Conditional Treatment-Averaged Causal Effect (C-TACE)

• C-TACE is the conditional treatment-averaged causal effect that quantifies treatment differences across covariate-defined strata, aligning with the standard CATE formulation.
• Methodologies like staged event trees, high-dimensional linear regression, and tree boosting are employed, each addressing unique estimation and inference challenges.
• Applications focus on estimating heterogeneous treatment effects while tackling issues such as positivity, exchangeability, and nonregular inference.

Searching arXiv for the specified papers and closely related work on conditional treatment effects. arXiv search query: (Varando et al., 30 Sep 2025) OR "Staged Event Trees for Transparent Treatment Effect Estimation" Conditional Treatment-Averaged Causal Effect (C-TACE) denotes a covariate-conditional causal contrast: the effect of treatment evaluated within strata of observed characteristics rather than averaged over the full population. In the literature represented here, the standard term for this object is the conditional average treatment effect (CATE). For binary treatment and outcome with categorical covariates, it is the contrast between $P(Y \mid \operatorname{do}(R=1), \mathbf Z)$ and $P(Y \mid \operatorname{do}(R=0), \mathbf Z)$; in potential-outcomes notation it appears as $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$; and in nonparametric notation it is written as $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$ (Varando et al., 30 Sep 2025, Kato et al., 2023, Munch et al., 14 May 2026). Across these formulations, the defining role of C-TACE is to represent treatment-effect heterogeneity that can be obscured by the average treatment effect alone (Munch et al., 14 May 2026).

1. Terminological status and mathematical formulations

Within the cited work, “C-TACE” is not the standard label. The established term is CATE, and the relevant estimand is consistently the treatment effect conditional on covariates. The staged event tree paper states that it considers “the task of estimating the ATE or its conditional version (CATE) for a binary treatment variable ($R$) and a binary outcome ($Y$) when a set of additional categorical covariates ($\bm Z$) is available” (Varando et al., 30 Sep 2025). The high-dimensional linear-regression paper likewise studies the conditional average treatment effect under the target

$f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$

and explicitly notes that it does not use the term “C-TACE” (Kato et al., 2023). The nonparametric heterogeneity paper defines the same object as

$\tau(P)(w)=\mu(P)(1,w)-\mu(P)(0,w),$

with $\mu(P)(a,w)=\mathbb{E}_P[Y\mid A=a,W=w]$ (Munch et al., 14 May 2026).

Source	Formulation	Setting
(Varando et al., 30 Sep 2025)	$P(Y \mid \operatorname{do}(R=0), \mathbf Z)$0	Binary $P(Y \mid \operatorname{do}(R=0), \mathbf Z)$1, binary $P(Y \mid \operatorname{do}(R=0), \mathbf Z)$2, categorical $P(Y \mid \operatorname{do}(R=0), \mathbf Z)$3, staged event trees
(Kato et al., 2023)	$P(Y \mid \operatorname{do}(R=0), \mathbf Z)$4	Two treatments, potential outcomes, high-dimensional linear regression
(Munch et al., 14 May 2026)	$P(Y \mid \operatorname{do}(R=0), \mathbf Z)$5	Nonparametric CATE function and heterogeneity summaries

These formulations differ in notation and modeling assumptions, but they coincide in treating the conditional effect as a function of observed covariates. In the binary-outcome case considered in staged event trees, the contrast reduces to a difference in conditional outcome probabilities within a covariate-defined subtree (Varando et al., 30 Sep 2025). In potential-outcome and outcome-regression formulations, the same object is expressed as a conditional mean difference (Kato et al., 2023, Munch et al., 14 May 2026).

2. Identification and causal assumptions

The identification conditions attached to C-TACE in these papers are standard but explicit. The staged event tree framework assumes consistency, positivity, and conditional exchangeability (Varando et al., 30 Sep 2025). It further states that, from a DAG perspective, conditional exchangeability is equivalent to assuming that the covariates $P(Y \mid \operatorname{do}(R=0), \mathbf Z)$6 are sufficient to control all paths from $P(Y \mid \operatorname{do}(R=0), \mathbf Z)$7 to $P(Y \mid \operatorname{do}(R=0), \mathbf Z)$8 except the direct link $P(Y \mid \operatorname{do}(R=0), \mathbf Z)$9 (Varando et al., 30 Sep 2025). The high-dimensional linear framework uses unconfoundedness,

$f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$0

and overlap,

$f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$1

for some $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$2 (Kato et al., 2023). The nonparametric heterogeneity paper states that under conditional exchangeability and positivity, the CATE is identifiable from observed data (Munch et al., 14 May 2026).

A central technical implication is that C-TACE is a conditional causal parameter, not merely a conditional associational contrast. The identifying role of the conditioning covariates is therefore substantive: they must be rich enough to justify exchangeability, and their support must be broad enough to justify positivity. This becomes especially important in stratified or tree-based representations where some strata can be sparse or structurally deterministic.

