Issues with Proc Mixed

[Jest]

Hi,

I am trying to run a proc mixed analysis for the following data:

ID            Score        year         group        othervariables

1                80             0              1

1                90             1              1

1                77             2              1

2                66             0              0

2                90             3              0

2                82             6              0

.                  .

.                  .

.                  .

So the years with scores are from 0 to 7 but some have scores years 0, 1,2, 3, 6 and some 0, 3, 6 etc. It varies. The dataset is fairly large and includes 17155 observations and 223 variables...I need to adjust for some variables in the model (around 7 etc).

When running the unadjusted model using the following code:

proc mixed data=cup.cup;

class year group;

model F393MSE = group year group*year;

repeated year/ subject=id type=ar(1);

run;

I get the following error:

WARNING: ODS graphics with more than 5000 points have been suppressed. Use the PLOTS(MAXPOINTS= )

option in the PROC MIXED statement to change or override the cutoff.

WARNING: Stopped because of infinite likelihood.

I added the  PLOTS(MAXPOINTS= NONE) in the statement, but it didn't work.

Surprisingly, when I added covariates in the model (around seven) the model runs. However, I do need to run the unadjusted.

Also, one other question. How does one run a GLM with REML method and specify the random effects...

Thank you for your help!

[WarrenKuhfeld]

You have two independent warnings.  This one is only related to graphs:

WARNING: ODS graphics with more than 5000 points have been suppressed. Use the PLOTS(MAXPOINTS= )

option in the PROC MIXED statement to change or override the cutoff.

This is why the procedure stopped:

WARNING: Stopped because of infinite likelihood.

Changing the maximum number of points plotted will not affect the likelihood.  Changing the model will.

[Ksharp]

Try 
repeated year/ subject=id type=chol ;

[Jest]

Thank you ksharp. I tried repeated year/ subject=id type=chol;

and I don't think chol is an option for the variance covariance structure. I got the following error:

ERROR 22-322: Syntax error, expecting one of the following: ANTE, AR, ARH, ARMA, CS, CSH, FA,
FA0, FA1, HF, LIN, LINEAR, SIMPLE, SP, TOEP, TOEPH, UN, UNAR, UNCS, UNR, VC.

any additional ideas? Is it possible for me to try proc glm with REML?

Thanks.

[sld]

If you have a mixed design, the MIXED (or GLIMMIX) procedure is much better than GLM; in fact, you generally do not want to use the GLM procedure with a mixed model.

Is the dataset sorted by ID? In the current model ID is a continuous variable, and because you are using it as a SUBJECT in the REPEATED statement, the dataset has to be sorted by ID.

What are the distributional properties of F393MSE? Is this variable measured on a continuous scale? Is normality a reasonable assumption?

Would it make any sense to regress on year? 

Do you still get an error: 

1) if you add ID to the CLASS statement

2) if you use 

repeated year/ subject=id type=cs;

If none of these provide any insight into your problem, then probably you'll need to provide data.

[Jest]

Thank you sld...

Sorry for the delay in sending an update...

The model worked when transforming the outcome which is a continuous scale to a logarithmic scale.

However, what are the options in running the model on the logarithmic scale and then generating the lsmeans and plot on the regular scale?

Thanks..

[sld]

The ILINK option does not function for DIST=LOGNORMAL. To get what you want, you can re-transform estimates produced (on the log scale) by the model "by hand"--either literally by hand, or save estimates to a SAS dataset and do the conversion in a data step or even in Excel.

See the GLIMMIX > MODEL > DIST= documentation http://documentation.sas.com/?docsetId=statug&docsetVersion=14.2&docsetTarget=statug_glimmix_syntax1...

which says:

When you choose DIST=LOGNORMAL, the GLIMMIX procedure models the logarithm of the response variable as a normal random variable. That is, the mean and variance are estimated on the logarithmic scale, assuming a normal distribution, . This enables you to draw on options that require a distribution in the exponential family—for example, by using a scoring algorithm in a GLM. To convert means and variances for  into those of , use the relationships

Also see this thread https://communities.sas.com/t5/General-SAS-Programming/ilink-will-not-return-inverse-values-in-the-l...

HTH 🙂