Just an FYI, I ran into the kernel/bounded continuous problem with ordbetareg, so I wrote a hybrid posterior predictive plot that visualized the continuous (open interval) responses and the discrete responses separately.
That discrete bar plot is useful for illustrating the proportions, but useless for predictive checking as we discuss in our paper. It would be good to use calibration plots for the end points. That KDE is also clearly oversmoothing compared to the data histogram shown earlier
Kernel density estimates (KDE) and histogram are commonly used, but may hide interesting structure in the data. We show how probability integral transform (PIT) ECDF diagnostic can detect when KDE or histogram is oversmoothing. Quantile dot plot is often a better choice
Predictive checks compare model predictions and data
We propose an overlay approach for quantile dot plots to be used when comparing predictive distribution with uncertainty to data
KDE is misleading for categorical data. Bar graph is useless for binary and categorical data, and almost always useless for ordinal data (when using ordinal model with C-1 intercepts, where C is the number of categories). We recommend calibration plots for categorical data
Comments
Just an FYI, I ran into the kernel/bounded continuous problem with ordbetareg, so I wrote a hybrid posterior predictive plot that visualized the continuous (open interval) responses and the discrete responses separately.
I.e. see package vignette:
https://saudiwin.github.io/ordbetareg_pack/vignettes/package_introduction.html#posterior-predictive-plot
We propose an overlay approach for quantile dot plots to be used when comparing predictive distribution with uncertainty to data