Weighted Average Treatment Effects

• WATEs are causal estimands that generalize the average treatment effect by averaging heterogeneous treatment effects with user-defined weights, defining specific target populations.
• Different weight functions (e.g., for ATT, ATC, and ATO) emphasize various subpopulations, permitting tailored causal interpretation and addressing overlap challenges.
• Robust estimation strategies, including inverse probability weighting and doubly robust methods, enhance precision and efficiency while managing model uncertainty.

Weighted Average Treatment Effects (WATEs) are causal estimands that generalize the average treatment effect by averaging heterogeneous treatment effects with a chosen set of weights. In the standard potential-outcomes setup, the central idea is that different causal questions correspond to different target populations induced by a nonnegative tilting function or by a function of the propensity score, so the estimand is not a single fixed “ATE” but a family of effects indexed by weighting choice. This perspective encompasses familiar targets such as the ATE, ATT, ATC, and overlap-population effects, and it also appears in broader frameworks such as generalized average treatment effects, rank-weighted treatment-effect metrics, weighted survival contrasts, and variance-weighted panel estimands (Tao et al., 2018, Tian et al., 10 Jun 2026, Kallus et al., 2019).

1. Formal definitions and parameterizations

A standard definition writes the WATE as

$$\tau_h=\frac{E\!\left[h(\mathbf X)\{Y(1)-Y(0)\}\right]}{E\{h(\mathbf X)\}},$$

where the nonnegative target function $h(\mathbf X)$ defines a target population with density proportional to

$$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$

In this form, WATE is the average causal effect among people represented by the weight function $h$ (Tao et al., 2018).

A closely related parameterization specializes the weight to a function of the propensity score $e(X)=P(A=1\mid X)$. With $\tau(X)=E\{Y(1)-Y(0)\mid X\}$ and $h(X)=\lambda\{e(X)\}$,

$$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$

so the choice of $\lambda$ determines which subpopulation is emphasized (Wang et al., 18 Sep 2025, Liu et al., 31 Mar 2026). In this sense, WATEs are not a different causal model; they are different target populations induced by different weights (Tian et al., 10 Jun 2026).

A finite-sample version appears in the generalized average treatment effect framework:

$$\tau_V=\sum_{i=1}^n V_i \bigl(g_1(X_i)-g_0(X_i)\bigr),$$

where $h(\mathbf X)$0 and $h(\mathbf X)$1 are user-chosen weights defining the target population. In that formulation, WATEs are a special case of a broader GATE family, and $h(\mathbf X)$2 “is chosen to target any of the common causal estimands” (Kallus et al., 2019).

These formulations make explicit a structural distinction between the estimand and the estimator. In the GATE notation, $h(\mathbf X)$3 defines the estimand, whereas $h(\mathbf X)$4 defines the estimation weights; the relationship between the two is the object of optimization in kernel optimal matching (Kallus et al., 2019). A plausible implication is that WATE methodology is best understood as joint work on target selection and estimation, not merely as reweighting for balance.

2. Canonical target populations and special cases

The most common WATEs are obtained by particular choices of $h(\mathbf X)$5 or $h(\mathbf X)$6. These choices determine the target population and therefore the scientific question being answered.

Estimand	Weight function	Population emphasized
ATE	$h(\mathbf X)$7	Full study population
ATT	$h(\mathbf X)$8 or $h(\mathbf X)$9	Treated subpopulation
ATC	$$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$0 or $$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$1	Control subpopulation
ATO / overlap weights	$$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$2	Overlap population
ATM / matching weights	$$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$3	Matched population
ATEN / entropy weights	$$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$4	Entropy-weighted population
Trimming	$$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$5	Region of adequate overlap

These special cases are explicitly identified across the recent WATE literature (Li et al., 11 Aug 2025, Liu et al., 31 Mar 2026, Wang et al., 18 Sep 2025).

