Closed ito4303 closed 8 years ago
These look great. I made some comments, but there's not really anything you need to change. Are you making sure for these that you recover roughly the same estimates as the original code did with BUGS?
Oh, just let me know if it's OK to merge or go ahead and merge it yourself (I don't know if you have that permission).
Thank you for reviewing. I always confirm that results are similar to true values (for simulated data) or the results of BUGS (for real data).
However, I usually do not compare running times. In case of the model in section 11.6 (ipm_hoopoe.stan), Stan took 97 seconds including compilation (4 chains in parallel, 12000 iterations/chain) while WinBUGS took 123 seconds (via Wine, no parallelization, 3 chains, 20000 iterations/chain) on my machine (MacBook Pro, 2.5 GHz Intel Core i7).
I don't think I have the permission to merge. Please merge the code if there is no problem.
That's probably way more iterations than you need for Stan.
You only need to run up to some reasonable effective sample size for whatever inferences you need to make.
To compare times, you really want to measure time per effective sample size. And compute those effective sample sizes with multiple chains using Stan's more conservative n_eff estimator. But no worries here --- you don't need to do that for this issue, I was just curious.
I translated models in Chapter 11 of BPA. Would you please review the code?