priorities vs. quotas
TRANSCRIPT
Priorities vs. Quotas
Oguzhan Celebi∗
MIT
Joel P. Flynn†
MIT
August 18, 2021
Abstract
How should authorities concerned with match quality, fairness, and diversity, but un-
certain over the distribution of agents’ characteristics, allocate a resource? We show
that, when preferences over these dimensions are separable, a new monotone subsidy
schedule (MSS) mechanism requires no knowledge of the state and is ex post optimal.
We rationalize the common priority and quota mechanisms as limits of MSS when risk
aversion over diversity is low and high, respectively. Echoing lessons from price vs.
quantity regulation (Weitzman, 1974), priorities positively select agents over states,
while quotas guarantee a level of diversity, but MSS achieve both.
∗MIT Department of Economics, 50 Memorial Drive, Cambridge, MA 02142. Email: [email protected]†MIT Department of Economics, 50 Memorial Drive, Cambridge, MA 02142. Email: [email protected] are grateful to Daron Acemoglu, George-Marios Angeletos, Jonathan Cohen, Roberto Corrao,
Stephen Morris, Anh Nguyen, Parag Pathak, Karthik Sastry, Tayfun Sonmez, Alexander Wolitzky, andparticipants in the MIT Theory Lunch for helpful comments. First posted version: March 27, 2020.
1. Introduction
Authorities in charge of resource allocation in institutions often face conflicting objectives.
On the one hand, they want to allocate resources to the most suitable individuals to maximize
match quality or respect some notion of fairness. On the other hand, they want to ensure that
those who receive the resources are diverse according to a variety of characteristics such as
socio-economic background, religion, race, age, or gender. To this end, such authorities have
broadly used two classes of policies: quotas,1 where a certain portion of the resource is set
aside for given groups; and priority subsidies (or simply priorities), where such individuals
are given higher scores than an underlying index. This motivates two related questions:
What constitutes optimal policy in this setting? And when, if ever, are priorities or quotas
optimal?
These questions are important because priorities and quotas are the primary policies that
have been used for resource allocation in non-price contexts. Quotas have been introduced
across a range of markets, for example: Chicago Public Schools employs reserves for students
from different socio-economic groups at its competitive exam schools; Boston Public Schools
used reserves for walk-zone students at all schools; and universities such as University of
California, Davis instituted a quota system for minority students. Priority subsidies have
also been widely employed, for example: the widely used New York State Task Force on
Ventilator Allocation guidelines give differential priority to agents with differing mortality
risk; church-run schools in the UK give explicit admissions points to students from various
religious groups; the University of Michigan and the University of Texas have used different
admissions scales for minority students; and the Vietnamese university entrance exam has
given explicit exam points boosts to students from certain disadvantaged groups. However,
there is currently no formal understanding of whether these policies are optimal, how these
policies are different, or which policy an authority should pursue.
These settings all have four features in common. First, the individuals have an underlying
score (e.g., exam score, index of clinical need) that allocates them property rights over the
resource. Second, the authority is endowed with some power to affect these property rights by
designing a set of rules that transform the scores of individuals with certain characteristics
(e.g., quotas or priority subsidies for certain groups). Third, this set of rules is usually
designed when there is substantial uncertainty over the economy (e.g., the score distribution
of the students, number of individuals and doctors who need treatment during a pandemic).
Fourth, once the rules are designed, the authority implements an outcome that is fair with
respect to the (transformed) scores in the sense that they never allocate the resource to a
1We use quota as a general term that includes the widely used reserve policies (see Definition 3).
1
lower score individual when individuals with higher score are not allocated the resource.
In this paper, we therefore formulate and solve the optimal mechanism design problem of
an authority who allocates a resource to agents who are heterogeneous in their suitability for
the resource and other attributes. The authority cares separably about an index of match
quality and the numbers of agents of different attributes who are allocated the resource.
However, they are uncertain about the distribution of scores and attributes in the population,
which varies arbitrarily across states of the world.
We introduce a new class of mechanisms, monotone subsidy schedules, which proceed in
two steps. First, monotonically transform an agent’s score s when they have attribute m into
a new score Am(ym, s) that depends on the measure of agents who have the same attribute
and a higher score ym. Second, allocate the resource in order of transformed scores until the
resource is exhausted. This mechanism is fair in the sense that, for any set of agents with
the same attributes, the highest-scoring agents are allocated the resource.2
Addressing our first question, we derive in closed-form a monotone subsidy schedule
mechanism that implements the ex post optimal allocation in every state of the world while
requiring only knowledge of the authority’s own preferences (Theorem 1). We argue that
policymakers could use the class of fair mechanisms we propose to improve outcomes relative
to priority and quota mechanisms which are, by contrast, generally suboptimal.
This result not only suggests an improvement for policymakers relative to priority and
quota mechanisms, but also allows us to answer our second question and provide natural
sufficient conditions under which priorities and quotas are optimal. This allows us to offer
the following simple rationalizations of priority and quota policies (Corollary 1): first, if an
authority is certain of the composition of the population, then both priorities and quotas
are optimal. Second, if an authority is risk-neutral or highly risk-averse over the number of
assigned agents of different attributes, they can optimally use priority or quota mechanisms,
respectively.
To both illustrate and develop the intuition behind these results, we study a detailed
example that allows for a closed-form comparison of priorities, quotas, and the optimal
subsidy schedule mechanism. We do this in the spirit of the seminal analysis of Weitzman
(1974), who compares price and quantity regulation in product markets. In the example,
the resource corresponds to seats at a school and there are two groups of students (minority
and non-minority students). The authority is uncertain over the relative scores of minority
and non-minority students, and has linear-quadratic preferences over the scores of admitted
2This is equivalent to assigning agents in order of their transformed scores evaluated at the measureof already admitted agents of the same attribute. Thus, our proposed mechanism requires only that theauthority is able to rank all students within each group at the point of assignment, and requires no knowledgeof the state.
2
students and the number of minority students admitted to the school.
