University of Southern California Dissertations and Theses

Perspectives on state capacity and the political geography of conflict

(USC Thesis Other)

Perspectives on state capacity and the political geography of conflict

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Perspectives on state capacity and the political geography of con ict by Therese Anders A Dissertation Presented to the Faculty of the Graduate School University of Southern California In Partial Fulllment of the Requirements for the Degree Doctor of Philosophy (Political Science and International Relations) May 2020 Dedication F ur meine Eltern, deren Liebe, Vertrauen, und R uckhalt nicht in Worte zu fassen sind und Kontinente uberspannen ii Acknowledgements First and foremost, I extend my gratitude to my dissertation committee: Patrick James, Brian Rathbun, Jerey Nugent, and Christopher Fariss. I am especially thankful for the freedom and trust my committee granted me. Their tireless guidance and support allowed me to shift the focus of my dissertation to pursue a new and exciting research agenda surrounding the measurement of territorial control | research that now forms the center of my scholarly work. I thank Pat for helping me navigate life as an academic, and for always encouraging me to pursue opportunities for new research projects and collaborations. I am grateful to Brian for his advise and for extend- ing opportunities to co-author and learn the ropes of academic publishing. I would like to express my deep appreciation to Je for his thoughtful and eye-opening commentary throughout various stages of this dissertation. I owe special thanks to Chris, whose mentorship has been instrumental for my academic development. He is a role model and I am forever grateful for his generosity with time, kindness, feedback, reassurance, introductions to other scholars, and wisdom. Working with Jonathan Markowitz, Benjamin Graham, Megan Becker, and Chris has been a dening chapter on my path towards the Ph.D. I thank them for supporting and trusting me early on, and for giving me exceptional opportunities to grow as a researcher and teacher. Their collegiality and friendship are exemplary. Special thanks go to Pablo Barber a for being extraordi- narily generous with his time and guidance for all things research and career related. I would also like to express my appreciation to James Lo and Andrew Coe for the insightful feedback they have provided to me and other graduate students throughout the years. They all demonstrate what it means to be a critical thinker and supportive colleague, and I hope to emulate their example and pay it forward to future colleagues and students. I would also like to express my deep gratitude to Veri Chavarin whose patience, kindness, knowledge, and support have been invaluable. There are many brilliant colleagues and friends that I wish to thank for the advise, laugh- ter, and conversations we have shared throughout the years. Rod Albuyeh, Miriam Barnum, Juve Cortes, Taylor Dalton, Stephanie Kang, Xinru Ma, Suzie Mulesky, Quynh Nguyen, Hai-Vu Phan, Sara Sadhwani, G uez Salinas, Jihyun Shin, Eric Stollenwerk, Gregor Walter-Drop, and Shiming Yang all made this quest special. I am particularly grateful to Joey Huddleston and Tom Jamieson for their friendship and kindness | I truly missed them in the last two years of graduate school and will forever look forward to spending time with them at conferences and beyond. I want to give special recognition to Vic Chonn Ching, Whitney Hua, Ste Neumeier, iii Jenn Rogl a, and Anne van Wijk for being amazing friends and travel companions, adventurers, erce women, and incredibly smart human beings. I am grateful for my talented and inspiring friends Christin, Janina, Sarah R. Kelly, Joe, and Sarah M. | they were all a source of happiness and support. The Hanabaluza gang has been the absolute best and I am beyond grateful for the glorious memories we have created throughout the years. They are the reason a part of my heart will forever be in Berlin. I am unbelievably grateful to call Louise and Sven both family and close friends. Words cannot describe how much I appreciate them giving me a home while living thousands of kilometers away. My sister's love and presence meant the world to me on this journey. I am thankful to my grandmas for their unconditional love and support, and especially their patience when I did not call as often as I should have. I would also like to pay tribute to my grandpas | I am sad that I do not get to share the joy over this achievement with them, because their education, devotion, and encouragement have shaped me into the person and scholar I am today. I would like to express my deep gratitude and love to Justin. He has been an incredible source of care, comfort, and love. Justin encouraged me when I was frustrated, held me when I was sad, stayed up with me when I worked through the night, and celebrated every small and large feat with me. You have been my rock. I am grateful to Chuck, Michelle, Joelle, James, Sam, Duc, Julia, Helen, Red, Milky, and everyone for welcoming me into their family with open arms and open hearts, for their friendship, and for giving me a home away from home. Their reassurance has lifted me up and carried me through the highs and the lows. Last but not least, I would like to thank my parents. I could not have done any of this without their tremendous support and encouragement throughout the years. They have been cheerleaders, critics, carers, counselors, and condantes, and tirelessly helped me to navigate life decisions small and large. I cannot thank them enough. Their trust and love are inspiring, and I dedicate this work to them. iv Table of Contents Dedication ii Acknowledgements iii List Of Tables viii List Of Figures x Abbreviations xvi Chapter 1: Introduction: Measuring state capacity in con ict 1 1.1 Measurement innovations for con ict processes . . . . . . . . . . . . . . . . . . . . 6 1.2 From `guns versus butter' to `hearts and minds' . . . . . . . . . . . . . . . . . . . . 10 1.3 Actor behavior in asymmetric con icts . . . . . . . . . . . . . . . . . . . . . . . . . 13 Chapter 2: Territorial control in civil wars: Theory and measurement using ma- chine learning 17 2.1 Territorial control in civil war . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.2 Tactical choice in asymmetric civil war . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.3 Modeling territorial control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.3.1 Measuring rebel tactics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.3.2 Mapping rebel tactics onto territorial control . . . . . . . . . . . . . . . . . 29 2.3.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.3.3.1 Transition probabilities . . . . . . . . . . . . . . . . . . . . . . . . 32 2.3.3.2 Emission probabilities . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.3.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.3.5 Case selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.4 Hidden Markov Model (HMM) estimates of territorial control . . . . . . . . . . . . 36 2.4.1 Colombia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.4.2 Nigeria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.6 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 2.A Measuring con ict exposure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 2.A.1 Spatial and temporal decay functions . . . . . . . . . . . . . . . . . . . . . 50 2.A.2 Comparison between discrete and continuous aggregation . . . . . . . . . . 54 2.B Estimation procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 2.C Summary statistics for deforestation model . . . . . . . . . . . . . . . . . . . . . . 58 2.D Additional gures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 2.D.1 Transition probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 2.D.2 Case selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 v 2.D.3 ACLED validation data for Northeast Nigeria . . . . . . . . . . . . . . . . . 62 Chapter 3: How does Insecurity Aect Government Welfare Spending? Insights from Colombian Municipalities 63 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 3.2 Insecurity and Non-Coercive Government Action . . . . . . . . . . . . . . . . . . . 67 3.3 Three Logics Connecting Violence and Welfare Investments . . . . . . . . . . . . . 71 3.3.1 The intensity of violence perspective . . . . . . . . . . . . . . . . . . . . . . 71 3.3.2 State-centric strategic perspective . . . . . . . . . . . . . . . . . . . . . . . 72 3.3.3 Citizen-centric strategic perspective . . . . . . . . . . . . . . . . . . . . . . 73 3.3.4 Scope condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 3.4 Intrastate Con ict in Colombia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 3.5 Municipal Expenditure in Colombia . . . . . . . . . . . . . . . . . . . . . . . . . . 79 3.6 Data and estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 3.6.1 Measuring municipal investment in social welfare . . . . . . . . . . . . . . . 81 3.6.2 Measuring insecurity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 3.6.3 Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 3.6.3.1 Base model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 3.6.3.2 Alternative explanations . . . . . . . . . . . . . . . . . . . . . . . 85 3.6.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 3.7 Empirical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 3.7.1 Main regression results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 3.7.2 Alternative explanations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 3.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 3.9 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 3.A Summary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 3.B Regression Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 3.C Auxiliary results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Chapter 4: Bread before guns or butter: Introducing Surplus Domestic Product (SDP) 113 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 4.2 Subsistence and Surplus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 4.2.1 Converging and diverging trends in the international system: Gross Domes- tic Product (GDP) vs. Surplus Domestic Product (SDP) . . . . . . . . . . 121 4.3 Measuring Relative Power-Resources and Potential Threat . . . . . . . . . . . . . . 126 4.3.1 Measuring the dierence in relative power-resources between states using SDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.3.2 Relative power-resource relationships between one state and all other states 128 4.4 Potential Threat using SDP vs. GDP . . . . . . . . . . . . . . . . . . . . . . . . . 130 4.4.1 Assessing the most potentially threatening states using SDP vs. GDP . . . 130 4.4.2 Modeling arming and power projection . . . . . . . . . . . . . . . . . . . . . 134 4.5 Evaluating Military Burdens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 4.7 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 4.A Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 4.B GDP = surplus + subsistence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 4.C Rankorder Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 4.D Coverage of Composite Index of National Capabilities (CINC) variables . . . . . . 152 4.E Comparing SDP with GDP and GDP per capita . . . . . . . . . . . . . . . . . . . 156 4.F Measuring potential threat in the strategic environment . . . . . . . . . . . . . . . 158 vi 4.F.1 Geographic proximity and the loss of strength gradient . . . . . . . . . . . . 165 4.F.2 Construction of the potential threat measure . . . . . . . . . . . . . . . . . 166 4.F.3 Correlation between alternative potential threat measures . . . . . . . . . . 167 4.F.4 Global trends of potential threat . . . . . . . . . . . . . . . . . . . . . . . . 168 4.F.5 Spatiotemporal variation of potential threat . . . . . . . . . . . . . . . . . . 171 4.F.6 Top 20 states facing the most threatening strategic environment . . . . . . 172 4.F.7 Economic-based potential threat versus population-based potential threat . 173 4.G Dependent variables: Military investments . . . . . . . . . . . . . . . . . . . . . . . 175 4.G.1 Evolution of the military investments over time . . . . . . . . . . . . . . . . 175 4.G.2 Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 4.G.3 Comparing individual countries over time . . . . . . . . . . . . . . . . . . . 177 4.H Regression models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 4.H.1 Dropping the population-based potential threat variable . . . . . . . . . . . 179 4.H.2 Bivariate regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 4.H.3 Post-WWII sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 4.H.4 Alternative interest compatibility measures . . . . . . . . . . . . . . . . . . 181 4.H.5 Summary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 4.I GDP, Population, and GDPpc Component Datasets . . . . . . . . . . . . . . . . . 184 4.J Latent Variable Model Specication . . . . . . . . . . . . . . . . . . . . . . . . . . 191 Chapter 5: Re ections and future research 194 5.1 Territorial control and subnational con ict processes . . . . . . . . . . . . . . . . . 195 5.2 Subnational governance in con ict . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 5.3 Threat and arming in the international system . . . . . . . . . . . . . . . . . . . . 202 Bibliography 205 vii List Of Tables 2.1 Set of possible states Q =fS1;S2;S3;S4;S5g of the latent variable territorial control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.2 Heuristic to translate the observed exposure to terrorist attacks and conventional war acts into the categorical variable of rebel tactics O. . . . . . . . . . . . . . . . 28 2.3 Transition matrix , as inspired by Kalyvas (2006). Rows sum to one. Values across the diagonal indicate the probability of a grid cell remaining in the same state across two periods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.4 Matrix of emission probabilities used in the estimation of the HMM. Rows sum to one. The probability value in each cell of this matrix answers the following question: \Given that the true state of an areai at timet is, for example,S1, what is the probability of observing, for example, A from the data?" . . . . . . . . . . . 33 2.5 Relationship between rebel territorial control and deforestation in Colombia. . . . 39 2.6 Summary statistics for the logistic regression model of deforestation in Colombia on changes in territorial control as a result of the 2016 peace agreement. The unit of analysis for territorial control is annual averages of monthly-level estimates for 0.25 degree hexagonal grid cells. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 3.2 Description and data sources for key municipal-level indicators capturing insecurity. 83 3.3 Summary statistics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 3.4 Regression results for the con ict variable: Total fatalities per 100,000 population. 101 3.5 Regression results for the con ict variable: Total non-con ict related fatalities per 100,000 population. Con ict variable is standardized (mean zero, sd one). . . . . . 102 3.6 Regression results for the con ict variable: Armed forces fatalities by non-state armed actors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 3.7 Regression results for the con ict variable: Armed forces fatalities by guerrillas (FARC and ELN). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 3.8 Regression results for the con ict variable: Oensives against armed forces by non- state armed actors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 3.9 Regression results for the con ict variable: Oensives against armed forces by guerrillas (FARC and ELN). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 3.10 Regression results for the con ict variable: Political homicides by non-state armed actors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 3.11 Regression results for the con ict variable: Political homicides by guerrillas (FARC and ELN). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 3.12 Regression results for the con ict variable: Attacks against civilians by non-state armed actors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 viii 3.13 Regression results for the con ict variable: Attacks against civilians by guerrillas (FARC and ELN). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 4.1 Regression models relating dierent specications of the potential threat variable to investments in arming and power projection. Power-resources are measured using SDP at a $3 per diem subsistence level. The loss of strength gradient is conceptualized as curvilinear using the formula 1 log(distance) . Interest compatibility based joint democracy using Polity scores. . . . . . . . . . . . . . . . . . . . . . . . 142 4.2 Regression models relating dierent specications of the potential threat variable to investments in arming and power projection. Power-resources are measured using GDP. The loss of strength gradient is conceptualized as curvilinear using the formula 1 log(distance) . Interest compatibility based joint democracy using Polity scores. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 4.3 Concepts and operational denitions of each of the component parts of the country- year potential threat measure based on economic resources. . . . . . . . . . . . . . 162 4.4 Concepts and operational denitions of each of the component parts of the country- year potential threat measure based on economic resources. . . . . . . . . . . . . . 163 4.5 Hypothetical example of a three-state system. The example demonstrates how each component part of the potential threat variable is combined into the nal value for this country-year variable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 4.6 Summary statistics for key variables. . . . . . . . . . . . . . . . . . . . . . . . . . . 183 4.7 Component Measures for GDP, GDP per capita, and Population Latent Variable Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 4.8 Prior Distribution for Latent Variables and Model Level Parameter Estimates . . 193 ix List Of Figures 1.1 Overview of the linkage between dissertation chapters. . . . . . . . . . . . . . . . . 5 2.1 Spatial distribution of terrorist attacks and conventional war acts associated with Boko Haram in Nigeria in 2014. Data on terrorist attacks come from the Global Terrorism Database (GTD) data (START, 2016). Data on the location of conven- tional war acts come from the Georeferenced Event Dataset (GED) data (Sundberg and Melander, 2013). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.2 The map illustrates the relationship between territorial control and Boko Haram tactical choice in civil war in Nigeria in 2015. . . . . . . . . . . . . . . . . . . . . . 26 2.3 Theorized relationship between observed variation in rebel tactical choice and levels of territorial control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.4 Graphical representation of a HMM as a Bayesian network. . . . . . . . . . . . . . 31 2.5 Estimated levels of territorial control in Colombia for hexagonal grid cells with a minimum diameter of 0.25 degrees, excluding the Amazon and Orinoco natural regions (N = 851). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 2.6 Estimated levels of territorial control in Northeast Nigeria for hexagonal grid cells with a minimum diameter of 0.25 degrees (N = 942). . . . . . . . . . . . . . . . . . 42 2.7 Spearman's rank-order correlation coecients for a annual-level correlations be- tween HMM estimates and the Armed Con ict Location and Event Data (ACLED) validation data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 2.8 The schematic illustrates how shifting the location of the grid cell centroids from their original (randomly sampled) location (panel A) by just 25% (panel B) can result in vastly dierent conclusions about the coding of rebel tactics. Red dots indicate the location of conventional events; blue triangles those of terrorist at- tacks. This is a simplied example | in the analysis, Tactics it is computed using probabilities from Poisson distributions under application of a margin parameter. . 49 2.9 Logistic function that describes the decay of the in uence of an event in relation to a centroid in the spatial and temporal dimensions. . . . . . . . . . . . . . . . . . 52 2.10 In uence of the slope parameter and in ection parameter on the shape of the logistic decay curve for an example of distances varying from 1km to 150km. . . . 52 2.11 In uence of the slope parameter and in ection parameter on the shape of the logistic decay curve for an example of event ages from 0 to 12 months. . . . . . . . 53 2.12 Comparing the dynamics of continuous versus discrete aggregation of events in the spatial dimension using simulated data. . . . . . . . . . . . . . . . . . . . . . . . . 56 2.13 The gure compares the distribution of transition probabilities between the empir- ical observations from Kalyvas (2006) and the modied transition probabilities in this paper. The graph shows that while the transition probabilities dier slightly, the patterns of transitions between states from t 1 to t remain unchanged. . . . . 59 x 2.14 Number of cases that the measurement strategy can be applied to based on dierent thresholds of power asymmetry between the rebels and the government. Data on power asymmetry come from Polo and Gleditsch (2016). . . . . . . . . . . . . . . . 60 2.15 The graph illustrates the selection of cases for which the measurement strategy is applicable based on thresholds in average and maximum rebel-to-government troop ratios over the course of the con ict. Plotted in red are cases that would be included based on a 0.5 threshold indicating rebels that are half as strong as the government forces. Future work will investigate the determination of the most appropriate threshold. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 2.16 Yearly averages of monthly-level ACLED validation data values. Values that are closer to 0 indicate full rebel control; values closer to 1 full government control. 0.5 indicates cells that are highly contested. . . . . . . . . . . . . . . . . . . . . . . . 62 3.1 Theoretical relationship between insecurity and government spending on welfare- related public goods. The theoretical considerations focus on explaining empirical settings that fall on the solid part of the curve, namely lower levels of insecurity. . 77 3.2 The plot shows the geographical and temporal evolution of the Colombian con ict over time. Dark shaded areas represent municipalities in which guerrillas (Fuerzas Armadas Revolucionarias de Colombia{Ej ercito del Pueblo (FARC) and/or Ej ercito de Liberaci on Nacional (ELN)) are present. The maps illustrate that in the past 25 years, Colombia experienced the highest level of civil con ict in the early 2000s. In 2003, guerrillas were operating in over 60% of municipalities. In 2013, this number dropped to 18 percent of municipalities in which guerrillas were present. . . . . . . 79 3.3 Eect of guerrilla presence on municipal expenditure in education, health care, and water/sanitation. The points plot the coecient for a regression of munici- pal welfare investments on the interaction between a dummy for one-year lagged guerrilla presence and the year (including lower order terms) with 95% condence intervals. The model is estimated via Ordinary Least Squares (OLS) and includes municipality xed eects. Standard errors are clustered by municipality. 1994 is the excluded category for the year dummies. Positive coecients denote eects of guerrilla presence on welfare investments that are larger than in 1994; negative co- ecients smaller eects than in 1995. Shown in the background by a thick grey line is the standardized number of municipalities with guerrilla presence in a given year. Positive values denote above average numbers of municipalities in which guerrillas are present. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 3.4 Coecient plot for three dependent variables: Annual per capita municipal in- vestment in education, health care, and water/sanitation. Dependent variables are measured in thousands of per capita constant 2010 pesos and log-transformed. Bud- getary controls are included as contemporaneous and one-year lagged regressors. All other explanatory variables are lagged by one year. All insecurity indicators have been standardized to have a mean of zero and a standard deviation of one. All models include the full set of baseline controls. 0.0001 is added to all variables be- fore logging. Horizontal lines indicate the 90% and 95% condence intervals. Solid shapes correspond to coecients that are signicant at the minimum 5% level of signicance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 xi 3.5 Coecient plot illustrating the eects of control variables capturing the three al- ternative explanations for increased municipal social welfare investments in light of insecurity. Point estimates represent the coecients of the insecurity indicators for education and welfare per capita investments. Investments in the water sector are excluded from the graph, but can be found in the full regression tables in the appendix. Dependent variables are measured in thousands of per capita constant 2010 pesos and log-transformed. Budgetary controls are included as contempora- neous and one-year lagged regressors. All other explanatory variables are lagged by one year. All insecurity indicators have been standardized to have a mean of zero and a standard deviation of one. All models include the full set of baseline controls. 