>>>>> Peter Dalgaard
>>>>> on Wed, 20 May 2026 15:16:21 +0200 writes:
>> On 19 May 2026, at 15.25, Kurt Hornik
>> <[email protected]> wrote:
>>
>>>>>>> Martin Maechler writes:
>>
>>>>>>> Stephanie Evert on Sun, 17 May 2026 09:03:52 +0200
>>>>>>> writes:
>>
>>>>> I will revise my feature request to "have an
>>>>> easily-discoverable fast p-value-only calculation" for
>>>>> those who aren't using the odds ratio or confidence
>>>>> interval.
>>
>>>> fisher.pval() in the "corpora" package.
>>
>>>> Best, Steph
>>
>>> Thank you Steph;
>>
>>> However, I still agree with Chris that it would be nice
>>> to have an easy way to basically just get the p-value,
>>> in a case where neither the ESTIMATE nor the confidence
>>> interval is needed.
>>
>>> Indeed, I've not unfrequently seen R code snippets
>>
>>> <some>.test(.........)$p.value
>>
>>> also in CRAN package R code and I think base R should
>>> provide a simple documented way to get only the pvalue
>>> in cases where the full test costs a factor of 1000
>>> more, even if is still only microseconds in case of
>>> small data sets.
>>
>>> My svn commit to the R sources, > 22 years ago (!)
>>
>>> r29551 | maechler | 2004-05-24 21:34 |
>>
>>> did introduce the optional argument 'conf.int = TRUE'
>>> for the 2x2 case _and_ the NEWS entry (of svn rev 29551)
>>> has been
>>
>>>> o fisher.test(*, conf.int=FALSE) allows to skip the
>>>> confidence interval computation
>>
>>> However, Chris is (indirectly) correct -- he did not
>>> mention the 'conf.int' option -- but in this very large
>>> case, conf.int=FALSE does not help much, as indeed, the
>>> unconditionalized
>>
>>> logdc <- dhyper(support, m, n, k, log = TRUE)
>>
>>> is already expensive, and also the computation of the
>>> sample OR (odds ratio)
>>
>>> ESTIMATE <- mle(x)
>>
>>> is costly. Indeed, I will add a new argument to
>>> fisher.test() to allow skipping the expensive parts and
>>> get the p.value and *still* a valid result of class
>>> "htest". For now, this will only be fast when as by
>>> default the null hypothesis is `or = 1`
>>
>> Perhaps we could simply take `or = NULL` or `or = NA` to
>> say "do not even estimate the OR"?
>>
> Hm, no. A simple testonly=TRUE would be
> better. Double-dutying an argument just makes it trickier
> to explain. And you might actually want to get the p-value
> for other values of OR.
Indeed, the default or = 1 specifies H_0, the null
hypothesis; but one can use or = 1.1 or 1.5 , 0.5
In my not-yet-committed but finished version of fisher.test(),
I have used pval.only (or pvalue.only),
as we have several precedences of logical <something>.only
function arguments (and "." there is *not* looking like S3
_and_ is typed faster than say "_")
-M
> -p
>> -k
______________________________________________
[email protected] mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel
