On Sat, 07 Dec 2002 14:26:25 +0000, Robert J. MacG. Dawson wrote:

>> In my case, I wanted to know whether my data was normally distributed
>> or not because I wanted to set bounds on what would be considered
>> normal (as in healthy) and
>> what not. If the data was normal, I could use the normal distribution
>> to determine these bounds.
>       OK, that makes some sort of sense - but still, you want to find out if
> the distribution IS, appproximately, normal, not that it ISN'T exactly
> so.
>       In principle, I could wreck any "positive" finding by dropping a
> truckload of perfectly good data off unsolicited at your lab & insisting
> that you use it!

Sure you could, and in fact, I would be glad. If I have lots of good data,
there is no use of using some named distribution instead of the real data.
With all that data, I could just as well determine the bounds based on
that. Only in the case where my dataset is limited, I would be interested
in approximating it by some distribution so that I can set bounds
presumably more accurate than when I base it on the real data.

So, in case of a small dataset and a test that shows it is unlikely to be
a sample from a normal distribution, I would use the data. If the test
doesn't think there is reason to assume it is unlikely, I use the
distribution. If I have lots of data, the outcome matters less and less,
since the values of the bounds determined on the data will come closer to
those of the underlying distribution.

So, I am not interested in the question whether my data is normally
distributed, but I want to use the answer of that question for further
analysis. In case of large sample size, the further analysis would be the
same for either known or unknown distribution and the normality test
wouldn't matter anymore.

If you have any doubts, please let me know! Without some opposition, it's
easy to just stop thinking and assume that what you is indeed the right
thing...

Koen
.
.
=================================================================
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