IV. Differences across Groups - Faculty of...

Fortin – Econ 560 Lecture 4A

IV. Differences across Groups

1. Theory of Discrimination 1.What is Economic Discrimination? 2. Theories of Discrimination a. Taste-Based Discrimination b. Statistical Discrimination c. Other Models

3. Evidence 1. Direct Evidence 2. Indirect Evidence

a. Estimating simple models of wage determination b. Methodologies for decomposing wage changes between groups

c. Variants of Oaxaca-Blinder


1.1. What is Economic Discrimination? The study of labour market discrimination comes from the observation that there

are long-lasting differences in the average wage rates or income among groups of workers, who are presumed to be equally productive or have equally productive capacity.

Glass ceilings (sticky floors) which correspond to the under-representation (over-

representation) of certain groups in the upper (lower) tail of the wage distribution have recently received more attention.

In the United States, studies have traditionally focused on racial (black-

white), ethnic (Hispanic) and gender wage differentials (O’Neill and O’Neill, 2005).

The study of other economic outcomes, such as employment, housing, etc.

appeal to the concept of occupational/industrial segregation and spatial segregation, which is interlinked with discrimination.


More recently there has been some interest with regards to persons with disabilities. The role of beauty, height and weight has also been studied (e.g. Hamermesh and Biddle, 1994, 1998; Mobius and Rosenblatt, 2006; Case and Paxson, 2008).

Drawing from psychology, recent experimental studies have also examined

the effects of observed characteristics, such as status or group identity, on the way people are treated differently (e.g. Fershtman and Gneezy, 2001).

In Canada, studies have traditionally focused on gender and language

(French-English) wage differentials, and more recently on wage differentials across more ethnic groups including aboriginals (Pendakur, Pendakur and Woodcock, 2006) and persons with disabilities.

Racial or ethnic profiling has introduced another dimension in which

instances of discrimination can be invoked.

In Europe, the poor labour market outcomes of 2nd generation immigrants have provoked severe incidents of civil unrest. Discrimination on the basis of religious beliefs (e.g. Muslims in the UK) has come to the fore.


Sexual orientation is also another type of group membership that is included

in most human rights codes which protect against discrimination through litigation.

Discrimination raises problems of inequity and of inefficiency.

Many countries also have responded to these perceived inequities by enacting

equality promoting legislation.

In the United States, legislations have taken the form of “affirmation action” legislations which focus on “equality of opportunities” rather than “equality of outcomes”.

More recently however, there have been reactions and limitations to

affirmative action perceived as “reverse discrimination”. In Canada, both types of legislations have been enacted under the

“employment equity” federal legislation and “pay equity” (“comparable worth”, “pay fairness”) mostly provincial legislation.


Canada’s two more populous provinces (Quebec and Ontario) have been

world leaders in enacting pro-active pay equity legislation that apply to the private, as well as public sectors. Most other provinces have complaint-based legislation applying to the public sector. The efficacy of these legislations is doubtful (Baker and Fortin, 2004).

440 Michael Baker and Marie Drolet

Figure 7 Female-Male Wage Ratios, Full-Time Workers Age 25-54, by Marital Status

Married/Common Law

Year

Source: Authors'calculations from the Survey of Work History (SWH), Survey of Union Membership (SUM), Labour Market Activity Survey (LMAS), Survey of Labour and Income Dynamics (SLID), and Labour Force Survey (LFS).

Table 3 The Average Gender Wage Ratio by Province for Selected Periods

1981 1987-1989 1996-1998 2006-2008

nf 0.75 (9) 0.73 (9) 0.79 (9) 0.83 (8) PEI 0.83 (1) 0.81 (1) 0.88 (1) 1.00 (1) ns 0.76 (6) 0.79 (2) 0.83 (3) 0.89 (3) nb 0.82 (2) 0.74 (8) 0.81 (7) 0.87 (6) qc 0.81 (3) 0.77 (4) 0.83 (3) 0.89 (3) on 0.76 (6) 077 (4) 0.84 (2) 0.85 (7)

mb 077 (5) 0.78 (3) 0.83 (3) 0.90 (2) sk 0.81 (3) 0.75 (6) 0.80 (8) 0.88 (5) ab 071 (10) 0.75 (6) 0.78 (10) 0.78 (10)

bc 0.76 (6) 0.72 (10) 0.82 (6) 0.83 (8)

Notes: Provincial rank in parentheses.

Source: Authors'calculations from the Survey of Work History, Survey of Union Membership, Labour Market Activity Survey, Survey of Labour and Income Dynamics, and Labour Force Survey.

Canadian Public Policy - Analyse de politiques, vol. xxxvi, no. 4 2010

This content downloaded from 137.82.185.238 on Tue, 12 Nov 2013 14:23:02 PMAll use subject to JSTOR Terms and Conditions

Source: Baker and Drolet (2010)

Nicole

Rectangle

Source: Jacobsen (2007)

Nicole

Rectangle

Nicole

Rectangle

Nicole

Rectangle


In economics, labour market discrimination is defined as a situation in which persons who provide labour market services and who are equally productive in a physical or material sense are treated unequally in a way that is related to an observable characteristic such as race, ethnicity or gender.

More formally following Cain (1986), let Y represent the labour market outcome

of interest, X represent a vector of productivity characteristics that are exogenous to the process of discrimination, and 0Z denote the fact that a person is in the majority group and 1Z if in the minority group.

Assuming a linear and additive model, ZXY , (1)

if 0 we have evidence of discrimination

We define as labour market discrimination (the economist’s view) 0,1, ZXYEZXYED A


But what if all X characteristics are endogeneous, any difference in X across groups is attributed to the process of discrimination under study and the unconditional difference may be a more appropriate measure of discrimination (the layman’s view)

01 ZYEZYED

If a labour market determinant such as educational attainment depends of inputs supplied earlier in life, called pre-market characteristics, then differences in labour market outcomes will not be attributable to labour market discrimination.

These pre-market factors are increasingly including non-only cognitive skills (such

as IQ, test scores, etc.) but also non-cognitive factors (self-confidence, competiveness, etc.)

Discrimination studies are also concerned with issues of “glass ceiling”, the

inability of certain groups to reach the higher pay echelons and “sticky floors”, the concentration of certain groups at longer pay echelons.

Source: Altonji and Blank (1999)

Nicole M Fortin

Nicole M Fortin

Nicole

Rectangle

Nicole

Rectangle

Nicole

Rectangle

Nicole

Oval

Nicole

Oval

Nicole

Rectangle


1.2. Theories of Discrimination The main models of discrimination can be classified into 2 broad classes depending

on the basis for discrimination: The first is based on taste or prejudice some members of the majority group

A against interacting with members of the minority group B The second is statistical information by employers in the presence of

imperfect information about the skills or behavior of the minority group More recent theories (efficiency wage models, principal-agent models)

introduce imperfect information into taste-based models of discrimination The neoclassical theories of discrimination are generally set within a competitive

framework with the exception of the monopsonistic model and are demand-side driven.

The supply-side of the labour market is effectively neutralized by the assumption

that minority and majority groups are equally productive.


Discrimination can lead to 1) unequal treatment between individuals from the two groups 2) unequal outcomes (e.g. wages) for all members of a particular group 3) segregation by establishments or occupations, that is the unequal

representation of different groups with equal productivity and motivation

a. Taste-Based Discrimination Taste-based perspectives of discrimination (Becker, 1957) are based on the notion

that discriminating parties have a preference for some groups and an aversion or a prejudice towards others, even when they are equally productive.

Segregation is often an outcome of his models or a means of mitigating

discrimination. These types of prejudice can be lead employer, employee or customer

discrimination o Which may call for different types of remedy


When an employer is prejudiced against a minority group, denoted B , he incurs a psychic cost 0d that acts to rise the cost of hiring the minority member to dwB

That is, firms will maximize: U = p F(NB + NA) − wANA − wBNB – dNB

where p is the price level, F is the production function, Nj is the number of workers of group j = {A,B}, and wj is the wage paid to members of each group.

A prejudiced employer will hire members of the B group only if dww BA .

Let G (d) denote the CDF of the prejudice parameter d in the population of

employers. Employers discriminate to varying degrees so that there is a distribution of di.

The optimal number of workers hired at each firm is determined by the solutions to pF’(NA) = wA

pF’(NB) = wB + d


The market demand for A and B workers will depend on the distribution of distribution of di.

Implications of the model:

1) B workers will be employed by the least prejudiced firms and that A and B workers will be segregated by employer in the labour market.

o The model can be extended for the prejudice to apply only when B workers work in certain types of jobs, then we will get occupational segregation.

2) If discriminating tastes are widespread and there are many B workers seeking employment, some B workers will have to find jobs at discriminating firms.

o A wage differential BA ww will arise if and only if the fraction of discriminating employers (or discriminating jobs) is sufficiently large that the demand for B workers when BA ww is less than the supply.

wN, G (d)), w (wN AsABA

dA )(

wN, G (d)), w (wN dsABA

dB )(


3) The more prevalent and the stronger employers’ discriminatory taste and the larger the number of B workers seeking employment the larger the market wide wage gap )( BA ww

o If there are enough non-discriminating employers, then discrimination is competed away.

