Distributional Treatment Effects

• Distributional Treatment Effects are defined as the difference between cumulative distribution functions of potential outcomes under different treatments, capturing heterogeneity and risk beyond mean effects.
• They leverage regression-adjusted augmented inverse probability weighting under covariate-adaptive randomization to efficiently estimate shifts across the full outcome distribution.
• Practical applications, such as in microcredit studies, demonstrate DTEs' ability to reveal reductions in downside risk even when average outcomes show minimal change.

A distributional treatment effect (DTE) quantifies how an intervention shifts the entire outcome distribution, not merely the mean, and is a central object for understanding heterogeneity, risk, and the non-mean consequences of social programs, clinical interventions, and policy changes. For a pair of treatment arms $w, w'$, the DTE at threshold $y$ is defined as

$$\Delta^{\rm DTE}_{w,w'}(y) = F_{Y(w)}(y) - F_{Y(w')}(y), \quad y \in \mathcal Y$$

where $F_{Y(w)}(y)$ is the cumulative distribution function of the potential outcome under treatment $w$ (Byambadalai et al., 6 Jun 2025).

1. Formal Definition and Characterization

The DTE identifies how the probability of outcomes below any level $y$ differs between treatment regimes, revealing not only average (mean) effects but changes in quantiles, tail probabilities, and probabilities of critical outcomes (e.g., “zero revenue”). Unlike the average treatment effect (ATE), which summarizes only E[Y(1) – Y(0)], DTEs can diagnose whether a treatment reduces risk or increases the probability of avoiding highly undesirable outcomes (Byambadalai et al., 6 Jun 2025, Kallus et al., 2022). Extensions include conditional DTEs (CDTEs), which evaluate distributional effects within strata or given covariates (Kallus et al., 2022).

2. Covariate-Adaptive Randomization and Identification Challenges

Covariate-adaptive randomization (CAR) designs—including stratified block randomization and Efron's biased-coin design—do not assign treatments as i.i.d. Bernoulli – π draws. Instead, participants are grouped into strata $S_i$ and randomized within each stratum to achieve balance (Byambadalai et al., 6 Jun 2025). This induces nontrivial dependence in treatment indicators and within-stratum structure that must be explicitly accommodated.

Key assumptions for identification under CAR are:

• The joint collection $\{Y_i(w), S_i, X_i\}_{i=1}^n$ is i.i.d. across units;
• Given strata, treatment assignment is independent of potential outcomes and auxiliary covariates;
• Empirical treatment assignment probabilities $\widehat\pi_w(s)$ converge to their targets $\pi_w(s)$ at rate $o_p(1)$ (Byambadalai et al., 6 Jun 2025).

Statistical complications arise because standard empirical-process arguments (valid under i.i.d. Bernoulli assignment) will fail for CAR, requiring adjustment for dependence and the use of stratified weights and explicit modeling of within-stratum assignment probabilities.

3. Regression Adjustment and Efficient Estimation

Distribution regression frameworks provide efficiency and flexibility in estimating DTEs, especially in the presence of auxiliary covariates $X$ beyond the stratification. The regression-adjusted augmented inverse probability weighted (AIPW) estimator is

$$\widehat F^{\rm adj}_{Y(w)}(y) = \frac{1}{n}\sum_{i=1}^n \left[ \frac{\mathbf1\{W_i = w\}}{\widehat\pi_w(S_i)} \left(\mathbf1\{Y_i \le y\} - \widehat\mu_w(y, S_i, X_i)\right) + \widehat\mu_w(y, S_i, X_i) \right]$$

where $\widehat\mu_w$ is a flexible model-based estimate of the conditional distribution function (e.g., via random forests, boosting, neural nets) (Byambadalai et al., 6 Jun 2025).

The influence function for the estimator has two terms:

• The main term $\phi_w(y)$ addressing the estimation error conditional on strata and auxiliary covariates;
• The between-arm term $\zeta(y)$, quantifying mean shifts across arms within strata.

Under CAR and regular nuisance estimation conditions (uniform $n^{-1/2}$-rate convergence and Donsker/VC-type function class control), the estimator is asymptotically normal: $$\sqrt n \left[ \widehat\Delta^{\rm DTE,adj}_{w,w'}(y) - \Delta^{\rm DTE}_{w,w'}(y) \right] \rightsquigarrow \mathcal G(y)$$ with $\mathcal G(\cdot)$ a mean-zero Gaussian process and explicit covariance kernel (Byambadalai et al., 6 Jun 2025).

4. Inference: Confidence Bands and Semiparametric Efficiency

Pointwise and uniform inference for DTEs is available from plug-in variance estimation and multiplier (wild) bootstrap. The estimator is semiparametrically efficient: it attains the variance bound

$$\Omega(y, y) = \mathbb{E}[\pi_w(S)\phi_w^2(y)] + \mathbb{E}[\pi_{w'}(S)\phi_{w'}^2(y)] + \mathbb{E}[\zeta^2(y)]$$

and is thus optimal among regular estimators (Byambadalai et al., 6 Jun 2025). Confidence intervals and bands are constructed via the empirical AIPW process and standard Gaussian process approximations.

Simulation studies confirm substantial efficiency gains, with regression-adjusted ML estimators exhibiting 30–50% lower RMSE relative to empirical estimators for continuous outcomes (n=1,000), and confidence interval lengths correspondingly shorter. The method provides near-nominal coverage even in finite samples (Byambadalai et al., 6 Jun 2025).

5. Practical Applications and Interpretation

DTE analysis uncovers distributional heterogeneity often missed by mean-focused analysis. For example, in a microcredit field experiment in Mongolia, regression adjustment revealed:

• A statistically significant reduction (10 percentage points, SE ≈4.6 pp) in the probability of zero revenue, though average revenue shifted little;
• The implication that the main benefit was hedging against extreme downside risk, not uniformly raising the revenue distribution (Byambadalai et al., 6 Jun 2025).

More broadly, DTEs can be interpreted as showing:

• Where in the outcome distribution treatment modifies events (e.g., reduces risk of catastrophic loss);
• How tail probabilities and quantiles change under intervention, with or without mean improvement (Kallus et al., 2022).

6. Connections to Extensions and Related Theory

The DTE framework is extensible:

• Conditional DTE (CDTE) generalizes the effect to any functional of the conditional outcome distribution given covariates, with robust “pseudo-outcome” regression yielding agnostic best-in-class learning independent of model specification (Kallus et al., 2022).
• Risk-based DTEs such as conditional value-at-risk (CVaR) for the individual treatment effect distribution can be partially identified via bounds deriving from the CATE function, and efficiently estimated via orthogonal debiased procedures tolerant to black-box ML (Kallus, 2022).
• Structural or support restrictions (monotone, concave, Roy selection) can further sharpen identification and inference, particularly under partial identification settings and optimal transport duality (Kim, 2014).
• Extensions to multi-arm designs, high-dimensional covariates, and complex experimental designs (e.g., machine-assisted CAR) have been implemented via flexible machine learning estimators and cross-fitting, securing both robustness and efficiency (Byambadalai et al., 2024, Hirata et al., 10 Jul 2025).

7. Summary Table: Key Aspects of DTE Estimation under CAR

Aspect	Description	Reference
Definition	$$\Delta^{\rm DTE}_{w,w'}(y) = F_{Y(w)}(y) - F_{Y(w')}(y)$$	(Byambadalai et al., 6 Jun 2025)
Identification	Explicit stratification, within-stratum randomization, auxiliary covariates	(Byambadalai et al., 6 Jun 2025)
Efficient Estimation	Regression-adjusted AIPW, ML methods, cross-fitting	(Byambadalai et al., 6 Jun 2025)
Asymptotic Distribution	Uniform CLT in $L^\infty(\mathcal Y)$, mean-zero Gaussian process	(Byambadalai et al., 6 Jun 2025)
Semiparametric Efficiency	Attains variance bound; optimal among regular estimators	(Byambadalai et al., 6 Jun 2025)
Practical Advantages	30–50% RMSE reduction, shorter CIs, improved detection of tail and quantile effects	(Byambadalai et al., 6 Jun 2025)

DTE analysis under covariate-adaptive randomization enables granular, distribution-level causal inference that is robust to assignment mechanism complexities and adaptable to high-dimensional covariate information, supporting valid, efficient inference and nuanced policy decision-making.