Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

Is it possible to apply respondent weighting to the upper model used in HB estimation?

We haven't seen an option for this in the HB software and there's probably a good reason for this but it would be good to understand why this is. We have a MaxDiff on occasion-based needs when consuming food with a main and boost sample (boosted on certain food categories). The main sample is nat rep so ideally we would apply a weight to also make the main + boost sample nat rep when it comes to the upper model used in the HB estimation.

Many thanks!
asked Jan 2, 2018 by anonymous

1 Answer

0 votes
This is something that our Bayesian stats people here have looked at over the years.  We did some investigation many years back, but didn't find an approach that we felt was satisfactory.

When you do face a situation with HB in which a group has been oversampled by a relatively large factor, then it's a good idea to run HB with covariates, then reweight the results post hoc.  (That way the oversampled group doesn't bias the undersampled people via the upper-level model.)

For example, imagine you oversample cat-lovers by a large factor, so they are no longer 25%, but 60% of the population.  You should create a 2-category covariate variable (1=cat lovers, 2=other) and run HB estimation with this covariate specified.  Then, after obtaining the results (posterior utilities), you should weight the sample as you are intending to do so, such that the cat lovers again make up 25% of the weight of the population.

(BTW, I don't know the actual % of cat lovers in the USA.  My daughter is certainly one of them.)
answered Jan 2, 2018 by Bryan Orme Platinum Sawtooth Software, Inc. (159,360 points)
Thanks very much for your quick reply (and happy new year!). We suspected that might be the case but it's good to know that it's been investigated before.