THE MEASUREMENT OF RACE AND GENDER

WAGE DIFFERENTIALS:
EVIDENCE FROM THE FEDERAL SECTOR

GEORGE J. BORJAS*

Thispaperpresents an empiricalanalysisofwagedifferentials
basedon
race and genderin thefederalbureaucracy. By focusingon thestudyof
interagency variationsin wage differentials,
theauthorshows,first,that
theuse in earningsfunctions ofa simpledummyvariabletoindicaterace
and genderleads to downward-biased estimates ofthestandardizedwage
differential.
Second,acrossfederalagenciesthereis a positivecorrelation
betweenwage differentials based on race and thosebased on gender.
Finally,the low relativewage of black femalesis morea resultof their
genderthanof theirrace. This varietyof empiricalfindingsshowsthe
promiseof futurestudiesthatconcentrate on the interfirm
variancein
employment policiesthataffectwomenand minorities.

mHE analysisof wage differentialsbased into a portion "explained" by observable

on race and gender has evolvedinto a characteristicsand an "unexplained" resi-
voluminousliteraturein the past decade.- dual customarilylabeled discrimination.2
Mostofthisworkhas focusedon measuring Labor economists usually find a sizable
standardizedwage differencesamong the unexplained wage differentialamong the
variousracial groups and the two genders. races and sexesin the data setscurrentlyin
The statisticalprocedure used is a decom- general use.
positionof the observed wage differential This paper presentsan empiricalanalysis
of race and gender differencesin earnings
*The author is Professorof Economics at the in the federal bureaucracy. This "firm"
University
of California,Santa Barbara. His research employedover2.4 millionworkersin 1978,
was funded by a grant from the Employmentand of whom 31.1 percentwere women and 22
TrainingAdministration, U.S. Departmentof Labor. percentwereclassifiedas minorities.3These
'The literatureon wage differentialsby race is
workerswere employed by a myriad of
surveyedby Richard B. Freeman, "Labor Market federalagencies. To the extentthat agen-
Discrimination:Analysis,Findings,and Problems,"in
Michael D. Intriligatorand David A. Kendrick,eds., 2Theproblemsassociatedwiththisinterpretation are
Frontiers of Quantitative Economics (Amsterdam: welldocumentedin the literature.A particularlyclear
North-Holland,1974), pp. 501 - 62; and Ray Marshall, discussionof these problemsis given by Solomon W.
"The Economicsof Racial Discrimination:A Survey," Polachek,"PotentialBiases in MeasuringMale-Female
Journal of Economic Literature, Vol. 12, No. 3 Discrimination,"Journal ofHuman Resources,Vol. 10,
(September 1974), pp. 849 - 71. The literature on No. 2 (Spring 1975), pp. 205 - 29.
genderdifferences is surveyedbyCynthiaB. Lloyd and 3U.S. Office of Personnel Management, Equal
Beth T. Niemi, The Economics of Sex Differentials Employment Opportunity Statistics (Washington,
(New York:Columbia UniversityPress, 1979). D.C.: GPO, November1978), p. 2.

Industrialand Labor RelationsReview,Vol. 37, No. 1 (October 1983). C 1983 by CornellUniversity.

0019-7939/83/3701$01.00

79

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
80 INDUSTRIAL AND LABOR RELATIONS REVIEW

cies varyin theirpersonnelpolicies, a new full-timecivilians working in the United

dimensioncan be added to the empirical States whose records report the key vari-
study of wage differences,namely, the ables under study.
existence of interfirmvariations in the Table 1 presents statisticsdesigned to
hiring and placement of blacks and measure the race and gender wage dif-
women.4 ferentialin both the pooled sample and in
The objectiveofthispaper is to showthat each of the 31 largestfederalagencies. The
the availabilityof interfirmdata on race 31 agencies employ over 95 percent of the
and gender wage differentialsprovides a total civilian work force of the federal
unique opportunityto expand our under- [Link] firstcolumn of the table
standingof the measurementand interpre- indicates the average log of annual (full-
tation of these [Link] particular, time) earnings forwhite males (In rem) in
the empirical work below will analyze the each sample. Since the average annual
relationship among various measures wage forwhitemen is relativelyhigh (e9 9718
ofwage discriminationcurrentlyemployed = $21,414), it should be clear that even
by [Link] analysiswill also focus small percentage wage differentialsac-
on therelationshipamong the "discrimina- tuallyreflectsizable dollar differentials.
tioncoefficients"forthe different race and Initially,a simple regressiontechnique
gender groups. serves to measure the wage differential
betweenwhitemales and each of the other
Measurement of Wage Differentials race and gender groups, holding observa-
ble skills constant. In particular, the
This study analyzes data from the followingearningsfunctionis estimatedfor
Central PersonnelData File (CPDF) com- each federal agency (and for the pooled
piled by the Officeof Personnel Manage- sample):
ment, specificallydata on civilian workers
employed by the federal governmenton (1) In rh = Yhce + fRh + E h,
December 31, 1979. The sample is com- where rh denotes the annual full-time
posed of a one percentrandom sample of earningsofindividualh; Yhis a vectorof h's
personnel records in the eight largest socioeconomiccharacteristics;and Rh is set
federal agencies and of a 10 percent equal to unityifthe individualis a member
random sample for all other agencies.5 of a particularrace-gendergroup and zero
Oversamplingthe smalleragencies ensures ifhe is a whiteman. The initialestimatesof
that even the relatively small federal unexplained wage differentials- as
agencies are representedin the analysis. measured by the coefficient j - simply
Each recordoutlinessuch personal charac- allow membership in a particular race-
teristicsof the individual as education, gender group to shiftthe interceptof the
race, and gender and also notes such equation. Equation 1 will be estimated
employmentcharacteristicsas agency of threetimesin each agencyto make pairwise
employment, annual full-time earnings, comparisonsbetweenwhitemen and each
and years of government service. The of the othergroups.
analysis here is restrictedto permanent, In otherwords,a regressionwill firstbe
calculated for the sample of men, to
4The fact that federal agencies differ in their
personnel policies is discussed and documented in
estimatethe unexplained wage differential
GeorgeJ. Borjas,"Wage Determinationin the Federal between white men and black men.6 The
Bureaucracy:The Role of Constituentsand Bureau- second regressionis run on the sample of
crats,"Journal of Political Economy, Vol. 88, No. 6 whites,thus providingan estimate of the
(December 1980), pp. 1110-47. unexplained wage differential between
5Theeightlargestagenciesare theU.S. Departments
of Defense; Agriculture;Justice; Health, Education whitemen and whitewomen. Finally,the
and Welfare;Transportation;Treasury;and the Postal sample forthe thirdregressionis restricted
Service and the VeteransAdministration.A coding
errorin the creationof the data led to an 11 percent
random sample from the U.S. Department of the Throughout the paper, the sample of blacks
Interior. includesindividualsclassifiedas minorityin theCPDF.

