home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!dtix!darwin.sura.net!convex!news.utdallas.edu!corpgate!bnrgate!stl!sbil!wet!spitws100!pbond
- From: pbond@spitws100.sbil.co.uk (Philip Bond)
- Newsgroups: sci.math.stat
- Subject: Re: Testing for Normality
- Message-ID: <1992Sep15.092214.13671@sbil.co.uk>
- Date: 15 Sep 92 09:22:14 GMT
- References: <Bu5suB.HxM@mentor.cc.purdue.edu>
- Sender: news@sbil.co.uk
- Reply-To: pbond@spitws100.sbil.co.uk
- Organization: Salomon Brothers, Ltd.
- Lines: 37
-
- Normality Testing.
-
- If you have a lot of samples, a straightforward and powerful test of normality
- that does not require the use of tables is the method of moments. Compute
- the moments of the distribution ( it is probably sufficient to compute the
- first four moments but you can go higher if you so choose ). The normal
- distribution has the property that :-
-
- if m == mean , v is the variance of the distribution then
-
- define mn = the nth moment ie : E[ (x-m)^n ]
-
- where E is the expectation, ^ means "to the power of "
-
- we have the following handy result :-
-
- mn = 0 if n is odd
-
- mn = (n-1)(n-3)...3.1.v^n for n even.
-
- Furthermore, this property is unique to the normal distribution.
-
- Application :-
-
- compute the mean and variance of the sample distribution, then
- compute 1/n*sum( x-m)^n ) where x is the random variable ( I am assuming
- large n ; you need to adjust this to get an unbiased estimator for small
- n : use 1/Num Degrees of Freedom ).
-
- Then see if the 3rd , 5th etc. moments are small or vanishing, and whether the
- 4th, 6th ... follow the formula above. This is a pretty good test of
- normality. It will differentiate between normal and lognormal samples
- even for quite small sample sizes ( say, 30-50 observations ).
-
- All my own humble opinion of course.
-
- Hope this helps, Phil.
-