Big question in this lecture

  Stockholm Doctoral Course Program in Economics Development Economics II: Lecture 4

  Civil Conflicts Masayuki Kudamatsu

  IIES, Stockholm University

20 November, 2013

  Big question in this lecture

  • Why do civil conflicts occur?
  • 56% of countries experienced civil conflicts (25+ deaths per annum)
  • 20% of countries had 10+ years of

  civil conflicts

  • Source: Armed Conflict Database

  1960-2006 (Fig. 1 in Blattman and Miguel 2010)

  • Conflicts: detrimental to welfare
  • Convincing evidence hard to come by,

  though ⇐

  Pre-conflict micro data hardly available Why do civil conflicts occur? (cont.)

  Literature (mainly) proposes the following three reasons: Economic factors (sec.1)

  • Asymmetric information (sec.2)
  • Commitment problem (sec.3)
  • Also ethnic diversity is often emphasized as a source of conflict

1. Economic factors

  Poverty often cited as a cause of

  • civil wars Is this true theoretically
  • Use a contest model to illustrate
  • Is this true empirically?
  • Search for exogeneous variation
  • ⇐ Conflict may cause poverty

1.1 Model

  • Collective action problem: ignored

  Players: groups A & B

  • Endowment: 1 unit of labor for both
  • A & B Actions: both groups
  • simultaneously allocate labor between production and fighting

1.1 Model (cont.)

  Technologies Production: group i produces

  • w (1 − l )

  i i i l : labor allocated to fighting

  • w : labor productivity for group i i

  Fighting: probability of group A

  • winning given by l

  A

  • l l A B

1.1 Model (cont.)

  • Resources unaffected by labor •

    allocation (e.g. natural resource

    export revenue, foreign aid, etc.)
  • Main results: robust to including labor

  Victory payoff: V

  output in victory payoff (see Garfinkel & Skaperdas 2006)

  • A B

  If no fighting (l = l = 0), V is

  shared by the two groups according to some exogenous formula

  e.g. A gets sV and B gets (1 − s)V The contest model has been used in other contexts: Rent-seeking (Tullock 1980)

  • Endogenous property rights
  • (Skaperdas 1992)

1.2 Analysis

  = l = 0) cannot be an Peace (l A B equilibrium

  

⇐ If l = 0, l = ε > 0 makes group i

j i win for sure.

1.2 Analysis (cont.)

  • Group A solves max

  l A

  l A l

  • l

  A

  B

  V + (1 − l A )w A

  • Group B solves max

  l B

  l

  B

  l A + l B V + (1 − l B )w B

1.2 Analysis (cont.)

  • A

  FOC for l :

  l

  B

  V = w A

  2

  (l + l )

  A B

  • B

  FOC for l :

  l

  A

  V = w B

  2 � � l (l A + l B ) 1 l

1.2 Analysis (cont.)

  V 2 Deleting yields

  • (l +l ) A B

  l w

  

A B

  = l w

  

B A

  Less productive group spends more

  • time in fighting

  ⇐ Lower opportunity cost of labor

1.2 Analysis (cont.)

  w

  B ∗

  l =

  V A

  2

  (w A + w B ) w

  A ∗

  l =

  V B

  2

  (w A + w B )

  V ∗ ∗

  Conflict level (l + l = ) ↑ if

  A B (w +w ) A B

  V ↑ (resource curse)

  • (w + w ) ↓ (poverty as a cause of
  • A B

1.3 Evidence

  Tons of cross-country regressions using GDP per capita as a proxy for

  • poverty natural resource export as a proxy
  • for V Endogeneity is a major concern

  Conflict reduces GDP and resource

Miguel et al. (2004)

  • Rainfall as instruments for GDP per capita in African countries
  • GDP in Africa depends on rain-fed agriculture
  • Find GDP ↓ ⇒ civil conflicts ↑
  • But the mechanism is not clear

  e.g. GDP ↓ ⇒ Govt forces weaker

  • Also it’s LATE
  • Income ↓ due to other reasons may
Also the resource curse channel is

  not tested Recently 3 papers tackle both

  • channels

i. Besley & Persson (2009): use export

  commodity price as resource effect &

import price as poverty effect

ii. Besley & Persson (2011): use UN

  Security Council membership as resource effect & natural disaster as poverty effect

