Am I doing something wrong or is it a bug in autocorrelation implementation?
The very same model that computes normally in five minutes, takes hours when autocorrelation term is added with a single lag. In fact the original model is a distributed lag one, so it does pretty much the same (single lag of dependent variable is among predictors). Can’t understand why such difference in computation time.
Also, the first time I try fit it (with autoregression term added instead of explicitly specified lag variable), chains did not converge and error summary included suggestion to increase max_treedepth. I have set it to 15, and now it just takes forever, despite I have reduced number of chains to just two.
Specification is as follows:
mod1 = brm( bf(y ~ x1 + x2 + ar(p = 1, cov = TRUE), phi ~ x2), data = data1, family = Beta(link = "logit", link_phi = "log"), prior = prior1, warmup = 2500, iter = 5000, chains = 2, control = list(adapt_delta = 0.9, max_treedepth = 15), seed = 1234 )
Latest R, latest rstan, latest brms (stable release). Mac OS 11.1