I am trying to fit a model with a (two, actually) relatively large (>1.5 million) dataset with a hierarchical model of the type that A. Gelman often suggests to do MRP.
I have found it impossible to run a HMC sampler on it-I killed the job after seeing little progress for more than 1 day. I have tried to use the
meanfield algorithm proposed in
stan_glmer and it works remarkably well-it’s fast, and it converges, etc.
I wondered, however, if I can trust its estimates -I have seen all kinds of warnings. I have tried to compare its results with less complicated models/smaller samples and it seems to be ok, in general. But for the larger model, I have no benchmark to compare. I can I present these estimates as ok? Are there any diagnostics I can perform? Any workaround you can suggest?