Numerical stability of beta_mix_trunc_lpdf

[andrjohns]

This PR updates your package's tests to temporarily skip one of the scale factors when testing. This was causing a test failure due to a bug in Stan 2.26, and will be safe to re-add on the release of rstan 2.31 (which will quickly follow the release of rstan 2.26).

I've also updated your package's NAMESPACE to avoid a NOTE about not importing from rstantools.

If you could submit a new release to CRAN soon with this PR, it would be very much appreciated - we're trying to reduce the number of downstream failures so that we can submit

[kkmann]

will look into it today/tomorrow (and the mac fail x/) - could you give me a link to the stan bug for reference?

[andrjohns]

Unfortunately I haven't been able to easily reproduce it to be able to identify specifically where in Stan the issue is coming from, just that it is present with rstan 2.26 but not 2.31. Additionally, the failure only occurs when the tests are run through rcmdcheck(args="--as-cran"), not when the tests are run manually or through testthat.

If you have any suggestions for the different behaviours under rcmdcheck vs testthat, I could look into more, but given that it's resolved by 2.31 I'm not too concerned

If you want to verify this locally as well, you can compare the results of running rcmdcheck() under 2.26:

install.packages("rstan", repos = c("https://mc-stan.org/r-packages/", getOption("repos")))

Against the results from 2.31:

remotes::install_github("stan-dev/rstan@experimental",subdir="StanHeaders")
remotes::install_github("stan-dev/rstan@experimental",subdir="rstan/rstan")

[andrjohns]

While I still don't know exactly what caused it, I narrowed the issue down to a loss of precision in your beta_mix_trunc_lpdf function. By updating the function to operate entirely on the log-scale, the tests pass under all versions of rstan.

Hopefully that's a better solution for you than just removing tests!

[andrjohns]

@kkmann apologies to double-ping, but we're looking to make the submission soon. Would it be much trouble to merge and release this fix? Thanks!

[kkmann]

oh wow, cool - thanks!

[kkmann]

@andrjohns on its way to CRAN - there was/is a weird bug with ggplot2 on Mac and I do not have a Mac testing environment... fingers crossed.