Agreed. You are probably right about the outcome. It is easy to just add an interaction of nominal score group (high, med, low) vs. mismatch/match and see if there is an effect.

At our startup we get depersonalized credit data but we don't have a nice panel like the Fed :) so cannot really do that analysis on our end. I would be nice if they released this data publicly -- once it is setup for modeling (with just a few variables), it is so hard to re-identify so the risk of privacy invasion is low. I doubt that will happen though.