Honest Causal Tree Architecture

• Honest Causal Tree Architecture is a recursive partitioning method that estimates heterogeneous treatment effects by splitting data into separate samples for tree building and effect estimation.
• The approach employs an honest sample-splitting mechanism to eliminate adaptive bias, yielding unbiased leaf-level estimates with valid confidence intervals.
• It faces uniform convergence limitations due to boundary cell effects, making it well-suited for integrated MSE minimization in moderate-depth settings.

The Honest Causal Tree (CT-H) architecture is a recursive partitioning methodology designed for the estimation and inference of heterogeneous treatment effects in experimental and observational studies. Distinguished by its use of data-splitting for "honesty," this architecture prevents adaptive overfit by separating the sample into disjoint subsets for tree construction and effect estimation. CT-H methods are founded on the potential outcomes framework and achieve unbiased leaf-level CATE estimates with asymptotically valid confidence intervals, but are subject to fundamental limits in uniform convergence rates due to partitioning behavior and boundary cell effects (Cattaneo et al., 14 Sep 2025, Athey et al., 2015).

1. Formal Definition and Problem Setup

The CT-H framework observes an i.i.d. sample $D = \{(y_i, d_i, x_i) : i=1, \ldots, n\}$ consisting of covariates $x_i \in \mathbb{R}^p$, binary treatment $d_i \in \{0,1\}$, and outcome $y_i = d_i y_i(1) + (1-d_i) y_i(0)$. It is formulated within the Rubin-Neyman potential outcomes model, targeting the conditional average treatment effect (CATE) $\tau(x) = \mathbb{E}[y(1) - y(0) | x]$ (Athey et al., 2015).

Key causal inference assumptions:

• Unconfoundedness: $d_i \perp \{y_i(0), y_i(1)\}\ |\ x_i$
• Overlap: $0 < \mathbb{P}(d_i=1|x_i=x) < 1$ for all $x$

The central statistical objective is the estimation of $\tau(x)$ for arbitrary $x$, yielding a partition of $x_i \in \mathbb{R}^p$0 into leaves with piecewise-constant CATEs.

2. Honest Sample Splitting Mechanism

The hallmark of CT-H is its "honest" partitioning of data into two non-overlapping subsamples:

• Training subsample $x_i \in \mathbb{R}^p$1: Used exclusively for tree construction (partitioning the covariate space).
• Estimation subsample $x_i \in \mathbb{R}^p$2: Used solely for within-leaf estimation of CATEs and associated variances.

This separation eliminates adaptive bias in treatment effect estimation. Unlike no-sample-splitting (NSS) variants, CT-H prevents the tree structure from overfitting to outcome idiosyncrasies of the estimation data. Leaf estimates are thus conditionally unbiased with respect to the partition (Cattaneo et al., 14 Sep 2025, Athey et al., 2015).

3. Tree Construction and Splitting Criteria

Tree-formation proceeds recursively on $x_i \in \mathbb{R}^p$3, with splits determined by maximizing an honest-splitting criterion over all variable–threshold pairs $x_i \in \mathbb{R}^p$4:

• Difference-in-means (DIM):

$x_i \in \mathbb{R}^p$5

where $x_i \in \mathbb{R}^p$6 (difference of sample means for treated/controls in node $x_i \in \mathbb{R}^p$7).

• Inverse-probability-weighted (IPW):

$x_i \in \mathbb{R}^p$8

with $x_i \in \mathbb{R}^p$9 and $d_i \in \{0,1\}$0.

• Sum-of-squared-errors (SSE):

Nodes are split to minimize total within-node treatment/outcome regression squared error.

Splitting proceeds until a minimum node size or maximum tree depth is reached.

Split Criterion	Gain Function	Estimator Used
DIM	$d_i \in \{0,1\}$1	Difference-in-means
IPW	$d_i \in \{0,1\}$2	Inverse-probability weighting
SSE	Minimize SSE	Linear regression in leaves

After growing the tree, $d_i \in \{0,1\}$3 is used to estimate effects in each leaf, independent of how the partition was chosen (Cattaneo et al., 14 Sep 2025).

4. Estimation of Leaf-Wise Treatment Effects and Standard Errors

In each terminal node ("leaf") $d_i \in \{0,1\}$4, CT-H computes CATE and attaches standard error estimates:

• DIM: $d_i \in \{0,1\}$5
• IPW: $d_i \in \{0,1\}$6
• SSE: Estimate $d_i \in \{0,1\}$7 by OLS of $d_i \in \{0,1\}$8 on $d_i \in \{0,1\}$9 in $y_i = d_i y_i(1) + (1-d_i) y_i(0)$0; $y_i = d_i y_i(1) + (1-d_i) y_i(0)$1

Standard errors approximate the variance in each group within the leaf. E.g., for DIM,

$y_i = d_i y_i(1) + (1-d_i) y_i(0)$2

where $y_i = d_i y_i(1) + (1-d_i) y_i(0)$3 is sample variance of $y_i = d_i y_i(1) + (1-d_i) y_i(0)$4 for group $y_i = d_i y_i(1) + (1-d_i) y_i(0)$5 within $y_i = d_i y_i(1) + (1-d_i) y_i(0)$6 (Cattaneo et al., 14 Sep 2025, Athey et al., 2015).

