• Nenhum resultado encontrado

The Principal-Agent Problem

N/A
N/A
Protected

Academic year: 2023

Share "The Principal-Agent Problem"

Copied!
51
0
0

Texto

(1)

Leandro Gorno

October 29, 2013

(2)

Principal-agent problems:

I Uncertainty might develop after the establishment of a contractual relation:

- when one party has to make unobservable choices.

- when one party obtains private information.

I Principal-agent framework:

- one party (principal) designs the contract.

- the other party (agent) decides whether to accept it or not.

I Two extreme cases:

- Hidden actions (moral hazard).

- Hidden information (or monopolistic screening).

(3)

Specification:

I The owner of a firm (principal) wants to hire a manager (agent).

I Manager can exert two levels of effort: eL= 0 and eH = 1.

I The revenue of the firm X is stochastic and depends on exerted effort:

- Uncertainty is described by PDFs f0(x) andf1(x) -f1(x) first order stochastically dominates f0(x).

I Effort is not observable, but realized revenue is:

- Manager compensation can depend on firm performance:

w0 and w1.

(4)

Ex-post payoffs:

I Principal: X −W

I Agent: u(W,e) =v(W)−c(e) -v increasing and concave.

-c increasing.

- If she does not accept any contract, she gets u.

Ex-ante payoffs:

I Principal: E{X −W|e}=R

(x−w(x))fe(x)dx

I Agent: E{u(W,e)|e}=R

v(w(x))fe(x)dx−c(e)

(5)

Observable actions benchmark:

I The problem of the principal is:

maxe,w

Z

(x −w(x))fe(x)dx Z

v(w(x))fe(x)dx −c(e)≥u

I We can divide it in two subproblems:

- How to optimally implement any fixed action?

minw

Z

we(x)fe(x)dx Z

v(we(x))fe(x)dx−c(e)≥u

- Which action is optimal?

(6)

Observable actions benchmark:

I Lagrangian:

L= Z

we(x)fe(x)dx+γe

u+c(e)− Z

v(we(x))fe(x)dx

I FOC:

1

v0(we(x)) =γe a.s.

(7)

Observable actions benchmark:

I With agent strictly risk averse, optimal wage is constant:

we(x) = (v0)−1(1/γe) =we.

I The wage level is pinned down by reservation utility:

v(we) = Z

v(we)fe(x)dx =u+c(e).

I If agent is risk neutral, any compensation scheme which gives her exactly her reservation utility is optimal.

(8)

Observable actions benchmark:

I Then, the principal maximizes:

Z

(x −w(x))fe(x)dx = Z

xfe(x)dx −v−1(u+c(e)).

Example 1:

I Supposev(w) = ln(x) and c(e) =e.

I Then high effort is optimal for the principal if and only if E{X|e = 1} ≥E{X|e = 0}+ exp (u+ 1)−exp (u).

(9)

Non-observable actions:

I Incentives require inducing risk: efficient action vs.

insurance trade-off.

I If the agent is risk-neutral, this is not a trade-off!

I Hence, the outcome of the optimal contract under full information is still implementable when actions are not observable.

(10)

Risk-neutral agent:

I The principal sells the firm: w(x) =x −p.

I The agent doing effort e will now get a payoff of:

Z

w(x)fe(x)dx−c(e) = Z

xfe(x)dx−p−c(e) which she will willing to take as long as

p ≤pˆe = Z

xfe(x)dx −c(e)−u

I Note that ˆpe is the expected payoff of the principal under full information.

(11)

Risk-neutral agent:

I Then,p = max ˆpe is the optimal effort under full information.

I Hence, if the principal choosesp =p, then the agent will accept the contract and do an optimal effort.

I The intuition is simple: fully incentivizing the agent is optimal because she does not demand a premium for bearing the risk.

(12)

Risk-averse agent:

I Fully incentivizing the agent is now costly.

I Implementing low effort is still easy:

w(x) =w0 =v−1(u+c(0))

I Implementing high effort is more tricky.

(13)

Implementing high effort:

I To optimally implemente = 1, the principal must solve:

minw

Z

w1(x)f1(x)dx subject to the participation constraint:

Z

v(w1(x))f1(x)dx−c(1) ≥u and the incentive compatibilityconstraint:

Z

v(w1(x))f1(x)dx −c(1) ≥ Z

v(w1(x))f0(x)dx −c(0).

(14)

Implementing high effort:

I Lagrangian:

L= Z

w1(x)f1(x)dx+γ1

u+c(e)− Z

v(w1(x))f1(x)dx

1,0

c(1)−c(0)− Z

v(w1(x))(f1(x)−f0(x))dx

, whereγ1 ≥0 and µ1,0 ≥0 are the multipliers.