The preference of the authority between priority and quota mechanisms is governed by
its risk aversion over the number of admitted minority students: there is a cutoff value
such that quotas are preferred when risk-aversion exceeds this threshold and priorities are
otherwise preferred (Proposition 1). On the one hand, quotas guarantee a level of diversity by
mandating a minimal level of minority admissions. On the other hand, priorities positively
select minority students across states of the world as relatively more minority students
receive the resource in the states in which minority students have relatively higher scores,
improving match quality. Monotone subsidy schedules optimally exploit the guarantee effects
of quotas and the positive selection effects of priorities, and are always optimal. Thus, our
paper shows how standard price-theoretic lessons regarding instrument choice carry over to
markets without an explicit price mechanism.3 Finally, we leverage the example to provide
insights into optimal precedence order design and a recent debate regarding the allocation
of medical resources (Pathak, Sonmez, Unver, and Yenmez, 2021).
Related Literature Of most relevance to our analysis are the studies of quotas by Ko-
jima (2012), who shows how affirmative action policies that place an upper bound on the
enrollment of non-minority students may hurt all students, Hafalir, Yenmez, and Yildirim
(2013) who introduce the alternative and more efficient minority reserve policies, and Ehlers,
Hafalir, Yenmez, and Yildirim (2014) who generalize reserves to accommodate policies that
have floors and ceilings for minority admissions.4 The issue of priority design has also been
studied. Erdil and Kumano (2019) and Echenique and Yenmez (2015) study the effect of a
certain class of substitutable priorities, while Celebi and Flynn (2021) study how to opti-
mally coarsen underlying scores into priorities. Our focus on comparing priorities, quotas,
and optimal mechanisms distinguishes our analysis from this literature which considers the
properties of each policy in isolation and without an explicit treatment of uncertainty.
Finally, our result regarding the optimal order in which to process quotas (Corollary 2)
contributes to the literature that studies the effects of changes in precedence order (Dur,
Kominers, Pathak, and Sonmez, 2018; Dur, Pathak, and Sonmez, 2020; Pathak, Rees-Jones,
and Sonmez, 2020a,b). The difference between our paper and these is that we analyze the
optimal precedence order under uncertainty when the level of quotas is also under the control
of the authority.
3Spiritually, this builds on Azevedo and Leshno (2016) who introduced the price-theoretic analysis ofstable matchings.
4The slot-specific priority model of Kominers and Sonmez (2016) embeds these previous models. Furtherrelated papers study quota policies in university admissions in India (Sonmez and Yenmez, 2021, 2020a,b),in Germany (Westkamp, 2013) and in Brazil (Aygun and Bo, 2021), and simultaneous processing of quotas(Delacretaz, 2020).
3
2. Optimal Mechanisms
An authority allocates a single resource of measure q ∈ (0, 1) to a unit measure of agents.
Agents differ in their type θ ∈ Θ = [0, 1]×M comprising their scores s ∈ [0, 1] and personal
attributes m ∈ M, where their score denotes their suitability for the resource and M is
a finite set comprising potential attributes such as race, gender, or socioeconomic status.
The true distribution of types is unknown to the authority. The authority’s uncertainty
is paramaterized by ω ∈ Ω, which the authority believes has distribution Λ ∈ ∆(Ω). In
state of the world ω, the type distribution is Fω ∈ ∆(Θ) with density fω. An assignment
µ : Θ → 0, 1 specifies for any type θ ∈ Θ whether they are assigned to the resource.
The set of possible assignments is U . An assignment is feasible if it allocates no more than
measure q of the resource. A mechanism is an ω−measurable function φ : ∆(Θ) → U that
returns a feasible assignment for any possible distribution of types. The authority is an
expected utility maximizer with Bernoulli utility ξ : U × Ω→ R. Given a mechanism φ, let
µφ(ω) be the assignment in state of the world ω. A first-best mechanism is any mechanism
that attains the value:
supφ
∫Ω
ξ(µφ(ω), ω)dΛ(ω) (1)
Is there a set of rules the authority can design without knowledge of the state of the world
that yields the same value as a first-best mechanism? There is, of course, no guarantee that
this is possible. Nevertheless, whenever the authority’s payoff derives from match quality
and diversity, and is separable in these desiderata, we derive an explicit first-best mechanism
that is fair,5 and can be implemented with no knowledge of ω on the part of the authority.
To place some structure on preferences, we first assume that the authority cares only
about (i) an index of match quality
sh(µ, ω) =
∫Θ
µ(s,m)h(s)dFω(s,m) (2)
for some continuous, strictly increasing function h : [0, 1]→ R+, which determines the extent
to which the authority values agents with higher scores, and (ii) the measure of agents of
each attribute allocated the resource x(µ, ω) = xm(µ, ω)m∈M
xm(µ, ω) =
∫[0,1]
µ(s,m)fω(s,m)ds (3)
5Agents with higher scores are allocated before agents with lower scores and the same characteristics.
4
Assumption 1. There exists some ξ : R|M|+1 → R such that:
ξ(µ, ω) ≡ ξ (sh(µ, ω), x(µ, ω)) (4)
We next assume that the authority’s preferences are separable in match quality and
diversity:
Assumption 2. The authority’s utility function is separable:
ξ (sh, x) ≡ g
(sh +
∑m∈M
um(xm)
)(5)
for some continuous, strictly increasing function g : R→ R and differentiable, concave, and
weakly increasing functions um : R→ R for all m ∈M.
Here, um determines their preference for assigned agents of attribute m, and g determines
their risk preferences over their utility over scores and diversity across states of the world.
We maintain these assumptions throughout our analysis.
2.1. Monotone Subsidy Schedule Mechanisms Are Optimal
We now introduce a new and simple class of subsidy schedule mechanisms that we will
demonstrate are first-best optimal.
Definition 1 (Subsidy Schedule Mechanisms). A subsidy schedule mechanism A = Amm∈M,
where Am : R× [0, 1]→ R, transforms the score s of agents with attribute m who have mea-
sure ym higher scoring agents of the same attribute into Am(ym, s). Agents are allocated the
resource in order of their transformed scores Am(ym, s) until it reaches capacity.
We will say that a subsidy schedule mechanism A is monotone when Am(·, s) is a de-
creasing function for all m ∈ M, s ∈ [0, 1] and Am(ym, ·) is an increasing function for all
m ∈M, ym ∈ R. Observe that monotone subsidy schedule mechanisms are fair in the sense
that they preserve the ranking of agents within any attribute. Moreover, they can be im-
plemented in the following “greedy” fashion: within each attribute m, rank all agents in
order of their score and assign agents in order of their transformed scores evaluated at the
measure of already admitted agents of the same attribute. Thus, A can be specified ex ante
without any contingency on the unknown state, and all that is required in the interim to
implement it is knowledge of the scores that individual agents have – a necessary condition
for performing any form of prioritized assignment.
Theorem 1. The subsidy schedule mechanism Am(ym, s) ≡ h−1(h(s)+u′m(ym)) is monotone
and a first-best mechanism.
5
The proofs of all results are provided in Section 4. Observe that A requires only that the
authority knows its preferences over match quality h and diversity um.6 To gain intuition
for the form of this mechanism, suppose that the authority has linear utility over scores
h(s) ≡ s. In this case, Am(ym, s) = s + u′m(ym), so an agent with attribute m is awarded a
subsidy of u′m(ym) when there are ym higher scoring agents of the same attribute, their direct
marginal contribution to the diversity preferences of the authority. This is optimal, because
this subsidy precisely trades off the marginal benefit of additional diversity with the marginal
costs of reduced match quality, which are constant. To generalize this beyond linear utility
of scores, consider the following observation: we can map agents’ scores from s to h(s), and
consider the optimal subsidy mechanism in this space. As h is monotone, this preserves
the ordinal structure of the optimal allocation, and the authority has linear preferences over
h(s). Thus, in this transformed space, the optimal subsidy remains additive and given by
u′m(ym). To find the optimal transformed score in the original space, we simply invert the
transformation h and apply it to the optimal score in the transformed space, yielding the
formula for the optimal mechanism in Theorem 1.
2.2. Rationalizing Priority and Quota Mechanisms
As we have discussed, the primary classes of mechanisms that have been used in practice
are priority and quota mechanisms. Priority mechanisms give each agent a priority based
on their score and personal attributes, and allocate the resource in order of the priority.7
Quota mechanisms reserve some portion of the resource for agents with different attributes
and allocate each portion in order of the score. Formally, we define these mechanisms as:
Definition 2 (Priority Mechanisms). A priority mechanism P : Θ→ R awards each student
θ ∈ Θ a priority P (θ), and then allocates the resource in order of priority until measure q
has been allocated.
Definition 3 (Quota Mechanisms). A quota mechanism (Q,D) reserves Qm measure of the
capacity for agents of each attribute m ∈ M such that∑
m∈MQm ≤ q, with the remaining
capacity QR = q−∑
m∈MQm allocated to a merit slot R in which all student types are eligible.
The mechanism arranges these slots via a bijection D :M∪R → 1, 2, . . . , |M|+ 1 (the
precedence order). The mechanism then proceeds by allocating the measure QD−1(k) agents of
attribute D−1(k) to the resource in ascending order of k until measure q has been allocated.
In general, as Theorem 1 makes clear and the example in the next section will demon-
strate, neither priority or quota mechanisms are optimal. This is because they fail to adapt
6At the end of Section 4.1, we show how separability (Assumption 2) is necessary for this conclusion.7We allow priority mechanisms to reverse the scores of agents with the same attribute, but this is never
optimal as they always prefer to allocate to individuals with higher scores, all else equal.
6
to the state of the world: P and (Q,D) are both fixed ex ante and depend only on individual
characteristics. The subsidy schedule circumvents this issue by using a rank-dependent score
adjustment which allows the mechanism to adapt to the state of the world without needing
to know it.
Nevertheless, there are simple sufficient conditions on the uncertainty and diversity prefer-
ences of the authority that allow us to rationalize priority and quota mechanisms as optimal.
Corollary 1. The following statements are true:
1. If there is no uncertainty (i.e., |Ω| = 1), then there exist first-best priority and quota
mechanisms.
2. If the authority is risk-neutral over the measures of assigned agents of different at-
tributes (i.e., um(xm) is linear for all m ∈ M), then there exists a first-best priority
mechanism given by P (s,m) = h−1(h(s) + u′m).
3. If the authority is highly risk-averse over the measures of assigned agents of different
attributes around a diversity target (i.e., u′m(xm) ≥ km for xm ≤ xtarm and u′m(xm) = 0
for xm > xtarm where km is sufficiently large for all m ∈ M and∑
m∈M xtarm < q), then
there exists a first-best quota mechanism in which Qm = xtarm and D(R) = |M|+ 1.
This result formalizes the idea that the suboptimality of priority and quota mechanisms
stems from their inability to adapt to the state. However, it also provides conditions on
preferences such that this inability is not problematic. On the one hand, if the authority
is risk-neutral over the measure of agents of different attributes, then they can perfectly
balance their match quality and diversity goals without regard for the state of the world as
there is a constant “exchange rate” between the two, so priorities are optimal. On the other
hand, if the authority is highly risk-averse as to the prospect of failing to assign xtarm agents
of attribute m, then a quota allows them to always achieve this target level of assignment in
all states of the world while minimally sacrificing match quality.
This offers the following simple rationalizations of priority and quota policies. First, if
an authority is certain of the composition of the population, then both priorities and quotas
are optimal. Second, if an authority is risk-neutral or highly risk-averse over the measure of
assigned agents of different attributes, they can optimally use priority or quota mechanisms,
respectively.
3. Priorities vs. Quotas: A Closed-Form Example
We now illustrate and clarify the intuition underlying these results in a simple example
in which the welfare gains and losses from using priorities or quotas can be derived in closed
form. To do so, we follow an intellectual approach similar to that of Weitzman (1974) in
7
his seminal comparison of price and quantity mechanisms in product markets. We use the
example to study optimal precedence order design and medical resource allocation.
3.1. The Setting of the Example
A single school has capacity q. Students are of unit total measure and either minority or
majority students. The authority has linear-quadratic preferences ξ : R2 → R over students’
total scores s and the number of admitted minority students x:8,9
ξ(s, x) = s+ γ
(x− βx
2x2
)(6)
where γ ≥ 0 indexes their general concern for admitting minority students relative to en-
suring high scores and βx ≥ 0 indexes the degree of risk-aversion regarding the measure of
admitted minority students. The minority students are of measure κ and have a distribution
of underlying scores that is uniform over [0, 1]. The majority students are of the residual
measure and all have common underlying score ω ∈ [ω, ω] ⊆ [0, 1]. Finally, we assume that
the affirmative action preference is neither too small nor too large with the following two
conditions minκ, q > 1+γ−ω1κ
+γβx+κ(ω−ω) and κ(1−ω) < 1+γ−ω
1κ
+γβx. These conditions ensure that
optimal affirmative action policies will neither be so large as to award all slots to minority
students in some states nor so small that there is no affirmative action in some states.
The authority can either implement a subsidy schedule mechanism (which will be op-
timal), an additive priority subsidy mechanism, or a quota mechanism. In this setting, a
subsidy schedule mechanism awards an additive score subsidy of A(y) to a minority student
when measure y other minority students have higher scores, and then allocates the school to
students in order of their transformed scores. An additive priority subsidy α ∈ R+ increases
uniformly the scores of minority students for the purposes of gaining admission: the score
used in admissions becomes uniform over [α, 1 + α]. The authority then admits the highest
scoring measure q students. A quota policy Q ∈ [0,minκ, q] sets aside measure Q of the
capacity for the minority students. The measure Q highest scoring minority students are
first allocated to quota slots, and all other agents are then admitted to the residual q − Qplaces according to the underlying score.10
8Alternatively, if the authority cares about both the average score and the proportion of minority students,
all of the analysis goes through. For example, Equation 7 becomes ∆ = κ2q
(1− κ
q γβx
)V[ω].
9To nest this in our more general setting, set h(s) = s, uminority(x) = γ(x− βx
2 x2)
(which is weakly
increasing over the relevant region given our parametric assumptions) and umajority ≡ 0.10This corresponds to a precedence order that processes quota slots first. We discuss the importance of
precedence orders in section 3.3 together with how our model can produce insights about their design.
8
3.2. Comparing Mechanisms
Let the authority’s expected utility be V ∗ under any first-best optimal mechanism, VS
under an optimal subsidy schedule mechanism, VP under an optimal priority mechanism,
and VQ under an optimal quota mechanism. The following proposition characterizes the
relationships between these mechanisms:
Proposition 1. The following statements are true:
1. The comparative advantage of priorities over quotas is given by:
∆ ≡ VP − VQ =κ
2(1− κγβx)V[ω] (7)
Thus, priorities are preferred to quotas if and only if:
1
κ≥ γβx (8)
2. The monotone subsidy schedule mechanism A(y) = γ(1 − βxy) is first-best optimal,
V ∗ = VS. The comparative advantage of subsidy schedule mechanisms over priorities
and quotas is given by:
∆∗ ≡ minV ∗ − VP , V ∗ − VQ =
12
(κγβx)2 κV[ω]
1+κγβx, κγβx ≤ 1,
12
κV[ω]1+κγβx
, κγβx > 1.(9)
Which is increasing in κγβx for κγβx ≤ 1, decreasing in κγβx for κγβx > 1, and equals
zero when κγβx = 0.
To develop intuition for the comparative advantage of priorities over quotas, observe
the following. First, a quota of Q admits measure Q minority students in all states of
the world under our assumptions. However, a priority policy induces variability in the
measure of admitted minority students across states of the world. This costs a priority
policy κ2
(1 + κγβx)V[ω] in payoff terms. Second, a priority policy positively selects minority
admissions across states of the world. In the proof of the result, we show that minority
admissions in state ω under the optimal priority policy are x(α, ω) = x(α) + ε(ω) where
x(α) = κ(1 + α − E[ω]) and ε(ω) = κ (E[ω]− ω). Thus, the optimal priority policy admits
more minority students when minority students score relatively well and fewer when minority
students score relatively poorly. This benefits a priority policy by −C[ω, ε(ω)] = κV[ω] in
payoff terms. Which is preferred then depends on the risk preferences of the authority over
the measure of admitted minorities. If the authority is close enough to risk-neutral and
9
1κ> γβx, then priorities are strictly preferred as positive selection dominates guarantees. If
the authority is sufficiently risk-averse and 1κ< γβx, then quotas are strictly preferred as the
guarantee effects dominate positive selection. Finally, the extent of uncertainty V[ω] may
intensify an underlying preference but never determines which regime is preferred.11
In this example, the optimal subsidy schedule is linear in the minority students’ ranks,
with slope given by the authority’s risk aversion over minority admissions. This allows the
subsidy schedule to optimally balance the positive selection and guarantee effects, and imple-
ment the first-best allocation in every state. From this, we learn that the loss from priority
and quota policies relative to the optimum is greatest when the authority is indifferent be-
tween the two regimes. Echoing our rationalization from earlier, the loss from restricting to
priority or quota policies is zero when the authority is risk-neutral or there is no uncertainty
regarding relative scores, and decreases as the authority becomes highly risk averse.
3.3. Optimal Precedence Orders
In this example so far, we modelled quotas by first allocating minority students to quota
slots and then allocating all remaining students according to the underlying score. However,
we could have instead allocated q−Q places to all agents according to the underlying score
and then allocated the remaining Q places to minority students. The order in which quotas
are processed is called the precedence order in the matching literature and their importance
for driving outcomes has been the subject of a large and growing literature (see e.g., Dur
et al., 2018, 2020; Pathak et al., 2020a). Our framework can be used to understand which
precedence order is optimal, a question that has not yet been addressed.
