topic Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors in Statistical Procedures

Multiple Imputation in GLIMMIX with continuous and categorical predictors

halladje — Fri, 05 Jul 2019 18:09:29 GMT

Hello!

I am trying to pool my effect estimates across my imputed datasets using a 3-level logistic regression using GLIMMIX. However, only my class variable effect estimates are saving.

Here is my current code:

proc glimmix data=lpa method=rspl noclprint;
class idschl jclass mhserv (Ref=first) class_mem (ref=first) female(ref=first) black (ref=first)
esa (ref=first) other (Ref=first);
model mhserv=class_mem female age esa black other assets/CL dist=binary link=logit solution ddfm=bw;
by _imputation_;
random intercept/ type=un subject=idschl;
random intercept/Type=un subject=jclass (idschl);
ods output solutionf=imputeout;
run;
quit;

**THESE RUN PROPERLY**

PROC DATASETS NOLIST NODETAILS;
CONTENTS DATA=imputeout ;
RUN;

**THS DATASET ONLY CONTAINS EFFECT ESTIMATES THAT APPEARED IN THE CLASS LINE**

proc mianalyze parms=a_mvn;
class idschl jclass mhserv class_mem female black
esa other;
modeleffects intercept class_mem female age esa black other assets;
STDERR class_mem;
run;

**THIS DOESN'T WORK BECAUSE EFFECT ESTIMATES FOR assets and age ARE NOT SAVED**

Thanks in advance!

Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors

SAS_Rob — Fri, 05 Jul 2019 18:40:21 GMT

It would be helpful to see the LOG messages you get from the two procedures as well as a snippet of the output from GLIMMIX that shows the parameters.

Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors

halladje — Fri, 05 Jul 2019 18:47:15 GMT

Hi Rob,

These are the variables saved in imputeout. As you can see, no continuous variable parameters are saved so therefore there is nothing to pool on the subsequent command.

13	Alpha	Num	8	BEST6.
10	DF	Num	8	BEST5.
2	Effect	Char	9
8	Estimate	Num	8	D8.4
14	Lower	Num	8	D8.4
12	Probt	Num	8	PVALUE6.4	Pr > \|t\|
9	StdErr	Num	8	D8.4	Standard Error
15	Upper	Num	8	D8.4
1	_imputation_	Num	8
5	black	Num	8	BEST.
3	class_mem	Num	8	BEST.
6	esa	Num	8	BEST.
4	female	Num	8	BEST.
7	other	Num	8	BEST.
11	tValue	Num	8	7.2	t Value

Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors

SAS_Rob — Fri, 05 Jul 2019 18:55:35 GMT

It seems you are right, but figuring out why without the GLIMMIX output and LOG containing all the code would be practically impossible.

Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors

halladje — Fri, 05 Jul 2019 19:15:13 GMT

See attached for the log output from the "by _imputation_" run.

Let me know your thoughts! Thanks so much.

Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors

SAS_Rob — Fri, 05 Jul 2019 19:25:19 GMT

I see a WARNING message that the table SolutionF was not created. Proc GLIMMIX does not have a table with that name the way MIXED does. Instead you should be using the ParameterEstimates table. Try changing that and see if the problem goes away. If not, then please post the entire LOG that includes the MIANALYZE part as well.

Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors

halladje — Fri, 05 Jul 2019 20:08:15 GMT

I have been having the same issues with ParameterEstimates as well. See attached. I did my imputation in BLIMP - an external 3-level multi-level imputation software - so the actual MI was not done in SAS. However, I renamed the imputation variable the same as the SAS default and it is producing accurate estimates in the stratified analyses.

Thanks again, I really appreciate your time.

Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors

halladje — Tue, 09 Jul 2019 12:14:35 GMT

I realized I had some errors in the last code. See attached for an updated. Same issue but I figured out what is going on (just not how to fix it). The results of my imputation are saving for each variable but where each variable and each imputation is listed as an observation. The headings/variable labels are related to the classes (to indicate the coding for that effect estimate), followed by the estimate, standard error, degrees of freedom, etc. So the continuous effects are being saved but as an observation and not as a heading.

Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors

SAS_Rob — Tue, 09 Jul 2019 13:40:19 GMT

When you use a PARMS= data set for input, you should not use the STDERR statement. All you need is the MODEFFECTS statement.

Re: Multiple Imputation in GLIMMIX with continuous and categorical predictors

halladje — Tue, 09 Jul 2019 15:15:22 GMT

Great! That worked thank you so much.

Jillian