Primary Commodity Exports and Civil War
Author(s): James D. Fearon
Source: The Journal of Conflict Resolution, Vol. 49, No. 4, Paradigm in Distress? Primary
Commodities and Civil War (Aug., 2005), pp. 483-507
Published by: Sage Publications, Inc.
Stable URL: http://www.jstor.org/stable/30045128 .
Accessed: 24/10/2013 17:09
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].
.
Sage Publications, Inc. is collaborating with JSTOR to digitize, preserve and extend access to The Journal of
Conflict Resolution.
http://www.jstor.org
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
PrimaryCommodityExportsand Civil War
JAMES D. FEARON
Departmentof Political Science
StanfordUniversity
Collier and Hoefflerreportedthatcountrieswith a higherpercentageof nationalincome fromprimary
commodityexportshavebeen moreproneto civil war,aninterestingfindingthathasreceivedmuchattention
frompolicy makersandthe media.The authorshows thatthis resultis quite fragile,even using Collierand
Hoeffler'sdata.Minorchangesin the sampleframingandthe recoveryof missing dataundermineit. To the
extentthatthereis anassociation,it is likely becauseoil is a majorcomponentof primarycommodityexports
andsubstantialoil productiondoes associatewith civil warrisk.The authorarguesthatoil predictscivil war
risknot becauseit providesan easy sourceof rebel start-upfinancebutprobablybecauseoil producershave
relativelylow statecapabilitiesgiven theirlevel of percapitaincome andbecauseoil makesstateor regional
control a tempting"prize."An analysis of data on governmentobservanceof contractsand investor-perceived expropriationrisk is consistentwith this hypothesis.
Keywords: civil war; naturalresources;oil; rebelfinance
Paul CollierandAnkeHoeffler'spaper"GreedandGrievancein CivilWar"hashada
majorimpacton bothpublicdebateand social science researchon contemporarycivil
wars. First posted to a World Bank Web site in early 2000,1 Collier and Hoeffler undertook a cross-national statistical analysis of civil war onset in 161 countries since 1960
and found that "the extent of primary commodity exports is the strongest single influence on the risk of conflict" (Collier and Hoeffler 2000). By way of explanation, they
argue that primary commodity dependence creates better opportunities to finance
rebel groups and so enables rebellion.
This finding received widespread press coverage, garnering articles and editorial
comment in The Economist, the New YorkTimes, the Washington Post, and the Financial Times, among many other newspapers and magazines. Indeed, the study's main
1. The earliestdraftI have is datedApril 26, 2000. Severalsubsequentversions were posted to the
Bank's "Economicsof Civil War,Crime, and Violence" Web site (http://econ.worldbank.org/programs/
conflict), andthe publishedversionis CollierandHoeffler(2004). The core statisticalmodel for the "Greed
andGrievance"paperis presentedin CollierandHoeffler(2002b), andthe dataused forthatarticleareavailable at Hoeffler's Web site http://users.ox.ac.uk/~ball0144/.
Collier et al. (2003) reportsome of the main
resultsof the "Greedand Grievance"paperalong with manyothersin book form.
AUTHOR'SNOTE:I wish to thankNils PetterGleditsch,SteveKrasner,DavidLaitin,Piivi Lujala,and
JamesRon fortheircomments;andPaulCollierandAnke Hoefflerfor helpfuldiscussions.The dataused in
this article will be posted at http://www.stanford.edu/~jfearon/
and also at http://www.yale.edu/unsy/jcr/
jcrdata.htm/.
JOURNAL OF CONFLICT RESOLUTION, Vol. 49 No. 4, August 2005 483-507
DOI: 10.1177/0022002705277544
© 2005 Sage Publications
483
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
484
JOURNALOF CONFLICTRESOLUTION
findingandthe authors'interpretationof it may be the most widely reportedresultof
anycross-nationalstatisticalstudyof civil conflict,ever.2As TheNew YorkTimessummarizedthe first version of the paper,the Bank
found thatthe single biggest risk factorfor the outbreakof warwas a nation'seconomic
dependenceon commodities.Eagernessto profitfrom coffee, narcotics,diamondsand
othergemstonesboth promptsoutbreaksof violence and determinestheirstrengthover
time, says the study."Diamondsare the guerrilla'sbest friend,"said Paul Collier, the
authorof the studyanddirectorof researchat the Washington-basedWorldBank'seconomics department."Civilwarsarefarmorelikely to be causedby economic opportunities thanby grievance."3
Collier andthe mediadrewmajorpolicy implicationsfromthe finding.Forexample, "To reduce the occurrenceof these wars, [Collier] suggests, countries should
diversifytheireconomies and visibly channelcommodityincome into social service
programs,to underminepublic supportfor rebelswho seize the mines andfarmlands.
The world community,meanwhile, can help by refusing to do business with rebel
groups."4In December 2002, the New YorkTimes listed policy implications of the Collier and Hoeffler paper in its "The Year in Ideas" section: "Throughout Africa and in
parts of Asia and Latin America, guerrillas finance their armies through the illegal
export of commodities: timber, diamonds, oil and coca. Policy makers are now trying
to encourage guerrillas to give up their fight by cutting off the money spigot. An
embargo, they believe, can put rebels out of business or drive them to the negotiating
table."5Begun in May 2000 and endorsed by the United Nations, the Kimberley Process to end trade in "conflict diamonds" is an important policy initiative consistent
with Collier and Hoeffler's argument.6
For academic research agendas, Collier and Hoeffler's work has brought the question of rebel financing to the fore, a highly valuable contribution. In the literature on
"contentious politics" and social movements, the idea that the "opportunity structures" facing would-be rebels are at least as important for explaining rebellion as relative deprivation or social grievances has been common for some time (Eisinger 1973;
2. See, for examples,SathnamSanghera,"RebelsFight for Loot Not Causes,"FinancialTimes,June
16, 2000, Londoned., p. 12; John Burgess, "CivilWarsLinkedto CertainEconomies;WorldBank Cites
Commodities'Role,"WashingtonPost, June 16, 2000, final ed., p. E03; JosephKahn,"WorldBankBlames
Diamondsand Drugs for ManyWars,"New YorkTimesJune 16, 2000, late ed., p. A14; G. PascalZachary,
"MarketForces Add Ammunitionto Civil Wars-Research Suggests Rebels Have 'Greed'as Motive;PrimaryExportsCount,"WallStreetJournal,June 12, 2000, Easterned., p. A21; MartinWolf, "HowCivil War
PlaguesthePoor,"FinancialTimes,December27, 2000, Londoned., p. 13;PaulCullen,"MoneyIs Still atthe
Rootof WarWorldwide,"IrishTimes,May 11,2001, p. 55; SebastianMallaby,"AJ.FK. Approachto Terrorism,"WashingtonPost,November26, 2001, finaled., p. A25; TinaRosenberg,"ToPreventConflicts,Lookto
CommoditiesLike Diamonds,"New YorkTimes,July 15, 2002, late ed., p. A6; TinaRosenberg,"TheYearin
Ideas;Peace throughEmbargo,"New YorkTimes,December 15, 2002, late ed., sec. 6, p. 108; "TheGlobal
Menace of Local Strife-Civil Wars,"TheEconomist,May 24, 2003, U.S. ed., p. 23; and DaphneEviatar,
"Howto Save Africa,"Newsweek,September22, 2003, Atlanticed., p. 44.
3. Kahn,"WorldBank Blames Diamondsand Drugs for ManyWars."
4. John Burgess, "CivilWarsLinkedto CertainEconomies."
5. TinaRosenberg,"TheYearin Ideas;Peace throughEmbargo,"citing Collier andHoeffler'sstudy
directly.
6. See www.kimberleyprocess.com.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 485
McCarthyandZald 1977;Tilly 1978). But "opportunitystructures"wereneverclearly
specifiedfor empiricalexaminationin the case of civil war.Collier andHoefflertook
an importantstep forwardby proposingto examinerebel organizationsas businesses
of a sort and to measurepossible determinantsof theirfinancialviability.7
In this article, I use the data from Collier and Hoeffler (2002b) to reassess the
impactof primarycommodityexportson the probabilitythata countryhas a civil war.
othercrossDespite the claims made on behalf of this proxy for rebel "opportunity,"
nationalstatisticalstudieshave not noteda similarlystrongrelationship.In particular,
Fearonand Laitin (2003) find no supportfor an independenteffect of primarycommodityexportdependenceon civil waronset, despiteusing the same measureof commodity exportsas Collier and Hoeffler and a fairly similarlist of civil wars.
What accountsfor the differentresults?Because so many things can vary across
two statisticalstudies with "the same"dependentvariable-such as specifics of the
sample, model specification(which variablesare includedand how), measures,and
estimationmethods-it can be difficultto identifythe reasonsfor conflictingfindings.
My procedureis to startby replicatingthe main regressionresults in Collier and
Hoeffler(2002b). I then examinethe impactof changingminoraspectsof the sample
frameand the model specification,one aspect at a time.
I find that one does not have to departmuch from Collier and Hoeffler (2002b)
before primarycommodity exports cease to matterin statisticalterms. Collier and
Hoeffler group their data in five-year intervals,asking whethercharacteristicsof a
countryat the startof the five-yearperiodpredictwhethera civil warbeganduringthe
period.Since the dependentvariable,civil waronset,is measuredatleastannually,and
becausethe choice of five-yearperiodsis arbitrary,thereis a strongcase for using the
country-yearas the unit of observationratherthanthe country-five-year-period
(for
instance,Kenyain 1980 ratherthan Kenyafrom 1980 through1984). Even without
suchanargument,theresultson primarycommodityexportsshouldnot dependon this
choice of sampleframing.Butwhenthe dataareanalyzedin a country-yearformat,the
apparentimpactof primarycommodityexportslargely evaporates.
There appearto be two main reasons. First, the country-yearformat makes lag
times for severalindependentvariablesmore consistentand allows more consistent
treatmentof quickly renewed wars (see below). Second, the country-yearformatis
less subjectto "list-wise deletion"of civil war onsets due to missing data.Although
thereare seventy-ninecivil war startsin Collier and Hoeffler'scivil war list, twentyseven of these, or aboutone-third,arenot used in theiranalysisdue to missing dataon
an independentvariable.WhenI reformatthe data,sixteenwarstartsreturnto the estimationsample,mainlybecause thereareless missing dataon prioreconomic growth
rateswhen I use countryyearsas observations.Withthese warsin the sample,primary
commodityexportsno longersignificantlyimprovethe model's abilityto predictcivil
7. The "greedversusgrievance"contrasthas also frameddiscussionat a numberof academic/policy
makerconferenceson civil war,for example,the InternationalPeace Academy-sponsoredconferencethat
producedBerdalandMalone(2000) andSherman(2001). The BerdalandMaloneproject,whichincludesan
earlierpaperby Collier(1999), reflectsparallelbutmorecase-studybasedresearchby Keen(1998) andReno
(1998), among others.See also Ballentineand Sherman(2003).
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
486
JOURNALOF CONFLICTRESOLUTION
war onsets. In addition,the estimatesfor other independentvariableschange markedly. The effect of "socialfractionalization"disappears,8and the estimatedeffect for
primarycommoditiesbecomes smallerand more dubious when fractionalizationor
other statisticallyinsignificantvariablesare droppedfrom the model.
A more systematicway of assessing the impactof list-wise deletionon Collierand
Hoeffler'sestimatesis to use multipleimputation,a proceduredevelopedby Donald
Rubin (1987) thatallows one to use all the informationin a data set, includingcases
thathavesome missingdata.WithCollierandHoeffler'sfive-yearsampleframing,the
estimatesfor primarycommodityexportsweakenconsiderablywhen multipleimputation is used. They become insignificantby conventionalstandardswhen multiple
imputationis used in the country-yearframing.
Anotherfinding is thatthe data do not supportthe propositionthat civil war risk
increases as primarycommodity exports increase to about 35 percentof GDP and
declines thereafter.CollierandHoeffleruse such a parabolicspecification,notingthat
few countrieshavemorethan35 percentof GDPin primarycommodityexports,so the
relationshipbetween civil war risk and commoditydependenceis mainly positive. I
use nonparametricmethodsto examinewhatthe datathemselves"say"aboutthe functional form of the relationshipand find almost no supportfor an increasing-thendecreasingpattern.Rather,it looks as if primarycommodityexportsincreasecivil war
risk at a decreasingrate, so that using the logarithmis betterjustified than fitting a
upside-down U. This also requires fewer contortionsin the theoretical argument
advancedto rationalizethe impactof primarycommodityexports.Nonetheless, the
log of primarycommodity exports remains an insignificantinfluence on civil war
outbreakin the country-yearmodel.
Whatshouldone makeof all this?Despite the lack of a sharprelationshipbetween
primarycommodityexportsandcivil waroutbreak,I believe thatthe maintheoretical
claim thatCollier and Hoeffler sought to supportis most likely correct.Betterrebel
fundingopportunitiesprobablydo imply a greaterriskof civil war,otherthingsequal.
But thereis little reasonto expect thatprimarycommodityexportsas a percentageof
GDP are a good measureof rebel financingpotential.
Contraryto the implication of the World Bank's press releases summarizing
"Greedand Grievance,"the WorldBank measureused by Collier and Hoeffler does
not include diamondsand othergems, and of course it does not reflect drugproduction. Instead,the measureseems to pick up mainly cash crops (like coffee or wheat)
and oil exports.
Largeprofitsfromcash cropsor oil requirecontrolof a nationaldistributionor production system, which rebels lack.9 To profit substantiallyfrom these sources of
income, rebels must turnto smaller-timeextortion,or to what Michael Ross (2004)
calls "bootyfutures"(rebelsselling futureresourceexploitationrightsto foreigncompanies or states).
8. A measureof the combinationof ethnic and religiousheterogeneity,discussed below.
9. On this point, see also Snyderand Bhavnani(2005 [this issue]).
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 487
Small-timeextortion(from the rebels' perspective,taxation)of local agricultural
producerscertainlydoes occur.But my own readingof the case literature,for whatit is
worth,does not suggest that cash crops or oil productionare criticalfor this.10And
looking at a set of thirteen"mostlikely cases,"Ross (2004) finds only one in which
booty futuresin oil helped finance a rebel group's start-upcosts (Congo Republicin
1997); it seems unlikely thatthis mechanismis very common.1"There is, then, little
reasonto thinkthatcash crops or oil productionaregood measuresof the availability
of rebel financing. In addition,at least one theoreticalconsiderationcuts the other
way: primarycommodityexportsprovidegovernmentswith a relativelyeasy source
of tax revenue,which may counterbalanceor offset increasedextortionpossibilities
for rebels.
I arguethatan empiricallymore plausibleand internallyconsistentexplanationis
thatoil exportersare more proneto civil war because they tend to have weakerstate
institutionsthanothercountrieswith the same per capitaincome (Fearonand Laitin
2003). Stateswith high oil revenueshaveless incentiveto developadministrativecompetence and control throughouttheir territory.So while oil revenues help a state
againstinsurgentsby providingmorefinancialresources,comparedto othercountries
with the samepercapitaincomethey shouldtendto havemarkedlyless administrative
and bureaucraticcapacity.Furthermore,easy riches from oil make the state a more
temptingprize relativeto workingin the regulareconomy.
Empirically,I show thatit is difficultto distinguishthe apparentimpactof primary
commodityexportsfrom oil exportsandthatoil exportsare somewhatmore strongly
relatedto civil war risk than other primarycommodity exports in the country-year
framing.A similar theoreticalargumentmay apply for nonfuel primarycommodities-heavy dependenceon taxationof primarycommoditiesat the bordermay associate with weakerstate institutions,on average.Using a measureof investorperceptions of the risk of different governments'reneging on contracts, I find that oil
producersareindeedseen as havingless reliablestateinstitutionsgiven theirincomes,
while the effect of otherprimarycommoditiesis marginalbut in the right direction.
The articleis structuredas follows. In the second section, I introducethe CollierHoefflerdataset (fromCollier and Hoeffler2002b) and its measureof primarycommodity exports,known in short as sxp. The thirdsection comparesthe results with
five-yearperiodsto countryyearsandundertakesseveralotherrobustnesschecks. The
fourthsection considers results when missing values are multiply imputed,and the
fifth looks at the effect of controllingfor oil exportsin additionto primarycommodities in general.The sixth sectiondevelopsthe alternativehypothesismentionedabove:
the (weak) relationshipbetweenprimarycommodityexportsandcivil waronset may
be due to primarycommodityexportsbeing a noisy measureof low state capability
given income, especially via fuel exports. A brief conclusion considers policy
implications.
10.In noneof MichaelRoss' (2004) thirteen"mostlikely"cases was thereevidencethatrebel"looting"
based on legal agriculturalcommoditiesor oil helped financerebel start-upcosts.
11. Fora nuancedanalysisof the effect of oil on the Republicof Congoconflict, see EnglebertandRon
(2004).
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
488
JOURNALOF CONFLICTRESOLUTION
SXP AND COLLIER AND HOEFFLER'S DATASET
THE SAMPLE AND THE DEPENDENT VARIABLE
Collier and Hoeffler (2002b) employ a data set consisting of 161 countries
observedfor each of the eight five-yearperiodsbeginningin 1960-64 and ending in
1995-99. This makes for a "potentialsample"of 161 x 8 = 1,288 observations.The
161 countriesare 150 of the 152 countriesfrom the Penn WorldTables(PWT) economic database,12plus 11 additionalcountriesfor which CollierandHoefflercoded a
civil warbeginningat some point in 1960 to 1999. These lattercases do not appearin
the estimationsample (the cases with complete data used in the statisticalanalysis)
because they lack economic data. There is some reason to worry aboutthe implicit
selection rule here. Big civil wars cause states to have troublecollecting economic
data, which in turnslowers the odds thatthe countrywill appearin PWT.Thus, the
cases are partly selected on the dependentvariable,which will tend to bias effect
estimates.
