outliers using CPS data. In addition, a visual inspection of cases for which respondents
rep 0 rte d f°u{1y wages between $100 and $200 was conducted for reasonableness (e.g., yenficatron that
additj “me income was consistent with these wage levels), and resulted in the classrficatron of one

01131 outlier.

 

82

set particularly appropriate to address the issues discussed above. These data provide a
rich source of information on labor market histories, including an extensive collection of
variables to control for differences in human capital and job characteristics. For example,
MCSUI provides measures of past work experience related to one’s current job, specific '
job tasks performed, and information on the racial/ethnic composition of the job.

Measures are taken in the following analyses to ensure that the data can be used to
draw inferences regarding the underlying metropolitan populations. First, analysis
weights for respondents are used, which were set inversely proportional to the household
sampling weight. Analysis weights also reflect nonresponse (if nonresponse is not
unifonnly distributed) and the number of persons eligible for interview in the respective
household. Second, in all analyses robust standard errors are calculated that are also
adjusted for the clustering and stratification of the survey design.18

A shortcoming of these data concerns the size of the sample and whether it is
representative of the US. labor market in general. The empirical estimates arrived at in
this study probably provide good first approximations of the impact of racial job
segregation on blacks in metropolitan areas. However, since estimates of the impact of
ethnic JOb segregation are driven primarily by observations on Hispanics in the Los
Angeles area, the same cannot be said for Hispanics. At best, empirical results on ethnic
segregation may only be representative of metropolitan areas with relatively high

concentrations of Hispanics.

Measurin Se e ation

J0b segregation is measured by a series of dummy variables created from a survey

question asking “What (is/was) the race and ethnicity of most of the employees doing the

83

kind of work you (do/did) at this location?” The potential survey responses to this
question are non-Hispanic white, non-Hispanic black, Hispanic, Asian, mixed racial
group (in Atlanta and Boston only), or “other.” A series of five dummy variables were
created from the responses to the job segregation question, combining the mixed-racial-
group response with “other.” Although the survey question does not refer to a specific
level of occupational disaggregation, the resulting segregation measure is likely to
capture whether a fairly detailed occupation within an establishment is segregated by race
or ethnicity.

Throughout this paper I refer to the segregation measure as an indication of “job
segregation.” However, since these data do not allow me to explicitly control for
establishment-level segregation, the segregation measure used here probably captures the
impact of segregation at the establishment level as well as at the job level.19 The
potential determinants of segregation at the establishment level are likely to be similar to
those affecting job-level segregation, and include employer and consumer discrimination
(i.e., discrimination in job hiring at the establishment due to either employer or customer

PTCJUdiCB), “quality sorting,” or personal choice. Spatial mismatch, on the other hand, is

\
18 .
19 Thiswas accomplished using Stata survey (svy) commands (see StataCorp, 1997, pp. 305-312).
thelshe Inclusion of establishment-level controls for firm size and industry may reduce this possibility if
datae controlspick up dimensions of racial and ethnic segregation at the firm level. Furthermore, lack of
se e011 establishment-level segregation may actually imply that the following estimates of the job
1e f; $39011 e.fl‘ect are biased towards zero. Bayard et al. (1999) find that segregation at the establishment
Count: SSSQCtated with higher wages and thus reduces the estimated wage gap for minorities. This seems
OWDednmmthe, but perhaps their establishment measures are picking up wage effects from minority
(1998) :nd oDerated businesses. In contrast to the findings of Bayard et a1. (1999), Carrington and Troske
of blackmd llttle establishment-level segregation - that within metropolitan areas the interfirm distribution
imp] thaam‘l White workers is close to what would be suggested by random assignment. Their fmdrngs

y t llttle bias is likely to result from the omission of establishment-level segregation data.

 

84

more likely to impact segregation at the establishment level than at the job level, unless it
is coupled with customer or employer discrimination.20
Empirical Methods

I first investigate the magnitude of racial and ethnic wage gaps for men and

women using OLS estimates of log wage regressions of the following general form:

(1) ln(W) = p, + BlBLACK + BZHISP + HCB3 + PERSONl34 + INSTHB5 + Joan, + e,

 

where W is the hourly wage, BLACK is a dummy variable equal to one if the individual is
black, and HISP is a dummy variable equal to one if the individual is Hispanic. HC is a
vector of human capital controls including educational attainment, age, previous
experience doing similar work and job tenure. In some specifications, HC also includes
measures of English fluency and US. citizenship, which may be related to productivity.21
PERSON includes personal controls for whether the individual is married and number of
children under 18. INST IT contains institutional controls for firm size and whether the
individual is covered under a collective bargaining agreement. JOB controls for job
characteristics such as whether the individual works part time (less than 35 hours a
week), Whether they have the authority to supervise others and, if so, whether they have
the ability to set the pay of those they supervise. JOB also contains dummy variables

indicating Whether specific job tasks are performed on a daily basis, including talking

\

20
Segem:fiempirical results that follow, however, spatial mismatch should have a similar impact on the job
Were Ev fin measure used here as it would on a measure of establishment segregation, if such a measure
21 Tr . a1 able, since the segregation question refers to one’s job at a specific location.

F10 (1997) finds that English language deficiencies are an important source of lower earnings for

HiZXlCan-Ameficans. Moreover, US. citizenship may independently affect labor market outcomes for
panics, Whether or not they are legally permitted to work.

85

with customers face to face, talking with customers on the telephone, reading paragraphs,
writing paragraphs, using a computer and doing arithmetic. Finally, controls for industry
and occupation are included in some specifications.

The impact of racial and ethnic segregation is then examined by adding the
controls for job segregation (the categorical variable indicating whether one works with

mostly whites, blacks, Hispanics, Asians or “other”). I estimate these wage equations by

.- 19:?!

race/ethnicity and sex to allow the effects of job segregation, and other characteristics, to

 

vary by demographic group.22 4_
In order to assess the impact of race and ethnicity on the number of job benefits

received, I specify ordered logit equations for men and women of the following form:

(2) BEN = a0 + alBLA CK + azHISP + HCa3 + PERSONOL4 + INSTITas + JOBa6 + u,

where BEN is a variable ranging from O to 4 based on the total number of four
employment benefits an individual receives through their job: health insurance for
themselves, health insurance for their family, a retirement plan, and/or paid sick leave.

The independent variables correspond to those in equation (1). The job segregation

_\

22

Some Of the variables discussed above meant to serve as proxies for productivity, such as supervisory
39mm}? and the responsibility of specific job tasks, may be affected by discrimination. For example, if
ghosties do not have the same opporttmity for attaining valuable control over supervisory authority and
VariZbihty to set pay as do whites with the same qualifications, then specifications that include these
to job es W111 “over control,” and underestimate discrimination. Minorities may also have unequal access
mino S éSSOCIated with certain job tasks, such as jobs working with computers. In addition, like women,
sung?“ may be segregated into lower-paying industries and occupations. Consequently, the job tasks,
sets ofisory authority, industry and occupation variables are not included in all specifications. Two other
Citize ”Enables that are only included in later specifications are the controls for English fluency and US.

nshlp. While these variables are likely to be related to productivity, there is also the strong possibility

of ' ' - ' ' '
discnmmation based on differences in language or crnzenshlp-

 

86

controls are then added to equation (2) to evaluate the impact of racial and ethnic job
segregation on the number of benefits received.

Equation (2) is then run as a probit equation with BEN respecified as a dummy
variable for each of the four employment benefits, indicating whether the individual
receives the benefit in question. That is, separate probit equations are rim to determine
whether job segregation affects the probability of receiving each of the four benefits (as
opposed to the number of job benefits received), since job segregation may affect the
likelihood of receiving the various benefits differently.

Lastly, I explore the potential causes of job segregation. While equations (1) and
(2) follow from the basic human capital model, there are many potential reasons why
minorities are likely to work in highly-segregated jobs. Although I do not explicitly
model the process of job segregation, the theories presented above provide some
direction for choosing likely determinants. For example, customer discrimination and the
“spatial mismatch hypothesis” are two theories that may help explain racial and ethnic
segregation in the labor market. The consumer discrimination theory suggests that
minorities who work in jobs with direct face-to-face customer contact may segregate
themselves into firms that sell to nondiscriminatory customers only in order to avoid
potential wage penalties associated with such discrimination. Similarly, spatial mismatch
suggests that minorities who live in hi ghly-segregated, inner-city areas, without the
necessary means to follow the movement of jobs to the suburbs, may be segregated into
lower ‘Paying jobs. If these models accurately characterize the segregation process,
measures of residential segregation, commute time and customer contact are likely to be

correlated with job segregation.

87

Hirsch and Macpherson’s (1994) “quality sorting” theory implies that higher-
skilled workers will be sorted into occupations with a low proportion of minorities, while
lower-skilled workers are sorted into occupations with higher concentrations of
minorities. Thus if some sort of “quality sorting” plays a role, human capital levels are
likely to be important determinants of job segregation. Finally, as discussed above,
metropolitan area and firm size may capture the effects of employer discrimination on job
segregation.

To more formally explore the potential determinants of racial and ethnic
segregation in the labor market, I estimate probit equations for blacks and Hispanics of

the following form:

(3) JOBSEG = 90 + 91RESSEG + 92C0MM+ 93CUS + HC94 + INSTIms + CITY96

+ 67MALE+ n,

where JOBSE G is a dummy variable equal to one if the individual is black and works
with mostly blacks or Hispanic and works with mostly Hispanics (whites are omitted
from this analysis). RESSE G measures residential segregation in one’s census tract as the
Proportion black if the respondent is black, and as the proportion Hispanic if the
rev‘spondent is Hispanic. COMM is a measure of average commute time in minutes and
CUS is a dllrnmy equal to one if the individual deals with customers or clients face to
face on a daily basis. Due to sample size concerns, men and women are pooled together,

and MALE is included to control for sex. The other variables are as defined above.

88

Results

Table 1 reports descriptive statistics for white, black and Hispanic male and
female workers in MCSUI. The data on average log hourly wages show that racial and
ethnic differences are larger for men (-O.28 for blacks and -O.58 for Hispanics) than for
women (-0.20 for blacks and -O.54 for Hispanics), although the differences in wage gaps
between the sexes in MCSUI are much more modest than those found in other data (see,
for example, Cain, 1986 and Bayard, et a1, 1999). The data on the total number of
employment benefits received (of 4 possible) indicate that racial differences are slightly
larger for men than for women, but the ethnic differences are virtually identical by sex.
In general, however, blacks are close to parity with whites in terms of the level of
employment benefits, while Hispanics receive significantly fewer employment benefits
through their jobs.

Whites in the sample are slightly older, more likely to have higher educational
degrees, have more tenure and previous experience than minorities, all of which are
associated with higher wages and benefits (although the difference in tenure between
white women and black women is not significant).23 Minority men are more likely than
white men to work part time, whereas white women are more likely than minority women
to work part time. This may partly explain the larger unadjusted wage differences among
men compared to women.

AS indicated by Table 1, a significant proportion of Hispanics do not speak
English “very well” (65.8 percent of Hispanic men and 62.1 percent of Hispanic women).
Notably , a rather large proportion of Hispanics are not US. citizens, which may coincide

\

23
leverlrless Othervnse noted, all differences discussed in this section are (at least) Significant at the lO-percent

89

with the large representation of Hispanics from the Los Angeles area (a region known for
a relatively large amount of immigration fi'om Mexico). The descriptive statistics also
reveal that Hispanics work in significantly smaller establishments, while blacks tend to
work in larger establishments (although not significant in these data, this latter result is
explored by Holzer, 1997).

In terms of job tasks performed on a daily basis, Table 1 shows that whites are
significantly more likely than both minority groups to use a computer on the job, a
characteristic known to be associated with higher wages. White men are also more likely
than black men to write paragraphs, while white women are more likely than black
women to do arithmetic on a daily basis. Whites of both sexes are significantly more
likely than Hispanics to perform virtually all of the remaining tasks on a regular basis,
and to work in jobs with supervisory authority. Finally, white women are more likely
than minority women to have jobs with the ability to set the pay of others.

Table 1 also reports descriptive statistics on segregation by race and ethnicity. A
remarkable degree of job segregation for minorities is revealed when investigating the
survey question that asks about the race/ethnicity of most employees doing similar work.
Blacks are most likely to report that the majority of their coworkers are also black, with
36 percent of black men and 41 percent of black women working in predominantly black
lObS. This represents a significant amount of job segregation considering the relatively
small PTOPOI‘tion of blacks in the labor markets of these metropolitan areas, with blacks
accounting for only 13.2 percent of the overall sample of workers. Perhaps the most
striking flIlding is the degree of segregation among Hispanics, with 72 percent of men

and 69 Percent of women working in jobs with predominantly Hispanic coworkers.

