On 17 Dec 2003 15:54:09 -0800, [EMAIL PROTECTED] (Doug) wrote:

> Hi,
> 
> I was sitting through a presentation of some research yesterday about
> the results of a pilot study where a control group was compared with a
> treatment group and I found that I was rather confused about the use
> of the term standard error.  The researcher appeared to be saying
> something about having too small a standard error was not a good
> thing.
> 
> Could someone please explain or link me to a page where I could get
> some detailed information about use of the standard error.  I think
> that it is a measure of the standard deviation of the sample means,

 - you are correct, that is *one* of the uses -

> but this doesn't appear to be the context that was used (from my
> understanding).

 - and it could be the standard deviation of some other statistic.

> 
> When is it good to have a large standard error?  When is it good to
> have a small standard error?  Do these answers differ for different
> tests and different experimental designs?

The replies posted so far seemed fine.  I hope that my 
re-stating of the message may add something.

It is *great*  to achieve a small standard error, as a result
of good experimental design.  It is *bad*  to have a 
statistical analysis produce a standard error that you 
know is too small to be realistic.   That makes you ask,
"What went wrong?"  -  It is apt to indicate overfitting, 
when you see it in a pilot sample that is too small for
the analysis that is intended for a larger sample.

When you have a *huge*  sample - is the other case
where the SE becomes "too small".    That is, it provides
a thumbnail-test that might tell you that *something*  could
be happening;  but there are two problems that could lead
to the conclusion that the term is 'too small.' 
 - One, it could falsely imply a difference, if this is not a
randomized trial, and the assumption of independent
sampling  has (merely) come out to be untrue.  (In other
words, you ought to (a) control for confounding  variables,
and  perhaps (b) use a bigger error term.)
 - Two, it could show that the nominal p-level that you are
using is inappropriate for the size of effects that you were
interested in.  If that's the case, you should adjust the p-level.

-- 
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