Michael Betancourt discusses the trade offs here. Interestingly, in this section Betancourt uses a heuristic criteria to assign parameters with fewer observations (< 25) a non-centered parameterization and the rest a centered parameterization. In his toy example, the mixed parameterization is better suited (i.e. shows no divergences) than the uniformly centered or non-centered parameterizations.
I’m still trying to wrap my head around it myself, but this is something I am trying to apply to my own problem at the moment.