Herman Rubin wrote:
> For regression, linearity is the
> important assumption, and homoscedasticity is of some
> importance, but normality only somewhat affects significance
> levels...
Well, yes and no - at least if by "regression" you mean
ordinary least-squares regression. (If you don't, then neither
linearity nor homoscedasticity is automatic either...)
The idea of the mean permeates our culture to the point
where many somewhat-numerate people [I point out in some haste that this
does not refer to anybody on this list, let alone the learned Herman!]
cannot conceive that anything *but* the mean could be "fair" to use in
summarizing a group or comparing groups. Nonetheless, unless you have a
reason to use it, adding a bunch of numbers and dividing by N is a
meaningless ritual.
Normality, in combination with the "maximum-likelihood" principle,
gives a reason to minimize the sum of squared vertical offsets. (This is
linked to the use of the mean by the fact that
the mean is the point around which the sum-of-squares variation is
least.)
-Robert Dawson
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
