How the IQ Taboo Promotes Statistical Discrimination

Last week I wrote a post on the importance of acknowledging the existence of a racial IQ gap. In the comments, I mentioned that I believe that the taboo on the IQ gap promotes statistical racial discrimination. The reason for this is that IQ scores tend to predict job performance reasonably well. All else being roughly equal, it is in an employer's best interest to hire the candidate with the highest IQ.

But because of the IQ taboo, IQ tests are presumed to be racist, and employers are barred from using them in most cases. Hence an employer who wants to select the smartest candidate is forced to rely on some proxy for IQ. Here blacks are doubly disadvantaged. First, race is a proxy for IQ. Given two job candidates, it's a good--albeit far from sure--bet that the white candidate is smarter than the black candidate. On top of that, an employer who attempts to gauge candidates' intelligence through non-technical interviews (technical interviews are often thinly veiled IQ tests) may unintentionally end up underestimating the intelligence of black candidates due to cultural factors such as the use of black vernacular English and/or subconscious stereotyping.

In short, when employers are denied the right to use IQ tests to gauge candidates' intelligence, racial discrimination becomes a +EV strategy.

One might object that most employers don't know about the IQ gap. This is probably true. But it's likely that many have noticed a performance gap. If an employer notices that his black hires have more often than not underperformed his white hires, or that this was true of his coworkers before he opened up shop himself, he will be inclined to discriminate on the basis of race. Worse, if he remains in the dark about the explanation for this phenomenon, he may drift towards outright bigotry.

Another objection one might raise is that education serves as a good proxy for IQ. This is true, but it is an imperfect proxy made even more imperfect by affirmative action. At many elite universities, the median black SAT score is 200-300 points below the median white SAT score--a difference of 1-1.5 standard deviations. Because blacks face lower admissions standards than whites at schools which practice affirmative action, it's a reasonable--though again not sure--bet that a white candidate is more intelligent than a black candidate who graduated from the same college.

That said, I suspect that employers hiring college-educated workers are generally better able to measure the qualifications of job candidates, so I would expect racial discrimination to be less attractive to such employers. Consider instead the portion of the labor force which has only a high school diploma. Assume that this consists of everyone with an IQ between 80 and 105. If the median black IQ is 85 and the median white IQ is 100, this means that the IQs of white high school graduates will tend to cluster at the high end of that range, while the IQs of black high school graduates will tend ot cluster at the low end. Although the IQ gap is somewhat narrower than it is in the full population, it still persists--perhaps it is 10 points rather than 15--and it still pays to discriminate.

Statistical discrimination is how we make judgments when we lack the information necessary to judge others as individuals. The antidote is not obscurantism, but enlightenment--the easier it is to judge others as individuals, the less statistical discrimination there will be.

Share this

Reverse affirmative action

At GNXP they say that's not quite enough. Eliminating statistical discrimination would require reverse affirmative action.

Also, I don't know if interviews would be that disadvantageous for blacks. Rushton claims they have a winning personality not captured by tests. I've heard it speculated that in the future machines/outsourcing will do most of our thinking for us and the remaining jobs will be ones that rely on good interpersonal skills, which if Rushton is right could mean the Supreme Court's dream of affirmative action becoming unnecessary could be realized.

Maybe I'm missing something

Maybe I'm missing something in the linked GNXP post, but the point seems to be rather obvious: a college diploma is not as accurate an indicator of intelligence as an IQ score. Well, duh. And the post doesn't conclude that eliminating statistical discrimination would require reverse affirmative action; that would only be true if the presence or absence of a college diploma was the only piece of information an employer could legally obtain about a potential employee. Rather, the goal of eliminating statistical discrimination could be achieved by more accurate employee recruit testing.

And speaking of winning personalities, citing Rushton as a legitimate source doesn't make me think very highly of yours.

Statistical discrimination really unjust?

Granted it might be in some cases, and granted it may depend on your definition of "unjust", but let me define justice as follows: the greater the degree of error in each individual case, the greater the injustice. Add up the errors to get the total degree of injustice. A procedure that reduces total injustice is arguably not any more unjust than an alternative procedure that increases the total injustice.