Several extensions refine this menu. In the GATE framework, SATE, TATE, CATE, and OWATE correspond to different choices of $$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$6. For CATE, weights restrict attention to units in a subgroup $$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$7. For OWATE, $$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$8 is chosen to produce a population with good overlap and thus better estimability, and the authors emphasize that “OWATE $$\frac{f(\mathbf X)h(\mathbf X)}{E\{h(\mathbf X)\}}.$$9 is variable and chosen so to be most easily estimable” (Kallus et al., 2019).

A treated-oriented variant appears in the overlap weighted average treatment effect on the treated. In that construction,

$h$0

but the target remains ATT-like because the treated units are the reference group and the modified control weight becomes

$h$1

The stated motivation is to keep the “treated as reference” logic while avoiding the $h$2 instability of standard ATT weighting (Liu et al., 2023).

WATEs also include rank-based targets. Rank-weighted average treatment effects are WATEs in which the weights depend on the rank induced by a score $h$3, rather than directly on covariates or score values. In that setting,

$h$4

so the weight is determined by the rank percentile $h$5 (Yadlowsky et al., 2021).

3. Identification, overlap, and interpretation

The canonical identifying assumptions for WATEs are the familiar potential-outcomes conditions: consistency or SUTVA, strong ignorability or unconfoundedness, and positivity (Tao et al., 2018, Orihara, 2023). In the GATE framework for transport or sample generalization, these are supplemented by ignorable treatment assignment,

$h$6

and ignorable sample assignment,

$h$7

together with non-interference and positivity (Kallus et al., 2019).

Overlap occupies a central interpretive role. For the overlap estimand,

$h$8

the weight

$h$9

is proportional to the conditional variance of treatment given covariates,

$e(X)=P(A=1\mid X)$0

It is largest when $e(X)=P(A=1\mid X)$1 and vanishes near $e(X)=P(A=1\mid X)$2 or $e(X)=P(A=1\mid X)$3, so it emphasizes covariate regions where treatment and control both occur with meaningful frequency and downweights extreme-propensity units where overlap is weak (Tian et al., 10 Jun 2026).

This emphasis yields a sharp comparative interpretation. If

$e(X)=P(A=1\mid X)$4

is nondecreasing, then

$e(X)=P(A=1\mid X)$5

whereas if $e(X)=P(A=1\mid X)$6 is nonincreasing, the inequalities reverse. Under the linear special case

$e(X)=P(A=1\mid X)$7

$e(X)=P(A=1\mid X)$8 is a convex combination of ATT and ATC (Tian et al., 10 Jun 2026). This suggests that overlap weighting is not only a stability device but also an estimand with a structural location between treated- and control-targeted effects.

The same logic extends beyond the simplest binary-treatment setting. In weighted local average treatment effects with a binary instrument and a binary treatment, local versions of ATE, ATT, ATC, and ATO are defined by weighting the complier effect $e(X)=P(A=1\mid X)$9 with both the treatment propensity score and the principal score $\tau(X)=E\{Y(1)-Y(0)\mid X\}$0, and analogous bracketing results hold under monotonicity of the local propensity-score conditional mean effect $\tau(X)=E\{Y(1)-Y(0)\mid X\}$1 (Tian et al., 10 Jun 2026).

4. Estimation strategies and efficiency theory

The basic inverse probability weighting estimator for the propensity-score-weighted class is

$\tau(X)=E\{Y(1)-Y(0)\mid X\}$2

with

$\tau(X)=E\{Y(1)-Y(0)\mid X\}$3

This estimator is the direct IPW template for WATEs such as ATE, ATT, and ATO (Orihara, 2023).

Augmented estimators incorporate outcome regression. For a fixed target function $\tau(X)=E\{Y(1)-Y(0)\mid X\}$4, the augmented inverse probability weighting estimator $\tau(X)=E\{Y(1)-Y(0)\mid X\}$5 is doubly robust when $\tau(X)=E\{Y(1)-Y(0)\mid X\}$6 is known or correctly specified: it is consistent if the propensity score model is correct or if the conditional mean model is correct (Tao et al., 2018). A second estimator, $\tau(X)=E\{Y(1)-Y(0)\mid X\}$7, extends double robustness to the case $\tau(X)=E\{Y(1)-Y(0)\mid X\}$8, thereby covering ATT and ATC (Tao et al., 2018). More recent semiparametric analysis sharpens the boundary of this phenomenon: DR is essentially limited to linear $\tau(X)=E\{Y(1)-Y(0)\mid X\}$9 weights, whereas ATO, ATEN, and some beta-weighted estimands admit rate doubly robust estimators under product-rate conditions on nuisance estimators (Wang et al., 18 Sep 2025).