(will be discussed in Political Recently 3 papers tackle both

  • channels

iii. Dube & Vargas (2013): look at

  subnational-level evidence from Colombia

1.4 Dube & Vargas (2013)

  • Use 978 municipality panel data on conflict incidents from Colombia (1988-2005)

  (based on newspapers & reports by Catholic priests)

  • where civil wars ongoing since 1960s
  • 3-way conflict (govt, guerilla, &

  paramilitary), but govt & paramilitary collude

  • Columbia exports oil & coffee
  • Oil price: proxy for V
y = λ(OP ∗ Oil ) + ρ(CP ∗ Cof ) jrt t j t j

  �

  ρ + α + β

  

j t

  • x jt
  • δ t + γ(Coca ∗ t) + ε

  r j jrt

  • jrt : # of guerilla attacks,

  y

  paramilitary attacks, clashes, or y jrt = λ( OP t ∗ Oil j ) + ρ(CP t ∗ Cof j )

  �

  ρ + α j + β t

  • x jt
  • δ + γ(Coca ∗ t) + ε

  r t j jrt

  • t

  OP : int’l price of oil in year t

  • j

  Oil : oil production level in y jrt = λ(OP t ∗ Oil j ) + ρ( CP t ∗ Cof j )

  �

  ρ + α j + β t

  • x jt
  • δ + γ(Coca ∗ t) + ε

  r t j jrt

  • t

  CP : domestic coffee price in year t

  • j

  Cof : hectares of land devoted to

  = λ(OP ∗ Oil ) + ρ(CP ∗ Cof ) y jrt t j t j

  �

  ρ + α β + +

  

j t

x jt

  • δ t + γ(Coca ∗ t) + ε

  r j jrt

  • jt

  : vector of controls, incl. log

  x

  population y jrt = λ(OP t ∗ Oil j ) + ρ(CP t ∗ Cof j )

  �

  ρ + α j + β t

  • x jt

  δ + + γ( ∗ t ) + ε r t Coca j jrt

  • r

  δ t: region-specific linear trends

  • j

  Coca ∗ t: linear trends for

1.4 Dube & Vargas (2013) (cont.)

  λ ∗ Oil ρ ∗ Cof

  y jrt = (OP t j ) + (CP t j )

  �

  ρ + α + β

  j t

  • x

  jt

  • δ t + γ(Coca ∗ t) + ε

  r j jrt

  Theoretical predictions λ > 0: resource curse effect

jrt

  ∗ t) + ε

  � jt

  j

  t + γ(Coca

  r

  t

  j

  ρ + α

  )

  y jrt = λ (OP

  j

  ∗ Cof

  t

  ) + ρ (CP

  j

  ∗ Oil

  t

  • β
  • x
  • δ
    • Now OLS estimation of this equation would lead to bias in

1.4 Dube & Vargas (2013) (cont.)

  • t

  No need to instrument OP

  Int’l price is used • Columbia produces <1% of world oil

  • production
  • j

  No need to instrument Oil

  

measured at the 1st year of the

  • sample

  ⇒ No endogenous change in response to conflict

1.4 Dube & Vargas (2013) (cont.)

  Instrument for CP t Int’l coffee price: not suitable

  • Columbia is a major coffee exporter
  • Log foreign coffee exports from 3
  • largest producers (Vietnam, Brazil, Indonesia), denoted by FE

  t Is this really exogeneous? •

1.4 Dube & Vargas (2013) (cont.)

  • j

  Need to instrument Cof

  Measured in 1997, when CP peaked • t (i) Usually low-production municipalities may produce coffee a lot in 1997 ⇒ Non-classical measurement error

  (ii) Municipalities w/ downward pre-1997 trend in ε mt e.g. Govt effort in security may have started coffee production by C 1997, because P � t ⇒ Spurious corr. btw CP & y for t jrt

1.4 Dube & Vargas (2013) (cont.)

  Instrument for Cof :

  j

  rainfall and temperature at j

  ⇐ Coffee grows well if m

T < 26 (degrees Celcius)

