11. Labour Market Discrimination

Transcription

1 IV. Wage Dierentials across Groups and Labour Market Discriination 11. Labour Market Discriination B. Evidence 1.Direct Evidence a. Audit and correspondence studies b. Field and lab experients 2. Indirect Evidence a. Estiating siple odels o wage deterination b. Methodologies or decoposing wage changes between groups c. Variants o Oaxaca-Blinder

2 11.B.Evidence 1.Direct Evidence: a. Audit and correspondence studies Audit and correspondence studies, as well as ield and lab experients have recently been used to detect racial and gender discriination in eployent and housing. Audit studies involve sending trained auditors who are identical in qualiications at prospective job interviews or housing openings while randoizing on race or gender. Correspondence studies involve sending ictitious resues with and without a perceived inority trait to potential eployers and coparing the outcoes or the ictitious applicants. The ost coon way to anipulate perceived inority status has been through the naes o the ictitious applicants, but soe studies have anipulated other traits such as location o study and experience (Oreopoulos, 2011).

3 Audit and correspondence studies in labour arket settings easure dierences in job oer rates or callbacks by eployers, but they do not show how uch o the dierence in wages between groups is due to discriination as copared with other actors. Riach and Rich (2002) note that the indings ro ield studies appear to be ore consistent with the ajority white populations having a general distaste or inorities, but that statistical discriination, or arketers using observable characteristics to ake statistical inerence about productivity or reservation values o arket agents, or exaple, cannot be ruled out, ex ante or ex post. These types o audit studies are crucially dependent on an unstated hypothesis: that the distributions o unobserved (by the testers) productivity characteristics o ajority and inority workers are identical. Because they are not double-blind, we ay suspect that inority testers ay unconsciously take actions that reduce their odds o receiving a job oer to conir the existence o discriination The set o eployers audited is likely to be randoly chosen Also saple sizes are invariably sall because audit studies are costly More careul studies attept to avoid this proble

4 Pager, Western, Bonikowski (2008) is a particularly careul audit study It used atched teas o testers who applied or 341 entry-level jobs in New York City over nine onths in 2004; they sent atched black, white, and Latino job seekers. To calibrate the agnitude o racial preerences, they copare inority applicants to white applicants just released ro prison ater serving 18 onths or a drug elony (possession with intent to distribute cocaine), and ind that White elons do about as well as Black and Hispanic applicants with no (reported) criinal record. To circuvent the proble o sall saple sizes, uch recent work in discriination has used correspondence studies Correspondence studies have convincingly deonstrated dierences in callback rates tied to eployer perceptions o gender (Firth, 1982; Riach and Rich, 1987, 2002), race (Bertrand and Mullainathan, 2004) and other attributes (Eriksson and Rooth, 2011; Krot et al., 2013). Bertrand and Mullainathan (2004) sent about 5,000 resues to eployers where the resues are identical except or a nae that is ale or eale or with white or black sounding naes.

5 Correspondence studies have also been criticized. For exaple, Fryer and Levitt (2004) argue that distinctively black naes, which were used in the Bertrand and Mullainathan study to signal race to eployers, signal ore than just race. However, Bertrand and Mullainathan (2004) has popularized the ethod and the nuber o correspondence studies in econoics has greatly increased (See Table 1 and 2 o Bertrand and Dulo (2016). It is also the case that, like audit studies, correspondence studies track outcoes that are an interediate step in the process that generates wage variation rather than inal outcoes. The expansion o online plators allows researchers to extent the correspondence ethod, at ties ollowing the transactions up to copletion They can ocus on custoer discriination and attept to identiy echanis at play Soe criticiss o experiental studies o discriination are that they typically exaine only interediate outcoes.

6 List (2004) perored an experient in the sportscard arket and ound that there is a strong tendency or inorities to receive initial and inal oers that are inerior to those received by ajorities, but overall, the data indicate that the observed discriination is not due to anius, but represents statistical discriination. Zussan (2013) studies ethnic discriination in the online arket or used cars in Israel. 1) He sent about 8,000 paired eails to sellers o second-hand cars, one coing ro Jewish-sounding nae and one ro Arab-sounding nae; the irst one was 22 percent ore likely to receive a response 2) In a ollow-up phone survey, he elicited sellers attitudes about inorities to tease out potential echaniss and ound that Jewish car sellers who strongly disagree with the stateent that the Arabs in Israel are ore likely to cheat than the Jews do not discriinate against the Arab buyer, while others sellers do. Doleac and Stein (2013) exaine the eect o race on arket outcoes by selling ipods through local online classiied advertiseents throughout the US. Each advertiseent eatures a photograph including a dark or light-skinned hand, or one with a wrist tattoo; the ipod is held by the hand o the apparent seller.

