On the blurring of the color line: Wages and employment
for Black males of dierent skin tones∗
Daniel Kreisman
Gerald R. Ford School of Public Policy
University of Michigan
Marcos A. Rangel
Department of Economics, USP
PRC-NORC, University of Chicago
and
PLAS, Princeton University
October, 2013
Abstract
We evaluate the role skin color plays in earnings and employment for early career Black
males in the NLSY97. By applying a novel, scaled measure of skin tone to a nationally representative sample, and by estimating the evolution labor market dierentials over time, we
bridge a burgeoning literature on skin color with more established literatures on Black-White
wage dierentials and labor market discrimination. Our results suggest that a blurring of the
color line elicits subtle yet meaningful variation in the evolution of inter and intra racial wage
gaps over time.
This research project and the data collection attached to it would not have been possible without the support
from Dan Black and Robert T. Michael. We are also grateful to a number of the University of Chicago's administrative
sta who were kind enough to share their views on the role of complexion within the African American community
in South-Chicago. We wish to thank Cynthia Cook-Conley in particular. Thanks are due to Dan Black, Kerwin
Charles, William Darity Jr., Jerey Grogger, Ricardo Madeira, Ann Morning and Heleno Pioner for comments on
earlier drafts. Comments from participants of the Population Association of America Meeting (2011), and workshops
at the Harris School and the University of Sao Paulo are gratefully acknowledged. The usual caveat applies. Contact:
[email protected]; [email protected].
∗
1
1
Introduction
Nearly a half century since the passage of Civil Rights legislation in America it remains
clear that Black workers earn less on average than their White peers with similar observable
skills, education, background characteristics and experience. While numerous contributions
to the economics literature benchmark these labor market differentials across race, significantly fewer studies address disparities that emerge within the African American community
itself.1 Some recent evidence implies associations between perceived “Whiteness” for Black
Americans and labor market outcomes, including racially distinct first names (Bertrand &
Mullainathan, 2004; Fryer & Levitt, 2004), facial features (Belleti et al., 2008), speech patterns (Grogger, 2011) or simply “acting White” (Austen-Smith & Fryer Jr., 2005; Fryer &
Torelli, 2009). We add to this literature by estimating the relationship between labor market
outcomes and the most salient indicator of race in the US: skin tone.
To accomplish this, we employ novel data on skin tone collected as part of this project in
the 2008 wave of the National Longitudinal Study of Youth (NLSY97). While this is not the
first study to estimate the relationship between skin color and labor market outcomes (see
Hersch, 2006, 2008; Hughes & Hertel 1990; Goldsmith et al., 2007) we present three significant
advantages over previous efforts: first, we make use of a standardized visual measure of skin
tone used by all interviewers; second, we apply this measure to a nationally representative
sample of early career Black and White males with detailed records of skill accumulation
and labor market outcomes; third, we exploit the panel nature of our data to observe the
evolution of wage gaps over the early careers of these respondents.
We begin, following a tradition initiated with Neal and Johnson (1996), by including
measures of skills accumulated before individuals become active in the labor market and observing the residual differences in employment and wages that emerge during the interaction
between suppliers and demanders of labor. We build on these results, following Oettinger
1
See Carneiro, Heckman & Masterov, 2005; Lang & Lehmann, 2012; Neal & Johnson, 1996; O’Neil, 1990;
Ritter & Taylor, 2011.
2
(1996) and Altonji and Pierret (2001), by interacting race and skin color with measures of
labor market experience, allowing us to view the evolution of wage disparities as workers
accumulate skill and match quality is revealed.
We find that while the lightest third of the Black complexion distribution earn more
than the median group, there is no skin color differential between the median and darkest
terciles, suggesting a positive, non-linear relationship between skin color and wages for Black
males. While controlling for a host of background characteristics, childhood circumstances,
education and a measure of skill reduces the Black-White gap by nearly half, these controls
have smaller effects on the intra-racial light-dark gap. Turning our analysis to employment,
we find virtually no difference in the share of weeks employed across skin color, despite
significant differences across race. Lastly, longitudinal exercises indicate that the skin color
gap widens as respondents accrue labor market experience. These results suggest that viewing
racial disparities in stark Black and White ignores subtitle differences that emerge within
race from what Alba (2009) calls a blurring of the color line.
The remainder of the paper is organized as follows. Section 2 briefly summarizes theories
of labor market discrimination and previous research on skin color. Section 3 describes the
NLSY97 sample, our measure of skin color, and the relationship between skin color ratings
and interviewer characteristics. Section 4 describes our empirical strategy and relates it to
previous work. In Section 5 we discuss results and assess the robustness of our findings.
Lastly, conclusions and suggestions for future research are laid out in Section 6.
2
Background
2.1
Theories of Black-White labor market differentials
Two prominent economic theories have been advanced for the existence of racial gaps in
wage offers.2 The first, termed taste based prejudice, is attributable to Becker (1971). In this
2
see Lang & Lehmann, 2012 for a full discussion
3
view, prejudiced employers gain disutility from interacting with minority workers and thus
offer lower wages for comparable work. In this scenario, prejudiced employers and minority
workers sort away from one another in the labor market. Thus, wage differentials between
minority and non-minority workers of the same skill and experience arise when minority
workers are unable to find employment with non-prejudiced employers. The likelihood of
this occurring is increasing both in the minority share of the labor force, and the share of
employers that are prejudiced. In the case that sorting is not possible, the level of prejudice
of the marginal discriminator determines the wage gap. Charles and Guryan (2008) test
Becker’s theory by combining regional measures of racial attitudes with estimates of the size
of the Black labor force, finding that Black-White wage inequality increases with both Black
labor supply and White racial distaste as the theory predicts.
A second hypothesis suggests that employers are faced with incomplete information about
potential employees. In this case, employers base decisions on the average characteristics of
the group to which potential employees belong. Of course, this reasoning is not restricted to
race and can be applied equally to gender, age, sexual orientation or any other feature. Such
ignorance or limited information generate what is referred to as statistical discrimination in
the literature (Aigner & Cain, 1977; Phelps, 1972). Empirical research on this version is more
robust and largely relies on the analysis of employers’ learning about actual productivity as
workers accumulate experience or tenure in combination with labor market search models
In this spirit, Oettinger (1996) proposes a two-period model where Black and White
workers have identical average productivity, but employers and employees only observe a
noisy ex ante signal of match quality. After the first period, true productivity is realized and
workers either choose to stay, and earn the wage associated with their realized match, or to
leave and draw again from the distribution of employers. The key assumption here, a hallmark
of the statistical discrimination literature, is that the signal of match quality is less precise
for Black workers. Thus, sorting is more difficult and unproductive job changes are more
likely. From this, the author’s model yields four key predictions. First, that the Black-White
4
wage gap will be zero at entry; second, that a gap emerges at high levels of experience; third,
that Blacks have smaller returns to experience; and fourth, that Blacks have larger returns
to tenure.3 Testing these predictions using the NLSY79, the author confirms hypotheses 1-3,
but fails to find evidence of hypothesis 4. We conduct similar tests with respect to both race
and skin color in our empirical exercises below.
Similarly, Altonji and Pierret (2001) argue that observable measures of skill should decrease in importance as employers learn about employees’ productivity over time, while the
importance of productivity measures observable only to the econometrician should increase
over time. Accordingly, if employers use race as a source of information, adding interactions
between experience and unobservable skill measures (such as AFQT) should lead to a more
positive coefficient on the interaction term between race and experience, and should increase
the main effect of race. Conversely, if employers do not, or only partially, use race as a source
of information, then the race gap should be small when experience is zero, should widen over
time if race is negatively correlated with the unobserved productivity measure, and controlling for unobservable productivity should decrease the race gap in experience. The authors
confirm that the importance of productivity measures observable at time of hire decrease
over time, and that the importance of measures unobservable at the time of hire increase
in importance. Apposite to our case, they conclude that there is little evidence that firms
statistically discriminate on the basis of race.
Additionally, several studies look at specific traits associated with race for Blacks and
Whites that employers might use to elicit information about potential productivity. For
example, Grogger (2011) find that Black NLSY97 respondents who were more consistently
identified as Black using a minimal amount of speech by a group of listeners earned less than
those not consistently identified as Black. In an audit study, Bertrand and Mullainathan
(2004) find that fictional resumes with “White” names received nearly fifty percent more
callbacks than those with “Black” sounding names. In response, Fryer and Levitt (2004) find
3
Two other predictions are contingent on assumptions about wage contracts, which are not necessarily
relevant here.
