Applications of Screening

Shota Ichihashi∗

ECON 813

In this lecture we study applications of monopoly screening and the techniques developed there.
There are many important applications, and we cover three of them: auctions, insurance, and
taxation.

1 Optimal Auctions (Myerson, 1981)

A seller (or an auctioneer) has a single individual good to sell. There are n ≥ 2 bidders. Each bidder
i has value θi for the good, which is distributed according to a positive density fi on [θi , θi ] with
θi > θi ≥ 0. Bidders’ values are independent. What auction rule maximizes the seller’s expected
revenue, subject to the constraint that each bidder obtains a non-negative expected payoff in the
auction?
This problem may look quite different from the screening model, as there are multiple bidders
and the seller is not choosing quantity or quality. However, it turns out to be similar to what we
have learned. Recall that in the previous lecture, a buyer’s utility from the goods is θv(q) where
q represents quantity or quality. For the auction problem, we can interpret q as the probability
with which a buyer wins the good. We can then write a buyer’s utility as θi qi , where qi is the
probability for bidder i of obtaining the good (i.e., winning the auction). The fact that variable qi
is a probability introduces two feasibility constraints:
X
qi (θ) ≤ 1 and qi (θ) ∈ [0, 1], ∀i, ∀θ = (θ1 , . . . , θn ). (1)
i

The first condition means that the probability that the seller allocates the good to some bidder is
at most 1; the second condition means that the probability that bidder i wins the good is between
0 and 1.
We now formally set up the optimal auction problem. Let Θ = i [θi , θi ] ⊂ Rn denote the space
Q
Q
of value profiles, and let f (θ), which equals i fi (θ) due to the independence assumption, denote

∗
Various sections of this note draw heavily on notes written by Alexander Wolitzky and “Microeconomic Theory”
by Mas-Colell, Whsinston and Green. The opinions expressed in this article are the author’s own and do not reflect
the views of the Bank of Canada.

1
the joint density of n bidders’ values. We also write θ−i for a profile of values of bidders other
Q Q
than i, Θ−i for the set j6=i Θj of value profiles excluding bidder i, and f−i = j6=i fj for the joint
density of θ−i .
Revelation principle implies that, to maximize revenue, the seller can focus on a direct mecha-
nism (q1 (θ), . . . , qn (θ)) and (t1 (θ), . . . , tn (θ)), where qi (θ) is the probability that bidder i wins the
object given value profile θ, and ti (θ) is the payment from bidder i to the seller.
The problem does not look like an auction problem, because we apply revelation principle
and work on direct mechanism. However, once we derive an optimal direct incentive compatible
mechanism, we can check whether common auction formats such as first-price or second price
auctions would attain the optimal revenue. For now, we focus on the task of characterizing the
optimal direct incentive compatible mechanism.
R R
Let Qi (θi ) = Θ−i qi (θi , θ−i )f−i (θ−i )dθ−i and Ti (θi ) = Θ−i ti (θi , θ−i )f−i (θ−i )dθ−i denote the
expected allocation and expected transfer for bidder i with type θi . The seller’s problem is as
follows:
Z X
max Ti (θi )fi (θi )dθi
q(θ),t(θ) Θ
i

subject to
(ICθi ) θi Qi (θi ) − Ti (θi ) ≥ θi Qi (θi0 ) − Ti (θi0 ), ∀θi , θi0 ∈ [θi , θi ], i = 1, . . . , n
(IRθi ) θi Qi (θi ) − Ti (θi ) ≥ 0, i = 1, . . . , n
q(θ) is feasible, i.e., it satisfies (1).

Note that ICθi and IRθi are exactly the same as screening problem. That is, we can show that the
IC is equivalent to the monotonicity of Qi (·) and the integral version of the local IC constraint:
Z θi
Ui (θi ) = Ui (θi ) + Qi (x)dx
θi

where Ui (θi ) = θi Qi (θi ) − Ti (θi ). Using Ui (θi ) = 0 (i.e., the IR constraint for the lowest type is
binding), we obtain
Z θi
Ti (θi ) = θi Qi (θi ) − Qi (x)dx. (2)
θi

Plugging this into the seller’s objective and using the integration by parts, we can rewrite the
seller’s problem as follows:
Z X 
1 − Fi (θi )
max θi − Qi (θi )fi (θi )dθi .
q(θ) Θ fi (θi )
i
R
Note that Qi (θi ) = Θ−i qi (θi , θ−i )f−i (θ−i )dθ−i , so we can write the objective in terms of qi , not

2
Qi :
Z X 
1 − Fi (θi )
max θi − qi (θ)fi (θ)dθ
q(θ) Θ fi (θi )
i

subject to
Monotonicity: Qi (θi ) is non-decreasing in θi
q(θ) is feasible, i.e., it satisfies (1).

