Description
Estimates a causal effect from observational data with many confounders by using flexible ML to partial out confounding, with valid inference via cross-fitting.
With rich observational data, analysts often regress outcomes on treatment plus many controls, or use ML predictions naively, both of which bias estimates. DML fits ML models for the outcome given covariates and for the treatment given covariates, then regresses the residualised outcome on the residualised treatment. Neyman orthogonality makes the estimate insensitive to small errors in the ML models, and cross-fitting avoids overfitting bias, so standard errors are valid. It still relies on no unmeasured confounding, so it pairs with DAG reasoning and sensitivity analysis. For UX analytics with hundreds of behavioural features, it is a principled upgrade on 'controlled for everything' regressions.
Attributes
- Discipline
- Experimentation & Causal Inference
- Loss type
- Unmigrated
- Goal
- Prove causation
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Weeks
Steps
- Draw a DAG and choose covariates that block backdoor paths; exclude mediators and colliders.
- Split data into K folds (typically 5) for cross-fitting.
- In each fold, fit flexible models (e.g. gradient boosting) for outcome and treatment on covariates using the other folds.
- Compute residuals and regress outcome residuals on treatment residuals to estimate the effect.
- Aggregate across folds and compute robust standard errors.
- Run sensitivity analysis for unmeasured confounding and compare with simpler estimators.
Sources
- Chernozhukov et al., Double/debiased machine learning for treatment and structural parameters (2018), Econometrics Journal
Also known as
- DML
- Debiased machine learning
- Orthogonalised regression
Categories
- Analytics & Behaviour
- Experimentation & Causal Inference