Re: diehard and ent results quesion

From: Stephen Avis (271828@virgilio.it)
Date: 02/26/03 

From: "Stephen Avis" <271828@virgilio.it>
Date: Wed, 26 Feb 2003 10:35:02 GMT

"Cristiano" <cristiano.pi@nsquipo.it> ha scritto nel messaggio
news:ch07a.236629$0v.6632871@news1.tin.it...
>
> Can you tell me, please, what is the smallest N for which the p-value
> given by the chi-square statistic is good without any correction? I'd
> like at least 4 exact figures.

I wish I could! Unfortunately I do not even have a closed formula which
gives a good approximation to the mean for a given N and df pair!
Observations based on a goodly number of inverse MDI tables
that I have "calculated" (by generating "large numbers" of N-partitions
and calculating the info_freq) suggest that the mean and variance are
both very close (< 0.1%) to those for chi-square at about N=10*df.
Presumably,
from that point on the p-distribution is also "very close" to chi-square

You should bear in mind that chi square is the asymptotic distribution and
so there will always
be a correction to apply - though, of course, this correction will get
smaller as N becomes larger..

Relevant Pages
Quantcast