Product-Rate Conditions for Valid AIPW Inference

Why nuisance rates matter for AIPW inference

The augmented inverse-probability weighted estimator (AIPW) combines an estimated propensity score with estimated treatment-specific outcome means. Its double robustness means that, under the identification assumptions, the population score can still target the average treatment effect when one nuisance model is correct. But that property alone does not guarantee valid root-n inference when both nuisance functions are estimated imperfectly. The size of their joint error matters.

This tutorial derives the relevant second-order remainder. The key result is that the remainder is controlled by a product: propensity-score error multiplied by outcome-regression error. If that product is smaller than \(n^{-1/2}\), the remainder is negligible on the scale used for root-n inference. The condition permits one nuisance estimate to converge relatively slowly if the other is sufficiently accurate.

The causal interpretation of \(\tau\) requires the usual identification conditions: consistency of the observed outcome with the potential outcome under the received treatment, conditional exchangeability of treatment and potential outcomes given \(X\), and positivity, so that \(0<e_0(X)<1\) on the target population. For stable inference, it is common to impose a stronger overlap condition: for some fixed \(\epsilon>0\), \(\epsilon\leq e_0(X)\leq1-\epsilon\) almost surely. These conditions identify the estimand; they do not by themselves ensure that estimated nuisance functions converge quickly enough.

The AIPW score and its remainder

For nuisance functions \(\eta=(e,\mu_1,\mu_0)\), define the AIPW score

The estimator averages this score, using estimated nuisance functions. For the derivation below, consider nuisance estimates fitted on data independent of the observation at which the score is evaluated. Cross-fitting can provide this separation; the earlier material on cross-fitting AIPW discusses its role. The product-rate result concerns the bias in the score’s population mean, and does not by itself control every source of sampling error.

Condition on \(X\), and take expectations of the score under the true data distribution. Since \(E[Y\mid A=a,X]=\mu_{a0}(X)\), the treated residual term has conditional expectation \(e_0(X)\{\mu_{10}(X)-\hat\mu_1(X)\}/\hat e(X)\). The control residual term has conditional expectation \((1-e_0(X))\{\mu_{00}(X)-\hat\mu_0(X)\}/\{1-\hat e(X)\}\), with a minus sign in the score. Subtracting the true conditional treatment effect gives the exact remainder.

This identity shows why the error is second order: if either the propensity score is exactly correct or both outcome means are exactly correct, the displayed remainder is zero. When neither is exact, it is their errors acting together—not simply the size of either error alone—that determines this component of bias.

From the remainder to a product-rate condition

For a function \(f(X)\), write \(\|f\|_2=\{E[f(X)^2]\}^{1/2}\). Suppose the estimated propensity score is bounded away from zero and one with probability tending to one: for a fixed \(\epsilon>0\), \(\epsilon\leq\hat e(X)\leq1-\epsilon\). Applying the triangle inequality and Cauchy–Schwarz to the remainder yields

Thus a sufficient product-rate condition is

Here \(o_p(n^{-1/2})\) means that multiplying the quantity by \(\sqrt n\) gives a term that converges to zero in probability. This makes the population remainder negligible relative to the estimator’s usual root-n fluctuation. The product condition is a sufficient bound; the exact remainder can be smaller because errors may cancel, but relying on cancellation is generally unsafe.

A convenient symmetric sufficient condition is that the propensity-score error and each outcome-regression error are all \(o_p(n^{-1/4})\). Multiplying any two such rates gives \(o_p(n^{-1/2})\). Symmetry is not required: one nuisance component can converge more slowly if the other converges quickly enough. In each case, the rates must be understood in the stated \(L_2\) norm, not as pointwise rates at selected covariate values.

Why the boundary rate is different

The difference between \(O_p(n^{-1/2})\) and \(o_p(n^{-1/2})\) is essential. A product merely bounded at order \(n^{-1/2}\) need not vanish after multiplication by \(\sqrt n\); it can leave a nonzero first-order bias. The next example gives a concrete illustration using the exact remainder.

Conditions beyond the product rate

A product-rate condition controls the population bias term, not the entire estimator. For the usual root-n normal approximation, one also needs a valid asymptotic linear expansion with the efficient influence function

In particular, the influence function must have finite, nonzero variance, and the remaining empirical fluctuation from estimating nuisance functions must be negligible. With cross-fitting, a standard route is to require suitable \(L_2\) convergence of the fitted score to the true score, along with moment and overlap conditions; conditional on the training data, the held-out observations can then control the empirical fluctuation. Without sample separation, additional restrictions on the complexity of the fitted functions may be needed. The product-rate calculation alone does not supply these conditions.

The overlap bound on \(\hat e\) is also substantive. If estimated scores approach zero or one, the inverse factors in the score can amplify outcome errors and may make the remainder bound uninformative. Truncating scores can limit extreme weights, but it changes the nuisance estimate and may introduce bias; it is not an automatic proof of the required rate or of valid inference.

Diagnostics, alternatives, and limitations

The true nuisance functions are unknown, so their actual \(L_2\) errors and convergence rates cannot generally be verified from one data set. Diagnostics can still reveal warning signs: propensity-score estimates concentrated near zero or one, weak predictive performance for treatment or outcomes, substantial instability across training samples, or strong sensitivity of the AIPW estimate to reasonable fitting choices. Held-out prediction measures and comparisons across learners can inform model assessment, but good predictive performance does not certify the product-rate condition for the causal estimand.

A parametric approach may obtain fast nuisance rates when its models are correctly specified and sufficiently regular, but misspecification can invalidate that rate argument. Flexible learners can reduce dependence on restrictive functional forms, yet may converge more slowly; the product-rate condition explains why their joint rates matter. If one nuisance function is known or estimated especially accurately, the other may be allowed to converge more slowly. In contrast, simply choosing a doubly robust estimator does not ensure small finite-sample error when both nuisance estimates are poor.

The condition is a sufficient asymptotic guarantee, not a finite-sample error bound. It does not ensure good overlap, correct identification, low variance, or adequate sample size. Nor does a negligible remainder make an invalid causal estimand valid: exchangeability, consistency, and positivity remain separate requirements. The earlier material on Neyman orthogonality explains why small nuisance perturbations have no first-order effect under the relevant conditions; the product-rate result quantifies how quickly their combined error must shrink for that protection to support root-n inference.