5. Cross-Validation and Complexity Control

Overfitting in the tree-building phase is mitigated by honest cross-validation. The training subsample is further split into $y_i = d_i y_i(1) + (1-d_i) y_i(0)$7 folds, and for each choice of complexity parameter (such as a per-leaf penalty $y_i = d_i y_i(1) + (1-d_i) y_i(0)$8), trees are pruned and evaluated on held-out folds using an unbiased estimate of the honest EMSE:

$y_i = d_i y_i(1) + (1-d_i) y_i(0)$9

The value of $\tau(x) = \mathbb{E}[y(1) - y(0) | x]$0 maximizing $\tau(x) = \mathbb{E}[y(1) - y(0) | x]$1 is selected, finalizing tree complexity (Athey et al., 2015).

6. Theoretical Properties: Risk Bounds and Consistency

CT-H achieves valid inference at the leaf level with unbiased CATE estimation and asymptotically correct standard errors, conditional on the tree:

• Minimax lower bound (sup-norm risk): With non-negligible probability, smallest cells produce errors at least $\tau(x) = \mathbb{E}[y(1) - y(0) | x]$2. Polynomial rates in $\tau(x) = \mathbb{E}[y(1) - y(0) | x]$3 are unattainable for uniform error, regardless of sample splitting. Tiny boundary leaves are the mechanism (Cattaneo et al., 14 Sep 2025).
• Integrated mean squared error (MSE): For trees of depth $\tau(x) = \mathbb{E}[y(1) - y(0) | x]$4,

$\tau(x) = \mathbb{E}[y(1) - y(0) | x]$5

up to logarithmic factors. The decay rate for integrated (global) risk is near-parametric, because small cells affect a negligible measure of the data space.

• Sup-norm inconsistency with depth: If tree depth grows $\tau(x) = \mathbb{E}[y(1) - y(0) | x]$6, pointwise sup-norm risk remains bounded away from zero. Deep trees, even with honesty, suffer pointwise inconsistency from arbitrarily small leaves (Cattaneo et al., 14 Sep 2025).
• Unbiasedness and inference: Estimates are unbiased (conditional on partition) with valid Gaussian confidence intervals (Athey et al., 2015).

Property	Honest CT Guarantee	Underlying Mechanism
Leaf unbiasedness	Yes	Data-splitting for estimation
Leafwise valid inference	Yes	Standard error estimation with held-out data
Uniform sup-norm rates	No (polynomial unattainable)	Small boundary leaf phenomenon
Integrated MSE rate	$\tau(x) = \mathbb{E}[y(1) - y(0) | x]$7 (up to logs)	Small-cell impact is localized
Consistency under depth	Only for shallow trees	Deep/adaptive trees inconsistent

7. Practical Considerations and Implications

CT-H’s sample splitting architecture provides robust protection against overfitting while enabling valid inference on heterogeneity across covariate-defined subgroups (Cattaneo et al., 14 Sep 2025). However, this comes with measurable costs and limitations:

• Data efficiency: Each stage (split selection, estimation) receives only half the data, effectively doubling sample requirements.
• Uniform error limitations: Worst-case (sup-norm) errors can persist, especially due to small boundary leaves, with non-shrinking lower bounds as $\tau(x) = \mathbb{E}[y(1) - y(0) | x]$8.
• L2-risk appeal: CT-H is effective when integrated error is the primary concern, as in risk minimization over $\tau(x) = \mathbb{E}[y(1) - y(0) | x]$9—but not when uniform accuracy is required across the covariate space.
• Depth selection: Growing trees beyond $d_i \perp \{y_i(0), y_i(1)\}\ |\ x_i$0 yields pointwise inconsistency; practical deployments must trade off granularity against risk of extreme errors.

A plausible implication is that CT-H approaches are most suitable for moderate-depth, moderate-dimensional settings where population-level heterogeneity is sought with valid inference, and uniform accuracy across all subgroups is not required (Cattaneo et al., 14 Sep 2025, Athey et al., 2015). Multiplicity corrections are necessary if multiple hypothesis testing over leaves is performed, but inference remains standard because splits are independent of estimation data (Athey et al., 2015).

• "The Honest Truth About Causal Trees: Accuracy Limits for Heterogeneous Treatment Effect Estimation" (Cattaneo et al., 14 Sep 2025).
• "Recursive Partitioning for Heterogeneous Causal Effects" (Athey et al., 2015).