Is “statistical” discrimination a useful concept?

During a meeting with one of my students to discuss the topic of her final paper, I was asked, “What’s the difference between statistical discrimination and simple discrimination?” I had to pause on that one. My understanding was that discrimination as it is generally used is “taste-based.” So an employer deciding not to hire you because she would prefer not to work with other women is taste-based discrimination. Even though you may have the most impressive CV, she is willing to pay or incur a penalty (the loss of your productivity) to not have you as an employee based on dislike of women.

How does this differ from statistical discrimination? Your employer has no personal feelings against women – she just “knows” that women are more likely to be the ones to take days off from work because of sick children and although your CV looks great, she believes that your level of productivity will be lower than a slightly less impressive male’s because you have (or may someday have – she isn’t allowed to ask you in the interview) children.

My student then asks, “Well how is it really different if the outcome is the same?” I think that’s a brilliant question! And it prompts another interesting question, again resulting from interactions with students. We discussed Bobo,’s paper on Laissez Faire Racism in class today. The authors contend that current racial attitudes of White Americans, rather than reflecting a decline in racism point to a shift from what they term Jim Crow racism to laissez faire racism. Jim Crow racism was characterized by overt bigotry, demands for strict segregation in virtually all domains, advocacy of government mandated discrimination, and the adherence of beliefs that Blacks were categorically inferior to Whites. Enter the Civil Rights Movement, the dismantling of Jim Crow, and the shift away from biological interpretations of the disadvantaged status of Blacks in our society. We shift from the age of Jim Crow to that of laissez faire racism in which we have the end of state-enforced inequality, a move towards race-neutral and anti-discriminatory state practices, but a reliance on informal racial bias is pervasive.

Is statistical discrimination then simply the evolution of racism from Jim Crow to laissez faire? From a taste-based discrimination that denoted a dislike or distaste for minorities based on negative stereotypes to statistical or information based discrimination that is stripped of animus or dislike but based on stereotypes as well?

Is there a case for keeping the “statistical” in statistical discrimination and if so, why? Does its inclusion better inform us about inequality within organizations – particularly race, ethnic, and gender inequality?

No, this is not the essay I plan to assign to my students on their midterm. I truly would like a well-reasoned explanation for why this term is helpful.

Written by lhinkson

October 5, 2009 at 8:29 pm

Posted in uncategorized

14 Responses

Subscribe to comments with RSS.

  1. John List has some good empirical papers on this.

    In one, he sends disabled/non-disabled people around to get quotes for car repair. The disabled are quoted higher prices. Discrimination, right? But then he has the disabled shoppers mention that they have shopped around previously (under the assumption that salesmen are assuming that disabled customers in general don’t shop around). That eliminates the gap in quoted prices.

    Obviously, every market will differ. But it is possible to test differences between taste-based and statistical discrimination.



    October 5, 2009 at 9:18 pm

  2. First of all I would object to the use of the term ‘statistical’. It seems to lend credence to discriminating claims by suggesting that they are based on real statistics, not mere assumptions.



    October 5, 2009 at 9:40 pm

  3. Sure.

    Statistical discrimination is based on average characteristics of a group (e.g. assuming that a female applicant is shorter and lighter than a male applicant b/c women are shorter and lighter on average than men).

    Taste based discrimination is based on your reaction to a group (e.g. not hiring female applicants b/c they make you uncomfortable).

    So statistical discrimination picks up on characteristics that might affect productivity, whereas taste-based discrimination picks up on things that affect the welfare of the person doing the hiring.



    October 5, 2009 at 10:05 pm

  4. In one case, we know the motivations of actors doing the discriminating. In the other case, we do not. The underlying motivations may or may not be the same but it is probably naïve to assume that there is not overlap between the two or that one does not reinforce the other.



    October 5, 2009 at 10:10 pm

  5. AS Thorfinn notes, the basic idea is that one type is merely an information issue, while the other is not. Thus, there is a test: the discrimination persists in the face of information, it’s taste based. If it changes, it might be statistical. That can be tested empirically.

    I’d also add that you can turn the issue into a non-falsifiable proposition: you can always claim that taste based discriminators are going on an unobserved variables that your regression didn’t catch. I’ve seen a few proponents of stat disc. weasel using this tactic.



