Measuring unconcious bias

[madd0ct0r]

I took part in the study here: https://implicit.harvard.edu/implicit/index.jsp and it came out with a strong bias to white people over black people.
It is quite an American orientated one (the political spectrum is called liberal-conservative, for a start), but it was interesting to do.

I'd like to reject their results, because I don't like the idea. This feeling dosen't surprise me, nor does my implicit bias. Time to stick to careful metrics instead of gut feelings to keep that pinned down I guess.

[Simon_Jester]

Since the website is down for maintenance at the moment, Maddoc, could you actually tell us what their methodology is?

[Kingmaker]

It's an implicit association test. Basically, they show pictures of white or black faces (or whatever it is they are testing), along with positive or negative concepts (e.g. joy, agony) and measure how quickly you are able to categorize the word. The notion being that quickly associating black faces with negative concepts and white faces with positive concepts (or vice versa) is indicative of implicit bias.

Also, the website is back up.

[madd0ct0r]

yah. you have to register in order to take the test. It is carried out exactly as kingmaker said. There is a detailed faq that follows the test but I didn't want to directly quote that in case knowing too much going in skews the results.

[Simon_Jester]

I'm not particularly interested in taking the test because I don't trust that the underlying mechanism of the test actually works.

Also because even if it does measure "subconscious bias," it's a bias I can't do anything about, and offers no insight as to how I came to be biased in the first place.

[Alyrium Denryle]

I'm not particularly interested in taking the test because I don't trust that the underlying mechanism of the test actually works.

Also because even if it does measure "subconscious bias," it's a bias I can't do anything about, and offers no insight as to how I came to be biased in the first place.

They work. It is a standard test in social psychology, and they are properly controlled.

What they are really useful for is things like police, and subconscious bias of this type accounts for a lot of the disparity in police shootings by suspect race. Police are more likely to over-estimate the threat posed by a black person, and are faster to reach the decision to shoot.

[Simon_Jester]

I can easily see how you'd test for that directly (show video of black people doing random stuff, and see if the officer goes for a gun in training or whatever).

I'm skeptical of the word-association thing unless it can be proven to correlate with more direct measures (like a cop in VR training pulling a gun more readily).

[madd0ct0r]

It's a well established technique, but skeptisim seems reasonable.

Googling one way found an negative result from 2001: http://psycnet.apa.org/journals/psp/81/5/774/ and a positive non paywalled 2007 report with a short lit review: https://hal.archives-ouvertes.fr/hal-00571656/document

I expect this Hazard study is trying to gather data beyond the usual pool of students in testing.

[Simon_Jester]

To be clear, maddoctor, are you saying that there are a number of studies that show connections between the word-association metric and other metrics that are more obviously connected to racism?

[madd0ct0r]

I'm saying that 30 minutes of googling and reading social papers that are a bit beyond what I can reliably understand and spot mistakes in, suggested 1) it's a common technique, 2) not everyone agrees bias shown translates to real world action but 3) most results indicate it predicts real world actions.

The second paper reported connections of the type you describe.

[Esquire]

To be clear, maddoctor, are you saying that there are a number of studies that show connections between the word-association metric and other metrics that are more obviously connected to racism?

To confirm what maddoctor said, yes; it's a reasonably well-established and validated metric in the social sciences. I can pull some articles if you like, I'd just need to wait until I get back to work (institutional journal access is a wonderful thing).

That said - as usual in the social sciences, there are conceptual and metaconceptual validity concerns. The HAL study (the one with the positive finding), particularly, sampled only undergraduate psychology students, controlled for no moderating variables, and failed to establish adequate adherence to statistical assumptions. And, of course, all it really says is that college students who are prejudiced against a group are at most 41% more likely to act in a prejudiced* fashion, which isn't exactly an earth-shaking claim. All of which the study authors acknowledge, which is nice to see.

*Very loosely-defined, at that. "Would reduce funding for a student group that by definition doesn't serve me" is not the same as "wouldn't hire a black/Asian/Jewish person," or whatever.

[Simon_Jester]

Yeah.

Another problem is that even if you identify this kind of bias... what do you do about it? If I find out I have high blood pressure, that's actionable. If I find out I have acrophobia, there are fairly well defined techniques I can use to overcome that at least partially.

But if I have a reflexive bias that makes it easier for me to associate the word "depressing" with "bad" than "happy" with "good" whenever I just spent five seconds looking at the picture of a black person... what do I even do about that? What does it mean, in terms of real consequences that we can clearly measure?

[Esquire]

Nothing whatsoever, of course. The ivory tower is especially tall and its doors especially well-guarded in sociology/anthropology/some psychology subfields; as far as I can work out they essentially spend a lot of time telling minorities what they already know and inventing jargon that paradoxically keeps the majority from being able to engage productively with social issues.

