Distributional Kink Designs in Causal Inference

• Distributional Kink Designs are causal inference methods that exploit policy kinks to reveal changes in the full outcome distribution rather than just the mean.
• They combine bunching-based strategies, where sorting produces mass points at thresholds, with derivative-based formulations that capture heterogeneous effects.
• These approaches enable nonparametric causal analysis by extrapolating counterfactual distributions under assumptions like bi-log-concavity, informing treatment effects and policy evaluations.

Distributional kink designs are a family of causal research designs that exploit a policy kink—a change in marginal incentives at a known threshold—to study how the entire outcome distribution responds, rather than only the conditional mean. In the bunching-based formulation, a kink in a budget set induces sorting and possibly a mass point at the threshold, and identification proceeds from densities, cumulative distribution functions, quantiles, and excess mass near the kink (Goff, 2022). In derivative-based formulations, the design generalizes regression kink logic from means to distributional objects such as conditional CDFs, conditional quantiles, and Wasserstein distances, thereby targeting heterogeneous and shape-changing effects that mean-based RKD can miss (Chiang et al., 2016). Across these formulations, the central idea is that a kink reveals causal information through distributional irregularities or derivative discontinuities localized at the threshold.

1. Conceptual scope and relation to regression kink designs

Distributional kink designs differ from conventional regression kink designs in the object of identification and in the way they treat behavior at the threshold. Mean-based RKD identifies local causal effects from changes in the slope of conditional expectations at a kink under smoothness and regularity assumptions. In the classical setup, the conditional mean must remain differentiable from each side, and bunching at the kink is typically ruled out because a point mass destroys the differentiability needed for the standard RKD estimand (Goff, 2022).

The bunching-based DKD literature turns this failure into an identification strategy. When a convex kink in the budget schedule causes agents to sort to the threshold, the researcher observes not only a spike at the kink but also censored densities on each side. This design uses the mass at the kink, the left and right limits of the density, and the CDF mass on either side to identify or bound treatment effects among those induced to locate at the kink (Goff, 2022). In this sense, bunching is not a nuisance to be removed; it is itself the identifying signal.

A second strand of the literature generalizes RKD from means to distributional objects without requiring the bunching interpretation. Quantile regression kink design defines a quantile-specific Wald ratio using left and right derivatives of the conditional quantile function at the kink, thereby recovering heterogeneous marginal effects at different conditional quantiles under suitable structural assumptions (Chiang et al., 2016). More recent work casts sharp RKD as a general framework for partial effects on Hadamard-differentiable functionals of the outcome distribution, including means, quantiles, moments, and inequality measures (Wang et al., 13 Jun 2025). A further extension evaluates the derivative of the 2-Wasserstein distance at a policy kink, interpreting it as the flow of probability mass through the kink and decomposing it into orthogonal contributions from location, scale, skewness, and higher-order shape components (Schindl et al., 22 Feb 2026).

This suggests that “distributional kink design” is best understood as an umbrella term for kink-based causal designs that target the distribution of outcomes rather than only its first moment. Within that umbrella, the bunching-based and derivative-based approaches are closely related but conceptually distinct: the former exploits sorting and censoring generated by the kink, whereas the latter exploits slope changes in distributional functionals at the kink.

2. Bunching-based DKD: potential outcomes, censoring, and the buncher ATE

The nonparametric bunching formulation in the overtime-pay application provides the most explicit treatment-effects interpretation of DKD (Goff, 2022). Let $h$ denote weekly hours and let the kink point be $k=40$, induced by overtime regulation. The wage schedule is piecewise linear:

$$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$

with $p>1$, and in the Fair Labor Standards Act application $p=1.5$ (Goff, 2022).

For each worker-week $it$, define potential choices under two linearized counterfactual schedules. Let $h_{0it}$ be the hours chosen if the firm paid $w$ for all hours, and let $h_{1it}$ be the hours chosen if the firm paid $pw$ for all hours. Denote their CDFs by $k=40$0 and $k=40$1, with densities $k=40$2 and $k=40$3 where they exist. Under a general convex choice model, realized hours under the actual kinked schedule are

$k=40$4

so the left side of the observed distribution reveals the density of $k=40$5 censored at $k=40$6, the right side reveals the density of $k=40$7 censored at $k=40$8, and the point mass at the kink is

$k=40$9

The bunching mass can also be written as

$$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$0

These formulas formalize the central DKD observation: the kink reveals two partially observed counterfactual distributions plus a mass point generated by “straddling” behavior (Goff, 2022).

