Augmented Standardization Estimator

• The paper introduces an augmented standardization estimator that begins with a plug‐in outcome regression and augments it with residual weighting to eliminate first-order bias.
• It employs a minimax linear strategy over function classes to optimize residual weights, enhancing efficiency and robustness in estimating continuous linear functionals.
• The estimator unifies frameworks like AIPW and AutoDML and adapts to settings such as cluster-randomized and stepped-wedge designs for improved marginal treatment effect estimation.

Searching arXiv for recent and foundational papers on augmented standardization estimators and related estimators. An augmented standardization estimator is an estimator that begins with a plug-in standardization, or outcome-regression, estimate of a target functional and then augments it with a residual-weighted correction. In the most general formulation discussed in the recent literature, the target is a continuous linear functional of a conditional expectation function, the baseline estimator is obtained by integrating a fitted regression over an empirical or target covariate distribution, and the augmentation is chosen so that first-order bias is removed while semiparametric efficiency is approached or attained under stated conditions. This general pattern appears in “Augmented Minimax Linear Estimation” (Hirshberg et al., 2017), in model-robust standardization for cluster-randomized trials (Li et al., 25 May 2025), in its stepped-wedge extension (Fang et al., 23 Jul 2025), and in the linear AutoDML characterization of augmented balancing weights (Bruns-Smith et al., 2023).

1. General definition and canonical form

In modern semiparametric language, an augmented standardization estimator is a doubly robust estimator written in standardization form. One starts from an estimate of the regression function and evaluates the target functional by standardization, then adds or subtracts a residual term weighted by an estimated Riesz representer, inverse-probability weights, or other balancing weights (Bruns-Smith et al., 2023).

For i.i.d. data $(Y_i,Z_i)$, with regression function

$m(z)=\mathbb E_P[Y_i\mid Z_i=z],$

the baseline plug-in estimator is

$$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$

or, in evaluable cases, $\psi(\hat m)$. The augmented minimax linear form is

$$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$

so that the augmentation term

$$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$

is subtracted from the standardization estimator (Hirshberg et al., 2017).

The same structure can be written in the more familiar efficient-influence-function form

$$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$

where $h$ defines the linear functional and $\hat\alpha$ estimates the Riesz representer (Bruns-Smith et al., 2023). In this representation, the estimator is naturally decomposed into a standardization component and an augmentation component. The core idea is invariant across applications: standardize predicted outcomes to a target distribution, then correct remaining first-order error with a residual-based term.

2. Target estimands as linear functionals of regression

The general setup treats the estimand as a continuous linear functional of the regression $m$,

$m(z)=\mathbb E_P[Y_i\mid Z_i=z],$0

with $m(z)=\mathbb E_P[Y_i\mid Z_i=z],$1 linear on a function space $m(z)=\mathbb E_P[Y_i\mid Z_i=z],$2. In evaluable cases there is a Riesz representer $m(z)=\mathbb E_P[Y_i\mid Z_i=z],$3 such that

$m(z)=\mathbb E_P[Y_i\mid Z_i=z],$4

This formulation covers missing-data estimands, treatment effects, average partial effects, and distribution-shift functionals (Hirshberg et al., 2017).

A central example is the missing-at-random mean. With $m(z)=\mathbb E_P[Y_i\mid Z_i=z],$5, observed outcome $m(z)=\mathbb E_P[Y_i\mid Z_i=z],$6, and regression

$m(z)=\mathbb E_P[Y_i\mid Z_i=z],$7

the target is

$m(z)=\mathbb E_P[Y_i\mid Z_i=z],$8

Under MAR, this equals $m(z)=\mathbb E_P[Y_i\mid Z_i=z],$9. The corresponding Riesz representer is

$$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$0

The standardization estimator

$$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$1

is classical outcome regression or standardization (Hirshberg et al., 2017).

The same framework also includes the average partial effect for a continuous treatment,

$$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$2

and the conditionally linear model

$$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$3

For binary treatment, $$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$4 is the conditional treatment effect and $$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$5 is the usual ATE. It also includes generalized policy or distribution-shift effects,

$$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$6

again a linear functional with a Riesz representer (Hirshberg et al., 2017).

The significance of this formulation is that it makes standardization primary rather than incidental. The object of interest is not a regression coefficient, but a marginal functional defined by integrating a regression over a specified distribution or contrast. This distinction becomes decisive in settings where regression coefficients and marginal causal estimands diverge.

3. Augmentation, minimax correction, and semiparametric efficiency

The central defect of pure plug-in standardization is that its bias is of order $$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$7 or worse, and with high-dimensional or rough $$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$8, the plug-in often converges too slowly to be $$\hat\psi^{\text{plug}}=\frac{1}{n}\sum_{i=1}^n h(Z_i,\hat m),$$9-consistent. It is also not semiparametrically efficient in general unless $\psi(\hat m)$0 is very accurate, for example $\psi(\hat m)$1 in $\psi(\hat m)$2 norm and undersmoothed (Hirshberg et al., 2017).

The augmented minimax linear construction addresses this by choosing the residual weights through a minimax problem over a class $\psi(\hat m)$3 of plausible regression errors. With

$\psi(\hat m)$4

the weights are defined as

$\psi(\hat m)$5

Here $\psi(\hat m)$6 is the worst-case linear approximation error over the regression-error class, and the $\psi(\hat m)$7 penalty controls variance (Hirshberg et al., 2017).

This yields an estimator that is best understood exactly as an augmented standardization estimator for a large class of linear functionals. The augmentation follows the same pattern as AIPW—outcome regression plus residual weighting—but the weights are chosen by targeted minimax balancing rather than by direct propensity or Riesz modeling. The function class $\psi(\hat m)$8 may be a Hölder ball, a Sobolev ball, a RKHS unit ball, an $\psi(\hat m)$9-type hull of basis functions, or, in the conditionally linear APE setting, the class

$$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$0

If the Riesz representer $$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$1 were known, the oracle augmented estimator would be

$$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$2

Its influence function is

$$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$3

Under the paper’s conditions, the AML weights converge in $$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$4 to $$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$5, the estimator is asymptotically linear with this same influence function, and it is semiparametrically efficient under the stated variance condition (Hirshberg et al., 2017).

A recurring source of confusion is the status of robustness. The AML construction has the same functional form as doubly robust AIPW, but it is not advertised as doubly robust in the usual sense. It demands a mild consistency rate on $$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$6, imposes essentially no smoothness requirement on $$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$7, and controls the bilinear error term by approximate orthogonality or balance over $$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$8. This suggests that the defining feature is not classical nuisance-model redundancy, but targeted debiasing of a standardization estimator.

4. Relation to AIPW, AutoDML, and linear regression representations

The augmented standardization form coincides algebraically with the standard DR or AIPW representation,

$$\hat\psi_{\mathrm{AML}}=\frac{1}{n}\sum_{i=1}^n\left[h(Z_i,\hat m)-\hat\gamma_i\{\hat m(Z_i)-Y_i\}\right],$$9

with $$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$0 an estimate of the Riesz representer (Hirshberg et al., 2017). In the missing-at-random mean example, the familiar AIPW estimator is

$$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$1

If one explicitly constrains the AML weights to the form $$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$2, AML reduces to AIPW (Hirshberg et al., 2017).

The linear AutoDML analysis sharpens this connection by showing that, when both the outcome model and the weighting model are linear in a basis, the augmented estimator is exactly a single linear regression plug-in (Bruns-Smith et al., 2023). With a linear outcome model $$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$3 and linear balancing weights, the augmented estimator can be rewritten as

$$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$4

where each coefficient of $$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$5 combines the regularized outcome-regression coefficient and the unpenalized OLS coefficient. In transformed coordinates,

$$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$6

This produces several exact equivalences. Double ridge—ridge for the outcome model and ridge-type balancing for the weights—is numerically equal to a single undersmoothed ridge regression. The paper states that double kernel ridge is likewise equivalent to a single undersmoothed kernel ridge, numerically in finite samples (Bruns-Smith et al., 2023). With lasso-type balancing, the augmented estimator exhibits a double selection property: the final active set is the union of variables selected by the outcome lasso and variables selected by the balancing stage.

The same analysis also identifies collapse cases. If the outcome model is unregularized OLS, the augmented estimator reduces to OLS. If the weights achieve exact balance, the augmented estimator reduces to the OLS plug-in. If the weight penalty is tuned to $$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$7, the estimator can again become numerically equal to OLS standardization. The Lalonde re-analysis in the paper is used to illustrate these phenomena (Bruns-Smith et al., 2023). A plausible implication is that, in linear and RKHS settings, the distinction between “augmented” estimation and “undersmoothed” plug-in estimation is often representational rather than substantive.