0.0001 is added to all variables before logging. Horizontal lines indicate the 90% and 95% condence intervals. Solid shapes correspond to coecients that are signicant at the minimum 5% level of signicance. . . . . . . . . . . . . . . . . 96 4.1 Evolution of economic power-resources based on a state's share of global SDP versus GDP. SDP is based on a $3 per diem subsistence level. A state either has positive surplus or no surplus at all. If a state has no SDP, then it would need to extract from the available subsistence resources, which reduces the ability of individuals to produce economic surplus. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.2 Top 10 powers ranked by their average share of global SDP or GDP. With the exception of China and a few other powers, the membership in the top 10 club is similar between the two measures of economic income. What does change is the rank-ordering of the countries. SDP produces a more historically valid ranking of the great powers than GDP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 4.3 Yearly correlation coecients with 95% condence intervals. In each panel, we assess the degree to which a state's share of global SDP and GDP correlate with each of four component variables of CINC. Note that we utilize an alternative to CINC's population measure (Supplementary Appendix 4.D). . . . . . . . . . . . . 124 4.4 Top 10 potentially threatening states for the United States by decade for $3 per diem subsistence level relative SDP (upper panel), relative GDP (middle panel), and relative population (lower panel). Power-ratios are weighted by the inverse of the logged distance between Washington D.C. and the other state's capital. Countries that are potentially threatening to the United States are not democratic, which is denoted through darker shading (not jointly democratic). . . . . . . . . . 131 4.5 China's and Russia's contribution to the total potential threat faced by the United States for SDP versus GDP. Preference compatibility is measured using Polity scores, and supplemented with data from Boix et al. (2013) to reduce the number of missing values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 4.6 Coecients and 95% condence intervals of standardized potential threat variables for regression models of two dependent variables|the military expenditure index and naval tonnage index|on potential threat and control variables. See Tables 4.1 and 4.2 for model specication information. . . . . . . . . . . . . . . . . . . . . . 136 4.7 Change in military burden over time for regions: the Americas (including the US and Canada), Europe (including Russia), Asia, the Middle East and North Africa, and Sub-Saharan Africa. Lines represent the smoothed average over all countries in the region for two indicators of military burden: military expenditure as a pro- portion of SDP versus as a proportion of GDP. . . . . . . . . . . . . . . . . . . . . 137 4.8 Change in military burden over time for select countries. . . . . . . . . . . . . . . 138 xii 4.9 The dollar values displayed on the x-axes and y-axes in the panels above are in billions of $US. Suppose a country with a population of 2,739,726 people. Such a country needs to generate 3 billion $United States (US) dollars (365 days $3 per-day 2,739,726 people) per year to healthfully sustain each member of the population over the long term, which is v, the minimum surplus value. Such a country is consuming all of its income for subsistence up until it generates income surpassing this minimum surplus value v. Once such a country generates income greater than v, the country is generating positive surplus income which it can invest in items other than \bread" (e.g., \butter" or \guns"). Poor and under- developed countries do exist today and in earlier periods of history with income levels at and below this threshold. Indeed, some state governments have worked diligently to develop extractive institutions to take even the subsistence income of the population. However, these states do not maintain the levels of healthy adults necessary for other state-making tasks (e.g., conscription) to sustain such a strategy over the long run. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 4.10 The top row of panels shows the yearly correlation between GDP and surplus in- come (SDP). The middle row of the panels shows the yearly correlation between GDP and subsistence income. The bottom row of panels shows the yearly pro- portion of countries that generate enough income to pass above the subsistence threshold at $1, $2, or $3 per person per day. The columns indicate these subsis- tence thresholds for each set of panels. . . . . . . . . . . . . . . . . . . . . . . . . . 148 4.11 Top 10 potentially threatening states for Japan by decade for $3 per diem sub- sistence level relative SDP on the upper, standard relative GDP in the middle, and relative population on the lower panel. Dyads that are not jointly democratic are potentially threatening and denoted through opaque shading. Dyads that are jointly democratic are not potentially threatening and denoted through brighter shading. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 4.12 Top 10 potentially threatening states for the United Kingdom by decade for $3 per diem subsistence level relative SDP on the upper, standard relative GDP in the mid- dle, and relative population on the lower panel. Dyads that are not jointly demo- cratic are potentially threatening and denoted through opaque shading. Dyads that are jointly democratic are not potentially threatening and denoted through brighter shading. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 4.13 Annual correlation with 95% condence intervals between the original CINC score and a re-computation of CINC that drops China from the global sums of iron and steel production, primary energy consumption, total population, urban population, military expenditure, and military personnel. The annual observation for China is dropped from both series. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 4.14 The plots display yearly correlation coecients with 95% condence intervals. In each of the panels, we assess the degree to which SDP (orange) and GDP (grey) cor- relate with the iron and steel production and primary energy consumption variables of CINC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 4.15 The graph plots the natural logarithm of GDP against the natural logarithm of SDP for select years. As time progresses and countries develop, SDP and GDP correlate highly. An exception are least developed countries, mostly in Sub-Saharan counties, who do not have a positive surplus in 2010. The SDP measure is based on a $3 per day subsistence threshold and is truncated to 1 for countries with no surplus resources in order to allow for a transformation via the natural logarithm. . . . . . 156 xiii 4.16 The graph plots GDP per capita against SDP across all country-years in the sam- ple. The linear patterns of dots show individual countries' trajectories over time. Labeled are observations for select countries in 1990. . . . . . . . . . . . . . . . . 157 4.17 Comparing the binning of countries by a) distance, b) a linear transformation of distance computed as max(Distance)Distance max(Distance) , and c) the inverse of the logged distance from the perspective of the United States, France, and China. . . . . . . . . . . . 165 4.18 The graph illustrates the construction of the potential threat measure for the USin 1935. The SDP is based on a $3 per diem subsistence value. . . . . . . . . . . . . . 166 4.19 Correlation plot for alternative potential threat measures, using SDP ($3 subsis- tence level) to measure economic resources. Colored cells denote values that are signicant at the minimum 5% level of signicance. . . . . . . . . . . . . . . . . . 167 4.20 The graphs show the evolution of potential threat over time for alternative indica- tors of preference compatibility (see the caption of Figure 4.19 for data sources). Plotted is each country-year observation with the line denoting the loess smoothed trend over time across all countries. . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 4.21 Evolution of potential threat over time for alternative indicators of preference com- patibility. Plotted is each country-year observation with the lines denoting the loess smoothed trend over time for three groups of states: a) states that enter the international system before 1900, b) states that enter between 1900 and 1945, and c) states that enter after 1945 based on Gleditsch and Ward (1999). . . . . . . . . 169 4.22 Evolution of potential threat over time for each indicator of preference compatibility (see the caption of Figure 4.19 for data sources). Plotted is each country-year observation with the lines denoting the line of best t over time for the rst, second, third, and fourth quartile of states based on the distribution of the per capita SDP (using a $3 per diem subsistence level). . . . . . . . . . . . . . . . . . . . . . . . . . 170 4.23 Maps plotting the spatiotemporal distribution of the natural log of the potential threat variable for the years 1965 and 2000. Potential threat is measured through joint democracy based on the Polity, Boix et al., and Unied Democracy Scale (UDS) scores, respectively. Power-resources are measured using the SDP indicator with a $3 per diem subsistence level. Grey shaded areas denote missing values. The maps are based on the borders for 1 January 1965 and 2000, respectively, using data from thecshapes library inR (Weidmann et al., 2010). The operationalization of each indicator is based on substantive choices of each coding team. Therefore, the coverage does not always perfectly map, either spatially or temporally. For example, some geographical spaces such as Greenland or former colonies in Africa are missing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 4.24 Top 20 states facing the most potentially threatening strategic environment in 1816, 1910, 1935, 1965, 1990, and 2010. The red dots show our estimate of the total level of potential threat each country faces when using the Polity2 score to measure preference compatibility; green triangles the Boix et al. estimates, and blue squares the UDS potential threat scores. Countries are ranked based on the potential threat variable that measures preference compatibility via the Polity2 score. Error bars indicate the 95% condence intervals for the average of all alternative potential threat measures.All potential threat variables are standardized; hence, the x-axis measures are expressed in standard deviations. . . . . . . . . . . . . . . . . . . . . 172 4.25 Relationship between country-year values of potential threat based on population on the y-axis and potential threat based on economic resources for various subsis- tence thresholds on the x-axis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 xiv 4.26 Annual correlation between three alternative Potential Threat (PT) measures over time. We vary how power resources are measured across the three indicators: using GDP, using SDP ($3 subsistence threshold), and using population. For all indicators, preference compatibility is measured via joint democracy (Polity) and power resources are weighted by the inverse of the logged distance between capital cities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 4.27 The two upper plots illustrate the temporal evolution of the military expenditure index (military expenditure/SDP in 2011 constant US Purchasing Power Parity (PPP) dollars) and naval tonnage index (naval tonnage/SDP in constant 2011 international PPP dollars), respectively. . . . . . . . . . . . . . . . . . . . . . . . . 175 4.28 The plot illustrates the correlations between the two alternative dependent variables over time. Points denote the correlation coecients for each year between 1965 and 2007. Lines represent the Loess smooth over those points. . . . . . . . . . . . . . . 176 4.29 The plot demonstrates the ability of our measurement strategy to obtain scores for the level of potential threat that individual countries face at any given point in time (granted data availability). Plotted in the rst row are economic resource- based potential threat scores using alternative regime type indicators to measure preference compatibility for the United States, the United Kingdom, France, Japan, China, and Brazil in the 20th century | the line representing a smoothed trend across all variables. In the rows below, we graph the time trends for the two dependent variables military expenditure as a proportion of SDP and naval tonnage as a proportion of SDP. All variables are shown on a logarithmic scale with base 10. 177 4.30 The graph shows the evolution of the potential threat across world regions for alternative measures of preference compatibility. . . . . . . . . . . . . . . . . . . . 178 4.31 Coecients and 95% condence intervals of standardized potential threat variables for regression models of two dependent variables | the military expenditure index and naval tonnage index | on potential threat and control variables. All models include controls for the natural log of income (SDP or GDP) and a country's Polity2 score. Standard errors are clustered by country; right-hand side variables are lagged by one year. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 4.32 Regression models upon dropping all control variables. . . . . . . . . . . . . . . . . 180 4.33 Regression models for a post-WWII sample. For model specication details, see the caption of Figure 4.31. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 4.34 Regression models for alternative preference compatibility measures. For model specication details, see the caption of Figure 4.31. . . . . . . . . . . . . . . . . . . 181 4.35 Regression models for alternative preference compatibility measures. For model specication details, see the caption of Figure 4.31. . . . . . . . . . . . . . . . . . . 182 xv Abbreviations ACLED Armed Con ict Location and Event Data AF Armed Forces AU African Union AUC Autodefensas Unidades de Colombia BACRIM bandas criminales CEDE Centro de Estudios sobre Desarrollo Econ omico CINC Composite Index of National Capabili- ties COIN Counterinsurgency COW Correlates of War DANE Departamento Administrativo Nacional de Estad stica DNP Departamento nacional de planeaci on ELN Ej ercito de Liberaci on Nacional FARC Fuerzas Armadas Revolucionarias de Colombia{Ej ercito del Pueblo GDP Gross Domestic Product GED Georeferenced Event Dataset GNI Gross National Income GNP Gross National Product GTD Global Terrorism Database HMM Hidden Markov Model HMRF Hidden Markov Random Field ICB International Crisis Behavior IDEAM Instituto de Hidrolog a, Meteorolog a y Estudios Ambientales IDP Internally Displaced Person IGO Intergovernmental Organization ISA International Studies Association JAGS Just Another Gibbs Sampler LSG Loss of Strength Gradient MAPS Monitoring Attitudes, Perceptions and Support MAUP Modiable Areal Unit Problem MCMC Markov Chain Monte Carlo NMC National Material Capabilities NP National Police NSA Non-State Actors in Armed Con ict Dataset NSAA Non-state armed actor OECD Organisation for Economic Co- operation and Development OLS Ordinary Least Squares PPP Purchasing Power Parity PRIO Peace Research Institute Oslo PT Potential Threat PWT Penn World Tables SDP Surplus Domestic Product SGP Sistema General de Participaciones UCDP Uppsala Con ict Data Program UDS Unied Democracy Scale UNGA United Nations General Assembly US United States WHO World Health Organization xvi Chapter 1 Introduction: Measuring state capacity in con ict Acquiring accurate data on actors engulfed in violent con ict is one of the great challenges for the study of con ict processes. These data are often of strategic importance and thus actors have incentives to actively hide them from the public eye. Detailed information on economic or military capability risks exposing actors' vulnerability. The publication of con ict-related data may uncover war crimes and cause scrutiny that persists long after the con ict ended. In active war zones, the danger of bodily harm, displacement of the population, and destruction of infrastructure renders the collection and communication of con ict-related data a colossal challenge. The resulting dearth of data in the eld of con ict processes increases the importance of inno- vative approaches to measurement. Measurement forms the connective tissue between empirical observation and theoretical constructs. This dissertation contributes to a better understanding of con ict processes at the intrastate and interstate levels by oering improved strategies toward the measurement of key concepts of interest, such as territorial control in civil war and economic power in the international system. A guiding question for this dissertation is how to measure the capacity of states engulfed in con ictual interactions with internal and external adversaries? I center on the link between states' administrative/bureaucratic and military capacity. 1 In interstate con ict, a state's ability 1 I follow Hendrix's denition of military capacity as \the state's ability to deter or repel challenges to its authority with force" and administrative capacity as states' \ability to collect and manage information" (Hendrix, 2010, 274) | a precondition for the ability to extract revenues and redistribute resources. 1 to extract resources from the land or the population is a central determinant of its ability to develop military might. At the intrastate level, low state reach and the under-provision of goods and services | a consequence of low administrative capacity | increase the risk of violent rebellion and potentially trigger a test of the government's military capacity (Hendrix and Young, 2014, 333). Military and administrative capacity can thus be thought of as coercive versus non-coercive, highly correlated, dimensions of a state's ability \to implement preferred policies" (Buhaug, 2010, 109). The three articles that comprise this dissertation take dierent perspectives on the interplay between administrative and military capacity in con ict. Chapter 2 introduces a new measure of territorial control in civil war, dened as the \extent to which actors are able to establish exclusive rule on a territory." (Kalyvas, 2006, 111) A low level of administrative capacity ren- ders subnational areas more susceptible to violent contestation and increases the diculty of maintaining territorial control. A minimum level of military capacity, on the other hand, is ben- ecial toward the goal of establishing stable and non-fragmented territorial control needed to erect non-coercive governance regimes. Chapter 3 focuses on governments' incentives to leverage their administrative capacity, in particular the use of public good provision, to enhance their war ghting capacity in intrastate con icts. I demonstrate that governments use welfare spending strategically when faced with an internal armed challenger to their statehood. Chapter 4 adds an interstate perspective, which challenges the common use of GDP to measure states' capacity to invest in military capabilities or other goods and services (i.e. aspects of administrative capacity). The analysis demonstrates that comparisons between states' military capacity, as operationalized by the percentage of income spent on the military, are more concurrently valid when accounting for subsistence resources required to ensure the survival of their population. I use a model-based approach toward measurement to compensate for shortcomings in the quantity and quality of empirical data on con ict processes. The specication of an underlying model allows me to leverage observed data for the quantication of theoretical constructs that are 2 otherwise not easily observed (Tal, 2017). For example, the model of territorial control in civil war in Chapter 2 stipulates observed variation in rebel tactics as an avenue to measure who commands control in a given area. The model of potential threat in Chapter 4 is used to translate three observable indicators, namely economic power, distance, and preference compatibility, into levels of the highly complex and hard-to-measure concept of threat in a state's geopolitical environment. Measurement thus sits at the center of the process of scientic inquiry in this collection of essays, providing the link between data collection and theoretical constructs. Measurement is itself a deeply theoretical process. As Tal (2017) states, \some level of theory- ladenness is a prerequisite for measurements to have any evidential power." Theory guides the selection of relevant component indicators of interest that are used to measure complex constructs. Theory also informs the model that is formulated as a blueprint for how to combine the component parts for measures of interest that are comprised of more than one empirical indicator. As an example, it is a theory of non-state armed actor behavior in asymmetric intrastate con ict 2 that suggests the distinction between terrorist and conventional guerrilla tactics as a fruitful avenue toward identifying rebel versus government strongholds (Chapter 2). While innovative measurement can compensate for some shortcomings, the underlying data is still the key input for any empirical research endeavor. In a eld that is hamstrung by a shortage of ne-grained data, the publication of systematic information on actor and con ict characteristics, such as the Correlates of War (COW) project (Singer and Small, 1972) and the International Crisis Behavior (ICB) data (Brecher et al., 2017; Brecher and Wilkenfeld, 2000) for interstate con ict, or the Uppsala Armed Con ict Data for intrastate violence, 3 spurred both new empirical insights into con ict processes, and novel theoretical arguments. Further data collection eorts allowed scholars to move from cross-country analyses to more disaggregated perspectives. Over the last decade, a number of projects started to publish con ict event data that record the 2 See for example Polo and Gleditsch (2016); de la Calle and S anchez-Cuenca (2015). 3 Later extended as the Uppsala Con ict Data Program (UCDP)/Peace Research Institute Oslo (PRIO) Armed Con ict Database and related datasets (Gleditsch et al., 2002). 3 longitude and latitude of individual con ict events, as well as detailed information on the timing, duration, actors, mode of violence, and fatalities associated with an incident | giving rise to a new research program exploring micro-dynamics of intrastate con ict. 4 These large-scale data acquisition eorts form the empirical foundation for the improvements in measurement presented in this dissertation. This collection of essays contributes a number of new insights and methods to the study of con ict processes at the intrastate and interstate levels. In Chapter 2, I develop and validate a measurement model for the estimation of territorial control in asymmetric civil war. For most contemporary and historic con icts, we lack detailed information on who controls which subna- tional area at a given point in time. I advance a theoretical model of actor behavior in asymmetric civil war and utilize geo-coded con ict event data in a machine learning framework to compute territorial control estimates for insurgencies in Nigeria and Colombia. In addition, I develop an improved method to estimate the exposure of subnational geographic units to con ict events. In Chapter 3, I engage with one of the dominant paradigms in the study of asymmetric civil war: the notion that counterinsurgents seek to capture the `hearts and minds' of the population in order to gain the upper hand in the con ict and establish territorial control. Using data from Colombian municipalities, I highlight the strategic importance of welfare spending by showing that governments do indeed react to violence with higher investment in citizen welfare, but only if violence presents a direct threat to the survival of the state. Chapter 4 champions a measure- ment correction for one of the most widely used theoretical constructs in international relations research: economic power in the international system. Together with my co-authors, I show that the mis-measurement of economic power via GDP can be mended by accounting for subsistence, that is resources a state's population needs for survival. We introduce SDP as a better measure for the resources that states can invest in the military or citizen welfare. In addition, the chapter provides an improved method to estimate potential threat in states' geopolitical environment and 4 See for example the GED (Sundberg and Melander, 2013) or the ACLED project (Raleigh et al., 2010). 