[R] glmnet inclusion / exclusion of categorical variables

[David Winsemius]

On Aug 9, 2013, at 6:44 AM, Kevin Shaney wrote:

> 
> Hello -
> 
> I have been using GLMNET of the following form to predict multinomial logistic / class dependent variables:
> 
> mglmnet=glmnet(xxb,yb ,alpha=ty,dfmax=dfm,
> family="multinomial",standardize=FALSE)
> 
> I am using both continuous and categorical variables as predictors, and am using sparse.model.matrix to code my x's into a matrix.  This is changing an example categorical variable whose original name / values is {V1 = "1" or "2" or "3"} into two recoded variables {V12= "1" or "0" and V13 = "1" or "0"}.

You set their penalty factors to be 0 to at least observe the case where inclusion is performed. And setting the penallty factor for both to be small would allow you to "honestly" use 0 as the estimated coefficient in such cases where one was estimated and the other not. 

> 
> As i am cycling through different penalties, i would like to either have both recoded variables included or both excluded, but not one included - and
> can't figure out how to make that work.   I tried changing the
> "type.multinomial" option, as that looks like this option should do what i want, but can't get it to work (maybe the difference in recoded variable names is driving this).

Doesn't the 'family' argument, used to set what I think you are calling 'type', just refer to the y argument, rather  than the predictors. You may want:

   mglmnet=glmnet(xxb,yb ,alpha=ty,dfmax=dfm, type.multinomial="grouped",
                 family="multinomial",standardize=FALSE)

> 
> To summarize, for categorical variables, i would like to hierarchically constrain inclusion / exclusion of recoded variables in the model - either all of the recoded variables from the same original categorical  variable are in, or all are out.

I do understand that I am possibly not directly answering your question, but in some respect I wonder if it deserves an answer. I think it is meaningful if some factor levels are "penalized-out" of models.

-- 
David Winsemius
Alameda, CA, USA