🚀 go-pugleaf

In article <1992Aug16.211142.27499@mailhost.ocs.mq.edu.au> wskelly@laurel.ocs.mq.edu.au (William Skelly) writes:
>In article <1992Aug14.172833.11844@cbfsb.cb.att.com> rizzo@cbnewsf.cb.att.com (anthony.r.rizzo) writes:
>>Can someone explain why calculating the Standard Deviation (SD),
>>for small samples, with (n-1) in the denominator is better than
>>doing so with (n) in the denominator?  I'm sure that there's
>>a perfectly good reason for doing so.  But we, lowly engineers
>>aren't usually told the reason.

Since the n-1 problem has generated considerable interest, it it
perhaps worth a historical note.  The problem is a special case of the
degrees of freedom in the residual sum of squares in a linear
regression problem (for the special case take the regression matrix to
be a vector of all ones).  The correct expression--$n-p$, where $p$ is
the number of parameters and $n$ is the number of observations--is due
to Gauss, who more than any one person was responsible for the
introduction of the variance as a measure of dispersion.  His
definitive treatment of the subject is contained in a series of three
memoirs, beginning with "Theoria Combinationis Observationum Erroribus
Minimis Obnoxiae, Pars Prior" which appeared in 1821 and can be found
in Volume 4 of his works.  Here is what he says about the problem in
the "Pars Posterior" (1823).

   In Articles~15 and~16 we gave a method for approximating the
   precision of observations\symbolnote{1}{)An inquiry into the same
   problem, which I published in an earlier memoir ({\it Bestimmung der
   Genauigkeit der Beobachtungen.\ Zeitschrift f\"ur Astronomie und
   vervandte Wissenschaften\/} Vol.~I, p.~185), was based on the same
   hypothesis about the probability function of the experimental error
   that I used to construct the method of least squares in the theory of
   the motion of heavenly bodies.}).  But this method presupposes that a
   sufficient number of the errors themselves are known exactly, a
   condition which seldom, if ever, holds in practice.  However, if some
   observeded quantities depend on one or more unknowns according to a
   known law, we may find the most reliable values of the unknowns by
   the method of least squares.  If the values of the observed
   quantities are then computed from them, they will be felt to differ
   very little from the true values, so that the greater the number of
   observations the more surely we may take the differences as the true
   observation errors.

   This procedure is used in actual problems by all calculators who try
   to estimate precision a posteriori.  But it is theoretically unsound;
   and although it is good enough for practical purposes in many cases,
   in others it can fail spectacularly.  Therefore, it is well worth
   while to examine the problem more closely.

Gauss goes on to point out that the mean of the residual sum of
squares is by definition less than or equal to the mean of the sum of
squares of the errors and hence must underestimate the latter, i.e.,
the variance of the error.  He then derives the divisor $n-p$
mentioned above.

Pete Stewart