On Racial Discrimination in Hiring

When Quillian et al. (2017) meta-analyzed the research on hiring discrimination they found that black applicants received 36% fewer call backs than white applicants even when the submitted applications were identical in every way other than the race implied by the applicant’s name. This sort of evidence is often pointed to as among the most compelling empirical justifications for the view that racism holds back minorities in the American economy. On its face, this research is indeed persuasive, but upon closer examination it turns out to be very unconvincing .

In order to interpret this evidence as a demonstration of racism, we must assume that black and white applicants with the same resume qualification are equally productive employees. If white applicants are, on average, better employees than black applicants with the same resumes, then there will be a non-racist reason for employers to prefer white employees over black ones even when they have the same qualifications.

There are several reasons to think that black employees will be less good employees than equally qualified white ones and absolutely no reason to think otherwise. The first reason is statistical in nature.

Job Performance and Distributions

Ideally, employees will be hired based on their ability to perform in their job. On average, Black Americans score .35 standard deviations below whites on measures of job performance.

4

Roth et al. (2003)

Some might worry that subjective measures of job performance will be racially biased, but these racial differences are even larger if job performance is measured via an objective criterion rather than via the subjective impressions of supervisors and employers.

5

Roth et al. (2003)

For “call back” studies to be valid measures of discrimination, these differences in job performance must disappear once we control for the sorts of qualifications one finds on a resume. Because variables like education, job experience, and reference checks only weakly correlate with job performance, this is extremely unlikely to be true.

Work.jpg

(Roth et al., 2005Schmidt and Hunter, 1998)

Actually, it almost surely isn’t true due to some simple statistical reasons. Suppose that the distributions of job performance among blacks and whites consist of two overlapping normal distributions, like this:

2Now suppose that qualifications on a resume require a certain level of skill and ability to obtain such that bad employees cannot easily acquire them. As is hopefully evident in the below example, there is no possible threshold for job performance, or any other relevant trait (e.g. job related knowledge, cognitive ability, self discipline, etc.,) where the white mean is above the black mean in general, but not still above the black mean aong those who exceed that threshold. 
3Because most of the variables typically contained within a resume are not direct measures of job performance but, rather, qualifications that anyone can acquire passed a certain threshold of ability and determination, it is almost certain that white job performance still exceeds black job performance even among those with the same qualifications. It is therefore rational, and not racist, to prefer white employees over black ones when considering applicants with the same qualifications.

The Limits of Resumes

Employers will also have reasons to prefer white employees if whites differ from blacks on some trait relevant to job performance that is not perfectly correlated with the information contained on resumes. As noted above, cognitive ability is among the best predictors of job performance, and there is no direct measure of cognitive ability contained in the typical resume. Something similar could be said of work ethic or conscientiousness.

Employers have no reason to assume that people are the same with respect to these variables just because they are the same with respect to the variables measured on resumes. Large meta-analyses show that racial groups differ in variables like cognitive ability, so this is a plausible non-racist explanation for the results of call back experiments (Roth et al., 2001).

Affirmative Action

Non-racist incentives for discriminating in hiring will be even more powerful if we allow black people to acquire the same qualifications as white people at lesser thresholds than what is required for white people. In this scenario, among applicants with any such qualifications, white job performance will exceed black people’s job performance among those with equal qualification even if there is no mean difference in job performance between black and white people in the total population. (There is.)6This situation is exactly what happen when a society institutes affirmative action and affirmative action for African Americans is widespread in contemporary America. For instance, in elite American law schools black students score 2.3 standard deviations below white students, implying that the average black student is comparable to the 10.7th percentile of the white students (Sander, 2004). Black students at elite schools even score lower than do white students in mid-range private schools.

1.jpg

The obvious implication of this is that there is a rationale for employers not only to prefer a white employee when qualifications are matched but, in fact, to prefer a white student from a mid-range law school to a black student with a law degree from a more elite university.

We see a similar situation with respect to students in medical school. A majority of black applicants get into medical schools with grades that would leave white applicants with only an 8% chance of acceptance.

2.jpg

Similarly, in the admissions offices of elite universities African Americans are given a boost roughly equivalent to a 230 point increase in their SAT score and are several times more likely than whites to be admitted given the same qualifications (Espenshade et al., 2004).

3

This was famously shown by Princeton sociologist Thomas Espenshade using data from the 1980s and 90s. This work is somewhat dated and US laws concerning college admission policies have changed with time. Some people doubt that admissions offices still practice this kind of affirmative action. To answer such skepticism, I was able to find 20 studies utilizing more recent data. Aggregated, they suggest that, when comparing people of equal qualifications, Black applicants are roughly 21 times more likely to be admitted, while Hispanics are 3 times as likely, and Asians are 6% less likely.

(The race columns show the odds of admission compared to those of white applicants when qualifications are held constant.)

Citation School Type Black Hispanic Asian
Nagai (2008) Arizona State Law 1115.43 84.95 2.18
Lerner and Nagai (2002) University of Virginia Law 730.8 1.1 1.86
Nagai (2008) University of Nebraska Law 442.39 89.63 5.78
Armor (2004) William and Mary Law 267 0.66 0.66
Nagai (2008) University of Arizona Law Law 250.03 18.15 2.54
Lerner and Nagai (2002) William and Mary Law 167.51 2.47 3.29
Danielson and Sander (2014) Berkeley Law 121.6 18.2 1.6
Armor (2004) University of Virginia Undergrad 106 2.81 0.94
Nagai (2006) University of Michigan Undergrad 62.79 47.82 0.81
Lerner and Nagai (2002) University of Maryland Medical 20.63 2.51 0.68
Armor (2004) North Carolina State Undergrad 13 1.93 0.64
Lerner and Nagai (2001) SUNY Medical 9.44 4.08 0.76
Nagai (2011) Miami University Undergrad 7.99 2.16 2.14
Danielson and Sander (2014) UCLA Undergrad 5.15 1.92 0.85
Lerner and Nagai (2006) US Naval Academy Military 4.44 3.32 0.67
Lerner and Nagai (2001) University of Washington Medical 4.01 4.86 0.9
Nagai (2011) Ohio State Undergrad 3.33 4.3 1.47
Lerner and Nagai (2006) US Military Academy Military 1.94 1.2 0.68
Lerner and Nagai (2002) George Mason Law 1.13 1.09 1.74
Median All All 20.63 2.81 0.94
Mean All All 175.51 15.43 1.59

This data implies two important conclusions. First, it reinforces the fact that affirmative action leads to unequal requirements for whites and non-whites to acquire the same qualifications.  Secondly, it calls into question the sweeping generalizations about economic opportunity that are often made on the basis of discrimination in hiring.

Suppose it’s true that black people get 36% fewer call backs than white people and suppose that this is due to unfair discrimination. At the same time, black people have a 2100% advantage with respect to getting into a good college. This pro black bias is 58 times greater than the supposed anti-black bias. If people were given the choice between having to apply to a third fewer jobs and being 21 times more likely to get into an elite school, it isn’t obvious that they would chose the former. It therefore seems wrong to simply describe this situation as “white privilege”, as is often done.

Turing from education to employment, since the 1960s, all employees of the federal government and most employees of state and local governments, including contracted workers, have been required to engage in affirmative action programs aimed at increasing the prevalence of minorities in their work forces. As of 2017, there are 14 million government employees, accounting for roughly 9% of all workers in the economy.

Affirmative action is not mandated in the private sector but, just as in the case of education, it is done voluntarily by the vast majority of important US businesses. A 2017 report on all the companies in the S&P 100 found over 90% of them had engaged in diversity initiates and 75% of them had gone as far as setting specific hiring targets for minority employment. The same report found that such practices are rapidly gaining in popularity.

8We thus have three different theoretical reasons, one dealing with affirmative action, one concerning the limits of resumes, and another dealing with the inherent nature of qualifications and statistical distributions, to think that blacks and whites with the same job qualifications won’t exhibit the same level of job performance.

Direct Evidence on Qualifications and Performance

The most important evidence here comes from Roth et al. (2003) who meta-analyzed data from 19 previous studies and found that black employees scored .30 standard deviations lower than white employees on measures of job performance even when they were working the same job at the same organization. Now, as liberals are fond of pointing out, for a black applicant to get the same job as a white applicant the black applicant often needs to have greater qualifications. The conjunction of these two facts have three important implications.

First, employers should not treat job experience the same way when they see it on black and white resumes since the white applicant will have, on average, performed better at their previous job than the black applicant. This fact alone invalidates call back experiments.

Second, black employees with qualifications which are better than the qualifications of white employees, to the degree that this is typically true among people with the same job, still perform less well at their job.

Thirdly, black people are currently being hired into jobs at lower levels of job skill than what is being required for white applicants, meaning that if firm owners were economically self interested, they would discriminate more heavily against black applicants. Of course, this finding is what we would expect if firms were pursuing affirmative action policies in their hiring processes, either because they are not economically self interested or because they fear the social and legal consequences of discriminating against black applicants to the degree that they would have to such that black job performance equalized with that of whites hired into the exact same job. Thus, this evidence suggests that firms are discriminating in favor of black applicants rather than against them.