5. Cluster-randomized trials and estimand alignment

In cluster-randomized trials, augmented standardization is developed as a model-robust estimator of marginal treatment effects defined by explicit cluster weights rather than by regression coefficients. With cluster-level treatment $$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$8, cluster size $$\hat\Delta=\frac{1}{n}\sum_{i=1}^n \hat\gamma_i\{\hat m(Z_i)-Y_i\}$$9, and cluster mean potential outcome

$$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$0

the paper defines

$$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$1

Choosing $$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$2 yields the cluster-average treatment effect, and choosing $$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$3 yields the individual-average treatment effect (Li et al., 25 May 2025).

The proposed estimator is

$$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$4

with $$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$5 the known randomization probability. This has exactly the structure of an augmented standardization or AIPW estimator: standardization of outcome-model predictions plus an IPW residual augmentation (Li et al., 25 May 2025).

The paper’s central claim is that coefficient estimators from GLMMs and GEEs are not generally estimand-aligned with $$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$6 or $$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$7 under model misspecification or informative cluster size, whereas the standardized augmented estimator is. Outcome models may be cluster-level GLMs, LMMs, GLMMs, or GEEs; the working model is used only to generate predictions of cluster means, and consistency does not require correct specification of that model. The Appendix derives the efficient influence function for $$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$8, and the resulting estimator is consistent even if the outcome regression model is misspecified, asymptotically normal under standard regularity conditions, and semiparametrically efficient if the outcome model is correctly specified (Li et al., 25 May 2025).

This framework also yields a deletion-based jackknife variance estimator and a natural test for informative cluster size based on the contrast between $$\hat\psi=\frac{1}{n}\sum_{i=1}^n\Big\{h(X_i,Z_i,\hat m)+\hat\alpha(X_i,Z_i)\big(Y_i-\hat m(X_i,Z_i)\big)\Big\},$$9 and $h$0. The methods are implemented in the MRStdCRT R package (Li et al., 25 May 2025). A common misconception is that the estimator is “doubly robust” in the classical observational-study sense. The paper explicitly states that, because the randomization probabilities are known by design, there is no scenario in which both the outcome and treatment models are allowed to be misspecified; robustness is model-assisted with respect to the outcome model.

6. Stepped-wedge designs, informative size, and practical scope

For stepped-wedge cluster-randomized trials, the same logic is extended to staggered rollout and multiple non-equivalent marginal estimands. With cluster-period treatment $h$1, cluster-period size $h$2, and potential outcomes $h$3, the paper defines four super-population treatment effects by varying the weights $h$4: horizontal-individual ATE, horizontal-cluster ATE, vertical-individual ATE, and vertical-cluster ATE. When cluster-period sizes are non-informative these coincide; when sizes are informative, they differ numerically (Fang et al., 23 Jul 2025).

The observed cluster-period mean is

$h$5

and the period-specific augmented estimator is

$h$6

aggregated over rollout periods by

$h$7

The corresponding ATE is

$h$8

This is presented as the central estimator of the paper (Fang et al., 23 Jul 2025).

The augmentation term is zero in expectation under stepped-wedge randomization whether or not the working model is correct, so consistency is preserved under arbitrary misspecification of the working outcome model. If the working model is the true conditional mean, the augmentation achieves the minimum asymptotic variance within the class considered. The paper allows individual-level LMMs, GLMMs, marginal GEEs, and cluster-period models for $h$9, and emphasizes that the model affects efficiency rather than consistency (Fang et al., 23 Jul 2025).

This stepped-wedge development also clarifies the relation between augmented standardization and existing design-based estimators. The paper states that the ANCOVA estimators of Chen and Li are special cases of the augmented construction when $\hat\alpha$0 is a linear model with period-centered covariates. It also emphasizes several limitations: no exposure-time heterogeneity, only cluster-level SUTVA, no fully integrated treatment of missing data in the main theory, and continuing small-sample challenges when the number of clusters is small (Fang et al., 23 Jul 2025).

Across these literatures, the augmented standardization estimator is best viewed as a family of estimators rather than a single formula. Its defining structure is stable: a standardization of fitted potential outcomes to the target distribution, combined with a residual correction whose role is to preserve identification under design-based or semiparametric conditions and to recover efficient first-order behavior. The exact form of the augmentation—inverse-probability weighting, Riesz-representer weighting, minimax linear balancing, or cluster-period residual correction—depends on the design and estimand, but the conceptual object remains the same.