Bimodal clusters: meaning and application

You measure associations around a theme and notice something striking: many people are clearly positive, many others clearly negative, and hardly anyone is in between. The average then lands near neutral, although almost nobody is neutral. This article briefly explains what such a bimodal cluster can mean and how to investigate it.

A bimodal cluster is a Neurofactor working concept for a theme of associations in which valence has two peaks, one positive and one negative, with few or no associations in the neutral middle.

It may point to two subgroups with different meanings, or to ambivalence: the same people are positive and negative at once. Which applies becomes clear only when you look at each person and probe further. It signals a need for further research, not proof of two target groups.

Associations that belong together form a cluster or theme (see What are association clusters?). Each association also has a valence: positive, neutral or negative.

In a bimodal cluster, the valences gather around two peaks: clearly positive and clearly negative associations, with little in between. "Bimodal" comes from statistics and means a distribution with two peaks. "Bimodal cluster" is a Neurofactor working concept, not a validated classification.

Look at the distribution of valence, not only at a summary figure. For each cluster, plot how often each valence value occurs, for example in a bar chart. Two peaks at the ends and a dip in the middle point to a bimodal pattern.

Also weigh association strength: two peaks of weak associations say less than two peaks of strong ones.

Averaging the valences often puts a bimodal cluster near neutral, a figure that describes almost nobody. Kaplan pointed to this problem in attitude measurement: a neutral position on a scale from negative to positive can mean indifference, but also ambivalence. He proposed measuring positive and negative feelings separately (Kaplan, 1972).

DiMaggio, Evans and Bryson described opinion polarisation as a movement towards separate peaks and showed that division can grow while the average barely shifts. So inspect distributions, not only means (DiMaggio et al., 1996).

Perhaps the cluster contains two groups who give the same theme different meanings. One group links it to something they want, the other to something they avoid. Each person is then fairly clear-cut; the two peaks appear because you look at both groups together.

That is a reason to investigate whether different associative target groups exist, not proof of it. First test whether the difference recurs in other themes and relates to choices.

Perhaps the same people mention both positive and negative associations within the theme. They are not neutral but divided: ambivalent. Priester and Petty describe how an attitude can have positive and negative sides at once, and how a neutral answer on a bipolar scale loses information about that conflict (Priester & Petty, 1996).

That is tension within people, not between groups, and it calls for a different follow-up question.

The cluster distribution itself does not show the difference. Go back to the participants.

1. Look at each person. Do participants mainly mention one side, or both? The first points more towards subgroups, the second towards ambivalence. A mix is also possible.
2. Compare with other themes and characteristics. Are the two camps related to a situation, an experience or patterns in other clusters?
3. Hold in-depth interviews. Ask people from both peaks what the associations mean to them and when they tip the balance.
4. Test statistically where possible. The dip test measures how far a distribution departs from the best-fitting distribution with a single peak (Hartigan & Hartigan, 1985). With few participants and a short valence scale, such a test is of limited value.

See also How do you measure associations?.

A bimodal cluster is a pattern; subgroups and ambivalence are possible explanations.

This example is fictional and contains no research outcome.

A fictional provider of home-delivered meal kits collects associations from people considering a subscription. Within the theme "planning and convenience", "no grocery stress" and "something new every week" are strongly positive; "being tied to a subscription" and "less freedom in what I eat" are strongly negative. There are hardly any neutral associations.

The provider draws no conclusion from the average. It first checks for each participant whether people sit on one side or on both, and then holds interviews. Only then does it know whether there are two groups, or people who want the convenience but fear the commitment.

• Not a neutral theme. An average near zero does not mean people find the theme unimportant.
• Not proof of two target groups. Two peaks can also arise within the same people.
• Not a validated classification. It is a working concept for this measurement and this group, with no fixed threshold for "bimodal".
• Not a prediction of what one person will choose.

A bimodal cluster shows that a theme divides people rather than leaving them indifferent. So look at the distribution, not only the average. Then investigate, person by person and in interviews, whether you are dealing with two subgroups or with ambivalence, before you base target groups or messages on it.