Inverse Probability of Treatment Weighting (IPW / IPTW)
Also known as: IPW, IPTW, inverse probability of treatment weighting, marginal structural model weighting, Ters Olasılık Ağırlıklandırma (IPW / IPTW)
Inverse Probability Weighting is a causal-inference method that assigns each observation a weight equal to the inverse of its probability of receiving the treatment it actually received. Introduced by Robins, Hernán and Brumback (2000) for marginal structural models, it builds a pseudo-population in which treatment is independent of measured confounders, balancing selection bias.
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.
+55 more
When to use it
Use IPW for observational causal questions where you want to estimate a treatment effect free of confounding by measured covariates, with at least about 100 observations. It applies to binary, continuous (generalized IPW), or time-varying treatments and works best when the conditional independence (no unmeasured confounding) and common support (positivity) assumptions hold. It is less suitable when the propensity model is poorly specified or when some units have near-zero probability of treatment, which produces extreme weights.
Strengths & limitations
- Removes confounding by measured covariates by building a balanced pseudo-population.
- Handles binary, continuous (generalized IPW), and time-varying treatments, making it the natural fit for marginal structural models.
- Can be combined with an outcome model into a doubly robust estimator that stays consistent if either model is correct.
- Requires the conditional independence and common support (positivity) assumptions, which cannot be fully verified from data.
- Extreme weights from units with near-zero treatment probability inflate the variance; trimming helps but only partly.
- Misspecification of the propensity model biases the weights and therefore the estimated effect.
Frequently asked
What is a propensity score in IPW?
It is e(X), the estimated probability of receiving treatment given the measured covariates, usually from a logistic regression. IPW uses the inverse of this probability as the weight for each observation.
Why do extreme weights matter?
When a unit had a probability of treatment close to 0 or 1, its inverse weight becomes very large and can dominate the estimate, inflating the variance. Trimming or stabilising the weights mitigates this, but severe positivity violations cannot be fully rescued.
What is the difference between IPW and propensity score matching?
Both use the propensity score to address confounding, but IPW re-weights every observation to form a pseudo-population, while matching pairs treated and untreated units with similar scores. For small samples or unstable propensity estimates, matching is often a safer alternative.
How does IPW relate to doubly robust estimation?
IPW can be combined with an outcome regression to form an augmented (doubly robust) estimator, which stays consistent as long as either the propensity model or the outcome model is correctly specified — a useful safeguard against misspecification.
Sources
- Robins, J. M., Hernán, M. A., & Brumback, B. (2000). Marginal Structural Models and Causal Inference in Epidemiology. Epidemiology, 11(5), 550-560. DOI: 10.1097/00001648-200009000-00011 ↗
- Cole, S. R., & Hernán, M. A. (2008). Constructing Inverse Probability Weights for Marginal Structural Models. American Journal of Epidemiology, 168(6), 656-664. DOI: 10.1093/aje/kwn164 ↗
How to cite this page
ScholarGate. (2026, June 1). Inverse Probability of Treatment Weighting (IPW / IPTW). ScholarGate. https://scholargate.app/en/causal-inference/inverse-probability-weighting
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 Mediation AnalysisCausal inference↔ compare
- DAG Causal IdentificationCausal inference↔ compare
- Doubly Robust EstimationCausal inference↔ compare
- Logistic RegressionResearch Statistics↔ compare
- Propensity Score MatchingResearch Statistics↔ compare