test_NUTS test_ADVI discussion page

test_NUTS

test_NUTS will create a file so that:

Models included:

• Can run in NUTS
• Have Rhat under 1.1
• Chooses model with the better parametrization

Info included:

• Rhat
• n_eff
• posterior means
• posterior sds
• inits (in case we want to cheat for test_ADVI)

###Progress stan_demo stops at model 48

test_ADVI

test_ADVI uses info from file created from test_NUTS to:

• Run ADVI for meanfield and fullrank
• Return a matrix with the NUTS_info and ADVI means and ads
• Report on the abs max z-scores by:

• Creating a histogram for each algorithm
• Reporting on abs max z-scores above 1, below 1, and no z-scores

###Progress Code is done and waiting for test_NUTS

##Models that didn't converge: ARM/Ch.12/radon_group.stan - there exists a different model with a better parametrization (NOT INCLUDED)

ARM/Ch.13/earnings_latin_square.stan - tried with tree depth 12, accept delta 0.99; still had issues and took long time to run (NOT INCLUDED)

ARM/Ch.13/earnings_vary_si_chr.stan - Tree depth 20, accept delta 0.999 55 minutes; still had issues and took long time to run (NOT INCLUDED)

ARM/Ch.13/radon_inter_vary.stan - Need non-centered parameterization, but no priors on sigmas, low effective sample size so probably not good candidate (NOT INCLUDED)

ARM/Ch.13/radon_vary_si_chr.stan - mu_a had Rhat above 1.1, but is not used anywhere in code; removed mu_a (INCLUDED)

ARM/Ch.17/17.1_radon_correlation.stan - no prior on sigma_a, sigma

ARM/Ch.17/17.1_radon_vary_inter_slope.stan - Need non-centered parameterization