90

Comparatively, Hispanics comprise only 21.3 percent of the sample of workers. Finally,
more than 80 percent of whites work in jobs dominated by whites, while whites make up
61.7 percent of the overall workforce in the three-city sample.

The figures reveal comparatively less racial and ethnic segregation along industry
and occupation lines. There are no striking relationships in terms of the distribution of
workers by the eight broad categories of industries used here, with the exception of
Hispanics being less likely to work in the service industry than whites and blacks. A
notable relationship does appear, however, when looking at the distribution of workers by
occupation: whites are significantly more likely to be in managerial, professional or
technical positions than minorities. Among men, 51 percent of whites are in these higher
paying positions, compared to only 28 and 14 percent of blacks and Hispanics,
respectively. Similarly, 45 percent of white women are in managerial, professional or
technical positions, compared to only 18 percent of black women and 14 percent of
Hispanic women. Blacks and Hispanics are more likely to be in lower-paying service
and labor occupations than whites, and Hispanics are significantly more likely to be in
crafis/operative positions.

Table 2 presents mean hourly wages (in panel A) and mean log wages (in panel
B) for the six demographic groups, by the race and ethnicity of the majority of one’s
coworkers. The table also reports median wages in brackets to avoid having results
driven by tails in highly skewed distributions (this is especially important given the small
sample sizes for some cells). For example, comparing mean hourly wages of black males
Working with mostly whites ($15.56) to that of those working with mostly blacks

($ 14.17) suggests an unadjusted wage differential of just under 10 percent. However,

91

performing a similar comparison using median hourly wages or log wages (mean or
median) indicates that black males segregated into jobs with mostly blacks earn more
than 30 percent less than those working with mostly whites.24 A cross-wise comparison
using log wages in panel B reveals that the unadjusted wage gap between black and white
men working with mostly whites is 17 percent, while it is 26 percent for those working in
jobs with mostly blacks.

Overall, the comparison of wages by race/ethnicity of coworkers in Table 2
reveals an interesting relationship: Blacks and Hispanics seem to incur the largest wage
penalties when they are segregated into jobs dominated with other blacks and Hispanics,
respectively. In other words, blacks earn the least when they work in jobs dominated by
blacks, and Hispanics earn the least when they work in jobs dominated by Hispanics.
This result holds for both men and women. White men, on the other hand, earn the most
when they work in jobs with predominantly white coworkers. Although it is rare for
whites of either sex to work in jobs with mostly Hispanics (as seen in Table 1), those that
do earn significantly less than those that work in “white jobs.” Comparatively, the small
fraction of blacks that work with mostly Hispanics seem to fare quite well, as do
Hispanics who work with mostly blacks.

How Job Segregation Affects Wages

Table 3 reports baseline OLS log wage regressions for men and women that

describe the relationships between the variables listed in Table 1. Columns (1) and (4)

include controls for demographic group; personal controls for marital status and number

 

 

2‘ A visual inspection of certain characteristics of high wage earners (such as education, occupation and
reported family income) was performed to assure the reasonableness of these observations. The inspection
resulted in a decision not to omit any additional high-wage observations from the sample other than those
110th in footnote #17.

92

of children under eighteen; basic human capital controls for age, educational attainment
(categorized as the highest degree earned), prior experience doing similar work and
tenure with firm; and institutional and job controls for part-time work, firm size, and
collective bargaining coverage. The estimated coefficients on these variables are
consistent with human capital theory, with positive returns to age (a proxy for general
human capital), educational attainment, job-specific experience and tenure, while those in
part-time jobs receive lower wages.25 It is interesting to note that relative to high-school
dropouts, women in the sample earn higher returns for all levels of educational attainment
than men earn. Working in a larger firm benefits women in the sample, yet firm size is
insignificant for men. For both sexes, the unadjusted black-white differential indicated in
Table 1 falls by almost half after controlling for these basic characteristics. For men, the
Hispanic-white differential falls by more than one-third, while the estimated Hispanic—
white differential for women falls by more than one-half. Overall, the wage gaps remain
larger for men.

In columns (2) and (5) I add controls for English fluency and citizenship.
Although these variables may account for some productivity differences, it is important
to keep in mind that there may also be discrimination based on one’s ability to speak
English and citizenship. The addition of these controls reduces the Hispanic-white wage
differential for men from -O.353 to -O.299, and considerably reduces the adjusted

Hispanic-white wage gap for women, which falls from a significant -0.223 to an

¥

25 Tests allowing the returns to a high-school diploma to differ fi'om returns to a GED revealed that for
men, returns to a GED were generally lower but that the GED dummy was consistently insignificant in the
male wage models. All else equal, women with a GED appeared to earn higher wages than those with a
high-school diploma, however only 2 percent of the sample reported having a GED. More importantly,
including a separate dummy variable for a GED did not affect the estimated coefficients on the
race/ethnicity dummies or on the job segregation dummies in any of the wage models.

93

insignificant -0.096. Relative to being able to speak English “very well,” a lack of
English fluency seems to have a significant negative impact on wages. While the
associated wage penalties are fairly consistent across other levels of fluency for men,
women’s wages are not significantly penalized for speaking English less than “very
well,” but are penalized to a greater extent than men’s for not being able to speak English
at all. Likewise, women experience a much greater wage disadvantage than men for not
being a US. citizen by birth (in fact, men’s wages do not appear to be penalized at all for
not being a US. citizen by birth). The dramatic differences in these estimated returns by
sex calls into question the contention that English fluency and citizenship are only
picking up productivity differences and are not subject to discrimination. Perhaps
Hispanic women face some sort of “double jeopardy” in the labor market, and are
discriminated against for lack of citizenship and the ability to speak English to a greater
extent than men. It is also possible that Hispanic women without US. citizenship and the
ability to speak English do not have the social networks necessary to find well-paying
jobs, whereas similar Hispanic men have better network systems.

Columns (3) and (6) include controls for daily job tasks, supervisory authority,
and industry and occupation. Since these characteristics may themselves by affected by
discrimination, these results should be interpreted with some care. For men, the inclusion
of these controls reduces the racial and ethnic wage differentials considerably (from -
0.144 to -0.078 for blacks and from -O.299 to -0.176 for Hispanics).26 Working with a

computer on a daily basis and having supervisory authority to set pay are associated with

 

2° Controls for industry and occupation alone explain a significant amount of the wage gaps. When log
wages are regressed on racial and ethnic dummies with the addition of industry and occupation controls
only, the black-white wage gap is -.l37 for men and -.158 for women, while the Hispanic-white wage gap
is -.361 for men and -.313 for women.

94

higher wages for men. For women, the inclusion of these controls reduces the black-
white wage gap from -O.119 to -0.086, while the Hispanic-white wage differential is
reduced to an insignificant -0.025. Having supervisory authority to set pay and working
with a computer on a daily basis are also associated with higher wages for women, as are
the daily tasks of writing paragraphs and talking with customers on the phone. Having
face—to-face customer contact on daily basis, however, is associated with lower wages for
women.

The impact of job segregation on wages is now investigated to better understand
the sources of racial and ethnic differences in wages. Table 4 reports results of wage
regressions corresponding to the columns used in Table 3, but with the inclusion of job-
level segregation controls. The results indicate that men working in jobs with
predominantly black or Hispanic coworkers earn less than those working with mostly
whites. Based on columns (1) and (2), men who work with mostly blacks earn about 13
percent less, and men who work with mostly Hispanics earn 17-21 percent less,
compared to those who work in “white jobs.” Column (3), which includes control for
daily job tasks, supervisory authority, industry and occupation, estimates the wage
disadvantage associated with working with mostly blacks or Hispanics at roughly 11
percent.

While the included productivity proxies were shown to explain a significant
amount of the unadjusted racial and ethnic wage gaps in Table 3, it seems that job
segregation has some additional explanatory power. Columns (1) and (2) of Table 4
indicate that the black-white wage gap for men is reduced from roughly -0. 14 to

approximately -0.09 when I control for the effects of job segregation. The Hispanic-

95

white wage gap for men is reduced by more than one-quarter, but remains fairly large at
roughly -0.22. When controls for daily job tasks, supervisory authority, industry and
occupation are included in column (3), the black-white wage gap for men is reduced to
-0.05 (a difference that is statistically insignificant). The Hispanic—white wage gap for
men is also reduced considerably with the addition of these controls, to roughly -0. 13.

Like men, women also experience wage disadvantages associated with working in
jobs with predominantly black or Hispanic coworkers. Based on columns (4) and (5),
women who work with mostly blacks earn about 10 percent less, and those who work
with mostly Hispanics earn 20—23 percent less than women who work with mostly whites.
Column (6) for women estimates the wage disadvantage associated with working with
mostly blacks at 9 percent, and the disadvantage associated with working with mostly
Hispanics at roughly 12 percent.

The inclusion of controls for job segregation reduces the black-white wage
differential for women from a significant -0. 12 to about -0.07 (an estimate that is not
quite significant), based on columns (4) and (5). When controls for daily job tasks,
supervisory authority, industry and occupation are also included in column (6), the racial
wage gap for women falls further, to an insignificant -.05. The job segregation controls
reduce the Hispanic-white wage gap by more than half to roughly -O.10 based on column
(4). When controls are added for English fluency and citizenship in column (5), there is
no significant difference in wages between Hispanic and white women.

Overall, the results of Table 4 seem to provide some evidence of “quality sorting”
at the job level. The inclusion of additional personal and job characteristics (particularly

the controls for daily job tasks, supervisory authority, industry and occupation) is

96

associated with a reduction in the estimated impact of job segregation, especially for
those working with Hispanics.

Next I allow the estimated effects of job segregation on wages to differ by race
and Hispanic ethnicity. Table 5 reports regression results by demographic group, with
Table 5a containing the results for blacks, and the results for Hispanics and whites in
Tables 5b and 5c, respectively. The controls included in the first column under each
demographic-group heading correspond to those in column (1) of Table 3, while the
controls included in the second column for each demographic group are the same as those
in column (3) of Table 3. Note that the English fluency and citizenship controls are
included only in the regressions for Hispanics.

Overall, Table 5 confirms the general result we saw in Table 2: minorities who
are segregated into jobs with other minorities of the same race/ethnicity earn less than
those who work in jobs with predominantly white coworkers. Table 5a shows that black
men whose jobs are racially segregated seem to be penalized more than black women
who are segregated. Black men who work with mostly blacks earn 14—19 percent less
than otherwise identical black men who work with mostly whites, while black women
who are racially segregated earn about 9 percent less than those who work with mostly
whites.

Turning to Table 5b, the results for Hispanics indicate that ethnic job segregation
is associated with wages that are almost 15 percent lower for Hispanic women, compared
to women with the same characteristics who work in jobs with mostly whites. For
Hispanic men, column (1) suggests that ethnic job segregation is associated with wages

that are almost 20 percent lower. However, when controls for English fluency,

97

citizenship, job tasks, supervisory authority, industry and occupation are added in column
(2), the effect of ethnic job segregation on wages is cut in half, indicating that Hispanic
men who work with mostly Hispanics earn 10 percent less than those that work with
mostly whites (a difference that is not statistically significant). Table 5b also seems to
suggest that Hispanic men who work in jobs dominated by blacks are penalized
considerably, earning over 30 percent less than those working in “white jobs.” However,
these estimates are based on differences across small cells.27

Table So, which contains the results for whites, indicates that once I control for
differences in basic human capital and personal characteristics, the wage disadvantages
that white men seemed to incur when working with minorities (based on Table 2 results)
disappear. On the other hand, white women that work in jobs dominated by Hispanics
appear to earn significantly less, even after including the full set of controls.

Keeping in mind the wage implications of the various models discussed in the
theory section, I am now in a position to evaluate the overall findings of Table 5 (a more
complete discussion of the merits of these models follows below in the subsection
assessing the likely determinants of job segregation). Although no firm conclusions can
be drawn since several of the models do not provide clear predictions, the finding that the
considerable wage disadvantage white men experience when working in minority
dominated jobs disappears once differences in basic productivity measures are eliminated

lends support to the hypothesis that some sort of “quality sorting” may take place at the

 

27 Table 5 indicates that the only demographic group for whom face-to-face customer contact seems to
affect wages is black women. As mentioned in an earlier footnote, a better measure of the possibility of
consumer discrimination would be an indication of the proportion of minorities living in the area where the
job is located. Short of that, in regressions not shown here, residential segregation measures were
interacted with the face-to-face dummy to determine if the wage effect of customer contact depends on the
proportion minority in one’s neighborhood. In all specifications, this interaction variable was insignificant
and had little impact on the face-to-face dummy.

