Let’s be clear. This is a numerical problem caused by overflow in the log-density calculations. JAGS is perfectly capable of slice sampling from the beta(0.5, 0.5) distribution, even when starting from an extreme value. The problem seems to arise when using a beta or Dirichlet distribution in a hierarchical model. In some cases, the sum of the contributions to the log-density seems to overflow, yielding an apparent infinite density.

So what can we do in such cases? Earlier versions of JAGS just stayed put. Sometimes the chain can recover at the next iteration, because the values of the other parameters have changed. This is the case for the “classic bugs” LITTERS example. But there could be models in which the slice sampler gets permanently stuck. The user would know nothing about it unless they were monitoring that node and the output would potentially be invalid.

Truncating beta distributions away from the boundary seems to work, but is clearly a sticking plaster solution. In the above-mentioned LITTERS example, I use

p[i,j] ~ dbeta(a[i], b[i]) T(,0.9999)

More constructive suggestions are welcome because this problem is a lot more common than I would like.