Distributional Treatment Effects (DTE)

• Distributional Treatment Effects (DTE) are methods that compare full potential outcome distributions to reveal treatment heterogeneity beyond average effects.
• They employ regression-adjusted and distribution regression estimators, using cross-fitting and machine learning to handle complex experiments like CAR.
• Simulation and empirical studies show that DTE methods can significantly reduce estimation error and confidence interval lengths, enhancing inference in high-dimensional settings.

Distributional Treatment Effects (DTE) characterize the impact of treatment by comparing entire potential outcome distributions between experimental arms, rather than just their expectation. This approach underlies a granular understanding of treatment heterogeneity, revealing differences in quantiles, tail probabilities, or outcome dispersion that are invisible to conventional average treatment effect (ATE) analyses. Rigorous estimation and partial identification of DTEs have become a focal point in contemporary causal inference, particularly with the integration of machine learning, complex experimental designs such as covariate-adaptive randomization (CAR), and semiparametric efficiency theory.

1. Definition and Theoretical Framework

Let $W_i \in \{1, \ldots, K\}$ denote the treatment arm assigned to unit $i$, $S_i \in \{1,\ldots,S\}$ the stratum under CAR, $X_i \in \mathbb{R}^{d_x}$ any additional pre-treatment covariates, and $Y_i(w)$ the potential outcome under arm $w$. The marginal distribution function for the $w$-th potential outcome is

$F_{Y(w)}(y) = P(Y_i(w) \leq y).$

The Distributional Treatment Effect comparing arms $w$ and $w'$ is

$i$0

This functional captures the difference in cumulative probabilities at each $i$1 over the entire range of outcomes, rather than aggregating over $i$2.

Under CAR (including stratified block or Efron’s biased-coin design), the marginal CDF is a mixture over strata,

$i$3

This allows for flexible modeling of outcome distributions that respects both the CAR design and observed covariate distributions (Byambadalai et al., 6 Jun 2025).

2. Distribution Regression and Augmented IPW Estimation

Estimation of DTEs exploits the fact that only $i$4 is observed per subject. Efficient procedures construct regression-adjusted estimators using the "distribution regression" (DR) approach:

• Nuisance function for each $i$5: $i$6.
• $i$7 is estimated by a flexible supervised learner (e.g., LASSO, random forest, gradient boosting, DNN), typically with cross-fitting to mitigate overfitting and ensure robustness.

For a CAR-stratum-specific propensity $i$8 (known by design or estimated empirically as $i$9),

$S_i \in \{1,\ldots,S\}$0

The regression-adjusted estimator of $S_i \in \{1,\ldots,S\}$1 is

$S_i \in \{1,\ldots,S\}$2

and the corresponding DTE is $S_i \in \{1,\ldots,S\}$3 (Byambadalai et al., 6 Jun 2025).

3. Asymptotic Theory and Semiparametric Efficiency

The regression-adjusted DTE estimator admits a first-order expansion in terms of influence functions:

• Let $S_i \in \{1,\ldots,S\}$4,
• $S_i \in \{1,\ldots,S\}$5,
• $S_i \in \{1,\ldots,S\}$6.

The estimator satisfies

$S_i \in \{1,\ldots,S\}$7

The limit law is a tight, mean-zero Gaussian process $S_i \in \{1,\ldots,S\}$8 in $S_i \in \{1,\ldots,S\}$9 with covariance kernel

$X_i \in \mathbb{R}^{d_x}$0

A Hahn-style tangent space calculation shows that any regular estimator of the DTE has asymptotic variance at least $X_i \in \mathbb{R}^{d_x}$1 for each $X_i \in \mathbb{R}^{d_x}$2, and the regression-adjusted estimator attains this semiparametric efficiency bound (Byambadalai et al., 6 Jun 2025).

4. Variance Estimation and Inference

The plug-in sample variance estimator is

$X_i \in \mathbb{R}^{d_x}$3

This is used for constructing pointwise confidence intervals: $X_i \in \mathbb{R}^{d_x}$4 Uniform confidence bands are constructed via a multiplier (wild) bootstrap:

• Draw iid multipliers $X_i \in \mathbb{R}^{d_x}$5 with mean zero and variance one,
• Form the bootstrap process $X_i \in \mathbb{R}^{d_x}$6,
• Repeat $X_i \in \mathbb{R}^{d_x}$7 times, estimate the $X_i \in \mathbb{R}^{d_x}$8-quantile $X_i \in \mathbb{R}^{d_x}$9 of $Y_i(w)$0,
• Form the uniform band $Y_i(w)$1 (Byambadalai et al., 6 Jun 2025).

5. Simulation and Empirical Evidence

Simulation studies (e.g., $Y_i(w)$2 with 4 CAR strata, $Y_i(w)$3-dimensional $Y_i(w)$4, and nonlinear outcomes) benchmark empirical, linear, and machine learning (ML, via gradient boosting) regression adjustment:

• Linear adjustment reduces root mean squared error (RMSE) by approximately 10–30%,
• ML adjustments yield up to 50% RMSE reduction (especially in nonlinear, high-dimensional settings),
• 95% CI lengths shrink by 5–15%,
• Nominal coverage is maintained for all methods.

Application to microcredit data in Mongolia (n=611) with 16 baseline covariates and 5 strata shows that gradient boosting adjustment reduces standard errors by 1–13% (mean 7%). The DTE and PTE analyses detect a significant 10 percentage point drop in $Y_i(w)$5 (SE 4.6 pp), indicating that group lending reduces zero-revenue risk. At higher revenue levels, inference is limited by sample size and covariate informativeness, as indicated by wide confidence bands (Byambadalai et al., 6 Jun 2025).

6. Methodological Extensions and Practical Considerations

Distribution regression frameworks under CAR are compatible with diverse machine learning methods for conditional distribution estimation, provided that cross-fitting and regularization ensure the necessary $Y_i(w)$6 rate for Donsker or VC-type function classes.

Plug-and-play variance estimators and functional delta-method-based inference are standard. The methodology extends—without loss of efficiency—to multi-arm designs, as all influence function and variance expressions are generic in $Y_i(w)$7.

Empirical results highlight the substantial gains (particularly for complex outcome-covariate relationships) rendered by flexible adjustment relative to unadjusted or linear methods. The required elements for efficient and valid inference are notably explicit: regression adjustment, cross-fitted nuisance estimation, and multiplier bootstrap confidence bands.

7. Connections to Broader DTE Literature

This semiparametric-efficient, regression-adjusted framework for DTE estimation under CAR (Byambadalai et al., 6 Jun 2025) complements a broader literature of DTE identification and estimation strategies:

• Nonparametric and semi-parametric estimation frameworks (see (Byambadalai et al., 2024, Oka et al., 2024)),
• Local IV and distributional LATE definitions in the presence of compliers (Shaw, 15 Jun 2025),
• Machine learning–automated nuisance adjustment (gradient boosting, random forests, deep networks) (Hirata et al., 10 Jul 2025). It supplies a rigorous efficiency benchmark for designs with covariate-adaptive randomization and provides empirical researchers with a concrete template for high-performance DTE estimation in both moderate and high-dimensional settings.