The Implicit Association Test

People say they love the new packaging. Their reaction times can disagree. The Implicit Association Test tries to measure that gap: instead of asking what you think, it times how fast you sort words or pictures when two ideas share a key. More than forty million IATs have been taken on one research website alone, and it is one of psychology's most-used and most-disputed measures.

What is actually happening

The test has seven short blocks. First you sort only cats and dogs. Then only good and bad words. Then all four at once, with dogs sharing a button with good and cats with bad. Then the sides swap, and you sort all four again with dogs now paired with bad.

If dogs and good already go together for you, the first combined task feels easy and the second clumsy. The speed difference between the two is the measure. The very first IAT, in 1998, showed it with flowers and insects: people sort "flower + pleasant" faster than "insect + pleasant".

Because some people are simply faster or more erratic than others, raw speed differences are rescaled into a standard score called D, which accounts for each person's own variability and penalises errors. It has been the norm since 2003.

Why one score means little, an average means more, and what neither fixes

One score is noisy. Take the same IAT twice and the two scores agree only moderately. That's enough to compare groups, but the field's own best-practices paper says it is "not adequate to justify treating a single IAT observation as accurately diagnostic" of a person.

Averaging removes noise, not doubts about what is being measured. Think of a bathroom scale that jumps around by a couple of kilos and may also be slightly miscalibrated. Weighing hundreds of people cancels the jumping around; it does nothing about the calibration. The IAT's sharpest critics are arguing about calibration: whether the score measures anything distinct from what people would simply tell you. That's why even a precise group average needs careful interpretation.

It predicts behaviour modestly, and the size is disputed. Its developers' review found it predicts behaviour, and adds a little beyond what people say about themselves, especially on sensitive topics such as race. A critical re-analysis of the race studies found it predicted poorly and no better than simple questionnaires.

Changing a score doesn't change behaviour. Brief exercises can shift IAT scores a little, but across hundreds of studies the shift didn't carry through to behaviour.

In marketing and consumer research: a second opinion, not a replacement

Here is the result the sales pitches leave out. In the largest review, across about 40 consumer studies and more than 3,000 people, simply asking people predicted their brand choices better than the IAT did (r = .55 against .32). For everyday brands, people mostly know and will tell you what they prefer.

So why bother? Two reasons, each with limits:

• It can add a little on top of asking. In the same review, the IAT and self-report each predicted behaviour beyond what the other did. A recent independent study of toothpaste brands found IAT-measured brand image predicted actual market share, more strongly when shoppers chose from only a few brands.
• It is a check on polite answers. Where people may tell you what they think you want to hear, a disagreement between the IAT and the survey is worth investigating.

Know the limits before you pitch it:

• Established brands only. An unfamiliar brand or a new pack design looks "negative" just because it is slower to sort, which rules out much new-product testing.
• Two at a time. The standard IAT compares a pair. Five pack options means several IATs, or a different method.
• It takes time. A standard IAT is 190 sorting trials plus instructions. On a phone panel, run one per session.
• Plan the sample. As a rough guide (80% power, 5% significance) for a small effect (d = 0.3): about 90 people to detect which way a sample leans, and about 175 per group to compare two segments.
• Report averages, not people. Everything above about individuals applies to shoppers too.

If you are going to run it, get these right

These come from the best-practices paper written by 22 IAT researchers.

• Use the standard seven blocks and lengths: 20, 20, 20, 40, 30, 20 and 40 trials. Block 5 was lengthened from 20 to 30 to reduce the effect of which pairing comes first.
• Counterbalance which pairing comes first and which side each category starts on.
• Use stimuli people sort instantly: at least three per category, each clearly in its category. Pilot them and drop anything people hesitate over.
• Use familiar categories only.
• Score with D and apply the standard exclusions: drop trials over 10 seconds, and drop participants for whom more than 10% of trials were faster than 300 ms.
• Need a score per person? Run the IAT twice and average.
• Report everything: block order, counterbalancing, stimuli, exclusions and internal consistency.

Build a standard seven-block Implicit Association Test.

Targets: CATEGORY A vs CATEGORY B (replace with your own; both must be
familiar to participants; at least 4 easy-to-sort items each, as words or images).
Attributes: Good (wonderful, pleasant, excellent, joy, love) vs
Bad (terrible, awful, horrible, nasty, hate).

Blocks and trial counts: 1 targets only (20), 2 attributes only (20),
3 combined (20), 4 combined (40), 5 targets reversed (30),
6 combined reversed (20), 7 combined reversed (40).
In combined blocks, strictly alternate target and attribute trials, and never
allow more than four same-key-correct trials in a row.
Use E/I keys on desktop and two large on-screen buttons on touch devices.

Counterbalance across participants which pairing comes first, and which side
each target starts on. Show a red X on errors; the participant must give the
correct response to continue.

Before the IAT, ask which of the two they expect to lean towards
(A / B / No difference). After it, ask a 1–7 liking question for each,
so the measured and stated preferences can be compared.

Score with the D algorithm: drop trials over 10,000 ms, and flag participants
with more than 10% of trials under 300 ms. Save every trial with its block,
pairing order and stimulus. Report results as group averages, not individual scores.