Asymptotic Linearity and Efficiency of AIPW

From an AIPW Estimate to a Sampling Distribution

An AIPW estimate is useful not only because it combines outcome regression and propensity weighting, but also because its large-sample behavior can often be described by an average of independent contributions. That description is called an asymptotic linear expansion. It explains why a normal approximation can apply and identifies the variance an efficient estimator can achieve.

Let $O_i=(X_i,A_i,Y_i)$ denote the data for unit $i$, where $X_i$ is a vector of pretreatment covariates, $A_i\in\{0,1\}$ is treatment assignment, and $Y_i$ is the observed outcome. The target is the average treatment effect, $\tau=E[Y(1)-Y(0)]$. Define the true propensity score $e_0(x)=P(A=1\mid X=x)$ and the treatment-specific outcome regressions $m_{a0}(x)=E[Y\mid A=a,X=x]$, for $a=0,1$.

For this causal interpretation, assume consistency, conditional exchangeability $ (Y(1),Y(0))\perp A\mid X$, and positivity: for some $\epsilon>0$, $\epsilon\leq e_0(X)\leq 1-\epsilon$ with probability one. Also assume independent and identically distributed observations and finite second moments for the relevant outcomes and influence function. These assumptions identify the causal target and support the large-sample results below; they are not consequences of the AIPW formula.

Definition: Write $m_a(x)$ and $e(x)$ for nuisance functions used in the estimator. The AIPW score is $$ \psi(O;m,e)=m_1(X)-m_0(X)+\frac{A}{e(X)}\{Y-m_1(X)\}-\frac{1-A}{1-e(X)}\{Y-m_0(X)\}. $$ The AIPW estimator is $\hat{\tau}=n^{-1}\sum_{i=1}^n\psi(O_i;\hat m,\hat e)$. An estimator is asymptotically linear with influence function $\phi$ if $$ \hat{\tau}-\tau=\frac{1}{n}\sum_{i=1}^n\phi(O_i)+o_p(n^{-1/2}), $$ where $E[\phi(O)]=0$ and $E[\phi(O)^2]<\infty$.

At the true nuisance functions, the score’s centered version is the efficient influence function for the ATE:

\phi_0(O)=\psi(O;m_0,e_0)-\tau.

If the estimator has the expansion with $\phi_0$, the central limit theorem gives

\sqrt{n}(\hat{\tau}-\tau)\ \xrightarrow{d}\ N\{0,V_{\mathrm{eff}}\}, \qquad V_{\mathrm{eff}}=E[\phi_0(O)^2].

Thus the estimator’s error is, to first order, the average of mean-zero contributions with variance $V_{\mathrm{eff}}$. The corresponding large-sample variance of $\hat{\tau}$ is approximately $V_{\mathrm{eff}}/n$. This statement is about a sampling distribution across repeated samples; it does not say that a particular estimate has a probability of being close to $\tau$.

Why the Influence Function Determines Efficiency

In the nonparametric model for the observed data, the variance $E[\phi_0^2]$ is the semiparametric efficiency bound for regular, asymptotically linear estimators of the ATE. An estimator is efficient when it is regular and its first-order influence function is $\phi_0$; its asymptotic variance then equals this bound. The bound is a lower limit on asymptotic variance in the stated model, not a guarantee that every AIPW implementation reaches it.

The bound has a useful decomposition. Let $\tau(X)=m_{10}(X)-m_{00}(X)$, and let $\sigma_a^2(X)=\operatorname{Var}(Y\mid A=a,X)$. Conditional on $X$, the residual terms in the influence function have mean zero, and only one treatment arm’s residual is observed for each unit. Taking conditional variances and then averaging gives

V_{\mathrm{eff}} = E\left[ \frac{\sigma_1^2(X)}{e_0(X)} +\frac{\sigma_0^2(X)}{1-e_0(X)} +\{\tau(X)-\tau\}^2 \right].

The first two terms reflect outcome variability amplified when the corresponding treatment arm is uncommon at a covariate value. The final term reflects variation in conditional treatment effects across covariate values. This expression also shows why positivity matters for precision: when $e_0(X)$ approaches zero or one, one of the residual-variance contributions can become large.

