Policy Evaluation Doubly Robust Estimation
Doubly Robust Estimation for Policy Evaluation · Also known as: DR estimation for policy, augmented IPW for policy evaluation, AIPW policy evaluation, doubly robust policy analysis
Policy Evaluation Doubly Robust Estimation applies the doubly robust (DR) estimator to assess the causal effect of a public policy or programme. It combines a model of treatment assignment (propensity score) with a model of the outcome, and requires only one of the two models to be correctly specified to produce a consistent estimate of the average treatment effect, making it a resilient tool for programme evaluation.
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 evaluating a policy or programme with observational data where treatment is not randomly assigned and you are uncertain whether the propensity score model or the outcome model is correctly specified. It is appropriate when the outcome is continuous or binary, sample sizes are at least moderate (n > 100 for stable weighting), and the positivity assumption holds (all units have non-zero probability of receiving treatment). It is preferable to plain IPW when outcome model information is available. Do not use it when there is near-violation of positivity (propensity scores near 0 or 1), when the policy has universal coverage (no untreated comparison group), or when sample sizes are very small, as extreme weights can destabilize the estimator.
Strengths & limitations
- Doubly robust: consistent if either the propensity model or the outcome model is correctly specified, not both.
- Semiparametrically efficient when both models are correct, yielding the lowest possible asymptotic variance in its class.
- Compatible with flexible machine learning models for nuisance estimation, enabling data-adaptive bias reduction.
- Directly applicable to standard policy evaluation targets: ATE, ATT, and population subgroup effects.
- Residual correction terms reduce mean squared error relative to pure outcome regression when outcome model is imperfect.
- Allows transparent reporting of both component models, making the evaluation audit trail clear for policy audiences.
- Requires the positivity assumption: every unit must have a non-zero probability of treatment; near-violations cause extreme weights and inflated variance.
- Double robustness does not protect against violations of the unconfoundedness (no unmeasured confounding) assumption; omitted confounders still bias the estimate.
- When both models are misspecified, consistency is lost and the estimator may perform worse than simpler alternatives.
- More complex to implement and explain to non-technical policy audiences than regression adjustment or simple matching.
- Extreme propensity score weights can produce numerically unstable estimates and inflated standard errors in finite samples.
Frequently asked
What does 'doubly robust' actually mean?
It means the estimator is consistent if either the propensity score model or the outcome regression model is correctly specified — but not necessarily both. This gives two chances to get the estimation right, which is valuable when model uncertainty is high.
How is this different from plain inverse probability weighting?
Plain IPW uses only the propensity model; if that model is misspecified, the estimate is inconsistent. Doubly robust estimation adds an outcome model correction that provides a second layer of protection. It also tends to have lower variance than IPW when the outcome model is informative.
Can I use machine learning models to estimate the nuisance functions?
Yes, but you should use cross-fitting (sample splitting): fit the nuisance models on one part of the data and compute the DR estimator on the other, then average across folds. This prevents bias from overfitting and validates the use of flexible learners.
What is the positivity assumption and why does it matter?
Positivity requires that every unit in the study population has a non-zero probability of both receiving and not receiving treatment. Without it, some units cannot be compared, and inverse probability weights become undefined or extreme, destabilizing the estimator.
Is doubly robust estimation the same as TMLE?
They are related but distinct. Both use outcome and propensity models and achieve double robustness and semiparametric efficiency. TMLE (targeted maximum likelihood estimation) uses a targeting step that updates the outcome model to directly optimize for the causal estimand, which can offer better finite-sample performance. Standard AIPW and TMLE are complementary approaches in the doubly robust family.
Sources
- Bang, H., & Robins, J. M. (2005). Doubly robust estimation in missing data and causal inference models. Biometrics, 61(4), 962-973. DOI: 10.1111/j.1541-0420.2005.00377.x ↗
- Robins, J. M., Rotnitzky, A., & Zhao, L. P. (1994). Estimation of regression coefficients when some regressors are not always observed. Journal of the American Statistical Association, 89(427), 846-866. DOI: 10.1080/01621459.1994.10476818 ↗
How to cite this page
ScholarGate. (2026, June 3). Doubly Robust Estimation for Policy Evaluation. ScholarGate. https://scholargate.app/en/causal-inference/policy-evaluation-doubly-robust-estimation
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.
- Doubly Robust EstimationCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare
- Policy Evaluation Propensity Score MatchingCausal inference↔ compare
- Propensity Score WeightingCausal inference↔ compare