Date:         Fri, 14 Jan 2000 20:18:46 GMT
Reply-To:     Dale McLerran <dmclerra@FHCRC.ORG>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Dale McLerran <dmclerra@FHCRC.ORG>
Organization: Fred Hutchinson Cancer Research Center
Subject:      Re: Prob. with analysis for many groups

[posted and mailed]

lbadiella@my-deja.com wrote in <85d96n$hda$1@nnrp1.deja.com>:

> Hi all,
> First of all, I apologize for all my spelling mistakes.
>
> Here is my litle problem.
>
> I have a dataset with 3000 of groups corresponding to different offices
> and about 30 observations each group corresponding to people asked. I
> would like to compare groups and study if the values for a continuous
> variable (satisfaction with an entity) are different or not.
>
> What kind of analysis is more appropiate for this kind of problem with
> many groups. Anova shold work for few groups. But in practice, does
> Anova work for many groups?
>
> In the other hand, I know there are some other variables that would
> influentiate to the continuous variable. How can I delete
> this influence?
>
> Thanks in advance,
>
> Llorenç Badiella
>
>
> Sent via Deja.com http://www.deja.com/
> Before you buy.

Llorenç,

You are on the right track here.  ANOVA as introduced in almost any
first course in statistics is not appropriate.  That is to say, a
fixed-effect analysis of variance should not be used in this situation.
Rather, you should use what is called a random-effects ANOVA.  The
random effects ANOVA considers a population of randomization units
(in your case worksites), and asks if there is variation among the
units of randomization.  What you get out of the random-effects ANOVA
is a variance estimate.

The MIXED procedure is the best procedure for fitting a random-effects
ANOVA.  The syntax would be something like the following:

PROC MIXED DATA=mydata RATIO COVTEST;
CLASS worksite <categorical covariates>;
MODEL response = <covariates> / S;
RANDOM worksite;
RUN;

This model has two variance components, the variance among the worksites
and the between person variance after accounting for worksite effects.
Now, with 3000 worksites, you can examine whether the worksite variance
differs from zero.  If it does, then there are worksite effects present
in your sample.  The COVTEST option above will test the variance
estimates to see if they depart from zero.

Inclusion of covariates results in a mixed-effects model (hence the name
of the procedure).  This addresses the second part of your posting,
where you would like to control for the influence of other variables.
These other variables get tested as fixed-effects because they do not
appear on the RANDOM statement.

Hope this helps.

Dale McLerran                            dmclerra@fhcrc.org
Fred Hutchinson Cancer Research Center   ph:  (206) 667-2926
Seattle, WA 98109                        fax: (206) 667-5977