Date: Thu, 25 Mar 2010 12:43:28 -0700
Reply-To: Dale McLerran <stringplayer_2@YAHOO.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Dale McLerran <stringplayer_2@YAHOO.COM>
Subject: Re: GLIMMIX warning
In-Reply-To: <OFEFBB0803.4D9D18F7-ON862576F1.004D2883-862576F1.004DF065@unmc.edu>
Content-Type: text/plain; charset=iso-8859-1
Robin demonstrates an analysis where a reparameterization of the
variance of the random effect might result in a better behaved
solution. The constraint on the parameter space near the
final estimated value may be adversely affecting the ability
to take a derivative. We can reparameterize the variance in
the NLMIXED model. Unfortunately, I don't believe that any
such reparameterization of the variance is possible when using
the GLIMMIX procedure.
Instead of parameterizing the NLMIXED model in terms of the
variance of the random effect, try parameterizing the model in
terms of log(V) or log(sqrt(V)). In NLMIXED, we could write
Robin's model with code as shown below (although a part of
the code is incomplete as I don't know what distribution was
employed):
proc nlmixed data=mydata;
parms _t1-_t6
_m1-_m3
LOG_S 0;
_t7=0;
_m4=0;
array _t {7} _t1-_t7;
array _m {4} _m1-_m4;
eta = b0 + _t{trt) + _m{mix} + U;
model y ~ ...;
random U ~ normal(0, exp(2*LOG_S)) subject=subj_id;
run;
Note that I have capitalized every part of the parameterization
of the variance of the random effect (U). The parameter which
we estimate is LOG_S. The variance is simply the function
exp(2*LOG_S). Note that by computing the variance as the
exponent of a parameter which has interpretation as log(sqrt(V)),
we ensure that the variance remains positive. It could also
help in computing the derivative of the parameter since our
parameter is not confined to a non-negative space but has
domain of the entire set of real numbers.
Dale
---------------------------------------
Dale McLerran
Fred Hutchinson Cancer Research Center
mailto: dmclerra@NO_SPAMfhcrc.org
Ph: (206) 667-2926
Fax: (206) 667-5977
---------------------------------------
--- On Thu, 3/25/10, Robin R High <rhigh@UNMC.EDU> wrote:
> From: Robin R High <rhigh@UNMC.EDU>
> Subject: Re: GLIMMIX warning
> To: SAS-L@LISTSERV.UGA.EDU
> Date: Thursday, March 25, 2010, 7:11 AM
> Holly,
>
> It likely means one or more your parameter and variance estimates
> have not fully converged. When I ran a similar analysis with GLIMMIX
> (with method=quad) that produced this warning and then ran it with
> NLMIXED I received the following output:
>
> GLIMMIX
>
> Covariance Parameter Estimates
>
> Standard
> Cov Parm Subject Estimate Error
> Intercept blk 0.004511 0.004165
>
>
> Solutions for Fixed Effects
>
> Standard
> Effect trt mix Estimate Error
> Intercept 3.4980 0.06333
> trt 1 -1.1689 0.07881
> trt 2 -0.7097 0.06683
> trt 3 -0.8683 0.07058
> trt 4 -0.5676 0.06380
> trt 5 -0.2461 0.05793
> trt 6 -0.3003 0.05883
> trt 7 0 .
> mix 1 0.4434 0.05361
> mix 2 0.3627 0.05449
> mix 3 0.1230 0.05743
> mix 4 0 .
>
>
>
> The NLMIXED Procedure
>
> Parameter Estimates
>
> Standard
> Parameter Estimate Error Gradient
> b0 3.4974 0.06334 -0.00118
> _t1 -1.1688 0.07882 -0.0003
> _t2 -0.7095 0.06683 -0.00002
> _t3 -0.8681 0.07058 -0.00079
> _t4 -0.5674 0.06380 -0.00013
> _t5 -0.2459 0.05793 -0.00032
> _t6 -0.3000 0.05883 -0.00038
> _m1 0.4438 0.05362 -0.00029
> _m2 0.3631 0.05449 -0.00014
> _m3 0.1234 0.05744 0.00045
> sig 0.004509 0.004158 -0.01312 ****
>
> The estimates here are quite close, even the offending variance estimate.
> How serious it would be in your situation is not known, though is a sign
> that you don't have all that much variation in the random factor section,
> and if possible, to acquire more sections.
>
> And just as an aside, you might also try dist=negbin with ddfm=contain
> (rather than bw).
>
> Robin High
> UNMC
>
>
>
>
>
> From: Holly Kalyn Bogard <holly.bogard@EC.GC.CA>
> To: SAS-L@LISTSERV.UGA.EDU
> Date: 03/24/2010 02:57 PM
> Subject: GLIMMIX warning
> Sent by: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
>
> Hello.
>
> I am a graduate student using GLIMMIX to analyze count data. I am
> investigating the effects that the point count distance to industrial
> disturbance has on bird abundance. Several of my models are bringing
> up a note which I have not encountered before. Has anyone encountered
> this warning? Any advice? I would appreciate any advice.
>
> Thanks,
>
> Holly
>
>
> 2173 title 'Point Count Distance To DIRT Model';
> 2174 proc glimmix data=merged method=LAPLACE IC=pq;
> 2175 class section pasture;
> 2176 model species=logPCDT_DIRT/ dist=p link=log solution DDFM=BW;
> 2177 random intercept / subject=section(pasture);
> 2178 run;
>
>
> NOTE: Convergence criterion (GCONV=1E-8) satisfied.
> NOTE: At least one element of the gradient is greater than 1e-3.
> NOTE: PROCEDURE GLIMMIX used (Total process time):
> real time
> 0.28 seconds
> cpu time
> 0.25 seconds
>