I FOC:

1

v0(w1(x)) =γ11,0

1− f0(x) f1(x)

(15)

Implementing high effort:

I γ1 >0:

- Since effort affects profits,f0(x)>f1(x) on some open set.

- Then, ifγ1 = 0, the FOC would imply v0(w(x))<0.

- This is impossible, so γ1 >0.

I µ1,0 >0:

- Ifµ= 0, then the FOC would imply a constant wage.

- But facing that schedule, the agent would choosee = 0.

I Hence, both PC and IC bind at the optimum.

(16)

The likelihood-ratio:

I Given an optimal contract, define ˆw = (v0)−1(1/γ1).

I Define thelikelihood ratio as r(x) =f0(x)/f1(x).

I Then, the FOC can be written:

1

v0(w1(x)) = 1

v0( ˆw) +µ1(1−r(x))

I Thus,w1(x)>wˆ whenever r(x)<1 and viceversa.

I Intuitive way of providing incentives: pay relatively more in states more likely to occur under high effort.

(17)

The likelihood-ratio:

I But also surprising implication: optimal wages may decrease with performance!

I What do we need to obtain motonicity?

- MLRP:r(x) must be decreasing.

(18)

Example 2:

I Supposev(w) = 1−exp (w), c(e) = e/2, u = 0 and fe(x) = exp (−x/(1 + 5e)).

I Then, low effort is implemented by w0 = 0.

I To implement high effort, note that the FOC yields:

w1(x) = ln (γ11,0(1−r(x))),

wherer(x) = 6 exp (−5x/6). Note that the MLRP holds.

I Finally, it is a matter of finding multipliers such that both constraints hold with equality.

(19)

Example 2 (cont’d):

I The following picture shows those values of γ1 and µ1,0 for which the constraints bind:

(20)

Optimal effort level:

I Low effort is implemented through full-information wages.

I In order to implement high effort, higher wages (in expected value) are required to compensate the agent for the risk.

I Therefore, non-observability distorts effort downwards (this result does not generalize to multiple effort levels).

I Whenever high effort is optimal under full information, non-observability generates a welfare loss.

(21)

Multiple effort levels:

I Suppose nowe ∈[0,∞).

I Then, a typical IC constraint would look like this:

Z

v(we(x))fe(x)dx −c(e)≥ Z

v(we(x))fe0(x)dx−c(e0).

I But there are [0,∞)×[0,∞) of them!

(22)

The first-order approach:

I If e >0 and fe(x) is smooth in e, the following FOC is necessary:

Z

v(we(x))

∂fe(x)

∂e

dx =c0(e).

I Then, we can use it instead of the double continuum of ICs in the cost-minimization problem of the principal and hope for the best!

I Obvious problem: it may not be sufficient.

(23)

The first-order approach:

I Lagrangian:

L= Z

we(x)fe(x)dx+γe

u+c(e)− Z

v(we(x))fe(x)dx

e

c0(e)− Z

v(we(x))

∂fe(x)

∂e

dx

,

I FOC:

1

v0(we(x)) =γee

∂fe(x)

∂e

fe(x)

!

(24)

Example 3:

I If v(w) = ln(w) and fe(x) = exp (−λ(e)x)/λ(e), where λ(e) = 1/(1 +ke), the FOC derived using the first-order approach yields:

w(x) = γee

(1−λ(e)x)k λ(e)(1 +ke)2

I Hence, an optimal contract implementing positive effort should be linear!

(25)

A dynamic model:

I Based on “Optimal unemployment insurance” by Hopenhayn and Nicolini (JPE 1997).

I Unemployed worker seaches for jobs by exerting efforte ≥0.

I Probability of finding a job this month isp(e)∈(0,1), with p0(e)>0 and p00(e)<0.

I Worker’s monthly payoff isu(c(ht))−e(ht) after historyht while searching and u(w) after becoming employed.

I Government wants to ensure the unemployed worker (expected) utility V in the cheapeast possible way.

I Common discount rateβ ∈(0,1).

(26)

Worker’s problem:

I Value after finding a job:

VtE =VE = u(w) 1−β

I Optimal effort satisfies VtU = max

e

u(ct)−e+β p(e)VE + (1−p(e))Vt+1U )

I FOC for e >0:

βp0(e)(VE −Vt+1U ) = 1

(27)

Government’s problem:

I Problem doesn’t look recursive, but can be made so by using promised utility as a forward-looking state variable (“APS trick”).

I Bellman equation:

C(V) = min

c,e,VU

c +β(1−p(e))C(VU) subject to apromise keeping constraint

u(c)−e+β(p(e)VE + (1−p(e))VU ≥V

I and the (first-order) incentive compatibility constraint βp0(e)(VE −VU) = 1.

