Machine Learning-Augmented Propensity Score Matching
Machine Learning-Augmented Propensity Score Matching Estimator · Also known as: ML-PSM, boosted propensity score matching, ML-augmented PSM, nonparametric propensity score matching
Machine learning-augmented propensity score matching (ML-PSM) replaces the traditional logistic regression used to estimate propensity scores with flexible machine learning algorithms — such as gradient boosted trees, random forests, or LASSO — to better capture complex, nonlinear relationships among covariates. The resulting richer propensity scores improve covariate balance and reduce bias in the estimated average treatment effect on the treated (ATT).
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 ML-PSM when you have observational data with a binary treatment and a large covariate set whose relationships with treatment assignment are likely nonlinear or involve higher-order interactions that a logistic regression would miss. It is most valuable with at least several hundred observations per arm and at least 10–15 potential confounders. Do not use it as a substitute for domain knowledge about confounders — ML cannot recover causal identification if key confounders are unmeasured. Avoid it when the sample is very small (the ML model will overfit the propensity score) or when logistic regression already achieves excellent balance, in which case the added complexity is unnecessary.
Strengths & limitations
- Captures nonlinear and interaction effects among covariates without requiring the researcher to specify the functional form, reducing propensity score model misspecification.
- Often achieves better covariate balance than logistic-regression-based PSM in high-dimensional or complex covariate settings.
- Widely available implementations in R (twang, MatchIt with glmnet) and Python (sklearn + causalinference) make adoption feasible.
- Compatible with any downstream matching algorithm (nearest-neighbour, kernel, radius) and with doubly robust estimators as a natural extension.
- Reduces sensitivity to analyst degrees of freedom in model specification, since ML model selection can be guided by cross-validated balance criteria.
- ML propensity score models require tuning hyperparameters (tree depth, regularisation strength, number of trees); poor tuning can lead to over- or under-smoothed scores.
- Bootstrap inference that repeats the entire ML estimation is computationally expensive for large datasets.
- Cannot address unmeasured confounding: if important covariates are missing from X, no ML technique will rescue the causal identification.
- Interpretability of the propensity score model is reduced relative to logistic regression, which may be a concern for regulatory or replication contexts.
- Overfitting to the training data can occur in small samples, producing propensity scores that are extreme (near 0 or 1) and cause poor overlap.
Frequently asked
Which ML algorithm works best for propensity score estimation?
Gradient boosted trees (e.g., the TWANG approach by McCaffrey et al.) and LASSO logistic regression are the most validated options. Gradient boosting tends to excel when there are complex nonlinearities; LASSO is preferable when interactions are sparse. In practice, choose based on cross-validated balance criteria rather than predictive accuracy alone.
Does ML-PSM solve the problem of unmeasured confounding?
No. Like all propensity score methods, ML-PSM requires the unconfoundedness (ignorability) assumption: all variables that jointly determine treatment and outcome must be included in X. ML only improves how well the measured confounders are handled; it cannot substitute for missing data.
How should I report uncertainty for ML-PSM estimates?
Use a bootstrap that re-runs the full pipeline — ML model training, propensity score estimation, matching, and outcome estimation — on each bootstrap sample. Standard analytic variance formulas that ignore the first stage will understate uncertainty.
When should I prefer doubly robust estimation over pure ML-PSM?
Doubly robust methods (such as augmented IPW or targeted learning) add an outcome model as a second layer of protection: the ATT estimator is consistent if either the propensity or the outcome model is correctly specified. If you expect either model might be partially misspecified, doubly robust estimation offers a safety net that plain ML-PSM does not.
Is a large sample required?
ML methods for propensity score estimation are most reliable with at least a few hundred treated observations and a comparable or larger control group. In small samples, the ML model is prone to overfit the training data, producing propensity scores that are too extreme, which degrades matching quality. With very small samples, a well-specified logistic regression is usually safer.
Sources
- McCaffrey, D. F., Ridgeway, G., & Morral, A. R. (2004). Propensity score estimation with boosted regression for evaluating causal effects in observational studies. Psychological Methods, 9(4), 403-425. DOI: 10.1037/1082-989X.9.4.403 ↗
- Westreich, D., Lessler, J., & Funk, M. J. (2010). Propensity score estimation: neural networks, support vector machines, decision trees (CART), and meta-classifiers as alternatives to logistic regression. Journal of Clinical Epidemiology, 63(8), 826-833. DOI: 10.1016/j.jclinepi.2009.11.020 ↗
How to cite this page
ScholarGate. (2026, June 3). Machine Learning-Augmented Propensity Score Matching Estimator. ScholarGate. https://scholargate.app/en/causal-inference/machine-learning-augmented-propensity-score-matching
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.
- Coarsened Exact MatchingCausal inference↔ compare
- Doubly Robust EstimationCausal inference↔ compare
- Entropy BalancingCausal inference↔ compare
- Machine learning-augmented doubly robust estimationCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare
- Propensity Score WeightingCausal inference↔ compare