[Rd] plogis (and other p* functions), vectorized lower.tail
m@tth|@@-gond@n @end|ng |rom gmx@de
Thu Dec 9 20:12:32 CET 2021
Ok, I see. It’s a nontrivial change, since the present setup (DEFMATH3_2 & friends) only allows the real-valued function arguments to be vectorized. If this is to be changed, all these macros would need to be rewritten. Case closed.
Von: Martin Maechler
Gesendet: Donnerstag, 9. Dezember 2021 16:03
An: Matthias Gondan
Cc: r-devel using r-project.org
Betreff: Re: [Rd] plogis (and other p* functions), vectorized lower.tail
>>>>> Matthias Gondan
>>>>> on Wed, 8 Dec 2021 19:37:09 +0100 writes:
> Dear R developers,
> I have seen that plogis silently ignores vector elements of lower.tail,
and also of 'log'.
This is indeed the case for all d*, p*, q* functions.
Yes, this has been on purpose and therefore documented, in the
case of plogis, e.g. in the 'Value' section of ?plogis :
The length of the result is determined by ‘n’ for ‘rlogis’, and is
the maximum of the lengths of the numerical arguments for the
(note: *numerical* arguments: the logical ones are not recycled)
The numerical arguments other than ‘n’ are recycled to the length
of the result. Only the first elements of the logical arguments
(above, we even explicitly mention the logical arguments ..)
Recycling happens for the first argument (x,p,q) of these
functions and for "parameters" of the distribution, but not for
lower.tail, log.p (or 'log').
>> plogis(q=0.5, location=1, lower.tail=TRUE)
>  0.3775407
>> plogis(q=0.5, location=1, lower.tail=FALSE)
>  0.6224593
>> plogis(q=c(0.5, 0.5), location=1, lower.tail=c(TRUE, FALSE))
>  0.3775407 0.3775407
> For those familiar with psychological measurement: A use case of the above function is the so-called Rasch model, where the probability that a person with some specific ability (q) makes a correct (lower.tail=TRUE) or wrong response (lower.tail=FALSE) to an item with a specific difficulty (location). A vectorized version of plogis would enable to determine the likelihood of an entire response vector in a single call. My current workaround is an intermediate call to „Vectorize“.
> I am wondering if the logical argument of lower.tail can be vectorized (?). I see that this may be a substantial change in many places (basically, all p and q functions of probability distributions), but in my understanding, it would not break existing code which assumes lower.tail to be a single element. If that’s not
> possible/feasible, I suggest to issue a warning if a vector of length > 1 is given in lower.tail. I am aware that the documentation clearly states that lower.tail is a single boolean.
aah ok, here you say you know that the current behavior is documented.
> Thank you for your consideration.
As you mention, changing this would be quite a large endeavor.
I had thought about doing that many years ago, not remembering
details, but seeing that in almost all situations you really
only need one of the two tails (for Gaussian- or t- based confidence
intervals you also only need one, for symmetry reason).
Allowing the recycling there would make the intermediate C code
(which does the recycling) larger and probably slightly
slower because of conceptually two more for loops which would in
99.9% only have one case ..
I'd have found that ugly to add. ... ...
... but of course, if you can prove that the code bloat would not be large
and not deteriorate speed in a measurable way and if you'd find
someone to produce a comprehensive and tested patch ...
> With best wishes,
> [[alternative HTML version deleted]]
> R-devel using r-project.org mailing list
[[alternative HTML version deleted]]
More information about the R-devel