  • m

  R ≥ 1800 (mm)

1.4 Dube & Vargas (2013) (cont.)

  ⇒ ∗ Cof

  CP t j is instrumented by: t

  • t j

  FE

  ∗ R FE

  • FE ∗ T • t j t j j FE ∗ R ∗ T
  • R : mean annual rainfall in municipality j j T : mean annual temperature in j

  municipality j First-stage (page 20)

  • Kleibergen-Paap F statistic: 15.94
  • Exceeds Stock-Yogo critical value
Digression: Detecting weak instruments

  • Stock & Yogo (2005) provide critical values of F-statistic or Cragg-Donald statistic (if # of endogenous variables > 1) for
  • Bias in 2SLS is > 10% of bias in OLS (available only for # of instruments ≥ 3)
  • Actual size of 5% test being less than 10%
  • • (Baum, Schaffer,

  Stata ivreg2 authors

  Stillman 2007: 489-491) take the i.i.d.

  assumption seriously, and if you cluster standard errors, it reports Kleibergen-Paap statistics Not clear if Stock & Yogo (2005)

  • critical values are relevant in this case
Findings (Tables II, III): ρ < 0 for all types of conflict ˆ

  • ˆ

  λ > 0 for paramilitary attacks only

  • ρ > 0 if y ˆ is agricultural wage
  • jrt

  

⇒ Evidence for opportunity cost

mechanism

  ˆ λ > 0 if y is municipality govt

  • jrt

  revenue

  ⇒ Evidence for resource mechanism

1.5 Recent studies on

  opportunity cost mechanism

  Jia (2012): Sweet potato mitigated

  • the impact of droughts on peasant revolts in China (1470-1900) Bueno de Mesquita (2013):
  • Guerrilla & terrorism should have an inverse U-shaped relation with opportunity cost

1.6 Limitation of the contest

  model

  Fighting is always an equilibrium

  • Does not distinguish arming from
  • fighting

  ∗ ∗ ∗

  • A A B

  /(l By setting s = l + l ), both

  parties should prefer peace ⇒ Need for other explanations for

2. Asymmetric information

  War breaks out if you overestimate your strength

  • you underestimate the opponent’s
  • strength

  Below we illustrate the latter with a simple model of bargaining

2.1 Model (Fearon 1995)

  Players: groups A & B

  • V : Pie to share or fight over
  • Actions:
  • A proposes V − x as transfer to B
  • B then decides whether to accept
  • V − x or go to war

  p: Probability of group A winning if

  • war breaks out
  • A B : Cost of war for A & B,

  , c c

2.1 Model (cont.)

  {u , u }: Payoffs A B

  If B accepts A’s offer, {x, V − x}

  • If B fights, {pV − c , (1 − p)V − c }
  • A B

  With complete info., peace achieved by x = pV + c (which makes B indifferent)

  B

2.2 Analysis

  • B : only

  Suppose A does not know c

  knows its distribution F (c B ) B accepts peace if

  • − x ≥ (1 − p)V − c

  V B ⇐⇒ c ≥ x − pV

  B

  A solves

  • max [1 − F (x − pV )]x

  x FOC (assuming interior solution) 1 − F (K )

  = K + c

  A

  f (K ) where K ≡ x − pV RHS: increases w/ K

  • LHS: decreases w/ K if F yields the
  • monotone increasing hazard rate.

  ⇒ Unique K satisfying FOC

2.3 Limitation

  Information story may explain the

  • onset of a war But both parties should learn about
  • each other over time

  ⇒ cannot explain a long-lasting war Which is typically the case for civil • wars, e.g. Colombia, Sudan, Angola, DR Congo

3. Commitment

  In the previous model, peace is

  • achieved w/ complete info. This, however, assumes group A’s
  • ability to commit to x If A’s winning probability goes up
  • >tomorrow, it will renege on the promise B then decides to fight t

3.1 Model (Powell 2006)

  • 2-period model
  • Players: groups A & B
  • V : Pie to share or fight over in each period
  • If war breaks out in period t:
  • A wins w/ prob. p t
  • Fraction (1 − d) of V : destroyed forever
  • The loser eliminated forever (payoff 0)
Each period t

