On 22 Nov 2003 09:46:20 -0800, [EMAIL PROTECTED] (Pat) wrote:

> Hello,
> 
> Say you use a Monte Carlo process to get a confidence interval. 
> Rather than running 100 iters (or 1000, or whatever) and settling for
> a fixed p, the idea is to get an arbitrary p value.

What you describe, below, is *not*  an iteration to get
a confidence interval.  You describe iterating to get a 
*point-estimate*.  

> 
> So you iterate ad infinitum.  As you go, the p goes down, and the
> interval gets wider.  

How unlikely is it to flip 20 heads in a row?  I can get an estimate
as you describe it, by flipping a coin (or having a computer
fake it) about a  million times.

>              Say you have a measured value, which initially
> lies outside the interval.  You iterate the MC until it becomes so
> wide it encompasses your value.  You then have a p of < 1/(n-1) where
> n is the # of iterations it took to make the interval this large.

Where is the *interval*?  Whose interval?  What interval?

I get an estimate that it takes about a million flips to get that
event, "20 in a row";  so I conclude that the p is about 1 in a
million.  Is the time-to-first occurrence a good estimate of the p?
Sometimes it is.  I should be able to compute a better estimate 
if I let the computer keep flipping until it has found dozens or
hundreds of the event.

One per million, using n,  is a "point estimate."   I have to repeat 
that experiment -- of flipping a million coins -- 500 times 
or more in order to generate a confidence interval (with 
any confidence as to the accuracy of its size).

> 
> I have a gut feeling something is wrong here.  It seems a bit too
> "sporty", a procedure that lets you dial up any desired p value is
> suspect.  However, I can't put my finger on why it is wrong.

I hope this clarifies how you were oversimplifying things.

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
"Taxes are the price we pay for civilization." 
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to