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
 RACE AND GENDER WAGE DIFFERENTIALS 81

am
CO
00
m
0-)
0
a
00 tr.
00 v~
en
00
00 -
-4
-0
00 LC) tr-. C
in 00
a) 00 a) t- t- an O - 0 - am 0 0q

CO - - - a) a) -
0 Co - 0 0n tq t
r- 00 tr0 in 0
t t,01 CO co 0a co 'IO CO

o a) CO -. 'IO 00 LO a - co
t~~~ C~
01 -
0 -
0D
01 a)~~~~~~~~
tr.
01
In 0 0
0 a)
aqOCO 0

~~~~~
r- CO) 00 00
~~~~~~~~~a
to(
C
Cl C C2 Cl 0101 1 0

- I I I I~~~~~~~~~~~~~~~~~~~~~~~Q
co CZI I ItI

~~~4-4~~~~~~~
V CO 01 01 (0
CO~~~~~~~~~~~~~0
CO -
00 (0
co ~- CO
(~~~~~~~~~~~04
' a) 04
0
C0
0
q
- C
-4
a)
CI co
0 C
r-

t CO 10 c 00CO
1 01 0(0 LOl~ M
00 ~ oC ' -q 0~
- C~O CO coa
.
OJJ .0 - 0
I -
(7
-
I'
01
I
I'
-~01 IC -
0 I'- 06
I'
-f
I'
01
I I Ci :6 :

a) 0 t. 01 01 00 (0o t0 00 01 co co
~~~
ul0

t~~~~.000C~~i
0 o t' a)0n110101aqco C0o00 Ca) (m Z1 o o0000 co 00a) LO 00 co
cl C0~ 7 71 lz s 010
000CCl!(: 7!c iO0 -~CO0lt~(1 OCla0
- CO - ~~~~~
~~~~-4
01 0 - I

V ~~~~~~~401~~-v C 00t 'cq0 .-'4 00914 ' 00

C 0)00 t-.C 01 lLo a) tq ra) CO00 OC o401 - tr
.0-a) 0MC)0 Q0 r- 0100 -40 -.400 00 OCO 00 ,1 0a) CO 0010010 C-

00 01 - 0 a) LO 00 in 01 00 1- 00 00 a) t-
-..
t.-
00 01
0
It - co
)
a) 0 01 in 0 00
-
co
(
01
t-
00 In
co in (0o LO -. t-. a) In0 0 tr.

.Sa) ~~~~~~~~~~~
~~~~~
-
-4 1- -4 - -4 - -4 - -4 -4 -4 r-

C)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~V C

4-~~~~~~ ~ ~ ~ ~ ~
~~~~ 4-J 4

?) - ,44CZ
o~~~C
~~ ~ 0 - 4) -~~ c,, H H ~ ~
bk

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
82 INDUSTRIAL AND LABOR RELATIONS REVIEW

- c(e s 00 c's c) (s e t> 0 00 00 t> dq

0~~~~~~~~~~C0a

O.4 r C. Cm - 00

- 0
C0 00 m
an
00
0 0 - C r m
< 0 0 0 C( 0
in C.0 m en in > d m d.
't ca 0 n in -o

M w l 00 t- <D M - CO ~
C(4

c.C0 d C( O
t<D 00
00 c( C( O

9 n in to 00 an <D 00 00 00~O
an
O O-
. in~
an O.4
0 m'
m 00
> an
d
in an
.0 -

0q 10 0 00 00
O O
c(t
dq 00

r- 00 C(0 - - m m 0 D C

CO C4 0 O C
d q cec Ol LO
I I I II I I I I I I C
I I 0
I 0
I I
a0 d t to. t Cq C) CZ

Co) 0aaCC0 (4 d

* * * * * * * * * * * * * * *
SS
X~~~~~~~~~~~~~~~~~~~~~~~~.
cl CD mt 0 CD t- CO C s C) 00O
qeNoo>c
qe-

.t I 00K I e00 I NIL e I > I am

I00)sIsI - I - II I II I I I I_ I
~~~~~~~~~~~~~~~~~~~~~~I
oo 0 0 cZ 0 0C ZO 0 C Co 00
I
LO

