[EMAIL PROTECTED] (Dalby, James WLAP:EX) wrote in
news:[EMAIL PROTECTED]: 

> I have been refreshing my knowledge of outliers and influential
> observations in regression analysis and could use some clarification
> on the difference between the two.  I'm aware that some outiers are
> influential while others are not, but I'm wondering whether all
> influential observations are outliers.  Is it possible for an
> influential observation to not be an outlier?  If you know of a graph
> that would answer this, please refer me to it.

The problem is that "outlier" is a rather poorly-defined term.  By the 
definitions I use, it is in fact possible for an influential observation 
not to be an outlier.  By my definition, an outlier is an observation whose 
value was generated by a different process than a supermajority of the 
observations, i.e. it's an indicator of inhomogeneity in the system.  An 
influential observation, OTOH, is simply an observation that, if excluded 
from the modelling process, would result in a substantially different 
model.  There's no logical reason that a completely homogeneous process 
(which by definition would not produce outliers) could not produce 
influential observations.

The presence of influential observations suggests that any model that takes 
them into account may be biased from the "true" model for the underlying 
population.  The presence of outliers (assuming that the different process 
isn't simply observation error), OTOH, suggests that there may not *be* One 
True Model for the underlying population and that a model that takes them 
into account may be an average of apples and oranges (take a look at the 
classic Hertzsprung-Russell illustration, where inhomogeneous observations 
result in a model that's physically impossible).
.
.
=================================================================
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