Estimating AIPW Standard Errors from Influence Functions

From an AIPW Estimate to Its Standard Error

An AIPW estimate is accompanied by uncertainty, not just a point estimate. A standard error quantifies the sampling variability expected under repeated sampling from the target population. For AIPW, a direct way to estimate that variability is to calculate an influence-function value for each observation and measure how dispersed those values are.

This tutorial concerns inference for the average treatment effect (ATE). Let $A_i\in\{0,1\}$ denote the treatment received by individual $i$, $Y_i$ the observed outcome, and $X_i$ the measured covariates. Define the propensity score $e(X)=P(A=1\mid X)$ and the treatment-specific outcome regressions $m_a(X)=E(Y\mid A=a,X)$, for $a=0,1$. The target is $\tau=E\{Y(1)-Y(0)\}$, identified from observed data under the usual causal identification assumptions.

Definition: The AIPW score for one observation, evaluated at nuisance functions $e,m_1,m_0$, is $$ \psi(O;e,m_1,m_0)=m_1(X)-m_0(X)+\frac{A}{e(X)}\{Y-m_1(X)\}-\frac{1-A}{1-e(X)}\{Y-m_0(X)\}. $$ The empirical AIPW estimate is $\hat{\tau}=n^{-1}\sum_{i=1}^n\hat{\psi}_i$, where $\hat{\psi}_i$ uses the estimated nuisance functions evaluated for individual $i$. The estimated influence-function value is $\hat{\phi}_i=\hat{\psi}_i-\hat{\tau}$.

The subtraction of $\hat{\tau}$ centers the empirical values: $\sum_i\hat{\phi}_i=0$. Under the asymptotic linearity conditions for AIPW, these values estimate the individual contributions to sampling variation in $\hat{\tau}$. The relevant earlier material is the standard subject of asymptotic linearity and efficiency of AIPW; the variance calculation here uses that expansion rather than deriving a different estimator.

Assumptions Behind the Calculation

A numerical standard error is meaningful only if the estimator has the sampling behavior it is intended to measure. For the ATE and this influence-function calculation, the key conditions include:

Identification: treatment is consistent with the potential outcome corresponding to the treatment received, there is no interference between individuals, and treatment is conditionally exchangeable given $X$.
Positivity: treatment probabilities are strictly between zero and one in the covariate population relevant to the target. For stable inference, probabilities should not be so close to the boundaries that the influence-function terms have extreme variability.
Sampling and moments: observations are independent and identically distributed, and the influence function has finite variance.
Asymptotic linearity: the fitted AIPW estimator admits the expansion $\hat{\tau}-\tau=n^{-1}\sum_i\phi(O_i)+o_P(n^{-1/2})$, with nuisance estimation sufficiently controlled for the empirical influence-function variance to be consistent.

These conditions separate two questions. Identification and positivity support interpretation of the target as a causal ATE. Asymptotic linearity and finite variance support the standard error and confidence interval. If the last conditions fail, plugging values into the standard-error formula does not make the resulting inference valid.

Variance estimate: Let $s_{\hat{\phi}}^2=(n-1)^{-1}\sum_{i=1}^n(\hat{\phi}_i)^2$ be the sample variance of the centered estimated influence-function values. Then $$ \widehat{\operatorname{Var}}(\hat{\tau})=\frac{s_{\hat{\phi}}^2}{n} =\frac{1}{n(n-1)}\sum_{i=1}^n(\hat{\psi}_i-\hat{\tau})^2,\qquad \widehat{SE}(\hat{\tau})=\sqrt{\widehat{\operatorname{Var}}(\hat{\tau})}. $$

Why the Formula Estimates Sampling Variability

The justification follows from the asymptotic linear representation. If $\phi(O)$ has mean zero and variance $\sigma_\phi^2$, then the leading random part of the estimator is the sample average of $n$ influence-function values. Consequently, its variance is approximately $\sigma_\phi^2/n$. The empirical centered variance estimates $\sigma_\phi^2$.

