Group Average Treatment Effects (GATEs)

• Group Average Treatment Effects (GATEs) are subgroup-specific averages of individual treatment effects that summarize heterogeneity by conditioning on moderators, score bins, or rank thresholds.
• They are estimated using methods such as doubly robust pseudo-outcomes, matching with covariate adjustment, or local regression techniques to ensure valid causal interpretation.
• Interpreting GATEs requires careful consideration of subgroup definitions and aggregation methods, as mis-specification can lead to bias and misleading conclusions.

Group Average Treatment Effects (GATEs) are subgroup-specific averages of treatment effects, commonly written in the modern literature as $\mathbb{E}[Y(1)-Y(0)\mid G=g]$. In contemporary causal machine learning and program evaluation, the subgroup may be indexed by a moderator $Z$, by score-defined impact groups, or by rank-based score thresholds. Relative to individualized or fully covariate-conditioned effects, GATEs provide a lower-dimensional summary of treatment-effect heterogeneity, but their meaning depends sharply on the subgroup definition, the conditioning set used for identification, and the way individual-level effects are aggregated (Bearth et al., 2024, Jacob, 2019, Li et al., 2023).

1. Definitions, parameterizations, and relation to adjacent estimands

A standard formalization appears in the moderation literature, where the individualized treatment effect target is

$\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$

and the GATE is

$$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$

In this formulation, the group is defined by the moderator $Z$, and the GATE averages the fully conditioned treatment effect over the covariate distribution that naturally occurs within that subgroup (Bearth et al., 2024).

A second parameterization sorts units into ex ante fixed impact groups using an estimated treatment-effect score. In observational studies, this is written as

$$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$

where $K$ is fixed in advance and the intervals $I_k$ are quantile-based bins of an out-of-sample score. This formulation presents GATEs as a practical compromise between the global average treatment effect and a fully nonparametric CATE, while also limiting the danger of finding spurious heterogeneity due to small subgroups in the CATE (Jacob, 2019).

A third formulation is the sorted GATES curve in randomized experiments,

$\Psi(p)=E[\psi_i\mid F(S_i)\ge p],$

where $S_i=f(X_i)$ is a treatment-prioritization score and $Z$0 indexes a cutoff in the score distribution. Here the group is the upper tail of the score distribution rather than a prespecified demographic or institutional subgroup (Li et al., 2023).

The treatment-status averages ATT and ATU are also GATE-like objects. With binary treatment $Z$1,

$Z$2

These are subgroup average effects where the subgroup is defined by treatment status rather than by an ex ante moderator such as age, region, or score bins (Słoczyński, 2018).

2. Identification and causal interpretation

In observational studies, identification of GATEs typically rests on unconfoundedness given the full covariate vector $Z$3, not merely the subgroup variable $Z$4. In the matching-based nonparametric framework,

$Z$5

with

$Z$6

and under SUTVA,

$Z$7

The identified GATE is then

$Z$8

where $Z$9 may be a subvector of $\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$0 or any real-valued function of $\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$1 (Wu et al., 25 Aug 2025).

Grouped observational data with unobserved group-level heterogeneity require an additional layer of adjustment. Under group-level unconfoundedness,

$\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$2

and an exponential-family model for the within-group distribution of $\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$3, aggregate balancing statistics

$\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$4

become sufficient to absorb latent group heterogeneity, yielding

$\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$5

This identifies overlap-restricted effects of the form

$\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$6

rather than a separate nonparametric treatment effect for each raw group label when groups are small (Arkhangelsky et al., 2018).

A recurring identification point in the GATE literature is that the subgroup-defining variable and the confounder-adjustment variable play different roles. The subgroup variable $\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$7 determines which heterogeneous effect is being summarized; the full covariate vector $\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$8 determines whether the subgroup-specific effect is causally identified. This distinction is explicit in both the matching-based GATE estimator and the moderation literature’s BGATE construction (Wu et al., 25 Aug 2025, Bearth et al., 2024).

3. Estimation strategies in observational studies

A prominent observational estimator is the doubly orthogonal GATES procedure. Its first stage constructs a doubly robust pseudo-outcome,

$\tau(x,z)=E[Y_i^1-Y_i^0\mid Z_i=z,X_i=x],$9

then regresses this transformed outcome on covariates to obtain a smoothed score $$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$0. In the second stage, units are sorted into $$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$1 groups using quantiles of $$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$2, and group effects are estimated via an orthogonalized partial linear projection,

$$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$3

The resulting coefficients are interpreted as a best linear predictor for effect heterogeneity based on impact groups, and the paper recommends repeated sample splitting and median aggregation as a bagging-type stabilization device (Jacob, 2019).

A distinct nonparametric route is matching on the full confounder vector $$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$4, followed by smoothing on the low-dimensional group variable $$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$5. The matching estimator imputes potential outcomes using nearest neighbors from the opposite treatment arm and then applies local constant regression,

$$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$6

The bias-corrected version augments the matched imputation with outcome-regression adjustments $$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$7 and $$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$8, and the paper shows consistency, double robustness, and asymptotic normality for the bias-corrected estimator under stated conditions. The associated software is the R package MatchGATE (Wu et al., 25 Aug 2025).

The same broad logic extends to staggered-adoption panel settings. There the target becomes the group-time conditional average treatment effect

$$\theta^G(z)=E[Y_i^1-Y_i^0\mid Z_i=z]=E[\tau(X_i,z)\mid Z_i=z].$$9

identified by a doubly robust conditional DiD signal and estimated by local polynomial regression in a scalar pre-treatment covariate $Z$0. The paper develops uniform confidence bands for the entire function $Z$1 and provides implementation in the R package didhetero (Imai et al., 2023).

4. Sorted groups, subgroup discovery, and inference in randomized experiments

Score-sorted GATEs occupy a distinct place in the literature. In observational studies, score-defined GATES use fixed quantile groups $Z$2, with the number of groups chosen in advance in a manner the paper describes as analogous to pre-analysis plans. This yields an interpretable low-dimensional summary with coefficients, p-values, and confidence intervals from a linear model, rather than a fully nonparametric CATE surface (Jacob, 2019).

In randomized experiments, subgroup discovery based on an estimated prioritization score creates a multiple-testing problem because the same data are used to search over subgroup cutoffs and to evaluate subgroup effects. The design-based solution is uniform inference for the whole sorted GATES curve

$Z$3

The paper develops one-sided uniform confidence bands that hold simultaneously over all $Z$4, so that a subgroup can be chosen adaptively—by maximizing a lower confidence bound or by selecting the largest subgroup whose effect exceeds a threshold $Z$5—without invalidating the guarantee. The validity relies only on random sampling of units, complete randomization of treatment, continuity conditions on the score, and a finite second moment condition on individual treatment effects (Li et al., 2023).

The practical distinction is consequential. Fixed quantile GATES are primarily descriptive summaries of heterogeneity across coarse score strata. Uniform-band methods instead support subgroup identification with a statistical guarantee after searching across many candidate cutoffs. Both are GATE procedures, but they answer different questions: one emphasizes stable reporting of heterogeneous effects; the other emphasizes post-selection-valid subgroup discovery (Jacob, 2019, Li et al., 2023).

5. Moderation analysis and balanced subgroup comparison

Raw differences in GATEs can be difficult to interpret as moderation effects because each subgroup average is taken over a different covariate distribution. In the binary-moderator setup,

$Z$6

compares

$Z$7

so both the conditional effect function and the averaging distribution differ across groups. To address this, the balanced group average treatment effect is defined as

$Z$8

where $Z$9 is a set of a priori determined balancing covariates whose distribution is held fixed across groups. The associated contrast

$$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$0

is intended to isolate subgroup differences in treatment effects net of distributional differences in $$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$1 (Bearth et al., 2024).

Estimation is based on a two-stage DML procedure with cross-fitting. The first stage constructs the doubly robust pseudo-outcome

$$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$2

and the second stage regresses this pseudo-outcome on $$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$3 within moderator groups using

$$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$4

along with the moderator propensity

$$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$5

Under the stated overlap, nuisance-consistency, risk-decay, and stability conditions, the estimator is $$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$6-consistent and asymptotically normal (Bearth et al., 2024).

