|
Abstract:
|
In this article we discuss identifying variables that are important for adapting or personalizing treatment. Most variable selection techniques focus on variable selection for the prediction of the response in a supervised learning setting. However, as noted by many, very few of these variables are likely to be useful for deciding which treatment to provide to which patient. In constructing personalized treatments or dynamic treatment regimes, we are most interested in variables that have qualitative interaction with the treatment. Variables that qualitatively interact with treatment not only inform us about the magnitude of treatment effect, but also differentiate between patients who should be offered different treatments. We propose to identify such variables through value of information, and discuss how to make the statistical inference on the variables selected.
|
