Presentation by Sylvain Chassang "Corruption, Intimidation and Whistleblowing" at the SITE Corruption Conference, 31 August 2015.
Find more at: https://www.hhs.se/site
2. Motivation
Can we use reports from informed parties (monitor) to
target corrupt agents more efficiently?
Difficulty
In many environments, few competent parties able to
inform on corruption
Anonymous phone lines provide no real anonymity
Principal’s use of information becomes signal which lets
the agent discipline monitor
Take-away
Effective policy must garble the information content of the
monitor’s messages
How much and how?
3. The Model
Players and Actions
1. Single agent, corrupt or not, c ∈ {0, 1}
2. Single monitor, observes c, reports m = 1 (corrupt), m = 0
(not corrupt)
3. E.g. pair of cops, judge and clerk, boss and subordinate,
firm and accountant
4. Following message m, principal triggers intervention
i ∈ {0, 1} with probability σm ∈ [0, 1]; σ = (σ0, σ1) policy
dimension of interest
5. Agent retaliates at level r ≥ 0
Timing and commitment
1. Principal commits to intervention policy (σ0, σ1)
2. Agent commits to a retaliation policy r as a function of
observables (intervention + possible leaks)
4. The Model – Payoffs
Reduced-form consequences of intervention
uA = c × πA + i × vA(c) − kA(r)
uM = c × πM + i × vM(c, m) − r
uP = c × πP + i × vP
Assumption 1.
πA ≥ 0, vA ≤ 0
vM(c, m = c) ≥ vM(c, m = c)
πP < 0, vP < 0
Allows for malicious monitors
(benefit from intervention against honest agent)
5. The Model – Information
Leaks
conditional on intervention i = 1, outcome z ∼ f(z|m, c)
(i, z) observed by the agent
revelation principle does not hold
Arbitrary incomplete information
Agent doesn’t know monitor’s preferences vM, belief ΦA
Principal has incomplete information over types
(vM, πA, vA, kA, ΦA)
bounded support for payoffs
6. Exogenous vs Endogenous Information
Proposition 1 (basic trade-off).
(i) Assume messages exogenously informative, i.e. m(c) = c
If optimal policy = 0, then σ0 = 0 and σ1 > 0
(ii) Assume messages endogenous
∃λ > 0 st whenever σ1/σ0 ≥ λ
the agent is corrupt and commits to retaliate
the monitor sends message m = 0
Intuition
intervention has informative content
responsive intervention makes it easy for agent to
incentivize the monitor
should worry about silent corruption
7. Example: UK Accounting Authority
Financial Reporting Review Panel
Investigates records of public companies
Uses tips from competent informants
Likely whistleblower: auditing firm
Policy Change
1999–2004, purely reactive
2005–2010, proactive
12. Inference from Unverifiable Messages
monitor has type τM = vM
agent type τA = (πA, vA, kA, ΦA) (where ΦA belief over τM)
true distribution µT over types τ = (τM, τA) ∈ T
unknown to principal
Intervention and retaliation policies induce message profile
m : τM → ∆ ({0, 1})
13. Two Properties
Proposition 2.
Given corruption decision c, message profile m(·) constant
along ray {(σ0, σ1)|σ1 = λσ0}
Proposition 3.
(i) The set of profiles (σ0, σ1) such that a given agent is
corrupt is star-shaped around 0
(ii) Fix λ = σ1
σ0
. The mass of corrupt agents τA
cσ(τA)dµT (τA)
is decreasing in σ0
16. Inference – Single Policy
Proposition 4 (no inference).
Single policy experiment puts no restrictions on corruption
For any distribution of reports at a single policy σ, range of
consistent corruption rates is [0, 1].
Take policy σ and distribution µT yielding report
T
m∗
(σ, τ)dµT (τ)
We have that
TA
c∗(σ, τA)dµT (τA)
µT s.t T m∗(σ, τ)dµT (τ) = T m∗(σ, τ)dµT (τ) = [0, 1]
17. Inference – Two Policies
Benchmark policy (σB
0 , σB
1 ); New policy (σN
0 , σN
1 )
σO
0 < σN
0 and
σO
1
σO
0
=
σN
1
σN
0
Proposition 5 (bounds).
T
[1 − cN
(τ)]dµT (τ) ≥
T
mN
(τ)dµT (τ) −
T
mO
(τ)dµT (τ)
#honest agents at new policy ≥ drop in reported corruption
T
cO
(τ)dµT (τ) ≥
T
mN
(τ)dµT (τ) −
T
mO
(τ)dµT (τ)
#corrupt agents at old policy ≥ drop in reported corruption
18. Take-Away
In many environments no herd-anonymity, need to provide
anonymity through garbled response
Can be implemented through noisy surveys – related, but
quite distinct from randomized response surveys
Rule of thumb
first provide sufficient anonymity that people are willing to
complain
then scale up enforcement