Contributions of Non-Cognitive Skills to the Racial Wage Gap Throughout the Distribution ∗†

Melinda Petre

This Version: August 28, 2017

Abstract Analyzing the distributions of wages for white, black, and Hispanic men reveals dierences in wages throughout the distribution. There are also clear cognitive and non-cognitive skill dierences across the groups. Do dierences in the distributions of these skills explain dierences in the distributions of wages? Conditional on observed skills, are blacks, whites, and Hispanics in the same quantiles of the wage distribution rewarded dierently? How much of the observed wage gaps are explained by cognitive and non-cognitive skills? Using data from the NLSY79, I look at the impacts of noncognitive skills on wages for blacks, Hispanics, and whites throughout the distribution. I use within unconditional quantile (Firpo, Fortin, and Lemieux, 2010) Oaxaca-Blinder decompositions to understand how much of the explained wage gap can be attributed to cognitive and non-cognitive skills. I also use reweighted (DiNardo, Fortin, and Lemieux, 1996) within unconditional quantile Oaxaca-Blinder decompositions to understand dierences in the predicted wage gap and whether skills help explain these dierences. I nd that non-cognitive skills explain 16-27% of the wage gap between blacks and whites throughout the distribution and 26-33% of the wage gap throughout the distribution between Hispanics and whites while the portion of the gap explained by cognitive skills decreases substantially in the right tail of the distribution for both comparisons. Between blacks and Hispanics, non-cognitive skills explain 29.8% of the ∗

School of Education, University of California-Irvine, [email protected].

†

The research reported here was supported by the Institute of Education Sciences, U.S. Department

of Education, through Grant R305B120013 to the University of California, Irvine (PI: Greg Duncan). The opinions expressed are those of the authors and do not represent views of the Institute or the U.S. Department of Education. I would like to thank Marianne Bitler, Sandra Black, Jerey Denning, Daniel Hamermesh, Qian Lu, Gerald Oettinger, Chester Polson, Maria Fernanda Rosales Rueda, and Stephen Trejo, as well as numerous participants in the University of Texas seminars and the SOLE/EALE World Conference for their helpful comments. All remaining errors are my own.

1

wage gap in the 10th percentile and 18.9% to 19% in the 40th and 50th percentiles. The explanatory power of non-cognitive skills at other points in the distribution is smaller in magnitude. Evidence from predicted reweighted wages suggests that conditional on skills and education, the wage gaps between blacks and whites should be smaller than they we actually observe throughout the distribution. Likewise, especially in the bottom half of the wage distribution, the reweighted estimates predict a smaller wage gap between whites and Hispanics. Looking at Hispanics and blacks, a larger gap is predicted than the gap we actually observe and the majority of this gap can be explained by cognitive and non-cognitive skills. 1

Introduction

Do dierences in the distributions of skills explain dierences in the distributions of wages? Plotting the distributions of wages for whites, blacks, and Hispanics reveals the existence 1

of a wage gap throughout the entire distribution, as seen in Figure 2.

This is also evid-

ent from the literature establishing that wage gaps exist between blacks, Hispanics, and whites (Carneiro et al. (2005), Cain (1987), Altonji and Blank (1999), and Urzua (2008), for example). In addition, there are clear cognitive skill dierences, as seen in Figure 3 and conrmed in the literature (Carneiro et al. (2007), for example). We know that, on average, both black and Hispanic males make less than white males, but what happens to the wage gap when we compare those with similar skills and look across the entire distribution of wages?

How much of the wage gap at dierent points in the distribution is explained by

cognitive and non-cognitive skills? If whites had the same distribution of skills as blacks and Hispanics, would we observe dierences in wages throughout the distribution? How would these dierences impact the predicted wage gaps throughout the distribution? While other papers examine the role of cognitive and non-cognitive skills in explaining wages and wage gaps (Murnane et al. (2001) and Urzua (2008), for example), this is the rst paper to incorporate both cognitive and non-cognitive skills in explaining the wage gap at dierent points throughout the distribution. I expand upon the literature by decomposing the observed wage gap between whites, blacks, and Hispanics the portion explained by cognitive skills and multiple non-cognitive skills using data from the National Longitudinal Survey of Youth 1979 cohort (NLSY79) and unconditional quantile Oaxaca-Blinder decompositions. Then, I reweight the data using methods from DiNardo, Fortin, and Lemieux to predict wages throughout the distribution that whites would earn based on the distributions of skills for other groups.

These results have implications for understanding how skill gaps might

1 This gure is constructed using data from the NLSY79.

2

contribute to wage gaps throughout the distribution.

Specically, since wage gaps exist

to varying extents throughout the distribution of wages, it is important to understand if the sources of the gap vary across the distribution as well. Moreover, studying wage gaps between racial groups is becoming increasingly important as the racial composition of the U.S. shifts toward historically minority groups and wage disparities persist. Here, I dene cognitive skills as IQ, book smarts, and raw intelligence and non-cognitive skills (personality traits, soft skills) as resilience, motivation, self esteem, people skills, internal control, and other related skills. The NLSY79 oers some psychological tests which serve as proxies for non-cognitive skills. These include the Pearlin Mastery Score, Coding Speed Score, Rosenberg Score and Rotter Internal Locus of Control Scale. I include all of these measures and evaluate their collective impact on wage gaps.

The Pearlin Mastery

Scale measures alienation and anomie which is dened as a subjective sense of powerlessness and state of meaninglessness (Seeman (1991)). Coding Speed Scores represent a measure of motivation (eort) because the test requires eort to achieve a good score. Rosenberg Scores measure self esteem and Rotter Scores measure the degree to which an individual views life outcomes are their own doing versus their environments. These tests are described in detail in the main text. As is convention in the literature, I use the Armed Forces Qualication Test (AFQT) as a measure of cognitive skills. Variation in cognitive and non-cognitive test scores across individuals provides one possible explanation for observed dierences in wages across individuals and racial groups. This paper strives to quantify the impact of these skill dierences on the wage gap at dierent quantiles. Consistent with existing work, I nd evidence that non-cognitive skills are a determinant of wages.

The evidence in this paper also suggests that non-cognitive skills are import-

ant throughout the distribution of wages, especially in explaining wage gaps throughout. I nd that non-cognitive skills explain 16-27% of the wage gap between blacks and whites throughout the distribution and 26-33% of the wage gap throughout the distribution between Hispanics and whites while the portion of the gap explained by cognitive skills decreases substantially in the right tail of the distribution for both groups. Between blacks and Hispanics, non-cognitive skills explain 29.8% of the wage gap in the 10th percentile and 18.9% to 19% in the 40th and 50th percentiles. Non-cognitive skills do not seem to explain as much of the gap between blacks and Hispanics at other points. Evidence from predicted wages suggests that conditional on skills and education, the wage gap between blacks and whites should be smaller than we observe throughout the distribution. Likewise, estimated wages predict a smaller wage gap throughout the bottom half of the distribution between whites and Hispanics. Finally, reweighting methods predict a larger wage gap between blacks and Hispanics than is actually observed conditional on skills and education.

3

The remainder of this paper is as follows. The related literature is discussed in Section 2, the methods are described in Section 3, and a description of the data follows in Section 4.

Results are presented in Section 5 and conclusions and discussion follow in Section 6.

Descriptions of skill measures, Figures, Tables, and some methodological refreshers are in the Appendix.

2

Literature Review

The relevant literatures highlight three points: (1) wage gaps exist amongst blacks, whites, and Hispanics, (2) skills gaps exist amongst blacks, whites, and Hispanics, and (3) noncognitive skills are important.

This paper contributes to all three of these strands by in-

vestigating dierences in cognitive and non-cognitive skills as contributors to the wage gap throughout the distribution.

2.1

Wage Gaps Exist

There is a large literature establishing the existence of wage gaps between blacks, whites, and Hispanics.

Most generally, Cain (1987) and Altonji and Blank (1999) survey a wide

range of literature establishing the existence of both wage and skills gaps. In addition, using the NLSY79, Oettinger (1996) nds that no wage gap between blacks and whites exists at the beginning of careers, but that one develops over time, mostly as a result of mobility dierences between blacks and whites.

Also using the NSLY79, Neal and Johnson (1996)

nd that dierences in AFQT scores account for most of the wage gap between young male blacks and whites and that gaps in test scores can be traced back to observable dierences in family backgrounds and school environments between blacks and whites.

Fryer (2010)

and Fryer et al. (2013) nd that educational attainment helps explain the wage gap for blacks and Hispanics. Carneiro et al. (2005) look at the relative signicance of cognitive skill dierences and expectations about discrimination in wage gaps, nding that both factors are not plausible explanations for the wage gaps that are observed. Urzua (2008) investigates the relationship between ability, schooling choices, and black-white wage gap, controlling for enogeneity between ability and schooling, nding that even after controlling for ability, racial wage gaps exist.

Specically, Urzua looks at the aggregate wage gap by education

level, whereas I look at the wage gap at dierent points in the distribution and decompose this gap into portions explained by dierent types of skills. Reimers (1983) nds evidence that the wage gap results from discrimination between whites and Hispanics. Grenier (1984) uses data from the 1976 Survey of Income and Edu-

4

cation to nd that a language handicap explains a large portion of the wage dierential between whites and Hispanics.

This paper builds on the literature by decomposing the

wage gap across all three racial groups at dierent points in the distribution into the pieces explained by cognitive and non-cognitive skills, as well as predicting wage distributions conditional on skills and other individual characteristics. In addition, I pay special attention to the black Hispanic wage gap.

2.2

A Skills Gap Exists

This literature establishes that there are cognitive and non-cognitive skill dierences between blacks and whites and that these dierences emerge from an early age. Carneiro and Heckman (2003) and Cunha and Heckman (2007) present evidence of an early gap in both cognitive and non-cognitive skills. Fryer and Levitt (2004) nd that after controlling for individual and environmental characteristics, there is no black-white cognitive achievement gap when children enter kindergarten but a gap emerges during kindergarten and rst grade. Carneiro et al. (2007) nd that the impact of non-cognitive ability does not vary systematically when dierent parental socio-economic status and education subgroups of the population are considered using the British National Child Development Survey.

Carneiro et al. (2005) hypothesize

that minority students and parents might have pessimistic expectations about whether they receive fair rewards for their education relative to their white counterparts and that these expectations might lead to a lower investment in skill formation, nding that dierences in cognitive ability begin before formal schooling starts. Murnane et al. (2001) examines academic skills, the ability to complete tasks quickly, and self-esteem and the impacts of these on predicting wages for black, white, and Hispanic men, nding that these three measures are of varying importance across groups. Lundberg (2013) and Lundberg (2014) nd that socio-economic status, which is correlated with race, impacts which skills people need to complete education, nding clear dierences across socioeconomic statuses. I provide evidence from the NLSY79 of dierences in distributions of skills and evidence that this variation might help explain the wage gaps that exist throughout the distribution.

2.3

Non-cognitive Skills are Important

There is a growing literature establishing that non-cognitive skills are important in determining life outcomes. Farkas (2003), Bowles and Gintis (1976), and Bowles and Gintis (2002) summarize early ndings in the literature, claiming that only 20% of earnings are due to cognitive ability and the remaining 80% could be attributed to non-cognitive skills. Lindqvist

5

and Vestman (2011) nd that the returns to cognitive and non-cognitive skills vary throughout the distribution for men in Sweden. This paper adds to this literature by exploring the contribution of skills to the wage gap across the wage distribution among men from dierent races. Heckman et al. (2001) and Heckman and Rubinstein (2001) use non-cognitive skills to explain why GED recipients earn less, work at lower hourly rates, and have lower levels of schooling than other dropouts. Heckman et al. (2006) look at the eects of cognitive and non-cognitive skills on wages, schooling, work experience, occupational choice, and participation in risky adolescent behaviors, demonstrating a correlation between these abilities and educational choice. Lleras (2008) uses the National Education Longitudinal Study to look at the impact of cognitive and non-cognitive skills on educational attainment and earnings 10 years after high school graduation, nding that those with better social skills, work habits, and extracurricular activities have higher educational attainment and earnings.

3

Empirical Strategy

The empirical strategy consists of two parts. First, I use an unconditional quantile OaxacaBlinder decompositions (Firpo et al. (2007), Fortin et al. (2010)), hereafter FFL-OB to look at the portion of the wage gap explained by skills throughout the distribution. Then, I use a reweighted, counterfactual distribution technique from DiNardo et al. (1996) combined with FFL-OB to predict the wage gap and the portion explained by skills throughout the distribution if whites had the same skill distributions as blacks (Hispanics). I begin by discussing 2

FFL-OB and DFL.

These methods where chosen due to their ease of implementation, the

simplicity of the reweighting method, and because the literature has established that they are ecient under certain assumptions (e.g. Hirano et al. (2003) and Firpo (2007)). These advantages are discussed in further detail below.

3

From FFL-OB, we can learn about the

portion of the wage gap at dierent points in the distribution that is explained by cognitive and non-cognitive skills. Using DFL combined with FFL-OB tells us about predicted wage gaps conditional on skills, oering insight into where inequality might be present.

2 I provide a general refresher on OB decompositions in Appendix B. 3 There are also limitations associated with these methods including general equilibrium eects and the potential for reweighting to have undesirable properties in small samples, especially if the common support assumption is a stretch.

6

3.1

Unconditional Quantile Oaxaca Decompositions (FFL-OB)

FFL uses re-centered inuence regressions to estimate conditional distributions and unconditional quantile re-centered inuence regressions (RIF). Dene RIF as interest).

Y

replaced by a re-centered inuence function (based on the statistic of

Then, IF(y; v) is the inuence function corresponding to the observed wage

y

v(Fy ) and RIF(y; v) = v(Fy ) + IF(y; v) aggregates back to the statistic of interest, RIF(y; v)dF (y) = v(Fy ). RIF(y; v) can be modelled as a function of explanatory variables: E [RIF(y; v)|X] = Xγ +  where γ is estimated by OLS. τ −1(y≤Qτ ) In the case of quantiles, the re-centered inuence function, IF(y; Qτ ) is given by fY (Qτ ) where 1(·) is an indicator function and fY (Qτ ) is the density of the marginal distribution of Y . It follows that for the distributional statistic of interest

´

RIF(y; Qτ )

= Qτ + IF(y, Qτ ) = Qτ +

τ − 1(y ≤ Qτ ) = c1,τ 1(y > Qτ ) + c2,τ fY (Qτ )

c1,τ = f (1Q ) and c2,τ = Qτ − c1,τ (1 − τ ). This Y t 1(y ≤ Qτ ) on X estimated at y = Qτ using a link where

of

can be estimated from a regression function from a linear probability

model. The RIF is computed using the sample quantiles kernel methods. Then,

ˆ (Y ; Qτ ) RIF

is obtained using

ˆ τ , the estimated density at τ Q ˆ τ and fˆ(Q ˆ τ ) in RIF(y; Qτ ). Q

using

The coecients from the unconditional quantile RIF regressions are:

!−1 γˆb,τ =

X

X

xi xTi

i

ˆ (Yb ; Qτ )xi RIF i

i

Then, the OB decomposition follows:

∆τ = x¯b=1 (ˆ γb=1,τ − γˆb=0,τ ) + (¯ xb=1 − x¯b=0 )ˆ γb=0,τ where

∆τ

is the dierence in wages at

τ.

This gives the portion of the wage gap within each quantile explained by characteristics, as estimated in Section 5.1. Here, the unconditional quantile decompositions are limited by the linear approximation. It could be that using linear approximations at dierent points of the distribution loses some of the informational richness in the distribution at any given point which has implications for the goodness of t of the approximation. In addition, these methods assume that the distribution of skills has a common support across all races. Based on the distribution of skills as discussed in Section 4, it is reasonable to assume that the

7

4

support is common.

3.2

Estimating the Counterfactual Densities of Wages

In order to predict the distribution of wages conditional on skills, I use a reweighing approach from DiNardo et al. (1996). This technique is outlined below. For simplicity, it is explained using the whole distribution rather than unconditional quantiles as detailed in Section 3.1. For details of DFL using RIF regressions for unconditional quantiles, please see Fortin et al. (2010). Individual characteristics are observed and written as(Y, X, b) where individual attributes and

X

b

Y

are wages, 5

is an indicator of whether or not the individual is black.

X

are

Here,

includes cognitive and non-cognitive skills as well as other individual characteristics. The

joint distribution of wages is written as value of

b

is

F (Y, X, b) and the joint distribution given a particular

F (Y, X|b).

Given the joint distribution of wages and the conditional distribution of wages for a particular value of

b,

the density of wages conditional on

b

can be written as a function of

the joint distribution of wages. For example, for a black individual

(b = 1),

the distribution

of wages is as follows:

ˆ f (Y |X, b = 1)dF (X|b = 1) = f (Y ; bY = 1, bX = 1)

f (Y ) = X∈Ω In this notation, distribution of

X

bY

is the distribution of wages for a given value of

characteristics for a given value of

b

and

bX

is the

b.

Then, the distribution of wages over white individuals can be written as a function of the distribution of characteristics of those black individuals as follows:

ˆ f (Y |X, bY = 0)dF (X|bX = 1)

f (Y ; bY = 0, bX = 1) = ˆ

f (Y |X, bY = 0)dF (X|bX = 1)

= ˆ =

dF (X|bX = 0) dF (X|bX = 0)

f (Y |X, bY = 0)ψX (X)dF (X|bX = 0)

where

4 This assumption is required for these methods to be ecient. 5 That is,

( 1 b= 0

if an individual is black if an individual is not black or Hispanic (white)

8

ψX (X) =

dF (X|bX = 1) dF (X|bX = 0)

is a reweighing function that can be estimated from the data derived using Bayes' rule as follows:

dF (X|bX = 1) dF (X|bX = 0) Pr(bX = 1|X) × = Pr(bX = 0|X)

ψX (X) =

Pr(bX

= 0) Pr(bX = 1)

In the reweighing function, Pr(bX

= 1|X) and Pr(bX = 0|X) can be estimated from the data using a probit specication, and Pr(bX = 1) and Pr(bX = 0) are observed directly in the data. Once estimates of ψˆX (X) are obtained from the sample probabilities and conditional probability estimates, kernel density estimation is used to back out the counterfactual distribution.

This gives the distribution of wages that would prevail for white workers if

they had the distribution of characteristics of the black workers. Estimates use FFL-OB to decompose the skill predicted wage gap (that is the wage gap predicted for white workers if they had the skills of black/Hispanic workers) into portions that are explained by cognitive and non-cognitive skills. The key limitation of this approach is that it applies a partial equilibrium analysis in a general equilibrium setting. This is also a limitation of FFL-OB. This leaves questions about the endogeneity versus exogeneity of the labor market. Applying this approach requires maintaining the assumption that labor market institutions which dictate the prices of labor in the market are exogenous. However, in reality, these prices might be dependent on the changing composition of skills in the labor market.

Similarly, another

limitation is that this method requires assuming a common support among the distributions of skills across races. However, provided this assumption holds, the literature (e.g. Hirano 6

et al. (2003) and Firpo (2007)) has shown that these methods are ecient.

4

Data

This paper uses data on males from the National Longitudinal Survey of Youth, 1979 co7

hort (NLYS79).

Key variables include: race, urban residence, census region of residence,

wages, potential experience, and educational attainment (measured in years). Non-cognitive

6 It is also possible that schooling and labor market participation are endogenous. See Urzua (2008) and Heckman (1979) for discussion.

7 Women are omitted due to questions about their labor force attachment.

9

measures are discussed at length below and include the Rotter Internal Locus of Control Score, the Rosenberg Score, the Pearlin Mastery Scale, and the Coding Speed Test Score. These measures are discussed at length below. AFQT scores are also recorded and used as a measure of cognitive skills. Table 1 reports summary statistics including whether a residence is urban, region of residence, log hourly wages broken down into ve year age ranges, and potential experience. Potential experience is dened as age minus years of schooling minus six.

The table also

summarizes AFQT scores and the Rotter, Rosenberg, Pearlin, and Coding Speed measures and highest grade completed.

Only individuals with more than 8 years of schooling are

included. For consistency, hourly wages are converted to 1990 dollars.

8

The distribution of log

wages for whites, blacks, and Hispanics appears in Figure 2. The distribution of log wages for whites is slightly higher than the distribution for Hispanics, which is slightly higher than the distribution of log wages among blacks.

Potential experience is largest for Hispanics,

then blacks across most age groups. Whites on average attend two thirds of a year more of school than blacks do on average who on average attend a fth more of a year of school than Hispanics.

4.1

Measures of Cognitive Skills

The AFQT test was given as part of the NLSY79. AFQT scores are standardized by birth year, as is convention in the literature. Although study participants were born in dierent years, the test was administered to all subjects at the same time and thus, standardization by birth year corrects for any gain in test scores that results from being older. Average standardized AFQT scores by race are reported in Table 1.

9

The average for

whites 0.41 (standard deviation 0.88) in the sample is larger than the average for blacks 0.72 (standard deviation 0.91). Hispanics fall in the middle: -0.20 (standard deviation 0.90). The densities of AFQT scores for whites, blacks, and Hispanics are displayed in Figure 3. Note that the density of scores among whites is more highly concentrated around the mean than it is for the other two races.

8 100 dollars in 2009 is approximately 61 dollars in 1990 dollars. 9 This standardization does not account for possible endogeneity by education level as in Urzua (2008).

10

4.2

Measures of Non-Cognitive Skills

Table 1 summarizes measures of non-cognitive skills for blacks, Hispanics, and whites. measures are standardized by birth year.

The Rotter Locus of Control Scale

10

All

11

The Rotter Locus of Control Scale measures the

amount of control individuals believe that they have over their own lives. That is, whether individuals feel they have control over their own outcomes or whether their environment determines them. The version of the test administered in 1979 as part of the NLSY79 is an abbreviated version containing four questions. Each question has between 1 and 4 points so scores can range from 4 to 16. A score of 4 on a question means that an individual feels that internal elements control life outcomes whereas a score of 1 indicates that an individual feels as though their environment has control. Questions are asked in pairsan internal and an external questionand respondents scores indicate which statement they more closely relate to. A higher the score represents an individual with more internal control. The list of questions can be found in Appendix A.1. According to Christie (1991), the Rotter Locus of control scale is the most widely used and cited measure of locus of control.

12,13

Raw averages for the Rotter Locus of Control Scale are reported in Table 1, as well as the birth year standardized averages and standard deviations. The average scores for whites are slightly larger than those of blacks, which are slightly larger than those of Hispanics: this means that Hispanics and blacks are more likely to believe that their environment has

10 The literature on non-cognitive skills often uses psychologist interviews and teachers evaluations to assess non-cognitive skills and look at their impact on lifetime outcomes. Segal (2008) uses teacher surveys from NELS, where teachers were surveyed about tardiness, inattentiveness, disruptiveness, homework completion and absenteeism to nd that classroom behavior is related to family background variables for boys: higher educated and higher income families are linked to better classroom behavior.

