I am working with a relatively small data set where the presence of a response is binary coded (1 = present). The design was nested such that there were unique participants within unique groups within 2 conditions.
My formula is:
CR_fmla <- bf(CR_Present ~ Training_Lev + Category + Category:Training_Lev + (1 | Training_Lev / Group / Grp_Spkr))
I have set family to “bernoulli”
I am trying to create some visualizations of the fitted model by showing the predicted probabilities of the response being present (1).
add_fitted_draws the correct function for this? Is the returned
.value the predicted probability?
A box plot of the data:
“CR” is the response I’m interested in.
When I plot the fitted model with 'add_fitted_draws` as follows:
Data %>% select(-Speaker) %>% distinct() %>% add_fitted_draws( Mod, allow_new_levels = TRUE, value = "Probability" )
I get the following figure:
In this figure it looks as though CR being present is more likely in the postgrads in the IF category.
pp_check seems okay:
To try a different approach, I calculated the proportions of CR present responses for individuals and groups and then fitted a
zero_one_inflated_beta model and plotted the fitted proportions, again using
add_fitted_draws, which resulted in the following figure:
The pp_check for the beta model does not look great:
Are these the right pp_checks to be examining?
Is there something else I should be doing to manually calculate posterior probabilities?
Due to collapsing the data to conduct the beta regression on proportions, I’m not able to subject the two models to a
loo_compare since the number of observations is different.