Augmented inverse probability weighting
Estimate the average treatment effect by combining propensity adjustment with outcome prediction.
Concepts and reasoning
Two ways to reconstruct a missing outcome
An outcome regression predicts what each person would experience under each treatment. A propensity model instead adjusts for how people entered treatment groups. Augmented inverse probability weighting, AIPW, combines these approaches: start with predicted potential outcomes, then correct those predictions using weighted residuals from the outcomes actually observed.
Let m₁(X) predict the outcome under treatment and m₀(X) predict it under control. Their difference estimates a person's conditional contrast. But a fitted outcome model can be systematically wrong. The residual correction uses observed treated people to repair the treated prediction and observed controls to repair the control prediction.
The important result is double robustness: under causal identification assumptions, the ATE estimator can be consistent if either the propensity model is correct or both relevant outcome regressions are correct. This is an asymptotic property, not a guarantee that any finite data set produces an accurate answer. Both model families can be wrong.
Why the correction can work
If an outcome model is correct, its residual has conditional mean zero in its treatment group. The weighted correction then averages toward zero, leaving the correct standardized prediction. If the propensity score is correct, inverse weighting reconstructs the relevant outcome mean even when the prediction model is imperfect.
This logic still depends on having comparable treated and control observations. AIPW does not recover outcomes in regions with no treatment overlap and cannot remove an unmeasured confounder. The introduction to augmented inverse probability weighting develops the distinction between statistical model robustness and causal assumptions.
Mathematical foundations
The estimating score
For a continuous outcome or a binary outcome modeled on the probability scale, define
The estimate is the average of these scores:
Only one residual term is active for each person. A treated person contributes the treated residual; a control contributes the negative of the control residual. Both predicted potential outcomes are present for everyone.
For a correctly specified propensity score, conditional on X,
That equality explains the inverse-weighting correction. Under exchangeability, the observed conditional mean identifies the treated potential-outcome mean. The same reasoning applies to control.
With suitable regularity conditions, the centered score acts like an influence function and supports uncertainty estimation. Flexible machine learning may require cross-fitting and additional convergence conditions. “Doubly robust” alone is not sufficient justification for a particular standard-error formula.
Worked example
Worked example: individual corrections
Suppose four illustrative participants all have an estimated propensity of 0.5. Their fitted potential outcomes and observed outcomes are:
| Person | A | Observed Y | Predicted m₁ | Predicted m₀ | AIPW score |
|---|---|---|---|---|---|
| 1 | 1 | 10 | 8 | 5 | 7 |
| 2 | 1 | 8 | 9 | 6 | 1 |
| 3 | 0 | 4 | 7 | 5 | 4 |
| 4 | 0 | 7 | 10 | 6 | 2 |
For person 1, the predicted contrast is 8 − 5 = 3. The residual correction is (10 − 8)/0.5 = 4, so the score is 7. For person 3, the predicted contrast is 2 and the control correction is −(4 − 5)/0.5 = 2, giving 4.
The AIPW estimate is (7 + 1 + 4 + 2)/4 = 3.5 outcome units. Pure outcome standardization would average the predicted contrasts, (3 + 3 + 2 + 4)/4 = 3. The augmentation changes the estimate because observed outcomes reveal prediction errors.
Now change person 1's estimated propensity from 0.5 to 0.05. Their correction becomes 40, and their score becomes 43. Augmentation has not neutralized the overlap problem. A small denominator can give one residual enormous influence.
This four-person calculation illustrates the estimator, not a credible small-sample causal study. In a real analysis the models are estimated, their uncertainty matters, and outcomes must be compatible with the chosen model family.
Interpreting the results
Understand the two parts of the estimator
Augmented inverse-probability weighting combines predicted outcomes under each treatment with weighted corrections for observed prediction errors. The average prediction contrast supplies one route to the effect, while the residual correction uses treatment probabilities to compensate for outcome-model errors.
The estimated average treatment effect is a treatment-minus-control mean difference. For binary outcomes, it is a risk difference on the absolute probability scale. It should be interpreted for the target population represented by the supplied observations and adjustment assumptions.
Double robustness is a consistency property: under suitable conditions, one correctly specified nuisance model can suffice even if the other is misspecified. It does not mean both models may be wrong, that small-sample bias disappears, or that confidence intervals are automatically valid under every model failure. Poor overlap can make the correction highly unstable despite a plausible outcome model.
Assumptions and limitations
Two statistical models do not replace causal identification
Conditional exchangeability, positivity, consistency, and correct time alignment remain necessary for a causal interpretation. Both nuisance models rely on the measured confounders; neither corrects an unrecorded common cause of treatment and outcome simply by being present.
Outcome models need support for predicting both potential outcomes across the target population. Extrapolation to covariate profiles absent in one arm may dominate the answer. Propensity estimates close to zero or one amplify residuals and can give a few observations excessive influence.
Simple parametric fitting also assumes appropriate functional forms and independent observations. Flexible nuisance estimation often needs sample splitting, cross-fitting, and suitable uncertainty theory. Those refinements are separate methodological choices, not automatic consequences of using an augmented estimator.
Diagnostics and sensitivity analysis
Examine both nuisance models and their interaction
Review propensity overlap, covariate balance, and the concentration of inverse-probability weights. Then inspect outcome-model fit within each arm, including systematic residual patterns and predictions in poorly represented regions. Good balance alone cannot establish that both nuisance models are adequate.
Compare plausible prespecified covariate forms and, where appropriate, simpler weighting or outcome-regression estimates. Disagreement can identify sensitivity to extrapolation or model choices; agreement does not rule out a shared confounding bias.
Check bootstrap convergence and extreme estimates. If many resamples fail, the data may not support the model complexity. Report the estimand, nuisance specifications, interval procedure, and overlap limitations. A doubly robust label is not a substitute for these details or for sensitivity to unmeasured confounding.
References and other material
Read the methodological references and supporting tutorials below. The examples in this lesson are original teaching calculations, not findings from these sources.
- Augmented inverse probability weighting and double robustness — A free introduction to combining propensity and outcome models.