Neyman Orthogonality of the AIPW Score

Why First-Order Sensitivity Matters

AIPW combines outcome-regression predictions with inverse-propensity-weighted residuals. Its double-robustness property says that the population estimating equation remains correct when either the outcome regression or the propensity score is correct. Neyman orthogonality describes a more local feature: when both nuisance functions are at their true values, small perturbations in them have no first-order effect on the population score. This matters when nuisance functions are estimated flexibly, because it explains why modest estimation errors need not translate directly into first-order bias in the target equation.

The claim concerns the expectation of the score, not the value of the score for each observation. It also does not say that an arbitrary nuisance error is harmless. We will derive the exact mean error first; its form shows both the cancellation at the truth and what can remain when both nuisance functions are wrong.

Setup and Definition

Let \(O=(Y,A,X)\) denote an observation, with binary treatment \(A\in\{0,1\}\), outcome \(Y\), and covariates \(X\). Write \(e^\star(X)=P(A=1\mid X)\) for the true propensity score and \(m_a^\star(X)=E(Y\mid A=a,X)\) for the true outcome regression under treatment level \(a\). Under consistency, conditional exchangeability given \(X\), and positivity, the average treatment effect is \(\tau^\star=E\{m_1^\star(X)-m_0^\star(X)\}\).

For candidate nuisance functions \(\eta=(m_0,m_1,e)\), the AIPW score for \(\tau\) is

Assume \(0<e^\star(X)<1\) almost surely, with overlap bounded away from zero when needed for finite expectations, and that the outcomes and score terms are integrable. Perturbations of \(e^\star\) must keep the propensity in \((0,1)\); directions for the outcome regressions must yield integrable score terms. These conditions allow the expectations and directional derivatives below to be evaluated.

Deriving the Mean-Score Error

Set \(q_a(X)=m_a(X)-m_a^\star(X)\), the error in the fitted outcome regression, and let \(e(X)\) be any candidate propensity function. Conditional on \(X\), the expected treated residual is \(E[A\{Y-m_1(X)\}\mid X]=e^\star(X)\{m_1^\star(X)-m_1(X)\}\). Similarly, the expected control residual is \(E[(1-A)\{Y-m_0(X)\}\mid X]=(1-e^\star(X))\{m_0^\star(X)-m_0(X)\}\). Substituting these conditional expectations into the score gives

To compare this with the truth, subtract the true conditional contrast \(m_1^\star-m_0^\star\). The treated-regression contribution simplifies as \(q_1-(e^\star/e)q_1=(e-e^\star)q_1/e\). The control contribution includes the minus sign in front of the control residual block: \(-q_0+(1-e^\star)q_0/(1-e)=(e-e^\star)q_0/(1-e)\). Thus the exact mean error is

The plus sign between the two regression-error terms is important: the control residual is subtracted in the score, and that subtraction produces a positive sign in this factorized mean-error identity. The formula immediately shows that the expected score is zero if \(e=e^\star\), or if both outcome regressions are correct. That is the population double-robustness cancellation. Orthogonality sharpens the local implication when all nuisances are correct.

Why the First Derivative Is Zero

Consider a differentiable path through the truth: \(e_t=e^\star+t h\) and \(m_{a,t}=m_a^\star+t g_a\), where \(h,g_0,g_1\) are admissible directions and \(t\) is sufficiently close to zero. Along this path the exact mean-error formula becomes

The right-hand side has a factor \(t^2\). If the denominators stay bounded away from zero and the terms are integrable, the expression divided by \(t\) converges to zero as \(t\to0\). Therefore the derivative of the population moment at zero is zero. This is a direct justification of orthogonality for simultaneous perturbations. The formula also shows why the remaining error can be second order in the path size: it involves the propensity perturbation multiplied by outcome-regression perturbations.

The same result can be checked one nuisance block at a time. If only the outcome regressions move while \(e=e^\star\), the mean error is exactly zero for every admissible outcome perturbation. If only the propensity moves while \(m_a=m_a^\star\), it is again exactly zero for every admissible propensity perturbation. In either case the derivative is zero, and the simultaneous-path calculation verifies that combining the directions does not introduce a first-order term.

Locality, Diagnostics, and Limits

Orthogonality is a derivative property at the true nuisance functions. It is not a guarantee of small error for large perturbations, a guarantee of overlap, or a statement that each observation’s score is insensitive to nuisance estimates. AIPW may still be unstable when estimated propensities are near zero or one, and truncating weights can change the estimating equation. The result also cannot repair failures of consistency, conditional exchangeability, or positivity, which are needed to interpret \(\tau^\star\) causally.

A useful diagnostic is to examine whether the conditions supporting the derivation are plausible: fitted propensities should not routinely approach the boundaries, and the weighted residual terms should be well behaved. One can also calculate the exact mean-error expression in a simulation where the true nuisance functions are known, or numerically perturb nuisance functions around known truth and check whether the change is approximately quadratic rather than linear. In real data, the true functions are unknown, so an observed fit cannot establish orthogonality empirically. Cross-fitting helps separate nuisance fitting from score evaluation, but it does not itself prove positivity or remove the need for appropriate nuisance accuracy.

The main alternatives have different protections and limitations. Outcome regression relies on a correct outcome model for its causal interpretation; inverse-probability weighting relies on a correct propensity model and can be sensitive to extreme weights. AIPW combines both blocks and has the double-robust population property, while Neyman orthogonality describes its local sensitivity at the joint truth. Matching is a different strategy for constructing comparisons and does not use this AIPW score identity. None of these properties removes the need to justify the causal assumptions or assess the data’s support.

Common Mistake

A frequent sign error is to write the mean-error bracket as \(q_1/e-q_0/(1-e)\). That expression fails because the control block’s minus sign in the score acts on the residual \(m_0^\star-m_0=-q_0\); the resulting contribution to the factorized error is \(+(e-e^\star)q_0/(1-e)\). Another mistake is to interpret a zero derivative as zero bias for arbitrary nuisance estimates. The correct conclusion is narrower: the first-order population change vanishes at the truth, while simultaneous errors can still leave a nonzero remainder.