\sqrt{n}(\hat{\tau}-\tau) =\frac{1}{\sqrt{n}}\sum_{i=1}^n\phi(O_i)+o_P(1) \ \Longrightarrow\ N(0,\sigma_\phi^2).

Replacing the unobserved $\phi(O_i)$ by $\hat{\phi}_i$ is justified when the estimated influence-function values consistently capture the first-order contributions. The law of large numbers then gives $s_{\hat{\phi}}^2$ as an estimate of $\sigma_\phi^2$, and Slutsky's theorem gives the estimated standard error for $\hat{\tau}$. A two-sided asymptotic $100(1-\alpha)\%$ Wald interval is

\hat{\tau}\ \pm\ z_{1-\alpha/2}\widehat{SE}(\hat{\tau}),

where $z_{1-\alpha/2}$ is the corresponding standard normal quantile; for a 95% interval, it is approximately 1.96. This is an asymptotic normal interval, not a claim that there is a 95% probability that a fixed parameter lies in this particular interval.

Calculate each fitted score.
For every individual, substitute the observed treatment and outcome and the fitted propensity and outcome predictions into the AIPW score.

Compute and center the estimate.
Average the $n$ scores to obtain $\hat{\tau}$, then subtract $\hat{\tau}$ from each score to obtain $\hat{\phi}_i$.

Estimate dispersion and uncertainty.
Calculate the sample variance of the centered values, divide by $n$, and take the square root to obtain the standard error.

Construct and assess the interval.
Use the standard normal quantile for a Wald interval, then check whether the assumptions and empirical influence-function values make that approximation credible.

Worked Example: Calculating a Standard Error from Four Scores

Suppose a hypothetical sample of four individuals has outcome units of milligrams. After fitting the nuisance functions, the following table gives the outcome-regression contrast $m_1(X_i)-m_0(X_i)$, the combined residual-correction term, and their sum, $\hat{\psi}_i$. The correction term is the sum of the two treatment-specific residual terms in the AIPW score.

Individual	Regression contrast	Residual correction	$\hat{\psi}_i$ (mg)
1	2	0	2
2	3	1	4
3	1	-1	0
4	5	1	6

The AIPW estimate is the average of the score values:

\hat{\tau}=\frac{2+4+0+6}{4}=3\text{ mg}.

The centered values are $-1,1,-3,3$ mg. Their squared deviations sum to $1+1+9+9=20$ mg$^2$. Thus the estimated influence-function variance is $20/\{4(4-1)\}=1.6667$ mg$^2$, and

\widehat{SE}(\hat{\tau})=\sqrt{1.6667}=1.291\text{ mg}.

The 95% Wald interval is $3\pm1.96(1.291)$, or approximately $[0.470,5.530]$ mg. Under the stated asymptotic assumptions, this procedure has approximately 95% coverage across repeated samples; it does not assign a probability to the fixed ATE after observing these data.

Examples of Centering and Sensitivity

Worked Example: Using the Sample Variance Correctly

For a second hypothetical AIPW analysis, suppose the five fitted score values are $1,2,3,4,5$ days. The AIPW estimate is

\hat{\tau}=\frac{1+2+3+4+5}{5}=3\text{ days}.

The centered score values are $-2,-1,0,1,2$, and the sum of their squares is $4+1+0+1+4=10$ days$^2$. The sample variance is $s_{\hat{\phi}}^2=10/(5-1)=2.5$ days$^2$. Therefore

\widehat{SE}(\hat{\tau})=\sqrt{\frac{2.5}{5}} =\sqrt{0.5}=0.7071\text{ days}.

The 95% interval is $3\pm1.96(0.7071)$, which is approximately $[1.614,4.386]$ days. The denominator $n-1$ estimates the variance of the influence-function values; dividing that variance by $n$ estimates the variance of their sample average.