4 an extended data series of military expenditure as a proportion of economic resources (GDP or SDP) from 1816{2012. Interstate level Chapter 3 Subnational welfare spending Chapter 2 Territorial control in civil war Actor behavior in asymmetric civil wars Intrastate level Measurement innovation for conict processes Chapter 4 Surplus Domestic Product Use of economic resources in conict Military & administrative state capacity Figure 1.1: Overview of the linkage between dissertation chapters. Figure 1.1 illustrates the linkages between and among the three articles that constitute the core of this dissertation, and their contribution to our understanding of con ict processes at interstate and intrastate levels. The chapters dier regarding the nature and content of their con- tribution. First, Chapters 2 and 4 provide new, theoretically founded, measurement approaches to key concepts of interest in the study of con ict processes: territorial control in civil war and economic power in the international system. These studies enhance both the accuracy and the spatiotemporal coverage of their respective measures of interest, as compared to existing work. Second, Chapters 3 and 4 contribute new insights into governments' use of resources in con ictual 5 interactions and the conversion of economic power into military power. The measurement inno- vation in Chapter 4 allows us to more accurately measure the resources that states have available for investment and yields new insights into the distribution of economic capability, threat, and military burdens in the international system. Chapter 3 adds an intrastate perspective on the interaction between con ict and states' resource use by showing that subnational governments use investments in citizen welfare as a strategic tool in con ict. Comprising the third and nal side of the triangle depicted in Figure 1.1, Chapters 3 and 2 improve our understanding of spa- tiotemporal dynamics of actor behavior in civil war, in particular asymmetric (or irregular) civil wars that are fought between the state and much weaker non-state armed actors. Both studies focus on localized dynamics by utilizing ne-grained units of analysis, such as municipalities in Chapter 3 and small grid cells in Chapter 2. In the following sections, I elaborate in more detail on the connections between the chapters of this dissertation. 1.1 Measurement innovations for con ict processes Chapters 2 and 4 tackle the challenge of improving the measurement of two widely used indicators in international relations: territorial control in civil war and economic power in the international system. Both measures are central variables for the understanding of con ict processes at the intrastate and interstate levels, respectively, and both are notoriously dicult to measure accu- rately. The detailed measurement of territorial control is made dicult primarily by \messy patch- work" patterns exhibited by asymmetric intrastate con icts (Kalyvas, 2006, 88) and a dearth of localized information in con ict zones. Con icts like the Boko Haram insurgency in Nige- ria, the communist insurgency in the Philippines, or the Maoist insurgency in India rarely see 6 clearly dened frontlines. Instead, actors often command varying levels of control on a village-by- village basis. 5 Mapping territorial control in these asymmetric con icts requires detailed localized knowledge | a scarce resource in areas that, in many cases, are characterized by remote location, inaccessible terrain, and an under-provision of infrastructure, that foster insurgency. In contrast, conventionally fought civil wars, such as the con ict between pro-Russian separatists and the Ukrainian military in the Donbass, tend to feature comparatively smooth frontlines that make it easier to assess territorial control at a given location and point in time. The diculty of measur- ing territorial control, in particular in asymmetric con icts, results in a shortage of high quality data on territorial control | a source of information that is crucial for a better understanding of subnational con ict dynamics. The measurement approach in Chapter 2 is rooted in conceptualizing territorial control as a la- tent variable that is not observed directly, but whose consequences, namely the relative frequency of terrorism versus guerrilla ghting, can be detected and measured. For most asymmetric con- icts, we cannot directly observe the status of territorial control on the ground. However, thanks to data collection eorts such as the GED (Sundberg and Melander, 2013) and the GTD, 6 we have information on the location and timing of individual con ict events. 7 I demonstrate that when faced with obstacles to observing territorial control directly, we can instead estimate the indicator of interest via a) characteristics of observed con ict events and b) a model that species the relationship between event characteristics as observed emissions of territorial control as the underlying latent construct of interest. The labeling of con ict events in the GED and GTD databases allows me to distinguish between cases in which rebels use terrorist, as opposed to conventional guerrilla, tactics. The theoretical model underlying the estimation strategy species that rebels will use terrorism less in areas where they command higher levels of control, and will 5 See for example the analysis of village-level territorial control of the New People's Army of the Communist Party of the Philippines (Rubin, 2020). 6 See National Consortium for the Study of Terrorism and Responses to Terrorism (START) (2016). 7 Due to the diculty of collecting and communicating information on the occurrence of violent events in con ict zones, these data are likely subject to reporting biases (Reeder, 2018; Weidmann, 2015). These data are, however, the best source of information available to date. 7 instead resort to guerrilla tactics. Spatiotemporal variation in the co-occurrence of these two types of con ict events thus allows inferences about the underlying level of territorial control. The measurement of a state's power in the international system is a challenge both on con- ceptual and empirical terms. The question of what characterizes state power is one of the most debated questions in international relations. As a consequence, denitions of power vary widely. At their core is the capacity of actors to determine their and others' circumstances and actions (Barnett and Duvall, 2005). Understandings of power dier with regard to which traits make actors powerful, in particular material resources versus social in uence. Here, I focus on a ma- terialistic conceptualization of power as derived from economic wealth. This is certainly not the only dimension of what makes a state a powerful actor in the international sphere (Hart, 1976). However, economic might is a central dimension of states' power, in particular, because economic power can be converted into military power | albeit not one-to-one, for example due to increasing opportunity costs as a result of a high specialization in production and domestic resistance, as captured by the `guns vs. butter' dilemma (Poast, 2019). 8 Chapter 4 highlights another, to date under-appreciated, reason for why economic wealth cannot be converted one-to-one into military power. In a nutshell, not all resources that states produce are freely available for investment, because a signicant portion has to be used to sustain the population. The analysis in Chapter 4 suggests an amount of three dollars per day per person as the minimum resources needed to ensure the survival of the population. Historically, few states produced economic resources surpassing this minimum subsistence threshold, and many devel- oping countries today are struggling to generate wealth above subsistence levels. In the existing literature, a state's economic power is often measured via GDP, which is problematic because it assumes that all economic resources can be converted into military power or other material goods, and does not take the subsistence needs of the population into account. As a result, the relative material capacity of poor and populous states is dramatically overestimated when using GDP to 8 See section 1.2 below. 8 measure states' economic power. This systematic mis-measurement also biases conclusions regard- ing states' relative investment in the military and other goods and services. The commonly used indicator of measuring defense burdens as military expenditures as a percentage of GDP is fun- damentally awed, 9 because it tends to underestimate the relative amount of economic resources that states with high subsistence needs and lower levels of income invest in their militaries. Chapter 4 oers a number of improvements over existing measurements of economic power in the international system, and related indicators. First and foremost, we propose to separate GDP into surplus and subsistence income. We introduce Surplus Domestic Product (SDP) as a more accurate measure of the resources that states have available for investment. SDP is computed by subtracting from GDP the income necessary to meet the basic caloric needs of the population, which we specify at a level of three dollars per day per person. We present a number of derivative indicators using SDP that more accurately describe relative power relationships between coun- tries, potential threat states face in their geopolitical environment, and states' military burdens, than comparable indicators incorporating GDP. Second, the measurement strategy in Chapter 4 improves an existing measure of potential threat in the international system by a) incorporat- ing the new SDP indicator (as opposed to GDP) and b) specifying relative power relationships between states via dyadic, as opposed to global, ratios. 10 While existing research stipulates a decreasing trend of the global level of geopolitical competition as a result of more states entering the international system (Markowitz and Fariss, 2018), the adjusted measure instead shows an increase in average global levels of competition throughout the 20th century. 11 Decreasing trends of system-level geopolitical competition are only observed after the end of the Cold War, mainly due to democratization. As state democratize, they become less threatening to their peers in the international system. Third, Chapter 4 extends the spatiotemporal coverage of data on military 9 See for example the use of military expenditure as a percentage of GDP in the studies by Yesilyurt and Elhorst (2017) and Collier and Hoeer (2002). 10 Rather than assessing a dyads' power dierential by comparing states' share of total global power resources (see Markowitz and Fariss 2018), in Chapter 4, we instead compute an indicator of states' share of total dyadic power resources. 11 See Figures 4.20, 4.21, and 4.22 in the appendix to Chapter 4. 9 burdens by computing indicators of military expenditure as a percentage of economic resources in constant US PPP dollars; made possible by revised estimates of states' GDP developed in a related project (Fariss et al., 2017). 1.2 From `guns versus butter' to `hearts and minds' The `guns versus butter' tradeo holds that states must decide to allocate scarce resources to either welfare needs (`butter') or the military (`guns'). The dilemma for states is that while `butter' spending yields direct benets, military investments will be benecial only indirectly and as a function of the threat states face in the international system (Powell, 1993, 114). If a state invests too little in defense and gets defeated by another state, all previous and future welfare investments are at stake. High military investments, on the other hand, divert resources away from domestic consumption, and may ultimately result in an eciency loss in the production of warfare-related goods (Poast, 2019, 225). Chapter 4 demonstrates that existing estimates of the severity of the `guns versus butter' tradeo are biased, because the commonly used measure of the resources available for welfare versus warfare spending, namely GDP, does not take into account the basic subsistence needs of the population. In Chapter 4, this minimum level of resources needed to ensure the survival of the population is termed the basic `bread` that has to be provided in equilibrium, before investments in goods exceeding basic subsistence, such as the military or citizen welfare, can be made. Not accounting for subsistence needs led existing research to overestimate the resources available for `guns versus butter' spending. As a consequence, the relative severity of the `guns versus butter` tradeo between countries with low versus high subsistence needs are biased. As an example, in 2000, Finland and Nigeria had a similar GDP of 182 and 195 billion 2011 PPP US dollars, respectively. While Nigeria had to spend an estimated 66% of this amount on the basic subsistence needs of its population of 117 million, Finland needed to muster a mere 3% to ensure 10 the basic survival of its ve million inhabitants. 12 After accounting what countries need to spend on `bread,' Nigeria faced a much more severe `guns versus butter' tradeo than Finland, despite having similar levels of total income. 13 The mechanisms of the `guns versus butter' tradeo from International Relations theory cannot be translated directly to the realm of subnational con ict for three main reasons. First, the incentive structure for welfare versus warfare spending is dierent when facing an internal armed challenger to statehood, as opposed to a common external threat. Many intrastate con icts are fueled by grievances of the local population. Horizontal and vertical inequality in particular have been found to give rise to popular resistance, including violent rebellion (Cederman et al., 2013). The (under-)provision of `butter' can therefore become a variable contributing to the outbreak of violence itself, not a line item to be negotiated with the electorate. Second, the provision of welfare-related goods and services has been found to in uence the dynamics and intensity of intrastate con ict. For actors ghting insurgencies, local popular sup- port, and in particular information sharing, are crucial factors for success | and con ict parties might seek to foster it with public good provision (Arjona, 2016; Berman et al., 2011). Increasing the welfare of local populations and decreasing poverty also increases the opportunity cost for insurgent recruitment (Khanna and Zimmermann, 2017). The impact of public good provision on con ict dynamics is well documented in the existing literature, however, no consensus exists regarding the direction of the eect. For example Khanna and Zimmermann (2017) nd that police attacks against Maoist insurgents and rebel violence against civilians increased after the introduction of an anti-poverty program in India. Crost et al. (2016), on the other hand, show that insurgent violence in the Philippines decreased as a result of a cash transfer program. 12 This is based on a minimum subsistence threshold of $3 per person per day. 13 Note that SDP is dierent from GDP per capita in that it measures the total amount of resources, not the per person resources, that a country has available for `guns' or `butter' investments. Measuring the total amount of surplus income is relevant when seeking to scale national investment by available resources, such as for example the commonly used indicator of military expenditure as a percentage of GDP to measure defense burdens. This is relevant because two countries with diering levels of SDP would still need to spend a similar amount of resources to achieve, for example, military capabilities of a similar size. The cost of a ghter jet is approximately the same for Finland and Nigeria, but this purchase signies a much higher defense burden for Nigeria than it does for Finland due to its higher surplus income. 11 Finally, at the intrastate level, the authority over the use of welfare versus warfare resources is often subject to various levels of scal and functional decentralization, which implies that the seemingly unitary actor decision of `guns versus butter' purported by International Relations theory, on the subnational level, is a spatially disaggregated multi-actor game. The importance of accounting for decentralization is additionally elevated when considering that asymmetric in- trastate con icts are often locally bound and do not in uence all areas of a country equally. In many countries, local governments are responsible for the provision of welfare services such as education, health care, or sanitation, while the security apparatus remains centralized at the na- tional level. 14 Thus, to the extent that a relationship between welfare spending and con ict exists, as posited in this dissertation and existing research, 15 at the local level, and in particular under conditions of scal decentralization, it is unlikely to be driven by the `guns versus butter' trade-o. `Guns` are provided by a dierent entity, and local governments are left with the decision of a decision how and how much to spend on `butter.' `Butter' is generally described as \the government's provision of non-security goods aimed at enhancing societal welfare." (Poast, 2019, 225) However, in intrastate con icts where actors depend on the support of the local population for military success, service provision and coer- cive counterinsurgency eorts are strategic complements (Berman et al., 2011, 781), and `butter' spending becomes a security-related good. In fact, it has been shown that `butter' investments are not a monopoly of governments, and that in many cases, rebels vying to overtake the state engage in strategic public good provision as well (Stewart, 2018; Arjona, 2016; Mampilly, 2011). The insight that `butter' spending in intrastate con ict has a strategic value is at the core of the `hearts and minds' doctrine. Advocates of this counterinsurgency approach stipulate that \[d]efeating an insurgency requires a blend of both civilian and military eorts." (U.S. 14 For example, in Colombia and Indonesia, many aspects of public policy and service provision are highly de- centralized while defense and security remain within the authority of the central government (Bazzi and Gudgeon, 2015; Eaton, 2006). Thus, the considerations that municipal governments in Colombia face regarding welfare investments are not a true `guns versus butter' tradeo (Chapter 3). Instead, the spending on `guns' in Colom- bia is largely centralized, while decentralization reforms since the 1980s progressively increased the authority of municipalities over `butter' spending. 15 See for example Sexton et al. 2019; Taydas and Peksen 2012. 12 Army/Marine Corps, 2014, 1) Under the `winning hearts and minds' doctrine, the counterinsur- gent is to use tools such as the provision of development aid as a means to gain the support of the local population. However, counterinsurgents are not the only actor seeking to gain legitimacy through `butter' spending. It has been demonstrated that rebels will also use a mixture between a `hearts and minds' and a `coercion for cooperation' strategy in their interactions with the civilian population (Arjona, 2016, 5). Chapter 3 complements the interstate-level analysis of the `guns versus butter' tradeo in Chapter 4 with an intrastate-specic perspective on the use of economic resources in con ict. Using data from over 1,000 Colombian municipalities, I investigate how local governments adjust their welfare investments in the sectors of education, health, and sanitation as a reaction to the type of insecurity experienced on the ground. A testament to the strategic value of `butter` spending in civil wars, I demonstrate that welfare investments are not reactive to general insecurity experienced by the civilian population. Instead, welfare investments vary as a function of attacks that immediately undermine the statehood of local governments. The analysis yields more nuanced insights into the functioning of the `hearts and minds' approach. While welfare provision has been shown to decrease the incidence and intensity of violence in many con icts (Cort es and Montolio, 2014; Taydas and Peksen, 2012; Berman et al., 2011) and it is desirable for governments to improve welfare provision to under-served communities, in practice governments limit its use to situations in which they expect benets for their ability to fend o an armed challenger to statehood. 1.3 Actor behavior in asymmetric con icts The setting of asymmetric con icts is an important scope condition for the analyses in Chapters 2 and 3. Asymmetric (or irregular) con icts are situations in which \government forces have a clear advantage over rebels in coercive capacity" (Berman and Matanock, 2015, 444) The relative weakness of insurgents vis- a-vis the government is one of the central characteristics of insurgencies, 13 in particular at the beginning of a con ict (Fearon and Laitin, 2003, 79). According to Kalyvas and Balcells (2010, 415), these asymmetric con icts account for approximately 54% of all civil wars between 1944 and 2004, and many recent strifes, for example in Afghanistan or Myanmar, can be characterized as such. 16 At the core of why asymmetric con ict is an important scope condition for Chapters 2 and 3 is the centrality of the local population for con ict dynamics and actor behavior in clashes characterized by a high power asymmetry between combatants. As a eld manual for the US Armed Forces states, \[t]he struggle for legitimacy with the population is typically a central issue of an insurgency." (U.S. Army/Marine Corps, 2014, 9) The main reason lies in the \consequential role that civilians play in sharing information." (Berman and Matanock, 2015, 444) Asymmetric wars are not fought along trenches with heavy weaponry. 17 Instead, insurgents stage small-scale attacks against armed forces and escape counterattacks by blending back into the civilian popu- lation (Lyall and Wilson, 2009). This makes it dicult for the government to identify insurgents, especially because civilians have incentives to misrepresent their allegiance in an eort to mini- mize the risk of retribution from either side (Matanock and Garcia-Sanchez, Forthcoming). As a consequence, the success of both insurgents and counterinsurgents depends on their capacity to elicit civilian cooperation. Governments seek information to identify and prosecute combatants. Insurgents strive for cover and logistical support from the local population. As mentioned in Section 1.2, armed actors are faced with the challenge of striking a balance between leveraging their military capacity, i.e. coercion, and their administrative capacity, i.e. cooptation, to achieve their respective goals. The theoretical underpinnings of Chapters 2 and 3 are rooted in the central role that civil- ian cooperation plays for the success of insurgents and counterinsurgents in asymmetric civil war. Chapter 3 focuses on governments' attempts to win the allegiance of the local population, 16 See Figure 2.15 for an overview of con icts that can be characterized as asymmetric, based on data from Polo and Gleditsch (2016) 17 In fact, a higher level of mechanization in weaponry has been found to decrease the ability of armed forces to distinguish insurgents from civilians and selectively apply repression (Lyall and Wilson, 2009). 14 specically conditions under which a government will use welfare spending as a strategic tool to complement military counterinsurgency eorts. The three logics for why states invest in social welfare as a reaction to insecurity in Chapter 3 center around the function of welfare spending as a tool to win the support of the local population. I demonstrate that governments use welfare investments to enhance their ghting ability, and not out of a concern for the wellbeing of the population. In the theory of rebel tactics in Chapter 2, the danger of losing civilian support is one of the main reasons why, within their strongholds, rebels prefer guerrilla tactics, as opposed to more indiscriminate terrorist attacks. Indiscriminately endangering the lives and livelihoods of the civilians within their sphere of in uence risks alienating the population that rebels depend on for establishing and maintaining control. This is not to say that in conventionally fought civil con icts civilians do not in uence con ict dynamics, or that their victimization is not deliberate, on the contrary. 18 However, the logic of civilian victimization in conventional civil wars that feature clearly dened frontlines is much more akin to that in interstate con icts (Krcmaric, 2018). The warring parties in conventionally fought intrastate con icts are not as dependent on information from the civilian population for success (Berman and Matanock, 2015, 445), and civilian defection is harder when actors' control is rm and stable behind their respective frontlines (Krcmaric, 2018, 23). As Balcells argues, \violence in irregular war is a direct consequence of the competition for control of territory, for which information is crucial; control of information is, on the contrary, less crucial in conventional wars, where frontlines are non-porous and where the outcome of the war is mostly determined by the evolution of battles." (Balcells, 2017, 25) Thus, the measurement model in Chapter 2 and conclusions on the strategic importance of welfare spending in con ict in Chapter 3 are dependent on the context of asymmetric con ict. 18 See for example Balcells (2017) on civilian agency in the Spanish and Ivorian civil wars, or Lidow (2016) on the determinants of civilian victimization in Liberia. See Valentino (2014) for an overview of the literature. 