98

 

' Pu.

job level for white men. However, even after including the full set of controls, racial and
ethnic job segregation is associated with lower wages for white women and minorities.
These findings are consistent with predictions made by the spatial mismatch hypothesis
and the two discrimination models, but suggest that the quality sorting model alone
cannot explain racial and ethnic job segregation for these groups of workers (unless the
empirical specifications do not adequately control for skill differences). The results of
Table 5 also provide some support for the “personal choice” model — the idea that
minorities may choose to work with coethnics and accept lower wages to do so.
However, this model does not explain why the wages of white women should be
penalized for working with Hispanics, or why Hispanic males should experience a wage
disadvantage for working with mostly blacks. Overall, Table 5 suggests that not only is it
likely that several forces influence the amount of job segregation in the labor market, but
that these forces may affect the different demographic groups to varying degrees.

Does Job Segregation Affect Benefits?

The results presented above provide compelling evidence that racial and ethnic
segregation at the job level has a detrimental impact on the wages of minorities and may
explain a portion of the persistent wage differentials found between whites and
minorities. A related question is whether job-level segregation affects the level of
employment benefits associated with one’s job, in addition to wages.

Table 6 reports baseline ordered logit estimates that describe the relationships
between the number of benefits received (measured by an index ranging from O to 4) and
the variables listed in Table 1, first without controlling for job segregation. Controls

included in columns (1)-(6) of Table 6 correspond to those of columns (1)-(6) of Table 3.

99

Based upon column (1) of Table 6, which includes the basic personal and human
capital controls, there is no statistical difference in the distribution of employment
benefits between black men and white men. Black women, Hispanic men and Hispanic
women, however, all receive fewer job benefits than their white counterparts with the
same basic personal and human capital levels. When controls are added for English
fluency and citizenship, the black-white benefits differential remains for women, while
differences in benefit levels between whites and Hispanics are eliminated for both men
and women. Again, it is important to note that while the English and citizenship controls .,
may measure important productivity differences, it is also possible that some
discrimination is based on differences in these characteristics. It is interesting to note
that, as in the log wage regressions, the negative effects associated with the ability to
speak English and US. citizenship differ a great deal by sex. Women are penalized in
terms of the level of benefits received to a much greater extent than men for not being
able to speak English and not being a US. citizen (although these results fall apart in
column (6), which does not seem to fit the data for women well in general).

The impact of most human capital variables on the level of employment benefits
received is as one would predict. Age, educational attainment, prior job experience and
tenure are all associated with a higher level of employment benefits, while part-time jobs
are associated with significantly lower benefits. Being unionized and working in larger
firms are associated with higher benefits. Overall, there seems to be little relationship
between job benefits and supervisory attainment and job tasks, although for men jobs that
entail working with computers or reading paragraphs on a daily basis are associated with

a higher level of benefits.

100

Table 7 investigates the impact of job segregation on the level of job benefits.
Ordered logit estimates of coefficients on the job segregation measures and the
race/ethnicity dummies are reported for specifications similar to those of Table 6. For
men, column (1) indicates that the Hispanic-white benefits differential is completely
eliminated once I control for job segregation. But there is a significant negative effect of
being segregated into jobs with predominantly Hispanics on the level of employment
benefits for men. This estimated effect remains significant in all three columns.

Column (4) for women indicates that the inclusion of job segregation measures
does nothing to eliminate the racial and ethnic differentials in employment benefits.
While the Hispanic-white differential for women is eliminated with the inclusion of
controls for English fluency and citizenship in columns (5) and (6), the black-white
differential for women remains significant even with the additional controls for job tasks,
supervisory authority, industry and occupation in column (6). Table 7 does not seem to
indicate that job segregation negatively affects the likelihood of receiving employment
benefits for women. In fact, column (6) indicates that women working in minority
dominated jobs are likely to receive a greater number of benefits than those working in
“white jobs.”

To investigate this issue further, Table 8 contains results from estimating separate
probit equations for individual job benefits, since arguably, benefits such as health
insurance may be of greater value than, say, sick leave. Panel A contains the results for
men, and shows that men segregated into jobs with Hispanics are significantly less likely
to receive retirement insurance and health insurance for their family. The results in panel

B indicate that women segregated into jobs with minorities (either blacks or Hispanics)

101

 

are more likely to receive health insurance for themselves and their family. Considering
the high cost of health insurance, these benefits may somewhat offset the lower wages
paid to women in minority dominated jobs. (Unfortunately, the data do not provide an
indication of the level of coverage associated with benefits, so further speculation as to
the economic significance of these results is not pursued here.)
The Determinants QfJob Segregation

I now turn to the investigation of the most probable sources of segregation at the
job level. Since policy makers are likely to be more interested in the effects of
segregation on the labor market outcomes of minorities, I focus on the determinants of
job segregation for blacks and Hispanics, omitting whites from the following analyses.
As a first pass at attempting to understand the various potential sources of j ob
segregation, Table 9 presents several of the characteristics discussed above by
race/ethnicity and whether one works in a segregated job (being in a segregated job is
defined as working with mostly blacks if the individual is black, or working with mostly
Hispanics if the individual is Hispanic). Cutting the data this way reveals some
interesting relationships.

Table 9 indicates that both blacks and Hispanics who work in segregated jobs
also live in more segregated neighborhoods (defined at the census tract level), while those
who are not segregated by job commute longer distances to work. For blacks, those who
work in a segregated job are more likely to have daily face-to-face customer contact,
while Hispanics are less likely to have direct customer contact if they work in a
segregated job. The distribution of educational attainment is lower for both groups of

minorities who are segregated at the job level. Being unionized, working in a larger firm,

102

 

and having supervisory authority are all characteristics that seem to decrease the
likelihood of job segregation for minorities.

For Hispanics, job segregation is strongly related to English fluency and
citizenship status. Table 9 shows that Hispanics who work in segregated jobs are much
more likely to have poorer English skills and are much less likely to be a US. citizen.
The metropolitan area dummies reveal yet another interesting relationship. Blacks in
Atlanta appear to be more likely to be segregated by job. Since the south is a region with
a long history of intense racial discrimination, this may indicate the existence of some
residual employer and/or consumer discrimination. Likewise, Hispanics living in the Los
Angeles metropolitan area, a region known for ethnic tensions, are more likely than not to
be segregated by job. Finally, the majority of blacks and Hispanics living in Boston are
not working in segregated jobs.

Table 10 further investigates these relationships in a multivariate setting. The
results confirm the first striking relationship we saw above in the summary statistics of
Table 9: Minorities who face more residential segregation are more likely to be
segregated in the job market. Note also that blacks and Hispanics who commute longer
distances to work are significantly less likely to be in a segregated job. These two
relationships seem to support the theory regarding the impact of spatial mismatch on job
segregation.28

The impact of residential segregation is particularly sizable for blacks, with a one

hundred percent increase in the proportion black in one’s census tract implying roughly a

 

28 In models not reported here, an interaction term between commute time and residential segregation was
included and indicated that the effect of residential segregation on job segregation declines significantly
with commute time. While most estimates of other included controls did not vary significantly with the
addition of this interaction term, commute time itself lost its significance in several of the specifications.

103

40 percent increase in the likelihood of working in a segregated job. For Hispanics,
columns (3)-(5) indicate that a one hundred percent increase in the proportion Hispanic in
one’s neighborhood implies an increase in the probability of job segregation on the order
of 14-20 percent. (These estimates are calculated at the mean of the data.) Commute
time, on the other hand, seems to have a much smaller impact on the likelihood of job
segregation.

Once controlling for the appropriate covariates, whether one works in a job with
daily face-to-face customer contact does not seem to affect the likelihood of job
segregation. Although column (3) for Hispanics indicates that Hispanics in jobs with
customer contact are less likely to work in a segregated job, the significance of this result
falls apart in columns (4) and (5), when controls for English fluency and citizenship are
included. Nevertheless, these results alone should not necessarily be taken as convincing
evidence that customer discrimination does not affect the probability of job segregation.
Again, a better measure of the likelihood of facing customer discrimination would be a
measure of customers’ racial and ethnic composition, interacted with face-to-face contact,
since one would expect less discrimination by customers of one’s same race or
ethnicity.29

Columns (4) and (5) indicate that Hispanics who cannot speak English “well” or
“at all” are significantly more likely to be segregated into jobs dominated by Hispanics.
The impact of these variables is quite sizable, with the inability to speak English “well”
or “at all” suggesting a 14-24 percent increase in the probability of job segregation,

compared to those who can speak English “very well.” The citizenship status of

104

 

Hispanics also seems to play a significant role in the determination of whether one is
segregated by job, with non-US. citizens 7-11 percent more likely to be segregated than
US. citizens by birth. Such segregation may result from personal choice, employer
discrimination or customer discrimination, or some combination of the three.

While part-time work status does not significantly affect the likelihood of j ob
segregation once I control for other covariates, being unionized seems to significantly
reduce job segregation for blacks and Hispanics. (Perhaps unions serve to provide an
important source of labor market information, especially for newly immigrated
Hispanics). Supervisory authority and the ability to set the pay of coworkers, in general,
does not appear to be strongly associated with the likelihood of being in a segregated job,
although column (2) for blacks indicates that those in supervisory positions are less likely
to be segregated.

Table 10 also investigates the impact of firm size and region on whether
minorities are likely to be racially or ethnically segregated by job. Both of these
variables may be indicative of employer discrimination. Since larger firms are subject to
Affirmative Action laws and more likely to have employees that specialize in labor law,
we might expect to see a lesser degree of racial and ethic segregation along job lines
among larger firms. This is essentially what is seen in Table 10, which shows that an
increase in firm size is associated with a significant decrease in the likelihood of job
segregation for both Blacks and Hispanics. Table 10 also confrrrns that region of the
country (for which metropolitan area serves as a proxy) is likely to play a significant role

in the likelihood of a minority being segregated. Blacks in Atlanta are more likely to be

 

29 To test whether the effect of meeting customers face to face depends on the proportion minority in one’s
neighborhood, measures of residential segregation were interacted with the customer contact dummy. This

105

segregated by job than blacks in other regions, and Hispanics are more likely to be
segregated by job if they work in Los Angeles. The metropolitan dummy variables may
be picking up both employer and customer discriminatory tastes, which are likely to be
correlated by region. However, Hispanics in Los Angeles may be more likely to work in
segregated jobs than those in other regions simply due to the relatively high proportion of
Hispanics living in the area.
Potential Econometric Problems

Several econometric problems need to be addressed before we can draw any firm
conclusions from the above empirical results. The first set of potential problems stems
from limitations of the data used in this study. As discussed above, job segregation is
measured from a survey question asking individuals the race/ethnicity of most of their
fellow coworkers doing similar work. The survey question does not refer to a specific
level of occupational disaggregation (although most respondents likely refer to a fairly
detailed occupation in their establishment). Measurement error in the binary indicators of
job segregation therefore may be an issue. If the measurement error is classical (i.e.,
mean-zero “white noise”), however, the estimated effects of job segregation presented in
the above tables are biased upwards towards zero.30

The lack of segregation data by industry and occupation is unlikely to introduce a

significant amount of bias to the results of this paper. The findings of Bayard et a1.

(1999) show that while segregation at the establishment and job level is severe and has a

 

interaction term was generally positive, but insignificant in all specifications.

3° Another potential problem with this measure of job segregation is that one can not determine how many
other people do the job in question. This is only a concern for cases in which there are very few coworkers
doing similar work, in which case the measure will not be very meaningful. As an extreme example,
suppose there are only two people performing a particular job within a firm, both of whom are Hispanic.
The respondent in question will be classified as working in a segregated job.

106

 

significant impact on wages, there is relatively little racial or ethnic segregation along
occupation or industry lines. Overall, they find that the negative segregation effect for
minorities stems primarily from job-level segregation. Moreover, any such bias is likely
reduced by the inclusion of individual controls for industry and occupation.

As with most empirical studies, the possibility of omitted variables is another
potential econometric problem that needs to be addressed. Individuals who are in
segregated jobs may be different than those who are not in segregated jobs in ways not
captured by the included controls. Such unobserved heterogeneity will bias the above
estimates of the effect of job segregation on earnings and benefits if such differences are
related to labor market outcomes. For example, those in segregated jobs may be less
skilled than those in non-segregated jobs (as predicted by the “quality sorting”
hypothesis). Differences in job segregation will therefore reflect not only forms of
discriminatory behavior, but also unmeasured quality differences. Omitted differences in
skill will be captured by the error term and cause the above estimates of the segregation
effect to be downward biased (away from zero).

I deal with this omitted variable problem by using family background
characteristics as proxy variables for unmeasured skill. As noted above, Neal and
Johnson (1996) conclude that the disadvantages young black workers face in the labor
market arise mostly from obstacles they faced as children in acquiring productive human
capital. It is likely that personal human capital levels are correlated with both parental
human capital and family structure during one’s upbringing. As such, controls for living
with both parents until age 16, parental education in years, and whether parents worked

when the individual was age 16 are included to control for unobserved heterogeneity.

