Quantile Welfare Analysis

• Quantile Welfare is defined as assessing welfare distributions using rank-based quantiles and curves rather than conventional mean measures.
• It encompasses multiple approaches including distributional compensating variation, treatment effect quantiles, and market quantile intersections for policy evaluation.
• This framework enables nuanced analysis of inequality and poverty by applying lower-tail measures and integrated quantile functionals across diverse applications.

Quantile welfare is a family of welfare concepts that evaluates policies, allocations, or social states through quantiles or quantile-based functionals of welfare distributions rather than through means alone. In the literature, the term covers several distinct but related constructions: welfare as the area under excess-demand quantile curves in large two-sided markets; distributional compensating variation $CV_{\alpha,\tau}$, defined as a quantile of individual welfare effects; policy learning based on $Q_\tau(Y_1-Y_0\mid X)$, the conditional quantile of individual treatment effects; and social evaluation of health policy through $Q_\alpha[QALY(k)]$ rather than $E[QALY(k)]$ (Muir et al., 2016, Echenique et al., 2024, Cui et al., 2023, Manski et al., 5 Sep 2025). A closely related line replaces point quantiles by lower-tail averages such as superquantiles, while poverty and inequality research recasts lower-tail welfare objects directly in quantile-function form (Lambros et al., 24 Dec 2025, Nair et al., 8 Apr 2025).

1. Core formulations and conceptual scope

Quantile welfare has no single canonical object. In one strand, the welfare object is a quantile of the population distribution of individual welfare effects. In the discrete-choice framework of "Utilitarian Social Choice and Distributional Welfare Analysis" (Echenique et al., 2024), the central definition is distributional compensating variation,

$$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$

where $$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$. In that formulation, quantile welfare is literally the $\tau$-quantile of the distribution of individual compensating variations under a weighted preference distribution.

In a second strand, quantile welfare is a decision criterion over treatment policies. "Policy Learning with Distributional Welfare" (Cui et al., 2023) defines the welfare-relevant object as

$Q_\tau(Y_1-Y_0\mid X=x),$

the $\tau$-quantile of the distribution of individual treatment effects within covariate group $x$, and studies policies maximizing

$Q_\tau(Y_1-Y_0\mid X)$0

This is explicitly not the usual quantile treatment effect $Q_\tau(Y_1-Y_0\mid X)$1; the relevant quantile is the quantile of the gain distribution itself.

In a third strand, quantile welfare is a social planner’s criterion over population welfare distributions. "Utilitarian or Quantile-Welfare Evaluation of Health Policy?" (Manski et al., 5 Sep 2025) studies a planner who ranks feasible policies by a chosen population quantile,

$Q_\tau(Y_1-Y_0\mid X)$2

rather than by the population mean. The paper presents this as an alternative to utilitarian evaluation of health policy and ties it to Manski’s contrast between quantile utility and expected utility.

In a fourth strand, the word “quantile” does not refer to a social quantile of persons but to quantile curves describing market primitives. "Approximating welfare in large efficient markets" (Muir et al., 2016) shows that, in a large two-sided market, welfare can be represented as the area between empirical demand and supply quantile curves up to their empirical crossing: $Q_\tau(Y_1-Y_0\mid X)$3 The paper describes this representation as “quantile welfare” because welfare is an integral of excess quantile demand over traded mass.

A plausible implication is that “quantile welfare” is best understood as an umbrella term for welfare criteria and welfare representations that replace or supplement average-based aggregation by rank-based evaluation of distributions. The common element is not one specific formula, but the use of quantiles, quantile curves, or quantile integrals as the primitive welfare object.

2. Social-choice foundations and aggregation of quantile concerns

A major question is whether quantile welfare has an axiomatic foundation comparable to the standard utilitarian one. "Utilitarian Social Choice and Distributional Welfare Analysis" (Echenique et al., 2024) addresses this by moving from social welfare functions to social choice functions. With regular random expected utility distributions $Q_\tau(Y_1-Y_0\mid X)$4 for agents and $Q_\tau(Y_1-Y_0\mid X)$5 for the planner, the paper defines weighted utilitarian social choice by

$Q_\tau(Y_1-Y_0\mid X)$6

and shows that this is equivalent to local behavioral utilitarianism and to a menu-level Pareto condition. Within that framework, quantile welfare is not ad hoc: the key proposition states

$Q_\tau(Y_1-Y_0\mid X)$7

so the planner’s stochastic compensating variation is exactly the $Q_\tau(Y_1-Y_0\mid X)$8-quantile of the distribution of individual compensating variations. The same paper also proves a negative result: median compensating variation does not admit a Harsanyi-style social preference foundation satisfying the Pareto property if one insists on a social welfare function over outcomes. Quantile welfare therefore depends on the shift from preference aggregation to choice aggregation.

A distinct aggregation problem arises when individuals themselves evaluate risky acts by quantiles. "Spectral Aggregation of Quantile Preferences" (Nguyen, 29 Jun 2026) models individual $Q_\tau(Y_1-Y_0\mid X)$9 as ranking acts by $Q_\alpha[QALY(k)]$0. Social evaluation is spectral: $Q_\alpha[QALY(k)]$1 Its central support theorem states that $Q_\alpha[QALY(k)]$2 satisfies weak Pareto at profile $Q_\alpha[QALY(k)]$3 if and only if

$Q_\alpha[QALY(k)]$4

where $Q_\alpha[QALY(k)]$5, and satisfies strong Pareto if and only if, in addition, $Q_\alpha[QALY(k)]$6 for every represented level $Q_\alpha[QALY(k)]$7. The implication is sharp: social weight on a “phantom level” not represented by any individual violates Pareto. The representative-quantile case $Q_\alpha[QALY(k)]$8 is then Pareto-compatible only if $Q_\alpha[QALY(k)]$9, and strong Pareto requires unanimous agreement on that same quantile.

An ordinal variant appears in "Quantile agent utility and implications to randomized social choice" (Caragiannis et al., 19 Feb 2025). There, each agent evaluates a lottery $E[QALY(k)]$0 by the $E[QALY(k)]$1-quantile representative $E[QALY(k)]$2, the least preferred realizable outcome such that the probability of worse-ranked outcomes is at most $E[QALY(k)]$3. Lotteries are compared exactly by comparing these representatives. The paper proves that stochastic dominance is equivalent to agreement of the induced representative ordering at every $E[QALY(k)]$4. This yields an ordinal percentile welfare language for randomized outcomes, distinct from both expected utility and full stochastic-dominance comparison.

These papers jointly establish that quantile welfare can be grounded in several incompatible but rigorous normative architectures: choice-based utilitarianism over random utility distributions, spectral aggregation of heterogeneous quantile concerns, and ordinal representative-outcome evaluation of lotteries. They also show that quantile welfare is not merely “utilitarianism with a different statistic.” Its admissible social aggregation rules can be much more restrictive, especially under Pareto conditions.

3. Quantiles of welfare effects, treatment gains, and policy comparison

In applied policy analysis, quantile welfare most often refers to welfare effects across persons rather than welfare levels of a single market outcome. "Policy Learning with Distributional Welfare" (Cui et al., 2023) formulates this directly in a binary-treatment model. The planner’s criterion is based on

$E[QALY(k)]$5

and the unrestricted first-best rule is

$E[QALY(k)]$6

The paper emphasizes that this is not the quantile treatment effect $E[QALY(k)]$7, because the latter need not refer to the same individual in treatment and control states. It also gives the welfare interpretation of $E[QALY(k)]$8: low $E[QALY(k)]$9 is prudent, $$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$0 yields a majority-benefit logic, and high $$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$1 is negligent. For $$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$2, when outcomes are continuous,

$$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$3

The same paper shows that the relevant welfare object is generally only partially identified because it depends on the joint distribution of $$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$4. Under marginal identification, the identified set is

$$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$5

with Makarov-type bounds. To address this, the paper proposes minimax-regret policies and derives closed-form robust rules for unconstrained policy classes. A key consequence is that quantile welfare in treatment choice is inseparable from counterfactual dependence uncertainty.

A closely related but distinct lower-tail framework is developed in "Welfare at Risk: Distributional impact of policy interventions" (Lambros et al., 24 Dec 2025). The paper does not identify welfare quantiles themselves. Its central object is the lower superquantile

$$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$6

interpreted as the average outcome among the bottom $$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$7 of realizations. For latent individual welfare effects $$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$8, the paper shows

$$CV_{\alpha,\tau}(p^0,p') := T_{\hat\pi_\alpha}(\tau\mid p^0,p'),$$9

where $$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$0 is the identified conditional average welfare effect. Under continuity,

$$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$1

Thus the object is an integrated quantile functional rather than a point quantile. The paper explicitly interprets itself as welfare-at-risk or tail-distribution welfare analysis rather than direct identification of welfare quantiles.

