Struct DmlFit 

pub struct DmlFit { /* private fields */ }

A fitted DML model: estimates, standard errors and the per-repetition scores and nuisance predictions behind them.

Arrays indexed by treatment have length k = n_treat; per-repetition arrays have shape (k, n_rep).

Implementations

impl DmlFit

pub fn bootstrap( &self, method: BootstrapMethod, n_boot: usize, seed: u64, ) -> Result<Bootstrap>

Multiplier bootstrap with n_boot draws per repetition: t*_b = Σ_i ξ_bi · ψ_i / (J · n · se).

pub fn bootstrap_with_weights( &self, weights: &[Array2<f64>], ) -> Result<Bootstrap>

Bootstrap with caller-supplied weights, one (n_boot, n) matrix per repetition (e.g. to reproduce DoubleML’s draws exactly).

impl DmlFit

pub fn model_name(&self) -> &str

pub fn treatment_names(&self) -> &[String]

pub fn n_obs(&self) -> usize

pub fn n_treat(&self) -> usize

pub fn n_rep(&self) -> usize

pub fn n_folds(&self) -> usize

pub fn n_clusters(&self) -> Option<usize>

Number of clusters (two-way: the smaller dimension) for a clustered fit, None otherwise.

pub fn coef(&self) -> ArrayView1<'_, f64>

Point estimates θ̂ (median over repetitions).

pub fn se(&self) -> ArrayView1<'_, f64>

Standard errors (aggregated over repetitions).

pub fn t_stat(&self) -> Array1<f64>

pub fn p_value(&self) -> Array1<f64>

Two-sided p-values for H0: θ = 0 from the aggregated t-statistic, as DoubleML’s pval. (Multiple-testing adjustments in p_adjust instead take the median over repetitions.)

pub fn confint(&self, level: f64) -> Result<Array2<f64>>

Pointwise Wald confidence intervals at level, shape (k, 2) (median over repetitions of each repetition’s interval).

pub fn p_adjust(&self, method: PAdjust) -> Array1<f64>

Multiple-testing adjusted p-values across treatments (applied per repetition, then the median is taken). For the bootstrap-based Romano–Wolf procedure see Bootstrap::romano_wolf.

pub fn all_coef(&self) -> ArrayView2<'_, f64>

Estimates for every repetition, shape (k, n_rep).

pub fn all_se(&self) -> ArrayView2<'_, f64>

pub fn psi(&self) -> ArrayView3<'_, f64>

Score values ψ at θ̂, shape (n, k, n_rep).

pub fn psi_elements(&self) -> (ArrayView3<'_, f64>, ArrayView3<'_, f64>)

Score elements ψ_a and ψ_b (ψ = ψ_a·θ + ψ_b), shape (n, k, n_rep).

pub fn cate(&self, basis: ArrayView2<'_, f64>) -> Result<Blp>

Best linear predictor of the CATE on basis (shape (n, p), include a constant column for an intercept), with HC0 inference (Semenova and Chernozhukov, 2021), as DoubleML’s cate. Available for a single treatment of Plr (outcome residual on the basis scaled by the treatment residual), Irm with the ATE score and Apo (orthogonal signal on the basis).

pub fn gate(&self, groups: ArrayView2<'_, f64>) -> Result<Blp>

Group average treatment effects for mutually exclusive groups given as 0/1 dummy columns (shape (n, n_groups)), as DoubleML’s gate.

pub fn policy_tree( &self, features: ArrayView2<'_, f64>, depth: usize, min_samples_leaf: usize, ccp_alpha: f64, ) -> Result<PolicyTree>

Policy tree recommending treatment where the estimated effect is positive, as DoubleML’s policy_tree (for Irm with the ATE score, one treatment and one repetition). DoubleML’s defaults: depth = 2, min_samples_leaf = 8, ccp_alpha = 0.01.

pub fn nuisance_loss(&self, name: &str) -> Option<ArrayView2<'_, f64>>

Out-of-sample loss of a nuisance learner, shape (n_treat, n_rep): RMSE for regression learners, log loss for classifiers (DoubleML’s nuisance_loss). Computed on the rows where the nuisance applies.

pub fn nuisance_losses( &self, ) -> impl Iterator<Item = (&str, LossKind, ArrayView2<'_, f64>)>

(name, loss kind, losses) for every evaluated nuisance learner.

pub fn nuisance_targets(&self, name: &str) -> Option<ArrayView3<'_, f64>>

Targets of a nuisance learner, shape (n, n_rep, n_treat), with NaN where the nuisance does not apply (to evaluate custom metrics).

pub fn nuisance_names(&self) -> impl Iterator<Item = &str>

Names of the stored nuisance predictions (e.g. ml_l, ml_m).

pub fn predictions(&self, name: &str) -> Option<ArrayView3<'_, f64>>

All predictions of one nuisance, shape (n, n_rep, k).

pub fn prediction( &self, name: &str, rep: usize, treatment: usize, ) -> Option<Array1<f64>>

Out-of-fold predictions of one nuisance for a repetition and treatment.

impl DmlFit

pub fn concat(fits: &[DmlFit]) -> Result<DmlFit>

Stack fits of the same data and sample splits into one fit with all their estimates (DoubleML’s concat), for joint inference: the bootstrap, simultaneous confidence bands and linear combinations then account for the correlation between the estimates.

Predictions and learner losses are stacked by nuisance name (NaN where a fit has no such nuisance); sensitivity elements are kept if every fit has them.

impl DmlFit

pub fn sensitivity_analysis( &self, params: &SensitivityParams, ) -> Result<SensitivityResult>

Bounds on θ, their confidence bounds and robustness values under the confounding scenario params (PLR, IRM and DID models).

pub fn sensitivity_elements( &self, ) -> Option<(ArrayView2<'_, f64>, ArrayView2<'_, f64>)>

Sensitivity elements (σ², ν²), each of shape (n_treat, n_rep): the residual variance of the outcome and the variance of the Riesz representer. None for models without sensitivity analysis.

pub fn sensitivity_benchmark( &self, short: &DmlFit, ) -> Result<SensitivityBenchmark>

Calibrate a confounding scenario against observed covariates: short must be the same model refitted on the same sample splits without the benchmarking covariates (see sensitivity_benchmark on the models).

Trait Implementations

impl Clone for DmlFit

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for DmlFit

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Display for DmlFit

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.