ghurault / huraultmisc Goto Github PK

Some Stan-related function in the package achieve similar outcomes as function in tidybayes.

Maybe mark them as obsolete or mention the difference in the doc?

Extracting draws

extract_parameters_from_draw(fit, param, draw) roughly equivalent to fit %>% tidy_draws() %>% gather_variables() %>% filter(.draw == draw & .variable %in% param), but the tidybayes version is a bit inefficient as all draws and all parameters are extracted then filtered (also the draw IDs are not exactly the same).

spread_draws would probably be more efficient when only a handful of parameters are extracted (but can only specify which draw afterwards, or select the number of draws that are extracted).

summary_statistics

Slightly equivalent to fit %>% tidy_draws() %>% gather_variables %>% mean_qi(), although we don't have information about R_hat or n_eff and indexes are not processed.

spread_draws with mean_qi (or equivalent) would be more useful to summarise the distribution of a handful of parameters.

extract_distribution with type = "hdi" or "eti" for plotting fan charts

Could use stat_lineribbon in tidybayes

In order to test stan functions, we need to manipulate a stanfit object.
To do this rapidly, simple solutions include:

• providing a pre-compiled modelled as data (but not compatible with every machine)
• providing a stanfit object as data

However, these files are too big (around 5MB) triggering a CRAN note.
It would be nice to find a way around this.

A potential solution could be to compile the models during the build as described here.
Or use a package using pre-compiled Stan models, like rstanarm (include the package as Suggests).

cf. redundant code