1

Economic Shocks and Battle Deaths in Civil Wars

THORSTEN JANUS† and DANIEL RIERA-CRICHTON††

†Department of Economics, Dept. 3985, 1000 E. University Ave.

Laramie, WY 82071, United States. (email: tjanus@uwyo.edu)

††Department of Economics, Bates College, Pettengill Hall, Room 273

Lewiston, ME 04240, United States. (email: drieracr@bates.edu)

Abstract

This paper estimates the effects of exogenous income shocks - in the form of commodity terms

of trade (CTOT) shocks - on battle deaths in civil wars. We show that CTOT growth generally

decreases conflict, but that in fuel exporting economies with intermediate ethnic fractionalization

levels, dominant, or polarized ethnic groups, both negative and positive shocks increase conflict.

The positive effects come from fossil fuel windfall in fuel-exporters. The evidence is consistent

with opportunity cost and rent-seeking motives for conflict as well as the resource-curse

literature

JEL Classification: D74, O11, O17

Keywords: civil war, conflict, ethnicity, ethnic diversity, commodity prices, terms of trade

Wordcount: 12,073 +250 (Figure 1) = 12,323

2

I. Introduction

During the course of civil wars, the death toll can vary immensely. For example, the dataset we

present below shows that the beginning and end phases of the El Salvadoran, Guatemalan, and

Nicaraguan civil wars generated less than a thousand fatalities per year. However, the death toll

in the intervening years sometimes exceeded 10,000.* In this paper, we estimate the effects of

exogenous income shocks in the form of commodity terms of trade (CTOT) shocks – the change

in national commodity export relative to import prices - on the annual deaths tolls in civil wars.

We argue that countries with intermediate ethnic fractionalization, a dominant ethnic group, and

a high level of ethnic polarization are more likely to respond adversely to commodity-income

shocks and that the net effect of the shocks depends on the balance between the rent-seeking and

peaceful economic opportunities they generate. We show that CTOT growth generally decreases

conflict, but that in fuel exporting economies with intermediate ethnic fractionalization levels,

dominant, or polarized ethnic groups, both negative and positive shocks increase conflict.

Further, the positive effects come from fossil fuel windfall in fuel-exporters. The evidence is

consistent with opportunity cost and rent-seeking motives for conflict as well as the resource-

curse literature. We estimate that a standard deviation increase in the growth rate of our 3-year

moving average CTOT index decreases the death toll per conflict year by 34-41%. However, a

standard deviation increase in the fossil-fuel-specific CTOT increases the death toll in fuel

exporters with adverse ethnic compositions by about 32%. The evidence is, therefore, consistent

with opportunity cost and rent-seeking motives for conflict as well as the resource-curse

literature.

* Bosnia and Herzegovina’s civil war from 1992-95 killed, respectively, 17,000, 23,000, 11,000, and 1,300 individuals. Appendix A depicts the time-series for the four countries.

3

In order to estimate these effects, we address three identification challenges. First,

incomes can be endogenous to conflict. This is why we study the effects of CTOT as opposed to,

for example, per-capita-GDP growth. The focus on commodity prices allows us to isolate the

most volatile and plausibly-exogenous component of the general terms of trade. The

international-trade and macroeconomics generally predicts that terms of trade growth increases

national income growth in term of goods, that is, real GDP (Dixit and Norman 1980, Obstfeld

and Rogoff 1996, Agenor and Montiel 2008, Krugman et al. 2014, Feenstra 2015).

Second, wartime economies may respond differently than peacetime economies to

economic shocks. For example, the decline of the rule of law may change the production

structure away from contract, transport, and capital intensive sectors, such as formal

manufacturing, to informal, labor-intensive, and easier-to-protect sectors, such as the production

of primary resources for exports, mobile services, trade networks that procedure goods from

abroad, or non-seasonal agricultural crops that cannot be expropriated as easily (Mayshar et al.

2015). It can allow warlords, criminals, and even elements in the government and rebel armies to

pursue illegal activities, such as extortion, kidnapping, smuggling, trafficking, foreign-aid

appropriation, and illegal natural resource extraction (Keen 2000, Rubin 2000, Le Billon 2001,

Bannon and Collier 2003, Dube and Vargas 2013, Nunn and Qian 2014), and to create self-

governed (as well an ungoverned) regions, like in Afghanistan, Colombia, Somalia, and Syria in

recent decades. The fact that there can be many spillovers from the conflict and both the

government and the rebels can target civilians (Berman et al. 2011, Valentino et al. 2004) forces

households to protect themselves and their assets. The fact that most of these economic and

institutional changes only occur after the conflict begins suggests that the conflict can change the

economy’s structural equations and, therefore, its response to economic shocks. In order to

4

address this concern, we purposely only estimate the death toll variation within the conflict years

and omit the peace years. Appendix B presents formal evidence that - in contrast to our finding

that positive fossil-fuel CTOT shocks predict more battle deaths in fuel exporters – they are

negatively (but insignificantly) related to the risk that civil wars begins in these countries in the

first place.

Finally, different income shocks could have different effects in different environments

(Dal Bó and Dal Bó 2011, Dube and Vargas 2013, Bazzi and Blattman 2014, Nunn and Qian

2014, Janus and Riera-Crichton 2015). Although we cannot test all possible non-linear

hypotheses, the literature suggests that ethnic dominance and polarization can increase conflict

(Horowitz 1985, Collier and Hoeeffler 2004, Esteban and Ray 2011). Moreover, the fact that

commodity windfalls can encourage rent-seeking (Tornell and Lane 1999, Ross 2012, Dube and

Vargas 2013) suggests that terms of trade growth should not necessarily decrease conflict. Even

if it increases the opportunity cost conflict inputs, perhaps by creating jobs, it can increase the

return to rent-seeking and the parties’ ability to finance the war effort (Bannon and Collier 2003).

We, particularly, hypothesize that the rent-seeking and conflict-finance effects are larger relative

to the opportunity cost effect after positive relative to negative terms of trade shock. The paper

shows that the data strongly supports this asymmetric-effects idea for economies with adverse

ethnic compositions.

The paper belongs to the literature that relates income shocks to civil war. We contribute

to this literature in three ways. First, we estimate the intensity of ongoing civil wars as opposed

to the onset and time-series incidence of the wars, which have often been studied (Hegre and

Sambanis 2006, Blattman and Miguel 2010). Second, in contrast to most previous studies that

relate internationally-determined commodity prices to conflict (Brückner and Ciccone 2010,

5

Dube and Vargas 2013, Aguirre 2016) we estimate the effects of export prices in terms of import

goods rather than currency, that is, the effects of commodity terms of trade rather than

commodity export price shocks. Janus and Riera-Crichton (2015, 2017) show that that CTOT

declines predict a higher onset risk for civil wars; regressing the onsets separately on the log-

export and log-import price changes in the numerator and denominator of the CTOT-index yields

similar absolute coefficients; and failing to control for the change in import prices can changes

the export price estimates, which could, therefore, suffer omitted variables bias.

Third, we provide evidence that positive and negative income shocks can both increase

conflict. The fact the positive effects come from fossil fuel windfalls in fuel-exporter economies

with adverse ethnic compositions bridges the empirical conflict literature with the literature on

the natural resource curse (Smith 2004, Fearon 2005, Basedau and Lay 2009, Ross 2012) by

suggesting that fuel booms in fuel-dependent economies may only increase conflict in countries

with adverse ethnic compositions. A simple interpretation is that both fuel (mainly, oil)

dependence and certain ethnic compositions (having a dominant ethnic group or polarized

groups) encourage rent-seeking. The combination of the two during a boom in easily-

appropriable fossil fuel revenues crates a powerful rent-seeking effect.

In the remainder of the paper, Section 2 explains why we believe that commodity terms

of trade shocks and ethnic compositions can influence conflict. Section 3 presents our data and

econometric model. Section 4 presents the main results. In Section 5, we estimate the effects of

positive relative to negative and fossil fuel relative to other terms of trade shocks. Section 6

concludes the paper.

6

II. Theoretical background

In this section, we detail the reasons we choose to relate terms of trade shocks and ethnicity to

conflict in the paper We also explain why positive and negative shocks could have different

conflict effects.

(1) The effects of terms of trade shocks

Although several studies have estimated the effects of commodity export price changes on

conflict (Brückner and Ciccone 2010, Dube and Vargas 2013, Bazzi and Blattman 2014, Aguirre

2016), the theory of international economics only relates changes in export prices relative to

import prices, that is, terms of trade changes, to national income (Matsuyama 1988, Easterly et

al. 1993, Turnovsky 1993, Mendoza 1995, Rodrik 1999, Agenor and Montiel 1999, Krugman et

al. 2014). For example, in a macroeconomic model a representative export good and a

representative import good, but, for simplicity, without a non-tradable good, real GDP in

international currency is Y C I G NX= + + + , where ( )x mNX p X p M= − in the standard

notation. If initially 20x mp X p M= = , a 10% decrease in export prices and a 10% increase in

import prices both decrease the trade balance and GDP by 2% on impact. A 10% price drop for

both export and import prices leave GDP unchanged; if the import prices fell 15%, income

increases on impact despite the decline in export prices. If export and import price changes have

these types of opposing, but symmetric income effects, then, regressions relating linking

economic outcomes to export prices also could suffer omitted variable bias (Janus and Riera-

Crichton 2015, 2017). Lederman and Porto (2016) discuss the impact of commodity prices on

household welfare based on survey data from African and Latin America as well as a literature

review. The conclude that households spend large budget fractions on commodities, often

7

depend on commodities to earn income, and international price changes pass through to

households. Thus, households should be directly exposed to CTOT shocks.

(2) The effects of ethnic dominance and polarization

Although the empirical conflict literature has not established a linear effect of ethnic diversity or

fractionalization – that moving from a single to many small ethnic groups monotonically

increases the conflict risk (Fearon and Laitin 2003, Blattman and Miguel 2010) - there is

evidence that countries with either (a) a single large ethnic group that lives together with smaller

groups or (b) at least two significant ethnic groups may be conflict-prone. Our reading of the

literature suggests that the problem, frequently, is that one of the large ethnic groups can

dominate the political system and choose policies that, directly or indirectly, expropriate the

excluded groups. On the other hand, the groups that are currently excluded or fear they will be

excluded in the future can secede, rebel, and conduct military coups (Gellner 1983, Horowitz

1985, Smith 1986, Posen 1993, Gurr and Harff 1994, Collier and Hoeffler 2004, Ross 2005,

Østby 2008, Fearon and Laitin 2011, McGarry and O’Leary 2013, Weiner 2015). In Chad and

Sudan, the Arab population has historically dominated the smaller groups. The government’s

neglect of the countries’ peripheries has encouraged conflict. In Indonesia and Russia, the

traditional dominance of the Javanese and Russians have encouraged the ethnic minorities in

Aceh, East Timor, West Papua, and Chechnya to secede. Sri Lanka’s Tamil Tigers fought to

secede from the Sinhala-dominated central government. In Burundi, Iraq, and Syria, the Tutsi,

Sunni Muslim, and Alawite minorities (or powerful elements within these groups) conducted

military coups in the 1960s-1970. In these cases, the problem seems to be that a large group lives

8

together with smaller groups. Following Collier and Hoeffler (2004), we call this situation ethnic

dominance.

Another potentially adverse ethnic configuration is ethnic polarization (Esteban and Ray

1994, 1999, 2011). The polarization concept relates the degree of polarity in the size distribution

of social groups and the social distance between individuals, to the level of conflict between

them. When the social distance is one between members of different groups and zero between

same-group members, as most empirical studies assume (Montalvo and Reynal-Querol2005,

Estbeban et al. 2012), then, the Esteban and Ray (1994) polarization index is maximized when

there are two equally large groups. More generally, societies with at least two significant ethnic

groups should experience more conflict. This could, potentially, explain the ethnic conflict

histories of Afghanistan, Angola, Bosnia, Croatia, Guatemala, Iraq, Israel, Lebanon, and Sri

Lanka. Montalvo and Reynal-Querol (2005) and Estbeban et al. (2012) show that the

polarization index predicts conflict incidence across countries.

(3) The asymmetric effects of positive and negative income shock

Surprisingly, the evidence suggests that not only negative (Miguel et al. 2004, Miguel 2005,

Brückner and Ciccone 2010, Bazzi and Blattman 2014, Blattman and Miguel 2010), but positive

income shocks as well can increase conflict. In the 1990s, US food aid may have increased

warlord conflict in Somalia as the warlords fought to control the scarce commodity (Dowden

2009, Albright 2013, Nunn and Qian 2014). Angrist and Kugler (2008) link increases in coca

production to violence in Colombia. During Sri Lanka’s civil war, the Tamil Tigers used

remittances to purchase military equipment; Angola’s UNITA rebels and Sierra’s Leone’s

Revolutionary United Front may have relied on diamonds. Timber and minerals may have

9

financed the wars in Cambodia and the Eastern Congo in the 1990s and 2000s (Le Billon 2001,

Bannon and Collier 2003, Ross 2003, Janus 2012, Dube and Vargas 2013, Global Witness 2015).

In Iraq after 2003, at least two of the main rebel groups, - al-Qa’ida in Iraq and the Mahdi Army

- relied on extortion, theft, and black market sales (Bahney et al. 2010, Stanford Mapping

Militant Organizations Project 2015†). After al-Qa’ida in Iraq renamed itself the Islamic State of

Iraq and Syria (ISIS) and entered the Syrian civil war in 2013, it took control of oil fields that by

September 2014 may have earned it $1-2 million per day (Byman 2015).‡ Dube and Vargas

(2013) link price gains for capital-intensive goods, such as oil, to violence in Colombia.

Our reading of this literature suggests that income growth has ambiguous conflict effects.

On one hand, it can increase the opportunity cost of conflict inputs like labour (Hirshleifer 1999,

Miguel et al. 2004, Chassang and Padró-i-Miquel 2009). We call this the opportunity cost effect.

On the other hand, there is a conflict-finance effect: marginal income can be used to procure

more conflict inputs Bannon and Collier (Rustad and Binningsbø 2012, Berman et al. 2011).

Additionally, there is a rent-seeking effect: as income increases, so does the pool of contestable

wealth and the return to fighting (Bannon and Collier 2003, Parker and Vadheim 2017).§

Whether the rent-seeking and conflict-finance effects can dominate the opportunity cost effect is

an empirical question. Although we lack the data we would need to quantify and compare the

different effects, however, we think it is reasonable to hypothesize that the rent-seeking and

† See http://web.stanford.edu/group/mappingmilitants/cgi-bin/, accessed November 28, 2015 ‡ This estimate comes from an expert assessment cited in the New York Times, 16 September, 2014: “How ISIS Works.” Oil revenues may be important for the Iraqi Kurds currently fighting ISIS: “Strapped for cash and increasingly frustrated with Baghdad’s stingy disbursement of the federal budget… the Kurdistan Regional Government, which governs the Kurdish region in northern Iraq, has been ramping up independent oil sales. The KRG says it needs the oil revenue because it is weighed down by the costs of fighting Islamic State militants.” (Washington Post, August 16, 2015, “Iraq oil feud renewed as cash-strapped Kurds turn backs on deal with Baghdad.”). § During recessions, we should expect the same effects in reverse. Additionally, economic downturns can, potentially, increase psychological distress and cause individuals to look for “scapegoats” and violent social identities that can restore their sense of belonging and empowerment (Miguel 2005, Cramer 201, Zivin et al. 2011).