5
that for Blacks born in the 1980’s and 1990’s, distinctively Black names are not negatively
correlated with later life outcomes once childhood circumstances are accounted for.
Lastly, a separate literature estimates labor market differentials between Black and White
workers by observing residual differences in employment or earnings after accounting for measures of education, skill, experience and background characteristics. We draw specific attention to seminal work by Neal and Johnson (1996), who conduct these exercises with respect
to wages, and to Ritter and Taylor (2011), who conduct similar exercises with employment.
While our data do not allow us to clearly distinguish between the theories described above,
we conduct similar empirical exercises allowing us to frame our results within the existing
literature while bridging this work with a burgeoning literature on differentials within race
and across skin color as described below.
2.2
Skin color in previous literature
While few data sets contain information on the skin color of Americans, a small number of
specialized studies have sustained research on the topic with mixed results. The National Survey of Black Americans (NSBA) collected information on three generations of 2,107 adults
between 1978 and 1980. Skin color in the sample was collected from non-standardized interviewer perceptions on a Likert scale with five categories: 1. Very Dark Brown; 2. Dark Brown;
3. Medium Brown; 4. Light Brown/Light Skinned; 5. Very Light Brown/Very Light Skinned.
Using these surveys, Hughes and Hertel (1990) indicate that, conditional on parents’ socioeconomic status, lighter Blacks were more likely to hold more prestigious occupations than
their darker peers. Goldsmith et al. (2007) revisit the same data comparing two definitions
of race - the first coined “the one-drop rule” in which only race matters, and the second
“the rainbow rule” in which complexion matters. They find that even after controlling for
demographic characteristics, education and workplace characteristics, the average African
American earns over 16 percent less per hour than a lighter complexioned peer.
Similarly, using the Multi-City Study of Urban Inequality (1992-1994), administered to
6
approximately 8,500 households in Atlanta, Boston, Detroit and Los Angeles, also using a
likert scale to classify skin color (light, medium and dark), the same authors find that wages
decline with darkness of complexion. Hersch (2006) also explores these data to discuss differences in labor market outcomes according to skin color, but her findings indicate that
among males, wage gaps are seemingly unrelated to variation in complexion. More recently,
Akee and Yuksel (2010) use the Coronary Artery Risk in Young Adults study (CARDIA),
which is the richest example regarding the objective collection of skin color measures based
on readings from a reflectance spectrometer, finding that the salience of skin tone has decreased for women over time, but not for men. Finally, the New Immigrant Survey (NIS,
2003) collected information on the skin color of recent legal immigrants to the U.S. using
the same eleven-point scale applied to respondents of the NLSY97 who are the focus of
this paper, as described below. Hersch (2008) analyzed these data finding significant labor
market differentials across skin color for immigrants of all races. We note that skin color
for immigrants may play a different role in labor market outcomes than it does for African
Americans.
3
Data: The NLSY97
The National Longitudinal Study of Youth 1997 cohort is administered annually to a nationally representative sample born between 1981 and 1985. Initial rounds were designed to
provide a detailed description of each respondent’s home environment and behaviors with
a strong focus on human capital accumulation, school-to-work transitions and subsequent
labor market outcomes.
In 2006 a proposal for the inclusion of skin color was submitted by one of the authors
to the NLSY’s Principal Investigators and to the Bureau of Labor Statistics (BLS), the
government agency responsible for the NLSY surveys. The scale was adopted for the 2008
wave of data collection (Round 13); field work was facilitated by the fact that NORC also
7
conducted interviews for the New Immigrant Survey, for which the original skin color scale
was drawn five years earlier.
The scale ranges from 0 (lightest) to 10 (darkest), and presents color images of human
hands with identical forms but different skin tonalities; a facsimile version of the scale can be
seen in Figure A.1 in the Appendix.4 Interviewers are instructed to review and/or memorize
the scale pre-interview in order to rate respondents without presenting the card directly. Of
the 3,455 Black and White male respondents in the NLSY survey, 2,837 were interviewed in
2008. 2,417 of these respondents received a skin color rating in 2008; the majority of those
not rated were interviewed by phone. Although 95 non-rated respondents received skin color
ratings in 2009 or 2010, we omit these respondents from our primary analysis as interviewer
effects are arguably different across waves.5 Figure 1 shows a histogram of skin color ratings
for all 2,417 Black and White males who received a skin color rating in 2008. Although the
scale is different, the distribution of skin tone across and within race is strikingly similar to
that found from using a reflectance spectrometer in the CARDIA study.6
[Figure 1 about here.]
3.1
Interviewer characteristics and measured skin color
In general, information in the NLSY is collected in person. NORC interviewers were largely
matched to the demographic majority of the surveyed area in an attempt to maximize
response rates. Regarding racial composition, 63 percent of interviewers are White and 19
percent are Black. Most are female, have some college education and are above fifty years of
age. Column 1 of Table 1 shows means of interviewer characteristics for the 167 interviewers
who completed skin color ratings for Black and White respondents in 2008. Columns 2-4 show
4
See Massey et al. (2003).
We do include them in one of our robustness checks with no change to results.
6
Figure 1, pp.35, in Akee and Yuksel (2010)
5
8
mean interviewer characteristics weighted by the 2,417 NLSY respondents who received skin
color ratings in 2008.
[Table 1 about here.]
Table 2 shows means of skin color ratings by interviewer characteristics for Black and
White male respondents. While differences across interviewer age and gender are trivial, we
find that the average skin color rating of Black respondents interviewed by Black interviewers
(5.9) is more than one unit lower (lighter) than it is for those interviewed by non-Black interviewers (6.7). The former also report skin color with higher variance among Blacks than the
latter. These simple correlations raise an important issue: despite the use of a standardized
skin color measure, skin color gradations may still depend on subjective evaluations correlated with interviewer race or the racial composition of the respondent group interviewers
contact in the field. We address this concern by evaluating the reliability of interviewer skin
color ratings.
[Table 2 about here.]
3.2
Assessing the reliability of interviewer ratings
Ideally we would have several interviewers rate each respondent and conduct inter-rater
reliability tests. In lieu of this, we approximately test the reliability of skin color ratings
using siblings. As a thought experiment, imagine that each respondent has only one twin
sibling. Now imagine that all siblings who live apart are interviewed by different interviewers
while all siblings living together are interviewed by the same interviewer, and that living apart
is an idiosyncratic phenomenon. We could then examine the correlation between ratings by
these different interviewers as if we had two measures for the same individual. That is, we
9
would estimate (γ1 ) in a dyadic linear regression between the skin color of sibling i, as rated
by interviewer r, and the skin color of sibling j, as rated by interviewer s:
colorir = γ0 + γ1 colorjs + ωi
(1)
cov(colorir , colorjs )
plim γˆ1 =
var(colorjs )
(2)
While this quantity would be ideally suited for an evaluation of classical measurement error,
this is unlikely to be our case. If observed skin color is equal to the true color plus some
error, and skin color measures are discrete and bounded from above and below, 0-10 in our
case, then the measurement error is necessarily negatively correlated with the true color
value and should be considered non-classical. Moreover, the truncation of the distribution
of errors generates a non-zero correlation between errors of different raters.7 In other words,
for respondent i rated by interviewer r, we observe: colorir = color∗ + εri , and can rewrite the
slope coefficient from equation (2) above as:
plim γˆ1 =
cov(colori∗ , colorj∗ ) + cov(εri , colorj∗ ) + cov(εsj , colori∗ ) + cov(εri , εsj )
var(colorj∗ ) + 2cov(εsj , colorj∗ ) + var(εsj )
(3)
Assuming identical covariance between the error and the true measure, independent of rater,
Equation 3 simplifies to:
plim γˆ1 =
cov(colori∗ , colorj∗ ) + 2cov(εsj , colorj∗ ) + cov(εri , εsj )
var(colorj∗ ) + 2cov(εsj , colorj∗ ) + var(εsj )
(4)
Equation 4 reveals opposing forces. On one hand, a positive relationship between errors by
different raters, cov(εri , εsj ) > 0, would overestimate the reliability of our measure. On the
other, legitimate variation in phenotypes expected within a given family, as predicted by
the genetics of skin color determination, cov(colori∗ , colorj∗ ) < var(colorj∗ ), would move the
estimated value of the slope parameter away from unity, even if correlated measurement
7
See Black et al. (2000), section 2.2 in particular, for a thorough discussion.