In this example, the same factors that determine whether one should prefer priorities or
quotas determine whether one should prefer processing quotas second or first. By virtue of
uniformity of scores, it can be shown in the relevant parameter range that a priority subsidy
of α is equivalent to a quota policy of κα when the quota slots are processed second. Thus,
the comparative advantage of priorities over quotas is exactly equal to the comparative
advantage of processing quotas second over first. The intuition is analogous: processing
quotas second allows for positive selection while processing quotas first fixes the number of
admitted minority students. Thus, on the one hand, when the authority is more risk averse,
they should process quota slots first to reduce the variability in the admitted measure of
minority students. On the other hand, when they are less risk averse, they should process
11There is in fact a formal mapping between ∆ in our setting and that of Weitzman (1974), whichcorresponds to the comparative advantage of prices over quantities. Mapping Weitzman’s C ′′−1 7→ κ,B′′ 7→ −γβm, V[α(θ)] 7→ V[ω], we have that Weitzman’s ∆ coincides with our own. The positive selectioneffect is equivalent to the effect that price regulation gives rise to the greatest production in states wherethe firm’s marginal cost is lowest. Moreover, the guarantee effect is equivalent to the ability of quantityregulation to stabilize the level of production.
10
quotas second to take advantage of the positive selection effect such policies induce. These
results are summarized in the following corollary:
Corollary 2. The optimal quota-second policy achieves the same value as the optimal pri-
ority policy; quota-second policies are preferred to quota-first policies if and only if 1κ≥ γβx.
3.4. Beyond Affirmative Action: Medical Resource Allocation
The lessons of this paper apply not only to affirmative action in academic admissions, but
much more broadly to other settings in which centralized authorities must allocate resources
to various groups. One prominent such context is the allocation of medical resources during
the COVID-19 pandemic. An important issue faced by hospitals is how to prioritize frontline
health workers (doctors, nurses and other staff) in the receipt of scarce medical resources:
hospitals wish to both treat patients according to clinical need and ensure the health of
the frontline workers needed to fight the pandemic. To map this setting to our example,
suppose that the score s is an index of clinical need for a scarce medical resource available
in amount q, the measure of frontline health workers is κ, and ω indexes the level of clinical
need in the patients currently (or soon to be) treated by the hospital, which is unknown.
The risk aversion of the authority γβx corresponds to both a fear of not treating sufficiently
many frontline workers and excluding too many clinically needy members of the general
population.
In practice, both priority systems and quota policies have been used, as detailed exten-
sively by Pathak et al. (2021). The primary concern that has been voiced is that if a priority
system is used, some attributes (or characteristics) may be completely shut out of allocation
of the scarce resource and that this is unethical, so quotas should be preferred. Our frame-
work can be used to understand this argument; if there is an unusually high draw of ω, a
priority system would lead to the allocation of very few resources to frontline workers, and
vice-versa. Our Proposition 1 implies that if the authority is very averse to such outcomes
(γβx is high), quotas will be preferred and for exactly the reasons suggested. However, we
also highlight a fundamental benefit of priority systems in inducing positive selection in
allocation: when ω is high, it is beneficial that fewer resources go to the less sick medical
workers and more to the relatively sicker general population. More generally, as per Theorem
1, we argue that a subsidy schedule mechanism that awards frontline workers a score subsidy
that depends on the number of more clinically needy frontline workers could further improve
outcomes.
Finally, an important additional consideration in this context arises if the hospital or
authority must select a regime (priorities or quotas) before it understands the clinical need
of its frontline workers κ, after which it can decide exactly how to prioritize these workers
11
or set quotas, but before ultimate demand for medical resources ω is known. It follows from
Proposition 1 that the comparative advantage of priorities over quotas is:
E[∆] =1
2
(E[κ]− (V[κ] + E2[κ])γβx
)V[ω] (10)
Thus, an increase in uncertainty V[κ] regarding the need of frontline workers leads to a greater
preference for quotas. This is for the reason that the volatility in the number of frontline
workers is convex in κ, the sensitivity of the measure allocated to medical workers to the
underlying demand for medical resources ω. This highlights a further advantage of quotas in
settings where a clinical framework must be adopted in the face of uncertainty regarding the
clinical needs of frontline workers, as was the case at the onset of the COVID-19 pandemic.
4. Proofs
4.1. Proof of Theorem 1
Proof. We characterize the optimal allocation for each ω ∈ Ω and show that the claimed
subsidy schedule implements the same allocation. Fix an ω ∈ Ω and suppress the dependence
of Fω and fω thereon, and define the utility index of a score as s = h(s) with induced densities
over s given by fm for all m ∈ M. Let the measure of agents with any attribute m ∈ Mthat is allocated the resource be xm ∈ [0, xm] where xm =
∫ h(1)
h(0)fm(s)ds. Observe that,
conditional on fixing the measures of agents of each attribute that are allocated the resource
x = xmm∈M, there is a unique optimal allocation (i.e., ξ-maximal µ). In particular, as g
and h are continuous and strictly increasing, the optimal allocation conditional on x satisfies
µ∗(s, m;x) = 1 ⇐⇒ s ≥ sm(xm) for some thresholds sm(xm)m∈M that solve:∫ h(1)
sm(xm)
fm(s)ds = xm (11)
We can then express the problem of choosing the optimal x = xmm∈M as:
maxxm∈[0,xm], ∀m∈M
∑m∈M
∫ h(1)
sm(xm)
sfm(s)ds+∑m∈M
um(xm) s.t.∑m∈M
xm ≤ q (12)
where a solution exists by compactness of the constraint sets and continuity of the objective.
