Sawtooth Software: The Survey Software of Choice

What is the Latent Class Segmentation Module?

Choice-based conjoint data have traditionally been analyzed by estimating average part worths for groups of respondents. However, if there are distinct segments, a model that recognizes them can produce more accurate results than an aggregate (single group) solution. Although one can conduct separate analyses for subgroups differing by demography or product-usage, it has been difficult to do segmentation based on choices themselves.

The CBC Latent Class Segmentation Module is an add-on analytical tool that is integrated within the Lighthouse Studio and SMRT interfaces. Data are automatically available in SMRT when using its CBC component, and may be easily imported from Lighthouse Studio. It uses choice data for the simultaneous development of segments and estimation of part worth utilities. For example, one segment might be composed of price-sensitive shoppers, and another might be composed of those who usually select premium brands. Each respondent has a probability of belonging to each segment, but can be classified into the most likely segment for subsequent tabulation.

This module has other features that may be of interest to CBC users: it permits weighting of respondents, and you can specify attributes for which utilities should be monotonic, such as for levels of price or quality. Choosing the number of segments is facilitated by specifying a range to investigate, such as from 2 through 30 segments. Statistics are provided for assessing goodness of fit for each solution, and resulting group membership for different solutions is tabulated with one another.

If the market is truly segmented, conducting Latent Class analysis prior to using hierarchical Bayes estimation might be beneficial. CBC/HB can be run separately within groups of relatively homogeneous respondents identified by Latent Class to estimate individual-level part worths. This approach has theoretical merit, but evidence from practitioners suggests that segmenting prior to running HB has not improved overall model fit to holdout criteria.