The potential sample also contains 151 observationssuch as "Azerbaijan196064"--country five-yearperiodsduringwhich the "country"was not an independent
state.It is difficultto imaginea good rationalefor includingsuch territoriesseparately
from the states thathad sovereigntyover them at the time, and doing so raises questions aboutwhy includethese territoriesandnot others(for instance,why not California, or Chechnya?).Fortunately,the questionis largelymoot since all but 15 of these
cases aredroppedfromthe estimationsampledueto missingeconomicorotherdata.13
CollierandHoefflerwish to examinethe determinantsof civil waronset and,thus,
code a dependentvariablethatequals 1 if a civil war startedin the country-five-yearperiodin question.They drop from the sample country-periodsin which a civil war
was ongoing,providedthatno newwarstartedin thesamefive-yearperiod. Forexample, Afghanistan1990-94 is coded as having a civil war start,due to the post-Soviet
warfor Kabulthatbeganin 1992. If, however,this warhad not occurred,Collier and
Hoeffler would have coded Afghanistan 1990-94 as missing data, since the antiSoviet-backed-regimewarthatends in 1992 was ongoing duringpartof 1990-94, and
this is the rule they apply otherwisefor country-periodsin which a civil war is ongoing. For instance, the war in Cambodiais coded as ending in 1991, and Cambodia
1990-94 is coded as missing datadue to the ongoing warat the startof the period.This
procedureremoves from the potential sample the sixty-six countryperiods with a
somewhat durable post-war peace, while twelve country periods with a quickly
renewed(ornew) warareincluded.It is difficultto say whatkindof bias thisprocedure
might introduce,if any.14But it is worthnoting thatthe treatmentof periodswith an
12. East Germanyand Dominica were dropped.
13. The nonindependentterritoriesthatappearin the estimationsample are Cape Verde1965-74, Fiji
1965-69, Namibia 1965-89, PapuaNew Guinea 1965-74, Reunion1975-89, andSuriname1965-74. All fifteen of these cases arecoded as "atpeace,"althoughCorrelatesof War(COW)codes an "extrastate"
warthat
startedin 1975 in Namibia.Since other"extrastatewars"are included,this may be a coding error.
14. Droppingperiodsof ongoing war artificiallyincreasesthe mean of the dependentvariable(onset)
for countriesthathad a war,and especially for countrieswith multipleonsets. The experienceof countries
thathad at least one civil war onset may thus be overweightedin the logit analyses.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 489
ongoing civil waris not symmetricanddependson whetherpeace prevailedfor one to
five years afterthe war's conclusion.15
The estimationsample consists of only 750 out of the 1,288 countryperiodsthat
makeup thepotentialsample-42 percentof the "potentialsample"is lost due to "listwise deletion"of cases thataremissing dataon one or moreindependentvariables(31
percentif one considersonly postindependencecountryperiods).The main sourceis
missing economic data, and particularlythat needed to estimatethe averagegrowth
ratein per capitaincome in the previousfive-yearperiod (the period 1960-64 disappears from the estimation sample for this reason).16As Rubin (1987) and, more
recently in political science, King et al. (2001) have stressed, list-wise deletion is
always inefficient and causes biased estimates except in the unlikely event that the
missingness is "completelyat random."Twenty-sixof the 78 civil war onsets in the
potentialsample (33 percent)are missing cases, which is worrisomegiven that civil
war onsets arerelativelyrarein these dataand thus have a greaterimpacton the estimates than "zero"cases do.17Twenty-fourof the 26 cases lack data on economic
growthin the previousfive-yearperiod.
SXP
The main independentvariableof interestin "Greedand Grievance"is a World
Bank measureof primarycommodity exportsas proportionof GDP,known as sxp.
Measuredat five-yearintervalsstartingin 1960 andmissing for only 125 of the 1,288
potentialcases, sxp has a meanof. 16 anda medianof. 11, indicatingthe highly skewed
distributionshownin Figure 1.18Most of the variationin sxp (75 percent)is due to differentaveragelevels of primarycommodityexportsacrosscountries.The remaining
quarteris due to variationover time within countries.
The measurecovers exportsof goods classified in five categoriesof the Standard
InternationalTradeClassification(SITC),listedin Table1, which also gives examples
15. Why dropthe eighty-oddperiodswith an ongoing civil warat all? A naturalalternativeis to code
theseas zeros,since no new civil warstartedalthoughone couldhave.Concurrentcivil warsin a single countryappearin mostcivil warlists, thoughnotin the COW-basedlist usedby CollierandHoeffler.(Theywould
occurin this list if anticolonialwarswerecodedin the colonialempireratherthanin the colony.)Of course,if
one civil waris alreadyin progressin a country,the riskof anotherone startingwill probablybe lower.But if
so, why not estimatethis effect ratherthandropthe cases (as in FearonandLaitin2003)? This happensto be
impossiblewith the COW-basedlist usedby CollierandHoefflerbecause"civilwarin progress"wouldperfectly predictthe lack of onset of a new war,which causes a problemfor the maximum-likelihoodmethods
of logit or probit.So this is perhapsa pragmaticargumentfor CollierandHoeffler'scodingproceduregiven
the absence of concurrentwarsin theirwar list.
16. Note also thatthe five-yearperioddesign also implies thatthe "lagged"growthratecan be rather
farfromthe warstartin question,for instance,if the warstartsin 1994 the growthrateestimatecomes from
1985-89.
17. King andZeng (2001) note thatwith a "rareevent"such as a civil war onsets (relativeto periods
withoutonset), additionalevents aremore statisticallyinformativein thatthey have a biggerimpacton the
estimatedvariance-covariancematrix.
18. Six country-periodshaveansxp estimategreaterthan1 (Bahamas1960-80andBahrain1980).The
Bahamas'valuesareprobablya WorldBankcodingerror,as the valuesgo from2.16 in 1980 to 0.26 in 1985,
and stay below 0.26 afterwards;possibly before 1985 the decimalpoint is one space too farto the right.In
any event, all six cases are droppedfrom the estimationsample due to missing dataon othervariables.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
490
JOURNALOF CONFLICTRESOLUTION
500
Frequency
200
0
0.2
0.0
0.6
0.4
0.8
1.0
sxp
250
Frequency
100
0
-6
-5
-4
-3
-2
-1
0
log(sxp)
TI
Figure 1: Histograms of sxp and Log(sxp)
of goods foundin each category.The measuredoes not includepreciousgems such as
diamonds (which are found in SITC 66, nonmetallicmineralmanufactures),or, of
necessity, contrabandnarcotics,both of which have been noted as sources of rebel
fundingin the case-studyliterature.
SXP AND FUEL EXPORTS
An oil-exportingcountry'sper capitaincome does not reflect its level of bureaucratic developmentas well as would the same income level for a nonoil country.
Fearon and Laitin (2003) argue that if per capita income is related to civil peace
becauseit is mainlya measureof statepolice andbureaucraticcapability,this implies
thatoil exportersshouldhavea highercivil warriskthanone wouldexpecton the basis
of income. Additionally,oil exportsare an easy sourceof huge revenuesfor whoever
controlsthe stateor the oil-producingregion (in contrastto, say, developingan effective income tax apparatus).This increasesthe valueof the "prize"of secession or state
control.19
FearonandLaitinfindthatcountrieswithmorethanone-thirdof theirexport
revenuesfromfuels haveroughlytwice as greatan annualriskof civil waronset (con19. Humphreys(2005 [this issue]) discusses a numberof otherhypotheticallypossible mechanisms.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
EXPORTSAND CIVILWAR 491
Fearon / PRIMARYCOMMODITY
TABLE 1
sxp Components
SITC 0
E.g.
SITC 1
E.g.
SITC 2
E.g.
SITC 3
E.g.
SITC4
E.g.
SITC 68
E.g.
Food and live animals
All foodstuffssuch as wheat, coffee, sugar,livestock
Beveragesand tobacco
Self-explanatory
Crudematerials,inedible, except fuels
Textiles,rubber,wood products
Mineralfuels, lubricants,and relatedmaterials
Oil, coal, naturalgas
Animal and vegetableoils, fats, and waxes
Self-explanatory
Nonferrousmetals
Silver,copper,nickel, aluminum,lead, tin
NOTE:SITC = StandardInternationalTradeClassification.
trolling for income and other covariates). Using new data on oil productionand
reserves(versusoil exportrevenues),Humphreys(2005 [this issue]) likewise finds a
stronglink. Similarly,CollierandHoeffler(2002, 12) note that"of the manypotential
disaggregationsof primarycommodityexportspermittedby [their]data,only one was
significantwhenintroducedintoourbaselineregression,namelyoil versusnon-oil."
Perhapsprimarycommodity exports appearto matternot because they indicate
rebelfinancingpotentialbutbecausethey correlatewith oil production,which marks
stateweaknessconditionalon income, or affectscivil warrisk via some othermechanism. Alternativelyor in addition,the samemechanismmay operatefor primarycommodity exports that Fearon and Laitin (2003) postulatefor oil. Conditionalon per
capitaincome, a countrythatderivesa greatershareof nationalincome fromprimary
commodityexportsmay have a relativelyweak stateapparatus.High dependenceon
primarycommodityexportsmay associatewith the thin "extractiveinstitutions"that
Acemoglu, Johnson,and Robinson(2001) argueexplainpoor economic growthperformance.Collier and Hoeffler (2002a) find thatsxp seems to mattereven when they
controlfor the fuel exportcomponentof sxp, an issue I reconsiderbelow.
The bivariaterelationshipbetweenprimarycommodityexportsas a percentageof
GDP andfuel exportsas a percentageof GDP is fairly strong.Forthe 632 cases in the
Collier and Hoeffler potentialsample with data on both variables,the correlationis
.79; see Figure2 for the scatterplot.20
As Figure2 suggests, this fairly strongcorrelation is dueto the factthatbothvariablesareskewedandthecountriesthatscorehighest
on bothmeasuresaremajoroil producers.However,the correlationremainsnontrivial
even if one restrictsattentionto countrieswith less than,say,40 percentof theirGDP
20. Iconstructedfuel exportsas a percentageof GDPby multiplyingfuel as a percentageof exportrevenues by exportsas a percentageof GDP,bothfromthe WorldBank'sWorldDevelopmentIndicators.Notice
thata numberof observationslie below the 45 degreeline in Figure2, which shouldbe impossiblesince fuel
exportsshouldbe a componentof totalprimarycommodityexports(sxp).Evidentlytherearesome discrepancies in the fuel exportrevenuedataused for sxp and for the publishedfuel exportmeasure.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
492
JOURNALOF CONFLICTRESOLUTION
0.8
0.6
Exp./GDP
0.4
Command
0.2
Primary
0.0
0.0
0.2
0.4
0.6
0.8
Fuel ExDorts/GDP
Figure 2:
Fuel Exports and sxp
fromfuel exports(thecorrelationis then .43). As I arguefurtherbelow, in so faras sxp
appears related to civil war onset, this may be because of its relationshipto oil
production.
PRIMARY COMMODITY EXPORTS AND CIVIL WAR ONSET
Model 1 in Table2 replicatesthe mainmodel in CollierandHoeffler(2002b). The
control variablesare the log of per capita income in the first year of the five-year
period;averageannualgrowthrate of income for the priorfive-yearperiod;"social
fractionalization,"
a combinationof two measuresof ethnicandreligiousheterogeneity;21"ethnic dominance,"a dummy variable marking states whose largest ethnic
group is between 45 and 90 percentof total population;peace years, the numberof
monthssince any priorcivil war ended (begunat 172 in 1962 for stateswith no prior
war,this being the numberof monthssince the end of WorldWarII); the log of total
populationin the first year of the five-yearperiod;and a 0 to 1 measureof the geographicdispersionof the populationwithinthe country,highervalues of which indi21. CollierandHoeffler(2002b, 26-27) constructthis measureby multiplyingtogethertwo 0-to-100
scales of ethnicandreligiousheterogeneity,andthenaddingthe maximumof the two scales. They arguefor
this functionalform on the groundsthatthey foundit to be (empirically)"superiorto variants"in an early
version of the "Greedand Grievance"paper.I divide by 100 so that the estimatedcoefficient is more
readable.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 493
TABLE2
Civil WarOnset and PrimaryCommodityExports(sxp)
in Two DifferentSampleFrames
Model 1:
Five-Year
Periods
sxp
2
sxp
16.773**
(.001)
-23.800*
(.018)
Log(sxp)
Log(income)
Growth
Peace years
Log(population)
Fractionalization
Ethnicdominance
Geo. dispersion
Constant
N
N (wars)
Model 2:
Five-Year
Periods
-0.950**
(.000)
-0.098*
(.018)
-0.004**
(.000)
0.510**
(.000)
-2.479**
(.007)
0.480
(.144)
-0.992
(.275)
-3.437
(.167)
750
52
Model 3:
Country
Years
8.225*
(.035)
-14.307
(.075)
1.033**
(.000)
-0.991**
(.000)
-0.102*
(.014)
-0.004**
(.000)
0.591**
(.000)
-2.629**
(.005)
0.513
(.120)
-1.120
(.216)
-0.473
(.842)
750
52
-0.535**
(.003)
-0.146**
(.000)
-0.004**
(.000)
0.311**
(.001)
-0.822
(.244)
0.396
(.122)
-0.167
(.820)
-2.502
(.097)
4,466
69
Likelihoodratiotest forjoint significanceof sxp and sxp2:
5.05
16.56
yf=2
.081
.000
Prob>
Model 4:
Country
Years
6.827
(.062)
-12.528
(.104)
Model 5:
Country
Years
6.154
(.091)
-11.531
(.132)
Model 6:
Country
Years
5.870
(.091)
-10.993
(.132)
-0.432**
(.005)
-0.147**
(.000)
-0.004**
(.000)
0.281**
(.001)
-0.409**
(.007)
-0.130**
(.000)
-0.004**
(.000)
0.253**
(.003)
-0.441**
(.002)
-0.132**
(.000)
-0.004**
(.000)
0.258**
(.001)
0.398
(.120)
-0.017
(.982)
-3.091*
(.028)
0.170
(.815)
-2.846*
(.039)
-2.567
(.052)
4,515
70
3.82
.148
4,605
70
3.15
.207
5,008
73
3.31
.192
NOTE:Logitestimateswithcivil waronset as the dependentvariable;observationsarecountryfive-yearperiodsin models 1 and2 andcountryyearsin models 3 through6. Incomeandpopulationarelagged;income
growthis for the priorfive-yearperiodor last five years.p-values are in parentheses.
*p < .05. **p < .01.
cate moreconcentratedsettlementpatterns.See CollierandHoefflerfor detailson the
sourcesand constructionof these variables,and argumentsfor why they shouldbe in
the model as predictorsof civil war onset.
The mainconcernhereis with the resultsfor the primarycommoditymeasures,sxp
and its square.A likelihoodratiotest confirmswhat the individualp-values suggest:
the odds that adding primarycommodity exports to the model in this mannerhas
improvedthe fit of the model by chance are estimatedat less thanone in a thousand.
The coefficients attenuatea bit more than 20 percentif one drops ethnic dominance
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
494
JOURNALOF CONFLICTRESOLUTION
TABLE3
Proportionof Five-YearPeriodswith Civil WarOnset
by Populationand CommodityExports(Treciles)
PrimaryCommodityExports/GDP
Population(millions)
Less than 3.5
3.5 to 12.2
More than 12.2
Total
Low
.043
69
.032
94
.057
174
.047
337
Medium
High
Total
.020
100
.081
135
.160
100
.087
335
.029
172
.090
100
.178
45
.069
317
.029
341
.070
329
.107
319
.068
989
andgeographicdispersionfrom the model, thoughstatisticalsignificanceremains.If
one has only sxp and its squarein the model and nothingelse, the estimatedrelationship appearsconsiderablyweaker,though still quite unlikely to be due to chance.
When the log of countrypopulationis added,the estimatedcoefficients on primary
commodityexportscome close to theirvaluesin model 1. This is becauselargercountries tend to have low commodityexportsas a percentageof GDP,but a greatercivil
warrisk. Table3 takesa less parametricapproach,showingthe proportionof country
five-year-periodsthathad a civil war start,by trecileson both countrypopulationand
primarycommodityexportsas a shareof GDP.It shows thatthe associationbetween
commodityexportsand civil war onset holds only for largercountries.
The impliedsubstantiveimpactof primarycommodityexportsbasedon model 1 is
considerable.Holdingothervariablesat theirmedianvalues,the estimatedprobability
of civil waronset rises from .011 at the 10thpercentileof sxp, to .031 at the median,to
.11 at the 90th percentile. The slight decline at high levels of sxp implied by the
squaredterm only begins aroundthe 90th percentileof sxp, which is 39 percentof
GDP from primarycommodityexports.
DO VERY HIGH LEVELS OF COMMODITY EXPORTS LOWER CIVIL WAR RISK?
AlthoughCollier and Hoeffler's parabolicspecificationof the impactof primary
commodity exports on the log odds of civil war produces statistically significant
results, so does using, for example, the log of sxp. Statisticianshave developedless
parametricmethodsthatenableone to examinewhatthe datathemselves"say"about
the form of the relationshipbetweencivil warrisk andprimarycommodityexports.22
Figure3 plotstheresultsof a generalizedadditivemodel(GAM)versionof Collierand
22. Conditional,of course,on the restof the model. See Beck andJackman(1998) for an introduction
writtenfor political scientists.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 495
10
changeinlog(odds)of onset
95% confidence interval
onset5
war
of
0
log(odds)
in
-5
change
0.0
0.2
0.4
0.6
0.8
Primary commodity exports/GDP (sxp)
Figure 3:
Change in Log(Odds) of Civil War as a Function of sxp
Hoeffler'score specification,in which I estimatethe form of the effect of sxp on the
log-odds of civil war directly.
Giventhe size of the 95 percentconfidenceintervalabove 50 percentof GDP from
primarycommodityexports,thereis scarcely any indicationof a parabolic,increasing-then-decreasingrelationship.Figure 3 also makes clear why adding squaredsxp
improvesthe fit of the model. A straightline fit to the datawould averageacross the
steep partwhere most of the cases are and the flat partwhere the small numberof
extremelyhigh sxp cases are.Figure3 suggeststhat,basedon the CollierandHoeffler
model, the probabilityof civil waris increasingin sxp up to about20 or 25 percentof
nationalincome and then essentially flat above this. So the actualrelationshipin the
datais betterapproximatedby the log of sxp. Model 2 in Table2 shows the relevant
results. A 1 percentincrease in primarycommodityexports in GDP is estimatedto
associate with a 1 percentincreasein the odds of civil war outbreak.23
CollierandHoefflersuggestedthatthe increasing-then-decreasing
patternwas due
to higher primarycommodity export shares initially favoringrebels but ultimately
favoring the state due to higher tax revenues.While one can propose a model that
makes functional form assumptionsto rationalizethis pattern,one can also easily
23. The overallfit of the model is betteras well, as judged by the AkaikeInformationCriterion.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
496
JOURNALOF CONFLICTRESOLUTION
design a model in which the two effects simply offset each other,yielding no correlation. If one thinksprimarycommodityexportsindicatean economy thatmore easily
supportsrebelfinancing,then it is easierto explaina logarithmicrelationship.Still, if
one also thinksthatprimarycommodityexportsincreaseresourcesavailableto governments, then one still needs an explanation for why this would not offset the
additionalsupportprovidedto rebels.
In whatfollows, I will considerboth CollierandHoeffler'sparabolicspecification
and the empiricallybetter supportedlogarithmicspecificationwhen looking for an
effect of primarycommodities.
COUNTRY YEARS VERSUS FIVE-YEAR PERIODS
Why choose five-yearperiodsstartingin 1960 ratherthan1961, or 1962 (etc.)?Are
the resultssensitiveto this arbitrarychoice? Civil waronset,the dependentvariable,is
measuredat least annuallyin this andmost othercivil warlists, as areincome, income
growth, population,and peace years. The available measures of social fractionalization,ethnicdominance,and geographicdispersionaretime invariantor close to it.
Thus, using countryyears as the unit of analysis would seem more appropriate.In
addition,a country-yearframingavoids the problemswith coding quickly renewed
warsdiscussedabove andallows consistentandbrief lag times for income per capita,
economic growth,population,and peace years.24
The one considerationthat arguablyfavors using five-year periods is that sxp is
measuredat five-yearintervalsbeginning in 1960. However,as noted above, threequartersof the variationin sxp is acrosscountries,and sxp in year t is well correlated
with sxp in yeart - 5 (r = .85). Thus,filling in the missingyearswith a linearinterpolation or spline is certainlya defensible approachto dealing with the missing data.
I constructeda country-yearversionof the CollierandHoefflerdata,interpolating
values for primarycommodityexports.25Model 3 in Table2 shows the resultsusing
the country-yearframingand the same set of independentvariablesas in Collier and
Hoeffler (2002b). The estimatedcoefficients cannot be directly comparedto those
reportedin the five-yearperioddesign, since the dependentvariableis different(log
odds of war onset in the next year versusin the next five years). The changedsignifi24. Seventeenwarsin theCollierandHoefflerlist startin a yearthatis divisibleby five. Forthese,there
is no lag at all in the five-yearset up forincomepercapitaandpopulation,whichraisesa mildconcernabout
endogeneity.
25. To avoid implausibleextrapolations,I extendedthe first or last measuredvalue forwardor backwardas appropriate.With few exceptions, this means using the sxp value for 1995 for years after 1995.
WhenI triedextrapolatingfor these years,the resultsareless favorableto primarycommodityexports.Following CollierandHoeffler,I replicatedor interpolatedthe valuesfor social fractionalization,ethnicdominance,andgeographicdispersionintomissingyears,since theseeitherdo notvaryin time orvaryonly in one
year(geographicdispersion).Formissingincome,growth,andpopulationdata,I employedthe sameprocedureas CollierandHoeffler,usingWorldBank-estimatedgrowthratesto extendPennWorldTablesincome
dataforwardpast 1992 (see FearonandLaitin[2003] for a discussionof the income andpopulationsources
used here).In constructingthe new dataset, I foundsome codingerrorsin the CollierandHoefflerdata,particularlyin the "peaceyears"variable.It appearsthatno consistentrule is appliedfor whetherpeace years
shouldbe countedto the startof the five-yearperiodor the startof a war thatbegins duringthe five-year
period.I did not correctthese errorsin the country-yearversionto ensurecomparability.I also doubtthat
doing so would have much effect.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 497
TABLE4
Civil WarOnset and Log(sxp) in the Country-YearFraming
Model 1:
CountryYears
Log(sxp)
Log(income)
Growth
Peace years
Log(population)
Fractionalization
Ethnicdominance
Geo. dispersion
Constant
N
N (wars)
.290
(.111)
-.549**
(.002)
-.146**
(.000)
-.004**
(.000)
.309**
(.001)
-.772
(.274)
.387
(.131)
-.082
(.909)
-1.074
(.479)
4,466
69
Model 2:
CountryYears
Model 3:
CountryYears
.202
(.218)
-.457**
(.002)
-.146**
(.000)
-.004**
(.000)
.278**
(.002)
.159
(.303)
-.471**
(.001)
-.133**
(.000)
-.004**
(.000)
.258**
(.002)
.384
(.135)
.065
(.928)
-1.900
(.144)
4,515
69
-1.523
(.203)
5,008
73
NOTE:Logit models with civil war onset as the dependentvariableand countryyears as observations.
Incomeandpopulationarelaggedoneyear;incomegrowthis forthepriorfiveyears.p-valuesarein parentheses.
*p < .05. **p < .01.
cance levels, however,suggest thatthe model may fit the dataless well. Note thatthe
estimatedcoefficient for fractionalizationfalls by a factor of three and is no longer
remotely significant,and the estimatedcoefficient for geographicdispersionis now
much closer to zero. Thep-values for sxp and its squarehave increased,and the relevantlikelihoodratiotest fails to rejectthe null hypothesisthattheircontributionto the
model is due to chance at the 5 percentlevel (p = .081).
Also worrisomeis thatthe estimatedcoefficientsfor sxp andsxp2arenow quitesensitiveto minorchangesin the model's specification.Models4, 5, and6 in Table2 drop,
in succession, the fractionalization,ethnic dominance, and geographic dispersion
variables,which showed little or no sign of a statisticallysignificantrelationshipto
civil waronsetin models 1 or 3. In fact, droppingany of these threevariables,singly or
in any combination,leaves a model in which one cannotrejectthe null hypothesisthat
primarycommodityexportsandtheirsquarearestatisticallyunrelatedto civil warrisk
(usingthe likelihoodratiotest anda 10 percentthreshold).As models 5 and6 in Table
2 illustrate,the lack of statisticalsignificancecan be substantial.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
498
JOURNALOF CONFLICTRESOLUTION
Table4 presentsmodelsusing the logarithmof sxp ratherthantheparabolicspecification adopted by Collier and Hoeffler. Primarycommodity exports now cease to
show a statisticallysignificanteffect at the 5 percentlevel even when all controlvariables are included (model 1), with even smallercoefficients and significancelevels
when I dropthe more dubiousof the controls(models 2 and 3).
Theseresultsdo notruleoutthepossibilitythathigherlevels of primarycommodity
exports(at least up to a point) cause a countryto have a higherrisk of civil waronset.
However,the strengthof the observedassociationdependsheavily on the choice of
sampleframingandcontrolvariables,somethingthatis farless trueof othervariables
in the model, such as per capitaincome, countrypopulation,years of peace, and economic growth rate. Given the great uncertaintyabout what variables should be
includedin a statisticalmodel of civil warriskin principle,one can say thathigherprimary commodity exports are not robustlyassociatedwith higher risks of civil war,
even in Collier and Hoeffler's own data.26
MULTIPLYIMPUTING MISSING DATA
What accountsfor the weakerresultsin the country-yearframingof the data set?
One possibility is that merely by making the lag times and treatmentof successive
warsconsistent,I havegottenridof accidentsthatproducedapparent"significance"in
the five-yearperiodset up. Anotherpossibilityis thatthe country-yearframingallows
more efficient use of the data (that is, fewer missing observations).Notice that the
numberof wars in the estimationsamplerises from fifty-two in the five-yearperiod
analysis to as many as seventy-three(of seventy-eight)in the country-yearanalysis.
This is largelybecause the latterloses fewer cases due to missing dataon the lagged
economic growthrate.
By using multipleimputation,one can gain insight into how much of the "significance"of primarycommodityexportsis due to list-wise deletionof observationsdue
to missingdata.Table5, model 1, uses the five-yearperiodframingandthe same specificationas in Table2, model 1, but multiplyimputesmissing values so thatinformation from all 1,288 cases in the "potentialsample"can be used.27The dependentvariable-the log odds of civil war outbreakin the next five years-is the same, so the
26. It is particularlystrikingthatonly by includingcontrolsthatdo notthemselvesappearstronglyor at
all relatedto civil war risk do primarycommodityexportsbegin to appearto have a consistent impact.
Anothersharpdifferencebetween resultsfor the five-year-periodandcountry-yearframingsconcernsthe
applicationof conditionalfixed effects: sxp andsxp are stronglysignificanteven with fixed effects in the
five-yearperiodset up, butnot at all with fixed effects and country-years.
27. Only 1,063 cases areused in the logit. Before makingthe datasets with imputedvalues,I dropped
the 121 preindependencecountryperiodsand the threecountriesthatwere neverformallyindependent.I
also dropfromthe estimationthe countryperiodsduringwhich a waris in progress,since they are"missing
in principle"in CollierandHoeffler'ssetupratherthanforlackof information.I thinkthesearethe mostjustifiabledecisions, but in any event, the resultswere slightly worse for primarycommodityexportswhen I
imputedon thebasis of all 1,288cases in thepotentialsample.See the appendixfordetailson the imputation
model.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 499
TABLE5
MultiplyImputedMissing Data Models for Civil WarOnset
Model 1:
Five-Year
Periods
sxp
sxp2
Log(income)
Growth
Peace years
Log(population)
Fractionalization
Ethnicdominance
Geo. dispersion
Constant
N
N (wars)
8.320*
(.033)
-11.775
(.120)
-.581**
(.002)
-.091**
(.006)
-.003**
(.002)
.387**
(.000)
-.626
(.436)
.432
(.118)
.060
(.937)
-4.500*
(.022)
1,063
75
Model 2:
Five-Year
Periods
7.156*
(.049)
-10.319
(.155)
-.477**
(.004)
-.087**
(.007)
-.003**
(.001)
.354**
(.000)
Model 3:
Country
Years
-4.502*
(.017)
6.088
(.104)
-9.965
(.195)
-.509**
(.001)
-.085**
(.000)
-.003**
(.001)
.291**
(.001)
-.554
(.409)
.395
(.116)
.073
(.922)
-2.859*
(.040)
1,063
75
5,168
76
Model 4:
Country
Years
5.195
(.142)
-8.789
(.233)
-.424**
(.002)
-.080**
(.000)
-.003**
(.000)
.265**
(.001)
-3.040*
(.020)
5,168
76
NOTE:Combinedlogit estimatesfrom sixteen multiplyimputeddata sets with civil war onset as the dependentvariable;observationsarefive-yearperiodsin models 1 and2 andcountryyears in models 3 and4.
Incomeandpopulationarelagged one year;income growthis for the priorfive-yearperiodor five years.pvalues are in parentheses.
*p < .05. **p < .01.
coefficientsmay be directlycompared.(The percentageof cases with a civil waroutbreakis also very similarfor the observedand imputeddata,at about7 percent.)
Overall,the Collier-Hoefflerstatisticalmodel performsworse when all cases with
some dataareused in the estimation.Severalof the coefficientestimatesmove noticeablytowardzero, althoughstandarderrorsalso fall a bit, reflectingthe efficiencygains
from using more data.Among the initially "significant"variables,primarycommodity exportsandtheirsquareshowthe largestpercentagechanges-the estimatesfor the
effect of primarycommodityexportsdropa factorof two when I avoid list-wise deletion.28Substantively,whereasgoing fromthe 10thto 90th percentileon sxp associates
with a change in risk from 1.1 to 11 percentin Collier and Hoeffler's set up, with the
missing datamodel the correspondingspreadis 2.3 to 7.4 percent.Statisticalsignifi28. The marginaleffect of sxp on the log odds of civil war outbreakat the medianlevel of sxp (.11)
reducesfrom 11.54 (Table2, model 1) to 5.72 (Table5, Model 1).
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
500
JOURNALOF CONFLICTRESOLUTION
cance also weakens somewhat,especially for sxp2.Model 2 in Table5 shows further
deteriorationin the estimatedeffects for primarycommodityexportswhen I dropstatistically the insignificant measures for fractionalization,ethnic dominance, and
geographicdispersion.
Models 3 and 4 in Table5 show the analogouslogits for multiplyimputeddatain
country-yearformat.Hereone sees a moremarkedreductionin the significancelevels
of the primarycommodityexportvariables.It appearsthatbothlist-wise deletionand
idiosyncrasiesof the five-yearperiodformatcontributedto the apparentstrengthof
the associationbetweenprimarycommodityexportsand civil warrisk in Collier and
Hoeffler's analysis.
OIL OR PRIMARY COMMODITY EXPORTS?
As discussed above, oil exports are a majorcomponentof primarycommodity
exportsfor a numberof countries,and the two variablesare moderatelywell correlated. Collier and Hoeffler (2002a) reportthat high oil dependenceassociates with
highercivil warrisk thanone would expect on the basis of otherprimarycommodity
exportsalone, althoughthe effect for the othercommoditiespersists.Using the data
from Collier and Hoeffler (2002b) in the five-yearperiodformat,I find thatthe measures of primarycommodityexportsremainreasonablystablewhen I add a variable
for fuel exports as a percentageof total exports (see Table 6, model 1). This result
dependsto some extenton havingthe statisticallyinsignificantvariablescoding geographicdispersionof populationand ethnic dominancein the model, as shown in
Table6, model 2.
In the country-yearframing,however,oil and primarycommodity exports trade
places. Table6, model 3, shows thatneitheroil nor sxp and its squareareparticularly
impressivewhen all threeincludedin the countryyearset up;model 4 shows a similar
resultwhen I use the log of sxp. Table6, model 5, shows thatwhen one dropsthe more
marginalvariables,primarycommodity exports are statisticallyinsignificant,while
there is strongersupportfor a nonrandomassociationbetween high oil exports and
civil war risk.
PRIMARY COMMODITY EXPORTS AND STATE STRENGTH
Collier and Hoeffler's theoreticalargumentfor expecting a strong relationship
between primarycommodity exports and civil war risk is that primarycommodity
exports measurerebel financing opportunity.This argumentis problematicfor two
main reasons. First, sxp is composed mainly of cash crops and fuel exports. Both
requirecontrolof a nationaldistributionsystemto exploit, which rebelgroupsalmost
neverhave (Fearonand Laitin2003). It is conceivablethatcash crop and oil exports
happento be correlatedwith the presenceof drugsandpreciousgems thathave been
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 501
TABLE6
Oil versus PrimaryCommodityExports
Model 1:
Five-Year
Periods
sxp
2
sxp
13.552*
(.013)
-19.380
(.052)
Model 2:
Five-Year
Periods
7.127
(.114)
-11.070
(.178)
Model 3:
Country
Years
6.800
(.116)
-13.446
(.106)
Log(income)
Growth
Peace months
Log(population)
Fractionalization
Ethnicdominance
Geo. dispersion
Constant
N
N (wars)
0.007
(.332)
-1.017**
(.000)
-0.084*
(.045)
-0.003**
(.001)
0.397**
(.007)
-2.307*
.011
0.546
(.095)
-1.126
(.223)
-0.915
(.769)
672
52
0.008
(.229)
-0.630**
(.002)
-0.089*
(.027)
-0.003**
(.001)
0.240
(.058)
-1.467
(.599)
722
53
Model 5:
Country
Years
3.286
(.389)
-8.750
(.237)
0.011
(.068)
-0.663**
(.001)
-0.144**
(.000)
-0.003**
(.000)
0.247*
(.029)
-1.081
.142
0.496
(.061)
-0.325
(.672)
-1.012
(.591)
0.205
(.375)
0.010
(.090)
-0.690**
(.000)
-0.143**
(.000)
-0.003**
(.000)
0.258*
(.031)
-1.089
.142
0.488
(.066)
-0.212
(.779)
0.035
(.984)
3,972
65
3,972
65
Log(sxp)
Fuel exports
Model 4:
Country
Years
0.012*
(.030)
-0.525**
(.001)
-0.128**
(.000)
-0.003**
(.000)
0.157
.112
-1.050
(.519)
4,339
69
NOTE:Logit estimateswith civil waronset as the dependentvariable;observationsarefive-yearperiodsin
models 1 and2 andcountryyearsin models 3, 4, and5. Incomeandpopulationarelagged one year;income
growthis for the priorfive-yearperiod or five years. Fuel exportsare as a percentageof total exports,pvalues are in parentheses.
*p < .05. **p < .01.
observedto financerebelgroupsin a numberof conflictsaroundthe world.Butthereis
no good reason to expect that this is the case.29
Second, even if production of primary commodity exports does raise opportunities
for "war taxation" by rebel groups, it also provides the government with a relatively
29. Moreover,recent work by Lujala, Gleditsch, and Gilmore (2005 [this issue]) and Humphreys
(2005) casts doubton the strengthandnatureof the linkbetweendiamondsandcivil war.Lujala,Gleditsch,
andGilmorefindlittle evidencethatdiamonddepositsaresystematicallyassociatedwith anelevatedriskof
onset (althoughalluvialdiamondsmay be relatedto higherrisksof "ethnic"waronset). Humphreysfinds a
strongrelationshipboth in Africa and globally between level of diamondproductionand civil war risk,
thoughhe arguesthatthe datado not supportthe mechanismsuggestedby CollierandHoeffler(easierrebel
finance).
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
502
JOURNALOF CONFLICTRESOLUTION
easy sourceof tax revenue.Why shouldrebelsbe morefavoredby this sourceof revenue than the government?Theremay be an argumenthere for drugsor alluvial diamonds, but if anythingit would seem to work the otherway for cash crops and fuel
exports.
An alternativehypothesisis thathigherdependenceon primarycommodityexports
for nationalincome marksweakerstate institutions,on average,for a given level of
income. This argumenthas often been madewith respectto oil producers(Karl 1997;
FearonandLaitin2003; Wantchekon2000). As TerryKarl(1997, 61) writes,"Given
theiraccess to easy revenuesfrompetroleum,few [oil exporters]soughtto supplement
state income throughsubstantialincreasesin domestictaxation. . . high statenessin
this arena[runningtheoil industry]occurredatthelong-termexpenseof theircapacity
to build extensive, penetrating,and coherentbureaucraciesthat could successfully
formulateand implementpolicies."
The same generalargumentmightextend,with less force, to nonfuelprimarycommodities. As with oil, other primarycommodity exportsrequirenationalcollection
and marketingsystems thatmake statetaxationrelativelyeasy. Acemoglu, Johnson,
andRobinson(2001) arguethatwheresettlementwas not a good optiondueto disease,
colonial regimes set up "extractiveinstitutions"designed to do little more than tax
commodity exports. RobertBates (1981) describes how the commodity marketing
boardsdesigned by the colonial regimes in Africa were often the sole basis for state
revenue and patronageafter independenceand provided little incentive for urbanbased elites to pursueconstructiveruraldevelopmentstrategies.Althoughthe institurequiredfor cash cropsis moreextensiveandsocially penetrating
tionalinfrastructure
than for "enclave"mining industries,it could be that high dependenceon primary
commodityexportsmarksstateswith less developedadministrativecapacitiesgiven
theirincome.
Unfortunately,good direct measures of a state's administrativecapability and
integrityarelacking.Acemoglu, Johnson,andRobinson(2001) and othershave used
informationfrom surveysof investorson the risk of expropriationandrepudiationof
governmentcontractsin differentcountries(KnackandKeefer1995). These aremeasuredfor about120 countriesbetween 1981 and 1994 on a scale of 0 to 10, withhigher
valuesindicatinglowerrisksof expropriationor repudiationof contracts.Not surprisingly,these measuresarehighly correlatedwith percapitaincome.If, for a given level
of income, primarycommodityexportsassociatewith weakerstates,then one would
expect the partialcorrelation(controllingfor income) between the contractrepudiation or expropriationmeasuresand sxp to be negative.
Table7, model 1, reportsthe regressionof the measureof contractobservanceon
logged primarycommodityexportsandlogged percapitaincomefor the 115 countries
with data.30As expected,for a given level of income, stateswith greaterprimarycommodity dependenceare perceivedas less reliableby investors,on average,although
30. Contractobservanceis averagedfor each countryoverthe availabledatafrom 1982 to 1995;there
is very little variationovertime withincountries,andthereis no theoreticalreasonto expectthatshort-term
fluctuationsin primarycommodityexportsmeasurechangesin statecapacityin the sense described.Log of
income is averagedoverpost-1979 valuesfor each country,andlog of sxp is averagedover all values.I also
consideredthe measureof risk of expropriation,which gives nearlyidenticalresults.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 503
TABLE7
GovernmentObservanceof Contractsand sxp
Log(sxp)
Log(income)
Model 1
Model 2
-0.351**
(2.64)
1.361**
(12.28)
-0.161
(1.13)
1.429**
(13.04)
-0.014**
(2.98)
-5.435**
(6.25)
Fuel exports
Constant
N
R2
-5.558**
(6.19)
115
.604
115
.633
NOTE:Ordinaryleast squareswith a 0 to 10 measureof perceivedgovernmentobservanceof contractsas
the dependentvariable,averagedfor each countryover availabledatafrom 1982 to 1995. Log(sxp)is averaged for each countryover availabledatafrom 1960 to 1995, andlog(income)andfuel exportslikewise for
datafrom 1980 to 1999. t-statisticsare in parentheses.
**p < .01.
the substantiveimpactis not large.31Whenone addsfuel exportsas anothermeasureof
stateweaknessconditionalon income (see Table7, model 2), the estimatedcoefficient
for primary commodity exports more than halves and becomes statistically
insignificant.
So oil exportershave significantlyweakerstatesgiven income per capita,according to this measureof stateweakness.But while exportersof otherprimarycommodities have marginallyless reliable governmentson average, the effect is not very
consistent.
CONCLUSIONS AND POLICY IMPLICATIONS
Four main conclusions emerge from the preceding analysis. First, the empirical
association between primarycommodity exports and civil war outbreakis neither
strongnorrobust,even using CollierandHoeffler's(2002a, 2002b) civil warcodings
and model specifications.Second, insofaras thereis some association,this is due in
partto the inclusion of fuel exports in the primarycommodity measure,which are
morerobustlyrelatedto conflictonset.Third,it seems unlikelythatoil exports(orcash
crops)predicthighercivil warriskbecauseoil providesbetterfinancingopportunities
for would-berebels. It seems more likely thathigh oil exportsindicatea weakerstate
given the level of per capitaincome and possibly a greater"prize"for state or seces31. Movingfromthe 10thto the 90th percentileon sxp is estimatedto associatewith a dropof 0.75 on
the 10-pointscale for contractrepudiationrisk, or abouthalf of one standarddeviation.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
504
JOURNALOF CONFLICTRESOLUTION
sionistcapture,bothof which mightfavorcivil war.Similarconsiderationsmay apply
for nonfuel commodityexports.
Fourth,thereis directevidence thatoil exportershave less reliableand competent
states given their income levels and weakerevidence that this is true on averagefor
exportersof otherprimarycommodities.
What implicationsdo these findings have for policy regardingnaturalresources
and civil war?We should begin by distinguishingtwo questions.First, what policy
measuresshouldbe takento preventnew civil warsfrombreakingout?This question
canbe informedby researchon the causesof civil waronset,whichis the focus of Collier and Hoeffler's work on primarycommodityexports. Second, what policy measuresshouldbe takento shortenan ongoing civil war?Answersto this questionmight
be informedby researchon whatdeterminescivil warduration.CollierandHoeffler's
researchon primarycommodity exports and civil war onset is not directlyrelevant
here.
Regardingthe second policy question-What shouldbe done to shortenan ongoing civil war?-it is just common sense thatif a particularrebel groupis financedby
the "looting"of naturalresourcerents,thena policy optionforhelpingto endthe waris
to try to cut off this source.No fancy cross-nationalresearchis needed to supportor
make such an argumentplausible.32Thus, policies to eliminatethe sale of "conflict
diamonds,"to secure oil pipelines, or to reduce demandand supply of narcoticsare
natural possibilities for such cases.33
Regardingthe firstquestion-What shouldbe done to preventcivil waroutbreaks
in general?-I have arguedhere thatthereis no clearevidence thathigh levels of primary commodity exportscause higher risk of civil war onset by making for easier
rebelstart-upfinance.If this is correct,thenpolicy measuresintendedto "diversify...
economies and visibly channel commodity income into social service programs"would not be expected to reduce civil war risks, at least not via the posited
mechanism.34
Oil exportersdo seem to havebeen moredisposedto civil waronset,butit is not yet
clearwhatthe most importantmechanismsare.If, as arguedhere,oil proxiesfor weak
stateadministrativecapabilitiesat a given level of income, or if oil makesfor trouble
by raisingthe "prize"value of stateor regionalcontrol,thenpolicies to involve internationalinstitutionsin the monitoringand managementof weak states'oil revenues
32. Cross-nationalresearchmay be useful to establishhow prevalentis the phenomenonand potentially for estimatinghow effective are differentmeasuresto cut rebel fundingfrom naturalresourcerents.
Fromcase studies,it is quiteclearthatrebels sometimesdo obtainfinancethis way (see, for example,Ross
2004). Fearon(2004) finds thatrebel financefrom contrabandis stronglyassociatedwith longer civil war
duration.In theirstudy of civil war duration,Collier,Hoeffler,and Soderbom(2004) find thatfalls in the
priceof primarycommoditiesassociatewithshorterdurationin countriesthatarehighlydependenton them,
althoughit is unclearif this is due to an effect on rebelor governmentfunding.Humphreys(2005) findsthat
oil anddiamondproductionare associatedwith shorterratherthanlongercivil wars.My pointin the text is
thateven if this is correcton average,if one observesrebelsliving off illicit diamondrevenuesin a particular
case (like Angola),thentryingcut off this sourcewouldremainanobviouspolicy optionfortryingto endthe
war.
33. Providedthatit is apparentthatthe rebel groupoffers less prospectof good governancethanthe
governmentin power,which is not always the case.
34. JohnBurgess,"CivilWarsLinkedto CertainEconomies."
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 505
couldhelpbreakthe link. Internationalmonitoringandinfluenceon the distributionof
oil revenuesmight reducethe payoffs for an extractive,exploitativestrategyof state
captureandcontrol,while increasingpoliticians'incentivesto competeon the basis of
service and infrastructureprovision.
For example,the WorldBank has recentlyattemptedto negotiatemonitoringand
managementarrangementsas a condition for supportingpipeline developmentin
southernChad.Outsideexpertsare skepticalthatthis specific deal will work,but the
generalidea seems worthpursuingbased on the empiricalfindingsaboutoil andconflict.35Forweak statesthatalreadyexportlargeamountsof oil, the IMF,WorldBank,
or a new internationalinstitutioncould offer a standardizedexternalmonitoringand
managementservice thatthe state could publicly commit to. If a weak-stateoil producerholds elections,the existenceof such an optioncould inspireor drivecandidates
to competefor votersupportby declaringa willingnessto committo the international
package.36
While the empiricalcase must be deemed "notproven,"it remainsreasonableon
theoreticalandanecdotalgroundsto thinkthatthe availabilityof rebelfinanceaffectsa
country'sprospectsfor civil war.We still do not know, however,if rebel finance is a
"criticalconstraint"thatvariesa greatdealfromplaceto place, orif it is easily satisfied
providedotherconditionsare satisfied(e.g., weak centralgovernmentwith little rural
presence, a sudden increase in grievancesdue to changed governmentpolicy, or a
prevalenceof unemployedyoung males).Indeed,we stillknow little aboutthe sources
of rebel groups'incomes. How much comes from local donationsand "revolutionary
taxes,"how much from foreign governmentsor companies,how much from diasporas? This is difficult informationto gather,althoughinterestingwork by Weinstein
(2005 [this issue]) and Humphreysand Weinstein(2004) is beginningto explorethe
effect of differentincome sourceson rebel organizationand strategies.While Collier
andHoeffler'sproposalthatthe availabilityof rebelfinanceis a key determinantof the
prospectsfor civil war remainsto be demonstrated,it has without doubt helped to
generatean importantresearchagenda.
APPENDIX
The Imputation Model
basedon a
I usedJ. L. Schafer'ssoftware,NORM,for producingmultipleimputations
multivariatenormalmodel of the data(Schafer2000). The centralideais thatthe variablesin the
imputationmodel are assumed to have a multivariatenormaldistribution,the parametersof
which areestimatedfromthe observeddata.Imputedvaluesformissing dataaredrawnfromthe
relevantconditionaldistributionto make a set of n imputed data sets. Coefficient estimates
35. On the Chadproject,see for exampleDaphneEvitar,"StrikingIt Poor:Oil as a Curse,"New York
Times,June7, 2003, p. B9; andKrasner(2004). JeffreySachsanda teamfromColumbiaUniversity'sEarth
Institutehave been involved in helping to design institutionsfor the managementof oil revenuesin Sao
Tom6e Principe.See http://www.earthinstitute.columbia.edu/cgsd/STP/.
36. Krasner(2004) proposesthis and severalother"sharedsovereignty"mechanismsto make weak
stateoil revenuesless easily appropriableby corruptpoliticians.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
506
JOURNALOF CONFLICTRESOLUTION
based on these n data sets are then combined using "Rubin's rules" to yield the numbers reported
in Table 5.
I limited the variables in the imputation model to the variables in Collier and Hoeffler's
(2002b) main specification: civil war onset, primary commodity exports, the log of per capita income, the log of population, fractionalization, ethnic dominance, years of peace, and geographic dispersion. To make the multivariate normal assumptions more accurate, primary commodity exports and fractionalization were logged for use in the imputation model. Imputed values for peace years were rounded to the nearest integer, and imputed values for the dichotomous
variables civil war onset and ethnic dominance were rounded to zero or one, whichever was
closer.
The sixteen imputed data sets were created by saving the estimates from every 5,000th round
of 80,000 rounds of data augmentation. Auto-correlation plots suggested strongly that 5,000
rounds was far more than enough to ensure that convergence in distribution had occurred (the
EM algorithm for estimating means and covariances of the imputation model data converged in
less twenty-four iterations for both the country-year and five-year-period versions).
REFERENCES
Acemoglu, Daron, Simon Johnson,and James A. Robinson. 2001. The colonial origins of comparative
development:An empiricalinvestigation.AmericanEconomicReview91 (5): 1369-1401.
Ballentine, Karen,and Jake Sherman.2003. Thepolitical economyof armed conflict: Beyondgreed and
grievance. Boulder,CO: LynneRienner.
Bates, RobertH. 1981. Marketsand states in tropicalAfrica: Thepolitical basis of agriculturalpolicies.
Berkeley:Universityof CaliforniaPress.
Beck, NathanielL., and Simon Jackman.1998. Beyond linearityby default:Generalizedadditivemodels.
AmericanJournalof Political Science 42:596-627.
Berdal,Mats R., andDavid M. Malone,eds. 2000. Greedand grievance:Economicagendas in civil wars.
Boulder,CO: LynneRienner.
Collier,Paul. 1999. Doing well out of war.Washington,DC: WorldBank.
Collier, Paul, V. L. Elliott, HavardHegre, Anke Hoeffler,MartaReynal-Querol,and Nicholas Sambanis.
2003. Breakingthe conflicttrap:Civil war and developmentpolicy. Washington,DC: WorldBank and
OxfordUniversityPress.
Collier,Paul,andAnke Hoeffler.2000. Greedandgrievancein civil war.WorldBankWorkingPaperSeries
2000-18, April 26, Washington,DC.
---.
2002a. Greed and grievance in civil war. World Bank, DECRG. http://econ.worldbank.org/
programs/conflict.
. 2002b. On the incidence of civil war in Africa.Journalof ConflictResolution46 (1): 13-28.
---.
2004. Greedand grievancein civil war. OxfordEconomicPapers56 (4): 563-96.
Collier, Paul, Anke Hoeffler, and Mans Soderbom.2004. On the durationof civil war.Journal of Peace
Research41 (3): 253-74.
Eisinger,Peter. 1973. The conditionsof protestbehaviorin Americancities. AmericanPolitical Science
Review67:11-28.
Englebert,Pierre,and James Ron. 2004. Primarycommoditiesand war:Congo-Brazzaville'sambivalent
resourcecurse. ComparativePolitics 37 (1): 61-81.
Fearon,JamesD. 2004. Whydo some civil warslast so muchlongerthanothers?Journalof Peace Research
41 (3): 275-301.
Fearon,JamesD., and David D. Laitin.2003. Ethnicity,insurgency,andcivil war.AmericanPolitical Science Review97 (1): 75-90.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions
Fearon / PRIMARYCOMMODITY
EXPORTSAND CIVILWAR 507
Humphreys,Macartan.2005. Naturalresources,conflict, and conflict resolution:Uncoveringthe mechanisms. Journalof ConflictResolution49 (4): 508-37.
Humphreys,Macartan,andJeremyM. Weinstein.2004. Whatthe fighterssay:A surveyof ex-combatantsin
SierraLeone. Reportfor the Post-ConflictReintegrationInitiativefor Educationand Empowerment
(PRIDE).New York:ColumbiaUniversity/Stanford,CA: StanfordUniversity.
Karl,TerryL. 1997. Theparadox of plenty.Berkeley:Universityof CaliforniaPress.
Keen, David. 1998. Theeconomicfunctions of violence in civil wars. London:OxfordUniversityPress.
King, Gary,JamesHonaker,Anne Joseph,and KennethScheve. 2001. Analyzingincompletepolitical science data:An alternativealgorithmfor multipleimputation.AmericanPolitical Science Review95 (1):
49-69.
King, Gary,andLangcheZeng. 2001. Logistic regressionin rareevents data.PoliticalAnalysis9 (2): 13763.
Knack,Stephen,andPhilipKeefer.1995.Institutionsandeconomicperformance:Cross-countrytests using
alternativeinstitutionalmeasures.Economicsand Politics 7 (3): 207-27.
Krasner,Steven D. 2004. Sharingsovereignty:New institutionsfor collapsed and failing states. International Security29 (2): 85-120.
Lujala,Piivi, Nils PetterGleditsch,andElisabethGilmore.2005. A diamondcurse?Civil waranda lootable
resource.Journalof ConflictResolution49 (4): 538-62.
McCarthy,John,andMayerN. Zald. 1977. Resourcemobilizationand social movements:A partialtheory.
AmericanJournalof Sociology 82:1212-41.
Reno, William. 1998. Warlordpolitics and Africanstates. Boulder,CO: LynneRienner.
Ross, Michael. 2004. Whatdo we know aboutnaturalresourcesandcivil war?Journalof Peace Research
41:337-56.
Rubin,Donald B. 1987. Multipleimputationfor nonresponsein surveys.New York:Wiley.
Schafer,J. L. 2000. NORMfor Windows95/98/NT,version2.03. November.http://www.stat.psu.edu/-jls/
misoftwa.html.
Sherman,Jake. 2001. The economics of war: The intersectionof need, greed, and creed: A conference
report.September10. WoodrowWilson InternationalCenterfor Scholarsand the InternationalPeace
Academy,Washington,DC.
Snyder,Richard,andRaviBhavnani.2005. Diamonds,blood, andtaxes:A revenue-centeredframeworkfor
explainingpolitical order.Journalof ConflictResolution49 (4): 563-97.
Tilly, Charles. 1978. Frommobilizationto revolution.Reading,MA: Addison-Wesley.
Wantchekon,Leonard.2000. Whydo resourcedependentstates have authoritariangovernments?New
Haven,CT:Yale UniversityPress.
Weinstein,JeremyM. 2005. Resourcesand the informationproblemin rebel recruitment.Journalof Conflict Resolution49 (4): 598-624.
This content downloaded from 171.65.249.4 on Thu, 24 Oct 2013 17:09:01 PM
All use subject to JSTOR Terms and Conditions