10
errors were not an issue.8 Therefore, we cannot consider the estimated quantity as an exact
indicator of the relevance of the measurement error, but only as a raw approximation of it.
Despite differences between the ideal experiment and the variation available in our data,
we estimate the slope coefficient from a dyadic regression of White, Black or Hispanic pairs
of individuals (N=1,082) and find point estimates between 0.71 and 0.73 that are statistically
significant (and different than one) at conventional levels. These results are robust to taking
into consideration that discreteness in the skin color scale generates a non-trivial amount of
correct classification between siblings simply by chance. In addition, we computed several
traditional measures of reliability using the same sample.9 Despite the absence of consensus
regarding benchmark values, the overall impression is that these numbers indicate a low,
yet not extremely low, reliability of our skin color measure. Finally, we investigate whether
interviewers are affected by relative reference groups when rating skin color by examining
whether the correlation is higher between siblings who co-reside and are interviewed by
the same interviewer than it is for siblings that live apart and who are rated by different
interviewers. Correlations are indeed stronger among the former, but we are still able to
reproduce equivalent reliability measures between those same-rater and different-rater pairs
of siblings once we account for interviewer characteristics using a fixed-effects specification.
Therefore, if this type of subjectivity is a factor in our estimates, it should be eliminated
when interviewer characteristics are controlled for.
3.3
Analysis sample
[Table 3 about here.]
Table 3 describes the sequence of restrictions used in order to reach our working sample.
We restrict our analysis to non-Hispanic Black and White males, avoiding the modeling
8
See Banerjee et al (1999) on inter-rater reliability indicators, and Barsh (2003), Diamond (1994), Parra
et al. (2004) and Strum et al. (1998) on the genetics of skin color.
9
The resulting Cohen κ-statistic is 0.24, Scott’s π-statistic is 0.24, and Holley and Guilford’s G-index is
0.32.
11
of female labor force decisions which are often substantively different for Black and White
women.10 We limit analysis to respondents who have AFQT scores, a measure of highest grade
completed, and who have entered the labor market, defined as the first of four consecutive
semesters of non-enrollment. The employment sample is restricted to respondents with labor
force participation measures, and the wage sample is restricted to non-self employed, nonmilitary, 30-plus hour per week jobs.
We divide Black respondents into terciles of the Black skin color distribution, where
skin tones 1-5 comprising the lightest third, and 8-10 the darkest. Table 4 shows summary
statistics across race and skin tone for Black and White respondents in our sample; results
for the employment sample are similar. Although not statistically different at conventional
levels, Table 4 reveals subtile differences in AFQT, maternal education, region of residence
and family wealth measured in 1997 across skin tone within the Black community.
[Table 4 about here.]
4
Empirical specification
Our interest is in wage and employment differentials that emerge within race and across
skin color that cannot be accounted for by differences in background, experience or skill. We
model this relationship first on the pooled sample and then add interactions with experience
to analyze the evolution of inter- and intra-racial wage gaps as productivity and match
quality are observed. In all specifications, we assume that skin color has no significance for
White respondents in the labor market and accordingly set color equal to zero for Whites.
10
See Neal, 2006.
12
4.1
Cross-sectional analysis
We begin our analysis by pooling observations across waves, estimating versions of:
yitj = α + β ′ f (Blacki , Colori ) + γ1′ X1,it + γ2′ X2,i + δ ′ Sit + τt + ηj + εit
(5)
Where yitj is either the natural log of real wages or the share of weeks employed 30+ hours
in the past interview year for individual (i) in year (t) rated by interviewer (j) in 2008. Sit
includes highest grade completed at time t and AFQT standardized over Black and White
males with skin color ratings. The vector X1,it includes age, MSA status and region at time
t. X2,i includes age at labor market entry and contains time invariant measures (as of 1997)
of childhood circumstances, including: household poverty ratio, if the family ever received
government aid, if respondent lived with both parents at age 6, lived in the South or outside
an MSA at age 12, mother’s highest grade completed, height and weight, and proxies for
behavioral problems constructed using principal component analysis (school suspension, use
of alcohol or drugs, characteristics of peers, and participation in illicit acts).11
In each exercise we present two parameterizations of f (Blacki , Colori ). The first, which
we call the blurred color line, divides the Black population into terciles of the Black skin
color distribution. The second, termed the sharp color line, excludes skin tone and compares
Blacks and Whites in the traditional sense.
Blurred Color Line: f1 = β1 LightBlacki + β2 M edBlacki + β3 DarkBlacki
f˜1 = β˜1 LightBlacki + β2 Blacki + β˜3 DarkBlacki
Sharp Color Line: f2 = βˇ1 Black
f˜1 re-estimates the blurred color line replacing the median tercile with a main effect for
Black, allowing the reader to interpret coefficients on the first and third terciles as relative
11
See the Data Appendix for definitions and a list of all constructs.
13
to the median Black respondent. In this case, β˜1 = β1 − β2 , β˜3 = β3 − β2 , and β2 is identical
in f1 and f˜1 . We also show outcomes from a non-parametric specification with dummy
indicators for each separate skin tone in the Appendix. In our preferred specification, we
include interviewer fixed effects (ηj ) to account for subjective variation in skin color ratings
across interviewers. Otherwise, we include controls for interviewer characteristics listed in
Tables 1 and 2 above. We follow with series of robustness checks to test for sensitivity to
alternative sample definitions.
de
Since employment shares are bounded by 0 ≤ [ dtt ] ≤ 1, we estimate a generalized linear
model (GLM) using a quasi-maximum likelihood estimator based on a logistic transformation, as described in Papke and Wooldridge (1996), allowing us to abstract the share of
weeks employed between interviews from the number of weeks with valid labor market outcomes during that time.12 To facilitate interpretation of these coefficients, marginal effects
are reported for employment.
4.2
Longitudinal analysis
We extend this analysis to investigate how intra-racial wage differentials evolve by including
interactions between skin tone and experience as described in Equation 6 below:
yitj = α + β ′ f (Blacki , Colori ) + γ1′ X1,it + γ2′ X2,i + δ ′ Sit + ρExperienceit
+ ω1′ f (Blacki , Colori )∗Experienceit + ω2′ Sit ∗Experienceit + τt + ηj + εit
(6)
This strategy builds on previous work by Altonji and Pierret (2001). The authors argue that
at time of hire wages will be strongly correlated with observable measures, such as education,
and the relationship between unobservable skill measures, such as AFQT, will be low. Over
time, the importance of the former should decrease and the value of latter should increase,
12
We test several other specifications, including a dummy indicating more than 75% of weeks employed,
the total number of weeks employed, share of weeks unemployed and an alternate estimator accounting for
mass points at 0 and 1 with little substantive change to results. Results are available upon request.
14
evidenced by negative coefficients on interactions between education and experience, and
positive coefficients on interactions between experience and AFQT. They proceed to test
whether employers elicit information from race until true productivity is revealed. In this
case, interactions between Black and experience, or dark skin and experience for intra-racial
comparisons, should be positive despite a negative main effect.
Analogously, these specifications approximate exercises in Oettinger (1996). While Altonji
and Pierret use potential experience to avoid the endogenous nature of tenure or actual
experience, Oettinger specifically compares interactions between race and both experience
and tenure, predicting a higher returns to tenure for Blacks and to experience for Whites,
as described in Section 2 above. Translating this to skin tone differentials, we would expect
little difference in wages at labor market entry and higher returns to tenure for dark skinned,
relative to lighter skinned, Black respondents, and higher returns to experience for light
skinned Blacks, compared with their darker skinned peers if skin color is used as a source
of information by employers. Accordingly, we model two versions of interactions between
skin color and experience. The first interacts race and skin color with potential experience,
defined by the number of years since labor market entry. The second adds to this main effects
and interactions with job-specific tenure, allowing us to interpret results with respect to the
existing literature.
