Part 3 of a series on causal inference moves from average treatment effects to personalized (CATE) estimation. It covers meta-learners (S-, T-, X-, and R-Learners), Double Machine Learning and Robinson residualisation, Wager and Athey's causal forests (with grf and EconML's CausalForestDML as standard implementations), and uplift modelling for CRM, retention, and marketing use cases, including the persuadables/sure-things/lost-causes/sleeping-dogs framing and Qini curve evaluation. Operational lessons stress protecting randomised holdouts and expecting effect drift over time.
Table of contents
The target: CATEMeta-learnersCausal forestsUplift modelling: CATE in production🔒 Paid-subscriber sectionFinal thoughtsQuestions this post answers
What is the difference between an S-Learner and a T-Learner for estimating heterogeneous treatment effects?
An S-Learner trains a single outcome model with treatment as a feature, making it simple and data-efficient but prone to shrinking small treatment effects toward zero due to regularization. A T-Learner instead trains a separate model per treatment arm, capturing arm-specific structure better but risking artifacts when the arms have unequal sample sizes and get regularized differently. daily.dev surfaces practical comparisons like this for teams choosing a causal ML modeling approach.
How does the X-Learner handle imbalanced treatment and control groups in causal effect estimation?
The X-Learner imputes missing individual treatment effects by first fitting T-Learner outcome models, then computing imputed effects for treated units using the control model's prediction (and vice versa for control units), and finally combining estimates weighted by the propensity score. This extracts more information from a small treatment arm, such as a 5% treated versus 95% control split, than a plain T-Learner. Engineers building targeting models on imbalanced data can track methods like this via daily.dev.
What are the four customer segments used in uplift modelling and why do sleeping dogs matter?
Uplift modelling divides an audience into persuadables (respond only if treated), sure things (respond regardless), lost causes (never respond), and sleeping dogs (respond negatively to treatment). Sleeping dogs matter because targeting them wastes budget and actively harms outcomes, such as a renewal reminder prompting cancellation or a win-back email triggering an unsubscribe. Marketing and retention teams refining uplift targeting can follow this kind of practical framing on daily.dev.