[R-pkg-devel] Problems on R-devel because of changes in stats family function

[Duncan Murdoch]

On 11/04/2023 3:37 a.m., Rossum, Bart-Jan van wrote:
> Hi,
> 
> When submitting a new version of my LMMsolver package to CRAN, it fails on the automated tests, see:
> 
> Windows: https://win-builder.r-project.org/incoming_pretest/LMMsolver_1.0.5_20230407_120945/Windows/00check.log
> 
> I managed to track this down to a change in the output of the family function in the stats package:
> https://search.r-project.org/R/refmans/stats/html/family.html
> 
> More specifically, it is about a new output in the call to family():
> 
> dispersion
> (optional since R version 4.3.0) numeric: value of the dispersion parameter, if fixed, or NA_real_ if free.
> 
> Even though the documentation seems to suggest this output is optional, I don't see a way to suppress it. In all my calls to family I get an extra output dispersion with value NA.
> I need to store the family information in my own outputs for further processing. Also my output are quite complex by themselves so I tend to check my full output objects with outputs previously stored.
> Exactly this is giving me problems now. In R 4.3 all of my outputs will have an extra item as well, causing the tests to fail.
> 
> I am not sure how to solve this. I could suppress the affected tests when an installation of stats with version >= 4.3 is detected, but that doesn't feel good.
> Do you have any other suggestions on how to solve this.

One way to solve it would be to set the dispersion parameter to NULL 
before doing your comparison.  That should work across R versions.  For 
example,

    f <- gaussian()
    f$dispersion <- NULL
    expect_equal_to_reference(f, "family")

Another possibility would be to save two reference values, e.g.

    f <- gaussian()
    expect_equal_to_reference(f, if (getRversion() < 4.3.0) "oldfamily" 
else "newfamily")

You should do some sort of manual comparison when those two reference 
values are first created to make sure there are no other changes.

Duncan Murdoch