5
5.1
Results
Wages
Our main specification, Table 5, shows the impact of various controls on inter and intra racial
wage gaps, as described in Equation 5 above. We begin with raw differences in Column 1,
controlling only for year effects, and find that accounting for educational attainment and
AFQT in Column 2 reduces the sharp Black-White gap (Panel B) by roughly one-third,
from 21.5 to 13.5 percentage points, and reduces the intra-race light-median gap by one15
quarter, from 8.1 to 6.1 percentage points (Panel C), and has no discernible effect on the
medium-dark gap, which remains statistically indistinguishable from zero. The effects for
each complexion group can be observed in Panel A.
Column 3 replaces Sit (AFQT and education) with a host of background controls accounting for individual characteristics, geography and childhood circumstances. Comparatively,
these controls account for a larger share of the Black-White wage gap, nearly one-half, than
do controls for education and AFQT. The effect of controlling for background characteristics
on the light-dark gap is relatively similar to controlling for only education and pre-market
skills. This implies larger differences in Black-White background characteristics correlated
with wages than skill accumulation, particularly in the case of lighter Blacks. Column 4
includes both sets of controls which explain a larger share of the inter-racial gap in wages.
Intra-racial gaps remain less sensitive to the addition of covariates.
Columns 5 and 6 add interviewer characteristics and interviewer fixed effects respectively.
Our preferred specification, in Column 6, reduces the Black-White wage gap by approximately half, from 21.5% to 11.8%. These estimates are in line with recent literature, highlighting the fact that differences in pre-market skills and background characteristics explain
a significant share of observed wage disparities.13 In contrast with the effect on inter-racial
wage gaps (Panel B), the inclusion of subsequent controls on intra-racial wage gaps (Panel
C) has little impact, leading us to conclude that controls for background, demographics, skill
and education, explain a larger share of the inter-race wage gap (roughly one-half) than they
do the intra-race gap (one-third) in the most conservative estimate.14 Importantly, we are
not in a position to rule out: i) that the accumulation of skill is itself a response to expected
labor market discrimination (see Coate & Loury, 1993); or ii) that our skill measures are
racially/color-biased or incomplete in nature (see Goldsmith et al, 1997; Darity & Mason,
1998).
13
see Black et al. (2008), Carneiro, et al. (2005) and Lang & Manove (2011).
In an earlier version we also explored an alternative estimation based on heteroskedasticity of measurement errors following contributions by Lewbel (2012). The results confirmed findings presented here.
14
16
[Table 5 (Reg 1) here.]
5.2
Employment
In Table 6 we estimate the same set of regressions on the share of weeks employed in the
past year. As above, the inclusion of individual characteristics reduces the Black-White
gap by nearly half, from 17% to 8%, comparable to findings in Ritter and Taylor (2011).
Yet, we find little evidence of employment differentials across skin color. Moreover, the
sequential inclusion of control variables leaves the intra-racial employment gap virtually
unchanged. We conclude that while inter-racial differences in employment correspond with
Black-White differences in wages, the same is not true for differences within race, across
skin color. Therefore, a unifying theory explaining skin color differentials will have to be
distinct, at least in this dimension, from one that applies to racial differences in labor market
outcomes.
[Table 6 (Reg 2) here.]
5.3
Sensitivity analysis
In Table A.1 in the Appendix we replicate our preferred specification, Column 6 of Table 5,
over various sample restrictions. Results are largely robust to these changes with two exceptions. First, restricting the sample to observations after 2006 increases the light-dark wage
gap indicating a widening of the intra-racial gap either in later years or among more experienced workers. Second, in Columns 7–9 we address possible effects of mixed-race ancestry,
defined as an identifiable non-White (for Whites) or non-Black (for Blacks) biological parent.
A recent working paper by Arcidiacono et al. (2012) finds that mixed-race children born to
White mothers are advantaged in several academic and labor market outcomes. In Column
7, which includes a mixed-race dummy and an interaction with Black, the inter-racial gap
17
in wages increases, particularly for lighter Blacks. Excluding respondents with identifiable
mixed-race parentage in Column 8 reduces the intra-racial advantage for the lightest tercile
relative to median Black from 0.072 to 0.058; moreover, although still positive, the intraracial coefficient on Light is no longer statistically different from zero. Nonetheless, it is still
the case that differences are statistically significant between lightest and darkest terciles.
We are reluctant to classify respondents by their parent’s race on two accounts. First,
our sample consists of respondents classified by the NLSY as non-Hispanic Black or White,
which explicitly excludes a small sample of “mixed” individuals. Thus, by our definition here,
several “mixed” respondents were omitted when we restricted to non-Hispanics. Also, since
most of our presumed mixed-race respondents here have an Hispanic parent, the “mixed”
classification should be interpreted with caution. Second, the NLSY was not able to locate
both biological parents for all respondents, raising concerns about selection. If respondents
for whom both parents can be contacted are advantaged over those for whom they cannot,
all else equal, those respondents should be more likely to find a mixed race parent. We in
fact find that having two identifiable biological parents increases the likelihood of finding a
mixed race set for Blacks. Omitting presumed mixed-race respondents does not affect our
longitudinal results, although we do not show all of these robustness checks for brevity.
5.4
Longitudinal analysis
In Table 7 we interact skin tone with both potential experience (Columns 1-2) and tenure
(Columns 3-4) to show the evolution of inter and intra-racial wage gaps, as described in
Equation 6 above. Both potential experience and tenure are mean centered in all specifications such that level differences in wages correspond to those from our non-interacted models.
Interactions between experience and education (HGC, mean centered) and AFQT are also
included to address conditional differences in returns to experience by skin color and race.
Each model also includes our full set of controls.15
15
We model experience linearly in all cases as these data contain only early career experience - mean
potential experience is 3.9 and mean tenure is 2.1 in the wage sample - average marginal effects using higher
18
Results in Table 7 indicate that as workers accrue experience two phenomena occur. First,
the wage gap between Black and White workers widens as potential experience increases.
Second, as workers accrue job specific tenure, the composite Black-White gap narrows. Yet,
a blurring of the color line reveals that these results are driven by a divergence in hourly
wages within the Black population. Column 2 indicates that the entire Black-White gap in
potential experience is driven by the darkest tercile, for whom the Black-White gap widens
by 2 percentage points per year of potential experience, while wages for the lighter twothirds of the color distribution remain roughly constant over time. Columns 3 and 4, which
combine potential experience and tenure, indicate that the lightest Black respondents gain
3.6 percentage points on Whites with each year of tenure, while the darkest skinned Blacks
do not advance significantly relative to Whites as tenure accumulates. Intra-racial differences
in returns to tenure exist, although they are not statistically different from zero.
In Table A.2 in the Appendix we repeat our exercises using a non-parametric specification
of skin color by including dummies for each value of skin tone, with skin tones 1-3 and
9-10 combined for sufficient sample size. Column 1 re-estimates static intra-racial gaps,
confirming larger average differentials for the darkest two terciles compared with the lightest
third. Although Blacks with skin colors 1-3 have worse labor market outcomes, these outliers
account for a small number of observations (47 individuals in total). While some of these
exceedingly light ratings may be attributable to albinism, it is unlikely that most of them
are. Columns 2 and 3 repeat exercises for potential experience and tenure. In both cases it
is clear that the darkest tercile of Blacks earn less as experience accrues compared to their
lighter peers.
[Table 7 (Reg 3) about here.]
order polynomials are similar and are omitted for ease of interpretation and brevity.
19
5.5
Discussion
Static versions of statistical discrimination models seem at odds with our findings. We discover little evidence of convergence in wages across skin tone for Blacks as labor market
experience accrues. In fact, our evidence suggests that both Black-White and dark-light
wage gaps increase over time. These findings add to the evidence of growth in racial differences in wages across the life course presented by Tomaskovic-Devey et al. (2005), indicating
that a widening happens within the African American population as their careers progresses
and improved wage offers are obtained (or not).