Procedure 1: Suppose you have two populations, white and black youths, and suppose black youths are significantly more criminal than white youths. Suppose that when dealing with an individual youth, you first identify his race, and then treat him as if he were as criminal as the mean of his race. Then for any given white youth and any given black youth, you will treat the black youth as if he is significantly more criminal than the white youth.

Procedure 2: Suppose you are race blind and treat all individual youths of any race as if they were as criminal as the mean of all youths. Then for any given white youth and any given black youth, you will treat the black and the white youth as if they are equally criminal.

It is likely, probable even, that Procedure 1 reduces the average error that you will commit in dealing with youths, as compared with Procedure 2. This is despite that fact that, were we to carefully select a white youth and a black youth who are equally criminal, they would be treated unequally. Procedure 1 is likely to produce less error despite that case, because that is not the only type of case.

To be sure, it it not always

To be sure, it it not always true. If there are 50% white, 50% black everyone is equally criminal except for one black guy who is extremely criminal and one white guy who is a saint, statistical discrimination produce more injustice than non discrimination. (I know you just said likely).

This exemple may seem ad-hoc but it can be made more realistic by thinking of a small criminal subset of the black population that would bring an overall bad image, an argument that has probably already been made.

The way to bet

I acknowledged this at the start of the comment. I think, nevertheless, that it is more likely than not that stat disc reduces error usually.

Mean versus median

I was restricted in my previous reply by the cumbersome way the iPhone deals with commenting on this blog. To expand further, in the scenario you envision I would say the real source of non-mimization of error is not statistical discrimination, but the difference between median and mean. You essentially produced a situation where the whole dataset is not skew, but each of the two subsets are skew (median and mean are not the same). If you simply change "mean" to "median" in my own presentation, then your counterexample ceases to work, because the median guess is the guess that minimizes average error (I think). At a guess, mean might minimize the average of the square of error, but don't quote me on that.

Meanwhile, the rational (and entirely justified) way of going about this is not to use either the median or the mean, but instead to use whatever assumption optimizes your own outcome. That will depend on the cost to you of various possible errors.

Top-notch debating being

Top-notch debating being delivered via iPhone? We have the best commenters ever!

but let me define justice

but let me define justice as follows: the greater the degree of error in each individual case, the greater the injustice.

Why define it that way? Why not define justice as fairness? Or justice as [fill in the blank]?

Suppose you didn't know whether you would be born white or black. Would you prefer Procedure 1 or Procedure 2?

The problem with your argument is that it is circular. You associate justice with predictive value, and then conclude that statistical discrimination is just. Well of course it is! That's the whole point of statistical discrimination: that in the absence of perfect information, we use whatever limited information we have to make assumptions. But the predictive usefulness of these assumptions does not necessarily make these assumptions just or fair.


Why define it that way?

Because it makes sense. If someone commits a crime, then the pursuit of justice is the pursuit of the actual criminal, not the pursuit of an entirely innocent person. It is just if an innocent be treated as if he were innocent and that the guilty be treated as if he were guilty. Surely it is not just for an innocent person to languish in jail for a crime he did not commit. Conversely, surely it is not just for a criminal to get away with his crime. Both of these are errors.

Why not define justice as fairness?

It is void for vagueness. Furthermore, if you defined "fairness" clearly, my guess is that I could demonstrate that it was not well-defined, giving rise very quickly to massive contradictions.

One kind of fairness is sameness of treatment. But there are different kinds of sameness. For example, one way of treating a murderer and an innocent person "the same way" is to send them both to jail. Another way to treat them "the same way" is to treat them in the same conditional way: if-you-murder-then-you-go-to-jail-otherwise-not. And there are any number of other ways of treating two people "the same way".

No, Constant, it doesn't

No, Constant, it doesn't make sense, and your response completely ignores the heart of my criticism. Racial profiling may have some predictive value in a world where the only factor we have access to is race, but predictive value is not the same as justice. Treating people as presumed guilty because of their race and not their individual characteristics is abhorrent to justice.