[Simon_Jester]

Well, that's what worries me.

If I consistently associate "car crash" with "bad" faster than I associate "vacation" with "good" or whatever after looking at black faces, clearly that is a metric of something. The problem is that it's not obviously measuring anything useful or actionable, and it's very hard to show correlations between that and other things on a large scale. And it's certainly not measuring anything I can realistically do something about.

[Purple]

So why does it worry you? I mean, you just gave a good explanation of why it shouldn't worry you. So why does it?

[madd0ct0r]

Well for me, its showing my assumptions about myself are not quite true. It means that when im in a posistion where that bias comes through, i need to be aware of it and takes steps to control for it.
Things like reviewing cvs, asming for the names to be hidden before i review them, that sort of thing.

[Simon_Jester]

Since I haven't taken the test, it isn't about worrying over the test result. The point is that a test which provides a measurement with ill-defined consequences and no way to do anything about the result of the measurement... is not a good measurement.

Suppose you had a little machine that whenever you pointed it at someone gave a numerical reading from one to ten. You know ten is better than one, but you don't know why exactly, except for ideological reasons. You don't know what the numerical scores correlate to, either- it's not any obvious directly measurable thing. There's nothing anyone can do to alter their score significantly in the long run.

Frankly, the machine isn't worth very much. If anything it's likely to cause more trouble than it's worth, because it "rates" people in a way that will cause confusion and discord.

That's my criticism here.

[Purple]

If that is the source of your worries than I am indeed in agreement. And thus withdraw my question/objection from before.

[Alyrium Denryle]

Since I haven't taken the test, it isn't about worrying over the test result. The point is that a test which provides a measurement with ill-defined consequences and no way to do anything about the result of the measurement... is not a good measurement.

Suppose you had a little machine that whenever you pointed it at someone gave a numerical reading from one to ten. You know ten is better than one, but you don't know why exactly, except for ideological reasons. You don't know what the numerical scores correlate to, either- it's not any obvious directly measurable thing. There's nothing anyone can do to alter their score significantly in the long run.

Frankly, the machine isn't worth very much. If anything it's likely to cause more trouble than it's worth, because it "rates" people in a way that will cause confusion and discord.

That's my criticism here.

Ok. The information gained has a lot of applications. What you do about it is situation dependent. On an individual level, there is not much you can do directly. The bias is unconscious, buried pretty deep in how your brain processes language and identifies threats. But it is probably socialized. Taking steps to make the bias irrelevant in decision making is something you can do. But ultimately, the social environment human brains are raised in needs to change.

[Simon_Jester]

I went back and read over the papers maddoctor linked to. And I noticed that the second one, the one that actually talks about there being a positive correlation between IAT (Implicit Association Test) results and overt acts of bias... Well, what it really says is that attitude IAT results are a better predictor of what the experiment declares to be discriminatory behavior, compared to self-reporting of prejudices.

That strikes me as a no-brainer, because I doubt that self-reporting of prejudices is even slightly effective as a measure of determining whether people are prejudiced.

It doesn't make attitude IAT results useful or reliable.

I kept digging and honestly I'm having trouble parsing the results mentally, but unless I'm badly misunderstanding the conclusions, the attitude IAT didn't seem to do a particularly good job of predicting racist behavior or actions in general, even if it maybe-sorta works when dealing specifically with whether college undergraduates want high or low funding for minority groups on campus, making that our definition of a 'racist action.'

Now, the stereotype IAT they reference seems to work better- but that's a different test.

[Arthur_Tuxedo]

I'm not sure why you insist there is nothing you can do once you've realized you have unconscious biases. There are plenty of things you can do to minimize the impact of biases, like grading homework and tests with the names of the students hidden to you. I've caught myself making split-second assumptions based on stereotypes and corrected myself, assumptions I probably wouldn't have noticed if I hadn't been thinking about unconscious bias.

[Purple]

Shouldn't you be doing that sort of stuff anyway? I mean, as a teacher odds are what ever bias toward race or what ever else is going to be massively outshone by the fact you know these students individually and thus have personal biases towards them as individuals. As in you are for example far more likely to forgive a mistake to a student you like and think "oh, he knows this." rather than to one who consistently under performs.

[Simon_Jester]

See, the kind of stereotyping that is tested for with the stereotype association tests, I can see how that's something you can counteract.

But the kind that involves this totally subconscious, non-explicit thing you can't even vocalize... exactly how does one fight that without randomly overcompensating?

[Purple]

Well to my understanding the standard approach of modern society is indeed to randomly overcompensate wildly and not only act as if it knows what its doing but outright chastise everyone who fails to fall into line.

[Simon_Jester]

Which compensations do you deem "wild" and "over," exactly?