Dynamic models of statistical discrimination provide a better explanation for these findings. Our evidence suggests that experience accumulation may generate smaller gains for
darker Blacks if they are representative unproductive job changes. Therefore, if Oettinger’s
(1996) argument can be extended from White-Black to Light-Dark differences in the quality
of prospective job-market match signal, the wage differences we uncover within the Black
population are result of mismatches and job instability. In an additional empirical exercise,
not reported here, using the same specification as the wage regressions above, we find significant differences between dark and light Blacks in the probability of holding a job for one year
or more. Combined with our results showing no differential in employment probabilities, this
indicates that the effects of potential experience accrue differentially for lighter and darker
skinned Blacks, representing differential patterns of job mobility. Thus, despite having similar employment likelihoods in a given year, darker skinned Blacks face a higher rate of job
switches than lighter Blacks, resulting in differential trajectories of wage accumulation. It is
important to keep in mind that these effects may be related to time changes in labor market
churning associated with a boom-bust pattern of the U.S. business cycle during the time
period studied here; thus, we make these claims with caution.
20
6
Conclusion
We present evidence of the role skin color plays in earnings and employment for African
American males based on a novel and representative sample of American youth. By differentiating skin tone within race, we add to a robust literature on racial differentials in labor
market outcomes in the tradition of Neal and Johnson (1996), Ritter and Taylor (2011) and
others. Moreover, by employing a uniform measure of skin tone to estimate skin color differentials in a representative sample, and by estimating the evolution of these differentials
over time, we add to a nascent literature on the relationship between skin color and earnings
building on earlier work by Hersch (2006, 2008), Goldsmith et al. (2007) and others.
While we find no impact of complexion on employment, we show that measured differences
in earnings within race, across skin color cannot be explained away by skills accumulated
before labor market entry. In fact, we show that accounting for pre-market skill accumulation increases the empirical importance of intra-racial differences relative to inter-racial
differentials. In particular, while close to two-thirds of racial wage gaps are explained by
pre-market factors, this is true only for at most one-third of the color gap between light
and medium skinned African Americans. The color gap in wages not explained by skills
brought to the market amounts to 5.4 percentage points (or half the racial gap) based on our
most conservative estimates. Additional analyses comparing the evolution of wage profiles
as respondents accumulate labor market experience are equally revealing. We find that as
Black workers accumulate experience, the gap between light and darker skinned Black respondents widens, while the gap between the lightest skinned Black workers and their White
counterparts remains constant.
In addition, we provide a novel description of the distribution of skin tone among Black
and White American youth and investigate the relationship between skin color ratings and
interviewer and respondent characteristics. We find that Black NLSY respondents were rated
with both lighter skin tones and higher variance by Black interviewers by White interviewers.
21
To account for potential effects from interviewer heterogeneity, we include interviewer fixed
effects in our preferred specification, finding little meaningful effect on results.
Our results are subject to limitations. It is important to consider that when examining
the accumulation of labor market experience we only observe early careers (the average age
among respondents is 28 in 2010). Moreover, although our skin color measure is a significant
improvement on previous work, it is not entirely objective, and cannot be treated as cardinal. We highlight the importance of perception in racial classifications and underscore the
inherent subjectivity of racial classifications both in surveys and in day-to-day interactions.
Finally, we are not in a position to rule out the possibilities that the accumulation of such
skills is in itself a response to expected labor market discrimination; or that our skill measures
are racially/color-biased or incomplete in nature. Nonetheless, by presenting evidence on the
relevance of skin tone in labor market outcomes beyond traditional dichotomous racial classifications, we challenge the notion of race as Black and White as it is traditionally employed
in economics.
22
References
Aigner, D. and Cain, G. G. (1977). Statistical theories of discrimination in labor markets.
Industrial and Labor Relations Review, 30(2).
Akee, R. and Yuskel, M. (2010). Skin tone’s decreasing importance on employment: Evidence
from a longitudinal dataset, 1985-2000. IZA Discussion Paper Number 5120.
Alba, R. (2009). Blurring the Color Line: The New Chance for a More Integrated America.
Boston: Harvard University Press.
Altonji, J. G. and Pierret, C. R. (2001). Employer learning and statistical discrimination.
Quarterly Journal of Economics, 464(2).
Arcidiacono, P., Beauchamp, A., Hull, M., and Sanders, S. (2012). Isolating mechanisms for
the racial divide in education and the labor market: Evidence from interracial families.
Unpublished manuscript, Duke University.
Arrow, K. (1998). What has economics to say about racial discrimination? Journal of
Economic Perspectives, 12(2).
Austen-Smith, D. and RolandFryer Jr. (2005). An economic analysis of ‘acting white’. The
Quarterly Journal of Economcis, May.
Banerjee, M., Capozzoli, M., McSweeney, L., and Sinha, D. (1999). Beyond kappa: A review
of interrater agreement measures. Canadian Journal of Statistics, 27:3–23.
Barsh, G. (2003). What controls variation in human skin color? PLoS Biol, 1(1).
Becker, G. (1971). The Economics of Discrimination, 2nd ed. Chicago: The University of
Chicago Press.
Belleti, N. and Wade, T. (2008). Racial characteristics and female racial attractiveness
perception among united states university students. In Kauffman, K. D., editor, Racism
in the 21st Century. Lansing: Springer.
Bertrand, M. and Mullainathan, S. (2004). Are emily and greg more employable than lakisha
and jamal? a field experiment on labor market discrimination. The American Economic
Review, 94(4).
Black, D., Haviland, A., Sanders, S., and Taylor, L. (2008). Why do minority men earn
less? a study of wage differential among the highly educated. Review of Economics and
Statistics, 88(2):300–313.
Black, D. A. (1995). Discrimination in an economic search model. Journal of Labor Economics, 13(2):309–334.
Black, D. A., Berger, M., and Scott, F. (2000). Bounding parameter estimates with nonclassical measurement error. Journal of the American Statistical Association, 95(451):739–748.
23
Black, D. A., Kolesnikova, N., and Taylor, L. J. (2009). The role of location in evaluating
racial wage disparity. Federal Reserve of St. Louis Working paper 2009-043B.
Carneiro, P., Heckman, J., and Masterov, D. (2005). Labor market discrimination and racial
differences in premarket factors. Journal of Law and Economics, 48(1):1–40.
Charles, K. and Guryan, J. (2008). Prejudice and wages: An empirical assessment of becker’s
the economics of discrimination. Journal of Political Economy, 116(5):773–809.
Coate, S. and Loury, G. (1993). Will affirmative-action policies eliminate negative stereotypes. The American Economic Review, 83(5).
Darity Jr., W. and Mason, P. (1998). Evidence on discrimination in employment: Codes of
color, codes of gender. Journal of Economic Perspectives, 12(2).
Diamond, J. (1994). Race without color. Discover, 15(11).
Faldon, J. (2009). Statistical discrimination and the implications of employer-employee racial
matches. Unpiblished manuscript, Vanderbilt University.
Fryer Jr, R. and Levitt, S. (2004). Causes and consequences of distinctive black names.
Quarterly Journal of Economics, 119(3):767–805.
Fryer Jr, R. and Torelli, P. (2009). An empirical analysis of ’acting white’. NBER Working
Paper Number 11334.
Goldsmith, A., Hamilton, D., and Darity Jr., W. (2007). From dark to light: Skin color and
wages among african-americans. Journal of Human Resources, 42(4):702–731.
Grogger, J. (2011). Speech patterns and racial wage inequality. Journal of Human Resources,
46(1):1–25.
Gyimah-Brempong, K. and Price, G. (2006). Crime and punishment: And skin hue too? The
American Economic Review, 96(2):246–250.
Helmreich, J. (2010). Richard alba, blurring the color line. Society, 47(4).
Herrig, C., Keith, V. M., and Horton, H. D. (2004). Skin Deep: How Race and Complexion
Matter in the Color-Blind Era. University of Illinois Press: Chicago, IL.
Hersch, J. (2006). Effects among african americans: perceptions and reality. American
Economic Review Papers and Proceedings, 96(2):251–255.
Hersch, J. (2008). Profiling the new immigrant worker: The effects of skin color and height.
Journal of Labor Economics, 26(2).
Hughes, M. and Hertel, B. R. (1990). The significance of color remains: A study of life
chances, mate selection, and ethnic consciousness among black americans. Social Forces,
64(4).
24
Krieger, N., Sidney, S., and Coakley, E. (1998). Racial discrimination and skin color in the
cardia study: Implications for public health research. American Journal of Public Health,
88(9).
