Occasionally I get the warning message "Adaptation incomplete. Recreate the
model with a longer adaptive phase." Usually I follow the advice. But out of
curiosity, what triggers this warning? (I know in general terms what it means
for Gibbs samplers to not converge; I'm wondering what specific criteria JAGS
uses for determining this). Thanks!
See section 3.4 of the user manual: "Adaptation and burn-in".
The warning is most likely to come from Metropolis-Hastings samplers: there
are theoretical results on what the optimal value for the acceptance
probability should be. If the acceptance rate is too far away from this value
then the M-H sampler is not operating at optimal efficiency.