get_refmodel: latent projection cannot be combined with observation weights

[pintubread]

Hi! Not sure if this is an issue, but just wanted to ask for clarification. I wanted to perform projpred on my multilevel binomial brms model, so following the vignette, I ran get_refmodel(my_model, latent=TRUE) as a first step. I received an error message as follows:

Error in extract_model_data(object, newdata = NULL, extract_y = TRUE) :
Currently, the latent projection may not be combined with observation weights (other than 1).

As far as I'm aware, my model has no observation weights. May I know what may be causing the issue and how I can remedy it? Thanks!

[fweber144]

In a binomial model, the observation weights are given by the numbers of trials. In your case, I guess you have a number of trials greater than 1 for at least one observation? If yes, projpred should actually mention in the error message that you can de-aggregate the dataset (so that each new observation has a number of trials of 1) to circumvent this. However, since you mentioned that you have a brms model, we might have to adapt an error message in brms. Do you have a reprex?

[pintubread]

Hi! Many apologies for the late reply.

Yes, the number of trials in my dataset is greater than 1. Please find the reprex below`

library(projpred)
library(brms)
set.seed(1410)

data <- data.frame(n_subjects = sample(11:1000,size = 50,replace = T),
                   dummy_var_1 = rnorm(50),
                   dummy_var_2 = rnorm(50),
                   dummy_var_3 = rnorm(50),
                   dummy_var_4 = rnorm(50),
                   age_group =  sample(1:6,size = 50, replace=T))
data$n_positives <- sapply(data$n_subjects,function(x){sample(11:x,1)})

brms_model <- brm(formula = n_positives|trials(n_subjects) ~ 1 + (1|age_group) + (0+dummy_var_1|age_group) + dummy_var_2 +
                    dummy_var_3 + dummy_var_4,
                  data = data,
                  family=binomial())

get_refmodel(brms_model, latent=TRUE)

Thank you!