I’m fitting a hurdle lognormal model with an interaction like this:
fit1_ann <- brm(bf(biomass ~ depth_cat + sst_mean_ann + sst_anom_ann + sst_mean_ann:sst_anom_ann, hu ~ depth_cat + sst_mean_ann + sst_anom_ann + sst_mean_ann:sst_anom_ann), data = dat, family = hurdle_lognormal())
So both the lognormal and the binomial components have the same covariates.
Odd thing is the marginals for the terms that are interacting (sst mean and sst_anom) appear to be mirror images of each other.
Even stranger is that there is no correlation in the chains when you plot the samples for the terms that look to be reflections.
See below figure. For instance, the b_sst_anom_march (lognormal coefficient) term is pretty much a reflection of the b_hu_sst_anom_march term (coefficient for the logit hurdle). Same goes for the other terms related to sst.
- Operating System: Windows 10
- brms Version:2.11.1
Is this behaviour expected of a lognormal hurdle? I’ve never seen this before in other applications of hurdle models.