I have found myself doing a large amount of Bayesian analysis and I am
uncertain what framework learn and I'm posing this question to get a sense
of perspective. As I am doing a lot of applied work right now, I need to do
a fair amount of "standard data analysis" (eg mixed models, glm, etc)
but I
also have to have the flexibility to write more complicated models. As near
as I can tell, there are three routes I could take:
1) Use OpenBUGS. This has the benefit of being well established but there
are a number of things I dislike. While the Windows OS can be addressed
through Wine, my biggest complaint is how restrictive the language is.
Trying to do anything complex requires using 'tricks' (the zeros trick
for
specifying an arbitrary likelihood, or the delta trick for specifying a
dirichlet distribution using gammas) that often require already knowing them
to be able to find them in the examples. The 'cut' function is pretty
awesome, though.
2) Use JAGS. Currently this seems like the most appealing because it looks
like the easier models can be implemented using the BUGS language and it
seems like the modules will provide a very nice way to specify more
complicated models. The downside is that I have no idea if it has a future
and I'm somewhat reluctant to put my eggs into a basket that might go
nowhere.
3) The MCMCpack library in R. This seems like it would be flexible enough
for the complicated models, but I don't know if is automated enough for the
easier analysis. For example, a hierarchical model has a likelihood that
has a large number of pieces and it is far too easy to make simple coding
mistake. Because of it's flexibility, I also wonder if I will find myself
always wrestling with what sampler to use.
Ideally I would write a model in BUGS and for part of the hierarchy have it
call an ugly likelihood function that I wrote in R and if the default
samplers weren't working, I could start tweeking them to figure out what was
happening. (I suppose while I'm wishing I might as well throw out something
about world peace and a cushy faculty position.)
Any advise and/or perspective would be appreciated.
Derek Sonderegger
--
View this message in context:
http://n4.nabble.com/What-Bayesian-Framework-tp1573084p1573084.html
Sent from the R help mailing list archive at Nabble.com.
