topic Re: PROC LOGISTIC, SELECTION=BACKWARD Variable Selection in Statistical Procedures

PROC LOGISTIC, SELECTION=BACKWARD Variable Selection

KyleM_Corrie — Sun, 06 Mar 2016 07:50:56 GMT

I've been looking at the sample "SAS Statistical Business Analysis" exam questions at http://support.sas.com/certify/creds/samples/

Question #6 is as follows:

Question 6

When selecting variables or effects using SELECTION=BACKWARD in the LOGISTIC procedure, the business analyst's model selection terminated at Step 3.

What happened between Step 1 and Step 2?

DF increased.
AIC increased.
Pr > Chisq increased.
- 2 Log L increased.

correct_answer = "D"

My question is: Shouldn't the answer be "C"? My understanding is that variable selection is based on a variable's p-value. The backward selection process, by default, eliminates variables one-by-one which don't meet the 0.10 criterion to "stay" in the model.

Could someone please clarify if my understanding is wrong? If so, please explain why "D" is the correct answer.

Thanks.

Re: PROC LOGISTIC, SELECTION=BACKWARD Variable Selection

Norman21 — Sun, 06 Mar 2016 09:14:25 GMT

The p values of all variables in the model will change when one is removed. One or more of the remaining might become "significant".

The question specifically asks about the step before variable selection. As variables are removed, - 2 Log L will increase. So the suggested answer is correct.

Norman.

Re: PROC LOGISTIC, SELECTION=BACKWARD Variable Selection

Ksharp — Mon, 07 Mar 2016 01:44:15 GMT

Oh, Contra .

Pr > Chisq( H0: beta = 0 ) is smaller is better (which means model is more significant. i.e. model can explain more variance of Y )

- 2 Log L is bigger is better(which is likelihood of X. more is better model).

Re: PROC LOGISTIC, SELECTION=BACKWARD Variable Selection

Norman21 — Mon, 07 Mar 2016 07:36:26 GMT

Bigger is not better.

The more variables that are included, the better the model fit (and the lower the deviance). However, the idea is to construct a model with only as many variables as are required.

Removing variables will decrease the model fit, but this, in itself, is not a bad thing.

Norman.