Latent Class Analysis with Dual-Response None

Question

Dear all,
based on the data of a CBC I would like to perform a latent class analysis. In the CBC a Dual-Response None question was implemented, asking the respondents if they would switch from their current contract to the selected one. This allows us to look at the willingness of respondents to switch. I did several runs with and without the dual none as a parameter in the estimation. The segments are similar to some extent in their preferences, but there are still differences, for example in the segment sizes.
Is it better to perform the estimation of the preferences using only the product attributes and their levels or should I consider the Dual-Response None as a parameter?
How much does the integration of the "Dual-Response None" influence the estimation of the preferences?

Thank you!

Answer 1

This is a good and also challenging question.  If the purpose of running latent class were only to improve the accuracy of market simulator predictions (compared to an aggregate MNL solution), then the answer would be to include the None tasks in the latent class estimation and to use a large number of classes (say, 12+).

However, I suspect that the purpose of your latent class analysis isn't to improve market simulation predictions, but rather just to detect market segments for strategic segmentation purposes.  If that's the case, then it's hard to say which latent class segmentation would be more useful to you (one that included the None parameter or one that ignored the None information).

You can get a sense of how much the None parameter is influencing your latent class segmentation result by comparing the utility of the None across your latent class segments.  If the None utility is about the same, then you would think that it isn't driving the latent class results much.

Comments

Thank you for the answer!
You are right that i would like to detect market segments for strategic segmentation purposes. So I understand that it is hard to choose between the options.

The CBC consisted of 8 choice tasks with 3 concepts each. There were 4 attributes with 2, 3, 4 and 4 levels. The adjusted data set consists of 777 participants.
In the latent class analysis, for the 4 segment solution,  I get approximately the following values for the Dual-None Response parameter:
Raw Utilities: -0.85/ 3.13/ 0.50 and 5.07.
Zero-centered diffs: -50.71/ 215, 03/ 39.13 and 310.78

How would you interpret the influence of these values?

Thanks!