Causal Inference
calibratedDML
implements calibrated doubly robust estimators for mean
potential outcomes and treatment-versus-control contrasts. The
standard workflow fits or accepts cross-fitted nuisance
estimates, calibrates them, and then forms point estimates and
confidence intervals.
Overview
The standard implementation estimates arm means and treatment-versus-control contrasts. You can run the full procedure from raw data or start from cross-fitted nuisance estimates prepared elsewhere.
Primary method
The standard estimator is the main entry point for interval valid inference on arm means and treatment-versus-control contrasts.
CalibratedDML(...).fit(X, A, y)
calibrated_dml(data, outcome, treatment, covariates, ...)Calibration
Standard doubly robust estimators mainly protect point estimates, but valid confidence intervals often still depend on both nuisance functions being estimated well enough. This package adds a calibration step before debiasing so confidence intervals can remain valid under weaker nuisance-rate conditions.
In practice, calibration can also stabilize extreme inverse weights and reduce variance, especially when overlap is limited; see Stabilized Inverse Probability Weighting via Isotonic Calibration. This does not remove the population-level overlap assumptions required by the theory.
Step 1
Estimate outcome and treatment models
Step 2
Post-hoc one-dimensional maps align fitted scales
Step 3
Form the final calibrated estimator and interval
Outcome Regression Calibration
Calibration adjusts the outcome-regression scale so predicted values better match observed outcomes among comparable units.
Propensity Calibration
Calibration helps ensure that inverse-propensity weights are large only when they need to be. If calibrated propensities are still very small, that points to real limited overlap rather than avoidable estimation error.
In calibrated DML, these calibration steps are what connect the practical procedure to the inferential guarantees.
Installation
The repository provides both Python and R interfaces for the standard calibrated DML estimator.
Python
pip install calibratedDMLPython
pip install "git+https://github.com/Larsvanderlaan/calibratedDML.git"R
remotes::install_github("Larsvanderlaan/calibratedDML")Workflows
The package supports estimation from raw data, estimation from supplied nuisance functions, and multi-arm analyses.
Choose this path if the package should fit the nuisance models, perform calibration, and report inference.
Call: CalibratedDML(...).fit(...) or calibrated_dml(...)
Choose this path when cross-fitted nuisance estimates come from an external pipeline and only calibration, debiasing, and inference should be run here.
Call: fit_from_nuisances(...) or calibrated_dml_from_nuisances(...)
The standard estimator also handles treatments with more than two levels, returning arm means and contrasts relative to a specified control arm.
Call: standard calibrated DML with control_level
References
The main paper, repository, and related references are collected here.
Companion package
ppi_aipwSemisupervised mean inference with AIPW, calibration, and uncertainty quantification.
Companion package site