Typically I do not suggest breaking the sample into groups that are analyzed separately via HB unless the sample size per group is at least 400. A paper many years back examined the issue of breaking HB into subgroup analysis vs. running HB on the entire sample, and the results (given the sample sizes employed over multiple studies) showed no improvement (in hit rate) for HB by subgroup vs. HB overall...followed by post hoc dividing into groups.
However, if the main goal is to explore the differences between groups and to more appropriately test for differences between subgroups, then the covariates approach for HB is the more technically supported way to go about this. It involves specifying the grouping variable as a covariate and then examining the resulting alpha file (the draws from the upper-level model) after convergence.
However, this is not typically done unless there are at least 50 to 100 respondents per segment. 28 respondents vs. 7 respondents seems awfully thin.