I have a model of the form
x0 ~ normal(-1, 1); for (t in 1:(T-1)) z[t+1] ~ normal(f(z[t], x0), sig); data ~ normal(z, eps);
where HMC shows Rhat values near 1 for
x0 & high Rhat values for
z, but I only care about the
Is it correct to assume that because the Rhat of
z is poor, and that the only link between
z, that the estimate of
x0 is bad despite its Rhat being low?
After rereading the section of BDA on Rhat (and a bunch of forum posts), that alternative interpretation I thought of is it, running the chains longer would improve estimate of
z, but not
x0… but I can’t make heads or tails of that given the dependencies of the variables in this model. Any words of wisdom are welcome.