Question

ACBC- Separate Participant Groups, One Market Model, Separate ACBC?

We are building an excel based market share model for alternative car vehicles using an ACBC with HB and a survey. We know that the socio demographic data and vehicle preferences may differ a lot within our participant groups (e.g. University students vs. older car buyers). We even want to prove these differences with the market model.
Is it therefore useful to have separate ACBC parts with separate HB estimation for each group in the survey? We are currently planning to use only one ACBC for the whole experiment. We reckon that we will receive better estimates from higher numbers of participants in the HB process, which would be lower, if we separate the participants in separate ACBC.

Do you agree?

Thank you for your help.

Answer 1

This is that issue of Bayesian Shrinkage to a global population mean for conjoint data that has been debated at our conferences over the last 10 years.  The findings for CBC have been that Bayesian shrinkage indeed can decrease the estimated differences between people and repondent groups.  But, even if that happens, the market simulation results and predictions are usually no better by including covariates in CBC, or segmenting the respondents into different groups and running HB separately within the groups.

But, the issue is even less a concern for ACBC, because there is more information at the individual-level from ACBC than CBC.  So, there is less Bayesian shrinkage happening for ACBC than CBC (because there is more certainty at the individual level and less need to borrow information across the sample).

If this really concerns you for ACBC, you can take additional steps to reduce the Bayesian Shrinkage, by increasing the Degrees of Freedom and Increase the Prior Variance assumptions during ACBC/HB estimation.  If you follow the recommendations in the software, you'll be getting robust individual-level information, allowing for stable individual-level estimation.  You'll be OK.