Fixes for JAGS 5.0.0. Reduce variance of diffuse normal priors.

[martynplummer]

In preparation for the release of JAGS 5.0.0 I am going through the reverse dependencies of the rjags package to ensure they work with the new version.

Changes to the way that initial values are calculated mean that JAGS does not work well with the extremely diffuse priors found in the missingHE package. In JAGS 4.x.y, initial values were set to deterministic "typical" values. For the normal distribution, the typical value is the prior mean. This allows the use of extremely diffuse normal priors like dnorm(0, 0.0000001) with a guaranteed starting value of 0.

In JAGS 5.0.0, initial values depend on the prior distribution. Ideally, initial values are drawn from the prior distribution but this is not compatible with diffuse priors. As a compromise, JAGS 5.0.0 sets the initial value to be the mean of N draws from the prior where by default N=10000. This works for, e.g. dnorm(0, 0.001) where it generates intial values with mean 0 and variance 0.1. However, this compromise still does not work for the very diffuse priors used in the missingHE package as it leads to the error message about parameter values being incompatible with the data. My suggestion is that you increase the prior precision of mean parameters. I don't think this will make a material difference to your parameter estimates but allows stable initial values to be generated.