[Rd] model.weights and model.offset: request for adjustment
Martin Maechler
m@ech|er @end|ng |rom @t@t@m@th@ethz@ch
Thu Feb 3 18:09:29 CET 2022
>>>>> Martin Maechler
>>>>> on Thu, 3 Feb 2022 15:21:45 +0100 writes:
>>>>> tim taylor
>>>>> on Thu, 3 Feb 2022 11:30:17 +0000 (GMT) writes:
>>> On 03/02/2022 11:14 Martin Maechler <maechler using stat.math.ethz.ch> wrote:
>>>
>>>
>>> >>>>> Ben Bolker
>>> >>>>> on Tue, 1 Feb 2022 21:21:46 -0500 writes:
>>>
>>> > The model.weights() and model.offset() functions from the 'stats'
>>> > package index possibly-missing elements of a data frame via $, e.g.
>>>
>>> > x$"(offset)"
>>> > x$"(weights)"
>>>
>>> > This returns NULL without comment when x is a data frame:
>>>
>>> > x <- data.frame(a=1)
>>> > x$"(offset)" ## NULL
>>> > x$"(weights)" ## NULL
>>>
>>> > However, when x is a tibble we get a warning as well:
>>>
>>> > x <- tibble::as_tibble(x)
>>> > x$"(offset)"
>>> > ## NULL
>>> > ## Warning message:
>>> > ## Unknown or uninitialised column: `(offset)`.
>>>
>>> > I know it's not R-core's responsibility to manage forward
>>> > compatibility with tibbles, but in this case [[-indexing would seem to
>>> > be better practice in any case.
>>>
>>> Yes, I would agree: we should use [[ instead of $ here
>>> in order to force exact matching just as principle
>>>
>>> Importantly, because also mf[["(weights)"]]
>>> will return NULL without a warning for a model/data frame, and
>>> it seems it does so also for tibbles.
>>>
>>> > Might a patch be accepted ... ?
>>>
>>> That would not be necessary.
>>>
>>> There's one remaining problem however:
>>> `$` access is clearly faster than `[[` for small data frames
>>> (because `$` is a primitive function doing everything in C,
>>> whereas `[[` calls the R level data frame method ).
>>>
>>> Faster in both cases, i.e., when there *is* a column and when there
>>> is none (and NULL is returned), e.g., for the first case
>>>
>>> > system.time(for(i in 1:20000) df[["a"]])
>>> user system elapsed
>>> 0.064 0.000 0.065
>>> > system.time(for(i in 1:20000) df$a)
>>> user system elapsed
>>> 0.009 0.000 0.009
>>>
>>> So that's probably been the reason why `$` has been prefered?
>> Would .subset2(df, "a) be preferable?
R> df <- mtcars
R> system.time(for(i in 1:20000) df[["hp"]])
>> user system elapsed
>> 0.078 0.000 0.078
R> system.time(for(i in 1:20000) df$hp)
>> user system elapsed
>> 0.011 0.000 0.010
R> system.time(for(i in 1:20000) .subset2(df,"hp"))
>> user system elapsed
>> 0.004 0.000 0.004
>> Tim
> Yes, I think that's a very good idea --
> notably, as interestingly it seems to work with tibble's very
> well, too.
Interestingly (or not), changing this also fixes a real (rare!) bug:
When digging for a regression test, I've stumbled over an lm() example,
which when modified to use the not so common "(weight)_2" as
*predictor* variable name it started to use that both as
predictor and also as weight (of some kind) such that the fit
changed.
This problem went away after apply the change,
[replacing `a$b` with `.subset2(a,b)]
Now committed to R-devel, svn rev 81650.
If there are no negative effects, this may also be backported to
R-patched.
Thank you both, once more!
Martin
More information about the R-devel
mailing list