Worked Example: Computing the Efficiency Bound

Suppose, hypothetically, that across the covariate distribution the components of the bound are $E[\sigma_1^2(X)/e_0(X)]=22.5$ outcome-units squared, $E[\sigma_0^2(X)/(1-e_0(X))]=26.7$ outcome-units squared, and $E[\{\tau(X)-\tau\}^2]=2.25$ outcome-units squared.

Calculate the bound. Adding the three terms gives

V_{\mathrm{eff}}=22.5+26.7+2.25=51.45

The bound is $51.45$ outcome-units squared per observation. For a sample size of $n=400$, the asymptotic variance of an efficient AIPW estimator is $51.45/400=0.1286$ outcome-units squared, and its asymptotic standard deviation is $\sqrt{0.1286}=0.3586$ outcome units. This is a large-sample benchmark, not a claim that a finite-sample estimate must have exactly this variability.

For comparison, a regular estimator in the same model with asymptotic variance $60$ outcome-units squared per observation would not attain the bound. An estimator with variance below $51.45$ would conflict with the bound’s assumptions if it were claimed to be regular and asymptotically linear for the same unrestricted model.

Why Plugging In Estimated Nuisance Functions Can Work

The influence-function limit concerns a score evaluated at the true nuisance functions, whereas AIPW uses estimated functions. The key issue is whether estimating those functions changes the estimator at order $n^{-1/2}$. A short calculation identifies the population remainder. For fixed candidate functions $m$ and $e$, conditional expectation given $X$ yields

E\{\psi(O;m,e)\}-\tau = E\left[ \frac{e(X)-e_0(X)}{e(X)}\{m_1(X)-m_{10}(X)\} + \frac{e(X)-e_0(X)}{1-e(X)}\{m_0(X)-m_{00}(X)\} \right].

For example, the treated residual has conditional expectation $e_0(X)\{m_{10}(X)-m_1(X)\}$. Combining it with $m_1(X)$ in the score gives the treated-arm contribution $m_{10}(X)+\{e(X)-e_0(X)\}\{m_1(X)-m_{10}(X)\}/e(X)$. Applying the same conditional-expectation calculation to the control residual gives the second remainder term. This establishes the displayed identity.

The remainder is a product of propensity-score error and outcome-regression error, rather than a first-order error from either nuisance function alone. But a small population remainder is not, by itself, a proof of asymptotic linearity: the empirical fluctuation from using estimated functions must also be controlled. A simple sufficient setting is correctly specified, sufficiently smooth finite-dimensional nuisance models estimated at root-$n$ rates, with regularity conditions that make their empirical-process contribution negligible. More generally, the needed conclusion is that nuisance estimation changes the score average by $o_p(n^{-1/2})$. This tutorial’s claim of efficiency applies only when that conclusion, as well as the causal and moment assumptions, holds.

Worked Example: Checking a Simple Nuisance-Rate Condition

Consider a hypothetical sequence of fitted nuisance functions for which both estimation errors, measured in suitable mean-square norms, are $O_p(n^{-1/2})$. Assume also that fitted propensities remain bounded away from zero and one, and that the empirical fluctuation condition just described holds.

Check the product order. The product of two $O_p(n^{-1/2})$ errors is $O_p(n^{-1})$. Since $n^{-1}=o(n^{-1/2})$, the population remainder is smaller than the first-order sampling scale. Under the stated empirical-process condition, the estimator therefore has the same first-order expansion as the score at the truth.

Consequently, its limiting variance is $E[\phi_0^2]$, so it attains the efficiency bound under these assumptions. The conclusion depends on both nuisance functions converging appropriately and on the empirical fluctuation condition; the product calculation alone is not enough to establish the result.

Reading the Limit and Diagnosing Failure

The asymptotic normal approximation has an immediate scale interpretation. If $V_{\mathrm{eff}}$ is finite and the influence-function expansion holds, then $\hat{\tau}-\tau$ is typically of order $n^{-1/2}$, and $\sqrt{n}(\hat{\tau}-\tau)$ has an approximately centered normal distribution for large $n$. A concrete hypothetical illustration: if $V_{\mathrm{eff}}=4$ outcome-units squared and $n=100$, the limiting standard deviation of $\hat{\tau}$ is $\sqrt{4/100}=0.2$ outcome units. The approximation does not establish that it is accurate at $n=100$; that depends on how well the assumptions and large-sample approximation apply.

Several failures can undermine the argument. Propensities near zero or one can make the influence function highly variable or even violate the finite-variance condition. Poorly estimated nuisance functions can leave a remainder that is not negligible at the root-$n$ scale. Misspecified nuisance models can also affect efficiency even when a consistency argument remains available. In addition, dependence between units, incorrect treatment or outcome definitions, and failure of conditional exchangeability undermine assumptions used to identify or analyze the target.

Diagnostics are clues, not proofs of the asymptotic conditions. Inspect the propensity-score range and treatment-arm support across covariates; assess whether fitted outcome predictions are plausible; and examine whether a few observations contribute disproportionately to the score. Severe problems suggest that the claimed normal limit may be a poor approximation. They cannot, however, verify exchangeability from observed data alone.

Worked Example: Consistency Does Not Guarantee Efficiency

Suppose, hypothetically, that the propensity score model is correct and bounded away from zero and one, but the fitted outcome regressions converge to fixed functions $m_1^*(X)$ and $m_0^*(X)$ that differ from the true regressions. The AIPW population remainder above is zero because $e=e_0$, so the score’s expectation still equals $\tau$.

That establishes a population consistency argument, but it does not establish the efficient influence-function expansion. At the probability limit of these nuisance functions, the score is generally $\psi(O;m^*,e_0)-\tau$, which need not equal $\phi_0(O)$. Its variance can therefore differ from $V_{\mathrm{eff}}$, and the estimator need not attain the bound. Efficiency requires the appropriate first-order behavior, not merely an unbiased population score or consistency.

Common Mistake and Practical Comparison

Common mistake: “AIPW is doubly robust, so it must be efficient whenever one nuisance model is correct.” This confuses a consistency property with a first-order variance result. Double robustness concerns when the estimator’s target is recovered; efficiency requires the estimator’s influence function to be the efficient one and requires the regularity conditions for the expansion. A correct propensity model paired with a misspecified outcome model can yield consistency without attaining the efficiency bound.

The main alternatives clarify what the efficiency result contributes. Outcome regression relies on modeling conditional means and can be efficient under suitable conditions, but it does not generally protect against a misspecified outcome model. Inverse-probability weighting relies on the treatment model and can be unstable when fitted propensities are extreme. AIPW combines both components and has the efficient influence function as its first-order contribution when its nuisance estimation and regularity conditions are adequate. None of these methods repairs unmeasured confounding or a failure of positivity.

The practical conclusion is conditional rather than automatic: AIPW can attain the semiparametric efficiency bound, but the formula alone does not ensure it. The causal assumptions identify the estimand; the influence-function expansion describes the sampling error; and nuisance and moment conditions determine whether that expansion is valid. The efficient influence function for the ATE supplies the benchmark against which the estimator’s first-order behavior is judged.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

What does an asymptotically linear expansion say about the leading sampling error of an estimator?
If $V_{\mathrm{eff}}=36$ outcome-units squared and $n=144$, what is the asymptotic standard deviation of an efficient AIPW estimate?
Why does a zero population remainder when the propensity score is correct not, by itself, prove efficiency?
Name two assumptions needed for the causal ATE interpretation and one condition needed for the normal-limit argument.
What does a propensity score near zero imply for the treated-residual term in the efficiency bound?

Solutions: (1) The estimator’s error equals an average of mean-zero influence-function contributions plus a term negligible relative to $n^{-1/2}$. (2) The standard deviation is $\sqrt{36/144}=0.5000$ outcome units. (3) The estimator can be consistent while its first-order influence function differs from the efficient influence function, so its variance need not equal the efficiency bound. (4) Examples are consistency, conditional exchangeability, and positivity; a finite influence-function second moment or valid nuisance-estimation remainder control supports the normal limit. (5) The term $\sigma_1^2(X)/e_0(X)$ can become large, increasing variance and potentially violating finite-variance conditions.