search

mjskay / tidybayes

Bayesian analysis + tidy data + geoms (R package)
http://mjskay.github.io/tidybayes
GNU General Public License v3.0
726 stars 59 forks source link

Mangled variable names from rstan::As.mcmc.list interferes with tidybayes #248

Open psmits opened 4 years ago

psmits commented 4 years ago

I've already posted this on the stan discourse, but I figured this repo would be a better place to take this.

I'm working with the posterior samples from a model fit using RStan. I started my normal routine of using tidybayes to extract posterior samples from the stanfit object and start plotting. However, I ran into an issue with the variable names that come out of tidy_draws() where instead of the expected bracket notation e.g. x[1], I get the results in an unexpected dot notation form e.g. x.1. This mangling of the names does happen when i use dimnames() on the stanfit object.

I hunted for a long time to find the source of the problem, reading the github repo for tidybayes and rstan and I think I found the source of the problem. rstan::As.mcmc.list appears to be the source of the unexpected renaming.

You can compare the outputs from the following

library(rstan)
library(tidybayes)

csvfiles <- dir(system.file('misc', package = 'rstan'),
                pattern = 'rstan_doc_ex_[0-9].csv', full.names = TRUE)
fit <- read_stan_csv(csvfiles)
dimnames(fit)
As.mcmc.list(fit)
tidy_draws(fit)

I'm not sure how to resolve this issue, because it appears to break a lot functionality with tidybayes. For example, this now longer works because the intermediate mcmc.list created by rstan::As.mcmc.list doesn't have the expected variable names.

gather_draws(fit, z[i, j])
mjskay commented 4 years ago

Hmm this looks related to https://github.com/mjskay/tidybayes/issues/132? You might find the workaround there helpful in the short term.

Long term I plan to look at this bug in detail and fix it, though given my current constraints this may not happen until this summer.

psmits commented 4 years ago

thanks for pointing out that work around. I was able to use the .repair_names function jburos wrote as a workaround.

Ben Goodrich replied to my post on the Stan forums about how there might be changes to rstan::As.mcmc.list down the line to address this incompatibility. I replied with a link to #132.