107

 

 

The inclusion of family background variables should also reduce any omitted variable

bias in the estimates of the coefficients on the black and Hispanic dummy variables.3 ‘
Table 11 lists coefficient estimates on the black and Hispanic dummy variables

and on the job segregation controls, replicating the models shown in Table 4 with the

addition of family background proxies for quality. The results provide little evidence of

omitted variable bias affecting the job segregation coefficient estimates. In fact, when I l?
i

controlling for family background, the estimates of penalties associated with working in 1‘:

jobs dominated by Hispanics actually increase slightly. On the other hand, omitted L '

variable bias may have caused the wage gaps to be overestimated in Table 4. When
family background variables are included, the black-white wage gap for men is no longer
significant, the Hispanic-white wage gap for men is reduced by about one-third, and the
Hispanic-white wage gap for women also loses its significance.32
Table 12 investigates this issue by demographic group, reporting coefficient
estimates on the job segregation controls from models similar to those in Table 5, but
with family background controls. Again, there is little change in the job segregation

coefficient estimates and little evidence of omitted variable bias.33 If the family

background variables are indeed controlling for skill differences not captured by the other

 

3' Neal and Johnson (1996) find such family background variables to be strongly correlated with AF QT
scores, and to explain a portion of the black-white gaps in AF QT.

32 The family background variables themselves were consistently insignificant all wage equations (with the
exception of mother’s education implying a significant increase in wages for men only). For this reason,
coupled with the fact that coefficients on the job segregation variables of interest did not changed
significantly and the family background variables are missing for a significant fraction of the sample, these
background controls were left out of the wage equations in previous tables.

’3 It is important to note that the estimates presented in Tables 10 and 11 are based on significantly smaller
samples than those reported in Tables 4 and 5 due to missing information on family background. For
example, roughly 25 percent of the full sample reported that they did not know the number of years their
father was in school. This causes some of the estimates (in Table 12 in particular) to be estimated from
very small cells, which is likely to explain most of the significant variations between Table 5 results and
Table 12 results.

108

 

controls, a significant reduction in the job segregation effects would be expected if
quality sorting at the job level were the only force behind segregation.

Although the results presented in Table 10 suggest that spatial mismatch (as
proxied by residential segregation) contributes to job segregation, it is worth investigating
this relationship a bit further before drawing any firm conclusions. First of all, it is
important to note that even in the absence of discrimination and spatial mismatch,
workforces will still resemble neighborhoods in terms of racial and ethnic composition.
Jobs in more segregated neighborhoods will be more segregated, since the probability
that an employer hires a minority should rise with the proportion of minority applicants.34
If residential segregation is merely picking up such a neighborhood effect and not spatial
mismatch, however, there are no obvious wage implications. Therefore, the fact that
working in minority dominated jobs was found to significantly affect wages calls into
question the possibility that job segregation is simply due to the fact that neighborhoods
tend to be racially and ethnically segregated.

As a firrther probe, Table 13 provides estimates of coefficients on measures of job
segregation and residential segregation for wage models similar to those depicted in
Table 5. If job segregation is due solely to the ethnic makeup of one’s neighborhood, and
not to labor market discrimination or spatial mismatch, then holding constant residential
segregation, it should not affect wages. Table 13 shows that job segregation still

negatively affects wages even after controlling for residential segregation. For male

 

3" This logic also predicts a negative relationship between firm size and job segregation if larger firms are
likely to hire more people into any given job. Note also that if there is residential segregation but no spatial
mismatch problems (i.e., firms and workers are similarly distributed across metropolitan areas), commute
time will still be negatively related to job segregation. However, since the relocation of firms to the
suburbs is well documented, it is unlikely that such findings between commute, residential segregation and
job segregation should be characterized as arising due to neighborhood composition effects only.

109

 

 

minorities, measures of the proportion minority in one’s neighborhood generally has a
negative but insignificant effect on wages, and does not have an appreciable effect on
either the magnitude or the significance of the job segregation variables. Likewise,
residential segregation measures do not have a significant impact on wages for Hispanic
women, nor do they significantly alter the effect of job segregation. The only minority
group for whom residential segregation seems to substantially reduce wages, and reduce
the negative impact of j ob segregation, is black women. Interestingly, (with the
exception of black women) Table 13 indicates that living in minority dominated
neighborhoods has the biggest negative wage effect for whites.

Since not everyone works where they live, Table 14 takes this analysis a step
further by controlling for not only whether one lives in a minority dominated
neighborhood (now measured by a dummy variable for a neighborhood that is at least 50
percent black or Hispanic), but also whether they are likely to work in a minority
dominated neighborhood. This is accomplished by adding an indication of a short
commute (less than or equal to 10 minutes), and a measure of short commute interacted
with minority dominated residence.35 The results support the contention that job
segregation is not due solely to the makeup of one’s neighborhood. Comparing the
results of Table 5 with those of Table 14, the negative effects of j ob segregation on wages
remain even after controlling for the racial and ethnic makeup of where one lives and

works, leaving spatial mismatch, discrimination, and/or other factors as potential causes.

 

3’ Several different specifications of these variables were explored, with a short commute defined as less
than or equal to 5, 10, 15, or 20 minutes, and a minority dominated neighborhood defined as having at least
30, 50, or 75 percent blacks or Hispanics. The job segregation effects did not change significantly under
these different scenarios; however, the estimates on the dummy variables themselves (and the interaction
term) were somewhat sensitive to how they were defined (although they were generally insignificant). For
this reason, I do not put much emphasis on the dummy variable estimates.

110

There is one more implication of this simple test. If spatial mismatch (i.e.,
relatively low labor demand in minority dominated areas) were the only reason for job
segregation, then otherwise-identical workers working in equally segregated areas should
earn the same amounts, whether or not their jobs are segregated. Thus, Table 14 also
suggests that job segregation is likely due to more than just spatial mismatch, since job
segregation still matters even after controlling for neighborhood segregation. In addition,
if the interaction term accurately picks up whether one works in a minority dominated
neighborhood, the results seem to indicate that spatial mismatch is only a significant
problem for black women in terms of its wage effects}6 However, it is important to keep
in mind that past studies (which tend to focus on blacks) find that spatial mismatch has a
stronger effect on employment rates than on wages.37

Another econometric issue concerns the estimated effects of commute time on job
segregation. Underlying the results presented in Table 10 is the assumption that
minorities take spatial mismatch as a given constraint, and then decide whether or not
they can commute longer distances to overcome its negative effects on crowding. As
predicted by the discussion outlined in the theory section, I find that commute time is
negatively related to the likelihood of job segregation. However, Ihlanfeldt and Sjoquist
(1990) interpret high average commute times in a neighborhood as indicative of more
mismatch and poorer job access under the assumption that individuals facing more spatial
mismatch (particularly inner-city minorities) will have to commute more. Since more

spatial mismatch would imply more job segregation, this interpretation suggests that a

 

36 It is interesting to note that living in a segregated neighborhood actually seems to be related to higher
wages for Hispanic females. This may be indicative of the importance of social contacts for new
immigrants.

111

 

measure of average commute time may be positively correlated with job segregation.
Therefore, if higher commute times reflect more spatial mismatch, in addition to the
efforts to overcome its negative effects, the estimated coefficients on commute time in
the above job segregation probit equations are muddied by these opposing relationships
and biased downward towards zero.

It is also possible that commute time may be simply capturing the empirical
finding that more-educated and hi gher-wage individuals generally commute longer
distances. However, the correlation between commute time and income is strongest for
whites, who are much less likely to be constrained by the discrimination and zoning
problems that often restrict the residential choices of minorities (Holzer, 1991). Thus
estimating separate equations for blacks and Hispanics likely mitigates the severity of this
problem.38 (While instrumental variables provides another potential way to deal with
these potential econometric problems associated with commute time, variables that
influence commute time but that do not directly influence the likelihood of being in a
segregated job would be required. Unfortunately, there are no compelling choices for
such instruments available in MCSUI”)

Lastly, I turn to the issue of whether the negative effect of job segregation on

wages merely represents a compensating differential that minorities are willing to pay to

 

’7 Another possibility is that, since the sample sizes are rather small, there is simply not enough
independent variation between residential segregation and job segregation.

’8 Ihlanfeldt and Sjoquist (1990) find substantially larger effects of travel time on employment rates for
blacks than for whites, supporting the notion that the bias (downward in their case) resulting from
endogeneity is reduced by running separate equations. The inclusion of controls for educational attainment
and other wage determinants in the logit segregation equations should also reduce the likelihood of the
commute time measure capturing the fact that more-educated and higher-wage people generally commute
over further distances.

’9 Cutler and Glaeser (1997) use this IV approach to deal with the possibility of residential segregation
being endogenous with respect to their outcome measures of education, income and single motherhood.
The two instruments they use for residential segregation are the number of municipal and township

112

 

work with coethnics. That is, how much of the job segregation we see is a product of
personal choice, as opposed to a real constraint due to spatial mismatch or some form of
discrimination? This is an important question because self-selection into a segregated job
will lead to the same outcomes as discrimination, but with drastically different policy
implications. For instance, while some may view the significance of English fluency
controls in the job segregation models for Hispanics as indicative of employer or
customer discrimination, it is also possible that Hispanics with weak English skills may
choose to work along side those with whom they can more easily communicate.

To further gauge the importance of personal choice or self-selection in the job
segregation process, I use a series of variables that assesses racial attitudes as a proxy for
the preference to work with coethnics. Survey respondents were asked to use a scale
from 1 to 7 to express their views concerning the ease of “getting along with” whites,
blacks and Hispanics, where 1 means “tends to be hard to get along with” and 7 means
“tends to be easy to get along with.” A rating of 4 is neutral, meaning “the group is not
towards one end or the other.” For instance, blacks that give a high rating for blacks, and
relatively low ratings for whites and Hispanics, may be interpreted as individuals who
have a taste for working with coethnics.

An indicator of the preference to work with coethnics was created from these
variables to assess the relative importance of personal choice, or self-selection, in the job
segregation process. The self-selection dummy equals one if an individual gave their

own demographic group a higher rating than other demographic groups, meaning they

 

governments in the metropolitan area and the share of local revenue that comes from intergovernmental
sources.

113

”TI «O‘fhj

ll_im ' ,
1- _
1 4|

view their own race/ethnicity as being relatively easier to get along with.40 This is
admittedly a subjective measure, but given the lack of panel data necessary to difference
out such preferences, it should at least provide some indication of whether self-selection
into segregated jobs is a significant concern.

In Table 15, I present the results of including the self-selection dummy in the five
models depicted in Table 10. As an additional probe, I also include the individual scale
variables from which the self-selection dummy was created in separate specifications.
Overall, I found little evidence that job segregation is likely to be due solely to self-
selection as measured by racial attitudes. Panel A of Table 15 shows that the self-
selection dummy variable was generally insignificant for blacks and Hispanics.

When the scale variables were included separately, however, there is at least some
evidence that racial preferences may play a role in the job segregation process. Panel B
of Table 15 indicates that blacks who find Hispanics relatively easy to get along with are
less likely to be segregated into jobs with other blacks, while their racial attitudes towards
whites and other blacks are not significantly related to the likelihood of job segregation.
Perhaps counter-intuitively, it seems that Hispanics who find those of similar ethnic
background relatively easy to get along with are less likely to be segregated into jobs
with other Hispanics. Column (3) of Table 15 seems to indicate that Hispanics who find
blacks relatively easy to get along with are less likely to be in a segregated job, however
this result falls apart in the models indicated by columns (4) and (5). Racial attitudes of

Hispanics towards whites are not significantly related to the likelihood of job segregation.

 

‘0 An alternative self-selection measure was created to equal one if an individual gave their own
race/ethnicity a rating greater than 4, meaning they viewed their own demographic group as being
relatively easy to get along with, but gave other demographic groups a rating less than 4, indicating they

114

 

Overall, this somewhat crude test of the impact of racial attitudes implies that
further investigation of the importance of self-selection in job segregation is merited.41 It
is important to note, however, that the majority of the other covariates were not changed
significantly with the addition of the various measures of racial attitudes. This suggests
that spatial mismatch and discrimination, as indicated by the other determinants of job
segregation previously discussed, are likely to play a role in the job segregation process

above and beyond personal choice.

Conclusions and Policy Implications

The results of this paper can be summarized into three main findings. First, job
segregation is an important contributor to the lower wages paid to minorities than to
whites with similar individual characteristics. Job segregation explains a significant
portion of the black-white wage gap that remains after controlling for differences in
human capital and job characteristics for both men and women. On the other hand, while
job segregation explains a fraction of the Hispanic-white adjusted wage differential for
men, it still remains sizable at around 13 percent. The Hispanic-white wage gap for
women is virtually eliminated with the addition of controls for English fluency and
citizenship, before controlling for job-level segregation. However, job segregation has a
significant negative impact on the earnings of blacks and Hispanic women even after

controlling for differences in a variety of skill measures. For example, Hispanic women

 

 

viewed other races/ethnicities as being relatively difficult to get along with. The results of using this
alternative measure were very similar to those reported in Table 15.

