I am familiarizing myself with LASSO for
Poisson regression using the glars
package to model discrete outcomes in observational healthcare data sets.
The technique performs as advertized in terms of selecting reasonable subsets
of variables for inclusion in the models, picking from the large number of
variables that are available to me. I have, however, run into two issues:
1) I would like to set things up so that a particular variable, namely
treatment with Drug X, is always
included in the model. I can do this with stepwise regression [ step() ] by specifying treatment in my “lower” model,
but I don’t know how to do it with the available LASSO package. Can
this be done?
2) I am more interested in getting an accurate coefficient estimate for
the treatment variable than I am
in overall prediction accuracy. This coefficient is “my answer”.
The LASSO as currently configured will shrink the treatment coefficient so that there is some headroom within
the constraint box to share with other variables that improve overall accuracy.
Is there a way to remove the penalty constraint from one particular variable,
while keeping it for the others? Does it make sense to do this?
Thanks in advance,
Alan
Alan Hochberg
VP, Research
ProSanos Corporation
225 Market St. Ste. 502,
Harrisburg, PA 17101
Tel
717-635-2124 * Fax 717-635-2575