Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Conflict or compromise : theory and evidence from Africa and Asia Santarrosa, Rogerio Bianchi 2019

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata


24-ubc_2019_november_santarrosa_rogerio.pdf [ 5.49MB ]
JSON: 24-1.0383289.json
JSON-LD: 24-1.0383289-ld.json
RDF/XML (Pretty): 24-1.0383289-rdf.xml
RDF/JSON: 24-1.0383289-rdf.json
Turtle: 24-1.0383289-turtle.txt
N-Triples: 24-1.0383289-rdf-ntriples.txt
Original Record: 24-1.0383289-source.json
Full Text

Full Text

Conflict or Compromise: Theory and Evidencefrom Africa and AsiabyRogerio Bianchi SantarrosaB.A., University of São Paulo, 2008M.A., Fundação Getúlio Vargas (EESP-FGV), 2011a thesis submitted in partial fulfillmentof the requirements for the degree ofDoctor of Philosophyinthe faculty of graduate and postdoctoralstudies(Economics)The University of British Columbia(Vancouver)September 2019c© Rogerio Bianchi Santarrosa, 2019The following individuals certify that they have read, and recommend tothe Faculty of Graduate and Postdoctoral Studies for acceptance, the thesisentitled:Conflict or Compromise: Theory and Evidence from Africaand Asiasubmitted by Rogerio Bianchi Santarrosa in partial fulfillment of therequirements for the degree of Doctor of Philosophy in Economics.Examining Committee:Patrick Francois, EconomicsSupervisorThorsten Rogall, EconomicsSupervisory Committee MemberFrancesco Trebbi, EconomicsUniversity ExaminerChristopher Kam, Political ScienceUniversity ExaminerGustavo J. Bobonis, University of TorontoExternal ExamineriiAbstractCivil wars are a recurring phenomenon undermining development inweak states. Faced with the possibility of costly conflict, why don’t leadersshare power? I investigate the role of an unexplored commitment problem,both theoretically and empirically. The model features a leader who can ap-pease challengers by sharing power, but doing so increases their effectivenessat launching a rebellion. I show that commitment worsens as the opposi-tion becomes stronger, and derive testable non-monotonic implications ofgroup strength and distributional shocks on power-sharing and conflict. Aschallengers become stronger, the likelihood of inclusion (positive transfers)increases up to a threshold, beyond which the leader prefers to exclude anopposing group and face conflict. I test the model using data on politicallyrelevant ethnic groups in Africa and Asia, their access to executive power,and armed organizations claiming to represent them. To that end, I usethree complementary strategies: (i) within-country variation in populationshare to proxy for group strength; (ii) quasi-randomly split groups acrosscountries; and (iii) conflict-inducing distributional economic shocks withina country, by combining geo-referenced data on the ethnic homelands, crop-land maps and international prices. The empirical findings strongly accordwith predictions of the theory. I then structurally estimate the model pa-rameters and explore policy relevant counterfactuals, including the effects ofdemocratization, changes in military capacity, financial aid, sanctions andquotas.iiiLay SummaryCivil conflict has been the most predominant type of conflict sinceWorld War II. They are generally organized along ethnic lines, and typi-cally launched by groups underrepresented in the executive power. Whydon’t leaders share power instead of excluding opposing groups and provok-ing harmful wars? I develop a theory where leaders can appease opposinggroups by sharing power, but doing so makes the opposition more power-ful and more likely to win wars. The theory predicts when leaders includeopponents, when they opt for exclusion, and when conflict occurs. Usinga novel dataset on the degree of inclusion of ethnic groups into executivepower and ethnic conflicts in Africa and Asia, I find strong empirical sup-port for the theory. The estimation of the model gives insight on the mosteffective policies aimed at mitigating conflict.ivPrefaceThis dissertation is original, unpublished, independent work by the au-thor, Rogerio Santarrosa.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . xii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Bargaining over Power: A Model of Power-sharing andConflict . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112.2 A Simple Model . . . . . . . . . . . . . . . . . . . . . . . . . . 132.2.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . 132.2.2 Model Set-up . . . . . . . . . . . . . . . . . . . . . . . 142.2.3 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 152.2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 162.3 Model Extension: Dynamic Framework . . . . . . . . . . . . . 182.3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . 18vi2.3.2 Model Set-up . . . . . . . . . . . . . . . . . . . . . . . 192.3.3 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 202.3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . 212.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 Power-sharing and Conflict: Empirical Evidence fromAfrica and Asia . . . . . . . . . . . . . . . . . . . . . . . . . . 273.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.2 Data: EPR-2018 . . . . . . . . . . . . . . . . . . . . . . . . . 283.2.1 Key variables . . . . . . . . . . . . . . . . . . . . . . . 303.2.2 Descriptive overview of the data . . . . . . . . . . . . 333.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 343.2.4 Summary Statistics . . . . . . . . . . . . . . . . . . . . 353.3 Empirical Evidence . . . . . . . . . . . . . . . . . . . . . . . . 353.3.1 Within-Country Variation . . . . . . . . . . . . . . . . 353.3.2 Split groups . . . . . . . . . . . . . . . . . . . . . . . . 433.3.3 Comparative Statics: Evidence from relative economicshocks . . . . . . . . . . . . . . . . . . . . . . . . . . . 463.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504 Power-sharing and Conflict: Structural Estimation andPolicy Counterfactuals . . . . . . . . . . . . . . . . . . . . . . 524.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524.2 Additional data . . . . . . . . . . . . . . . . . . . . . . . . . . 534.3 Econometric specification . . . . . . . . . . . . . . . . . . . . 554.3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . 554.3.2 Strength . . . . . . . . . . . . . . . . . . . . . . . . . . 564.3.3 Conflict . . . . . . . . . . . . . . . . . . . . . . . . . . 574.3.4 Power-sharing . . . . . . . . . . . . . . . . . . . . . . . 584.3.5 Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . 604.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624.5 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . 62vii4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634.7 Counterfactuals . . . . . . . . . . . . . . . . . . . . . . . . . . 654.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96Appendix A Additional Tables . . . . . . . . . . . . . . . . . . 104Appendix B Additional Figures . . . . . . . . . . . . . . . . . . 114Appendix C Dynamic Model: Solution . . . . . . . . . . . . . 122viiiList of TablesTable 1 Summary Statistics (non-leading groups) . . . . . . . . . . 74Table 2 Probability of Inclusion and Conflict . . . . . . . . . . . . 75Table 3 Inclusion and group size (3rd order polynomial) . . . . . . 76Table 4 Discrimination and group size . . . . . . . . . . . . . . . . 77Table 5 Probability of Inclusion and Conflict, by group size . . . . 78Table 6 Probability of Inclusion and Conflict, by group size (Leader& Country-year FEs) . . . . . . . . . . . . . . . . . . . . . 78Table 7 Conflict and group size, by power status . . . . . . . . . . 79Table 8 Inclusion and Conflict by group size (Split groups) - Africaand Asia . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80Table 9 Inclusion and Conflict by group size (Split groups) - Africa 81Table 10 Effects of Relative Shocks on Conflict and Inclusion . . . . 82Table 11 Effects of Relative Shocks on Conflict and Inclusion (Noleader change) . . . . . . . . . . . . . . . . . . . . . . . . . 83Table 12 Structural Estimates . . . . . . . . . . . . . . . . . . . . . 84Table 13 Counterfactuals . . . . . . . . . . . . . . . . . . . . . . . . 85Table A1 EPR-2018: Descriptive Statistics (Africa and Asia) . . . . 104Table A2 Conflict and Power status . . . . . . . . . . . . . . . . . . 106Table A3 Probability of Inclusion and Conflict, by group size (con-trolling for ranking of group size) . . . . . . . . . . . . . . 107Table A4 Probability of Inclusion and Conflict, by group size (con-trolling for distance to capital and leader’s homeland) . . . 108ixTable A5 Probability of Inclusion and Conflict, by group size (con-trolling for rainfall) . . . . . . . . . . . . . . . . . . . . . . 109Table A6 Probability of Inclusion and Conflict, by group size (con-trolling for geography) . . . . . . . . . . . . . . . . . . . . 110Table A7 Probability of Inclusion and Conflict, by group size (con-trolling for intra-group fractionalization) . . . . . . . . . . 111Table A8 Probability of Inclusion and Conflict, by group size (con-trolling for light density) . . . . . . . . . . . . . . . . . . . 112Table A9 Probability of Inclusion and Conflict, by group size (ColdWar vs Post-Cold War) . . . . . . . . . . . . . . . . . . . . 113xList of FiguresFigure 1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86Figure 2 Results (Equilibrium path) . . . . . . . . . . . . . . . . . 87Figure 3 Probability of Inclusion and Conflict . . . . . . . . . . . . 88Figure 4 Inclusion - Predicted values (Country FE regression) . . . 89Figure 5 Effect of group size on conflict, by Power status . . . . . . 90Figure 6 Illustration: split groups . . . . . . . . . . . . . . . . . . . 91Figure 7 Inclusion - Predicted values (Ethnic FE regression) . . . . 92Figure 8 Effects of cc/pi ↓ (pi/cc ↑) . . . . . . . . . . . . . . . . . . 92Figure 9 Distributional shocks - Illustration (Togo) . . . . . . . . . 93Figure 10 Model fit - Probability of Inclusion and Conflict, by groupsize . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94Figure 11 Counterfactuals - Effects on Conflict Probability . . . . . 95Figure A1 Ethnic groups in Conflict over the control of the govern-ment (1945-2017) . . . . . . . . . . . . . . . . . . . . . . . 115Figure A2 EPR-2018: Kenya . . . . . . . . . . . . . . . . . . . . . . 116Figure A3 EPR: Politically relevant groups in Africa and Asia (2017) 117Figure A4 EPR: Politically relevant groups in Africa and Asia (2017) 118Figure A5 Probability of Inclusion and Conflict, by group size . . . . 119Figure A6 Political status distribution, by group size . . . . . . . . . 120Figure A7 Transfers to included groups, by group size . . . . . . . . 121xiAcknowledgmentsI am deeply indebted to my supervisor, Patrick Francois, for all of hisadvice and encouragement. I am particularly thankful for his generosity,which included time and financial support. I would also like to thank mycommittee members, Siwan Anderson, Thorsten Rogall and Paul Schrimpf,for their continuous support throughout the years.I have benefited from the support of many other UBC faculty and staff,including Francesco Trebbi, Felipe Valencia Caicedo, Munir Squires, DavidGreen, Matilde Bombardini. I am also grateful to my fellow PhD studentsfor their help and advice at multiple stages of this project, and their compan-ionship during this journey, especially Nathan Canen, Anderson Frey, JoaoFonseca, Arkadev Ghosh, Juan Felipe, Vinicius Pecanha, Michael Wiebe,Tom Cornwall, Ben Milner and Marcelo Sacchi de Carvalho. I could notforget close friends who have provided companion, joy, and priceless supportduring this journey: Mateus, Carol, Erich, Tais, Bruno, Kajle, Natasha, andWes.Last but not least, I am indebted to my family, for their longstandingsupport. Most of all, I am grateful to my wife Acza, who embraced ourdecision to embark on this path and have absorbed much of the burden thatcame from it. This achievement is pretty much yours!xiiChapter 1IntroductionCivil wars have been the most predominant type of conflict involving thestate since World War II, severely undermining development and institutionsin the developing world. More precisely, during this period 90% of conflictsfought by central governments were against organizations within the country(as opposed to conflicts against another state).1 Although societies may bedivided along many different lines - class, geography, ideology, for example -still the majority of civil conflicts are between groups defined by ethnicity.2One of the most consequential conflicts for political and economic sta-bility are those over the control or composition of the central government.And they are strongly associated with the ethnic composition of the gov-ernment cabinet: these conflicts are typically launched by ethnic groups un-derrepresented in the executive power (Denny and Walter [2014], Cedermanet al. [2010]). Such ethnopolitical movements, and the resulting violence, aremostly predominant in Africa and Asia, including the Middle East (Krauseand Suzuki [2005]). In fact, ethninicity is the fundamental channel through1Own calculation using data from the Uppsala Conflict Data Program (UCDP).2According to Denny and Walter [2014], 64% of civil wars (defined by conflicts withmore 1,000 battle deaths) have been divided along ethnic lines. Own calculation usingupdated data and a lower threshold at 25 battle deaths per year finds that around 54% ofintrastate conflicts are ethnic based. Denny and Walter [2014] argue that rebel movementsare more likely to organize around ethnicity because ethnic groups are more apt to beaggrieved, better able to mobilize, and more likely to face difficult bargaining challengescompared to other dimensions.1which competition over wealth and power is expressed in the region (Brown[2003], Roessler [2016], Bayart [2009]).3Examples abound historically, and even recently. The Burundian civilwar was the result of long standing ethnic divisions in the country. Sincecolonization Burundi had been historically governed by the minority ethnicTutsi, largely opposed by majority Hutu political organizations. In the 90’sa succession of bi-ethnic governments attempting to mitigate ethnic tensionsescalated into a civil war with an estimate of 300,000 casualties (mostly civil-ians)4. Likewise, the Sudanese civil wars have been the result of a strugglefor power and an inability on the part of contesting parties to reach agree-ment on how to share it. The Southern region had fought one of the longestwars since World War II against the northern, Arab-dominated government.Yet, the resulting independence of South Sudan precipitated further tensionsbetween President Kiir and his vice-president, Riek Machar, causing the lat-ter’s dismissal and igniting a new ongoing civil war between their respectiveethnic groups. Largely similar accounts describe civil conflicts in SouthAfrica, Liberia, Uganda, Rwanda, Iraq, Afghanistan, Syria and many othercountries where government and opposition (organized along largely ethnicor religious lines) struggled for power, but were unable to reach a peacefulagreement.A natural question arising in such settings is why, faced with the possi-bility of costly rebellion, leaders don’t share power instead of excluding orseverely under representing opposing groups? In fact, a large literature hassuggested that this is a realistic viable option. Since exclusion is the keyfactor behind ethnic conflicts, power-sharing has been often prescribed as apolitical solution to overcome deep divisions and mitigate conflict.5 Theserecommendations also find some empirical basis. Hartzell and Hoddie [2003]3Using Ethnic Power Relations-2018 Dataset, 30% of African and Asia countries havefaced an ethnic conflict over the control of the government in the post World War II period,compared to only 7% in the rest of the world. Figure A1 illustrates the ethnic groupsinvolved in this type of conflict and their concentration in Africa and Asia (particularly,in the Middle East)4, for instance, Spears [2000], Lemarchand [2006], Mehler [2009] and Binningsbøand Dupuy [2009] for a discussion on the topic.2shows that power-sharing is a frequent form of conflict resolution. Francoiset al. [2015] demonstrate for a sample of African countries that regimes aresurprisingly inclusive, and do so to avoid revolutions.Yet, conflict is widespread. Therefore, the more precise question is whyisn’t power-sharing always able to prevent conflict? Why do leaders some-times include opposing groups into the coalition and mitigate conflict, whilein other times opt for exclusion and hence invite it? A number of theoreticalreasons have been proposed to explain the failure to reach efficient, mutuallyadvantageous agreements, but very limited empirical evidence has been ableto pin down the mechanisms (Blattman and Miguel [2010]).In this thesis, I investigate the role of a unexplored commitment prob-lem, both theoretically and empirically. The theory developed in Chapter2 considers the perspective of a leader standing alongside contesting groupswho have the capacity to attempt to overthrow the leader via entering into a costly conflict. The leader can appease challengers6 by giving themsome of what they want, i.e., sharing power (which will here be thought ofas a division of the rents from office, but which typically involves positionsin the cabinet of the country and the military). However, the downside tosuch sharing is that, should the challenger subsequently decide to repudiatethe deal and rebel, the probability of the challenger succeeding will increasewith the share of power that the challenger has enjoyed under the powersharing agreement. Horowitz [1985] and Roessler [2016, 2011] have arguedthat access to the state’s coercive apparatus can easily be leveraged by con-testing parties into a better position to strike against incumbent leaders,and improving their bargaining power. 7 This effect of sharing power leads6I use "challenger" and "opposition" interchangeably throughout the thesis7Cabinet appointment entails authority and discretion over the distribution of resourcesin key areas. And there is strong evidence of ethnic favoritism on the distribution ofexpenditure (e.g., Kramon and Posner [2016]; Ejdemyr et al. [2018]). For example, ethnicgroups with co-ethnics in the cabinet can have road building in their districts (Burgesset al. [2015]). In turn, the roads can facilitate the mobilization of militias when attemptingto take control of the capital (Rogall [2014]). Alternatively, we may also think of politicalinclusion as an expansion of state capacity to the ethnic region, including policing, militarycontrol, recruitment to the army, public goods, etc. Again, all of these can be used toincrease the advantage of the rebels in a potential conflict.3to a commitment problem on the challenger’s part: they are unable to com-mit to not using their partial access to power (which is given as a form ofappeasement) as a platform from which to launch aggression against theleader.This reasoning leads to a clear non-monotonicity in the relationship be-tween group strength (defined as the probability of success in a rebellion),power sharing, and conflict. Weak groups remain peaceful despite receivingno access to the state, as they cannot credibly move against the leader, andtherefore do not need to be accommodated. As groups become stronger,they receive access to, and increasingly larger shares of, power but remainpeaceful until a threshold level of strength is reached. Beyond this thresh-old groups are too strong to be appeased without receiving large sharesand having their bargaining power substantially increased. Consequently,the leading group chooses to exclude these. However, facing exclusion, andgiven their high level of innate strength, these groups become the ones likelyto launch conflicts.These predictions are strongly confirmed in Chapter 3. I test the modelusing a innovative data on politically relevant ethnic groups in Africa andAsia, from 1946 to 2017.8 The dataset was introduced by Cederman et al.[2010] and motivated by the fact that the majority of ethnic conflicts werethe result of competing ethnic groups’ claims to the state, which in turn iscaptured by the representatives of some groups. Therefore, the collectionof data informing of the degree of representation of each ethnic group inthe executive power was crucial to understand ethnic conflicts. This uniquedataset provides information on the access of each ethnic group to executivepower. Furthermore, armed organizations in conflict over the governmentare linked to ethnic groups based on their claims of representation or recruit-ment of such groups. As a result, I am able to test the model’s theoreticalmechanisms not only via conflict outcomes, but also by examining the shar-ing of power with opposition groups under peaceful conditions.In addition, I proxy the strength of each group by the ethnic group’spopulation size as fraction of the country’s total population. In fact, a8Ethnic Power Relations Dataset (EPR-2018)4group’s capacity to overthrow the government may be a function of manydifferent determinants. However, probably the most important determinantis numbers. Rebel leaders with a higher number of co-ethnics are able todraw from a larger pool of potential recruiters and resources.I conduct the empirical analysis using three complementary strategies.First, I use within-country variation, by comparing conflict and inclusionprobabilities for groups of different sizes bargaining over control of the samestate. The probability of inclusion in a leader’s government coalition in-deed follows an inverted-U shaped relationship with respect to group size,which is to the best of my knowledge, a new empirical finding to this lit-erature. The probability of conflict, on the other hand, is low and initiallynon responsive to increases in group size. It increases abruptly beyond athreshold, becoming high for very large groups; again consistent with thepattern predicted by the theory. Results are robust to using leader andcountry-year fixed effects, and the inclusion of a battery of controls.9 Theempirics also support an indirect prediction of the model: larger groups aremore likely to be in conflict if excluded (because the excluded comprise boththose who are too weak to threaten the leader, and those who are too strongto be appeased). Nonetheless, again as predicted, there is no effect of groupsize on conflict for those included. Those included are accommodated viapower-sharing arrangements, and do not want to engage in conflict.Still there is the concern that ethnic groups of different sizes system-atically differ in some unobservable ethnic-level characteristics which areinstead driving the results.10 To address this, the second empirical strat-egy focuses on the sample of split ethnic groups who reside in more thanone country. In this specification, the empirical analysis compares the sameethnic group that has different sizes (as measured by the share of their pop-9Controls included are group size ranking, precipitation, distance to capital, distanceto the leader’s region, terrain, and intra-group fractionalization.10One could also imagine that ethnic groups in conflict were classified more broadlyas a politically relevant ethnic group. If this is the case, groups in conflict would befrom an ethnicity comprising a larger share of population just because of the arbitrarilyclassification in EPR. Comparing the same ethnic definition in two different countries rulesout any bias caused by this5ulation in their resident country). This identification strategy controls forany unobserved ethnic-level trait. I also restrict the analysis to the Africansub-sample, where borders were arbitrarily drawn in the context of the Eu-ropean colonization. Inclusion and conflict probabilities follow the sameinverted U pattern predicted by the model.A third set of reduced form results tests a further prediction of the modelrelating to the ratio between the opportunity costs of conflict and the valueof rents. In the model, when a group decides to enter into conflict, the groupconsiders the rents from office relative to the opportunity costs of conflict.Increases in the benefit to cost ratio affect equilibrium thresholds, makingchallengers more inclined towards conflict and, therefore, more expensive tobuy off. This results in greater inclusion of weaker groups, and simultane-ously greater exclusion of stronger groups, as well as a higher likelihood ofconflict precipitated by them. To test this comparative static, I combinecrop price data with geo-referenced data on cropland maps of each ethnicgroup, to construct ethnic-year price indicies. I use this variable to create aproxy for changes in the relative economic standing of a given ethnic groupcompared to the ruling ethnic group. The empirical findings are consistentwith the model, where price shocks which alter the relative economic ad-vantage of leaders over challengers have differential effects on inclusion andthe incidence of conflict, for weaker compared to stronger groups.In Chapter 4, I structurally estimate the model. In the econometricstructure, I assume that agents observe strength perfectly, but the econo-metrician only observes a noisy measure of strength: relative group size. Ialso allow for measurement error in conflict and inclusion. The model isestimated via simulated maximum likelihood, and closely recovers the mainempirical patterns in the data. The estimated parameters are then used toevaluate several policy-relevant counterfactuals. Each one of them can bepotentially initiated or influenced by international interventions aimed at re-ducing conflict. I evaluate the effects of democratization, military capacity,financial aid, sanctions and quotas.The counterfactual exercises provide the following results. Shifting theleadership to the largest ethnic group (an assumed consequence of democ-6ratization here) has a substantial impact in reducing conflict. This effectis driven by the fact that the strongest groups depart from the positionof challengers. Financial aid has ambiguous effects: it reduces conflict ifit is given to challengers, by increasing their opportunity cost, but it maysignificantly raise the chances of war if it increases rents from office. Onthe other hand, both sanctions contingent on conflict, and government aidconditional on peace, have a powerful effect in reducing the probability ofconflict. Lastly, quotas mandating group representation, though effective inincreasing inclusion, have a negligible impact on conflict.The thesis contributes to a number of literatures. Firstly, it relates to thevast theoretical literature on conflict (e.g Fearon [1995], Powell [2002], Chas-sang and Padro-i Miquel [2009, 2010], Dal Bó and Powell [2009]). In par-ticular, it is an example of a model of conflict resulting from a commitmentproblem (see Fearon [1995], Powell [2004], Powell [2006]). However, the na-ture of the commitment problem here is unique. Previous work has exploredthe lack of commitment to future transfers, when power shifts dynamically,such as when a leader cannot commit to transfer promised amounts to chal-lengers in the event of their becoming weaker via disarming (Powell [2004],Acemoglu and Robinson [2001]). Conflict in the theory studied in Chapter2 arises even when a leader can commit to future transfers. Here insteadtransfers shift the bargaining power between the conflicted parties, and chal-lengers are unable to commit to not utilizing that increased power. Fearon[1996] explores this type of problem; though in a distinct direction to thatdone here. In a dynamic framework, Fearon [1996] models the bargainingof two states over a territory in successive periods, with the current divi-sion determining the state’s military odds in the next period. In contrast,in my model the success probability of a war is a function of how muchpower the challenger enjoys at the moment of the conflict decision. Fromthis approach, I find novel implications that could not be found in Fearon’s,namely the worsening of the commitment problem for the strongest groups.Most importantly, not only do I explain the possibility of conflict, but Ialso rigorously test the new and unique implications of the key theoreticalmechanism against the