7 Black sellers receive ewer oers than white sellers, and the correspondence with black sellers indicates lower levels o trust. Black sellers receive average oers $5.72 (11%) lower and tattooed sellers $5.53 (10%) lower. Discriination in proessional sports has been used to study both consuer and eployee discriination Kahn and Sherer (1988) have ound that hoe attendance at NBA gaes is positively correlated with the raction o white players. Price and Wolers (2010) study NBA reereeing and ind that ore personal ouls are called against players when they are oiciated by an opposite-race reereeing crew than when oiciated by an own-race crew, even when conditioning on player and reeree ixed eects and speciic gae ixed eects. o Players earn up to 4% ewer ouls or score up to 2.5% ore points when they beneit ro a positive own-race bias Lab experients have recently been used to study preerence dierences between en and woen, ocusing on three actors that are relevant in the labor arket: risk taking, social preerences and reaction to copetition (Croson and Gneezy, 2004).

8 I woen preer jobs that are less risky, ore socially virtuous and less copetitive, then this could explain part o the gender dierences in the labor arket. Field experients include Goldin and Rouse (1996) which studied one o the cleanest test experient where orchestras where orchestras beginning in the 1970s and 1980s used a screen to hide an auditioning usician ro the jury. This increased the proportion eale aong new hires draatically. Direct productivity estiates A sall nuber o studies have attepted to provide direct estiates o ale and eale productivity and to copare ale and eale wages relative to productivity. Hellerstein, Neuark, and Troske (1999), or exaple, use ir level data ro the United States to estiate production unctions and the arginal productivity o eales relative to ales. o They ind that eales are soewhat less productive than ales but that the ratio o their wages is uch lower than the ratio o their productivity, suggesting a discriinatory pay gap.

9 2. Indirect Evidence: a. Estiating siple odels o wage deterination The siplest way o assessing the extent o discriination consists in an equation siilar to equation (1), ln wi = δ 0 + δ 0DDi + i + ε i (1 ) where D i = 1 denote that individual i belong to the inority group and D i = 0 is the oitted category. The coeicient δ 0D captures the wage disadvantage o the inority group that is not explained or accounted or by the productive characteristics. This ethod has the advantage that any potentially disadvantaged groups can be copared to the ajority group. One drawback is that is that it ay be diicult to separate the eect o various productive characteristics, or exaple education vs. experience.

10 b. Methodologies or decoposing wage changes between groups Blinder (1973) and Oaxca (1973) have proposed another widely used ethodology. The idea was to coe up with an adjusted wage gap that would take into account soe o the dierences in the productive characteristics o the two groups. Suppose that we have estiated our standard huan capital odel, we can write on average ln w = g g g ln wi = i g + εi or g =,, than This is because in population ters, taking the expectations over we get, EE EE ln ww gg gg = gg + EE εε gg where EE εε gg = 0 by assuption and where EE EE ln ww gg bb = EE(ln ww gg ) by the law o iterated expectations

11 I we think that the proper counteractual average wage is that woen would have earned at the ale returns, we would write w w ) ( ) ( ln ln = + = (2) where the irst ter in the last equality capture the ipact on the gender wage gap o dierence in average characteristics o en and woen, and the second ter easure dierences dues to dierential returns, soeties called the unexplained part soeties called the part due to discriination. On the other hand, we could think o the proper counteractual average wage as the average wage that en would have earned at the eale returns, we would write w w ) ( ) ( ln ln = + = (3)

12 Alternatively, the pooled wage structured coputed with a gender duy ro equation (1 ) provides a regression-copatible decoposition. The idea is to construct two counteractual average log wages, the average log wages that woen would have earned at the pooled returns,, and the average log wages that en would have earned at the pooled returns,, the decoposition is then written as ln ww ln ww = + [ ] (4) where the irst ter captures the ipact on the gender salary gap o dierences in the average characteristics o en and woen, evaluated at the pooled returns, and where the last ter in bracket will correspond to the paraeter α o equation (1). The sub-coponents o this last ter can be interpreted as the advantage o en,, and the disadvantage o woen,. Generally, either choice o counteractuals will give dierent results, so it is coon to report sensitivity o the decoposition to the choice o reerence wage structure.