15 In the following three chapters, I elaborate on the measurement of territorial control in asym- metric con icts (Chapter 2), the analysis of welfare spending as a strategic tool in civil wars (Chapter 3), and the establishment of Surplus Domestic Product (SDP) as a better measure of the resources states can invest in arming and other goods or services (Chapter 4). In the conclusion in Chapter 5, I detail the contributions of each essay for our understanding of con ict processes at the interstate and intrastate levels, in particular the interplay between actors' military and administrative capacity, and re ect on avenues for future research. 16 Chapter 2 Territorial control in civil wars: Theory and measurement using machine learning Securing control over territory is a key objective for actors in violent con ict. Those who exert control over an area have the opportunity to extract resources (Carter, 2015), pursue the col- laboration of the population (Arjona, 2016; Kalyvas, 2006), and increase their mobilization base (Stewart and Liou, 2017; de la Calle and S anchez-Cuenca, 2015). Areas of consolidated control can serve as safe havens for combatants and a home base from which future oensives can be coordinated and launched (Arjona, 2016). Gaining control over an area is a pre-condition for the establishment of non-violent political order. It gives actors the ability to govern non-coercively, for example via the provision of public goods (Stewart, 2018). Territorial control is thus a central variable of interest for the study of intrastate con ict dynamics (Staniland, 2012; Sambanis, 2004). Many intrastate con icts, in particular civil wars that feature a high power asymmetry between rebels and the government, do not exhibit clearly dened front lines. Territorial control in these asymmetric civil wars tends to be spatially fragmented and dicult to measure. To date, we lack data on territorial control that oer wide temporal and cross-sectional coverage and are suciently ne-grained to account for variation on the subnational level. This shortage of information causes territorial control to be astonishingly absent as a variable in micro-level studies of civil war | despite its centrality as a theoretical concept. Given the diculty of measuring territorial 17 control in asymmetric con icts via direct observation, how can we estimate changes in territorial control across time and space? I advance a novel measurement strategy for territorial control in asymmetric intrastate con ict. I show that we can estimate territorial control by translating a theory of actor behavior in civil war into a machine learning model and leveraging information on variation in rebel tactical choice based on geo-coded event data. Building on existing work regarding the relationship between territorial control and tactical choices of insurgents, I develop a theoretical model that links the relative frequency of terror- ist attacks and conventional war acts between government forces and the rebels to patterns of territorial control. The measurement strategy builds on two empirical relationships: 1) rebels use terrorism predominantly outside their strongholds; 2) preferring conventional guerrilla tactics when they command higher levels of control. Hence, we observe more insurgent terrorist attacks relative to conventional ghting in areas exhibiting a higher level of government control, and vice versa. Translating this theoretical relationship into measurement, I employ a function of an area's spatially and temporally weighted exposure to terrorism and combat events as observable emissions from the latent variable territorial control. I treat territorial control as a latent variable that can be modeled and estimated. Following Kalyvas (2006), territorial control is conceptualized as a categorical variable ranging from complete rebel control to complete government control, with levels of contestation in between. The latent variable territorial control is estimated via a Hidden Markov Model (HMM). HMMs compute the most likely evolution of territorial control in an area over time, conditional on observed rebel tactics and model priors. These priors specify beliefs about how likely a cell changes from one state of territorial control to another (transition probabilities) and how accurately observed rebel tactics measure the latent state of control (emission probabilities). Contrary to dominant applications in computer science, in this application, priors are not learned from training data but instead informed by theoretical arguments and out-of-sample empirical observations. As a proof of concept, I present estimates of territorial control for the con ict between the Fuerzas Armadas 18 Revolucionarias de Colombia{Ej ercito del Pueblo (FARC) rebels and the Colombian government from 2006 to 2017 and the Boko Haram insurgency in Nigeria from 2008 to 2017. Existing research considered the in uence of territorial control on aspects such as tactical choices by rebels and counterinsurgents (Carter, 2015; de la Calle and S anchez-Cuenca, 2015; Asal et al., 2012), the level of noncombatant victimization (Stewart and Liou, 2017; Aronson et al., 2017; Schutte, 2017; Bhavnani et al., 2011; Kalyvas and Kocher, 2009; Kalyvas, 2006), or the provision of public goods (Stewart, 2018; Mampilly, 2011). To the extent that territorial control is explicitly measured, it is operationalized as a binary variable indicating whether a given subnational actor commands territorial control or not (Stewart and Liou, 2017; Polo and Gleditsch, 2016; de la Calle and S anchez-Cuenca, 2015, 2012), studies are limited to a single in-depth case, 1 or territorial control is conceptualized as a function of distance to some power center (Schutte, 2017). These existing measures are plagued by a number of shortcomings. The operationalization as actor-specic binary indicator is oblivious to variation in the degree of territorial control that a group commands across space and over time. Sophisticated approaches utilizing con ict event data, information on geographical features such as road networks and terrain, or survey data provide ne-grained data of territorial control, but are extremely resource intensive to implement and to date applied to only a few con ict settings (Aronson et al., 2017; Tao et al., 2016). Lastly, measuring territorial control as a function of the distance from a government or rebel power center ignores that territories under control not usually evenly spaced around strongholds, but are often reminiscent of patchwork, with boundaries between zones of control being \blurred and uid." 2 (Kalyvas, 2006, 88) The project yields three main contributions. First, I provide ne-grained data on territorial control that accommodate high levels of spatial and temporal variation but are produced with a methodology that can be applied cross-nationally. Second, I show how con ict scholars can 1 See Aronson et al. (2017) on Nigeria, Schutte (2017) on Afghanistan, Tao et al. (2016) on Liberia, Garc a- S anchez (2016) on Colombia, Bhavnani et al. (2011) on Israel, the West Bank, and Gaza, Kalyvas and Kocher (2009) on Vietnam, and Kalyvas (2006) on Greece. 2 See also the maps created for Nigeria by Aronson et al. (2017). 19 use their rich theoretical knowledge to inform priors in machine learning applications. Finally, I develop a new approach toward the measurement of an area's exposure to con ict events. Rather than discretely assigning con ict events to grid cells, I compute a cell's exposure as the spatially and temporally weighted sum of con ict events. Not only does this approach account for the spatial dependence of con ict events in the estimation of the HMM | it presents a valuable methodology for event-based subnational con ict analyses beyond the measurement of territorial control. The paper proceeds as follows. First, I establish the centrality of territorial control for the study of intrastate con ict and discuss previous measurement attempts. I then formulate a theoretical account of actor behavior in civil war through which I relate territorial control to tactical choices of insurgents, specically the use of domestic terrorism versus conventional ghting. Leveraging con ict event data on the observed relative frequency of terrorist attacks and violent incidents that are indicative of guerrilla warfare, I obtain estimates of territorial control in a machine learning framework. Finally, I conduct out-of-sample validation to show that the methodology can recover ne-grained spatiotemporal patterns of territorial control. 2.1 Territorial control in civil war Territorial control is a crucial variable for understanding violent con ict, as it shapes armed actors' tactics and aspirations, in particular \the dynamics of bargaining, recruitment, and lethality" (de la Calle and S anchez-Cuenca, 2012, 583). Who commands what level of control has also been shown to condition civilian behavior in con ict zones, such as voting (Garc a-S anchez, 2016) and information sharing (Arjona, 2016). I dene territorial control as the \extent to which actors are able to establish exclusive rule on a territory." (Kalyvas, 2006, 111) Territorial control is not an end in itself, however, maximizing local monopoly on violence is a necessary condition toward the goal of establishing statehood. An actor who enjoys full statehood in a given area by denition 20 also has territorial control. For the purpose of measurement, I conceptualize territorial control as a ve-category variable ranging from full rebel to full government control. It is assumed to be zero-sum: In any given area, increases in the level of territorial control exercised by one actor equal decreases for competing actors. In the existing literature, territorial control is most prominently studied as a factor determining selective versus indiscriminate victimization of civilians by government or non-state actors (e.g. Stewart and Liou 2017; Quinn 2015; Kalyvas and Kocher 2009). Generally speaking, the less territorial control actors command, the more indiscriminate the violence they in ict will be, and vice versa. Other work considers the interplay between civilian cooperation and armed actor coercion in the establishment and consolidation of territorial control (Arjona, 2016). In cross-national studies, scholars frequently rely on a binary indicator from the Non-State Actors in Armed Con ict Dataset (NSA) to code whether a group exercised control over terri- tory (Cunningham et al., 2013). The NSA data additionally rate how eectively non-state armed actors exercise control. However, the information is supplied at the group-level and is unable to capture temporal and subnational variation. Alternative approaches that operationalize govern- ment territorial control as a function of the distance from a country's capital vary subnationally, but do not allow for temporal variation at a given location (Schutte, 2017). One of the key characteristics of asymmetric civil war is the absence of clearly dened front lines. Rebels that are weak compared to the government tend to avoid direct contact with state forces and \try to disperse as much as possible so that the state cannot respond to the multi- pronged challenge." (Arjona, 2016, 43) Neither the operationalization of territorial control via binary actor-level indicators, nor distance to the capital city, can account for the \messy patch- work" patterns (Kalyvas, 2006, 88) that are observed in empirical studies of asymmetric con ict. Published examples of country-specic accounts documenting this fragmentation include recent eorts to create estimates of territorial control using con ict event data for Liberia (Tao et al., 2016), post- or in-con ict surveys (Kalyvas and Kocher, 2009; Arjona, 2016), or the study of 21 historic military records and interviews (Kalyvas, 2006). The approach by Tao et al. (2016) and Aronson et al. (2017) to hand-code assault initiators and the status of territorial control after attacks based on media reports underlying the Georeferenced Event Dataset (GED) produces ne-grained estimates that can be constructed for a cross-section of con ict zones. However, it is extremely labor intensive and, at the time of writing, no data has been publicly released. While recent years have seen increased interest in studying territorial control in asymmetric civil war, the eld suers from a shortage of available data that vary subnationally and temporally, can recover \patchy" patterns of control, and are produced with a methodology that accommo- dates cross-country comparison. I improve upon existing approaches by conceptualizing territorial control as a latent variable that can be estimated for small subnational spatial and temporal units using publicly available event data. The methodology can feasibly be applied cross-nationally but provides sucient detail for within-country analyses. The model set-up is based on a theoretical account that links observable variation in rebel tactics | their choice between irregular terror- ist and more conventional guerrilla ghting strategies | to territorial control in the context of asymmetric civil wars. 2.2 Tactical choice in asymmetric civil war \...[I]nsurgency is almost entirely terroristic." (Schelling, 1960, 27) Terrorism and civil war are not separate phenomena and instead co-occur frequently (Bakker et al., 2016; Fortna, 2015). Only a small share of terrorist incidents are perpetrated by groups that are specialists in this tactic and terrorism is commonly used as one among many possible forms of violence within home territories (Tilly, 2004). In fact, it has been shown that\most incidents of terrorism take place in the geographic regions where civil war is occurring and during the ongoing war." (Findley and Young, 2012, 286) Within con ict zones, we observe large variation in the degree of overlap between terrorist attacks and events that are indicative of conventional 22 insurgent tactics. Consider the spatial distribution of terrorist attacks and war ghting by Boko Haram in Nigeria in 2014 (Figure 2.1). Terrorist incidents occur throughout the country. A signicant overlap between terrorist and non-terror violence can only be observed in the far Northeast|the region in which Boko Haram sought to established their rule. The micro-level separation of terrorist attacks and conventional war ghting can be explained by tactical choices of rebels. 3 Existing research suggests that when lacking the strength to ght the government directly, insurgents resort to terrorism. Figure 2.1: Spatial distribution of terrorist attacks and conventional war acts associated with Boko Haram in Nigeria in 2014. Data on terrorist attacks come from the Global Terrorism Database (GTD) data (START, 2016). Data on the location of conventional war acts come from the GED data (Sundberg and Melander, 2013). Once a dissident group has made the strategic choice between violent versus non-violent re- sistance, the question of terrorism versus non-terror violence is a matter of tactics (Bakker et al., 2016). The tactical choice for insurgents is to either attack states' armed forces directly or indi- rectly target the government via coercive action intended to spread fear among the public (Carter, 3 I use the term conventional war ghting with respect to tactics that are conventionally used by rebels in asymmetric civil wars, such as small battles with the government, ambushes, and hit-and-run attacks, not with regard to the usage of the term in international humanitarian law. 23 2015, 117). Polo and Gleditsch (2016, 816) state that while denitionally, the two concepts are not mutually exclusive and hard to delineate, \[t]errorism [:::] diers from conventional attacks in civil con icts in that the immediate targets or victims are typically non-combatants, and each individual victim is normally less important than the purpose of conveying a message to the in- tended audience." There is no single accepted way to dene terrorism, let alone distinguish it unambiguously from other forms of political violence. 4 I follow the previous literature in stipulat- ing three conditions that have to be met for a violent event to be coded as terrorism, rather than non-terror violence. To qualify as terrorism, the violent action must seek to convey a political message to an audience broader than the immediate targets of the attack, does not directly target the military capability of the state, and lie outside the realm of \legitimate warfare activities," including but not limited to, the targeting of noncombatants (START, 2016, Bakker et al. 2016; Chenoweth 2013). Territorial control is a key factor in explaining insurgents' use of terrorism as opposed to more conventional guerrilla tactics (Carter, 2015; de la Calle and S anchez-Cuenca, 2015). Terrorism arises from insurgents' inability to control territory (Asal et al., 2012). Rebel territorial control is associated with guerrilla tactics such as \hit-and-run attacks, ambushes, raids, and small-scale battles," however, when forced to remain underground, those same groups rely predominantly on bombings and assassinations (de la Calle and S anchez-Cuenca, 2015, 810). Tactical choices in civil war echo actors' maximization of benets and minimization of costs, subject to resource constraints and the actions of the opponent. All else equal, rebels prefer conventional tactics over terrorism for two reasons. First, in a quest to indirectly pressure the government by in icting pain and fear among the population, terrorist campaigns run a high risk of alienating civilians whose support rebels depend upon. Second, terrorism does not aid insurgents' immediate goal of securing territorial gains (Carter, 2015, 130). Terrorism is therefore a second-best choice of tactics for rebels that do not command control over a given area and are 4 See Asal et al. (2012). 24 unable to engage in direct ghting with government forces. Terrorism \allows dissidents to avoid direct, costly strikes on government forces that are typically superior in numbers and weaponry." (Hendrix and Young, 2014, 335) Territorial control is qualitatively dierent from rebel strength. Territorial control measures the degree to which rebels or the government rule over an area without interference from opposing actors. Rebel strength refers to the size of a group or its material capability. However, a group's military power and territorial control are related. Military power, the power to destroy, can be distinguished from coercive power, that is the power to hurt (Schelling, 1960). The less territory a group controls, the more it will rely on coercive, as opposed to military power (de la Calle and S anchez-Cuenca, 2015, 797). In environments characterized by low state capacity, armed actors are more likely to adopt conventional tactics, while groups facing more capable governments are likely to resort to terrorism (see de la Calle and S anchez-Cuenca in Asal et al. 2012, 482). The link between rebel tactical choice and territorial control can be observed empirically. Evi- dence from Nigeria suggests that once the government was able to re-capture insurgent strongholds in 2015, Boko Haram moved away from ghting for territory and intensied \its campaign of sui- cide bombings against soft targets." 5 Figure 2.2 overlays a coarse map of territorial control in Northeast Nigeria with the location of Boko Haram terrorist attacks (blue triangles) and con- ict events that are indicative of conventional ghting (red dots) within the two weeks following the measurement of territorial control. 6 The map shows conventional ghting to be clustered in contested areas and along the borders of insurgent-held territory. With the exception of isolated 5 https://reliefweb.int/report/nigeria/analysis-scrutinising-boko-haram-resurgence, accessed 18 Au- gust 2018. 6 The map is adapted from Reuters, see http://blogs.reuters.com/data-dive/2015/05/05/ mapping-boko-harams-decline-in-nigeria/, accessed 24 October 2018. To the best knowledge of the au- thor, at the time of writing, this is the most detailed information on territorial control that is publicly available for Nigeria at the height of the con ict. The map covers 32 local government areas in the Yobe, Borno, Adamawa states: Abadam, Askira/Uba, Bama, Bayo, Biu, Chibok, Damboa, Dikwa, Geidam, Gubio, Gujba, Gulani, Guzamala, Gwoza, Hawul, Jere, Kaga, Kala/Balge, Konduga, Kukawa, Kwaya Kusar, Madagali, Mafa, Magumeri, Maidugur, Marte, Michika, Mobbar, Monguno, Ngala, Nganzai, Shani. Gwoza is coded as being under rebel control on 24 April 2015 because it contains the Boko Haram stronghold in the Sambisa forest. 25 events in the border region with Cameroon in the West, terrorist events were limited to areas of government control. Figure 2.2: The map illustrates the relationship between territorial control and Boko Haram tactical choice in civil war in Nigeria in 2015. 26 2.3 Modeling territorial control I argue that the observed level of terrorism relative to conventional tactics is indicative of the unobserved distribution of territorial control in asymmetric civil wars. I translate a theoretical model of the relationship between rebel tactical choice and territorial control into a measurement model. The model rests on the insight that higher levels of rebel territorial control are associated with higher levels of conventional ghting, while higher levels of government control are associated with more terrorism. In areas of complete control of either actor, violent events will be scarce. In principle, territorial control could be operationalized along a continuum from full rebel to full government control. However, for the purpose of estimation via a discrete-state HMM below, I conceptualize territorial control as a categorical variable ranging from complete rebel (S1) to complete government control (S5), with levels of contestation in between (see Table 2.1). These states correspond to the categorization of zones of territorial control in the existing literature (Kalyvas and Kocher, 2009; Kalyvas, 2006). Territorial control Q Description S1 Full rebel control S2 Contested, closer to rebel control S3 Highly contested S4 Contested, closer to government control S5 Full government control Table 2.1: Set of possible states Q =fS1;S2;S3;S4;S5g of the latent variable territorial control. 2.3.1 Measuring rebel tactics Rebel tactics are observable emissions of the unobserved latent variable territorial control. I operationalize rebel tactics in a given area as a function of that area's relative exposure to terrorist attacks versus events that are indicative of conventional guerrilla ghting. This approach is 27 informative regarding the mixture of tactics used in a given area. I employ a heuristic that translates a function of the relative frequency of terrorist attacks T it and conventional war acts C it into values of the variable of observable emissions o it in area i at time t. Specically, I compare the probability of the observed exposure to terrorist events T it =P (bE [T] it c; [T] t ) to the respective probability of observed conventional ghting C it =P (bE [C] it c; [C] t ) from a zero-in ated Poisson distribution. E [T] it and E [C] it are continuous measures of an area's exposure to terrorist and conventional con ict events, respectively. [T] t and [C] t denote the expected number of events for each tactic in a given time periodt across all areasi within a country. There are four possible observations O =fA;B;C;Dg of the rebel tactic variable O, as outlined in Table 2.2. Tactics O Observation Description Comments o it =A E [T] it E [C] it 0 Little to no exposure to terrorism and conventional events. To account for coding errors and small margins, observed exposure values below a threshold xs, in the main specication xs = 0:1, are truncated to zero. o it =B jC it T it jm Similar non-zero exposure to terrorism and conventional ghting. In the main specication, overlap of zero-in ated Poisson probabilities of con ict exposure is specied as m = 0:025. o it =C C it < T it , and jC it T it j>m More exposure to terrorism than conventional ghting. o it =D C it > T it , and jC it T it j>m More exposure to conventional ghting than terrorism. Table 2.2: Heuristic to translate the observed exposure to terrorist attacks and conventional war acts into the categorical variable of rebel tactics O. I develop a continuous measure of areas' exposure to terrorist events E [T] it and conventional war ghting E [C] it . The in uence of individual con ict events on area i is modeled to dissipate continuously over space and time. I compute exposure as the sum of spatially and temporally weighted event counts for the centroid of area i at time t. 7 While the HMM computes the most 7 E it = P J j=1 w d ij wa jt , wherew d ij = 1=(1 +e 7+0:35d ij ) denotes weighted distances d ij from event j to the centroid of area i in kilometers, and wa jt = 1=(1 +e 8+2:5a jt ) weighted event ages a jt in months. Weighted distances or ages below w< 0:05 are truncated to zero. For more detail, see section 1 in the online appendix. 