A controversial aspect of Becker’s model of employer discrimination is the notion

that such discrimination should not survive in the face of market forces because In the short run, non-discriminating employers should out-perform

discriminating employers because they are willing to hire the cheaper but equally productive inputs.

Minorities that are discriminated against should move at the margin to employers that do not discriminate.

As capital moves to non-discriminating firms, in the long run these employers should expand and discriminating ones contract, until discriminatory wage differentials are eliminated and the law of one price prevails. That is, if the assumption of zero profit in the long-run equilibrium in competitive market holds.

prejudice and wages 779

Fig. 2.—Relationship between racial tastes and the relative wages and relative supply ofblacks and whites. The figure shows how the equilibrium ratio of black to white wagesresponds to three sets of market conditions. When the relative supply of black workers issmall relative to the number of unprejudiced employers, as is the case when supply is asdepicted by , the marginal discriminator is unprejudiced and there is no racial wageS1

gap in equilibrium. When the distribution of racial preferences among employers is heldconstant, a shift out in the relative supply of black workers (from to ) requires thatS S1 2

more prejudiced employers hire blacks, and the ratio of black to white wages falls fromone to R. When the relative supply of black workers is held constant, an increase inprejudice among employers likely to be the marginal discriminator (which causes therelative demand curve to rotate from ABD to ), further reduces the equilibrium ratio′ABDof black to white wages to .′R

the distribution of prejudice among employers are such that blacks canall be hired by totally unprejudiced employers.

The foregoing suggests that if racial wage gaps were empirically re-lated to the average and the marginal level of prejudice among em-ployers, Becker’s model predicts that only with the latter measure shouldthere be a systematic relationship. The effort to test this propositionempirically is complicated by the fact that, even if the complete distri-bution of prejudice among employers were known, it is impossible toknow ex ante which employer is the marginal one. Becker’s originaldiscussion does suggest one simple measure for the prejudice of themarginal employer that should hold under particular conditions. Spe-cifically, if firms are of equal size and if p is the percentage of blacks in

Source: Charles and Guryan (2008)

776 journal of political economy

Fig. 1.—Relationship between census division black-white wage gap and two prejudice-related questions from the GSS.

First, we find that racial wage gaps are much more closely related tothe level of prejudice of the “marginal” person in the distribution thanthey are to average levels of prejudice. We further show that it is onlyprejudice in the left tail of the prejudice distribution that seems tomatter for wage gaps; wages do not vary at all with the prejudice of themost prejudiced persons in a state. Importantly, the foregoing resultsare from regressions that control for the racial makeup of states. Finally,we show that the fraction of a workforce that is black, with prejudice

Source: Charles and Guryan (2008)


Becker realized that not all sectors of the economy are perfectly competitive. So that one might expect more discrimination in less competitive areas of the economy.

Goldberg (1982) models racial sentiment slightly differently than Becker, representing it not as distaste for blacks but rather as nepotism, or favoritism toward whites which may be more persistent.

This impact of competitive market forces at reducing discrimination need not to

apply to prejudices emanating from co-workers and customers. The basic idea of employee (or co-worker) discrimination is that some members

of the majority group A are prejudiced against group B members and do not like to work with members of the minority group, if forced to do so they will act as if their wage was A

wA

wA wdwdw )1( where 0wd is a prejudice coefficient.

Segregating workers should eliminate group wage differentials

We should get segregation by occupations but not necessarily a wage differential

unless workers A and B are forced to work alongside


With consumer discrimination, prejudiced consumers in group A get less utility if they purchase from a group B member than from a group A member. They act as if the price was )1( cdp .

Example: waiters in fancy French restaurants vs. waitresses in Diners The effect of such discrimination on wages is reduced to the extent that B members

can serve only B customers and unprejudiced A customers, or to the extent that B work in occupations without consumer contact.

Holzer and Ihlanfeldt (1998) use data from a survey of employers in four large metropolitan areas in the United States to show that the racial composition of an establishment's customers has sizable effects on the race of who gets hired, particularly in jobs that involve direct contact with customers and in sales or service occupations.

o They also find that the race of customers affects wages, with employees in establishments that have mostly black customers earning less than those in establishments with mostly white customers.

Taste-based gender discrimination may also emanate from families in gender preferences for off-springs or educational choices, unequal division of household tasks.

Washington Post-Kaiser Family Foundation-HarvardUniversity Gender Poll ResultsThursday, March 26, 1998

This Washington Post/Kaiser Family Foundation/Harvard University study is based onrandom telephone interviews with 1,202 adults on Aug. 14-Sept. 7, 1997. The margin oferror is plus or minus 3.0 percentage points.

………35. (Asked of half sample) For each of the following, would you personally prefer todeal with a man or a woman?

A. Elementary school teacher Doesn't Matter/ No Man Woman Same (vol.) opinion9/7/97 All 8 61 30 *9/7/97 Male 5 65 30 09/7/97 Female 11 58 31 *

B. Electrician or plumber Doesn't Matter/ No Man Woman Same (vol.) opinion9/7/97 All 62 5 32 *9/7/97 Male 61 3 36 09/7/97 Female 64 7 29 *

C. Nurse Doesn't Matter/ No Man Woman Same (vol.) opinion9/7/97 All 5 67 29 *9/7/97 Male 6 63 31 09/7/97 Female 4 70 26 *D. Insurance agent Doesn't Matter/ No Man Woman Same (vol.) opinion9/7/97 All 31 22 47 *9/7/97 Male 26 20 54 09/7/97 Female 35 25 40 1E. A supervisor at work Doesn't Matter/ No Man Woman Same (vol.) opinion9/7/97 All 38 24 38 *9/7/97 Male 39 15 46 *9/7/97 Female 37 33 30 1

Nicole

Text Box

Customer Discrimination?

Source: Holzer and Ihlanfeldt (1998)

Nicole

Rectangle


b. Statistical Discrimination The basic premise of this literature pioneered by Phelps (1972) and Arrow (1973) is

that firms have limited information about the skills and turnover propensity of applicants, in particular young workers.

Firms thus have an incentive to use easily observable characteristics such as race or

gender to “statistically discriminate” among workers if these characteristics are correlated with performance.

In the Unites States, it is illegal to make hiring, pay or promotion decisions based on

predictors about workers behavior (productivity, absenteeism, turnover, etc.) by race or gender even if such predictions are statistically rational forecasts given the information set available to the employer.

There are 2 main strands to the statistical discrimination literature, both based on

feedback effects from statistical discrimination


1) the first investigates how incorrect prior beliefs about the productivity of group members can influence pay and hiring decisions

Workers are offered a skilled job if they invest in training.

If firms initially think that fewer group 퐵 members are qualified to benefit from the training, this will influence the investment decision of group 퐵 in a way that confirms the firm’s prior

Even if the 퐴 and 퐵 workers have identical skills and the same training cost distribution and even if firms update priors in a sensible way then stereotypes that are initially negative may become self-confirming

2) the second concerns the consequences of group differences in the precision of the

information that employers have about individual productivity: the imperfect information takes the long of a noisy signal

Suppose that the true productivity of a specified group of workers is difficult

for firms to discern, perhaps because of cultural differences


Example 1: Assume that when workers apply for jobs, the employer sees the race of the

applicant X ={A, B} and some error-ridden signal 휂 of productivity.

Assume that employers have learned from experience that the actual productivity of each group, 휂 ∼ 푁(휂 ,휎 ) , so that for individual i , they assume that

휂 = 휂̅ + 휀

When a worker applies for a job, the employer observes the error-ridden productivity signal of the form: 휂 = 휂 + ί where ί ∼ 푁(0,휎ί ) with 휎ί > 0

And the signal is unbiased: 휂 = 휂̅ + 휀 + ί with 퐸(휂 |휂 ) = 휂 .

What is the expectation of productivity 휂 given the observed signal 휂 and group membership X? This is simply the regression equation

퐸(휂|휂,푋) = 휂̅ (1 − 훾) + 휂훾 = 휂 + (휂 − 휂̅ )훾 where 훾 = is the coefficient of the bivariate regression of 휂 on 휂 .


For whichever group has lower 휎 , 휂 will be more informative; employers will put less weight on the mean for this group and more weight on the signal. See Figures 1a and 1b of Cain and Aigner.

The difference in information quality will have three main implications: i) to the extent that productivity depends on the quality of the match between

the skills of the worker and the requirements of the job, the expected productivity will be lower for groups when the firm is more uncertain

ii) this leads to differences across groups in the return to job matching iii) the wages of group 퐵 members may be less responsive to performance

because firms have difficulty “seeing” their productivity. This would weaken the incentive of group 퐵 members to invest in skills

The feedback effects of this form of statistical discrimination implies productivity

differences and there is no longer “economic discrimination”


c. Other Models :

i. Efficiency wage models These models are based on segmented labour markets and on the idea that while

group 퐴 and group 퐵 are equal productive, there are differences in other productive characteristics such as effort, quit rate, mobility or manageability.

In these models, worker’s productivity is increased by increases in pay and it is

assumed that the detection of shirkers in more difficult in one of the two sectors, that is, there is imperfect information about the productivity of workers. o Two occupations are usually needed: one with high supervision and low wage

premiums and a second with low supervision and high premiums. o Men as outside workers and women as clerical workers are classic examples of

instances where these models could apply.


ii. Monopsony in the labour market A classic case of workers’ exploitation in neoclassical economics arises under

conditions of monopsony. Workers are captive in a market where there is only one employer or where a group

of employers collude to act as one buyer Assuming that the monopsonist cannot differentiate the reservation wage of each

workers (unlike the discriminating monopsonist) but can differentiate the reservation wages of broad groups of homogeneous workers.

In such circumstances, the monopsonist only has to pay the same wage within

groups: different wages can persist between groups. Consider the case where groups 퐴 and 퐵 workers are equally productive, but where

group 퐵 workers have lower reservation wages, which may in turn, reflects poor employment opportunities elsewhere or low subjective evaluation of their own labour market worth.

Source: Benjamin, Gunderson and Riddell (1998)

Nicole

Text Box

1

Nicole

Text Box

2

Nicole

Text Box

3

Nicole

Text Box

4

Nicole

Text Box

5

Nicole

Pencil

Nicole

Pencil


Then as illustrated in the next figure, group 퐵 labour supply curve is lower than

group’s 퐴, and each have their own marginal cost curve which lies above each supply schedule. The total marginal cost curve TMC is obtained by horizontally summing the 퐴 and 퐵 marginal cost curves. The total amount of labour demanded

TN is determined by the intersection of the TMC and the VMP at ( *, MCNT ). The number of workers of each group demanded will be taken from the group

specific inverse marginal cost curves evaluated at *MC , but the wage paid will be determined by the group specific supply curves.

The wage differential between the two groups emerges as a result of their different

reservation wages (or labour supply elasticities). This in turn can be due to the extent that women are more tied to their local labour market and their mobility is restricted because of household pressures.

While the empirical evidence on gender differences in wage elasticities is unclear,

there are some indications of differences in reservation wages.


iii.Rent-seeking Models Extensions of the monopsony model argue that one group in the economy can band

together to improve their well-being at the expense of the rest of society.

Akerlof (1980) social norms’ model argue that in small markets in which every participant is a potential trader, a social custom regarding hiring various groups can be stable.

A non-discriminating employer will lose profits because the discriminating majority

will boycott his business. If the markets are small, the loss of a few traders will be costly in foregone profits. This could apply in small communities where people can monitor each other’s behavior.

Rent-seeking models may apply more to racial or ethnic discrimination than to

gender discrimination because it is easier to conceive of these groups as being separable in society.


iv. Supply-Side Theories,Crowding The crowding hypothesis formalized by Bergmann (1971) implies that when group 퐵 members are segregated into only a few occupations.

This creates an excess supply in those occupations which reduces their wages.


1.3.Evidence

1.Direct Evidence: Audit studies and field experiments have recently been used to detect racial and

gender discrimination in employment and housing. This has involved sending trained auditors who are identical in qualifications

at prospective job interviews or housing openings while randomizing on race or gender.

Riach and Rich (2002) note that the findings from field studies appear to be more consistent with the majority white populations having a general “distaste” for minorities, but that statistical discrimination, or marketers using observable characteristics to make statistical inference about productivity or reservation values of market agents, for example, cannot be ruled out, ex ante or ex post.

These types of “audit studies are crucially dependent on an unstated hypothesis: that the distributions of unobserved (by the testers) productivity characteristics of majority and minority workers are identical.


Because they are not double-blind, we may suspect that minority testers may unconsciously take actions that reduce their odds of receiving a job offer to confirm the existence of discrimination

The set of employers audited is likely to be randomly chosen Also sample sizes are invariably small because audit studies are costly

More careful studies attempt to avoid this problem Goldin and Rouse (1996) studied one of the cleanest test form a field experiment

where orchestras beginning in the 1970s and 1980s used a screen to hide an auditioning musician from the jury. This increased the proportion female among new hires dramatically.

Bertrand and Mullainathan (2004) sent resumes to employers where the resumes

are identical except for a name that is male or female or with white or black sounding names.

Pager, Western, Bonikowski (2008) used matched teams of testers who applied

for 341 entry-level jobs in New York City over nine months in 2004; they sent matched black, white, and Latino job seekers.

Source: Goldin and Rouse (2000)

VOL. 90 NO. 4 GOLDIN AND ROUSE: ORCHESTRATING IMPARTIALITY 729

TABLE 6-LINEAR PROBABILITY ESTIMATES OF THE LIKELIHOOD OF BEING ADVANCED: WITH INDIVIDUAL FIXED EFFECTS

Preliminaries

Withoutsemifinals With semifinals Semifinals Finals

(1) (2) (3) (4) (5) (6) (7) (8)

Blind -0.017 0.003 0.109 0.224 0.026 0.102 -0.154 -0.060(0.039) (0.046) (0.172) (0.242) (0.089) (0.096) (0.150) (0.149)

Female X Blind 0.125 0.111 0.013 --0.025 -0.179 -0.235 0.308 0.331(0.068) (0.067) (0.215) (0.251) (0.126) (0.133) (0.196) (0.181)

Number of auditions attended -0.020 0.010 0.015 0.126(0.014) (0.010) (0.030) (0.028)

Years since last audition -0.005 -0.006 -0.005 0.016(0.007) (0.005) (0.013) (0.015)

Automatic placement -0.096 -0.069(0.064) (0.073)

"Big Five" orchestra -0.154 -0.059 0.006 -0.059(0.035) (0.024) (0.081) (0.084)

Total number of auditioners in -0.003 0.014 -0.371 -0.262round (-;- 100) (0.081) (0.031) (0.521) (0.756)

Proportion female at the audition 0.118 0.312 0.104 0.067round (0.139) (0.134) (0.218) (0.159)

Principal -0.079 --0.078 -0.082 -0.185(0.037) (0.019) (0.066) (0.076)

Substitute 0.165 0.123 0.167 0.079(0.081) (0.093) (0.183) (0.217)

p-value of Ho: Blind + (Female 0.053 0.063 0.342 0.285 0.089 0.170 0.222 0.042X Blind) = 0

Year fixed effects? No Yes No Yes No Yes No YesR2 0.748 0.775 0.687 0.697 0.774 0.794 0.811 0.878Number of observations 5,395 5,395 6,239 6,239 1,360 1,360 1,127 1,127

Notes: The unit of observation is a person-round. The dependent variable is 1 if the individual is advanced to the next roundand 0 if not. Standard errors are in parentheses. All specifications include individual fixed effects, an interaction for the sexbeing missing and a blind audition round, a dummy indicating if years since last audition is missing, and [in columns (3)-(8)]whether an automatic placement is missing.Source: Eight-orchestra audition sample. See text.

round and those that were not. In the evennumbered columns we include year and instrument fixed effects, as well as individualand audition covariates. The individual correlates are whether the musician had an automatic placement in a semifinal or final round,years since the last audition in the sample,and the number of previous auditions inwhich we observe the musician to have competed. We also control for the total number ofmusicians in the round, the proportion femaleamong contestants, and whether the auditionis for a principal or substitute position.

Because 42 percent of the individuals in oursample competed in more than one round in ourdata set (24 percent of the musicians competedin more than one audition) and 6 percent competed both with and without a screen for a

particular type of round (e.g., semifinal), we areable to use an individual fixed-effects strategyto control for contestant "ability" that does notchange with time. In all columns of 'Table 6 weinclude individual fixed effects, in which casethe identification is from individuals who auditioned both with and without a screen.38 The

38 There are 639 person-rounds comprised of individualswho auditioned at a preliminary round that was not followedby a semifinal round [columns (1) and (2) of Table 6], bothwith and without a screen; on average these individualscompeted in 2.7 such preliminary rounds. There are 55person-rounds comprised of individuals who auditioned at apreliminary round that was followed by a semifinal round[columns (3) and (4)], both with and without a screen; onaverage these individuals competed in 2.4 such preliminaryrounds. There are 223 person-rounds comprised of individuals who auditioned at a semifinal [columns (5) and (6)],

Nicole

Rectangle

Nicole

Rectangle

Nicole

Rectangle

Nicole

Rectangle

997 VOL. 94 NO. 4 BERTRAND AND MULUINATHAN: RACE IN THE LABOR MARKET

Percent callback Percent callback for Percent difference for White names African-American names Ratio (D-value)

Sample: All sent resumes

Chicago

Boston

Females

Females in administrative jobs

Females in sales jobs

Males

Notes: The table reports, for the entire sample and different subsamples of sent resumes, the callback rates for applicants with a White-sounding name (column 1) an an African-American-sounding name (column 2), as well as the ratio (column 3) and difference (column 4) of these callback rates. In brackets in each cell is the number of resumes sent in that cell. Column 4 also reports the p-value for a test of proportion testing the null hypothesis that the callback rates are equal across racial groups.

employers rarely, if ever, contact applicants via postal mail to set up interviews.

E. Weaknesses of the Experimen~

We have already highlighted the strengths of this experiment relative to previous audit studies. We now discuss its weaknesses. First, our outcome measure is crude, even relative to the previous audit studies. Ultimately, one cares about whether an applicant gets the job and about the wage offered conditional on getting the job. Our procedure, however, simply measures callbacks for interviews. To the extent that the search process has even moderate frictions, one would expect that reduced interview rates would translate into reduced job offers. How- ever, we are not able to translate our results into gaps in hiring rates or gaps in earnings.

Another weakness is that the resumes do not directly report race but instead suggest race through personal names. This leads to various sources of concern. First, while the names are chosen to make race salient, some employers may simply not notice the names or not recog- nize their racial content. On a related note, because we are not assigning race but only race-specific names, our results are not repre- sentative of the average African-American (who may not have such a racially distinct

name).28 We return to this issue in Section IV, subsection B.

Finally, and this is an issue pervasive in both our study and the pair-matching audit studies, newspaper ads represent only one channel for job search. As is well known from previous work, social networks are another common means through which people find jobs and one that clearly cannot be studied here. This omis- sion could qualitatively affect our results if African-Americans use social networks more or if employers who rely more on networks differentiate less by race.29

111. Results

A. Is There a Racial Gap in Callback?

Table 1 tabulates average callback rates by racial soundingness of names. Included in brackets under each rate is the number of resumes sent in that cell. Row 1 presents our results for the full data set. Resumes with White

As Appendix Table A1 indicates, the African-American names we use are, however, quite common among African-Americans, making this less of a concern.

29 In fact, there is some evidence that African-Americans may rely less on social networks for their job search (Hany J. Holzer, 1987).

1003 VOL. 94 NO. 4 BERTRAND AND MULLAINATHAN: RACE IN THE LABOR MARKET

TABLE&-EFFECT ADDRESS OF CALLBACKOF APPLICANT'S ON LIKELIHOOD

Dependent Variable: Callback Dummy

Zip code characteristic:

Zip code characteristic

Zip code characteristic* African-American name

African-American name

Fraction college or Fraction Whites more Log(per capital income)

0.020 0.020 0.054 0.053 0.018 0.014 (0.012) (0.016) (0.022) (0.03 1) (0.007) (0.010) - -0.000 - -0.002 - 0.008

(0.024) (0.048) (0.015) - -0.031 - -0.03 1 - -0.112

(0.015) (0.01 3) (0.152)

Notes: Each column gives the results of a probit regression where the dependent variable is the callback dummy. Reported in the table is the estimated marginal change in probability. Also included in columns 1, 3, and 5 is a city dummy; also included in columns 2,4, and 6 is a city dummy and a city dummy interacted with a race dummy. Standard errors are corrected for clustering of the observations at the employment-ad level.

resumes with African-American-sounding names. Taken at face value, these results suggest that African-Americans may face relatively lower individual incentives to invest in higher skills.36

C. Applicants ' Address

An incidental feature of our experimental de- sign is the random assignment of addresses to the resumes. This allows us to examine whether and how an applicant's residential address, all else equal, affects the likelihood of a callback. In addition, and most importantly for our pur- pose, we can also ask whether African-Ameri- can applicants are helped relatively more by residing in more affluent neighborhoods.

We perform this analysis in Table 6. We start (columns 1,3, and 5 ) by discussing the effect of neighborhood of residence across all applicants. Each of these columns reports the results of a probit regression of the callback dummy on a specific zip code characteristic and a city dummy. Standard errors are corrected for clustering of the observations at the employment-ad level. We find a positive and significant effect of neighborhood quality on the likelihood of a callback. Applicants living in Whiter (column I), more educated (column 3) ,or higher-income (column 5) neighborhoods have a higher probability of receiving a callback. For example, a 10-percentage-point increase in the fraction of college-educated in zip code of residence in-

36 This of course assumes that the changes in job and wage offers associated with higher skills are the same across races, or at least not systematically larger for African- Americans.

creases the likelihood of a callback by a 0.54 percentage point (column 3).

In columns 2,4, and 6, we further interact the zip code characteristic with a dummy variable for whether the applicant is African-American or not. Each of the probit regressions in these columns also includes an African-American dummy, a city dummy, and an interaction of the city dummy with the African-American dummy. There is no evidence that African- Americans benefit any more than Whites from living in a Whiter, more educated zip code. The estimated interactions between fraction White and fraction college educated with the African- American dummy are economically very small and statistically insignificant. We do find an economically more meaningful effect of zip code median income level on the racial gap in callback; this effect, however, is statistically insignificant.

In summary, while neighborhood quality affects callbacks, African-Americans do not benefit more than Whites from living in better neighborhoods. If ghettos and bad neighbor-hoods are particularly stigmatizing for African- Americans, one might have expected African- Americans to be helped more by having a "better" address. Our results do not support this hypothesis.

D. Job and Employer Characteristics

Table 7 studies how various job requirements (as listed in the employment ads) and employer characteristics correlate with the racial gap in callback. Each row of Table 7 focuses on a specific job or employer characteristic, with

Nicole

Rectangle

1004 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2004

TABLE7-EFFECT AND EMPLOYER ON RACIAL IN CALLBACKSOF JOB REQUIREMENT CHARACTERISTICS DIFFERENCES

Sample mean Marginal effect on callbacks Job requirement: (standard deviation) for African-American names

Any requirement? (Y = 1) 0.79 0.023 (0.41) (0.015)

Experience? (Y = 1) 0.44 0.01 1 (0.49) (0.013)

Computer skills? (Y = 1) 0.44 0.000 (0.50) (0.013)

Communication skills? (Y = 1) 0.12 -0.000 (0.33) (0.015)

Organization skills? (Y = I) 0.07 0.028 (0.26) (0.029)

Education? (Y = 1) 0.11 -0.031 (0.3 1) (0.017)

Total number of requirements 1.18 0.002 (0.93) (0.006)

Sample mean Marginal effect on callbacks Employer characteristic: (standard deviation) for African-American names

Equal opportunity employer? (Y = 1) 0.29 -0.013 (0.45) (0.0 12)

Federal contractor? (Y = 1) 0.11 -0.035 (N = 3,102) (0.32) (0.016) Log(emp1oyment) 5.74 -0.001 (N = 1,690) (1.74) (0.005) Ownership status: (N = 2,878) Privately held 0.74 0.01 1

(0.019) Publicly traded 0.15 -0.025

(0.015) Not-for-profit 0.11 0.025

(0.042) Fraction African-Americans in employer's zip code 0.08 0.1 17

(N = 1,918) (0.15) (0.062)

Notes: Sample is all sent resumes (N = 4,870) unless otherwise specified in column 1. Column 2 reports means and standard deviations (in parentheses) for the job requirement or employer characteristic. For ads listing an experience requirement, 50.1 percent listed "some," 24.0 percent listed "two years or less," and 25.9 percent listed "three years or more." For ads listing an education requirement. 8.8 percent listed a high school degree, 48.5 percent listed some college, and 42.7 percent listed at least a four-year college degree. Column 3 reports the marginal effect of the job requirement or employer characteristic listed in that row on differential treatment. Specifically, each cell in column 3 corresponds to a different probit regression of the callback dummy on an African-American name dummy, a dummy for the requirement or characteristic listed in that row and the interaction of the requirement or characteristic dummy with the African-American name dummy. Reported in each cell is the estimated change in probability for the interaction term. Standard errors are corrected for clustering of the observations at the employment-ad level.

summary statistics in column 2. Column 3 African-American dummy. The reported coef- shows the results of various probit regressions. ficient is that on the interaction term. Each entry in this column is the marginal effect We start with job requirements. About 80 of the specific characteristic listed in that row on percent of the ads state some form of require- the racial gap in callback. More specifically, ment. About 44 percent of the ads require some each entry is from a separate probit regression minimum experience, of which roughly 50 per-of a callback dummy on an African-American cent simply ask for "some experience," 24 per- dummy, the characteristic listed in that row and cent less than two years, and 26 percent at least the interaction of that characteristic with the three years of experience. About 44 percent of

Nicole

Rectangle


TABLE 8---CALLBACK RATE AND MOTHER'S BY FIRST NAME EDUCATION

White female African-American female

Name Percent callback Mother education Name Percent callback Mother education

Emily Aisha Anne Keisha Jill Tamika Allison Lakisha Laurie Tanisha Sarah Latoya Meredith Kenya Carrie Latonya Kristen Ebony

Average 91.7 Average Overall 83.9 Overall

Correlation -0.318 (p = 0.404) Correlation -0.383 (p = 0.309)

White male African-American male

Name Percent callback Mother education Name Percent callback Mother education

Todd 5.9 87.7 Rasheed 3.0 77.3 Neil 6.6 85.7 Tremayne 4.3 -Geoffrey 6.8 96.0 Kareem 4.7 67.4 Brett 6.8 93.9 Darnell 4.8 66.1 Brendan 7.7 96.7 Tyrone 5.3 64.0 Greg 7.8 88.3 Hakim 5.5 73.7 Matthew 9.0 93.1 Jamal 6.6 73.9 Jay 13.4 85.4 Leroy 9.4 53.3 Brad 15.9 90.5 Jermaine 9.6 57.5

Average 91.7 Average Overall 83.5 Overall

Correlation -0.0251 ( p = 0.949) Correlation -0.595 (p = 0.120)

Notes: This table reports, for each first name used in the experiment, callback rate and average mother education. Mother education for a given first name is defined as the percent of babies born with that name in Massachusetts between 1970 and 1986 whose mother had at least completed a high school degree (see text for details). Within each sextrace group, first names are ranked by increasing callback rate. "Average" reports, within each race-gender group, the average mother education for all the babies born with one of the names used in the experiment. "Overall" reports, within each race-gender group, average mother education for all babies born in Massachusetts between 1970 and 1986 in that race-gender group. "Correlation" reports the Spearman rank order correlation between callback rate and mother education within each race-gender group as well as the p-value for the test of independence.