The treatment effect at the unit level is $$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$1, the change in hours when moving from the global low-wage schedule to the global high-wage schedule. The principal estimand is the buncher ATE,

$$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$2

where $$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$3 indicates a “counterfactual buncher” with $$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$4 and $$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$5 is the share of such units. When $$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$6, this reduces to $$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$7 (Goff, 2022).

Under a rank condition weaker than full rank invariance, the buncher ATE can be written as an average quantile treatment effect over the bunching-rank window,

$$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$8

where $$w(h) = \begin{cases} w, & h \le k \ p\,w, & h > k \end{cases}$$9 is the quantile function of $p>1$0 (Goff, 2022). This representation connects bunching-based DKD directly to distributional causal analysis: the local average response at the kink is an integral of quantile differences over the ranks corresponding to bunchers.

The same paper also connects this estimand to elasticities in logs. Let

$p>1$1

and define an arc average wage elasticity among bunchers by

$p>1$2

In an isoelastic model this collapses to a common elasticity; in the general model it remains a local average among bunchers (Goff, 2022).

3. Identification through extrapolation and bi-log-concavity

The core identification problem in bunching-based DKD is an extrapolation problem. Because $p>1$3 is observed only to the left of the kink and $p>1$4 only to the right, the researcher does not observe the relevant counterfactual segments of the two distributions near $p>1$5. Identifying the buncher ATE therefore requires two extrapolations: one for $p>1$6 and one for $p>1$7 over the bunching-rank interval $p>1$8 (Goff, 2022).

Without structure, these missing segments are not point identified. The paper therefore imposes bi-log-concavity (BLC) on the counterfactual distributions in the unobserved regions. A CDF $p>1$9 is bi-log-concave if both $p=1.5$0 and $p=1.5$1 are concave. BLC implies the existence of a continuous density with locally bounded derivative, increasing hazard $p=1.5$2, and decreasing reverse hazard $p=1.5$3. The paper emphasizes that BLC is weak and partly testable on the observed tails, and that it nests common parametric densities often used implicitly in bunching practice (Goff, 2022).

Under CHOICE, CONVEX, RANK, and conditional BLC for $p=1.5$4 and $p=1.5$5 given $p=1.5$6, the buncher ATE lies in sharp bounds. Let $p=1.5$7 and denote left and right limits at the kink by $p=1.5$8, $p=1.5$9, $it$0, and $it$1. Then

$it$2

$it$3

with

$it$4

These bounds are sharp (Goff, 2022).

The intuition is that BLC constrains each missing quantile segment by log-linear envelopes pinned down by the local level and slope of the CDF at the boundary of the observed region. Integrating those envelopes over the bunching-rank window yields the $it$5 expressions. The width of the identified set shrinks with larger side densities and with smaller net bunching $it$6, and the paper notes that the bounds collapse in the small-kink limit, aligning with classical approximations (Goff, 2022).

This framework also clarifies why classical isoelastic point identification is fragile. In the isoelastic model, the bunching mass is treated as a missing segment of one counterfactual density of unknown length, and point identification relies on parametric interpolation across that segment. The nonparametric DKD interpretation recasts that strategy as an extrapolation problem and shows why it can fail empirically (Goff, 2022).

4. Estimation, inference, and empirical implementation

The overtime application supplies a practical estimation blueprint for bunching-based DKD (Goff, 2022). Densities and CDF limits at the kink are estimated using the local polynomial density estimator of Cattaneo, Jansson and Ma (2020). On the left side,

$it$7

with an analogous right-side fit for $it$8. The application uses a triangular kernel and MSE-optimal bandwidths averaged across sides (Goff, 2022).

A key implementation issue is estimation of the counterfactual bunching mass $it$9. The paper presents two strategies. The preferred design uses paid time off: PTO hours shift the overtime threshold in pay but not hours worked, so $h_{0it}$0 identifies the active bunching share in the PTO subsample. Under a representativeness assumption, $h_{0it}$1. Empirically, $h_{0it}$2 and $h_{0it}$3, implying $h_{0it}$4 (Goff, 2022). A second approach treats counterfactual bunchers as highly sticky and uses the non-changers upper bound $h_{0it}$5 (Goff, 2022).

For inference on partially identified parameters, the paper uses Imbens–Manski/Stoye adaptive critical values together with a firm-level cluster bootstrap with 500 draws, incorporating uncertainty in both the local distribution estimates and $h_{0it}$6 (Goff, 2022).

The empirical application uses payroll processor administrative data for 2016–2017. After restricting to weekly-paid hourly workers with observed overtime at some point and excluding California, the sample contains 630,217 paychecks for 12,488 workers across 566 firms. Hours are recorded at fine granularity; straight-time wages rarely change week to week, while hours do (Goff, 2022). The observed distribution exhibits a substantial mass at 40 hours, a secondary peak at 32 attributable to PTO, and holes around 40 consistent with straddling behavior. Virtually all paychecks with $h_{0it}$7 include overtime pay, supporting the institutional interpretation (Goff, 2022).

The estimated buncher ATE in levels depends sharply on the treatment of counterfactual bunchers. With PTO-based $h_{0it}$8, the preferred estimate is $h_{0it}$9 hours, approximately 40 minutes. With $w$0, attributing all bunching to the kink, the bounds widen to $w$1 hours. With $w$2 from non-changers, the estimate is approximately $w$3 (Goff, 2022). Expressed as an elasticity using $w$4, the preferred estimate is $w$5 (Goff, 2022).

The paper also reports broader policy effects under additional nonparametric assumptions. For all covered hourly workers, the ex-post average effect of FLSA on weekly hours lies in $w$6 hours under PTO-based $w$7, and among directly affected workers with $w$8 the effect is roughly double, $w$9 hours. A counterfactual double-time premium $h_{1it}$0 yields predicted average effects in $h_{1it}$1 hours, though less precisely estimated. Effects vary by industry, with the smallest in Health Care and in Professional/Scientific/Technical and the largest in Real Estate and Wholesale Trade (Goff, 2022).

A practical implication is that DKD can deliver informative empirical content without committing to a fully parametric choice model. In the overtime application, the isoelastic model is rejected because the implied global elasticity needed to rationalize the modest bunching is inconsistent with the observed density shape, which motivates the reduced-form and partially identified DKD approach (Goff, 2022).

5. Distributional derivatives at kinks: quantiles, partial effects, and Wasserstein DKD

A broader distributional literature studies kinks through derivative discontinuities of distributional functionals rather than through bunching masses. The foundational quantile formulation is the quantile regression kink design of Chiang and Sasaki. Its estimand is

$h_{1it}$2

Under flexible heterogeneity and endogeneity, this estimand equals a weighted average of heterogeneous marginal effects on the boundary set corresponding to the $h_{1it}$3-th conditional quantile, thereby giving a causal interpretation to quantile-specific kink effects (Chiang et al., 2016).

Later work generalizes this idea from quantiles to arbitrary smooth functionals of the conditional distribution in sharp RKD. Let a smooth intervention be $h_{1it}$4 with $h_{1it}$5 and define $h_{1it}$6. For a Hadamard-differentiable functional $h_{1it}$7, the local partial effect at the kink is

$h_{1it}$8

The distribution partial effect is

$h_{1it}$9

and the general functional effect is

$pw$0

This produces explicit formulas for means, quantiles, higher moments, interquartile range, and coefficient of variation, and it clarifies that the form of the policy intervention enters through the scalar $pw$1 (Wang et al., 13 Jun 2025).

The distributional discontinuity design framework extends the same logic to Wasserstein geometry and then to distributional kink designs (Schindl et al., 22 Feb 2026). For one-dimensional distributions,

$pw$2

At a kink, let

$pw$3

where $pw$4. The derivative at $pw$5 is

$pw$6

The paper interprets this derivative as the flow of probability mass through the kink (Schindl et al., 22 Feb 2026).

In the sharp design, the corresponding distributional kink estimand can be written as

$pw$7

with an analogous fuzzy version that replaces the known policy slope change by the kink in $pw$8 or more generally in $pw$9 (Schindl et al., 22 Feb 2026).

