Peter -
Your message is cryptic. I've just re-read help("RNGkind") in
R 1.6.1 (linux Redhat rpm) and it doesn't say anything I can see
about "an unfortuante interaction between the two methods that
are used for generating uniform and normal variables".
Are you referring to a thread in R-help that starts with a message
from Robin Hankin, Tuesday Nov 26 2002 ? On Jonathan Baron's site,
that would be
http://finzi.psych.upenn.edu/R/Rhelp02/archive/9058.html
but even that thread is very vague. (I don't understand a reference
to "PR#1664" in Thomas Lumleys's email (archive/9064.html).)
Maybe it's high time to put a paragraph describing this issue into
the help files for either rnorm() or RNGkind().
- tom blackwell - u michigan medical school - ann arbor -
On 28 Jan 2003, Peter Dalgaard BSA wrote:
> "Charles Annis, P.E." <[EMAIL PROTECTED]> writes:
>
> > Dear R-Aficionados:
> >
> > I realize that no random number generator is perfect, so what I report
> > below may be a result of that simple fact. However, if I have made an
> > error in my thinking I would greatly appreciate being corrected.
> >
> > I wish to illustrate the behavior of small samples (n=10) and so
> > generate 100,000 of them.
> >
> > n.samples <- 1000000
> > sample.size = 10
> > p <- 0.0001
> > z.normal <- qnorm(p)
> > # generate n.samples of sample.size each from a normal(mean=0, sd=1)
> > density
> > #
> > small.sample <- matrix(rnorm(n=sample.size*n.samples, mean=0, sd=1),
> > nrow=n.samples, ncol=sample.size)
> > # Verify that from the entire small.sample matrix, p sampled values are
> > below, p above.
> > #
> > observed.fraction.below <- sum(small.sample <
> > z.normal)/length(small.sample)
> > observed.fraction.above <- sum(small.sample >
> > -z.normal)/length(small.sample)
> >
> > > observed.fraction.below
> > [1] 6.3e-05
> > > observed.fraction.above
> > [1] 0.000142
> > >
> >
> > I've checked the behavior of the entire sample's mean and median and
> > they seem fine. The total fraction in both tails is 0.0002, as it
> > should be. However in every instance about 1/3 are in the lower tail,
> > 2/3 in the upper. I also observe the same 1/3:2/3 ratio for one million
> > samples of ten.
> >
> > Is this simply because random number generators aren't perfect? Or have
> > I stepped in something?
> >
> > Thank you for your kind counsel.
>
> You stepped in something, I think, but I probably shouldn't elaborate
> on the metaphor ... There's an unfortunate interaction between the two
> methods that are used for generating uniform and normal variables (the
> latter uses the former). This has been reported a couple of times
> before and typically gives anomalous tail behaviour. Changing one of
> the generators (see help(RNGkind)) usually helps.
>
> --
> O__ ---- Peter Dalgaard Blegdamsvej 3
> c/ /'_ --- Dept. of Biostatistics 2200 Cph. N
> (*) \(*) -- University of Copenhagen Denmark Ph: (+45) 35327918
> ~~~~~~~~~~ - ([EMAIL PROTECTED]) FAX: (+45) 35327907
>
> ______________________________________________
> [EMAIL PROTECTED] mailing list
> http://www.stat.math.ethz.ch/mailman/listinfo/r-help
>
______________________________________________
[EMAIL PROTECTED] mailing list
http://www.stat.math.ethz.ch/mailman/listinfo/r-help