Health-policy evaluation provides a further variant. "Utilitarian or Quantile-Welfare Evaluation of Health Policy?" (Manski et al., 5 Sep 2025) contrasts mean QALY evaluation with quantile-welfare evaluation, arguing that quantile welfare requires only an ordinal formalization of utility whereas expected utility requires a cardinal one. The paper proposes nonparametric bounds on the quantile welfare of health states using binary-choice time-tradeoff experiments. After normalizing benchmark health vectors $$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$2 by $$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$3, it derives interval statements such as

$$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$4

and

$$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$5

if

$$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$6

The paper’s broader claim is that mean-based health-policy analysis can be unduly influenced by problematic tail valuations, particularly for states judged worse than dead, whereas quantile welfare is much less driven by that tail behavior.

Taken together, these approaches show that quantile welfare can serve three different functions in policy analysis: a criterion for assigning treatment using the distribution of gains, a lower-tail welfare-at-risk criterion under latent heterogeneity, and a population-level alternative to mean QALYs in health economics. A common misconception is that these are all the same object. They are not. Some target point quantiles of welfare effects, some target integrated lower-tail quantiles, and some target quantiles of welfare levels induced by policy.

4. Quantile representations in markets, mechanisms, and allocation problems

In large-market theory, quantile welfare appears as a representation of efficient gains from trade. "Approximating welfare in large efficient markets" (Muir et al., 2016) studies a two-sided market with $$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$7 buyers and $$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$8 sellers, i.i.d. valuations $$T_\pi(\tau\mid p^0,p')=\inf\{z\in \mathbb R : G_\pi(z\mid p^0,p')\ge \tau\}$$9 and costs $\tau$0, and efficient quantity

$\tau$1

The paper reformulates the market through empirical quantile functions $\tau$2 and $\tau$3. Demand is represented by $\tau$4, supply by $\tau$5, and the efficient quantity is the empirical crossing of these curves. Welfare has the exact integral representation

$\tau$6

The population analogue uses

$\tau$7

and the population crossing $\tau$8 solves

$\tau$9

The resulting first-order deterministic welfare is

$Q_\tau(Y_1-Y_0\mid X=x),$0

The same paper derives a strong approximation based on empirical quantile processes and Brownian bridges. The empirical gap process satisfies

$Q_\tau(Y_1-Y_0\mid X=x),$1

and welfare admits the expansion

$Q_\tau(Y_1-Y_0\mid X=x),$2

Thus efficient quantity and welfare are jointly approximated by a centered bivariate normal vector after centering by their quantile-based deterministic approximations and scaling by $Q_\tau(Y_1-Y_0\mid X=x),$3. The paper also notes that the same machinery extends to mechanisms representable through transformed types $Q_\tau(Y_1-Y_0\mid X=x),$4 and $Q_\tau(Y_1-Y_0\mid X=x),$5, including the profit-maximizing mechanism.

A different use of quantile welfare appears in indivisible allocation. "Maximum Welfare Allocations under Quantile Valuations" (Aziz et al., 25 Feb 2025) considers agents who value a bundle $Q_\tau(Y_1-Y_0\mid X=x),$6 by a quantile of its item values: $Q_\tau(Y_1-Y_0\mid X=x),$7 The special cases $Q_\tau(Y_1-Y_0\mid X=x),$8 and $Q_\tau(Y_1-Y_0\mid X=x),$9 are worst-item and best-item valuation. Utilitarian welfare is

$\tau$0

and egalitarian welfare is

$\tau$1

The paper’s central finding is that complexity varies sharply with balancedness. Maximum balanced ESW is polynomial-time solvable even with heterogeneous quantiles, while balanced USW is NP-hard and hard to approximate within

$\tau$2

By contrast, unbalanced USW admits a polynomial-time

$\tau$3

approximation via the scapegoat algorithm, whereas unbalanced ESW has an interlaced tractability landscape depending on the quantile value $\tau$4.

These two literatures use the term “quantile welfare” in structurally different ways. The large-market paper studies welfare generated by the crossing and integration of empirical quantile curves. The indivisible-allocation paper studies welfare generated by agents whose preferences over bundles are themselves quantile-based. What unifies them is that rank positions, not sums alone, determine efficient or welfare-relevant outcomes.

5. Identification, regression, and empirical implementation

Quantile welfare requires access to welfare distributions or to rank-weighted summaries of those distributions. "Identifying the Distribution of Welfare from Discrete Choice" (Capéau et al., 2023) provides a nonparametric route in discrete-choice environments with unrestricted unobserved preference heterogeneity. For a finite choice set $\tau$5, welfare is defined through Fleurbaey-style nested opportunity sets. The key threshold characterization is

$\tau$6

which converts welfare threshold events into discrete-choice events under virtual prices. The marginal welfare CDF for the chosen alternative is then identified as

$\tau$7

Since the CDF is identified, all welfare quantiles follow by inversion: $\tau$8

The same paper identifies welfare-change distributions. For compensating variation,

$\tau$9

so

$x$0

is identified from cross-sectional data. With panel data, the paper further identifies joint distributions of baseline welfare and welfare changes, making conditional and incidence-type quantile welfare analysis feasible.