28 likely sequence of territorial control independently for each subnational area, the use of weights allows for spatial dependence in observed rebel tactics between spatial units. 2.3.2 Mapping rebel tactics onto territorial control Figure 2.3 illustrates the theoretical model how observed rebel tactics o it relate to unobserved levels of territorial control q it . The prevalence of the use of terrorism by rebels is theorized to be increasing in the level of government control. As previous research posits, \guerrillas resort to terrorist tactics when they act beyond their areas of control" (de la Calle and S anchez-Cuenca in Asal et al. 2012, 483). Hence, the use of terrorist tactics has an inverse relationship with the level of territorial control of that group in a given area. The use of conventional tactics is increasing in the level of rebel control | suggesting more direct confrontation and hence more conventional war ghting between the two actors. 8 Based on this theoretical model, zones of full rebel or full government control are associated with the relative absence of violence (o it =A). Control is either undisputed, or actors successfully established exclusive rule and prevent opponents from penetrating the area. Areas that are contested, but closer to rebel control, are expected to see relatively higher levels of conventional war ghting, as opposed to terrorism (o it = D). Here, rebels will limit their use of terrorism because they seek to reduce the amount of harm in icted on their constituent population in an eort to minimize the risk of denunciation and maximize popular collaboration (Polo and Gleditsch, 2016). In these areas, I expect a higher level of direct war ghting between insurgents and the government because when rebel control is high but not consolidated, government forces seek to confront rebels conventionally in a quest to regain control. Highly contested areas, that is regions where neither rebels nor the government command a high level of control, are characterized 8 The model is based on the simplifying assumption that there are only two parties to the con ict: a state and a non-state armed challenger that are both treated as unitary actors. This is of course a strong, and in many con ict settings unrealistic, assumption. However, because the unit of analysis in this project is a small grid cell, this assumption does not preclude an estimation of territorial control in con ict settings with more than one non-state armed actor, as long as groups' aspirations for control do not overlap signicantly. 29 Terrorist attacks Conventional fighting S1 Full rebel control S3 Highly contested S4 Contested, closer to government S2 Contested, closer to rebels S5 Full govern- ment control C it > T it C it ≈ T it ≠ 0 o it = D o it = B o it = C C it < T it o it = A E [T] it ≈ E [c] it ≈ 0 o it = A E [T] it ≈ E [c] it ≈ 0 Figure 2.3: Theorized relationship between observed variation in rebel tactical choice and levels of territorial control. by a relative parity of conventional and terrorist events (o it =B). Areas that are contested, but in which the government enjoys a high level of territorial control, are expected to exhibit relatively more terrorist attacks than conventional war ghting (o it = C), because insurgents ghting a highly capable government will substitute conventional war acts with terrorism (Carter, 2015). 2.3.3 Estimation I estimate territorial control via a Hidden Markov Model (HMM). 9 HMMs are graphical models that provide a method for uncovering the most likely sequence of unobserved states of a discrete latent variable given a set of observable outputs, transition, and emission probabilities. The ability to map four emission states onto a ve-category latent variable distinctly positions HMMs as an eective method to estimate the latent variable territorial control. 9 The model is implemented using the HMM package in R (Himmelmann, 2010). A computer science conference proceedings think piece discusses diculties of the estimation via HMM and extension to Hidden Markov Random Field Models (Anders et al., 2017). The present manuscript is the rst to develop a thorough theoretical model and compute HMM estimates of territorial control, made possible by accounting for the spatial dependence between grid cells through continuous spatial decay in observable emissions. 30 . . . Territorial Control 1 Territorial Control 2 Territorial Control t θ t =P(q t |q t-1 ) Transition probability φ t =P(o t |q t ) Emission probability Observed Hidden . . . Tactics 1 Hidden Tactics 2 Tactics t Figure 2.4: Graphical representation of a HMM as a Bayesian network. Figure 2.4 illustrates the conditional dependencies between the sequence of hidden states Q and the sequence of observed outputs O over time t. The model makes two assumptions. First, the value of the outputo it , that is the emission of celli at timet, is generated by a process whose true state q it is not observed. Second, the sequence of hidden states follows a Markov process, that is the state of the hidden variable at timet depends only on the state of the variable att 1, but no prior time periods (Ghahramani, 2001). The HMM computes the maximum likelihood path of territorial control over the entire period of observation. The most likely sequence of labels is decoded via the Viterbi algorithm. 10 While territorial control is operationalized as a ve-category variable Q =fS1;S2;S3;S4;S5g, the observable emissions only have four possible outcomes O =fA;B;C;Dg. A situation in which no terrorist attacks or conventional war acts are observed is theorized to be indicative of either full rebel or full government control. Hence, we cannot linearly map the observed indicator of rebel tactics onto territorial control. HMMs provide a solution to this problem via two unique 10 HMM maximizes the function vt(h) = max N g=1 v t1 (g) gh h (ot), where h indexes current state, g indexed previous state,v t1 (g) indicates the path probability of previous time step, gh denotes the transition probability fromqg toq h , and h (ot) the emission probability givenh. The Viterbi algorithm conducts the maximization step and via recursion returns the label for each unit of the most probable path (Jurafsky and Martin, 2017). 31 features. First, the model allows me to specify that the latent states of full rebel control (S1) and full government control (S5) are equally likely to produce an observable output of little to no violence (o it =A). Second, HMMs maximize the most probable path over the entire sequence of observations, not a single instance in time. This allows HMMs to sort out whether an area that experiences little to no violence is more likely to be under the full rebel or full government control, given transition probabilities, emission probabilities, as well as path probability of the previous time step. 2.3.3.1 Transition probabilities The transition matrix species the probabilities of an area transitioning from one latent state to another. Each cell in captures the probability of an area transitioning to a specic state of the latent variable territorial control q t , given its instance in the previous period q t1 . Conceptually, transition probabilities capture the volatility of territorial control. They are informative about assumptions on how the latent variable evolves over time, independent of the observed output. q t S1 S2 S3 S4 S5 q t1 S1 0.250 0.500 0.025 0.200 0.025 S2 0.250 0.150 0.075 0.500 0.025 S3 0.050 0.025 0.050 0.850 0.025 S4 0.025 0.075 0.150 0.125 0.625 S5 0.050 0.075 0.475 0.025 0.375 Table 2.3: Transition matrix , as inspired by Kalyvas (2006). Rows sum to one. Values across the diagonal indicate the probability of a grid cell remaining in the same state across two periods. I leverage existing research to construct the matrix of transition probabilities. 11 The transition matrix in Table 2.3 is inspired by observed transitions between zones of territorial control during the Greek civil war (Kalyvas, 2006, 277). In the original matrix, a number of possible transitions, for example from full rebel to full government control, are never observed. This would indicate a transition probability P (S5jS1) = 0. However, while areas are unlikely to transition from one 11 In principle, transition (and emission) probabilities could be learned from training data. However, in the present application this is infeasible because of a lack of publicly available ne-grained data on territorial control. 32 extreme on the spectrum of territorial control to another without at least temporarily experiencing contestation, it is not impossible. Therefore, the transition probabilities presented in Table 2.3 are modied from Kalyvas's empirical results to allow for all possible transitions between states to have non-zero probabilities. 12 2.3.3.2 Emission probabilities Emission probabilities guide the translation of observations into hidden states. The emission probabilities in Table 2.4 are derived heuristically. Each cell describes the probability of observing a specic output value o t given the true state of the latent variable territorial control q t . o t A B C D q t S1 0.600 0.175 0.050 0.175 S2 0.050 0.175 0.175 0.600 S3 0.050 0.600 0.175 0.175 S4 0.050 0.175 0.600 0.175 S5 0.600 0.175 0.175 0.050 Table 2.4: Matrix of emission probabilities used in the estimation of the HMM. Rows sum to one. The probability value in each cell of this matrix answers the following question: \Given that the true state of an area i at time t is, for example, S1, what is the probability of observing, for example, A from the data?" Here, I oer a brief justication for select probabilities in Table 2.4. If an area was in a state of complete rebel control S1, given the theoretical model, I do not expect to see any terrorist incidents or con ict events, hence observing evidence state o it = A has the highest emission probability with P (AjS1) = 0:6. However, since zones of complete rebel control in asymmetric con icts are expected to be rare, there is a non-zero chance of observing occasional ghting between the rebels and the government. Hence, the probability of observing similar levels of 12 Kalyvas's empirical transition matrix contains a row with transitions to a territorial control zone of value \0." No further explanation is given what this zone entails. Therefore, I spread the relative frequency of observations of a transition to zone 0 proportionately across zones 1 through 5. In addition, I make small adjustments in the numerical values to allow for a minimum transition probability of 2.5% between all possible states of territorial control. The overall patterns of possible transitions remain unchanged, see Figure 3 in the online appendix. 33 terrorism and conventional war ghting P (BjS1) = 0:175 or more conventional ghting than terrorism P (DjS1) = 0:175 when the true underlying state is complete rebel control are small, but not zero. The probability of observing more terrorism than conventional ghting in areas of complete rebel control is conceptualized to be extremely low with P (CjS1) = 0:05. For the latent variable state of contested territory with rebels having an advantage, S2, the most likely output to observe is o it =D with P (DjS2) = 0:6. If the situation is reversed and the government has the upper hand, the most likely observation is o it =C withP (CjS4) = 0:6. In case of high contestation without a clear advantage for either side, I expect to observe similar numbers of terrorist and conventional con ict events, such thatP (BjS3) = 0:6. Emission probabilities can be thought of as capturing our condence in the ability of the model to recover the true sequence of states. Here I assume that there is a 60% chance that given the true state of the latent variable territorial control q t , we observe the expected emission o t . 2.3.4 Data The unit of analysis is the grid cell-month. I leverage geo-coded event data to measure each area's monthly exposure to terrorist incidents and conventional war acts in hexagonal grid cells with a minimum diameter of 0.25 degrees (approximately 28km at the equator). Data on conventional war acts come from the GED version 17.1 (Croicu and Sundberg, 2017; Sundberg and Melander, 2013). To achieve the highest level of delineation between terrorist attacks and events that are indicative of conventional ghting between rebels and government, only observations that are cat- egorized by the GED as occurring within the realm of \state-based con ict" are considered. This excludes events that are classied as \non-state con icts" or \one-sided violence" directed against civilians. Further, only events that can be attributed to at least the second order subnational administrative division are included. Data on terrorist incidents come from the GTD (START, 2016). GTD codes whether there is any doubt that an event constitutes terrorism as opposed to other forms of violence, such as 34 conventional war acts or common crime. This variable is available from 1997 onward. I restrict my sample to GTD observations post-1997 that are unambiguously coded as terrorist attacks and that can be attributed to at least a second order administrative division. 2.3.5 Case selection The measurement strategy outlined above is applicable to cases characterized by a high power asymmetry in favor of the government. As a logical derivative of the theoretical framework, only rebels that are weak compared to the government should resort to terrorist tactics. Hence, I expect to observe signicant amounts of terrorism only when a high power asymmetry prevails throughout signicant stretches of the con ict. Data on troops ratios between 1997 to 2011 from Polo and Gleditsch (2016) suggest 37 intrastate con icts in which rebels are at most half as strong as the government | rendering them candidates for an estimation of territorial control via HMM. 13 Data to assess the validity of the estimates of territorial control in asymmetric civil wars are extremely sparse. In fact, it is the lack of ne-grained data on territorial control that motivates the development the new estimation strategy. As an initial proof of concept, I present estimates of territorial control for the con icts between the FARC rebels and the Colombian government from 2006 to 2017 and for the Boko Haram insurgency in Northeast Nigeria from 2008 to 2017. Colom- bia allows for an initial validation of the methodology by assessing the correlation of territorial control with deforestation in the aftermath of the 2016 peace agreement. Nigeria is included in the Armed Con ict Location and Event Data (ACLED) database, which allows for the construction of a coarse set of out-of-sample validation data on territorial control. 13 See Figures 4 and 5 in the online appendix. 35 2.4 HMM estimates of territorial control 2.4.1 Colombia The 2016 peace agreement between the Colombian government and the FARC ended a con ict that left approximately 250,000 people dead and displaced millions. 14 Colombia experienced decades of violence stemming from clashes between insurgent groups, paramilitaries, and ocial armed forces, as well as violence linked to and exacerbated by the presence of drug cartels in the country (Cort es and Montolio, 2014; Dube and Vargas, 2013). Figure 2.5 plots estimated levels of territorial control for the FARC guerrilla and the Colombian government from 2006 to 2017. 15 Graphed are annual averages of monthly territorial control estimates for 851 grid cells. I exclude the Amazon and Orinoco natural regions in the East because the low population density in these areas raises concerns over systematic under-reporting of con ict events. The shading of the grid cells in Figure 2.5 indicates estimated levels of territorial control from red denoting full rebel to blue denoting full government control. The maps demonstrate signicant spatial and temporal variation in territorial control, with a few persistent hot spots of rebel controlled areas along the border to Venezuela in the Northeast and the Southwestern coastal region. These patterns are corroborated by expert accounts of the con ict. Upon being elected to power in 2002, president Alvaro Uribe ignited a heavy military campaign against the rebels. \From 2002, the state saw a steady recovery of areas that had been under guerrilla control. [:::] Certain regions of the country, however, continued to exhibit high levels of violence, especially in the west and near the border with Venezuela." (Arjona, 2016, 92) 14 The Guardian, 23 June 2016, https://www.theguardian.com/world/2016/jun/23/ colombia-farc-rebel-ceasefire-agreement-havana, accessed 1 December 2017. 15 The full model is estimated on data from 1997 to 2017. Existing studies suggests that the guerrillas controlled signicant amounts of territory in the 1990s (Arjona, 2016, 91). At the time of writing, no information is available regarding the exact location of FARC strongholds in 1997. If such information became available, grid cells covering these areas could be initiated with a strong prior that suggests rebel control. Absent this information, all grid cells are initiated with at priors of 0.2 for all possible states of control. Based on my theory of rebel tactics, I expect little to no violence in guerrilla strongholds. This suggests that FARC control is likely underestimated in the late 1990s and early 2000s. The estimates are expected to be more valid after 2005 when the paramilitaries started to demobilize. 36 Figure 2.5: Estimated levels of territorial control in Colombia for hexagonal grid cells with a minimum diameter of 0.25 degrees, excluding the Amazon and Orinoco natural regions (N = 851). 37 The Colombian peace process provides a unique opportunity for out-of-sample validation of my estimates of territorial control. The signing of the peace agreement between the government and the FARC in 2016 introduced a sudden change in territorial control, because it forced rebels to disarm and abandon their strongholds. The timing of the eventual signing of the peace agreement in 2016 was plausibly unexpected, given the long history of failed peace negotiations between the government and the FARC. The unanticipated timing minimizes endogeneity bias when relating the pre- and post-peace dierences in proxy variables to the dierences in FARC territorial control. As a major source of revenue for the FARC, gold mining and coca production were heavily regulated within their territorial strongholds. It has also been reported that \guerrillas enforced strict limits on logging by civilians { in part to protect their cover from air raids by government warplanes." 16 Criminal organizations, bandas criminales (BACRIM), were quick to move into the power vacuum that the FARC left when it laid down its arms following the peace agreement. The BACRIM are reportedly less inclined to regulate mining, logging, and coca production and are instead intensifying these operations. This led to a rise in rates of deforestation, in particular in areas that were previously controlled by the FARC. Relating pre- and post-peace levels of forest cover to estimated changes in territorial control allows me to corroborate my estimates of territorial control. If my estimates of territorial control are valid, deforestation should be more likely in areas that saw larger changes in territorial control as a result of the peace accord. To test this, I estimate the following model. Deforestation i;t = 0 + 1 Control i;t + 2 Peace t + 3 (Control i;t Peace t ) + i Deforestation i;t is a dummy that equals one if a grid cell i experienced deforestation over the course of year t. Control denotes the change in the annual level of territorial control (averaged 16 https://www.theguardian.com/world/2017/jul/11/colombia-deforestation-farc, accessed 1 December 2017. 38 over monthly HMM estimates) between year t and t 1, with positive values indicating changes toward more government control. 17 Peace t is a dummy for the year 2016 in which the peace agreement was signed and FARC forces started to disarm. A positive coecient on the interaction between changes in territorial control and the the peace dummy Control i;t Peace t indicates a higher probability of deforestation in 2016 in areas that saw changes in territorial control as a result of the demobilization of rebel forces. Errors are clustered by grid cell. The probability of deforestation is estimated via logistic regression. Raster data on deforestation from 2013 to 2016 are obtained via the forest monitoring system from the Colombian Instituto de Hidrolog a, Meteorolog a y Estudios Ambientales (IDEAM). For each grid cell in each year, I code a binary indicator whether the cell experienced deforestation. In 2013, 26 grid cells saw their forest cover reduced. This number jumped to 51 cells in 2016. Table 2.5: Relationship between rebel territorial control and deforestation in Colombia. Deforestationi;t (1) (2) (3) Controli;t Peacet 3.59 3.29 (1.68) (1.63) Controli;t 0.63 1.51 1.56 (1.00) (0.97) (0.86) Peacet 0.29 0.22 0.07 (0.16) (0.16) (0.18) Deforestationi;t1 0.86 (0.28) Constant 3.03 3.04 2.94 (0.10) (0.10) (0.11) Observations 3,404 3,404 2,553 Log Likelihood 670.74 668.87 545.38 Akaike Inf. Crit. 1,347.48 1,345.73 1,100.75 Note: p<0.05; p<0.01; p<0.001 Logistic regression coecients with bootstrapped clustered standard errors by grid cell in parentheses. 17 To compute annual averages and changes in territorial, I recode the discrete territorial control variable with 0 indicating full rebel control, 0.25 indicating zones that are contested but closer to rebel control, 0.5 denoting highly contested areas, 0.75 indicating contestation with the government having the upper hand, and 1 signifying full government control. 39 The results in Table 2.5 support the hypothesis that changes in territorial control as a result of the peace agreement are associated with a higher probability of deforestation. The coecient on the interaction between changes in territorial control and the timing of the 2016 peace agreement in model 2 is positive and statistically signicant at the minimum 5% level. A change of 0.25 in the annual level of territorial control, for example that from an area that is on average contested but closer to government control to one that is on average fully controlled by the state, is associated with an increase in the predicted probability of deforestation from 5.6% to 9.1%. The coecient for the interaction remains statistically signicant upon including a lagged dependent variable in model 3. Thus, HMM estimates of territorial control produce results in line with observed empirical relationships between changes in territorial control and deforestation in Colombia. 2.4.2 Nigeria To demonstrate the applicability of the approach across country contexts, I provide monthly-level estimates of territorial control for Northeast Nigeria. The area has seen sustained ghting between the Nigerian government and Boko Haram since 2009. 18 In late 2014 and early 2015, reports suggest that the insurgents controlled 15 localities in the northeast border region with Cameroon, and had partial control over additional 15 local government areas. 19 In early to mid 2015, the Nigerian government and African Union multilateral troops launched attacks against Boko Haram that caused the insurgents to lose a majority of the territory they previously controlled. 20 However, recent reports cast doubt over the government's claim that it drove the insurgents out of the region, 18 See https://www.nytimes.com/2014/11/11/world/africa/boko-haram-in-nigeria.html accessed 10 Septem- ber 2018. 19 See https://www.amnesty.org/en/latest/news/2015/01/boko-haram-glance/, accessed 27 September 2018. Estimates suggests that the total area controlled by the insurgents amounted to approximately 20,000 square miles | approximately the area of Belgium (see https://www.telegraph.co.uk/news/worldnews/africaandindianocean/nigeria/11337722/ Boko-Haram-is-now-a-mini-Islamic-State-with-its-own-territory.html, accessed 27 September 2018). 20 http://web.stanford.edu/group/mappingmilitants/cgi-bin/groups/view/553?highlight=boko\%2Bharam, accessed 27 September 2018. 40 with an account from May 2018 suggesting that Boko Haram continues to control parts of Borno state via roadblocks, stop and search operations, and the collection of \taxes" for protection. 21 Figure 2.6 illustrates yearly averages for monthly-level HMM estimates for 15 Northeast Nige- rian states from 2009 to 2017 for grid cells whose centroids are 0.25 degrees apart (N = 942). 22 The estimates illustrate the onset of the Boko Haram insurgency in 2009. 23 Starting in 2011, Boko Haram is estimated to gain temporary areas of control in Borno state. The insurgents sub- sequently establish persistent strongholds that are at rst scattered throughout the study region. By 2014, the estimates show the consolidation of Boko Haram control in the Northeast. Following the oensive of the Nigerian government and the African Union (AU), Boko Haram is estimated to have lost signicant amounts of territory between 2015 and 2016. By 2017, its strongholds are limited to a few areas in the border region with Cameroon. The inclusion of Nigeria in the ACLED data version 8.0 allows me to construct an out-of- sample validation data set (Raleigh et al., 2010). Through event type labels, ACLED contains information on whether an event resulted in rebels gaining control or establishing a base (coded as S1, continuous value 0), battles with no changes in control (S3, continuous value 0.5), the government gaining control or establishing a base (S5, continuous value 1), and instances of remote violence (S2 for government remote violence with a continuous value 0.25; S4 for insurgent remote violence mapped to continuous value of 0.75). 