Oh! It occurred to me to take a look at the comparisons of the last post, but looking at truly-positive and truly-negative correlations separately. Here’s for the scenario where all correlations are rlkj_cor():
And a new plot, showing the absolute deviation of the posterior median from the true value:
A little strange that while the MVN demonstrates greater bias and lower uncertainty at high K for both positive and negative true correlations, but nonetheless has calibrated recovery at those parameter combinations. I also don’t really get why the Deviations don’t show the same pattern as the Bias.
Will post the mvn-plus-some-high-cors-data version shortly, but wanted to get these somewhat surprising results from the plain-ol’-mvn-data out for comment in the interim.