C-4
1:

Cl
0
C)

M
I 00
00

-
qO

La-40
O C0

-4 a) C0
000

0O
4

1C
0o 00 00Q cs _s
dq
dq
( _s
00O
00 _s > _s
m
_s
dq
s_s
en_s
O _4 c) _s4 $s-4
sd s>_ sO_

~~ ~~
L~~~C~) 0 C~(:~ai C(6 ( .0 ( .( O. .l . I (C'),~ t-

e ~~~~~~~~I
)<c00
I " I I C I CI OI
I 00
I
(Z d0 LO en
I I
C0
I
c
I I I I
0
O 0 0 0 0 0 0 0 00 0 00en dq O
0

C) O
- s -O e- e0 > ce- -- s o ce m- O
ce > s O s s m

o Lo L 00 to CDo ) 0 ( Z C~)tZ C(a) rC 00 00 0 0 Lor

LO r- 0 - LO C.0 LO co dq co 00 C) 111 r -ICQ

o co co 0CZ L CD

O~~~~~~~~~~a 00 LO 00 c C) 00
4d
4dq
d4dqcee

? cs dq ce m CQ CQ ce ce e oo ce o oo '~~~~~~~~~~~~~
C

S~~ ~.> ~ >- S *- - s

=~~~Oo o 0 o O O m O 0 .
,~~ { U, , _ U.

~~~~~~
.t

: bJD dC
~~~0 0 0 ~~~~~~- 0 CZ

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
 RACE AND GENDER WAGE DIFFERENTIALS 83

to whitemen and black women, yielding the estimated : for black males; 1wf is the
thestandardizedwage differential between statisticfor white females; and fbf is the
thesetwogroups. This pairwiseestimation statisticforblack females. The coefficients
procedure facilitatescomparison between in the pooled sample show that, holding
the empirical resultsin Table 1 and the skills constant, significantwage differen-
literatureon wage differentialsbased on tials by gender and race exist, such that
race and gender. whitemen earn approximately12 percent
The socioeconomicvectorY is composed more than black men and 22 to 27 percent
of education (defined as a vector of 21 more than all females.9
dummyvariables indicating both the ex- The regressiontechnique described by
tent and nature of educational attain- Equation 1 assumes that the differencein
ment7);years of experience in the federal earningsbetweenwhitemen and anyof the
sector; years of experience in the nonfed- other race-gender groups is caused by a
eral sector(definedas age minus education constantshiftin the level of the earnings
minus experience in the federal sector profile among the groups. It is likely,
minus 6); quadratics of and an interaction however,that race and gender differences
between the two experience variables; in thepayoffto such variables as education
regionof employment;statusas a veteran; and experiencealso determinetheextentof
and status as physically handicapped.8 race and gender wage [Link]
Clearly,the variable for work experience methodologyusuallyemployedto allow for
previousto federalemploymentis likelyto thispossibilityestimatesan earningsfunc-
be measured with error in the case of tionwithineach of the race-gendergroups.
women. The CPDF provides no informa- These regressionsare used to decompose
tion, however,on the workactivitiesof the the observed wage differentialinto ex-
federalworkforcebeforeenteringthe civil plained and unexplained parts. Formally,
[Link] potentialbiases thatarisefrom supposewe believethatthewhite-maleand
measuring gender wage differentialswith black-male wage structuresare different.
thisexperiencevariable willbe discussedin Since the regression equation passes
detail below. throughthe mean, we can write:
Columns 2 through 4 in Table 1 give
estimatesforf bm )+, (2) In rm =Ywm&wm
and 4bf, wherefbmis
and ln rbm= Ybm& bmi
7The definitionof the dummy variables in the
education vector is as follows: (1) some elementary wherethebars denotesample averages,and
school but did not complete; (2) elementaryschool the "^" denotes the estimated (vector of)
completedbut no highschool; (3) some highschoolbut The observedpercentagewage
coefficients.
did not graduate; (4) high school graduate or
certificateof equivalency; (5) terminaloccupational
differential
is givenby:
programbut did not complete; (6) terminaloccupa-
tional programcompleted; (7) some college but less (3) In rb -'In rm = Ybmbm Ywmwm
than one year; (8) one yearof college; (9) twoyearsof
college; (10) associate's degree; (11) three years of Adding and subtractinga term Ybm6awto
college; (12) four years of college; (13) bachelor's the right-handside of Equation 3 yields:
degree; (14) post-bachelor'swork but no additional
higherdegrees; (15) firstprofessionaldegree, such as - in rwm=
dentistry,law, or medicine; (16) a degree beyond the
(4) In rbm (Ybm
-
YWM) &wm
firstprofessionaldegree but no additional higher + Ybm (&bm -wm)
degree; (17) master'sdegree; (18) post-master'sbut no
additional higherdegree; (19) sixth-yeardegree, such Equation 4 decomposes the observed
as an advanced certificatein education, an advanced average wage differentialinto two parts:
master'sin education, or a certificateof advanced one the resultof differencesin characteris-
graduatestudy;(20) post-sixthyear but no additional
higher degree; (21) doctorate degree, including a
doctorof education or doctorofjuridical science, and [Link],Equal Pay in thePublic Sector:
Ph.D.; or (22) post doctoralwork. Fact or Fantasy(Princeton,N.J.: PrincetonUniversity
8See Borjas, "Wage Determination,"fora discussion Press,1977) fora comprehensivediscussionofhowrace
of theeffectof thesevariableson the wage structurein and gender wage differentialsin the federal sector
the federalbureaucracy. compare withthoseobservedin the privatesector.