The most obvious response to this argument would be to claim that measures of job performance are racially biased since they often include supervisor ratings, but, as already noted, objective measures of job performance indicate larger racial gaps than do subjective ones, suggesting that supervisor ratings are biased in favor of black employees, and that the real gap in job performance among those working the same jobs at the same firms is probably larger than what was estimated by Roth et al. since their data included a mix of objective and subjective measures.

We also have a lot of evidence suggesting that large racial gaps in cognitive ability persist when comparing people of similar educational attainment. Since, as noted above, cognitive ability is the single best predictor of job performance with the possible exception of a structured interview, this is relevant to the general argument I’ve been making.

To begin with, in the early 1990s the federal government carried out a massive study to measure people’s work-related cognitive abilities in terms of things like everyday math skills, writing skills, and the ability to efficiently use information taken from a document.

456

In all three domains, African American who were either graduate students or who had graduate degrees scored below white Americans with the same level of education as well as white Americans that only had a bachelor’s degree. At the same time, white Americans with associate degrees scored higher in all three domains that did African Americans with bachelor degrees. Thus, if employers care about work related cognitive abilities, it is rational for them to prefer white applicants over black applicants when said applicants have the same qualifications and even when black applicants have marginally better qualifications.

Similar results are found in Neill (1990), a paper which shows the mean AFQT percentile scores of black and white men aged 19-21 by education level for the years 1953-1958 and 1980. (The AFQT is a test designed by the military to measure cognitive skills relevant to job performance such as reasoning ability, mathematical ability, and reading ability.)

7

In 1980, Black people who had completed 3-4 years of college came in, on average, at the 49.7th percentile, or slightly below the average score unconditional on education. By contrast, the average percentile scores were 80.2 for whites with 3-4 years of college, 65.8 for whites with 1-2 years of college, and 46.5 for whites with 3-4 years of high school.

This was even more extreme in the 50s, when black people with 3-4 years of college completed scored lower than whites with 3-4 years of high school. Thus, we once again see a clear rational for employers preferring white applicants over black applicants even when black applicants have higher levels of educational attainment.

This line of reasoning might be objected to on the grounds that we aren’t comparing blacks and whites who went to the same schools. It may be that the schools white people go to are more elite than the schools black people go to, so much so that the average high-school a white person goes to is more cognitively demanding than the average college a black person goes to, and this explains why we see these results, but this can’t explain hiring discrimination since such studies involve applicants saying that everyone went to the exact same school.

A lot could be said about this objection. First, there’s no evidence that this is true, and I’ve offered several reasons for thinking that even within the same school whites will average higher cognitive ability than non-whites. Second, while white people do go to better schools than black people, the degree to which this would need to be true to make this objection valid seems implausible. Thirdly, cognitive differences within the same school almost certainty exist since universities have lower SAT requirements for minority students. Data from Temp (1971) is relevant to this conjecture, at is shows that even in the 1960s there was a significant gap in mean SAT score by race within all 13 universities for which Temp had data. In most universities, this gap was almost an entire standard deviation in size.

7

The same pattern was found when Goldman and Richards (1974) compared the SAT scores of white and Hispanic Americans attending the University of California, Riverside.

Thus, the data on cognitive ability is consistent with the data on job performance in suggesting that qualifications cannot be interpreted the same way across races, and both educational and economic institutions require less of black applicants than they do whites. Since black and white people compete for the same positions in economic and educational institutions, this implies widespread discrimination against whites in favor of blacks rather than the other way around.

More Problems

The way in which the results of call back experiments have changed with time provides further evidence against the view they they are measuring racism. Quillian et al.’s meta-analysis finds that call back discrimination against blacks has increased over time since the 1970s.

8

This is not what we would expect if application discrimination was due to racism since racism against blacks has obviously fallen since the 1970s. By contrast, this trend is easy to explain if application discrimination is partly a response to affirmative action since the proportion of African Americans who have benefited from affirmative action has surely increased since the 1970s.

Now, so far, I’ve been discussing these experiments as if they were being reported accurately in the research literature. This is almost surely not true. To understand how this is discernible, you need to understand the basic idea of testing a distribution of effects for publication bias.

