Hierarchical prior (for partial pooling) on correlation matrices?

There are things that can be done, but almost nothing besides the constructions that Stan uses guarantees positive definiteness. One thing is to utilize the fact that a convex combination of correlation matrices is a correlation matrix. So, if you have \mathbf{\Lambda}_0 as the common correlation matrix and \mathbf{\Lambda}_j as the deviation in the correlation matrix for the j-th group, then \left(1 - \alpha\right) \mathbf{\Lambda}_0 + \alpha \mathbf{\Lambda}_j is the correlation matrix for the j-th group with \alpha \in \left[0,1\right]. You could then put LKJ priors on \mathbf{\Lambda}_j with shape \frac{1}{\alpha} or e^\alpha to encourage the group-specific deviations to be small.

4 Likes