[Rd] [patch] ?confint: "assumes asymptotic normality"

peter dalgaard pdalgd at gmail.com
Thu Jul 20 23:47:07 CEST 2017


> On 20 Jul 2017, at 19:46 , Scott Kostyshak <skostyshak at ufl.edu> wrote:
> 
> On Thu, Jul 20, 2017 at 04:21:04PM +0200, Martin Maechler wrote:
>>>>>>> Scott Kostyshak <skostyshak at ufl.edu>
>>>>>>>    on Thu, 20 Jul 2017 03:28:37 -0400 writes:
>> 
>>>> From ?confint:
>>> "Computes confidence intervals" and "The default method assumes
>>> asymptotic normality"
>> 
>>> For me, a "confidence interval" implies an exact confidence interval in
>>> formal statistics (I concede that when speaking, the term is often used
>>> more loosely). And of course, even if a test statistic is asymptotically
>>> normal (so the assumption is satisfied), the finite distribution might
>>> not be normal and thus an exact confidence interval would not be
>>> computed.
>> 
>>> Attached is a patch that simply changes "asymptotic normality" to
>>> "normality" in confint.Rd. This encourages the user of the function to
>>> think about whether their asymptotically normal statistic is "normal
>>> enough" in a finite sample to get something reliable from confint().
>> 
>>> Alternatively, we could instead change "Computes confidence intervals"
>>> to "Computes asymptotic confidence intervals".
>> 
>>> I hope I'm not being too pedantic here.
>> 
>> well, it's just at the 97.5% border line of "too pedantic"  ...
> 
> :)
> 
>> ;-)
>> 
>> I think you are right with your first proposal to drop
>> "asymptotic" here.  After all, there's the explict 'fac <- qnorm(a)'.
> 
> Note that I received a private email that my message was indeed too
> pedantic and expressed disagreement with the proposal. I'm not sure if
> they intended it to be private so I will respond in private and see if
> they feel like bringing the discussion on the list. Or perhaps this
> minor (and perhaps controversial?) issue is not worth any additional
> time.

At any rate, it is important not to let the pedantry cause the text to become misleading. If you just write "assumes normality", readers may consider the procedure to be simply wrong when the estimator (or worse: the original data) is not normally distributed. And "computes asymptotic c.i." is just wrong, because they are sometimes exact. 

It may be necessary to spell things out more extensively. Something like "the default method assumes normality and that the s.e. is known. For asymptotically normally distributed estimators, it yields an asymptotic confidence interval."

-pd

  

> 
>> One could consider to make  'qnorm' an argument of the
>> default method to allow more general distributional assumptions,
>> but it may be wiser to have useRs write their own
>> confint.<foo>() method, notably for cases where
>> diag(vcov(object)) is an efficiency waste...
> 
> Thanks for your comments,
> 
> Scott
> 
>> Martin
>> 
>> 
>>> Scott
>> 
>> 
>>> -- 
>>> Scott Kostyshak
>>> Assistant Professor of Economics
>>> University of Florida
>>> https://people.clas.ufl.edu/skostyshak/
>> 
>> 
>>> ----------------------------------------------------------------------
>>> Index: src/library/stats/man/confint.Rd
>>> ===================================================================
>>> --- src/library/stats/man/confint.Rd	(revision 72930)
>>> +++ src/library/stats/man/confint.Rd	(working copy)
>>> @@ -31,7 +31,7 @@
>>> }
>>> \details{
>>> \code{confint} is a generic function.  The default method assumes
>>> -  asymptotic normality, and needs suitable \code{\link{coef}} and
>>> +  normality, and needs suitable \code{\link{coef}} and
>>> \code{\link{vcov}} methods to be available.  The default method can be
>>> called directly for comparison with other methods.
>> 
>> 
>>> ----------------------------------------------------------------------
>>> ______________________________________________
>>> R-devel at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-devel
> 
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel

-- 
Peter Dalgaard, Professor,
Center for Statistics, Copenhagen Business School
Solbjerg Plads 3, 2000 Frederiksberg, Denmark
Phone: (+45)38153501
Office: A 4.23
Email: pd.mes at cbs.dk  Priv: PDalgd at gmail.com



More information about the R-devel mailing list