Finding Segments via Latent Class Analysis

Top  Previous  Next

Note: Latent Class is often used to discover market segments (for use as banner points) from MaxDiff data.  Segment membership is reported on the Segment Membership tab of the output.

Lighthouse Studio includes a built-in Latent Class estimation routine for MaxDiff.  You may decide instead to analyze data separately using our standalone Latent Class software, but the two routines use the same procedure and it is more convenient to perform the analysis within Lighthouse Studio.

When you run Latent Class  (Analysis | Analysis Manager select Latent Class as the Analysis Type and click Run), the results are displayed in the report window and utility scores are saved into a subfolder within your project directory.  You can weight the data, or select subsets of respondents or tasks to process.  You can constrain certain utility scores to be higher or lower than other utility scores.


Interpreting Latent Class Scores


Lighthouse Studio reports the preference scores on three different scales:


Zero-Centered Interval Scores

Rescaled Scores (0 to 100 scaling) also known as Probability Scale

Raw Scores


When comparing differences across groups, we recommend you use either the Zero-Centered Interval Scores or the Rescaled Scores (0 to 100 scaling) rather than the Raw Scores.  Not every group might think the reference item (the final item) has equal utility, and yet Raw Scores scales all items with respect to that final item set to 0.  Moreover, with Raw Scores, the more consistent the answers are within a group (the better the within-group model fit), the larger the magnitude of the scores.  Just like you shouldn't compare the weight of a person standing on a planet of unknown size in another galaxy to a person standing here on the earth, you shouldn't compare latent class groups directly on the Raw Scores.  The scores for the two groups may be on different scales.  


Zero-Centered Interval Scores zero-centers and normalizes the raw utility scores to have a constant range (such as 100), allowing you better ability to compare results across different groups of respondents (who might have different degrees of consistency and who may have quite different preferences for the final reference item).


Rescaled Scores (0 to 100 scaling) uses the exponential transform to place the items on a ratio scale, where the sum of scores for each group is 100.  The ratio scale gives you the ability to say that an item with a score of 2.0 is twice as preferred as an item with a score of 1.0.  The Raw Scores and Zero-Centered Interval Scores do not support such ratio comparisons among items.  We should note that the consistency of different respondent groups can affect the relative sensitivities of their ratio scales.  Click here for more details regarding the Rescaled Scores (0 to 100 scaling) procedure.


Suggestions Regarding the Use of Latent Class with MaxDiff


Latent Class is not guaranteed to get the globally optimal (maximum likelihood) answer each time it runs.  There is a random component to the process and it is possible for Latent Class to find a suboptimal solution, so you should ask the software to run multiple replicates from different starting points to make sure you achieve a high quality solution.  By default, 5 replications are performed.  You can increase this for even greater precision and confidence in your final solution.


Latent Class includes fit statistics that help indicate which number of groups best fits the structure of the data (e.g. 2-group, 3-group, etc.).  The standalone Latent Class manual. provides details regarding how to use these statistics to judge the quality of different solutions.


Selecting the right number of groups to use in your final solution involves a balancing act of judging the quality of the fit statistics for different solutions as well as considering their managerial usefulness.  The segments should be reasonably sized and have useful interpretation.


Some Details Regarding Latent Class Analysis


Latent Class divides respondents into segments having similar importances/preferences based on their choices in MaxDiff questionnaires.  It uses latent class analysis for this purpose, which simultaneously estimates utility scores for each segment and the probability that each respondent belongs to each segment.  Latent Class is an integrated analytical component within Lighthouse Studio (and we also provide a standalone Latent Class software system).  


Latent class has a role analogous to that of CBC's logit program, but rather than finding average utility scores for all respondents together, it detects subgroups with differing preferences and estimates utility scores (with logit scaling) for each segment.  The subgroups have the characteristic that the respondents within each group are relatively similar but the preferences are quite different from group to group. You may specify how many groups are to be considered, such as the range of 2 through 6.  A report of the analysis is shown on the screen, with multiple tabs.  Across the different tabs of the report, you'll find utility scores for subgroups along with each respondent's probabilities of membership in the groups.


The latent class estimation process works like this:


1. Initially, select random estimates of each group's utility values.
2.Use each group's estimated utilities to fit each respondent's data, and estimate the relative probability of each respondent belonging to each group.
3. Using those probabilities as weights, re-estimate the logit weights for each group.  Accumulate the log-likelihood over all groups.
4.Continue repeating steps 2 and 3 until the log-likelihood fails to improve by more than some small amount (the convergence limit).  Each iteration consists of a repetition of steps 2 and 3.  


Latent class reports the utility scores for each subgroup or "segment."   Latent class analysis does not assume that each respondent is "in" one group or another.  Rather, each respondent is considered to have some non-zero probability of belonging to each group.  If the solution fits the data very well, then those probabilities approach zero or one.


For more information on our latent class analysis approach, please see the full standalone Latent Class manual.

Page link: