Predict risk given success/trials data

Interfaces brms
1

Hello,

I’m fitting a logistic risk model on hospital data, given Region and Class predictors, with cases and denominators for each observation, with the aim of making simulations.

the model is:
brm(cases | trials(num) ~ (1 | Region / Class), family = binomial(), data = DF, chains = 8, cores = 8, iter = 8000)

I would like to predict the per hospital risk, but predict.brmsfit() produces only the predicted count of events, and does not take a type = 'response' argument as predict.glm() to predict risk. I could use posterior_linpred() and then invlogit() but it doesn’t include the residual error in the prediction, which I need.

What am I missing?

2

You want to use the fitted method.

3

Uhm, fitted() returns to me a non-integer (?) number of successes, not the probability of success:

fitted(model, nsamples = 1, summary = F) %>% as.vector() %>% summary()
    Min.  1st Qu.   Median     Mean  3rd Qu.     Max. 
  0.4842   3.7819   9.2655  15.9890  17.0964 142.8172
4

Because you didnt set trials to 1. There are multiple discourse threads or github issues that explain how you do it.

5

I didn’t know about it. I try searching a lot on google before asking here but nothing came up.

Thanks!

6

Hmm strange. Essentially you do conditions = data.frame(ntrials = 1) if ntrials is the name of you trials variable in your data set.

8

The uncertainty will increase proportionally.

1 Like