about sas warning : obtaining minimum variance..

[aliceyang1102]

Hi, all.

i am trying to get results using sas university v.94

in the study, I want to know effect of sadness to acting.

variables are

sad: sadness(0,1)

sadcb: between subject in sadness

sadcw: withins subject in sadness

timec: time

when i try this, i got the results.

proc glimmix data=work.import1;

class id time;

model acting(event="1") = sad_lag1 sadcb_lag1 timec / link=logit

dist=binary ddf=95, 94, 95 solution cl;

random intercept / subject=id type=un g solution cl;

random time/ subject=id type=ar(1) residual;

run;

but when i try this syntax

PROC GLIMMIX DATA=work.import1;

CLASS id time;

MODEL acting (EVENT="1")=sadcw sadcb timec/LINK=logit

DIST=binary DDF=95, 94, 95 SOLUTION CL;

RANDOM intercept/SUBJECT=id TYPE=un g SOLUTION CL;

RANDOM time/SUBJECT=id TYPE=ar(1) RESIDUAL;

run;

i got this warning.

WARNING: Obtaining minimum variance quadratic unbiasedestimates as starting values for the covarianc...

please help if you know about this problem.

thank you

[StatsMan]

Using a PARMS statement to specify your own starting values can often get around this message.  GLIMMIX usually does a pretty good job in coming up with starting values, but you occasionally need to try your own values if you get a message like this.  The message can also indicate that your model is either too complicated or a poor fit to your data.  If you cannot get the model to converge no matter what starting values you try, then your model is telling you that it will just not work with this particular set of data.

Also, you do not need the TYPE=UN on the first RANDOM statement.  With a single RANDOM effect on the G-side, TYPE=UN is no different than specifying no TYPE= effect.  

[aliceyang1102]

Thank you for your reply.

when i search about this problem, i read about the solution which suggest PARMS.

But i dont know which PARMS i have to put in my syntax.

Is there any guide to get PARMS which can solve my problem?

[StatsMan]

You can try fitting a simpler model and using the final estimates from that model as starting values for the parameters in your new model.  You can use a range of values for the new parameters, if you do not have any insight as to what those values should be.

Always review the iteration history to make sure your model estimation has converged cleanly.  If you see the optimization jumping around a lot during the iteration history, that can be a sign that your model may not have converged to the true optimal solution. 

[aliceyang1102]

Thank you. I will try!!!!

Have a good day