Date: Tue, 27 Jan 2004 14:02:51 -0500
Reply-To: Peter Flom <flom@NDRI.ORG>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Peter Flom <flom@NDRI.ORG>
Subject: Re: build a model with only the significant factorsandinteractions
Content-Type: text/plain; charset=US-ASCII
Well that's why I said it's ALMOST never a good idea.....:-)
However, wouldn't the situation below be better handled using a
hierarchical linear model with PROC MIXED ? Seems like either a two
level model (cows nested in herds) or possibly a 3 level model, if it's
the same cows being measured at different ages.
Peter
Peter L. Flom, PhD
Assistant Director, Statistics and Data Analysis Core
Center for Drug Use and HIV Research
National Development and Research Institutes
71 W. 23rd St
www.peterflom.com
New York, NY 10010
(212) 845-4485 (voice)
(917) 438-0894 (fax)
>>> Nada Nadarajah <knada@ACESAG.AUBURN.EDU> 1/27/2004 11:07:41 AM >>>
Peter & Kevin:
There are SOME unique situations in Animal breeding where we often
re-code and fit "interaction terms" as main effect ( sometime as fixed
or random) in estimating parameters. Look at a simple classic
situation
where if I want to estimate the AGE effect of cows (2, 3, 4,5... yr.
olds) on milk production records collected from different herds (#100,
101, 102..) in different years ( 90, 91 ,92..) and different seasons (
fall=1, winter=2, spring=3 and summer=4). We are aware that because
of
the large variations in management levels in herds, good & bad
years
for pature/feeding, season (cold, hot) these effect would be be
significant. Our objective is to get a best estimate for Age effect
on
milk production. We want to fit a model that accounts for all the
above
effects and produce the best estimate for Age.
So, the herd -Year-season (the interaction) could be re-coded in some
fashion as (herd*1000000 + year*100 + season ), gives us a unique
code
for that interaction. example : herd #101, yr= 91, season=1 would be
10109101, is the effect for that combination. Here I am not interested
in that interaction, I knew that will be significant, all I want is to
fit the efficient model to get the best estimate for Age effect on
milk
production.
Just to give some thoughts on interactions as main effect.
Nada.
Peter Flom wrote:
>Kevin
>
>
>
>There may be SOME situations where it's a good idea to have just the
>interaction, but I cannot think of any, offhand
>
>HTH
>
>Peter
>
>
>Peter L. Flom, PhD
>Assistant Director, Statistics and Data Analysis Core
>Center for Drug Use and HIV Research
>National Development and Research Institutes
>71 W. 23rd St
>www.peterflom.com
>New York, NY 10010
>(212) 845-4485 (voice)
>(917) 438-0894 (fax)
>
>
>
>
>
>>>>Kevin Roland Viel <kviel@EMORY.EDU> 1/27/2004 9:09:42 AM >>>
>>>>
>>>>
>Peter,
>
> I try to follow almost of your posts. I apologize if I may have
>missed some or only perused others. While I am sure that any
question
>concerning selection procedures based on significance will elicit
this
>response from you, David, and perhaps Dale. I was wondering if I
>could
>not pull more from you.
>
> Consider the situation for which you have reason to suspect the
>importance of an interaction term. If the resulting regression
>suggests that it is significant (at an arbitrary alpha value) yet the
>main
>effect that is a component of the interaction is insignificant, do
>you drop it from the model?
>
>Thanks,
>
>Kevin
>
>
>Kevin Viel
>Department of Epidemiology
>Rollins School of Public Health
>Emory University
>Atlanta, GA 30322
>
>
--
*****************************************
* Nada K. Nadarajah, PhD *
* Sr. Research Fellow *
* Dept. Of Animal Sciences *
* 214 Upchurch Hall *
* Auburn University, AL 36849-5415 *
* USA *
* Tel:334-844-1502 Fax: 334-844-1519 *
* e-mail knada@acesag.auburn.edu *
* URL:http://www.auburn.edu/~nadarka *
* *
*****************************************
| 
