Convergence issues with multivariate model in brms

General
,
1

Hello,

I am new to modelling and attempting my first multivariate model in brms. I am modelling three different covarying DVs as a function of two shared predictors, though an important feature of this model is that my first DV becomes a predictor in two subsequent submodels. I set rescor to FALSE to help convergence. Main effect estimates look fine and stable, but residual standard deviations of the three DV do not converge. The three submodels are formulated as follows:

d_t ~ Aff + Ag + (1 | id)

d_t.1 ~ d_t + Aff + Ag + (1 | id)

e_t.1 ~ e_t + Aff + Ag + (1 | id)

My main questions:

  1. Why are the residual standard deviations not converging (rhats > 1.05; posterior distributions look highly bimodal) and could it have to do with the fact that one of my responses becomes a predictor in two subsequent models?
  2. Is there anything I can try to fix this convergence issue that goes beyond adjusting adapt_delta and number of iterations?
  3. Can I trust the fixed effect estimates of the model to be interpretable given the described convergence issues and why?

I believe this is the best representation of my causal system, but might have to give it up and go univariate if the model is generally problematic since I can’t model in Stan.

Any input is highly appreciated and would help me finish my PhD!

Many thanks

2

Hey @sofia-pereira. It looks like this is your first post. Welcome to the Stan forums!

  1. Given you’ve set rescor = FALSE, I’m not sure.
  2. adapt_delta often helps, but I wouldn’t start there in this case. I noticed you said you’re new to modeling, and you’re wrestling with a sophisticated-looking model. When I run into problems like this, I scale my model way back, get the simpler variants working, and see where exactly the problems pop up. Try doing that in your case.
  3. No. Don’t trust the output until the model is running smoothly. Good textbooks like SR2 will help clarify why.
1 Like
3

Hey @Solomon, thank you for the welcome and the advice!

It really helped to re-build up the model from scratch. My random effects seem to have been causing the convergence issues. Thank you!

1 Like