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
84 INDUSTRIAL AND LABOR RELATIONS REVIEW

tics,(Ybm - wm) &wm, and one the resultof (the Postal Service)to about 58 percent(the
differencesin structure,Ybm(&bma- Departmentof Labor and the Government
The latter part measures the additional PrintingOffice).
earningstheaverageblack man would earn These resultsclearlyindicate that inter-
if he were "treated like" a white man. agency differencesin personnel policies
Define:0 play an importantrole in the hiring and
placement of blacks and women in the
(5) A1= &wm) j = (bm, wf, bJ).
-

[Link] understandingof
Columns 5 through7 in Table 1 give the the causes of such differencesrequires
calculated As for 30 of the agencies specifying a behavioralmodel ofthefederal
analyzed here." Again, the pooled-sample [Link] a companion paper, these
estimates reveal that white men earn differences are explained by assumingthat
substantiallymore than the other race- different federalagencies cater to different
gendergroups. constituentsand that these constituents'
Perhaps the most strikingfinding in tastesfordiscriminationvary.'2In the case
Table 1 is the verylarge variance in the of racial discrimination,forexample, it is
unexplained wage differentials(whether assumed that white (black) constituents
measured by fj or A1) across federal preferto "see" whites(blacks) employedby
agencies. For example, the wage gap the [Link] the extentthat
betweenwhitemen and black men ranges theracial compositionofthe constituencies
from approximately zero percent (the offederalagenciesvariesacrossagencies, it
Department of Agriculture,the Environ- is clear that a governmentthat wants to
mental ProtectionAgency,and the Equal ensure political support will rationally
EmploymentOpportunityCommission) to employblacks, or whites,in those agencies
over 30 percent (the Federal Home Loan whereitis politicallyadvantageousto do so.
Bank Board, the Government Printing In fact, by identifyingthe agency's
Office,the InternationalCommunications constituencyas the set of employees of
Agency, and the Interstate Commerce industriesaffectedby the agency'sredistri-
Commission). Similarly,the standardized bution policies, the empirical analysis in
log wage differentialbetween white men the companion paper revealed that the
and white women (as measured by AW) demand functionforblacks and women in
ranges from about 3 percent (the Postal federalagencies depends on such factorsas
Service) to 52 percent (the Government the percentage of blacks in the agency's
PrintingOffice). Finally,the standardized the percentage of women in
constituency,
log wage differentialbetween white men
the agency'sconstituency, and the degreeto
and black women ranges from 2 percent
which the agency enforces affirmative
actionprogramsin [Link]
findings,therefore,show that the "unex-
'00f course, it is also possible to measure the
unexplained wage differentialby calculating how plained" wage differentials documented in
much lower the earnings of the average white male Table 1 can, in fact, be explained by the
would be if he faced the black wage structure. characteristicsof the political sphere in
Generally,these two methodologies yield different which the agencyoperates.
measures of wage [Link] conceptually
correctmethodologycalculates the earningsof a black
male if he faced the whitemale wage structurethat also be noted that the explanatory power of the
wouldexistin theabsence [Link] is likely, earnings functionssummarized in Table 1 is much
however,that if black males are a numericallysmall higher than those reported in the literature. For
minority, theyhave a minorimpact on the white-male example, the white-male earnings function in the
wage [Link], thesecond termin Equation 4 is pooled sample yieldedan R2 of .588. Generally,the R2
a good approximationof the true unexplained wage in all the regressionswas greater than .5, with the
differential. exception of the Postal Service, which consistently
"The decomposition procedure cannot be con- showedan R2 in the .2 to .3 range.
ducted forthe Equal EmploymentOpportunityCom- '2GeorgeJ. Borjas, "The Politics of Employment
mission,since, of the 257 observationsin thatsample, Discriminationin theFederal Bureaucracy,"Journal of
only33 are ofwhitemales. This led to a veryimprecise Law and Economics, Vol. 25, No. 2 (October 1982),
estimateof the white-malewage [Link] should pp. 271- 99.

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
 RACE AND GENDER WAGE DIFFERENTIALS 85

Table 2. RelationshipAmong Alternative

Measures of Discrimination(N = 30).
(t-ratiosin parentheses)

Independent Variable
Dependent
Variable Constant flbm fWf bf R2

Abm -.0056 1.1635* - - .926

(- .56) (18.72)
AWf -.0313* - 1.0592* - .917
(-2.08) (17.61)
Abf .0052 - - 1.0599* .904
(.23) (16.22)

*Significantat the .05 level in a two-tailedtest.