The problem is simple to solve when virtual values areinon-decreasing in θi . For any θ = (θ1 , . . . , θn ),
P h 1−Fi (θi )
what (q1 (θ), . . . , qn (θ)) maximizes i θi − fi (θi ) qi (θ)? If bidder i’s virtual θi − 1−F i (θi )
fi (θi ) is
negative for all i, we should set qi (θ) = 0 for all i. If some bidder has a positive virtual value,
then we should set qi (θ) = 1 for bidder i who has the highest (positive) virtual value. Note that
this (q1 (·), . . . , qn (·)) satisfies the monotonicity constraint (i.e., Qi (θi ) is non-decreasing) because of
monotone virtual values.
To characterize the optimal mechanism, we also need to construct (t1 (θ), . . . , tn (θ)). One pos-
sible ti is Z θi
ti (θ) = θi qi (θ) − qi (x, θ−i )dx. (3)
θi

Indeed, if we take expectation of both sides of the above equation with respect to θ−i (and switching
the order of integrals with respect to x and θ−i ), we obtain equation (2).
As an example, suppose that each θi is uniformly distributed on [0, 1]. The virtual value is
1−Fi (θi )
θi − fi (θi ) = 2θi − 1. Thus, the optimal auction allocates the object to bidder i who has the
1
highest value θi , provided that it exceeds 2. Equation (3) tells us bidder i’s payment. If bidder
1
i does not win the object (namely, if θi is less than 2 or less then the highest value among other
bidders), qi (x, θi ) = 0 for all x ≤ θi , so ti (θ) = 0. If bidder i wins the object (namely, if θi is greater
1
than 2 and the highest value among other bidders), she pays
Z θi
ti (θ) = θi · 1 − 1dx
max{ 12 ,θ−i
∗
}
 
1 ∗
= max , θ−i ,
2
∗ is the highest values among the bidders excluding bidder i. This auction format is a
where θ−i
1
2nd price auction with reserve price 2. (To be precise, the truthful bidding equilibrium of the
second-price auction with reserve price 21 implements (q1 , . . . , qn ) and (t1 , . . . , tn ).)
We already have necessary ingredients to prove one of the most important results in auction
theory, revenue equivalence theorem.

Theorem 1. Suppose bidder i’s type is drawn from an interval [θi , θi ] with strictly positive density
fi , independently across bidders. Then any auction format and equilibrium in which

3
 1. the good is always allocated to the bidder with the highest value, and

2. a bidder with her lowest possible type (θi ) gets payoff 0,

the seller’s expected revenue is the same.

To see this, note first that any auction format and an equilibrium generates some (q1 (θ), . . . , qn (θ))
and (t1 (θ), . . . , tn (θ)). The induced mechanism is incentive compatible: If bidder i has true type θi
but strictly prefers to report θi0 6= θi , the bidder would deviate and imitate type θi0 in the auction,
which contradicts that we focus on an equilibrium of the auction. Point 2 of the theorem implies
Ui (θi ) = 0, and Point 1 determines (q1 (θ), . . . , qn (θ)), which in turn determines Ti (θi ) from (2).
Thus, the seller’s expected revenue is
Z X 
1 − Fi (θi ) ∗
θi − qi (θ)fi (θ)dθ,
Θ fi (θi )
i

where qi∗ (θ) is the allocation rule whereby the highest value bidder obtains the good with probability
1. Finally, in any auction and equilibrium that satisfies the conditions in Theorem 1, bidder i’s
expected payoff is uniquely determined as Ui (θi ) = θi Qi (θi ) − Ti (θi ). Thus all auction formats that
satisfy these properties are equivalent not only in the seller’s expected revenue but also in each
bidder’s expected payoff.
An important remark is that Theorem 1 is one version of revenue equivalence theorem, and it
is not necessarily about optimal auctions. Indeed, we have proved that an auction that satisfies
Point 1—i.e., it allocates the good even when all bidders have negative virtual values—does not
maximize revenue. A more general revenue equivalence theorem, which considers optimal auctions,
is as follows.

