Policy using Treatment Classes with DynamicDML

[titubs]

@kbattocchi Hi Keith! I was curious to learn if the DynamicDML class is able to use Treatment Classes (like the DRPolicyForest) instead of discrete/continous variables? I am looking to understand what the optimal NEXT treatment in sequence for a given user would be? My treatment classes look like this:

Treatment A : consists of a distribution of possible values Treatment B : consists of a distribution of possible values

Example Policy: User X received Treatment B on day 1 and Treatment A on day 2. What is the next best Treatment for the next time?

Is this something the DynamicDML can assist with (or any alternative)?

[esbraun]

One possibility is to train two different DynamicDML learners: one for treatment A and the other for treatment B. For the treatment A learner, include features in X that describe the past history of treatment B, and vice versa for the treatment B learner. You could then compare the expected effect of the two different treatments for a given future time. You'd have to assume your characterization of the past history of treatment controls for all related confounding.

Someone please correct me if I'm wrong!