Problem
An analyst has observational data and has to estimate how much changing one variable, the treatment, changes an outcome. There is no randomized experiment, and there are many candidate control variables. Plain multiple regression gives a biased treatment coefficient when controls are chosen badly. ML models predict the outcome well but do not recover the structure of the effect: LASSO gives a good forecast without guaranteeing that its coefficient is close to the true one.
Example from the article: how a country's trade openness (trade/GDP) affects deviations from uncovered interest rate parity, with a small sample and many candidate macro controls.
Solution in brief
Model the outcome as a linear effect of the treatment plus an unknown function of the controls, and model the treatment as an unknown function of the same controls. Fit both unknown functions with any ML method, then estimate the treatment coefficient from the residuals using an orthogonalized score. This is double machine learning (DML). The orthogonality condition makes the estimate insensitive to errors in the ML fits.
Questions this case answers
- How to estimate the effect of a continuous treatment on an outcome from observational data with many candidate controls, without biasing the treatment coefficient when machine learning models are used for the nuisance parts
- How to choose control variables with a causal graph so that adjustment removes confounding without opening collider bias
- How to estimate the causal effect of a treatment from observational data with many potential confounders
- How to choose control variables for a causal regression using a causal graph
Tools and methods
- Double machine learning (DML) — treatment coefficient with ML nuisance fits.
- Neyman orthogonality, sample splitting — remove regularization bias.
- Partially linear model — the two-equation setup.
- DAG and backdoor criterion — choose controls in low dimensions.
- DoWhy, CausalML, causal-learn, EconML, CausalImpact — packages the author names.
Google or a one-time email code. No password.