Worked Example: An Influence-Function Outlier and a Wide Interval

Suppose a third hypothetical analysis produces five AIPW score values of $-1,0,1,2,8$ milligrams. Their mean is $\hat{\tau}=10/5=2$ mg. The centered values are $-3,-2,-1,0,6$ mg, whose squared deviations sum to $9+4+1+0+36=50$ mg$^2$. Hence

s_{\hat{\phi}}^2=\frac{50}{4}=12.5\text{ mg}^2,\qquad \widehat{SE}(\hat{\tau})=\sqrt{\frac{12.5}{5}}=1.581\text{ mg}.

The 95% Wald interval is $2\pm1.96(1.581)$, approximately $[-1.099,5.099]$ mg. The value 8 mg makes a substantial contribution to the variance estimate: its squared deviation alone is 36 of the total 50 mg$^2$. This does not by itself prove the analysis is invalid, but it is a reason to inspect the observation, fitted propensity, residual, and score calculation. With only five observations, the normal approximation is also especially uncertain.

Diagnostics, Alternatives, and Limitations

The influence-function calculation is simple, but its diagnostics matter. Inspect the distribution of $\hat{\phi}_i$, not just the final standard error. A small number of very large values can dominate the estimated variance. In AIPW, the residual corrections divide by $e(X_i)$ or $1-e(X_i)$, so observations with fitted treatment probabilities near zero or one can produce large score contributions. Check those probabilities, the corresponding residuals, and the arithmetic of the score. Also verify that every prediction is for the correct treatment level and that the outcome units are consistent throughout.

A conventional model-based standard error from an outcome regression alone is not automatically a standard error for the AIPW estimator: it may omit the variability contributed by the weighting and augmentation terms. Likewise, reporting only the standard deviation of the score values is insufficient; uncertainty in their average uses that standard deviation divided by $\sqrt{n}$. A bootstrap is a possible alternative, but it must resample observations and repeat the entire estimation procedure, including nuisance fitting. It does not remove identification or positivity requirements, and in a small or unstable sample it need not provide reliable inference.

The Wald interval relies on an approximately normal sampling distribution. It can perform poorly with small samples, highly skewed influence-function values, or severe positivity problems. Its width is not a diagnostic test of causal identification: a narrow interval cannot establish exchangeability, correct measurement, or a well-defined treatment. Nor does a calculated standard error correct bias when the estimator's first-order approximation is inappropriate.

Common mistake: Treating $\sqrt{\sum_i\hat{\phi}_i^2/(n-1)}$ as the standard error of $\hat{\tau}$. That quantity estimates the standard deviation of individual influence-function values, not the sampling standard deviation of their average. The standard error is $s_{\hat{\phi}}/\sqrt{n}$. Also, centering is essential: calculate deviations from $\hat{\tau}$, rather than using an uncentered second moment as if it were a variance.

The practical conclusion is conditional: when the AIPW estimator is asymptotically linear, the influence-function values have finite variance, and their empirical distribution is informative about that variance, the centered-score formula provides a direct standard error and an asymptotic confidence interval. When those conditions are doubtful, the appropriate response is to investigate the estimator and data-generating conditions, not merely to report more decimal places.

Check Your Understanding

Work through the questions, then compare your reasoning with the solutions.

For fitted AIPW score values $2,4,0,6$, what are the estimate and standard error?
Why are the empirical influence-function values centered by subtracting $\hat{\tau}$?
A report gives the sample standard deviation of the influence-function values as 3 units for $n=9$. What is the corresponding standard error?
What does an unusually large fitted influence-function value suggest, and what should be examined?
Does calculating this standard error establish that the ATE is identified? Explain.

Solutions. (1) The mean is $3$; the centered values are $-1,1,-3,3$, with squared deviations summing to 20. Thus $\widehat{SE}=\sqrt{20/\{4(3)\}}=1.291$ units. (2) Centering estimates the mean-zero influence function and makes the empirical deviations represent dispersion around the estimated ATE. (3) The standard error is $3/\sqrt{9}=1$ unit. (4) It may indicate a high-variance contribution, perhaps associated with a propensity near a boundary or a large residual; inspect the fitted propensity, residual, observation, and score calculation. (5) No. The standard error concerns sampling variability under the required conditions; identification additionally depends on causal assumptions such as consistency, no interference, exchangeability, and positivity.