double machine learning (DML)

You would love to use a flexible learner to soak up dozens of confounders, but plugging a regularized model straight into an effect estimate poisons it: the model's bias, designed to help prediction, leaks into the causal number. Double machine learning is the discipline that lets you keep the flexible learners yet recover an effect estimate that is unbiased to first order and converges at the usual square-root-n rate with valid confidence intervals.

Two ideas do the work. First, Neyman orthogonality: the estimating equation is reshaped so its sensitivity to small errors in the nuisance functions (the outcome model and the propensity model) vanishes to first order, so moderately inaccurate learners do not move the effect. In the partially linear model this appears as residual-on-residual regression, predicting treatment and outcome from covariates and relating what each learner could not explain. Second, cross-fitting: nuisances are fit on one data fold and the effect evaluated on another, removing the overfitting bias that would otherwise creep back in.

The payoff is a principled bridge between modern prediction and classical inference: any sufficiently accurate learner can serve as a nuisance estimator, while the target parameter retains parametric-rate guarantees. The assumptions remain causal, not magical, unconfoundedness, overlap, and nuisance learners that converge fast enough (faster than the fourth root of n).