A complementary econometric framework is "Weighted-average quantile regression" (Chetverikov et al., 2022). It models a rank-weighted conditional distributional summary as linear in covariates: $x$1 When $x$2, the left-hand side is the conditional mean. When $x$3 is decreasing, the paper states that the object can be interpreted as a social welfare regression because lower quantiles receive larger weight; when $x$4 is signed, it can represent inequality contrasts. The paper develops an estimator based on an influence-function representation and sample splitting, proves $x$5-consistency and asymptotic normality, and emphasizes that the target is a welfare- or inequality-relevant weighted average of conditional quantiles rather than one quantile at a time.

A plausible implication is that empirical quantile welfare analysis splits into two broad tasks. One task is structural or nonparametric identification of the welfare distribution itself, followed by inversion or functional evaluation. The other is semiparametric regression of a chosen quantile-weighted welfare functional on observables. The first is distribution recovery; the second is covariate decomposition of a specified rank-sensitive welfare index.

6. Lower-tail welfare, poverty, and robust inequality measurement

A large adjacent literature uses quantiles to analyze lower-tail welfare and poverty even when the word “welfare” is not the headline term. "Revisiting poverty measures using quantile functions" (Nair et al., 8 Apr 2025) recasts standard poverty measures directly in terms of the income quantile function $x$6. With $x$7, the additively separable poverty functional becomes

$x$8

and the Foster-Greer-Thorbecke family becomes

$x$9

The paper derives quantile expressions for the poverty gap ratio,

$Q_\tau(Y_1-Y_0\mid X)$00

the mean income of the poor,

$Q_\tau(Y_1-Y_0\mid X)$01

the income gap ratio,

$Q_\tau(Y_1-Y_0\mid X)$02

the Gini index of the poor,

$Q_\tau(Y_1-Y_0\mid X)$03

and the Sen index,

$Q_\tau(Y_1-Y_0\mid X)$04

Its deeper claim is that poverty analysis can be reframed as lower-tail welfare analysis indexed by the population share $Q_\tau(Y_1-Y_0\mid X)$05, and that objects such as $Q_\tau(Y_1-Y_0\mid X)$06, $Q_\tau(Y_1-Y_0\mid X)$07, and $Q_\tau(Y_1-Y_0\mid X)$08 can determine the underlying income quantile function under stated conditions.

Robust inequality measurement supplies another quantile-based welfare perspective. "Quantile Versions of the Lorenz Curve" (Prendergast et al., 2015) defines three quantile Lorenz analogues,

$Q_\tau(Y_1-Y_0\mid X)$09

and corresponding inequality coefficients

$Q_\tau(Y_1-Y_0\mid X)$10

The paper shows that these curves are scale invariant, satisfy $Q_\tau(Y_1-Y_0\mid X)$11, and respond monotonically to median-preserving transfers, yielding $Q_\tau(Y_1-Y_0\mid X)$12 under such transfers. It emphasizes bounded influence functions and robustness to outliers, but also notes two departures from the classical Lorenz framework: lack of universal convexity and replacement of the usual mean-centered transfer principle by a median-centered version.

"Decomposing the Quantile Ratio Index with applications to Australian income and wealth data" (Prendergast et al., 2017) studies the quantile ratio index

$Q_\tau(Y_1-Y_0\mid X)$13

which averages relative gaps between symmetric lower and upper quantiles. For a symmetric $Q_\tau(Y_1-Y_0\mid X)$14-partition of the rank space, it defines conditional indices $Q_\tau(Y_1-Y_0\mid X)$15 and proves the exact decomposition

$Q_\tau(Y_1-Y_0\mid X)$16

with weights $Q_\tau(Y_1-Y_0\mid X)$17. The paper is explicit that this is not a standard within-group/between-group subgroup decomposition. It is a rank-based decomposition over symmetric quantile regions. That distinction matters: the object is a robust inequality diagnostic by distributional location, not a complete welfare criterion.

These literatures are closely related to quantile welfare but not identical to it. They generally analyze lower-tail welfare, poverty severity, or inequality through quantile functions, quantile Lorenz curves, or integrated symmetric quantile ratios. A plausible implication is that they provide the measurement and diagnostic infrastructure for quantile welfare analysis even when they do not directly specify a social planner’s objective over policies. In that sense, they form the lower-tail and distributional-statistics side of the quantile-welfare program.