Quantile Treatment Effects in Causal Inference

• Quantile Treatment Effects (QTEs) are a framework that measures how interventions impact different points of an outcome distribution under key assumptions such as unconfoundedness and overlap.
• Estimation methods for QTEs include inverse probability weighting, weighted quantile regression, and penalized approaches that efficiently handle variable selection in high-dimensional settings.
• QTEs extend to specialized designs including dynamic, spatial, and privacy-preserving contexts, with robust simulation studies highlighting their empirical performance.

Quantile Treatment Effects (QTEs) provide a rigorous framework for characterizing how a treatment or intervention impacts the entire distribution of an outcome, not merely its average. QTEs quantify how treatment shifts the outcome at specific quantiles, detecting distributional heterogeneity and addressing scientific or policy questions about impacts on individuals at different points of the outcome spectrum. The QTE concept is crucial for causal inference in randomized experiments, observational studies, high-dimensional settings, dynamic and spatial domains, privacy-constrained data, and in the presence of endogeneity or complex experimental designs.

1. Definition, Identification, and Core Assumptions

Let $Y(1)$ and $Y(0)$ denote the potential outcomes under treatment and control. For any $q\in(0,1)$, the q-th quantile of the treated and control potential outcome distributions is defined as

$$\xi^{(a)}(q) = \inf\{y: F^{(a)}(y) \geq q\}, \quad F^{(a)}(y) = P(Y^{(a)} \leq y), \quad a\in\{0,1\}$$

The QTE at quantile level $q$ is

$\Xi(q) = \xi^{(1)}(q) - \xi^{(0)}(q)$

QTEs are identified under standard assumptions:

• SUTVA (no interference or hidden versions of treatment),
• Strong ignorability (unconfoundedness): $(Y(1), Y(0)) \perp A \,|\, X$, and
• Overlap: $0 < \pi(x) < 1$ for all covariate values $x$, $\pi(x) = P(A=1 | X=x)$.

Under these, the potential outcome CDFs are identified by

$$F^{(1)}(y) = E\left[\frac{A \, I(Y \leq y)}{\pi(X)}\right],\quad F^{(0)}(y) = E\left[\frac{(1-A) \, I(Y \leq y)}{1-\pi(X)}\right]$$

Estimation leverages either direct plug-in empirical quantiles, inverse probability weighting (IPW), or modeling strategies; variance estimation and confidence band construction follow from influence function or resampling theory (Wu et al., 28 Mar 2025, Sun et al., 2021).

2. Methodological Advances: Estimation and Variable Selection

Standard estimation strategies include:

• IPW and CDF-based plug-in estimation: Estimate the propensity score, form IPW empirical CDFs as above, extract plug-in quantiles, and difference across groups to yield QTEs (Wu et al., 28 Mar 2025).
• Weighted quantile regression: Minimize the weighted sum of check-loss functions with estimated IPW or overlap weights (Sun et al., 2021, Shoji et al., 2024).
• Penalization and covariate selection: High-dimensional and network-feature-aware selection can be incorporated using penalized outcome regression and graphical interaction models, followed by adaptively penalized propensity score estimation (Wu et al., 28 Mar 2025, Liu et al., 2023, Shoji et al., 2024).

Covariate selection for the propensity score is crucial. Data-driven variable selection, particularly using quantile outcome-adaptive lasso (QOAL), quantile regression outcome-adaptive lasso (QROAL), or graphical interaction penalization, can minimize bias and variance, and guarantee oracle consistency under appropriate regularity (Wu et al., 28 Mar 2025, Liu et al., 2023, Shoji et al., 2024).

3. Advanced Topics: High-dimensional, Functional, and Personalized QTEs

In high-dimensional settings, modern strategies involve penalized quantile regression with sparsity-inducing penalties, followed by bias correction via projection or debiasing techniques for valid inference. The individualized QTE (IQTE)—the QTE at covariate value $x$—is of particular interest in precision medicine and optimal policy (Sun et al., 24 Mar 2025): $$\mathrm{IQTE}(\tau \mid X=x) = Q_{Y(1)|X}(\tau|x) - Q_{Y(0)|X}(\tau|x)$$ Debiased estimators constructed as linear projections of penalized quantile regression achieve minimax-optimal rates for CI length and hypothesis testing and yield valid uniform-in-$\tau$ inference over quantile regions. The methodology accommodates both personalized and quantile-wise heterogeneity, with uniform Gaussian process limits for the estimated IQTE curves (Sun et al., 24 Mar 2025).

4. Specialized Designs: Panel, Spatial, Dynamic, and Privacy-Constrained QTEs

• Panel and synthetic control QTEs: For panels with few treated units and many controls, QTEs are estimated via latent factor quantile regressions fit to controls, then plug-in quantile estimation for the treated. Inference accommodates serial dependence via block bootstrapping, and factor structures are allowed to vary across quantiles (Xu et al., 1 Apr 2025).
• Spatial and spatiotemporal QTEs: Deep spatial quantile regression models link outcome, treatment, spatial location, and covariates by mixture-of-splines neural networks. Plug-in spatial QTEs are defined by integrating conditional quantile contrasts over the covariate distribution at each spatial site. Local adjustment strategies address spatial confounding (Gong et al., 2 Sep 2025).
• Dynamic and pathwise QTEs: In time-series and dynamic settings with sequential actions, dynamic conditional QTEs can be decomposed into period-wise effects under specific structural monotonicity (Li et al., 2023).
• Privacy-preserving QTEs: In distributed or data-minimization contexts, QTEs are estimated by histogram aggregation plus differential privacy mechanisms, with formal accuracy-privacy tradeoffs depending on bin choice and Laplace noise magnitude (Yao et al., 2024).

5. Experimental Design and Resampling for QTEs

• Matched pairs and covariate-adaptive randomization: The asymptotic variance of QTE estimators is sensitive to design. In matched-pair experiments, variance is strictly lower than simple randomization due to negative within-pair dependence; consistent inference requires matched-aware bootstrapping (gradient or IPW multiplier bootstrap) (Jiang et al., 2020).
• Covariate-adaptive randomization (CAR): Under CAR, regression-adjusted QTE estimators with auxiliary regressions increase efficiency. Their asymptotic distribution is a tight Gaussian process, with direct multiplier or covariate-adaptive bootstrap yielding valid inference and semiparametric efficiency bounds under correct adjustment (Jiang et al., 2021).
• Rerandomization: Rerandomization using Mahalanobis distance improves QTE estimator efficiency by reducing covariate imbalance and generates a non-Gaussian asymptotic distribution for the QTE, represented as a linear combination of Gaussian and truncated Gaussian random variables (Han et al., 18 Jan 2026).

The table below summarizes representative estimation and inference strategies for QTEs in major experimental designs:

Experimental Design	Estimation Strategy	Valid Inference Procedure
Simple Randomization	IPW or plug-in quantiles	Influence function/Bootstrap
Matched Pairs	Pairwise quantile estimator	Gradient/IPW multiplier bootstrap
Covariate-Adaptive	Regression-adjusted QTE	Multiplier/CAR-aware bootstrap
Rerandomization	Covariate-balanced quantile	Conservative sandwich CI

6. Extensions and Robustness: Endogeneity, Extremes, and Functional QTEs

• Endogeneity and extreme QTEs: In the presence of endogeneity, local IV or RD approaches construct identified CDFs for compliers; regular variation in outcome tails enables stable Pareto extrapolation of extreme QTEs, with inference calibrated via subsampling rather than bootstrapping (Sasaki et al., 2024).
• Back-transformed and functional QTEs: For interpretability, back-transformed QTEs (BQTEs) express treatment effects on measurement units, rather than quantile levels, using probability integral transforms and piecewise interpolation, with nonparametric bootstrap providing confidence bands (Hemilä et al., 2023).
• Dynamic, spatial, and functional settings: Deep learning, spline mixtures, and plug-in approaches are used for QTE estimation in spatial, dynamic, or functional settings, often leveraging neural-network architectures and advanced model averaging for conditional distribution flexibility (Xu et al., 2022, Gong et al., 2 Sep 2025).

7. Simulation, Empirical Performance, and Practical Implementation

Across settings, QTE estimators with variable selection, penalization, or nonparametric modeling demonstrate controlled bias and variance, nominal coverage, and robust performance even in high-dimension or with complex dependencies (Wu et al., 28 Mar 2025, Gong et al., 2 Sep 2025, Shoji et al., 2024). Recommendations for practical use include:

• Model covariates and network/interaction features with penalized outcome and PS models.
• Conduct cross-validation, BIC tuning, or balance measures for penalty selection.
• Apply bootstrapped, block, or adaptive resampling for valid inference given design.
• Use local adjustment, spatial basis, or debiasing when facing spatial or endogeneity issues.
• Translate QTEs to the original measurement scale or to targeted subpopulations as appropriate.

Methods are implemented via standard convex optimization (e.g., coordinate descent for penalized regression, neural-network training for deep QTE, cross-fitted orthogonalization for efficient machine learning estimation) (Wu et al., 28 Mar 2025, Kallus et al., 2019). Empirical applications demonstrate sensitivity to model specification, choice of quantile, and variable selection, and underscore the necessity of aligning estimand, estimation procedure, and experimental/observational design for valid QTE analysis.

Key References:

• (Wu et al., 28 Mar 2025) A Unified Approach for Estimating Various Treatment Effects in Causal Inference
• (Sun et al., 24 Mar 2025) Minimax Rate-Optimal Inference for Individualized Quantile Treatment Effects in High-dimensional Models
• (Gong et al., 2 Sep 2025) Causal Spatial Quantile Regression
• (Sun et al., 2021) Causal Inference for Quantile Treatment Effects
• (Shoji et al., 2024) Quantile Outcome Adaptive Lasso: Covariate Selection for Inverse Probability Weighting Estimator of Quantile Treatment Effects
• (Jiang et al., 2020) Bootstrap Inference for Quantile Treatment Effects in Randomized Experiments with Matched Pairs
• (Xu et al., 1 Apr 2025) Quantile Treatment Effects in High Dimensional Panel Data
• (Han et al., 18 Jan 2026) Rerandomization for quantile treatment effects
• (Liu et al., 2023) Quantile regression outcome-adaptive lasso: variable selection for causal quantile treatment effect estimation
• (Jiang et al., 2021) Regression-Adjusted Estimation of Quantile Treatment Effects under Covariate-Adaptive Randomizations
• (Sasaki et al., 2024) Extreme Quantile Treatment Effects under Endogeneity: Evaluating Policy Effects for the Most Vulnerable Individuals

The QTE framework constitutes the foundational machinery for robust distributional causal inference, supporting a wide array of extensions for heterogeneous, high-dimensional, spatial, dynamic, privacy-aware, and structurally complex treatment effect analysis.