I have an issue with a relatively small sample (N=500) for a new tariff plan launch test.
The new product is much more appreciated with the group of operator's subscribers and some of less liked (but still liked well) by competitors' subs.
It would be ok if there was natural distribution on the operators subs in the sample. But it's not - there are 50% our's to 50% competitors.
So HB sometimes returns the patern of our operator's subs to competitor ones.
Some group memberships, or weights would do the job.
In a least liked scenario I'll splt the HB estimation by 2 groups and estimate 2 different models.