  1. A proposes a transfer schedule

  2

  {s } to B

  τ τ =t

  

2. B decides whether to accept or to

  go to war In period 1, A cannot commit to s

  • 2

  < p Suppose p

  1 2 : A is temporarily weak In period 2: A’s payoff from war: p (1 − d)V

  • 2

  ⇒ Credible s must be at most

  2

  (1 − p

  2 (1 − d))V

  In period 1: A’s maximum credible concession

  • to avoid war:

  = V s

  1 B’s payoff from war

  • (1 − p )(1 − d)(1 + δ)V

  1

⇒ War cannot be avoided in period 1 if

  (1 − p )(1 − d)(1 + δ)V

  1

  > [1 + δ(1 − p (1 − d))]V

  2

3.3 Intuition

  Key: large shifts in future

  • distribution of power (p

  2 >> p 1 )

  Party whose power will increase

  • cannot commit to a large transfer in future

  

They’d rather fight otherwise

  • Opponent then prefers a
  • preemptive attack, even if it’s costly (d > 0)

3.4 Evidence

  Yet to come

  • How to measure future shifts in relative power?

3.5 Implications on political

  institutions Acemoglu & Robinson (2000, 2001, 2006): •

model democratization as commitment

device to future transfer Bruckner & Ciccone (2011): rainfall shock • led to democratization in Africa (1980-2004)

  • Chaney (2013): when the Nile flooded,

    religious leaders less likely to be replaced

4. Other explanations of war

  Leader bias (Jackson and Morelli

  • 2007)

  Jones & Olken (2009) for evidence

  • that leaders matter

  Grievances (key to solve collective

  • action problem?) Multiple threat points (Ray 2009,
  • Morelli and Rohner 2010)

5. Ethnic diversity

  Viewed as the leading source of

  • civil conflict (& govt failure) Is this empirically true? (sec. 5.1 -
  • 5.3) What’s the theoretical rationale?
  • >(sec. 5.4-5.5) Why ethnicity, rather than c

5.1 Measurement

  Ethnic diversity often measured by

  • the prob. that 2 randomly chosen individuals in a country belong to different ethnic groups
  • � � �

    N π i (1 − π i ) = π i π j

      i i j

    =1 �=i

    i π : pop. share of ethnic group i

    This measure has become popular

    • among economists because of its presumed exogeneity

      Political scientists do not agree, • though. (e.g. Posner 2004)

    Michalopoulos (2012) provides

    • evidence against exogeneity

      Found to be negatively correlated

    • w/ govt & economic performances

      cf. Easterly & Levine (1997), La Porta et

    al. (1999), Alesina et al. (2003)

    5.1 Measurement (cont.)

      But no robust association w/ civil

    • conflicts

      cf. Collier & Hoeffler (1998), Fearon and Laitin (2003)

      But is this the correct measure of

    • ethnic diversity to predict conflict?

      This measure: highest if many ethnic

    • groups of equal size Conflict: most likely if two groups of
    • equal size (ie. polarization)

    5.2 Polarization

      Esteban & Ray (1994) derive an index of polarization from two ideas:

      

    1. Identity (how many people you are

      identified with) ↑ ⇒ Conflict ↑

      

    2. Alienation (how far other people are

      from you) ↑ ⇒ Conflict ↑

    5.2 Polarization (cont.)

    • Polarization index:

    • α i

      K

       n i =1

    n

    j

      

    =1

      π

      1

      π

      j

      ν

      ij

      : pop. share of group i

    • π i
    • ν ij
    • α ∈ (0, α
    • n: # of groups
    • K > 0: constant for normalizing index

      : distance between groups i & j

      ](α ≈ 1.6)

      π i π

      j π k

      ν ij ik π

      i > π j

      = π

      k

      π i π

      j π k

      ν ij ik π

      i > π k v ij > v ik

      − v

      ij

      π i π

      

    j

    π k

      ν ij ik ν

      ij

      = ν

      ik

      2

      (2005)

      Adopt this polarization index into

    • ethnic conflict by setting

      ν = 1, ∀i, j (distance between any • ij

    two ethnic groups: same)

    α = 1 (perhaps for simplicity) • K = 4 (⇒1 is maximum: ie. • N = 2, π = π = 1/2) 1 2

      ⇒

    N

      2

      

    5.3 Montalvo & Reynal-Querol

    (2005) (cont.)

