Thanks for a very informative answer!
- I’ll (re-)try a non-centered parametrisation.
- Indeed there are many variables in the model, and quite a few three-way interactions, so yes, the model is very complicated. But I think I need that level of complexity to answer the research questions. Basically, it concerns how wealth and culture (here “religion”) affects the probability of educational deprivation for boys and for girls (and thus the gender gap in education), and wealth and culture each have separate measurements for the household level, the community level (which has a random intercept) and the national level (which has both random intercept and a random slope for the effect of gender). I’ll check up your links. The sample is from the general population taken over a period of 30 years in 80 countries.
- I started by running the model in lme4, where the opportunities for mistakes is fewer, I’ll double check that the model has the same parameters and that the estimates are in the same ballpark. And simulating from the priors is probably a very good idea! (I am new to pymc so I haven’t tried that yet)