A notable property of the Wasserstein formulation is a mean bound. At a boundary, and by extension at a kink,

$k=40$00

with equality if and only if the effect is purely additive across quantiles. The same logic is stated for slopes at kinks. This turns the gap between mean and Wasserstein effects into a diagnostic for heterogeneity (Schindl et al., 22 Feb 2026).

The same paper further shows that the Wasserstein distance admits an orthogonal decomposition into squared differences in $k=40$01-moments,

$k=40$02

with $k=40$03 corresponding to location, $k=40$04 to scale, $k=40$05 to skewness, and higher $k=40$06 to higher-order shape components (Schindl et al., 22 Feb 2026). A plausible implication is that distributional kink analysis can summarize high-dimensional heterogeneity with a single scale-interpretable index while still decomposing that index into familiar distributional components.

6. Extensions, diagnostics, and methodological boundaries

Several adjacent literatures refine or broaden the DKD idea. Robust uniform inference for quantile treatment effects in regression discontinuity and kink designs develops a unified robust-bias-corrected local polynomial and multiplier-bootstrap framework for sharp and fuzzy mean and quantile RD/RKD, including uniform inference for CDF and quantile processes in fuzzy designs (Chiang et al., 2017). In the fuzzy RKD case, the distributional CDF effect at threshold $k=40$07 takes the derivative-ratio form

$k=40$08

and the associated CDF process can be inverted to obtain quantile treatment effects with uniform confidence bands (Chiang et al., 2017).

A separate line studies kinks with manipulation and interior responses. In “Identifying Causal Effects under Kink Setting: Theory and Evidence,” the running variable $k=40$09 is manipulated around a kinked schedule, and the observed density to the right of the kink is shifted leftward by interior responders rather than only by local bunchers (Lu et al., 2024). The paper reconstructs a counterfactual density $k=40$10 and outcome schedule $k=40$11 by relocating shifters, then identifies both a slope change and a level change at the kink. In that framework, the slope change identifies $k=40$12, the effect of payments on the outcome, whereas the level change identifies $k=40$13, the direct effect of the assignment variable on the outcome (Lu et al., 2024). This suggests that when manipulation around a kink is substantial, a DKD analysis may need to distinguish between pure derivative discontinuities and additional level effects generated by endogenous sorting.

For implementation, the literature repeatedly emphasizes visual and nonparametric diagnostics. In bunching designs, fine-grained histograms near the kink should display a spike and holes on each side; left and right densities and CDF limits should be estimated locally; and BLC can be partly tested by examining concavity of $k=40$14 CDF and $k=40$15 survival in the observed ranges (Goff, 2022). In derivative-based RKD, standard diagnostics include density continuity of the running variable, covariate smoothness, placebo kinks, bandwidth sensitivity, and robust bias correction (Chiang et al., 2017). In Wasserstein DKD, quantile monotonicity and rearrangement are relevant because the analysis proceeds through estimated quantile processes (Schindl et al., 22 Feb 2026).

Methodological boundaries are equally clear. Classical RKD is not well defined when bunching creates a mass point at the kink, which is precisely the setting in which bunching-based DKD becomes relevant (Goff, 2022). Conversely, derivative-based DKD requires smoothness of conditional distributions and derivatives around the kink, and causal interpretation in fuzzy designs depends on relevance and exclusion restrictions analogous to those in standard RKD (Chiang et al., 2017). The generalized sharp RKD framework for partial effects assumes deterministic assignment $k=40$16 and does not cover fuzzy RKD without further development (Wang et al., 13 Jun 2025). The manipulation-based kink literature, for its part, stresses that notched bunching methods relying on local “donut” manipulation are not valid for many kink settings because interior responses alter the entire right-side density (Lu et al., 2024).

A final ambiguity concerns terminology. “Distributional kink design” may denote bunching-based nonparametric treatment-effects analysis at a kink (Goff, 2022), quantile or CDF derivative analysis in RKD (Chiang et al., 2016), general partial-effect analysis of smooth functionals at a sharp kink (Wang et al., 13 Jun 2025), or Wasserstein-based distributional effects and their decomposition (Schindl et al., 22 Feb 2026). Rather than a contradiction, this reflects a common substantive aim: recovering causal information about how a kink changes the distribution of outcomes, including its location, dispersion, tails, and local mass.