Influence-Function Inference for AIPW

Why Influence-Function Inference Matters

An AIPW estimate combines an outcome regression with inverse-probability weighting. Its point estimate alone does not show how much sampling uncertainty remains. Influence-function inference supplies a practical variance estimate by describing how each observation contributes to first-order variation in the estimated average treatment effect (ATE). It also makes clear why uncertainty from fitting the nuisance models can be handled by the standard formula under suitable conditions.

The formula is not a substitute for assumptions. The causal interpretation requires identification assumptions, and the normal confidence interval requires regularity and adequate nuisance-model estimation. The result below is for independent, identically distributed observations and a binary treatment.

Setup and Assumptions

For each unit, observe \(O=(Y,A,X)\), where \(Y\) is the outcome, \(A\in\{0,1\}\) is treatment status, and \(X\) is a vector of pre-treatment covariates. Let \(Y(1)\) and \(Y(0)\) denote the potential outcomes, and define the ATE as \(\psi=E[Y(1)-Y(0)]\). The nuisance functions are the propensity score \(e(X)=P(A=1\mid X)\) and the conditional outcome means \(m_a(X)=E[Y\mid A=a,X]\), for \(a=0,1\).

The causal interpretation uses consistency, conditional exchangeability, and positivity. Consistency means \(Y=AY(1)+(1-A)Y(0)\). Conditional exchangeability means \((Y(1),Y(0))\) is independent of \(A\) given \(X\). Positivity means \(0<e(X)<1\) almost surely; for stable inference, probabilities should also not be arbitrarily close to zero or one in the target population. These assumptions identify the ATE as \(E[m_1(X)-m_0(X)]\).

The ATE Influence Function

The influence function describes the first-order effect of a small change in the data distribution on the target parameter. For the ATE under the conditions above, it is the centered AIPW contribution:

The formula has two parts. The term \(m_1(X)-m_0(X)-\psi\) reflects variation in covariates and in the conditional treatment contrast. The other terms correct for outcome residuals: treated units supply information about \(m_1\), and control units supply information about \(m_0\). The factors \(1/e(X)\) and \(1/\{1-e(X)\}\) account for the fact that only a fraction of units at a given \(X\) are observed in each treatment group.

A short derivation helps explain why these corrections appear. Write \(\tau(X)=m_1(X)-m_0(X)\), so the identified functional is \(\psi=E[\tau(X)]\). Along a smooth perturbation of the data distribution with score \(s(O)\), the change in the distribution of \(X\) contributes \(E[(\tau(X)-\psi)s(O)]\). The conditional outcome mean changes contribute \(E[\dot m_1(X)-\dot m_0(X)]\), where the conditional-mean derivatives are \(E[(Y-m_1(X))s(O)\mid A=1,X]\) and \(E[(Y-m_0(X))s(O)\mid A=0,X]\).

The residual terms in \(\phi\) reproduce exactly those conditional derivatives when multiplied by the score and averaged. For example, \(E[A\{Y-m_1(X)\}s/e(X)\mid X]\) equals \(E[(Y-m_1(X))s\mid A=1,X]\). The control term gives the corresponding control derivative with a minus sign. The propensity mechanism has no separate derivative term because the identified target is an average over \(X\), not an average over the treatment assignment mechanism. Thus the displayed \(\phi\) has mean zero and its inner product with any score gives the derivative of the ATE functional, which is the influence-function characterization.

Estimating the Variance With Fitted Nuisance Models

In practice, \(e(X)\) and \(m_a(X)\) are unknown, so the influence function is estimated by substituting fitted nuisance functions and the AIPW estimate. For observation \(i\), define

Because fitted quantities and finite-sample calculations need not make the values sum to exactly zero, center them before estimating their variance. The usual empirical estimate of the influence-function variance and the corresponding standard error are

The variance estimate for \(\hat\psi\) is \(\hat V_\phi/n\); its square root is the standard error. The \(n-1\) denominator makes \(\hat V_\phi\) the sample variance of the estimated influence values. All quantities in the influence function have outcome units, so its variance has squared outcome units and the standard error has outcome units.

Conditions for the Standard Error to Be Valid

Plugging fitted nuisance functions into the influence function is not automatically sufficient for valid inference. A standard route to the stated asymptotic result is that both nuisance estimates consistently estimate their target functions, positivity is controlled, the outcome has adequate moments, and the product of the nuisance estimation errors is small enough. A commonly used rate condition is that the product of their relevant estimation errors is \(o_p(n^{-1/2})\). This permits each nuisance estimate to converge more slowly than \(n^{-1/2}\) individually, provided their combined error is sufficiently small.

When flexible machine-learning methods are used, cross-fitting is often used: nuisance models are fitted on other folds of the data, then evaluated on the held-out observations to form the AIPW contributions. This reduces bias from using the same observations both to fit complex nuisance functions and to evaluate their residuals. Cross-fitting does not by itself guarantee valid inference; consistency, rate, moment, and overlap conditions remain necessary. With regular, correctly specified finite-dimensional nuisance models, standard fitted-model inference can also be justified under the corresponding regularity conditions.

Double robustness concerns consistency of the AIPW point estimator when either the propensity model or outcome model is correctly specified, under the assumptions previously stated for that result. It does not mean that the simple plug-in influence variance is automatically valid whenever just one nuisance model is correct and the other converges to a wrong limit. In that situation, fitting the misspecified nuisance model can affect first-order variation; valid inference may require accounting for that fitting step. The variance formula here is justified in the setting where the nuisance estimates support the stated influence-function expansion.

The independent-observation formula also assumes independent sampling. If observations are clustered or otherwise dependent, treating every unit's influence value as independent generally understates or misstates uncertainty. The variance calculation must then reflect the sampling dependence, for example by aggregating contributions at the independent cluster level when that is justified by the design.

Common Pitfall

Influence-function inference matters because it connects the estimator to a transparent unit-level measure of uncertainty. Large absolute influence values can signal that a few observations dominate the estimated effect, often because of large inverse-probability weights or unusual residuals. That is a reason to investigate overlap and model fit; it is not, by itself, permission to change the target population or discard observations without justification.