Problem
Analysts who target promotions or communications need the effect of an action on each user, not only the average effect. The basic learners (T-learner, S-learner, Class Transformation) lose accuracy when one group is much larger than the other, when the groups differ in covariates, or when customers choose themselves whether to take part. Real data has no known true effect, so methods can be compared only on synthetic data.
Example from the article: a promotion "three chocolates for the price of two". Stores are split into those with and without the promotion, and each customer in a promo store decides whether to take part.
Solution in brief
The talk covers "advanced" estimators of the conditional average treatment effect (CATE) as implemented in EconML and CausalML. They fit outcome and propensity models on separate data parts and combine them, which stays stable under imbalance and confounding. DRIV handles self-selection through an instrument, and uplift trees (DDP, CTS) split directly on the effect.
Questions this case answers
- How to estimate heterogeneous treatment effects when treated and control groups are unbalanced or differ in covariates
- How to estimate a causal effect when units choose their own treatment and only an encouragement can be randomized
- How to choose a meta-learner (X, R, DR) for estimating heterogeneous treatment effects?
- How to estimate the effect of a promotion when customers choose themselves whether to take part?
Tools and methods
- EconML, CausalML — libraries implementing all methods below
- X-learner — cross-imputes effects between groups, weights them by propensity score
- R-learner — cross-validated nuisance models, then minimizes a residual loss for the effect
- DR-learner — doubly robust pseudo-outcome averaged over three data parts
- DRIV learner — local average treatment effect with an instrument
- DDP and CTS trees — uplift trees for two or many treatments
Google or a one-time email code. No password.