      In pooled sample of country by

    • 5-year-period (1960-1999): Ethnic polarization ↑ ⇒ Prob. of civil conflicts ↑

    5.4 Esteban & Ray (2011)

      Provide a theory that links

    • fractionalization and polarization indices to conflict Theory also tells when which
    • measure is more important to predict conflict The key is whether the prize from
    • winning is public or private goods.

    5.4.1 Model: Players

      N agents

      m groups of agents

    • Size of group i: N ( ⇒ N = N)
    • i i i

      Denote group i’s population share

    • by n ≡ N /N

      i i

    5.4.1 Model: Actions

      Agent k of group i expends effort

    • r (k ) to fight over a budget of size N

      i

    • i � �� ���

      (r (k )) with Effort is costly: c

      c > 0, c > 0, c > 0 ���

      c > 0 ensures the uniqueness of the

    • equilibrium Group i wins w. prob.
    • r i (k ) R

      i k ∈i � � ≡ R is our measure of conflict

    • intensity

    5.4.1 Model: Public good

      Winning group spends fraction λ

    • (fixed) of the budget to produce the public good they prefer Group i’s payoff from public good is
    • λu if they win ii λu if group j �= i wins ij
    • ij ii ij : “distance” between i

      δ ≡ u − u

      and j

    5.4.1 Model: Private goods

      Fraction 1 − λ of the budget will be

    • shared equally by winning group members to produce private goods Group i member’s payoff is:
    • (1 − λ)N/N if group i wins i 0 otherwise

      

      Each individual’s payoff is therefore:

    • 1 − λ

      π (k ) = p + i i p j λu ij −c i (r i (k )) n i

      j

      We assume player k ∈ i maximizes:

    • π (k ) + α π (l)

      i i l

    ∈i;l�=k

      = (1 − α)π i (k ) + α π i (l)

    • α is an extent of altruism to other fellow members of the same group
    • α can also be the bargaining power of the group leader who maximizes

      l ∈i

      π

      i

      (l)

    • In a political economy model, govt

      maximizes a weighted sum of

      individual utilities (cf. Persson and Tabellini (2000)) with weights different by political institutions etc.

    5.4.2 Analysis

      max

      r i (k )

      (1 − α)π i (k ) + α

      l ∈i

      π i (l) = σ i p i 1 − λ n i

    • λ

      j

      p j u ij − c(r

      i

      (k )) − α

      l ∈i;l�=k

      c (r

      i

      (l))

      • λ

      p

      l ∈i;l�=k c (r i

      (k )) − α

      i

      − c(r

      ij

      u

      j

      p

      j

      i

      1 − λ n

      i

      i

      Ignore the last term as it doesn’t depend on r

      (l) = σ

      i

      π

      l ∈i

      (k ) + α

      i

      (1 − α)π

      r i (k )

      (l) are given max

      i

      (k ) and r

      i

      (l))

      1 − λ n i

    • λ
      • Rewrite this term by using

      i

      c (r

      l ∈i;l�=k

      (k )) − α

      i

      − c(r

      j p j u ij

      i p i

      max

      (l) = σ

      i

      π

      l ∈i

      (k ) + α

      i

      (1 − α)π

      r i (k )

      (l))

      � � 1 − λ σ + λ

      i p i p j u ij

      n

      i � � � � j

      1 − λ 1 − λ

    • = σ p λu − + λp u

      i j ij i ii

      n n

      i i j � � � � �=i 1 − λ

      1 − λ

    • = σ p − λδ − λu +

      i j ij ii n n i i j �=i So agent k of group i’s maximization problem becomes: � � � �� 1 − λ max σ p − λδ −