13 c. Variants o Oaxaca-Blinder Firpo, Fortin and Leieux (2009) propose an unconditional quantile regression (RIF regression) ethodology that allows one to peror Oaxaca-Blinder decoposition or any quantile, such as the 90 th percentile to detect glass ceiling eect or at the 10 th percentile to detect sticky loor eect. The RIF regression, E[RIF(Y ; qτ ) ] = γτ, is the regression o RIF(Y; qτ ) on, where the RIF is the recentered inluence unction o the qτ quantile and has the property that E(E [RIF(Y ; qτ = x]) = qτ. Thus, in the sae way as OLS estiates can be used to decopose the overall ean because conditional eans (estiated by OLS) average up to the unconditional (overall) ean: E(Y)=E[E(Y )], the RIF regression estiates can be used to decopose quantiles or other distributional statistics. The inluence unction provides the inluence, or contribution o each data point to a statistic (ean, edian, Gini, etc.) o Widely used in robust statistics/econoetrics: inluential observations (outliers?) have a big ipact on the statistic o interest

14 Running a regression o the inluence unction on the s then tells us the ipact on on the inluence unction and then, in turn, on the statistic o interest. In the case o the ean, the inluence unction is just Y-μ, thus the regression o (recentered) inluence unction or the ean, which is Y, on is just a standard OLS regression (o Y on ). Inluence Function: (ore oral deinition) Start with distribution F(Y) and ocus on distributional statistic ν(f). The IF is the directional derivative: IFν(Y)= li ε 0 [ν((1-ε)f+ ε δy) - ν(f)] / ε where δy is a ass point distribution at Y For the τth quantile qτ this is: IFτ(Y)= 1 [τ 1(Y qτ) ]/(qτ) where (qτ) is the density o Y evaluated at qτ, and where 1(.) is the indicator unction. For exaple, or the edian we have IF.5(Y)= -.5/(q.5) i Y q.5 IF.5(Y)=.5/(q.5) i Y > q.5

15 In practice, we run regressions o the recentered inluence unction, RIF, on. This way the average itted value is equal to the statistic o interest Mean: RIFμ(Y) = μ + IFμ(Y) = μ + (Y-μ) = Y Quantile: RIFτ(Y)= qτ + IFτ(Y) = qτ + [τ 1(Y qτ) ]/(qτ) And then we can apply the Oaxaca-Blinder ethodology to apportion parts o the group dierences to various explanatory actors A ore coputationally ore deanding and less lexible ethodology has been proposed by Machodo and Mata (2005) and ipleented by Albrecht et al. (2003) to study glass ceiling eects (to be discussed later).

16 For cross-country coparisons, Blau and Kahn (1996) have proposed to use the (Juhn, Murphy, and Pierce, 1993) JMP s approach to incorporate dierences in the residual dierences. ln wi ln wk = i( i k ) + k ( i k ) ( ε i ηik ) + ( ηik ε k ) (3) where i and k index countries (with k denoting the benchark country), and reer to country averages and dierences between en and woen, respectively, or the atrix o observable endowents and characteristics, or the vector o estiated coeicients ro the ale regressions, ε or the residuals ro these regressions and η or the theoretical residuals that would be obtained in country i i it had the sae residual wage structure as country k. This decoposition assues that the ranking o individuals relects the distribution o unobserved characteristics, and that the distribution o unobserved characteristics in the ale population is the sae in all countries. o These are very strong assuptions

17 The latter η ik are obtained by calculating or each individual o country i the residual that an individual with the sae ranking position with respect to the distribution o ale ull-tie wage and salary eployees would have in the benchark country k. The eect o country ixed eects has to be netted out ro the wage gap in the benchark country beore ipleenting the decoposition described in equation (3). This way, the cross-country average o each ter o the equation is approxiately equal to zero. The irst ter on the right-hand side represents the contribution o cross-country dierences in gaps in observed characteristics to the gender (or aily) wage gap, netting out the eect o cross-country dierences in arket prices or these characteristics that is relected in the second ter. The su o the third and the ourth ter represents cross-country dierences in the residual and is split into the eects o cross-country dierences in unobserved characteristics (third ter) and crosscountry dierences in their arket prices (ourth ter). Hence, the su o the second and ourth ters represents the total eect o cross-country dierences in the wage structure, or given gender gaps in characteristics.

18 Conversely, the su o the irst and third ters represents cross-country dierences in the ender (or aily) wage gap adjusted or the whole wage structure. As with the other JMP decoposition, this requires very strong rank preserving assuptions across wage structures. The OLS wage regression that is at the heart o the JMP technique provides a odel or the conditional ean o the wage distribution and its results do not extend naturally to wage quantiles.

19 Basic readings: Altonji, J. and R. Blank Race and Gender in the Labor Market Chapter 48 in Ashenelter and Card, Handbook o Labor Econoics, vol 3C (Elsevier, North Holland, Asterda 1999) pp O Neill, J. and D. O Neill, What Do Wage Dierentials Tell Us about Labor Market Discriination? NBER Working Paper (April 2005).