Elastic Net selection with interaction effects

[delgaa07]

Hi all,

I am a healthcare investigator and am looking to have a model I am trying to make include interaction terms. My only problem is that when I include the interaction effects into the MODEL statement, I sometimes will get the interaction effects without the main effects in the model. I want to know how I might be able to keep the main effects in the model along with an interaction effect. 

Here is my code:

%macro elasticnet (l2_cust_low, l2_cust_high);
ods select CoefficientPanel ASEPlot ParameterEstimates FitStatistics;
proc glmselect data=data.resilience3 plots=all seed=712;
title1 "SELECTION=ELASTICNET(L2SEARCH=GOLDEN L2LOW= &l2_cust_low L2HIGH= &l2_cust_high))";
title2 "Primary Analysis";
partition fraction(test=0.3);
class 	sex(ref='0') nonhisp_black(ref='0') nonhisp_white(ref='0') hispanic(ref='0') Married(ref='0') single(ref='0')
		ged_hsdiploma(ref='0') Assoc_Bachelor(ref='0') graduate(ref='0') Tetra_abc(ref='0') para_abc(ref='0')
		sleep(ref='0') pain_interfere2(ref='0') Healthsts_pos(ref='0') fyear_1(ref='0') fyear_5(ref='0') fyear_10(ref='0') 
		fyear_15(ref='0') fyear_20_25(ref='0') fyear_30plus(ref='0');
model BRsTScr2 = age sex nonhisp_black nonhisp_white hispanic Married single  
					ged_hsdiploma Assoc_Bachelor graduate Tetra_abc para_abc
					fyear_5 fyear_10 fyear_15 fyear_20_25 fyear_30plus
					BSFBMTS2 BSFSCTS2 BSFFMoTS2 sleep pain pain_interfere2 Healthsts_pos 
					BCHPITot2 BCHMbTot2 BCHOpTot2 BCHSocIn2 depression life_sat pain*depression / 
					SELECTION=ELASTICNET(L2SEARCH=GOLDEN L2LOW= &l2_cust_low L2HIGH= &l2_cust_high choose = cv) 
					cvmethod=split(10) hierarchy=none showpvalues;
					
run;
%MEND elasticnet;

%elasticnet(0,0.1);
%elasticnet(0.1,0.2);
%elasticnet(0.2,0.3);
%elasticnet(0.3,0.4);
%elasticnet(0.4,0.5);
%elasticnet(0.5,0.6);
%elasticnet(0.6,0.7);
%elasticnet(0.7,0.8);
%elasticnet(0.8,0.9);
%elasticnet(0.9,1);

Any and all help will be greatly appreciated!

[Rick_SAS · Accepted solution]

For some selection methods, you can use HIERARCHY=SINGLE or INCLUDE=n to force terms to stay in the model. However, the elastic net method does not support those options. 

I think the best you can do is to run the procedure twice. The first time, use the SELECTION=ELASTICNET method to choose effects for the model. Then run the procedure again using SELECTION=NONE to fit the model, but force the main effects that you want in the model. The first run creates the _GLSIND macro variable, which contains the selected effects.  Therefore, if the pain*depression effect is selected, the second model will be

MODEL BRsTScr2 = pain depression &_GLSIND / selection=none;

For your example, in your