data.7The empirics exploited in Chapter 3 are related to, and partly informedby, the vast empirical literature on conflict. A number of papers have exam-ined the effect of exogenous economic shocks on the probability of conflict(e.g, Miguel et al. [2004], Bazzi and Blattman [2014], Berman et al. [2017],Berman et al. [2017], Burke et al. [2015]), with no direct evidence of thecausal theoretical mechanisms. In contrast, I am able to identify relativeeconomic shocks and their effects on conflict. Adhvaryu et al. [2018], Guar-iso and Rogall [2017], and Mitra and Ray [2014] are limited examples of suchapproach. Yet, the approach taken here differs due to the particular theo-retical results guiding the empirics. One important and unique distinctionis the estimation of the impact of economic shocks not only on conflict, butalso on power-sharing (which is used to accommodate potential conflicts).Observing how power is shared when conflict does not occur (or at least isoff the equilibrium path) can be particularly informative of the theoreticalmechanisms.11Mayoral and Ray [2017] also explain and document how conflict overpolitical power is initiated by large groups. The explanation is driven bythe assumption that political power generates a public good: given someinitial allocation of public goods, large groups have more incentives to fightbecause the prize will not be split, but enjoyed by the entire group andproduce higher total value. In contrast, here, I allow that a leader canfreely choose a fully transferable allocation of rents to appease challengers.I also extend the analysis by testing the predicted power-sharing decisionempirically.In addition, the thesis contributes to a growing literature on the po-litical organization and institutions of weak states. In particular, I studyhow leaders survive in power when facing challengers who they fear maydisplace them violently (De Mesquita et al. [2005], Gandhi and Przeworski[2006], Egorov and Sonin [2011], Francois et al. [2014], Francois et al. [2015],11Dube and Vargas [2013] also test the implications of a theoretical model predictingviolence levels, in the context of Colombia. In this thesis, I find evidence of a commitmentproblem not studied before in the context of bargaining over rents from political power inAfrica and Asia.8Acemoglu et al. [2008], Acemoglu et al. [2010], Arriola [2009]), and evenfrom within their inner circle (Kudamatsu and Besley [2008]). For instance,Boix [2003] and Acemoglu and Robinson [2001] are concerned with howbottom-up forces shape political transition and stability. North et al. [2009]provides a seminal analysis of societies where violence is limited by politicaland economic manipulation, mostly based on group identities and personalties (resembling the ethnic dimension featured in this paper).Notably, I contribute to the still limited empirical research on the alloca-tion of political power in non-democratic regimes. Evidence of entrenchmentof political elites is found in Indonesia (Martinez-Bravo [2014], Martinez-Bravo et al. [2017]), Sierra Leone (Acemoglu et al. [2014]), China (Francoiset al. [2016]), Russia (Schleiter [2013], Tunisia (Buehler and Ayari [2018]),and Haiti (Naidu et al. [2015]). This study is closely related to Francoiset al. [2015] who demonstrate that African ruling coalitions are large andgenerally proportional to population shares. They show how this may ariseas a peace-inducing response by leaders to violent threats perceived againstthem or their regimes. Conflicts do not occur along the equilibrium pathof that model, and their patterns cannot be addressed with it. Here, thetheoretical model generates conflict along the equilibrium path and estab-lishes a set of new empirical facts regarding how power is shared in bothAfrica and Asia. Acemoglu et al. [2010] studies the incentives of a civiliangovernment to create a strong army to defeat a rebellious faction, with thedownside of increasing the influence and coup opportunities of its military.According to their model, civilian governments endogenously choose weakarmies incapable of ending insurrections provided the rebels are not strong.On the contrary, the model presented here concludes that civil wars tend tobe fought between the strongest groups. As we will see, this accords wellwith observed patterns of civil conflict.Related to this literature, a line of research emphasizes the risks andlimitations of co-opting elites in to power-sharing arrangements due to theadded capabilities that may be acquired from within government (Haber[2006], Magaloni [2008], Debs [2007]). Here, I formalize this problem andderive conditions under which power-sharing attempts will fail. Roessler9[2011, 2016] argue that a leader considers the risk of coups and civil warswhen deciding between including or excluding an ethnic group. He con-cludes that the leader will include larger groups because the risk of civil warbecomes higher than the risk of coups. In this thesis, the insight is similarbut distinct in the following manner: leaders do face a threat from within,but it depends on the share of power enjoyed by the challenger. This keydistinction implies a different relationship between group size and inclusion,which (to the best of my knowledge) has not been documented before.1212In a contemporary working paper, Chaturvedi and Das [2018] also show an inverted-Urelationship between inclusion and population size for Majority Representation systems(MR). Their theoretical approach and empirics is done in the context of a specific demo-cratic institution. In this thesis, I document this relationship for mostly non-democraticcountries (and periods) and its resulting effect on civil conflict.10Chapter 2Bargaining over Power: AModel of Power-sharing andConflict2.1 IntroductionCivil conflicts have been a recurring phenomenon across the globe. Andthey are generally organized along ethnic lines. In particular, these conflictsare launched by ethnic groups excluded from, or underrepresented in, theexecutive power (Cederman et al. [2010]). Why would a leader ever chooseto exclude a group who is likely to respond to exclusion by initiating conflict?A vast theoretical literature attempts to explain why a costly conflictwould occur when there is another, mutually advantageous, efficient alloca-tion. A number of reasons have been proposed by the literature (indivisibil-ities, information or commitment problems). I investigate theoretically therole of an unexplored commitment problem generated by a reasonable im-plication of power-sharing agreements. In the model, power-sharing, whichtypically involves offering positions in the cabinet of a country, may be usedto appease a challenger willing to overthrown the government through con-11flict. However, it also increases the chances the challenger will prevail in theevent of a conflict.Several examples can illustrate this feature. Cabinet appointments entailauthority and discretion over the distribution of resources in key areas. Andstrong evidence of ethnic favoritism on the distribution of expenditure hasbeen found in the literature (e.g., Kramon and Posner [2016]; Ejdemyr et al.[2018]). For instance, Burgess et al. [2015] find that districts occupied byethnic groups with co-ethnics in the cabinet were more likely to be favoredwith road building in Kenya during non-democratic periods. In turn, roadsmay facilitate the mobilization of militias and masses towards an organizedviolence against the government. In fact, the findings of Rogall [2014] areconsistent with this. Alternatively, we may also think of political inclusionas an expansion of state capacity to the ethnic region, including policing,military control, recruitment to the army, public goods, among other ben-efits. At the same time that these included ethnic groups have a strongerpresence of the state, they may also have greater access to weapons andproximity to the security forces. Furthermore, cabinet appointments maydirectly influence the probability of success of a rebellion just because ofbetter access to inside information. Likewise, inclusion raises the possibilityof attacking the regime from within, in which case the chances of successwould likely increase with the share of posts occupied by the group (Roessler[2016]).This effect of power-sharing has been often mentioned in the theoreticalliterature but have not been analyzed before (e.g., Haber [2006], Magaloni[2008], Debs [2007]).1 Others have pointed out how access to the state’scoercive apparatus may be leveraged by ethnic groups in several contexts(Horowitz [1985]; Roessler [2016]).There is no information asymmetry in the model. Indeed, informationproblems, like uncertainty about the value of resources or military capac-ity of conflict actors seem implausible explanations for long intra-state wars(Fearon [1995]). The theory fits the class of models of conflict resulting from1Fearon [1996] investigates a similar problem, but with a different perspective. Seediscussion on the literature in the Introduction.12commitment problems. Most of this literature has focused on the commit-ment problem on the side of the leaders - for instance, when leaders cannotcommit to make future transfers(e.g., Acemoglu and Robinson [2001],Fearon[2004]). I depart from the literature by focusing on a commitment on chal-lenger’s side, motivated by the examples above. The model generates veryunique and distinctive implications that can be tested against the data (anexercise that is pursued in the following chapters).2.2 A Simple Model2.2.1 OverviewI start with a brief overview of the model. A leader of a country enjoysthe prestige and the rents of holding power. Nevertheless, the leader facesa challenger that can potentially overthrow the regime by violent means.Conflict is costly; that is, if leader and challenger fight over control of thegovernment, both parties will face losses. Losses here can be thought of asthe destruction of infrastructure, disruption of economic activities, oppor-tunity costs of resources diverted to fighting, etc. Because of the potentialcosts incurred by both parties, there is the potential for welfare-improvingsharing. For example, the leader could offer the challenger a share of powerroughly calibrated to the leader’s probability of losing the war. This wouldmake both parties better off compared to a conflict situation, since the ex-pected payoffs would approximate the probabilities of victory but the piewould include the associated costs that are lost when a conflict occurs.However, if the leader were to attempt such a power-sharing deal withthe challenger he would also be granting access to resources that could beused to overthrow the leader. A commitment problem thus emerges. Thechallenger cannot commit to not use these extra capabilities against theleader. So the leader must decide between offering a share of power in orderto compensate a challenger at the risk of strengthening them, or opt forstrategic exclusion and face a higher likelihood of war. Next, I present themodel in details.132.2.2 Model Set-upThere is a static bargaining problem between a leader (l) and a challenger(c) over political power. This rules out the possibility of a leader includinga subgroup of challengers and excluding the remainder, for example. It ismotivated by the strong ethnic ties that shape group identity, and that weconceive of as primordial and hence take as primitives for the rest of thisstudy. For a complementary approach that allows groups to be endogenouslyconnected in the midst of conflict see König et al. [2017].The total value of political power is pi, capturing both economic spoilsand personal rents of being in a prestigious position2. The challenger is onlyable to seize power by violent means. If conflict occurs, both the leader andchallenger face losses cl and cc, respectively. The challenger has strength θc,which denotes the challenger’s probability of success from a conflict when itstarts from a position of complete exclusion from power. The leader observesthe challenger’s strength perfectly.Assume that pi is fully divisible so that the leader may offer a share ofrents τ ∈ [0, 1] to appease the challenger.3 Power-sharing, however, increasesthe potency of a challenging group’s threat, if it subsequently decides to rebeland move against the leader. Specifically, the subsequent probability ofchallenger’s success in conflict with the leader is now given by θc+ατ , whenthe challenger receives share τ . The parameter α > 0 captures the extent towhich the likelihood of success in conflict increases as a response to an extraunit of power shared. This is the key novel assumption introduced into themodel and captures a number of important features of reality. Power-sharinginvolves the allocation of cabinet and military positions to a potentiallycontesting group. Holders of such positions are empowered in many ways:they may control key positions in the administration, have some control over2Anedoctal evidence linking office-holding to benefits abounds. Benefits may includewealth, resources, patronage, corruption opportunities, prestige and the prerogatives ofoffice. See Arriola [2009] for discussion.3Divisibility is reasonable in this context. Arriola [2009] and Francois et al. [2015]show the allocation of cabinet posts as means of power-sharing. Powell [2006] and Bidneret al. [2014] show that a lottery can also be used as a way of sharing power in case ofindivisibilities.14resource management within it, have access to (and inside information on)the leader, and/or knowledge about vulnerabilities within the regime, or themilitary. All of these factors, and more, may offer an advantage to the groupwere it to attempt a rebellion. We bound parameters so that θc+α < 1 ∀ c,so that the probability of success is bound below one.The timing of the game is as follows. The leader chooses an offer τ .After observing τ , the challenger decides to rebel or not. If not, the leaderand challenger obtain the proposed allocation (1− τ)pi and τpi, respectively.If conflict is chosen, the winner of the conflict obtains pi minus the costoccurred; the loser just suffers the cost. In expectation, under allocation τfor challenger c, the value of rebelling to the challenger is (θc + ατ)pi − cc,whereas the leader’s expected payoff under a challenge is (1−θc−ατ)pi− cl.Figure 1 illustrates the game.2.2.3 AnalysisWe are looking for a Subgame Perfect Nash Equilibrium. Using backwardinduction, we first look at the challenger’s optimal strategy given an offer τby the leader. The challenger’s payoff of peace is higher than conflict if:τpi ≥ (θc + ατ)pi − cc ⇒ τ ≥ θc − cc/pi1− αWe then define τminc as the minimum transfer needed to be offered bythe leader to appease the challenger:τminc ≡θc − cc/pi1− α (2.1)Note here the nature of the commitment problem. If power-sharing didnot shift the challenger’s capabilities, the leader could appease the potentialrebel by just offering the latter’s probability of success minus a discountassociated with the incurred cost of conflict. However, that is not enoughfor appeasement since the offer also increases the chances of victory forthe rebels. The appeasing offer must be high enough to at least make thechallenger indifferent to going to war, given that the probability of success15also depends on the transfer (θc + ατ). Therefore, the leader must find thetransfer that solves this fixed-point problem. This results in a multiplyingeffect. The minimum appeasing offer is θc−cc/pi (the optimal transfer in theabsence of any commitment problem) multiplied by 11−α , which is increasingin α.It is straightforward to see that, conditional on the challenger’s decision,τ always decreases the leader’s payoff. A positive offer is made only if itinduces the challenger to peace. Therefore, regarding the leader’s optimalstrategy, it is sufficient to evaluate τc = 0 and τc = τminc .If θc ≤ cc/pi, the leader can sustain peace even if offer is zero. Theprobability of a rebel’s victory is so small that the expected gains does notexceed the cost of conflict. If θc > cc/pi, then the leader must decide tooffer zero and face war, or to give just enough to buy peace. The value ofexclusion will be larger than a peaceful power-sharing if:(1− θc)pi − cl > (1− τminc )pi ⇒ (1− θc)pi − cl > (1−θc − cc/pi1− α )pi ⇒θc >(1− α)cl + ccαpi2.2.4 ResultsThis results in a unique equilibrium4 described by Proposition 1.Proposition 1. There exist a unique Subgame Perfect Nash Equilibrium.It comprises the following strategies:i) The challenger rebels if and only if τc < θc−cc/pi1−α ; playing peace whenτc ≥ θc−cc/pi1−α ;ii) The leader offers:τc = 0, if θc ≤ θ∗c ≡ cc/piτc = θc−cc/pi1−α , if θ∗c < θc ≤ θ∗∗c ≡ (1−α)cl+ccαpiτc = 0, if θc > θ∗∗c .4Technically, there exist multiple equilibria in a measure zero set of parameter valueswhere this applies. Specifically, there exist multiple equilibria only if θc is exactly equalto (1−α)cl+ccα/pi. This is a knife edge case.16Because conflict is costly, a weak challenger is not a credible threat.The leader can exclude such a group from power and sustain peace. Chal-lenging groups are able to commit to rebellion as they become stronger,and anticipating this, the threat of costly wars makes the leader willing toshare power to obtain peace. The offer, however, must be high enough todeal with the commitment problem (on the side of the challenger) generatedby the increased chances of success from access to political resources. Theminimum transfer required to appeasement increases substantially with theinitial strength of the group (due to the multiplier 11−α). Beyond a threshold(θ∗∗c ), the challenger is so expensive to buy off that a costly war is prefer-able for the leader. The reason is that stronger groups require progressivelylarger transfers, causing greater shifts of power with their inclusion. Figure2 provides a visual illustration of the equilibrium path.One consequence of this result is the heterogeneity of excluded groups.Both the weakest and strongest groups are left out of coalitions, but fordifferent reasons. The weakest are excluded due to the lack of credible threatthey pose, while the strongest are excluded because the transfer needed tosubdue them is so large and would empower them so much, that it is betterto keep them cut-off altogether.The results also shed some light on the importance of relative economicshocks. An aggregate shock that changes the costs of conflict (cc and cl)and rents from holding power (pi) simultaneously, and in the same direction,has no effect on final outcomes (if the ratios are kept constant). In contrast,relative shocks do affect the thresholds, and change the distribution of groupsincluded and in conflict.Though the mechanism proposed here is simple, it yields quite strongempirical predictions. In summary: (1) there is an inverted-U relationshipbetween strength and inclusion, and an increasing convex relationship be-tween strength and conflict. (2) there is heterogeneity in excluded groups– both very weak and very strong are excluded. And the strong excludedgroups are likely to be observed in conflict. There should be no such re-lationship between strength and conflict probabilities for included groups.(3) relative shocks that decrease the cc/pi ratio move both thresholds to the17left. This implies higher probability of inclusion for relatively weak groups,and higher probability of exclusion, and hence conflict, for strong ones. Inthe next section I elaborate on the empirical setting in which to test theseresults of this model.2.3 Model Extension: Dynamic Framework2.3.1 OverviewI extend the model to a dynamic setting. Differently from the staticmodel analyzed above, the game does not end after conflict or acceptanceof the leader’s offer. Instead, groups move to a next period when the willbe played again. The leader in the next period is determined by the resultsobtained in the previous period. If peace was played, the previous leaderkeeps power and starts next period as a leader. On the contrary, if conflictwas played, the winner of conflict starts as a leader.The main implication of this feature is the following. When decidingto play conflict or peace, a challenger now considers the payoff of being afuture leader. In turn, the payoff of being a future leader will depend onthe strength of the current leader who will in the position of challenger incase he is overthrown. Winning a war against a strong leader implies highercosts of accommodation or a more powerful rebel in the future. Therefore,the strength of the leader will affect the chances of peace. In particular,strong leaders will make the value of conflict lower and groups more likelyto appeased.In the empirical part of the thesis (Chapter 3), I use relative group size(share of country’s total population) as a main proxy for group strength.The dynamic model implies the same empirical patterns suggested by thestatic one: groups of a larger population share face a greater commitmentproblem, and are more likely to be in conflict. The distinction is that thedynamics adds another cause for the relationship. Groups of larger relativesize also face (by construction) challengers of smaller size, which imposes a18lower disciplinary effect, a higher value of seizing power, and consequentlyleads to more conflict. Next, I present the details of this extension.2.3.2 Model Set-upConsider an infinite horizon, discrete time economy, with per perioddiscount rate δ. There are two groups, denoted by A and B, with strengthθA and θB, respectively. At time 0, I assume group A is the leader and groupB is the challenger, without loss of generality. Similar to the static model,groups bargain over the rents pi. The leader (group A) starts by choosinga share τB of rents to be offered to group B. Group B, in the position of achallenger, observes τB, and decides to start a conflict, or accept the offerand stay in peace.If peace is played, groups A and B consume the proposed allocation:(1 − τB)pi and τBpi, respectively. The period ends, the leader keeps power,and the game is played again next period with A as the leader. If conflictoccurs, the probability of rebel’s success is given by θB + ατB, α < 1. Theleader and the challenger face costs cl and cc (costs are indexed by status,and not by group identity). The loser realizes the cost of war, and dies(continuation value is zero). The winner consumes pi minus the associatedcost, and starts next period as the leader. The deceased is replaced by anew player from the same group, who starts as a challenger.The motivation for this last assumption is the following. First, death,imprisonment and exile are the post tenure fate of 80% of leaders replacedirregularly (e.g., by conflict).5 It is reasonable to assume similar outcomesfor challengers. On the other hand, the ethnic demographic composition(one of the most important dimensions along which conflict is organized)does not disappear or even change too much. A new leader will certainlyneed to deal with the demands of rival ethnic groups, and their replacedleadership.5Own calculation using Archigos dataset192.3.3 AnalysisWe are looking for a stationary Markov Perfect Equilibrium, where play-ers decide on the offer (when in the "leader" state) conditional on the strengthof the of the challenging group, and on conflict or peace conditional on thepresent offer (when they are in the "challenger" state).The value of being leader for player i ∈ [A,B] is given by:V li,t = max{τj,t}∞t=t{((1− τj,t)pi + δV li,t+1)(1− Cj,t(τj,t))+((1− θj − ατj,t)(pi + δV lj,t+1)− cl)(Cj,t(τj,t))},where Cj,t(τj,t) ∈ (0, 1) is the probability of conflict decided by player j 6= iconditional on observing an offer τj,t, in period t. The leader i chooses τj,tto maximize his payoff, taking into account player j’s response.The value of being a challenger for player j is given by:V Cj,t = max{Cj,t(τj,t)}∞t=t{V CCj,t (τj,t), V CPj,t (τj,t)},where V CCj,t (τj,t) and V CPj,t (τj,t) are the payoffs of playing conflict and peace,respectively.The values of peace and conflict are given by:V cpj,t = τj,tpi + δV cj,t+1V ccj,t = (ατj,t + θj)(pi + δV lj,t+1)− cjNote that in the static model the conflict decision was solely based onthe challenger’s strength and the offer made. The key distinction here fromthe static model is that the incentives for conflict for, say, group A now alsodepend on its continuation value of being a leader. This, in turn, dependson the strength of group B (since, as a new leader, group A will face thethreat of the deposed group B). In turn, the incentives of conflict for groupB will also depend on the strength of group A, and so on.20The conflict decisions impose discontinuities on the value function. Thevalue of being leader assumes distinct functions of the parameters dependingon the strategies played in the subgame. Therefore, we must solve the modelby finding conditions for each possible equilibrium scenario. In the AppendixC, I find the solution for every possible equilibrium.2.3.4 ResultsThe results are summarized by Proposition 2.Proposition 2. There exist a unique stationary Markov Perfect Equilib-rium. It generates the following outcomes on the equilibrium path for aleader A and challenger B:i) If θB ≤ θ∗B: Leader A offers τ∗B = 0, and challenger B plays peace(C∗(τ∗B) = 0);ii) If θ∗B < θB ≤ θ∗∗B : Leader A offers τ∗B > 0, and challenger B plays peace(C∗(τ∗B) = 0);iii) If θB ≥ θ∗∗B : Leader A offers τ∗B = 0, and challenger B plays conflict(C∗(τ∗B) = 1),where θ∗B and θ∗∗B are functions of the model parameters (δ, α, cc/pi, cl/pi),and the other group strength (θA).In summary, a weak challenger is not included and keeps peace. As itbecomes stronger, power-sharing is required so the group can be appeased.The leader is willing to do that since conflict destroys some surplus. Asthe challenging group becomes too strong crossing a second threshold, theleader will prefer to exclude the group altogether and face conflict. Again,similar to the main force of the static model, power-sharing increases theconflict capabilities of the opposition, and creates a commitment problem.This commitment worsens for the strongest groups, since they are the oneswho demand the highest offers.The reader can follow the derivation of conditions for each equilibriumfor the specific values of θ∗B, θ∗∗B and τ∗B (Appendix C). The functions presentdiscontinuities and can be very long and intractable expressions (that’s whythey are not presented here). Under some parameter space, the first thresh-21old (θ∗B) may depend only on the cost of conflict (if, for instance, the chal-lenger B does not share power or face conflict in the leader state). If groupA is strong enough to be appeased with some share in equilibrium whengroup B is in power, then θ∗B will also depend (negatively) on α and θA.This is because they are determinants of how much group B will have toshare if he becomes leader through conflict. Finally, consider that θA is sohigh that the equilibrium path for a leader B would be the exclusion and theconsequential conflict from group A. Then θ∗B will also depend on, but willbe a different function of, θA. In the former case, θA affects the sharing thatthe group would give away in the state of leader; in the latter, it would affectthe payoffs of being a future leader through the costs of war that occurs inthat state.The main novel result from the dynamic model is the elucidation ofthe importance of the leader’s strength. Formally, it can be shown that∂θ∗B∂θA≥ 0, ∂θ∗∗B∂θA ≥ 0, and∂τ∗B∂θA≤ 0. That is, the strength of the leaderimposes a disciplinary effect on the challenger’s behavior. A group whowants to depose a strong leader anticipates the high costs of accommodationor conflict in the future. This decreases the incentives for conflict, makesthe group cheaper to be appeased, and raises the likelihood of peace.In the empirical analysis, I compare ethnic groups of different sizes as asource of variation for group strength. The main idea is that a bigger groupB is in expectation stronger than a smaller group A. In the light of thedynamic model, there is another difference. Group B is not only stronger,but also faces a weaker challenger (namely group A). Similarly, group A isnot only weaker, but also faces a stronger challenger (namely, group B). Inany case, the key predictions of the dynamic framework supports the onesof the static model. There is just one additional channel in play besides theown group strength: as the other group becomes weaker, the incentive forconflict grows, initially increasing the likelihood of inclusion, but eventuallyleading to exclusion and conflict.222.4 DiscussionI briefly discuss some modelling assumptions here, and some possibledirections for future research.Functional form I assume the probability of success in conflict isa linear function of the leader’s offer. If this function were convex, higheroffers would change rebel victory probability even more abruptly. This wouldmagnify the commitment problem. On the other hand, if the function wereconcave, the marginal effect of the offer would be decreasing. Therefore, asgroups become stronger, the marginal effect of increasing the offer is lower.However, strong groups still receive larger shares and have their strengthincreased more. The concavity operates just in the margin. In this scenario,strongest group are not necessarily the ones excluded and in conflict. Therewill be two forces: one given by nature of the model, and the other by thefunction form. And they operate in opposite directions in generating theexclusion and conflict from the part of the strongest groups.That being said, the same results are still derived if the functional formare θc + ατ1/2 or θc + ατ1/3, for example. In addition to that, strong con-cavities seem unreasonable in this setting. It would imply, for instance,that group strength would increase substantially with the first cabinet seatsoffered, but less if the group already had other seats.In chapter 4, the model is structurally estimated assuming the linearspecification. It would be possible estimate a more flexible function. Never-theless, the linear specification provides a good and parsimonious fit. FigureA7 shows how the allocation of cabinet appointments has a closely linear re-lationship with population share (the key proxy for group strength). Figure10 shows how the structural model closely recovers the empirical patternsof the data (Figure 3).Secessionist wars The model is explicitly elaborated to explain con-flict over the control of the government. In particular, the empirical analysisin the next chapter relies on the availability of data on the divisions of power,and a good proxy for the capabilities to overthrow the government (popu-lation share of the ethnic group).23The mechanism proposed in this chapter could also be present in conflictsover territory. In this case, a government could expand regional autonomyto an ethnic group, but such concession would possibly empower the groupeven more in its attempt of secession. Nevertheless, I do not test for thiscase empirically in this thesis. The main reason is that population share -the key proxy for strength in governmental conflict - is not a good proxy forsecessionist wars. In this type of conflict, groups do not need to mobilizetowards the capital and overthrow the regime. Instead, they simply need tokeep control of a piece of the territory (generally, the ethnic homeland).Yet, there is a way to think about conflict over territory in the contextof bargaining over the executive power. Whatever game leaders and ethnicgroups are playing when dealing with the autonomy of their region, this willbecome an outside option for ethnic groups. Leaders could potentially offersome sharing of power to accommodate groups for that reason. However, thesame implication of the model is kept. Ceteris paribus, groups stronger intheir capabilities to overthrow the government require higher sharing (andconsequently, enhance commitment problems).From a similar perspective, weak groups (in their capabilities of over-throwing the government) expect very little from the bargaining over therents of the executive power. For these groups, the outside option of fight-ing for regional autonomy is higher. Indeed, empirically secessionist conflictsare commonly led by small groups.Multilateral bargaining Most countries have multiple ethnicgroups. The application of this model to a setting with multiple groupsassumes that leaders bargain with each group individually. There are a fewreasons why this is an acceptable assumption in this context.First, the key motivation for this thesis is that conflicts are generallyorganized along ethnic lines. Groups tend to be concentrated over differentparts of the territory, and organize politically along this dimension. As aconsequence, even the structure of the data (to be presented in Chapter 3) isdyadic in nature, and identify conflicts fought by particular groups (and notby a collusion of groups). In fact, 71% of episodes of ethnic governmentalconflicts in Africa and Asia are launched by a single group. When more than241 group is involved, it is never the case that the conflict was launched byall groups at the same time. For this reason, a natural theoretical directionhere is to understand the difference between the outcomes of each bilateralbargaining separately.Second, the model presented here posits a specific relationship betweengroup strength (which, again, is proxied by group size in Chapter 3) and finaloutcomes. It can be shown that conditional on group’s own size, the size ofother groups has no explanatory power for conflict and inclusion outcomes,confirming that I am probably focusing on the most important elements ofthe question in hand.Given these facts, focusing on the commitment problem in a bilateralbargaining environment seems to be a plausible first approach to understandthe exclusion and conflict of ethnic groups. Indeed, this exercise yields novelinteresting and strong results to be tested. However, exploiting this gamein a multilateral setting can be a very fruitful avenue for future research.2.5 ConclusionIn this chapter, I have built a model based on a commitment problemthat has been often mentioned and recognized anecdotally, but had notbeen modelled before. The key feature of the model is that power-sharing(though it is meant to appease) increases the strength of the opposition.Strong implications are derived: leaders obtain peace from weaker groups,but the commitment problem worsens for the strongest groups.The theory predicts a U-inverted relationship between group strengthand its probability of inclusion in the coalition. Moreover, conflict prob-abilities do not react immediately to increases in strength since groups ofintermediate strength are accommodated in power-sharing agreements. Nev-ertheless, when they are too strong, they are the ones likely to be excludedand in conflict. Additional implications with respect to relative economicshocks are also derived. Relative economic shocks against a group leads tomore inclusion if the group is weak, but to more exclusion and conflict if itis strong.25In the next chapter, I elaborate on the ideal setting to test the keypredictions of the model.26Chapter 3Power-sharing and Conflict:Empirical Evidence fromAfrica and Asia3.1 IntroductionI test the predictions of the model developed in Chapter 2 in the contextof ethnic conflict over control of government using the EPR-2018 dataset(described in detail below). There are two reasons to focus on ethnic groupsas the unit of the analysis. The first is data availability. As described below,we are able to observe for each ethnic group: their degree of representationin the executive (inclusion), their representation by rebel organizations inconflict with the state, and a good proxy for group strength (populationshare). The presence of all these elements provides an ideal setting to studyempirically the importance of the proposed theoretical mechanism.Second, ethnicity plays a key role in politics in Africa and Asia (Brown[2003], Roessler [2016]). This does not mean that ethnicity is the only ornecessarily the most important social identity in all contexts, neither doesit necessarily imply that ethnicity is the unique dimension political mobi-lization. Yet, ethnicity is frequently the channel through which competition27over wealth and power is expressed (Bayart [2009]). Shared ethnicity (basedon religious, language, or ancestry commonality) facilitates cooperation anddetermines the extent of influence of political elite actors. Ethnicity oftenbecame even more salient as a result of colonial institutions which oftengranted to tribal leaders control of important state resources. Moreover,upon independence, may of the new post-colonial states combined an ab-sence of strong coercion capacity with strong ethnicity based political insti-tutions. This resulted in an political environments where power could wasnaturally contested by elites based on ethnicity identification that leveragedthe strength of their ethnic networks.Given the predominance of this historical pattern in African and Asiancountries, and the unique preponderence of ethnic conflicts in these countriessince World War II, I restrict the analysis to Africa and Asia in what follows.3.2 Data: EPR-2018The Ethnic Power Relations (EPR-2018) datasetwas originally intro-duced by Cederman et al. [2010] and provides annual data from 1946 to2017 on politically relevant ethnic groups, their relative sizes (as share ofthe country’s population), and their access to executive state power in allcountries of the world with population of at least 250,000 (Vogt et al. [2015]).Previous literature had analyzed ethnic conflict largely ignoring the roleof the state, focusing, for instance, on the effects of ethnic fractionalizationor polarization (Easterly and Levine [1997], Montalvo and Reynal-Querol[2005]). In contrast, Cederman et al. [2010] argue that the state is not aneutral actor, but in fact the central object of, and participant, in ethnicconflicts. In this context, the state is analyzed as an institution captured bythe representatives of particular ethnic groups, and ethnic wars are conceivedas the result of their competing claims over it. This understanding exactlymotivated the collection of this very novel dataset with the goal of measuringthe ethnic configuration of the state.The data collection involved nearly one hundred country and regionalexperts who were asked to identify the ethnic categories most salient for28national politics in each country. For the purpose of this data, ethnicity isbroadly defined by any sense of commonality based on a belief in commonancestry and/or shared culture. This may include common languages, phe-notypical features, and religion. A group is considered politically relevant,and thus observable in the dataset, if at least one active political organiza-tion (for example, but not necessarily, a political party) claims to representthe group’s interests in national politics, or if its members are subjected tointentional state-led discrimination in the domain of public politics. Indirectdiscrimination, in the form of disadvantages in the education or economicspheres for example, are not included in this definition. A2 shows a map ofpolitically relevant groups in Kenya, as an example.All politically relevant ethnic groups were categorized according to thedegree of access to executive power obtained by those who claimed to repre-sent them1. Depending on where political power is effectively exercised, thiscan be the presidency, the cabinet, and senior posts in the administrationof ostensibly democratic regimes; the army command in military dictator-ships; or the ruling party leadership in one-party states. Importantly, datais coded based on groups’ absolute access to power, rather than on theirunder- or over-representation relative to demographic size.The EPR dataset records the political status of each group categori-cally as either: "monopoly", "dominant", "senior partner", "junior partner","powerless", "discriminated" or "self-exclusion". The first two categories de-scribe groups that control executive power alone largely on their own. Incontrast to monopoly, dominant indicates that there is inclusion of "token"members from other groups who do not, however, have any real influence ondecision making. Senior and junior partners are indicators of power-sharingarrangements that are delineated according to their level of influence. Thisis measured by the number, and importance, of the positions controlledby group members (again, irrespective of group size). The remainder areexcluded groups. The powerless delineate the groups that are simply not1Access to legislative and judicial institutions are disregarded. In the African literature,it is well established that the executive is where main power, both political and economic,resides (Francois et al. [2015], Posner [2005])29represented, discriminated indicates groups subject to active, intentional,and targeted discrimination by the state in the domain of public politics.An example of which is the prohibition against holding or contesting for po-litical offices. The final category of self-exclusion applies to groups who haveexcluded themselves from central state power, in the sense that they con-trol a particular territory of the state which they have declared independentfrom the central government.3.2.1 Key variablesInclusion For the empirical analysis, it is necessary to identify theleader, the challengers, and whether each challenger is included or not in thecoalition. Given the structure of political status described above, I assigna group as "leader" if it is coded as "monopoly", "dominant", or "SeniorPartner". I assign a challenger as included in the coalition if its coded as"junior partner". Lastly, "powerless", "discriminated", and "self-excluded"are assigned as excluded.2 Figure A3 shows a map of politically relevantethnic group in Africa and Asia, according to the power status assignedby the EPR dataset. Figure A4 presents the same map highlighting theirclassification into leaders, and included or excluded challengers.Conflict The dataset also links groups to rebel organizations in theUCDP/PRIO actor database (ACD2EPR)3. Conflict is attributed to an eth-nic group if the actor of violence recruits from, and claims to, represent thatgroup. The UCDP/PRIO dataset identifies conflict as a contested incom-patibility that concerns government and/or territory where the use of armedforce between two parties, of which at least one is the government of a state,results in at least 25 battle-related deaths in a calendar year4. There aretwo important clarifications here. First, I use only ethnic conflict. In thischapter, the unit of observation is the ethnic group and, therefore, conflictis not coded as having occurred if the rebel organization does not recruit2Self-excluded groups comprise a very small fraction of the dataset (1.51%). Resultsare similar if these groups are dropped.3See Wucherpfennig et al. [2012] for a reference.4UCDP/PRIO Armed Conflict Dataset Codebook. claim to represent any ethnic group. Second, I restrict attention to civilconflict over the government, where the contest between actors concerns ei-ther replacement of the central government or a change in its composition.The measures of transfers (power conceded) and strength (group size) arenot appropriate to study conflict when incompatibility concerns territorysuch as that over inter-state wars, or where a group demands secession orautonomy. The key conflict incidence variable is then coded as 1 for allobservations in which there is an ongoing conflict (0, otherwise).5Strength In the model, the innate strength is the probability of suc-cess if a group starts a conflict against the leader’s group (absent of anypower-sharing). Note also that the final potential probability of success in aconflict is a function of this pre-determinant innate strength and the shareof rents offered by the leader. The ideal measure of innate strength wouldcapture the former, and should be insensitive to the later.Generally speaking, we could conjecture that the probability of a rebel’svictory is a function of two main inputs: manpower and resources. The firstprovides for the necessary numbers and popular support, while the secondgrants firepower.I then use group size, measured by the ethnic share of the country’s totalpopulation, as a proxy for group strength. Population share is by no meansa perfect measure of group strength. Yet, it is a key determinant of thebargaining power of ethnic elites contesting the spoils of the state. As statedby Roessler, in natural states, elites are only as powerful as their supporternetworks6. A larger group means a larger pool of potential recruits andsupport, a larger share of territorial occupation, greater access to naturalresources, and ultimately a greater capacity of mobilization. In addition,the political claims of larger ethnic groups can enjoy more legitimacy andattract more support (Cederman et al. [2010]). Ultimately, group size is themost realistic proxy for manpower.5Irredentism may be categorized as either intra-state secessionist conflict or inter-stateconflict depending on the main actor initiating the conflict. These conflicts are not usedin the analysis.6Roessler [2016]31Other measures reflecting the resources available for each group sufferfrom significant caveats. For example, income or access to roads could alsobe determinants of success in a conflict. However, they are endogenous innature. They may be the result of the access of their co-ethnics to power(a key outcome of the model), and also reflect opportunity costs (anotherinput of the model). In contrast, population share is not subject to politi-cal decisions, and had been largely determined by historical processes thatanticipated the formation of these states. Furthermore, my identificationstrategy is extended to an ethnic-fixed effect specification, where I exploitthe quasi-natural size variation in groups split across national borders.7Population share also has some convenient properties that aid in em-pirical investigation. First, it is a continuous measure, which is importantbecause of the non-linear effects that we wish to test. Second, it is variabledefined over a broad range, including both very weak and very strong groups(the dataset exhibits groups with population share as low as 0.01%, and ashigh as 98%). Third, the population share measure is available for everygroup in the data, and is consistently calculated and not greatly subject tomeasurement error. Other measures that could proxy for group power, suchas pre-colonial centralization, access to natural resources, presence of coeth-nics in neighboring countries, are dominated in at least one of the aspectsabove, and are definitely weaker determinants of manpower.GeoEPR The GeoEPR (Wucherpfennig et al. [2011]) provides geo-spatial information on the settlement patterns of groups. This map of ethnicgroups is used to construct several geo-referenced pieces of information atthe ethnic-level. In particular, I use the map to identify crops that eachgroup grows, which is used in the identification strategy of relative economicshocks.Additional variables I include a battery of controls in several ro-bustness checks throughout the analysis: leader’s ethnic group size, within-7Another threat would be a case where population share varies as a result of conflict.As argued later in the empirical strategy, such case is unlikely in this setting though.Although, there is a reasonable amount of deaths in civil conflicts, it is very small tocause any real change in the ethnic composition of the population, let alone change thesize order of the groups.32country size ranking of groups, distance between the ethnic homeland andthe national capital, distance between the group’s and the leader’s homeland,different measures of precipitation and drought, measures of elevation andmountainous terrain, within group fractionalization and light density. Thesevariables are all constructed in the ethnic level. They are obtained (directly,or indirectly by using own calculations) from the GROWup project. Theproject uses geo-referenced information from several sources and constructethnic-specific variables based on the settlement patterns of ethnic groups(Girardin et al. [2015]).8 In the Appendix, I provide more details on eachvariable and the estimation results using them.3.2.2 Descriptive overview of the dataTable A1 presents a general picture of the EPR dataset, indicating thenumber of politically relevant groups per country, the distribution of groupsby political power, and the incidence of conflict. Countries have 4.7 relevantgroups on average. For the analysis, at least two groups are needed as weneed to identify a leader and at least one challenger, and then estimatethe probability of inclusion of the challenger conditional on its strength.Conditional on having at least two politically relevant groups, we obtain5.5 ethnic groups per country: 2.5 groups comprise the leader and includedon average, while 3 are excluded. Ethnic conflict over the control of thegovernment (the conflicts we are uniquely focused on) occur in 5.7% ofcountry-year observations, and 38.1% of African and Asian countries had atleast one episode historically, suggesting that this is a prevailing and criticalissue in many settings. Leaders who do not share power (monopoly anddominant) are around 9% of groups; 10.9% of groups are leaders in power-sharing arrangements; included groups comprise 25.4%; the remaining 54.6%are excluded, from which a substantial amount are discriminated against(36.6% are powerless and 16.5% discriminated).8Data can be accessed at DiscussionThere are some caveats worth mentioning here. The first is selection.Non-politically relevant groups are not observed in the dataset. This impliesthat we observe relatively stronger groups that have managed to exert atleast some political influence. The counterpart of this is that there are agreat number of small groups with limited or no representation in power (andwho we would predict to never be in conflict) in the dataset. This yieldsenough variation to test the implications of being a weak group. Twenty-five percent (25%) of ethnic groups in the data have population shares of1% or less; 8.7% of groups have a size less than or equal to 0.1% of the totalpopulation.Next, the definition of ethnic groups is broader than usually used in otherdatasets (e.g. GREG; Francois, Rainer and Trebbi, 2015). The reason forthis is that even though there is a multitude of ethnic groups, they tend toalign in to more general categories when bargaining at the level of nationalpolitics. The same is true for militias organized along ethnic lines with thegoal of taking control of the state.Furthermore, we do not know how and why these ethnic groups be-came politically relevant. Here, I take the ethnic group unit as given. Thekey assumption for the identification strategy of the next sections to makesense is that differences in population share between two politically relevantethnic groups affects inclusion and conflict (only) through their underlyingstrength, and not through any unobserved factors that may have resultedin those groups becoming different sizes. This assumption seems to be par-ticularly justifiable for the countries of our sample. In most cases, ethnicidentities were predetermined well before colonization and the formation ofthe independent national states.Despite these minor caveats, the dataset provides an ideal setting tostudy empirically the causes of conflict. The reason is the availability ofmeasures for all key elements of the theory: conflict, transfers, and groupstrength. Consequently, we are able to test not only implications over actual34conflict, but also over power-sharing (when conflict occurs off the equilibriumpath) - which is a significant contribution to the conflict literature.3.2.4 Summary StatisticsTable 1 presents the summary statistics of the sample used in the esti-mations. The sample include non-leading groups only (see discussion on theEmpirical Strategy below). Out of the sample of these challengers, 31.7%are included in the government coalition. The other key outcome is conflict,and around 2.07% of the group-year observations are in rebellion aiming tooverthrow the government.The key proxy variable for group strength, and the main regressor, isgroup size. The average size of a non-leading ethnic group is around 9.6%,varying from as low as 0.01% to as high as 98% .3.3 Empirical Evidence3.3.1 Within-Country VariationEmpirical StrategyThe model posits a unique relationship between the strength of chal-lengers, the transfers received, and the incidence of conflict in equilibrium.To test such a relationship, the ethnicity of the leader and of the challengersmust be defined. As described in the data section, I assume the ruling ethnic-ity is the most powerful group in the country-year (i.e monopoly, dominantor senior partner, necessarily). Given the EPR coding procedure (see above),the most powerful group in the country is the one holding the most influ-ential and decisive positions, and very likely will be the leader’s ethnicity.99There is a strong correlation between the most powerful group in the country in EPRand the ethnicity of the leaders in Francois, Rainer and Trebbi (2015). The few exceptionsof discrepancies are mostly ambiguous, for example: i) leader’s ethnicity (FRT) is notpolitically relevant, but targets support from other politically relevant groups (EPR); ii)In Democratic Republic of Congo, Laurent-Désiré Kabila (Luba ethnic group, and codedas so in FRT) led ethnic Tutsis (EPR) in overthrowing Mobutu; iii) In Nigeria’s history,35For the remainder of the (non-leader) groups, I estimate the probabilitiesof being included in the government coalition (as Junior Partners), and ofconflict, conditional on population size:Ycet = β1sizecet + β2size2cet + γc + γt + cet, (3.1)where Ycet is conflict or inclusion, dummy variables of ethnicity e, in countryc, at period t; γc and γt are country and year fixed effects; and sizecet is thegroup size measured by the ethnic share of the country’s total population.The identification strategy relies on comparing groups of different sizesin the same country. The key assumption is that differences resulting ininclusion and conflict are caused by the strength of the group. So, forexample, if group size is correlated with natural resources, or the portionof territory occupied historically, there will be no threat to identification aslong those correlates affect conflict and inclusion through group strength.The insight of the identification is to compare how contenders for powerin the same country (same rents from office), under the same institutionalsetting, differ in the likelihood of being appeased as a function of the size oftheir ethnic network and capabilities.Note there is a selection here. By excluding the leader from the regres-sion, we are comparing groups that may not be strong enough to hold powerin the first place. The identification is still valid if under this condition biggergroups are stronger. As it will be shown, the findings are inconsistent witha different a story where bigger groups would be weaker conditional on notbeing leaders. Certainly, the relationship between strength and populationshare conditional on not being leader is different than the unconditional one(and so the estimated coefficients). However, it is the former relationshipthat matters for us, given that leaders do not launch conflicts.A final observation is that most of the within-country variation comesfrom comparing group of different sizes, and not from within-group variationover time. Group size is rarely updated in the dataset and generally onlyPresidents of different ethnicities (FRT) were arguably chosen or kept in check but themilitary dominated by the Hausa-Fulani.36marginally so. One possible concern is that conflict episodes may themselvescause variation in group size. But this is not the case. First, group sizeis rarely updated during conflict periods (because it requires populationsurveys/census updates). Second, the number of deaths related to conflictare not high enough to cause significant changes in the population share of anethnic group. Estimates of battle-related deaths in civil conflicts between1989 and 2017 (from UCDP) point to an average of 1,163 causalities perconflict-year.Main ResultsBefore presenting the results with country fixed-effects, I show the pureunadjusted relationship between population share and the likelihood of in-clusion and conflict. Figure 3 fits a nonparametric (kernel-weighted localpolynomial) regression of inclusion and conflict on group size. 10 It depictswhat is, to the best of my knowledge, a novel empirical pattern. The pat-tern closely resembles the established theoretical results (see Figure 2). Theprobability of inclusion (positive transfers) increases, while the probabilityof conflict is virtually unchanged as a group becomes stronger. This pro-ceeds until a threshold, beyond which the probability of inclusion declines,and the incidence of conflict spikes.The same relationship holds comparing groups within the same coun-try. Table 2 reports results from estimations that include country fixed-effects and a second-order polynomial of group size. The likelihood of beingincluded in the governing coalition has a hump-shaped relationship withgroup size, implying that leaders concentrate more power when faced withan opposition supported by a very large fraction of the population (column1). The peak of inclusion is reached when group size is around 40% of thepopulation. Similar to the nonparametric regression, the marginal effects ofgroup size on conflict are initially small but spike for the largest groups. Theeffects of group size and group size squared are jointly significant (column2). Column 3 shows the result of regressing conflict incidence on group size10Figure A5 includes binned scatterplot37squared only (without group size). The estimated coefficient is large andsignificant.11 12Robustness checksMinorities in power Another alternative explanation for the resultsis the presence of minorities in power. The identification strategy comparesgroups of different size within the same country. However, by construction,the sum of population share of all groups must be less than 100%. There-fore, the outcomes of a larger group may be driven by the fact the remaininggroups are smaller. In particular, a non-leader of large size necessarily im-plies a minority ethnic group in control of the state. One could argue thatconflict may be the result of something special about minorities in power(for instance, if minorities adopt more discriminatory policies and excludemore groups). Nevertheless, I can control for the size of leader. Since mi-norities can be in power and face challengers of different sizes, we can stillhave variation in both the size of the challengers and the leader size. Table5 shows that the main relationship between group size, and inclusion andconflict, is robust to controlling for leader size.Leader and Country-year fixed-effects Part of the variation fromthe estimation with country fixed-effects comes from shifts in leadership.For example, assume one country with two groups: A and B. The specifi-cation with country-fixed effects compares the outcomes for A when A isthe challenger and B is the leader, with the outcomes for B in a differentperiod, when A is the leader and B is the challenger. It is clear that, bydoing so, the leadership is not constant and the outcomes are not necessarilycompared in the same period. In Table 6, I show the results are robust toincluding leader and country-year fixed effects. In practice, we are removingfrom the analysis countries with only 2 groups (because they have only onechallenger per period). Although these cases provide an important source11Table 3 reports the results of a regression which includes a 3rd order polynomial ofgroup size. The implied pattern remains the same.12Figure 4 plots the predicted value of inclusion on group size, assuming value 0 for thefixed-effects.38of variation, since they are the ones likely to have the large groups, we arestill able to find significant effects without them.Controls The results are also robust to the inclusion of a batteryof controls.13 Assume for a moment that our findings are picking up theeffect of some other(s) determinant(s) of inclusion or conflict. For this to betrue, this other determinant should either have the established non-linearrelationships with inclusion and conflict, or a non-monotonic associationwith group size. I estimate the main specification including several possibledeterminants of inclusion and conflict.One confounding factor would be if leaders make inclusion decisionsbased on the size ranking of groups, and not necessarily on their purestrength (Table A3). Leaders might also include groups based on theirdistance to the capital, or the leader’s own homeland (Table A4), and possi-bly choose to fight against big groups that are far away (e.g., Buhaug et al.[2009]).I also control for precipitation and frequency of droughts (Table A5),which has a long established relationship with conflict in the literature.Terrain ruggedness is another important correlate of conflict, state capacityand ethnic diversity (Jimenez-Ayora and Ulubaşoğlu [2015], Michalopoulos[2012], Fearon and Laitin [2003]). In Table A6, results are shown includingelevation’s standard deviation, and the proportion of mountains. Further,regressions in Table A7 include measures of intra-group religious and lan-guage fractionalization.A last concern addressed here is the underlying relationship betweenstrength and income. Groups might be included, or in conflict, as a resultof their income. Adding income in the specification raises two major issues.The first is data availability. This can be overcome by using night lightdensity in ethnic homelands. But this is available only from 1992 onward,which reduces the power of our estimation. Second, and more important, in-come is a classically problematic control to include in this case since conflictincidence may have a meaningful impact on income. With this caution in13Summary Statistics for all controls are reported in Table 139mind, results are reported in Table A8. The results persist with coefficientslargely preserved when these additional controls are added.Alternative ExplanationsStrength concave in group size One threat to the results is thatactual group strength may be concavely related to a group’s populationshare. For example, very large groups may face more coordination prob-lems. Alternatively, they may comprise a larger share of the population asa symptom of their inability to organize politically to a finer more coherentlevel. Similarly, large challenging groups may be particularly week, sincethey were not able to become leaders despite their size. Against this pos-sible concern, if the reason for exclusion of big groups is indeed their lowerstrength, we should observe conflict responses similar to the small groups.But we observe something entirely different. While the likelihood of inclu-sion decreases for the biggest groups, the probability of conflict increases.This suggests that the effect is not driven by the fact that they are weaker,but that groups are excluded despite being stronger. In fact, the model pro-vides a mechanism rationalizing why the leader would exclude such groups- that is, because they are stronger. Furthermore, even if the largest groupsare politically weak in the sense that they were unable be leaders, their bignumbers can still affect the success of a rebellion directly.Large groups included as Senior Partners Another concern thatcan be raised is that strong groups are included as senior partners, and thatis why we observe a decline in the probability of their being junior partners.This is unlikely for three reasons. First, Senior Partner and Junior Partnersare coded taking into account not only the number of positions, but alsotheir importance. Therefore, it is the Senior Partners who are most likelyto be the actual leaders taking decisions on the allocation of cabinet posts.In fact, there are very few cases in which more than one Senior Partner ispresent per country-year (around 3%). Taking these, even if I assign thelargest senior partner as an included group, the results persist - I still finda high probability of exclusion for large groups in general.40Second, if large groups were being appeased by the offer of senior partner-ship, we should not observe an increase in conflict. This last piece suggeststhat, in fact, those groups are not being fully accommodated.Third, this alternative explanation cannot account for the fact that we donot observe excluded groups moving from exclusion to inclusion positions.14To reinforce this point, Table 4 presents results for the estimated probabilityof discrimination conditional on group size. Consistent with the previousresults, the probability of discrimination is high for small and big groups,and lower for groups of intermediate size. Thus, bigger groups who are notleaders not only have their chances of inclusion reduced, but their chancesof state-based political discriminated are also increased.Cold WarCivil conflicts across Africa and Asia are known to have received signifi-cant international interventions, and may be a result of the struggle betweenforeign powers and their attempt of political influence. This was the casemainly during the Cold War period, when the United States and SovietUnion could potentially instigate rebel movements in order to replace in-cumbent governments with aligned regimes. Definitely, the proposed modelis not the only explanation for civil conflicts, which can also be the result ofsuch geopolitical factors. However, we still must conjecture whether thosefactors may explain the patterns in the data that I claim to be the result ofthe key mechanism of the model.First, foreign interventions are not a concern to the empirical test ofthe model if they are occurring based on country characteristics aside ofthe ethnic distribution (since we use country fixed effects and use withincountry variation in the ethnic population share). Second, it is also not aconcern if the additional strength leveraged by the group with the supportof foreign allies could be accommodated in the power-sharing agreements.In this case, we could think of foreign alliances as one of the many factors14Figure A6 shows the composition of power status by group size. It can be seen thatas groups become larger they are more likely to be leaders. However, we do not see thesame pattern in terms of mobility from exclusion to included groups.41influencing group strength for which we observe only a proxy (populationshare). Nevertheless, it would be a threat if conflicts are initiated by foreignpowers, and those foreign powers only pick the largest ethnic groups - withprobabilities increasing in the population share in a convex form.In Table A9, I report the results splitting the sample between the periodof 1946-1991 and 1992-2017. The key empirical patterns are the same forboth periods, ruling out the hypothesis of the Cold War being the maindriver of the results.Further PredictionsEffects of group size conditional on status Here, I test one ad-ditional model prediction. As explained above, the leader’s decision impliesheterogeneity in excluded groups. They are composed of those who are tooweak to have any credible threat, and those who are too strong to committo peace if included. We should thus expect a positive relationship betweenour measure of strength (group size) and conflict for excluded groups. Butwe should not expect this for included groups whose additional strength isaccommodated by larger transfers in the power-sharing arrangement.15Table 7 presents the results of the empirical test of this prediction. Col-umn 1 shows the impact of group size on conflict probability conditionalon the group being excluded from the coalition. As always, specificationsinclude country and year fixed-effects. Consistent with the model, the ef-fect of group size is positive and statistically significant. In columns 2 and3, results are reported for only included groups. In column 2, leaders areincluded in the sample, while in column 3 I restrict the sample to only non-leaders included as junior partners. For these groups, the impact of groupsize is much smaller and not distinguishable from 0.15Francois et al. [2015] collect detailed information on the ethnicity of ministers for asample of African countries. I combine the EPR data with theirs to obtain the share ofcabinet appointments for each included group. Indeed, I find a strong positive relationshipbetween group size and the share of appointments for included groups, which is alsoconsistent with the model’s prediction. The results are reported by Figure A7423.3.2 Split groupsA series of robustness checks ruled out alternatives explanations. Still wemight conjecture that ethnic groups of different sizes differ in ethnic-levelcharacteristics that are unobservable to the econometrician, which mightbe the true underlying causes of conflict. For instance, a possible threatfor my previous identification strategy is the following. Imagine that theexperts responsible to identify the politically relevant groups classified theethnic groups more broadly when they observed a conflict. If this is thecase, groups in conflict may be from an ethnicity comprising a larger shareof population just because of the arbitrary classification in EPR.In order to deal with this possibility, in this section I use the existence ofethnic groups split across different countries. Then, I exploit the differencesin the sizes of the same group across those countries to identify the effect ofgroup strength.Empirical StrategyFor this empirical strategy, I use the Trans-border Ethnic Kin (EPR-TEK) dataset. This is a companion to the main EPR-2018 which recordsall politically relevant ethnic groups living in at least two countries (i.e.ethnic groups with transnational ethnic connections) and whose settlementarea is split by an international border.16 I restrict the sample only to ethnicgroups living in at least two countries, and I estimate the following flexiblespecification that includes ethnic fixed effects:Ycet = β1sizece + β2size2ce + β3size3ce + γe + ce, (3.2)where Ycet is our outcome of interest (conflict incidence or inclusion), incountry c, ethnic group e, and year t.1716See Vogt et al. [2015] for details on the data. See Cederman et al. [2013] for a referenceof application.17I include the cubic term in the estimation of the effect of group size on inclusion. Thisis necessary because the peak of inclusion probability occurs for groups around 20% ofless (see Figure 7).43Under this specification, we are comparing the same ethnic group thatis present in different countries, and as a consequence, has different sizesmeasured as the share of country’s total population.18 This specification,therefore, controls for any determinant of conflict and inclusion that is in-herently constant to the ethnicity. For example, groups from a particularnon observed type of ancestry, history, religion or social norms may be moreor less likely to be in conflict or included. If these characteristics are corre-lated with group size, they can be a source of bias in the country fixed-effectspecification. By using within-ethnicity variation, we assure the comparisonof groups of similar culture, history and geographical region. Perhaps, mostimportantly, this strategy is robust to selection bias in the definition of thepolitically relevant groups by the data coders.Figure 6 depicts the empirical strategy for the example of Shia Arabs;who do not fit the model exactly but serve as a clear illustration. The figureshows the presence of Shia Arabs in 4 different countries. Our data is annual,but in the figure I make it explicit whether the group is generally excludedor included when they are not leaders, and whether the group has been inconflict against the government. They are only 3% of the Iranian population.In Iran, the Shia Arabs are mostly excluded and always in peace. The sameis true in Saudi Arabia, where they are 15% of the population. In Kuwait,on the other hand, they are 11% of the population and are often representedin executive power. A very different scenario occurs in Iraq, where althoughthe Shia Arabs are the vast majority of the country’s population (63%), they18It is possible to conceive that ethnic groups might have support from their co-ethnicsfrom the other side of the border in ways that the control (smaller) group would becontaminated by the treatment (larger co-ethnic group). In fact, the predicted pattern ofgroup size and inclusion is shifted to the left (Figure 4 versus Figure 7, suggesting that thissample of groups is on average stronger. Suppose we have a group split in two countries.In country A, the group has 10% of population share, while it has 40% in country B. Thekey assumption is that the group will be stronger in country B, where it has a larger shareof population in the territory (40%) with the little help of its co-ethnics on the other sideof the border, than in country A, where it has only 10% of population despite havinga better support across the border. More generally, by including ethnic fixed effects wecompare the same ethnic group, having by construction the same number of co-ethnicsacross the world, and use only the variation of the population inside the borders of eachcountry.44have been often excluded and are in a constant attempt to seize power. Thisexample is not cherry-picked; the model predicts less likelihood of inclusionin the case of Kuwait as opposed to Saudi Arabia. But the true efficacy ofthe model is whether it explains conflict and exclusion patterns for the dataas a whole.ResultsTable 8 confirms the evidence found using within-country variation forAfrica and Asia. In column 1, I estimate our main specification for the in-clusion outcome. All coefficients of the flexible specification of group size aresignificant. The estimated values imply a positive effect of increasing groupsize on inclusion probability up to around a 20% population share. Afterthis, increasing group size decreases the probability of inclusion. Columns 2and 3 report an increasing and convex relationship between group size andarmed rebellion.In columns 4 and 5, I test once more the additional model implication ofmodel that strength is predictive of conflict only if the challenger is excludedfrom power. The results using ethnic fixed-effects confirm this prediction.Conditional on being excluded, a group comprising 10% more of its country’spopulation is 1.96 p.p. more likely to be in conflict, compared to its otherco-ethnics (column 4). The effect is very large given the average incidenceof 1.5 % in the dataset (2.1% if considering challengers only). The effect ofgroup size on rebellion is much smaller and only significant at the 10% levelif the group is included in the coalition.Africa as natural experiment Africa provides a unique opportu-nity for a further robustness check. The reason is that national borders werehistorically arbitrarily drawn, and therefore it provides exogenous variationin ethnic sizes within each country – often referred to as the “The Scramblefor Africa”. This term denotes the partitioning of the continent by Eu-ropean powers during the period of New Imperialism, between 1881 and1914. The most significant event was the Berlin Conference of 1884 thatlaid down the principles that would be used subsequently to divide the rest45of the continent. Asiwaju [1985] argues that the Berlim conference had noAfrican representation or concern for local conditions, resulting in acciden-tal marking of African borders. For instance, Alesina et al. [2011] documentthat eighty percent of African borders follow latitudinal and longitudinallines. The artificial borders caused ethnic groups to be split between dif-ferent countries. At the time of independence, new rulers in Africa decidedto keep the borders drawn by former colonizers to avoid conflict with othercountries, making the artificial divisions persist over time. Thus, the ran-domness of African borders split groups in bigger and smaller sizes that areuncorrelated with national characteristics.Table 9 presents the results for Africa only. The first panel includes allpolitically relevant ethnic groups in African countries. As a robustness check,I remove non-native ethnic groups (e.g White Europeans) and “umbrellagroups” in the the second panel. Umbrella groups refer to cases where oneethnic group in a country is related to two or more sub-groups in another.For example, “Blacks” in Mali are coded as the same ethnicity of bothMandingue and Pulaar in Senegal. In both panels, the findings attest to theprevious results and model predictions.3.3.3 Comparative Statics: Evidence from relativeeconomic shocksNow I test a further prediction of model given by the comparative staticsimplied by Proposition 1. The model postulates that the ratio between costof conflict and the value of rents matters. Income variations changing costsand rents simultaneously in the same direction have an ambiguous effect;only changes in their ratio matters. Increasing the opportunity cost maymake groups less willing to rebel and cheaper to buy off, but it can be offsetby increasing the value of the prize (rapacity effect).Variations in the opportunity cost, and value of rents ratio provide auseful way to test the model. According to the model, dθ∗pi/cc < 0 anddθ∗∗pi/cc< 0.In words, an economic shock that makes the challenger relatively poorer(i.e. the value of rents increases relative to the opportunity cost), lowersboth thresholds. Intuitively, this makes conflict more attractive for the46contesting groups, but causes (testable) nonlinear effects on conflict andinclusion probabilities. Specifically, conflict should increase, but more sofor bigger groups (θ∗∗ moves to the left increasing the support of groups inconflict). The effect on inclusion is ambiguous. The distribution of groupschanges with more inclusion of weaker groups and exclusion of stronger ones.Figure 8 illustrates the effect of such economic shocks. These somewhatarcane implications of distributional shocks don’t seem to naturally followfrom some other cause, and allow for a further test of the model.Data: Commodity price shocks and crop mapsAs I explain in detail below, I identify relative economic shocks from rel-ative price variation of crops grown by different ethnic groups. I constructcommodity price indexes using data from Bazzi and Blattman [2014]. Theycollect data on prices of 65 commodities, the share of each commodity in to-tal exports, and the value of exports to GDP for the period of 1957 to 2008.The primary source of commodity trade was extracted from the United Na-tions Commodity Trade Statistics Database (UN Comtrade) assembled bythe United Nations Statistics Division (UNSD 2010). The dataset providestrade values by country, year and commodity. Additional data were col-lected from regional and individual country statistical yearbooks. Prices aretaken primarily from IMF International Financial Statistics (IFS), followedby US Bureau of Labor Statistics (BLS) and Global Financial Data (GFD).I then utilize individual maps of 32 crops that can be combined withthe international price index. I collect geo-referenced information for allcrop commodities listed in Bazzi and Blattman [2014]. Each crop map wasproduced by Monfreda et al. [2008]. It compiles land use information fromnational and sub-national census statistics in an updated global dataset ofcroplands on a 5 minute by 5 minute (roughly 10 km by 10 km) latitude/lon-gitude grid during 1997-2003. From these maps, I obtain average productionper land-area in an ethnic homeland. Finally, I use GeoEPR shapefile (eth-nic homeland maps) in combination with the map of each individual cropto construct the weight of each crop at the ethnic group level.47Empirical StrategyThere are two main challenges to estimating distributional economicshocks. First, there is no consistent annual income variable available atthe ethnic level for the whole sample.19 Second, even if a good proxy wasavailable, there would be endogeneity concerns. In order to overcome thesechallenges, I use exogenous variation in international prices of the cropsgrown by each ethnic group to identify differential (ethnic-specific) economicshocks. An ethnic-year price index is calculated as follows:Peit =∑cPctwceixciwci,where Pct is a real price index (corrected by CPI) of commodity (crop) c,at time t, and xci is a time-invariant share of commodity c in total exportsof country i. This weight is calculated by the average of the annual shareof that particular commodity in total exports for the period over which wehave data available (1957-2008).The ratio wceiwci is a measure of ethnic advantage, where wcei is the av-erage production by area of crop c, at the homeland of the ethnic group e,in country i. The same is done at the country-level to calculate wci. Thevariables wcei and wci are constructed combining maps for each crop c (Mon-freda et al. [2008]), and the country and ethnic homeland maps (GeoEPR).I then define an ethnic-specific economic shock variable:S˜eit = ∆(logPeit) ∗ (Xi/GDPi),where Xi/GDPi is a constant export per GDP ratio (we also take the av-erage over the sample period). Intuitively, the economic shock reflects thevariation in international prices weighted by the importance of each com-modity in the export basket. This is weighted by the relative importanceof the ethnicity production of that commodity. In addition, we expect thatshocks may be more important the higher is the magnitude of exports inthe country’s GDP. The commodity price index is always included in the19Night-time light density is only available after 1992.48regression analysis using first differences (because commodity prices followapproximately a random walk process).I proxy increases in pi/cc with price variation that makes the ethnic grouprelatively poorer with respect to the leader. In order to calculate relativeshocks, I take the difference between the ruling ethnic group (l) and eachgroup e in the country:Seit = ∆log(Plit/Peit) ∗ (Xi/GDPi), (3.3)Finally, I estimate the following specification:∆Yeit = φei + γt + β0Seit + β1Seit−1 + β2Seit−2 + eit, (3.4)where Yeit is the outcomes of interest: conflict over the government, orinclusion. The parameters φei and γt are ethnic and time fixed-effects. Asalways, robust standard errors are clustered at the country-level.Figure 9 illustrates the identification strategy. Togo is an African countrywith two politically relevant groups: Ewe living mostly in the south, andKabre that are more predominant in the northern-central areas. The Ewe’shomeland is relatively more productive in coffee, while the Kabre’s homelandis relatively more productive in cotton. When the price of coffee increasesmore than the price of cotton, I assume that the ratio of opportunity cost torents decreases for the Kabre (who are more abundant in cotton), increasingtheir value of fighting. The opposite effect would occur for the Ewe’s group.ResultsTable 10 presents results for regressions testing the implications of themodel illustrated in Figure 8. Exogenous economic shocks inducing leadersto become relatively richer have marginally small positive effects on theprobability of conflict (Column 1). This weak causal effect hides strongheterogeneity. The effect of distributional shocks in favor of the leader onconflict onset are larger for bigger groups (Column 2). A 10% increasein the leader’s prices relative to a challenger with a population share of4950% increases the probability of conflict by around 2.5 p.p. (comparedto a sample average of 2.1%). Consistent with the model’s predictions,commitment worsens for strong groups. Increasing rebellion incentives forsuch groups leads to bargaining failures, while weaker groups are more likelyto be accommodated in power-sharing agreements.The effect of distributional shocks on inclusion is not significant (Col-umn 3). In column 4, there is some evidence (although the coefficients aresignificant only at 10% level) that the effect is heterogeneous. Specifically,a relative shock favoring the leader will have a maximum positive effect forgroups that are around 19.4% of the country’s population. For groups biggerthan this, positive effects diminish and eventually become strongly negative,implying a lower probability of appeasing an opposition more disposed toconflict.One potential concern with the relative shocks variable constructed isthe possibility of variation due to leader transition, and not internationalprices only. To address this concern, Table 11 reports the same analysisexcluding periods of leader transition. The results are preserved.3.4 ConclusionIn this chapter, I have presented strong empirical support to the the-ory developed in Chapter 2. I use a novel dataset combining informationon access to executive power by ethnic groups, conflict over the control ofthe government launched them, and a reasonable proxy for group strength.The key predictions of the model are tested using a multitude of empiricalstrategies. First, I compare groups of different sizes within the same coun-try to show evidence for the strong non-linearities between group strength,and conflict and inclusion outcomes, predicted by the model. The resultsare robust to the inclusion of a series of fixed-effects and a battery of con-trols. Second, the results persist using a quasi-natural variation induced bygroups split across borders. Third, I test a comparative statics implicationof model, identifying conflict-inducing relative economic shocks in the ethniclevel.50The analysis has provided novel empirical patterns of the distributionof power among ethnic groups in weak states - particularly, the evidenceof exclusion of strong groups. Also, the new documented facts are tightlylinked to the model. This moves forward the needed effort of testing theoriesin the conflict literature (Blattman and Miguel [2010]).51Chapter 4Power-sharing and Conflict:Structural Estimation andPolicy Counterfactuals4.1 IntroductionThe model has successfully accounted for key patterns in the data, andsurvived several empirical tests. Now we can take one step forward. Underthe assumed structure of the theory, the model parameters can be estimated.This allows me to quantify the effects of policies that can be influencedand executed by the international community aimed at mitigating conflicts.Most of them are fruit of heated debate, and generally lack any scientificdiagnosis about their potential impact.The policy-relevant counterfactuals I evaluate are democratization, mili-tary capacity, financial aid, sanctions and quotas. The relationship betweendemocratization and conflict has been a topic of long debate and little em-pirical consensus (e.g., Mansfield and Snyder [2007], Bernhard et al. [2017]).Most findings suffer from endogeneity issues, and also by the fact that de-mocratization itself may be a leader’s decision of power-sharing. Withouttaking any normative stand, here I simply estimate the effect of democra-52tization in a key outcome of interest - intrastate conflict. Alternatively, Iinvestigate the effects of enhancing the state capacity of governments, po-tentially increasing the military superiority of incumbents.Financial aid, another relevant policy, has been widely used by the in-ternational community to provide assistance to developing countries. Stateswith conflict history have been target of substantial amount of resourcesgiven the size of their economies. For example, official development as-sistance to historically conflict-prone countries, like Afghanistan, Liberia,Burundi, and Rwanda, are 13%-21% of their gross national income.1 Nev-ertheless, the effects of such interventions on conflict have also been a topicof contradictory results in the literature. The model estimated here will beable to illuminate how financial aid can have positive or negative resultsdepending on its design (depending particularly on the recipient group).I also investigate the role of sanctions and conditional aid in discipliningleaders and rebel groups. This is of first order importance. A common viewof why violent leader transitions occur so frequently in weak states is thatnew leaders are promptly recognized by other national leaders once theyassume office (Roessler [2016]). What would be the effect of sanctions - atrade restriction, for instance - on countries in conflict? How would leadersand challengers adapt their power-sharing and conflict decisions? Wouldthis be quantitatively important?Furthermore, quotas - numerical requirements of representatives of par-ticular groups in political positions - have been increasingly implementedworldwide in order to achieve higher inclusion and equality. Could weachieve peace by enforcing some minimum inclusion of opposing groups?I evaluate the effects of such policy in mitigating conflicts.4.2 Additional dataFor the purpose of estimating the model, I make use of additional databeyond the standard EPR data presented in Chapter 3. The reason for col-lecting additional data is the following. The model has implications beyond1Data obtained from the World Bank database.53conflict and inclusion probabilities. The parameters of the model define athreshold (θ∗∗) that determines who is in conflict and who is appeased in apower-sharing agreement. The distribution of strength not only generatesthe probability of conflict but also the probabilities of rebel victory con-ditional on fighting. Furthermore, power-sharing transfers also depend onthe distribution of strength conditional on being between the two thresholdsthat determine inclusion. Thus, having information on the shares of powertransfered to challengers (not available in the EPR) and the outcomes ofcivil conflict can provide additional variation to help in identifying struc-tural parameters.Information on shares of power obtained by included groups comes fromthe dataset collected and used by Rainer and Trebbi [2012] and Francoiset al. [2015]. They employ a complete dataset on the ethnic affiliation ofeach national minister for 15 African countries since independence until2004. The ethnicity categories in their dataset are not the same as in EPR.Thus, I first match the ethnicities in the two datasets and then calculatethe proportion of cabinet seats occupied by each included group in EPR.I will use this variable as a noisy measure (more on this below) of thetransfer conceded by the leader. Note that I only have this informationfor a sub-sample of the whole dataset. In the econometric specification(below), I derive the likelihood of observing a particular value of transfer foran included group if it is observed, and also the likelihood of being included(independent of how much is shared) for groups that are not part of thissub-sample.Next, I construct a binary variable indicating if an insurgent challengerwon or lost the war. Data on the outcomes of conflict are coded manually,case by case, for every group identified as in conflict against the government.The main source of information is the UCDP Conflict Termination Dataset(Kreutz, 2010). As explained above, ethnic groups in conflict are matchedwith rebel actors that claim to represent them, and then with the outcomeof the conflicts that those actors are launching.There are 6 possible outcomes in the UCDP Termination Dataset (rebelvictory, government victory, low activity, actor ceases to exist, peace agree-54ments and ceasefire. I follow the dataset ruling if the outcome is defined asrebel victory or government victory. In most of the cases the outcome iscoded as low activity (because UCDP codes conflict only for episodes withmore 25 battle deaths per year). Then, I search in the same dataset for thefinal result if the conflict eventually resumed. Otherwise, the final outcomeis obtained from other online sources. If the actor ceases to exist, I identifythe new actor name that represents the ethnic group (which is generally thecase) and the resulting outcome of that conflict event. The final outcome iscoded as missing for Peace agreements and Ceasefire agreements, unless theconflict eventually resumes and a rebel victory or defeat can be clearly estab-lished. Finally, outcomes of ongoing conflicts are also missing. Here again, Idefine the likelihood of conflict (for those with a missing outcome) and thelikelihood of victory (or defeat) for groups from whom I could identify thefinal outcome of conflict.In sum, the data we have consists of every non-leader group, whether it isincluded in the coalition, and whether it is in conflict against the government.For a subsample, we will also have the shares allocated to included groups,and the final outcome (victory/defeat) of conflict.4.3 Econometric specification4.3.1 OverviewI give a brief overview of the structural specification before showing thedetails. In general, I assume players are playing consistently according tothe model. Information is complete: Challengers know their own strength.Leaders know it as well. The game is static and played repeatedly everyperiod. The econometrician does not observe group strength perfectly, butstrength depends on the group’s population share. Strength varies betweengroups, but does not vary over time. Group inclusion (and power-sharingtransfers) and conflict can vary over time. This temporal variation comprisesdeviations from outcomes predicted by the model which are due either to55measurement error or arise due to reasons outside the model. Next, I elab-orate on the econometric specification.4.3.2 StrengthEach ethnic group i has a time-invariant2 innate strength θi. We cannotobserve θi perfectly, but observe the population share (si) of each group as aproxy. Specifically, θi is drawn from a cumulative distribution F (θ|si), anddensity function f(θ|si), with support [0, 1−α]. As clarified above, θi is theprobability of victory if the group launches a conflict and is only defined (inthe context of the theoretical model) over this support.3I assume F (.) and f(.) are functions with beta distribution, such thatθi ∼ Beta(γ, γ(1 − si)/si, 0, 1 − α), where γ > 1. The shape parametersγ and γ(1 − si)/si depend on the population share of the group, and 0and 1 − α are the minimum and maximum values in the support of thedistribution.45 These parameters are defined to generate E[θi|si] = (1−α)si.This captures the hypothesis of group size being a proxy for strength. Everygroup may draw any innate strength from 0 to 1−α; however, in expectation,arbitrarily small groups have negligible chances of winning, while groupscomprising almost the whole population have probability of success close tothe maximum. Beta is a suitable distribution in this case for being defined2The time-invariant innate strength of the group captures the strong persistence ofconflict and power-sharing in the dataset.3θi must be not greater than 1 − α to guarantee that the probability of success ofconflict does not surpass 1 for any power-sharing transfer4This distribution will be estimated for the sample of non-leaders. Again, note there isa selection since we might expect that leaders are stronger in expectation even conditionalon population share. Nevertheless, the distribution of strength conditional on not beingleader is exactly what we are interested in (and, it is the one estimated here). The expectedstrength unconditional on being leader and challenger is not relevant since leaders don notgo to war, and their strength is not determinant of the equilibrium. We want to knowwhat population share reveal about the group strength for those who are challengers.5I don’t impose any country heterogeneity because reduced-form estimates of the effectsof population share with and without country fixed-effects are very similar. Technically,maximum likelihood estimates of country-specific distributions of θ could result in im-plausible corner solutions. Groups in countries with only exclusion and peace would beestimated as weak as possible. Meanwhile, groups in countries where the challenger is avictorious rebel would have their strength estimated as high as possible.56in a limited interval.6 The assumption γ > 1 guarantees a bell shape for thedistribution.74.3.3 ConflictConflict initiated by group i, at period t, is observed through a binaryvariable Cit ∈ {0, 1}. According to the model, a group with strength θi > θ∗∗is in conflict in equilibrium. In the data, such groups will be observed inconflict, in a particular period t, with probability φc. There are a few reasonswhy there may be time variation in measured conflict even if the underlyingdata generating process is constant. For instance, conflict episodes are onlycoded as such in the dataset if they resulted in at least 25 battle deaths inthe year. Similarly, active insurgent groups may occasionally stop fightingto regroup (while the underlying contest with the government still existsunabated), or the events of conflict may have not been observed by datacoders.If θi < θ∗∗ group i is not in conflict in equilibrium, according to themodel. In the data, such groups are assumed to be in conflict with probabil-ity φ0. These episodes are pure errors not explained by the model. We canthink of them as uninsurable conflicts (i.e., ones that cannot be appeased bypower-sharing)8. It will be shown that the estimated value of φc is substan-tially greater than φ0, which turns out to be fairly small. Practically, in thecontext of the likelihood estimation, this positive probability of conflict dealswith conflict outliers (e.g. included groups in short conflicts). Formally:P (Cit = 1) =φ0 if θi ≤ θ∗∗ = (1−α)cl/pi+cc/piαφc if θi > θ∗∗ = (1−α)cl/pi+cc/piα(4.1)6For examples and discussion of the use of the beta distribution in this literature,see Merlo (1997), Diermeier, Eraslan, and Merlo (2003), Adachi and Watanabe (2007),Francois, Rainer and Trebbi (2015)7A higher γ indicates lower variance. In practice, variance of groups in the extreme(with population share close to 0 or 1) is substantially low for any γ. This is because theexpectation of thetai is already close to one of the support limits.8For example, there may be short-term instability due to disagreements in the alloca-tion of power-sharing agreements.57Let Wit be a binary variable determining the outcome of a conflict. Ittakes value 1 in case of a rebel’s victory (with probability θi) and value 0 incase of a leader’s victory (with probability 1− θi).Conflict4.3.4 Power-sharingFinally, every player in the game receives (and perfectly observes) τias a function of the group’s innate strength θi, according to Proposition1 (τi = 0 if θi ∈ [0, cc/pi] ∪ [ (1−α)cl/pi+cc/piα , 1 − α], or τi = θi−cc/piα if θi ∈[cc/pi, θi−cc/piα ]). We imperfectly observe τi for included groups, where theobservation includes time-varying group-specific errors, vit. I define thefollowing latent variable:X∗it = τi + vit,where vit is identically distributed across time and ethnic groups, and hassymmetric distribution around zero, with density function g(.) and cumu-lative function G(.). The empirical ministerial allocation observed in thedata is Xit = X∗it if X∗it ≥ 0 and Xit = 0 if X∗it < 0. In what follows, I setvit ∼ N(0, η).Therefore, an included group (τi > 0) is erroneously observed out ofthe coalition in the data with probability P (Xit = 0|τi > 0) = G(−τi).This probability goes to zero as τi increases. Meanwhile, the probability ofobserving such groups as excluded approaches 50% as τi approaches zero.I finally allow for the possibility of error in assigning excluded groups(they could be empirically assigned as included in some periods). For exam-ple, ethnic groups may have unobservably different factions or subgroups -one of whom may be included, and the other excluded; or a group’s strengthmay eventually move away from its innate condition. The probability of sucherrors will also be governed by the parameter η, and will depend on the dis-tance to inclusion thresholds. For simplicity of notation in the derivation ofthe likelihood, we define the probability of empirical inclusion of excludedgroups as η˜:58η˜ = P (Xit > 0|τi = 0) = η(1− Φ(d/η)), where d = min(|θi − θ∗|, |θi − θ∗∗|)(4.2)The component (1−Φ(d/η)) follows the same structure as the measurementerror for included groups. It is the probability that a normally distributedrandom variable, with standard deviation equal to η, is greater than d. Ifthe group strength is far from what is needed to be included, the empiricalprobability of inclusion is lower. To put it another way, as the group strengthbecomes closer to the inclusion threshold it is more likely to be observed asincluded. In the limit, as d approaches zero, (1− Φ(d/η)) converges to 0.5.For this reason, the term is multiplied by η, making it dependent on thevariance of the measurement error.Note how the measurement error is governed by η. As η goes to zero,there is no measurement error. We would observe power-sharing allocationsperfectly, and excluded groups would never be mistakenly observed as in-cluded in the data. As η increases both error types are more likely (seeinginclusion of an actually excluded group, or empirically assigning exclusionto a group included - in the model - in the coalition).The probability density function of observing a specific ministerial allo-cation for an excluded group is given by:P (Xit = x > 0|τi = 0) = η˜P (τi + vit = x|τi = 0, x > 0)= η˜P (vit = x|vit > 0)= 2η˜g(x)(4.3)This group-specific idiosyncratic component has two purposes in our esti-mation. First, it captures temporal deviations from a long-run constanttransfer (explained by the model). This is consistent with the data thatshows that power-sharing allocations are very persistent with small fluctua-tions over a historical average. Second, it captures the unusual cases wherewe observe transitory inclusion of groups, some of them in conflict (which isnot predicted by the model for any group type). In this econometric spec-59ification, I allow for cases where θi > θ∗∗, and even though a group is inconflict and excluded according to the model, transitory transfers may beobserved due to idiosyncratic shocks.9 This is also consistent with the data:79% of included groups in conflict are found to be excluded in other periods;whereas only 38% of peaceful included groups are eventually excluded.4.3.5 LikelihoodFor the structural estimation I assume all groups face the same cost ofconflict cc. For identification purposes, I also set cc = cl = c (both α and cldetermine the value of the second threshold, so they cannot be separatelyidentified). It is also clear from the solution of the model that all outcomesdepend only on cc/pi and cl/pi, and not on cc, cl and pi separately; thus, onlythe ratios are identified.Given the set of model parameters Θ = [α, c/pi, γ, η, φ0, φc], the jointlikelihood of observing a power-sharing allocation Xit, conflict decision Cit,and conflict outcome Wit conditional on θi is defined as:lit(Xit, Cit,Wit|θi; Θ) =(Citφ0 + (1− Cit)(1− φ0))[(1− η˜)1(Xit = 0) + 2η˜g(x)1(Xit = x > 0)+η˜1(Xit > 0, x = .)], if θi < cc/pi(Citφ0 + (1− Cit)(1− φ0))[G(− θi − cc/pi1− α )1(Xit = 0) + g(x−θi − cc/pi1− α )1(Xit = x > 0)+(1−G(− θi − cc/pi1− α ))1(Xit > 0, x = .)], if cc/pi ≤ θi ≤(1− α)cl/pi + cc/piα[(1− Cit)(1− φc) + Citφc1(Wit = .) + Citφc1(Wit ∈ {0, 1})(Witθi + (1−Wit)(1− θi))]∗[(1− η˜)1(Xit = 0) + 2η˜g(x)1(Xit = x > 0) + η˜1(Xit > 0, x = .)],if θi >(1− α)cl/pi + cc/piα(4.4)The first case gives the likelihood conditional on group strength beinglower than the first threshold (τi = 0). In this case, the probability of conflict9An examination of the cases of included groups in conflict points to several reasons(not inconsistent with the key elements of the model, but due to the limitations of thedata): i) ethnic group has two factions, one included and the other excluded and inconflict; ii) conflict in anticipation of a future exclusion; iii) short-term instability due todisagreements in the allocation of power-sharing agreements.60is φ0 (Equation (4.1)). The probability of observing this group excluded is1− η˜ (analogously, the inclusion probability is η˜), as given by equation (4.2).We observe cabinet allocations for a subsample of included groups. In thiscase, the probability density function of observing a particular allocation xis given by 2η˜g(x), as shown in equation (4.3).In the second case it is shown the likelihood conditional on group strengthbeing between the two thresholds. In this interval, the group is predicted tobe in peaceful inclusion. Again, the conflict probability is φ0. The proba-bility of exclusion depends on the distribution of measurement errors for in-cluded groups. As explained above, it is given by G(−τi), where τi = θi−cc/pi1−α .If we observe the cabinet allocation of an included group, we use the densityfunction g(vit = Xit − τi).Lastly, the third case has the likelihood for groups above the secondthreshold, and therefore predicted to be in conflict and exclusion. The prob-ability of observing conflict empirically for this group is φc. For a subsampleof rebel groups, we have conflict outcomes (Wit ∈ {0, 1}). If this is the case,the probability of conflict is also multiplied by the probability of victory (θi)or defeat (1−θi), depending on the realization (Wit). Inclusion probabilitiesfollow the same structure as the groups in the first block. Note that η˜, G(.),and g(.) depend fundamentally on the parameter η to be estimated.The likelihood of the time series of observations for each ethnic group iconditional on θi is:li(Xi, Ci,Wi|θi; Θ) =T∏lit(Xit, Cit,Wit|θi; Θ). (4.5)Integrating over the distribution of unobserved strengths, we obtain thelikelihood of each group conditional on population share:li(Xi, Ci,Wi|si; Θ) =ˆ 1−α0li(Xi, Ci,Wi|θi; Θ)f(θ|si)dθ (4.6)61Finally, the log-likelihood is given by:lnL(Θ) =∑iln li(Xi, Ci,Wi|si; Θ) (4.7)4.4 EstimationI estimate the parameters Θ of the model using Maximum SimulatedLikelihood (MSL). Note that the conditional probability li(Xi, Ci,Wi; Θ)involves an intractable integral. In the absence of a closed-form solution,the likelihood of each observation is estimated as:lˆi(Xi, Ci,Wi; Θ) =1RR∑r=1li(Xi, Ci,Wi|θri ; Θ), (4.8)where θri , r = 1,= 1, ..., R, are R draws with marginal distribution f(θ|si).The estimator ΘˆMSL minimizes the negative simulated log-likelihood:ln Lˆ(Θ) = −∑iln lˆi(Xi, Ci,Wi|si; Θ) (4.9)The MSL estimator is consistent with usual Maximum Likelihood (ML)asymptotic distribution if N → ∞, R → ∞, and √N/R → 0, where N isthe number of observations.4.5 IdentificationI aim to identify the following parameters [α, c/pi, γ, η, φ0, φc]. Everygroup draws a unique and persistent strength θi. This generates a time in-variant τi. Therefore, η can be identified from within group time variationin the allocation of cabinet seats, as well as group shifts in inclusion/ex-clusion. Conditional on the value of the second threshold (c/pi and α) andstrength distribution, we can compute the probability of falling within eachrange (before or above the threshold) for each group conditional on popu-lation share. The resulting conflict probability will depend on φ0 and φc.More precisely, the probability of conflict for a group of size si is equal to62F (θ∗∗|si)φ0 +(1−F (θ∗∗|si))φc. Variation in group sizes and their respectiveempirical likelihood of conflict identify φ0 and φc.Given the conditional distribution of strength, c/pi is identified to matchthe probabilities of exclusion and low conflict. The second threshold is afunction of both c/pi and α. Having the value of the first as given, α isidentified to match the empirical probabilities of inclusion/low conflict andexclusion/high conflict. A bigger α implies a higher threshold, and higherlikelihood of inclusion/peace compared to exclusion/conflict. Furthermore,the second threshold determines the distribution of the strength of groupsin conflict - because it defines the minimum strength to have conflict inequilibrium. Consequently, the empirical probabilities of success also providevariation for the identification of α.Lastly, γ is identified from the variance in equilibrium outcomes; thatis, how likely groups of the same size are distributed in different equilib-rium zones. A higher γ implies lower variability and more precision in theprediction of inclusion and conflict probabilities.4.6 ResultsTable 12 presents the maximum simulated likelihood estimates of thestructural model parameters [α, c/pi, γ, η, φ0, φc]. The cost of conflict is esti-mated to be 3.6% of the value of the rents of capturing the government. Thisimplies a relatively low cost, and explains the steep response to the inclusionprobability from increments in the population share for small groups. Thissmall value is consistent with the low levels of income of the average citi-zen, and the fact that opportunities of economic success are concentrated inholding political positions. This is a common view in many African coun-tries. For example, salaries of ministers in Zaire at the end of Mobutu’s ruleare estimated to be 40 to 60 fold greater than those received by doctors andjudges (Francois et al. [2015];Gebrewold [2009]).The marginal shift in success probability of conflict due to power-sharingtransfers (α) is estimated at 0.17. Combined with the estimated value of thecost, it implies θ∗∗ = 39% - the maximum strength of groups that leaders63are willing to accommodate in the coalition. If the threat is higher, groupscannot be included without significantly increasing their power and makingthe cost of inclusion larger than the expected losses from conflict. In fact,69% of conflict episodes, for which we observe the final outcome, end withthe rebel’s victory, suggesting high likelihood of success for insurgents.The estimate of γ is 1.0004. Because the structure of the strength distri-bution imposes expectation of strength equal to population share (adjustedby 1 − α), the variance of distribution is always low for groups close tothe boundary of the distribution (group size close to 0 or 1). However, anestimate estimate of γ close to 1 suggests a high variance for groups withpopulation share in the middle range, and the importance of other non-observed explanatory factors for group strength.The parameter η governing the distribution of inclusion and exclusionmeasurement error is relatively low (0.08). This captures the temporal stan-dard deviation of cabinet appointments within group. It also governs theprobability of empirical inclusion for groups expected to be excluded. Theestimate implies a probability of 2% of inclusion for a group whose strengthis far from the inclusion threshold by 0.05. For groups that are further than0.1 from the inclusion threshold, the inclusion likelihood is less than 1%.Finally, the estimates of φ0 and φc highlight the explanatory power ofthe model. The probability of conflict for groups below θ∗∗, and therefore,unaccounted by the model is 0.9%. On the other hand, groups who are es-timated to be above the threshold, and subsequently, strategically excludedby the leader, are in conflict 32.8% of the time. This is also suggestive thatthe model explains longer and more persistent conflicts.Figure 10 shows the simulated probability of inclusion and conflict bygroup size generated by the estimated parameters. The non-linear relation-ships observed in the pure data (as shown in Figure 3) are closely recoveredby the structural model.644.7 CounterfactualsI use the model to study several policy relevant counterfactuals. I firstexplain how the model parameters, or the structure of the game, may beheavily influenced by international interventions or institutional policies (de-mocratization, military capacity, financial aid, sanctions and quotas). Then,I examine how these changes affect conflict probabilities and other outcomes,by simulating the model in the dataset under the new parameters or struc-ture. Results are shown in Figure 11 and Table 13.All counterfactuals are compared to the results of the baseline model.The probability of conflict and inclusion are 2.52% and 29%, respectively.Using the inclusion probability and the size of the the leaders, I compute thatthe coalition has ethnic representation comprising 66% of the population (inexpectation). Back-of-the-envelope calculations find the probability of aneventual leader’s removal through conflict to be 13.5%. This is calculated asthe sum of probabilities of challengers being strong enough to be excludedand in conflict, multiplied by their expected strength conditional on the firstcase.Democratization Several politically unstable regimes in Africa andAsia are associated with autocracies headed by ethnic minorities. In manycases, these ethnic minorities have held power by excluding large segmentsof society. This is the case of Liberia that was ruled by a small minority offreed American slaves repatriated there until 1980. Other examples includewhite minority rule in Zimbabwe and South Africa, the Tutsi-controlledgovernment in Burundi, and Saddam Hussein’s Sunni-dominated regime inIraq.In this counterfactual experiment, I arbitrarily move the leadership ofthe government to the largest group. This is not necessarily the strongestgroup but is likely to be so. Subsequently, the composition of challengersis replaced by one with challengers of smaller size, who are more likely tobe appeased. I call this experiment "democratization" since it approximatesthe largest group advantage that arises in democratic elections.65Democratization generates a large drop in conflict of 35% (from 2.52%to 1.63%). The probability of inclusion only slightly moves from 28.9% to29.6.1%. The effect is largely driven by having a leader who would be apotentially excluded rebel otherwise. This increases the size of the coalition(from 66% to 73.7%). Under democratization the regime is considerablymore stable. The probability of an eventual leader’s removal by violentmeans falls from 13.5% to 5.3%. Higher stability is driven by both a smallerlikelihood of conflict and the fact that rebel groups are now weaker.Military Capacity The second counterfactual experiment examinesthe effect of increasing military capacity. State capacity, conceptualized asthe coercive control and the monopoly of violence, is clearly associated withcivil conflict. Conflict-prone countries are those where armed groups mountcredible threats against weak regimes. The implied policy here is one wherethe international community would provide military and organizational as-sistance to governments. Increasing the military capacity of a country wouldenhance the control of the government forces over the territory and weakenanti-government militias. In particular, I examine the effect of decreasingthe probability of rebel victory by 20% (∆θ/θ = −20%). I find that conflictwould be reduced by 20%. The inclusion probability also decreases. A lowerstrength induces the inclusion of groups who would have previously been ex-cluded and in conflict. On the other hand, this effect is overcompensated bythe exclusion of groups who lose their credible threat. As a result, stabilityincreases with a lower probability of the leader’s removal (8.4%).Financial aid to groups Governments and international organiza-tions have extensively used development assistance to reduce violence andimprove lives in conflict-prone regions. Development assistance has the maingoal of improving welfare by expanding heath care, schooling, job oppor-tunities, and income. Increasing the economic opportunities and futureprospects of citizens makes them less vulnerable to be recruited by rebelgroups. In the context of the model, I assume this would increase the op-portunity costs for rebel groups due both to higher salaries required forrecruitment and the larger losses of infrastructure, and interruption of in-vestments and economic activities. Here I evaluate the effect of financial aid66that increases the opportunity cost of engaging in conflict (cc) by 20%. Theprobability of conflict is reduced by only 10%; notably less than the changesassociated with most of the other counterfactual policies. The small responseis due to the fact that the opportunity cost parameter has been estimated tobe relatively small compared with the estimated prize of capturing the gov-ernment. Consequently, increasing costs by 20% does not have a substantialeffect on cc/pi. The inclusion probability decreases are driven by the lowernumber of groups with incentives to fight.Unconditional Government Aid (or Reduced Arming) Finan-cial assistance may have a very different effect if instead of increasing theopportunity cost, it in fact raises the rents from office. The implied pol-icy here is one that provides financial resources directly to the state, wherepolitical elites have discretion on their allocation, and can use them fortheir own benefit. Here I evaluate the effect of a 25% increase in the rentsfrom office (∆pi/pi = 25%) due to the higher revenues received by the coun-try. Because the results depend only on the ratios cc/pi and cl/pi, this isequivalent to reducing the costs of conflict cc and cl by 20%. This can beinterpreted as a "Reduced Arming" scenario. This would be a scenario withhigher restrictions to the production and trade of weapons. In this case, thedestructive potential of conflict, proliferated by weapons, troops, training,and resources from foreign allies, would be reduced. Consequently, the costsof conflict would be also diminished.Results show an estimated increase in conflict of 23%. There are acouple of reasons for this large effect. First, it entails an effective decreasein both the leader’s and challengers’ relative costs, diminishing the scope fora peaceful bargain. Second, a decrease in cost can raise conflict more thanan increase in the cost diminishes violence because there are more groupsimmediately to the left of θ∗∗ than to the right. The probability of inclusionincreases to 30.1%, while the coalition size declines to 64.8%. This is due tothe inclusion of smaller groups and exclusion of bigger groups.These results help reconcile the apparently contradictory results of theimpact of financial aid on conflict in the literature. Some studies have shownfinancial aid increases conflict (e.g Crost et al. [2014], Weintraub [2016], and67Nunn and Qian [2014]). Others have found zero or negative effects on re-cruitment and insurgent support when benefits were designed to be less sub-jected to rebel targeting by directly giving them to community individuals(Crost et al. [2016], Blattman and Annan [2016], Lyall et al. [2018]). Theseresults are consistent with ours that point to positive effects on conflict oflarger rents, and modest negative effects of greater opportunity costs.Government Aid Conditional on Peace What if, instead of giv-ing unconditional financial aid, the international community gives financialassistance that is conditional on peace? I estimate the effect of increasingthe rents from office pi via such a peace premium of 5%. It is assumed thatthe international community can credibly commit that the increment is re-moved if a conflict breaks out. The change in the structure of the game hasinteresting implications. First, because insurgents lose these additional rentsif they start a conflict and win, groups can be appeased with a lower shareof total power (which includes the extra rents). Second, leaders now havea higher opportunity cost of excluding groups and facing war (i.e., the lossof aid). As a result, conflict is reduced to 1.72% (32% drop). The inclusionprobability and coalition size increase (31.4% and 69.8%, respectively), atthe same time that the probability of an eventual leader’s removal drops to8.4%.Sanctions against regimes in conflict The next policy experimentexplores the effect of an international sanction against a regime in conflict.I assume this intervention to be a drop of 5% in the rents of governments inwar. It implies ∆cl/pi = 5%. Raising the costs of conflict for leaders enlargesthe scope for peaceful bargaining and induces inclusion. This interventionhas sizable effects. Conflict is reduced by 39%, the inclusion probabilityincreases to 32%, coalition size reaches 70.9% of population, and leaders areremoved with probability as low as 6.9%.It is worth noting why this intervention has a larger effect in reducingconflict than conditional aid. In fact, leaders also face an implied opportu-nity cost if they refuse the aid and go to war. However, as long as powermust be shared to appease the challenger, the leader does not enjoy theincrements of rents to their full extent. For this reason, the effective op-68portunity cost is higher under sanctions than under conditional (on peace)aid.Sanctions against government established through conflictAnother type of sanction is to punish not the leader in conflict, but thenew regime established irregularly (through conflict). I assume this as a 5%drop in the value of the rents for successful rebels. This counterfactual ex-periment has an effect similar to conditional aid. Because challengers wouldreceive a lower value of rents in the case of a successful rebellion, they re-quire lower transfers to be appeased. The lower value of transfers causesa downward shift in group capabilities, diminishing the commitment prob-lem. Additionally, this type of sanction increases the ratio of opportunitycost to the actual prize of overthrowing the government. The result is adecline of 35% in conflict. The inclusion probability goes to 30.7% (lowerthan inclusion under conditional aid).Quotas Quotas have been a widely advocated policy in order to ensuregreater and fairer representation of groups in society. How effectively does asystem of quotas reduce conflict? I explore the effects of a quotas with thefollowing rules: (i) every group that comprises more than 20% of populationmust be included; and (ii) the minimum share of rents to be allocated to anyincluded group is 25% of group’s population share. For example, a groupthat is 40% of the population should be included in the coalition and haveat least 10% of the seats.Quotas have a direct and indirect effect on conflict. First, the mini-mum share required by the system may be more than enough to appease achallenger who would be in conflict otherwise (direct effect). Second, if theminimum required is not enough for appeasement, the leader must decidebetween giving the minimum required by law and facing conflict, or com-plementing the transfer of power to buy peace. Note that the leader nowfaces a lower utility value in facing conflict. If there is no appeasement, theleader will face a war where he is more likely to lose because of the minimumshare he is required to give to the opposition. This, in turn, may induce fullappeasement (an indirect effect on conflict).69The simulation for the specified quota system results in 2.17% proba-bility of conflict (14% drop), 35.2% probability of inclusion, and 78.1% ofthe population represented in the coalition. Quotas increase inclusion andcoalition size much more than any of the other policies. However, the effecton conflict is fairly modest. The probability of an eventual leader’s removalmarginally drops to 11.8%. On the one hand, stability is increased becausethere are fewer wars; but on the other , leaders are more likely to be removedby persisting conflicts.4.8 ConclusionIn this chapter, the model was structurally estimated. The major benefitof the exercise is to quantify the effects of relevant policies commonly consid-ered as potential peace-keeping interventions. The results have shed somelight into the discussion. Democracy is found to be effective in reducingconflict. Financial aid, consistent with previous literature, has ambiguouseffects. It essentially depends on which way it goes. However, conflict inci-dence reacts more to growth in the rents from office (when conflict incidenceis increased) than to growth in financial aid to groups directly (when con-flict incidence is decreased). In fact, the effect of latter is very limited giventhe small estimated opportunity costs of war in our context. Altogether,financial assistance is looks like a high risk policy for a potential marginalimpact. Sanctions over countries in conflict is very effective in discipliningactor that may be potentially involved in war. Lastly, quotas is kind ofirrelevant for conflict (but obviously good for inclusion).Structural estimation of conflict models are still rare in the literature.10Other attempts to do so in light of other models will provide very fruitfulinsights to policy. This is paramount given the obvious difficulties of doingexperimental interventions to identify best practices in these contexts.10König et al. [2017] is one of the very few examples.70Chapter 5ConclusionWhy does conflict occur? Why does bargaining fail leading to violentrebellion? In this thesis, I have investigated the role of a simple feature,arguably very present and salient in the context of ethnic politics in weakstates. Sharing power (by allocating cabinet and military positions) withcontesting groups grants them access, privileges, information, and adminis-trative capabilities. This, in turn, increases their likelihood of success shouldthey choose to move against the leader. I show in a simple model how thisleads to a inverted-U shape distribution of included groups in the govern-ment coalition based on their strength. In particular, I show that leaderschoose to exclude and ultimately face conflict with the most powerful groups.Based on the theoretical predictions, I have documented a novel empir-ical pattern showing that the largest ethnic groups have a declining prob-ability of inclusion, despite their more credible threats against the regime.This newly uncovered stylized fact, rationalized by the model, is found com-paring groups in the same country, as well as groups of the same ethnicityquasi-randomly split across countries. In addition, I identify exogenous dis-tributional shocks within countries that provide further evidence in favor ofthe model.Identifying the underlying theoretical mechanisms causing conflict is offirst order importance for policies. After providing reduced-form evidencesupporting the model, I structurally estimate the model parameters. Sev-71eral policy relevant counterfactuals are examined. Results show the limitsof financial aid and quotas. Democratization, conditional assistance (onpeace), and sanctions contingent on conflict both yield promising prospectsin reducing the chances of rebellion.There are numerous additional questions, not fully addressed here, thatare important avenues for future research. First, all of the analysis is per-formed using ethnic conflict over the government. This followed from thepossibility of observing (with noise) offers, and conflict outcomes simultane-ously, and the importance of ethnicity as a determinant of political identityin the context of Africa and Asia. Nevertheless, nothing in the theoreticalor empirical framework indicates that the results should be expected only inthis particular setting. The same mechanism may play a role in non-ethnicor secessionist wars too, but we cannot provide such evidence here. Second,the analysis was restricted to only politically relevant groups and assumedtheir existence in an exogenous way. The formation and configuration ofsuch groups is still an open question for future research. Finally, the prob-lem here was ultimately studied as a game between elites: in particular,political entrepreneurs who can bargain over political spoils under the spec-tre of violent threats. The analysis has nothing to say about other importantconsequences of these power allocations. For example, a fruitful research av-enue might be to understand their effects on governance, accountability toethnic supporters, delivery of public goods and other development outcomes.Yet, the theory and empirical findings do explain the incidence of conflict,which clearly has an important relationship with underdevelopment.72Tables73Table 1: Summary Statistics (non-leading groups)Variables Mean Std. Dev Min Max ObsIncluded 0.317 0.465 0 1 20,597Conflict 0.0207 0.142 0 1 20,597Group size 0.0955 0.143 0.0001 0.980 20,597Rank (field) 5.867 5.635 1 32 20,597Rank (track) 5.098 5.879 1 36 20,597Distance to the capital 734.6 626.1 23.25 3,361 16,527Leader distance 612.8 504.5 0 2,335 16,801Precipitation 1,145 761.4 4.900 4,342 10,765Drought (SPEI CRU) 0.0495 0.0616 0 0.600 15,832Drought (SPEI GDM) 0.0556 0.0712 0 0.913 16,077Drought (SPI) 0.0445 0.0651 0 0.606 10,648Elevation (std) 302.0 259.0 6.424 1,739 17,922Mountains (proportion) 0.354 0.323 0 1 17,452Night light (calibrated) 0.0624 0.0634 0 0.659 6,553Night light 1.656 3.620 0 32.69 6,875Fractionalization (religion) 0.312 0.271 0 0.749 19,788Fractionalization (language) 0.196 0.271 0 0.750 19,842Log (Night light) 0.585 0.740 0 3.517 6,875Peit 165.2 4,960 0 245,536 12,898Seit = ∆log(Plit/Peit) ∗ (Xi/GDPi) 0.000986 0.0556 -1.386 0.807 11,955Notes: The table shows summary statistics of the non-leading groups used in the estimation. Includedis a dummy variable assuming value 1 if a non-leader is included as junior partner in the coalition (0 ifgroup is powerless, discriminated or self-exclusion). Conflict incidence is coded as 1 for every year wherethere is ongoing conflict from a rebel organization claiming to represent the ethnic group. Groupsize ismeasured as the fraction of countryâs total population.Variable rank (field) ranks group sizes within country and assumes value equal to 1 + the number ofgroups with higher population. Rank (track) is 1 + the number of groups with lower population.Leader distance is the distance in kilometers from the centroid of the ethnic group homeland to thecentroid of the leaderâs ethnic homeland.Drought (SPI) gives the proportion of months in the growing season that are part of the longest streakof consecutive months in that growing season with SPI1 values below -1.5. The growing season isthe growing season for the cellâs main crop, defined in the MIRCA2000 dataset v.1.1. SPI1 indexmeasures deviation from long-term normal rainfall for that month. The values are standardized wheredeviation estimates less than 1 standard deviation indicate near normal rainfall. Drought (SPEI CRU)uses the Standardized Precipitation and Evapotranspiration Index SPEI-1 from the SPEIbase v.2.3.SPEIbase is based on precipitation and potential evapotranspiration from the Climatic Research Unit ofUniversity of East Anglia CRU v.3.22. Drought (SPEI GDM) uses the Standardized Precipitation andEvapotranspiration Index SPEI1 from the SPEI Global Drought Monitor. SPEI GDM uses the GPCCâfirst guessâ product and GHCN/CAMS, while using the Thornthwaite potential evapotranspiration(PET) estimation. Precipitation gives the yearly total amount of precipitation (in millimeter) in thecell, based on monthly meteorological statistics from the Global Precipitation Climatology Centre. Allvalues are weighted (by area) mean of all grid values in the ethnic homaland. Values at grid value arefrom the PRIO-GRID dataset.Elevation (std) is the standard deviation of gridded elevation measurements (0.008330 decimal degreeresolution) intersecting with group polygon. Mountains (proportion) is the group-level area-weightedmean of the gridded proportion of mountainous terrain within the cell based on elevation, slope andlocal elevation range, taken from a high-resolution mountain raster developed for UNEPâs MountainWatch Report.EPR-2018 provides information on the 3 largest religions within group and the fraction of group as-sociated with each. Fractionalization is measured as 1 − religion21 − religion22 − religion23 − (1 −∑3ireligioni)2. The same is done for language.Night light measures average nighttime light emission from the DMSP-OLS Nighttime Lights Time SeriesVersion 4. The calibrated variable accounts for intersatellite differences and interannual sensor decayusing calibration values from Elvidge et al (2014).74Table 2: Probability of Inclusion and Con-flict(1) (2) (3)Included Conflict Conflictgroupsize 2.3176*** -0.0349(0.417) (0.093)groupsize2 -3.1275*** 0.2390 0.1937***(0.544) (0.164) (0.064)Country FE Y Y YYear FE Y Y YF test (p-value) 0.000 0.0062 -N 20,597 20,597 20,597R2 0.481 0.141 0.141Notes: Data is at the group-year level. Estimation for thesample of non-leaders. Included is a dummy variable assum-ing value 1 if a non-leader is included as junior partner inthe coalition (0 if group is powerless, discriminated or self-exclusion). Conflict incidence is coded as 1 for every yearwhere there is ongoing conflict from a rebel organizationclaiming to represent the ethnic group. Groupsize is mea-sured as the fraction of country’s total population. Robuststandard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.175Table 3: Inclusion and group size(3rd order polynomial)(1)Includedgroupsize 3.5484***(0.830)groupsize2 -8.1579***(2.571)groupsize3 4.3896**(1.945)Country FE YYear FE YN 20,597R2 0.486Notes: Data is at the group-year level. Esti-mation for the sample of non-leaders. Includedis a dummy variable assuming value 1 if a non-leader is included as junior partner in the coali-tion (0 if group is powerless, discriminated orself-exclusion). Groupsize is measured as thefraction of country’s total population. Robuststandard errors clustered at the country level inparentheses.*** p<0.01, ** p<0.05, * p<0.176Table 4: Discrimination and group size(1)Discriminatedgroupsize -1.0062***(0.308)groupsize2 1.3895***(0.400)Country FE YYear FE YN 20,597R2 0.390Notes: Data is at the group-year level. Estimation forthe sample of non-leaders. Discriminated is a dummyvariable assuming value 1 if a non-leader faces statediscrimination (0 if group is coded as junior partner,powerless, or self-exclusion). Groupsize is measuredas the fraction of country’s total population. Robuststandard errors clustered at the country level in paren-theses.*** p<0.01, ** p<0.05, * p<0.177Table 5: Probability of Inclusion and Conflict, by group size(1) (2) (3)Included Conflict Conflictgroupsize 2.3068*** -0.0379(0.418) (0.092)groupsize2 -3.0877*** 0.2502 0.2002***(0.545) (0.161) (0.065)leader size 0.0766 0.0216 0.0191(0.211) (0.077) (0.079)Country FE Y Y YYear FE Y Y YN 20,597 20,597 20,597R2 0.481 0.142 0.141Notes: Data is at the group-year level. Estimation for the sample of non-leaders.Included is a dummy variable assuming value 1 if a non-leader is included asjunior partner in the coalition (0 if group is powerless, discriminated or self-exclusion). Conflict incidence is coded as 1 for every year where there is ongoingconflict from a rebel organization claiming to represent the ethnic group. Group-size is measured as the fraction of country’s total population. Leader size is thegroupsize of leading group (most powerful group) in the country-year. Robuststandard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.1Table 6: Probability of Inclusion and Conflict, by group size (Leader &Country-year FEs)(1) (2) (3) (4) (5) (6)Included Conflict Conflict Included Conflict Conflictgroupsize 2.3648*** 0.0502 2.3649*** 0.0368(0.438) (0.073) (0.598) (0.090)groupsize2 -3.1689*** 0.0546 0.1267** -3.2246*** 0.1004 0.1602**(0.589) (0.092) (0.049) (0.951) (0.137) (0.076)Leader FE Y Y Y N N NCountry-Year FE N N N Y Y YN 20,597 20,597 20,597 20,597 20,597 20,597R2 0.542 0.182 0.182 0.617 0.454 0.454Notes: Data is at the group-year level. Estimation for the sample of non-leaders. Included isa dummy variable assuming value 1 if a non-leader is included as junior partner in the coalition(0 if group is powerless, discriminated or self-exclusion). Conflict incidence is coded as 1 forevery year where there is ongoing conflict from a rebel organization claiming to represent theethnic group. Groupsize is measured as the fraction of country’s total population. Leaderfixed-effects is dummy for the ethnic identity of the leading group. Robust standard errorsclustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.178Table 7: Conflict and group size, by power status(1) (2) (3)Conflict Conflict ConflictExcluded groups Leaders + Included Included (non-leaders)groupsize 0.1495*** 0.0294 0.0130(0.054) (0.019) (0.061)Country FE Y Y YYear FE Y Y YMean Dep.V ar 0.0267 0.0081 0.0078N 14,062 11,670 6,535R2 0.160 0.112 0.198Notes: Regression of conflict incidence on group size conditional on political status. Groupsize measured as fraction of country’s total population. Estimation is done for excluded groupsonly (powerless, discriminated, self-exclusion) in column (1), for leaders and included groups(monopoly, dominant, senior and junior partners) in column (2), and for included groups only(junior partners) in column (3). Robust standard errors clustered at the country level in paren-theses.*** p<0.01, ** p<0.05, * p<0.179Table 8: Inclusion and Conflict by group size (Split groups) - Africa andAsia(1) (2) (3) (4) (5)Included Conflict Conflict Conflict Conflictif excluded if includedgroupsize 2.7138*** 0.0121 0.1960*** 0.0696*(0.826) (0.094) (0.066) (0.036)groupsize2 -7.6834*** 0.1897 0.2031***(2.722) (0.154) (0.069)groupsize3 4.7720**(2.118)Ethnic FE Y Y Y Y YN 15,204 15,204 15,204 10,182 5,022R2 0.370 0.159 0.159 0.191 0.152Notes: Estimation with sample of ethnic groups present in at least two countries, using ethnicfixed-effects. Estimation for the sample of non-leaders. Included is a dummy variable assumingvalue 1 if a non-leader is included as junior partner in the coalition (0 if group is powerless,discriminated or self-exclusion). Conflict incidence is coded as 1 for every year where there isongoing conflict from a rebel organization claiming to represent the ethnic group. Groupsizeis measured as the fraction of country’s total population. A cubic term is included in theestimation of the effect of group size on inclusion. This is necessary because the peak ofinclusion probability occurs for groups around 20% of less (see Figure 7). In column (4),estimation for the sample of excluded groups. In column (5), estimation for junior partnersonly. Robust standard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.180Table 9: Inclusion and Conflict by group size (Split groups) - AfricaAll Selected groups(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)Included Conflict Conflict Conflict Conflict Included Conflict Conflict Conflict Conflictif excluded if included if excluded if includedgroupsize 2.0493* -0.0548 0.2479** 0.0263 2.5109* -0.0042 0.2626** -0.0178(1.171) (0.127) (0.100) (0.045) (1.319) (0.136) (0.104) (0.021)groupsize2 -8.2142** 0.2805 0.2243** -9.4630** 0.2248 0.2206**(3.475) (0.200) (0.087) (3.843) (0.206) (0.094)groupsize3 5.8070** 6.6638**(2.533) (2.811)Ethnic FE Y Y Y Y Y Y Y Y Y YN 5,789 5,789 5,789 3,576 2,213 4,815 4,815 4,815 3,079 1,736R2 0.421 0.199 0.198 0.220 0.237 0.457 0.202 0.202 0.227 0.254Notes: Estimation with sample of ethnic groups present in at least two countries, using ethnic fixed-effects. Estimation for the sample of non-leaders. Included is a dummy variable assuming value 1 if a non-leader is included as junior partner in the coalition (0 if group is powerless,discriminated or self-exclusion). Conflict incidence is coded as 1 for every year where there is ongoing conflict from a rebel organization claimingto represent the ethnic group. Groupsize is measured as the fraction of country’s total population. In columns (4) and (9), estimation for thesample of excluded groups. In column (5) and (10), estimation for junior partners only. Columns 1-5 report estimation for all African groups. InColumns 6-10, a selected sample of groups is used, where non-native and umbrella groups are removed. An ethnic group in one country is calledumbrella if it is related to two or more subgroups in another country. Robust standard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.181Table 10: Effects of Relative Shocks on Conflict and Inclusion(1) (2) (3) (4)Onset Onset Included IncludedSt(Pleader/Pgroup) 0.0564* 0.0058 -0.0754 -0.4094*(0.029) (0.035) (0.085) (0.245)St(Pleader/Pgroup) ∗ size 0.4958** 6.0691*(0.220) (3.303)St(Pleader/Pgroup) ∗ size2 -15.6417*(8.832)Ethnic FE Y Y Y YYear FE Y Y Y YN 14,708 14,687 6,761 6,761R2 0.058 0.064 0.072 0.078Notes: Estimation of the effects of relative economic shocks on conflict and inclusion.Estimations use first difference. Onset is 1 if an ethnic conflict started in that year. In-cluded is measured by change in inclusion status (1 if group moved from excluded tojunior partner; -1 if group moved from junior partner to excluded; and 0 otherwise). Rel-ative economic shocks are first-difference of the ratio between leader-specific and ethnicgroup-specific prices. Relative economic shocks are interacted with group size and groupsize square when indicated. Regressions include first and second lags (not shown) ofprice shocks. Group size measured as fraction of country’s total population. All regres-sions include ethnic fixed-effects. Robust standard errors clustered at the country level inparentheses.*** p<0.01, ** p<0.05, * p<0.182Table 11: Effects of Relative Shocks on Conflict and Inclusion (Noleader change)(1) (2) (3) (4)Onset Onset Included IncludedSt(Pleader/Pgroup) 0.0434* -0.0062 -0.0386 -0.6142*(0.025) (0.015) (0.164) (0.343)St(Pleader/Pgroup) ∗ size 0.4575*** 10.0974*(0.164) (5.258)St(Pleader/Pgroup) ∗ size2 -26.7154*(13.820)Ethnic FE Y Y Y YYear FE Y Y Y YN 14,212 14,191 6,431 6,431R2 0.054 0.059 0.107 0.111Notes: Same estimation as Table 10, but excluding years of changes in the leading group.Estimation of the effects of relative economic shocks on conflict and inclusion. Estimationsuse first difference. Onset is 1 if an ethnic conflict started in that year. Included is measuredby change in inclusion status (1 if group moved from excluded to junior partner; -1 if groupmoved from junior partner to excluded; and 0 otherwise). Relative economic shocks arefirst-difference of the ratio between leader-specific and ethnic group-specific prices. Relativeeconomic shocks are interacted with group size and group size square when indicated.Regressions include first and second lags (not shown) of price shocks. Group size is measuredas the fraction of country’s total population. All regressions include ethnic fixed-effects.Robust standard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.183Table 12: Structural EstimatesResults Std. Dev.α 0.1678 (0.0014)c/pi 0.0358 (0.0002)γ 1.0004 (0.0192)η 0.0760 (0.0002)φ0 0.0088 (0.0002)φc 0.3282 (0.0074)Obs 20,597Log-Likelihood -5,062Notes: The table presents the results of structuralestimation. Parameters are estimated by MaximumSimulated Likelihood (MSE). 50,000 draws were usedin the simulation. Standard errors computed byOuter-Product Gradient Approximation.84Table 13: CounterfactualsProb ofConflictProb ofInclusionCoalitionsizeProb of leader’sremovalModel 2.52% 28.9% 66.0% 13.5%Democratization(largest group = leader) 1.63% 29.6% 73.7% 5.3%Military capacity(∆θ/θ = −20%) 2.02% 26.7% 66.6% 8.4%Aid to groups (increased opportunity cost)(∆cc/cc = +20%)2.27% 27.2% 66.1% 12.2%Unconditional Gov Aid/Reduced Arming( ∆pi/pi = +25% or∆cl/cl = ∆cc/cc = −20%)3.10% 30.1% 64.8% 16.3%Gov Aid Conditional on Peace( ∆pi/pi = +5%) 1.72% 31.4% 69.8% 8.4%Sanction against regimes in conflict(∆cl/pi = +5%)1.54% 32.0% 70.9% 6.9%Sanction against gov established throughconflict (∆pi/pi = −5%) 1.63% 30.7% 69.9% 7.7%Quotas(min 25% of pop share for groups > 20%) 2.17% 35.2% 78.1% 11.8%Notes: The table shows the predictions given by the estimated model parameters and for eachcounterfactual experiment. Coalition size is calculated by multiplying the probability of inclusionand group size (measured as the share of country’s population). Probability of leader’s removalis calculated by the sum of probabilities of challengers being strong enough to be excluded andin conflict, multiplied by their expected strength conditional on the first case. Democratizationchanges the composition of challengers by making the largest ethnic group always the leader.Military capacity decreases the strength of every group by 20% (∆θ/θ = −20%). Aid togroups decreases the opportunity cost of conflict for challengers by 20% (∆cc/cc = +20%).Unconditional Gov Aid increases the value of rents from office by 25% (∆pi/pi = +25%),which is equivalent to reduce costs for leader and challengers by 20% (∆cc/cc = −20% and∆cl/cl = −20%). Gov Aid Conditional on Peace increases the rents (pi) by 5% conditionalif there is no conflict. If there is conflict, leader and challenger fight over the non-incrementedvalue of pi. Sanctions against regimes in conflict increases the relative cost of conflict for theleader by 5% (∆cl/pi = +5%). Sanctions against gov established through conflict decreasesthe value of rents by 5% for successful rebels. Quotas changes the structure of the game imposinga minimum enforced allocation of power: every group comprising more than 20% of the country’spopulation must be given at least 25% of the group size85FiguresFigure 1: ModelNotes: Extensive-form representation of the model.86Figure 2: Results (Equilibrium path)Notes:Illustration of the model results on the equilibrium path. θ is the innate strength of thechallenger, defined as the winning probability of a conflict; τ is the endogenous equilibrium power-sharing transfer made by the leader to the challenger; cc and cl are the costs of conflict forchallengers and leader, respectively; pi is the value of the bargaining object; α is the marginalchange in the winning probability in response to τ . "Peace" or "Conflict" is the endogenousdecision taken by the challenger.87Figure 3: Probability of Inclusion and Conflict0.2.4.60 .2 .4 .6 .8 1Group size (share of country’s population)95% CI lpoly smoothkernel = gaussian, degree = 2, bandwidth = .1, pwidth = .07Probability of Inclusion0.1.2.30 .2 .4 .6 .8 1Group size (share of country’s population)95% CI lpoly smoothkernel = gaussian, degree = 2, bandwidth = .1, pwidth = .12Probability of Conflict88Figure 4: Inclusion - Predicted values (Country FE regression)0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1−0.500.5groupsizeInclusion (predicted value)  QuadraticCubicNotes: This figure reports the predicted values of inclusion probability from estimations fromTables 2 and 3. It shows how predicted values of inclusion probability vary with group size.Values for country and year fixed-effects are set as zero. Therefore, the plotted predicted valuesmust be interpreted as the expected increase in probability of inclusion compared to a group ofsize zero (or marginally above zero) in the same country and year.89Figure 5: Effect of group size on conflict, by Power status−. groupsizeExcluded groups Included groupsNotes: This figure shows the results of the estimation reported in columns 1 and 3 of Table7. Regression of conflict incidence on group size conditional on political status. Group sizemeasured as fraction of country’s total population. Estimation is done for excluded groups only(powerless, discriminated, self-exclusion) in the panel on the left, and for included groups only(junior partners) in the panel on right. A 95% confidence interval is shown based on robuststandard errors clustered at the country level.90Figure 6: Illustration: split groupsNotes: Illustration of the identification strategy of the estimations with ethnic fixed-effects. TheShia Arabs are present in different countries with varying size, measured by the share of eachcountry’s total population. For each case, it is reported the most frequent power-status outcomewhen the group was not the leader (included or excluded), and whether the group was ever inconflict.91Figure 7: Inclusion - Predicted values (Ethnic FE regression)0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−0.7−0.6−0.5−0.4−0.3−0.2− (predicted value)  Africa and AsiaAfricaAfrica (selected groups)Notes: This figure reports the predicted values of inclusion probability from estimations fromTables 8 and 9. It shows how predicted values of inclusion probability vary with group size.Values for ethnic fixed-effects are set as zero. Therefore, the plotted predicted values must beinterpreted as the expected increase in probability of inclusion compared to a co-ethnic group ofsize zero (or marginally above zero).Figure 8: Effects of cc/pi ↓ (pi/cc ↑)Notes: Illustration of the model’s comparative statics for relative economic shocks. In this exam-ple, the shock is an increase in the ratio between rents and opportunity cost of conflict for thechallenger (pi/cc).92Figure 9: Distributional shocks - Illustration (Togo)05010015020025030035040019571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009Price Indexcoffee cottonNotes: Illustration of the variation behind the relative economic shocks. Ewe’s homeland growsrelatively more coffee, while Kabre grows relatively more cotton. Coffee and cotton prices vary indifferent rates.93Figure 10: Model fit - Probability of Inclusion and Conflict, by groupsizegroupsize0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 of InclusionProbability of ConflictNotes: The figure shows the simulated probability of inclusion and conflict conditional on popu-lation share using the estimated parameters (as reported in Table 12).94Figure 11: Counterfactuals - Effects on Conflict Probability0% 1% 2% 3% 4%ModelDemocratization (largest group = leader)Military capacity (Δθ/θ = -20%)Aid to groups (∆cc/cc = +20%)Unconditional Gov Aid (∆π/π = +25%)Gov Aid Conditional on Peace (∆π/π = +5%)Sanctions against regimes in conflict (∆cl/π= +5%)Sanctions against gov established through conflict (∆π/π = -5%)Quotas (min 25% of pop share for groups >20%)-35%-20%-10%+23%-32%-39%-35%-14%Notes: The figure shows a bar chart with the probability of conflict predicted by the modelparameters and for each counterfactual experiment. Percentage variation with respect to themodel is shown on the right side of each bar. Democratization changes the composition ofchallengers by making the largest ethnic group always the leader. Military capacity decreasesthe strength of every group by 20% (∆θ/θ = −20%). Aid to groups decreases the opportunitycost of conflict for challengers by 20% (∆cc/cc = +20%). Uncondtional Gov Aid increases thevalue of rents from office by 25% (∆pi/pi = +25%), which is equivalent to reduce costs for leaderand challengers by 20% (∆cc/cc = −20% and ∆cl/cl = −20%). Gov Aid Conditional on Peaceincreases the rents (pi) by 5% conditional if there is no conflict. If there is conflict, leader andchallenger fight over the non-incremented value of pi. Sanctions against regimes in conflictincreases the relative cost of conflict for the leader by 5% (∆cl/pi = +5%). Sanctions againstgov established through conflict decreases the value of rents by 5% for successful rebels. Quotaschanges the structure of the game imposing a minimum enforced allocation of power: every groupcomprising more than 20% of the country’s population must be given at least 25% of the groupsize95Bibliography[1] D. Acemoglu and J. A. Robinson. A theory of political transitions.American Economic Review, 91(4):938–963, 2001. → pages 7, 9, 13[2] D. Acemoglu, G. Egorov, and K. Sonin. Coalition formation innon-democracies. The Review of Economic Studies, 75(4):987–1009,2008. → page 9[3] D. Acemoglu, A. Vindigni, and D. Ticchi. Persistence of civil wars.Journal of the European Economic Association, 8(2-3):664–676, 2010.→ page 9[4] D. Acemoglu, T. Reed, and J. A. Robinson. Chiefs: Economicdevelopment and elite control of civil society in Sierra Leone. Journalof Political Economy, 122(2):319–368, 2014. → page 9[5] A. Adhvaryu, J. E. Fenske, G. Khanna, and A. Nyshadham.Resources, conflict, and economic development in Africa. Technicalreport, National Bureau of Economic Research, 2018. → page 8[6] A. Alesina, W. Easterly, and J. Matuszeski. Artificial states. Journalof the European Economic Association, 9(2):246–277, 2011. → page 46[7] L. R. Arriola. Patronage and political stability in Africa. ComparativePolitical Studies, 42(10):1339–1362, 2009. → pages 9, 14[8] A. I. Asiwaju. The conceptual framework. Partitioned Africans, pages1–18, 1985. → page 46[9] J.-F. Bayart. The State in Africa: the politics of the belly. Polity,2009. → pages 2, 28[10] S. Bazzi and C. Blattman. Economic shocks and conflict: Evidencefrom commodity prices. American Economic Journal:Macroeconomics, 6(4):1–38, 2014. → pages 8, 4796[11] N. Berman, M. Couttenier, D. Rohner, and M. Thoenig. This mine ismine! how minerals fuel conflicts in africa. American EconomicReview, 107(6):1564–1610, 2017. → page 8[12] M. Bernhard, Ö. F. Örsün, and R. Bayer. Democratization in conflictresearch: How conceptualization affects operationalization and testingoutcomes. International interactions, 43(6):941–966, 2017. → page 52[13] C. Bidner, P. Francois, and F. Trebbi. A theory of minimalistdemocracy. Technical report, National Bureau of Economic Research,2014. → page 14[14] H. M. Binningsbø and K. Dupuy. Using power-sharing to win a war:The implementation of the lomé agreement in Sierra Leone. AfricaSpectrum, 44(3):87–107, 2009. → page 2[15] C. Blattman and J. Annan. Can employment reduce lawlessness andrebellion? A field experiment with high-risk men in a fragile state.American Political Science Review, 110(1):1–17, 2016. → page 68[16] C. Blattman and E. Miguel. Civil war. Journal of EconomicLiterature, 48(1):3–57, 2010. → pages 3, 51[17] C. Boix. Democracy and redistribution. Cambridge University Press,2003. → page 9[18] D. Brown. The state and ethnic politics in Southeast Asia. Routledge,2003. → pages 2, 27[19] M. Buehler and M. Ayari. The autocratâs advisors: Opening theblack box of ruling coalitions in Tunisia’s authoritarian regime.Political Research Quarterly, 71(2):330–346, 2018. → page 9[20] H. Buhaug, S. Gates, and P. Lujala. Geography, rebel capability, andthe duration of civil conflict. Journal of Conflict Resolution, 53(4):544–569, 2009. → page 39[21] R. Burgess, R. Jedwab, E. Miguel, A. Morjaria, and G. Padró iMiquel. The value of democracy: evidence from road building inKenya. American Economic Review, 105(6):1817–51, 2015. → pages3, 12[22] M. Burke, S. M. Hsiang, and E. Miguel. Climate and conflict. AnnualReview Economics, 7(1):577–617, 2015. → page 897[23] L.-E. Cederman, A. Wimmer, and B. Min. Why do ethnic groupsrebel? New data and analysis. World Politics, 62(01):87–119, 2010. →pages 1, 4, 11, 28, 31, 105[24] L.-E. Cederman, K. S. Gleditsch, I. Salehyan, and J. Wucherpfennig.Transborder ethnic kin and civil war. International Organization, 67(2):389–410, 2013. → page 43[25] S. Chassang and G. Padro-i Miquel. Economic shocks and civil war.Quarterly Journal of Political Science, 4(3):211–228, 2009. → page 7[26] S. Chassang and G. Padro-i Miquel. Conflict and deterrence understrategic risk. The Quarterly Journal of Economics, 125(4):1821–1858,2010. → page 7[27] S. Chaturvedi and S. Das. Group size and political representationunder alternate electoral systems. Mimeo, 2018. → page 10[28] B. Crost, J. Felter, and P. Johnston. Aid under fire: Developmentprojects and civil conflict. American Economic Review, 104(6):1833–56, 2014. → page 67[29] B. Crost, J. H. Felter, and P. B. Johnston. Conditional cash transfers,civil conflict and insurgent influence: Experimental evidence from thePhilippines. Journal of Development Economics, 118:171–182, 2016.→ page 68[30] E. Dal Bó and R. Powell. A model of spoils politics. AmericanJournal of Political Science, 53(1):207–222, 2009. → page 7[31] B. B. De Mesquita, A. Smith, J. D. Morrow, and R. M. Siverson. Thelogic of political survival. MIT press, 2005. → page 8[32] A. Debs. On dictatorships. PhD thesis, Massachusetts Institute ofTechnology, 2007. → pages 9, 12[33] E. K. Denny and B. F. Walter. Ethnicity and civil war. Journal ofPeace Research, 51(2):199–212, 2014. → page 1[34] O. Dube and J. F. Vargas. Commodity price shocks and civil conflict:Evidence from colombia. The Review of Economic Studies, 80(4):1384–1421, 2013. → page 898[35] W. Easterly and R. Levine. Africa’s growth tragedy: policies andethnic divisions. The Quarterly Journal of Economics, 112(4):1203–1250, 1997. → page 28[36] G. Egorov and K. Sonin. Dictators and their viziers: Endogenizingthe loyalty–competence trade-off. Journal of the European EconomicAssociation, 9(5):903–930, 2011. → page 8[37] S. Ejdemyr, E. Kramon, and A. L. Robinson. Segregation, ethnicfavoritism, and the strategic targeting of local public goods.Comparative Political Studies, 51(9):1111–1143, 2018. → pages 3, 12[38] J. D. Fearon. Rationalist explanations for war. InternationalOrganization, 49(3):379–414, 1995. → pages 7, 12[39] J. D. Fearon. Bargaining over objects that influence future bargainingpower. Mimeo, 1996. → pages 7, 12[40] J. D. Fearon. Why do some civil wars last so much longer thanothers? Journal of peace research, 41(3):275–301, 2004. → page 13[41] J. D. Fearon and D. D. Laitin. Ethnicity, insurgency, and civil war.American Political Science Review, 97(01):75–90, 2003. → page 39[42] P. Francois, I. Rainer, and F. Trebbi. The dictator’s inner circle.Technical report, National Bureau of Economic Research, 2014. →page 8[43] P. Francois, I. Rainer, and F. Trebbi. How is power shared in Africa?Econometrica, 83(2):465–503, 2015. → pages3, 8, 9, 14, 29, 42, 54, 63, 121[44] P. Francois, F. Trebbi, and K. Xiao. Factions in nondemocracies:Theory and evidence from the chinese communist party. Technicalreport, National Bureau of Economic Research, 2016. → page 9[45] J. Gandhi and A. Przeworski. Cooperation, cooptation, and rebellionunder dictatorships. Economics & Politics, 18(1):1–26, 2006. → page8[46] B. Gebrewold. Anatomy of violence: understanding the system ofconflict and violence in Africa. Routledge, 2009. → page 6399[47] L. Girardin, P. Hunziker, L.-E. Cederman, N.-C. Bormann, andM. Vogt. Growup-geographical research on war, unified platform.ETH Zurich. http://growup. ethz. ch, 2015. → page 33[48] A. Guariso and T. Rogall. Rainfall inequality, political power, andethnic conflict in Africa. 2017. → page 8[49] S. Haber. Authoritarian government. In The Oxford Handbook ofPolitical Economy. 2006. → pages 9, 12[50] C. Hartzell and M. Hoddie. Institutionalizing peace: Power sharingand post-civil war conflict management. American Journal of PoliticalScience, 47(2):318–332, 2003. → page 2[51] D. Horowitz. Ethnic Groups in Conflict. University of CaliforniaPress, 1985. → pages 3, 12[52] P. Jimenez-Ayora and M. A. Ulubaşoğlu. What underlies weak states?The role of terrain ruggedness. European Journal of PoliticalEconomy, 39:167–183, 2015. → page 39[53] M. D. König, D. Rohner, M. Thoenig, and F. Zilibotti. Networks inconflict: Theory and evidence from the great war of africa.Econometrica, 85(4):1093–1132, 2017. → pages 14, 70[54] E. Kramon and D. N. Posner. Ethnic favoritism in education inKenya. 2016. → pages 3, 12[55] V. Krause and S. Suzuki. Causes of civil war in Asia and Sub-SaharanAfrica: A comparison. Social Science Quarterly, 86(1):160–177, 2005.→ page 1[56] M. Kudamatsu and T. Besley. Making autocracy work. Institutionsand economic performance, pages 452–510, 2008. → page 9[57] R. Lemarchand. Consociationalism and power sharing in Africa:Rwanda, Burundi, and the Democratic Republic of the Congo.African Affairs, 106(422):1–20, 2006. → page 2[58] J. Lyall, Y.-Y. Zhou, and K. Imai. Can economic assistance shapecombatant support in wartime? Experimental evidence fromAfghanistan. Mimeo, 2018. → page 68100[59] B. Magaloni. Credible power-sharing and the longevity ofauthoritarian rule. Comparative Political Studies, 41(4-5):715–741,2008. → pages 9, 12[60] E. D. Mansfield and J. Snyder. Electing to fight: Why EmergingDemocracies go to War. MIT Press, 2007. → page 52[61] M. Martinez-Bravo. The role of local officials in new democracies:Evidence from Indonesia. American Economic Review, 104(4):1244–87, 2014. → page 9[62] M. Martinez-Bravo, P. Mukherjee, and A. Stegmann. Thenon-democratic roots of elite capture: Evidence from Soeharto mayorsin Indonesia. Econometrica, 85(6):1991–2010, 2017. → page 9[63] L. Mayoral and D. Ray. Groups in conflict: Size matters, but not inthe way you think. Mimeo, 2017. → page 8[64] A. Mehler. Peace and power sharing in Africa: a not so obviousrelationship. African Affairs, 108(432):453–473, 2009. → page 2[65] S. Michalopoulos. The origins of ethnolinguistic diversity. AmericanEconomic Review, 102(4):1508–39, 2012. → page 39[66] E. Miguel, S. Satyanath, and E. Sergenti. Economic shocks and civilconflict: An instrumental variables approach. Journal of politicalEconomy, 112(4):725–753, 2004. → page 8[67] A. Mitra and D. Ray. Implications of an economic theory of conflict:Hindu-muslim violence in India. Journal of Political Economy, 122(4):719–765, 2014. → page 8[68] C. Monfreda, N. Ramankutty, and J. A. Foley. Farming the planet: 2.geographic distribution of crop areas, yields, physiological types, andnet primary production in the year 2000. Global biogeochemical cycles,22(1), 2008. → pages 47, 48[69] J. G. Montalvo and M. Reynal-Querol. Ethnic polarization, potentialconflict, and civil wars. American economic review, 95(3):796–816,2005. → page 28[70] S. Naidu, J. A. Robinson, and L. E. Young. Social origins ofdictatorships: Elite networks and political transitions in Haiti.Unpublished Manuscript, 2015. → page 9101[71] D. C. North, J. J. Wallis, B. R. Weingast, et al. Violence and socialorders: A conceptual framework for interpreting recorded humanhistory. Cambridge University Press, 2009. → page 9[72] N. Nunn and N. Qian. US food aid and civil conflict. AmericanEconomic Review, 104(6):1630–66, 2014. → page 68[73] D. N. Posner. Institutions and Ethnic Politics in Africa. CambridgeUniversity Press, 2005. → page 29[74] R. Powell. Bargaining theory and international conflict. AnnualReview of Political Science, 5(1):1–30, 2002. → page 7[75] R. Powell. Bargaining and learning while fighting. American Journalof Political Science, 48(2):344–361, 2004. → page 7[76] R. Powell. The inefficient use of power: Costly conflict with completeinformation. American Political science review, 98(2):231–241, 2004.→ page 7[77] R. Powell. War as a commitment problem. International organization,60(1):169–203, 2006. → pages 7, 14[78] I. Rainer and F. Trebbi. New tools for the analysis of political powerin Africa. Technical report, National Bureau of Economic Research,2012. → pages 54, 121[79] P. Roessler. The enemy within: Personal rule, coups, and civil war inAfrica. World Politics, 63(02):300–346, 2011. → pages 3, 9[80] P. Roessler. Ethnic Politics and State Power in Africa: The Logic ofthe Coup-Civil War Trap. Cambridge University Press, 2016. → pages2, 3, 10, 12, 27, 31, 53[81] T. Rogall. Mobilizing the masses for genocide. UnpublishedManuscript, 2014. → pages 3, 12[82] P. Schleiter. Democracy, authoritarianism, and ministerial selection inRussia: How presidential preferences shape technocratic cabinets.Post-Soviet Affairs, 29(1):31–55, 2013. → page 9[83] I. S. Spears. Understanding inclusive peace agreements in Africa: theproblems of sharing power. Third World Quarterly, 21(1):105–118,2000. → page 2102[84] M. Vogt, N.-C. Bormann, S. Rüegger, L.-E. Cederman, P. Hunziker,and L. Girardin. Integrating data on ethnicity, geography, andconflict: The ethnic power relations data set family. Journal ofConflict Resolution, 59(7):1327–1342, 2015. → pages 28, 43[85] M. Weintraub. Do all good things go together? Developmentassistance and insurgent violence in civil war. The Journal of Politics,78(4):989–1002, 2016. → page 67[86] J. Wucherpfennig, N. B. Weidmann, L. Girardin, L.-E. Cederman, andA. Wimmer. Politically relevant ethnic groups across space and time:Introducing the geoepr dataset. Conflict Management and PeaceScience, 28(5):423–437, 2011. → page 32[87] J. Wucherpfennig, N. W. Metternich, L.-E. Cederman, and K. S.Gleditsch. Ethnicity, the state, and the duration of civil war. WorldPolitics, 64(1):79–115, 2012. → page 30103Appendix AAdditional TablesTable A1: EPR-2018: Descriptive Statistics (Africa and Asia)Variables Mean Std. Dev Min Max ObsCountry-year level:Number of relevant groups 4.692 4.717 0 37 5,484Number of relevant groups (if more than 1) 5.454 4.659 2 37 4,718Number of included groups 2.474 2.332 1 15 4,718Number of excluded groups 2.981 4.265 0 36 4,718Governmental Ethnic conflict incidence (ongoing) 0.0566 0.231 0 1 4,718Governmental Ethnic conflict history 0.381 0.486 0 1 4,718Group-year level:Population share 0.162 0.224 0.0001 0.981 25,732Monopoly (leader) 0.0339 0.181 0 1 25,732Dominant (leader) 0.0566 0.231 0 1 25,732Senior partner (leader) 0.109 0.312 0 1 25,732Junior partner (included) 0.254 0.435 0 1 25,732Self-exclusion (not included) 0.0158 0.125 0 1 25,732Powerless (not included) 0.366 0.482 0 1 25,732Discriminated (not included) 0.165 0.371 0 1 25,732Governmental conflict incidence (ongoing) 0.0183 0.134 0 1 25,732Governmental conflict history 0.129 0.336 0 1 25,732Notes: The table shows summary statistics of the main dataset (Ethnic Power Relations 2018).The dataset is organized at the ethnic group level. The first panel shows statistics collapsedat the country-year level. The second panel presents statistics at the ethnic group-year level.Included groups are the number of groups coded as "monopoly", "dominant", "senior partner",and "junior partner". Excluded groups are "powerless", "self-exclusion" and "discriminated".Conflict incidence is coded 1 for every period of an ongoing ethnic conflict over the control ofthe government (and 0, otherwise). Each power category is a dummy indicating if the grouphas that level of power in that year.104Throughout the analysis, I use ethnicity as the unit at which I observeconflict and inclusion decisions. Because ethnicity is not a perfect measureof political identity, it is possible that we may observe inclusion of somemembers of a group while at the same time a distinct sub-set of the groupmay be excluded and perhaps even in conflict with the included members ofthe same group. For example, a rebel organization that claims to fight fora particular group may have no effective relationship with their co-ethnicswho are in power. Despite this possibility, Table A2 validates the powerand salience of ethnic identity in this context. There is a strong relationshipbetween exclusion and conflict.1 The correlation holds comparing both dif-ferent groups in the same country and the same group at different points oftime (ethnic fixed-effects).Clearly, this result does not imply causality. Included and excludedgroups may differ in several aspects, and, in fact, the proposed theory impliesincluded groups and excluded ones will differ in their threat capabilities. Thedata though shows that there is a correlation between the endogenous choiceof exclusion on the part of the leader, and the decision to start a conflict onthe part of a challenger.1Similar results were also obtained by Cederman et al. [2010]105Table A2: Conflict and Power statusIncluding leaders Excluding leaders(1) (2) (3) (4) (5) (6) (7) (8)Incidence Incidence Onset Onset Incidence Incidence Onset OnsetExcluded 0.0157** 0.0431*** 0.0028* 0.0096** 0.0137* 0.0277*** 0.0020 0.0069*(0.006) (0.013) (0.002) (0.004) (0.008) (0.010) (0.002) (0.004)Ethnic FE N Y N Y N Y N YCountry FE Y N Y N Y N Y NYear FE Y Y Y Y Y Y Y YMean Dep.V ar 0.0182 0.0182 0.0034 0.0034 0.0207 0.0207 0.0037 0.0037N 25,732 25,732 25,172 25,172 20,597 20,597 20,101 20,101R2 0.112 0.271 0.019 0.078 0.132 0.321 0.026 0.113Notes: Regressions of Conflict (Incidence or Onset) on exclusion category. Group is consideredexcluded if it is not included in the coalition either as a leader (monopoly, dominant, senior partner)or included in a power-sharing arrangement (junior partner). Thus, excluded groups are the onesclassified as either powerless, self-exclusion or discriminated. Columns 1-4 include all groups in theregression. Estimations in Columns 5-8 do not include the leaders in the sample, thus it comparesjunior partners and the excluded categories. All regressions include year-fixed effects, and Countryfixed-effect or Ethnic fixed-effects as indicated. Robust standard errors clustered at the country levelin parentheses.*** p<0.01, ** p<0.05, * p<0.1106Table A3: Probability of Inclusion and Conflict, by group size (control-ling for ranking of group size)(1) (2) (3) (4) (5) (6)Included Conflict Conflict Included Conflict Conflictgroupsize 2.0952*** -0.0470 2.0370*** -0.0296(0.385) (0.099) (0.412) (0.098)groupsize2 -2.8849*** 0.2522 0.1927*** -2.8097*** 0.2330 0.1953***(0.489) (0.170) (0.065) (0.506) (0.169) (0.065)rank (field) -0.0069 -0.0004 -0.0002(0.009) (0.000) (0.000)rank (track) 0.0079 -0.0001 -0.0003(0.009) (0.000) (0.000)Country FE Y Y Y Y Y YYear FE Y Y Y Y Y YN 20,597 20,597 20,597 20,597 20,597 20,597R2 0.483 0.142 0.141 0.485 0.141 0.141Notes: Data is at the group-year level. Estimation for the sample of non-leaders. Included is adummy variable assuming value 1 if a non-leader is included as junior partner in the coalition (0if group is powerless, discriminated or self-exclusion). Conflict incidence is coded as 1 for everyyear where there is ongoing conflict from a rebel organization claiming to represent the ethnicgroup. Groupsize is measured as the fraction of country’s total population. Variable rank(field) ranks group sizes within country and assumes value equal to 1 + the number of groupswith higher population. Rank (track) is 1 + the number of groups with lower population.Robust standard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.1107Table A4: Probability of Inclusion and Conflict, by group size (con-trolling for distance to capital and leader’s homeland)(1) (2) (3) (4) (5) (6)Included Conflict Conflict Included Conflict Conflictgroupsize 2.3994*** -0.1260 2.4321*** -0.0049(0.509) (0.127) (0.501) (0.110)groupsize2 -3.3913*** 0.3719* 0.1987** -3.4118*** 0.1725 0.1661***(0.736) (0.220) (0.077) (0.673) (0.156) (0.056)distance to capital -0.0001 -0.0000 -0.0000(0.000) (0.000) (0.000)leader distance -0.0001 -0.0000 -0.0000(0.000) (0.000) (0.000)Country FE Y Y Y Y Y YYear FE Y Y Y Y Y YN 16,527 16,527 16,527 16,801 16,801 16,801R2 0.494 0.145 0.144 0.504 0.150 0.150Notes: Data is at the group-year level. Estimation for the sample of non-leaders. Includedis a dummy variable assuming value 1 if a non-leader is included as junior partner in thecoalition (0 if group is powerless, discriminated or self-exclusion). Conflict incidence iscoded as 1 for every year where there is ongoing conflict from a rebel organization claimingto represent the ethnic group. Groupsize is measured as the fraction of country’s totalpopulation. Leader distance is the distance in kilometers from the centroid of the ethnicgroup homeland to the centroid of the leader’s ethnic homeland. Robust standard errorsclustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.1108Table A5: Probability of Inclusion and Conflict, by group size (controlling for rainfall)(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)Included Conflict Conflict Included Conflict Conflict Included Conflict Conflict Included Conflict Conflictgroupsize 2.7198*** -0.0671 2.5757*** -0.1267 2.5741*** -0.1195 2.8459*** -0.0553(0.580) (0.154) (0.549) (0.122) (0.545) (0.123) (0.608) (0.160)groupsize2 -3.8850*** 0.3781 0.2778** -3.5777*** 0.3715* 0.1991** -3.5794*** 0.3714* 0.2085** -4.0036*** 0.3672 0.2858**(0.929) (0.258) (0.117) (0.783) (0.215) (0.078) (0.783) (0.217) (0.080) (0.949) (0.260) (0.121)precipitation -0.0000 -0.0000 -0.0000(0.000) (0.000) (0.000)Drought (SPEI CRU) -0.1180 -0.0017 -0.0023(0.120) (0.031) (0.031)Drought (SPEI GDM) -0.1010 0.0159 0.0170(0.070) (0.033) (0.033)Drought (SPI) -0.0476 -0.0231 -0.0233(0.076) (0.026) (0.027)Country FE Y Y Y Y Y Y Y Y Y Y Y YYear FE Y Y Y Y Y Y Y Y Y Y Y YN 10,765 10,765 10,765 15,832 15,832 15,832 16,077 16,077 16,077 10,648 10,648 10,648R2 0.532 0.210 0.210 0.488 0.146 0.145 0.494 0.146 0.145 0.533 0.210 0.209Notes: Data is at the group-year level. Estimation for the sample of non-leaders. Included is a dummy variable assuming value 1 if a non-leaderis included as junior partner in the coalition (0 if group is powerless, discriminated or self-exclusion). Conflict incidence is coded as 1 for everyyear where there is ongoing conflict from a rebel organization claiming to represent the ethnic group. Groupsize is measured as the fractionof country’s total population. Robust standard errors clustered at the country level in parentheses. Drought (SPI) gives the proportion ofmonths in the growing season that are part of the longest streak of consecutive months in that growing season with SPI1 values below -1.5.The growing season is the growing season for the cell’s main crop, defined in the MIRCA2000 dataset v.1.1. SPI1 index measures deviationfrom long-term normal rainfall for that month. The values are standardized where deviation estimates less than 1 standard deviation indicatenear normal rainfall. Drought (SPEI CRU) uses the Standardized Precipitation and Evapotranspiration Index SPEI-1 from the SPEIbasev.2.3. SPEIbase is based on precipitation and potential evapotranspiration from the Climatic Research Unit of University of East AngliaCRU v.3.22. Drought (SPEI GDM) uses the Standardized Precipitation and Evapotranspiration Index SPEI1 from the SPEI Global DroughtMonitor. SPEI GDM uses the GPCC ’first guess’ product and GHCN/CAMS, while using the Thornthwaite potential evapotranspiration(PET) estimation. Precipitation gives the yearly total amount of precipitation (in millimeter) in the cell, based on monthly meteorologicalstatistics from the Global Precipitation Climatology Centre. All values are weighted (by area) mean of all grid values in the ethnic homaland.Values at grid value are from the PRIO-GRID dataset. Robust standard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.1109Table A6: Probability of Inclusion and Conflict, by group size (control-ling for geography)(1) (2) (3) (4) (5) (6)Included Conflict Conflict Included Conflict Conflictgroupsize 2.3964*** -0.0624 2.5374*** -0.1019(0.500) (0.103) (0.502) (0.119)groupsize2 -3.2275*** 0.2799 0.2013*** -3.5737*** 0.3461 0.2056***(0.642) (0.174) (0.066) (0.738) (0.211) (0.078)Elevation (std) -0.0001 0.0000 0.0000(0.000) (0.000) (0.000)Mountains (proportion) -0.1493 -0.0038 -0.0026(0.169) (0.016) (0.015)Country FE Y Y Y Y Y YYear FE Y Y Y Y Y YN 17,922 17,922 17,922 17,452 17,452 17,452R2 0.491 0.149 0.148 0.494 0.140 0.139Notes: Data is at the group-year level. Estimation for the sample of non-leaders. Included is adummy variable assuming value 1 if a non-leader is included as junior partner in the coalition (0if group is powerless, discriminated or self-exclusion). Conflict incidence is coded as 1 for everyyear where there is ongoing conflict from a rebel organization claiming to represent the ethnicgroup. Groupsize is measured as the fraction of country’s total population. Elevation (std) isthe standard deviation of gridded elevation measurements (0.008330 decimal degree resolution)intersecting with group polygon. Mountains (proportion) is group-level area-weighted mean ofthe gridded proportion of mountainous terrain within the cell based on elevation, slope and localelevation range, taken from a high-resolution mountain raster developed for UNEP’s MountainWatch Report. Robust standard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.1110Table A7: Probability of Inclusion and Conflict, by group size (controllingfor intra-group fractionalization)(1) (2) (3) (4) (5) (6)Included Conflict Conflict Included Conflict Conflictgroupsize 2.2584*** -0.0568 2.2648*** -0.0585(0.426) (0.092) (0.425) (0.094)groupsize2 -3.0400*** 0.2577 0.1848*** -3.0197*** 0.2540 0.1789***(0.549) (0.163) (0.066) (0.544) (0.165) (0.066)Fractionalization (religion) -0.0043 -0.0010 -0.0018(0.054) (0.008) (0.008)Fractionalization (language) -0.1099** 0.0213 0.0212(0.051) (0.015) (0.015)Country FE Y Y Y Y Y YYear FE Y Y Y Y Y YN 19,788 19,788 19,788 19,842 19,842 19,842R2 0.468 0.145 0.145 0.471 0.146 0.146Notes: Data is at the group-year level. Estimation for the sample of non-leaders. Included is adummy variable assuming value 1 if a non-leader is included as junior partner in the coalition (0if group is powerless, discriminated or self-exclusion). Conflict incidence is coded as 1 for everyyear where there is ongoing conflict from a rebel organization claiming to represent the ethnicgroup. Groupsize is measured as the fraction of country’s total population. EPR-2018 providesinformation on the 3 largest religions within group and the fraction of group associated with each.Fractionalization is measured as 1−religion21−religion22−religion23− (1−∑3ireligioni)2. Thesame is done for language. Robust standard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.1111Table A8: Probability of Inclusion and Conflict, by group size (controlling for light density)(1) (2) (3) (4) (5) (6) (7) (8) (9)Included Conflict Conflict Included Conflict Conflict Included Conflict Conflictgroupsize 2.9161*** 0.1213 2.7571*** 0.1250 2.8984*** 0.1291(0.686) (0.207) (0.661) (0.209) (0.680) (0.208)groupsize2 -4.0236*** -0.0725 0.1289 -3.7547*** -0.0785 0.1288 -4.0116*** -0.1009 0.1124(1.221) (0.277) (0.105) (1.202) (0.279) (0.105) (1.192) (0.268) (0.099)Night light 0.0244 -0.0004 -0.0003(0.018) (0.001) (0.001)Log (1+Night light) 0.1393 -0.0032 -0.0019(0.085) (0.004) (0.004)Night light 1.4933 -0.0651 -0.0668(calibrated) (1.157) (0.048) (0.044)Country FE Y Y Y Y Y Y Y Y YYear FE Y Y Y Y Y Y Y Y YN 6,875 6,875 6,875 6,875 6,875 6,875 6,553 6,553 6,553R2 0.565 0.245 0.244 0.569 0.245 0.244 0.563 0.254 0.254Notes: Data is at the group-year level. Estimation for the sample of non-leaders. Included is a dummy variableassuming value 1 if a non-leader is included as junior partner in the coalition (0 if group is powerless, discriminated orself-exclusion). Conflict incidence is coded as 1 for every year where there is ongoing conflict from a rebel organizationclaiming to represent the ethnic group. Groupsize is measured as the fraction of country’s total population. Night lightmeasures average nighttime light emission from the DMSP-OLS Nighttime Lights Time Series Version 4. The calibratedvariable accounts for intersatellite differences and interannual sensor decay using calibration values from Elvidge et al(2014). Robust standard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.1112Table A9: Probability of Inclusion and Conflict, by groupsize (Cold War vs Post-Cold War)1946-1991 1992-2017(1) (2) (3) (4) (5) (6)Included Conflict Conflict Included Conflict Conflictgroupsize 1.8432*** -0.1104 2.7810*** 0.0790(0.503) (0.104) (0.525) (0.202)groupsize2 -2.2500*** 0.3072 0.1613* -3.8741*** 0.1610 0.2882*(0.666) (0.220) (0.087) (0.872) (0.405) (0.165)Country FE Y Y Y Y Y YYear FE Y Y Y Y Y YN 11,005 11,005 11,005 9,592 9,592 9,592R2 0.515 0.114 0.113 0.548 0.236 0.236Notes: Data is at the group-year level. Estimation for the sample of non-leaders. Included is a dummy variable assuming value 1 if a non-leader isincluded as junior partner in the coalition (0 if group is powerless, discriminatedor self-exclusion). Conflict incidence is coded as 1 for every year where thereis ongoing conflict from a rebel organization claiming to represent the ethnicgroup. Groupsize is measured as the fraction of country’s total population.Robust standard errors clustered at the country level in parentheses.*** p<0.01, ** p<0.05, * p<0.1113Appendix BAdditional Figures114Figure A1: Ethnic groups in Conflict over the control of the govern-ment (1945-2017)Notes: The figure highlights the ethnic homeland of groups supporting or been represented by rebelorganizations in conflict against the government during 1946-2017. Only conflict over the controlor composition of the government is highlighted (it does not include secessionist movements).Different shades reflect the frequency of conflict episodes.115Figure A2: EPR-2018: KenyaNotes: The figure shows the geographical distribution of politically relevant ethnic groups inKenya. The highlighted areas show where groups are concentrated. Blank areas may be eitheruninhabited regions or areas inhabited by several groups with no concentration of a particularethnicity.116Figure A3: EPR: Politically relevant groups in Africa and Asia (2017)Notes: The figure shows the geographical distribution of politically relevant ethnic groups in Africaand Asia. Political status reflects the degree of representation in the executive power, as definedby EPR-2018.117Figure A4: EPR: Politically relevant groups in Africa and Asia (2017)Notes: The figure shows the geographical distribution of politically relevant ethnic groups in Africaand Asia. Political status is defined as used in the estimations. Leaders are the most powerfulgroup in the country (dominant, monopoly, or senior partner). Included groups are junior partners.Excluded groups are those categorized as Powerless, Discriminated or Self-Exclusion.118Figure A5: Probability of Inclusion and Conflict, by group size0. .2 .4 .6 .8 1bins lpoly smoothProbability of Inclusion0. .2 .4 .6 .8 1bins lpoly smoothProbability of ConflictNotes: Binned scatterplot and Local polynomial regression of inclusion (left) and conflict incidence(right) on group size for non-leader groups. Inclusion assumes value 1 if group is included as juniorpartner, or 0 if group is excluded (powerless, discriminated or self-exclusion). Conflict incidenceis dummy variable indicating if the ethnic group is conflict over the control of the government inthat year. Group size is measured as the fraction of country’s total population.119Figure A6: Political status distribution, by group size0204060801000 .2 .4 .6 .8 1groupsizeExcluded Included (Junior Partner)LeaderNotes: Distribution of ethnic groups according to their political status (leader, included andexcluded) by group size. A group is leader if political status is monopoly, dominant or seniorpartner. Included are junior partners. Excluded are groups categorized as powerless, discriminatedor self-exclusion. Group size is measured as the fraction of country’s total population.120Figure A7: Transfers to included groups, by group size01020304050govshare0 .1 .2 .3 .4groupsizeNotes: This figure plots the relationship between group size and share of cabinet appointments(govshare) for included groups only (junior partners). Data is constructed by combining EPR-2018 with ethnic affiliation of cabinet members from Rainer and Trebbi [2012] and Francois et al.[2015] for available countries.121Appendix CDynamic Model: SolutionThis Appendix presents the conditions for each possible equilibriumstrategy of the dynamic model studied in Section 2.3. We are looking for astationary Markov Perfect Equilibrium. In what follows, I absorb subscriptt, since value function will be the same every period.Equilibrium with Exclusion and Peace of groups A and BI aim to find the conditions for an equilibrium where:• group A excludes group B (τB = 0) when leader, and stay in peace forany offer when challenger (CA(τA) = 0 ∨ τA ≥ 0);• group B excludes group A (τA = 0) when leader, and stay in peace forany offer when challenger (CB(τB) = 0 ∨ τB ≥ 0)In this equilibrium, V lA = V lB = pi1−δ . For group A, the value of peace ishigher than the value of conflict if:V cpA =piτA1− δ ≥ VccA = (ατA + θA)(δV lA + pi)− cc (C.1)Since we are looking for a stationary equilibrium, the first equality aboveassumes that an optimal strategy in period t will also be a best response forall periods. A similar condition must be satisfied for group B. We can then122derive that groups A and B play peace for any offer if:θA ≤ (1− δ)cc/piθB ≤ (1− δ)cc/pi (C.2)Under these conditions, the leader’s optimal strategy is to offer zero. Heobtains peace, and the value of being leader is consistent with the one usedwhen computing the conflict payoffs.Formally, if θA and θB satisfy conditions (C.2), the equilibrium strategiesare:i) Group A: τ∗B = 0 and C∗A(τA) = 0 ∨ τA;ii) Group B: τ∗A = 0 and C∗B(τB) = 0 ∨ τB.Equilibrium with Exclusion and Peace of group B and Inclusionand Peace of group AI aim to find the conditions for an equilibrium where:• group A excludes group B (τB = 0) when leader, and group B playspeace;• group B includes group A (τA > 0) when leader, and group A acceptsthe offer.In this scenario, V lA = pi1−δ . The value of conflict for group A is:V ccA = (ατA + θA)(δV lA + pi)− ccIf peace is the optimal strategy, it will be the optimal strategy everyperiod. Therefore, the value of peace for group A would be:V cpA =τApi1− δPeace is played if the value of peace is higher than the value of conflict,which is true when:τA ≥ θA − (1− δ)cc/pi(1− α) = τminA123Now we must find the conditions where group B’s best response is toinclude group A. First, we must have τminA ≥ 0. This is obtained when:θA ≥ (1− δ)cc/pi (C.3)Second, the value of appeasing group A must be higher than excludingthem and facing conflict. The value of appeasing challenger A for the leaderB is given by:V lB(τminA ) =(1− θA−(1−δ)cc/pi)1−α pi1− δMeanwhile, the playoff of exclusion (and, consequently, conflict) is:V lB(τA = 0) = (1−θA)(pi+δV lB(τA = 0))−cl ⇒ V lB(τA = 0) =(1− θA)pi − cl1− δ(1− θA)(C.4)Finally,V lB(τminA ) ≥ V lB(τA = 0)⇒θA ≤−α− ccδ2 + ccδ +√(α+ δ(cc(δ − 1)− 1))2 + 4(δ − 1)δ(cc(δ − 1) + (α− 1)cl) + δ2δ(C.5)It remains to find the conditions where group B is excluded and stillplays peace, which is consistent with the assumed value of the leader A. Forthat, the value of conflict must be lower than the value of peace when theoffer is zero:V ccB = θB(pi + δV lB)− cc = θB(pi + δ(1− θA−(1−δ)cc/pi)1−α pi)1− δ )− cc ≤ VcpB = 0⇒θB ≤ (1− α)(1− δ)cc/pi(1− α− δθA) + (1− δ)δcc/pi(C.6)It is worthy to compare this conditions with the ones from the previousequilibrium, where both players exclude the challenging group and obtainpeace. Now the strength of group A is higher which requires group B to124react and offer some power-sharing. In turn, the stronger group A is, thelower the value of being leader is for group B. And, consequently, the higherthe maximum θB that supports the peaceful exclusion of group B.Therefore, if θA and θB satisfy conditions (C.3), (C.5) and (C.6) theequilibrium strategies are:i) Group A: τ∗B = 0 and C∗A(τA) =0 if τA ≥θA−(1−δ)cc/pi(1−α)1 if τA < θA−(1−δ)cc/pi(1−α);ii) Group B: τ∗A =θA−(1−δ)cc/pi(1−α) and C∗B(τB) = 0 ∨ τB.Equilibrium with Exclusion and Peace of group B and Exclusionand Conflict of group AI aim to find the conditions for an equilibrium where:• group A excludes group B (τ∗B = 0) when leader, and group B playspeace;• group B excludes group A (τ∗A = 0) when leader, and group A playsconflict.This scenario is similar to the one just analyzed, except for a couple ofdifferences. First, now group B must find optimal to exclude group A. Thismeans that condition (C.5) is reverted:θA >−α− ccδ2 + ccδ +√(α+ δ(cc(δ − 1)− 1))2 + 4(δ − 1)δ(cc(δ − 1) + (α− 1)cl) + δ2δ(C.7)Second, since it is optimal to exclude group A and face conflict, the valueof being leader for group B is now given by equation (C.4). Group B willplay peace for any offer if:V ccB = θB(pi + δV lB)− cc = θB(pi + δ(1− θA)pi − cl1− δ(1− θA) )− cc ≤ VcpB = 0⇒θB ≤ (1− δ + δθA)cc/pi1− δcl/pi(C.8)125When θA equals to the right-hand side of the expressions in (C.5) and(C.7), then the inequalities (C.6) and (C.8) are equivalent. As soon as θAcrosses that threshold, the optimal strategy for leader B is to face war againstgroup A, instead of sharing power. This makes the value of being leaderdepend on θA in a different way. The reason is that θA now determinesthe outcomes of conflict and not the offer in the subgame played on theequilibrium path. That’s why the functions (C.6) and (C.8) are different. Inturn, group B can have higher strength and still play peace under exclusion.Essentially, the payoff of conflict is low because in case of victory the groupwill face another conflict against a strong challenger.Finally, if θA and θB satisfy conditions (C.7) and (C.8) the equilibriumstrategies are:i) Group A: τ∗B = 0 and C∗A(τA) =0 if τA ≥θA−(1−δ)cc/pi(1−α)1 if τA < θA−(1−δ)cc/pi(1−α);ii) Group B: τ∗A = 0 and C∗B(τB) = 0 ∨ τB.The conditions where there is exclusion and conflict of group B, andexclusion and peace of group A on the equilibrium path are just symmetricto this one.Equilibrium with Inclusion and Peace of group B and Inclusionand Peace of group AI aim to find the conditions for an equilibrium where:• group A includes group B (τ∗B > 0) when leader, and group B playspeace;• group B includes group A (τ∗A > 0) when leader, and group A playspeace.In this scenario, the values of being leader for groups A and B are givenby:V lB =(1− τ∗A)pi1− δV lA =(1− τ∗B)pi1− δ126where τ∗A and τ∗B are the optimal sharing offered by the leaders to theirrespective challengers.Given the values of being leader, we must find the minimum offers thatappease each group. The value of conflict for group A is given by:V ccA (τA) = (ατA + θA)(pi + δV lA)− cc = (ατA + θA)(pi + δ(1− τ∗B)pi1− δ )− ccGroup A prefers to play peace if:V cpA (τA) =τApi1− δ ≥ VccA (τA) = (ατA + θA)(pi + δ(1− τ∗B)pi1− δ )− cc ⇒τA ≥ θA(1− δτ∗B)− (1− δ)cc/pi1− α(1− δτ∗B)(C.9)Similarly, group B plays peace if:τB ≥ θB(1− δτ∗A)− (1− δ)cc/pi1− α(1− δτ∗A)(C.10)In a equilibrium where the leader makes an appeasing offer, conditions(C.9) and (C.10) are satisfied with equality. This is because the leaderdoes not have any reason to offer more than the minimum share necessaryto appease the challenger. Solving the system of equations, we obtain theoptimal transfers:τ∗A =−α2 + αδθA − αδθB + 2α+ δ2θAθB − 12αδ(1− α− δθB) +√F +G2αδ(1− α− δθB)andτ∗B =−α2 − αδθA + αδθB + 2α+ δ2θAθB − 12αδ(1− α− δθA) +√F +G2αδ(1− α− δθA) ,where F =(α2 + α(δ(θB − θA)− 2)− δ2θAθB + 1)2and G = 4αδ(α+ δθB − 1)(θA(α+ δθB − 1) + (δ − 1)(α+ δθA − 1)cc/pi)(C.11)127Now we must find conditions where the leader chooses to offer a positivetransfer to the challenger. First, this requires the right-hand side of theexpressions (C.9) and (C.10) to be positive. That is, peace can only beobtained with some power-sharing. This is satisfied if:θA ≥ (1− δ)(1− α)cc/pi1− α+ δ((1− δ)cc/pi − θB)θB ≥ (1− δ)(1− α)cc/pi1− α+ δ((1− δ)cc/pi − θA)(C.12)Second, the leader must prefer to appease the challenger instead of ex-cluding the opposing group and facing conflict. A leader A who excludes arebel group B gets:V lA(τB = 0) = (1−θB)(pi+δV lA(τB = 0))−cl ⇒ V lA(τB = 0) =(1− θB)pi − cl1− δ(1− θB)Then, the payoff of inclusion is higher than exclusion if:V lA(τB = τ∗B) ≥ V lA(τB = 0)⇒(1− τ∗B(θB, θA))pi1− δ ≥(1− θB)pi − cl1− δ(1− θB) (C.13)If θB is below the minimum required by expression (C.12), and conse-quently satisfies inequality (C.6), a leader A could obtain peace by offeringnothing to challenger B. As θB marginally increases above the minimum re-quired by the inequality (C.12), the value of inclusion is clearly greater thanthe value of exclusion. The reason is that the minimum τB that achievespeace is just marginally greater than zero, and the leader would then losejust a marginal share of the total surplus. On the other hand, in case ofexclusion, the leader’s payoff drop discontinuously because of the costs ofwar.Solving the inequality (C.13) for θB is difficult since the optimal sharingis a complicated function of the parameters of the model (Equation (C.11)).However, it is possible to inspect this inequality for the whole range of theparameter space. This is possible since all parameters of this model must128fall in a very specific range for them to make any sense: δ < 1, cc/pi, cl/pi <1, θA, θB < 1−α, α < 1. The inspection of the inequality (C.13) shows thateither: i) the value of a peaceful inclusion is always greater than the valueof exclusion; or ii) there is θ∗∗B , such that V lA(τB = τ∗B) ≥ V lA(τB = 0) ⇐⇒θB ≤ θ∗∗B , conditional on (C.12).By analogy, the same result is replicated for the values of θA. Therefore,conditional on (C.12), θ∗∗A and θ∗∗B (if they exist) satisfy the following:(1− τ∗B(θ∗∗B , θA))pi1− δ =(1− θ∗∗B )pi − cl1− δ(1− θ∗∗B )and(1− τ∗A(θ∗∗A , θB))pi1− δ =(1− θ∗∗A )pi − cl1− δ(1− θ∗∗A )(C.14)In conclusion, if θA and θB are such that (C.12) is satisfied, θA ≤θ∗∗A (ifit exists), and θB ≤θ∗∗B (if it exists), the equilibrium strategies are:i) Group A: τ∗B and C∗A(τA) =0 if τA ≥ τ∗A1 if τA < τ∗Aii) Group B: τ∗A and C∗B(τB) =0 if τB ≥ τ∗B1 if τB < τ∗B,where τ∗A and τ∗A are given by (C.11).Equilibrium with Inclusion and Peace of group B and Exclusionand Conflict of group AI aim to find the conditions for an equilibrium where:• group A includes group B (τ∗B > 0) when leader, and group B playspeace;• group B excludes group A (τ∗A = 0) when leader, and group A playsconflict.129In this scenario, the values of being leader are:V lB = (1− θA)(pi + δV lB)− cl ⇒ V lB =(1− θA)pi − cl1− δ(1− θA)V lA =(1− τ∗B)pi1− δ ,where τ∗B is the optimal sharing offered by leader A to challenger B.Group A plays peace if:V cpA (τA) =τApi1− δ ≥ VccA (τA) = (ατA + θA)(pi + δ(1− τ∗B)pi1− δ )− cc ⇒τA ≥ θA(1− δτ∗B)− (1− δ)cc/pi1− α(1− δτ∗B)= τminA(C.15)Group B plays peace if:V cpB (τB) =τBpi1− δ ≥ VccB (τB) = (ατB + θB)(pi + δ(1− θA)pi − cl1− δ(1− θA) )− cc ⇒τB ≥ (1− δ)((δ(1− θA)− 1)cc/pi + θB − δθBcl/pi)1 + αδ − α+ δθA − δ + α(δ − 1)δcl/pi = τ∗B(C.16)First, we must find conditions where groups would be in conflict if theywere offered nothing. For group B, this is:V ccB (τB = 0) = θB(pi + δ(1− θA)pi − cl1− δ(1− θA) )− cc ≥ 0⇒ θB ≥(1− δ + δθA)cc/pi1− δcl/pi(C.17)andV ccA (τA = 0) = θA(pi + δ(1− τ∗B)pi1− δ )− cc ≥ 0⇒θA ≥(1− δ)2δ(1− (1− δ)cc/piδ)[α + δ2cc/pi + δθB − αδcl/pi − δ2θBcl/pi − 1+√(α + δ2cc/pi + δθB − αδcl/pi − δ2θBcl/pi − 1)2 − 4δcc/pi((δ − 1)δcc/pi − 1)(α(δcl/pi − 1) + 1)] (C.18)130Second, group A must prefer the inclusion of group B than facing conflictagainst them:V lA(τB = τ∗B) ≥ V lA(τB = 0)⇒(1− τ∗B(θB, θA))pi1− δ ≥(1− θB)pi − cl1− δ(1− θB) ⇒θB ≤ 12(1− δ)δ(1− δclpi)[−α+ δ + cc/piδ + δcl/pi + αδ + αδcl/pi − δ2 − 2cc/piδ2 − 2δ2cl/pi−αδ2cl/pi + cc/piδ3 + δ3cl/pi + δθA + cc/piδ2θA − cc/piδ3θA +√C +D],where C = 4δ(δ − 1)2(δcl/pi − 1)(cc/pi(δ − 1)(δ(θA − 1) + 1) + cl/pi(δ(−θA) + δ+α(δ − 1)(δcl/pi − 1)− 1))and D = (δ(cc/pi(δ − 1)(δ(θA − 1) + 1) + δ − θA + (δ − 1)2(−cl/pi)− 1)+α(δ − 1)(δcl/pi − 1))2(C.19)On the other hand, in this equilibrium leader B prefers the exclusion ofA than its inclusion:V lB(τA = τminA ) < V lB(τA = 0)⇒(1− τminA (θA, θB))pi1− δ <(1− θA)pi − cl1− δ(1− θA)(C.20)Again, it is difficult to find a closed-form solution for θA. I use the sameapproach taken in the previous section. Similarly, it is possible to show thateither: i) there is no θA that satisfies the inequality; or ii) there is θ∗∗A suchthat the inequality is satisfied for any θA ≥ θ∗∗A .In sum, if θA and θB are in the region given by (C.17), (C.17), (C.20)and (C.19), the equilibrium strategies are:i) Group A: τ∗B and C∗A(τA) =0 if τA ≥ τminA1 if τA < τminAii) Group B: τ∗A = 0 and C∗B(τB) =0 if τB ≥ τ∗B1 if τB < τ∗B,where τminA and τ∗B are given by (C.15) and (C.16).131The conditions where there is inclusion and peace of group A, and ex-clusion and conflict of group B on the equilibrium path are symmetric tothis one.Equilibrium with Exclusion and Conflict of group B andExclusion and Conflict of group AI aim to find the conditions for an equilibrium where:• group A excludes group B (τ∗B = 0) when leader, and group B playsconflict;• group B excludes group A (τ∗A = 0) when leader, and group A playsconflict.In this scenario, the values of being leader are:V lA = (1− θB)(pi + δV lA)− cl ⇒ V lA =(1− θB)pi − cl1− δ(1− θB)V lB = (1− θA)(pi + δV lB)− cl ⇒ V lB =(1− θA)pi − cl1− δ(1− θA)The minimum shares that appease the respective challenging groups aregiven by:V cpA (τA) =τApi1− δ ≥ VccA (τA) = (ατA + θA)(pi + δ(1− θB)pi − cl1− δ(1− θB) )− cc ⇒τA ≥ (1− δ)((δ(1− θB)− 1)cc/pi + θA − δθAcl/pi)1 + αδ − α+ δθB − δ + α(δ − 1)δcl/pi = τminAandV cpB (τB) =τBpi1− δ ≥ VccB (τB) = (ατB + θB)(pi + δ(1− θA)pi − cl1− δ(1− θA) )− cc ⇒τB ≥ (1− δ)((δ(1− θA)− 1)cc/pi + θB − δθBcl/pi)1 + αδ − α+ δθA − δ + α(δ − 1)δcl/pi = τminB(C.21)Now we must find the conditions for which exclusion is better than anappeasing inclusion. This is satisfied for group A if:132V lA(τB = τminB ) < V lA(τB = 0)⇒(1− τminB (θB, θA))pi1− δ <(1− θB)pi − cl1− δ(1− θB) ⇒θB >12(1− δ)δ(1− δclpi)[−α+ δ + cc/piδ + δcl/pi + αδ + αδcl/pi − δ2 − 2cc/piδ2 − 2δ2cl/pi−αδ2cl/pi + cc/piδ3 + δ3cl/pi + δθA + cc/piδ2θA − cc/piδ3θA +√C +D],where C = 4δ(δ − 1)2(δcl/pi − 1)(cc/pi(δ − 1)(δ(θA − 1) + 1) + cl/pi(δ(−θA) + δ+α(δ − 1)(δcl/pi − 1)− 1))and D = (δ(cc/pi(δ − 1)(δ(θA − 1) + 1) + δ − θA + (δ − 1)2(−cl/pi)− 1)+α(δ − 1)(δcl/pi − 1))2(C.22)Similarly for group B:θA >12(1− δ)δ(1− δclpi)[−α+ δ + cc/piδ + δcl/pi + αδ + αδcl/pi − δ2 − 2cc/piδ2 − 2δ2cl/pi−αδ2cl/pi + cc/piδ3 + δ3cl/pi + δθB + cc/piδ2θB − cc/piδ3θB +√C˜ + D˜],where C˜ = 4δ(δ − 1)2(δcl/pi − 1)(cc/pi(δ − 1)(δ(θB − 1) + 1) + cl/pi(δ(−θB) + δ+α(δ − 1)(δcl/pi − 1)− 1))and D˜ = (δ(cc/pi(δ − 1)(δ(θB − 1) + 1) + δ − θB + (δ − 1)2(−cl/pi)− 1)+α(δ − 1)(δcl/pi − 1))2(C.23)In sum, if θA and θB are in the region given by (C.23) and (C.22), theequilibrium strategies are:i) Group A: τ∗B = 0 and C∗A(τA) =0 if τA ≥ τminA1 if τA < τminAii) Group B: τ∗A = 0 and C∗B(τB) =0 if τB ≥ τminB1 if τB < τminB ,where τminA and τminB are given by (C.21).It can be verified that all conditions for each equilibrium are mutuallyexclusive. In fact, when values of θA and θB reach the limit of the inequalityconditions of one equilibrium, they will necessarily satisfy the conditions of133the next equilibrium. For instance, suppose the parameters are such thatthere is neither sharing nor conflict on the equilibrium path. When thestrength of a group reaches the upper limit that supports such equilibrium,this upper limit will coincide with the lower limit for the group to be includedand in peace.Taking all the above results together, a result very similar to the staticmodel can be obtained. This is summarized by Proposition 2.134


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items