Issues using posterior_predict function with weighted models


#1

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

•Operating System: Windows 10
•brms Version: 2.2.0

Hi Paul
I have used the brm function to run hierarchical models and then used the posterior_predict function on the generated models without issues (details on how we have defined the model can be found here: Running time for hierarchical model ). However, I have now calculated the same models with the addition of using weights, using this formula in the brm function: yi | weights(wei) ~ predictors (i.e. the only change in the model formulae is the inclusion of the weights, all other arguments are identical to the unweighted models). When trying to use posterior_predict on the weighted models I am given the following error message: “Something went wrong. Did you transform numeric variables to factors or vice versa within the model formula? If yes, please convert your variables beforehand. Or did you set a predictor variable to NA? If no to both, this might be a bug. Please tell me about it.” I did not transform the variables and there are no NA’s in the data used.
Since the inclusion of weights is the only change in the weighted model I assume that this is what is causing trouble, but I cannot figure out how or why. Do you have any suggestions?
Thanks in advance for your reply.
Tiril


#2

Hi Tiril,

can you provide a minimal reproducible example? A basic weighted model does not create this issue (with the github version of brms):

df <- data.frame(y = rnorm(100), w = 1:10, x = rnorm(100))
fit <- brm(y | weights(w) ~ x, data = df)
posterior_predict(fit)

Edit: Maybe you sould update brms to the latest CRAN version (2.4.0) and see if that solves your problem.


#3

I am about to submit the next version of brms to CRAN, but I want to be sure that this issue is resolved before submitting. Could you please try out brms 2.4.0 from CRAN or 2.5.0 from github and if it still fails provide a reproducible example? That would help me a lot! :-)

Update: I did some tests and I can’t reproduce your error not even with brms 2.2.0. Is the weighting variable also used somewhere else in the model?


#4

Sorry for being slow here (I am working with Tiril). We also could not reproduce the error with a different data set. So it looks that something went wrong with the particular dataset Tiril was working with.


#5

Thanks for the quick response! Good to know it does not seem to be a general problem.