24 The events are aggregated to grid cells on a monthly level. Territorial control is assigned based on the occurrence of control-related events within a grid cell. New events cause a cell to update the coded level of territorial control based on event type. In the case of multiple events occurring in the same grid cell month, I average across the continuous values that are indicative of the status of territorial control associated with 21 https://www.dw.com/en/boko-haram-islamists-still-control-parts-of-northeastern-nigeria/ a-43851013, accessed 27 September 2018. 22 The coverage of the HMM results mirrors the 15 states included in the study by Aronson et al. (2017) that are most subjected to Boko Haram violence, namely Adamawa, Bauchi, Benue, Borno, Gombe, Jigawa, Kaduna, Kano, Katsina, Nassarawa, Niger, Plateau, Taraba, Yobe, and the Federal Capital Territory. 23 The model is estimated starting in 2008 and initiated with a strong prior of full government control in the rst month of observation. 24 ACLED contains a small number of events for which manual coding is necessary to determine the actor gaining control, in particular for occurrences of remote violence. The respective documentation is available upon request. 41 Figure 2.6: Estimated levels of territorial control in Northeast Nigeria for hexagonal grid cells with a minimum diameter of 0.25 degrees (N = 942). 42 each event. Cells that experience zero or one event over the entire period of observation 2008 to 2017 are assumed to be under full government control. Similarly, if a cell does not experience any violence in the previous six months, it is assumed to be under government control. 25 I create two sets of validation data. The rst validation set adopts the assumption that remote violence is indicative of areas that are contested but closer to either rebel or government control, depending on the perpetrator (denoted \full sample" below). The second data set drops this assumption and considers only ACLED events that make explicit reference to changes in territorial control (denoted \restricted sample" below). 26 Due to the strong assumptions necessary to construct a testing set from ACLED as well as concerns regarding reporting error in these data (Eck, 2012), conclusions from a comparison with the HMM estimates should be taken with a grain of salt. For example, based the media reports above, the validation data likely underestimate the extent of Boko Haram control in 2014. However, the testing set constructed from ACLED data oers the best opportunity for out-of-sample validation of territorial control in Nigeria available to date. Figure 2.7: Spearman's rank-order correlation coecients for a annual-level correlations between HMM estimates and the ACLED validation data. 25 The con ict in northeast Nigeria is highly active during the period between 2009 and 2017 | rendering six months a reasonable upper bound for the stationarity of control in a given cell. 26 Event types in the restricted sample are the government gaining control or establishing a base, Boko Haram gaining control or establishing a base, and battles with no changes in control. Figure 6 in the online appendix plots yearly averages for both validation data sets. 43 Figure 2.7 plots annual Spearman's rank-order correlation coecients between monthly terri- torial control estimates and the ACLED validation data. The correlation ranges between 0.21 and 0.35. The validation data that relies on instances of remote violence to indicate areas of partial control of either Boko Haram (S2) or the government (S4), on average shows a higher correlation with the HMM estimates. Compared with the validation data, the HMM appears to be overes- timating the level of rebel complete control and to be underestimating the level of contestation between the government and Boko Haram. However, the HMM estimates in Figure 2.6 uncover general spatial patterns and major temporal trends in the distribution of territorial control for the government and Boko Haram in the Northeast of the country. In particular the ability of the model to capture the severe reduction of insurgent territorial between 2015 and 2016 is striking. 2.5 Conclusion I propose a novel measurement approach for the estimation of territorial control in asymmetric civil wars. I leverage observed variation in the co-occurrence of terrorist attacks and conventional ghting within a machine learning framework to obtain estimates of the latent variable territorial control. As a proof of concept, I present estimates of actors' control over territory for the ght between the FARC and the Colombian government and the Boko Haram insurgency in Nigeria. A validation of the Colombia estimates using patterns of deforestation in the aftermath of the 2016 peace agreement suggests that the model is able to recover general trends in the evolution of territorial control across time and space. Newly developed validation data for the Boko Haram insurgency in Nigeria show that the estimates correlate reasonably with alternative measures of territorial control. The results demonstrate that Hidden Markov Models (HMM) are a fruitful approach to ad- dress the lack of data on territorial control in asymmetric civil wars. The methodology outlined in this paper allows for the generation of territorial control estimates that utilize publicly available 44 con ict event data, are easy and fast to estimate, capture ne-grained spatiotemporal subnational patterns, and can be computed for a cross-section of countries. The estimates thus yield a valuable source of information for subnational analyses of civil war dynamics both for within-country and comparative studies. As an example, cross-nationally available estimates of territorial control are crucial for enhancing our understanding of how belligerents' local provision of public goods inter- acts with their level or territorial control. The estimates will also be instrumental for investigating cross-border contingencies in asymmetric civil wars. Rebel groups like Boko Haram operate across international borders and often leverage the remoteness of border regions to their advantage upon gaining strongholds beyond the reach of their primary enemy's armed forces. HMM estimates are produced with a methodology allows for such localized cross-country analyses. The derivation of the transition probabilities that specify prior beliefs regarding the temporal evolution of territorial control illustrates how existing empirical or theoretical research can be leveraged to inform model parameters in machine learning applications. All parameters can be adapted based on additional cross-sectional information or case-specic knowledge. As an example, the Nigeria model is initiated with a strong prior for complete government control because the start of the Boko Haram insurgency in 2009 is captured by the data. In other con ict settings, one could initiate the model with a prior of full rebel control for grid cells that are known insurgent strongholds in the rst period of observation. Future work should also explore the inclusion of covariates such as terrain or landcover to introduce heterogeneity across grid cells regarding our expectation how likely rebels are able to capture and sustain a territorial base. The measurement strategy incorporates a new method to gauge subnational areas' exposure to con ict events. Rather than discretely assigning events to grid cells, I allow the impact of violent events for a given area to dissipate continuously over space and time. Con ict exposure is computed as the sum of spatially and temporally weighted events. HMMs estimate territorial control for each grid cell individually. They explicitly model time dependencies but do not account for the correlation of the latent variable between close spatial units. The continuous measure 45 of areas' exposure to con ict events captures spatial dependency in observed rebel tactics and thus overcomes a major limitation of standard HMMs. This methodological innovation renders HMMs a suitable tool to estimate latent constructs that feature spatial and temporal variation in subnational-level research. 46 2.6 Acknowledgements A version of the article presented in this chapter is forthcoming in Journal of Peace Research, as part of a special issue \Innovations in concepts and measurement for the study of peace and con ict," edited by James Lo (University of Southern California, USC) and Christopher Fariss (University of Michigan). An earlier version of this manuscript received the International Studies Association (ISA) Peace Studies section Kenneth Boulding Award for the best graduate student paper presented at the 2018 Annual Convention. I am grateful to the audiences at meetings of New Faces in Political Methodology XI, the Peace Science Society, the Society for Political Methodol- ogy, ISA, ISA West, ISA-FLACSO, American Political Science Association, the USC Networked Democracy Lab and USC CIS Working Paper Series, and to Lisa Argyle, Adam Badawy, Pablo Barber a, Sam Bell, Charity Butcher, Christopher Fariss, Michael Findley, Benjamin Graham, Joseph Huddleston, Kyosuke Kikuta, Satish Kumar, Brett Ashley Leeds, James Lo, Jonathan Markowitz, Ashly Townsen, James Walsh, Hong Xu, and Xin Zeng for helpful comments and discussion. 47 Appendix 2.A Measuring con ict exposure Spatially and temporally disaggregated con ict event data is a key source of information in the study of subnational violence. However, many covariates of interest operate on an areal level, for example economic wealth, terrain, the ethnic composition of the population, the availability of natural resources, or the provision of public services. Individual con ict events are thus typically aggregated to grid cells or administrative units such as districts or municipalities to match the unit of measurement of the covariates and to measure the exposure of these subnational areas to con ict events. Each con ict event is commonly assigned discretely to the subnational area within which it is located. This standard practice is problematic for two main reasons. First, scholars are frequently only accounting for events that fall within the boundaries of a chosen subnational unit and do not account for events that happen in the vicinity. Second, inferences regarding an area's exposure to con ict are highly sensitive to the drawing of boundaries | widely cited in the literature as the Modiable Areal Unit Problem (MAUP). A similar issue arises in the temporal domain when con ict events are discretely assigned to the calendar month or year in which they occurred. An event's in uence on local con ict dynamics is unlikely to abruptly stop at the chosen spatial or temporal boundaries; nor will it homogeneously aect the entirety of the space. Rather, its impact dissipates continuously over space and time. A simplistic approach to coding areas' exposure to terrorism con ict events would sum the number of events that fall within a given grid cell. This procedure faces the problem that the 48 assignment of con ict events to grid cells is highly dependent on the sampling of centroid loca- tions. MAUP describes the discrepancy between real world spatial patterns of events and patterns created via aggregation of events into homogenous spatial units (Openshaw and Taylor, 1979). Shifting the location of the centroids can have a severe in uence on the number of events that are assigned to a particular cell. This is particularly concerning when the drawing of grid cell boundaries leads clusters of events to be broken up into smaller groups | causing the relative frequency of terrorist events and conventional war acts to change dramatically. Figure 2.8 il- lustrates this issue. Based on the location of grid cell centroids in panel A, we would code the relative frequency of rebel and conventional war ghting to correspond to the values of the variable Tactics it = [D;D;A]. If the centroids were shifted by 25% relative to the location of the events, we would conclude the emissions of these three cells to have values of Tactics it = [B;D;C]. Shifting centroids by 25% C 1 1T, 0C: Tactics it = D C 2 5T, 4C: Tactics it = D C 3 0T, 0C: Tactics it = A C 1 2T, 3C: Tactics it = B C 2 3T, 0C: Tactics it = D C 3 1T, 1C: Tactics it = C C 1 C 1 C 2 C 2 C 3 C 3 A B Figure 2.8: The schematic illustrates how shifting the location of the grid cell centroids from their original (randomly sampled) location (panel A) by just 25% (panel B) can result in vastly dierent conclusions about the coding of rebel tactics. Red dots indicate the location of conventional events; blue triangles those of terrorist attacks. This is a simplied example | in the analysis, Tactics it is computed using probabilities from Poisson distributions under application of a margin parameter. To alleviate this problem, I propose a novel measurement model for rebel tactics in civil war that uses spatial and temporal weights to associate con ict events with grid cells rather than 49 relying on discrete assignment. The importance of individual violent events for the estimation of territorial control decreases over time and space. I model this intuition by assigning space- and time-varying weights to each event. 2.A.1 Spatial and temporal decay functions For each grid cell centroid-monthc it ,i = 1;:::I indexes centroids andt = 1;:::T indexes months. For each con ict event e jm , j = 1;:::J indexes individual events and m = 0;:::M indexes the calendar month in which the event occurred. Let lon i and lat i denote the longitude and latitude of each grid cell centroid c i in radians, respectively. Similarly, let lon j and lat j denote the longitude and latitude of each con ict event e j in radians. Then the spatial distance d ij in km between centroid c i and event e j is computed as the geodesic distance between two points using the Haversine formula, d ij = 2r arcsin s sin 2 lat j lat i 2 + cos(lat i ) cos(lat j ) sin 2 lon j lon i 2 ! ; where r 6371 denotes the earth mean radius in kilometers. The temporal distance (in the following called age)a tm =tm measures the months between when evente jm occurred and the time of observation of the grid cell-month c i t. An event occurring in the month of observation has an age of a tm = 0, while an event that occurred four months ago has an age of a tm = 3. For each centroid-month unit c it I measure the spatial distance d ij in km and the temporal distance a tm to each con ict event e jm , resulting in a total number of centroid-event-month observations of size K = IJT . Specically, for each grid cell c i in each month t, I create a vector D of spatial distances and a vector A of temporal distances to each event. Events that occur in the future from the time of observation t (i.e. where u > t) receive a missing value. I then weight both vectors to allow the impact of con ict events on grid cells to dissipate over space and time. 50 I assume the impact of an event to dissipate following a logistic decay function of the general form w = 1 1 +e + x ; where x denotes the decaying quantity (here event age or distance between the event and a centroid), determines the slope of the curve and denes its in ection point. To describe a decay function, both the slope parameter and he in ection parameter have to be positive real numbers. To model spatial decay, assume the slope parameter to be d = 7 and the in ection parameter to be d = 0:35. To model temporal decay in months, I use a steeper sigmoid curve. I assume the temporal slope parameter to be a = 8 and the in ection parameter to be a = 2:5. 27 Figure 2.9 plots the decay functions using the parameter values above. Based on the shape of the logistic decay functions above, an event that occurs at the location of the centroid of a grid cell receives a spatial weight of 1. An event that occurs 10km away from the centroid receives a weight of 0.97 and an event 25km away is weighted by 0.15 | after which its in uence tends toward 0. The temporal weight features a dierent rate of decay. In the rst month, the event receives a temporal weight of 1, followed by 0.95 in the second, 0.62 in the third, and 0.11 in the fourth month; after which the weight approaches zero. The exposure of grid cell c it to con ict events E it is computed as the sum over all temporally and spatially weighted events J. E it = J X j=1 w dij w ajt (2.1) Thus, the resulting unit of observation is the grid cell-month, i.e. a vector of exposure values of size E =IT . 27 Figures 2.10 and 2.11 illustrate the shape of the logistic decay functions for alternative parameter values. 51 (a) Spatial (b) Temporal Figure 2.9: Logistic function that describes the decay of the in uence of an event in relation to a centroid in the spatial and temporal dimensions. Figure 2.10: In uence of the slope parameter and in ection parameter on the shape of the logistic decay curve for an example of distances varying from 1km to 150km. 52 Figure 2.11: In uence of the slope parameter and in ection parameter on the shape of the logistic decay curve for an example of event ages from 0 to 12 months. 53 2.A.2 Comparison between discrete and continuous aggregation Figure 2.12 presents a visual comparison between the discrete and continuous aggregation of events to hexagonal grid cells with a minimum diameter of 0.25 and 0.5 degrees, using simulated data. 28 Note that for illustration, the aggregation of events is performed only in the spatial dimension. No temporal weight is applied. Dots denote the location of events. The shading of cells indicates the computed con ict exposure value for the discrete assignment of events to cells (rst row), the continuous assignment with a logistic decay curve that falls below a weight of 0.9 at a distance of approximately 14km (second row), and continuous assignment that dips below 0.9 at approximately 36km (third row). The plot in Figure 2.12 highlights the shortcomings of the discrete assignment approach. For grid cells of size 0.25 degree and 0.5 degree, discrete aggregation leads to cells that are neighboring con ict hotspots to receive a zero exposure value. From the plot it is also evident that the same underlying locations would have received a vastly dierent value of con ict exposure if the centroids of the cells were moved a little, or the cell size was increased or decreased. It is upon the researcher to decide the appropriate rate of decay of con ict events | a choice that is likely context and application specic. Figure 2.12 illustrates that the choice the rate of decay is highly in uential for the resulting computed con ict exposure value. Applying a rate of decay in that dips below 0.9 at approximately 36km (third row) leads to a sum of weighted events that is signicantly higher than the maximum count in the discrete assignment case. 29 Moreover, the maximum con ict exposure max(E i ) = 13:1 for the 0.25 degree resolution grid is computed for a cell that itself does not experience any con ict events within its borders. This cell's high con ict exposure value is caused by the abundance of con ict events in neighboring cells that, given the slow rate of spatial decay, accumulate in the centering cell. The ability of a con ict event to contribute to the total con ict exposure beyond grid cell borders is in principle a desired 28 Event locations are a subset of the simulated data from Schutte and Donnay's (2014) mwa R package. 29 The maximum values for the discrete aggregation in this example are max(E i ) = 5 for the 0.25 degree grid and max(E i ) = 7 for the 0.5 degree resolution. 54 feature of the methodology to overcome the MAUP. However, as shown in Figure 2.12, a slow rate of decay leads to a potential over-aggregation of con ict events. Once the appropriate rate of decay is determined, the cells size appears to be somewhat less in uential for the computed con ict exposure of underlying locations. This is not to say that cell size does not matter when using spatial or temporal weights to compute con ict exposure. The fact that cells feature homogenous con ict exposure within their boundaries means that larger cells will mask heterogeneity in the intensity of violence in the underlying locations. However, cell size appears to introduce a smaller bias when impact of events on grid cell centroids is allowed to dissipate continuously, as compared to the discrete aggregation case. For the sample of event locations in Figure 2.12, the slope parameter = 7 and in ection parameter = 0:35, combined with the 0.25 degree grid resolution, appear to be an appropriate choice. The conservative rate of decay prevents an over-accumulation of events. Since events can impact locations that are not contained within the bounds of the grid cells within which they occur, cells that do not themselves experience events but are in the vicinity of con ict hot spots receive a non-zero con ict exposure value. These parameter and resolution choices thus strike a balance between preventing an over-accumulation of events, yet minimizing the bias that a re-sampling of grid cell centroids would cause on the resulting con ict exposure values. Future research should explore a relaxing of the assumption of a homogenous rate of decay for all con ict events and in all locations. The methodology could be extended by incorporating covariates that operate at the event-level or the level of the underlying subnational areas. For example, one could assume a dierent rate of decay for large-scale terrorist attacks versus instances of communal violence. Similarly, the rates of decay in the spatial and/or temporal dimensions could be computed as a function of population density or terrain. 55 Figure 2.12: Comparing the dynamics of continuous versus discrete aggregation of events in the spatial dimension using simulated data. 56 2.B Estimation procedure For each grid cell i, the following procedure is used to estimate the most probable sequence of territorial control over all time periods t. 1. Compute the exposure of the grid celli in montht to terrorist eventsE T it and to conventional war acts E C it to all events J, by (a) computing the spatial distance d ij of each event j to the centroid of grid cell i in kilometers and weighting it using a logistic decay function, (b) computing the temporal distance (each event's age)a jt between the monthm when the event occurred and the time of observation of the grid cell t in months and weighting it using a logistic decay function (note that only positive temporal distances a jt are considered), and (c) summing the product of the spatial and temporally weighted distances for terrorist and conventional events for each grid cell-month to arrive at E T it and E C it . Note that spatially- and temporally weighted sums of under 0.05 in a grid cell-month are set to zero to avoid later grid-cell months having in ated cumulative event exposures. 2. For each grid cell i in each month t, determine the value of the variable o it =f E T it ;E C it . 3. For each grid celli in each montht, create a sequence of observed outputs O2fA;B;C;Dg, where an individual observation o it is determined by Tactics it . 4. For each grid cell i compute the most probable sequence of latent states Q 2 fS1;S2;S3;S4;S5g over all time periods t, given the sequence of observed indicator of rebel tactics O over all time periods t, the time-invariant matrix of transition probabilities , and the time-invariant matrix of emission probabilities via a Hidden Markov Model (HMM). 57 2.C Summary statistics for deforestation model Table 2.6: Summary statistics for the logistic regression model of deforestation in Colombia on changes in territorial control as a result of the 2016 peace agreement. The unit of analysis for territorial control is annual averages of monthly-level estimates for 0.25 degree hexagonal grid cells. Statistic N Mean St. Dev. Min Pctl(25) Pctl(75) Max Controli;t 3,404 0.9783 0.0755 0 1 1 1 Controli;t 3,404 0.0062 0.0895 1 0 0 1 Peacet 3,404 0.2500 0.4331 0 0 0.2 1 Deforestationi;t 3,404 0.0496 0.2172 0 0 0 1 58 2.D Additional gures 2.D.1 Transition probabilities Figure 2.13: The gure compares the distribution of transition probabilities between the em- pirical observations from Kalyvas (2006) and the modied transition probabilities in this paper. The graph shows that while the transition probabilities dier slightly, the patterns of transitions between states from t 1 to t remain unchanged. 59 2.D.2 Case selection Figure 2.14: Number of cases that the measurement strategy can be applied to based on dierent thresholds of power asymmetry between the rebels and the government. Data on power asymmetry come from Polo and Gleditsch (2016). 60 Figure 2.15: The graph illustrates the selection of cases for which the measurement strategy is applicable based on thresholds in average and maximum rebel-to-government troop ratios over the course of the con ict. Plotted in red are cases that would be included based on a 0.5 threshold indicating rebels that are half as strong as the government forces. Future work will investigate the determination of the most appropriate threshold. 61 2.D.3 ACLED validation data for Northeast Nigeria (a) Full set of ACLED event types (b) Subset of ACLED event types (more restrictive) Figure 2.16: Yearly averages of monthly-level ACLED validation data values. Values that are closer to 0 indicate full rebel control; values closer to 1 full government control. 0.5 indicates cells that are highly contested. 62 Chapter 3 How does Insecurity Aect Government Welfare Spending? Insights from Colombian Municipalities 3.1 Introduction `Winning the hearts and minds' of people is a widely-used dictum for the insight that governments in insurgent con icts are unlikely to prevail on the battleeld alone, and instead need to gain the support of local population. `Winning hearts and minds' has become a virtually universally accepted policy recommendation for governments facing an insurgent challenger. The high level of acceptance of the strategy has tempted many researchers to use the concept of `winning hearts and minds' as a blanket statement for all theoretical mechanisms that aim to capture government attention to citizen needs in times of con ict. So little attention has been paid to the details in the application of the strategy, that `winning hearts and minds' is at the verge of becoming a truism in the literature on peace and con ict. In this study, I develop and test a set of falsiable propositions that guide the investigation of how governments employ non-coercive investments in citizen welfare as strategic complements to coercive con ict management. Recent research establishes a robust positive relationship between a government's welfare- mindedness and the reduction of the incidence or intensity of violence (Crost et al. 2016; Dasgupta et al. 2015; Cort es and Montolio 2014; Taydas and Peksen 2012; Beath et al. 2012; Berman et al. 63 2011; Burgoon 2006). Yet, these works only establish that welfare-oriented spending can reduce or prevent violence, not when governments are likely to employ it. The conditions under which governments decide to invest in social welfare as a con ict management tool remain opaque. Not only are we lacking insight into the underlying strategic rationale of governments, the dearth of research regarding the decision to invest in social welfare exposes the nding of a positive relation- ship between welfare investments and reduced violence to a potential selection bias: Governments might use investments in social welfare as a con ict management tool predominantly in cases in which the promise of success is particularly high. Governments are acutely aware of the interconnection between security and development, and some explicitly recognize welfare investments as a strategic con ict management tool. Referring to the Maoist insurgents, the then prime minister of India, Manmohan Singh, said: We must, however, recognize that naxalism is not merely a law and order issue. In many areas, the phenomenon of naxalism is directly related to underdevelopment. [. . .] Our strategy, therefore, has to be to `walk on two legs' to have an eective police response while at the same time focusing on reducing the sense of deprivation and alienation." 1 However, not all policymakers prioritize welfare investments in times of con ict. In 2004, at the height of the most recent wave of violence in the Colombian civil war, then president of Colombia, Alvaro Uribe, exhibited opposite priorities: Of course we need to eliminate social injustice in Colombia ... but what is rst? [:::] Peace. Without peace, there is no investment. Without investment, there are no scal resources for the government to invest in the welfare of the people." 2 1 Former Prime Minister of India, Dr. Manmohan Singh: \PM's speech at the Chief Minister's meet on Naxal- ism." http://archivepmo.nic.in/drmanmohansingh/speech-details.php?nodeid=302, accessed 9 March 2016. 2 BBC News: \Uribe defends security policies." http://news.bbc.co.uk/2/hi/americas/4021213.stm, accessed 10 March 2016. 64 Uribe's statement illustrates that investing in social welfare is not a universally accepted con- ict management tool. Singh highlighted the necessity of a parallel employment of welfare- and security-related strategies. Uribe, on the other hand, endorsed a clear prioritization: Focus on security rst, then development will follow. 3 Uribe's prioritization is puzzling in light of the academic research on the security-development nexus. If development programs and public good provision are as successful in reducing violence as academic scholarship suggests, why are not all governments whole-heartedly sending teachers, doctors, and social workers along with their tanks and police forces to the areas of persistent insecurity? This puzzle motivates the central question in this paper: What explains variation in the degree of a government's welfare-mindedness in light of insecurity? Social welfare is a fundamentally citizen-centric aspect of domestic policy. My argument builds on the conjecture that in order to succeed in ghting insecurity, governments employ welfare spend- ing to gain the support of the population. Welfare spending has a strong redistributive component and reduces con ict via improving the citizens' living conditions, thereby alleviating grievances that motivate rebellion and collaboration with the armed challenger. I argue that governments are not benevolent spenders and instead employ welfare spending as a highly strategic tool to ensure their survival in cases when their fate is intimately linked that of the civilian population. The eect of insecurity on welfare spending should be most pronounced when governments are expecting a strategic benet from civilian cooperation and when civilians are caught in the front lines between the government and an armed challenger. I investigate the strategic use of non-coercive resources in civil war using subnational data from Colombia. Civil con icts experience a high degree of spatiotemporal variation regarding the intensity of con ict. In many civil wars, parts of the country are in complete turmoil while other 3 Please note that Uribe's statement pertains only to the strategy of the national government of Colombia at a time when the presence of armed non-state actors in the country was at the highest level in the 21st century. It does not necessarily translate to the strategies of the municipalities, the unit of analysis in this study, who after major decentralization reforms since the 1980s, are chie y responsible for the provision of public services such as education, health care, and water and sanitation. 65 areas experience a relative absence of ghting. An investigation at the national, as opposed to the subnational, level of analysis would likely mask governments' use of welfare provision as a reaction to localized violence. Colombia provides a suitable testing ground for two reasons. First, Colombian municipalities enjoy a considerable amount of discretion and spending power with regard to citizen welfare | causing the main source of variation in welfare spending to operate at approximately the same geographical level as variation in subnational con ict dynamics. Second, Colombian municipalities oer a high quality and quantity of data on both welfare investments and con ict dynamics, in particular in comparison with other con ict-ridden countries in the Global South. While the focus on a single country limits the generalizability of the ndings, Colombia provides a unique opportunity for the testing of the theoretical propositions developed in this study. Using data from over 1,100 Colombian municipalities for the period 1994{2010, I investigate how dierent types of insecurity in uence expenditure patterns on three categories of public goods: education, health care, and water/sanitation services. Prior to the signing of the peace accord in 2016, Colombia experienced decades of intrastate con ict between the armed forces, guerrilla groups, and paramilitaries. In addition, the country has endured one of the highest homicides rates in in the world. I test my argument employing four sets of indicators measuring dierent facets of insecurity: the overall intensity of violence, violence directed against armed government forces, violence directed against political agents of the government, and instances where civilians are the victims of the ght between the government and an organized armed challenger. I nd that governments use welfare spending in a highly strategic way. The results support the hypothesis that government investment in social welfare varies as a function of the size of the threat that an armed challenger poses to the survival of the government. This threat, and therefore the incentive to increase welfare spending, is higher in cases where government actors are under immediate attack, as measured for example through the number of fatalities among armed forces or oensives launched by non-state actors. The threat, and consequently welfare 66 spending, is also higher when civilians | whose intelligence and support is crucial for the success of a government's counterinsurgency campaign | are caught in the crossre; such as in cases attacks against the civilian population. This paper makes two major contributions to the existing literature. First, it opens up the black box of the `winning hearts and minds' concept established in existing research by asking not whether, but under which conditions, governments resort to these strategies. Second, it engages in theory renement by developing alternative mechanisms that might motivate governments to utilize welfare spending as a con ict management tool, and testing these mechanisms empirically. The results of this research allow us to better understand the motivations that drive government public good provision in con ict zones. 3.2 Insecurity and Non-Coercive Government Action This paper is motivated by two ndings in the recent literature on the security-development nexus that tackle the relationship between violence and public good provision from two contrasting causal perspectives. On one hand, existing research documented the detrimental eects of violence on social welfare, including but not limited to, education (Oyelere and Wharton 2013; Rodr guez and S anchez 2012; Wald and Bozzoli 2011) and health care (Sexton et al. 2019; Iqbal 2006). On the other hand, scholars nd increasingly robust evidence that improvements in social welfare can have a preventative eect for the onset of civil war (Taydas and Peksen 2012; Thyne 2006), reduce the incidence of transnational terrorism (Burgoon 2006), and decrease incidents of riots (Justino 2015) or levels of violence in ongoing intrastate con icts (Crost et al. 2016; Dasgupta et al. 2015; Cort es and Montolio 2014; Beath et al. 2012; Berman et al. 2011). Recent work also establishes a link between the preferential access to welfare-related material goods such as electricity and a reduction in the likelihood of violent opposition to the regime (De Juan and Bank 2015). The 67 provision of welfare-related public goods such as education or health care can thus become a strategic asset in violent con icts. The ndings of these two literatures suggest that governments should put a premium on utilizing public good provision as a strategic complement to coercive con ict management tools. However, little is known about governments' motivation to utilize social services to `win the hearts and minds' of the population and reduce popular support for the armed non-state challenger. One reason for why the use of social welfare-related instruments in con ict is not well understood might be the high level of empirical endogeneity between public good provision and insecurity. Public good provision is found to in uence violence, but the latter also has an eect on the former. Another reason for why governments' motivation to employ welfare spending as a con ict management tool remains understudied is that public goods can be supplied by providers other than the government | encompassing the entire range of private and public, for-prot and not- for-prot, international and national non-state actors (Lloyd et al. 2019; Stewart 2018; Berman and Matanock 2015; Krasner and Risse 2014; Mampilly 2011) | thereby making the attribution of outcomes to actors and actions dicult. The insight that welfare investments aect violence is captured by the literature establishing the con ict-reducing or -preventing eect of public good provision cited above. The second direc- tion of the relationship | that violence can aect welfare investments | is studied less frequently and to date has most prominently been captured by the `guns vs. butter' literature. 4 This re- search focuses on trade-os between military and social spending under conditions of war and establishes that military spending may crowd-out other expenditure. However, by and large, this literature is based predominantly on cases of large-scale interstate wars and focuses on the budget allocations of Western developed states. In addition, the majority of the works within the `guns versus butter' debate are limited to establishing that the trade-o between military and welfare spending exist, but do not seek to explain variation across space and/or time. 4 For a short review of this literature, see Sexton et al. (2019) and Obinger and Petersen (2016). 68 An exception to this trend is the recent paper by Sexton et al. (2019). Using Shining Path dissident violence in Peru as an example, the authors nd that military spending crowds-out health care expenditure. These results appear to stand in direct contrast to the argument made here, i.e. that insecurity should be associated with higher, rather than lower, levels of welfare spending. Dierences in the administrative setting between the study by Sexton et al. (2019) and my own stand out when speculating about the reasons for the divergent results. Peru is scally centralized, while Colombia features a high degree of administrative decentralization, but with centralized armed and national police forces (Cardenas et al., 2016). Sexton et al. (2019) employ Peruvian central government expenditures that might be much more susceptible to the guns versus butter eect than municipal expenditures in Colombia. This is not to say that local politics matter. In fact, Sexton et al. recognize that, choosing coercion can require a resource-constrained government to cannibalize spend- ing on social services in favor of other priorities | shifting resources from `butter' to `guns.' [...] [S]uch budget shifts cause short- and long-term deterioration in civilian welfare, including increasing loss of life, particularly at the local level, where cuts are most proximate to the services provided to people in need. (Sexton et al., 2019, 354) Hence, the incentive of local governments to invest in citizen welfare as a reaction to insecurity might be be much higher, compared to the national administration. Investment decisions by lower level administrative units are therefore a crucial puzzle piece to gaining an in-depth understanding of the functioning of `hearts and minds' strategies. Furthermore, the nature of insecurity diers between the two studies. Sexton et al. (2019) explore the eect of a very specic type of inse- curity: dissident violence in the context of drug tracking and looting. This type of violence is conceptually very dierent from the insurgency-style insecurity that I argue causes governments to increase welfare spending. Future medium- or large-N comparative research should investigate 69 further the intricacies of the eect of dierent administrative settings and types of violence on government investments in social welfare. The provision of security in itself is a public good (Lee et al. 2014; Risse 2011); and indeed one could ask how dierent types of insecurity impact the state's investment in its armed forces or police apparatus. However, I exclude security from the list of public good investments in this study to focus on non-coercive measures that governments take in light of persistent insecurity. If popular grievances and opportunity-seeking are the central root causes of insecurity, 5 then coercive measures such as military operations and increased policing will only tackle the symp- tom of the problem. Coercive measures might oer symptomatic relief in weakening the state's armed challengers. However, these symptomatic actions are likely to be unsustainable in the long run, because they fail to confront the root causes. Non-coercive measures that address popular grievances and opportunity-seeking, such as investments in the welfare of citizens, are therefore much more likely to yield sustained reductions in violence. The conviction that coercive military measures will only address the symptoms of insecurity, and that a more sustainable strategy to address root cases, lies at the heart of dual (or hybrid) Counterinsurgency (COIN) campaigns we observe in Afghanistan, Iraq, and India. 6 The paper thus makes the assumption that, while the provision of security is crucial, governments may also gain support by providing services to the local population. 5 A erce debate exists regarding the causal mechanisms that link underdevelopment and insecurity. Most arguments structure around the causes of a) low opportunity cost due to persistent grievances such as poverty (Collier and Hoeer 2004) or relative deprivation (Cederman et al. 2013; Stewart 2008), and b) politico-military moments in countries with low state reach (Holtermann 2012; Fearon and Laitin 2003). 6 For a review of the recent and historical literature on the `hearts and minds' approach, see Berman and Matanock (2015). 70 3.3 Three Logics Connecting Violence and Welfare Investments I distinguish between three logics for governments to invest in social welfare as a reaction to insecurity. They are rooted in the conjecture that welfare spending benets the population, and that governments invest in social welfare to secure citizen support. All three logics therefore t within the framework of the government seeking to `win the hearts and minds' of the population, but they probe separate underlying incentives. The logics dier in terms of which strategic situations trigger this eect and induce governments to use welfare spending as an instrument to combat insecurity. 3.3.1 The intensity of violence perspective First, welfare spending could simply be a function of the intensity of insecurity, as measured for example by the number of total fatalities in a given year, irregardless of the victim, perpetrator, or type of violence causing them. 7 According to this logic, governments use welfare spending as a reaction to insecurity in order to improve citizens' living situation and increase public support, no matter whether the insecurity stems from con ict or crime. Here, the primary goal of welfare spending is to increase popular support by improving citizen security which prevails if \a state's citizens [are able] to live free from immediate danger to their lives and livelihood" (Schr oder 2010, 19). Based on this logic, if governments use investments in social welfare to increase popular support, I expect an overall positive correlation between insecurity and welfare spending (H1). H1 Intensity of insecurity: Higher levels of insecurity are associated with higher levels of government investment in social welfare. 7 Fatalities are not the only corollary of insecurity. In addition to people dying, we see injuries, traumatization, displacement, and physical damages as a result of insecurity. However, since data on fatalities suers from a much lower underreporting bias, I focus on intentional homicides and deaths due to violent con ict as my main measure for the intensity of insecurity. 71 Civil war is only one of many causes of insecurity. In fact, it is not the most common source of non-natural fatalities. World Health Organization (WHO) data from the early 2000s suggests that globally, homicides exceed war-related deaths by a factor of approximately three | a ratio that shows an upward trend (data cited in Fox and Hoelscher 2012). The same is true for Colombia. Despite an ongoing civil war in the country, signicantly more Colombians fell victim to homicides than killings by insurgents, paramilitaries, or government armed forces. In 1987, a commission on the causes of violence in Colombia concluded: \Much more than that of the mountains, it is the violence in the streets that is killing us" (cited in S anchez and N u~ nez 2007, 28, own translation). The overall intensity of insecurity in a country is therefore operationalized as fatalities stemming from both con ict and common or organized crime, as well as communal violence. 8 3.3.2 State-centric strategic perspective The second logic raises doubts about the conceptualization of government spending as tool that is universally used by governments as a reaction to general insecurity. Governments are facing budget constraints and have to be strategic in their usage of funds. This second perspective is rooted in the insight that civil con icts can be characterized as a struggle over the statehood between a government and an organized armed non-state opponent (B orzel 2012). Increasing service provision in a given region is likely to augment the provider's ability to project control, and therefore its strength vis- a-vis the armed non-state challenger. Improving public goods and service provision may thus increase the ability of a government to combat its opponent. Anecdotal evidence suggests that governments are acknowledging the fact that improving the quality of service provision is related to the dynamics of violent con ict. For example, Berman et al. (2013) cite a report by the United States (US) army, according to which \increasing the 8 I adopt the denition of common crime as \acts of violence committed by individuals or groups that do not actively re ect an attempt to contest the authority of a state" (Fox and Hoelscher 2012, 433). Organized crime is dened as a \phenomenon comprising hierarchically organized groups of criminals with the ability to use violence, or the threat of it, for acquiring or defending the control of illegal markets in order to extract economic benets from them" (Kalyvas 2015, 1518, based on Reuter 2008). Based on Tajima (2013), communal violence is dened as "group violence between two or more communities, including both ascriptive identity groups (e.g., ethnic or religious groups) and locational groups (e.g., village or neighborhood groups)" (p. 104). 72 quantity or quality of government services will reduce violence" (1). This is the central logic of so- called hybrid counterinsurgency campaigns that pursue a coupling of coercive and non-coercive means, based on the observed \complementarity between public service provision and security provision in generating stability" (Berman et al. 2013, 3). If governments used welfare spending as a response to attacks against their agents, but not as a reaction to overall levels of insecurity, this would show that governments employ welfare investments to counter strategic situations on the ground. However, this insight alone does not establish whether governments do indeed use welfare spending to improve the their ghting capacity (or ability to project control). After all, gov- ernments could seek to improve their overall ability to govern, not their odds of defeating their armed challenger. In order to trace out the military-strategic aspect of the state-centric logic, I distinguish between violence directed against armed versus unarmed agents of the state. This dis- tinction yields a second theoretical expectation. If governments use investments in social welfare in an attempt to improve their technical ability to project control, then we should see a positive correlation between con ict and welfare investments only for violence experienced by armed state forces. H2 State-centric: Higher levels of violence against armed state actors, but not vio- lence directed against unarmed agents of the state, are associated with higher levels of government investment in social welfare. 