We can derive necessary and sufficient conditions on the solution(s) to this problem by
12
considering the Lagrangian:
L(x, λ, κ, κ) =∑m∈M
∫ h(1)
sm(xm)
sfm(s)ds+∑m∈M
um(xm)
+ λ
(q −
∑m∈M
xm
)+∑m∈M
κm(xm − xm) +∑m∈M
κmxm
(13)
The first-order necessary conditions to this program are given by:
∂L∂xm
= −s′m(xm)sm(xm)fm(sm(xm)) + u′m(xm)− λ− κm + κm = 0 (14)
λ∂L∂λ
= λ
(q −
∑m∈M
xm
)= 0 (15)
κm∂L∂κm
= κm(xm − xm) = 0 (16)
κm∂L∂κm
= κmxm = 0 (17)
for all m ∈M. By implicitly differentiating Equation 11, we obtain that:
s′m(xm) = − 1
fm(sm(xm))(18)
Thus, we can simplify Equation 14 to:
∂L∂xm
= sm(xm) + u′m(xm)− λ− κm + κm = 0 (19)
Observe that all constraints are linear. Thus, if the objective function is concave, the
first-order conditions are also sufficient. To this end, as all cross-partial derivatives of the
objective function are zero, it suffices to check that ∂L∂xm
is a decreasing function of xm for all
m ∈M. Observe by Equation 18 that sm(xm) is a decreasing function of xm. Moreover u′m
is a decreasing function of xm by virtue of the assumption that um is concave for all m ∈M.
Thus, the objective function is concave.
Thus, to verify that our claimed subsidy schedule is a first-best mechanism, it suffices to
show that the allocation it implements satisfies Equations 14 to 17. The subsidy schedule
Am(ym, s) = h−1 (h(s) + u′m(ym))) in the transformed score space yields transformed scores
h (Am(ym, s)) = s+ u′m(ym). Define xm as the admitted measure of students of attribute m
under this mechanism. Agents of attribute m ∈ M are allocated the resource if and only if
13
s+ u′m(xm) ≥ sC for some threshold sC that solves:
∑m∈M
∫ h(1)
maxminsC−u′m(xm),h(1),h(0)fm(s)ds = q (20)
We can therefore partition M into three sets that are uniquely defined: (i) interior MI =
m ∈ M|sC − u′m(xm) ∈ (h(0), h(1)); (ii) no allocation M0 = m ∈ M|sC − u′m(xm) ≥h(1); (iii) full allocation M1 = m ∈ M|sC − u′m(xm) ≤ h(0). For all m ∈ M0, we
implement xm = 0. For all m ∈M1, we implement xm = xm. For all m ∈MI , we implement
xm ∈ (0, xm). For any m ∈MI , the allocation threshold is sm(xm) = sC − u′m(xm). For any
m ∈M0, the allocation threshold is h(1). For any m ∈M1, the allocation threshold is h(0).
We now verify that this outcome satisfies the established necessary and sufficient condi-
tions. For all m ∈MI , by the complementary slackness conditions we have that κm = κm =
0. Substituting the above into Equation 14 for all m ∈MI we obtain that:
sC − λ = 0 (21)
which is trivially satisfied for λ = sC . As q =∑
m∈M xm, the complementary slackness
condition for λ is then satisfied. For all m ∈M0, by complementary slackness we have that
κm = 0 and Equation 14 is satisfied by:
κm = λ− h(1)− u′m(0) (22)
For all m ∈ M1, by complementary slackness we have that κm = 0 and Equation 14 is
satisfied by:
κm = h(0) + u′m(xm)− λ (23)
This completes the proof.
In Footnote 6, we comment that Assumption 2 is necessary for this result. Following
the same steps as above but without Assumption 2 (while assuming that ξ is differentiable
and weakly increasing), an optimal generalized subsidy schedule necessarily depends on the
entire vector y = ymm∈M and the state ω:
Am(y, s;ω) ≡ h−1
(h(s) +
ξxm (sh(y, ω), y)
ξsh (sh(y, ω), y)
)(24)
where sh(y, ω) is the match quality index in state ω when the highest scoring y = ymm∈Magents of each attribute are allocated. Observe that this generally depends on ω via the joint
14
distribution of attributes and scores and is therefore not implementable without knowledge
of the state. Observe further that this collapses to the optimal subsidy schedule we derive
when Assumption 2 is imposed.
4.2. Proof of Corollary 1
Proof. Part (i): Suppose |Ω| = 1 and let x∗m denote the measure of attribute m agents in the
optimal allocation, with x∗ = x∗mm∈M. A priority policy P (s,m) = h−1(h(s) + u′m(x∗m)) =
Am(x∗m, s) implements the same allocation as the optimal subsidy schedule mechanism and
by Theorem 1, is optimal. A quota mechanism with (Q,D) where Qm = x∗m implements x∗
for all D. Part (ii): When um is linear, u′m is constant and the first-best optimal subsidy
schedule mechanism is a priority mechanism P (s,m) = h−1(h(s) + u′m). Part (iii): When
u′m(xm) ≥ km for xm ≤ xtarm and u′m(xm) = 0 for xm > xtar
m where km is sufficiently large
for all m ∈ M and∑
m∈M xtarm < q, observe that the optimal mechanism admits xm ≥ xtar
m
for all m ∈ M in all states of the world, but conditional on xm ≥ xtarm for all m ∈ M
admits the highest scoring set of agents. A quota Qm = xtarm and QR = q−
∑m∈M xtar
m , with
D(R) = |M|+ 1 implements this allocation and is first-best optimal.
4.3. Proof of Proposition 1
Proof. Part (i): First, if we admit all minority students over some threshold s, the total score
of admitted minority students is κ∫ 1
ssds. Moreover, when we admit measure x minority
students where x ≤ minκ, q, this admissions threshold is defined by x = κ∫ 1
sds = κ(1− s).
Thus, we have that s = 1− xκ. Finally, the residual measure q−x admitted majority students
all score ω. Thus, the total score is given by s = qω+ (1−ω)x− 12κx2 for 0 ≤ x ≤ minκ, q.
As both quota and priority policies always admit the highest-scoring minority students, the
authority’s utility is given by:
U = qE[ω] + E[(1 + γ − ω)x]− 1
2
(1
κ+ γβx
)E[x2] (25)
We now derive the admitted measure of minority students. In the absence of a priority
or quota policy, α = 0 or Q = 0, we have that x = κ(1 − ω) measure minority students
are admitted. Thus, under a quota policy Q, measure x = maxQ, κ(1 − ω) minority
students are admitted. Under a priority policy, the measure of admitted minority students
is x = κ∫ 1
ω−α dx = κ(1 + α− ω). In each case x is capped by minκ, q and floored by 0.
