The Implicit Association Test

Continuation of: Bogus Pipeline, Bona Fide Pipeline
Related to: The Cluster Structure of Thingspace

If you’ve never taken the Implicit Association Test before, try it now.

Any will do. The one on race is the “classic”, but the one on gender and careers is a bit easier to watch “in action”, since the effect is so clear.

The overwhelming feeling I get when taking an Implicit Association Test is that of feeling my cognitive algorithms at work. All this time talking about thingspace and bias and categorization, and all of a sudden I have this feeling to attach the words to...

...which could be completely self-delusional. What is the evidence? Does the Implicit Association Test work?

Let the defense speak first1. The Implicit Association Test correctly picks up control associations. An IAT about attitudes towards insects and flowers found generally positive attitudes to the flowers and generally negative attitudes to the insects (p = .001), just as anyone with their head screwed on properly would expect. People’s self-reports were also positively correlated with their IAT results (ie, someone who reported loving flowers and hating insects more than average also had a stronger than average IAT) although these correlations did not meet the 95% significance criterion. The study was repeated with a different subject (musical instruments vs. weapons) and similar results were obtained.

In the next study, the experimenters recruited Japanese-Americans and Korean-Americans. Japan has been threatening, invading, or oppressing Korea for large chunks of the past five hundred years, and there’s no love lost between the two countries. This time, the Japanese-Americans were able to quickly match Japanese names to “good” stimuli and Korean names to “bad” stimuli, but took much longer to perform the opposite matching. The Korean-Americans had precisely the opposite problem, p < .0001. People’s self-reports were also positively correlated with their IAT results (ie, a Korean who expressed especially negative feelings towards the Japanese on average also had a stronger than average IAT result) to a significant level.

There’s been some evidence that the IAT is pretty robust. Most trivial matters like position of items don’t much much of a difference. People who were asked to convincingly fake an IAT effect couldn’t do it. If the same person takes the test twice, there’s a correlation ofabout .6 between the two attempts2. There’s a correlation of .55 between the Bona Fide Pipeline and the IAT (the IAT wins all competitions between the two; it produces twice as big an effect size). There’s about a .24 correlation between explicit attitude and IAT score, which is significant at the 90% but not the 95% level; removing certain tests where people seem especially likely to lie on their explicit attitude takes it up to 95. When the two conflict, the IAT occasionally wins. In one study, subjects were asked to evaluate male and female applicants for a job. Their observed bias against women correlated more strongly with their scores on a gender bias IAT than with their own self-report (in other experiments in the same study, explicit self-report was a better predictor. The experimenters concluded both methods were valuable in different areas)

Now comes the prosecution. A common critique of the test is that the same individual often gets two completely different scores taking the same test twice. As far as re-test reliability goes, .6 correlation is pretty good from a theoretical point of view, but more than enough to be frequently embarrassing. It must be admitted: this test, while giving consistent results for populations, is of less use for individuals wondering how much bias they personally have.

Carl Shulman would be heartbroken if I didn’t mention Philip Tetlock, so here goes. This is from Would Jesse Jackson Fail the Implicit Association Test?, by Tetlock and Arkes (2004):

Measures of implicit prejudice are based on associations between race-related stimuli and valenced words. Reaction time (RT) data have been characterized as showing implicit prejudice when White names or faces are associated with positive concepts and African-American names or faces with negative concepts, compared to the reverse pairings. We offer three objections to the inferential leap from the comparative RT of different associations to the attribution of implicit prejudice: (a) The data may reflect shared cultural stereotypes rather than personal animus, (b) the affective negativity attributed to participants may be due to cognitions and emotions that are not necessarily prejudiced, and (c) the patterns of judgment deemed to be indicative of prejudice pass tests deemed to be diagnostic of rational behavior.

In other words, there are a bunch of legitimate reasons people might get negative IAT scores. Any connection whatsoever between black people and negative affect will do. It could be the connection that black people generally have low status in our society. It could be that a person knows of all the prejudices against black people without believing them. It could be that a person has perfectly rational negative feelings about black people because of their higher poverty rate, higher crime rate, and so on. Or it could be somethng as simple as that, for whites, black people are the out-group.

...this actually isn’t much of a prosecution at all. I consider myself a moderate believer in the IAT, and I think it all sounds pretty reasonable.

What most IAT detractors I’ve read want to make exquisitely clear is that you can’t hand someone an IAT, find an anti-black bias, and say “Aha! He’s a racist! Shame on him!”3

I think this is pretty obvious4. You can hold beliefs on more than one level. A person may believe there is a dragon in his garage, yet not expect an experiment to detect it. A skeptic may disbelieve in ghosts, but be afraid of haunted houses. A stroke victim may deny an arm is hers while admitting it is attached to her body. And it’s supposed to be news that you can give black people some sort of vague negative connotation on a nonconscious level without being Ku Klux Klan material?