The staged event tree paper emphasizes positivity in a particularly concrete way: “The staged event tree clearly illustrates a violation of the positivity assumption in the group of high-risk, referred patients, as all of them receive treatment” (Varando et al., 30 Sep 2025). This makes visible a common misconception: conditional effect estimation is not guaranteed merely because covariates are observed. If a stratum assigns treatment deterministically, the conditional causal contrast is not identified there under the stated assumptions.

3. Staged event trees and localized conditional effects

In staged event trees, C-TACE is represented structurally rather than only as an abstract functional. The causal ordering is

$f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$3

with covariates first, then treatment, then outcome (Varando et al., 30 Sep 2025). A staged event tree is an $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$4-compatible rooted tree in which stages encode equality constraints on conditional probabilities: $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$5 if the corresponding nodes are in the same stage (Varando et al., 30 Sep 2025). Nodes in the same stage are colored the same, and same-colored nodes have equal conditional probabilities (Varando et al., 30 Sep 2025).

This representation gives C-TACE a localized interpretation. The paper states that “Conditional treatment effects are directly observable from the transition probabilities for the outcome variable” (Varando et al., 30 Sep 2025). In other words, once a path through the covariates $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$6 is fixed, the relevant contrast is read from the treated and untreated outcome branches in the corresponding subtree. Context-specific independencies become visible because different covariate contexts can share stages or split into distinct stages, allowing the effect of $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$7 on $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$8 to vary across $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x]$9-contexts (Varando et al., 30 Sep 2025).

The interventional semantics are encoded directly in the tree. For intervened variables $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$0 set to values $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$1, the staged-tree causal model defines

$$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$2

(Varando et al., 30 Sep 2025). For treatment-effect estimation, the paper introduces a randomized staged event tree $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$3, where treatment labels are replaced by a “naive” randomized assignment while all other conditional structure is retained. It states that this “is equivalent to standardization with respect to the levels of $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$4 ... also known as adjusting through subclassification” (Varando et al., 30 Sep 2025).

A second construction, the ps-stratified staged event tree $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$5, uses the treatment stages themselves to induce strata that are then propagated to the outcome subtree (Varando et al., 30 Sep 2025). The paper describes this as a propensity-score stratification analogue naturally encoded in the stages’ structure of the treatment variable. In both constructions, C-TACE remains a stratum-specific contrast in outcome probabilities, but the strata are obtained either from the covariate paths directly or from treatment-assignment similarity.

The same paper compares these staged-tree estimators with classical estimators including AIPW, q.model, BN_tabu, BN_pc, full, and oracle (Varando et al., 30 Sep 2025). Its stated implementation for ATE is: learn stages from data, randomize $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$6 in the fitted tree, and compute $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$7. For conditional effects, it states that confidence intervals for the outcome probabilities given treatment values across different strata can be obtained using standard inferential methods for categorical data, and that bootstrap is a possible solution to post-selection inference problems (Varando et al., 30 Sep 2025). This suggests a view of C-TACE in which interpretation and inferential uncertainty are both attached to visually identifiable local contexts.

4. High-dimensional linear models and implicit sparsity

A distinct line of work studies C-TACE in high-dimensional linear regression, again under the standard CATE formulation. The model has two treatments $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$8, two potential outcomes $$\tau(P)(w)=\mathbb{E}_P[Y\mid A=1,W=w]-\mathbb{E}_P[Y\mid A=0,W=w)$$9, covariates $R$0, and observed outcome

$R$1

(Kato et al., 2023). The estimand is

$R$2

Under the paper’s linear potential-outcome model,

$R$3

the conditional effect becomes

$R$4

(Kato et al., 2023).

The central structural assumption is implicit sparsity. The coefficients are decomposed as

$R$5

so that a common component $R$6 is shared across treatment arms while treatment-specific components differ (Kato et al., 2023). Equivalently, if $R$7,

$R$8

and therefore

$R$9

The common terms cancel in the contrast, leaving only the treatment-specific differences (Kato et al., 2023). The paper emphasizes that sparsity is thus imposed on the CATE contrast rather than on either potential-outcome regression separately.

The proposed estimator penalizes the difference between the two treatment-specific coefficient vectors: $Y$0 (Kato et al., 2023). Rewriting with $Y$1 yields a two-stage procedure: estimate $Y$2 by an interpolating or minimum-norm least-squares estimator, then estimate the contrast by Lasso, and finally set

$Y$3

(Kato et al., 2023).

The theoretical results are oracle inequalities and consistency for the contrast parameter rather than for $Y$4 and $Y$5 individually. In the fixed-design sub-Gaussian setting, the paper proves high-probability bounds of the form

$Y$6

(Kato et al., 2023). Under random design with sub-Gaussian covariates and $Y$7, it further gives

$Y$8

implying $Y$9 (Kato et al., 2023). A plausible implication is that this framework is especially tailored to settings in which effect modification is sparse even when the arm-specific outcome surfaces are not.