Let me qualify myself and

Let me qualify myself and retract a bit. The more I think about the issue of statistical discrimination, the more conflicted I become. I certainly don't think it should be legally prohibited. Whether it deserves moral opprobrium is less clear; on the one hand, it seems tremendously unfair to be judged negatively by others, not for your own individual merits, but for the demerits of your race, gender, etc.; on the other hand, we live in a world of imperfect information, where statistical discrimination is often a necessity. I'm thinking specifically of the case of insurance, a business model based entirely on imperfect information. I'll have to give this issue a lot more thought. If anyone reading this comment knows of any good articles or books on the ethics of statistical discrimination, I'd appreciate your recommendations.


Treating people as presumed guilty because of their race and not their individual characteristics is abhorrent to justice.

I cannot accept your axiom because it is unworkable. Suppose that you witness a murder in locked room (a really locked room, not one of those tricky "locked rooms" of detective fiction). All you see (through a window, peephole, whatever) is that a black man murdered an asian man. You can't identify individuating features. You open the room. Inside, you see Tom, Dick, and Harry. Tom is a dead asian man lying on the floor. Dick is a black man standing in the room. Harry is a white man standing in the room. Based on what you saw, it is entirely reasonable for you to conclude that Dick is the murderer. But all you have to go on is the murderer's race.

So, it's not abhorrent to justice to consider someone guilty based on no visible evidence other than their race. I raised the probability to one in this scenario to drive home the point, but we commonly convict based on less than unity probability; if employing race is okay in a probability-one scenario, why not a lower-than-one probability scenario? (The same can be said of sex, and so forth; though I am not black, I am a man, and I understand perfectly the suspicion under which I sometimes fall, as a man, and have no real problem with the person who subjects me to the suspicion, though it does sometimes inconvenience me. It is not their fault, not the fault of the sexual profiler; the "unfairness" is the "fault" nature and of other men.)

Race is merely an identifiable marker which a person has no, or little (e.g. Michael Jackson) control over, and as with any such marker, it is entirely justified to use that marker when calculating probability of guilt. Here's another example, hypothetical: suppose you visit two countries. In one country, half the young men are criminals and the other half are upstanding citizens. In the other country, all the young men are upstanding citizens. In the first country, it is entirely reasonable and justified for you to feel nervous in the presence of a random young male stranger. In the second country, you have no real reason to feel nervous. Now imagine two young men, both upstanding citizens, who are different from each other only in that they happen to live in those two countries, a fact about them which is not under their control (if you want to be a stickler, imagine that the border is firmly closed between the countries). One of the young men will come under suspicion from strangers frequently, and the other young man will not. It is "not fair" that these two essentially identical young men, different only in a factor outside their control (the country they reside), are treated differently by strangers. But this "unfairness" is not the fault of those strangers. It is the "fault" of nature (the luck of the draw, who was born where), and other young men.

If you want to insist on not profiling, and if you do not want to limit this principle to race but want to generalize this principle to all forms of coming to probabilistic conclusions on the basis of outward appearances, then you will render yourself almost completely blind. Indeed, you may as well poke your eyes out, so useless will they become.

I cannot accept your axiom

I cannot accept your axiom because it is insane.

Haha. Nice way of putting it. :)

I revised my previous comment right before you posted your own, so I won't respond in detail to this one.

I agree with you that the unfairness of statistical discrimination isn't (necessarily) the fault of those making the present judgements. My interest isn't so much in assigning blame for the existing unfairness, but in determing how we should act from this point on. One thing to consider is the self-fulfilling prophecy of this sort of unfairness. Consider a convicted felon recently released from prison. Although this felon was convicted for a non-violent drug offense, potential employers - and society at large - will treat the felon as if the offense commited was the average felony, the felon an average felon. The stigma so attached reduces the available opportunities as well the usual disincentives for violent behavior. After all, if society is going to treat me like a violent criminal, then I have little to lose if I act like one. And so the cycle continues. The issue here is not so much where to place the blame for the unfairness (obviously, in this example, the blame should be on drug prohibition) but on how to solve the problem of the self-fulfilling prophecy.

