[prev in list] [next in list] [prev in thread] [next in thread] 

List:       sas-l
Subject:    Re: Testing Trends when Equal Variances are not Assumed
From:       SUBSCRIBE SAS-L Anonymous <pdhart () LIVE ! COM>
Date:       2010-02-27 16:43:07
Message-ID: 201002271643.o1RBowjg018079 () mailgw ! cc ! uga ! edu
[Download RAW message or body]

On Fri, 26 Feb 2010 23:58:13 -0800, Dale McLerran
<stringplayer_2@YAHOO.COM> wrote:

>--- On Fri, 2/26/10, SUBSCRIBE SAS-L Anonymous <pdhart@LIVE.COM> wrote:
>
>> From: SUBSCRIBE SAS-L Anonymous <pdhart@LIVE.COM>
>> Subject: Re: Testing Trends when Equal Variances are not Assumed
>> To: SAS-L@LISTSERV.UGA.EDU
>> Date: Friday, February 26, 2010, 2:40 PM
>> On Fri, 26 Feb 2010 13:00:17 -0800,
>> Dale McLerran
>> <stringplayer_2@YAHOO.COM>
>> wrote:
>>
>> >--- On Fri, 2/26/10, SUBSCRIBE SAS-L Anonymous <pdhart@LIVE.COM>
>> wrote:
>> >
>> >> From: SUBSCRIBE SAS-L Anonymous <pdhart@LIVE.COM>
>> >> Subject: Testing Trends when Equal Variances are
>> not Assumed
>> >> To: SAS-L@LISTSERV.UGA.EDU
>> >> Date: Friday, February 26, 2010, 11:24 AM
>> >> The proc glm trend contrasts give me
>> >> tests under equal variances.  If I
>> >> repeat using same coefficients in spss, I get two
>> sets of
>> >> tests - one for
>> >> equal variances assumed and one for equal
>> variances not
>> >> assumed.  Can I
>> >> get these latter tests in glm?
>> >>
>> >> Thanks
>> >>
>> >> peter
>> >>
>> >
>> >Peter,
>> >
>> >Can you clarify your question, including a statement of the
>> >model which you are fitting and where the variance homogeneity
>> >assumption may be suspect?  Your question is so sparse in
>> >detail that it is difficult to respond to.
>> >
>> >Dale
>> >
>> >---------------------------------------
>> >Dale McLerran
>> >Fred Hutchinson Cancer Research Center
>> >mailto: dmclerra@NO_SPAMfhcrc.org
>> >Ph:  (206) 667-2926
>> >Fax: (206) 667-5977
>> >---------------------------------------
>>
>> Hey Dale,
>> Thanks.  This is not a stats question - I'm just simply trying to get
sas
>> to produce output similar to what I know I can get in another package.
>> With this code:
>>
>> data a;
>> input mem time $;
>> cards;
>> 2 1
>> 3 1
>> 1 1
>> 2 1
>> 0 1
>> 4 1
>> 6 2
>> 8 2
>> 5 2
>> 3 2
>> 7 2
>> 7 2
>> 6 3
>> 8 3
>> 10 3
>> 5 3
>> 10 3
>> 9 3
>> 11 4
>> 10 4
>> 7 4
>> 9 4
>> 8 4
>> 9 4
>> ;
>> proc glm data = a;
>> class time;
>> model mem = time;
>> contrast 'linear' time -1.5 -.5 .5 1.5;
>> run;
>>
>> I get a t-test for my linear trend.  This is a pooled t-test (using
MSE).
>> I would like to know if I can get a non-pooled t-test for this trend
>> also.  I've searched documentation, with no luck.
>>
>> Thanks
>>
>
>Ah, but it is a stats question and there is a model involved.
>Without providing detail, it was not clear whether there were
>other variables involved.  There could have been variance
>heterogeneity over some other variables.  A more specific
>presentation allows a better the response.
>
>I don't believe that GLM will produce a trend test in which
>the assumption of variance homogeneity is relaxed.  However,
>you can get such a test using the MIXED procedure.  For the
>data and model indicated above, MIXED procedure code for unequal
>variances can be obtained as:
>
>  proc mixed data = a;
>    class time;
>    model mem = time;
>    repeated / group=time;
>    contrast 'linear' time -1.5 -.5 .5 1.5;
>  run;
>
>
>The MIXED procedure would allow other tests.  For instance,
>a model in which the variance is a power of the mean could
>be constructed for the MIXED procedure (although such models
>are more easily fit using the NLMIXED procedure). In general,
>the MIXED procedure is favored over the GLM procedure.  The
>MIXED procedure has much more flexibility for alternate
>specifications of the variance model than are allowed by the
>GLM procedure.
>
>The data which you present above do not seem to conform to a
>power-of-the-mean model.  But were such a model appropriate,
>the POM model would be more efficient than a model in which
>the variance is estimated separately for each level of the
>predictor variable.
>
>In my experience, variances do either remain constant over
>all expected values or variances increase with expected values.
>In the data presented above (which may be dummied up), the
>mean at time=4 is larger than the mean at any other time.
>Mean values are monotonically related to time.  The variances
>do seem to increase with the mean until we get to time=4
>at which point the variance is as small as at time=1 which,
>as indicated by the monotonicity of the mean with time, has
>the lowest mean value.  These observations would lead me to
>question whether there is a real difference in residual
>variance over time or whether the difference in variances is
>some sort of artifact.
>
>Dale
>
>---------------------------------------
>Dale McLerran
>Fred Hutchinson Cancer Research Center
>mailto: dmclerra@NO_SPAMfhcrc.org
>Ph:  (206) 667-2926
>Fax: (206) 667-5977
>---------------------------------------

Thanks, Dale.  Great information.  I really appreciate it.

Peter
[prev in list] [next in list] [prev in thread] [next in thread]