In the them of “Opportunity for Comment!” some would argue (e.g. Dan Simpson) that you really can’t separate the prior and the data generating model (likelihood) but need to keep the full joint model intact. This would suggest the posterior is not _sufficient_ for Bayesian visualization.

I have argued that both the prior and posterior should always be visualized if not also the likelihood (posterior/prior). I also went further and suggested the individual unit of analysis likelihoods that multiply up to the likelihood should also be visualized http://statmodeling.stat.columbia.edu/wp-content/uploads/2011/05/plot13.pdf

The paper tends to receive two responses, nice ideal but almost trivial and makes absolutely no sense at all.

]]>https://ctg2pi.wordpress.com/2015/02/24/principles-of-posterior-visualization/

Good someone else is writing proper papers on this topic. ]]>