Re: [Rd] binomial()$linkinv no longer accepts integer values
On 1/8/25 17:26, Ben Bolker wrote: Thanks, that makes sense. I guess if it never worked for integers (or hasn't worked in a long time, at least) then it doesn't need to be fixed/changed ... Still you found out that the type check and use of REAL() in the function is in wrong order. Instead of "REAL() can only be applied to a 'numeric', not a 'integer' " one should get "Argument eta must be a nonempty numeric vector" Fixed in R-devel, Best, Tomas cheers Ben On 2025-01-08 11:20 a.m., Ivan Krylov wrote: On Wed, 8 Jan 2025 10:57:47 -0500 Ben Bolker wrote: I haven't done the archaeology to figure out when this broke/exactly what change in the R code base broke it: it happened within the last month or so binomial() itself exhibits this property even in R-4.2.2 from more than two years ago: R -q -s -e 'getRversion(); binomial()$linkinv(1L)' # [1] ‘4.2.2’ # Error in binomial()$linkinv(1L) : # REAL() can only be applied to a 'numeric', not a 'integer' It's the `etas` [1] that suddenly became integer due to a change in seq.int(): R -q -s -e 'str(seq.int(-8, 8, by=1))' # num [1:17] -8 -7 -6 -5 -4 -3 -2 -1 0 1 ... R-devel -q -s -e 'str(seq.int(-8, 8, by=1))' # int [1:17] -8 -7 -6 -5 -4 -3 -2 -1 0 1 ... __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
[Rd] binomial()$linkinv no longer accepts integer values
As of r87537, binomial()$linkinv no longer accepts integer arguments.
binomial()$linkinv(1.0) ## 0.7310586
binomial()$linkinv(1L)
Error in binomial()$linkinv(1L) :
REAL() can only be applied to a 'numeric', not a 'integer'
Since R is usually so permissive/sloppy with the distinction between
integers and numeric/doubles, I'd argue that this is a bug ...
(This came up because it happens to break lme4's tests.)
I haven't done the archaeology to figure out when this broke/exactly
what change in the R code base broke it: it happened within the last
month or so and it's certainly something upstream of the linkinv
function itself, as (according to 'git blame') nothing in that function
has changed in the last 4 years ... (if it were of interest I could 'git
bisect' to figure it out but maybe someone will save me the trouble ...)
Thoughts, insights? Should I post this to r-bugzilla? Or suck it up
and accept it as the new reality?
cheers
Ben Bolker
binomial()$linkinv is:
function (eta)
.Call(C_logit_linkinv, eta)
Here is the body of the function in C code:
https://github.com/r-devel/r-svn/blob/195ab0870a8131d01492f8f1d3a2ad514bc7e040/src/library/stats/src/family.c#L72C1-L89C2
SEXP logit_linkinv(SEXP eta)
{
SEXP ans = PROTECT(shallow_duplicate(eta));
int i, n = LENGTH(eta);
double *rans = REAL(ans), *reta = REAL(eta);
if (!n || !isReal(eta))
error(_("Argument %s must be a nonempty numeric vector"), "eta");
for (i = 0; i < n; i++) {
double etai = reta[i], tmp;
tmp = (etai < MTHRESH) ? DBL_EPSILON :
((etai > THRESH) ? INVEPS : exp(etai));
rans[i] = x_d_opx(tmp);
}
UNPROTECT(1);
return ans;
}
__
R-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] binomial()$linkinv no longer accepts integer values
On Wed, 8 Jan 2025 10:57:47 -0500 Ben Bolker wrote: > I haven't done the archaeology to figure out when this broke/exactly > what change in the R code base broke it: it happened within the last > month or so binomial() itself exhibits this property even in R-4.2.2 from more than two years ago: R -q -s -e 'getRversion(); binomial()$linkinv(1L)' # [1] ‘4.2.2’ # Error in binomial()$linkinv(1L) : # REAL() can only be applied to a 'numeric', not a 'integer' It's the `etas` [1] that suddenly became integer due to a change in seq.int(): R -q -s -e 'str(seq.int(-8, 8, by=1))' # num [1:17] -8 -7 -6 -5 -4 -3 -2 -1 0 1 ... R-devel -q -s -e 'str(seq.int(-8, 8, by=1))' # int [1:17] -8 -7 -6 -5 -4 -3 -2 -1 0 1 ... -- Best regards, Ivan [1] https://github.com/lme4/lme4/blob/54c54a320c23b34fea2f7e613928d1ebe7a3fd37/tests/testthat/test-glmFamily.R#L10C5-L10C25 __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] binomial()$linkinv no longer accepts integer values
Thanks, that makes sense. I guess if it never worked for integers (or hasn't worked in a long time, at least) then it doesn't need to be fixed/changed ... cheers Ben On 2025-01-08 11:20 a.m., Ivan Krylov wrote: On Wed, 8 Jan 2025 10:57:47 -0500 Ben Bolker wrote: I haven't done the archaeology to figure out when this broke/exactly what change in the R code base broke it: it happened within the last month or so binomial() itself exhibits this property even in R-4.2.2 from more than two years ago: R -q -s -e 'getRversion(); binomial()$linkinv(1L)' # [1] ‘4.2.2’ # Error in binomial()$linkinv(1L) : # REAL() can only be applied to a 'numeric', not a 'integer' It's the `etas` [1] that suddenly became integer due to a change in seq.int(): R -q -s -e 'str(seq.int(-8, 8, by=1))' # num [1:17] -8 -7 -6 -5 -4 -3 -2 -1 0 1 ... R-devel -q -s -e 'str(seq.int(-8, 8, by=1))' # int [1:17] -8 -7 -6 -5 -4 -3 -2 -1 0 1 ... -- Dr. Benjamin Bolker Professor, Mathematics & Statistics and Biology, McMaster University > E-mail is sent at my convenience; I don't expect replies outside of working hours. __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] binomial()$linkinv no longer accepts integer values
> On 9 Jan 2025, at 05:56, Tomas Kalibera wrote: > > > On 1/8/25 17:26, Ben Bolker wrote: >> Thanks, that makes sense. >> >> I guess if it never worked for integers (or hasn't worked in a long time, >> at least) then it doesn't need to be fixed/changed ... > > Still you found out that the type check and use of REAL() in the function is > in wrong order. Instead of > > "REAL() can only be applied to a 'numeric', not a 'integer' " > > one should get > > "Argument eta must be a nonempty numeric vector" > But that is in practice a bit ambiguous, because > is.numeric(1L) [1] TRUE and integers are also of the mode numeric, so I’d argue that the previous error was more informative. I would either re-word the error to explicitly exclude integers, or coerce integers to doubles. Technically the latter would be more consistent with R, but if someone explicitly passes an integer to the inverse logit function in a real application then chances are it’s not intentional. Cheers, Simon > Fixed in R-devel, > > Best, > Tomas > > >> >> cheers >>Ben >> >> >> On 2025-01-08 11:20 a.m., Ivan Krylov wrote: >>> On Wed, 8 Jan 2025 10:57:47 -0500 >>> Ben Bolker wrote: >>> I haven't done the archaeology to figure out when this broke/exactly what change in the R code base broke it: it happened within the last month or so >>> >>> binomial() itself exhibits this property even in R-4.2.2 from more than >>> two years ago: >>> >>> R -q -s -e 'getRversion(); binomial()$linkinv(1L)' >>> # [1] ‘4.2.2’ >>> # Error in binomial()$linkinv(1L) : >>> # REAL() can only be applied to a 'numeric', not a 'integer' >>> >>> It's the `etas` [1] that suddenly became integer due to a change in >>> seq.int(): >>> >>> R -q -s -e 'str(seq.int(-8, 8, by=1))' >>> # num [1:17] -8 -7 -6 -5 -4 -3 -2 -1 0 1 ... >>> R-devel -q -s -e 'str(seq.int(-8, 8, by=1))' >>> # int [1:17] -8 -7 -6 -5 -4 -3 -2 -1 0 1 ... >>> >> > > __ > R-devel@r-project.org mailing list > https://stat.ethz.ch/mailman/listinfo/r-devel > __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