4' A preliminary analysis of the effect of racial attitudes on wages revealed that the self-selection dummy
was insignificant with respect to wages, and the negative effect of job segregation on wages actually
increased when this variable was included. Given the substantial proportion of the sample that chose not to
respond to the racial attitudes questions, however, these results should be viewed with caution due to
potential sample selection issues. A complete analysis of the self-selection issue awaits future research.

115

who are segregated into jobs with mostly Hispanics earn roughly 15 percent less than
those who work with mostly whites. These findings suggest that while “equal pay” laws
may offer some hope for reducing the Hispanic-white differential for men, policies
targeted at alleviating segregation into lower-paying jobs may be more effective at
reducing pay gaps for minorities overall.

Second, job segregation plays a much smaller role in explaining differences in the
number of employment benefits received between minorities and whites than is does in
explaining wage differentials. Starting out with significantly less disparity in unadjusted
benefit differentials between whites and minorities to begin with, in general, job
segregation seems to have little negative impact on the level of benefits received
(although being segregated into jobs with mostly Hispanic men is associated with lower
benefit levels). However, it is important to note that the data used here tell us nothing
about the quality of the benefits packages received, so the above analysis should be
viewed only as an important first step in understanding the impact of j ob segregation on
employment benefits.

Finally, this paper sheds a considerable amount of light on the likely sources of
racial and ethnic job segregation. I find that minorities facing more residential
segregation are significantly more likely to work in a segregated job, while that those
who commute longer distances to work are less likely to be segregated by job. Since
these findings are consistent with the notion of spatial mismatch, this research suggests
that policies aimed at improving the accessibility of public transportation, reducing
housing market discrimination, or encouraging employers to locate in inner-city areas

may help reduce job segregation.

116

This paper also finds that minorities who work in larger firms are less likely to
work in segregated jobs, perhaps because larger employers are less likely to engage in
employment discrimination due to a greater awareness of the legal consequences or to
stronger enforcement of Affirmative Action laws. In addition, English fluency and
citizenship status are found to be strongly associated with the likelihood of j ob
segregation for Hispanics. This suggests that improving language skills and access to
labor market information for Hispanics may help reduce the considerable earnings
disparities we find between Hispanics and whites.

While I view the findings of this study as an important first step in understanding
these issues, more definitive policy conclusions should await further analysis of other
data that improve on some of the shortcomings discussed above. In particular, a more
rigorous examination of the role of personal choice, or self-selecting into segregated jobs,

is clearly needed before any firm conclusions can be drawn.

117

 

REFERERNCES

118

REFERENCES

Angrist, Joshua D. and Alan B. Krueger (1998), "Empirical Strategies in Labor
Economics," Working Paper #401, Princeton University, Industrial Relations Section,
June.

Bates, Timothy (1993), Banking on Black Enterprise, Washington, DC: Joint Center for
Political and Economic Studies.

Bayard, Kimberly, Judith Hellerstein, David Neumark and Kenneth Troske (1998), “Why
are Racial and Ethnic Wage Gaps Larger for Men than for Women? Exploring the Role
of Segregation Using the New Worker-Establishment Characteristics Database,”
Working Paper No. 98-12, Columbia, MO: University of Missouri — Columbia.

Bayard, Kimberly, Judith Hellerstein, David Neumark and Kenneth Troske (1999), “Why
are Racial and Ethnic Wage Gaps Larger for Men than for Women? Exploring the Role
of Segregation Using the New Worker-Establishment Characteristics Database,” NBER
Working Paper #6997, March.

Becker, Gary S. (1971), The Economics of Discrimination, Chicago: the University of
Chicago Press (Original edition, 1957).

Borjas, George J. (1998), “To Ghetto or Not to Ghetto: Ethnicity and Residential
Segregation,” Journal of Urban Economics, 44: 228-253.

Bratsberg, Bemt, and Dek Terrell (1998), “Experience, Tenure, and Wage Growth of
Young Black and White Men,” Journal of Human Resources, 33(3): 658-682.

Cain, Glen G. (1986), “The Economic Analysis of Labor Market Discrimination: A
Survey,” in Ashenfelter, O. and Richard Layard, eds., Handbook of Labor Economics,
New York: Elsevier Science Publishers.

Carrington, William J ., and Kenneth R. Troske (1998), “Interfinn Segregation and the
Black/White Wage Gap,” Journal of Labor Economics, 16(2): 231-260.

Chiswick, Barry R. (1973), “Racial Discrimination in the Labor Market: a Test of
Alternative Hypotheses,” Journal of Political Economy, 81 : 1330-1352.

Chiswick, Barry R., and Paul W. Miller (2001), “Immigrant Earnings: Language Skills,
Linguistic Concentrations and the Business Cycle,” Unpublished Working Paper.

Cordero-Guzman, Hector (1990), “Sociological Approaches to Employment
Discrimination,” Unpublished Manuscript, University of Chicago.

Cutler, David M., and Edward L. Glaeser (1997), “Are Ghettos Good or Bad?” The
Quarterly Journal of Economics, August: 827-871.

119

 

Darity and Mason (1998), “Evidence on Discrimination in Employment: Codes of Color,
Codes of Gender,” Journal of Economic Perspectives, 12(2): 63-90.

Hamermesh, Daniel S., and Albert Rees (1993), The Economics of Work and Pay, (5th
edition), New York: Harper Collins Publishers.

Hirsch, Barry T., and Edward J. Schumacher (1992), “Labor Earnings, Discrimination,
and the Racial Composition of Jobs,” Journal of Human Resources, XXV 11(4): 602-628.

Hirsch, Barry T., and David A. Macpherson (1994), “Wages, Racial Composition, and
Quality Sorting in Labor Markets,” Discussion Paper no. 1038-94, Madison, WI: Institute
for Research on Poverty.

Holzer, Harry J. (1991), “The Spatial Mismatch Hypothesis: What Has the Evidence
Shown?” Urban Studies, XXV 111: 105-122.

(1997), “Why Do Small Establishments Hire Fewer Blacks than Large
Ones?” Journal of Human Resources, XXXII(4): 896-916.

 

Holzer, Harry J ., and Keith Ihlanfeldt (1998), “Customer Discrimination and
Employment Outcomes for Minority Workers,” Quarterly Journal of Economics, August:
833-867.

Ihlanfeldt, Keith R., and David L. Sjoquist (1990), “The Role of Space in Determining
the Occupations of Black and White Workers,” Regional Science and Urban Economics,
XXI: 295-315.

Kain, John F. (1968), “Housing Segregation, Negro Employment, and Metropolitan
Decentralization,” Quarterly Journal of Economics, 82(2): 175-197.

Ladd, Helen F. (1998), “Evidence on Discrimination in Mortgage Lending,” Journal of
Economic Perspectives, 12(2): 41-62.

Macpherson, David A., and Barry T. Hirsch (1995), “Wages and Gender Composition:
Why Do Women’s Jobs Pay Less?” Journal of Human Resources, 13(3): 429-71.

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.

Ortiz, Vilma (1996), “The Mexican-Origin Population: Permanent Working Class or
Emerging Middle Class?” in Roger Waldinger and Mehdi Bozorgrnehr, eds., Ethnic Los
Angeles, New York: Russell Sage Foundation.

Sorensen, Elaine (1989), “Measuring the Effect of Occupational Sex and Race
Composition on Earnings,” In Robert T. Michael, Heidi I. Hartrnann, and Brigid

120

O’Farrel, Eds, Pay Equity: Empirical Inquiries, Washington, DC: National Academy
Press.

Stoll, Michael, A., Harry J. Holzer, and Keith R. Ihlanfeldt (1999), “Within Cities and
Suburbs: Racial Residential Concentration and the Spatial Distribution of Employment
Across Submetropolitan Areas,” Discussion Paper no. 1189-99, Madison, WI: Institute
for Research on Poverty.

Stata Statistical Software: Release 5.0, College Station, TX: Stata Corporation.

Trejo, Stephen J. (1997), “Why Do Mexican Americans Earn Low Wages?” Journal of
Political Economy, 105(6): 1235-1268.

Turner, Margery Austin, Raymond J. Struyk, and John Yinger (1991), Housing
Discrimination Study: Synthesis, Washington DC: US. Department of Housing and
Urban Development.

Turner, Margery Austin, and Maris Mickelsons (1992), “Patterns of Racial Steering in
Four Metropolitan Areas,” Journal of Housing Economics, September (2): 199-234.

Watts, Martin J. (1995), “Trends in Occupational Segregation by Race and Gender in the
U.S.A., 1983-92: A Multidimensional Approach,” Review of Radical Political
Economics, 27(4): 1-36.

Yinger, John (1995), Closed Doors, Opportunities Lost: The Continuing Costs of
Housing Discrimination, New York: Russell Sage Foundation.

Yinger, John (1998), “Evidence on Discrimination in Consumer Markets,” Journal of
Economic Perspectives, 12(2): 23-40.

121

 

 

Table 1
Sample Descriptive Statistics by Demographic Group

 

 

 

Men Women
White Black Hispanic White Black Hispanic
Log hourly wages 2.76 2.48 2.18 2.49 2.29 1.96
(.040) (.053) (.034) (.034) (.040) (.037)
# of employment benefits 2.96 2.74 1.87 2.54 2.43 1.62
(.116) (.138) (.133) (.117) (.127) (.128)
Age 39.42 35.76 34.39 39.70 37.04 35.60
(.767) (1.11) (.597) (.861) (.749) (.775)
Married .604 .426 .590 .601 .306 .457
Number of children < 18 .604 .656 .998 .738 .736 1.218
(.063) (.109) (.092) (.062) (.067) (.083)
No high school degree .024 .065 .443 .049 .068 .443
High school/GED .360 .524 .369 .400 .467 .315
Some college .147 .135 .076 .204 .275 .148
Bachelor’s degree .293 .192 .090 .251 .134 .076
Advanced degree .176 .084 .022 .096 .056 .018
Tenure (in years) 7.41 5.79 4.93 6.71 6.20 4.70
(.510) (.541) (.305) (.531) (.413) (.428)
Prior experience injob 4.64 3.53 2.61 4.24 2.85 2.12
(in years) (.472) (.583) (.349) (.486) (.245) (.333)
Part-timejob .104 .189 .156 .331 .190 .235
Collective bargaining .169 .260 .181 .139 .229 .132
Speaks English:
Very well .918 .786 .342 .909 .835 .379
Well .077 .184 .243 .077 .1 18 .167
Not well .007 .029 .370 .014 .047 .366
Not at all .000 .000 .045 .000 .000 .088
Citizenship status:
By birth in US. .929 .760 .298 .900 .932 .298
By naturalization .036 .047 .077 .064 .025 .132
Not a citizen .035 .193 .625 .037 .042 .618
Atlanta .137 .284 .012 .168 .327 .013
Boston .404 .134 .048 .408 .126 .049
Los Angeles .459 .582 .940 .423 .548 .938

 

122

 

._ .. —£

a?

t'

Firm size

Job tasks performed daily:

Customer contact
Telephone
Reading

Writing

Use computer
Mathematics

Supervise others
Set the pay of others

Coworkers are
predominantly:
White
Black
Hispanic
Asian
Other (or mixed)

Indust_ry:
Construction

Durable manufacturing
Non-durable mfg.
Transportation
Wholesale/retail trade
F inance/insurance/
real estate

Service

Public/self employed/
other

Occupation:
Managerial

Professional/technical
Sales
Clerical/administration
Service/labor
Crafts/operative

Self employed/other

Table l (cont’d)

 

 

hden VVonunr
White Black Hispanic White Black Hispanic
453 473 219 481 562 203
(72.7) (71.0) (80.6) (78.06) (70.8) (39.2)
.594 .549 .445 .627 .665 .469
.554 .549 .245 .619 .616 .322
.590 .649 .485 .558 .561 .355
.429 .376 .285 .459 .412 .260
.556 .440 .181 .575 .516 .261
.700 .674 .395 .662 .566 .366
.365 .320 .260 .287 .253 .171
.115 .100 .096 .070 .041 .033
.801 .338 .185 .838 .389 .215
.034 .355 .028 .023 .414 .021
.059 .119 .723 .032 .072 .688
.010 .004 .033 .027 .042 .023
.097 .184 .032 .080 .082 .053
.066 .063 .142 .013 .008 .004
.102 .040 .137 .077 .047 .105
.063 .124 .106 .043 .023 .157
.098 .123 .090 .034 .083 .037
.122 .117 .237 .196 .142 .180
.075 .086 .035 .089 .106 .046
.385 .368 .209 .505 .527 .432
.152 .125 .184 .052 .070 .031
.196 .078 .069 .131 .079 .053
.310 .201 .071 .317 .198 .089
.105 .057 .042 .126 .138 .101
.083 .169 .076 .261 .296 .163
.101 .284 .270 .110 .223 .303
.188 .176 .444 .044 .059 .283
.016 .019 .024 .009 .001 .008

 

Note: Standard errors are reported in parentheses for continuous variables. Standard errors for binary variables with
mean p can be calculated by taking the square root of p(l-p)/n. Sample sizes are 616, 476 and 610 for white males,
black males and Hispanic males, respectively, and 644, 916, and 633 for white females, black females and Hispanic
females, respectively. Some estimates are based on slightly smaller samples due to missing values.