1 9 8 6 . ~ ~ name and, in that gender-race cell, whose moth- For each first name in our experiment, we compute the fraction of babies with that ers have at least completed a high school

degree.-In Table 8, we display the average callback

46 This longer time span (compared to that used to assess name frequencies) was imposed on us for confidentiality rate for each firstname with this proxy for reasons. When fewer than 10 births with education data social background. Within each race-gender available are recorded in a particular education-name cell, group, the names are ranked by increasing call- the exact number of births in that cell is not reported and we back rate, Interestingly, there is significantimpute five births. Our results are not sensitive to this imhutation. One African-American female name (Latonya) and two male names (Rasheed and Hakim) were imputed in this way. One African-American male name (Tremayne) tatively similar when we use a larger data set of California had too few births with available education data and was births for the years 1989 to 2000 (kindly provided to us by therefore dropped from this analysis. Our results are quali- Steven Levitt).


TABLE AI-FIRST NAMESUSEDIN EXPERIMENT

White female Name L(WI/L(BI Percevtion White

Allison m

Anne m

Carrie m

Emily m

Jill Y2

Laurie m

Kristen m

Meredith m

Sarah m

Fraction of all births:

3.8 percent

White male Name L(W)k(B) Perception White

Brad m

Brendan m

Geoffrey m

Greg cc

Brett m

Jay a

Matthew a

Neil m

Todd m


1.7 percent

African-American female Name L(BI/L(WI Perce~tion Black

Aisha Ebony Keisha Kenya Lakisha m

Latonya m

Latoya m

Tamika 284 Tanisha 32


7.1 percent

African-American male Name L(B)/L(W) Perception Black

Darnell Hakim Jamal 257 Jermaine 90.5 Kareem m

Leroy 44.5 Rasheed m

Tremayne m

Tyrone 62.5 Fraction of all births:

3.1 percent

Notes: This table tabulates the different first names used in the experiment and their identifiability. The first column reports the likelihood that a baby born with that name (in Massachusetts between 1974 and 1979) is White (or African-American) relative to the likelihood that it is African-American (White). The second column reports the probability that the name was picked as White (or African-American) in an independent field survey of people. The last row for each group of names shows the proporrion of all births in that race group that these names account for.

REFERENCES

Aigner, Dennis J. and Cain. Glenn G. "Statistical Theories of Discrimination in Labor Mar- kets." Industrial and Labor Relations Re- view, January 1977, 30(1), pp. 175-87.

Altonji, Joseph G. and Blank, Rebecca M. "Race and Gender in the Labor Markey," in Orley Ashenfelter and David Card, eds., Handbook of labor economics, Vol. 30. Amsterdam: North-Holland, 1999, pp. 3143-259.

Arrow, Kenneth, J. "The Theory of Discrimina- tion," in Orley Ashenfelter and Albert Rees, eds., Discrimination in labor markets. Princeton, NJ: Princeton University Press, 1973, pp. 3-33.

. "What Has Economics to Say about Racial Discrimination?Voumal of Economic Perspectives, Spring 1998, 12(2), pp. 91-100.

Becker, Gary S. The economics of discrimination, 2nd Ed. Chicago: University of Chicago Press, 1961.

Brown, Colin and Gay, Pat. Racial discrimination 17 years after the act. London: Policy Studies Institute, 1985.

Cornell, Bradford and Welch, Ivo. "Culture, In- formation, and Screening Discrimination." Journal of Political Economy, June 1996, 104(3), pp. 542-7 1.

Council of Economic Advisers. Changing America: Indicators of social and economic well-being by race and Hispanic origin. September 1998, http://w3.access.gpo.gov/eop/ca/pdfs/ca.pdf.

Cross, Harry; Kenney, Genevieve; Mell, Jane and Zimmerman, Wendy. Employer hiring prac- tices: Differential treatment of Hispanic and Anglo job applicants. Washington, DC: Ur- ban Institute Press, 1990.

http://w3

Excerpts from Pager, Western, Bonikowski (2008) In one case, for example, Zuri, an African American tester, reports his experience applying for a job as a warehouse worker: “The original woman who had herded us in told us that when we finished filling out the application we could leave because “there’s no interview today, guys!”…When I made it across the street to the bus stop …the woman who had collected our completed applications pointed in the direction of Simon, Josue and myself [the three test partners] motioning for us to return. All three of us went over…. She looked at me and told me she “needed to speak to these two” and that I could go back.” Zuri returned to the bus stop, while his white and Latino test partners were both asked to come back at 5pm that day to start work. Simon, the white tester, reports, “She said she told the other people that we needed to sign something—that that’s why she called us over—so as not to let them know she was hiring us. She seemed pretty concerned with not letting anyone else know.”

In one case, for example, the three testers inquired about a sales position at a retail clothing store. Joe, one of our African American testers, reports: “[The employer] said the position was just filled and that she would be calling people in for an interview if the person doesn’t work out.” Josue, his Latino test partner, was told something very similar: “She informed me that the position was already filled, but did not know if the hired employee would work out. She told me to leave my resume with her.” By contrast, when Simon, their white test partner, applied last, his experience is notably different: “…I asked what the hiring process was—if they’re taking applications now,interviewing, etc. She looked at my application. ‘You can start immediately?’ Yes. ‘Can you start tomorrow?’ Yes. ‘10 a.m.’ She was very friendly and introduced me to another woman (white, 28) at the cash register who will be training me.”

13

testers had extended interaction with the employer. If testers were acting in ways that fulfill their expectations of discrimination, we would expect outcomes for those tests conducted with interaction to show greater evidence of differential treatment than those without. If the results are consistent, or show weaker evidence of differential treatment, we can be more confident that experimenter effects are not driving the results. The results of this test indicate that personal contact served to weaken the effect of race on hiring decisions (see Table 1).11 When applicants has little chance to interact with the employer, for example, whites were 9.6 times more likely to receive a callback or job offer than their black partners (and 1.8 times more likely than their Latino partners). By contrast, when applicants had the opportunity to interact with employers, whites were just 1.9 times more likely to receive a callback or job offer than blacks, and equal to the outcomes of Latinos. Rather than enacting expectations of discrimination, then, interaction between testers and employers resulted in a substantial reduction in discrimination. The friendly, appealing qualities of the testers appear to mediate the effects of racial stereotypes, reducing the negative bias evident in more superficial reviews.

Table 1. Percentage of positive responses and race differences, by level of personal contact

White Latino Black Race Differences

Subsample (W) (L) (B) W / L W / B L / B Total 31.0 25.2 15.2 1.2 2.0 1.7 No personal contact 14.4 8.0 1.5 1.8 9.6 5.3 Personal contact 44.2 42.9 23.8 1.0 1.9 1.8 White felon Latino Black Race Differences Subsample (Wf) (L) (B) Wf / L Wf / B L / B Total 17.1 15.9 12.9 1.1 1.3 1.2 No personal contact 9.4 10.6 3.4 0.9 2.8 3.1 Personal contact 27.0 22.4 34.0 1.2 0.8 0.7 Note: Personal contact varies across testers within teams. Tests involving personal contact represent 56% by white testers, 49% by Latino testers, and 61% by black testers in the first team (N=171); 44% of white testers, 45% of Latino testers, and 31% of black testers in the second team (N=170).

11 Note: Because of the vastly different baseline positive response rates across groups, we calculate ratios as an indicator of disparate treatment rather than differences.

Source: Pager, Western, Bonikowski (2006)

Nicole

Rectangle

Nicole

Rectangle

Nicole

Rectangle

35

Table 1. Job Channeling by Race original job title suggested job Blacks channeled down Server Busser (324) Counter person Dishwasher/porter (102) Server Busboy (189) Assistant manager Entry fast food position (258) Server Busboy/runner (269) Retail sales Maintenance (399) Counter person Delivery (176) Sales Stockboy (831) Sales Not specified(a) Hispanics channeled down Server Runner (199) Sales Stock (2) Steam cleaning Exterminator (79) Counter person Delivery (176) Sales Stock person (503) Whites channeled down Server Busboy (192) Hispanics channeled up Carwash attendant Manager (1058) Warehouse worker Computer/office (1001) Whites channeled up Line Cook Waistaff (254) Mover Office / Telesales (784) Dishwasher Waistaff (858) Driver Auto detailing (948) Kitchen job “Front of the house” job (5) Receptionist Company supervisor (347)

(a) employer told tester “sales might not be right for you…” Note: numbers in parentheses refer to employer ID codes.

