Targeted Maximum Likelihood Estimation (TMLE)
Also known as: Targeted Learning, TMLE, Targeted MLE, Hedeflenmiş Maksimum Olabilirlik Tahmini
Targeted Maximum Likelihood Estimation (TMLE) is a semiparametric, doubly robust causal inference method introduced by Mark van der Laan and Daniel Rubin in 2006. It combines flexible machine learning models for both the outcome and the treatment assignment mechanism, then applies a targeting step that re-fits the initial outcome model specifically to reduce bias for a pre-specified causal estimand such as the average treatment effect. TMLE is widely used in epidemiology, biostatistics, and health economics when estimating causal effects from observational data.
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 TMLE when estimating causal effects — average treatment effects, dose-response curves, or policy-relevant contrasts — from observational data with many confounders. It is appropriate when you want to leverage machine learning for nuisance parameter estimation without sacrificing valid inference. TMLE requires the positivity assumption (every unit has nonzero probability of each treatment level), no unmeasured confounding, and a well-defined intervention. When positivity is near-violated, near-violations inflate variance. Alternatives include double/debiased machine learning (DML) and augmented inverse probability weighting (AIPW), which share the double-robustness property.
Strengths & limitations
- Doubly robust: consistent if either the outcome model or the propensity score model is correctly specified, providing a safety net against model misspecification
- Compatible with any machine learning or ensemble method for nuisance parameter estimation, enabling data-adaptive confounding adjustment without imposing parametric restrictions
- Provides asymptotically valid standard errors and confidence intervals through the efficient influence function, unlike naive plug-in ML estimators
- Targets estimation error for the specific causal estimand of interest rather than minimizing overall prediction error, leading to reduced bias for the quantity that matters
- Positivity violations — regions of the covariate space where the propensity score is near zero or one — can inflate variance and destabilize the estimator substantially
- Requires correct specification or good approximation of both the outcome model and the propensity score model; if both are misspecified, double robustness does not save the estimate
- The targeting step assumes a particular parametric fluctuation form (logistic submodel), which may not be optimal for all outcome types without adaptation
- Implementation complexity is higher than simple regression adjustment; practitioners must manage nuisance model selection, cross-fitting, and influence-function inference carefully
Frequently asked
How does TMLE differ from inverse probability weighting (IPW)?
IPW reweights observations by the inverse of the propensity score without incorporating an outcome model, making it sensitive to extreme weights and propensity score misspecification. TMLE augments an outcome model estimate with a propensity-score-based correction, yielding double robustness and typically lower variance than IPW alone, especially when the outcome model is informative.
What is the role of Super Learner in TMLE?
Super Learner is a cross-validated ensemble that combines many candidate learners to minimize prediction error for a given loss function. In TMLE, it is used to estimate both the outcome model and the propensity score, maximizing the chance that at least one nuisance parameter is well-approximated and that the double-robustness property can be leveraged in practice.
Does TMLE require large samples to work well?
TMLE's theoretical guarantees are asymptotic, so performance in small samples depends on the complexity of the nuisance models used. In small to moderate samples, using parsimonious learners or parametric models for nuisance estimation is advisable. Cross-fitting and regularization help, but analysts should report honest uncertainty and consider sensitivity analyses when sample sizes are limited.
Sources
- van der Laan, M. J., & Rubin, D. (2006). Targeted maximum likelihood learning. The International Journal of Biostatistics, 2(1). DOI: 10.2202/1557-4679.1043 ↗
How to cite this page
ScholarGate. (2026, June 2). Targeted Maximum Likelihood Estimation (TMLE). ScholarGate. https://scholargate.app/en/causal-inference/targeted-maximum-likelihood
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.
- Double Machine LearningCausal inference↔ compare
- Doubly Robust EstimationCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare