conditional average treatment effect (CATE)

A drug that helps the young and harms the old has a misleading average. The conditional average treatment effect refuses to average everyone together; it reports the treatment effect as a function of who you are, so the same intervention can read as strongly beneficial for one profile and useless for another. It is the natural target whenever the point is to decide whom to treat rather than simply whether a treatment works on average.

Formally the CATE is the function tau(x) = E[Y(1) - Y(0) given X = x], the ATE restricted to the subpopulation with covariates x. Estimating it well is a machine-learning problem because tau(x) can be smooth and high-dimensional. Meta-learners decompose it into standard regressions: the S-learner fits one model with treatment as a feature, the T-learner fits separate models per arm, and the X-learner and R-learner use residualized, orthogonalized targets that are more robust. Causal forests adapt random forests to estimate tau(x) with valid confidence intervals.

The catch is that CATE estimation inherits all the identification assumptions of the ATE (unconfoundedness, overlap) and adds a statistical hazard: effect heterogeneity is a faint signal sitting on top of strong outcome variation, so naive models overfit noise as if it were structure. Honest estimation, sample splitting, and regularization are what keep heterogeneity claims credible.