10

conflict-finance effects may be larger relative to the opportunity cost effects after positive

relative to negative terms of trade shocks. If this is true, the coefficient estimate on positive terms

of trade shocks in the conflict regressions should exceed the coefficient on negative shocks,

which is easily testable.

Although we postpone a formal justification, there are several ways to motivate the idea

that positive and negative income shocks can have different conflict effects. For example, the

desire to avoid starvation and bankruptcy thresholds after negative shocks can encourage risk-

taking and make the opportunity cost of fighting close to zero the failure to fight could push the

individual below the threshold and cause a large utility drop with certainty. In a behavioral

framework as well, loss-aversion can encourage risk-taking (Thaler and Johnson 1990, Tversky

and Kahneman 1991, Jervis 1992, Thaler et al. 1997). Related, economic downturns could lead

to psychological distress and cause individuals to look for “scapegoats” and violent social

identities (Miguel 2005, Cramer 201, Zivin et al. 2011). This effect should keep the subjective

return to violence high during recessions even if the monetary return decreases. Alternatively,

positive CTOT shocks could generate windfall earnings that come are psychologically salient

(Chowdury et al. 2014) and encourage “irrationally exuberant” or other myopic behaviors

(Khwaja and Mian 2011, Shiller 2015). This should increase the subjective return to rent-seeking

during commodity-income booms. Finally, we believe it is possible that political elites

appropriate the income gains from commodities in boom times but share the losses with society

during busts. Particularly, we find below that it is only positive fossil fuel (mainly, oil) generated

income shocks in fuel exporting countries that increase conflict. The literature suggests that oil-

dependence increases corruption and weakens state capacity (Fearon 2005, Besley and Persson

2010, Ross 2012). This suggests that oil booms increase not only the maximum rent pool

11

political elites can appropriate but the actual rents the appropriate. When the terms of trade

decline, however, the corruption level, which partly depends on the theft of fuel revenues from

the public coffers, can only fall to the zero lower bound, unless the leaders liquidate their

accumulated wealth (Acemoglu and Robinson 2001)

III. Data and Estimation

In this section, we, first, explain the data sources and variable definitions we use. We, then,

outline the estimation procedure. Tables 1 and 2 display the summary statistics for the data.

[Table 1 Goes Here]

[Table 2 Goes Here]

Armed Conflict: We use the conflict data for internal and internationalized internal armed

conflicts for the 1946-2008 period provided by the Uppsala Conflict Data Program and the Peace

Research Institute of Oslo (UCDP/PRIO). Focusing on internal wars allows us to exclude

interstate conflicts and extra-systemic conflicts, which involve a state fighting a non-state group

abroad. Since, in these cases, one of the conflict actors is based abroad and may be another

government, the commodity price shocks we study may have different effects than in internal

conflicts. The definitions of armed conflict and internal armed conflict are as follows (Gleditsch

et al. (2002) and Themnér & Wallensteen (2011), Codebook for the UCDP/PRIO Armed

Conflict Dataset, Version 4, 1, 9; Lacina and Gleditsch (2005), Battle Deaths Dataset, Codebook

for Version 3, 2009, 2)

12

‘[An armed conflict is] a contested incompatibility that concerns government or territory or both where the

use of armed force between two parties results in at least 25 battle-related deaths. Of these two parties, at

least one is the government of a state.”

“Internal armed conflict occurs between the government of a state and one or more internal opposition

group(s) without intervention from other states…Internationalized internal armed conflict occurs between

the government of a state and one or more internal opposition group(s) with intervention from other states

(secondary parties) on one or both sides.’

Battle-related Fatalities: The battle-related fatalities data comes from Lacina and Gleditsch

(2005) and includes 1957 observations of battle-related fatalities from 1946-2008. We use the

version of the dataset that is compatible with the conflict dataset we described above. 1717 of the

battle death observations in this dataset are linked to internal or internationalized internal armed

conflicts rather than interstate and extra-systemic conflict. The definition of battle-related

fatalities is (Lacina and Gleditsch (2005), Battle Deaths Dataset, Codebook for Version 3, 2009,

2)

‘…those deaths caused by the warring parties that can be directly related to combat over the contested

incompatibility. This includes traditional battlefield fighting, guerrilla activities (e.g. hit-and-run

attacks/ambushes) and all kinds of bombardments of military bases, cities and villages etc. Urban warfare

(bombs, explosions, and assassinations) does not resemble what happens on a battlefield, but such deaths

are considered to be battle-related. The target for the attacks is either the military forces or representatives

for the parties, though there is often substantial collateral damage in the form of civilians being killed in the

crossfire, indiscriminate bombings, etc. All fatalities – military as well as civilian – incurred in such

situations are counted as battle-related deaths.’

13

Due to the difficulty of establishing the exact number of battle-related fatalities per year, Lacina

and Gleditsch (2005) provide a “low” and a “high” estimate for all the observations as well as a

“best” estimate for about 70% of the observations. Since they provide the data at a country-year-

conflict level, we add the low, high and best estimates for every country and year to compute

country and year specific low, high and best estimates. Table 1 shows that the low estimates

range from 10 to 50,000 with a mean of 1,478. The high estimates range from 25 to 250,000 with

a mean of 7,319. The best estimates average 4,061 with a standard deviation of 9,132.

Our empirical battle deaths measure is either the “best” country-year estimate, provided it

exists, or, since it does not exist for 30-40% of the observations, an “imputed best” estimate. This

imputed best estimate is the sum across the ongoing conflicts within a country year of either the

best conflict-specific estimate or, if it does not exist for that conflict, the average of the low and

high estimates for the conflict. This methodology follows Bazzi and Blattman (2014). If we

alternatively dropped the country-year observations that are based on missing-best estimates, we

could be dropping a non-random sample of observations and get selection bias. For example,

countries that have multiple ongoing conflicts –making missing data more likely ceteris paribus

- may have many ethnic groups and poor data collection. Moreover, as we show below, our

qualitative results are robust to using either the “low” or the “high,” instead of the “best” and

“imputed best” estimates, and to focusing on the observations for which Lacina and Gleditsch

(2005) use year-specific sources, which automatically excludes most of the imputed

observations.

14

Ethnicity: In order to study the effects of ethnic dominance and polarization, we first observe

that they are conceptually distinct. The ethnic dominance concept identifies countries where

there is a single large ethnic group that lives together with smaller group(s) that it can potentially

dominate. The polarization concept identifies countries where two or more large groups vie for

dominance. Unfortunately, Janus and Riera-Crichton (2015) show that it is difficult to

distinguish countries with dominant ethnic groups (defined as when the largest groups represents

about 50-85% of the population), countries with a higher-than-median Esteban and Ray (1994)

polarization index, and countries with an intermediate (25th-75th percentile) Herfindahl-

Hirschman ethnic fractionalization index. This is the case using the popular Fearon (2003) cross-

country ethnicity dataset as well as the equally-popular Alesina et al. (2003) data. In order to

illustrate the problem, Figure 2 copies Janus and Riera-Crichton’s (2015) Figure 1. Janus and

Riera-Crichton (2015, 25) explain that:

‘A linear regression [of ethnic fractionalization on the population share of the largest ethnic group, which is

used to define the ethnic dominance measure] yields an R2 of 0.96. On the other hand, there is also a close

quadratic relationship between ethnic fractionalization and polarization. Regressing polarization on

fractionalization and its square yields an R2 of 0.92. …regressing polarization on the ethnic plurality and its

square yields a similarly high R2 of 0.92.’

The fact that high polarization levels are correlated with ethnic dominance status and

intermediate fractionalization levels makes it difficult to separate the empirical effects. In order

to avoid favoring either the ethnic dominance or the polarization idea – given that we lack the

cross-country ethnicity variation we need to distinguish their effects convincingly - we prefer to

(a) mainly distinguish between countries in-and outside the 25th-75th Herfindahl-Hirschman

15

ethnic fractionalization percentiles, but (b) show that distinguishing the countries based on

whether they have dominant or polarized ethnic groups gives similar results. Specifically, we

import the following three indicator variables from Janus and Riera-Crichton (2015): (1) The

intermediate-ethnic fractionalization indicator is a dummy that equals one when the Herfindahl-

Hirschman ethnic fractionalization index is in 25th-75th percentiles in Fearon’s (2003) ethnicity

dataset. For brevity, we call these the intermediately-fractionalized/diverse (ID) and non-

intermediately-fractionalized/diverse (NID) countries. (2) The ethnic dominance indicator is a

dummy that equals one when the largest ethnic group represents 50-85% of the population. (3)

The ethnic polarization indicator is a dummy that equals one when the polarization index in

Esteban and Ray (1994) - under the parameter assumption in Montalvo and Reynal-Querol

(2005) and Esteban et al. (2012) that the social distance between individuals from different

ethnic groups is one and the within-group distance between individuals is zero - exceeds the

sample median.

As Janus and Riera-Crichton (2015) explain, the Fearon (2003) ethnicity dataset relies on

seven criteria, including, particularly, common ancestry and a sense of community and self-

consciousness as a group, to define a prototypical ethnic group. The paper identifies 822 ethnic

groups in 160 countries after consulting the CIA World Factbook, Encyclopedia Britannica,

Library of Congress Country Studies, as well as country-specific sources. We doubt that the

ethnicity measures is endogenous to conflict intensity. First, the ethnicity data is not time-

varying and social identities are unlikely to change much over the relatively short time horizon

we study. Second, even if political elites can manipulate the ethnic boundaries, they are likely to

be constrained by preexisting social categories (Horowitz 1985, Smith 1986, Chandra 2007,

Eifert et al. 2010). Third, as we show below, our results are robust to using the race-based

16

ethnicity data from Alesina et al. (2003). Finally, there is abundant evidence that ethnicity affects

developing-country politics (Bates 1981, Horowitz 1985, Chandra 2007). Kramon and Posner

(2012), Franck and Rainer (2012), and Burgess et al. (2015) link ethnic favoritism to education,

infant mortality and road construction (De Luca et al. 2015, Francois et al. 2015).

Commodity Terms of Trade: The dataset for commodity terms of trade (CTOT) covers the

period from 1970-2008. The CTOT index was originally developed by Ricci et al. (2008) and

Spatafora and Tytell (2009) and is defined as

( / ) / ( / )i ij jX M

jt it t it ti i

CTOT P MUV P MUV= Π Π (1)

where is the CTOT index for country in year ; is a common price index for each

of six commodity categories (food, fuels, agricultural raw materials, metals, gold, and

beverages); is the average share of exports of commodity in GDP from 1970 to 2006;

is the corresponding average share of imports; and the commodity prices are deflated by a

manufacturing unit value index (MUV). The fact that and are averaged over the sample

year ensures that the CTOT index is invariant to changes in export and import volumes in

response to conflict outcomes, thus isolating the effect of commodity price fluctuations. If we

compute the change in the log CTOT we can get the approximate CTOT growth rate per year

( 1)

( 1) 1 ( 1) 1

( / ) ( / )ln ln ln ln

( / ) ( / )

i ij j

i ij j

X M

it t it ti i

jt j t X M

i t t it ti i

P MUV P MUVCTOT CTOT

P MUV P MUV−

− − − −

Π Π − = − Π Π

(2)

jtCTOT j tjtP

i

jX i i

jM

i

jX i

jM

17

Following Brückner and Ciccone (2010), Bazzi and Blattman (2014), and Janus and Riera-

Crichton (2015), we note that the annual commodity price shocks may be serially correlated and

have lagged conflict effects. In order to address this concern, we include the growth rate of the

three-year moving average of the terms of trade index,

1

2 3

ln / 3 ln / 3t t

jt js js

s t s t

CTOT CTOT CTOT−

= − = −

∆ = −∑ ∑ (3)

Since the growth rate of the three-year moving average CTOT index approximately equals the

average annual growth rate over the three-year period (see the appendix to Janus and Riera-

Crichton 2015), we can interpret every % (0.01) increase in the index as a mean increase of 1%

per year over three years. The standard deviation of CTOT∆ is 0.012. In order to formally test

whether it is appropriate to include the shock to the moving average in the regressions - instead

of include the three component annual CTOT shocks - we first include the annual shocks and

test whether the coefficients differ statistically. As we show below, we find no clear evidence

that this is the case. Nor is there a clear pattern suggesting that one of the terms is a better battle-

deaths predictor. Due to the fact that it is difficult to include the separate annual shocks together

with their interactions with our ethnic measures - it gives us six terms to interpret an potential

multi-collinearity problems - we prefer to only include the growth in moving-average, CTOT∆ ..

Estimation: The estimation regresses the logarithm of the number of battle-related fatalities on

the 3-year-movig-average CTOT growth rate in equation (3). in a linear specification with

18

country and year fixed effects, country-specific time trends, and robust standard errors which we

cluster at the country level to control for serial correlation. Following Bazzi and Blattman

(2014), we also control for the duration of the conflict and add a dummy for the first conflict-

year. The regressions take the form

jt jt jt j jt jt j t j jtb CTOT CTOT ID d f z tα β γ µ ρ ε= + ∆ + ∆ × + + + + + + , (4)

where is the natural logarithm of the number of battle-related fatalities in country in year ,

jtCTOT∆ is the growth rate of the three-year moving-average CTOT index, jID is the time-

invariant dummy for the ID countries, andj t

d andjt

f represent the duration of the conflict since

the onset year and a dummy for the first conflict year (Bazzi and Blattman 2014) Finally, j

µ

and tz are the country and year fixed effects, jtρ is the country-specific time-trend, and

j tε is

the error term.**

IV. Results

Table 3, Column (1) presents the results of regressing the natural logarithm of the annual battle-

related death toll on the annual CTOT shocks for periods t, t-1, and t-2 as well as their

interactions with the ID dummy. The CTOT coefficients are insignificantly different and their

sum is significant and negative. Thus, terms of trade growth appears to decrease the death toll in

** Appendix C explains why we estimate a linear fixed-effects probability model instead of an interval regression model to account for the fact that some observations lack a point estimate for battle deaths (Bazzi and Blattman 2014). The reasons are that the interval estimator makes it difficult to include country fixed effects and non-normally-distributed errors can bias the estimates. Appendix C also explains why we believe that the country-level may be the best unit of analysis for civil war outcomes, despite the fact that there are several recent studies of subnational violence.

jtb j t

19

the NID economies. In the ID economies, however, the sum of the six CTOT terms (including

the interactions) is insignificant. Thus, we find no evidence that CTOT growth monotonically

decreases the death toll in ID countries.

