Multivariate logistic regression: univariate regression to select variables?

[michellemabelle]

How do I do univariate analysis for variable selection, say p<0.1, to enter into the final multivariate model?

Do I simply run it one by one per variable, and choose? Or is there a way to loop the variables, or to integrate it into proc logistic (eg, stepwise?)

proc logistic data=outfiles.trial2_miout descending;
	class AGE_above65 SEX BMI_obese ASACLASS_above4 FNSTATUS2 SMOKE HxDYSPNEA HYPERMED STEROID BLEEDDIS HxDIABETES HxCAD HxNEURO HxCHF HxRENALFAIL HxCOPD HxPVD ASCITES DISCANCR WNDINF TRANSFUS;
	model majormorbid=AGE_above65 SEX BMI_obese ASACLASS_above4 FNSTATUS2 SMOKE HxDYSPNEA HYPERMED STEROID BLEEDDIS HxDIABETES HxCAD HxNEURO HxCHF HxRENALFAIL HxCOPD HxPVD ASCITES DISCANCR WNDINF TRANSFUS;
	by _imputation_;
	ods output ParameterEstimates=outfiles.trial2_logistic_parms;
run;

proc mianalyze parms=trial2_logistic_parms;
	class AGE_above65 SEX BMI_obese ASACLASS_above4 FNSTATUS2 SMOKE HxDYSPNEA HYPERMED STEROID BLEEDDIS HxDIABETES HxCAD HxNEURO HxCHF HxRENALFAIL HxCOPD HxPVD ASCITES DISCANCR WNDINF TRANSFUS;
	modeleffects age_above65 sex bmi_obese asaclass_above4 smoke etoh diabetes_n steroid bleeddis cadhx_n neurohx_n neuro_deficit hxchf
	acuterenalfail Fnstatus2 HYPERMED DYSPNEA_n HXCOPD ascites discancr pvd_n wndinf_n transfus_n;
	ods output parameterestimates = trial2_mianalyze_parms;
run;

data outfiles.trial2_ORs;
	set trial2_mianalyze_parms;
	OR=exp(estimate);
	LCL_OR=exp(LCLMean);
	UCL_OR=exp(UCLMean);

proc print;
	var  OR LCL_OR UCL_OR;
run;

[Ksharp]

Check PROC HPGENSELECT .

[michellemabelle]

Thank you! Do I simply run HPGENSELECT one by one per variable, and choose those with p<0.1? Or is there a way to loop the variables, or to integrate it into proc logistic (eg, stepwise?) 

I also notice that HPGENSELECT is done through linear regression. My outcome variable is binary, eg. complication or no complication, so I thought logistic would be most appropriate. Regarding the convention for variable selection, I understand I need to run a univariate analysis first - would this be a univariate logistic regression or a univariate linear regression? 

Cheers!

[Ksharp]

No. You should put all the variables together. PROC HPGENSELECT also can handle logistic model .

Why do you want to do univariate analysis ? you should also consider the interaction effect between them .

[michellemabelle]

My supervisor insists that choosing variables to enter into the final multivariate model by doing univariate analysis is the standard in medical research for risk predictors. 

[Ksharp]

I think your supervisor want you to do some description statistic , like : max min , outlier 

and if it conform to Normal distribution ?