by David Winsemius <doe_snot@[EMAIL PROTECTED]
>
Feb 18, 2008 at 12:49 PM
Bob <rdsutterjr@[EMAIL PROTECTED]
> wrote in news:6e8ef07b-4f72-4e32-96df-
bf089333680c@[EMAIL PROTECTED]
> How im****tant is the homoscedasticity assumption in multiple linear
> regression if the purpose of the analysis is to identify which process
> factors (independent variables) have a significant relation****p with
> the dependent variable? No predictions will be made and the residuals
> are normally distributed.
Perhaps no extrapolations are planned, but you are doing "prediction".
> In this situation, would heteroscedasticity
> invalidate interpretation of the coefficient estimates? Thanks!
The heteroscedasticity does not invalidate the estimates of the means,
but does cause problems for inference about how much trust you can put in
their validity. And it definitely causes problems when the goal is model
selection. You generally make such choices on the basis of some sort of
variance reduction, and the heteroscedasticity has a large influence on
those comparisons. Since your hope is to narrow down the list of
candidates, your alternatives are either use suitable transformations or
include model comparison methods that are robust to violations of
homoscedasticity (or if the data permits, use probability models that
allow for alternate relation****ps between mean and variance.)
--
David Winsemius
