Machine Learning-Augmented Synthetic Control Method
Also known as: ML-augmented SCM, augmented synthetic control, ASC, penalized synthetic control
The machine learning-augmented synthetic control method extends the classical synthetic control estimator by using penalized regression or other ML algorithms — such as lasso, ridge, or random forests — to construct the donor weights and to model pre-treatment outcome trajectories. The augmentation corrects for residual imbalance left by the standard weighting step, yielding lower bias when no perfect synthetic control exists.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use this method when you have a single treated unit (or a small number) observed over many time periods, a set of untreated donor units, and a clearly defined intervention with a known start date. It is especially valuable when the standard synthetic control does not achieve a good pre-treatment fit — a common situation with many donors or high-dimensional covariates. Do not use it when you have many treated units simultaneously (prefer DiD), when the pre-treatment period is very short (too little information to train the outcome model), or when donor units may have been partially affected by the intervention (spillovers).
Strengths & limitations
- Corrects the bias of the classical synthetic control when no perfect donor combination exists, through a principled ML-based augmentation term.
- Handles large donor pools where standard convex weighting breaks down, by applying penalization to select and weight donors automatically.
- Retains the interpretability of the synthetic control weighting step while adding the flexibility of a data-driven outcome model.
- Doubly robust: consistent if either the weight model or the outcome model is correctly specified.
- Compatible with standard placebo-based inference procedures already established for synthetic controls.
- Requires a sufficiently long pre-treatment time series to train the ML outcome model; short pre-periods yield unreliable augmentation.
- The bias correction can overfit if the ML model is too flexible relative to the number of pre-treatment periods, requiring careful cross-validation or regularization.
- Inference via placebo tests is only valid when enough donor units are available; with fewer than about 20 donors, permutation p-values are coarse.
- Interpretation of the augmentation component is less transparent than the original synthetic control weights, which can complicate communication to non-technical audiences.
- Like all synthetic control approaches, it is designed for settings with one or very few treated units; scaling to many treated units requires different methods.
Frequently asked
How does this differ from the classical synthetic control?
The classical synthetic control finds donor weights that minimize pre-treatment imbalance but cannot correct for any residual imbalance remaining after weighting. The augmented version adds an ML-estimated bias term that accounts for this residual, yielding lower bias when a perfect synthetic match does not exist — which is the common case in practice.
Which ML algorithm should I use for the outcome model?
Ridge regression is the default recommended by Ben-Michael et al. (2021) because it is fast, stable, and its regularization parameter can be selected by cross-validation on pre-treatment folds. Lasso or elastic net can be useful when many covariates must be screened. Nonparametric methods require long pre-treatment periods to avoid overfitting.
How do I perform inference?
The standard approach is a permutation (placebo) test: run the exact same procedure on each donor unit, treating it as the pseudo-treated unit. The p-value is the fraction of donor units whose post-treatment fit error is as large as or larger than the actual treated unit's. This requires at least 20 donors for informative p-values.
What if I have multiple treated units?
The synthetic control framework — augmented or not — is designed for one or very few treated units. With many treated units adopting treatment at the same or staggered times, difference-in-differences or staggered DiD estimators are more appropriate.
How long should the pre-treatment period be?
At least as long as the post-treatment period, and ideally several times longer. A short pre-treatment period limits the ML model's ability to learn the outcome relationship and makes the augmentation correction unreliable. As a rough guide, aim for at least 20 pre-treatment periods.
Sources
- Ben-Michael, E., Feller, A., & Rothstein, J. (2021). The augmented synthetic control method. Journal of the American Statistical Association, 116(536), 1789-1803. DOI: 10.1080/01621459.2021.1929245 ↗
- Abadie, A. (2021). Using synthetic controls: Feasibility, data requirements, and methodological aspects. Journal of Economic Literature, 59(2), 391-425. DOI: 10.1257/jel.20191450 ↗
How to cite this page
ScholarGate. (2026, June 3). Machine Learning-Augmented Synthetic Control Method. ScholarGate. https://scholargate.app/en/causal-inference/machine-learning-augmented-synthetic-control-method
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Causal Impact AnalysisCausal inference↔ compare
- Difference-in-DifferencesEconometrics↔ compare
- Machine learning-augmented difference-in-differencesCausal inference↔ compare
- Panel Data Synthetic Control MethodCausal inference↔ compare
- Synthetic Control MethodCausal inference↔ compare