The efficient influence function is the organizing device for this modern theory. In the general WATE model of $h(X)=\lambda\{e(X)\}$0, one-step TMLE along the universal least favorable path is shown to be well-defined, to solve the estimating equation in finite time, and to yield an asymptotically efficient estimator under explicit regularity conditions on the weight function and initialization (Liu et al., 31 Mar 2026). Cross-fitted DML estimators similarly achieve asymptotic normality under orthogonal-score constructions and nuisance-rate conditions, without Donsker restrictions (Wang et al., 18 Sep 2025).

Kernel Optimal Matching takes a different route. For the weighted estimator $h(X)=\lambda\{e(X)\}$1, the conditional mean squared error decomposes into a squared imbalance term plus a variance term:

$h(X)=\lambda\{e(X)\}$2

KOM therefore chooses weights that reduce imbalance over an RKHS function class while controlling precision. In the fixed-$h(X)=\lambda\{e(X)\}$3 problem, it minimizes

$h(X)=\lambda\{e(X)\}$4

and with a kernel Gram matrix this becomes a linearly constrained convex-quadratic optimization problem. If $h(X)=\lambda\{e(X)\}$5 is itself optimized jointly with $h(X)=\lambda\{e(X)\}$6, the result is the kernel-optimal WATE or OWATE, formally the target population whose causal contrast is most easily estimable under the chosen kernel class (Kallus et al., 2019).

A further alternative replaces inverse propensity weights with matching weights. Augmented match weighted estimators for ATE and ATT use the factor $h(X)=\lambda\{e(X)\}$7 in place of the inverse-probability factor, retain double robustness and local efficiency, and exploit an unfixed-$h(X)=\lambda\{e(X)\}$8 regime to restore asymptotic smoothness and bootstrap validity (Xu et al., 2023).

5. Extensions beyond binary single-timepoint treatment

The weighted-average logic extends naturally to more general treatment regimes. For binary-valued $h(X)=\lambda\{e(X)\}$9, the average linear regression function

$$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$0

recovers the ATE. For multi-valued mutually exclusive treatments, it recovers the vector of ATEs. For continuous or mixed treatments under a linear heterogeneous potential response, it equals the average partial effect, and under a general nonlinear response with scalar continuous $$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$1 it becomes a weighted average of gradients (Graham et al., 2018). This places WATE-type reasoning inside semiparametric regression, not only binary-treatment program evaluation.

In survival and competing-risks settings, the target is a weighted cumulative causal effect rather than a weighted mean difference at a single endpoint. The weighted RMST difference is built from

$$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$2

and the competing-risks analogue uses weighted cause-specific restricted mean time lost. Cross-fitted one-step DML estimators based on efficient influence functions are shown to be consistent, asymptotically linear, and efficient, and under the stated homoskedasticity condition the overlap weight

$$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$3

minimizes asymptotic variance in the balancing-weight class (Xu et al., 2023).

Longitudinal work introduces stochastic flip interventions. In single-timepoint data, the interventional flip effect

$$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$4

is exactly the WATE

$$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$5

With longitudinal data, the same construction permits weighting or trimming on non-baseline covariates and remains identifiable under arbitrary positivity violations as long as the weight is zero whenever the target propensity score is zero (McClean et al., 10 Jun 2025).

Some WATEs are evaluative rather than target-defining. RATE metrics assess how well a prioritization rule ranks units by treatment benefit. The targeting operator characteristic compares the effect among the top $$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$6 fraction of a ranking with the population ATE, and RATE is a weighted average of that curve. The resulting estimand is a WATE whose weight depends on rank percentile rather than directly on covariates (Yadlowsky et al., 2021).

