Date: Thu, 18 Nov 1999 09:45:37 -0500
Reply-To: "Howard L. Kaplan" <h.kaplan@VENTANA-CRC.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: "Howard L. Kaplan" <h.kaplan@VENTANA-CRC.COM>
Organization: Interlog Internet (Toronto)
Subject: Re: MIXED models & underlying assumptions
Content-Type: text/plain; charset=us-ascii
"D.Stephen" wrote, in part:
>
> We have carried out these lake experiments using a fully factorial
> randomised ANOVA block design with repeated measurements (running for 8
> weeks - data collected weekly)... we often find that the assumptions of
> sphericity are broken (for within subject effects).
One alternative to worrying about the sphericity problem is to reduce
your data so that you are not dealing with week-to-week differences
within the analysis; you deal with them outside the analysis. For
example, if your main interest is in the rate change over weeks, then
you might want to replace the eight weekly measurements with the slope
of the best-fitting line. On the other hand, if your main interest is
in the maximum value that the dependent variable can attain, you might
want to analyze the average of the best two or best three weeks (which
should be somewhat more stable, in the sense of less noisy, than the
single largest value). This would let you analyze one or more models
having no repeated measurements, which gets around the sphericity
problem; also, some of these summaries may be rather more powerful than
analyzing the whole vector of eight weeks, since you aren't diluting
your effect among 7 df for weeks. On the other hand, if you replace the
single, large model having eight weeks with a series of separate models,
you may lose some of the protection that global tests give you against
excessive type I errors in multiple comparisons.
--
Howard L. Kaplan
Director, Research Technology and Methods
Ventana Clinical Research Corporation
947 - 76 Grenville Street
Toronto, Ontario, Canada M5S 1B2
(416)323-6400 ext 4915# voice; (416)323-7553 fax
h.kaplan@ventana-crc.com
| 
