New Latent Class Feature: Scale Constraints

On January 20, 2014, we released a new version of our stand-alone latent class software, version 4.7.0. It introduces a new optional feature called “scale constraints.” It constrains the solutions to ones where the vector of utilities for each group has the same standard deviation (where standard deviation is meant to be a proxy for scale factor).

Scale-constrained latent class avoids segmentation solutions where two segments have essentially the same pattern of preferences, but where one segment’s utilities are just larger in magnitude than another. Researchers employing CBC and MaxDiff should recognize that when respondents answer with less error, their utilities uniformly stretch in magnitude (scale). Respondents answering less consistently (more error) see the scale of their utilities shrink.

It isn’t a given that all researchers want to avoid segments that differ mainly in terms of scale. Some may find it useful to discover and target segments that are more consistent and certain of their answers than others.

We implement scale-constrained latent class in a very simple manner. It is just another type of constraint, similar to Utility Constraints. The user just clicks a check-box. In each iteration of the routine, we find the multiplier for each group’s vector of part-worth utilities that sets each group’s scale equal to the average scale of all groups, where the standard deviation of the utilities is the proxy for scale.

Certainly scale in logit-based analysis is more complex than the standard deviation across a set of part-worth utilities. Still, we hope this new feature in our latent class software gives you a simple and useful new option for developing segmentation solutions that inform effective managerial strategy.

Scale-constrained latent class always involves at least a slight loss in model fit, due to the constrained nature of the solution. If the goal of the latent class analysis is to predict respondent choices, then scale constraints are detrimental to that aim. Since most of our users use HB (rather than latent class) to predict respondent choices within market simulators, this is not a concern.

Most of our users who do employ latent class use it principally to find segments and assign respondents to those segments. Those segment memberships are useful as filters and banner points within crosstabs and market simulators. Most of our users typically rely on HB to estimate individual-level part-worth utilities for use within market simulators.

Read more about our implementation of Scale Constrained Latent Class:

The standalone Latent class software is currently available in our Downloads section. The SSI Web version that will include this new feature (version 8.3) will be released during Q1 of 2014.