Rstanarm(optimize = TRUE) diagnostics

[lauren]

@Xinyi_Chen, @andrewgelman and I are working on updating the MI package from a backend of bayesglm to a backend of rstanarm with the “optimize” setting.

Are there any diagnostics that we ought to be looking at?

[bgoodri]

@avehtari added a few to 2.19

New releases of rstanarm and rstantools R packages

Pareto diagnostic and Pareto smoothed importance sampling https://arxiv.org/abs/1507.02646 is used to diagnose and improve inference when using stan_glm with algorithm=‘optimizing’. Demonstration of timings with n=100,000 and p=100, and Gaussian and logistic regression at https://avehtari.github.io/RAOS-Examples/BigData/bigdata.html Using PSIS and importance resampling means also that we can use PSIS-LOO also when algorithm=‘optimizing’ (I recommend to increase the number of draws from the def…

[andrewgelman]

My main recommendation is that we should be using much stronger default priors. Perhaps something like Cauchy(0, 0.5) for all coefs except for the constant term?