The same paper distinguishes associational moderation from causal moderation. BGATE and $$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$7BGATE are offered as more interpretable moderation-analysis tools, but causal moderation requires stronger assumptions on the moderator itself and leads to the causal balanced group average treatment effect parameter

$$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$8

The empirical application on Swiss labor-market programs illustrates the importance of balancing: a significant raw nationality-based GATE difference largely disappears once age, gender, marital status, labor market history, mother tongue, and then all covariates are sequentially balanced (Bearth et al., 2024).

6. Aggregation bias and interpretational pitfalls

A central warning in the GATE literature is that common regression summaries may not target any researcher-chosen subgroup average under heterogeneous treatment effects. In the binary-treatment linear model

$$E[\tau(X)\mid G_k], \qquad G_k:=\{S(X)\in I_k\}, \quad k=1,\ldots,K,$$9

the OLS treatment coefficient satisfies

$K$0

under the paper’s ignorability-in-mean and linearity conditions, with weights inversely related to group size. In the equal-within-group-variance special case,

$K$1

so the treated share $K$2 weights ATU, not ATT. The paper’s core lesson is that OLS is an implicit, nontransparent aggregation of subgroup effects, and smaller groups get larger implicit weights (Słoczyński, 2018).

That result matters directly for GATE interpretation. ATT and ATU are already subgroup average effects, and the paper shows that adding or changing covariates alters the implicit OLS weights through the linear propensity score $K$3. The recommended diagnostics are $K$4 for bias relative to ATT and $K$5 for bias relative to ATE; software is provided in the R and Stata package hettreatreg (Słoczyński, 2018).

A second pitfall arises even when treatment is randomized and individualized CATE models are well specified. Aggregating predicted CATEs to the group level does not, in general, recover the corresponding experimentally identified GATE. On the additive scale, the paper defines group bias as

$K$6

where $K$7 is the model-implied GATE and $K$8 is the true experimentally identified GATE. The paper develops an asymptotically normal estimator, a Wald test for $K$9, and shrinkage-based bias correction with closed-form expressions. The general message is that subgroup causal interpretation is not inherited automatically from individualized causal prediction; it must be audited at the group level (Persson et al., 23 Feb 2026).

7. Extensions and competing uses of the acronym

Several adjacent literatures extend the subgroup-average logic without estimating standard covariate-defined GATEs. In IV settings with endogenous treatment, the relevant objects are grouped local average treatment effects: groups of instruments with equal first-stage propensities correspond to common complier strata and therefore identify the same LATE. This is a grouped-LATE analogue of GATEs, not a standard $I_k$0 estimand (Apfel et al., 2022).

Longitudinal and policy-based extensions go further. The generalized ATT framework defines

$I_k$1

so the subgroup is defined by natural treatment values under a longitudinal modified treatment policy rather than by baseline covariates. This is GATE-like in the sense of subgroup-average causal estimation, but the group is treatment-defined rather than covariate-defined (Susmann et al., 2024). In score-threshold assignment settings, a structural model with treatment rule $I_k$2 yields ATT and estimated ITE/CATE objects that can then be aggregated into score-bin or demographic GATEs, although the formal theory in that paper is developed for ATT rather than for arbitrary GATEs (Wibisono et al., 23 Apr 2025).

The acronym itself is overloaded. In some work, GATE means Generalized Average Treatment Effect,

$I_k$3

a weighted estimand that includes subgroup effects as special cases but is not the same as Group Average Treatment Effects (Kallus et al., 2019). In network-interference and marketplace settings, GATE instead means Global Average Treatment Effect, the contrast between an all-treated and an all-control world under interference (Han et al., 2023, Yu et al., 26 May 2026). For the subgroup-heterogeneity literature, the relevant meaning remains the modern one: averaging treatment effects within a chosen subgroup, whether that subgroup is defined by moderators, score bins, or sorted score thresholds.