home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!stanford.edu!agate!ucbvax!stat!chadster
- From: chadster@statBerkeley.EDU (Chad Heilig)
- Newsgroups: sci.math.stat
- Subject: Re: Standard Deviation.
- Keywords: (n) versus (n-1)
- Message-ID: <44620@ucbvax.BERKELEY.EDU>
- Date: 17 Aug 92 22:15:41 GMT
- References: <1992Aug14.172833.11844@cbfsb.cb.att.com> <c48nbgtf@csv.warwick.ac.uk> <thompson.714070323@daphne.socsci.umn.edu>
- Sender: nobody@ucbvax.BERKELEY.EDU
- Organization: Statistics Dept., U. C. Berkeley
- Lines: 27
-
- It is true that the `usual estimator' for the standard deviation is biased
- for the population SD. However, it might remain popular for the following
- reason:
-
- Suppose we want to test the hypothesis Null: mean=0 vs. Alt: mean != 0,
- but we do not know the population variance. Then our test takes the form:
-
- reject the null if Xbar/sqrt(S^2) > some constant
-
- where S^2 is the unbiased estimator for the variance and the constant is
- chosen so that we reject with a certain probability.
- So where am I going with this? Well, under the assumption that the
- data are normally distributed with mean 0, the test statistic above (sometimes
- called T squared) has a t-distribution with n-1 degrees of freedom; this is
- only true, however, under the additional contstaint that we use the unbiased
- estimator for the variance in forming the test-statistic. Thus the "N-1".
-
- That said, let me make two more notes:
- (1) S^2 is unbiased for population variance, regardless of the underlying
- dsitribution (as long as we have finite variance).
- (2) The distribution of the T-statistic is somewhat robust against non-
- normal underlying distributions.
-
- I realize that this gets awat a bit from the original question, but I think
- it rounds out the defense for using `N-1' in the SD estimate.
-
- Chad
-