New Work on Implicit Bias

Erica Roedder and Dan Kelly have just posted a very interesting and valuable paper on implicit bias called, "Racial Cognition and the Ethics of Implicit Bias."  It discusses recent work in psychology on implicit racial bias and considers whether it is morally problematic, in an of itself, to have implicit biases.  The paper considers work by Garcia, Blum and Corlett to determine whether and how their accounts of racism take a stand on implicit bias and find that there are important unaswered questions.

If you haven't taken the Implicit Association Test (IAP), you should take it before you read this paper.  Don't try to learn about how the test works first, just take it and see what happens.  I myself have qualms about whether the test is effective for those who know they have implicit biases and try to guard against their impact, but we can talk about that in a later series of posts.

Comments

[jj]

I thought that a large factor in the scoring of the IAP test is the time one takes. Wouldn't that make it sensitive to "guarding against," which would involve time-consuming monitoring?

[Sally Haslanger]

JJ, I'm so glad you raised that issue! I didn't know that speed was an issue when I took the test, so I took time to be sure that I "guarded against" possible racial biases (not biases against whites...that didn't matter much to me). I scored quite poorly on the test. I found this upsetting. Although I am sure I am not exempt from implicit racial bias, I find it hard to believe, given my work and my life, that I'm an extreme offender. My hypothesis has been that at least some of the problem came from my attempts to be careful. If this is not addressed in the test design, I think it is a poor tool to determine implicit bias. Do any of those familiar with the test design know if time consuming "guarding against" is taken into account?

[bryce huebner]

Hey Sally and JJ...

I don't think that there is a way to take such techniques as "guarding against bias" into account in the test. Briefly, for the online tests, the algorithm works by comparing latency of responses on the crucial tasks (so, comparing the latency of responses for your black face/good responses to the latency of responses for your black face/bad responses). In the published work on the task, latencies over 1500-1800ms (depending on the analysis) are considered too long to be evaluable, so are typically removed from the analysis. However, even within the 1500ms range, there is still a time to do a bit of conscious reflecting (or at least I would assume that is the case). I think that at least one of the significant reasons why we should see the IAT tasks as a fairly course measure of bias.

Hope this is helpful...

[jj]

Yes, it is a crude measure, but they do highlight time as the major variable. It might well have accounted for Sally's score.

For me the eye-opener was, sadly, the women in science test. I could happily associate "physics" with "grandfather", but when it came to "grandmother" my biases were palpable. Really, I felt as though my brain might break or something. And I've spent a lot of time and effort promoting women in science.

It made it clear to me that even with a group one favors, one might have areas where prevailing stereotypes prevail.

To continue the anecdotal: I scored unbiased on the racial one, and of course I was very happy, since I've worked on that for decades. Still, given our society and the prevalence of racism, I did worry that it might show that I don't much like anyone. Or that I was very good at taking tests. That was another reason why the women in science was revealing. It felt entirely different.

[Dan Kelly]

Offhand I don't know if or how the "guarding against" technique would systematically influence results, but I agree that it's an interesting question. (I can barely help BUT doing it when I take the tests!)

Here's one thought, though: maybe guarding against would have a blanket effect on how quickly one was able to sort and respond. So to the extent that employing the technique influences response time, it could have the same effect across the board, i.e. in responding to associated stimuli and in responding to unassociated stimuli. Since test results turn on *relative* response speed, adding what amounts to a constant amount of time to responses to both associated and unassociated stimuli could just wash out in the end. Of course, that's pretty speculative. It's also possible that guarding against could have other interesting effects, like systematically increasing or decreasing the number of errors one made (fewer errors is taken to indicate a strong association between concepts).

I agree the IAT itself is an imperfect measure, in and of itself (aren't they all!). It is worth pointing out, though, that it's not the only indirect method that's turned up evidence of implicit racial bias (startle eye-blink tests, simple evaluative priming, EMG measures, etc.) As always, we would ideally want as many indices as we can get to help shed light on these murkier areas of racial cognition.

What we (speaking for Erica) see as so interesting (and frustrating and disturbing) about what the results of these kinds of tests suggest is that what we intuitively call "bias" isn't a single thing. People might, on the one hand, harbor implicit biases, while at the same time explicitly and honestly endorsing views that are diametrically opposed to those implicit biases, on the other. Given this, and that a) the implicit biases appear to link up to behavior (and perhaps judgment) in different and perhaps more subtle and nefarious ways than our more reflectively held beliefs do, and that b) the presence of implicit biases are not easily accessible to unaided introspection, one begins to wonder more and more about these things. Not only about what they are, exactly, but what we should do and say about them from a moral point of view. That it's not immediately clear, and that the questions merit some thought, was what we are arguing in our paper, anyway.

Great blog, by the way!

[Bryce Huebner]

Dan,

A quick couple of quick comments

On your suggestion that "guarding against would have a blanket effect on how quickly one was able to sort and respond", I don't think that I can see how this would be the case. Here's why this seems strange to me. The most plausible account of effects like the ones that are tracked by the IAT are likely to be spelled out in terms of a dual-process theory of cognition. Now, take the unconscious, automatic process that yields immediate intuitive responses. If you implicitly associate white faces with GOOD (for example), this automatic process will facilitate faster sorting of white faces with GOOD than sorting of black faces with GOOD, etc. This generates the initial sort of implicit bias. However, suppose you engage a conscious and deliberative processes dedicated to monitoring outputs to check for bias. In order to get the leveling effect that you suggest, it would have to take more time to evaluate white face-GOOD automatic responses than black face-GOOD automatic responses; but, this seems strange.

Second, at least as far as I've seen the published results on IAT style tasks, people who make more than a couple of errors are tossed from the data pool. So, when you say "fewer errors is taken to indicate a strong association between concepts" I'm a bit worried about whether there would be any noticeable difference in the person who was 'guarding against bias' and the person who was going slower for some irrelevant reason. More to the point, I guess that the fact that many people who take the web-based version of these tasks do not make errors at all suggests that this differences is unlikely to make a difference--though it will depend heavily on the interpretation algorithm employed for analyzing the data.

Finally, this brings me back to my initial claim above that IAT is a coarse-measure of bias. I think that my worry here is best seen as noting that there are many ways in which one can fail IAT--the mathematical models used to generate a judgment about how biased you are cannot distinguish between reflective inhibitory strategies (such as the one described by Sally) and genuine implicit bias. Moreover, the statistical analyses that demonstrate a high degree of correlation between the results of IAT style tests and overt bias collapse across individual difference in a way that also fails to allow for a distinction between these strategies. That said, I think that you are exactly on the mark when you note that the theoretical robustness generated by IAT, blink-tasks, priming experiments, etc., provide incredibly good evidence that there is a significant effect here that needs to be treated more thoroughly in the philosophical literature on bias.

By the way, I really enjoyed the paper, and it's great to see people struggling with the ethical implications of implicit bias!

[Dan Kelly]

Hi Bryce,

I think we're actually in agreement in our suspicions about the guarding against strategy on the IAT - that it wouldn't work very well because the reflective evaluation would equally slow down responses to both types of stimuli. So any differences in speed of the underlying, automatic mechanisms would still emerge. Sound right?

That said, I worry that too much focus on the IAT alone can lead to missing the forest for the trees. I'm finding myself more and more impressed by not just the convergence of different kinds of evidence gather in controlled experimental setting with indirect tests, but also by the type of systematic evidence of racial biases in real world situations like the resume study and the NBA study. To me, these seem complementary, like pieces of a larger puzzle.

Glad you enjoyed the paper!

[jj]

Is more than this being said: there are many ways to fail to follow the instructions on the test, and to the extent that one does, the results will be less valid. People who don't aim to do it as quickly as possible (or whatever the relevant instructions are) will learn less about themselves, though they may not know it.