      

    i j ij

    r (k ) i n i j �=i

      − c(r (k ))

      i

      Now we have ∂p j R j p j

      = − = −

      2 Therefore, the FOC is σ i R

      j �=i

      p j ∆

      ij

      = c

      �

      (

      r i (k ) )

      where ∆ ij ≡ λδ ij +

      1 −λ n i

      for j �= i

    • LHS is the same for all k ∈ i

      ⇒ r i (k ) = r i , ∀k ∈ i

    5.4.3 Linking conflict to

      

    population distribution indices

      σ

      i �

      p ∆ = c (r )

      j ij i

      R

      j �=i

      Now we transform this FOC under the assumption of p i = n i , to derive per capita conflict intensity ρ ≡ R/N as a function of Gini,

      � σ

      i �

      p j ∆ ij = c (r i ) R

      j �=i

      If we assume p = n , ∀i,

      i i

      R R R R R R

      i i

      r i = = · = p i · = n i = ≡ ρ N i R N i N i N i N so we have

      σ i

      �

      σ

      i

      �

      = ρc

      ij

      ∆

      

    j

      n

      i

      n

      j �=i

      σ i N

      over all i,

      i

      i and sum

      ρn

      (ρ) Now multiply both sides by

      �

      = c

      

    ij

      ∆

      j

      n

      j �=i

      R

      (ρ)

      � � σ

      i �

      n n ∆ = ρc (ρ)

      i j ij

      N

      i j �=i

      RHS: Monotonically increases with

    • per capita conflict intensity ρ LHS: Linear combination of Gini,
    • Polarization, and Fractionalization
    Substituting σ (≡ (1 − α) + αN ) and

      i i

      1 −λ

      ∆ + ij (≡ λδ ij ) into LHS yields: � � n i

      n i n j [ (1 − α) + αN i ][ λδ ij + (1 − λ)/n i ] N i j �=i � �

    • i j ij i j �=i

      Gini n n δ

      Multiplied by (1 − α)λ/N

    • (1 − λ)/n

      � i j �=i n i n j

      [(1 − α) + αN

      i

      ][ λδ

      ij

      i

      ]

      N

    • Polarization:
    • Multiplied by αλ

      i

    j �=i

      n

      2 i

      n

      j

      δ

      ij

      � i j �=i n i n j

      [(1 − α) + αN

    • (1 − λ)/n

      i

      ][λδ

      ij

      i

      ]

      N

      n i n j

    • Fractionalization:

      i j �=i

    • Multiplied by α(1 − λ)
    In summary, conflict increases with (1 − α)λ

    • α[λP + (1 − λ)F ] G N If α > 0 (altruism to other members
      • of the same group), P & F matter P matters more, the higher λ (the
      • prize is more public)

      

    5.5 Esteban, Mayoral, & Ray

    (2012)

      Estimate this equation with

    • cross-country data To measure δ
    • ij , use linguistic

      distances on language trees

      cf. Desmet et al. (2012) To measure λ by country, use

    • PUB ∗ GDP

      PUB ∗ GDP + OIL

      PUB: degree of un-democraticness of

    • government OIL: per capita value of oil reserves
    • >GDP: per capita GDP
    To measure α by country, use World

    • Value Surveys in which respondents answer to the questions on:

      Adherence to social norms

    • Identification to local community

    • >Importance of helping others
    Then estimate

      P F

      ρ =β α λ + β α (1 − λ )F

      c c c P c c c c G �

    • β λ c (1 − α c )G c /N c κ + ε c
    • x c

      P F

      Results: β & β positive and

    • significant Coefficients on P
    • c & F c : becomes

      insignificant once these interaction

      For the onset of conflict,

      polarization does correlate, but fractionalization does not robustly (Table 6)

      But Esteban and Ray (2011) use the

    • contest model, which has no

      predictive power on the initiation of conflict

      Why do people tend to start fighting

    • along the ethnic divisions, instead

    5.6 Esteban & Ray (2008)

      Why is ethnic conflict more

    • prevalent than class conflict?

      Many conflicts these days are ethnic

    • in nature

      Show higher income inequality

    • increases the likelihood of ethnic conflict Go beyond 2-player conflict m
    each group

      Model: Players 1. ph: Poor ethnic majority 2. rh: Rich ethnic majority 3. pm: Poor ethnic minority 4. rm: Rich ethnic minority e.g

    h: Hindu, m: Muslim in India

    • No collective action problem within
    • n ij : Population share of class i of ethnicity j

      ≡ n ph + n

    pm

    • n p
    • n r

      ≡ n rh + n

    rm

      ≡ n ph + n

    rh

    > n m ≡ n pm + n rm

    • n h
    • Per-capita income
    • Rich y r
    • Poor y p (< y

      r

      

    )

      rh

      /n

      h

      = n

      rm

      /n

      

    m

    ⇒ Same per-capita income for each

    • n
    Model: Public goods

    • C: class budget
    • used for funding class public good
    • health care
    • education
    • infrastructure
    • E: ethnic budget
    • used for funding ethnic public good
    • religious festivals
    • temples
    • job reservations in govt.
    Model: Group ij’s Payoff

      If peace is achieved,

    • u (y i ) + s i C + s j E

      i ∈ {r , p}, j ∈ {h, m}

    • s
    • i ∈ [0, 1]: class i’s share of class

      budget in peace time ∈ [0, 1]: ethnicity j’s share of s

    • j
    Model: Group ij’s Payoff (cont.)

      If class alliances form,

    • w i A i A i

      (y − ) + + s u i C j E n A + A

      i p r

    • i

      A : # of activists financed by class i

    • i : compensation for each activist

      w

      in class alliance i Total compensation shared equally

    Model: Group ij’s Payoff (cont.)

      If ethnic alliances form

    • w A A

      j ij j

    • u (y − ) + s C E

      i i

      n A + A

      ij h m

    • ij : # of activists financed by class i

      A

      in ethnic alliance j ≡ A + A

    • j ij (−i)j

      A

    • j

      w : compensation for each activist

      Notice: in ethnic alliances,

    • compensation is shared equally within each class of an ethnic group

      ⇐ Otherwise, forming an ethnic

      alliance involves regressive redistribution, which is unlikely Utility cost of being an activist: fully

    • compensated by w i or w j ⇒ Doesn’t show up in the payoff

    1. Players form alliances.

      Randomly chosen player: either

    • proposes (i) class alliance, (ii) ethnic alliance, or (iii) peace If an alliance proposed, the other
    • player in the proposed alliance decides whether to accept or reject

      e.g. If rh proposes class (ethnic) alliance,

    then rm (ph) responds

      If accepted, move to stage 2 (ie. • conflict). If rejected or peace proposed, a new

    Model: Timing of events (cont.)

    1. Players form alliances (cont.)

      If all the 4 players propose peace or if

    • proposals rejected endlessly, peace payoffs realize
    • Players reject or never make a

      Assumption D:

    • proposal yielding the worst possible payoff

      ie. Delaying such a proposal so the worst payoff is discounted

      Players accept a proposal yielding the

    Model: Timing of events (cont.) 2a. If class alliances are formed,

      each alliance simultaneously chooses A (k = {p, r })

      k ⇐ There’s no asymmetry between h & m in terms of payoffs.

      2b. If ethnic alliances are formed,

      each class in each alliance simultaneously chooses A ik (i = {p, r }, k = {h, m}) Analysis: how to proceed

      

    1. Prove that, under Propositions 2-5,

      ethnic conflict is unique outcome of the game (Proposition 6)

      

    2. Check if a higher income inequality

      makes propositions 2-5 more likely to hold

      The paper proceeds in the reverse

    • order, but this way motivates us to care about propositions 2-5 better.

