Anja Bertsche

As you already mentioned, we sample from a discrete approximation to the underlying continous distribution. So the result is dependent on the chosen grid and its range. So the range of the prior needs to be wide enough for the posterior distribution! Unfortunately, the wider the range (i.e. more grid points), the slower the MCMC simulation. So I want to ask again, if the implementation of a user-defined function as e.g. dnorm() is possible without C++ knowledge?

Thank you, that works very well!

I admit my description was not really clear. So say you have an MCMC sample of a posterior distribution from a previous run (no conjugate model, maybe multimodal distribution). Now you have the following possibilites to use the resulting posterior as new prior: Fit a known distribution, e.g. normal, gamma, exponential, etc. I do NOT want to do this. Estimate the density with a kernel density estimate (in R, e.g., with the function density()). Can I use this density estimate as new prior for the next...

Hallo everybody, I have a sequential analysis, where I want to use the kernel density estimate of the posterior for the next update and so on. The first prior is a mixture prior and thereafter the kernel density estimate is used as prior. This is possible with SAS PROC MCMC using FCMP functions. However, I experienced that JAGS is the better sampler where PROC MCMC fails in the simple non-informative case. The mixing of the chains is very good and the results are similar to the frequentist random...