[R] Non-normal residuals.

Karl Ove Hufthammer karl at huftis.org
Thu Oct 29 09:22:08 CET 2009


On Wed, 28 Oct 2009 14:48:34 -0700 Bert Gunter <gunter.berton at gene.com> 
wrote:
> If generalities -- with the attendant risk of occasional specific caveats
> and violations -- can be tolerated, then George Box's (paraphrased) comments
> of circa 40-50 years ago seem apropos: why do statisticians obsess over
> normality, to which most analyses -- i.e. inference (especially from
> balanced designs)-- are robust, when lack of independence of the
> observations is the violation of assumptions that can reek the greatest
> havoc on the statistical analysis?

While I very much agree with the sentiment, ordinary linear regression 
is *not* very robust when dealing with very heavy-tailed errors, such as 
the t-distribution with low degrees of freedom. Of course, one could use 
'rlm' instead of 'lm' (and I often prefer to do so), but with 
proper/better modelling of the error distribution, you will better 
confidence and prediction intervals.

-- 
Karl Ove Hufthammer




More information about the R-help mailing list