Description
Combining an outcome model with a propensity model so the causal effect estimate is consistent if either model is correctly specified.
Doubly robust estimators such as AIPW combine regression adjustment (modelling the outcome) with inverse probability weighting (modelling treatment assignment). If either the outcome model or the propensity model is correct, the estimate is consistent, giving two chances to get it right. Modern versions use machine learning for both nuisance models with cross-fitting (double/debiased machine learning, Chernozhukov et al., 2018) to avoid overfitting bias, and allow heterogeneous effect estimation by segment. Rigour comes from a DAG-based covariate set, overlap checks and weight trimming, cross-fitting, and reporting the estimate alongside simpler estimators and a sensitivity analysis. It cannot fix unmeasured confounding.
Attributes
- Discipline
- Experimentation & Causal Inference
- Loss type
- Siloed
- Goal
- Prove causation
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Specify the causal question, time zero and DAG-based confounder set.
- Fit a propensity model and an outcome model, using flexible learners (e.g. gradient boosting) where appropriate.
- Use 5-fold cross-fitting so each unit's nuisance predictions come from models not trained on it (econml or causalml).
- Check overlap and trim or truncate propensity scores outside about 0.01 to 0.99.
- Compute the AIPW or DML estimate with robust or bootstrap confidence intervals.
- Compare with regression-only and weighting-only estimates, examine heterogeneity by segment, and add an E-value sensitivity analysis.
Sources
- causalgraphicalmodels
- dowhy
- causalml
- econml
- statsmodels
- Robins, J. M., Rotnitzky, A. & Zhao, L. P. (1994) Estimation of regression coefficients when some regressors are not always observed. JASA
- Chernozhukov, V. et al. (2018) Double/debiased machine learning for treatment and structural parameters. Econometrics Journal
- Hernán, M. A. & Robins, J. M. (2020) Causal Inference: What If
- The Inquiry Series, C5 — Causal Inference in Python — DAGs, Target Trial Emulation & Sensitivity Analysis
Also known as
- Augmented inverse probability weighting (AIPW)
- Doubly robust estimator
Categories
- Analytics & Behaviour
- Experimentation & Causal Inference