Skip to content

[R-pkg-devel] Errors in r-devel (Linux) and r-patched (Solaris)

3 messages · Sebastian Meyer, Helmut Schütz

Original
Helmut Schütz
# 
Dear all,

I was notified about errors:
https://cran.r-project.org/web/checks/check_results_replicateBE.html

Since I don't have access to those operating systems, I'm a little bit lost.
Here is what I get with 64bit R on Windows:
library(replicateBE)
x <- method.B(details = TRUE, print = FALSE,
               data = rds30, option = 1)[c(10, 19)]
y <- c(17.86418, 92.73371)
d <- as.numeric(signif(abs(x - y), 7))

With R 3.6.2
print(d)
[1] 4.431372e-06 1.182994e-07

With R 3.6.3
print(d)
[1] 1.078696e-05 1.182989e-07

x[10] are Satterthwaite's degrees of freedom obtained by package 
pbkrtest. In both R-versions I use its current version (0.4-8.6 of 
2020-02-20).

Any ideas / suggestions?
As a workaround I could reduce the tolerance of testthat's 
expect_quivalent() from to currently 5e-7 to a higher (_which one?_) 
value. But I still want to know what might be going on the CRAN's 
linux/solaris devel/patched installations since the other R-versions on 
all operating systems passed the tests.

Cheers,
Helmut
Sebastian Meyer
# 
Do both of your R installations use the same version of lme4 ?

A new lme4 version has been published on CRAN yesterday and some changes
regarding default numerical tolerances in optimizations could explain
the difference in your results. See the NEWS here:
https://cran.r-project.org/web/packages/lme4/news.html

Best regards,

	Sebastian Meyer


Am 08.04.20 um 13:43 schrieb Helmut Sch?tz:
Helmut Schütz
# 
Dear Sebastian,

Sebastian Meyer wrote on 2020-04-08 14:15:
THX, good point! On R3.6.2 I have lme4 1.1-21 (2019-03-05) and on R3.6.3 
indeed yesterday's lme4 1.1-23.
I see. Since the other 93 (!) tests passed, I will submit a fix only for 
_this_ data set. A little bit nasty since we recently published a paper 
about software validation: https://dx.doi.org/10.1208/s12248-020-0427-6
The supplementary material contains code for IQ (installation 
qualification). If a user on r-devel (Linux) or r-patched (Solaris) runs 
it ? with the current lme4 - he/she will be slapped in the face and the 
suppl. material will tell him/her:
"If a test fails [?] the package did not pass IQ in the user?s 
computational environment (hardware, operating system, R-release) and 
_must not_ be used."
Splendid.

Cheers,
Helmut