Segment finder

Identifying segments—groups of respondents who share similar attitudes, needs, and preferences—is a powerful tool for guiding strategy. A strong segmentation solution should group respondents who are alike within each segment and different between segments. Researchers often explore 3 to 6-segment solutions, as these typically strike the right balance between simplicity and strategic value.

To evaluate the usefulness of a potential segmentation solution, researchers examine the responses to other questions in the survey (e.g., via filters or cross-tabulation) by the different segments. This helps uncover actionable differences that can inform targeting strategies, product design, messaging, and more.

MaxDiff and CBC data are especially well-suited for segmentation because they’re based on choice rather than ratings:

• No scale bias: Unlike rating scales, choice data isn’t affected by individual differences in how people utilize scales (e.g., always rating high or low).
• More meaningful discrimination: Respondents can’t say everything is equally important—CBC and MaxDiff force trade-offs, encouraging thoughtful responses and reducing straightlining behavior.

Segment finder uses latent class MNL analysis to uncover meaningful segments, or groups, of respondents who share similar preferences based on their choices in a MaxDiff or CBC exercise. Before running Segment finder, we recommend you clean your data of inconsistent or fraudulent respondents.

Unlike HB analysis, which estimates a unique set of utilities for each respondent, latent class analysis identifies patterns in the data and groups respondents who tend to make similar choices. These groups, or segments, help you understand how different groups of people evaluate your products, features, or messages.

Latent class analysis estimates two key things:

• Part-worth utilities for each segment, summarizing how that group values the items in the exercise.
• Segment probabilities for each respondent, indicating the likelihood that they belong to each segment.

Once segments are identified, they’re saved as categorical variables. You can use them to segment your MaxDiff or CBC results, or create crosstabs with other survey responses.

Latent class analysis is an iterative estimation process that uncovers patterns in respondent choices. Here's a simplified overview of how it works:

1. Start with random utility estimates for each segment.
2. Calculate the probability (using the MNL equation) that each respondent belongs to each segment, based on their answers.
3. Re-estimate utilities for each segment using aggregate MNL, weighted by the respondent probabilities from the previous step.
4. Repeat steps 2–3 until the model stabilizes and improvements in fit stop (known as convergence).

Each respondent ends up with a probability of belonging to each segment. If the segmentation structure fits the data well, the highest of these probabilities is often close to 1, indicating a strong fit to a particular group.

Technical note: Segment finder uses a scale-constrained version of the Latent Class MNL model. This prevents solutions where two segments appear to differ only in response error (scale) rather than true preference. As part of step 3 above, we normalize the scale for each class to match the weighted average scale across all classes. While this constraint may result in slightly lower model fit compared to the traditional (unconstrained) LC-MNL model, the benefits of clearer, more meaningful segmentation often outweigh the tradeoff.

Segmentation is as much an art as a science. A statistically “strong” solution should also make sense from a strategic or managerial perspective. The best number of segments is often subjective, though we typically recommend staying around 5–6 groups or fewer. A measure of fit is described further below.

Segment finder automatically calculates all solutions starting from two segments up to your selected maximum.

For a deeper dive into latent class analysis, read our technical paper on the topic.

To run Segment finder:

1. Navigate to the Analysis section of Discover.
2. Select the Segment finder item under the exercise (CBC or MaxDiff) you'd like to analyze.
3. In the right panel, set the Maximum number of segments.
4. Click the Run button to begin analysis.

Once segment finder finishes running, each solution will appear in its own tab at the top of the screen.

At this point, each of the solutions is also available as an option for segmentation in the other areas of exercise analysis.

Each solution includes a set of tables to help you understand how respondents were grouped and to evaluate whether the solution is meaningful and actionable.

Hierarchical Bayes (HB) utilities segmented by latent class groups and the latent class utilities reported by Segment finder are not the same, though they typically tell a very similar story.

The table below compares Segment finder (latent class) utilities of a 2-group solution compared with the averaged HB utilities for the same groups of respondents.

At first glance, the values may appear quite different. However, the correlations between the corresponding segments (Segment 1 : Segment 1 and Segment 2: Segment 2) are both 99%, and the underlying patterns are consistent:

• Brand 1 is generally preferred, with Segment 1 showing a stronger preference than segment 2
• Segment 1 is much less likely to choose the none option
• Segment 2 is more price sensitive, with utility decreasing more as price increases

The differences in this example are expected and stem from several factors, including:

After rescaling, each respondent has a total utility range of 200 (100 × 2 attributes).

Now suppose Segment finder groups these two respondents into the same segment. After rescaling that segment to the same total range of 200, the resulting utilities would be:

The overall story is the same — both respondents prefer Blue to Red — but the numbers differ depending on whether rescaling happens at the individual level or from the pooled latent class utilities. Because HB and latent class are also different mathematical models, raw utilities will differ even before rescaling is applied.