I toned it down

I kept toning it down and adding further explanations, so if you don't mind reading the latest version that's my "final word" (for now). No more "insane". Now "unworkable" (in my mind, the same idea, but less over the top expression).

The issue here is not so much where to place the blame for the
unfairness (obviously, in this example, the blame should be on drug
prohibition) but on how to solve the problem of the self-fulfilling

That is certainly a real problem. It might be something like the "externalities" problem of the market. That is, an unfortunate outcome of people acting in perfectly understandable ways, for which there is no good solution. The obvious solution is to improve the knowledge we can gain about other people. Privacy issues there, but if you can call up someone's life history or a summary of it, credit history, employment history, letters of recommendation, details on exactly what they were convicted of if convicted (thus allowing you to distinguish peacable drug users from muggers) and so on, then some of this problem would be removed.

There are of course remaining problems. Even if people know a person was in for drug use, they may treat him as if he were a mugger because they don't know any better. And also, even if you remove all errors and stupidities, there still remains a cycle: a person who is genuinely a mugger might want to reform, but have his opportunities reduced by the stigma. This is, of course, ultimately because of his own actions, but still, there is a cycle which arguably makes it difficult for criminals to reform, and so criminals may be pushed back into criminality.

There was no need to tone it

There was no need to tone it down; I wasn't offended. There was poetry in the original "insane axiom" formulation.

It Doesn't Take a Genius

In my experience IQ is a much over rated characteristic for predicting job performance.Sure, all other things being equal, a person with a high IQ will be better at most jobs. Persons with a high IQ can learn faster and may be better at making borderline decisions. More highly intelligent persons tend to be better educated and it is difficult to access the contribution of IQ vs. a deeper knowledge base. Probably both are important. High IQ is more consistent with the ability to hold more professionally demanding jobs which are lucrative.
On the other hand, most jobs don’t require a high IQ. All jobs require that the job holder be acculturated to the demands of work, such as punctuality, stamina, honesty, ability to work with others and conformity to supervision. A really brilliant person would go nuts doing most jobs. Many geniuses are eccentric and neurotic. I can’t document it here but I think studies show that childhood geniuses achieve no more than others if not also gifted with creativity or strong drive.
There seems to be an underlying prejudice that high IQ makes persons superior workers. I think part of this has to do with the fact that poorly motivated, poorly educated and poorly acculturated persons do poorly on IQ tests and poorly as workers. A motivated person can substitute experience and effort for IQ. For instance, if you take a math course, you can do math problems better than smart person who has not taken math. In many other spheres, experience and/or motivation makes up for and overcomes a high IQ.
This seems to be borne out in practice in the area of race in which Black people from foreign countries win scholarships at Harvard and African doctors leave their homelands to practice in the USA or Britain. In the presence of motivation racial averages make no difference to individuals.

Statistical discrimination

Statistical discrimination is how we make judgments when we lack the information necessary to judge others as individuals. The antidote is not obscurantism, but enlightenment--the easier it is to judge others as individuals, the less statistical discrimination there will be.

I completely I agree. However, is the "IQ taboo" a sufficient explanation for existing statistical discrimination? It depends on how restrictive labor law is with regard to technical interviews.

You claimed that "IQ tests are presumed to be racist, and employers are barred from using them in most cases." I don't know enough about labor law to verify this claim. Let's assume it is true. What about SAT scores? Don't SAT scores track fairly well with IQ scores (as do GRE, LSAT, etc.) Are employers also barred from using SAT scores to screen job applicants? That would be quite strange, since SAT scores are perfectly legit for undergrad screening purposes. Or is it that SAT scores are much less accurate than IQ scores for tracking job performance?

What prevents employers from creating their own technical interviews (or hiring a third-party who specializes in test creation) specifically tailored to track future job performance?