Forward selection Method....

[HimaAalamuri]

Hi All,

could anyone knows please clarify my doubt,

i have one dept variable "depvar" and

10 covariates - covar1 covar2 covar3 covar4 covar5 covar6 covar7 covar8 covar9 covar10.

i have to display type3 with p-vlaues using proc glm by using forward selection method.

how can i start and select covariates by using forward seletion method ? in SAP i dn't have any info related to forward selection method so what should i do in this case?

please help me out from this!

and one more question what is ment by forwards selection method and why we have to use?? what is the purpose behind this?

Best Regards & Thanks in advance.

[SteveDenham]

Forward selection is a widely discredited model building tool that leaves you with a model that fits your data, but has almost no relevance to anything else.  It is available in a wide variety of procedures, although not in PROC GLM.

If it is not in your analysis plan, work with whoever thinks this is a good idea, and convince them to try some other method.  LAR, LASSO and ELASTICNET are all available in PROC GLMSELECT.  Combined with model averaging, you may develop a decent model.

Read the following "Stopping stepwise: Why stepwise and similar selection methods are bad, and what you should use".  Here is the link:

http://www.nesug.org/proceedings/nesug07/sa/sa07.pdf

Steve Denham

[HimaAalamuri]

Thank you so much SteveDenham!

Proc Glmselect: i used following code to display type3 p-value.

proc glmselect data=<datasetname>;

model deptval=covar1 covar2 covar3 ......................... covar20/ selection=forward(select=SL choose=AIC);

run;

my problems are:

1) in my SAP they won't mention any selection methoeds for forward like "AIC/Rsequre etc..", i only consider "AIC" in my model is it correct or not?

2) but in output they mentioned like this "the SAS type 3 p-values will be used to test (two-side) whether the impact of ach covariate is statistically significant at % significance level.  note that a type 3 p-value for a covariate  is comuted adjusting for all other coavariates in the model.

so based on above points my code is correct or not? may i modify anything please suggest !

Thanks in advance.

[SteveDenham]

The code is fine for forward selection.  My objection is to that particular methodology, which has been shown to lead to major problems.

Steve Denham

[HimaAalamuri]

hi SteveDenham & paigemiller,

by using proc glmselect we get pvalue at 0.20 but i want at 0.05 then how can i get?

so i am using sle=0.05 & sls=0.05, i am sure sle=0.05 but what is the use of sls=0.05 option? please let me know both options need to include or only one (Sle=0.05)?

pls suggest!

[PaigeMiller]

Another alternative to forward selection is PROC PLS.

(and I agree with @SteveDenham that forward selection has been discredited, and I wish SAS would remove it from PROC REG, but I doubt they will)