Programming the statistical procedures from SAS

Model Fit statistics in PROC MIANALYZE?

Reply
New Contributor
Posts: 2

Model Fit statistics in PROC MIANALYZE?

Trying to compare various models of data via AIC score and liklihood ratio testing, but I can't get these values when I pool my multipley imputed data in PROC MIANALYZE for each given model. In other words, I'm only getting the pooled estimate and standard error for each individual variable in my PROC MIANALYZE output, but can't seem to figure out how to get pooled global measures of model fit like pooled AIC score or pooled -2logliklihood score. Even though each indivdiual multipley imputed logistic regression has these values, I can't figure out how to pool them like I can pool the parameter estimates.  There doesn't seem like there is output option for these values in PROC LOGISTIC like there is for the parameters (parameterestimates = ) so I'm not sure if it's even possible to get pooled global estimates like AIC or -2logliklihood. 

 

Does anyone know how to get these pooled values? Thanks in advance!

 

My code so far (for one model as an example) is:

 

/*demographic MODEL*/
PROC LOGISTIC DATA= final.out_ipw_all outest = final.parms covout descending;
BY _imputation_ replicate;
WHERE sp2=1;
CLASS PARC (REF='2') depression (REF='1') BI_race_ethnicity (ref = '1') gender (ref = '1') / PARAM=REF;
MODEL PARC=depression age BI_race_ethnicity gender / link = glogit covb;
WEIGHT ipw_trim;
ODS OUTPUT ParameterEstimates=final.out_demographic ParmInfo=out.gmpinfo covb=final.out_cov_demographic;
RUN;

 

/*POOLING*/
proc sort data =final.out_demographic;
by variable ClassVal0 Response;
run;

 

PROC MIANALYZE data =final.out_demographic edf = 550 mult multivariate;
by variable ClassVal0 Response;
MODELEFFECTS Estimate;
StdErr StdErr;
ODS OUTPUT Parameterestimates=final.out_mianalyze_demographic_test;

RUN;

 

SAS Employee
Posts: 97

Re: Model Fit statistics in PROC MIANALYZE?

I haven't come across an effective way to pool measures like the AIC with multiply imputed data in the literature, but you could combine the likelihood ratio statistics, or more accurately, the Chi-Square statistics themselves using Dr. Paul Allison's %combchi macro.

http://www.ssc.upenn.edu/~allison/combchi.sas

 

Ask a Question
Discussion stats
  • 1 reply
  • 162 views
  • 0 likes
  • 2 in conversation