Lang, K. and Lehmann, J. K. (Forthcoming). Racial discrimination in the labor market:
Theory and empirics. Journal of Economic Literature.
Lang, K. and Manove, M. (2011). Education and labor market discrimination. American
Economic Review, 100.
Massey, D. and Martin, J. A. (2003). The nis skin color scale. NIS.
Neal, D. (2006). Why has black-white skill convergence stopped? In Hanushek, E. and Welch,
F., editors, Handbook of the economics of education, Vol. 1, chapter 9. Amsterdam: Elseiver
Science.
Neal, D. and Johnson, W. R. (1996). The role of premarket factors in black-white wage
differences. Journal of Political Economy, 104(5):869–95.
Oettinger, G. (1996). Statistical discrimination and the early career evolution of the blackwhite wage gap. Journal of Labor Economics, 14(1):52–78.
O’Neil, G. (1990). The role of human capital in earnings differences between black and white
men. Journal of Economic Perspectives, 4(Fall):25–45.
Papke, L. and Wooldridge, J. (1996). Panel data methods for fractional response variables
with an application to test pass rates. Journal of Econometrics, 145(1).
Parra, E. J., Kittles, R. A., and Shriver, M. D. (2004). Implications of correlations betweeen
skin color and genetic ancestry for biomedical reserach. Nature Genetics Supplement,
36(11).
Phelps, E. (1972). The statistical theory of racism and sexism. American Economic Review,
62(4):659–61.
Ritter, J. and Taylor, L. J. (2011). Racial disparity in unemployment. Review of economics
and statistics, 93(1).
Strum, R., Box, N., and Ramsay, M. (1998). Human pigmentation genetics: the difference
is only skin deep. BioEssay, 20:71–721.
Telles, E. and Lim, N. (1998). Does who classify race matter? self vs. social classification of
race and racial income inequality in brazil. Demography, 5(4).
Tomaskovic-Devey, D., Thomas, M., and Johnson, K. (2005). Race and the accumulation
of human capital across the career: A theoretical model and fixed effects application.
American Journal of Sociology, 111.
25
Table 1: Interviewer characteristics, means weighted by interviewers and respondents.
Weighted by
Interviewers
Interviewer gender
Male
Female
Missing gender
Interviewer race
White
Black
Other
Missing race
Interviewer age
Over 50
Under 50
Missing age
Interviewer education
HS or less
Some college
College
More than college
Missing education
Respondents per interviewer
N
Weighted by
Respondents
Black
White
All
0.150
0.796
0.054
0.149
0.827
0.024
0.150
0.819
0.032
0.149
0.822
0.029
0.629
0.192
0.120
0.060
0.547
0.332
0.091
0.030
0.746
0.131
0.087
0.035
0.677
0.201
0.089
0.034
0.617
0.311
0.072
0.619
0.349
0.032
0.656
0.299
0.044
0.643
0.317
0.040
0.144
0.305
0.216
0.275
0.060
39.07
0.096
0.365
0.287
0.222
0.031
56.54
0.178
0.327
0.167
0.292
0.036
49.91
0.150
0.340
0.209
0.268
0.034
52.20
167
834
1,583
2,417
Notes: Column 1 shows interviewer characteristics for the 167 interviewers who interviewed 834 Black and 1,583 White male respondents
in 2008. Columns 2-4 show mean interviewer characteristics by respondent race; i.e. interviewer means weighted by respondents.
26
Table 2: Skin color ratings for Black and White males
by interviewer characteristics.
Interviewer gender
Male
Female
Missing gender
Interviewer race
White
Black
Other
Missing race
Interviewer age
Over 50
Under 50
Missing age
Interviewer education
HS or less
Some college
College
More than college
Missing education
Observations
Skin Color, White
mean
sd
Skin Color, Black
mean
sd
1.979
1.772
1.680
1.609
0.867
1.115
6.403
6.491
6.000
1.987
1.911
2.271
1.849
1.870
1.304
1.732
0.887
1.673
0.575
1.070
6.781
5.935
6.618
6.160
1.749
2.065
1.993
2.154
1.807
1.791
1.734
0.861
1.317
0.996
6.479
6.481
5.909
1.898
1.975
2.202
2.004
1.774
1.826
1.714
1.614
0.846
0.945
1.515
0.810
1.098
7.100
6.487
6.490
6.124
6.500
1.769
1.978
1.946
1.800
2.195
1,583
834
Notes: Statistics are means and standard deviations of skin
tones by interviewer characteristics (rows) and respondent race
(columns). Sample is 2,417 Black and White Males with skin color
ratings in 2008.
27
Table 3: Sample restrictions.
All Respondents
Black and White Males
With Color Rating in 2008
With AFQT
Entered Labor Market
In Employment Sample
In Wage Sample
White
Black
Individuals
Observations
8,984
3,455
2,417
2,007
1,937
1,855
1,694
1,148
546
125,776
48,370
33,838
28,098
27,118
12,953
8,436
5,715
2,721
Notes: Entered labor market indicates if respondent was
ever observed after two consecutive non-enrolled terms.
Employment sample refers to respondents with nonmissing HGC and employment history. Wage sample
refers to respondents with wages in 30+ hour per week
jobs.
28
Table 4: Summary statistics - means and standard deviations by
race and skin color for wage regression sample.
AFQT (z-score)
HGC Ever
Entry Age
South
Non-MSA
Poverty ratio (1997)
Mom HGC < HS
Mom HGC > HS
Potential experience
Tenure
Wage observations
Observations
White
Black, 1-5
Black, 6-7
Black, 8-10
0.30
(0.94)
13.53
(2.72)
21.67
(2.63)
0.31
(0.44)
0.12
(0.25)
348.7
(275.8)
0.33
(0.47)
0.23
(0.42)
5.64
(1.86)
2.19
(1.97)
4.97
(2.55)
-0.49
(0.84)
12.50
(2.51)
21.23
(2.42)
0.54
(0.48)
0.11
(0.22)
208.3
(223.2)
0.55
(0.50)
0.14
(0.35)
6.43
(1.76)
1.79
(1.92)
4.80
(2.60)
-0.75
(0.73)
12.33
(2.40)
21.00
(2.33)
0.65
(0.47)
0.10
(0.22)
209.9
(226.6)
0.59
(0.49)
0.13
(0.34)
6.56
(1.73)
1.69
(1.41)
5.15
(2.43)
-0.72
(0.71)
11.88
(2.21)
20.84
(2.20)
0.75
(0.42)
0.11
(0.24)
180.0
(161.6)
0.58
(0.50)
0.09
(0.29)
6.88
(1.76)
1.57
(1.42)
4.95
(2.31)
1,148
161
207
178
Notes: AFQT is a z-score normalized to the working color sample of 2,417.
HGC is respondents’ highest grade completed. Entry Age is age at which
respondents first enter the labor market, defined by two consecutive nonenrolled years. South and Non-MSA are averaged over individuals. Poverty
ratio is for the family, measured in 1997. Wage observations counts the
number of observations respondents contribute to the wage regressions.
