STAN workshop Oct 24-25 Sempach Switzerland (Swiss Ornithological Inst.).

48 views
Skip to first unread message

Jeffrey Royle

unread,
Jul 8, 2014, 3:43:51 PM7/8/14
to unma...@googlegroups.com, Andy Royle
hierarchical modelers,

 This is a bit off-topic but probably of interest to many.

 The exciting new STAN software has been widely touted as one of the most serious competitors to become the successor of the BUGS software. This course is taught by two of the developers of STAN and hence offers a unique opportunity to kick-start your use of STAN.

regards,
andy


course_announcement_STAN 2014.pdf

Daniel Hocking

unread,
Jul 8, 2014, 9:56:40 PM7/8/14
to unma...@googlegroups.com, hmec...@gmail.com
Thanks for the info Andy. I know that for a while STAN couldn't handle discrete random effects, which are needed for all the occupancy and count models that account for imperfect detection. It put me off of learning STAN. Does anyone know if this ability has been improved or if anyone has tried one of these unmarked type models in STAN?

Thanks,
Dan
-----------------

Chris Sutherland

unread,
Jul 8, 2014, 9:59:23 PM7/8/14
to unma...@googlegroups.com

I dont know for sure but I spoke to someone last week who said that this is still the case.

--
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unmarked+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Kery Marc

unread,
Jul 9, 2014, 3:06:45 AM7/9/14
to unma...@googlegroups.com

Hi all,

 

yes, STAN still can’t  deal with discrete random effects (= latent variables, such as occupancy and abundance states in site-occ or Nmix models). However, I believe that the developers are aware that there is a great interest in this and may be able to deal with this in the future.

 

Kind regards  --- Marc

Dan Linden

unread,
Jul 9, 2014, 12:00:57 PM7/9/14
to unma...@googlegroups.com, hmec...@gmail.com
The link below contains Stan code for a working occupancy model and a suggestion from Bob Carpenter (a Stan developer) that it will be incorporated into an ecology portion of the manual.  He also references work on a Cormac-Jolly-Seber model in Stan. 

https://groups.google.com/d/topic/stan-users/EpX-wzOj7L4/discussion

The problem is that discrete random variables need to be marginalized out to calculate the gradient of the log probability function.  The Stan manual has a chapter devoted to Mixture Modeling where it discusses approaches to fit, for example, zero-inflated models.  The whole advantage of Hamiltonian Monte Carlo is the speed and convergence improvements of calculating that gradient.  But as pointed out in the discussion above, it seems like marginalizing out discrete parameters defeats the purpose of Bayesian estimation.

I suppose the details ultimately don't matter if the workaround still improves estimation metrics compared to BUGS models.  But it appears that this MCMC sampler is fundamentally not designed to handle discrete random variables.
Reply all
Reply to author
Forward
0 new messages