Nicole

Rectangle

Nicole

Rectangle


To “calibrate” the magnitude of racial preferences, they compare minority applicants to white applicants just released from prison after serving 18 months for a drug felony (possession with intent to distribute cocaine), and find that White felons do about as well as Black and Hispanic applicants with no (reported) criminal record.

Lab experiments have recently been used to study preference differences between

men and women, focusing on three factors that are relevant in the labor market: risk taking, social preferences and reaction to competition (Croson and Gneezy, 2004).

If women prefer jobs that are less risky, more socially virtuous and less competitive, then this could explain part of the gender differences in the labor market.

Discrimination in professional sports has been used to study both consumer and

employee discrimination Kahn and Sherer (1988) have found that home attendance at NBA games is

positively correlated with the fraction of white players. Price and Wolfers (2007) study NBA refereeing and find that more personal

fouls are called against players when they are officiated by an opposite-race refereeing crew than when officiated by an own-race crew, even when


conditioning on player and referee fixed effects and specific game fixed effects.

List (2004) performed an experiment in the sportscard market and found that “there is a strong tendency for minorities to receive initial and final offers that are inferior to those received by majorities, but overall, the data indicate that the observed discrimination is not due to animus, but represents statistical discrimination.”

Direct productivity estimates

A small number of studies have attempted to provide direct estimates of male and female productivity and to compare male and female wages relative to productivity.

Hellerstein, Neumark, and Troske (1999), for example, use firm level data from the United States to estimate production functions and the marginal productivity of females relative to males. o They find that females are somewhat less productive than males but that

the ratio of their wages is much lower than the ratio of their productivity, suggesting a discriminatory pay gap.


2. Indirect Evidence:

a. Estimating simple models of wage determination The simplest way of assessing the extent of discrimination consists in an equation

similar to equation (1), iiiDi XDw 00ln (1’)

where 1iD denote that individual i belong to the minority group and 0iD is the omitted category.

The coefficient D0 captures the wage disadvantage of the minority group that is

not “explained” or “accounted for” by the productive characteristics X . This method has the advantage that many potentially disadvantaged groups can be

compared to the majority group. One drawback is that is that it may be difficult to separate the effect of various

productive characteristics, for example education vs. experience.


b. Methodologies for decomposing wage changes between groups Blinder (1973) and Oaxca (1973) have proposed another widely used methodology.

The idea was to come up with an adjusted wage gap that would take into account

some of the differences in the productive characteristics of the two groups. Suppose that we have estimated our standard human capital model for mfg ,

ggggg XXEXw )|(ln

If we think that the proper counterfactual average wage is that women would have

earned at the male returns fm X , we would write

fmffmmfmfmffmmfm XXXXXXXww )()(lnln (2)

where the first term in the last equality capture the impact on the gender wage gap of difference in average characteristics of men and women, and the second term


measure differences dues to differential returns, sometimes called the unexplained part sometimes called the part due to discrimination.

On the other hand, we could think of the proper counterfactual average wage as the average wage that men would have earned at the female returns mf X , we would write

mmffmfmfmfffmmfm XXXXXXXww )()(lnln (3)

Alternatively, the pooled wage structured computed with a gender dummy from

equation (1’) provides a “regression-compatible” decomposition. The idea is to construct two counterfactual average log wages, the average log

wages that women would have earned at the pooled returns, 푋 훽, and the average log wages that men would have earned at the pooled returns , 푋 훽, the decomposition is then written as

ln푤 − ln푤 = 푋 − 푋 훽 + [푋 훽 − 훽 − 푋 훽 − 훽 ] (4)


where the first term captures the impact on the gender salary gap of differences in the average characteristics of men and women, evaluated at the pooled returns, and where the last term in bracket will correspond to the parameter α of equation (1). The sub-components of this last term can be interpreted as the advantage of men, 푋 훽 − 훽 , and the disadvantage of women, 푋 훽 − 훽 .

Generally, either choice of counterfactuals will give different results, so it is

common to report sensitivity of the decomposition to the choice of reference wage structure.

Table 10 Means and Partial Regression Coefficients of Explanatory Variables1) from Separate NLSY Log Wage Regressions

for Men and Women Ages 35-43 in 2000Means Female Male

Female MaleM2 M4 M2 M4

Coef. t-stat Coef. t-stat Coef. t-stat Coef. t-statRace

Hispanic (0,1) 0.182 0.193 0.063 2.57 0.060 2.61 -0.025 -1.02 -0.018 -0.75Black (0,1) 0.316 0.282 0.053 2.42 0.066 3.14 -0.022 -0.92 0.005 0.20

Education and skill level<10 yrs. 0.031 0.052 -0.089 -1.76 -0.078 -1.64 -0.028 -0.65 -0.025 -0.6010-12 yrs (no diploma or GED) * 0.103 0.124 --- --- --- --- --- --- --- ---HS grad (diploma) 0.300 0.326 -0.003 -0.10 -0.008 -0.27 -0.018 -0.65 -0.013 -0.50HS grad (GED) 0.045 0.056 -0.015 -0.34 -0.046 -1.12 0.027 0.63 0.015 0.38Some college 0.308 0.232 0.090 2.99 0.060 2.09 0.166 5.31 0.123 4.08BA or equiv. degree 0.153 0.155 0.276 7.61 0.216 6.19 0.373 10.23 0.260 7.08MA or equiv. degree 0.053 0.041 0.391 8.49 0.348 7.76 0.562 10.84 0.446 8.62Ph.D or prof. Degree 0.007 0.015 0.758 7.47 0.654 6.71 0.806 10.60 0.639 8.53

AFQT percentile score (x.10) 3.981 4.238 0.042 9.92 0.032 7.84 0.042 9.92 0.029 7.04

L.F. withdrawal due to family responsibilities (0,1) 0.549 0.130 -0.081 -4.16 -0.082 -4.46 -0.080 -3.14 -0.066 -2.74Lifetime Work Experience

Weeks worked in civilian job since age 18 ÷ 52 15.565 17.169 0.030 13.85 0.023 11.13 0.038 12.54 0.034 11.39Weeks worked in military since 1978 ÷ 52 0.062 0.573 0.046 3.53 0.040 3.22 0.025 5.15 0.020 4.46Weeks PT ÷ total weeks workd since age 22 0.137 0.050 -0.203 -4.24 -0.084 -1.81 -0.779 -7.90 -0.540 -5.70

Employment typeGov't employer (0,1) 0.215 0.144 -0.030 -1.50 -0.027 -1.13Non-profit employer (0,1) 0.100 0.049 -0.056 -2.13 -0.121 -3.20

OCC. Characteristics of Person's 3-digit OCC.SVP required in occup. (months) (DOT) 26.961 28.773 0.001 2.44 0.003 5.43Hazards (0,1) (DOT) 0.013 0.084 0.327 4.66 0.131 3.97Fumes (0,1) (DOT) 0.004 0.043 -0.293 -2.27 -0.075 -1.72Noise (0,1) (DOT) 0.080 0.307 0.005 0.18 0.019 0.83Strength (0,1) (DOT) 0.092 0.215 0.011 0.37 -0.049 -1.99Weather extreme (0,1) (DOT) 0.033 0.188 0.120 2.56 0.000 -0.01Prop. using computers (CPS) 0.557 0.415 0.157 2.19 0.045 0.49Prop. using computer for analysis (CPS) 0.143 0.139 0.497 4.62 0.258 2.22Prop. using computer for word proc. (CPS) 0.345 0.236 -0.255 -3.19 -0.007 -0.06Relative rate of transition to unemployment 0.772 1.092 -0.022 -1.11 -0.023 -1.91Relative rate of transition to OLF 1.046 0.789 -0.144 -7.30 -0.073 -3.57% female in OCC. X 0.1. (CPS ORG) 6.348 2.695 0.005 1.08 -0.019 -3.55

Adj. R-Square 0.392 0.464 0.403 0.467Dependent mean (Log Hourly Wage) 2.529 2.764Sample size 2704 26941) Model also controls for age, central city, MSA, region, and occupation missing.* Reference group.Source: National Longitudinal Survey of Youth (NLSY79) merged with measures of occupational characteristics (3-digit level) from the September 2001 CPS, the March CPS, the CPS ORG, and the Dictionary of Occupational Titles (1991).