Theorem 2. Suppose bidder i’s type is drawn from an interval [θi , θi ] with strictly positive density
fi , independently across bidders. Take two auctions formats, and pick an equilibrium from each
auction format. Suppose that the two equilibria have the same allocation rule and the same payoff
to a bidder with her lowest possible type. Then the two auction formats have the same expected
revenue for the seller.

2 Optimal Insurance (Stiglitz, 1977)

A risk-neutral principal (an insurance company) faces a risk-averse agent with utility function u
such that u0 > 0 and u00 < 0. The agent has initial wealth W > 0 but will incur a loss of L > 0
with probability θ ∈ (0, 1) (e.g., θ is the probability of a car accident). The probability θ is private
of information of the agent: With probability β, the agent has θ = θL ; with probability 1 − β, the
agent has θ = θH > θL . The principal chooses a contract that determines a payment from the
agent to the principal when the loss does not occur (i.e., premium), and a payment when the loss
occurs (L minus deductibles). What insurance contract maximizes the principal’s expected profit?

4
2.1 The First-Best Solution

Suppose that the principal can observe the agent’s type θ. Because W is commonly known, specify-
ing a payment from the agent to the principal is equivalent to specifying the agent’s consumption.
For example, asking the agent to pay t in case of no loss is equivalent to the agent consuming W −t.
To simplify notation, we suppose that a contract directly specifies the agent’s consumption (c1 for
the no loss event, and c2 for the loss event). The first-best problem is as follows.

max(1 − θ)(W − c1 ) + θ(W − c2 )

c1 ,c2

subject to
IRθ : (1 − θ)u(c1 ) + θu(c2 ) ≥ (1 − θ)u(W ) + θu(W − L).

In the objective function, (1 − θ)W and θW are constants and do not depend on (c1 , c2 ). Thus we
can rewrite the problem as

min(1 − θ)c1 + θc2

c1 ,c2

subject to
IRθ : (1 − θ)u(c1 ) + θu(c2 ) ≥ (1 − θ)u(W ) + θu(W − L).

Write the LHS of the IR constrains as U (θ). We show the first-best contract is

c1 = c2 = u−1 (U (θ)). (4)

A simple way is to observe that if c1 6= c2 , then the principal could instead offer c1 = c2 = c0 such
that u(c0 ) = (1 − θ)u(c1 ) + θu(c2 ). Since the agent is risk averse, this certainty equivalent c0 is less
than the expected consumption (1 − θ)c1 + θc2 in the original contract. Given c1 = c2 = c, IR
constraint becomes u(c) = U (θ), so c = u−1 (U (θ)).
An alternative way to show the same result is this: Because the object is increasing in (c1 , c2 ),
the IR constraint must bind at the optimum. Rewriting the constraint as c2 = · · · and plugging it
into the objective, we obtain the following:
  
U (θ) − (1 − θ)u(c1 )
min(1 − θ)c1 + θ u−1 .
c1 ,c2 θ

The first-order condition with respect to c1 becomes u0 (c1 ) = u0 (c2 ), or equivalently c1 = c2 . Thus
the optimal contract is c1 = c2 = u−1 (U (θ)). This is a standard result about optimal-risk sharing
under complete information: it is optimal for a risk-neutral party to fully insures a risk-averse one.

5
2.2 The Second-Best Solution

The principal’s problem is now

max β[(1 − θL )(W − cL L H H

1 ) + θL (W − c2 )] + (1 − β)[(1 − θH )(W − c1 ) + θH (W − c2 )]
cL L H H
1 ,c2 ,c1 ,c2

subject to
ICθH : (1 − θH )u(cH H L L
1 ) + θH u(c2 ) ≥ (1 − θH )u(c1 ) + θH u(c2 )

ICθL : (1 − θL )u(cL L H H
1 ) + θL u(c2 ) ≥ (1 − θL )u(c1 ) + θL u(c2 )

IRθH : (1 − θH )u(cH H
1 ) + θH u(c2 ) ≥ (1 − θH )u(W ) + θH u(W − L)

IRθL : (1 − θL )u(cL L
1 ) + θH u(c2 ) ≥ (1 − θL )u(W ) + θL u(W − L).