Tsai (2007) uses the 1988

NELS for premarket measures of non-cognitive skills. He uses the Rotter and Rosenberg tests and teacher evaluations. He nds some evidence that lower non-cognitive skills explain returns to the GED. Kuhn and Weinberger (2005) control for cognitive skills and nd that those who occupy leadership positions in high school earn 4-33% more as adults, using the Project TALENT (1960), NLS72 and High School and Beyond (82 seniors). Lindqvist and Vestman (2011) use Psychologist interviews from Swedish military enlistment to measure non-cognitive skills. They nd that those men with low earnings and face unemployment lack non-cognitive skills and that cognitive ability is a better predictor of earnings for more skilled workers above the median. This is not possible with the NLSY: there are no teacher evaluations and psychologist interviews in the data.

11 Descriptions of psychological tests were adapted from:

https://www.nlsinfo.org/content/cohorts/nlsy79/topical-guide/attitudes?nopaging=1. Accessed October 18, 2013.

12 Christie (1991) denes locus of control as: assumed internal states that explain why certain people

actively, resiliently and willingly try to deal with dicult circumstances while others succumb to a range of negative emotions.

13 There is a series of papers that looks at the Rotter Locus of Internal Control and Rosenberg Self Esteem

Score on lifetime outcomes. For example, Heckman et al. (2006) uses the NLSY79 and use AFQT scores as a measure of cognitive skills and the Rosenberg/Rotter test scores as a measure of non-cognitive skills.

11

more control over their lives than whites. The densities of the standardized Rotter Scores are found in Figure 4. The distributions of scores for whites, blacks, and Hispanics with this measure are fairly similar, providing some justication for the common support assumption.

The Rosenberg Self-Esteem Score

The Rosenberg Self-Esteem Scale describes the de-

gree of which a person either approves or disproves of their actions. Respondents are asked to agree or disagree with 10 statements of self-approval and disapproval. This test was ad14

ministered to individuals in the NLSY79 in 1980.

Items included are things like: as whole,

I am satised with myself  and at times, I feel as though I am useless. Scores range from 0 to 30, with a score of 30 representing the highest measurable level of self esteem. The list of questions can be found in Appendix A.2. According to Blascovich and Tamaka (1991), the Rosenberg Self-Esteem score is the most popular measure of global self esteem and is the standard with which developers of other measures seek convergence. It has also been shown to be highly internally consistent, with retest reliability contributing to its popularity. Raw averages as well as averages of standardized Rosenberg Self-Esteem Scale scores are reported in Table 1. The version of the test used here was administered in 1980; before most individuals in the sample entered the labor market. These statistics exhibit similar patterns to the Rotter Score: whites have higher self esteem than blacks on verge and blacks have higher self esteem than Hispanics on average. The densities of the standardized Rosenberg Scores can be found in Figure 5.

Once again, these distributions are similar across races

which provides some justication for the common support assumption.

The Pearlin Mastery Scale

The Pearlin Mastery Scale is a seven item test where each

item is a statement about the individuals' perception of themselves.

Respondents choose

strongly disagree, disagree, agree, and strongly agree for each statement. 15

administered to individuals in 1992.

This test was

Examples include: I have little control over what

happens to me and I often feel helpless in dealing with problems in life.

Total scores

are calculated on a scale of 7 to 28, where higher scores represent the perception of greater mastery over one's environment. The list of questions are in Appendix A.3. The psychology literature uses the Pearlin Scale as a measure of alienation and anomie. According to Seeman (1991), this scale measures the extent to which one regards one's life chances as being under

14 Although this test was administered to the 79 cohort in 1980, 1987 and 2006, the dataset used in this analysis only uses the administration from 1980 due to questions about whether non-cognitive skills are stable over time. The administration from 1980 occurs before most individuals enter the labor market, making it less likely that this measure is endogenous with individual's labor market status and wages.

15 Since the Pearlin Mastery Scale was administered in 1992, it is possible that scores are endogenous with

work histories. That is, how individuals respond might be impacted by their current satisfaction with their employment status and wages.

12

one's own control in contrast to being fatalistically ruled. Birth year standardized averages and raw averages are reported for the Pearlin Mastery Scale in Table 1. As was true with other non-cognitive measures, averages are slightly higher for whites than for blacks. blacks.

Averages for Hispanics lie in between averages for whites and

The densities of the Pearlin Score for each group are plotted in Figure 6.

These

densities are all similar: the main dierence being that a higher density of scores for whites are concentrated at the distribution's peak, again lending support for the common support assumption.

The Coding Speed Test

Segal (2012) establishes the Coding Speed Test (a section of

the ASVAB not used in the calculation of AFQT scores) as a measure of motivation. This study uses data from the NLSY, the U.S. military, and an experiment, providing evidence that the relationship between non-incentivized tests and economic success are not solely due to cognitive skills.

16

That is, the lack of performance based incentives on these tests for

civilians allows for non-cognitive skills to inuence test scores. Segal nds that an increase in coding speed is associated with an increase in earnings for male workers. Following suit, I use the Coding Speed Score as a proxy for motivation. The Coding Speed Test is a seven minute, 84 question test. At the beginning of each set, a list of words and a 4-digit code for each word are listed. Questions ask respondents to match the word to its code. A sample question page is found in Figure 1. A high score on the Coding Speed Test represents a more highly motivated individual than a lower test score. The distribution of Coding Speed Scores looks very similar to the distribution of AFQT scores, as evident from Figure 7. The mean of white scores is larger than the mean of Hispanic and black scores.

In addition, white scores seem to be more

concentrated around the mean, whereas black and Hispanic scores are more evenly spread throughout the distributions.

5

Results

Tables 2, 4, and 6 present results from the unconditional quantile Oaxaca (FFL-OB) decomposition evaluated at the 10th, 20th, 30th, 40th, and 50th percentiles of the wage distribution.

16 Participants took the test three times: twice for a xed payment and a third time with performance based monetary incentives.

She found that 38% of participants signicantly improved their scores under

the performance based incentive structure. These results support her hypothesis that if intrinsic motivation varies across individuals, then their ranking with non-incentivized exams might dier than their ranking on incentivized exams. This supports her ndings using the NLSY and military data: military recruits do better than civilians on the test and Coding Speed is correlated with earnings after controlling for cognitive ability and levels of education.

13

Tables 3, 5, and 7 present these same results evaluated at the 60th, 70th, 80th, and 90th percentiles of the wage distribution. These results are discussed in Section 5.1 and generally speaking, these decompositions give the portion of the wage gap at each percentile that is explained by both cognitive and non-cognitive skills.

.

In Section 5.2, these same results

are reported but reweighted using the reweighing factor from DFL to illustrate how wages would look if whites in these percentiles of the wage distribution had the same skills as blacks and Hispanics in these percentiles. Tables 8 and 9 report results comparing blacks with whites and Tables 10 and 11, Hispanics with whites. Tables 12 and 13 compare blacks and Hispanics.

5.1

Unconditional Quantile Oaxaca Decomposition Results

Table 2 reports results for the 10th, 20th, 30th, 40th, and 50th percentiles of the hourly log wage distribution decomposed within quantiles into the portion of the wage gap explained by characteristics and the portion unexplained by observed characteristics.

17

Table 3 does the

same for the 60th, 70th, 80th, and 90th percentiles. These observed characteristics include: years of education, cognitive skills, non-cognitive skills, whether an individual resides in a 18

city, region of residence, and a cubic in potential experience.

These decompositions are

estimated using FFL-OB and do not use the DFL reweighing factor.

Thus, the portion

explained represents how much of the observed wage gap within the quantile is explained by the given characteristics. Standard errors are reported in parentheses and are bootstrapped and clustered at the individual level. We observe that a wage gap exists between blacks and whites in each percentile of the distribution.

The gap is signicant and large throughout the distribution but reaches its

peak between the 60th and 70th percentiles.

For all percentiles, non-cognitive skills (self

esteem, internal control, alienation and anomie and motivation) are signicant in explaining the wage observed.

In the 90th percentile, the magnitude explained by cognitive skills

(AFQT) does not account for a signicant portion of the wage gap. Observed characteristics explain more of the wage gap at lower points in the distribution and explain less towards the upper tail, in the 80th and 90th percentiles. Wage gaps at the 90th percentile are largely attributed to unexplained factors related to cognitive skills and experience. For blacks in the 10th percentile, cognitive skills explain 59.5% of the observed wage gap whereas non-

17 Since the sample is restricted to those reporting wages, there might be selection issues related to labor force participation.

18 In controlling for education, I assume that education is exogenous. This might not actually be the case

because skills and education could be endogenous. That is, there is reason to think that skills play a role in education decisions and vice versa. Controlling for years of education in addition to skills might introduce multicolinearity problems.

14

cognitive skills explain 27.1%. In the 50th percentile, the portion of the wage gap explained by cognitive skills falls to 36.5% and 17.8% of the wage gap by non-cognitive skills. In the 90th percentile, cognitive skills hardly explain the wage gap, only 3.66% while non-cognitive skills still explain 20.5% of the observed wage gap. The same results for Hispanics and whites are reported in Tables 4 and 5. The wage gap within the percentiles is smaller between whites and Hispanics than the wage gap between whites and blacks. For Hispanics, dierences in observed characteristics explain most of the gap at each point in the distribution. of the gap.

Non-cognitive skills signicantly explain a portion

Cognitive skills seem to matter less at the top of the wage distribution, as

was the case when blacks and whites were compared. Across the board, not much within each quantile is left unexplained by observed characteristics. This might be evidence that Hispanics look more similar to whites in terms of their observed characteristics and wages than blacks do. In the 10th percentile, cognitive skills explain 66.3% of the wage gap whereas non-cognitive skills only explain 27.2% of the gap.

For Hispanics and whites in the 50th

percentile, cognitive and non-cognitive skills explain 52.3% and 28.3% of the wage gap, respectively. Finally, in the 90th percentile, cognitive skills only explain 7% of the wage gap whereas non-cognitive skills explain 30% of the wage gap. This pattern is similar to that observed between whites and blacksnon-cognitive skills seem to explain 26-33% of the wage gap throughout the distribution whereas the portion of the gap explained by cognitive skills decreases substantially in the right tail of the distribution.

19

Finally, Tables 6 and 7 report the results for Hispanics and blacks.

The wage gap is

signicant throughout the distribution, with Hispanics earning more than blacks. The gap is largest in the 60th, 70th, and 80th percentiles of the distribution. Throughout the distribution, cognitive and non-cognitive skills help explain some of the wage gap. However, the impact is not signicant in all quantiles but some exceptions exist.

In particular, in

the 10th percentile, non-cognitive skills signicantly explain 29.8% of the wage gap between blacks and Hispanics. In the 40th and 50th percentiles, non-cognitive skills explain 18.9% and 19.0% of the wage gap, respectively. Cognitive skills explain a signicant portion of the black-Hispanic wage gap except at the tails of the distribution (the 10th and 90th percentiles). Throughout the distribution, the total wage gap that is unexplained by the controls is large and signicant. This is dierent from the Hispanic-white and black-white unconditional quantile decompositions where the majority of the wage gap is signicantly explained by the controls.

19 This could be at least partially explained by clumping around the maximum score of the test. That is, if a lot of people towards the upper end of the wage distribution all have the top scores in the measures of cognitive and non-cognitive skills, there is likely to be less variation and thus less explanatory power towards the wage gap.

