Brms model with com_poisson family: interaction involving a categorical variable

Dear all,
I’m not being able to run some Bayesian models with the brms package and I would be grateful if you could help me. I’m new in Bayesian modeling. The problem is the following:

I have the attached dataset ‘dd.csv’ with the following seven variables:dd.csv (4.5 KB)

  • Subj: Participant ID
  • y: independent variable; count data
  • Group: a categorical variable with 9 levels (18 subjects in each level)
  • x1,…,x4: four dependent variables; count data

For each independent variable xi, I would like to assess the effect of the interaction “xi * Group” on y, for i=1,…4. Since y consists of counts, I am using the Conway-Maxwell-Poisson family. In the frequentist approach, I use the following models (which run ok):
mf <- glmmTMB(y ~ x1*Group + (1|Subj), data =dd, family=compois)

In the Bayesian framework, I would like to perform the following equivalent models
mb1 <- brm(y ~ 0+ x1*Group + (1|Subj), data=dd, family = brmsfamily(“com_poisson”), save_all_pars = TRUE, sample_prior=TRUE);

mb2,mb3,mb4 defined analogously.

However, all these mbi models run very, very slowly and do not converge!

Before these models, I ran similar brms models involving only the variable Group (not xi) and everything worked perfectly, including further analysis conducted with the hypothesis method produced very similar results and significance, compared to the frequentist results.

If I change to brmsfamily(“negbinomial”), the models mbi run o.k.; however, the significances of the hypothesis results do not agree with the frequentist approach, which is disappointing because it did not happen with the previous models where the com_poisson family converged.

Can you please help me running the bmi models with com_poisson family? Do you think that adding priors may help? Any suggestions?

Thank you very much!

  • Operating System: mac OS X El Capitan Version 10.11.6
  • brms Version: 2.8.0

There is a reason the COM-Poisson family is currently not officially supported in brms, which is that the family is still experimental. It is expected the COM-Poisson model will run very slowly for overdispersed data (in the parameterization currently used by brms). So if you have overdispersion, better go for another family (like neg-binomial). I would currently only recommend using the COM-Poisson distribution for underdispersed data.

Also, the significance of a hypothesis test should not guide any model selection, but I am sure you are aware of that.

1 Like

Thank you very much!

I am just worried about Bayesian significance because I would like to confirm the significance of several frequentist results. (I’m not interested in model selection.)

Taking into account that the brmsfamily(“com_poisson”) corroborated perfectly the results of the glmmTMBfamily(“compoisson”) in underdispersion cases, I believe that the same would happen for overdispersion situations, in case of convergence. I am sorry that the brmsfamily(“com_poisson”) is currently “under construction”.

I hope you can do this task in a near future. Once again, thank you very much for your wise and precious help, providing this powerful family (it worked perfectly in other situations)!!!

Dear Prof. Paul Buerkner,

Still on this issue, I would like to know your wise opinion about the following:

Since the models

mbi_1 <- brm(y ~ 0+ xi + (1|Subj), data=dd, family = brmsfamily(“com_poisson”));
mbi_2 <- brm(y ~ 0+ Group + (1|Subj), data=dd, family = brmsfamily(“com_poisson”));

work, but the models

mbi <- brm(y ~ 0+ xi * Group + (1|Subj), data=dd, family = brmsfamily(“com_poisson”));

do not (i=1,2,3,4), I was thinking that maybe the problem is due to the interaction ‘xi*Group’, and not because of overdispersion. (Even the models

mbi00 <- brm(y ~ 0+ xi + Group + (1|Subj), data=dd, family = brmsfamily(“com_poisson”));

have problems.)

Do you think that this hypothesis is plausible? In the affirmative case, do you know how can I solve the problem of the brm model including the interaction?

Kind regards,


I don’t know. The com_poisson family is still in the testing phase but we hope to find the time to run some more experiments with it in the next couple of months.