What constraints do we conjecture to be binding? One hint is that at the first-best, the agent is
fully insured and obtains a payoff of

(1 − θi )u(W ) + θi u(W − L). (5)

For example, if the agent has type θH and pretends to be θL , he gets a payoff of

(1 − θH )u(cFL B ) + θH u(cFL B ) = u(cFL B ) = (1 − θL )u(W ) + θL u(W − L).

Because θL < θH , both the high type and the low types strictly prefer the contract for the low type
than that for the high type. As a result, at the first-best, the high type wants to mimic the low
type, and the low type does not want to mimic the high type. Thus let’s conjecture that ICθH is
binding and ICθL is slack. Given ICθL is slack, the principal can reduce c1 and c2 to make IRθL
binding, without violating ICθH and IRθH .
We thus solve the problem assuming that ICθH and IRθL are binding; after solving this problem
we need to check that the solution satisfies ICθL and IRθH (which we leave as an exercise).
Note that cH H
1 and c2 appear only in the left-hand side of ICθH . Thus the principal can again
set cH H
1 = c2 (otherwise, the principal can replace the original contract for θH with its certainty
equivalent). This reduces our problem to

min β[(1 − θL )cL L H

1 + θL c2 ] + (1 − β)c2
cL L H
1 ,c2 ,c

subject to
ICθH : u(cH ) ≥ (1 − θH )u(cL L
1 ) + θH u(c2 )

IRθL : (1 − θL )u(cL L
1 ) + θL u(c2 ) ≥ U (θL ).

6
Using the binding ICθH and IRθL to eliminate cH and cL
2 , we obtain

U (θL ) − (1 − θL )u(cL
    
−1 1) θH θH − θ L
min β (1 − θL )cL
1 + θL u + (1 − β) U (θL ) − L
u(c1 ) .
cL
1
θL θL θL

The first-order condition in cL

1 is
1

u0 (cL
1) 1 − β θH − θL u0 (cL 1)
0 L
= 1 − 0 H
< 1. (6)
u (c2 ) β θ L (1 − θ L ) u (c )

u0 (cL
1)
Note that u0 (cL
< 1 implies cL L
1 > c2 , that is, the low-risk type is not fully insured. This contrasts
2)
with the first best in which both types are fully insured. Why? We know the intuition: The
principal makes the contract to the low type less attractive and reduce the information rent of the
high type. To see this formally, the binding ICθH (after eliminating cL
2 using IRθL ) implies

θH θH − θL
u(cH ) = U (θL ) − u(cL
1 ). (7)
θL θL

The equation implies that the high-risk type’s information rent is decreasing in cL
1 , which is the
agent’s consumption in case of no loss. With a fixed level of utility U (θL ) for the low type,
consuming more in case of no loss and consuming less in case of loss correspond to underinsurance,
which makes the contract less attractive to the high-risk type and reduces the information rent.

3 Optimal Taxation (Mirrlees, 1971)

A utilitarian government faces a continuum of individuals, each of whom is either low-ability (θ =
θL , fraction β) or high-ability (θ = θH , fraction 1 − β). Each individual chooses how much effort e
to exert, and produces output y = θe. An individual’s utility function is

u(y − t − c(e)),

where y is her production, t is the net tax she pays to the government, and c(e) is her cost of effort.
Assume u0 > 0, u00 < 0, c0 > 0, and c00 > 0. The government’s budget constraint is

βtL + (1 − β)tH ≥ 0.

The government’s problem is to design a tax system to maximize utilitarian social welfare

βu(θL eL − tL − c(eL )) + (1 − β)u(θH eH − tH − c(eH )).

1
Recall that the derivative of the inverse is the reciprocal of the derivative, i.e., (u−1 )0 (x) = 1
u0 (u−1 (x))
. Thus the
L
 
−1 U (θL )−(1−θL )u(c1 ) −1 0 L
derivative of cL
2 = u θL
with respect to cL
1 is u0 (cL ) · (1 − θL )u (c1 ).
2

7
3.1 Complete Information

If the government can observe ability and output (and thus effort), its problem is

max βu(θL eL − tL − c(eL )) + (1 − β)u(θH eH − tH − c(eH )).

eH ,eL ,tH ,tL

subject to the budget constraint. (We do not impose participation constraints, which does not
cause a problem given the government’s objective.) The budget constraint binds at the optimum,
so we can use βtL +(1−β)tH = 0 to eliminate tL or tH from the objective. This yields the first-order
conditions

u0L = u0 (θL eL − tL − c(eL )) = u0 (θH eH − tH − c(eH )) = u0H

c0 (eL ) = θL
c0 (eH ) = θH .