15

5.2

Reweighted Unconditional Quantile Results

Tables 8 and 9 report results for the 10th through 90th percentiles of the log wage distribution where observations are reweighted by the DFL counterfactual reweighing factor. Predictions use FFL-OB combined with the reweighting factor. This allows for making inferences about how dierent parts of the wage distribution would look if whites were rewarded for having the distribution of skills of blacks. That is, for whites, the predicted wages are the distribution of wages that would be observed if whites had the distribution of skills of blacks. For blacks, the predicted wages give the wages that would be predicted for blacks conditional on their skills.

Like before, all specications control for years of education, cognitive skills, non-

cognitive skills, whether an individual resides in a city, region of residence, and a cubic in potential experience. When white wages are predicted based on black skills, a wage gap is still anticipated throughout the majority of the distribution. However, the predicted gap is smaller than the observed wage gap throughout the distribution. In addition, lower wages are predicted throughout the distribution, across whites and blacks. This is likely because the predicted wages under the reweighting method shift individuals around in the distribution based on their skills, changing their rank in the wage distribution. This predicted wage gap is large and signicant beginning in the 40th percentile of the distribution and persisting through the rest of the right tail.

However, at points below the 40th percentile, dierent

patterns emerge. In particular, in the bottom 10th percentile, wages would actually be higher for whites having black skills than their given skills. This implies that whites in this part of the distribution, conditional on having the same skills as blacks, would earn less than blacks in the bottom 10th percentile. Especially in the lower portion of the distribution, cognitive and non-cognitive skills help explain large portions of the predicted wage gaps. Tables 10 and 11 reports the same specications for Hispanics and whites. These results give the predicted distribution of wages for whites if they had skills of Hispanics. For Hispanics, these results give predicted distribution of wages given their distribution of skills. The predicted wage gap is much smaller than the actual wage gap throughout the distribution. In addition, the predicted wage gap is only (slightly) signicant in the 50th percentile. The majority of the gap is once again explained by observed characteristics. Non-cognitive skills explain a signicant portion across all quantiles. The portion explained by cognitive skills is large and signicant except for those quantiles larger than the 70th percentile. Education is highly signicant in explaining the wage gap throughout the distribution. These results suggest that if whites had the same skills as Hispanics, the wage gap would be smaller throughout the distribution. These results also suggest that not much of the predicted gap between whites and Hispanics is left as unexplained by observed characteristics. Of the small gap observed under reweighting, a signicant portion is explained by observed char-

16

acteristics. Most notably, non-cognitive skills matter throughout the whole distribution and cognitive skills matter through the 60th percentile. Tables 12 and 13 present the same specications for blacks and Hispanics. These results give the predicted wages for Hispanics if they had the same distributions of skills as blacks. For blacks, it predicts wages conditional on the distributions of skills for individuals classied into that group. Reweighting observations predicts a larger gap than the gap that is observed. This suggests that Hispanics would earn higher wages if they had the same distributions of skills as blacks.

Throughout the predicted distribution, non-cognitive skills signicantly

explain a large portion of the wage gap. Taken together, evidence from predicted wages suggests that conditional on skills, wage gaps between whites and blacks and whites and Hispanics should be smaller than actually observed.

In particular, in the bottom 10th percentile, reweighting predicts that blacks

should earn more than their white counterparts, conditional on skills. Finally, reweighting predicts a slightly larger than actual wage gap between blacks and Hispanics throughout the distribution, although this predicted gap is not signicant.

6

Discussion

In this paper, I use unconditional quantile Oaxaca-Blinder decompositions to investigate cognitive and non-cognitive skills as an explanation of the wage gap observed throughout the log hourly wage distribution. I nd that the wage gap is largest around the 50th percentile between whites and blacks, whites and Hispanics, and blacks and Hispanics. In addition, I nd that cognitive skills do not add much to explaining wage gap at highest percentiles of the distribution for these comparisons. More specically, I nd that non-cognitive skills explain 16-27% of the wage gap throughout the distribution for blacks and 26-33% of the wage gap throughout the distribution for Hispanics while the portion of the gap explained by cognitive skills decreases substantially in the right tail of the distribution for both comparisons. Between blacks and Hispanics, non-cognitive skills explain 29.8% of the wage gap in the 10th percentile and 18.9% to 19% in the 40th and 50th percentiles. I also use reweighted unconditional quantile Oaxaca-Blinder decompositions to predict the wage gap within quantiles conditional on the distribution of skills. This method estimates the distribution of wages for whites if they had the distribution of skills of blacks and predicts wages across the distribution of blacks conditional on the whole distribution of skills. Similar results are presented for whites and Hispanics and blacks and Hispanics.

Evidence from

predicted wages suggests that conditional on skills and education, the wage gaps between blacks and whites should be smaller than it actually is throughout the distribution. Likewise,

17

the reweighted estimates predict a smaller wage gap between whites and Hispanics, especially in the bottom half of the wage distribution.

When the wage distribution is predicted for

Hispanics and blacks, a larger gap is predicted than the observed gap and the majority of this gap can be explained by cognitive and non-cognitive skills. This has important implications for policies promoting equal pay: policy promoting equal wages conditional on skills and education at the bottom of the distribution might make a dierence in the lives of black workers.

References Altonji, J. G. and R. M. Blank (1999). Chapter 48 race and gender in the labor market. Volume 3, Part C of Handbook of Labor Economics, pp. 3143  3259. Elsevier. Blascovich, J. and J. Tamaka (1991). Measures of self esteem. In P. R. S. John P. Robinson and L. S. Wrightsman (Eds.), Measures of Personality and Social Psychological Attitudes, Volume 1, pp. 115160. Academic Press. Bowles, S. and H. Gintis (1976). Schooling in Capitalist America. Basic Books. Bowles, S. and H. Gintis (2002). Schooling in capitalist america revisited. Sociology Educa-

tion 75, 118. Cain, G. G. (1987).

The economic analysis of labor market discrimination: A survey.

In

O. Ashenfelter and R. Layard (Eds.), Handbook of Labor Economics (1 ed.), Volume 1, Chapter 13, pp. 693781. Elsevier. Carneiro, P., C. Crawford, and A. Goodman (2007).

The impact of early cognitive and

non-cognitive skills on later outcomes. Open Access publications from University College London http://discovery.ucl.ac.u, University College London. Carneiro, P. and J. J. Heckman (2003, July). Human capital policy. IZA Discussion Papers 821, Institute for the Study of Labor (IZA). Carneiro, P., J. J. Heckman, and D. V. Masterov (2005, April). Labor market discrimination and racial dierences in premarket factors. Journal of Law and Economics 48 (1), 139. Christie, R. (1991). Locus of control. In P. R. S. John P. Robinson and L. S. Wrightsman (Eds.), Measures of Personality and Social Psychological Attitudes, Volume 1, pp. 413572. Academic Press.

18

Cunha, F. and J. Heckman (2007). The technology of skill formation. The American Eco-

nomic Review 97 (2), pp. 3147. DiNardo, J., N. M. Fortin, and T. Lemieux (1996, September). Labor market institutions and the distribution of wages, 1973-1992: A semiparametric approach. Econometrica 64 (5), 100144. Farkas, G. (2003).

Cognitive skills and noncognitive traits and behaviors in stratication

processes. Annual Review of Sociology 29, pp. 541562. Firpo, S. (2007, 01).

Ecient Semiparametric Estimation of Quantile Treatment Eects.

Econometrica 75 (1), 259276. Firpo, S., N. M. Fortin, and T. Lemieux (2007, July). Unconditional quantile regressions. Working Paper 339, National Bureau of Economic Research. Fortin, N., T. Lemieux, and S. Firpo (2010, June). Decomposition methods in economics. Working Paper 16045, National Bureau of Economic Research. Fryer, R. G. (2010, August). Racial Inequality in the 21st Century: The Declining Signicance of Discrimination. NBER Working Papers 16256, National Bureau of Economic Research, Inc. Fryer, R. G. and S. D. Levitt (2004, May). Understanding the black-white test score gap in the rst two years of school. The Review of Economics and Statistics 86 (2), 447464. Fryer, R. G., D. Pager, and J. L. Spenkuch (2013). Racial Disparities in Job Finding and Oered Wages. Journal of Law and Economics 56 (3), 633  689. Grenier, G. (1984). The eects of language characteristics on the wages of hispanic-american males. The Journal of Human Resources 19 (1), pp. 3552. Heckman, J. (1979).

Sample selection bias as a specication error.

Econometrica 47 (1),

15361. Heckman, J. J., J. Hsee, and Y. Rubinstein (2001, April). The ged as a mixed signal. Working paper, American Economic Association. Heckman, J. J. and Y. Rubinstein (2001). The importance of noncognitive skills: Lessons from the ged testing program. The American Economic Review 91 (2), pp. 145149.

19

Heckman, J. J., J. Stixrud, and S. Urzua (2006, July). The eects of cognitive and noncognitive abilities on labor market outcomes and social behavior. Journal of Labor Econom-

ics 24 (3), 411482. Hirano, K., G. W. Imbens, and G. Ridder (2003). Ecient estimation of average treatment eects using the estimated propensity score. Econometrica 71 (4), 11611189. Kuhn, P. and C. Weinberger (2005). Leadership skills and wages. Journal of Labor Econom-

ics 23 (3), pp. 395436. Lindqvist, E. and R. Vestman (2011). The labor market returns to cognitive and noncognitive ability:

Evidence from the swedish enlistment.

American Economic Journal: Applied

Economics 3 (1), pp. 101128. Lleras, C. (2008). Do skills and behaviors in high school matter? the contribution of noncognitive factors in explaining dierences in educational attainment and earnings. Social

Science Research 37 (3), 888  902. Lundberg, S. (2013). The College Type: Personality and Educational Inequality. Journal of

Labor Economics 31 (3), 421  441. Lundberg, S. (2014, July). Skill Disparities and Unequal Family Outcomes. IZA Discussion Papers 8346, Institute for the Study of Labor (IZA). Murnane, R. J., J. B. Willett, M. Braatz, and Y. Duhaldeborde (2001).

Do dierent di-

mensions of male high school students' skills predict labor market success a decade later? evidence from the nlsy. Economics of Education Review 20 (4), 311  320. Neal, D. A. and W. R. Johnson (1996, October). The role of premarket factors in black-white wage dierences. Journal of Political Economy 104 (5), 86995. Oettinger, G. S. (1996). Statistical discrimination and the early career evolution of the blackwhite wage gap. Journal of Labor Economics 14 (1), pp. 5278. Reimers, C. W. (1983). Labor market discrimination against hispanic and black men. The

Review of Economics and Statistics 65 (4), pp. 570579. Seeman, M. (1991). Alienation and anomie. In P. R. S. John P. Robinson and L. S. Wrightsman (Eds.), Measures of Personality and Social Psychological Attitudes, Volume 1, pp. 291372. Academic Press. Segal, C. (2008). Classroom behavior. Journal of Human Resources 43 (4).

20

Segal, C. (2012). Working when no one is watching: Motivation, test scores and economic success. Management Science 58 (8), 14381457. Tsai, J. (2007). Decoding the ged: The role of noncognitive ability and measurement error.

unpublished . Urzua, S. (2008).

Racial labor market gaps:

The role of abilities and schooling choices.

Journal of Human Resources 43 (4), 919971.

A

Non-cognitive Tests

A.1

The Rotter Locus of Control Scale Questions

There are pairs: internal and external item. 1. What happens to me is my own doing. (Internal) Sometimes I feel that I don't have enough control over the direction my life is taking. (External) 2. When I make plans, I am almost certain that I can make them work out. (Internal) It is not wise to plan too far ahead, because many things turn out to be a matter of good or bad fortune anyhow. (External) 3. In many cases, getting what I want has little or nothing to do with luck. (Internal) Many times, we might just as well decide what to do by ipping a coin. (External) 4. It is impossible for me to believe that chance or luck plays an important role in my life. (Internal) Many times I feel that I have little inuence over the things that happen to me. (External)

A.2

The Rosenberg Self-Esteem Scale Questions

1. I am a person of worth. 2. I have a number of good qualities. 3. I am inclined to feel that I am a failure.

21

4. I am as capable as others. 5. I feel I do not have much to be proud of. 6. I have a positive attitude. 7. I am satised with myself. 8. I wish I had more self respect. 9. I feel useless at times. 10. I sometimes think I am no good at all.

A.3

The Pearlin Mastery Scale Questions

1. I am a person of worth. 2. I have a number of good qualities. 3. I am inclined to feel that I am a failure. 4. I am as capable as others. 5. I feel I do not have much to be proud of. 6. I have a positive attitude. 7. I am satised with myself. 8. I wish I had more self respect. 9. I feel useless at times. 10. I sometimes think I am no good at all.

22

A.4

The Coding Speed Test

Figure 1: Sample Coding Speed Question

This is a sample of the coding speed test from the ASVAB, administered to all individuals in the NLSY79.

B

Oaxaca-Blinder Decomposition

Oaxaca-Blinder is a method for decomposing variances of the wage gap between two groups into the portion explained by characteristics (within group, composition eects) and that unexplained by characteristics (between group, wage structure eects). conditional expectation of wages,

Y,

given individual characteristics,

E(Y

|X) = Xβ

Then, the law of iterated expectations gives:

E(Y

) = E [E(Y |X)] = E(X)β

We can think similarly about the unconditional variance of

20 See Fortin et al. (2010) for further details.

23

Y:

20

X,

Assuming that the is linear gives:

var(Y

  |X)] + E [E(Y |X) − E(Y )]2   = E [var(Y |X)] + E [Xβ − E(X)β]2

) =

E [var(Y

=

E [var(Y

|

|X)] + {z }

within group

β |

0

var(X)β

{z

}

between group

where expectations are taken over the distribution of

X

and variance is decomposed into

the portion explained within groups and between groups. Suppose for simplicity that

b=

 1

if an individual is black

0

if an individual is not black or Hispanic (white)

Then, we can decompose the dierences in wages (at the mean) for groups using the unconditional variance as derived above.

0

0

∆v = E [varb=1 (X|b = 1)] − E [varb=0 (X|b = 0)]+βb=1 var(X|b = 1)βb=1 − βb=0 var(X|b = 0)βb=0 {z } | {z } | explained, composition eects

unexplained, wage structure eects

This gives the portion of the dierence explained (composition eects) and unexplained

βb=1 and βb=0 can be estimated easily, but the conditional variances, varb=0 (X|b = 0) are harder to back out. This standard OB decompos-

(wage structure eects). varb=1 (X|b

= 1)

and

ition only tells us about the wage gap at the mean. However, since we observe a wage gap throughout the distribution, it would be useful to understand if observed characteristics (and in particular, skills) helped explain the wage gap in quantiles.

21

The problem here is that the





law of iterated expectations does not hold within quantiles. That is, Qb,τ 6= EX Qb,τ (X) th where Qb,τ is the τ quantile of the unconditional distribution of Yb and Qb,τ (X) is the conditional quantile. We need to know the whole conditional distribution of compute Qb,τ . Specically, let

τ

Yb

given

X

to

be the conditional quantile:

ˆ   τ = FYb (Qb,τ ) = E FYb |Xb (Qb,τ |X) =

FYb |Xb (·) is the cumulative density of Y conditional on X in group b. of τ , we can solve for Qb,τ . This is where FFL becomes helpful.

where a choice

FYb |Xb (Qb,τ |X)dFXb (X) Then, given

21 The Oaxaca-Blinder approach is also limited by the choice of reference group and the linear specication.

24

C

Tables and Figures

Figure 2: Log Hourly Wages

This plot gives the density of the log of hourly wages for individuals. The data include the years 1979-2004 from the NLSY79.

Figure 3: AFQT

This plot gives the density of the standardized AFQT scores for individuals.

Scores are standardized by

birth year cohorts. This is considered a cognitive skill measure. The data include years 1979-2004 from the NLSY79.

25

Figure 4: Standardized Rotter

This plot gives the density of the standardized Rotter scores for individuals. The Rotter score measures the degree of internal control possessed by an individual and was administered in 1980.

This is considered a

non-cognitive skill measure and is standardized by birth year cohorts. The data include the years 1979-2004 from the NLSY79.

Figure 5: Rosenberg Density

This plot gives the density of the standardized Rosenberg scores for individuals.

The Rosenberg score

measures an individual's self esteem and was administered in 1980. This is considered a non-cognitive skill measure and is standardized by birth year cohorts. The data include the years 1979-2004 from the NLSY79.

26

Figure 6: Standardized Pearlin

This plot gives the density of the standardized Pearlin scores for individuals. The Pearlin score measures alienation and anomie, the subjective sense of powerlessness and state of meaningless than an individual possesses and was administered in 1992. This is considered a non-cognitive skill measure and is standardized by birth year cohorts. The data include the years 1979-2004 from the NLSY79.

Figure 7: Standardized Coding Speed

This plot gives the density of the standardized coding speed scores for individuals. The coding speed score represents a measure of motivation and was administered in 1979. This is considered a non-cognitive skill measure and is standardized by birth year cohorts. The data include the years 1979-2004 from the NLSY79.

27

Table 1: Basic Summary Statistics by Race Total

Whites

Blacks

Hispanics

Observations

41950

24082

10897

6971

Individuals

3738

2156

1008

577

57.41

25.98

16.62

79.27

73.60

83.78

91.81

Percentage

Urban residence (%) Region (%) Northeast

17.57

19.53

14.62

15.40

North Central

24.78

33.46

16.99

6.95

South

38.17

30.15

60.96

30.23

West

19.48

16.86

7.42

47.42

Log of real wage Ages <25

6.56

6.59

6.46

6.56

Ages 25-30

6.81

6.87

6.66

6.80

Ages 30-35

6.94

7.04

6.75

6.92

Ages >35

7.06

7.17

6.85

7.04

3.27

3.22

3.33

3.38

Potential Experience Ages <25 Ages 25-30

7.25

7.05

7.52

7.56

Ages 30-35

11.73

11.48

12.07

12.05

Ages >35

16.79

16.57

16.99

17.15

Mean

0

0.39

-0.73

-0.22

SD

1.00

0.85

0.91

0.91

AFQT

Rotter Score

11.44

11.69

11.21

10.96

Standardized Rotter Score

0

0.10

-0.10

-0.19

Std Deviation

1.00

1.00

.96

1.00

Rosenberg Score

22.75

22.94

22.69

22.19

Standardized Rosenberg Score

0

0.04

-0.01

-0.12

Std. Deviation

1.00

1.00

1.03

0.97

Coding Speed

40.35

44.52

31.66

39.59

Standardized Coding Speed

0

0.27

-0.57

-0.03

Std. Deviation

1.00

0.93

0.96

0.92

Highest Grade Completed

13.15

13.47

12.80

12.61

The data include the years 1979-2004 from the NLSY79. Only individuals with more than 8 years of schooling are included. Potential experience is dened as age minus years of schooling minus 6.

28

29

6.122***

0.173***

Mean Black Log Wage

RIF Wage Gap

Urban, Region

0.151***

0.0217

-0.185

30,156

Total Unexplained

Constant

Observations

(0.131)

(0.0311)

(0.0480)

(0.0455)

(0.0107)

(0.0216)

(0.111)

(0.0257)

(0.00250)

(0.00292)

(0.0120)

(0.0270)

(0.00426)

(0.0146)

(0.00990)

(0.00973)

30,156

-0.128

0.0754***

0.373***

-0.0437

-0.0246*

-0.0196

-0.0817

0.149***

-0.0174***

-0.00459

0.0574***

0.0882***

0.0251***

0.224***

6.284***

6.508***

(0.139)

(0.0266)

(0.0415)

(0.0523)

(0.0126)

(0.0206)

(0.104)

(0.0194)

(0.00322)

(0.00344)

(0.0130)

(0.0212)

(0.00485)

(0.0166)

(0.0129)

(0.0101)

20th percentile

30,156

-0.0563

0.105***

0.350***

-0.0408

-0.0268*

-0.00915

-0.112

0.160***

-0.0181***

-0.00462

0.0606***

0.0909***

0.0308***

0.265***

6.409***

6.674***

(0.130)

(0.0241)

(0.0377)

(0.0599)

(0.0139)

(0.0213)

(0.109)

(0.0230)

(0.00307)

(0.00355)

(0.0121)

(0.0230)

(0.00549)

(0.0173)

(0.0151)

(0.0103)

30th percentile

30,156

0.0438

0.109***

0.248***

-0.0576

-0.0330**

0.0102

-0.102

0.164***

-0.0182***

-0.00440

0.0578***

0.0922***

0.0368***

0.273***

6.538***

6.811***

(0.133)

(0.0248)

(0.0355)

(0.0631)

(0.0155)

(0.0227)

(0.119)

(0.0216)

(0.00326)

(0.00347)

(0.0140)

(0.0232)

(0.00627)

(0.0199)

(0.0156)

(0.0122)

40th percentile

30,156

0.0661

0.104***

0.203***

-0.0709

-0.0361**

0.0308

-0.0890

0.170***

-0.0173***

-0.00551

0.0487***

0.100***

0.0440***

0.274***

6.669***

6.943***

(0.160)

(0.0251)

(0.0387)

(0.0649)

(0.0150)

(0.0243)

(0.135)

(0.0196)

(0.00325)

(0.00369)

(0.0133)

(0.0196)

(0.00739)

(0.0214)

(0.0154)

(0.0147)

50th percentile

more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years of schooling minus 6.

Esteem Score, Pearlin Mastery Scale and the Coding Speed Test. The data include the years 1979-2004 from the NLSY79. Only individuals with

The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus of Control Score, Rosenberg Self

This table gives the Oaxaca-Blinder decompositions of the black white wage gap observed in unconditional quantiles of the log hourly wage distributions.