The Relationship Among The regressionsreveal two important

Discrimination Coefficients [Link],the simple dummy-variable
Although a strongbehavioral assump- coefficients(B3) systematicallyunderesti-
tion can offer important insights into mate the unexplained wage differential
federal personnel policies, the following measured by Ai. This fact is indicated by
analysisshows that even in the absence of the above unityestimatesof 01 and by the
behavioralassumptions,theresultsin Table significantly negativeconstanttermin the
1 suggesta number of empirical regulari- white-femaleregression.'3 Thus, the allow-
ties that can expand our understandingof ance of differential returnsto labor market
the discriminationprocess. In particular, characteristics(such as education and
the large variance in the flsand As across experience) by race and gender tends to
federalagencies providesa unique oppor- increase the standardizedrace and gender
wage differentials.4 Second, theregressions
tunityto analyze severalissuesraised in the
measurementof discrimination. show an extremelyhigh correlation be-
Table 1 presents two different tween the two measures of unexplained
methodologiesto estimaterace and gender wage gaps; in fact, the correlationcoeffi-
wage [Link] importantquestion cientbetweenA and I3 exceeds 95 percent
is whetherthereare importantdifferences for all three race-gender differentials.A
in the results obtained under the two crucial resultof the analysis,therefore,is
methods. The empirical literaturehas not that the variationin discriminationcoeffi-
yetansweredthisquestion, since analysisis cientsamong agencies is captured equally
usuallylimitedto estimatingfj and A1fora well by A and
single data set. Table 1 presentsboth of The discriminationmeasuresreportedin
these statisticsfor 30 differentagencies, Table 1 also allow us to investigatewhether
allowing us to ascertain whether the the gender wage differentialis related to
variation in the conceptually "better" the race wage differential measured by the
estimate, A, is systematicallyreflectedin wage gap between white men and black
the simpler statistic,A. This issue can be men. Again, previousempirical studies of
examined by estimatinga regressionof the
form: '31t should be noted, however,that of the three
(6) A = O0+0 13+yj, estimatesof0, onlyone (in the black-maleregression)
is significantly
greaterthan unity.
j= (bm, wf,bj), '4JoanG. Haworth,James Gwartney,and Charles
Haworth obtained a similar finding in "Earnings,
acrossthe 30 federalagencies. The estima- Productivity, and Changes in EmploymentDiscrimina-
ted regressionsforeach of the race-gender tion During the 1960's,"American Economic Review,
groups are presentedin Table 2. Vol. 65, No. 1 (March 1975), pp. 158-68.

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
86 INDUSTRIAL AND LABOR RELATIONS REVIEW

Table 3. RelationshipAmong Gender and Race

DiscriminationCoefficients(N = 30).
(t-ratiosin parentheses)

Independent Variable
Dependent
Variable Constant Abm 4bm AWf Of R2

Awf - .1766* .5496* - - - .209

(-4.48) (2.72)
~Wfiw- .1464* - .5581* - - .181
(-4.02) (2.48)
Abf -.2134* .7887* - - - .352
(- 5.41) (3.90)
fbf -.2177* - .8137* - - .319
(-5.97) (3.62)
Abf -.0546* .2943* - .8997* - .875
(-2.35) (2.89) (10.60)
fbf -.0976* - .3562* - .8198* .775
(- 3.65) (2.45) (7.41)

*Significantat the .05 level in a two-tailedtest.

race and gender discrimination cannot and significantcoefficientsof Abbmand f~bm

answer such questions, because they esti- reveal that agencies that have larger
mate thesestatisticsforonlyone data set. If standardized wage differentialsbased on
we view the federal governmentas com- race will have larger standardized wage
posed of agencies with differentpersonnel differentials based on gender. In fact, the
policies, we can analyze whetheragencies size of the coefficientsyl and y' indicates
that have large race wage differentialsalso that a 10 percentfall in the black relative
tend to have large gender wage differen- wage decreases the white-femalerelative
tials. Formally,thisanalysisis based on the wage by 5.5 percentand the black-female
regression:5 relativewage by 8 percent. Thus, Table 3
revealstheinteresting findingthatthereis a
(7) 'A 70 + l1Abm + V, (i = wf, bI strong positive correlation between race
and = ^'O + ^Y'Y bm+ V', ( = Wf, bI) and gender wage differentials.
The estimatedcoefficientsare presented This analysiscan be extendedbyaskinga
in rows1 through4 of Table 3. The results completelyunexplored question: to what
are [Link] negative and significant extent is the low relative wage of black
constant terms indicate that the unex- womendue to theirrace, to theirsex, or to
plained gender wage differentialwould be both? This question can be studied by
between15 and 20 percent(relativeto white estimatingregressionsof the form:
men) even in agencies whereno race wage = + XAbm + X2'Awf + 1U
differentialexists. Moreover,the positive (8) AbfJXO
f3bf = X)O + Xl O3bm+ X'2 + IL '

'5Equation7 is notderivedfromanystructuralmodel The estimatedregressionsare presentedin

of discriminatory behavior. Rather, its purpose is to rows5 through6 of Table 3, and theyalso
measure the correlationamong the various race and
genderwage differentials. Moreover,since the regres- yield interesting results. The relative
sorsand the dependentvariables in that equation are magnitudesof X2 and 21(or X 2 and X 1)
themselvesestimatedwith error,it is likely that the indicate thatthefactthatblack womenare
estimatorof the coefficientvector y is biased. The women is 2.3 to 3 times more important
directionof the bias cannot be ascertained,however, than thefacttheyare black in causing their
unlessassumptionsare made about the magnitudesof
y1and 'y'.
low relativewage. A major findingof this

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
 RACE AND GENDER WAGE DIFFERENTIALS 87

analysis,therefore,is the conclusion that resultingeither from employer behavior

genderis much more importantthan race (such as discrimination)or fromtechnolo-
in determiningthewage gap betweenwhite gical factorsthat differentiateblacks and
men and black women.'6 women [Link] resultsin
It is worthpausing at thispointto discuss Table 3 thereforeshow that our under-
thesubstantiveimplicationsof the findings standing of wage discriminationcan be
in Table 3. The positive correlation be- greatlyexpanded by the analysis of dif-
tween race and gender wage differentials ferencesin the characteristicsof firmsand
suggests that some set of factors leads of the marketsin which firmsoperate.
agencies systematically to pay lower wages
to bothblacks and women. In the context
of the politicalmodel discussedearlier,this Measurementof Labor-Force Experience
finding implies that common variables
One importantcriticismcan be made of
exist that enhance both blacks' and
the findingsin Tables 1 through 3. In
women's political value in some federal
particular,how robustare the resultswith
agencies and not in others;it also suggests
respectto theselectionand definitionofthe
that an understandingof the evolutionof
independent variables in the earnings
race and gender wage differentialsin the
federalbureaucracycould be obtained by a function?
careful analysis of the political forces For example, the regressionscontrolfor
influencingthe agency. yearsof federalserviceand not foryearsof
More generally,the results in Table 3 employmentin the specificjob held by the
implythat studies of interfirmdifferences individual. Similarly,none of the regres-
in discriminationin the privatesectorcan sionssummarizedin Table 1 controlsforthe
improveour understandingof discrimina- type of work conducted by the worker.
[Link] instance,it is well known Thus, it could be argued that many of the
that the standardizedwage differentials 13 interagency differences documented in
and A can be interpretedin two extreme Table 1 (and analyzedin Tables 2 and 3) are
[Link] eithermeasure truediscrimina- the result of differencesin occupational
tionor theycapture unobservedproductiv- characteristicsby race and gender across
ity differencesamong races and sexes federalagencies.
[Link] positivecorrela- It should be noted, however,that dis-
tion between race and gender wage dif- crimination,if it exists, can work either
ferentialsdocumentedin Table 3 indicates throughwage differentials withinthe same
that iff3and A- measure true discrimina- job or by placing selectedgroupsof people
tion, firmsthat discriminatemore than in particularlylow-payingor high-paying
otherfirmsagainstblacks also discriminate occupations. Controllingforoccupation in
more than others against women. Al- the regressions,therefore,would ignore an
ternatively, if:3 and A measure unobserved important channel in the discrimination
productivitydifferences,the results show process,and the resultingestimateswould
that to some extent the same unobserved not measure the full effectof discrimina-
characteristics are responsibleforboth race torybehavior in the federal sector. More
and genderwage differentials. generally,controllingforvariablesthat can
Note that under either interpretation, be used by the employerto accelerate or
the analysissuggestsa strategyforfurther hamper the progressof some employees
research. In particular,Table 3 provesthe some examples being job placement, years
existence of agency-specificfixed effects on the particular job, and promotion
decisions- biases measurementof the full
'6Note that this finding shows that the weights effectof wage discrimination.
attached to sex discriminationare greater than the A more seriousmeasurementproblem is
weightsattached to race discriminationin the deter- that the labor-forceexperience of women
mination of the black-female wage. A probable
explanation of this resultlies in the fact that gender
maybe measured [Link]
wage differentialsare, on the average, greater than particular,the CPDF providesno informa-
race wage differentialsin the federalbureaucracy. tionon theworkactivitiesofwomen before

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
88 INDUSTRIAL AND LABOR RELATIONS REVIEW

their entryinto the federal civil service. twosetsof estimatesare related by estimat-
Recentworkon the determinantsoffemale ing regressionsof the form:
earningsprovidesstrongevidence that the
(9) Ak = 60 + 6lAk + C
intermittentlabor-force participation of
marriedwomen can explain a substantial (k = wf,bj),
portionof the male/femalewage differen- '
Ok = 0'"+ 61k +
tial.7 AlthoughtheCPDF does not giveany
informationon the marital status or the The regressionsare presentedin Table 5.
experience of women outside the labor Note that the constanttermis not signifi-
force,it does give data on the number of cantly differentfrom zero in any of the
years of their employmentin the federal [Link],the coefficientof Ak
[Link] variable can be extremely or Sk is positive and significantand lies
usefulin calculatingan upper bound to the between .75 and .85. The results thus
bias created by ignoring the intermittent indicate that controllingfor the intermit-
labor-force participation of married tent labor-forceparticipation of married
women. women is responsiblefor,at most, 15 to 25
In particular, the earnings functions percentof the unexplained wage differen-
estimated in Table 1 hold constant a tials documented in Table 1.
measure of "previous"experience (PRE V) This estimate is likely to be an upper
definedas age minuseducation minusyears estimateof the true effectfortwo reasons.
of government service minus six. This First, the assumption that the negative
variable has, as expected, a positiveeffect depreciation effect will cancel out the
on the earnings of federal workers. The positiveeffectof labor-forceexperience is
averagevalue of PREV forwomen is about probably too strong, since there is some
ten [Link],a part of thistime debate about whether nonmarket time
was spent working in other jobs and indeed has a negative effecton current
another part of it was spent outside the earnings.'9 Second, the estimates suffer
labor market. The number of yearsin the from selectivitybias, since only women
labor marketwill have a positiveeffecton currentlyworking are included in the
currentearnings,whereas the nonmarket [Link] thereis a positivecorrela-
years may have a negative effect on tion over time in a woman's probabilityof
earnings, because of the depreciation of participatingin the labor force,such that
market skills. The simplest way to take womencurrently workingare morelikelyto
these facts into account is to assume that have workedin the past. This assumption
thetwoeffectsexactlyoutweigheach other, implies that women who are currently
so that women enter federal employment workingwillhave an above averagenumber
with zero years of labor-forceexperience. of yearsof labor marketexperiencebefore
The wage differentialsobtained for the startingfederalemployment.20
varioussamplesaftersettingPRE V equal to The results in Table 5 provide strong
zero for women are reported in Table 4. evidence, therefore,that the bias intro-
The reestimatedj3kand AO are somewhat duced by the intermittentlabor-force
smaller (in absolute value) than the more participationofmarriedwomen is not very
conventional estimates, Sk and Ak, pre-
sentedin Table 1.18
"9See,for example, Steven H. Sandell and David
In fact,we can analyze preciselyhow the Shapiro, "The Theory of Human Capital and the
Earnings of Women: A Reexamination of the Evi-
dence,"Journalof Human Resources, Vol. 13, No. 1
17Jacob Mincer and Solomon Polachek, "Family (Winter1978), pp.103 - 17.
Investments in Human Capital: Earningsof Women," 20James J. Heckman and Robert J. Willis provide
Journal of Political Economy, Vol. 82, No. 2, Part 2 empirical evidence of the serial correlation in the
(March/April1974), pp.S76-S108. probabilityof labor-forceparticipation in "A Beta-
"8Thisis not surprising,since assumingthat women Logistic Model for the Analysisof Sequential Labor
have no experience is likely to overstatethe actual Force Participationby Married Women,"Journal of
differencesin labor market experience and hence Political Economy, Vol. 85, No. 1 (February 1977),
reduce the unexplained wage differentials. pp.27 - 58.

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
 RACE AND GENDER WAGE DIFFERENTIALS 89

Table 4. Measurementof Gender Wage DifferentialsAssumingNo PreviousExperience.

