[Rd] apparently incorrect p-values from 2-sided Kolmogorov-Smirnov test (PR#14145)

[David John Allwright]

Dear Thomas, Thank you, yes, that sounds good, and I take the point about 
integer overflow.
Various questions:
(a) Is there some way I can try out the routine with this modification? (I 
am on a Linux system where I am just a user - I cannot install new 
versions of software myself) ?
(b) Is there a reference you can give me to a published paper where the 
method being used to compute the p-values is described?
Many thanks,
David.
------------------------------------------------------------------------------
On Fri, 18 Dec 2009, tlumley at u.washington.edu wrote:

>
>
> I've fixed this by adding 0.5/mn to q.  The problem (at least in principle) 
> with multiplying them all up is integer overflow.
>
> By the time 0.5/mn underflows to zero, missing one value in the distribution 
> won't matter.
>
>     -thomas
>
>
> On Fri, 18 Dec 2009, David John Allwright wrote:
>
>> Dear Thomas, Right, thank you. Yes, I haven't looked at the source code 
>> (because I don't know C) but something like what you mention could well 
>> cause the kind of problems I am seeing: a loop being exectued one too few 
>> or one too many times. And yes, I think those quantities should be 
>> multiplied up by m*n to all become integers so we escape rounding error 
>> problems. David.
>> 
>> ------------------------------------------------------------------------------
>> On Wed, 16 Dec 2009, tlumley at u.washington.edu wrote:
>> 
>>> On Tue, 15 Dec 2009, allwrigh at maths.ox.ac.uk wrote; (in part)
>>> 
>>>> 
>>>> x<-1:5
>>>> y<-c(2.5,4.5)
>>>> ks.test(x,y)
>>>> 
>>>> The value of the D_2,5 statistic is calculated as 0.4 correctly, but the
>>>> p-value is stated by R as 1, though in fact it should be 20/21=0.9524
>>> 
>>> 
>>> What we seem to have here is a rounding error problem.
>>> 
>>> In ks.c:psmirnov2x,  there is a double loop including
>>> 	    if(fabs(i / md - j / nd) > q)
>>> 		u[j] = 0;
>>> 
>>> where md=2, nd=5, and q=3/10.
>>> 
>>> Now,  to full precision  abs(1/2 - 4/5) > 3/10 is false, but at least on 
>>> my MacBook it is true in C double precision.
>>> 
>>> I'm not sure why the loop is working with doubles, since multiplying by 
>>> m*n should make everything an integer.
>>>
>>>     -thomas
>>> 
>>> Thomas Lumley			Assoc. Professor, Biostatistics
>>> tlumley at u.washington.edu	University of Washington, Seattle
>>> 
>>> 
>>> 
>> 
>
> Thomas Lumley			Assoc. Professor, Biostatistics
> tlumley at u.washington.edu	University of Washington, Seattle
>
>
>