In column (2), we use the growth rate of the three-year moving average CTOT, CTOT∆ ,

instead of the annual growth rates. Again, CTOT growth has negative effects in NID countries

but insignificant effects in ID countries. Column (3) replaces the ID dummy based on the

Fearon’s (2003) ethnicity data with the ID similarly-defined ID dummy based on the Alesina et

al.’s (2003) ethnicity dataset. Column (4) adds a lagged dependent variable to control for

persistence in the number of battle-related fatalities (Bazzi and Blattman 2014); to correct the

dynamic panel bias, we use the random-effects procedure developed in Hausman and Taylor

(1981) and Amemiya and MaCurdy (1986).†† In column (5), we restrict the sample to the

observations that are at least three years into the conflict in order to ensure that we do not

confound the effects of CTOT∆ shocks pre-and post-conflict initiation. In column (6), we again

follow Bazzi and Blattman (2014) and restrict attention to observations where Lacina and

Gleditsch (2005) observe year-specific deaths and, therefore, do not need to provide interpolated

or other noisy estimates (Lacina 2009, 5). Column (7) shows the results of imposing the two

restrictions jointly. Column (8) replaces the country fixed effects and country-specific time

trends with conflict-episode-fixed effects and a quadratic conflict-episode-specific time trend.

The idea is that either (i) the mean conflict-episode-specific death toll could vary within

countries or (ii) the conflict-specific death tolls have hump-shaped time-trends, as Figure 1 might

†† The Hausman-Taylor estimation declares the dynamic regressor endogenous. It is infeasible to use the alternative

General Methods of Moments (GMM) Arellano and Bond (1991, 1998) estimator because the number of time

periods in our panel does not exceed the number of panel units.

20

suggest, even after controlling for other factors. Through all of these robustness checks, the

results remain similar to the Column (2) results. Tables A2-A3 in Appendix D show that the

results also remain similar if we (a) replace the intermediate-diversity dummy with eight

alternative indicators for intermediate ethnic diversity, ethnic polarization, and ethnic dominance

(Table A1) and (b) control for the interaction between the CTOT shock and a range of

geographic, historical, and other, non-ethnic diversity measures (Table A2).

[Table 3 Goes Here]

In Table 4, we test whether negative terms of trade shocks have different effects than

positive shocks. In order to do so, we define a positive CTOT shock

{ ,0}jt jtCTOT Max CTOT+∆ = ∆ and a negative shock { ,0}jt jtCTOT Min CTOT−∆ = ∆ . The effect of a

positive shock is the coefficient on jtCTOT +∆ .. The effect of a negative shock, in contrast, is

minus the coefficient on jtCTOT −∆ . The full-sample results in Column (1) show that positive and

negative CTOT shocks have asymmetric effects: rather strikingly, both positive and negative

terms of trade shocks predict more an increased death toll.

In Columns (2)-(3), we try to explain the results in Column (1) by dividing the sample

into ID and NID countries. The results show that the positive full-sample effect of the positive

shocks come entirely from the ID countries. In other words, positive commodity-generated

income shocks only appear to increase the death toll in intermediately diverse countries. The

estimates for the NID countries in Column (3), in contrast, imply that positive and negative

CTOT shocks have symmetric effects: larger positive and less negative (closer-to-zero) shocks

21

both decrease the death toll. Since we cannot reject that the positive and negative shock

coefficients are the same in these countries, Column (4) reports the estimate with the original,

non-decomposed terms of trade shock. Although the estimate is negative, however, it is

insignificant at conventional levels (p=0.16).

In order to examine the robustness of these results, we inspected the partial residual plots

for the CTOT shocks and re-estimated the regressions after eliminating one country at a time.

For the ID countries, the magnitude of the positive shock estimate - the coefficient on

Pos∆ΤΟΤ(t) in Table 4 – is somewhat sensitive to removing a few of the oil-exporting ID

countries. In Tables 8-9 and Section 5 below, we show that the fact that the oil exporting ID

countries affect the results for the ID countries as a whole is not a coincidence. In fact, the fact

that positive terms of trade shocks increase the death toll in the ID countries as whole only

reflects that they increase he death toll in the fossil fuel (mainly, oil) exporters or about a quarter

of the ID countries. We explain these findings in more detail in Section 5.

The inspection of the partial residual plots and the re-estimation after eliminating one

country at a time in the case of the NID countries, on the other hand, shows that the results

remain stable unless when we omit Indonesia..‡‡ Excluding Indonesia nearly doubles the

coefficient estimate from -15.6 to -28.6 and it makes the t-statistic highly significant. In order to

better understand the Indonesia effect, Appendix E examines the case-study evidence for

Indonesia. The evidence suggests that a coalition of Indonesia’s largest two ethnic groups - the

Javanese and the Sundanese, which both live on Java - tend to dominate the central government

and that the government has historically fought three ethnic minorities that live on peripheral

‡‡ Dropping any of the 38 countries apart from Indonesia keeps the ∆TOT(t) coefficient in [-18.6, -12.2], closely

centered around the original -15.6.

22

islands; namely the Acehnese, East Timorese, and West Papuans. We believe that the conflict

problem in Indonesia is a good example of what an ethic dominance problem can look like in

practice. Although the single largest group in Indonesia, the Javanese, only represents 45% of the

population, Ross (2005) notes that several observers count the Javanese together with the 15%

Sundanese as a single group. If we followed that procedure (unlike Fearon (2003)), the

aggregated group would meet the 50%-population-share threshold we use to define ethnic

dominance an it would fall in the 25th-75th ethnic fractionalization percentile that defines the ID

countries. Due to Indonesia’s large effect on the NID estimate and its potentially-borderline

ethnicity classification, Table 4, Column (5) presents the NID estimates without Indonesia as

well. The -28.6 estimate implies that a standard deviation increase in the terms of trade shock

decreases the annual death toll in NID countries by 34%. This is comparable to the effect of

negative shocks in ID countries, which is about 41%. Nonetheless, we acknowledge that the

decision to exclude Indonesia from the NID sample is ultimately a judgment call.

[Table 4 Goes Here]

Figure 2 shows that the results remain similar when we divide the sample into countries

with low and high ethnic polarization and countries with and without a dominant ethnic group (a

group representing 50-85% of the population). Table 5 confirms that this is the case. In Tables 6-

7, we show that the results with alternative dependent variables look similar as well. In Table 6,

columns (1)-(3) re-estimate the Table 4, Column (2) regression but replace the dependent

variable with, respectively, the lowest estimates for the annual battle-related death tolls implied

by the Lacina and Gleditsch (2005) dataset; the corresponding highest annual estimates; and an

23

ordinal measure, which we coded to be equal to one when the best or imputed best estimate is at

most 1000 and two when the best or imputed best estimate exceeds 1000. In Table 7, similarly,

we replicate the Table 4, Columns (4)-(5) regressions for the NID countries with and without the

inclusion of Indonesia. The results again remain similar.

[Table 5 Goes Here]

[Table 6 Goes Here]

[Table 7 Goes Here]

V. The effects of fossil fuel dependence and fossil fuel terms of trade shocks on conflict

The results so far suggest that – consistent with the idea that income growth can increase

rent-seeking and help to finance civil wars (Angrist and Kugler 2008; Dube and Vargas 2013;

Aguirre 2016) - positive income shocks can increase conflict. The positive effects are, however,

restricted to intermediately ethnically fractionalized countries and countries with dominant and

polarized ethnic groups. In this section, ask whether the positive conflict effects in these

countries come from particular commodity groups and country types.

In order to address this question, we note that several studies have found that price

increases for capital-intensive natural resource sectors, such as the oil sector, increase conflict

(Dube and Vargas 2013; Aguirre 2016). The idea is that higher prices of capital-intensive goods

increase the return to rent-seeking effort relative to the opportunity cost because the conflict

sector is relatively labor intensive (Dal Bó and Dal Bó 2011). Additionally, fossil fuel resources

24

are often geographically concentrated, so ethnic groups that live in resource-rich areas may fight

to increase their autonomy and revenue shares (Le Billon 2001; Ross 2004). Alternatively, the

government can invade the regions preemptively (Ross 2005). The conflicts between the Iraqi

government and the Iraqi Kurds, Sudan’s Second Civil War, and Indonesia’s ethnic-secessionist

conflicts, for instance, pitted ethnic groups that were associated with the central government

against peripheral groups with access to oil and natural gas. Both the capital-intensity

geographic-concentration hypotheses suggest price increases for fossil fuels can increase

conflict.

Our second idea comes from the resource-curse literature that relates natural-resource

dependence and, particularly, oil dependence to economic development. Oil dependence can,

potentially, encourage the growth of undemocratic, corrupt, and repressive “rentier states,”

where the political elite uses resource income, such as royalties, to finance high living standards.

Further, since the state is relatively independent on tax collections, it may have little incentive to

invest in economic development and state capacity through improving the legal system, the

quality of the bureaucracy, the ability to collect income taxes, and so forth. (Smith 2004; Fearon

2005; Basedau and Lay 2009). The most extreme example may be Equatorial Guinea, whose

2015 PPP-based GDP per capita of $30,000 was about the same as Portugal’s, but whose Human

Development Index – a broader development measure that is tracked by the United Nations

Development Program, and which responds to health and education as well as income – looks

like Zambia’s. Once the repressive status apparatus fails and civil war begins in such countries,

the lack of social cohesion and institutional capacity could encourage a lot of rent-seeking

violence on average and, particularly, increases in rent-seeking after positive terms of trade

shocks.

25

On the basis of these two ideas, we estimate whether fossil-fuel-generated terms of trade

shocks have different effects than non-fuel terms of trade shocks and whether fuel and non-fuel

shocks have different effects in fuel-dependent economies. Thus, we decompose the change in

the log three-year-moving average terms of trade index in equation (3) into the change in the

fossil fuel component – the fossil fuel category in the CTOT index includes coke, coal, and

briquettes; petroleum and petroleum products; and gas (natural and manufactured) – and the

change in the remaining, non-fossil fuel component. Additionally, we define fuel-dependent

economies as economies whose average export share of fuels in GDP from 1970-2006 - which is

the year range we used to construct the CTOT-index weights - exceeds its import share.

Moreover, the average export share of fuel has to exceed 2% of GDP. The last requirement

excludes Afghanistan and Argentina, which are rarely considered fuel-dependent economies.

In Table 8, Column (1) presents the estimates for positive and negative CTOT shocks

generally (not yet disaggregated by commodity category) in ID fuel exporters, NID fuel

exporters, ID non-fuel exporters, and NID non-fuel exporters. The estimates imply that positive

CTOT shocks increase conflict intensity in ID as well and NID fuel exporters. Although it is

possible to interpret these estimates as suggesting that fuel dependence alone and not a country’s

ethnic composition that creates a positive relationship between conflict and terms of trade gains,

we do not believe this is the best interpretation. The reason is that 91% (89/98) of the

observations for NID fuel-exporters in the regression sample come from just four countries –

Angola, Azerbaijan, Indonesia, and Sudan - that have a history of ethnic conflict involving large

or dominant ethnic groups. Moreover, although their ethnic fractionalization levels are outside

the 25th-75th percentiles we use to define the ID economies, they are in the 15th-85th percentiles.

In Appendix F, we study the individual conflicts that generated these 89 observations and

26

conclude that they were ethnic conflicts that were, at least in part, either motivated or financed

by natural resources. On this basis, we believe that

(a) the evidence suggests that positive CTOT shocks increase the number of battle-related

fatalities in intermediately diverse fuel exporters; but that

(b) our study lacks enough civil war observations to allow us to estimate the effects in

highly ethnically homogenous and fractionalized fuel exporters - given that we only have 9

observations for countries below the 15th or above the 85th ethnic-fractionalization percentiles.§§

In Column (2) we show that, if we widen the ID definition from the countries in the 25th-

75th to the countries in the15th-85th ethnic fractionalization percentile, the coefficient on the

positive CTOT shocks in NID fuel exporters turns negative. Since the coefficient is based on

very few observations, however, it is hard to interpret. More importantly, the effect in the ID

economies is almost unchanged.

[Table 8 Goes Here]

In Table 9 we restrict attention to the intermediately diverse fuel exporter sample but

decompose the CTOT shocks sand ask whether the positive fossil fuel shocks specifically, as

opposed to the positive non-fossil fuel shocks and the negative shocks cause the conflict

§§ The 9 observations include two for highly diverse Cameroon (1984) and Nigeria (2004), and seven for relatively

homogenous Egypt (1993-98) and Tunisia (1980). The fact that we have so few observations for these countries,

however, suggests that they may be unlikely to experience civil war in the first place (Fearon and Laitin 2003;

Collier and Hoeffler 2004). Even if they are unlikely to experience civil war, however, they may be relatively

corrupt (e.g., Cameroon, Gabon, Nigeria) and autocratic (e.g., Libya, Syria, Tunisia, Yemen, and the Arab Gulf

states historically).

27

increases. The estimates in Column (1) suggest that this is the case. A standard deviation (0.011)

increase in the fossil fuel CTOT shock increases the annual death toll in the ID fuel exporters by

about 32%. Columns (2)-(3) show that the results for the ethnic dominance and polarized

countries are similar.

The positive relationship between the positive fossil-fuel CTOT shocks and conflict raise

the potential identification concern that conflict upticks in the fuel exporters increase the global

prices of fossil fuels, that is. In that regard, however, note that the reverse-causality hypothesis

can only explain the positive sign on the positive CTOT shocks in Table 9 and not the negative

sign on the negative shocks: if increases in conflict caused global fuel prices to increase, then

decreases in conflict should cause decrease in fuel prices. This implies that the regression

coefficients for both the positive and the negative CTOT shocks should be positive.

Second, we can address the reverse causality concern by inspecting the sample countries

and removing the ones that could plausibly affect global oil prices. Reviewing the sample

countries for the regression in Table 9, Column (1), the country that we should be most

immediately concerned with is Saudi Arabia. However, Saudi Arabia only contributes a single

conflict observation and dropping this observation gives virtually identical results. Nonetheless,

global energy and particularly oil markets can be volatile, so potentially even small events that

do not cause major contemporaneous supply disruptions can affect oil prices by affecting

expectations about future supplies and changing investors’ precautionary demand (Kilian 2014).

Although it is impossible to rule out such mechanisms, we can drop the countries that we believe

could plausibly affect global fuel prices either via current production or expectations. These

countries include Iran, Iraq, Mexico Oman, Russia, Saudi Arabia, and Venezuela.

28

Second, we can inspect the sample countries and remove the ones we believe could affect

global oil prices. Reviewing the Table 9, Column (1) sample immediately suggests that we

should omit Saudi Arabia. Dropping the single Saudi Arabia observation, however, gives

virtually identical results. Nonetheless, we cannot rule out that some of the other producers can

influence global prices and even events that do not affect the contemporaneous oil-supply could

affect oil prices by affecting investor expectations about future supplies and precautionary

demand (Kilian 2014). Although it is difficult to rule out such mechanisms entirely –

particularly, we do not observe market expectations regarding future supplies - we can drop the

countries that we intuitively believe could affect global fuel prices. These countries include Iran,

Iraq, Mexico Oman, Russia, Saudi Arabia, and Venezuela. Unfortunately, omitting these

countries only leaves us with 3 of the original 10 ID fuel exporters - Colombia, Malaysia, and

Trinidad and Tobago – and only 40 battle deaths observations. We hesitate to try to estimate

such a small sample; the year effects alone would perfectly explain one of the time-series. In

order to address this problem, the Columns (4)-(5) regressions redefine the ID countries as the

countries in the 15th-85th rather than 25th-75th ethnic-fractionalization percentiles. This implies

that we add the four countries we discussed earlier - Angola, Azerbaijan, Indonesia, and Sudan –

to the fuel exporter sample. Column (4) reports the estimates for all the ID fuel exporters in 15th-

85th fractionalization percentile. Column (5) reports the results without Iran, Iraq, Mexico Oman,

Russia, Saudi Arabia, and Venezuela. The results in Column (4) are almost identical to the

results in Column (1). The Column (5) results show that omitting the potentially-influential fuel

producers actually increases the positive-shock (as well as the negative-shock) estimates. These

results suggest that reverse causality is unlikely to explain the paper’s findings. Finally, we note

that the results are consistent with Dube and Vargas (2013), who show that higher international

29

oil prices increase conflict in Colombian municipalities. As the authors note, Colombia supplies

less 1% of the world’s oil, so the causal arrow is more likely to run from fuel prices to conflict

than in the other direction.***

[Table 9 Goes Here]

VI. Conclusion

In this paper, we study the effects of economic factors on the intensity of internal armed

conflicts. We show that adverse ethnic compositions in the form of ethnic dominance and

polarization critically affect how countries respond to commodity-generated income shocks.

Although negative shocks increase the death toll everywhere, positive shocks in countries with

an intermediate ethnic fractionalization index, ethnic dominance, and ethnic polarization also

increase the death toll. Finally, we show that the conflict-increasing effect of commodity

windfalls in these countries reflect fact that fossil fuel terms of trade gains increase conflict in

fuel exporters.

The results may reflect that income growth increases the opportunity cost of conflict

inputs as well as the return to fighting and the ability to finance the wars. In fuel-exporters with

adverse ethnic compositions, the rent-seeking and conflict-finance effects can, potentially,

dominate the opportunity-cost effect during fuel-price booms. (Dal Bó and Dal Bó 2011; Dube

and Vargas 2013). The literatures on the resource curse (Le Billon 2001; Fearon 2005) and

ethnicity (Horowitz 1985; Esteban and Ray 1999) suggest that fuel dependence, ethnic

*** In order to ensure that outlier effects do not explain the results, Appendix G depicts the partial residuals for the

positive and negative fossil fuel terms of trade shocks in the Table 9, Column (1) and (5) specifications. None of the

four panels, indicate that outliers affect the estimates.

30

dominance, and polarization can increase rent-seeking. The fossil fuel results could also be

relevant to the resource-curse literature as they suggest that resource booms may only increase

conflict in economies with structural weaknesses (adverse ethnic compositions and fuel

dependence).

Finally, our study suggests that economic stabilization policies could diminish the

number of human causalities and other welfare losses that are generated by civil wars. Moreover,

for a subset of particularly vulnerable economies, there may be no such thing as a “good”

economic shock. Thus, better management of both positive and negative economic shocks could

save lives.

Plain Language Summary

In this paper, we study the effects of commodity terms of trade shocks – the change in countries’

commodity export prices relative to their import prices - on the number of fatalities in civil wars.

We find that terms of trade growth, which should normally increase national income and create

more economic opportunities for individuals, decreases the annual numbers of fatalities.

However, when the fossil fuel terms of trade – the price of fuel exports relative to imports –

increase the death toll in fuel- (typically, oil-) exporting economies with certain ethnic

compositions that have been linked to conflict increases. The findings suggest that, in most

cases, national income growth decreases conflict. In some cases, however, conflict between

ethnic groups over the distribution of resources can undermine the positive effects of income

growth on peace.

31

References

Acemoglu, D., & Robinson, J. A. (2001). A theory of political transitions. American Economic

Review, 938-963.

Agénor, Pierre-Richard, and Montiel, Peter J. (2008). ‘Development macroeconomics’. 3rd ed.

New Jersey: Princeton University Press.

Aguirre, Alvaro. (2016). ‘Fiscal Policy and Civil Conflict in Africa’ Journal of African

Economies 25 (4): 614-36.

Alesina, A., Devleeschauwer, A., Easterly, W., Kurlat, S. and Wacziarg, A. (2003).

‘Fractionalization.’ Journal of Economic Growth 8 (2): 155-94

Amemiya, T. and MaCurdy, T. E. (1986). ‘Instrumental-Variable Estimation of An Error-

Components model.’ Econometrica 54 (4): 869-80.

Andersen, J., Frode, J., Nordvik, M., and Tesei, A. (2017). ‘Oil and Civil Conflict: On and Off

(Shore).’ Centre for Applied Macro- and Petroleum Economics (CAMP) Working Paper Series

No. 1/2017

Angrist, Joshua D., and Kugler, Adriana D. (2008). ‘Rural windfall or a new resource curse?

Coca, income, and civil conflict in Colombia.’ Review of Economics and Statistics 90 (2): 191-

215.

32

Arellano, M., and Bond, S. R. (1991). ‘Some Tests of Specification for Panel Data: Monte Carlo

Evidence and an Application to Employment Equations.’ Review of Economic Studies, 58(2):

277–97.

Arellano, M., and Bond, S. R. (1998). ‘Initial Conditions and Moment Restrictions in Dynamic

Panel Data Models.’ Journal of Econometrics 87(1): 115-43.

Aslaksen, S., and Torvik, R. (2006). ‘A Theory of Civil Conflict and Democracy in Rentier

States.’ Scandinavian Journal of Economics 108 (4): 571-85.

Bannon, I., and Collier P. (eds.). (2003). Natural resources and violent conflict: options and

actions. Washington, DC: World Bank Publications.

Basedau, M., and Lay, J. (2009). ‘Resource curse or rentier peace? The ambiguous effects of oil

wealth and oil dependence on violent conflict.’ Journal of Peace Research 46 (6): 757-76.

Bazzi, S., and Blattman, C. (2014). ‘Economic Shocks and Conflict: Evidence from Commodity

Prices.’ American Economic Journal: Macroeconomics 6 (4): 1-38.

Besley, T. and Persson, T. (2010). State capacity, conflict, and development. Econometrica,

78 (1):1–34.

33

Blattman, C., and Miguel E. (2010). ‘Civil War.’ Journal of Economic Literature 48 (1): 3–57.

Chandra, K. (2007). Why Ethnic Parties Succeed: Patronage and Ethnic Head Counts in India.

New York: Cambridge University Press.

Collier, P. and Hoeffler, A. (2004). ‘Greed and grievance in civil war.’ Oxford economic papers

56 (4): 563-595.

Cotet, Anca M., and Tsui, Kevin K. (2013). ‘Oil and conflict: What does the cross country

evidence really show?’ American Economic Journal: Macroeconomics 5 (1): 49-80.

Dal Bó, E. and Dal Bó, P. (2011). ‘Workers, Warriors and Criminals: Social Conflict in General

Equilibrium.’ Journal of the European Economic Association, 9 (4): 646–677.

Dixit, Avinash, and Victor Norman (1980). Theory of International Trade: A Dual, General

Equilibrium Approach. Cambridge: Cambridge University Press.

Dowden, R. (2009). Africa: Altered States, Ordinary Miracles. New York: Public Affairs

Dube, O. and Vargas, Juan F. (2013). ‘Commodity Price Shocks and Civil Conflict: Evidence

from Colombia.’ Review of Economic Studies 80 (4): 1384-1421.

34

Easterly, W., Kremer M., Pritchett, L. and Summers, Lawrence H. (1993). ‘Good Policy or Good

Luck?’ Journal of Monetary Economics 32 (3): 459-483.

Esteban, J. and Ray, D. (1994). ‘On the Measurement of Polarization.’ Econometrica 62 (4):

819–51.

Esteban, J. and Ray, D. (1999). ‘Conflict and distribution.’ Journal of Economic Theory 87 (2):

379-415.

Esteban, J. and Ray, D. (2011). ‘Linking conflict to inequality and polarization.’ American

Economic Review 101 (4): 1345-74

Esteban, J., Mayoral, L. and Ray, D. (2012). ‘Ethnicity and Conflict: An Empirical Study.’

American Economic Review 102 (4): 1310-42.

Fearon, James D. (2003). ‘Ethnic and Cultural Diversity by Country.’ Journal of Economic

Growth 8: 195-22.

Fearon, James D. (2005). ‘Primary commodity exports and civil war.’ Journal of conflict

Resolution 49 (4): 483-507.

Fearon, James D. and Laitin, D. (2003). ‘Ethnicity, Insurgency, and Civil War.’ American

Political Science Review 97 (1): 75-90.

35

Fearon, James D. and Laitin, D. (2011). ‘Sons of the Soil, Migrants, and Civil War.’ World

Development 39 (2): 199-211.

Feenstra, Robert C. (2015). Advanced international trade: theory and evidence. 3nd ed. New

Jersey: Princeton University Press.

Hausman, J. and Taylor, W. (1981). ‘Panel data and unobservable individual effects.’

Econometrica 49 (6): 1377– 1398.

Havard Strand. (2002). ‘Armed Conflict 1946-2001: A New Dataset.’ Journal of Peace Research

39 (5): 615-37. Codebook for Version 4-2009 available at https://www.prio.org/Data/Armed-

Conflict/UCDP-PRIO/Armed-Conflicts-Version-X-2009/

Hegre, Håvard and Nicholas Sambanis 2006. Sensitivity Analysis of Empirical Results on Civil

War Onset. Journal of Conflict Resolution 50, 4, 508-35

Hirshleifer, Jack. (1991). ‘The paradox of power.’ Economics and Politics, 3 (3): 177-200.

Horowitz, D. (1985). Ethnic Groups in Conflict. Berkeley: University of California Press.

Janus, T. and Riera-Crichton, D. (2015). ‘Economic Shocks, Civil War and Ethnicity.’ Journal of

Development Economics 115: 32-44.

36

Janus, T. and Riera-Crichton, D. (2017). ‘Controlling for Import Price Effects in Civil War

Regressions.’ Working Paper, University of Wyoming and Bates College.

Jervis, Robert. (1992). ‘Political implications of loss aversion.’ Political psychology 13 (2): 187-

204.

Keen, David. (2000). Incentives and disincentives for violence. Lynne Reinner Publishers;

International Development Research Centre.

Kilian, Lutz. (2014). ‘Oil price shocks: causes and consequences.’ Annual Review of Resource

Economics 6 (1): 133-154.

Krugman, Paul R, Obstfeld, M. and Melitz, M. (2014). International Economics: Theory and

Policy (10th ed.) Pearson, New York City, New York.

Lacina, B. (2006). ‘Explaining the severity of civil wars.’ Journal of Conflict Resolution 50 (2):

276-289.

Lacina, B. (2009). Battle Deaths Dataset 1946–2008. Codebook for Version 3.0. Centre for the

Study of Civil War (CSCW). International Peace Research Institute, Oslo (PRIO). See

https://www.prio.org/Data/Armed-Conflict/Battle-Deaths/The-Battle-Deaths-Dataset-version-30/

37

Lacina, B and Gleditsch, Nils P. (2005). ‘Monitoring Trends in Global Combat: A New Dataset

of Battle Deaths.’ European Journal of Population 21 (2–3): 145–166.

http://www.prio.no/CSCW/Datasets/Armed-Conflict/Battle-Deaths/

Le Billon, P. (2001). ‘The Political Ecology of War: Natural Resources and Armed Conflicts.’

Political geography 20 (5): 561-584.

Lederman, D. and Porto, G. (2016). ‘The price is not always right: on the impacts of commodity

prices on households (and countries).’ World Bank Research Observer 31(1): 168-197.

Matsuyama, K. (1988). ‘Terms-of-Trade, Factor Intensities and the Current Account in a Life-

Cycle Model.’ Review of Economic Studies 55 (2): 247-262.

Mayshar, Joram, Omer Moav, Zvika Neeman & Luigi Pascali (2015). \Cereals, Appropriability

and Hierarchy". CEPR Discussion Paper 10742, London, England: Centre for Economic Policy

Research

Mendoza, E. (1995). ‘The Terms of Trade, The Real Exchange Rate, and Economic

Fluctuations.’ International Economic Review 36 (1): 101-37

Miguel, E, Satyanath, S. and Sergenti, E. (2004). ‘Economic Shocks and Civil Conflict: An

Instrumental Variables Approach.’ Journal of Political Economy 112 (4): 725-53.

38

Montalvo, José G. and Reynal-Querol, M. (2005). ‘Ethnic Polarization, Potential Conflict and

Civil Wars.’ American Economic Review 95 (3): 796–816.

Obstfeld, M, and Rogoff, Kenneth S. (1996). Foundations of International Macroeconomics.

Cambridge, MA: MIT Press.

Østby, G. (2008). ‘Polarization, horizontal inequalities and violent civil conflict.’ Journal of

Peace Research 45 (2): 143-62.

Parker, D. P., & Vadheim, B. (2017). ‘Resource Cursed or Policy Cursed? US Regulation of

Conflict Minerals and Violence in the Congo.’ Journal of the Association of Environmental and

Resource Economists 4 (1): 1-49.

Posen, Barry R. (1993). The Security Dilemma and Ethnic Conflict. Survival 35 (1): 27-47.

Ricci, Luca A., Milesi-Ferretti, Gian M. and Lee, Jaewoo. (2008). ‘Real Exchange Rates and

Fundamentals: A Cross-Country Perspective.’ IMF Working Paper 08/13.

Ross, M. (2003). ‘Oil, drugs, and diamonds: How do natural resources vary in their impact on

civil war?’ In The political economy of armed conflict: Beyond greed and grievance, eds. Karen

Ballentine and Jake Sherman. Boulder: Lynne Rienner, 47-67.

39

Ross, Michael L. (2004). ‘What do we know about natural resources and civil war?’ Journal of

Peace Research 41 (3): 337-56.

Ross, Michael L. (2005). ‘Resources and Rebellion in Aceh, Indonesia.’ In Paul Collier and

Nicholas Sambanis (eds.) Understanding Civil War: Evidence and Analysis Vol. 2: Europe,

Central Asia, and Other Regions. Washington, DC: The World Bank.

Ross, M. (2012). The oil curse: how petroleum wealth shapes the development of nations. New

Jersey: Princeton University Press.

Rubin, Barnett R. (2000). ‘The political economy of war and peace in Afghanistan.’ World

Development 28 (10): 1789-1803.

Smith, Anthony D. (1986). The Ethnic Origin of Nations. Oxford: Blackwell Publishers.

Smith, B. (2004). ‘Oil wealth and regime survival in the developing world, 1960–1999.’

American Journal of Political Science 48 (2): 232-246.

Spatafora, N. and Tytell, I. (2009). ‘Commodity Terms of Trade: The History of Booms and

Busts.’ IMF Working Papers No. 09-205.

Thaler, Richard H., and Eric J. Johnson. (1990). ‘Gambling with the house money and trying to

break even: The effects of prior outcomes on risky choice.’ Management science 36 (6):643-660.

40

Thaler, Richard H., et al. (1997). ‘The effect of myopia and loss aversion on risk taking: An

experimental test.’ Quarterly Journal of Economics 112 (2): 647-661.

Turnovsky, Stephen J. (1993). ‘The Impact of Terms of Trade Shocks on a Small Open

Economy: A Stochastic Analysis.’ Journal of International Money and Finance 12 (3): 278-297.

Tversky, A., & Kahneman, D. (1991). Loss aversion in riskless choice: A reference-dependent

model. The quarterly journal of economics, 106(4), 1039-1061.

Valentino, B., Huth, P. and Balch-Lindsay, D. (2004). '‘Draining the sea’: mass killing and

guerrilla warfare.’ International Organization 58 (2): 375-407.

41

Tables and Figures

TABLE 1

Summary Statistics

Variable Obs Mean SD Min Max

Total Battle Deaths Low Estimate 906 1,196 2,899 14 37,000 Total Battle Deaths High Estimate 906 5,820 12,437 26 200,000 Total Battle Deaths Best Estimate 655 3,603 8,320 25 80,000 Commodity Terms of Trade Shock 906 0.001 0.016 -0.099 0.132 3 Year Moving Ave. CTOT Shock 906 0.001 0.012 -0.071 0.111 3 Year Mov. Ave. Fuel CTOT Shock 906 0.001 0.011 -0.076 0.097 Conflict Duration Up to Present Year 906 10.62 11.03 1 60 First Year of Conflict Dummy 906 0.16 0.37 0 1 Intermediate Ethnic Diversity Dummy 900 0.45 0.50 0 1 Ethnic Dominance Dummy 900 0.45 0.50 0 1 High Ethnic Polarization Dummy 900 0.45 0.50 0 1

TABLE 2

Sample Countries

Afghanistan Cuba Haiti Mauretania* Philippines Trin &Tob* Angola DR Congo India Mexico* Rep of Congo Tunisia Argentina Djibouti* Indonesia Morocco* Rwanda Turkey* Azerbaijan Dom Rep* Iran* Mozambique Saudi Arabia* Uganda Bangladesh Egypt Iraq* Nepal* Senegal Uruguay Bolivia El Salvador Ivory Coast Nicaragua* Sierra Leone Venezuela* Burkina Faso Eritrea* Kenya Niger* Somalia Vietnam Burundi* Ethiopia Laos* Nigeria South Africa Zimbabwe* Cambodia Gabon Lebanon Oman* Sri Lanka*

Cameroon Gambia Lesotho* Pakistan* Sudan Central African

Ghana Liberia Panama* Syria*

Chad Guatemala* Madagascar Papua New G Tajikistan*

Chile* Guinea* Malaysia* Paraguay Thailand*

Colombia* Guin.-Bissau Mali Peru* Togo Note: * indicates intermediately ethnically diverse countries

42

TABLE 3

The effects of commodity terms of trade shocks on battle-related deaths in civil wars (1) (2) (3) (4) (5) (6) (7) (8)

Estimation Method LSDV LSDV LSDV HTaylor LSDV LSDV LSDV LSDV

Dep. Variable: Ln (battle deaths)

dCTOT(t) -7.295

[5.910]

dCTOT(t-1) 0.283

[5.319]

dCTOT(t-2) -11.708**

[5.616]

dCTOT*ID(t) 11.349*

[6.603]

dCTOT*ID(t-1) 2.102

[6.127]

dCTOT*ID(t-2) 7.335

[6.046]

∆TOT(t) -17.107* -16.737* -26.620*** -16.534 -14.701 -32.917** -42.044***

[9.583] [9.687] [9.748] [12.422] [10.084] [16.089] [12.924]

∆TOT*ID(t) 20.116* 20.690* 23.843** 17.554 18.639 35.267** 38.146***

[10.995] [10.731] [10.490] [13.111] [11.871] [15.862] [13.236]

Duration -0.014 -0.013 -0.007 -0.027** -0.019 -0.028 -0.055* 6.991***

[0.023] [0.022] [0.021] [0.011] [0.032] [0.025] [0.031] [1.921]

First Year Dummy -0.950*** -0.927*** -0.893*** -0.937*** -0.155

[0.166] [0.165] [0.168] [0.193] [0.202]

AR(1) term 0.463***

[0.034]

Observations 900 900 885 756 664 742 536 900

R-squared 0.523 0.519 0.537 0.587 0.514 0.615 0.677

# countries/conflicts 79 79 78 59 50 76 45 154

p-val sum of shocks 0.06 0.08 0.09 0.01 0.19 0.15 0.05 0.00

p-val sum of interac. 0.08 0.07 0.06 0.02 0.19 0.12 0.03 0.00

p-val shocks+interac. 0.67 0.43 0.23 0.42 0.74 0.47 0.46 0.03

p-val shocks equal 0.35

p-val interac equal 0.6

Year dummies Y Y Y Y Y Y Y Y

Cntry/conf time trnds Y Y Y Y Y Y Y Y

Note: Robust standard errors clustered at the country-level (except in Column (1)) in brackets. * significant at 10%; ** significant at 5%; *** significant at 1%. ∆ denotes the change in the three-year moving average. Columns (1)-(2) estimate the effects of the current and two preceding years’ commodity terms of trade shocks on battle-related fatalities in countries with and without an intermediate ethnic diversity level. Column (2) estimates the same effects of the growth rate of the three-year moving average terms of trade shock. Column (3) replaces the intermediate diversity dummy based on the Fearon (2003) ethnicity dataset with the corresponding dummy based on the Alesina et al. (2003) ethnicity dataset. Column (4) includes a lagged dependent variable and uses Hausman and Taylor (1981) and Amemiya and MaCurdy (1986) to correct the dynamic panel bias. Column (5) restricts sample to observations which are at least three years into the conflict. Column (6) restricts it to observations for which Lacina and Gleditsch (2005) report year-specific battle-related fatalities. Column (7) imposes the two restrictions simultaneously. Column (8) estimates the column (2) specification with conflict fixed effects and a quadratic conflict-specific time trends rather than country fixed effects and country-specific time trends.

43

TABLE 4

The effects of positive and negative commodity terms of trade shocks in non-intermediately

diverse and intermediately diverse countries

(1) (2) (3) (4) (5)

Estimation Method LSDV LSDV LSDV LSDV LSDV

Sample Full ID NID NID NID

Pos∆ΤΟΤ(t) 19.171*** 20.118** -5.723

[7.208] [8.393] [36.590]

Neg∆ΤΟΤ(t) -24.477*** -34.393*** -25.754

[5.475] [7.495] [24.485]

∆ΤΟΤ(t) -15.605 -28.630***

[10.863] [9.895]

Duration -0.012 -0.050 0.000 0.000 0.007

[0.021] [0.045] [0.012] [0.012] [0.009]

First Year -0.928*** -1.023*** -0.926*** -0.920*** -0.863***

[0.159] [0.242] [0.219] [0.216] [0.216]

Observations 906 409 491 491 464

R-squared 0.524 0.486 0.614 0.614 0.614

p-val (Pos∆ΤΟΤ=-Neg∆ΤΟΤ ) 0.00 0.00 0.73

Number of countries 81 40 39 39 38

Year dummies Y Y Y Y Y

Country time trends Y Y Y Y Y

Note: Robust standard errors clustered at the country-level in brackets. * significant at 10%; ** significant at 5%; *** significant at 1%. ∆ denotes the change in the three-year moving average. Column (1) estimates the effects of positive and negative commodity terms of trade shocks in the full sample. Columns (2)-(3) estimate the effects in, respectively, the intermediately and non-ethnically diverse countries. Column (4) replaces the positive and negative shocks in the non- intermediately diverse countries with the original terms shock measure. Column (5) excludes Indonesia from the non-intermediately diverse sample.

44

TABLE 5

Results with polarization and ethnic dominance-based sample divisions

(1) (2) (3) (4) (5)

Estimation Method LSDV LSDV LSDV LSDV LSDV

Sample High

Polarization

Low

Polarization

Ethnic

Dominance

Non-Ethnic

Dominance

Non-Ethnic

Dominance

∆ΤΟΤ(t) -19.460*** -14.701 -26.134***

[6.011] [10.108] [9.639]

Pos∆ΤΟΤ(t) 23.783*** 18.726**

[4.716] [8.558]

Neg∆ΤΟΤ(t) -24.703*** -34.278***

[5.177] [8.287]

Duration 0.016 -0.065** -0.052 -0.002 0.005

[0.019] [0.032] [0.046] [0.013] [0.010]

First Year -0.711*** -1.235*** -0.976*** -0.989*** -0.923***

[0.161] [0.253] [0.251] [0.210] [0.209]

Observations 549 355 383 510 483

R-squared 0.579 0.554 0.468 0.624 0.625

Number of countries 48 32 37 41 40

Year dummies Y Y Y Y Y

Country time trends Y Y Y Y Y

Note: Robust standard errors clustered at the country-level in brackets. * significant at 10%; ** significant at 5%; *** significant at 1%. ∆ denotes the change in the three-year moving average. Columns (1)-(2) divide the sample into countries with high and low ethnic polarization, defined as a polarization index below and above the sample median. Columns (3)-(4) divide the sample into countries with and without a dominant ethnic group, defined as a group that represents 50-85% of the population. Column (5) reports the results without Indonesia.

45

TABLE 6

Robustness to alternative dependent variables (intermediately diverse sample)

(1) (2) (3)

Estimation Method LSDV LSDV LSDV

Dep. Var. Measure of Battle Deaths Low High Ordinal

Sample ID ID ID

Pos∆ΤΟΤ(t) 23.560** 21.119*** 3.669

[9.090] [7.123] [2.990]

Neg∆ΤΟΤ -19.120 -33.631*** -7.725**

[13.801] [7.579] [3.199]

Duration -0.057 -0.054 -0.018

[0.042] [0.043] [0.012]

First Year -0.920*** -0.848** -0.207**

[0.246] [0.317] [0.085]

Observations 409 409 409

R-squared 0.440 0.473 0.389

Number of countries 40 40 40

Year dummies Y Y Y Country time trends Y Y Y Note: Robust standard errors clustered at the country-level in brackets. * significant at 10%; ** significant at 5%; *** significant at 1%. ∆ denotes the change in the three-year moving average. The regression estimates apply to the intermediately ethnically diverse sample countries. Column (1) report the effects of positive and negative commodity terms of trade shocks on the best and imputed best battle deaths measure. In columns (2)-(4) we estimate the effects on the low and high estimates for the annual battle-related deaths in Lacina and Gleditisch (2005) and the effects on an ordinal measure which equals one when the best or imputed best estimate is at most 1000 and two when the best or imputed best estimate exceeds 1000.

TABLE 7

Robustness to alternative dependent variables (non-intermediately diverse sample)

(1) (2) (3) (4) (5) (6)

Estimation Method LSDV LSDV LSDV LSDV LSDV LSDV

Dep. Var. Measure of

Battle Deaths Low High Ordinal Low High Ordinal

Sample NID NID NID NID NID NID

∆ΤΟΤ(t) -13.940 -12.571 -3.310 -38.462** -28.558** -5.787**

[19.514] [12.778] [2.524] [15.174] [11.706] [2.495]

Duration 0.050 0.015 0.002 0.069*** 0.025** 0.001

[0.031] [0.016] [0.005] [0.021] [0.012] [0.005]

First Year -0.785*** -0.767*** -0.157** -0.664*** -0.671*** -0.158**

[0.269] [0.207] [0.074] [0.240] [0.192] [0.075]

Observations 491 491 491 464 464 464

R-squared 0.515 0.517 0.478 0.541 0.532 0.453

Number of countries 39 39 39 38 38 38

Year dummies Y Y Y Y Y Y Country time trends Y Y Y Y Y Y Note: Robust standard errors clustered at the country-level in brackets.* significant at 10%; ** significant at 5%; *** significant at 1%. ∆ denotes the change in the three-year moving average. The regression estimates apply to the non-intermediately ethnically diverse sample countries. Columns (2)-(4) report the effects of commodity terms of trade shocks using the low and high estimates for the annual battle-related deaths in Lacina and Gleditisch (2005) as well as an ordinal measure which equals one when the best or imputed best estimate is at most 1000 and two when the best or imputed best estimate exceeds 1000. Columns (3)-(6) reports the estimates without Indonesia.

46

TABLE 8

The effects of positive and negative shocks in intermediately and non-intermediately diverse net

fuel exporters and net fuel importers Estimation Method LSDV LSDV

Sample Full Full

Intermediate ethnic diversity definition Main (25th-75th percentile of

ethnic fractionalization index)

Extended (15th-85th percentile of

ethnic fractionalization index)

Pos∆ΤΟΤ(t) in ID net fuel exporter 23.070*** 24.710***

[6.183] [6.041]

Pos∆ΤΟΤ(t) in NID net fuel exporter 63.238* -5,440.035***

[33.034] [966.792]

Pos∆ΤΟΤ(t) in ID non-fuel exporter -13.219 -17.001

[25.606] [24.494]

Pos∆ΤΟΤ(t) in NID non-fuel exporter -41.426 -58.627*

[29.689] [34.689]

Neg∆ΤΟΤ(t) in ID net fuel exporter -17.170** -20.989***

[7.732] [6.678]

Neg∆ΤΟΤ(t) in NID net fuel exporter -78.935*** 262.301

[18.691] [178.779]

Neg∆ΤΟΤ(t) in ID non-fuel exporter -21.232 -18.187

[15.118] [13.170]

Neg∆ΤΟΤ(t) in NID non-fuel exporter -12.685 -11.021

[34.261] [59.105]

Duration -0.009 -0.009

[0.021] [0.021]

First year -0.924*** -0.904***

[0.160] [0.155]

Observations 900 900

R-squared 0.53 0.53

Number of countries 79 79

Year dummies Y Y

Country time trends Y Y

Note: Robust standard errors clustered at the country-level in brackets. * significant at 10%; ** significant at 5%; *** significant at 1%. ∆ denotes the change in the three-year moving average. ID and NID denote intermediately diverse and non-intermediately diverse countries. Column (1) reports the effects of positive and negative commodity terms of trade shocks in intermediately ethnically diverse (ID) and non-intermediately ethnically diverse (NID) net fossil fuel exporters and importers. Columns (1) and (2) define the intermediate ethnically diverse countries as countries with an ethnic fractionalization index in, respectively, the 25th -75th percentile and the 15th-85th percentile.

47

TABLE 9

The effects of fuel and non-fuel terms of trade shocks in fuel exporters

(1) (2) (3) (4) (5)

Estimation Method LSDV LSDV LSDV LSDV LSDV

Sample definition

Fuel exporters in

25th-75th percentile of ethnic fractionalization

Ethnic

dominance

High

polarization

Fuel exporters in 15th-

85th percentile of fractionalization

Small fuel

exporters in

15-85th

percentile of

fractionalization

Sample countries

Colombia Iran Iraq

Malaysia Mexico Oman Russia

Saudi Arabia Trinidad & Tobago

Venezuela

Colombia Iran Iraq

Malaysia Mexico Oman Russia

Saudi Arabia Venezuela

Angola Colombia

Iran Iraq

Malaysia Mexico Oman

Saudi Arabia Sudan

Trinidad & Tob Venezuela

Angola Azerbaijan Colombia Indonesia

Iran Iraq

Malaysia Mexico Oman Russia

Saudi Arabia Sudan

Trinidad & Tob Venezuela

Angola Azerbaijan Colombia Indonesia Malaysia

Sudan Trinidad & Tob

Pos fuel ∆ΤΟΤ(t) 28.69* 28.69* 37.34*** 32.42*** 119.32**

[13.605] [13.733] [6.319] [7.891] [37.996]

Pos non-fuel ∆ΤΟΤ(t) -24.66 -24.66 -17.06** -47.60 -173.12

[35.419] [35.752] [7.092] [28.210] [119.387]

Neg fuel ∆ΤΟΤ(t) -38.47*** -38.47*** -32.27*** -30.97*** -109.38***

[2.649] [2.674] [5.453] [7.001] [28.709]

Neg non-fuel ΤΟΤ(t) 16.80 16.80 -43.19 4.64 -76.12

[34.334] [34.657] [32.347] [40.085] [126.880]

Duration -0.09 -0.09 -0.01 -0.04 -0.02

[0.074] [0.075] [0.037] [0.036] [0.038]

First year -1.16* -1.16* -0.73*** -1.03*** -0.70

[0.547] [0.552] [0.185] [0.218] [0.518]

Observations 116 115 157 205 129

R-squared 0.723 0.72 0.71 0.644 0.740

No. countries 10 9 11 14 7

Year dumm. Y Y Y Y Y

Cntry time tr. Y Y Y Y Y

Note: Robust standard errors clustered at the country-level in brackets. * significant at 10%; ** significant at 5%; ***

significant at 1%. ∆ denotes the change in the three-year moving average. ID denote intermediately diverse. Column (1) reports the estimates for positive and negative fuel and non-fuel commodity terms of trade shocks in intermediately ethnically diverse net fuel exporters. Columns (2)-(3) repeat the analysis for net fuel exporters with ethnic dominance and high ethnic polarization. The Column (4) regression uses the countries with Herfindahl-Hirschman ethnic fractionalization index in the 15th-85th percentile instead of the 25th-75th percentile in the Column (1) regression. Column (5) omits fuel producers that could potentially influence global fossil fuel prices.

48

FIGURE 1

The Relationship Between Ethnic Fractionalization, Ethnic Polarization,

and the Population Share of the Ethnic Plurality

Source: Janus and Riera-Crichton (2015). Note: Fearon (2003) ethnicity data

0.5

1

0 .2 .4 .6 .8 1ethnic fractionalization

ethnic plurality

polarization

quadratic prediction of polarization,Rsq=0.91

linear prediction of ethnic plurality, Rsq=0.96

0.2

.4.6

.81

.2 .4 .6 .8 1ethnic plurality

polarization

quadratic prediction of polarization,Rsq=0.92

49

Appendix A

Death Toll Estimates for Guatemala, El Salvador, Nicaragua, and Bosnia and Herzegovina in the dataset used in the paper

FIGURE A1 Death Toll Estimates for Guatemala, El Salvador, Nicaragua, and Bosnia and Herzegovina

Note: Locally-weighted-regression (lowess) fits, varying bandwidths

02000

4000

6000

8000

10000

Death toll estimate

1950 1960 1970 1980 1990 2000Year

bandwidth = .4

Guatemala

-5000

05000

10000

15000

Death toll estimate

1970 1975 1980 1985 1990Year

bandwidth = .4

El Salvador0

2000

4000

6000

8000

Death toll estimate

1975 1980 1985 1990Year

bandwidth = .4

Nicaragua

05000

10000

15000

20000

25000

Death toll estimate

1992 1993 1994 1995Year

bandwidth = .8

Bosnia and Herzegovina

50

Appendix B

Do Positive Terms of Trade Shocks Increase the Onset Risk for Civil War?

In the main paper, we argue that the determinants of the onset and intensity of civil wars can

differ. In order to test this idea, we combine our dataset with the dataset for the onset of civil

wars in Janus and Riera-Crichton (2015). The latter paper shows that commodity terms of trade

declines predict the onset of civil wars in countries with intermediate ethnic diversity, ethnic

dominance, and high polarization. Civil war onset is measured with a dummy that equals zero in

peace years, one in the onset year, and missing otherwise. Since the ongoing-conflict

observations are excluded except for the onset years, there is almost no overlap in the country-

year coverage in the two datasets. The onset and intensity data both come from the UCDP/PRIO

(v. 4) civil conflict coding project, so in principle every onset in the onset data (Gleditsch et al.

2002; Themnér & Wallensteen 2011) has a corresponding battle-death time series in the intensity

dataset (Lacina and Gleditsch 2005) and vice versa. The onset dummy switches to one when the

annual battle-related death toll reaches a threshold of either 25 (for a minor conflict onset) or

1000 (a civil war onset).

In the main paper, we find that positive fossil fuel terms of trade shocks increase the

battle-related death toll in intermediately diverse net fuel exporters. In order to test whether these

shocks also increase the onset risk for civil conflict, we estimate equation (1’) in Janus and

Riera-Crichton (2015) for the intermediately ethnically diverse countries. Additionally, however,

we (a) decompose the commodity terms of trade growth rate into its positive and negative fossil

fuel and non-fuel components and (b) add an interaction between positive fossil fuel terms of

trade shocks and the dummy for fossil fuel exporters we defined above.

In contrast to what we found when we estimated battle-related fatalities conditional on a

prior war onset, the results in Table A1, Column (1), do not support the idea that positive fossil

51

fuel terms of trade shocks increase the onset risk for civil wars in the first place. Columns (2)-(4)

show that using the three alternative onset dummies in Janus and Riera-Crichton (2015) gives

similar results. This evidence is consistent with the idea that the determinants of the intensity and

onset of civil wars can differ. Nonetheless, Ross (2003) reviews both systematic and case-study

evidence and concludes that oil can increase the onset risk for civil wars and particularly

separatist conflict. It may be possible that one should distinguish between separatist and non-

separatist conflicts or for instance between off-and onshore oil (Andersen et al. 2017).

Nonetheless, the fact that estimating battle deaths conditional on an onset rather than the war

onsets gives different regression coefficients at least in the UCDP data and conditional on our

regression specifications suggest that, at least in our dataset, we should not estimate onsets and

battle deaths during the conflict years with the same regression model.

52

TABLE A1

Commodity terms of trade shocks and civil war onsets

(1) (2) (3) (4) (5) (6) (7) (8)

Estimation Method LSDV LSDV LSDV LSDV LSDV LSDV LSDV LSDV

Sample ID ID ID ID ID ID ID ID

Onset measures UCDP War

UCDP War

UCDP Conflict

UCDP Conflict

COW War

COW War

FL War

FL War

∆ΤΟΤ(t-1) -0.604* -0.715** -0.427** -0.375*

[0.306] [0.310] [0.189] [0.200]

Pos fuel ∆ΤΟΤ(t-1) -0.945 -1.473 -1.549 -1.637

[1.618] [1.779] [1.269] [1.352]

Pos fuel ∆ΤΟΤ(t-1)*nfe 0.351 -0.130 0.694 1.201

[1.645] [1.917] [1.229] [1.403]

Pos non-fuel ∆ΤΟΤ(t-1) 1.100* 0.360 -0.085 0.671

[0.589] [0.640] [0.429] [1.096]

Neg fuel ∆ΤΟΤ(t-1) -0.169 0.382 0.587 0.164

[0.688] [0.865] [0.590] [0.333]

Neg non-fuel∆ΤΟΤ(t-1) -6.552** -1.244 -2.276*** -2.919**

[2.746] [1.296] [0.608] [1.361]

Observations 2,149 2,149 2,238 2,238 2,454 2,454 1,504 1,504

R-squared 0.088 0.100 0.061 0.063 0.047 0.051 0.115 0.117

No. of countries 70 70 70 70 70 70 68 68

Year dummies Y Y Y Y Y Y Y Y

Country time tr. Y Y Y Y Y Y Y Y

p-val(Posfuel+Pf*nfe) 0.14 0.13 0.20 0.40

Note: Robust standard errors clustered at the country-level in brackets. * significant at 10%; ** significant at 5%; ***

significant at 1%. ∆ denotes the change in the three-year moving average. ID denote intermediately diverse. on civil war onsets. Columns (1) estimates the effects of the growth rate of the three-year moving average commodity terms of trade shock on the onset of civil wars in the Janus and Riera-Crichton (2015) civil war onset dataset. The onset variable equals one in a civil-war onset year and zero in peace years. The dummy is based on the Uppsala Conflict Data Program(UCDP). Column (2) separately estimates the effects of the positive and negative fuel and non-fuel terms of trade shocks as and allows the effects of positive fuel terms of trade shocks to differ for net fuel exporting countries.. Columns (3)-(4), (5)-(6), and (7)-(8) report the corresponding estimates for Janus and Riera-Crichton’s (2015) three alternative conflict onset dummies based on the Correlates of War dataset, Fearon and Laitin (2003), and the UCDP onset measure when the battle deaths threshold required for an onset is 25 instead of 1000 battle-related fatalities per year. We refer to Janus and Riera-Crichton (2015) for the detailed variable definitions and further discussion.

53

Appendix C

The estimation of linear compared to interval regressions and national vs. subnational data

Instead of using the linear fixed-effects model in the paper, we could alternatively follow Bazzi

and Blattman (2014), who estimate the effects of export price changes on battle-related fatalities

with an interval regression. This methodology, which is a non-linear estimation procedure, has

the advantage that it is precisely designed for situations where the researcher observes either an

interval or a specific value for the dependent variable. In our context, about a third of the battle

deaths observations are intervals. However, the interval regression model shares the potential

limitations of many other non-linear models. In particular, it does not allow us to include country

fixed effects, which are usually considered to be important in cross-country estimation.

Moreover, compared to the linear model it is more important, but it appears to be harder, to test

whether the errors are normally distributed. The reason why the error distribution is more

important than in the linear model is that the likelihood contribution of each interval observation

is the probability that the realized error term puts the dependent variable in the observed interval,

, (a1)

where and are the low and high estimates for battle deaths in country in year and

is the standard normal cdf. Therefore, the likelihood contribution and the likelihood

maximizing value of depend on assuming normality. In contrast, non-normally distributed

−Φ−

−Φ=<+<

σβ

σβ

εβ itlitithit

hitititlit

xyxyyxypr )(

lity hity i t

( ).Φ

β

54

errors do not bias the linear estimates (Arabmazar and Schmidt 1982; Lewbel and Linton

2002).†††

The fact that we estimate conflict intensity at the country level contrasts with the growing

study of subnational conflict outcome in the empirical conflict literature. data. Nonetheless, there

are three reasons why we doubt that the country-level approach will bias the estimates. First, the

main concern with cross-country studies is probably that countries have highly persistent but

unobservable cultural, historical, institutional, geographic, etc. conflict characteristics (Blattman

and Miguel 2010; Djankov and Reynal-Querol 2010; Cotet and Tsui 2013). In this paper,

however, we estimate a fixed-effects panel and not a cross-section of countries. The fixed effects

control for the mean effect of time-invariant, country-specific conflict determinants. Since the

fixed-effects estimator only uses the within-country deviations from the means to identify the

coefficients, ceteris paribus, it should only increase the bias compared to subnational panels if

the within-country deviations from the means are, loosely speaking, more highly correlated with

the error term than the subnational deviations from the means.

Second, due to the fact that subnational units are part of the national economy, they can

be exposed to correlated shocks, general equilibrium effects, and externalities. In that case, the

panel units lose their independence and the errors terms can become spatially correlated. The

creation of refugee flows and infrastructure destruction as well as supply and demand changes on

capita, labor, and goods markets can generate spillovers between neighboring municipalities or

provinces. Finally, it is important to note that the standard empirical civil war definitions in the

literature define a civil war as an armed conflict between the central government and a non-

governmental organization that kills a significant number of individuals (Lacina and Gleditsch

†††Greene (2004) finds that adding fixed effects in the Tobit model might not bias the estimates, but the error

variance and standard errors are incorrectly estimated.

55

2002; Fearon and Laitin 2003; Blattman and Miguel 2010). The fact that the central government

is involved suggests that there is a country-level unified actor that plays a strategic war-game in

which it engages strategically in several subnational battle theaters, and that operates under an

integrated national budget constraint. If a civil war is a strategic game that is played in the

country as a whole, the subnational units may be as inter-dependent as the battle theaters of

World War II.‡‡‡

References (uncited in the main paper)

Arabmazar, A., and Schmidt, P. (1982). ‘An investigation of the robustness of the Tobit

estimator to non-normality.’ Econometrica 50 (4): 1055-1063.

Ashraf, Q, and Galor, O. (2013). ‘The ‘Out of Africa’ hypothesis, human genetic diversity, and

comparative economic development.’ American Economic Review 103 (1): 1-46.

Greene, W. (2004). ‘Fixed effects and bias due to the incidental parameters problem in the Tobit

model.’ Econometric Reviews 23 (2): 125-147.

Lewbel, A., and Linton, O. (2002). ‘Nonparametric censored and truncated regression.’

Econometrica 70 (2): 765-779.

‡‡‡However, there are also examples of subnational violence where the central government is not obviously an

important decision maker (Dube and Vargas 2013; Bazzi and Gudgeon 2017).

56

Appendix D

Robustness of the Table 3 estimates to alternative ethnicity measures and control variables

In Table A2, we re-estimate the Table 3, Column (2) specification in the main paper with

alterative ethnicity measures: a dummy for ethnic diversity in the 15-85th instead of 25-75th

percentile; a dummy for high ethnic polarization with value one when the polarization measure

proposed in Esteban and Ray (1994) exceeds the sample median, under the assumption made in

Montalvo ad Reynal-Querol (2005) and Esteban et al. (2012) that the social distance between the

ethnic groups is one and zero otherwise; dummies for defining a dominant ethnic group as a

group that covers 50-85, 40-90, 45-90, ad 45-85 percent of the population; an above-median

polarization dummy using the alternative Alesina et al. (2003) ethnicity dataset; and a dummy

for ethnic diversity in the 25th-76th percentile using the alternative Soviet Atlas Narodov Mira

dataset. The results remain robust.

In Table A3, we re-estimate the Table 3, Column (2) specification in the main paper but

control for other factors - other than ethnicity - that could mediate the effects of commodity

terms of trade shocks. Column (1) includes an interaction between the terms of trade shocks and

a dummy for above-median mountainous terrain land cover, Column (2) adds an interaction with

a dummy for a non-contiguous state. Columns (3)-(4) interact the shock with dummies for

above-median and intermediate (1st-3rd quartile) religious fractionalization (Fearon and Laitin

2003). Columns (5)-(6) interact the shock with dummies for above-median and intermediate (1st-

3rd quartile) predicted genetic diversity (Ashraf and Galor 2013). Column (7) interact the shock

with dummies for British and French colonies (Fearon and Laitin 2003). Column (8) includes the

controls jointly except for including just one of the two religious diversity and one of the two

predicted genetic diversity measures. Except for the predicted genetic diversity measure, which

comes from Ashraf and Galor (2013), all the controls come from Fearon and Laitin (2003).

57

TABLE A2

Robustness to alternative ethnicity measures (1) (2) (3) (4) (5) (6) (7) (8)

Estimation Method LSDV LSDV LSDV LSDV LSDV LSDV LSDV LSDV

Dep. Variable: Ln (battle deaths)

∆TOT(t) -31.302 -15.914* -15.975* -32.470*** -17.872* -15.975* -34.180*** -17.107*

[19.082] [8.755] [9.300] [11.296] [10.277] [9.300] [9.222] [9.583]

∆TOT*EF1585 33.252

[20.016]

∆TOT*High polar 18.024*

[9.799]

∆TOT*Plural5085 18.812*

[10.770]

∆TOT*Plural4090 35.404***

[12.291]

∆TOT*Plural4590 20.515*

[11.428]

∆TOT*Plural4585 18.812*

[10.770]

∆TOT*Hi polar Ales 39.112***

[9.749]

∆TOT*ID(Sovi.dta) 20.116*

[10.995]

Duration -0.013 -0.013 -0.013 -0.013 -0.013 -0.013 -0.013 -0.013

[0.022] [0.022] [0.023] [0.022] [0.023] [0.023] [0.021] [0.022]

First Year Dummy -0.952*** -0.926*** -0.939*** -0.952*** -0.945*** -0.939*** -0.919*** -0.927***

[0.158] [0.160] [0.167] [0.162] [0.166] [0.167] [0.162] [0.165]

Observations 900 904 893 893 893 893 899 900

R-squared 0.518 0.518 0.519 0.521 0.519 0.519 0.523 0.519

# countries 79 80 78 78 78 78 79 79

p-val shocks 0.11 0.07 0.09 0.01 0.09 0.09 0 0.08

p-val interac. 0.10 0.07 0.08 0.01 0.08 0.08 0 0.07

p-val shocks+interac. 0.62 0.57 0.47 0.42 0.48 0.47 0.09 0.43

Year dummies Y Y Y Y Y Y Y Y

Cntry/conf time trnds Y Y Y Y Y Y Y Y

Note: Robust standard errors clustered at the country-level (except in Column (1)) in brackets. * significant at 10%; ** significant at 5%; *** significant at 1%. ∆ denotes the change in the three-year moving average. Columns (1)-(8) estimate the Table 3, Column (2) specification in the main paper with alterative ethnicity measures: a dummy for ethnic diversity in the 15-85th instead of 25-75th percentile; a dummy for high ethnic polarization with value one when the polarization measure proposed in Esteban and Ray (1994) exceeds the sample median, under the assumption made in Montalvo ad Reynal-Querol (2005) and Esteban et al. (2012) that the social distance between the ethnic groups is one and zero otherwise; dummies for defining a dominant ethnic group as a group that covers 50-85, 40-90, 45-90, ad 45-85 percent of the population; an above-median polarization dummy using the alternative Alesina et al. (2003) ethnicity dataset; and a dummy for ethnic diversity in the 25th-76th percentile using the alternative Soviet Atlas Narodov Mira dataset.

58

TABLE A3

Robustness to geographical, historical, and non-ethnic diversity measures (1) (2) (3) (4) (5) (6) (7) (8)

Estimation Method LSDV LSDV LSDV LSDV LSDV LSDV LSDV LSDV

Dep. Variable: Ln (battle deaths)

∆TOT(t) -34.110*** -56.613*** -18.351* -20.896** -21.045* -15.603 -37.573*** -38.500

[9.669] [14.136] [10.292] [10.268] [12.048] [12.164] [11.466] [25.998]

∆TOT*ID(t) 31.104*** 52.565*** 19.688* 14.819 17.204 21.311 31.975** 67.057***

[10.332] [14.800] [11.557] [10.998] [12.986] [14.340] [12.244] [22.844]

∆TOT*(Hi%Mnts) -2.541 -31.785

[6.199] [26.642]

∆TOT*(non-contig) 41.859* 44.609

[22.827] [28.664]

∆TOT*(Hi rel.frac) 2.432

[7.198]

∆TOT*(Int. rel.frac). 9.367 18.627

[11.020] [20.914]

∆TOT*(Hi gen.div) 8.003

[10.686]

∆TOT*(Int. gen.div). -2.826 -25.697

[17.521] [19.638]

∆TOT*Brit. colony 2.294 -24.350

[6.566] [26.385]

∆TOT*Fr. colony 15.652 33.512

[22.677] [26.074]

Duration -0.036 -0.035 -0.013 -0.012 -0.012 -0.013 -0.035 -0.031

[0.030] [0.030] [0.022] [0.022] [0.022] [0.022] [0.030] [0.029]

First Year Dummy -0.542*** -0.549*** -0.927*** -0.922*** -0.933*** -0.928*** -0.540*** -0.568***

[0.147] [0.144] [0.165] [0.166] [0.165] [0.164] [0.147] [0.147]

Observations 702 702 899 899 900 900 702 701

R-squared 0.637 0.641 0.519 0.519 0.519 0.519 0.638 0.646

# countries 75 75 78 78 79 79 75 74

p-val shock 0 0 0.08 0.05 0.08 0.20 0 0.14

p-val interac. 0 0 0.09 0.18 0.19 0.14 0.01 0

p-val shock+interac. 0.16 0.17 0.85 0.58 0.72 0.76 0.41 0.36

Year dummies Y Y Y Y Y Y Y Y

Cntry/conf time trnds Y Y Y Y Y Y Y Y

Note: Robust standard errors clustered at the country-level (except in Column (1)) in brackets. * significant at 10%; ** significant at 5%; *** significant at 1%. ∆ denotes the change in the three-year moving average. Columns (1)-(8) estimate the Table 3, Column (2) specification in the main but paper but controls for other factors than ethnicity that could mediate the effects of commodity terms of trade shocks. Column (1) includes an interaction between the terms of trade shocks and a dummy for above-median mountainous terrain land cover, Column (2) adds an interaction with a dummy for a non-contiguous state. Columns (3)-(4) interact the shock with dummies for above-median and intermediate (1st-3rd quartile) religious fractionalization. Columns (5)-(6) interact the shock with dummies for above-median and intermediate (1st-3rd quartile) predicted genetic diversity. Column (7) interact the shock with dummies for British and French colonies. Column (8) includes the controls jointly except for including just one of the two religious diversity and one of the two predicted genetic diversity measures. Except for the predicted genetic diversity measure, which comes from Ashraf and Galor (2013), all the controls come from Fearon and Laitin (2003).

59

Appendix E

Evidence that Indonesia could be considered an intermediately diverse country

In the main paper, we argue that it may be appropriate to categorize Indonesia as an

intermediately diverse country and include it in the intermediately diverse country group. In this

appendix, we explain why we believe this classification may be appropriate. We first provide an

argument based on the coding of the ethnic groups. We then provide an empirical argument that

shows that conflict intensity in Indonesia is poorly explained by the linear specification for the

non-intermediately diverse countries and well explained by the specification for the

intermediately diverse countries, that is, in Indonesia, both positive and negative income shocks

increase the conflict intensity.

In terms of the coding procedure, Indonesia does not quite match our empirical definition

of an intermediately ethnically diverse country: in order to be intermediately diverse, the ethnic

fractionalization index must be in the second to third quartiles or between 0.25-0.68. Indonesia’s

fractionalization index, however, is 0.77. Indonesia, similarly, does not quite satisfy our

definition of ethnic dominance: in order to have a dominant ethnic group, we require that the

largest ethnic group represents 50-85 percent of the population. The largest group in Indonesia,

however (the Javanese) only represent 45% of the population. Nonetheless, Indonesia is clearly

close to our somewhat arbitrary cut-off points for being intermediately diverse and having ethnic

dominance.

More importantly, the case-study of Indonesia’s historical civil conflicts in Ross (2005)

explains that (a) ethnic dominance is an important source of conflict in Indonesia. Moreover, (b)

the Javanese are often grouped with the second-largest group in the country, which is the

Sundanese (Ross 2005, 37):

60

‘Indonesia’s ethnic composition poses a civil war risk, however, because of the dominance of the

largest “ethnic” group, the Javanese. In 1976, the ethnic Javanese constituted 45 percent of the

population; the Sundanese, who are often grouped with the Javanese because they, like the

Javanese, are concentrated on the Island of Java, constituted another 15 percent of the population.

Whether they are treated as 45 percent or 60 percent of the population, the size of this group has

often contributed to antagonism between Indonesians who are indigenous to Java, and those from

other islands. Non-Javanese people see Indonesia’s government and military as Javanese-

controlled.’

If we group the Javanese and Sundanese together instead of separating the groups like in

Fearon’s original (2003) classification, Indonesia’s ethnic fractionalization index falls to 0.63.

This index puts it well within the 0.25-0.68 range which defines our intermediately diverse

countries. Since the largest group ethnic group, which now contains both the Javanese and the

Sundanese, now contains 60% of the population, Indonesia also satisfies the ethnic dominance

definition.

On the empirical side, if we believe that Indonesia resembles a country with intermediate

diversity and ethnic dominance, we should also expect its conflict intensity to follow the model

for the intermediately diverse rather than the non-intermediately diverse countries. Another

reason we might expect such a response is that the conflicts in Indonesia in our dataset pitted the

central government in Java against ethnic secessionists in East Timor, Aceh, and West Papua.

All three areas have natural resources that may be relatively easy to appropriate due to their

geographic concentration, and whose extraction process is likely to be capital-intensive and may

create relatively few employment opportunities that raise the opportunity cost of rebelling. They

include oil in East Timor (Dubois 2000; Le Billon 2007), oil and natural gas in Aceh (Robinson

1998; Dubois, 2000), and timber and minerals in West Papua (Heidbüchel 2007). As a result, we

61

should again expect that positive terms of trade shocks increase conflict in Indonesia rather than

decrease it like in the other non-intermediately diverse countries.

In order to test this idea, Table A4, Column (1) reports the regressions results for the non-

intermediately diverse countries when we allow terms of trade shocks in Indonesia to have

different effects than in the other countries. The results show that we can reject that terms of

trade growth has the same effect in Indonesia as in the other non-intermediately diverse

countries. Moreover, the -24.3 coefficient for the remaining non-intermediately diverse countries

is substantially larger in magnitude than the original Table 4, Column (4) estimate of -15.6.

In Column (2) we add Indonesia to the intermediately diverse sample and test whether, as

we should expect if Indonesia is effectively intermediately diverse, both positive and negative

terms of trade shocks increase its conflict intensity. The results support both hypotheses. In

Columns (3)-(4), we repeat the analysis when we split the sample according to the presence or

absence of ethnic dominance. The results in Colum (3) support that Indonesia responds

differently to terms of trade shocks than the other countries without a dominant group.

Controlling for the differential Indonesia response increases the coefficient magnitude for the

remaining non-dominance countries to -24.8 from -14.7 in Table 5, Column (4). Column (4)

suggests that both positive and negative terms of trade shocks increase conflict intensity in

Indonesia, so its shock responses resemble the responses of the original ethnic dominance

countries. On this basis, we believe that it may be reasonable to either control for the differential

Indonesia response in the main paper or to omit it from the non-intermediately diverse and non-

dominance samples.

62

TABLE A4

The effects of terms of trade growth in Indonesia

(1) (2) (3) (4)

Estimation Method LSDV LSDV LSDV LSDV

Sample NID ID

+Indonesia No ethnic dominance

Ethnic dom. +Indonesia

∆ΤΟΤ(t) -27.268** -24.768**

[10.172] [9.934]

∆ΤΟΤ(t)*Indonesia 42.678*** 37.392***

[13.622] [13.702]

Pos∆ΤΟΤ(t) 19.267** 17.891**

[7.684] [7.801]

Neg∆ΤΟΤ(t) -34.975*** -34.720***

[7.302] [8.062]

Pos∆ΤΟΤ(t))*Indonesia 30.218* 27.834*

[14.975] [15.967]

Neg∆ΤΟΤ(t))*Indonesia -76.787*** -75.869***

[14.575] [16.661]

Duration 0.002 -0.050 -0.001 -0.050

[0.011] [0.038] [0.013] [0.039]

First year -0.899*** -1.101*** -0.967*** -1.054***

[0.215] [0.242] [0.210] [0.250]

Observations 491 436 510 410

R-squared 0.617 0.515 0.626 0.502

Number of countries 39 41 41 38

p-val (Pos∆ΤΟΤ+Pos∆ΤΟΤ*Indon.) 0.02 0.01

p-val (Neg∆ΤΟΤ+Neg∆ΤΟΤ*Indon.) 0.00 0.00

Year dummies Y Y Y Y

Country time trends Y Y Y Y

Note: Robust standard errors clustered at the country-level in brackets. * significant at 10%; ** significant at 5%; *** significant at 1%. ∆ denotes the change in the three-year moving average. Column (1) replicates the Table 4, Column (3) regression model but allows the effect of terms of trade growth to differ in Indonesia compared to the remaining non-intermediately diverse countries. Column (2) reports the estimates when we add Indonesia to the intermediately diverse sample and estimate the separate effects of positive and negative terms of trade shocks. Columns (3) repeat the analysis but divide the sample into the countries without and with ethnic dominance.