There is a certain segment of society which interprets the sun rising in the morning as evidence of racism. It is not surprising that this segment of society also interprets the IAT as evidence for racism. I myself think racism is a bad word. Not in the way “shit” is a bad word, but in the way “wiggin” is a bad word. It divides experience in a perverse way, drawing a boundary such that Adolf Hitler ends up in the same category as the guy who feels a pang of guilty fear late at night when he sees a big muscular black guy walking towards him5. Taboo the word “racism”, “prejudice”, and any other anti-applause-light6, and a lot of the IAT debate loses its meaning.

Which is good, because I think the IAT is about much more than who is or isn’t racist. The IAT is a tool for measuring distances in thingspace.

Thingspace, remember, is the sort of space in which we draw categories7. “Chair” is a useful category because it describes a cluster of things that are close together in concept-space in a certain way: stools, rocking chairs, office chairs, desk chairs, et cetera. “Furniture” is another useful word because it describes another cluster, one that includes the chair cluster and other concepts nearby. Quok, where a “quok” is defined as either a chair or Vladimir Lenin, is a useless category, because Lenin isn’t anywhere near all the other members.

Speaking of communists, remember back when East and West Germany got reunited? And remember a little further back, when North and South Vietnam got reunited too? Those reunifications, no matter how you feel about them politically, were natural links between culturally and historically similar regions. But imagine trying to unite East Germany with South Vietnam, and West Germany with North Vietnam. The resulting countries would be ungovernable and collapse in a matter of weeks.

If you associate white people with good things, and black people with bad things, then forming the categories “white and good” and “black and bad” is like reuniting East and West Germany. You’re drawing a natural border around a compact area of the map. But being forced into the categories “white and bad” and “black and good” is about as natural as trying to merge East Germany and South Vietnam into the new country “Southeast Vietnermany”. You’re drawing an arbitrary boundary around two completely unrelated parts of the map and then begging in vain for the disgruntled inhabitants to cooperate with each other.

If you provoke a war between the reunified Germany and Southeast Vietnermany, and watch which side coordinates its forces better, you get the Implicit Association Test.

Why would we want to measure distance in thingspace? Loads of reasons. Take a set of pictures of famous cult leaders, mix them with a set of pictures of famous scientists, and test Less Wrong readers’ reaction times associating a picture of Eliezer Yudkowksy’s face with either set8. If it’s easier to place him with the scientists, or there’s no difference, that’s some evidence we haven’t become a cult yet. If it’s easier to place him with the cult leaders, we should start worrying.

Tomorrow: some more serious applications to rationality.


1: Most of these results taken from this, this, and this study.

2: There’s some evidence that priming can change your IAT score. For example, subjects shown a picture of a happy black family enjoying a picnic just before an IAT got lower bias scores than a control group who didn’t see the picture. And before condemning the test too much for its tendency to give different scores on different occasions, remember back to your school days when you’d have to take endless quizzes on the same subject. Occasionally just by chance you’d get a spread of ten point or so, and if you were on the borderline between passing and failing, you might very well pass one test and fail another test on the exact same material. This doesn’t mean grade school tests don’t really measure your knowledge, just that there’s always a bit of noise. The IAT noise is greater, but not overwhelmingly so.

3: There’s also a fear someone might use it for, say, evaluating applicants for a job. Due to its weakness as an individual measurement and the uncertainty about how well it predicts behavior, this would be a terrible idea.

4: Full disclosure: Despite strongly opposing prejudice on a conscious level and generally getting along well with minorities in my personal life, I get assessed as moderately biased on the racism IAT. I had some memorable bad experiences with certain black people in my formative years, so this doesn’t much surprise me.

5: In fact, Jesse Jackson (note for non-Americans: a well-known black minister and politician who speaks out against racism) himself admits to occasionally having these pangs of guilty fear—hence the name of Tetlock’s article.

6: I think Eliezer once coined a term for the opposite of “applause light”, for things like “racism” and “scientism” invoked only so people can feel good about hating them, but I can’t seem to find it. Can someone refresh my memory?

7: I was split on whether to use the term thing-space or concept-space here. Eliezer uses concept-space in a very particular way, but “good” and “black” seem much more concepts than things. I eventually went with thing-space, but I’m not happy about it.

8: This is a facetious example. It’s possible in theory, but there would be so much to control for that any result would be practically meaningless.