(t-ratiosin parentheses)
0
Agency
Agency O~~~f ~ f3
~~~~
3bf ~ i~f bf

Pooledt - .1946* - .2487* - .2035 - .2552

(-31.82) (-27.43)
CabinetAgencies
Agriculture -.1583* -.1960* -.1406 -.2032
(-6.06) (-4.46)
Commerce - .1636* - .2559* - .1834 - .2674
(-10.56) (-13.66)
Defense - .2776* -.3349* -.2714 -.3328
(-31.87) (-27.69)
Energy -.2670* -.3165* -.2680 -.3364
(-11.69) (-10.28)
Health, Education, -.1012* -.2006 -.1363 -.2122
and Welfare (-4.14) (-6.62)
Housing and Urban -.2591* -.3790* -.2905 -.4420
Development (-9.75) (-13.56)
Interior -.1543* -.2041* -.1614 -.2194
(-14.58) (-14.73)
Justice - .1723* - .2144* - .2029 - .2223
(-4.85) (-4.93)
Labor - .2177* -.4228* -.2680 -.4702
(-10.05) (-17.25)
State - .1197* -.2307* - .0892 - .1852
(- 3.58) (- 5.80)
Transportation -.4618* -.4989* -.4599 -.5035
(-10.96) (- 7.63)
Treasury -.2338* -.3792* -.4124 -.3288
(-8.92) (-11.31)
IndependentAgencies
Environmental -.0738* -.2123* -.0910 -.2305
ProtectionAgency (-3.04) (-6.53)
Equal Employment -.0330 -.1076 - -
OpportunityCommission (-0.23) (-0.99)
Federal -.1170 -.2759* -.1081 -.2040
CommunicationsCommission (-1.95) (-3.67)
Federal Deposit -.1798* -.2562* -.1257 -.2610
Insurance Corporation (-4.28) (- 3.83)
Federal Home Loan -.2922* -.4431* -.4057 -.5430
Bank Board (-3.03) (-4.30)
Federal Trade .0017 -.1081 -.0349 -.1550
Commission (0.02) (-1.02)
General Accounting -.0736* -.2099* -.0673 -.0867
Office (-2.47) (-4.68)
General Services - .1785* - .2782* - .1807 - .2934
Administration (-9.22) (-13.53)
Government Printing -.3688* -.4683* -.3911 -.4549
Office (- 7.82) (-13.15)
International Communications -.1008 -.3029* -.1331 -.3200
Agency (-1.84) (-5.26)
Interstate Commerce -.0702 -.4355* -.1915 -.3819
Commission (-0.86) (-4.91)

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
90 INDUSTRIAL AND LABOR RELATIONS REVIEW

Table 4 Continued.

Agency Of A4 A

National Aeronautics -.2409* -.2615* -.2482 -.2636

and Space Administration (-15.98) (-11.35)
National Labor - .1030* -.2436* -.1561 -.2141
RelationsBoard (-2.52) (-4.00)
Nuclear Regulatory -.0836 -.3015* -.1407 -.3227
Commission (-1.23) (-3.17)
Office of Personnel -.1456* -.1878* -.1239 -.2077
Management (-3.42) (-3.78)
Postal Service - .0444* - .0254* - .0464 - .0300
(-4.18) (-2.48)
Securities and - .0047 - .1379 - .0555 - .0703
Exchange Commission (-0.09) (-1.74)
Smithsonian -.1715* -.2677* -.1590 -.2770
Institute (-2.38) (-2.84)
Veterans - .0394 - .0803* - .0535 - .1128
Administration (-1.77) (-3.01)

tThe statisticsforthe pooled sample are estimatedfroma one percentrandom sample of all federalagencies.
*Significantat the .05 level in a two-tailedtest.

(N = 30).
Table 5. RelationshipBetween Ak(Qk) and A~k(03)
(t-ratiosin parentheses)

Independent Variable
Dependent
Variable Constant AWf Abf Owf bf R2

of .0257 .7835* - - - .674

(.85) (7.61)
A1bf .0208 - .8491* - - .861
(.87) (13.17)
!3wf .0129 - - .7843* - .703
(.53) (8.14)
?bf .0025 - - - .8285* .811
(.09) (10.94)

*Significantat the .05 level in a two-tailedtest.

importantin theCPDF data. Moreover,the

Summary
strong positive correlation between race
and gender wage differentialsis not af- The empirical analysis of interagency
fectedby the correctionused in Table 5. In race and gender wage differentialsin the
particular,Table 6 replicatesthe analysisin federal bureaucracy yields several impor-
Table 3 by relating the gender wage tant findings. First, the use of a simple
as measuredbyAQandf3Q
differentials to the dummy variable to indicate race and
wage gap between black men and white gender in the regressions leads to
men. As in Table 3, the most important downward-biased measures of the
result is the strong positive correlation standardized wage [Link],
among the various gender and race wage federal agencies that have larger race
[Link] other words, a common are likelyto have largergender
differentials
factor is partly responsible for the low [Link], the low relativewage
relativewages of both women and blacks. ofblack womenis moretheresultofthefact

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions
 RACE AND GENDER WAGE DIFFERENTIALS 91

Table 6. RelationshipAmong Gender and Race DiscriminationCoefficients,

AssumingNo PreviousExperience (N = 30).
(t-ratiosin parentheses)

Independent Variable
Dependent
Variable Constant A bm I3bm Wf Wf R2

wf - .1147* .4180* - - - .133

(-2.91) (2.07)
f?f -.1135* - .3542 - - .083
(-3.15) (1.59)
AOf - .1616* .6626* - - - .297
(-4.29) (3.44)
gf - .1774* - .6776* - - .261
(-5.08) (3.14)
Agf - .0701* .3293* - .7972* - .786
(-2.91) (2.83) (7.86)
gf - .0943* - .4181* - .7325* .683
(-3.47) (2.78) (5.99)

*Significantat the .05 level in a two-tailedtest.

they are women than the fact they are line of researchin discriminationanalysis:
black. Finally,the intermittent
labor-force the identification,at the firm level, of
participationofwomenis responsibleforno variables that account for interfirmdif-
more than 25 percent of the unexplained ferencesin employmentpolicies regarding
gender wage differentialin the federal both minoritiesand women.
sector. More generally, the type of research
It is worthwhileto emphasize an impor- conducted in this paper suggests that
tant implication of these findings for similar studies of interfirmdifferencesin
understandingfederal personnel policies therelativewages of the races and the sexes
and [Link] particular,as with any in the privatesectormay lead to extremely
empirical study of race and gender wage useful [Link] particular,if interfirm
differentials,there is no unambiguous, variation in these variables could be
ironclad proof that discriminationexists, explained by such factors as the racial
since, of course, wage differentialscan (gender) composition of the firm's cus-
always be dismissed as resulting from tomers,the degree of physical proximity
unobservedskilldifferentials.
Nonetheless, between the workersand the customers,
the empirical findingof a strongpositive and the race or sex of the firm'sownership
correlationbetweenrace and gender wage or management, the empirical analysisof
differentials
acrossfederalagencies implies wage discriminationwill have advanced
the existence of a common set of factors substantiallyfromits basically descriptive
responsible for the wage "discrimination state to a study of the determinantsof
coefficients."
That findingopens up a new discriminatory behavior.

This content downloaded from [Link] on Sun, 23 Feb 2014 11:28:17 AM

All use subject to JSTOR Terms and Conditions