123

 

Table 2
Hourly Wages and Log Wages
byDemogaphic Group and Job Segregation

 

 

 

 

 

A. Hourly Wages Men Women
Coworkers are White Black Hispanic White Black Hispanic
predominpntlv:
White 20.41 15.56 12.39 14.45 11.92 10.83
(1.519) (.914) (.857) (.839) (.637) (.627)
[16.92] [14.80] [11.52] [12.00] [11.00] [10.00]
Black 15.75 14.17 11.57 13.91 10.59 11.61
(2.497) (2.943) (3.066) (2.306) (.545) (1.886)
[14.00] [10.00] [11.00] [12.90] [8.50] [14.50]
Hispanic 12.61 15.72 9.33 7.26 16.21 6.95
(1.287) (4.168) (.439) (1.200) (6.86) (.362)
[10.50] [11.54] [7.50] [6.00] [7.50] [5.60]
B. Log Wages Men Women
Coworkers are White Black Hispanic White Black Hispanic
predommntlv:
White 2.81 2.64 2.40 2.50 2.37 2.28
(.046) (.061) (.072) (.03 8) (.055) (.059)
[2.83] [2.69] [2.44] [2.48] [2.40] [2.30]
Black 2.59 2.33 2.25 2.53 2.20 2.37
(. 134) (.093) (.266) (. 157) (.047) (.179)
[2.64] [2.30] [2 .40] [2.56] [2. 14] [2.67]
Hispanic 2.44 2.46 2.10 1.85 2.32 1.82
(.094) (.263) (.03 6) (. 149) (.254) (.040)
[2.36] [2.45] [2.01] [1.79] [2.01] [1.72]

 

 

 

Note: Standard errors of estimates are reported in parentheses. Median wages of distributions reported in
brackets. Wages of individuals working with mostly Asians or other/mixed available upon request.

 

Black

Hispanic

Age

Age2

High school/GED

Associates/

Vocational/Trade

Bachelor’s degree

Advanced degree

Part-time job

Married

# children <18

Prior experience

Tenure

Tenure2

Unionized

Ln (firm size)

Table 3
Log Hourly Wage Regressions

without Controlling for Job Segregation

 

 

Males
(1) (2) (3)
m149 al44 n078
(054) (055) (043)
a353 n299 m176
(052) (064) (050)
.016 .017 .022
(013) (013) (012)
a0002 n0002 «0003
(0002) (0002) (0001)
.190 .129 .114
(044) (048) (043)
.229 .164 .119
(061) (065) (056)
.504 .427 .265
(062) (064) (055)
.624 .518 .204
(091) (087) (083)
m067 a046 .005
(077) (075) (052)
.047 .054 .083
(050) (049) (037)
.033 .035 .034
(022) (022) (018)
.015 .015 .015
(004) (004) (003)
.054 .051 .038
(008) (008) (007)
a001 «001 a0008
(0003) (0003) (0002)
.069 .073 .164
(049) (048) (044)
.007 .008 .012
(012) (011) (010)

1225

Females
(4) (5) (6)
m118 n119 n086
(041) (042) (035)
a223 m096 n025
(050) (065) (043)
.027 .029 .020
(014) (014) (009)
a0003 «0004 a0002
(0002) (0002) (0001)
.362 .274 .139
(058) (068) (056)
.476 .393 .238
(056) (067) (061)
.679 .591 .363
(065) (077) (069)
.700 .629 .369
(103) (106) (098)
a062 m071 .032
(049) (048) (044)
.007 .020 mOOS
(039) (038) (029)
a007 a006 n005
(018) (017) (015)
.013 .012 .010
(004) (004) (003)
.020 .019 .007
(013) (013) (011)
.0003 .0003 .0006
(001) (0006) (0005)
a025 «042 .023
(.048) (.046) (.040)
.022 .019 .018
(009) (008) (008)

 

_ fi ls-a-l—q

Table 3 (cont’d)

 

Males Females
(1) (2) (3) (4) (5) (6)
Speaks English well -- -.l94 -.125 -- -.020 -.033
(.048) (.041) (.062) (.043)
Does not speak
English well -- -.180 -.129 -- -.047 -.091
(.071) (.061) (.099) (.060)
Does not speak
English -- -.268 -.081 -- -.275 -.320
(.092) (.109) (.109) (.085)
US. citizen by
nationalization -- .202 .003 -- -. 161 -. 181
(. 124) (.066) (.090) (.062)
Not a US. citizen -- .012 -.029 -- -.209 -.173
(.060) (.056) (.080) (.058)
Supervise others -- -- .034 -- -- -.O78
(.038) (.047)
Supervise and set
the pay of others -- -- .135 -- -- .323
(.068) (.067)
Daily job tasks:
F ace-to-face
customer contact -- -- -.050 -- -- -.093
(.035) (.036)
Talk w/ customers
on telephone -- -- .054 -- -- .098
(.042) (.039)
Reading
paragraphs -- -- .029 -- -- -.007
(.032) (.032)
Writing paragraphs -- -- .029 -- -- .130
(.039) (.036)
Work w/ computer -- -- .113 -- -- .096
(.042) (.041)
Arithmetic -- -- .01 1 -- -- -.017
(.039) (.034)
Occupation controls: No No Yes No No Yes
Industry controls: No No Yes No No Yes
N 1,640 1,638 1,438 2,059 2,059 1,913
QRZ .464 .479 .578 .458 .474 .623

 

 

Note: Robust standard errors of regression estimates reported in parentheses. Omitted categories are “high school
dropout” for educational attainment, “speaks English very well” for English fluency, and “US citizen by birth” for
citizenship. All regressions include a constant, a missing value dummy for part-time job, and control for
metropolitan area and whether reported job information is for most recent job. Sample sizes differ between
regressions due to missing values.

126

Table 4
Log Hourly Wage Regressions:
The Impact of Job Sggggation

 

 

 

Males Females
(1) (2) (3) (4) (5) (6)
Black -.094 -.091 -.051 -.O69 -.071 -.053
(.053) (.054) (.048) (.046) (.046) (.041)
Hispanic -.251 -.219 -.13 l -.099 -.001 .022
(.052) (.059) (.048) (.054) (.065) (.048)
Coworkers are
Predominantly:
Black -. 130 -.131 -.106 -.089 -.096 -.085
(.071) (.070) (.055) (.048) (.048) (.052)
Hispanic -.211 -. 167 -.108 -.233 -.204 -.118
(.05 8) (.060) (.049) (.064) (.067) (.061)
Asian -.022 -.010 -.051 -.016 -.002 -.068
(.161) (.138) (.101) (.058) (.064) (.054)
Other (or mixed) -.092 -.099 -.040 .063 .066 .039
(.066) (.064) (.065) (.053) (.052) (.050)
N 1,587 1,585 1,428 1,995 1,995 1,899
R2 .485 .499 .581 .477 .491 .627

 

Note: Included controls correspond to those listed in Table 3. Robust standard errors of regression

estimates are reported in parentheses. Sample sizes differ due to missing values.

127

 

Table 5a
Log Hourly Wage Regressions:
The Impact of Seggggation pyDemogaphic Grorm

 

 

 

 

 

 

Black
Males Females
Coworkers are (1) (2) (l) (2)
Predominaptlv:
Black -.192 -. 142 -.088 -.086
(.084) (.059) (.047) (.047)
Hispanic -. 198 -.078 .072 .049
(.127) (.090) (. 195) (.099)
Asian -.135 -.l63 -.026 -.031
(.174) (.245) (.061) (.074)
Other (or mixed) .066 .056 -.004 -.005
(.l 15) (.080) (.063) (.053)
High school/GED -.003 -.039 .193 .083
(.133) (.121) (.066) (.065)
Associates/
Vocational/Trade .038 -.040 .320 .153
(. 163) (.143) (.079) (.076)
Bachelor’s degree .472 .176 .664 .410
(.168) (.127) (.085) (.095)
Advanced degree .481 .343 .762 .563
(.206) (.162) (.091) (.103)
Part-time job -.082 -.020 -.102 .029
(.104) (.113) (.055) (.055)
Married -.073 .042 .055 .003
(.094) (.065) (.058) (.043)
# children <18 .019 -.008 .003 -.006
(.036) (.024) (.019) (.014)
Prior experience .021 .005 .011 .011
(.010) (.006) (.005) (.004)
Tenure .057 .049 .043 .028
(.015) (.012) (.011) (.008)
Tenure:2 -.001 -.001 -.0008 -0005
(.0005) (.0004) (.0003) (.0003)
Unionized .065 .126 .071 .148
(.073) (.050) (.052) (.047)
Ln (firm size) -.002 .008 .018 .021
(.019) (.014) (.011) (.009)

128

 

 

Table 5a (cont’d)

 

 

 

Black
Females
(l) (2) (3) (4)
Supervise others -- -.018 -- .049
(.055) (.043)
Supervise and set
the pay of others -- .144 -- .383
(.117) (. 156)
Daily iob tasks:
F ace-to-face
customer contact -- -.031 -- -.125
(.061) (.046)
Talk w/ customers
on telephone .028 .07 5
-- (.054) -- (.057)
Reading
paragraphs .159 -.O33
-- (.064) -- (.043)
Writing paragraphs -.017 .073
-- (.058) -- (.042)
Work w/ computer .141 .1 15
-- (.064) -- (.043)
Arithmetic
-- -.012 -- .034
(.050) (.038)
Occupation controls: No Yes No Yes
Industry controls: No Yes No Yes
N 448 414 843 821
R2 .464 .608 .505 .635

 

7“”!

. mu. 'a

tn
..1

Note: Robust standard errors of regression estimates are reported in parentheses. The omitted categories are “high
school dropout” for educational attainment, “speaks English very well” for English fluency, and “US citizen by birth”
for citizenship. All regressions include a constant, a missing value dummy for part-time job, and control for age,
metropolitan area and whether reported job information is for most recent job. Sample sizes differ between regressions
due to missing values.

129

Table 5b
Log Hourly Wage Regressions:

The Impact of Smtion by Demographic Group

 

Coworkers are
Predonnppntly:
Black

Hispanic

Asian

Other (or mixed)

High school/GED

Associates/

Vocational/Trade

Bachelor’s degree

Advanced degree

Part-time job

Married

# children <18

Prior experience

Tenure

Tenure2

Unionized

Ln (firm size)

 

 

 

Males
(1) (2)
«380 «347
(162) (097)
«196 «097
(071) (073)
«010 «018
(191) (111)
«002 .063
(096) (122)
.200 .098
(063) (051)
.201 .088
(090) (079)
.199 .089
(109) (076)
.649 .299
(199) (146)
.169 .103
(077) (064)
.094 .121
(065) (046)
«002 «0004
(018) (019)
.023 .014
(007) (006)
.071 .067
(016) (011)
«002 «002
(0009) (0007)
.210 .198
(074) (063)
.028 .048
(.019) (.017)

13()

Females
(1) (2)
«031 «234
(084) (093)
«148 «144
(081) (055)
«042 «049
(097) (124)
.091 «140
(096) (108)
.321 .127
(082) (078)
.330 «019
(101) (070)
.678 .167
(118) (084)
.170 «123
(299) (163)
.034 «041
(066) (059)
«032 .095
(061) (050)
«004 «026
(025) (018)
«007 «008
(007) (006)
.035 .012
(016) (013)
«0006 «0002
(0005) (0005)
.036 «035
(076) (061)
.029 .024
(015) (014)

 

Table 5b (cont’d)

 

 

Hispanic Males Hispanic Females
(1) (2) (3) (4)
Speaks English well -- «021 -- «005
(.059) (.065)
Does not speak
English well -- «005 -- « 164
(.074) (.063)
Does not speak
English -- .034 -- «314
(.104) (.066)
US. citizen by
nationalization -- -. 1 60 -- «214
(.110) (.080)
Not a US. citizen -- «O92 -- «120
(.072) (.070)
Supervise others
-- «123 -- «061
(.066) (.059)
Supervise and set
the pay of others -- .125 -- .063
(.097) (.138)
Daily job tasks:
F ace-to-face
customer contact -- .050 -- «013
(.046) (.066)
Talk w/ customers
on telephone .119 .113
-- (.061) -- (.064)
Reading
paragraphs «038 .041
-- (.047) -- (.050)
Writing paragraphs .063 .161
-- (.059) -- (.055)
Work w/ computer .165 «066
-- (.063) -- (.066)
Arithmetic -- .060 -- .033
(.043) (.046)
Occupation controls: No Yes No Yes
Industry controls: No Yes No Yes
N 580 537 577 545
R2 .428 .580 .470 .676

 

 

Note: Robust standard errors of regression estimates are reported in parentheses. The omitted categories are “high
school dropout” for educational attainment, “speaks English very well” for English fluency, and “US citizen by birth”
for citizenship. All regressions include a constant, a missing value dummy for part-time job, and control for age,
metropolitan area and whether reported job information is for most recent job. Sample sizes differ between regressions
due to missing values.

