How do you measure associations?

Want to know what people associate with your brand, employer or product? First clarify the question. Are you looking for spontaneous words, a direct evaluation, a relatively easy association or a relationship with a specific choice? These questions call for different measurements. Good association research connects the question, the task and the interpretation of the result.

Here, measuring associations means using a documented task to investigate the connections people make between a subject and other concepts, images or evaluations. The result depends on the task, stimuli, participants and context.

You can ask people to respond directly or investigate a connection indirectly through task performance, for example with an IAT. Define the result you need in advance. Keep mentioning an association, evaluating it and establishing its significance for a decision separately identifiable.

In this article, a measurement is explicit when someone directly reports the requested association or evaluation. In an implicit measurement, the relevant relationship is inferred indirectly from task performance.

Greenwald and Lai distinguish ‘indirectly measured’ from the theoretical claim ‘unconscious’. They also discuss how internal consistency and test–retest reliability assess different aspects of measurement precision. Greenwald & Lai, 2020.

Use concrete language in a research brief: what does the participant do, what data are recorded and what conclusion do you want to draw? The label ‘implicit’ should not replace that explanation. Treat a difference between methods as a finding requiring interpretation, without assuming that one answer reveals a hidden truth.

The table below is a practical planning aid. These are starting points for a research proposal, not a fixed ranking of methods.

For every proposed task, request evidence supporting the specific application. ‘Response-time measurement’ is not a complete method description. This article discusses the IAT in greater detail; it does not provide an executable priming protocol.

In free association, a person receives a cue, such as a brand name, and gives related words. In Small World of Words, De Deyne and colleagues collected multiple responses per cue to map word associations. De Deyne et al., 2019.

For your own study, you might ask: ‘Which words come to mind when you think about this employer?’ Decide beforehand how many responses you request and whether participants can explain them. Record original responses before combining them under a theme.

Retain the exact question too. ‘What do you find attractive about this employer?’ already asks for a positive selection. It is a different task from an open association question. If both are relevant, report them as separate components.

Then ask what a word means to that participant. In your study, ‘large’ might be explained as offering many opportunities or as being impersonal. Keep interpretations traceable to the explanation given. A general word database can support preparation; claims about your audience need appropriate data.

After an exploratory phase, you can ask people to evaluate specific attributes. Define:

1. The subject: is the participant evaluating a brand, product, advertisement or particular experience?
2. The attribute: does the question concern reliability, attractiveness or recognisability, for example?
3. The scale: what do the response categories mean, and how is ‘I don't know’ handled?
4. The context: what information has the participant already seen, and in what order?
5. The reporting: will you show percentages, distributions or a justified composite score?

This is an editorial planning aid. A scale from ‘--’ to ‘++’ can serve as an agreed coding scheme when its categories are clear. Without additional justification, treat it as ordinal: the distance between categories is not automatically equal. Do not silently convert the result into a percentage of association strength.

Greenwald, McGhee and Schwartz introduced the Implicit Association Test in 1998. The task compares classification performance when two target categories share response keys with two attribute categories in different pairings. Its result concerns a relative association within that comparison. Greenwald et al., 1998.

A research proposal should therefore name the comparison in full. Instead of writing only ‘we measure how trustworthy brand A feels’, also specify the other target category and both attribute categories. Without this information, a score is difficult to assess.

Do not automatically call a custom rapid yes/no question an IAT. For each task variant, ask which scientific method is followed and what has been adapted. Brand research also requires evidence for its own application; a classic publication does not automatically validate your words, brands or audience.

In 2003, Greenwald, Nosek and Banaji described improved IAT scoring. The D-score standardises differences between task blocks using response-time variability; the algorithm also specifies rules for errors and unusual responses. Greenwald et al., 2003.

Ask your brief to specify the algorithm variant, score direction, exclusion rules and analysis version. Request the numbers of observations and participants retained after quality checks. This article supplies neither a custom scoring algorithm nor universal thresholds.

A positive or negative value can only be interpreted alongside its coding. Include a sentence such as: ‘Within this task, a higher score indicates relatively greater pairing of category X with attribute Y compared with the specified alternative combination.’ Fill in X and Y according to the actual protocol. Do not label the score as purchase probability or market share.

Greenwald and colleagues recommend, among other things, familiar categories, easily classified exemplars and pilot testing with the intended participant population. They also call for transparent reporting of stimuli and procedures. Greenwald et al., 2022.

Create a dossier containing the final texts, images, instructions and task sequence. Record how you address potential order effects, for example through a justified distribution of starting orders. Also test the technical implementation on the devices that will actually be used.

In Dutch, English and German research, translation is a design choice requiring checks. Ask participants in each language whether categories and words are understood as intended. Specify whether you will describe findings by language or compare languages, and what evidence that comparison requires.

Use these working questions in the research plan:

• Internal consistency: how coherently do components of the measurement contribute to the intended score?
• Test–retest reliability: to what extent are differences between participants maintained when measurement is repeated under comparable conditions?
• Validity: what evidence supports the interpretation you want to give the score?

Request evidence appropriate to your application. A claim about a group average requires different support from a fixed classification of one person. Plan the sample around the question, comparison, required precision and expected exclusions. A general minimum sample size will not suit every design.

For quantitative conclusions, report relevant uncertainty; for qualitative interpretation, explain the data basis and divergent responses. If a score changes on repetition, consider both possible contextual change and measurement uncertainty. Do not use a reliability statistic as a substitute for substantive validation.

In this knowledge base, an associative target group is a bounded group of people who decide on the basis of similar associations within a specific choice context. This is a working definition. Association measures provide building blocks for this definition; their relationship with decisions also needs investigation.

Keep the steps visible: which associations were found, who shares them and which outcome relates to them? Collect an appropriate choice or behavioural measure for the last question. A relationship can be reported as a relationship. A causal claim requires a research design that adequately addresses alternative explanations.

If EEG is proposed, ask what additional research question, measure and interpretation it contributes. Do not convert an EEG result into an association, preference or purchasing score without separate validation. Combining methods is particularly useful when their roles are clear in advance.

A delivery service wants to understand what customers mean by reliability. This example is entirely fictional and contains no measured results or demonstrated commercial effects.

Suppose the team initially assumes that ‘fast’ is central to reliability. Exploration should allow other meanings, such as keeping promises or delivering without damage. Finalise the measurement task only after clarifying the concept under investigation.

An IAT is a possible next step here, not a mandatory component. First assess whether the categories and exemplars form an appropriate task. If a direct question or choice task answers the question better, organise the proposal around that.

The final advice should show which data support the recommendation. This example does not predict which attribute will matter most.

• Treating an unmentioned association as absent without considering the question asked.
• Presenting frequency, positive evaluation and choice relevance as the same score.
• Calling every rapid-response task an IAT.
• Choosing favourable exclusion rules after examining the results.
• Translating a group difference into a certain claim about every participant.
• Assuming a language version or new stimulus set is automatically comparable.
• Presenting a statistical relationship as proof of a cause of purchasing.

Use these points when reviewing the proposal, analysis and report. The research dossier should make the path from response or task performance to advice traceable.

A good plan names the association being investigated, the comparison, participants, task and interpretation limits. Request original stimuli, scoring rules, quality checks and a supported connection to the decision you want to make. This clarifies which conclusion the data support and which follow-up question remains open.