      

    Preference conditions for ethnic

    conflict to be unique outcome

      P2. ph: ethnic � class P3. If ph: class � peace

      ⇒ rh: ethnic � class

      P4. r : peace � class P5. p: class � peace If P5 does not hold, peace is also an

    • equilibrium, but class conflict is NOT.
    Peace won’t be an eq. outcome

      Suppose otherwise

    • ph prefer proposing class
    • � peace By P5, ph: class
    • pm accepts this proposal
    • ph: ethnic � peace (by P2, P5) ⇒ σ > s ⇒ σ < s

      h h m m

      ⇒ pm: peace � ethnic ⇒ pm’s best outcome: class (by P5) ⇒ Accepts immediately (by D)

      Suppose otherwise

    • r never initiate class conflict
    • rh: ethnic � class (by P5 & P3)
    • ⇒ rh’s worst outcome: class (by P4) ⇒ rh never proposes/accepts class (by D)
    • ph: ethnic � class (by P2)

      ph prefers proposing ethnic

    • Peace won’t be an eq. outcome

      rh accepts this proposal

      Poor ethnic majority: want ethnic

    • conflict (P2) Issue: whether rich ethnic majority
    • accept this even when rich prefer peace the most When poor can credibly threaten
    • rich with class conflict (1st part of the proof), peace no longer possible
    Propositions 2-5

      Paper shows P2-P5 hold under

    • reasonable set of parameter values Here we focus on why
    • within-ethnicity inequality helps satisfying P2 Proposition 1a: specifies condition
    • >for conflict � peace Proposition 1b: specifies condi
    • Assume that contributions x are small relative to income y i .

      ie. u (y i ) − u(y i − x) ≈ u (y i − x)x

    • Then class i in alliance k prefers conflict to peace iff

      λ

      σ

      2 k

    • (1 − λ

      ik

      ik

      )σ

      k

      > s

      k

    • λ ik

      ≡ A ik /A k Proposition 1a: Intuition

    • ik

      λ = 1 if k is class alliance

    • 2

      Then the condition boils down to

      σ > s

      k k

      Conflict should increase the budget

    • share sufficiently to compensate resources spent (A ik )
    Proposition 1a: Intuition (cont.)

    • ik = 0 if class i in ethnic alliance k

      λ

      does not contribute Then the condition boils down to

    • σ > s

      k k

      Conflict should increase the budget

    • share
    • Let G ∈ {C, E}

      Proof of Proposition 1a

    • Conflict � ik Peace if

      σ

      k

      − s

      k

      G > u

      

      y

      i

      − w k A ik n

      ik

      w k A ik n

      ik

    • LHS: Gain from conflict
    • RHS: Cost of conflict
    • � � � � ik Peace if

        � Conflict

        w A w A

        k ik k ik

        σ k − s k G > u y i − n ik n ik By FOC on A

      • � �

        ik , we have

        w k A ik w k A l

        �

        u y − = G

        i

        2

        n n (A + A )

        ik ik k l

      • ik = λ ik A k with both sides

        Multiply A

        of FOC and replace cost of conflict

      • � �

        � Conflict

        ik Peace if

        A A

        k l

        σ − s G > λ G

        k k ik

        2

        (A k + A l ) = λ σ (1 − σ )G

        ik k k

      • 2

        Rearranging this inequality yields:

        λ σ + (1 − λ )σ > s

        ik ik k k k Notice: this proof only uses the

      • inequality that compares the payoffs from conflict and peace

        ⇒ Difference between both sides of

        inequality: Net payoff of conflict This allows us to derive Proposition

      • 1b:
      Proposition 1b

        Class i of ethnicity j prefers ethnic

      • alliance to class alliance iff

        2

        2

        [λ + (1 − λ )n − s ]µ > σ − s

        ij n ij j j i j i

        where µ ≡ E/C LHS: Gain from ethnic conflict

      • relative to peace RHS: Gain from class conf

        Class i prefers class conflict (so

      • λ ik = 1) to peace iff

        

      2

        σ > s i

        

      i

        Class i of ethnicity j prefers ethnic

      • conflict to peace iff

        2

        λ n + (1 − λ )n > s

        ij ij j j j

        λ = 0 as an implication of pk within-ethnicity inequality

      • ik = 0 if class i in ethnic alliance k

        λ

        does not finance any activists: � � w A w A

        

      k ik k l

        u y − > E , ∀A ≥ 0

        i ik

        n ik n ik A k + A l i

        This is more likely if y is very small

      • ⇒ If income inequality very high

        (y >> y ), then r p

      • 2

        Condition for ethnic � class,

        2

        [λ n + (1 − λ )n − s ]µ > σ − s ,

        ij ij j j i j i