Logisitic PLS under SAS

[Celine_France]

I want to analyze a model which is composed of :
- 1 binary response (conform or not)
- 5 quantitative variables (time, temperature...etc)

My objective is to see which variables are linked to a conform response and which one are linked to a non-conform response.

So, I would like to make a PLS regression but under SIMCA software, the response must be quantitative (which is not my case here). How can I do a logistic PLS regression under SAS ?

By advance, thanks a lot,

Céline

[plf515]

> I want to analyze a model which is composed of :
> - 1 binary response (conform or not)
> - 5 quantitative variables (time, temperature...etc)
>
> My objective is to see which variables are linked to
> a conform response and which one are linked to a
> non-conform response.
>
> So, I would like to make a PLS regression but under
> SIMCA software, the response must be quantitative
> (which is not my case here). How can I do a logistic
> PLS regression under SAS ?
>
> By advance, thanks a lot,
>

I'm curious as to why you want a PLS model, instead of just an ordinary logistic model, which seems to me to answer your question. PLS, in my experience, can be really useful, but mostly when you have too many independent variables to interpret.

I beleive Paige Miller, who posts often on SAS-L, has posted about PLS with binary responses. You could try searching the archives of SAS-L.

Hope this helps

Peter

[Paige]

> I want to analyze a model which is composed of :
> - 1 binary response (conform or not)
> - 5 quantitative variables (time, temperature...etc)
>
> My objective is to see which variables are linked to
> a conform response and which one are linked to a
> non-conform response.
>
> So, I would like to make a PLS regression but under
> SIMCA software, the response must be quantitative
> (which is not my case here). How can I do a logistic
> PLS regression under SAS ?
>
> By advance, thanks a lot,

You could certainly set up your response as numeric 0 or 1 variable, pr the logistic transform of probabilities conform or not, and run it that way. PROC PLS would certainly accept that. Whether or not that's better than doing a logistic regression, as Peter suggests, depends on the exact nature of your data, and that is something I can't comment on. Message was edited by: Paige

[Celine_France]

Sorry to answer very late !

I've finally done a logistic regression on PLS components after having transformed my dataset .


Thanks a lot for your answers.