On 15 Jan 2003 14:05:24 -0800, [EMAIL PROTECTED] (dennis roberts) wrote:

> since i followed the model of m and m ... and, you seem to think their 
> wording is ok ... then, by inference ... so is mine

"followed the model"?  Well, M & M  used a known  variance and
did not make a probability statement about outcome, whereas you 
used an estimated variance, and stuck in a percentage -- so that
what you wrote   took the form of a power statement on the future
variance.
 - I hope to get to that library today, to read M & M.

> 
> what is puzzling about your original disagreement with my handout note is 
> the following:
> 
> Where it says,
> "how large of a SAMPLE  would I need,
> with 95% confidence, to produce an interval ...."  -  >>>> your words >>>>> 
> it ought to say,
> "... with 50% confidence, ..."
> 
> you claim that the method i illustrated essentially was working with a 
> fifty (50%) confidence interval (level) ... where did you come up with that?

The illustrated method wasn't  "working with"  a 50% CI.
The flat extrapolation, attempting to match prior results,
implicitly *has*   the 50% power (approx.) to match 
the narrowness of the  previous  95% CI.

Please notice that in your original, excerpted above
 - You are asking, how large a SAMPLE  is needed, 
with 95% confidence, 
to produce  <a Confidence Interval of a stated size (.05 gallons)
(also 95%)> .

I have to read that as a power statement
concerning the achievement of a CI of a certain size.
That is to say (given the earlier presentation), you hope
to achieve the same variance as just seen with N=16.
An alternate statement might have been, How large of
a sample would I need to have 80% confidence of 
achieving a 99% CI that is no more than X gallons
 - the 80%  is not quite the usual wording for it, but it is 
a power statement.

Now, let's turn to the essay:  It is  *all*  about achieving
certain estimates of mu.  (And the randomizations feature 
a fixed, known value of sigma; the consistency of estimates
of sigma, which should be the focus, is never mentioned.)

> 
> let's assume for the moment ... that we are taking samples from a 
> population where the mean is 100 and the sigma is 16 ... like iq scores
> 
> i generated 1000 samples of n=16 ... found the means ... then, using a t 
> value based on about 15 df ... when sample mean +/- 2.1315 units of error 
> (16/sqrt 16 = 4) ... and found that about 96 % of these intervals captured 
> 100 ... ie, about 96% of the intervals USING A MARGIN OF ERROR OF 2.1315 * 
> 4 = 8.5 ... CAPTURED the mu value ... so, the mu value was within + OR - 
> 8.5 about 96% of the time ...
> 
> NOT 50%

This randomization  is not strictly parallel to what you asked,
but this answer still has relevance:  
How many of the samples achieved sigma of 16 or less?

My statement (estimating power as 50%, for  achieving
the same standard deviation in a later sampling) was based 
on the common sense realization that half the outcomes
would be smaller variance than the other half.  Your original 
"power statement"  was estimated from *one*  random realization
of sigma  in a sample of 16.  

Yours  *is*  a statement about confidence in achieving 
a small size for the resulting CI, not about means.


[snip, stuff about estimating mu from fixed variance;
rather than estimating variance from an estimated variance.]

Hope this helps.

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
.
.
=================================================================
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