How to use macro to create a series of logistic regressions?

[PaigeMiller]

data a;

    set sashelp.class;

    if sex='M' then response=1; else response=0;

run;

proc pls data=a;

     model response=height weight;

run;

[Ksharp]

@PaigeMiller Thanks to post the code. I want know if the link function is LOGIT and the distribution of Y is conforming to Binomial Distribution ? If it was not, How could I say it is a logistic model .

[PaigeMiller]

It does not use the logit link, as I said the math is different.

If you really want to say that you are doing "logistic regression", then you probably can't say that using PROC PLS. If you want to say that you are modeling a 0/1 binary response, then you can say that using PROC PLS. As I have been pointing out, there are benefits when you use PROC PLS, and there are some disadvantages as well, specifically that the logit link isn't used. If the question is: which model fits best when you have a lot of correlated input variables, then it is my understanding, as demonstrated in this paper, (and my own personal experience) that PLS produces better fitting models than any variable selection method.

Also, here is a version of PLS that does use the logit link, although I have never found SAS code that implements this.

[Ksharp]

@PaigeMiller

I think PROC PLS is more like PROC REG, but better and more robust than PROC REG.

If you output the predict value by OUTPUT statement, you will see there is an obs 's predicted value is greater than 1, That could not be possible for a Probability .

[PaigeMiller]

Yes, there are drawbacks. I have always treated a predicted value >1 as a 1, and a predicted value <0 as a 0. PLS does a good job of modeling the "straight line" portion of the logistic curve.

The second link I gave overcomes this difficulty, it does use the proper logistic curve, while preserving the benefits of PLS in the case of correlated predictors, but as I said I am not aware of any SAS code that implements this.

[Ksharp]

Yeah. Hope one day PROC PLS could support    link=logit   and   dist=binomial options .

[huhuhu]

so smart!!

[Rick_SAS]

See the article "An easy way to run thousands of regressions in SAS." The example in that article is OLS, but the method would also work for a binary response.