63

References (uncited in the main paper)

Dubois, Brian. 2000. ‘The Timor Gap Treaty – where to now?’ Briefing Paper no. 25. Oxford,

UK: Oxfam Community Aid Abroad. Based on initial research by Monique Hanley and Kirsty

Miller.

Heidbüchel, E. (2007). The West Papua Conflict in Indonesia: Actors, Issues and Approaches.

Wettenberg: Johannes Herrmann Verlag.

Le Billon, P. (2007). ‘Geographies of war: perspectives on ‘resource wars’. Geography Compass

1 (2): 163-182.

Robinson, G. (1998). ‘Rawan is as Rawan Does: The Origins of Disorder in New Order Aceh.’

Indonesia 66: 126-57.

64

Appendix F

The role of ethnicity in the conflict observations for NID fuel exporters

In Section 5 in the main paper, we argue that most of the conflict observations for the fuel

exporters outside the intermediate-ethnic-fractionalization range (the 25th-75th percentile of the

Herfindahl-Hirschman ethnic fractionalization index) nonetheless included large or dominant

ethnic groups. Particularly, 91% or 89 of the total of 98 observations for NID fuel-exporters in

the Table 8, Column (1) regression come from just four countries - Angola (1975-2002, 2004,

2007), Azerbaijan (1992-95, 2005), Indonesia (1975-92, 1997-2005) and Sudan (1976, 1983-

2008) - that have a history of ethnic conflict involving large or dominant ethnic groups. The

Angola observations reflect the 1975-2005 postcolonial war to control the government and, from

1991, the Cabinda secession war. The Azerbaijan observations come from the Ngorno-Karabakh

conflict and two aborted coups in 1993 and 1995. The Indonesia observations come from a

mixture of the Acehnese, East Timorese, and West Papuan conflicts. The Sudan observations

come from the Islamic Charter Front coup in 1976, the Second Sudanese Civil War (for 1983-

2008), and the Western Darfur rebellion (2003-08). Below, we argue that most of these conflict

episodes can be characterized as ethnic conflicts and were either partly motivated or partly

financed by fossil fuels.

The main Angolan civil war from 1975-2002 started as a conflict between the largest

three ethnic groups over controlling the central government upon independence from Portugal.

During the war, the country’s oil revenues helped to finance the Angolan government’s war

effort against the UNITA rebels (Le Billon 2000; Bannon and Collier 2003; Le Billon 2000). In

addition, oil is an important factor in the ongoing Cabinda conflict, which we observe from 1991.

This conflict represents the attempt of the mainly ethnic Bakongo-inhabited, oil-rich, and

65

geographically separated Cabinda province in the north to secede (Porto 2003; Minorieties at

Risk 2017).

Although the historical roots of Sudan’s conflict go beyond the discovery of oil, the

discovery of oil reserves in the south and the perception that the northern government displaced

thousands of people to get access to the oil fields contributed to the Second Civil War from

1983-2008 (Johnson 2003; Collins 2005). Oil also helped to finance the Sudanese government

during the Western Darfur rebellion from 2003-08 (Patey 2010).

Indonesia’s conflict years in the sample are a combination of three distinct ethnic

conflicts that all reflected that the Javanese-dominated central government tried to establish

greater control of resource-rich peripheral areas where the ethnic minorities wanted to secede

(see Appendix D).

Azerbaijan’s 1992-94 and 2005 observations reflected that the mainly Armenian-

inhabited Ngorno-Karabakh region attempted to secede to Armenia. Although Ngorno-Karabakh

is not known to possess fossil fuels, oil extraction appears to have helped to finance Azerbaijan’s

war effort (Kaldor 2007, 163):

‘Because of the collapse of the official [Soviet] economy and because, in any case, taxation had

been centralised in the Soviet era, there was almost no official funding. On the Armenian side,

funding was almost entirely war related – diaspora support, Russian military assistance, loot and

pillage, contraband trade (especially petroleum products) and hostagetaking…On the Azeri side,

the government was able to commandeer crude oil from the Azerbaijan State Oil Company

(SOCAR) either for use at the front or for sale...’

66

Related, our 1995 observation for Azerbaijan is a coup attempt that Cornell (1999) argues can

most convincingly be explained be Russia’s desire to control the country’s oil (Cornell 1999,

57):

‘Moscow saw its control over Azerbaijan slipping away with the oil deal [that the

Azerbaijan’s state oil company had just re-negotiated with a western oil companies] and

therefore triggered a crisis that would bring its ally to power.’

References (uncited in the main paper)

Collins, R.O. (2005). ‘Civil Wars and Revolution in Sudan.’ Hollywood, CA: Tsehai Publishers

and Distributors.

Cornell, S. E. (1999). The Nagorno-Karabakh Conflict. Inst. för Östeuropastudier.

Johnson, D. H. (2003). The root causes of Sudan's civil wars. Vol. 601. Bloomington, IN:

Indiana University Press.

Kaldor, M. (2007). ‘Oil and conflict: the case of Nagorno Karabakh.’ In Mary Kaldor, Terry

Lynn Karl and Yahia Said (eds.) Oil Wars. London: Pluto Press,157-182.

Le Billon, P. (2000). ‘Angola's political economy of war: The role of oil and diamonds, 1975–

2000.’ African Affairs 100 (398): 55-80.

67

Minorities at Risk. (2017). Assessment for Cabinda in Angola. Accessed March 25, 2017,

http://www.mar.umd.edu/assessment.asp?groupId=54003

Patey, L. A. (2010). ‘Crude days ahead? Oil and the resource curse in Sudan.’ African Affairs

109 (437): 617-636.

Porto, J. G. (2003). Cabinda: Notes on a Soon-to-be-forgotten War. Institute for Security Studies

ISS Paper 77, August

Appendix G

Partial residual plots for the Table 9, Column (1) and (5) regressions.

In order to ensure that outlier effects do not explain the paper’s Table 9, Columns (1) and (5)

results, Figures A2-A3 plot the partial residuals for the positive and negative fossil fuel terms of

trade shocks in the two regressions.

68

FIGURE A2

Partial residual plots for the Table 9, Column (1) regression: the effects of positive and negative

fuel terms of trade shocks in intermediately ethnically diverse fuel exporters. Herfindahl-

Hirschman ethnic fractionalization index in the 25-75th percentile.

76Colombia

75Colombia

05Colombia

79Iraq

73Iraq78Iraq74Colombia

06Colombia

08Colombia

92Iran

81Colombia

93Iran

83Iraq

99Iran

75Malaysia

88Colombia

91Iran

03Russia77Iraq

85Iraq04Iraq

84Iraq

95Russia

87Colombia01Colombia

96Iran

07Iraq

94Russia00Iran

90Iran

82Iraq

86Iraq

80Iraq

97Iran

89Iraq

05Russia

92Iraq

96Russia

02Russia

75Oman73Oman

90Colombia

86Colombia

83Iran

87Iraq

00Russia

07Iran

01Russia

84Iran

95Iraq

96Mexico92Venezuela90Trinidad &Tobago79Saudi Arabia98Colombia82Venezuela94Mexico

06Russia

96Colombia08Iraq

07Colombia

93Iraq

91Iraq

74Iraq

90Iraq

81Malaysia

80Iran

85Iran

91Russia

82Colombia

04Russia02Colombia

82Iran

80Colombia94Iraq

94Colombia

99Russia89Colombia

97Colombia

74Oman

81Iran

93Colombia

04Colombia

88Iraq

07Russia

90Russia

85Colombia

00Colombia88Iran

79Iran

08Russia

06Iran

86Iran01Iran

95Colombia93Russia

81Iraq

08Iran

84Colombia

87Iran

91Colombia06Iraq

96Iraq

99Colombia

77Colombia

74Malaysia

05Iraq

03Colombia79Colombia

05Iran

83Colombia

78Colombia92Colombia73Colombia

76Iraq

75Iraq

-2-1

01

23

e( lNewtotbdeadbes_impute | X )

-.02 -.01 0 .01 .02 .03e( posdy3mactotfuel | X )

coef = 28.687887, (robust) se = 14.622514, t = 1.96

88Iraq

87Iraq

92Colombia

86Iraq91Colombia

77Colombia

83Colombia

87Iran

82Colombia

88Iran

90Colombia

75Iraq

79Colombia

86Iran

96Colombia80Colombia

95Colombia

00Colombia

74Iraq

81Colombia

85Iraq97Colombia

84Iraq93Iran

06Iraq

78Colombia

94Iraq

73Colombia

05Iraq08Russia

08Iraq

03Colombia

05Iran

07Colombia74Oman04Colombia

89Iraq

01Colombia

99Russia

96Iran

91Russia

07Russia

02Colombia

93Russia

76Colombia

85Iran

74Malaysia

01Russia06Iran

90Iran

00Russia

96Russia

81Malaysia

04Russia

96Mexico92Venezuela90Trinidad &Tobago98Colombia79Saudi Arabia82Venezuela94Mexico

99Iran

81Iraq93Iraq

95Russia

94Russia

75Malaysia

08Iran

76Iraq

90Russia

02Russia

73Oman75Oman07Iraq

91Iran

84Iran

99Colombia

05Russia

06Russia

89Colombia

04Iraq

07Iran

73Iraq92Iran

01Iran

94Colombia

80Iraq06Colombia

03Russia

79Iran

84Colombia78Iraq

05Colombia

81Iran97Iran

82Iran93Colombia

08Colombia

83Iran

85Colombia

80Iran

00Iran95Iraq

82Iraq

79Iraq

83Iraq

74Colombia75Colombia90Iraq96Iraq

77Iraq

91Iraq

92Iraq

86Colombia

87Colombia

88Colombia

-2-1

01

23

e( lNewtotbdeadbes_impute | X )

-.02 -.01 0 .01 .02 .03e( negdy3mactotfuel | X )

coef = -38.466011, (robust) se = 2.8476158, t = -13.51

69

FIGURE A3 Partial residual plots for the Table 9, Column (5) regression: the effects of positive and negative

fuel terms of trade shocks in intermediately ethnically diverse fuel exporters. Herfindahl-

Hirschman ethnic fractionalization index is in the 15th - 85

th percentile and large producers

omitted.

95Azerbaijan

76Sudan

75Colombia

05Sudan

76Colombia

98Indonesia

05Colombia

01Sudan

83Indonesia

99Indonesia

03Indonesia

79Indonesia

04Angola

78Indonesia

92Indonesia

01Colombia

84Indonesia85Indonesia

93Angola

89Indonesia

07Angola93Colombia

81Colombia88Sudan02Colombia

02Sudan92Angola

80Indonesia

93Sudan

82Indonesia75Angola

00Colombia

77Indonesia

91Indonesia

00Sudan

97Colombia

83Angola

06Sudan

94Angola

85Angola

96Colombia

97Indonesia

87Sudan

87Colombia

91Angola86Indonesia

86Angola81Angola

88Colombia

89Angola

97Sudan

77Angola84Angola

80Angola08Sudan

90Indonesia

90Angola

87Angola75Malaysia74Colombia94Colombia

90Trinidad &Tobago73Colombia81Malaysia90Colombia

74Malaysia

90Sudan08Colombia00Indonesia

94Sudan

86Colombia

82Angola

96Angola

91Colombia

78Angola

04Colombia

96Sudan

79Angola

04Sudan

94Azerbaijan

88Angola

86Sudan

06Colombia

99Angola

87Indonesia

92Colombia

07Sudan

98Angola82Colombia

91Sudan

02Indonesia

89Sudan

07Colombia

77Colombia

85Colombia

99Colombia

88Indonesia80Colombia

97Angola

89Colombia

76Angola

84Sudan

02Angola

00Angola

03Colombia

05Azerbaijan

98Colombia84Colombia85Sudan

92Azerbaijan

81Indonesia92Sudan

83Colombia

95Angola

99Sudan

01Angola98Sudan

03Sudan

83Sudan

04Indonesia

95Colombia

78Colombia95Sudan

79Colombia93Azerbaijan

01Indonesia

75Indonesia

05Indonesia76Indonesia

-3-2

-10

12

e( lNewtotbdeadbes_impute | X )

-.005 0 .005 .01e( posdy3mactotfuel | X )

coef = 119.31958, (robust) se = 39.467298, t = 3.02

93Azerbaijan

88Indonesia

87Indonesia

94Azerbaijan86Indonesia

92Sudan

83Sudan

92Colombia

95Sudan

79Colombia

95Colombia

05Indonesia91Sudan90Sudan

92Angola

80Colombia

75Indonesia82Colombia77Colombia

07Angola

88Angola

03Sudan

04Angola

95Angola98Sudan

87Angola91Colombia

96Sudan

90Colombia

76Indonesia

78Colombia

03Colombia

81Colombia

76Sudan83Colombia01Indonesia

86Angola

97Sudan

98Colombia

00Sudan85Indonesia02Sudan

00Colombia

84Sudan

75Angola

02Colombia74Malaysia

06Colombia77Angola79Angola

89Sudan

82Angola

89Colombia

96Colombia

78Angola

99Sudan

97Indonesia

99Colombia

08Colombia85Sudan81Malaysia

97Colombia

01Angola80Angola73Colombia90Trinidad &Tobago

81Angola85Angola84Colombia

04Sudan

84Angola

08Sudan

99Indonesia

99Angola

89Angola

04Colombia

01Colombia

00Angola

89Indonesia

76Angola

06Sudan84Indonesia

02Angola74Colombia98Angola75Malaysia

05Azerbaijan90Angola91Angola05Colombia07Colombia85Colombia

02Indonesia

01Sudan

00Indonesia

97Angola

92Indonesia

07Sudan81Indonesia

04Indonesia

83Angola

78Indonesia

05Sudan

96Angola

98Indonesia

76Colombia

75Colombia80Indonesia

94Angola

82Indonesia

77Indonesia03Indonesia

90Indonesia

93Angola

79Indonesia

94Colombia

91Indonesia

86Colombia

94Sudan

86Sudan

93Sudan

93Colombia

87Colombia

83Indonesia

88Colombia

87Sudan

88Sudan

95Azerbaijan

92Azerbaijan

-3-2

-10

12

e( lNewtotbdeadbes_impute | X )

-.005 0 .005 .01e( negdy3mactotfuel | X )

coef = -109.38375, (robust) se = 29.820848, t = -3.67