In response to a comment by Emmanuel Charpentier, I should write a few words about what has changed in JAGS 3.3.0.
Firstly, this is entirely a bug-fix release: there are no new features. As the number of JAGS users increases, people are using it in more diverse ways and uncovering bugs that were previously hidden. Feature development has been almost entirely on hold while I hunt down these bugs.
Bugs that cause JAGS to fail with an error message are the easiest to deal with. Even if the message makes no sense to you, I can usually track down the problem if you send me a worked example. The more concerning kind of bugs are silent. These are bugs that return an invalid set of samples without any warning. We have to wait until JAGS returns some egregiously wrong samples before such bugs are uncovered.
JAGS 3.3.3 fixes three silent bugs:
- It was possible for an initial value to overwrite a data value. This meant that you could get different results from the same model by supplying different initial values, (because in fact you were looking at different data). JAGS will now give an error if you try to overwrite an observed value during initialization.
- The uniform distribution was giving the wrong likelihood when used as an outcome variable. Note that there were no problems with uniform priors on hyper-parameters, which is the most common use of the uniform distribution.
- There were a number of problems with the IWLS sampler in the glm module. Fortunately, the IWLS sampler has low precedence so would normally not be used. This has now been more thoroughly tested and it will give correct samples, although it may still have very poor efficiency.
If you want to see more details of what has changed, see the NEWS file in the source distribution.