I am trying to fit a basic mixed-effects location scale model using brms. The data are from an experience sampling study (~4000 observations from ~300 participants). The outcome variable is continuous, but a bit wonky (right-skewed and zero-inflated).
I’ve tried fitting a null model, specified in brms as:
brm(bf(outcome ~ 1 + (1|ID1|ppno), sigma ~ 1 + (1|ID1|ppno)), data = data, control = list(adapt_delta = 0.95), iter = 4000, cores = 4, init = 0)
If I exclude participants that did not vary on the outcome variable (17 participants, 133 observations) the model converges without problems (all Rhat < 1.01, no divergent transitions). Including just one of these participants, however, reduces bulk ESS to ~4 and tail ESS to < 20 for all parameters. I’m a bit stumped by this sensitivity to individual cases and haven’t been able to pinpoint where specifically the issue arises. I’d appreciate any pointers on that.
Below, I’ve included pairs() plots for (A) no data exclusions, (B) exclusion of all but one invariant case, and (C) exclusion of all invariant cases.