[R] p-value from chisq.test working strangely on 1.8.1

[Jeffrey Chang]

On Dec 9, 2003, at 9:37 AM, Peter Dalgaard wrote:

> Marc Schwartz <MSchwartz at medanalytics.com> writes:
>
>> Confirmed on Fedora Core 1 with R Version 1.8.1 Patched (2003-12-07)
>> compiled using "gcc (GCC) 3.3.2 20031107 (Red Hat Linux 3.3.2-2)".
>>
>>
>>> chisq.test(matrix(c(0, 1, 1, 12555), 2, 2), simulate.p.value=TRUE)
>> ...
>> X-squared = 1e-04, df = NA, p-value = 1
>>
>>> chisq.test(matrix(c(0, 1, 1, 12556), 2, 2), simulate.p.value=TRUE)
>>
>> X-squared = 1e-04, df = NA, p-value = < 2.2e-16
>> ...
>>> chisq.test(matrix(c(0, 1, 1, 12557), 2, 2), simulate.p.value=TRUE)
>> ...
>> X-squared = 1e-04, df = NA, p-value = 1
>
> Ditto on RH8 with Martyn's RPM of 1.8.0 (yeah, I know...) and ditto
> with a reasonably current r-devel (gcc 3.2)
>
> Anyways, it is yet another fudge-factor issue: If you debug to the
> point in chisq.test where it calculates
>
> PVAL <- sum(tmp$results >= STATISTIC)/B
>
> you'll find that
>
> Browse[1]> any(diff(tmp$result))
> [1] FALSE
> Browse[1]> tmp$result[1]
> [1] 7.96432e-05
> Browse[1]> STATISTIC
> [1] 7.96432e-05
> Browse[1]> tmp$result[1] - STATISTIC
> [1] -1.355253e-20
>
> so PVAL becomes zero and yaddayaddayadda....
>
> The obvious fix would seem to be
>
> PVAL <- sum(tmp$results >= (1-1e-10)*STATISTIC)/B

Yes, this is also the behavior that I am seeing.  I do not know about 
numerical programming to comment on the fix, but it would solve the 
problem in my case.

Thanks to everyone who has looked into this!

Jeff