Bayesplot support for predictive checks with parameter dependence?


Since I like the bayesplot design I was wondering if there is support for predictive checks which use test statistics involving data and parameters?

I can only see functions for T(y) stuff while I want a T(y,theta) type of thing (page 147 BDA3).

OK, these plots are very simple, but the bayesplot thing is always by default pretty.



Hey Sebastian, there isn’t built in functionality for this yet, but maybe if you pass your own function to ppc_stat you might be able to give it access to the draws. If that doesn’t work, then if you think it would be good to have functionality for this and would have to implement on your own anyway, do you want to contribute it to bayesplot? If not, we can add to the list of things for me to add when I have the time.

I do not think this will work. The thing is, I need to know which draw goes with which yrep and I guess you are only passing the yrep into the function, but not the iteration number as well which you would need.

The plots I am looking for are really simple, so maybe we can leave this out for bayesplot…unless you guys have clever ideas aside from pair-wise scatter plots.

What I have in mind to do for my given problem is to resample a subset of my fitted data (resampling data sets of the same size as my hold-out data set), then calculate T(y_subsampled, theta) and calculate the mean over the draws. I then apply this to a hod-out data set and compare the respective mean calculated on that hold out data set to my reference distribution. At least that is my plan… haven’t seen this somewhere else, but I think it makes sense for my application.