The expected utility function over quotas is given by one of four cases. First, Q >
15
minκ, q and:
UQ(Q) = qE[ω] + (1 + γ − E[ω]) minκ, q − 1
2
(1
κ+ γβx
)minκ, q2 (26)
Second, Q ∈ [κ(1− ω),minκ, q) and:12
UQ(Q) = qE[ω] + (1 + γ − E[ω])Q− 1
2
(1
κ+ γβx
)Q2 (27)
Third, Q ∈ (κ(1− ω), κ(1− ω)) and:
UQ(Q) = qE[ω] +
∫ ω
1−Qκ
((1 + γ − ω)Q− 1
2
(1
κ+ γβx
)Q2
)dΛ(ω)
+
∫ 1−Qκ
ω
((1 + γ − ω)κ(1− ω)− 1
2
(1
κ+ γβx
)(κ(1− ω))2
)dΛ(ω)
(28)
Finally, Q ≤ κ(1− ω) and:
UQ(Q) = qE[ω] + E [(1 + γ − ω)κ(1− ω)]− 1
2
(1
κ+ γβx
)E[(κ(1− ω))2] (29)
We claim that the optimum lies in the second case. See that in case two the strict maximum
is attained at Q∗ = 1+γ−E[ω]1κ
+γβx∈ (κ(1 − ω),minκ, q), by our assumptions that minκ, q >
1+γ−ω1κ
+γβx+ κ(ω − ω) and κ(1− ω) < 1+γ−ω
1κ
+γβx. Moreover, in case three, the first derivative of the
payoff is given by:
U ′Q(Q) =
∫ ω
1−Qκ
((1 + γ − ω)−
(1
κ+ γβx
)Q
)dΛ(ω) (30)
Thus, checking that the sign of this is positive amounts to verifying that for all Q ∈ (κ(1−ω), κ(1− ω)), we have that:
Q <1 + γ − E[ω|ω ≥ 1− Q
κ]
1κ
+ γβx(31)
As the RHS is an increasing function of Q, it suffices to show that:
κ(1− ω) <1 + γ − ω
1κ
+ γβx(32)
12By our maintained assumptions we have that this interval has non-empty interior.
16
which we have assumed. We therefore have that:
VQ = qE[ω] + (1 + γ − E[ω])Q∗ − 1
2
(1
κ+ γβx
)Q∗2 (33)
We now turn to characterizing the value of priorities. There are three cases to consider.
First, when κ(1 + α− ω) ≥ minκ, q we have that x = minκ, q and:
UP (α) = qE[ω] + (1 + γ − E[ω]) minκ, q − 1
2
(1
κ+ γβx
)minκ, q2 (34)
Second, when κ(1 + α− ω) ≥ minκ, q ≥ κ(1 + α− ω) we have that:
UP (α) = qE[ω] +
∫ 1+α−min qκ,1
ω
((1 + γ − ω) minκ, q − 1
2
(1
κ+ γβx
)minκ, q2
)dΛ(ω)
+
∫ ω
1+α−min qκ,1
((1 + γ − ω)κ(1 + α− ω)− 1
2
(1
κ+ γβx
)[κ(1 + α− ω)]2
)dΛ(ω)
(35)
Finally, when minκ, q ≥ κ(1 + α− ω), we have that:
UP (α) = qE[ω] + E[(1 + γ − ω)κ(1 + α− ω)]− 1
2
(1
κ+ γβx
)E[(κ(1 + α− ω))2] (36)
We claim that the optimum under our assumptions lies only the third case. First, we argue
that there is a unique local maximum in the third case. Second, we show the value in the
second case is decreasing in α. By continuity, the unique optimum then lies in the third case.
First, it is helpful to write x(α) = κ(1 + α − E[ω]) and ε = κ (E[ω]− ω). The value in
the third case can then be re-expressed as:
UP (α) = qE[ω] + E[(1 + γ − ω) (x(α) + ε)]− 1
2
(1
κ+ γβx
)E[(x(α) + ε)2]
= qE[ω] + (1 + γ − E[ω])x(α)− E[ωε]− 1
2
(1
κ+ γβx
)x(α)2 − 1
2
(1
κ+ γβx
)E[ε2]
(37)
Finally, we have that E[ε2] = κ2V[ω] and E[ωε] = C[ω, ε] = −κV[ω]. Thus:
UP (α) = qE[ω] + (1 + γ − E[ω])x(α)− 1
2
(1
κ+ γβx
)x(α)2 +
κ
2(1− κγβx)V[ω] (38)
We then see that the optimal α∗ in this range sets x(α∗) = Q∗ < minκ, q. It remains only
17
to check that this optimal α∗ indeed lies within this case, or equivalently that κ(1+α∗−ω) ≤minκ, q. To this end, see that κ(1 + α∗ − E[ω]) = Q∗, and:
κ(1 + α∗ − ω) = Q∗ + κ(E[ω]− ω) ≤ Q∗ + κ(ω − ω)
≤ 1 + γ − ω1κ
+ γβx+ κ(ω − ω) < minκ, q
(39)
where the final inequality follows by our assumption that minκ, q > 1+γ−ω1κ
+γβx+ κ(ω − ω).
Second, in the second case we have that the first derivative of the payoff in α is given by:
U ′P (α) =
∫ ω
1+α−min qκ,1
d
dα
((1 + γ − ω)κ(1 + α− ω)− 1
2
(1
κ+ γβx
)[κ(1 + α− ω)]2
)dΛ(ω)
= κ
∫ ω
1+α−min qκ,1
((1 + γ − ω)−
(1
κ+ γβx
)(m(α) + ε(ω))
)dΛ(ω)
(40)
Checking that the sign of this is negative for all α such that κ(1 + α − ω) ≥ minκ, q ≥κ(1 + α− ω) then amounts to checking that:
x(α) >1 + γ − E[ω|ω ≥ 1 + α−min q
κ, 1]
1κ
+ γβx− E
[ε(ω)|ω ≥ 1 + α−min q
κ, 1]
(41)
for all x(α) ∈ [minκ, q−κ(E[ω]−ω),minκ, q−κ(E[ω]−ω)]. So it suffices to check that
the minimal possible value of the LHS exceeds the maximal possible value of the RHS. A
sufficient condition for this is that:
minκ, q − κ(E[ω]− ω) >1 + γ − ω
1κ
+ γβx− κ(E[ω]− ω) (42)
Which holds as we assumed that minκ, q > 1+γ−ω1κ
+γβx+ κ(ω − ω). We have now established
that:
∆ = VP − VQ =κ
2(1− κγβx)V[ω] (43)
Part (ii): By Theorem 1, we have that A(y) = u′(y) = γ(1 − βxy) is first-best optimal.
See in state ω that the payoff from admitting x(ω) minority students is given by:
qω + (1 + γ − ω)x(ω)− 1
2
(1
κ+ γβx
)x(ω)2 (44)
18
Thus, the x(ω) that solves the FOC is given by:
x(ω) =κ(1 + γ − ω)
1 + κγβx(45)
Under our maintained assumptions, we have that:
x(ω) =κ(1 + γ − ω)
1 + κγβx≤ 1 + γ − ω
1κ
+ γβx+ κ(ω − ω) < minκ, q (46)
and:
x(ω) =κ(1 + γ − ω)
1 + κγβx≥ 1 + γ − ω
1κ
+ γβx> κ(1− ω) (47)
Thus, this policy is feasible. Substituting, we have that:
V ∗ = qE[ω] +1
2
E[κ(1 + γ − ω)2]
1 + κγβx(48)
We have already shown that the value functions of priorities and quotas are given by:
VQ = qE[ω] +1
2
κ(1 + γ − E[ω])2
1 + κγβx, VP = VQ +
κ
2(1− κγβx)V[ω] (49)
We can therefore compute the loss from restricting to quota policies:
LQ =1
2
κV[ω]
1 + κγβx(50)
To find the loss from restricting to priority policies, we compute:
LP = LQ −∆ =1
2(κγβx)
2 κV[ω]
1 + κγβx(51)
Enveloping over these losses yields the claimed formula.
4.4. Proof of Corollary 2
Proof. We show that a quota-second policy Q is equivalent to a priority subsidy of α(Q) = Qκ
.
A quota-second policy admits the highest-scoring x = κ(1−ω)+Q minority students, floored
by zero and capped by minκ, q. A priority policy α(Q) = Qκ
admits the highest-scoring
x = κ(1 + α(Q) − ω) = κ(1 − ω) + Q minority students, floored by zero and capped by
minκ, q. Thus, state-by-state, quota-second policy Q and priority subsidy α(Q) = Qκ
yield
the same allocation. The claims then follow from Proposition 1.
19
References
Aygun, O., Bo, I., 2021. College admission with multidimensional privileges: The brazilian
affirmative action case. American Economic Journal: Microeconomics, Forthcoming .
Azevedo, E. M., Leshno, J. D., 2016. A supply and demand framework for two-sided matching
markets. Journal of Political Economy 124, 1235–1268.
Celebi, O., Flynn, J. P., 2021. Priority design in centralized matching markets. The Review
of Economic Studies, Forthcoming .
Delacretaz, D., 2020. Processing Reserves Simultaneously. Unpublished working paper, Ox-
ford University .
Dur, U., Kominers, S. D., Pathak, P. A., Sonmez, T., 2018. Reserve design: Unintended
consequences and the demise of boston’s walk zones. Journal of Political Economy 126,
2457–2479.
Dur, U., Pathak, P. A., Sonmez, T., 2020. Explicit vs. statistical targeting in affirmative
action: theory and evidence from chicago’s exam schools. Journal of Economic Theory p.
104996.
Echenique, F., Yenmez, M. B., 2015. How to control controlled school choice. American
Economic Review 105, 2679–94.
Ehlers, L., Hafalir, I. E., Yenmez, M. B., Yildirim, M. A., 2014. School choice with controlled
choice constraints: Hard bounds versus soft bounds. Journal of Economic Theory 153,
648–683.
Erdil, A., Kumano, T., 2019. Efficiency and stability under substitutable priorities with ties.
Journal of Economic Theory 184, 104950.
Hafalir, I. E., Yenmez, M. B., Yildirim, M. A., 2013. Effective affirmative action in school
choice. Theoretical Economics 8, 325–363.
Kojima, F., 2012. School choice: Impossibilities for affirmative action. Games and Economic
Behavior 75, 685–693.
Kominers, S. D., Sonmez, T., 2016. Matching with slot-specific priorities: Theory. Theoret-
ical Economics 11, 683–710.
Pathak, P. A., Rees-Jones, A., Sonmez, T., 2020a. Immigration lottery design: Engineered
and coincidental consequences of h-1b reforms. Tech. rep., National Bureau of Economic
Research.
Pathak, P. A., Rees-Jones, A., Sonmez, T., 2020b. Reversing reserves. Tech. rep., National
Bureau of Economic Research.
Pathak, P. A., Sonmez, T., Unver, M. U., Yenmez, M. B., 2021. Leaving no ethical value
behind: Triage protocol design for pandemic rationing. Tech. rep.
20
Sonmez, T., Yenmez, M. B., 2020a. Affirmative action with overlapping reserves. Tech. rep.,
Boston College.
Sonmez, T., Yenmez, M. B., 2020b. Constitutional implementation of reservation policies in
india. Tech. rep., Boston College Department of Economics.
Sonmez, T., Yenmez, M. B., 2021. Can economic theory be informative for the judiciary?
affirmative action in india via vertical and horizontal reservations. Tech. rep., Boston
College.
Weitzman, M. L., 1974. Prices vs. quantities. The Review of Economic Studies 41, 477–491.
Westkamp, A., 2013. An analysis of the german university admissions system. Economic
Theory 53, 561–589.
21