topic Re: PROC MIANALYZE procedure after PROC GLIMMIX binary outcome, 2-level model in Statistical Procedures

PROC MIANALYZE procedure after PROC GLIMMIX binary outcome, 2-level model

jtsfyu — Tue, 15 May 2018 03:40:13 GMT

Hello,

I used the blimp application to do multiple imputations for my missing data and was able to successfully run my 2-level hierarchical models using the imputed data. I can get pulled fixed effects, but I cannot figure out how to get my pulled Error Variance/Level-2 Intercept and my pulled model fit. Below is the program for one of my final models and my PROC MIANALYZE step. Thank you in advance for your help!!!

PROC GLIMMIX DATA=dissert.centered METHOD=LAPLACE NOCLPRINT;
CLASS cdscode ethn_new (ref=FIRST) sex (ref=FIRST);
MODEL m_health (EVENT=LAST)= p_ed sex ethn_new grade_new_c cumAA_c RJLevel YearsImp_c PeerRJ WhS/CL DIST=BINARY LINK=LOGIT SOLUTION
ODDSRATIO (DIFF=FIRST LABEL);
RANDOM INTERCEPT / SUBJECT=cdscode S CL TYPE=VC;
by _imputation_;
ods output ParameterEstimates=lgsparm1;
COVTEST /WALD;
RUN;

PROC MIANALYZE parms = lgsparm1;
CLASS ethn_new sex;
modeleffects Intercept p_ed sex ethn_new grade_new_c cumAA_c RJLevel YearsImp_c PeerRJ WhS;
RUN;

Re: PROC MIANALYZE procedure after PROC GLIMMIX binary outcome, 2-level model

SAS_Rob — Thu, 17 May 2018 17:57:06 GMT

Below is a sample program that shows how to combine the results from Proc GLIMMIX in MIANALYZE.

data test;
seed=621435234;
do id=1 to 30;
rc=rannor(seed); /* random effect on intercept for id */
rb=rannor(seed)*.7; /* random effect on slope for id */
do rep=1 to 20;
x1=ranuni(seed);
x2=ranuni(seed);
logit=-2 + (2+rb)*x1 + rc+.08*x2;
p=exp(logit)/(1+exp(logit));
if ranuni(seed)< p then y=1; else y=0;
output;
end;
end;
run;

data test;set test;
if ranuni(312)>.5 then x2=.;*create missing values;
run;
proc mi data=test out=out1 seed=10;
var y x1 x2;
run;
ods trace on;
proc glimmix data=out1;
by _imputation_;
class id;
model y=x1 x2/d=binomial solution or(label);;
random intercept/subject=id solution;
ods output parameterestimates=parms solutionr=rand(rename=(stderrpred=stderr));
run;

/*Combining Random Effects*/
proc sort data=rand;
by effect subject _imputation_;
run;

proc mianalyze parms=rand;
by subject;
modeleffects intercept;
run;

Re: PROC MIANALYZE procedure after PROC GLIMMIX binary outcome, 2-level model

jtsfyu — Sat, 19 May 2018 16:40:25 GMT

Thanks, SAS_Rob!

When I use the /*Combining Random Effects*/ code you provided, I just get 20 separate estimates for 20 imputation sets. I am looking for a way to have one pulled level 2 variance.

I am able to do that for the fixed effects by using this program--at the end of the analysis, I have the combined effects for the 20 models.

PROC GLIMMIX DATA=dissert.centered METHOD=LAPLACE NOCLPRINT;
CLASS cdscode ethn_r (ref=FIRST) sex (ref=FIRST)sex;
MODEL m_health (EVENT=LAST)= ethn_r sex grade_new_c /CL DIST=BINARY LINK=LOGIT SOLUTION
ODDSRATIO (DIFF=FIRST LABEL);
RANDOM INTERCEPT / SUBJECT=cdscode S CL TYPE=VC;
by _imputation_;
nloptions gconv=0;
ods output ParameterEstimates=lgsparm35;
COVTEST /WALD;
RUN;

proc mianalyze parms = lgsparm35;
CLASS ethn_r sex;
modeleffects Intercept ethn_r sex grade_new_c;
run;

Thanks for your help!

Jelena

Re: PROC MIANALYZE procedure after PROC GLIMMIX binary outcome, 2-level model

SAS_Rob — Mon, 21 May 2018 12:31:53 GMT

Because you use the SUBJECT= option on the RANDOM statement, you will get an estimate for each of the subjects. That is why you get multiple estimates and you have to use the BY statement in MIANALYZE.

In other words, there is not one overall estimate, but instead one overall estimate for each subject.