29
Table 5: Conditional differences in log hourly wages by race and skin color
Panel A: Blurred color line, compared with White
Black, light
Black, medium
Black, dark
(1)
(2)
(3)
(4)
(5)
(6)
-0.156***
(0.032)
-0.238***
(0.026)
-0.238***
(0.029)
-0.101***
(0.030)
-0.162***
(0.025)
-0.137***
(0.030)
0.058***
(0.007)
0.040***
(0.011)
-0.068**
(0.031)
-0.137***
(0.026)
-0.134***
(0.031)
-0.051*
(0.031)
-0.106***
(0.026)
-0.090***
(0.030)
0.050***
(0.007)
0.032***
(0.011)
-0.054*
(0.031)
-0.108***
(0.027)
-0.091***
(0.030)
0.050***
(0.007)
0.032***
(0.011)
-0.069**
(0.033)
-0.142***
(0.028)
-0.142***
(0.030)
0.045***
(0.006)
0.031***
(0.011)
(1)
(2)
(3)
(4)
(5)
(6)
-0.215***
(0.019)
-0.135***
(0.020)
-0.115***
(0.021)
-0.084***
(0.022)
-0.086***
(0.022)
-0.118***
(0.023)
HGC
AFQT
Panel B: Sharp color line
Black
Panel C: Blurred color line, compared with Black (main effect not shown)
Black, light
Black, dark
(1)
(2)
(3)
(4)
(5)
(6)
0.081**
(0.038)
-0.000
(0.035)
0.061*
(0.033)
0.025
(0.033)
0.069**
(0.034)
0.002
(0.033)
0.054*
(0.032)
0.015
(0.032)
0.054*
(0.032)
0.016
(0.032)
0.072**
(0.034)
-0.001
(0.030)
X
X
X
X
Both
X
X
X
X
X
X
Both
X
X
X
X
X
Both
X
8,436
1,694
0.267
8,436
1,694
0.267
8,436
1,694
0.338
Controls, All Panels
Interviewer FE
Interviewer Controls
Region & MSA
Background
Behaviors
Height & Weight
Age, Age at Entry
Year FE
N
Individuals
R2
X
X
Both
X
X
X
X
X
Age
X
8,436
1,694
0.134
8,436
1,694
0.236
8,436
1,694
0.225
Dependent variable is natural log of real hourly wage. Sample is respondents who have entered
the labor market (not enrolled and have at least two semesters out of school), with non-missing
HGC and AFQT, and valid wages (30+ hrs./week). Coefficients in Panels A and B are relative
to White, in Panel C coefficients are relative to medium Black. Controls are described in the
Data Appendix.
Standard errors clustered on individuals in parentheses.
∗p < 0.10, ∗∗ p < 0.05, ∗∗∗p < 0.01.
30
Table 6: Conditional Differences in Share of Weeks Employed in the Past Interview Year for Men by Race and Skin Tone (Coefficients are Marginal Effects).
Panel A: Blurred color line, compared to White
Black, light
Black, medium
Black, dark
(1)
(2)
(3)
(4)
(5)
(6)
-0.166***
(0.021)
-0.154***
(0.018)
-0.189***
(0.020)
-0.117***
(0.020)
-0.097***
(0.018)
-0.122***
(0.020)
0.028***
(0.006)
0.036***
(0.009)
-0.127***
(0.022)
-0.112***
(0.020)
-0.153***
(0.021)
-0.104***
(0.021)
-0.083***
(0.020)
-0.119***
(0.021)
0.024***
(0.006)
0.033***
(0.009)
-0.094***
(0.022)
-0.075***
(0.021)
-0.117***
(0.021)
0.024***
(0.006)
0.031***
(0.009)
-0.083***
(0.022)
-0.070***
(0.022)
-0.099***
(0.023)
0.022***
(0.005)
0.032***
(0.009)
(1)
(2)
(3)
(4)
(5)
(6)
-0.169***
(0.013)
-0.112***
(0.014)
-0.129***
(0.016)
-0.101***
(0.016)
-0.095***
(0.017)
-0.083***
(0.018)
HGC
AFQT
Panel B: Sharp color line
Black
Panel C: Blurred color line, compared with Black (main effect not shown)
Black, light
Black, dark
(1)
(2)
(3)
(4)
(5)
(6)
-0.011
(0.026)
-0.035
(0.025)
-0.021
(0.023)
-0.025
(0.023)
-0.015
(0.025)
-0.041*
(0.024)
-0.022
(0.023)
-0.036
(0.023)
-0.019
(0.023)
-0.041*
(0.023)
-0.012
(0.024)
-0.028
(0.024)
X
X
X
X
Both
X
X
X
X
X
X
Both
X
X
X
X
X
Both
X
1,855
12,953
1,855
12,953
1,855
12,953
Controls, All Panels
Interviewer FE
Interviewer Controls
Region & MSA
Background
Behaviors
Height & Weight
Age, Age at Entry
Year FE
Individuals
N
X
X
Both
X
X
X
X
X
Age
X
1,855
12,953
1,855
12,953
1,855
12,953
Notes: Dependent variable is the share of weeks employed 30+ hours in the past interview year
(weeks employed/weeks with non-missing labor force status). Coefficients are marginal effects
from a GLM with logit-link specification. Sample is respondents who have entered the labor
market (not enrolled and have at least two semesters out of school), with non-missing HGC and
AFQT. Coefficients in Panels A and B are relative to White, in Panel C coefficients relative to
medium Black. Controls are described in the Data Appendix.
Standard errors clustered on individuals in parentheses.
∗p < 0.10, ∗∗ p < 0.05, ∗∗∗p < 0.01.
31
Table 7: Conditional differences in wages by race and skin color, interactions with
potential experience and tenure.
(1)
(2)
(3)
(4)
-0.073**
(0.032)
-0.143***
(0.028)
-0.137***
(0.029)
0.020***
(0.005)
0.002
(0.009)
0.002
(0.007)
-0.018***
(0.006)
0.051***
(0.005)
0.030***
(0.011)
-0.000
(0.001)
0.007**
(0.003)
-0.072**
(0.030)
-0.138***
(0.027)
-0.135***
(0.030)
0.018***
(0.005)
-0.012
(0.008)
-0.008
(0.007)
-0.025***
(0.006)
0.049***
(0.005)
0.027**
(0.011)
-0.074**
(0.030)
-0.139***
(0.027)
-0.136***
(0.030)
0.018***
(0.005)
-0.012
(0.008)
-0.008
(0.007)
-0.025***
(0.006)
0.050***
(0.005)
0.025**
(0.011)
0.020***
(0.005)
0.036**
(0.015)
0.018*
(0.011)
0.007
(0.013)
0.021***
(0.005)
0.038***
(0.014)
0.022**
(0.011)
0.011
(0.013)
-0.003**
(0.001)
0.007
(0.005)
-0.015***
(0.004)
0.023***
(0.009)
-0.015***
(0.004)
0.027***
(0.009)
Panel A: Blurred color line, compared with White
Black, light
Black, medium
Black, dark
Potexp (centered)
Light*Potexp
Medium*Potexp
Dark*Potexp
HGC
AFQT
-0.069**
(0.033)
-0.142***
(0.028)
-0.135***
(0.029)
0.021***
(0.005)
-0.002
(0.009)
-0.004
(0.006)
-0.023***
(0.006)
0.051***
(0.005)
0.029***
(0.011)
HGC*potexp
AFQT*potexp
Tenure (centered)
Light*Tenure
Medium*Tenure
Dark*Tenure
HGC*tenure
AFQT*tenure
Panel B: Sharp color line (main effects not shown)
Black*Potexp
-0.009**
(0.005)
-0.004
(0.005)
Black*Tenure
Panel C: Blurred color line, compared with Black (main effect not shown)
Light*Potexp
Dark*Potexp
0.002
(0.010)
-0.019**
(0.008)
-0.000
(0.010)
-0.020**
(0.008)
-0.003
(0.009)
-0.016**
(0.008)
0.018
(0.017)
-0.011
(0.015)
-0.004
(0.009)
-0.016**
(0.008)
0.016
(0.016)
-0.011
(0.015)
8,436
0.339
8,436
0.340
8,312
0.361
8,312
0.362
Light*Tenure
Dark*Tenure
N
R2
Notes: Dependent variable is log of real hourly wage. Sample is respondents who have entered the labor market
with non-missing HGC and AFQT, and valid wages (30+ hrs./week). Panels A and B are relative to Whites,
Panel C coefficients relative to medium Black. Potential experience is (Age-HGC-6) in Columns 1-2, and is job
tenure (rounded to nearest year) in columns 4-6.
Standard errors clustered on individuals in parentheses. ∗p < 0.10, ∗∗ p < 0.05, ∗∗∗p < 0.01.
32
Figure 1: Histogram of skin color for Black and White
males, by race.
Notes: Sample is 2,417 non-Hispanic Black and White male respondents with skin color ratings in 2008.
33
Appendix
Figure A.1: NIS/NLSY skin color rating card.
Notes: NIS skin color rating scale. Users of the scale are requested to notify NIS staff by contacting Jennifer Martin at
[email protected].
