The Double-Robustness Property of AIPW

Why Double Robustness Matters

The augmented inverse-probability weighted estimator (AIPW) combines two ways to adjust for observed covariates: outcome regression and inverse-probability weighting. Its double-robustness property is a precise consistency result: under the causal assumptions stated below, the estimator converges to the average treatment effect if the propensity-score model is correct or if the outcome-regression models are correct. Both need not be correct.

This is a statement about the estimator’s large-sample target, not a claim that AIPW automatically detects model error or works under every form of bias. In particular, double robustness does not make unmeasured confounding disappear, repair a failure of positivity, or ensure stable estimates in a small sample.

Setup and Assumptions

Let \(D\) be a binary treatment, with \(D=1\) for treated and \(D=0\) for untreated. Let \(Y\) be the observed outcome, and let \(X\) denote pretreatment covariates. Write \(Y(1)\) and \(Y(0)\) for the potential outcomes under treatment and control. The target is the average treatment effect (ATE), \(\tau=E[Y(1)-Y(0)]\).

The causal interpretation requires consistency of the observed outcome with the potential outcome under the received treatment, no interference between units, conditional exchangeability \((Y(1),Y(0))\perp D\mid X\), and positivity \(0<e_0(X)<1\) almost surely in the target population. We also require the relevant expectations to exist. Positivity says that each covariate profile in the target population has a nonzero chance of receiving either treatment.

The fitted propensity and outcome regressions are denoted \(\hat e(X)\), \(\hat m_1(X)\), and \(\hat m_0(X)\). The AIPW estimator is

The first two terms compare predicted outcomes for the two treatment states. The weighted residual terms correct those predictions using observed outcomes. To state a consistency result, assume the fitted nuisance functions converge to limits \(e_*(X)\), \(m_{1*}(X)\), and \(m_{0*}(X)\). Require the limiting propensity to be bounded away from zero and one, and require sufficient moment and estimation regularity for the sample average to converge to the expectation of the limiting AIPW score. This last condition holds, for example, when nuisance estimates converge appropriately and their use in the sample average satisfies a law of large numbers. Correct specification means convergence to the corresponding true function, not merely that a model has been fitted.

The Population Argument

For fixed candidate functions \(q(X)\), \(r_1(X)\), and \(r_0(X)\), define the population AIPW functional

By consistency and conditional exchangeability, \(E[Y\mid D=a,X]=m_{a0}(X)\). Taking the conditional expectation of the score given \(X\) therefore gives

This identity contains both routes to the ATE. If \(q=e_0\), the coefficient of each outcome-regression residual is one, so the conditional expectation simplifies to \(m_{10}(X)-m_{00}(X)\), regardless of the limits \(r_1,r_0\). If instead \(r_1=m_{10}\) and \(r_0=m_{00}\), both residuals are zero, regardless of \(q\). Taking expectations then gives \(E[m_{10}(X)-m_{00}(X)]=\tau\).

For the propensity-correct route, the algebra is explicit:

For the outcome-correct route, substituting the true regressions makes both residual terms vanish:

Finally, averaging the resulting conditional expectation over \(X\) gives the ATE. The convergence assumptions ensure that the empirical average of the fitted scores approaches the corresponding population functional. Thus this is a consistency argument, not a guarantee that every finite-sample estimate equals or closely approximates the target.

What the Guarantee Does and Does Not Say

Outcome regression alone depends on correctly estimating both conditional means. Inverse-probability weighting alone depends on correctly estimating the propensity score. AIPW adds residual corrections to the outcome-regression contrast, giving it a consistency route through either nuisance specification. Which route applies is not identified merely by fitting the estimator: a plausible estimate cannot establish that a model is correct.

The assumptions about treatment assignment remain essential. If important confounders are absent from \(X\), conditional exchangeability may fail, and neither correct nuisance specification relative to the observed \(X\) repairs that failure. Positivity is also essential: if some covariate profiles have essentially no treated or no untreated units, the relevant comparisons are unsupported by the data. Even when the population theorem applies, estimated weights can be unstable in finite samples.

Practical diagnostics therefore remain necessary. Examine whether both treatment groups are represented across covariate profiles, inspect fitted propensity values for near-zero or near-one values, and assess outcome-model fit in each treatment group. Balance diagnostics can help assess how weighting has changed measured covariates, but balance on observed variables does not establish exchangeability. Diagnostics can reveal warning signs; they cannot prove that the chosen nuisance model is correct or that no unmeasured confounding exists.

A common mistake is to read “double robust” as “robust to any one modeling or design error.” The actual claim is narrower: under the causal assumptions and regularity conditions, one of the two nuisance components can converge to an incorrect limit while the other converges to its correct target. If both are wrong, if positivity fails, or if confounding remains, the theorem offers no general protection.