-0.0828*

0.361***

-0.0161

Non-cognitive Skills

Cubic in experience

-0.0408*

Cognitive Skills

Urban, Region

-0.0142

Education

Unexplained by characteristics

Total Explained

-0.0136***

-0.00169

Non-cognitive Skills

Cubic in experience

0.103***

0.0469***

Cognitive Skills

0.0161***

Education

Explained by characteristics

6.295***

Mean White Log Wage

10th percentile

Comparison Group: Black Males

Reference Group: White Males

Decomposition Method: RIF regressions without reweighing

Table 2: Oaxaca Decomposition on Unconditional Quantiles Explaining Black White Wage Gap Throughout the Distribution

30

0.302***

Mean Black Log Wage

RIF Wage Gap

30,156

Observations

(0.0129)

(0.158)

(0.0237)

(0.0338)

(0.0651)

(0.0152)

(0.0244)

(0.145)

(0.0193)

(0.00284)

(0.00361)

(0.0138)

(0.0209)

(0.00743)

(0.0198)

(0.0172)

30,156

0.00720

0.145***

0.0869**

0.0224

-0.0329*

0.0867***

-0.0255

0.144***

-0.0152***

-0.00837**

0.0544***

0.0551***

0.0581***

0.289***

6.939***

7.227***

(0.184)

(0.0275)

(0.0406)

(0.0788)

(0.0168)

(0.0272)

(0.170)

(0.0198)

(0.00332)

(0.00345)

(0.0152)

(0.0192)

(0.00948)

(0.0254)

(0.0204)

(0.0132)

70th percentile

30,156

-0.111

0.132***

0.0758*

0.0936

-0.0208

0.113***

-0.0178

0.125***

-0.0145***

-0.00975***

0.0443***

0.0415**

0.0633***

0.257***

7.128***

7.385***

(0.220)

(0.0236)

(0.0418)

(0.0935)

(0.0193)

(0.0306)

(0.201)

(0.0192)

(0.00317)

(0.00354)

(0.0125)

(0.0195)

(0.0109)

(0.0235)

(0.0214)

(0.0126)

80th percentile

30,156

-0.0800

0.144***

0.135***

0.211**

-0.0317

0.106***

-0.197

0.107***

-0.0157***

-0.00864***

0.0516***

0.00922

0.0710***

0.252***

7.367***

7.618***

(0.234)

(0.0314)

(0.0508)

(0.0997)

(0.0210)

(0.0282)

(0.214)

(0.0224)

(0.00353)

(0.00308)

(0.0174)

(0.0235)

(0.0107)

(0.0273)

(0.0243)

(0.0146)

90th percentile

more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years of schooling minus 6.

Esteem Score, Pearlin Mastery Scale and the Coding Speed Test. The data include the years 1979-2004 from the NLSY79. Only individuals with

The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus of Control Score, Rosenberg Self

This table gives the Oaxaca-Blinder decompositions of the black white wage gap observed in unconditional quantiles of the log hourly wage distributions.

0.0918

0.136***

Total Unexplained

Constant

0.149***

Cubic in experience

Non-cognitive Skills -0.0319

-0.0369**

Cognitive Skills

Urban, Region

-0.0841

0.0485**

Education

Unexplained by characteristics

0.166***

-0.0165***

Cubic in experience

Total Explained

0.0472***

-0.00737**

Urban, Region

Cognitive Skills

Non-cognitive Skills

0.0514***

0.0908***

Education

Explained by characteristics

7.086***

6.784***

Mean White Log Wage

60th percentile

Comparison Group: Black Males

Reference Group: White Males

Decomposition Method: RIF regressions without reweighing

continued

Table 3: Oaxaca Decomposition on Unconditional Quantiles Explaining Black White Wage Gap Throughout the Distribution,

31

6.214*** 0.0816***

Mean Hispanic Log Wage

RIF Wage Gap

Non-cognitive Skills

0.0944***

Total Explained

-0.0128

-0.166

26,684

Cubic in experience

Total Unexplained

Constant

Observations

(0.157)

(0.0230)

(0.0558)

(0.0810)

(0.0110)

(0.0140)

(0.125)

(0.0164)

(0.00252)

(0.00856)

(0.00534)

(0.0134)

(0.00516)

(0.0191)

(0.0182)

(0.0100)

26,684

-0.00605

0.00991

0.0546

-0.0203

-0.00927

-0.00473

-0.00432

0.0991***

-0.00574*

-0.00219

0.0292***

0.0452***

0.0326***

0.109***

6.399***

6.508***

(0.167)

(0.0239)

(0.0595)

(0.0808)

(0.00927)

(0.0188)

(0.137)

(0.0157)

(0.00331)

(0.00790)

(0.00693)

(0.0123)

(0.00586)

(0.0215)

(0.0188)

(0.0115)

20th percentile

26,684

0.000436

0.00245

0.0485

0.0337

-0.00939

0.000988

-0.0718

0.112***

-0.00598

-0.00101

0.0330***

0.0465***

0.0398***

0.115***

6.559***

6.674***

(0.160)

(0.0267)

(0.0451)

(0.0809)

(0.0102)

(0.0185)

(0.122)

(0.0165)

(0.00366)

(0.00780)

(0.00748)

(0.0119)

(0.00706)

(0.0249)

(0.0237)

(0.0121)

30th percentile

26,684

-0.0514

-0.0191

-0.00557

0.00931

-0.00587

-0.00116

0.0356

0.123***

-0.00594

0.000595

0.0340***

0.0472***

0.0471***

0.104***

6.708***

6.811***

(0.160)

(0.0267)

(0.0465)

(0.0947)

(0.0102)

(0.0166)

(0.125)

(0.0193)

(0.00376)

(0.00927)

(0.00716)

(0.0116)

(0.00889)

(0.0216)

(0.0213)

(0.0109)

40th percentile

26,684

-0.207

-0.0309

-0.0368

0.0113

-0.00128

0.000261

0.202

0.130***

-0.00560*

-0.000417

0.0282***

0.0522***

0.0561***

0.0995***

6.844***

6.943***

(0.171)

(0.0217)

(0.0450)

(0.0871)

(0.0115)

(0.0179)

(0.134)

(0.0181)

(0.00336)

(0.00966)

(0.00852)

(0.0119)

(0.00809)

(0.0232)

(0.0206)

(0.0113)

50th percentile

Only

6.

individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years of schooling minus

The data include the years 1979-2004 from the NLSY79.

The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus of Control Score,

Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test.

distributions.

This table gives the Oaxaca-Blinder decompositions of the Hispanic white wage gap observed in unconditional quantiles of the log hourly wage

-0.0163 0.151***

Urban, Region

0.00189

Non-cognitive Skills

0.0436 -0.0262*

Cognitive Skills

Education

Unexplained by characteristics

-0.00448*

Cubic in experience

0.00142

0.0222***

Cognitive Skills

Urban, Region

0.0211*** 0.0541***

Education

Explained by characteristics

6.295***

Mean White Log Wage

10th percentile

Comparison Group: Hispanic Males

Reference Group: White Males

Decomposition Method: RIF regressions without reweighing

Table 4: Oaxaca Decomposition on Unconditional Quantiles Explaining Hispanic White Wage Gap Throughout the Distribution

32

6.983*** 0.103***

Mean Hispanic Log Wage

RIF Wage Gap

-0.343*

26,684

Total Unexplained

Constant

Observations

(0.176)

(0.0251)

(0.0569)

(0.0829)

(0.0115)

(0.0211)

(0.156)

(0.0195)

(0.00347)

(0.00929)

(0.00698)

(0.0125)

(0.00968)

(0.0272)

(0.0219)

(0.0131)

26,684

-0.358*

-0.0200

-0.0443

-0.0539

0.00117

0.0243

0.411**

0.113***

-0.00484

-0.0116

0.0270***

0.0291**

0.0736***

0.0932***

7.134***

7.227***

(0.201)

(0.0309)

(0.0487)

(0.102)

(0.0129)

(0.0230)

(0.177)

(0.0158)

(0.00371)

(0.00906)

(0.00792)

(0.0118)

(0.0109)

(0.0297)

(0.0261)

(0.0124)

70th percentile

26,684

-0.196

-0.0329

-0.0315

-0.134

0.00647

0.0151

0.307

0.103***

-0.00462

-0.0180**

0.0230***

0.0228*

0.0796***

0.0699**

7.315***

7.385***

(0.228)

(0.0257)

(0.0542)

(0.105)

(0.0126)

(0.0203)

(0.191)

(0.0190)

(0.00372)

(0.00869)

(0.00818)

(0.0128)

(0.0119)

(0.0283)

(0.0252)

(0.0130)

80th percentile

26,684

-0.398

-0.0119

0.0594

-0.0420

-0.00160

0.0164

0.354

0.0986***

-0.00502

-0.0174

0.0260***

0.00608

0.0890***

0.0868***

7.531***

7.618***

(0.262)

(0.0278)

(0.0467)

(0.115)

(0.0125)

(0.0167)

(0.228)

(0.0195)

(0.00429)

(0.0109)

(0.00801)

(0.0127)

(0.0145)

(0.0281)

(0.0242)

(0.0146)

90th percentile

Only

6.

individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years of schooling minus

The data include the years 1979-2004 from the NLSY79.

The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus of Control Score,

Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test.

distributions.

This table gives the Oaxaca-Blinder decompositions of the Hispanic white wage gap observed in unconditional quantiles of the log hourly wage

-0.0190 -0.0258

Cubic in experience

0.00236

Urban, Region

-0.000477

0.0129

Cognitive Skills

Non-cognitive Skills

0.322**

Education

Unexplained by characteristics

-0.00531

-0.00579

Urban, Region

0.128***

0.0265***

Non-cognitive Skills

Total Explained

0.0474***

Cognitive Skills

Cubic in experience

0.0655***

Education

Explained by characteristics

7.086***

Mean White Log Wage

60th percentile

Comparison Group: Hispanic Males

Reference Group: White Males

Decomposition Method: RIF regressions without reweighing

continued

Table 5: Oaxaca Decomposition on Unconditional Quantiles Explaining Hispanic White Wage Gap Throughout the Distribution,

33

0.0912***

Mean Black Log Wage

RIF Wage Gap

0.00380

Urban, Region

0.0405**

-0.00139

Constant

(0.0151)

(0.172)

(0.0269)

(0.0617)

(0.0662)

(0.00660)

(0.00651)

(0.127)

(0.0204)

(0.00241)

(0.00810)

(0.0156)

(0.0137)

(0.00301)

(0.0180)

(0.0112)

16,230

-0.110

0.0701**

0.320***

-0.0290

-0.00674

-0.00357

-0.101

0.0451*

-0.0104***

0.00415

0.0221

0.0355**

-0.00631

0.115***

6.284***

6.399***

(0.168)

(0.0282)

(0.0517)

(0.0774)

(0.00781)

(0.00683)

(0.146)

(0.0238)

(0.00351)

(0.00947)

(0.0167)

(0.0164)

(0.00511)

(0.0217)

(0.0127)

(0.0161)

20th percentile

16,230

-0.0527

0.106***

0.304***

-0.0739

-0.00845

-0.00277

-0.0605

0.0445**

-0.0113***

-0.00302

0.0250

0.0426**

-0.00869

0.150***

6.409***

6.559***

(0.195)

(0.0302)

(0.0549)

(0.0911)

(0.00844)

(0.00941)

(0.178)

(0.0189)

(0.00375)

(0.0100)

(0.0172)

(0.0199)

(0.00678)

(0.0268)

(0.0163)

(0.0210)

30th percentile

16,230

0.105

0.121***

0.257***

-0.0676

-0.0144*

0.00381

-0.162

0.0486***

-0.0120***

-0.00229

0.0324**

0.0392**

-0.00878

0.170***

6.538***

6.708***

(0.195)

(0.0240)

(0.0453)

(0.0948)

(0.00759)

(0.00716)

(0.163)

(0.0188)

(0.00431)

(0.0103)

(0.0147)

(0.0176)

(0.00535)

(0.0244)

(0.0144)

(0.0215)

40th percentile

16,230

0.285*

0.120***

0.243***

-0.0837

-0.0175*

0.00925

-0.316**

0.0551**

-0.0116***

-0.000597

0.0330*

0.0428**

-0.00851

0.175***

6.669***

6.844***

(0.162)

(0.0299)

(0.0487)

(0.0868)

(0.00912)

(0.00810)

(0.147)

(0.0261)

(0.00395)

(0.0105)

(0.0185)

(0.0173)

(0.00594)

(0.0281)

(0.0185)

(0.0216)

50th percentile

Only

6.

individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years of schooling minus

The data include the years 1979-2004 from the NLSY79.

The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus of Control Score,

Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test.

distributions.

