When I run the model with transform = tr.log, I underestimate the value of alpha and beta.
According to Gelman, the expected means for alpha and beta should be 2.4 and 14.3 respectively. I can verify this by directly computing the posterior.
When I run this model, I estimate the means for alpha and beta to be 1.8 and 10.8. Running without transform = tr.log yields estimate closer to the true values. Since you take the log in logp_ab, is transform = tr.log needed again?