On Wed, 02 Apr 2003 22:28:59 GMT, "Arthur J. Kendall"
<[EMAIL PROTECTED]> wrote:

> The following SPSS syntax show hows to generate a sample with any  mean 
>   and sd  and then rescale to a given mean and sd.
> I just tried 3 kind of random variable functions from a couple dozen or 
> so that are available in SPSS
> 
> I can't think offhand how to adjust for skewness and kurtosis.
> Why do you want to do this?

It is my opinion that the original request was mis-aimed,
and we ought to show what *should* have been asked.
This is from the note that started this thread:

 " Is it possible to generate a sample that has a specific mean, SD 
 (also skewness, kurtosis)? To be more specific, I do not want to 
 sample from a given distribution like the normal distribution) that 
 has a specific mean and SD. If sampled from a normal distribution  
 with a given mean and SD, of course the mean of the sample is not the
 exact same as the mean of the normal distribution."

This is mainly a reply to that person - even though I have 
dropped the attribution, sorry.

Okay.  When we are doing 10,000 randomizations to validate, 
we *never* would  ask for the exact mean and SD.  That is 
not wrong in an obvious way; but it is not the way it is done.  
Right?  The *real-world*  does not give us exact means....

The only time to feed in one  "exact"  randomized set,  
would be for some sort of, say, "calibration".


Next:  The question about fixing the skewness and kurtosis 
reminds me that someone (probably) is trying to show 
how some procedure works, under various conditions.
All right:  Published claims of robustness do *not*  show
how something works with X amount of skewness as
the primary thing.  Instead, they show the results for
two or three *kinds*  of simple distributions, or for 
contaminated distributions.  

For example:  Here is how test T  works for Normal,
with large N and small N;  
how it works for uniform; 
how it works with a dichotomy;
how it works for data that are exponential;  
how it works with a Normal (0,1) that is contaminated 
with 10% of its cases coming from a second 
Normal where the variance is 10.  
(John Tukey liked to model "contamination" like 
this, claiming is was a realistic hazard for real data.)

Some Monte Carlo data once showed me that
skewness that was lognormal  (from taking exp(x), 
for normal(x)) damaged my subsequent testing  *more*  
than the same measured skewness when it was 
from squaring (taking (x+C)squared).  But I didn't 
think about that kurtosis at all.

Anyway, if you want to do randomizations
that are publishable -- or comparable to other
people's -- you want to come closer to matching
what is in the literature.


-- 
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