This table gives the Oaxaca-Blinder decompositions of the black Hispanic wage gap observed in unconditional quantiles of the log hourly wage

16,230

0.0507*

Total Unexplained

Observations

-0.0732 0.210***

Non-cognitive Skills

Cubic in experience

-0.00767

Cognitive Skills

Urban, Region

-0.0771 6.65e-05

Education

Unexplained by characteristics

Total Explained

-0.00669***

0.0272*

Non-cognitive Skills

Cubic in experience

0.0195

-0.00334

Cognitive Skills

Education

Explained by characteristics

6.214*** 6.122***

Mean White Log Wage

10th percentile

Comparison Group: Black Males

Reference Group: Hispanic Males

Decomposition Method: RIF regressions without reweighing

Table 6: Oaxaca Decomposition on Unconditional Quantiles Explaining Black Hispanic Wage Gap Throughout the Distribution

34

6.784***

0.199***

Mean Black Log Wage

RIF Wage Gap

0.0653**

-0.0108***

0.00828

-0.0162* -0.0360

0.170***

0.134***

0.439**

16,230

Cognitive Skills

Non-cognitive Skills

Urban, Region

Cubic in experience

Total Unexplained

Constant

Observations

(0.188)

(0.0291)

(0.0509)

(0.0992)

(0.00978)

(0.00958)

(0.175)

(0.0268)

(0.00389)

(0.0117)

(0.0192)

(0.0204)

(0.00655)

(0.0301)

(0.0182)

(0.0227)

16,230

0.397*

0.132***

0.134***

0.0685

-0.00900

0.0128

-0.472**

0.0637***

-0.0108***

0.0145

0.0196

0.0497***

-0.00935

0.196***

6.939***

7.134***

(0.226)

(0.0341)

(0.0511)

(0.105)

(0.0109)

(0.00871)

(0.184)

(0.0243)

(0.00387)

(0.0118)

(0.0193)

(0.0180)

(0.00585)

(0.0288)

(0.0183)

(0.0261)

70th percentile

16,230

0.154

0.135***

0.110*

0.211*

-0.00303

0.0227**

-0.360

0.0517*

-0.0105***

0.0285**

0.0106

0.0352**

-0.0121

0.187***

7.128***

7.315***

(0.257)

(0.0347)

(0.0586)

(0.123)

(0.0105)

(0.0110)

(0.243)

(0.0267)

(0.00377)

(0.0126)

(0.0198)

(0.0167)

(0.00834)

(0.0314)

(0.0217)

(0.0219)

80th percentile

16,230

0.393

0.151***

0.0761

0.249**

0.000316

0.0198**

-0.587**

0.0136

-0.00888**

0.0167

-0.00230

0.0215

-0.0134

0.165***

7.367***

7.531***

(0.301)

(0.0359)

(0.0598)

(0.125)

(0.0103)

(0.00945)

(0.260)

(0.0272)

(0.00381)

(0.0124)

(0.0184)

(0.0176)

(0.0105)

(0.0333)

(0.0238)

(0.0242)

90th percentile

Only

6.

individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years of schooling minus

The data include the years 1979-2004 from the NLSY79.

The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus of Control Score,

Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test.

distributions.

This table gives the Oaxaca-Blinder decompositions of the black Hispanic wage gap observed in unconditional quantiles of the log hourly wage

-0.431**

Education

Unexplained by characteristics

Total Explained

Cubic in experience

0.0292 0.00343

Urban, Region

0.0523**

Cognitive Skills

Non-cognitive Skills

-0.00888

Education

Explained by characteristics

6.983***

Mean White Log Wage

60th percentile

Comparison Group: Black Males

Reference Group: Hispanic Males

Decomposition Method: RIF regressions without reweighing

continued

Table 7: Oaxaca Decomposition on Unconditional Quantiles Explaining Black Hispanic Wage Gap Throughout the Distribution,

35

0.0496**

Non-cognitive Skills

Total Explained

0.206

-0.384***

Cubic in experience

Total Unexplained

(0.622)

(0.0828)

(0.194)

(0.342)

(0.0976)

(0.234)

(0.524)

(0.0589)

(0.00604)

(0.00909)

(0.0244)

(0.0464)

(0.00468)

(0.0803)

(0.0611)

(0.0516)

30,156

0.233

-0.167***

0.104

-0.195

-0.0742

-0.170

-0.0638

0.188***

-0.0108

-0.000447

0.0544***

0.144***

0.000919

0.0208

6.192***

6.213***

(0.539)

(0.0661)

(0.186)

(0.294)

(0.101)

(0.177)

(0.420)

(0.0417)

(0.0074)

(0.0045)

(0.0185)

(0.0300)

(0.00224)

(0.0677)

(0.0571)

(0.0361)

20th percentile

30,156

-0.0458

-0.0387

0.151

-0.0839

-0.0479

-0.0491

0.0376

0.145***

-0.00930

-0.00358

0.0443***

0.111***

0.00220

0.106

6.297***

6.403***

(0.552)

(0.0626)

(0.167)

(0.279)

(0.105)

(0.172)

(0.414)

(0.0344)

(0.00679)

(0.00352)

(0.0159)

(0.0227)

(0.00508)

(0.0657)

(0.0585)

(0.0298)

30th percentile

30,156

0.0262

0.0230

0.0981

0.0690

-0.0197

-0.0524

-0.0981

0.122***

-0.00979

-0.00882***

0.0360***

0.102***

0.00236

0.145**

6.419***

6.564***

(0.593)

(0.0638)

(0.155)

(0.295)

(0.112)

(0.180)

(0.468)

(0.0306)

(0.00700)

(0.00318)

(0.0143)

(0.0190)

(0.00524)

(0.0674)

(0.0622)

(0.0259)

40th percentile

30,156

0.187

0.0593

0.111

0.0408

-0.0197

-0.0284

-0.231

0.105***

-0.00930

-0.00948***

0.0268***

0.0942***

0.00268

0.164***

6.541***

6.705***

(0.620)

(0.0632)

(0.147)

(0.286)

(0.117)

(0.186)

(0.510)

(0.0273)

(0.00646)

(0.00322)

(0.0115)

(0.0175)

(0.00604)

(0.0668)

(0.0626)

(0.0232)

50th percentile

minus 6.

Only individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years of schooling

Score, Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test. The data include the years 1979-2004 from the NLSY79.

reweighted using DFL. The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus of Control

This table gives the Oaxaca-Blinder decompositions of the black white wage gap observed in unconditional quantiles of the log hourly wage distributions

30,156

-0.492

Urban, Region

Observations

-0.0710

Non-cognitive Skills

0.914

-0.408*

Cognitive Skills

Constant

-0.533

Education

Unexplained by characteristics

-0.00950

0.274***

Cubic in experience

0.0116

0.225***

Cognitive Skills

Urban, Region

-0.00207

Education

Explained by characteristics

-0.110

6.054***

Mean Black Log Wage

RIF Wage Gap

5.944***

Mean White Log Wage

10th percentile

Comparison Group: Black Males

Reference Group: White Males

Decomposition Method: RIF regressions with reweighing

Distribution

Table 8: Oaxaca Decomposition on Reweighted Unconditional Quantiles Explaining Black White Wage Gap Throughout the

36

RIF Wage Gap

0.0896***

Total Explained

(0.0204)

(0.623)

(0.0623)

(0.146)

(0.289)

(0.122)

(0.189)

(0.519)

(0.0241)

(0.0056)

(0.00286)

(0.00905)

(0.0146)

(0.00796)

(0.0656)

(0.0624)

30,156

0.464

0.167***

0.00415

0.137

-0.0285

0.0311

-0.442

0.0777***

-0.00645

-0.00721***

0.0282***

0.0599***

0.00334

0.245***

6.785***

7.030***

(0.693)

(0.0659)

(0.137)

(0.269)

(0.142)

(0.221)

(0.602)

(0.0206)

(0.00427)

(0.00235)

(0.00918)

(0.0117)

(0.00741)

(0.0685)

(0.0662)

(0.0174)

70th percentile

30,156

0.262

0.192***

-0.0136

0.239

-0.0558

0.0682

-0.308

0.0600***

-0.00537

-0.00830***

0.0232***

0.0470***

0.00349

0.252***

6.968***

7.219***

(0.768)

(0.0667)

(0.152)

(0.304)

(0.131)

(0.212)

(0.665)

(0.0182)

(0.00374)

(0.00292)

(0.00823)

(0.0106)

(0.00778)

(0.0683)

(0.0665)

(0.0154)

80th percentile

30,156

0.196

0.206***

0.0404

0.272

-0.109

0.105

-0.299

0.0395**

-0.00560

-0.00947**

0.0300***

0.0208**

0.00381

0.246***

7.233***

7.478***

(0.905)

(0.0585)

(0.147)

(0.221)

(0.115)

(0.177)

(0.752)

(0.0178)

(0.00412)

(0.00423)

(0.0104)

(0.00964)

(0.00870)

(0.0595)

(0.0577)

(0.0145)

90th percentile

minus 6.

Only individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years of schooling

Score, Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test. The data include the years 1979-2004 from the NLSY79.

reweighted using DFL. The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus of Control

This table gives the Oaxaca-Blinder decompositions of the black white wage gap observed in unconditional quantiles of the log hourly wage distributions

30,156

0.122**

Total Unexplained

Observations

0.0778

Cubic in experience

0.339

0.0669

Urban, Region

Constant

-0.0250

Non-cognitive Skills

-0.332 -0.00486

Cognitive Skills

Education

Unexplained by characteristics

-0.00840

Cubic in experience

-0.00910***

0.0231***

Non-cognitive Skills

Urban, Region

0.0804***

Cognitive Skills

Education

0.00350

0.212***

Mean Black Log Wage

Explained by characteristics

6.863*** 6.651***

Mean White Log Wage

60th percentile

Comparison Group: Black Males

Reference Group: White Males

Decomposition Method: RIF regressions with reweighing

Distribution, continued

Table 9: Oaxaca Decomposition on Reweighted Unconditional Quantiles Explaining Black White Wage Gap Throughout the

37

6.167*** 0.00770

Mean Hispanic Log Wage

RIF Wage Gap

0.0472***

Non-cognitive Skills

Total Explained

-0.119*

0.774

26,684

Total Unexplained

Constant

Observations

(0.603)

(0.0764)

(0.206)

(0.269)

(0.0604)

(0.103)

(0.416)

(0.0424)

(0.00525)

(0.0142)

(0.0193)

(0.0265)

(0.00930)

(0.0614)

(0.0535)

(0.0298)

26,684

0.687

-0.0307

-0.0424

-0.262

-0.0423

-0.0542

-0.317

0.0969***

-0.00541

-0.00224

0.0439***

0.0409***

0.0197***

0.0662

6.335***

6.401***

(0.592)

(0.0666)

(0.178)

(0.298)

(0.0610)

(0.109)

(0.401)

(0.0289)

(0.00648)

(0.00919)

(0.0160)

(0.0183)

(0.00692)

(0.0624)

(0.0583)

(0.0220)

20th percentile

26,684

0.572

0.00316

-0.0171

-0.133

-0.0380

-0.0114

-0.369

0.0923***

-0.00543

-0.00703

0.0443***

0.0320***

0.0285***

0.0954

6.466***

6.561***

(0.584)

(0.0676)

(0.060)

(0.325)

(0.0623)

(0.117)

(0.380)

(0.0248)

(0.00640)

(0.00758)

(0.0147)

(0.0153)

(0.00722)

(0.0666)

(0.0635)

(0.0199)

30th percentile

26,684

0.236

0.00847

-0.0178

-0.0775

-0.0293

0.00859

-0.112

0.0794***

-0.00544

-0.0164***

0.0450***

0.0224*

0.0339***

0.0879

6.615***

6.703***

(0.572)

(0.0649)

(0.141)

(0.313)

(0.0602)

(0.121)

(0.387)

(0.0225)

(0.00633)

(0.00727)

(0.0137)

(0.0131)

(0.00792)

(0.0645)

(0.0618)

(0.0183)

40th percentile

26,684

-0.136

0.0104

-0.00627

-0.0436

-0.0171

0.0436

0.170

0.0781***

-0.00501

-0.0177***

0.0340***

0.0287***

0.0382***

0.0885

6.742***

6.830***

(0.592)

(0.0635)

(0.130)

(0.320)

(0.0587)

(0.125)

(0.405)

(0.0215)

(0.00598)

(0.00692)

(0.0113)

(0.0128)

(0.00867)

(0.0635)

(0.0610)

(0.0175)

50th percentile

of schooling minus 6.

