Workflows
This page gives a concise map of the standard calibrated DML workflows: fitting from raw data, starting from supplied nuisance estimates, and multi-arm analyses. For tested end-to-end walkthroughs, use the Tutorials page, which links to Python notebooks, script mirrors, and R vignettes.
Workflows
Each card gives when to use it and the corresponding Python and R entry points.
Standard workflow
Use when the package should fit nuisance models, calibrate them, and report inference from raw data.
Python
CalibratedDML(...).fit(X, A, y)
R
calibrated_dml(data, outcome, treatment, covariates, ...)
Supplied nuisances
Use when cross-fitted nuisance estimates are already available and only calibration, debiasing, and inference should be run here.
Python
CalibratedDML(...).fit_from_nuisances(A, y, mu_mat, pi_mat, ...)
R
calibrated_dml_from_nuisances(A, Y, mu_mat, pi_mat, ...)
Multi-arm treatment
Use when treatment has more than two levels and the estimand is arm means plus treatment-versus-control contrasts.
Python
CalibratedDML(control_level=0, ...)
R
calibrated_dml(..., control_level = 0)
Custom learners
Use when you want to choose the nuisance learners yourself or bring in nuisance estimates from another pipeline.
Python
Pass sklearn-compatible regressors for outcome nuisance estimation and probabilistic classifiers for treatment nuisance estimation.
CalibratedDML(outcome_model=..., treatment_model=...).fit(X, A, y)
R
Use built-in learner names, a
SuperLearner library, or an
sl3 learner.
calibrated_dml(..., outcome_model = ..., treatment_model = ...)
For full control in either interface, use the supplied-nuisance
workflow with fit_from_nuisances(...) or
calibrated_dml_from_nuisances(...). R-side learner
support is documented on the
R package page.
Practical Notes
Supplied nuisance estimates should usually be cross-fitted. The inference layer should also be chosen deliberately.
Supplied nuisances
When you use the direct nuisance path, the expected input is a set of cross-fitted nuisance estimates rather than in-sample fits.
Inference choice
Standard calibrated DML uses jackknife intervals by default. They adapt to nuisance misspecification, are cheaper than the bootstrap, and worked better in our simulations. Use Wald only when both nuisances are consistent, even if one converges arbitrarily slowly.
Custom learners
Custom learners are interface-specific: Python accepts
sklearn-compatible estimators, while R accepts built-in model
specs plus optional SuperLearner and
sl3 backends.
Tutorials
The Tutorials page collects the tested Python notebooks, script mirrors, and R vignettes for the standard calibrated DML workflows when you want more than the concise workflow summaries on this page.