    October 6, 2009 at 1:52 am

  6. In economic models, the basic difference is that while taste-based discrimination will be penalized through market competition, statistical discrimination can persist because of the assumptions economists make about the actor’s information. We can get stuck in a lower equilibrium trap. Phelps (1972) and Spence (1973) wrote the big papers on this.

    Of course, homogenized knowledge/beliefs (whether it be perfect asymetrical, partial or no information) is another one of those silly assumptions economists make in order to simplify their models. I’m working on a paper now that frames it as a Hayekian knowledge problem that leaves room for heterogeneous local knowledge. I’m using Kirzner to argue that information gaps will be arbitraged away, but you guys might be more comfortable with the Burt’s idea of structural holes as a way overcome statistical discrimination.

    Even if the results are the same, the process by which it happens is different, and that alone is worth studying for policy reasons.



    October 6, 2009 at 2:19 am

  7. Well, motives certainly matter in trying to end something. You have to know what you’re fighting against to be effective.

    This is also related to the message in relatively enlightened people will write you off if you call them *-ist, but they might listen if you just accuse them of discriminating statistically and then tell them how their statistics are wrong.



    October 6, 2009 at 5:20 am

  8. It depends on what kind of statistical discrimination you’re talking about: economic or sociological. The economic version is considered the “strong” version and assumes that actual group differences in productivity exist. The sociological version notes that racial attitudes and beliefs along with stereotypes about a group’s productivity is more important than actual productivity. Therefore, employers only have to perceive productivity differences to justify their exclusion of members of a particular group. It also depends on job allocation patterns which could lead to employers gathering “information” about productivity of certain groups.

    The 1999 piece by Tomaskovic-Devey and Skaggs in Work and Occupations is pretty good at laying things out and testing for statistical discrimination. Here’s the cite:

    Tomaskovic-Devey, Donald and Sheryl Skaggs. 1999. “An Establishment-Level Test of Statistical Discrimination Hypothesis.” Work and Occupations 26: 422-445.



    October 6, 2009 at 12:12 pm

  9. Hillbilly raises an important distinction between economic and sociological statistical discrimination. It also begs the question, “Are differences in levels of productivity due to ‘actual’ levels of a group’s productivity or are they artifacts of organizational rules and policies that affect group productivity levels?

    I also agree with Trey that honing in one the motivations of actors can lead to more effective policy proscriptions.

    But if we buy the laissez faire racism argument, can we really distinguish between the two types of discrimination? Even if there are group-based average differences in levels of productivity say, can’t you simply be using that as a justification for excluding certain workers regardless of their individual-level attributes?



    October 6, 2009 at 2:35 pm

  10. I don’t think economists make a distinction between actual differences in the average group productivity and the perception of such.



    October 6, 2009 at 4:34 pm

  11. Does anyone here still disagree there are important differences?

    Other people have already mentioned that improving outcomes depends on understanding how much of each discrimination is happening.

    On a more personal note, for me, and I think for most people, there is also a moral difference.


    Michael Bishop

    October 6, 2009 at 7:55 pm

  12. Correll and Benard’s recent paper (;jsessionid=6638962430D481362DA5414B87E8065D?contentType=Book&contentId=1760425) is worth a look for those trying to distinguish between SD models and sociological models of discrimination. The core of the argument is that “discrimination in statistical models derives from an informational bias, while discrimination in status models derives from a cognitive bias.” A colleague of mine reacted to this paper with the cavil that some economic takes on SD are in fact (socio-) cognitive but offered that “C&B’s best point is that the performance of employees within the firm, after hire is the obvious place where status-characteristics theories have better explanatory power than statistical-discrimination theories. That is to say, the persistence and growth of inequality within many firms is completely counter to all varieties of statistical discrimination theory.”



    October 7, 2009 at 12:36 am

  13. This concept is wrong on so many levels. It justifies subtile forms of racism and is used in surveys to avoid the word indirect discrimination. As long as this term is used, no serious debate about biases of employers can be held. No change is needed because the bias are based on statistical knowledge. The “statistical” in the name give racists some credibility, without ever having the need to show statistics on real work behaviour of “minorities”. This “political correct” term protects the strong against the weak and should be banished from any scientific discourse.



    November 10, 2009 at 1:05 pm

  14. car repairs should only be done by professionally trained car technicians not by self-taught crews,`*


    Wire Gauge :

    October 26, 2010 at 11:28 am

Comments are closed.

%d bloggers like this: