Distributional Discontinuity Designs

• Distributional discontinuity designs are causal frameworks that analyze changes in the entire outcome distribution at a threshold rather than only in conditional means.
• They employ metrics like the 2-Wasserstein distance and quantile treatment effects to capture complex treatment effects and distributional heterogeneity.
• Advanced methods include local polynomial regressions, Bayesian nonparametric approaches, and extensions to kinks and distribution-valued outcomes for robust inference.

Distributional discontinuity designs are discontinuity-based causal frameworks that target changes in the full outcome distribution at a threshold rather than only discontinuities in conditional means. In the recent literature, this label is used both narrowly, for a framework that takes the 2-Wasserstein distance between limiting conditional outcome distributions as the central estimand, and more broadly, for a family of regression discontinuity, kink, Bayesian nonparametric, and random-distribution methods that study quantiles, cumulative distribution functions, densities, spectral structure, or other distributional features at a cutoff (Schindl et al., 22 Feb 2026, Cattaneo et al., 2021, Hinne et al., 2019, Dijcke, 4 Apr 2025).

1. Conceptual foundations

In the canonical sharp regression discontinuity setup, units have a running variable $X_i$, cutoff $c$, treatment assignment $T_i=1(X_i\ge c)$, potential outcomes $Y_i(1),Y_i(0)$, and observed outcome $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$. The standard sharp RD estimand under the continuity framework is the mean effect at the cutoff,

$$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$

Distributional extensions replace continuity of conditional means by continuity of conditional distributions or other conditional functionals at the cutoff, so that left and right limits identify $F_{Y(0)\mid X=c}$ and $F_{Y(1)\mid X=c}$, and hence quantiles, tail probabilities, densities, or other functionals (Cattaneo et al., 2021).

A formal modern definition appears in “Distributional Discontinuity Design” (Schindl et al., 22 Feb 2026). For a scalar outcome, let $P_{a\mid x_0}$ denote the conditional distribution of $Y(a)\mid X=x_0$. The central estimand is

$c$0

the 2-Wasserstein distance between the limiting conditional counterfactual distributions at the cutoff. In one dimension,

$c$1

so the design aggregates the entire quantile-effect curve into a single scale-interpretable measure of distribution shift (Schindl et al., 22 Feb 2026).

The broader literature also uses “distributional discontinuity” in other senses. “Bayesian nonparametric discontinuity design” defines discontinuities at the level of a latent regression function and allows level discontinuities, derivative discontinuities, and spectral or temporal discontinuities through Gaussian process covariance choice (Hinne et al., 2019). “A Toolkit for the Study of Treatment-Effect Discontinuities” studies sign switches of a Treatment Effects Curve $c$2 along the outcome axis, where the threshold is learned from the data rather than imposed exogenously; the paper emphasizes that its Vertical Discontinuity Analysis is RDD-inspired but not an RDD (Antognini et al., 26 Jun 2026).

2. Estimands and distributional decompositions

Quantile-based estimands remain a core part of the field. Under continuity of conditional distributions, the quantile treatment effect at the cutoff is

$c$3

identified by left and right local limits of conditional quantiles. The review literature places Frandsen–Frölich–Melly, Chiang–Hsu–Sasaki, Qu–Yoon, and Huang–Zhan within this quantile RD agenda, and treats robust bias-corrected uniform inference for quantile treatment effects as a central methodological extension of mean RD (Cattaneo et al., 2021).

The Wasserstein formulation adds a single global summary of the distributional shift. If

$c$4

then

$c$5

with equality if and only if the treatment effect is purely additive in distribution, meaning $c$6 for all $c$7. The same paper defines a heterogeneity index

$c$8

so $c$9 corresponds to a pure location shift and larger values indicate that the mean explains less of the total distributional change (Schindl et al., 22 Feb 2026).

A further decomposition uses $T_i=1(X_i\ge c)$0-moments. If $T_i=1(X_i\ge c)$1 and $T_i=1(X_i\ge c)$2 are the $T_i=1(X_i\ge c)$3-th $T_i=1(X_i\ge c)$4-moments of the limiting treated and untreated distributions at the cutoff, then

$T_i=1(X_i\ge c)$5

This yields component shares $T_i=1(X_i\ge c)$6 for location, scale, skewness, and higher-order shape changes; $T_i=1(X_i\ge c)$7 is the share due to $T_i=1(X_i\ge c)$8-location (Schindl et al., 22 Feb 2026).

A distinct but related estimand arises when the outcome is itself a distribution. In Regression Discontinuity Design with Distribution-Valued Outcomes, each unit has a random distribution $T_i=1(X_i\ge c)$9, and the target is the local average quantile treatment effect

$Y_i(1),Y_i(0)$0

This is the jump in the conditional average quantile function at the cutoff, and it coincides with the difference of conditional Fréchet means in 2-Wasserstein space (Dijcke, 4 Apr 2025).

3. Identification and assumptions

Identification in distributional discontinuity designs retains the local character of standard RD. The continuity framework requires continuity of the relevant counterfactual object at the cutoff: for scalar outcomes, continuity of $Y_i(1),Y_i(0)$1 or of quantile functions; for Wasserstein DDD, continuity of counterfactual distributions together with consistency, positive density of the running variable at the cutoff, and finite second moments; for R3D, continuity of average quantiles $Y_i(1),Y_i(0)$2 near the cutoff (Cattaneo et al., 2021, Schindl et al., 22 Feb 2026, Dijcke, 4 Apr 2025).

Under those conditions, the observed left and right limiting distributions isolate the causal effect of the treatment jump. In sharp DDD this gives

$Y_i(1),Y_i(0)$3

where $Y_i(1),Y_i(0)$4 are the limiting conditional quantiles of the observed outcome above and below the cutoff. In sharp R3D,

$Y_i(1),Y_i(0)$5

The same local logic underlies local randomization approaches, where within a narrow window around the cutoff treatment is treated as as-if randomized and distributional comparisons can be conducted with experimental tools (Schindl et al., 22 Feb 2026, Dijcke, 4 Apr 2025, Cattaneo et al., 2021).

Fuzzy designs introduce a local Wald structure. In fuzzy R3D,

$Y_i(1),Y_i(0)$6

where the denominator is the jump in treatment propensity. In persuasion-focused fuzzy RD, the target becomes a probability-of-causation-type parameter, and the paper derives sharp bounds under different observability scenarios using Fréchet–Hoeffding inequalities; the robust lower bound is

$Y_i(1),Y_i(0)$7

This remains a lower bound across full-data, ecological, and outcome-only scenarios, while tighter upper bounds require observing $Y_i(1),Y_i(0)$8 or externally knowing $Y_i(1),Y_i(0)$9 (Dijcke, 4 Apr 2025, Jun et al., 30 Sep 2025).

Bayesian nonparametric discontinuity design keeps the same quasi-experimental identification assumptions as RD and ITS, but changes estimation by comparing a continuous model $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$0 and a discontinuous model $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$1. Its main conceptual claim is that conditioning on a discontinuous model alone produces overconfidence; BNDD therefore uses Bayes factors and Bayesian model averaging so that small apparent discontinuities can be shrunk toward zero and the posterior can place positive mass on “no discontinuity” (Hinne et al., 2019).

4. Estimation and inference

Local polynomial methods remain the dominant frequentist machinery. For mean RD, the estimator is a difference in local polynomial intercepts on either side of the cutoff. For quantile RD, the least-squares criterion is replaced by the check loss, yielding local weighted quantile regressions on each side. For CDF-based effects, one can estimate RD with indicator outcomes $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$2, which turns distributional RD into a continuum of local mean RD problems (Cattaneo et al., 2021).

The Wasserstein DDD framework estimates bias-corrected local polynomial CDFs on each side of the cutoff, inverts them to obtain estimated quantile functions $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$3, and then plugs these into

$Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$4

Because inference for $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$5 is non-regular, the paper works largely with $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$6 and proposes an eigenvalue-based Monte Carlo test, a conservative Chebyshev-type test, and conservative confidence intervals for $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$7. The same framework estimates $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$8-moment contributions by plugging estimated quantiles into the $Y_i=T_iY_i(1)+(1-T_i)Y_i(0)$9-moment integrals (Schindl et al., 22 Feb 2026).

R3D proposes two estimators. The first applies local polynomial regression directly to unit-level quantile functions,

$$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$0

and the second projects the estimated quantile curve onto the space of monotone quantile functions, yielding a local Fréchet regression estimator in 2-Wasserstein space. The paper establishes asymptotic normality for both, proves that the Fréchet estimator has the same first-order limit as the local polynomial estimator, and develops multiplier-bootstrap uniform, debiased confidence bands together with a data-driven bandwidth selection procedure (Dijcke, 4 Apr 2025).

BNDD offers a different inferential route. It fits Gaussian process regressions under a continuous model and a block-independent discontinuous model, approximates marginal likelihoods via BIC, and reports Bayes factors, posterior model probabilities, and model-averaged effect posteriors. With appropriate covariance choice, it can detect discontinuities of any order and in spectral features; the empirical implementations use Python + GPflow 2.2 and tabular rather than deep reinforcement-learning machinery (Hinne et al., 2019).

5. Major extensions

Distributional discontinuity methods extend naturally from level discontinuities to kinks. In distributional kink design, the object is the Wasserstein derivative at a policy kink,

$$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$1

interpreted as the instantaneous rate at which probability mass flows through the policy kink when the treatment level is marginally increased. The same paper derives new identification results for fuzzy kink designs and an $$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$2-moment derivative decomposition analogous to the level-discontinuity case (Schindl et al., 22 Feb 2026).

A separate extension concerns outcomes that are themselves distributions. R3D is designed for cases where treatment is assigned at a higher level of aggregation than the outcome, such as a subsidy allocated based on a firm-level revenue cutoff while the outcome of interest is the distribution of employee wages within the firm. The paper shows that existing scalar quantile RD methods are biased and inconsistent in this setting because they do not accommodate the two-level randomness of random distributions (Dijcke, 4 Apr 2025).

Another extension relocates the threshold from covariate space to outcome space. The treatment-effect discontinuity toolkit defines the Treatment Effects Curve

$$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$3

and studies sign-switch points $$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$4 satisfying $$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$5 and $$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$6. It combines Horizontal Discontinuity Analysis, which partitions the outcome support into regions of opposite-signed effects and studies them with causal forests, with Vertical Discontinuity Analysis, which locates crossings and tests non-tangentiality with a bias-corrected Wald statistic. The paper stresses that the threshold is learned from the data and that a TEC crossing does not imply a discontinuity in the conditional average treatment effect (Antognini et al., 26 Jun 2026).

Distributional ideas also appear in persuasion analysis. “Persuasion Effects in Regression Discontinuity Designs” studies a binary-outcome local persuasion rate,

$$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$7

and derives sharp and fuzzy RD identification, lower and upper bounds, and complier analogues. The novelty is not a quantile or Wasserstein estimand, but a refined distribution-based causal parameter built from marginal probabilities and Fréchet–Hoeffding bounds (Jun et al., 30 Sep 2025).

6. Design choices, robustness, and limitations

Robustness procedures developed for mean RD carry over, but not without cost. Donut RD removes observations in an inner window around the cutoff. The econometric analysis shows that donut RD estimators can have substantially larger bias and variance than conventional RD estimators, and that the corresponding confidence intervals can be substantially longer. The same paper provides a formal testing framework for comparing donut and conventional RD estimation results. For CDF-based distributional RD, this logic transfers directly by treating $$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$8 as the outcome (Noack et al., 2023).

Discontinuities can also be treated as design variables rather than fixed institutional objects. “Designing Discontinuities” first estimates local RD effects at existing borders, then uses a quantization-theoretic objective, dynamic programming, and value iteration to redesign partitions such as time-zone borders. The paper explicitly proposes a two-step RD procedure—estimate with current discontinuities, design new borders, and reapply RD with the new discontinuities—and notes that this uses local RD evidence as a welfare weight for a global partition, which relies on an extrapolation step rather than on RD alone (Ferwana et al., 2023).

Partially randomized threshold rules illustrate another adjacent design logic. Tie-breaker designs replace a single threshold with a randomized buffer around the cutoff. In a nonparametric kernel analysis, if the goal is estimation of the mean treatment effect at merely one score value, the paper proves that about 2.8 times more subjects are needed for an RDD in order to achieve the same asymptotic mean squared error, and larger experimental radii choices for the TBD lead to greater statistical efficiency (Kluger et al., 2021). This suggests that embedding randomized variation around a cutoff can materially change the precision frontier, although the stated result is for mean treatment effects rather than for the distributional estimands considered above.

Several limitations recur across the literature. Wasserstein DDD, as formulated in (Schindl et al., 22 Feb 2026), is univariate and relies on finite second moments; inference for $$\tau_{\text{SRD}}=E[Y_i(1)-Y_i(0)\mid X_i=c] =\lim_{x\downarrow c}E[Y_i\mid X_i=x]-\lim_{x\uparrow c}E[Y_i\mid X_i=x].$$9 is non-standard and can be conservative. The treatment-effect discontinuity toolkit requires non-tangential crossings, and when crossings are flat it recommends treating thresholds as too weak for threshold-based analysis (Antognini et al., 26 Jun 2026). R3D requires enough units near the cutoff and enough within-unit observations to estimate unit-level empirical quantiles reliably (Dijcke, 4 Apr 2025). More generally, all of these designs remain local: absent additional structure, they do not by themselves justify extrapolation away from the cutoff or from the detected crossing (Cattaneo et al., 2021).