Just for the record. As I was saying above. The optimal frequency is proportional to the standard deviation of the estimator used in the leaf node.I use proportioning by sub-tree size. If you use the Peter1 depth-first search it uses proportioning by sub-tree count of the leaf nodes. I haven't yet started to see if other formulas with variance term in it would improve it but I suspect mean proportioning is pretty good too.
But since it might be too expensive to get a good estimation of the standard deviation
(you cannot reuse those searches for perft) using the subtree size is probably a good idea.