Source: O'Neil and O'Neil (2005)

Nicole

Rectangle

Nicole

Oval

Table 11

Gender Wage Gap: Decomposition Results (NLSY, 2000)

Using male coefficients Using female coefficients

M1 M2 M3 M4 M1 M2 M3 M4

Log Wage Gap (Male-Female) Attributable to:Age, race, region, central city, MSA 0.0044 0.0112 0.0089 0.0089 0.0040 0.0089 0.0064 0.0064AFQT 0.0132 0.0107 0.0073 0.0074 0.0143 0.0107 0.0081 0.0081Education level -0.0138 -0.0128 -0.0094 -0.0096 -0.0147 -0.0068 -0.0054 -0.0052L.F. withdrawal due to family responsibilities 0.0335 0.0272 0.0277 0.0340 0.0344 0.0343Lifetime work experience 0.1425 0.1135 0.1116 0.0901 0.0649 0.0655Nonprofit, government 0.0088 0.0081 0.0048 0.0050

Occupational characteristics: Investment related

SVP (Specific Vocational Preparation) 0.0062 0.0053 0.0020 0.0021Computer usage 0.0122 -0.0040 -0.0054 -0.0024

Compensating differencesDisamenities (physical) 0.0167 0.0040 0.0252 0.0267Unemployment risk; labor force turnover 0.0116 0.0028 0.0226 0.0259

TYP: % female in occupation 0.0721 -0.0137

Unadjusted log wage gap 0.2351 0.2351 0.2351 0.2351 0.2351 0.2351 0.2351 0.2351Total explained by model 0.0037 0.1851 0.2030 0.2342 0.0036 0.1370 0.1578 0.1526Unexplained log wage gap 0.2314 0.0500 0.0321 0.0009 0.2315 0.0981 0.0773 0.0825

Unadjusted hourly wage ratio (Female/Male) : 79.0 79.0 79.0 79.0 79.0 79.0 79.0 79.0Adjusted hourly wage ratio (Female/Male) : 79.3 95.1 96.8 99.9 79.3 90.7 92.6 92.1

Note: Decomposition results shown are derived from results of separate regressions for men and women. See Table 10 for variable means and coefficients using Model 2 and 4. Wage ratios are based on the exponentiated log hourly wage.Source: National Longitudinal Survey of Youth (NLSY79) merged with measures of occupational characteristics (3-digit level) from the September 2001 CPS, the March CPS, the CPS ORG, and the Dictionary of Occupational Titles (1991).


Nicole

Rectangle

Nicole

Rectangle

Nicole

Rectangle

Table 4

Means and Partial Regression Coefficients of Explanatory Variables1) from Separate Log Wage Regressions for Black, White, and Hispanic MEN Ages 35-43 in 2000 (NLSY)

Mean White Black Hispanic

White Black Hisp.M1 M2 M1 M2 M1 M2

Coef. t-stat Coef. t-stat Coef. t-stat Coef. t-stat Coef. t-stat Coef. t-stat

Education and skill level<10 yrs. 0.043 0.041 0.093 -0.051 -0.68 -0.036 -0.49 0.069 0.80 0.024 0.30 -0.064 -0.81 -0.082 -1.0810-12 yrs (no diploma or GED) * 0.083 0.149 0.198 --- --- --- --- --- --- --- --- --- --- --- ---HS grad (diploma) 0.328 0.358 0.274 0.064 1.33 0.009 0.19 0.072 1.51 0.005 0.12 -0.007 -0.12 -0.063 -1.10HS grad (GED) 0.041 0.079 0.062 -0.018 -0.24 0.031 0.43 0.042 0.62 0.078 1.22 -0.080 -0.87 -0.077 -0.89Some college 0.216 0.239 0.264 0.236 4.42 0.215 4.13 0.205 3.76 0.151 2.89 0.085 1.32 0.068 1.11BA or equiv. degree 0.207 0.109 0.079 0.419 7.31 0.427 7.66 0.335 4.88 0.294 4.51 0.355 3.77 0.369 4.13MA or equiv. degree 0.059 0.021 0.019 0.524 7.14 0.561 7.84 0.634 5.29 0.624 5.48 0.465 2.94 0.484 3.23Ph.D or prof. Degree 0.023 0.004 0.012 0.645 6.50 0.780 8.00 1.302 5.07 1.359 5.58 0.593 2.95 0.774 4.02

AFQT percentile score (x.10) 5.538 2.411 3.360 0.046 7.63 0.039 6.49 0.058 6.68 0.048 5.80 0.059 6.04 0.046 4.91Lifetime work experience (Year equivalents)

Weeks worked in civilian job since age 18 ÷ 52 17.828 15.865 17.279 0.047 9.17 0.040 9.20 0.049 7.55Weeks worked in military since 1978 ÷ 52 0.483 0.835 0.436 0.033 4.31 0.028 4.00 0.036 2.89

Adj. R-Square 0.296 0.337 0.287 0.359 0.262 0.335Dependent mean (Log Hourly Wage) 2.898 2.559 2.700Sample size 1416 759 519

1) Model also controls for age, central city, MSA and region. The analysis is restricted to wage and salary workers employed within the past month. * Reference group.Source: National Longitudinal Survey of Youth (NLSY79).


Nicole

Rectangle

Table 5White-Black and White-Hispanic Wage Gaps: Decompositon Results for MEN (NLSY)

White-Black Differential White-Hispanic Differential

Using black male coef.

Using white male coef.

Using hispanic male coef.

Using white male coef.

M1 M2 M1 M2 M1 M2 M1 M2

Log Wage Gap Attributable to:Age, region, central city, MSA 0.0622 0.0589 0.0354 0.0334 0.0282 0.0292 -0.0004 -0.0079AFQT 0.1800 0.1504 0.1435 0.1204 0.1276 0.1001 0.1000 0.0839Education 0.0731 0.0714 0.0663 0.0713 0.0709 0.0741 0.0768 0.0771Lifetime work experience 0.0691 0.0810 0.0286 0.0275

Unadjusted log wage gap 0.3387 0.3387 0.3387 0.3387 0.1982 0.1982 0.1982 0.1982Total explained by model 0.3153 0.3499 0.2451 0.3061 0.2267 0.2321 0.1764 0.1805Unexplained log wage gap 0.0234 -0.0112 0.0936 0.0326 -0.0285 -0.0339 0.0218 0.0177

Unadjusted minority/white hourly wage ratio: 71.3 71.3 71.3 71.3 82.0 82.0 82.0 82.0

Adjusted minority/white hourly wage ratio: 97.7 101.1 91.1 96.8 102.9 103.4 97.8 98.2

Note: Decomposition results shown are derived from results of separate regressions for men ages 35-43 by race and by model using NLSY79 data from the 2000 survey. See Table 4 for variable means and coefficients. Hourly wages are the exponentiated hourly log wages.

Source: National Longitudinal Survey of Youth (NLSY79).


Nicole

Rectangle

Nicole

Rectangle


c. Variants of Oaxaca-Blinder

Firpo, Fortin and Lemieux (2009) propose an unconditional quantile regression (RIF regression) methodology that allows one to perform Oaxaca-Blinder decomposition for any quantile, such as the 90th percentile to detect glass ceiling effect or at the 10th percentile to detect sticky floor effect.

The RIF regression, E[RIF(Y ; qτ )|X] = X’γτ, is the regression of RIF(Y; qτ ) on X, where the RIF is the recentered influence function of the qτ quantile and has the property that EX(E [RIF(Y ; qτ |X = x]) = qτ.

Thus, in the same way as OLS estimates can be used to decompose the overall mean because conditional means (estimated by OLS) average up to the unconditional (overall) mean: E(Y)=EX[E(Y|X)], the RIF regression estimates can be used to compose quantiles or other distributional statistics.

The influence function provides the “influence”, or “contribution” of each data point to a statistic (mean, median, Gini, etc.)

o Widely used in robust statistics/econometrics: “influential observations” (outliers?) have a big impact on the statistic of interest


Running a regression of the influence function on the X’s then tells us the impact on

X on the influence function and then, in turn, on the statistic of interest.

In the case of the mean, the influence function is just Y-μ, thus the regression of (recentered) influence function for the mean , which is Y, on X is just a standard OLS regression (of Y on X). Influence Function: (more formal definition)

Start with distribution F(Y) and focus on distributional statistic ν(F). The IF is the directional derivative: IFν(Y)= lim ε→0 [ν((1-ε)F+ ε δy) - ν(F)] / ε where δy is a mass point distribution at Y

For the τth quantile qτ this is: IFτ(Y)= 1 [τ –1(Y≤ qτ) ]/f(qτ) where f(qτ) is the density of Y evaluated at qτ, and where 1(.) is the indicator function.


For example, for the median we have IF.5(Y)= -.5/f(q.5) if Y ≤ q.5 IF.5(Y)= .5/f(q.5) if Y > q.5

In practice, we run regressions of the recentered influence function, RIF, on X. This

way the average fitted value is equal to the statistic of interest Mean: RIFμ(Y) = μ + IFμ(Y) = μ + (Y-μ) = Y Quantile: RIFτ(Y)= qτ + IFτ(Y) = qτ + [τ –1(Y≤ qτ) ]/f(qτ) And then we can apply the Oaxaca-Blinder methodology to apportion parts of the group differences to various explanatory factors

A more computationally more demanding and less flexible methodology has been proposed by Machodo and Mata (2005) and implemented by Albrecht et al. (2003) to study glass ceiling effects (to be discussed later).

Reference Group: Male Coef.

