dml-core

Double / debiased machine learning (Chernozhukov et al., 2018) for causal inference in Rust, modelled on and numerically validated against the DoubleML package.

• Models: partially linear regression (Plr, partialling-out and IV-type scores), interactive regression model (Irm, ATE / ATTE), partially linear IV (Pliv), interactive IV model (Iivm, LATE), difference-in-differences for panels (Did) and repeated cross-sections (DidCs)
• Inference: DML2 estimation, standard errors, confidence intervals, multiple treatments, repeated sample splitting, multiplier bootstrap, joint intervals, Romano–Wolf / Holm / Bonferroni
• Heterogeneous effects: S-, T-, X-, R- and DR-learners; best linear predictor of the CATE and group average treatment effects with robust inference
• Learners: RandomForestRegressor / RandomForestClassifier, Ols (ridge), Lasso, LogisticRegression, or your own via the Regressor / Classifier traits
• Fast: pure Rust, no BLAS; folds fitted in parallel with rayon (parallel feature, default)

cargo add dml-core

use dml_core::{DmlData, Lasso, Plr};

let data = DmlData::new(y, d, x)?;
let fit = Plr::new(Lasso::new(0.01), Lasso::new(0.01)) // learners for E[Y|X] and E[D|X]
    .n_folds(5)
    .n_rep(10)
    .seed(42)
    .fit(&data)?;
println!("{fit}");

Every model is tested for parity with Python DoubleML on identical data and sample splits (agreement to 1e-9). Python bindings, benchmarks and the full documentation live in the repository.

License

Licensed under either of Apache License 2.0 or MIT license, at your option.