In article <[EMAIL PROTECTED]>,
Spuzzz <[EMAIL PROTECTED]> wrote:
>OK all you statistical junkies are going to jump on my case now.  From
>what I understand this is one of, if not the most important proofs in
>all of statistics.  And I just don't get it.

>I remember learning it in college, where I must admit, I was more
>interested in getting a passing grade, than truly understanding the
>concept.  Now I would like to change that.

>So I understand the mathematical conclusions that are drawn -- that by
>drawing enough samples means, you will approach a normal distribution
>even if the population distribution is not normal.  It's kind of cool,
>but what I don't understand is, what is the practical implication?  We
>are talking about taking multiple samples and then looking at the
>distribution of the mean of each  sample.

You are correct.  The distribution of the sample mean 
for distributions with finite variances APPROACHES a
normal distribution, but is not a normal distribution
unless the original distribution is normal.

>This would be very useful if it said something like, the means and/or
>variances of a sample are similar to the means/variances of a
>population.  But it doesn't say that!  It only talks about the means
>of multiple samples having a normal distribution.

>Here's one thing that I read:
>"...we can use the normal distribution with virtually any population
>we are likely to encounter, all we have to do is grab a nice big
>sample and take the average of the sample..."

Whoever wrote this does not understand the CLT.

>So what is so great about using a normal distribution?  It must be the
>basis of later statistical techniques.  But statistician's are
>frequently excited about the CLT, as if by itself it has obvious
>practical applications (like say the pythagorean theorem)...

Those who understand probability are not that excited about
the CLT, as they know about the errors.  However, the CLT
and the results about well-behaved transformations does
give results about the limiting distribution of statistical
analyses, and this is often all that can be reasonably used.

For example, in regular problems, the twice the logarithm
of the likelihood ratio, in likelihood ratio tests, is
asymptotically chi-squared with the usual numbers of
degrees of freedom.  Tests for regression coefficients are
asymptotically correct.  Tests for correlations for values
other than 0 are not.

Least square and similar procedures are good despite lack of
normality, but they are not maximum likelihood, provided that
one does not disturb the underlying linear structure.  Any
attempt to make the observations more normal is likely to
destroy any useful structure.  Normality is NOT usually 
important, and probably never occurs in nature.

>Please excuse my ignorance...

Ignorance is not a sin.  Ignoring it is.
-- 
This address is for information only.  I do not claim that these views
are those of the Statistics Department or of Purdue University.
Herman Rubin, Department of Statistics, Purdue University
[EMAIL PROTECTED]         Phone: (765)494-6054   FAX: (765)494-0558
.
.
=================================================================
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