That is, the government has each individual exert first-best effort, and then redistributes income so
as to equalize marginal utilities (and hence total utilities: since u00 < 0, u0L = u0H implies uL = uH ).

3.2 Incomplete Information

If the government can only observe output, its problem is

max βu(θL eL − tL − c(eL )) + (1 − β)u(θH eH − tH − c(eH )) (8)

eH ,eL ,tH ,tL

subject to (9)
 
θL
(ICH) θH eH − tH − c(eH ) ≥ θL eL − tL − ceL (10)
θH
 
θH
(ICL) θL eL − tH − c(eL ) ≥ θH eH − tH − c eH (11)
θL
(Budget) βtL + (1 − β)tH ≥ 0. (12)

Note again that we do not have an IR constraint. To understand ICH, for example, note that if a
high-ability type takes the contract intended for the low-ability type, he must produce output θL eL ,
θL
which requires effort θH e L
on his part (since his true ability is θH ). Thus if a high-type reports to
 
be a low type, an individual gets utility θL eL − tL − c θθHL eL , the right-hand side of ICH. Note
 
also that the “correct” IC constraint would be u(θH eH − tH − c(eH )) ≥ u(θL eL − tL − c θθHL eL ),
but we write it in terms of consumption.
We conjecture that the binding constraints are the budget constraint and ICH. Note that ICH

8
is violated at the first-best, because at the first-best we have
 
θL
θH eH − tH − c(eH ) = θL eL − tL − c (θL ) < θL eL − tL − c eL .
θH

The binding budget constraint and ICH yield

  
θL
tH =β θH eH − c(eH ) − θL eL + c eL
θH
  
θL
tL = − (1 − β) θH eH − c(eH ) − θL eL + c eL .
θH

Substituting tH and tL out of the objective and taking first-order conditions with respect to eH
and eL yield

c0 (eH ) = θH
u0L − u0H
   
0 0 θ L 0 θL
c (eL ) = θL − (1 − β) c (eL ) − c eL
βuL + (1 − β)u0H
0 θH θH
< θL .

Hence, we have first-best effort for high-ability types and inefficiently low effort for low-ability types.
As usual, the reason why eL is distorted downward is that doing so realxes ICH. To understand
this in more detail, note that high-ability types’ information rent in consumption terms (i.e., the
difference in consumption net of effort costs between high- and low-ability types) is
 
θL
c(eL ) − c eL .
θH

This is intuitive: the benefit a high-ability type gets

 from pretending to be a low-ability type is
θL
that she can produce output θL eL at cost c θH eL rather than c(eL ). Convexity of c implies that
a marginal decrease in eL reduces high-ability types’ information rents in consumption terms by
u0L −u0H

0 θL 0 θL
c (eL ) − θH c θH eL > 0. The βu0 +(1−β)u0 term (i.e., the difference in marginal utilities between
L H
the low- and high-ability types, normalized by average marginal utility) then measures the benefit
of this reduction in inequality from the perspective of utilitarian social welfare. That is, by reducing
eL , the government can charge higher tax on high types and redistribute the tax revenue to low
types.
This model has interesting implication for optimal taxation. In particular, it suggests that the
marginal tax rate should be zero for the highest income and positive for lower incomes (although
there is of course redistribution through a lump-sum subsidy—indeed, the entire point of the
positive marginal tax on lower incomes is to make redistribution possible by getting high-ability
types to reveal themselves). However, it is important to recall that with a continuum of types, the
zero marginal tax at the top result only holds for the very highest income, and from our analysis

9
last time we know that the optimal marginal tax for very high incomes just below the top depends
1−F (θ)
on the inverse hazard rate of the distribution of abilities f (θ) . Diamond (1998) and Saez (2001)
aruge that, with a realistic ability distribution, the Mirrlees model is consistent with high marginal
taxes on the rich.

10