NLSY79. Only individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years

of Control Score, Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test. The data include the years 1979-2004 from the

distributions reweighted using DFL. The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus

This table gives the Oaxaca-Blinder decompositions of the Hispanic white wage gap observed in unconditional quantiles of the log hourly wage

-0.282 0.0814

Cubic in experience

Non-cognitive Skills

Urban, Region

-0.149 -0.0280

Cognitive Skills

-0.516

Education

Unexplained by characteristics

-0.00393 0.127***

Cubic in experience

0.0113

0.0715***

Cognitive Skills

Urban, Region

0.000884

Education

Explained by characteristics

6.175***

Mean White Log Wage

10th percentile

Comparison Group: Hispanic Males

Reference Group: White Males

Decomposition Method: RIF regressions with reweighing

the Distribution

Table 10: Oaxaca Decomposition on Reweighted Unconditional Quantiles Explaining Hispanic White Wage Gap Throughout

38

6.876*** 0.0985*

Mean Hispanic Log Wage

RIF Wage Gap

0.0626 0.00592 -0.0925 -0.0271 0.0216

-0.340

26,684

Cognitive Skills

Non-cognitive Skills

Urban, Region

Cubic in experience

Total Unexplained

Constant

Observations

(0.609)

(0.0617)

(0.125)

(0.327)

(0.0554)

(0.123)

(0.409)

(0.0202)

(0.00516)

(0.00673)

(0.00935)

(0.0115)

(0.0102)

(0.0620)

(0.0597)

(0.0164)

26,684

-0.256

0.0227

-0.0921

-0.159

0.0113

0.0642

0.454

0.0677***

-0.00377

-0.0177***

0.0247***

0.0148

0.0497***

0.0904

7.038***

7.129***

(0.621)

(0.0635)

(0.128)

(0.322)

(0.0574)

(0.126)

(0.417)

(0.0189)

(0.00446)

(0.00653)

(0.00902)

(0.0108)

(0.0105)

(0.0637)

(0.0618)

(0.0152)

70th percentile

26,684

-0.0664

0.00533

-0.111

-0.243

0.0177

0.0516

0.356

0.0571***

-0.00355

-0.0219***

0.0227***

0.00898

0.0509***

0.0624

7.237***

7.300***

(0.644)

(0.0632)

(0.136)

(0.296)

(0.0546)

(0.118)

(0.441)

(0.0190)

(0.00426)

(0.00730)

(0.00829)

(0.0117)

(0.0109)

(0.0625)

(0.0606)

(0.0152)

80th percentile

26,684

-0.168

0.0273

-0.0431

-0.146

-0.00557

0.0527

0.337

0.0476***

-0.00397

-0.0211***

0.0212***

-0.00153

0.0530***

0.0750

7.465***

7.540***

(0.624)

(0.0598)

(0.0786)

(0.273)

(0.0462)

(0.103)

(0.434)

(0.0209)

(0.00741)

(0.00877)

(0.00923)

(0.0149)

(0.0119)

(0.0565)

(0.0539)

(0.0168)

90th percentile

of schooling minus 6.

NLSY79. Only individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years

of Control Score, Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test. The data include the years 1979-2004 from the

distributions reweighted using DFL. The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus

This table gives the Oaxaca-Blinder decompositions of the Hispanic white wage gap observed in unconditional quantiles of the log hourly wage

0.412

Education

Unexplained by characteristics

-0.00433

Urban, Region

0.0769***

0.0263*** -0.0180***

Non-cognitive Skills

Total Explained

0.0255**

Cognitive Skills

Cubic in experience

0.0474***

Education

Explained by characteristics

6.974***

Mean White Log Wage

60th percentile

Comparison Group: Hispanic Males

Reference Group: White Males

Decomposition Method: RIF regressions with reweighing

the Distribution, continued

Table 11: Oaxaca Decomposition on Reweighted Unconditional Quantiles Explaining Hispanic White Wage Gap Throughout

39

6.074*** 0.130***

Mean Hispanic Log Wage

RIF Wage Gap

0.0863

-0.235

16,230

Total Unexplained

Constant

Observations

(0.268)

(0.103)

(0.120)

(0.0379)

(0.0604)

(0.0367)

(0.211)

(0.0267)

(0.00525)

(0.0102)

(0.0205)

(0.0094)

(0.00510)

(0.0387)

(0.0306)

(0.0236)

16,230

-0.309

0.124***

0.384***

-0.0654

0.0128

0.0701**

0.0320

0.0571**

-0.00153

0.0318***

0.0344**

-0.00888

0.00137

0.182***

6.227***

6.408***

(0.267)

(0.0420)

(0.103)

(0.110)

(0.0379)

(0.0367)

(0.211)

(0.0312)

(0.00916)

(0.0114)

(0.0220)

(0.0102)

(0.0107)

(0.0432)

(0.0328)

(0.0281)

20th percentile

16,230

-0.190

0.190***

0.321***

-0.114

0.0128

0.0822***

0.0784

0.0476*

-0.00249

0.0270*

0.0291*

-0.00793

0.00186

0.238***

6.350***

6.588***

(0.272)

(0.0423)

(0.101)

(0.120)

(0.0381)

(0.0369)

(0.220)

(0.0343)

(0.0098)

(0.0116)

(0.0213)

(0.0108)

(0.0150)

(0.0459)

(0.0335)

(0.0314)

30th percentile

16,230

-0.0160

0.139***

0.306***

-0.0384

-0.0179

0.0446

-0.140

0.0870***

-0.00433

0.0247***

0.0479***

0.0171**

0.00156

0.226***

6.481***

6.707***

(0.280)

(0.0390)

(0.104)

(0.127)

(0.0391)

(0.0368)

(0.229)

(0.0375)

(0.0107)

(0.0098)

(0.0232)

(0.0097)

(0.0127)

(0.0491)

(0.0349)

(0.0344)

40th percentile

16,230

0.0137

0.131***

0.286***

-0.0277

-0.0138

0.0320

-0.159

0.0908***

-0.00371

0.0233***

0.0356***

0.0340***

0.00173

0.222***

6.610***

6.832***

(0.304)

(0.0379)

(0.0893)

(0.131)

(0.0402)

(0.0376)

(0.253)

(0.0373)

(0.0098)

(0.0089)

(0.0120)

(0.0121)

(0.0139)

(0.0503)

(0.0361)

(0.0348)

50th percentile

of schooling minus 6.

NLSY79. Only individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years

of Control Score, Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test. The data include the years 1979-2004 from the

distributions reweighted using DFL. The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus

This table gives the Oaxaca-Blinder decompositions of the Hispanic white wage gap observed in unconditional quantiles of the log hourly wage

-0.0184 0.379***

Cubic in experience

Non-cognitive Skills

Urban, Region

0.0418 0.00961

Cognitive Skills

-0.0907

Education

Unexplained by characteristics

0.0436*

0.0225

Urban, Region -0.00152

0.0272*

Non-cognitive Skills

Total Explained

-0.00530

Cognitive Skills

Cubic in experience

0.000723

Education

Explained by characteristics

6.204***

Mean White Log Wage

10th percentile

Comparison Group: Black Males

Reference Group: Hispanic Males

Decomposition Method: RIF regressions with reweighing

the Distribution

Table 12: Oaxaca Decomposition on Reweighted Unconditional Quantiles Explaining Hispanic Black Wage Gap Throughout

40

0.207***

RIF Wage Gap

16,230

Observations

(0.347)

(0.0392)

(0.100)

(0.141)

(0.0431)

(0.0396)

(0.293)

(0.0382)

(0.00863)

(0.00673)

(0.0164)

(0.0174)

(0.0149)

(0.0519)

(0.0378)

(0.0353)

16,230

0.167

0.0922***

0.0582

0.213*

-0.0430

0.0508

-0.354

0.103

-0.00143

0.0198***

0.0476***

0.0353***

0.00199

0.195***

6.878***

7.074***

(0.409)

(0.0469)

(0.0999)

(0.154)

(0.0503)

(0.0456)

(0.375)

(0.0373)

(0.0079)

(0.0092)

(0.0215)

(0.0135)

(0.0166)

(0.0556)

(0.0428)

(0.0353)

70th percentile

16,230

-0.325

0.127***

-0.0407

0.396***

-0.00914

0.0908**

0.0153

0.0647**

-0.00162

0.0177*

0.0225

0.0236**

0.00255

0.192***

7.064***

7.256***

(0.505)

(0.0502)

(0.101)

(0.160)

(0.0498)

(0.0518)

(0.480)

(0.0368)

(0.00713)

(0.0108)

(0.0184)

(0.0127)

(0.0211)

(0.0570)

(0.0447)

(0.0352)

80th percentile

16,230

-0.337

0.153***

-0.00701

0.369***

-0.0136

0.106**

0.0350

0.0276

-0.00130

0.00885

0.00787

0.00873

0.00349

0.181***

7.342***

7.523***

(0.611)

(0.0571)

(0.133)

(0.174)

(0.0541)

(0.0571)

(0.579)

(0.0402)

(0.00696)

(0.0108)

(0.0182)

(0.0130)

(0.0284)

(0.0629)

(0.0491)

(0.0389)

90th percentile

of schooling minus 6.

NLSY79. Only individuals with more than 8 years of schooling are included. Experience here is potential experience, calculated as age minus years

of Control Score, Rosenberg Self Esteem Score, Pearlin Mastery Scale and the Coding Speed Test. The data include the years 1979-2004 from the

distributions reweighted using DFL. The measure of cognitive skills is the AFQT. The measures of non-cognitive skills are the Rotter Internal Locus

This table gives the Oaxaca-Blinder decompositions of the Hispanic white wage gap observed in unconditional quantiles of the log hourly wage

0.170

Constant

0.218***

0.0536

Urban, Region

0.0946***

-0.0297

Non-cognitive Skills

Total Unexplained

0.00345

Cognitive Skills

Cubic in experience

-0.321

Education

Unexplained by characteristics

-0.00231 0.113***

Urban, Region

Total Explained

0.0356*** 0.0211***

Non-cognitive Skills

Cubic in experience

0.0566***

Cognitive Skills

Education

0.00182

6.730***

Mean Hispanic Log Wage

Explained by characteristics

6.937***

60th percentile

Mean White Log Wage

Comparison Group: Black Males

Reference Group: Hispanic Males

Decomposition Method: RIF regressions with reweighing

the Distribution, continued

Table 13: Oaxaca Decomposition on Reweighted Unconditional Quantiles Explaining Hispanic White Wage Gap Throughout