Divergence with COM_poisson model

I am trying to use the brms package model some under dispersed data (for ages 0 and 1, var / mean is roughly 1, but the rest are all under 0.5). I have issues with using the COM_poisson family choice in terms of running into very high numbers of divergent transistions after warmup and low ESS. I have tried running more iterations and more chains, but it just increases the likelihood of crashing. I tried adjusting (at the sacrifice of time) adapt_delta but obviously it’s just not working very well.

I am relatively new to modeling using stan and brms, so I would appreciate any input to improve the the model while using stan/brms. The one I have included in the post I arrived at from using glmmTMB. The best (lowest AIC anyways) model I found was a 0 truncated model without ReaderID random effect.

brm(EstAgeInYrs ~ poly(KnwnAgeInYrs,3) + (1|ReaderID),  
              cores = 4, iter = 1000
              data = DataHI, family = "com_poisson")

Data set

DataHI.csv (39.7 KB)

Please also provide the following information in addition to your question:

  • Operating System: Windows 11
  • brms Version: Newest

Thanks in advance

@erlend_myhre. It is very doubtful that a GLMM or GAMM with CMP likelihood is appropriate for your data.

Reviewing your response variable simply doesn’t resolve that matter. The likelihood is conditional on the predictors in the model and not just the response variable.

Perhaps a GLMM or GAMM with zero-truncated Gaussian likelihood might be appropriate - but are the actual data values really integers like 8 instead of say 8.2?

Anyway, maybe a GAMM with zero-truncated Gaussian likelihood like this might be appropriate to consider for inference …

        backend="cmdstanr",  ## or "rstan"

plot(conditional_effects(fit))  ## covariate functional form seems intrinsically nonlinear to me

Meanwhile, there is little prospect of getting similar model estimates for a GLMM with CMP likelihood using glmmTMB compared to Stan (or the brms interface) as these two platforms use quite different parameterisations of that likelihood. They are not fitting the same parameterisation of a CMP regression model.

A useful CMP reference for you would be Huang A (2017) and applicable to glmmTMB:

1 Like

Thanks a lot for the response. I will look into all of this.