I'm really unclear that much more can or should be said about "guarding against." Clearly, some people will try it; most of us want to look good on tests and have them represent us as we think we are. It's a pretty neat point about the time parameter that 'guarding against' on this test is self-defeating.

On the other hand, it looks as thought 'guarding againt' might occur in a number of situation that the test can't tell apart. At the same time, it might be worth thinking about whether there is something common among the different reasons one might choose to 'guard against.'

[Josh Glasgow]

Really, really interesting paper. One thought I had concerns the potential recommendation of adjusting Black students' grades to compensate for one's own suspected or probable implicit bias. This is small, maybe too small to mention, but another strategy is to grade blindly (at least with typed work whose content does not reveal the racial identity of the author, and with written work, if the grader is ignorant of whose identities go with which handwriting). I know a number of professors who do this, not always with an eye towards undetected bias that is specifically racial, but just with respect to bias simpliciter. (Think of all of the factors, other than merit, that might lead you to favor one student over another.) This also has the virtue that one need not identify a specific percentage to compensate for specifically racial bias.

Another thought I had concerns the "skeptical" claim that "implicit biases have no influence on actual behavior" and that "tests like the IAT are simply measuring associations between otherwise inert mental representations." The response in the paper is to point out ways in which IAT performance can be predictive of racist judgments and behaviors (such as what prescriptions MD's issue). If I'm following the dialectic properly (and I haven't -- yet, hopefully! -- read those studies establishing the predictions, so please forgive me if I'm missing something here), it seems that there might be a gap in this response. The IAT, as I understand it, just captures our patterns of association. The fact that such patterns predict racist judgments and behaviors, though, does not show that the associative patterns have causal influence. Compare: if it turned out that having the physical features indicative of a White racial identity predicted likelihood of racist behaviors, that doesn't mean that having those physical features has a causal role to play in the production of those behaviors. So just going off of the prediction itself, I'm not sure how predictability is supposed to establish a causal link.

There's also the argument that studies like the NBA study can be usefully explained by appealing to implicit bias. But, if 'implicit bias' just means implicit association, I'm not sure that there's much explanatory work being done by mere association. The fact that White NBA refs might more readily associate White than Black physical features with, say, the word 'happiness' as a split-second reaction doesn't leave me feeling as though I really understand why the ref is more likely to call a Black person than a White person for a foul. I wonder if, instead, the end of the day explanatory story to tell here is just one that has more run-of-the-mill details, like White refs being raised to be more charitable to Whites or whatever. This is just an off-the-cuff reaction, and one that I'm not all that confident about, but might more complex developmental stories do the heavy lifting in explaining both the refs' behavior *and* their implicit associations?

[Erica Roedder]

Thanks so much much for the comment about blind grading. I think blind grading is an excellent, but partial, solution. Indeed, I think it's shocking for anyone to *not* grade blindly, given what we know about our psychology today.

However, as you allude to, the solution is only partial. That's because it's often possible, at least in my experience, to tell who has written a particular essay. For instance, I often discuss paper topics, beforehand, with my students. I may even have read early drafts. All of this means that blind grading is not always possible. To state it sharply: Good *teaching* (reading drafts, discussing paper topics) makes it tough to engage in good *grading* (unbiased grading).

I'm curious whether anyone out there actually *does* attempt to explicitly compensate for race in grading, e.g. by increasing a Black student's grade by a few points? For myself, it seems that the epistemic case to do so is quite good--but something still strikes me as "off" or "epistemically strange" about doing so. Nevertheless, I can't see a good argument to justify that brute sense that something's "off."

As a final note, of course, bias shows up in other areas where it's not possible to be "blind," e.g. hiring decisions, criminal trials, police confrontations and line-ups, even everyday evaluations of your neighbors or co-workers. Thus even if blind grading was always possible, we'd still need to answer questions about our epistemic obligation to compensate for racial bias in other contexts.