I'm new to jags and Bayesian inference, but I've run a fairly complex model in jags via rjags. However, I get the warning "adaptation incomplete". As far as I understand it, this means that the samplers couldn't be adapted. Thus, things such as step size aren't optimized.
Even with n.adapt set to 10000, the warning persists. But what does this mean for my model? Can I still trust the results when choosing high n.burn and n.iter?
What options do I have to get rid of the warning?
Welcome to CrossValidated!
You're right that the 'adaptation incomplete' warning means that the samplers still aren't optimised for your model after the adaptation phase. But it doesn't mean that they couldn't be adapted at all - just that JAGS doesn't think that it has adapted them to the optimal state. This tends to happen in complex models, where it is more likely that JAGS will need to fall back on "last resort" samplers that are less efficient but more general than the samplers it would prefer to use.
In practical terms, this means that you may experience higher autocorrelation and therefore slower mixing and reduced effective sample size compared to the same model with fully optimised samplers. To avoid this, you can try running a longer adaptation phase (although 10000 is already quite long). Or you could try manually specifying better initial values in the hope that this will improve things. Or you could just live with the sub-optimal sampler efficiency and just run the model for longer - as long as it has converged and you end up with a sufficiently high effective sample size (i.e. low enough Monte Carlo error) then you should arrive at the same result. In the real world it is sometimes a more pragmatic use of time to live with inefficient samplers rather than spending a lot of time trying to optimise them. The goal of MCMC inference is to get a robust answer: not necessarily to get the same answer in as few iterations as possible!
But (as always with MCMC) it is crucial to make sure that the samples are from the stationary posterior (i.e. the model has converged) and that the effective sample size is high enough. Check the trace plots carefully for all parameters.