131

Table 5c
Log Hourly Wage Regressions:
The Impact of Segregation by Demographic Group

 

 

White
Males Females
Coworkers are (1) (2) ( 1) (2)
Predominantly:
Black .030 .045 .029 .083
(.113) (.119) (.131) (.117)
Hispanic «137 «039 «503 «332
(.117) (.083) (.129) (.104)
Asian .094 «230 «035 «060
(. 100) (.086) (.098) (.086)
Other (or mixed) «125 «096 .100 .113
(.084) (.095) (.071) (.067)
High school/GED .074 .163 .434 .262
(.145) (.134) (.120) (.094)
Associates/
Vocational/Trade .105 .164 .519 .382
(.141) (.134) (.119) (.093)
Bachelor’s degree .384 .359 .681 .496
(.161) (.149) (.123) (.108)
Advanced degree .451 .273 .742 .457
(.165) (.142) (.145) (.135)
Part-time job «232 «106 «108 .033
(.140) (. 100) (.071) (.063)
Married .080 .097 .022 «033
(.074) (.060) (.059) (.041)
# children <18 .066 .089 «021 «013
(.041) (.033) (.029) (.025)
Prior experience .01 l .014 .017 .011
(.007) (.005) (.005) (.003)
Tenure .055 .032 .005 «002
(.012) (.009) (.016) (.013)
Tenurez «001 «001 .0009 .001
(.0004) (.0003) (.0007) (.001)
Unionized «030 .151 «034 .063
(.065) (.072) (.068) (.061)

132

 

 

 

Table 5c (cont’d)

 

 

 

 

White
Males
(1) (2) (3) (4)
Ln (firm size) .0004 .004 .021 .006
(.016) (.014) (.013) (.012)
Supervise others -- .067 -- «119
(.055) (068)
Supervise and set
the pay of others -- .159 -- .381
(.091) (.084)
Daily iob tasks:
Face-to-face
customer contact -- «081 -- «083
(.058) (.060)
Talk w/ customers
on telephone -- .041 -- .100
(.063) (.055)
Reading
paragraphs -- .01 l -- «039
(.053) (.048)
Writing paragraphs -- .050 -- .141
(.059) (.056)
Work w/ computer -- .061 -- .114
(.061) (.058)
Arithmetic -- «071 -- «010
(.065) (.052)
Occupation controls: N Y N Y
Industry controls: N Y N Y
N 559 479 575 533
R2 .406 .51 1 .395 .592

 

Note: Robust standard errors of regression estimates are reported in parentheses. The omitted categories are “high
school dropout” for educational attainment, “speaks English very well” for English fluency, and “US citizen by birth”
for citizenship. All regressions include a constant, a missing value dummy for part-time job, and control for age,
metropolitan area and whether reported job information is for most recent job. Sample sizes differ between regressions
due to missing values.

133

Table 6

Ordered Logit Estimates from Job Benefits Models

without Controllingfor Job Seggggation

 

 

Black

Hispanic

Age

2

Age

High school/GED

Associates/

Vocational/Trade

Bachelor’s degree

Advanced degree

Part-time job

Married

# children <18

Prior experience

Tenure

Tenure2

Unionized

Ln (firm size)

 

 

Males
(1) (2) (3)
«036 .075 2.57
(«185) (.373) (1.20)
-,430 «046 .025
(-2.68) («245) (.123)
.101 .095 .073
(«473) (2.33) (1.66)
«002 «002 «001
(-3.34) (-3. 17) (-2.44)
.984 .842 .631
(5.68) (4.42) (3.14)
1.41 1.24 .920
(5.94) (4.86) (3.41)
1.96 1.76 1.11
(9.23) (7.74) (4.27)
1.78 1.57 .989
(6.79) (5.73) (3.21)
-1.30 -l.34 -l.11
(-7.40) (-7.51) (-5.51)
.237 .268 .233
(1.71) (1.92) (1.56)
«060 «034 «015
(«968) («539) («224)
.010 .008 .004
(.887) (.707) (.277)
.212 .206 .219
(8.62) (8.33) (8.18)
«005 «004 «005
(-4.82) (-4.75) (-5.08)
1.24 1.29 1.54
(7.89) (7.98) (8.87)
.425 .421 .353
(11.77) (11.54) (8.88)

1341

Females
(4) (5) (6)
«276 «271 «234
(-1.74) (-1.71) (-1.39)
«461 .107 .137
(-3.22) (.640) (.773)
.074 .085 .086
(2.34) (2.67) (2.57)
«001 «001 «001
(-2.72) (-3.05) (-2.89)
.965 .537 .064
(6.08) (3.09) (.344)
1.04 .609 «069
(5.81) (3. 10) («325)
1.59 1.16 .445
(8.47) (5.75) (1.89)
1.40 1.07 .122
(5.28) (3.85) (.381)
-l.90 -l.98 -1.73
(-15.35) (-15.76) (-12.65)
.016 .069 «023
(.138) (.598) («189)
«046 «042 «036
(«930) («830) («688)
.022 .022 .022
(2.20) (2.17) (2.04)
.205 .207 .220
(9.43) (9.40) (9.18)
«005 «005 «006
(-6.91) (-6.88) (-7. 14)
1.01 .969 1.20
(6.82) (6.51) (7.33)
.387 .394 .427
(14.10) (14.16) (13.70)

 

 

 

Speaks English well

Does not speak
English well

Does not speak
English

US. citizen by
nationalization

Not a US. citizen

Supervise others

Supervise and set
the pay of others

Daily job tasks:
F ace-to-face

customer contact

Talk w/ customers
on telephone

Reading
paragraphs
Writing paragraphs

Work w/ computer

Arithmetic

Occupation controls:

Industry controls:

N
Pseudo R2

Table 6 (cont’d)

 

 

Males
(1) (2) (3)
-- «691 «577
(-3.92) (-3.11)
-- «789 «658
(-3.48) (-2.76)
-- .387 .989
(.860) (1.86)
-- «216 «430
(«740) (-1.38)
-- « 177 «145
(«940) («723)
-- -- «104
(«640)
-- -- «210
(«921)
-- -- «070
(«500)
-- -- .063
(.363)
-- -- .680
(4.61)
-- -- .177
(1.12)
-- -- .904
(5.66)
-- -- «299
(-2.06)
No No Yes
No No Yes
1,369 1,367 1,354
.202 .210 .252

Females
(4) (5) (6)

-- .089 .133
(.528) (. 176)

-- «706 «246
(-2.97) (.252)

-- -1.04 «700
(-2.82) (.403)

-- «550 «542
(-2.97) (.194)

-- «749 «692
(-3.59) (.219)

-- -- .294
(.137)

-- -- «299
(257)

-- -- «098
(. 137)

-- -- .122
(135)

-- -- .265
(130)

-- -- .150
(.136)

-- -- .478
(.137)

-- -- «256
(.120)

No No Yes

No No Yes
1,815 1,815 1,805
.217 .227 .250

 

Note: Z-statistics of regression estimates are reported in parentheses. The omitted categories are “high school
dropout” for educational attainment, “speaks English very well” for English fluency, and “US citizen by birth” for
citizenship. All regressions include a constant and control for metropolitan area Sample sizes differ between
regressions due to missing values.

135

 

 

Table 7

Ordered Logit Estimates from Job Benefits Models:

The Impact of Job Segregation on the Number of Benefits

 

Black

Hispanic

Coworkers are
Predominpntly:
Black

Hispanic

Asian

Other (or mixed)

N
Pseudo R2

 

 

Males
(1) (2) (3)
«018 .1 18 .272
(«086) (.542) (1.19)
«046 .291 .277
(«252) (1.43) (1.28)
.1 16 .057 .316
(.443) (.215) (1.13)
«735 «621 «317
(-4.15) (-3.39) (-1.64)
-1.61 -1.68 -1.88
(-3.52) (-3.86) (-4.46)
«215 «245 «477
(-1.01) (-1.13) (-1.99)
1,360 1,358 1,346
.208 .215 .259

Females
(4) (5) (6)
«370 «382 «439
(-2.04) (-2. 10) (-2.28)
«406 .069 «041
(-2 .42) (.371) («209)
.299 .3 10 .494
(1.35) (1.38) (2.07)
«1 l l . 142 .420
(«640) (.786) (2.19)
.919 1.00 1 .07
(2.86) (3.07) (3.17)
1.22 1 .23 1 .10
(5 .67) (5 .66) (4.93)
1,802 1,802 1,792
.227 .237 .259

 

Note: Included controls correspond to those listed in Table 6. Z-statistics of regression estimates are
reported in parentheses. Sample sizes differ between regressions due to missing values.

136

Table 8

 

A. Males

Black

Hispanic

Coworkers are
Predominpntly:
Black

Hispanic

Asian

Other (or mixed)

N
Pseudo R2

B. Females
Black
Hispanic
Coworkers are

Predominantly:
Black

Hispanic

Asian

Other (or mixed)

N
Pseudo R2

Retirement
(1) (2)
.106 .133

(.050) (.050)
.065 .086
(.047) (.056)
«069 «036
(.070) (.075)
«128 «102
(.049) (.053)
«314 «287
(.118) (.135)
«018 «065
( .062) (.070)
1423 1407
.332 .397

Retirement
(1) (2)
«008 «025

(.049) (.052)
«056 «017
(.048) (.056)
.017 .036
(.061) (.064)
«021 .055
(.049) (.052)
.173 .222
(.078) (.075)
.091 .043
(.056) (.061)
1884 1874
.375 .413

Essitlsarze
(1) (2)
.055 .101

(.045) (.042)
.015 .167
(.042) (.045)
«018 .041
(.061) (.056)
«082 .039
(.044) (.043)
«418 «479
(.101) (.108)
.055 .069
(.046) (.047)
1436 1420
.247 .327

Sick Leave
(1) (2)
«074 «079

(.046) (.049)
«100 «028
(.044) (.050)
.007 .040
(.054) (.055)
«090 «013
(.045) (.048)
.119 .112
(.070) (.073)
.124 .105
(.044) (.046)
1897 1887
.277 .333

Probit Estimates of the Effect of Job Segregation on Benefits

 

 

Person_al Health
(1) (2)
«098 «057
(.047) (.044)
«070 «016
(.037) (.037)
.003 .034
(.045) (.037)
«089 «015
(.037) (.033)
«297 «306
(.113) (.132)
.022 «012
(.039) (.043)
1436 1420
.300 .373
Personpl Health
(1) (2)
«099 «146
(.049) (.052)
«124 «047
(.046) (.053)
.100 .115
(.048) (.048)
.054 .140
(.042) (.041)
.246 .270
(.045) (.033)
.270 .262
(.026) (.026)
1899 1889
.340 .397

 

Family Health
(1) (2)
«134 «034
(.055) (.05 8)
«080 .001
(.048) (.05 8)
.034 .055
(.064) (.067)
«219 «164
(.048) (.053)
«402 «499
(.095) (.088)
«041 «055
(.058) (.066)
1375 1361
.288 .373
Farnil Health
(1) (2)
«063 «095
(.049) (.051)
«1 19 .0001
(.047) (.055)
.107 .152
(.059) (.063)
.058 .211
(.049) (.052)
.409 .453
(.052) (.049)
.317 .303
(.044) (.050)
1818 1802
.328 .400

 

Note: Table reports dF/dx for a discrete change of dummy variables from O to l, with standard errors in parentheses.
Included controls for model (1) correspond to those listed in Table 6, column (1); while those for model (2) above

correspond to those listed in Table 6, column (3). Sample sizes differ between regression due to missing values.