34
35
-0.118***
(0.023)
(1)
-0.069**
(0.033)
-0.142***
(0.028)
-0.142***
(0.030)
-0.117***
(0.029)
(2)
-0.041
(0.041)
-0.158***
(0.034)
-0.150***
(0.037)
(2)
8,436
0.338
5,361
0.328
X
0.118***
(0.043)
0.008
(0.038)
(2)
5,669
0.435
X
0.089**
(0.040)
0.025
(0.038)
(3)
-0.113***
(0.026)
(3)
-0.061
(0.037)
-0.150***
(0.033)
-0.126***
(0.034)
0.399***
(0.032)
(3)
8,634
0.341
X
0.071**
(0.034)
-0.002
(0.030)
(4)
-0.121***
(0.023)
(4)
-0.072**
(0.033)
-0.143***
(0.028)
-0.145***
(0.029)
(4)
9,039
0.326
X
0.067**
(0.034)
0.007
(0.030)
(5)
-0.117***
(0.023)
(5)
-0.074**
(0.033)
-0.141***
(0.028)
-0.134***
(0.030)
(5)
7,871
0.341
X
0.086**
(0.034)
0.005
(0.040)
(6)
-0.122***
(0.025)
(6)
-0.070**
(0.034)
-0.156***
(0.028)
-0.151***
(0.039)
(6)
8,345
0.340
0.056
(0.035)
-0.007
(0.030)
(7)
-0.130***
(0.024)
(7)
0.095
(0.065)
-0.088**
(0.034)
-0.144***
(0.028)
-0.151***
(0.030)
(7)
8,182
0.345
X
0.051
(0.036)
-0.006
(0.031)
(8)
-0.128***
(0.024)
(8)
-0.090**
(0.035)
-0.141***
(0.028)
-0.147***
(0.030)
(8)
8,002
0.347
X
0.058
(0.036)
0.001
(0.031)
(9)
-0.128***
(0.024)
(9)
-0.087**
(0.035)
-0.145***
(0.028)
-0.144***
(0.030)
(9)
Notes: All columns include all controls and sample restrictions in Column 6 of Table 5. Omit Entry Year omits year of labor market entry and includes
wage in that year as additional covariate. All Color Ratings includes 56 respondents with color ratings in 2009 or 2010. Hrs/Wk ≥20 includes wages
of respondents working 20+ hours per week. Black≥1 & White ≤6 drops outliers in color. Omit “Mixed” omits White respondents with an identifiable
non-White biological parent. Blk6=1 & Wht 6 omits Black respondents with skin color = 1 and White respondents with skin color = 7 or 8. Standard
errors clustered on individuals in parentheses; (∗p < 0.10, ∗∗ p < 0.05, ∗∗∗p < 0.01.)
N
R2
Year≥2006
Omit entry year
All color ratings
Hrs/wk ≥ 20
If Black≥ 2 & White≤ 6
Omit mixed Black
Omit mixed
0.072**
(0.034)
Black, dark
-0.001
(0.030)
Alternate Sample Specifications
Black, light
(1)
Panel C: Blurred color line, compared with Black
Black
Panel B: Sharp color line
Black*mixed
Entry wage
Black, dark
Black, medium
Black, light
(1)
Panel A: Blurred color line, compared with White
Table A.1: Robustness checks 1, variation in sample definitions.
Table A.2: Robustness checks 2, semi-parametric skin color.
(1)
Black, 1-3
Black, 4
Black, 5
Black, 6
Black, 7
Black, 8
Black, 9-10
HGC
AFQT
Potexp (centered)
HGC*potexp
AFQT*potexp
Black, 1-3*potexp
Black, 4*Cpotexp
Black, 5*potexp
Black, 6*potexp
Black, 7*potexp
Black, 8*potexp
Black, 9-10*potexp
Tenure (centered)
HGC*tenure
AFQT*tenure
Black, 1-3*tenure
Black, 4*tenure
Black, 5*tenure
Black, 6*tenure
Black, 7*tenure
Black, 8*tenure
Black, 9-10*tenure
N
R2
(2)
(3)
mean
std. err.
mean
std. err.
mean
std. err.
-0.146***
-0.052
-0.033
-0.171***
-0.114***
-0.133***
-0.145***
0.051***
0.030***
0.019***
(0.053)
(0.056)
(0.045)
(0.033)
(0.035)
(0.036)
(0.039)
(0.005)
(0.011)
(0.005)
-0.151***
-0.064
-0.035
-0.172***
-0.118***
-0.131***
-0.141***
0.051***
0.030***
0.020***
-0.000
0.006**
-0.019
0.021
-0.002
-0.008
0.010
-0.020***
-0.017*
(0.053)
(0.053)
(0.043)
(0.032)
(0.035)
(0.036)
(0.038)
(0.005)
(0.011)
(0.005)
(0.001)
(0.003)
(0.013)
(0.015)
(0.013)
(0.009)
(0.009)
(0.007)
(0.009)
-0.119***
-0.084**
-0.034
-0.161***
-0.115***
-0.125***
-0.143***
0.050***
0.025**
0.018***
(0.044)
(0.042)
(0.043)
(0.031)
(0.034)
(0.037)
(0.041)
(0.005)
(0.011)
(0.005)
-0.036***
0.003
-0.010
-0.013
-0.004
-0.024***
-0.026***
0.021***
-0.003**
0.007
0.070***
0.037
0.019
0.021
0.021
0.012
0.008
(0.010)
(0.012)
(0.013)
(0.010)
(0.009)
(0.008)
(0.008)
(0.005)
(0.001)
(0.005)
(0.016)
(0.028)
(0.016)
(0.015)
(0.014)
(0.017)
(0.017)
8,436
0.339
8,436
0.343
8,312
0.365
Notes: Dependent variable is natural log of real hourly wage. Sample is respondents who have
entered the labor market (not enrolled and have at least two semesters out of school), with
non-missing HGC and AFQT, and valid wages (30+ hrs./week). Coefficients in Panels A and B
are relative to White, in Panel C coefficients are relative to medium Black. Potential experience
(potexp) is (Age-HGC-6); HGC, potential experience and tenure are centered.
Standard errors clustered on individuals in parentheses.
∗p < 0.10, ∗ ∗ p < 0.05, ∗ ∗ ∗p < 0.01.
36
Data Appendix
Race
We define race using the NLSY’s RACE_ETHNICITY and ETHNICITY measures, excluding any
respondents who are classified as anything other than either non-Hispanic Black or White. From
this we exclude 127 White male respondents from the original sample.
Labor market entry
Labor market entry is defined as the first interview after four consecutive non-enrolled “semesters.”
To identify this, we use the NLSY’s monthly enrollment arrays, defining the fall semester as
September-November and the Spring semester as February-May, where enrollment in any month of
a semester would count as enrollment in that semester to account for missing values. In the case
that respondents return to school after four consecutive non-enrolled terms, we omit respondents
enrolled at the time of interview from analyses.
Employment
To determine the share of weeks employed in the past year, we use the NLSY weekly employment
arrays. Because these measures can be filled in retrospectively for non-interviewer years, for respondents who miss an interview year, we assign a temporary interview data half-way between the
prior and subsequent interviews. To account for differences across respondents in the number of
weeks since the most recent interview, the share of weeks employed is defined as the share of weeks
employed divided by the number of weeks since the last interview, as opposed to the past calendar
year. Like wages, we only estimate employment for non-enrolled respondents.
Wages
We use the wage in the most recent job at each interview, omitting wages below $2.00 and above
$100 per hour, and wages in military and self-employed jobs. Like employment, we only estimate
labor market outcomes for non-enrolled respondents.
Education and AFQT
Although education is measured at each interview and can be filled in retrospectively, some respondents have missing values. We recode these for respondents who have the same highest grade
completed on both sides of a missing value. AFQT is normalized to a sample with all Black and
White males with a skin color rating in 2008. Regression analysis is conducted only on respondents
with AFQT and highest grade completed in the year wages are reported.
Regression control variables
In the full specification, we include the following controls:
Geographic indicators include: Region (South omitted), MSA status (in MSA, not central city
omitted). Childhood background characteristics include: Family poverty ratio in 1997, Did family
ever receive government aid by 1997, Did respondent live with both biological parents at age 6,
Mother’s highest grade completed (high school omitted), did respondent live outside an MSA at age
12, Dod respondent live in the South at age 12. Height and Weight are measured in 1997. Behavioral
controls are the first two principal components from a PCA of a series of questions asked in 1997:
Did respondent ever smoke, drink, use marijuana, use a handgun, destroy property, steal, sell drugs,
do respondents peers smoke, use drugs, regularly attend church.
37