A: Raw log wage gap : Qτ[ln(wm )]-Qτ[ln(wf)] 0.170 ( 0.023) 0.249 ( 0.019) 0.258 ( 0.026)

Estimated log wage gap: Qτ[ln(wm )]-Qτ[ln(wf)] 0.192 ( 0.015) 0.239 ( 0.016) 0.276 ( 0.026) Total explained by characteristics 0.257 ( 0.028) 0.198 ( 0.027) 0.143 ( 0.019) Total wage structure -0.065 ( 0.027) 0.041 ( 0.024) 0.133 ( 0.025)

Mean RIF gap: E[RIFτ(ln(wm))]-E[RIFτ(ln(wf))] 0.180 ( 0.023) 0.241 ( 0.019) 0.260 ( 0.026)Composition effects attributable to Age, race, region, etc. 0.015 ( 0.005) 0.013 ( 0.004) 0.002 ( 0.004) Education -0.011 ( 0.005) -0.017 ( 0.006) -0.005 ( 0.01) AFQT 0.005 ( 0.02) 0.013 ( 0.004) 0.013 ( 0.005) L.T. withdrawal due to family 0.022 ( 0.021) 0.042 ( 0.014) 0.039 ( 0.017) Life-time work experience 0.234 ( 0.026) 0.136 ( 0.014) 0.039 ( 0.023) Industrial Sectors 0.008 ( 0.012) 0.020 ( 0.008) 0.047 ( 0.011) Total explained by characteristics 0.274 ( 0.035) 0.208 ( 0.025) 0.136 ( 0.028)

Wage structure effects attributable to Age, race, region, etc. -0.342 ( 0.426) 0.168 ( 0.357) 0.860 ( 0.524) Education 0.023 ( 0.028) -0.030 ( 0.031) 0.023 ( 0.045) AFQT -0.007 ( 0.03) 0.003 ( 0.042) 0.008 ( 0.062) L.T. withdrawal due to family -0.075 ( 0.032) -0.005 ( 0.025) 0.018 ( 0.032) Life-time work experience 0.084 ( 0.148) -0.085 ( 0.082) -0.078 ( 0.119) Industrial Sectors 0.015 ( 0.06) -0.172 ( 0.046) -0.054 ( 0.052) Constant 0.208 ( 0.349) 0.154 ( 0.323) -0.653 ( 0.493)Total wage structure -0.094 ( 0.044) 0.033 ( 0.028) 0.124 ( 0.036)

Table 4. Gender Wage Gap: Quantile Decomposition Results (NLSY, 2000)

10th percentile 50th percentile 90th percentile

Note: The data is an extract from the NLSY79 used in O'Neill and O'Neill (2006). Industrial sectors have been added to their analysis to illustrate issues linked to categorical variables. The other explanatory variables are age, dummies for black, hispanic, region, msa, central city. Bootstrapped standard errors are in parentheses. Means are reported in Table 2.

B: Decomposition Method: Machado-Mata-Melly

C: Decomposition Method: RIF regressions without reweighing

Nicole

Rectangle


For cross-country comparisons, Blau and Kahn (1996) have proposed to use JMP’s approach to incorporate differences in the residual differences.

)()()()(lnln kikikikikkiiki XXXww (3)

where i and k index countries (with k denoting the benchmark country), ¯ and ∆ refer to country averages and differences between men and women, respectively, X for the matrix of observable endowments and characteristics, β for the vector of estimated coefficients from the male regressions, ε for the residuals from these regressions and η for the theoretical residuals that would be obtained in country i if it had the same residual wage structure as country k.

This decomposition assumes that the ranking of individuals reflects the distribution of unobserved characteristics, and that the distribution of unobserved characteristics in the male population is the same in all countries.

o These are very strong assumptions


The latter ik are obtained by calculating for each individual of country i the residual that an individual with the same ranking position with respect to the distribution of male full-time wage and salary employees would have in the benchmark country k.

The effect of country fixed effects has to be netted out from the wage gap in the

benchmark country before implementing the decomposition described in equation (3). This way, the cross-country average of each term of the equation is approximately equal to zero.

The first term on the right-hand side represents the contribution of cross-country

differences in gaps in observed characteristics to the gender (or family) wage gap, netting out the effect of cross-country differences in market prices for these characteristics that is reflected in the second term.

The sum of the third and the fourth term represents cross-country differences in the

residual and is split into the effects of cross-country differences in unobserved characteristics (third term) and cross-country differences in their market prices (fourth term).

OECD EMPLOYMENT OUTLOOK – ISBN 92-64-19778-8 – ©2002

– Women at work: who are they and how are they faring?104

(cf. Annex Table 2.B.1 for estimation results). It must be noted that decomposition out-comes are only partially robust to the choice of the benchmark country (Blau and Kahn,1996) and a different choice might lead to somewhat different results from those presentedbelow. Similar problems arise as regard to the choice of the reference group for categor-ical variables (Oaxaca and Ransom, 1999).

Before proceeding further with the examination of the decomposition results, thereader deserves some guidance to their interpretation. On the basis of the available evi-dence, it is not possible to determine whether the residual term can be ascribed only togender differences in unobserved characteristics and/or in their remuneration or rather tolabour market discrimination.25 However, comparing the full decomposition with onefocussing on the first and second terms of equation [2] only, thus leaving the residualunexplained, provides estimates of upper and lower bounds to the effect of gender gaps inproductive characteristics and the effect of the wage structure. This comparison is high-lighted in Chart 2.7, which presents three different measures of the gender wage gap:i) the unadjusted wage gap, defined as the percentage difference between male and femaleaverage gross hourly wages; ii) the wage gap adjusted for cross-country differences inremuneration rates for observed characteristics, computed by subtracting the first term onthe right-hand side of equation [2] from the unadjusted wage gap; and iii) the wage gapadjusted for cross-country differences in the whole wage structure, computed by subtract-ing both the first and third terms of the right-hand side of equation [2]. This way, the

Chart 2.7. The gender wage gap adjusted for the effect of the wage structurea

Percentage difference between male and female average gross hourly wages, persons aged 20 to 64 yearsb

a) The gender wage gap adjusted for cross-country differences in the remuneration rates of observed characteristics is obtained as follows:, where i indexes countries, k denotes the benchmark country, ¯ and ∆ refer to country averages

and differences between men and women, respectively, W stands for gross hourly wages, X for the vectors of observed characteristics, andβ for the vector of estimated coefficients from the male wage regressions (cf. Annex Table 2.B.1). The gender wage gap adjusted for cross-country differences in the whole wage structure is obtained as follows: ,where ε stands for the residuals from the male wage regressions (defined as the difference between actual and predicted values) and η forthe theoretical residuals that would be obtained in country i if it had the same residual wage structure as country k.

b) Countries are ranked by decreasing hourly wage gap adjusted for the whole wage structure.Sources and definitions: See Annexes 2.A and 2.B respectively.

% %30

0

25

20

15

10

5

30

0

25

20

15

10

5

Netherl

ands

Austria

German

yFra

nce

Portug

al

United

Kingdo

mFin

land

Irelan

d

Denmark

Belgium

Greece

Spain

Italy

Hourly wage gap adjusted for the remuneration rates of observed characteristics

Hourly wage gap

Hourly wage gap adjusted for the whole wage structure

% %30

0

25

20

15

10

5

30

0

25

20

15

10

5

Netherl

ands

Austria

German

yFra

nce

Portug

al

United

Kingdo

mFin

land

Irelan

d

Denmark

Belgium

Greece

Spain

Italy


Hourly wage gap


% %30

0

25

20

15

10

5

30

0

25

20

15

10

5

Netherl

ands

Austria

German

yFra

nce

Portug

al

United

Kingdo

mFin

land

Irelan

d

Denmark

Belgium

Greece

Spain

Italy


Hourly wage gap


)(loglog kiiiadjobs

i XWW ββ∆∆∆ −−=

ikkiikikiiiadj

i XXWW η∆β∆η∆ε∆ββ∆∆∆ +=−−−−= )()(loglog

Nicole

Oval

Nicole

Oval


Hence, the sum of the second and fourth terms represents the total effect of cross-country differences in the wage structure, for given gender gaps in characteristics.

Conversely, the sum of the first and third terms represents cross-country differences

in the ender (or family) wage gap adjusted for the whole wage structure. As with the other JMP decomposition, this requires very strong rank preserving

assumptions across wage structures. The OLS wage regression that is at the heart of the JMP technique provides a model

for the conditional mean of the wage distribution and its results do not extend naturally to wage quantiles.


Basic readings: Cain, Glen. “The Economic Analysis of Labor Market Discrimination: a Survey,” in

Ashenfelter, O.C. and R. Layard, editors, Handbook of Labor Economics, North-Holland, vol 1, 1986, pp.709-730

Altonji, J. and R. Blank “Race and Gender in the Labor Market” Chapter 48 in Ashenfelter and Card Handbook of Labor Economics, vol 3C (Elsevier, North Holland, Amsterdam 1999) pp. 3143-3259

O’Neill, J. and D. O’Neill, “What Do Wage Differentials Tell Us about Labor Market Discrimination?” NBER Working Paper 11240 (April 2005).

IV. Differences across Groups - Faculty of...

Documents

Transcript of IV. Differences across Groups - Faculty of...