How to do automated variables selection when accounting for clustering effect in SAS

[Lao_feng]

Dear all,

I am doing an analysis to look at the association between a number of potential risk factors (age, gender, education, marital status, wealth index etc.) and risk of developing diabetes. Normally, logistic regression using "Proc logistic" should be fine for this analysis. But in my study, we used clustering sampling method (30 clusters in total) to enroll study participants. Therefore, I used " Proc glimmix" (SAS 9.4)instead to account for clustering effect. However, it seems that "Proc glimmix" cannot do automated variable selection like "Proc logistic" does. Does anyone know if there is a way to do automated variable selection while modelling clustering effect as well ?

Here are the codes:

proc glimmix data=diab NOCLPRINT;

class country2(ref='Pk')clu_num fam_num2 gender edu2g mar2g ;

model diab(ref='0')= country2 age_y9 gender edu2g mar2g wic/solution Link=logit dist=binary

oddsratio (label unit age_y9=5 UNIT WIC=21);

random int /sub=clu_num (country2);

random int /sub=fam_num2(clu_num country2) type=vc;

run;

Thanks a lot !

Feng Liang

[pau13rown · Accepted solution]

" information criteria can be used to select a best fitting covariance structure" https://communities.sas.com/t5/SAS-Statistical-Procedures/proc-glimmix-selection-stepwise-or-backwar...

automated is maybe not the best approach anywy, variable selection should be informed by clinical reasong and not entirely data-dependent, just my opinion