(28)

Analysis:

I Lagrangian:

L=c+β(1−p(e))C(VU) +µ 1−βp0(e)(VE −VU) +γ V −u(c) +e−β(p(e)VE + (1−p(e))VU

I FOCs forc,e and VU:

1 = γu0(c) C(VU) =−µp00(e)

p0(e)(VE−VU) +γ 1

βp0(e) −(VE −VU)

C0(VU) = γ−µ

p0(e) 1−p(e)

(29)

Analysis:

I Envelope theorem:

C0(V) = γ

I More on theVU FOC:

C0(VU) =C0(V)−µ

p0(e) 1−p(e)

<C0(V)

I So, if we assume conditions for C convex:

VU <V

I Sinceu0(c)C0(V) = 1, the replacement rate must be decreasing over time!

(30)

Analysis:

I More on thee FOC:

C(VU) = −µp00(e)

p0(e)(VE −VU)

Sincep00(e)<0, the optimal scheme requires VU <VE whenever the program is running a deficit.

I CombiningVU <V and βp0(e)(VE −VU) = 1, we can see that search effort increases over time!

(31)

Motivation:

I Suppose actions are observable but the agent only learns the cost of implementing different actions after signing the contract.

I In principle, the agent could communicate her private information to the principal, but the principal may not be able to verify the agent’s claim.

(32)

Assorted examples:

I Delegating construction.

I Debt restructuring.

I Public policy and constitutional design.

(33)

Model:

I Letx(e) be the (deterministic) gross profit of the principal when the agent exerts efforte ∈[0,∞).

I We assume x(0) = 0, x0(e)>0 and x00(e)<0 for alle.

I Suppose the uncertainty over agent’s costs is summarized by a state variableθ ∈ {θL, θH}.

I The probability ofθ =θH is λ∈(0,1).

(34)

Model:

I The utility of an agent receiving wage w and exerting effort e in stateθ is

v(w −g(e, θ)),

wherev is increasing and strictly concave and g satisfies:

-g(0, θ) = 0.

-ge(0, θ) = 0 and ge(e, θ)>0 for alle >0.

-gee(e, θ)>0.

-gθ(e, θ)<0.

-g(0, θ) = 0 andg(e, θ)<0 for all e >0.

(35)

Full information benchmark:

I Suppose information is verifiable.

I A contract consists of two wage-effort pairs: (wL,eL) and (wH,eH).

I The principal solves:

maxw,e λ(x(eH)−wH) + (1−λ)(x(eL)−wL) subject to the participation constraint:

λv(wH −g(eH, θH)) + (1−λ)v(wL−g(eL, θL))≥u.

(36)

Full information benchmark:

I Lagrangian:

L=λ(x(eH)−wH) + (1−λ)(x(eL)−wL)

+γ(λv(wH −g(eH, θH)) + (1−λ)v(wL−g(eL, θL))−u), whereγ ≥0 is the multiplier.

I Obtain FOCs.

(37)

Full information benchmark:

I From the FOCs:

v0(wH −g(eH, θH)) = 1

γ =v0(wL−g(eL, θL)).

I Strict risk aversion then implies

v(wH −g(eH, θH)) =v(wL−g(eL, θL)), so utility is fully equalized across states.

(38)

Full information benchmark:

I Show that optimal effort levels must be positive.

I Show that optimal effort levels satisfy:

x0(eH) =ge0(eH, θH) x0(eL) =ge0(eL, θL).

I Given optimal effort levels, wages are:

wH =g(eH, θH) +v−1(u) wL =g(eH, θL) +v−1(u).

(39)

Full information benchmark:

I Separation undelying optimal contract:

(40)

Unverifiable states:

I Insurance vs. incentives trade-off.

I The space of possible contracts is huge:

- The owner sets a performance-based wagew(x) together with some constraint on effort choices.

- The manager makes an annoucement ˆθ about the observed state and the owner sets a compensation w(ˆθ) without constraints on effort.

- The owner sets a wage w(e) depending on the level of effort the manager chooses.

I The optimal contract seems hard to find. However...

(41)

Revelation mechanisms:

I The manager annouces ˆθ after observing the true state θ.

I The contract specifies an outcome (w(ˆθ),e(ˆθ)) for each possible announcement.

I For every stateθ, the manager finds it optimal to report the true state.

The Revelation Principle:

I Every implementable outcome can be implemented through the revelation mechanism.

(42)

General program:

I The principal solves:

maxw,e E{x(e(θ))−w(θ)}

subject to the participation constraint

E{v(w(θ)−g(e(θ), θ))} ≥u and the incentive compatibility constraints

w(θ)−g(e(θ), θ)≥w(θ0)−g(e(θ0), θ).

(43)

Infinite risk-aversion:

I Consider the limit case in which the manager is infinitely risk-averse.

I Then, the participation constraint becomes:

minθ v(w(θ)−g(e(θ), θ))≥u.

I Equivalently,

w(θ)−g(e(θ), θ)≥v−1(u) for all θ∈ {θL, θH}.

(44)

Infinite risk-aversion:

I Any contract satisfying the constraints, must also satisfy w(θH)−g(e(θH), θH) ≥ w(θL)−g(e(θL), θH)

> w(θL)−g(e(θL), θL)

≥ v−1(u).

I So, we can ignore all but the lowest participation constraint.

I The lowest participation is binding (otherwise a contract with a slight constant reduction in wages would be feasible.)

(45)

Optimal contract:

I ˆeL ≤eL. Whenever eL >eL, the owner can get more profits from settingeL =eL.

I ˆeH =eH. Given eL ≤eL, tangency between the

θH-indifference curve and an isoprofit curve is necessary for optimality. And all such tangency points occur at eH =eH.

I ˆeL <eL. If not, a slight decrease in eL would allow a big increase ineH (the optimal level of eL is such that the marginal cost of reducing eL equals the marginal benefit of the highereH.)

(46)

Optimal contract:

I The optimal effort for θL satisfies the FOC:

[x0(ˆeL)−ge0(ˆeL, θL)] + λ

1−λ[ge0(ˆeL, θH)−ge0(ˆeL, θL)] = 0

I Differentiating and noting thatge0(e, θH)−ge0(e, θL) is negative and decreasing ine, we can get:

∂ˆeL

∂λ <0

I Moreover, when λ→0, we have ˆeL→eL and the solution approximates the first-best.

(47)

Price discrimination by a monopolist:

I Suppose consumers have difference preferences for a good.

I The monopolist can offer a menu of quantity-price (qi,pi) bundles (non-linear pricing).

I The utility of (qi,pi) for a consumer of type θ is given by U(qi,pi|θ) = v(qi, θ)−pi.

I The consumer gets 0 if she doesn’t buy.

I We assume that v satisfies properties analogous those of g.

(48)

Monopolist’s problem:

I The monopolist choosest(θ) and q(θ) to maximize E{t(θ)−cq(θ)}

subject to the participation constraint v(q(θ), θ)−t(θ)≥0 and the incentive compatibility constraint

v(q(θ), θ)−t(θ)≥v(q(θ0), θ)−t(θ0).

(49)

Two types:

I If θL and θH > θL, the binding constraints are:

v(qL, θL)−tL = 0 and

v(qH, θH)−tH =v(qL, θH)−tL.

I Relax and verify: ignoring non-binding constraints and solving this relaxed version of the original problem provides a solution.

(50)

Example:

I Supposev(q, θ) = 2θq1/2 and assumeλ∈(0, θLH).

I Then, ˆ qL =

θL−λθH (1−λ)c

2

ˆ qH =

λθH c

2

I Note that is possible that ˆqL>qˆH using the formulae above.

But in this case, the optimal contract is pools both types.

(51)

More than two types:

I Every IC contract is (weakly) monotonic.

I Only the participation constraint for the lowest type binds.

I Only downward local incentive compatibility constraints may bind in any IC contract.

I Hence, we can relax the problem by looking at monotonic contracts satisfying the relevant subset of constraints.

I Pooling some types might be optimal (since separation is costly).

Referências

Documentos relacionados

The integration of ServRobot in the DVA's system, as a Mobile Agent and Sensor Agent, could significantly improve DVA's performance, using it in the event's

Esse trabalho foi desenvolvido objetivando identificar a real importância que a problemática das ectoparasitoses têm, na visão de professores e gestores de duas escolas

(2015) Intracoronary Delivery of Human Mesenchymal/ Stromal Stem Cells: Insights from Coronary Microcirculation Invasive Assessment in a Swine Model.. This is an open access

In Argentina, there are periodic ILT outbreaks, especially in areas with high density of industrial poultry farms with poor management and biosecurity measures, in farms that

The present study concludes that Paederus dermatitis is more prevalent in rural than urban areas due to the strong affinity (positive association) of rove beetles with crops and

The probability of attending school four our group of interest in this region increased by 6.5 percentage points after the expansion of the Bolsa Família program in 2007 and

61 Figura 9: Captura de ecrã - Site do Movimento Natural Vibe 90 Figura 10: Captura de ecrã - E-mail Marketing Natural Vibe 91 Figura 11: Captura de ecrã - Perfil WhatsApp

The different agents are: user agent responsible for the user authentication and for presenting the user's information, the resource agent works as an access