Imagine that we had 10 groups of people flip a coin 10 times each. We then recorded the proportion of times each group got tails. We could use an elementary understanding of probability to figure out the likelihood different distributions of these proportions. We don’t need to do any math here to see what I mean. For instance, we now that the most probable outcome would be 50% tails, but if every group got exactly 50% tails we would surprised because this outcomes is highly unlikely. Instead, what we would expect is that something close to 50% would be the most common proportion, and proportions would be ever less common the further away we moved from 50%.

Using similar reasoning, we know that we should see studies converge on a single result as their statistical precision increases and that effect sizes should be symmetrically distributed around the point that is being converged upon. Sometimes, symmetry is not found, usually because a bunch of negative findings are missing where we’d expect them to be given what the positive findings look like. This is a sign of publication bias, meaning that the negative results are simply going unpublished. Nature has a nice graphic illustrating the idea:

Image result for funnel plot

Now, the distribution of effects found in the literature on hiring discrimination exhibits a pattern so suggestive of publication bias it is almost comical. Looking at these charts, it seems almost certain that the actual degree of hiring discrimination that takes place is lesser than what Quillian’s meta-analysis implies.

Comment Q2017 Figure 1

Summary

In summary, call back experiments are invalid because we have no reason to think that equally qualified blacks and whites will be equally productive employees. Actually, because of the way in which thresholds works with normal distributions, and because of affirmative action, we have reason to think that whites will be better employees than blacks when qualifications are held constant. This is what the direct data on job performance, as well as the literature on cognitive ability and educational attainment, suggests. This model explains why hiring discrimination against African Americans has increased with time while the “racism model” cannot. This is all assuming that the research literature has been reported honestly and it clearly hasn’t. The actually degree of bias against black Americans is less than what the published research implies, and this is unsurprising since a close look at the relevant literature actually suggests widespread discrimination in favor of blacks rather than against them.

Advertisements

9 thoughts on “On Racial Discrimination in Hiring

  1. Pingback: “Lived Experience” is Not Evidence | Ideas and Data

  2. I love this. You’re actually making the case that it’s OK to be racist, because blacks are actually really shitty (ie., genetically inferior). Wow. I suppose all of the housing studies also go out the window now?

    There’s NO TRAIT you can compare between blacks and whites that’s not hopelessly confounded by environment. None. Regressing out SES or some such are poor attempts at addressing this. And this has nothing to do with the degree that G is heritable across individuals, we’re talking about a group comparison. That being the case, arguing that it’s OK to not hire or house blacks because they’re inferior is basically a vicious racist circle that they can never exit. To use a gish gallop data to give the impression of genetic cause and effect as your site does over and over is simply nonsense. It’s dressing up a lot of big words and statistical concepts to give the impression that the fatal flaws have been overcome when they have not.

    When you can show behavioral differences (IQ here) linked to genes, and those genes can be shown to vary by group, you’ll have something. Until that point this is bigfoot “science”. A blurry image and a lot of talk about “plausability” because there’s food for bigfoot to eat in the forest, ergo he exists. To which I reply: go get me some bones.

    Like

  3. I was thinking about this where inference from data is concerned in a psychometric consulting job for a client and this is the piece that a lot of you seem to be missing.

    Let’s say that we have a psychometric test that, in part, measures some latent variable (let’s call it G!).
    (1) G appears to be, in part, heritable.
    (2) We observe stable differences in test scores on this psychometric between groups A and B.

    1 + 2 does not equal genetic differences between groups A and B. This should be obvious and would seem prima facie, as causes of variability on the individual and group levels are orthogonal, but I guess peopl have problems with this.

    The real problem seems to be that people ahve jumped into science where psychometrics and genetics (and geophysics – different motivated reasoning) are concerned to simply try and prove that their politics are “correct” – and they have difficulties with inferences we draw from data because they are not trained properly.

    Like

  4. Showed up here, but not on the other post? Weird. But I guess this isn’t the kind of stuff you discuss with people who aren’t “believers”. I think that’s how science works, right?

    So all the stuff the alt-right alway goes on about – dialogue, free exchange of ideas, facts and logic: BS, BS, and more BS. It’s nothing but a political product to be sold by any means necessary, including censoring opposing views in discussion spaces and lying, a la the information presented on your blog.

    Like

  5. HA! Really not gonna reply to reasonable criticism from a scientist, ha? OK. I’ll tell you what: you answer me and I’ll tell you who I am and we can debate in public. You don’t have to reveal yourself. What do you say?

    Like

  6. Pingback: Race and Educational Opportunity | Ideas and Data

  7. Pingback: How Racist are White Americans? | Ideas and Data

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s