I have a question regarding the HB estimation of max-diff utilities.

In HB estimation, information is used from both the respondent itself as well as the entire population (I assume that this is the case since the respondent itself has only seen a part of all possible combinations)?

I was wondering how much the share of the population is compared to the share of the respondent itself in the calculation of the utilities for each respondent.

When comparing max diff results across several breaks, e.g. age groups, we generally notice not that much differences between the groups. I was wondering if this has to do with the fact that information of the population is used as well in the HB estimation? Are the utilities mainly driven what the respondent has answered or has the population also a quite high influence in the estimation?

Many thanks,

Lieve Audoorn

1. To the degree that respondents answer few choice tasks relative to the number of parameters to be estimated, they will "borrow" a lot from the population means (be smoothed toward the population mens). We've recommended each respondent see each item 2x or 3x (preferably) for HB estimation. If you show each item 3x, the Bayesian smoothing should not be super great. But, if you show each item 1x (or fewer) to each respondent, the Bayesian smoothing will be fairly substantial.

2. If you want to be able to emphasize differences (avoid so much Bayesian smoothing) for certain respondent groups, you can include "covariates" in your HB modeling on the Advanced tab (e.g. large company vs. small company or age groups). This makes it so that respondents are smoothed toward the means of their covariate cohorts. SSI Web v8.2, to be released next week, will include covariates within its HB interface within the MaxDiff Analysis menu area. And, of course, our CBC/HB software has been supporting covariates for quite some time now.