Date:         Sun, 14 Jan 2007 23:46:14 -0800
Reply-To:     David L Cassell <davidlcassell@MSN.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         David L Cassell <davidlcassell@MSN.COM>
Subject:      Re: normality of residuals: opinions?
In-Reply-To:  <Pine.GSO.4.58.0701121625330.29461@leukothea>
Content-Type: text/plain; format=flowed

kviel@EMORY.EDU wrote back:
>
>On Fri, 12 Jan 2007, Reeza wrote:
>
> > Looks like your residuals do not pass the tests for normality, and
> > assuming your errors are in some order relative to the data collection
> > or observation there also appears to be some evidence of cyclical
> > patterns.  time series data?
>
>Thanks.  The data are a plasma protein level measured once.  There should
>be no period effect.  As for the normality tests, they may be sensitive to
>outliers, among other problems:
>
>data resid ;
>   do _n_ = 1 to 100 ;
>     e = rannor( 0 ) ;
>     output ;
>   end ;
>   e = 10 ;
>   output ;
>run ;
>
>proc univariate data = resid ( obs = 100 ) normal ;
>   var e ;
>run ;
>
>proc univariate data = resid normal ;
>   var e ;
>run ;
>
>
>Kevin Viel
>PhD Candidate
>Department of Epidemiology
>Rollins School of Public Health
>Emory University
>Atlanta, GA 30322

Okay, I see that lots of people have chimed in.  But let me add a
couple of points.

Yes, the standard normality tests are likely to go haywire over
a couple major outliers.  Just do a Q-Q plot and see where the
problems are coming from.

Since the residuals *should* be normal based on 'current
understanding of the model', as you (sort of) said, then you may
want to do some checking of the data.

In particular, I would start with some serious investigation of
the QC data for the test equipment.  Getting a 'spike' and a
glob on one tail may mean that there are a few batches or
sub-batches which did not meet QC standards and should not
have been passed onto the PI without re-analysis.  I have seen
a 'spike' in QC data when some calibration data were inadvertently
left in with the real data.  And that's just one thing that can go
wrong with this sort of lab analysis.

And if there is not adequate QC data to look at, that's a red
flag right there!

HTH,
David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330

_________________________________________________________________
Dave vs. Carl: The Insignificant Championship Series.  Who will win?
http://clk.atdmt.com/MSN/go/msnnkwsp0070000001msn/direct/01/?href=http://davevscarl.spaces.live.com/?icid=T001MSN38C07001