What are association clusters?

When choosing a service, people may mention ‘overview’, ‘control’ and ‘predictability’. You can investigate these words together, but why do they belong together? An association cluster becomes useful when you explain the relationship you mean, how you investigated it and what it means for the choice people make.

In this knowledge base, an association cluster is a bounded set of associations that are related within a defined context according to an explicit interpretive or analytical rule. This is our working definition. The rule might concern shared meaning, reported connections or similar measurement patterns.

Always specify what you are grouping: words, statements, measured associations or people. A cluster of associations does not yet constitute a target group. Defining an audience requires additional information about who shares these associations and how they relate to decisions.

We use ‘association cluster’ as an umbrella working term. The publications below support different building blocks; we do not identify any of their authors as the inventor of this term.

In 1975, Collins and Loftus described semantic processing through a model in which activation spreads between connected concepts. This provides historical background for thinking in terms of associative networks. Collins & Loftus, 1975.

John and colleagues introduced Brand Concept Maps, a method for eliciting individual brand association networks and combining them into a consensus map. It can represent direct and indirect connections around a brand. John et al., 2006.

For this article, we therefore distinguish the memory model, the research data and the final presentation. A drawn map is not a recording of neural connections. Referring to a memory model also does not validate a classification of your own.

Braun and Clarke discuss thematic analysis as a qualitative method for examining patterns of meaning. Analytical choices and the researcher's perspective need to be explained. A theme is therefore not automatically a statistically identified cluster. Braun & Clarke, 2006.

Jain discusses cluster analysis and its dependence on factors including data representation, similarity measures and algorithms. A grouping is therefore tied to the analysis chosen. Jain, 2010.

Use this table as a briefing aid. The approaches can complement one another, provided you explain how each finding was produced.

Here, an associative target group means a group of people who make decisions based on similar associations within a bounded choice context and can be delineated on that basis. This is also a working definition, not a universally validated classification of people.

A marketing target group might be described through industry, job role and company size. For an associative target group, you investigate the shared meaning that guides a choice. People in different roles may belong to the same group. People with the same role need not share a decision pattern.

Make the intermediate steps visible. First describe relationships between associations. Then investigate differences between people. Next, test whether those differences matter for the specific decision. If you only have a list of words, a proposed audience remains a hypothesis.

Where appropriate and permitted, retain responses at participant level. An average map across all participants cannot, without additional analysis, show which associations occur together for the same person. Also specify how you handle people who partly fit several proposed groups.

Use separate operational definitions in your research plan for:

• Frequency: how often an association is mentioned, specifying the denominator: participants, responses or measurement occasions.
• Strength: the result of the particular measure chosen for a connection. State the task and calculation.
• Valence: how positively or negatively an association is evaluated in the context studied.
• Choice relevance: the investigated relationship with a defined decision or outcome.

These recording rules prevent one score from carrying several meanings. A frequently mentioned objection might deserve discussion without being a positive brand association. If you use a scale from ‘--’ to ‘++’, define the categories precisely. Do not present that coding as a percentage or a universal neuroscientific unit.

The following steps are our practical editorial guidance. They do not prescribe a fixed Neurofactor measurement protocol.

1. Define the choice. Describe who chooses what, between which alternatives and in which situation. ‘Choosing a supplier when a contract expires’ is more specific than ‘understanding the customer’.
2. Document data collection. Record stimuli, questions, order, language and participant selection. Separate spontaneously mentioned associations from responses to suggested words.
3. Choose the unit of analysis. Do rows in your dataset represent associations, statements or people? Also name the variables being compared.
4. Make the grouping traceable. Retain original responses, merging rules and exceptions. Explain why ‘control’ is interpreted as oversight or personal agency.
5. Examine the boundaries. Consider counterexamples, overlap and sensitivity to reasonable alternative groupings. For statistical analyses, document settings and the treatment of missing values.
6. Test the application. Investigate the proposed relationship with choice behaviour using additional data. Specify in advance when you would revise or reject the audience definition.

Rousseeuw introduced the silhouette as a tool for assessing cohesion within and separation between statistical clusters. It assesses a grouping within the chosen data and distance measure. Rousseeuw, 1987.

For an audience decision, we also propose three practical questions. Does the cluster have an understandable meaning? Does the proposed boundary remain useful with new participants or a repeated measurement? Does it help explain or predict a previously specified choice or response?

Choose checks that fit your method. For qualitative work, request supported interpretations and discussion of divergent responses. For a statistical model, request a check using data that were not used to select the grouping. A good internal clustering metric does not, by itself, answer a question about purchasing behaviour.

An organisation wants to understand how team leaders assess scheduling software. The statements below were invented for this explanation. They are not customer quotations or research findings.

The researcher can investigate whether these themes fit under a broader label such as ‘control over scheduling’. An alternative is to keep them separate because overview, agency and flexibility may describe different needs. Both proposals must be traceable to responses and context.

Only then comes the audience question: is there a recognisable group that shares these associations and considers them when choosing software? People might request an overview but ultimately choose on integrations or price. Incorporate such findings instead of defending the initial cluster label.

One possible next step is to test two information variants based on the findings. Decide beforehand whether you measure understanding, preference or an actual choice. This example makes no claim about which variant performs better.

• Treating an appealing label as evidence. Include underlying responses and borderline cases in the report.
• Presenting a combined word list as segmentation of people. Show which data support the audience boundaries.
• Giving every connection the same meaning. State whether a network edge represents a reported relationship, co-occurrence or calculated similarity.
• Forcing a fixed number of clusters for presentation. Justify the grouping through the research question and analysis.
• Including AI suggestions as participant responses. Keep generated hypotheses separate and check classifications against the original data.
• Claiming a cause of choice from an association. Design an appropriate follow-up test before using causal language.

A useful report gives the context, original data, grouping rule, interpretation and uncertainties for each association cluster. Applying it in a target group profile adds a supported definition of the people and the relevant choice. This makes clear which claims have been investigated and which still need testing.