5. Potential-outcome prediction by tree boosting

C-TACE can also be estimated by directly predicting both potential outcomes and differencing them. C-XGBoost is a tree-boosting model built on XGBoost and adapted to causal inference by learning the pair of conditional outcomes

$\bm Z$0

for each instance $\bm Z$1 (Kiriakidou et al., 2024). The conditional effect estimate is then

$\bm Z$2

which the paper describes as the standard form of a conditional average treatment effect estimate at $\bm Z$3, often called CATE, ITE, or heterogeneous treatment effect depending on context (Kiriakidou et al., 2024).

The model uses a treatment-masked squared-error loss: $\bm Z$4 with

$\bm Z$5

(Kiriakidou et al., 2024). Only the factual branch is penalized for each sample, while the model maintains both potential-outcome predictors. The paper states that “at each iteration the C-XGBoost builds a multi-output tree with the size of leaf equals to the number of targets” (Kiriakidou et al., 2024). Reported implementation settings include $\bm Z$6, $\bm Z$7, $\bm Z$8, $\bm Z$9, and $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$0 (Kiriakidou et al., 2024).

Evaluation distinguishes average and heterogeneous performance. The ATE metric is

$f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$1

and heterogeneous-effect accuracy is assessed by

$f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$2

(Kiriakidou et al., 2024). Since PEHE measures error in effect differences, it is directly tied to C-TACE quality.

The empirical analysis uses a Synthetic semi-synthetic collection from Louizos et al. with 5000 samples and 1000 covariates, and ACIC data from the 2018 Atlantic Causal Inference Conference benchmark (Kiriakidou et al., 2024). Baselines are R-Forest, C-Forest, Dragonnet, and $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$3NN-Dragonnet. Statistical validation uses Dolan and Moré performance profiles, the Friedman Aligned-Ranks (FAR) test, and the Finner post-hoc test (Kiriakidou et al., 2024). On Synthetic datasets, C-XGBoost is reported as best on both $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$4 and PEHE, achieving the best score on 76% of datasets for $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$5 and on 58% for PEHE; on ACIC datasets it is best or tied-best on PEHE and achieves 62.5% of datasets with the best PEHE score (Kiriakidou et al., 2024). These findings support the interpretation of C-TACE as a prediction target recoverable from paired potential-outcome models, not only from explicit causal-function estimators.

6. Distributional summaries, uncertainty, and nonregular inference

A recurring difficulty with C-TACE is that a multivariate conditional effect surface can be hard to communicate when it depends on many covariates. One response is to summarize the distribution of the conditional effect in the population rather than to inspect the full function directly. The nonparametric heterogeneity paper defines the sublevel-set probability curve

$f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$6

and describes it as the cumulative distribution function of the random variable $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$7 (Munch et al., 14 May 2026). Each value $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$8 is the proportion of individuals whose expected treatment effect does not exceed threshold $f_P(x)=\mathbb{E}_P[Y^1-Y^0\mid X=x],$9; in particular, $\tau(P)(w)=\mu(P)(1,w)-\mu(P)(0,w),$0 is the fraction of the population for which treatment is not expected to be beneficial (Munch et al., 14 May 2026).

This construction does not replace C-TACE; it summarizes its heterogeneity. If treatment effects are nearly constant, $\tau(P)(w)=\mu(P)(1,w)-\mu(P)(0,w),$1 is almost a step function concentrated around the ATE; if effects vary widely, the curve spreads more smoothly (Munch et al., 14 May 2026). The paper’s technical contribution is that pointwise evaluation $\tau(P)(w)=\mu(P)(1,w)-\mu(P)(0,w),$2 is not pathwise differentiable under a nonparametric model, so standard regular asymptotically linear inference is not directly available (Munch et al., 14 May 2026).

To address this, the paper proposes three estimators. The first is a cross-fitted plug-in estimator,

$\tau(P)(w)=\mu(P)(1,w)-\mu(P)(0,w),$3

(Munch et al., 14 May 2026). The second is a Grenander-type estimator, constructed by first working with the antiderivative

$\tau(P)(w)=\mu(P)(1,w)-\mu(P)(0,w),$4

estimating it by a cross-fitted one-step procedure, and then differentiating the greatest convex minorant: $\tau(P)(w)=\mu(P)(1,w)-\mu(P)(0,w),$5 (Munch et al., 14 May 2026). The theory yields cube-root asymptotics involving Chernoff’s distribution. The third is a debiased machine-learning estimator for a best piecewise linear approximation

$\tau(P)(w)=\mu(P)(1,w)-\mu(P)(0,w),$6

which is pathwise differentiable under continuity at knot points and therefore amenable to standard debiased machine-learning inference (Munch et al., 14 May 2026).

Uncertainty quantification is also a major issue in staged event trees. The same literature states that “appropriate ATE and CATE estimators should be accompanied by standard error estimates” and notes that post-selection inference problems are inevitable in the proposed approach; it identifies bootstrap as a possible solution and also notes that standard inferential methods for categorical data can produce confidence intervals for outcome probabilities within strata, thereby enabling formal inference on CATE (Varando et al., 30 Sep 2025). Taken together, these results show that C-TACE is not only an estimand of heterogeneity but also a locus of nontrivial inferential challenges: conditioning sharpens substantive interpretation while simultaneously complicating identification, support conditions, and asymptotic theory.