137

 

 

Table 9
Characteristics of Minorities by Job Segregation

 

 

Blacks Hispanics

 

 

Segzegated Not Seggegated Seggegated Not Segigegated
Residential Segregation

(% own race/ethnicity) .596 .415 .606 .451
Commute time (minutes) 29.15 32.51 25.80 27.26
Face-to-face customer
contact .669 .625 .413 .550
Education:
No high school degree .092 .049 .542 .215
High school/GED .505 .481 .338 .367
Assoc JV oc ./T rade .210 .212 .058 .224
Bachelor’s degree .156 .167 .053 .157
Advanced degree .038 .091 .009 .037
Male .428 .490 .572 .531
Part-time job . 1 71 .202 .1 88 .202
Unionized .182 .290 .120 .252
Ln (firm size) 3.46 4.80 3.06 3.96
Atlanta .376 .260 .002 .038
Boston .078 .151 .027 .097
Los Angeles .545 .589 .971 .865
Indusfl:
Construction .049 .022 .092 .061
Durable manufacturing .039 .091 .160 .059
Non-durable mfg. .029 .055 .138 .091
Transportation .090 .l 12 .050 .103
Wholesale/retail trade .146 .124 .227 .182
F inance/Insurance/
real estate .089 .099 .038 .047
Service .503 .413 .261 .413
Public/self employed/
Other .090 .099 .1 19 .101

138

Occupation:
Managerial
Prof/technical
Sales
Clerical/admin.
Service/labor
Crafts/operative
Self employed/other

Speaks English:
Very well

Well
Not well
Not at all

Citizenship staLu_s_:
By birth in US.
By naturalization
Not a citizen

Supervise others

Supervise and set the
pay of others

Table 9 (cont’d)

 

 

Blacks
Seggegated Not Segzegated
.099 .067
.050 .234
.086 .110
.203 .264
.309 .205
.130 .103
.009 .010
.233 .314
.05 7 .074

Hispanics
Seggegated Not Seggegated
.042 .108
.039 .160
.060 .087
.092 .172
.3 30 .184
.423 .270
.015 .013
.258 .577
.195 .255
.457 .165
.089 .003
.198 .465
.064 .171
.738 .365
.196 .250
.064 .063

 

139

 

Table 10
Determinants of Job Segregation:
Probit Estimates of MaIginal Effects

 

 

Blacks Hispanics
(1) (2) (3) (4) (5)

Residential Segregation .391 .406 .212 .135 .143
(.046) (.047) (.053) (.054) (.059)

Commute Time «001 «002 «002 «002 «002
(.0006) (.0006) (.0007) (.0007) (.0008)

Face-to-face customer contact .017 .010 «071 «029 .028
(.031) (.033) (.030) (.030) (.034)

Male «024 «019 .032 .044 .023
(.030) (.033) (.029) (.029) (.034)

High school/GED «098 «112 « 147 «028 «016
(.061) (.060) (.035) (.038) (.041)

Associates/ Vocational/Trade «089 «086 «445 «301 «251
(.063) (.063) (.050) (.060) (.066)

Bachelor’s degree «090 «107 «378 «245 «190
(.065) (.066) (.059) (.065) (.073)

Advanced degree «204 «172 «435 «328 «178
(.061) (.068) (.109) (.127) (.136)

Part-time job «014 .040 «071 «085 «053
(.041) (.045) (.040) (.040) (.043)

Unionized «066 «056 «100 «084 «064
(.034) (.034) (.042) (.042) (.043)

Ln (firm size) «072 «051 «054 «048 «044
(.007) (.008) (.008) (.008) (.010)

Atlanta .024 .072 «520 «530 «548
(.035) (.036) (.156) (.171) (.173)

Boston «185 «162 «251 «216 «265
(.039) (.039) (.077) (.083) (.088)

Speaks English well -- -- -- «006 «025
(.038) (.041)

Does not speak English well -- -- -- .150 .139
(.038) (.042)

Does not speak English -- -- -- .244 .237
(.030) (.035)

140

US. citizen by nationalization

Not a US. citizen

Supervise others

Supervise and set the pay of
others

Industry controls included:
Occupation controls included:

N
Pseudo R2

Table 10 (cont’d)

 

 

Blacks Hispanics
(1) (2) (3) (4) (5)

-- -- -- «004 «030
(.055) (.060)

-- -- -- .1 14 .072
(.042) (.043)

-- «105 -- -- «045
(034) (043)

-- .0002 -- -- .081
(069) (060)

No Yes No No Yes

No Yes No No Yes
1280 1250 1144 1144 1087
.150 . 167 .202 .246 .279

 

Note: Dependant variable is a dummy equal to one if the individual is black and works with mostly
blacks or Hispanic and works with mostly Hispanics. The omitted categories are “high school dropout”
for educational attainment, “speaks English very well” for English fluency, and “US citizen by birth” for
citizenship. Standard errors of regression estimates are reported in parentheses. For dummy variables,
dF/dx is for a discrete change from 0 to 1.

141

 

Table 11
Log Hourly Wage Regressions:
The Impact of Job Sggregation Controllirg for Family Background

 

 

 

Males Females
(1) (2) (3) (4) (5) (6)
Black «070 «O80 «O47 «079 «088 «069
(.061) (.065) (.054) (.050) (.050) (.045)
Hispanic -. 162 « 159 «124 «068 «002 «075
(.060) (.066) (.05 1) (.076) (.082) (.063)
Coworkers are
Predominantly:
Black «131 «132 «146 «085 «085 «078
(.082) (.084) (.064) (.061) (.061) (.063)
Hispanic «222 «186 «134 «289 «276 «185
(.072) (.073) (.057) (.071) (.073) (.061)
N 1172 1170 1046 1401 1401 1330
R2 .492 .501 .583 .480 .491 .635

 

Note: Included controls correspond to those listed in Table 4, along with father’s education, mother’s
education (in years), and indicators of whether father usually worked and mother usually worked (when
respondent was age sixteen), and whether respondent lived with both parents (most of the time until age
sixteen). Robust standard errors of regression estimates are reported in parentheses. Sample sizes differ
due to missing values. Coefficient estimates on dummy variables for working with mostly Asians and

“other (or mixed)” available upon request.

142

A. Blacks
Coworkers are

Predominpntly:
Black

Hispanic

N
R2

B. Hispanics

Coworkers are

Predominantly:
Black

Hispanic
N
R2
C. Whites
Coworkers are

Predominpntlg
Black

Hispanic

N
R2

Table 12
Log Hourly Wage Regressions:

The Impact of Job Segregation by Demographic Group
Controflinglor Family Background

 

 

 

 

 

 

Males
(1) (2)
«175 «166
(.088) (.064)
«202 «037
(.143) (.091)
292 271
.503 .658
Males
(1) (2)
«394 «333
(. 193) (.11 1)
-. 191 .115
(.081) (.078)
384 354
.437 .637
Males
(1) (2)
«007 «047
(.122) (.122)
«134 «056
(.124) (.090)
496 423
.453 .548

Females
(1) (2)
«088 «100
(.049) (.048)
«035 .009
(.075) (.072)
543 526
.537 .646
Females
(1) (2)
«108 «015
(.128) (. 152)
«128 «131
(.094) (.071)
379 360
.522 .725
Females
(1) (2)
«003 .011
(.174) (.156)
«635 «386
(.125) (.101)
479 444
.415 .611

 

Note: Included controls correspond to those listed in Table 3, along with father’s education, mother’s

education (in years), and indicators of whether father usually worked and mother usually worked (when
respondent was age sixteen), and whether respondent lived with both parents (most of the time until age
sixteen). Robust standard errors of regression estimates are reported in parentheses. Sample sizes differ

due to missing values. Coefficient estimates on dummy variables for working with mostly Asians and
“other (or mixed)” available upon request.

143

Table 13

The Impact of Job Segregation and Residential Segregation on Wages

 

 

 

 

 

 

 

 

by Dempgraphic Group
A. Males Black Hispanic White
Coworkers are (l) (2) (l) (2) (l) (2)
Predominpntly:
Black «199 «149 «345 «330 .042 .055
(.081) (.057) (.158) (.097) (.112) (.117)
Hispanic «178 «068 «180 «095 «087 «018
(.129) (.092) (.075) (.073) (.118) (.084)
Pr0portion of
Neighborhood:
Black «106 «001 «122 «053 «093 «030
(.125) (.095) (.125) (.105) (.217) (.182)
Hispanic «392 «077 «168 «107 «672 «333
(.244) (.170) (.100) (.083) (.254) (.175)
N 447 413 580 537 559 479
R2 .478 .609 .435 .583 .421 .515
B. Females Black Hispanic White
Coworkers are (l) (2) (1) (2) (l) (2)
Predorm'gmtly:
Black «062 «064 «052 «245 .036 .089
(.047) (.047) (.083) (.097) (.132) (.117)
Hispanic .117 .086 «161 « 148 «466 «296
(.179) (.090) (.084) (.058) (.128) (.104)
Proportion of
Neighborhood:
Black «315 «248 .114 .080 «112 .012
(.132) (.084) (.107) (.102) (.225) (.154)
Hispanic «726 «523 .048 .068 «451 «463
(.224) (.155) (.116) (.089) (.224) (.171)
N 842 820 576 544 574 532
R2 .536 .649 .470 .675 .400 .598

 

Note: Included controls correspond to those included in the models depicted in Table 5. Robust standard errors

reported in parentheses. Sample sizes differ between regressions due to missing values.

144

Table 14

The Impact of Job Segregation on Wages by Demographic Group

A. Males

Coworkers are

Predominpntly:
Black

 

Hispanic

Neighborhood at
least 50% minority

Short commute
(<= 10 minutes)

Minority
neighborhood x
short commute

N
R2

B. Females

Coworkers are

Predominpntly:
Black

Hispanic

Neighborhood at
least 50% minority

Short commute
(<= 10 minutes)

Minority
neighborhood x

short cormnute

N
R2

Controlling for Neighborhood Effects

 

 

 

 

Black
(1) (2)
«202 «146
(082) (058)
«164 «064
(134) (092)
«056 «042
(081) (059)
«076 «045
(133) (096)
.137 .071
(167) (125)
447 413
.474 .610
Black
(1) (2)
«074 «074
(.047) (.044)
.039 .030
(145) (080)
«012 «016
(056) (050)
.434 .335
(182) (012)
«434 «341
(181) (112)
842 820
.541 .655

Hispanic White
(1) (2) (1) (2)
«382 «341 .046 .041
(158) (098) (115) (122)
«200 «097 «120 «020
(075) (076) (118) (083)
«O35 «031 «228 «093
(056) (050) (086) (077)
.126 .035 .008 «069
(104) (073) (082) (080)
«125 «042 «064 «133
(119) (099) (149) (169)
580 537 559 479
.437 .582 .410 .514
Hispanic White

(1) (2) (1) (2)
«066 «270 .003 .075
(082) (097) (130) (114)
«156 «153 «488 «320
(089) (058) (125) (105)
.124 .103 «146 «139
(064) (049) (097) (090)
.113 .067 .287 .201
(095) (066) (120) (118)
«108 «053 «093 «029
(138) (090) (052) (046)
576 544 574 532
.478 .680 .400 .593

 

Note: Included controls correspond to those included in the models depicted in Table 5.
reported in parentheses. Sample sizes differ between regressions due to missing values.

145

Robust standard errors

Table 15

Measures of Racial Attitudes in Job Segregation Models:

Panel A

Self Selection

N
Pseudo R2

Panel B

Ease of Getting Along with

(scale 1-7):
Blacks

Hispanics

Whites

N
Pseudo R2

Probit Estimates of Magginal Effects

 

 

 

 

Blacks Hispanics
(1) (2) (3) (4) (5)
.073 .082 .053 .010 .047
(.054) (.05 7) (.040) (.042) (.047)
569 555 665 665 631
.184 .220 .233 .287 .369
Blacks Hispanics
(1) (2) (3) (4) (5)
«005 .001 «026 «009 «017
(.017) (.017) (.011) (.011) (.012)
«026 «031 «030 «036 «023
(.015) (.016) (.012) (.013) (.014)
«012 «006 «005 «003 «001
(.01 1) (.011) (.009) (.009) (.010)
569 555 665 665 631
.194 .227 .250 .300 .375

 

Note: Dependant variable is a dummy equal to one if individual is black and works with mostly blacks or
Hispanic and works with mostly Hispanics. Additional controls correspond to those listed in Table 10.
Standard errors of regression estimates are reported in parentheses. For dummy variables, dF/dx is for a

discrete change from 0 to 1.

146

’ i
innit. .21.... . .....\
.2 n. liflytvluk. 1:36.955
2 1.; it 1. .

.ts...nm..u..u

e. «was.

9 c.

. , M25.
in
WIwWSr

is...

o 2..

.... .
:I-IE. 1
11:47.1

3b.».

211: .—
pry-a o '10..-—.
h: a.

4
.9;
ca .5
e