Beta-Binomial: Why not a default family in the package?

Thank you for the brms package. I’ve the custom family vignette using beta-binomial as an example. I’m curious why beta-binomial never made it into BRMS as one of the distributions?

I’m modeling

log( number of  people whose browser crashed +1 ) ~ offset(log(total # of people)) + other_vars

using a gaussian. I model both the mean and sigma of the guassian( I guess i could have also used a negbinomial family too)

To get fitted values i do something like

exp( predict(model) - log(total # of people))

to get a rate.

Sometimes this ratio is greater than 1. I would imagine the beta-binomial would fix this domain issue of the predictor and the overdisperson too, but i’m unwilling (my choice!) to implement the custom family so i live with my capping fitted to 1 approach.

An alternative is to model the logit of the rate i.e. logit( (#of people who crash+1)/(# of people)

Nevertheless, thanks much for the awesomeness that is brms.
Thanks in advance

Why not use a binomial model with a random effect over observations to account for potential overdispersion. This is conceptually very similar to the beta-binomial distribution. So if y is your response and obs is an indicator of observations (one value per observation), go for

formula = y | trials(<trial variable>) ~ <predictors> + (1|obs)

Since all the beta-binomial functions are laid out in the vignette, I have a hard time understanding you are unwilling to use it, but this is up to you of course.

1 Like

Thanks much. My reasons for not using it are not the best, and I will use it later but can’t right now . Alas, time is one of the reasons.

Much regards and thanks for accommodating