Double Machine Learning (DML)

Double machine learning (DML) is a causal inference method that uses machine learning to model the outcome and treatment from covariates, then estimates the treatment effect from the residuals using orthogonal scores and cross-fitting.

이중 머신러닝(DML)은 머신러닝으로 공변량에서 결과와 처치를 각각 모형화한 뒤 직교 점수와 교차적합을 이용해 잔차로부터 처치효과를 추정하는 인과추론 방법이다.

How do I estimate a treatment effect while controlling for many covariates with machine learning?

많은 공변량을 머신러닝으로 통제하면서 처치효과를 어떻게 추정하나요?

Double machine learning predicts both the outcome and the treatment from high-dimensional controls with flexible learners and estimates the effect from the remaining variation, which keeps the estimate approximately unbiased with valid confidence intervals. It fits observational data with many potential confounders, provided all relevant confounders are measured.

When to use Double Machine Learning (DML)

• Many covariates or nonlinear confounding in observational data
• Estimating an average treatment effect with valid confidence intervals
• Binary or continuous treatments with high-dimensional controls

When to use something else

• Important confounders are unmeasured: use instrumental variable methods or difference-in-differences
• Effect heterogeneity across individuals is the focus: use a causal forest

Assumptions and requirements

• Unconfoundedness: all confounders are observed
• Overlap: every unit has a chance of each treatment level
• Nuisance models that predict reasonably well
• Cross-fitting to avoid overfitting bias

What the analysis reports

• Treatment effect estimate with standard error and confidence interval
• Performance of the outcome and treatment nuisance models
• Learner choices and number of folds
• Sensitivity across learners

How to report it

Double machine learning with [learner] nuisance models and [k]-fold cross-fitting estimated an average treatment effect of [value] (SE = [value], 95% CI [[lower], [upper]]).

Related methods

• Instrumental Variable Forest
• Propensity Score Matching
• Causal Forest

Run Double Machine Learning (DML) on Quorant