The same pattern appears in panel models with latent heterogeneity. For a single regressor, both the principal components estimator and the interactive fixed effects estimator converge to

$$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$7

a variance-weighted average of unit-time-specific treatment effects (Juodis et al., 20 Apr 2026).

6. Variance estimation, diagnostics, and recurring controversies

Inference for WATEs is substantially more delicate than point estimation. For the common augmented estimator, currently used variance methods include the standard bootstrap, post-weighting bootstrap, sandwich estimation, and wild bootstrap. Their performance depends on the estimand and on whether the propensity score is ancillary to that estimand. In simulation studies, BOOT II and wild bootstrap are attractive for ATE, whereas standard bootstrap often remains the safest general-purpose choice for ATT, ATO, ATM, and ATEN, especially under treatment-effect heterogeneity or weak overlap (Li et al., 11 Aug 2025).

A central asymptotic issue is whether confidence intervals based on the usual robust sandwich variance that ignores propensity-score uncertainty are conservative. For the IPW estimator of

$$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$8

the exact asymptotic variance contains an outcome-variability term $$\gamma=\frac{E[\lambda\{e(X)\}\tau(X)]}{E[\lambda\{e(X)\}]},$$9 plus terms arising from estimation of the propensity score. The “simple CI” based only on the first term is always conservative for the ATE because $\lambda$0, but not uniformly so for ATT and ATO. For ATT and ATO, conservativeness holds under the Fisher sharp null, the conditional null, and homogeneous treatment effects; outside those cases, the paper proposes moment-inequality criteria to assess whether the omitted correction terms are nonpositive (Orihara, 2023).

Weak overlap also changes the tail behavior of weighted estimators. When $\lambda$1 approaches $\lambda$2 or $\lambda$3, the IPW variable

$\lambda$4

can become heavy tailed, possibly with infinite variance and slower-than-$\lambda$5 convergence. A tail-trimmed IPW estimator therefore trims the extreme realizations of $\lambda$6 itself rather than trimming covariates or estimated propensity scores, and then estimates and removes trimming bias under asymmetric heavy tails (Hill et al., 2024). This suggests a broader principle: robustness should target the weighted pseudo-outcome or influence variable, not only the propensity score.

A further controversy concerns “fixes” for poor positivity in ATT analysis. Trimming and truncation are often described as stabilizing devices, but the treated-oriented WATE literature argues that they alter the estimand by introducing a tilting function $\lambda$7, thereby producing scaled ATT or WATT estimands rather than the original ATT. The overlap weighted average treatment effect on the treated is proposed as a threshold-free alternative that keeps the treated-reference structure while replacing the unstable inverse-odds factor with $\lambda$8 (Liu et al., 2023).

Diagnostic work increasingly focuses on the relationship between heterogeneity and assignment. A recommended graphical device is the CP-plot, the plot of the estimated CATE against the estimated propensity score. Its role is to assess whether the monotonicity condition behind the bracketing relationship

$\lambda$9

or its reverse is plausible over the overlap region (Tian et al., 10 Jun 2026).

Privacy constraints add another inferential layer. For binary outcomes, differentially private WATE estimation has been developed for ATE, ATT, and ATC using truncated propensity scores, subsample-and-aggregate, Laplace noise, and Bayesian post-processing. In that setting, the global sensitivity of the point estimator is bounded by $$\tau_V=\sum_{i=1}^n V_i \bigl(g_1(X_i)-g_0(X_i)\bigr),$$0, and the resulting procedure is $$\tau_V=\sum_{i=1}^n V_i \bigl(g_1(X_i)-g_0(X_i)\bigr),$$1-differentially private while still delivering point and interval estimates for binary-outcome WATEs (Guha et al., 2024).

Across these developments, one theme is constant: WATEs are not interchangeable. ATT, ATC, ATO, ATM, ATEN, trimmed effects, rank-weighted effects, and kernel-optimal effects answer different questions about different target populations or rankings. The main methodological consequence is that estimation, efficiency, and inference cannot be separated from the choice of weight function.