lift/uplift/netlift modeling: CART analysis maximizing a t-statistic from a multivariate regression

Question

For tree based modeling most packages use some data element as the dependent variable. In uplift modeling the goal is to maximize the difference between treatment and control groups.

One way to implement this is with a CART which maximizes the t-statistic for an interaction term between the split (1 = left, 0 = right) and the treatment variable (1 = treated, 0 = control) b_3 below,

y = b_1 * T + b_2 * P + b_3 * T * P

as I understand it, the interaction term (T*P) captures difference in treatment effects between the two sides of a split.

being new to CART, I don't know which software packages would enable this analysis. I program in python (currently exploring orange), r and sas, so I would appreciate any pointers to example/documentation on similar projects.

Answer 1

R party package does support model-based recursive partitioning but I doubt that it allows what you want to do out of the box - you may have to hack the source code to make it do the splits based on the expected incremental effect.

By the way, this paper may save you some trial and error.