3.3.3 Citizen-centric strategic perspective Welfare spending is a fundamentally citizen-centric policy tool. As mentioned in the intensity of insecurity perspective above, investments in social welfare may be aimed at the political goal of fostering popular support. However, when engaged in a counterinsurgency campaign, popular support has a strategic dimension. It has been argued that \counterinsurgency is fundamentally 73 a struggle over people, not territory" (Berman et al., 2011, 773), because counterinsurgents rely on the local population deny support rebels and instead provide the government with intelligence. The third logic probes whether governments utilize welfare spending as a strategic tool to increase their ability to prevail in combat by reducing popular grievances | thereby disincentivizing re- bellion. This citizen-centric perspective diers from the rst logic in that it focuses on strategic, as opposed to general political, support. It diers from the second logic in that it focuses on the ability of welfare investments to improve a state's ability to prevail in con ict by increasing popular strategic support, as opposed to technical ghting capabilities. In this context, citizen security should only matter for welfare spending as a function of how much the survival of the state is threatened by the population's insecurity. I argue that it is not the concern for civilian support per se that determines the government's motivation to invest in welfare-related public goods. Instead, governments use welfare spending strategically if their survival is threatened as a result of violence experiences by their citizens. Rooted in the notion of insurgency and counterinsurgency as competitive statebuilding (Staniland, 2012, 244), this should only be the case if citizens are caught in the cross-re between the government and its armed challenger. All else equal, insecurity caused by in common crime | i.e. violence that is not targeted against the state in an eort to overthrow the government | should have a lower eect on welfare spending than insecurity in icted upon the population by a guerrilla group that seeks to overthrow the state. 9 The primary goal of welfare spending as a strategic tool is to improve state security, that is the state's freedom from immediate threat to its survival (Schr oder, 2010), not citizen security, as in the intensity of violence perspective. I assume a government invests in citizen support because it cares about its survival (Pierskalla 2012; Berman et al. 2011). The more impacted citizens are by the battle between the govern- ment and the rebels, the more likely they are to withdraw strategic support from the incumbent 9 Here, guerrilla-style warfare is dened as \a strategy of armed resistance that (1) uses small, mobile groups to in ict punishment on the incumbent through hit-and-run strikes while avoiding direct battle when possible and (2) seeks to win the allegiance of at least some portion of the noncombatant population" (Lyall and Wilson 2009, 70). 74 government. The higher the threat to the survival of the government, the higher its motivation should be to use scarce resources toward the reduction of the threat. Therefore, the level of government spending is hypothesized to increase as a a function of the level insecurity that the civilian population faces as a fall-out from insurgency. This yields the theoretical expectation that if governments use social welfare investments in an attempt to improve their ability to project control indirectly by decreasing popular strategic support for the rebels, civilian victimization on the hands of insurgents should be associated with higher welfare spending. H3 Citizen-centric: Higher levels of violence in icted upon civilians by armed chal- lengers of the government, but not overall levels of insecurity experienced by the pop- ulation, are associated with higher levels of government investment in social welfare. The citizen- and state-centric perspectives are intimately connected. For example, by improv- ing the chances of the population to nd employment in a sector that is legal, thus raising the opportunity costs for rebel recruitment, investing in education serves the dual goals of weaken- ing the armed challenger and/or decreasing the grievances of the local population (Taydas and Peksen, 2012). The interconnection is even more clear with respect to health care. Governments may invest in health care to improve their ability to care for sick or wounded agents of the state, thereby increasing their ghting capacity (state-centric perspective), and/or to boost popular support (citizen-centric perspective). Due to the diculty of drawing a clear line between the causal logics of the citizen- and state centric perspectives, the dierentiation between the latter two logics is a matter of gradation, rather than classication. Governments could plausibly use welfare spending as a tool in all three scenarios above. How- ever, assuming that governments are not benevolent spenders, but use investments in public goods strategically to ensure their survival, I do not expect to nd an eect based simply on the in- tensity of insecurity. Instead, rst, I expect the eect of insecurity on welfare spending to be increasing in the degree to which the state and its agents themselves are aected. Second, the 75 eect of insecurity on welfare spending should be increasing in the expected eectiveness of the expenditure. Governments should channel funds to those who suer the most from the violent battle between the state and the rebels, where additional funds might make the largest dierence in popular support. 3.3.4 Scope condition Each of the three theoretical expectations operates under the assumption that the government has a sucient amount of territorial control. I introduce this assumption to highlight that the ability of governments to provide services is conditional upon their access to the territory in question. Thus my hypotheses are limited to environments characterized by relatively low intensity con icts, where governance might be contested, but where the government has not lost signicant amounts of territorial control. 10 I focus on explaining the variation in governments' utilization of welfare spending to combat intrastate insecurity in the context of low to medium levels of insecurity; that is violence which typically does not reach the conventionally accepted threshold for civil war of 1000 battle-related deaths per year (Harbom et al. 2009). The rationale for this limitation is two-fold. First, empiri- cally speaking, political violence that reaches stages of civil war is the exception rather than the rule. Lower levels of violence are much more prevalent and aect much larger parts of the global population (Barron et al. 2009). For example, in 2014, only 11 out of 40 reported armed con- frontations reach the 1000 battle-related deaths threshold (Pettersson and Wallensteen 2015, 537). However, despite its empirical prevalence, few scholars have focused on lower levels of insecurity. Second, government welfare spending is not theorized to be linearly increasing in insecurity. Rather, the relationship between insecurity and government welfare spending is expected to follow an inverted-U shape, as depicted in Figure 3.1. The reason for this inverted-U shape is the intervening variable territorial control. Once a critical amount of territorial control is lost by the 10 In future work, I plan to use my new estimates of territorial control as an intervening variable. 76 Figure 3.1: Theoretical relationship between insecurity and government spending on welfare- related public goods. The theoretical considerations focus on explaining empirical settings that fall on the solid part of the curve, namely lower levels of insecurity. government, the state will not be able to provide welfare-related goods in these areas any longer and spending should go down. In addition, it could be argued that once a certain threshold of insecurity in surpassed, the trade-o between pro-citizen welfare spending and military spending could be lop-sided in favor spending that directly aides the survival of the state. 3.4 Intrastate Con ict in Colombia The 2016 peace agreement between the government of Colombia and the Fuerzas Armadas Rev- olucionarias de Colombia{Ej ercito del Pueblo (FARC) formally ended a decades-long armed con- frontation. 11 The country has experienced decades of intrastate violence stemming from clashes between politically-motivated insurgent groups and the ocial armed forces, paramilitaries (Car- denas et al. 2016; Cort es and Montolio 2014; Dube and Vargas 2013); and violence linked to and exacerbated by the presence of drug cartels in the country. 11 The Guardian, 23 June 2016, https://www.theguardian.com/world/2016/jun/23/ colombia-farc-rebel-ceasefire-agreement-havana, accessed 10 September 2016. 77 While the formation of the FARC and Ej ercito de Liberaci on Nacional (ELN) rebel move- ments traces back to the 1960s 12 , individual paramilitary groups did not start forming until the 1980s as armed counterinsurgency self-defense groups that were assembled and nanced by rural landowners and drug lords (Dube and Vargas 2013). The role of the paramilitaries for the con ict has changed over time and is subject to much debate both in academic as well as policy circles (Dube and Naidu, 2015). The paramilitaries became a major player in the con ict starting in the 1990s and their involvement is linked to the intensication of violence in the late 1990s and early 2000s. These right-wing groups formed an unocial alliance with Colombian government forces in their ght against the left-wing guerrillas and, while not being ocially funded by the gov- ernment, \they did receive informal assistance from military and police ocers through unocial channels" (Dube and Naidu, 2015, 252). Land inequality, insucient land titling, and rural poverty are central grievances in explaining the root causes of the con ict (Albertus and Kaplan 2013; Ib a~ nez and V elez 2008). In Colombia, the conditions that created inequitable land distribution trace back to the 1800s when the state engaged in colonizing the hinterland to alleviate land pressures in urban areas (Albertus and Kaplan 2013). Calls for land reform were at the center of demands made by the FARC and ELN guerrilla groups and were used to justify armed activity. The land reform that was undertaken by the Colombian government in the 1960s exacerbated, rather then solved, the problem due to elite capture of the reform process. Continued reform eorts throughout the decades have done little to tackle the underlying problems, \leaving many peasants still impoverished while large areas of cultivable land are underutilized" (Albertus and Kaplan 2013, 204). Research on the Colombian con ict has identied a number of issues that contributed to the continuation of intrastate violence well into the 21st century. The intensication of violence in Colombia since the 2000s (see Figure 3.2) has been linked to the considerable increase in coca 12 In fact, it has been argued that the roots of the FARC trace all the way back to the La Violencia movement that was active from the late 1940s to the late 1950s (Cort es and Montolio 2014; Rozo 2015; Albertus and Kaplan 2013). 78 Figure 3.2: The plot shows the geographical and temporal evolution of the Colombian con ict over time. Dark shaded areas represent municipalities in which guerrillas (FARC and/or ELN) are present. The maps illustrate that in the past 25 years, Colombia experienced the highest level of civil con ict in the early 2000s. In 2003, guerrillas were operating in over 60% of municipalities. In 2013, this number dropped to 18 percent of municipalities in which guerrillas were present. plantations since the late 1990s, as a consequence of the militarization of the war on drugs and the US-backed disruption of the air bridge of coca trade into Colombia that shifted the production of plants from Peru and Bolivia to Colombia. (Angrist and Kugler 2008; Dube and Vargas 2013) \By providing nancial resources to illegal armed groups, drug trade fueled the con ict and allowed its geographical expansion" (Ib a~ nez and V elez 2008, 660). Violence perpetrated by various armed groups against the civilian population and land expulsion of farmers has left millions internally displaced (Bozzoli et al. 2013). In addition, rent-seeking and predatory capture of proceeds related to the exploitation of Colombia's rich natural resources, including oil, gold, and coal, are cited as root causes of continued violence (Cort es and Montolio 2014; Dube and Vargas 2013). 3.5 Municipal Expenditure in Colombia As illustrated by Uribe's statement in the introduction, the Colombian federal government had little motivation to engage in economic development of insurgency-aected areas, and instead focused on eradicating coca plantations and eliminating the armed challengers with coercive 79 means. 13 However, there is a second side to this counterinsurgency strategy that has received much less attention in the existing literature: How do local level governments react to (changes in) the security environment? Does their reaction dier depending on the source of the insecurity they face? Colombia is a centralized state comprised of 32 departamentos and 1,122 municipios. Despite political centralization, municipalities enjoy a high degree of scal autonomy (Cardenas et al. 2016; Faguet and S anchez 2014). Thanks to a 25-year decentralization reform process starting in the 1980s, municipalities have considerable spending powers for local services such as education, health care, and sanitation. For example, in the area of education, \[m]ayors are in charge of setting need-based priority levels in schools with regard to investment, as well as the type of investment, whether the need is for infrastructure, educational material, feeding programs, or for more teachers, to mention just a few. Equally, in the water sector, the mayor is responsible for taking a decision as to who will provide the service, who will monitor the provider, and whether the coverage should be increased to new areas of the municipality." (Torres and Pach on, 2013, 10) Colombian military and police forces, on the other hand, remain highly centralized (Cardenas et al., 2016, 3). This allows me to study the eect of insecurity on state actors' welfare-mindedness in a context where governments do not have to worry as much about the `guns' part of the trade-o between military and social welfare spending. 14 The decentralization process was coupled with the introduction of a transfer scheme through which signicant funds are directed from the national to the local level. It has been noted that, \transfers have a substantial equalizing eect on the ability of local governments to deliver 13 Uribe received considerable criticism for his approach. Colombia later changed its strategy and invested in public services to combat its non-state armed challengers, see Sexton et al. (2019). 14 The fact that Colombian municipalities do not face a direct trade-o between military and security spending does, however, come with a caveat for generalizability. While this sub-national setting allows me to blend out the `guns' aspect and focus on the `butter,' the mechanisms discovered in this analysis may not travel to other environments where governments do face the trade-o between military and social spending. I do, however, attempt to control for a weak version of the `guns vs. butter' eect by including an indicator for municipal investments in the justice sector into one of my model specications. 80 comparable local service levels" (Chaparro et al. 2004, 2). However, the reforms placed signicant limitations on how the transferred funds have to be used at the local level, \mandating that the bulk should be spent on education and health care" (Faguet and S anchez 2014, 231). In particular, \30 per cent of municipal transfers are earmarked for education programs, 25 per cent for health, and 20 per cent for water and sewage projects" (Chaparro et al. 2004, 18). Initially, the transfer scheme was set up in a way that created incentives for municipalities to invest as little of their own resources as possible, in order to receive higher levels of royalties from the national government (Torres and Pach on 2013). With the reform of the transfer scheme statutes in the early 2000s, Colombia has seen an increase in the utilization of municipalities' own funds (Torres and Pach on 2013). For this investigation, this is an important policy change since it allows for a greater variation in the amounts that municipalities spend on education, health, and sanitation projects on top of centrally mandated levels in the latter years of the data. 3.6 Data and estimation 3.6.1 Measuring municipal investment in social welfare I employ a rich unbalanced panel data set of over 1,100 municipalities 15 from 1994 to 2010 to investigate the eect of insecurity on municipal welfare-mindedness in Colombia. My dependent variable \government investment in social welfare" is measured using the per capita municipal investments (inversi on) in the areas of education, health care, and water/sanitation. The Colom- bian ocial record's usage of the word inversi on cannot be equated with general \government spending" in these areas, nor does it refer to long-term future-oriented expenses with the expecta- tion of material returns. As Faguet and S anchez state, \Colombia's public accounts classify such items as teachers' and health workers' salaries as investments, and not running costs" (Faguet and S anchez 2014, 232). The Colombian spending category of `investments' is highly adequate 15 The number of municipalities in the sample varies due to the creation of new municipalities over the course of the sample period. 81 as an indicator for government willingness to provide public goods, because it is less likely to be comprised of non-volatile cost such as pensions and xed administrative costs, that are typically captured by government expenditure measures. Data on municipal investments in the areas of education, health care, and water/sanitation come from two sources: The Panel Municipal that is maintained by the Centro de Estudios sobre Desarrollo Econ omico (CEDE) institute of the Universidad de los Andes in Bogot a (covering expenditure data from 2000 to 2010) and the data employed by Cardenas et al. (2016) and Drazen and Eslava (2010). 16 All spending variables originate from the values collected by the Departamento nacional de planeaci on (DNP). 17 All scal variables are measured in thousands of per capita 2010 constant pesos and are log transformed. 18 3.6.2 Measuring insecurity To measure insecurity, I rely on indicators of violence from the CEDE panel. Most of these measures originally come from the Colombian Ministry of Defense and the National Police. 19 For some indicators, such as fatalities among armed forces, the data are ne-grained enough to distinguish between con ict events that are directly attributable to a certain actor. When appropriate, I employ both an indicator that contains only con ict events that are attributable to the guerrilla groups FARC and ELN versus an indicator measuring con ict events attributable to all non-state armed actors, to distinguish violence that is created through direct con ict between the government and its main armed challenger. 20 Table 3.2 summarizes the key variables to capture the eect of insecurity, arranged by the hypotheses that they pertain to. 16 I am much obliged to Marcela Eslava for graciously sharing her municipal expenditure data. In the models I am using data starting in 1994 when most con ict indicators (in their one-year lagged version) become available. 17 Drazen and Eslava (2010) mention that their series stops in 2002 because the way that the data is aggregated after 2003 changes, and because the Fiscal Transparency and Responsibility Law introduced in 2003 changed how municipal nances were reported and managed. While these changes complicate the merging of the two data series, two aspects allow me to combine the data. First, there is an overlap of two years between the CEDE panel and the data from Drazen and Eslava (2010), which allows me to compute an average for these three years to \stitch" the series together. This is only a rst very crude attempt at combining the two series. Future work on this project will use a more sophisticated imputation model to estimate a continuous trend for municipal spending and exploit the overlap in the data for the years 2000 to 2002. Second, the changes introduced in 2003 aected all municipalities 82 Variable Description Data Source H1 Homicides per 100,000 All annual fatalities not attributable to natural death or accidents, including con ict-related deaths. National Police (NP) H1 Not con ict-related homicides per 100,000 All annual fatalities that are not attributable to natural death, accidents, or the violent con ict. Computed by subtracting fatalities of Armed Forces (AF) and non- state armed actors from total annual homicides. Homicides from NP; bat- tle deaths from NP col- lected by CEDE and pub- lished by DNP. H2 AF fatalities (NSAA) Total fatalities among AF attributable to the actions of a Non-state armed actor (NSAA), that is FARC, ELN,Autodefensas Unidades de Colombia (AUC), and unidentied actors. CEDE (NP, DNP) H2 AF fatalities (GUER) Total fatalities among AF attributable to the actions of guerrillas, that is FARC and ELN. CEDE (NP, DNP) H2 Total oensives against AF (NSAA) All attacks by non-state armed actors to promote or generate a breakthrough in their ghting fronts. CEDE (NP, DNP) H2 Total oensives against AF (GUER) All attacks by guerrillas to promote or generate a break- through in their ghting fronts. CEDE (NP, DNP) H2 Political homicides (NSAA) Homicides of civilians in a political or ideological capac- ity, such as elected ocials, candidates, and community leaders (Dube and Naidu, 2015), attributable to the ac- tions of non-state armed actors. CEDE(NP, DNP) H2 Political homicides (GUER) Homicides of civilians in a political or ideological capac- ity, attributable to the actions of guerrillas. CEDE (NP, DNP) H3 Total attacks on civil- ians (NSAA) Includes a wide range of violent actions against civil- ians, including kidnappings, murders, and damage to property, attributable to the actions of non-state armed actors. CEDE (NP, DNP) H3 Total attacks on civil- ians (GUER) Includes a wide range of violent actions against civil- ians, including kidnappings, murders, and damage to property, attributable to the actions of guerrillas. CEDE (NP, DNP) Table 3.2: Description and data sources for key municipal-level indicators capturing insecurity. 3.6.3 Controls 3.6.3.1 Base model As mentioned in section 3.5, a signicant proportion of the expenditures in the areas of education, health care, and water/sanitation is mandated by the central government. While municipalities have signicant discretion over their local spending, they are not completely autonomous with respect to investments in education, health care, and sanitation. How much municipalities spend in these areas is dependent on how much of their resources are earmarked for a specic expenditure equally. Therefore, any changes in the reporting of municipal nances should be picked up by the year xed eects and do not threaten my identication strategy. 18 For all variables that are log transformed, I add 0.0001 to the value to prevent zero values from dropping out due to the transformation. 19 Unfortunately, the data on homicides in the CEDE panel does only start in 2003. I therefore supplement this series with homicide data from the national police that has previously been used in Rozo (2015). The correlation between the two series for the years in which both indicators are available is 0.99. 20 As mentioned above, existing research documents the existence of an unocial alliance between the paramil- itaries and the government. Employing a separate indicator that measures only violence perpetrated by guerrilla groups allows me to test whether it is the type of violence, or the actor perpetrating the acts, that yields the eect 83 category. For this reason, I employ an indicator which measures the level of expenditure required under the transfer mechanism Sistema General de Participaciones (SGP) within the respective issue area in the battery of controls. These mandated amounts per SGP are measured in thousands of per capita constant 2010 pesos and are log transformed. 21 Previous research suggest that governments adjust the resources they invest in social welfare based on transfers they receive Torres and Pach on (2013). I therefore control for the proportion of a municipality's revenue that comes from transfers and royalties from either the central government or the department level. A municipality's ability to make welfare investments in addition to the mandated amounts depends chie y on its tax revenue. As Perry et al. (2015) state, \[m]any municipalities make important contributions from local taxes to the nancing of water supply and basic health and education services" (6). I therefore also control for revenue from taxes as a proportion of a municipality's total income. To control for the total amount of resources that a municipality has available, I include the natural log of total municipal income in the battery of controls. Lastly, I include an indicator measuring whether or not the municipality was running a scal decit in a given year. All scal indicators stem from the CEDE municipal panel. As mentioned earlier, intrastate violence in Colombia has resulted in a high number of Inter- nally Displaced Person (IDP). Internal displacement puts additional burdens on municipalities. If people are forced to leave the area, a municipality might lose workers that are necessary to sustain its sources of revenue. In turn, a large in ux of internal refugees typically increases the number of those in need of public services without increasing the resources available to provide them. One could argue that the municipality with a net in ow of IDPs gains more workers and should therefore have more resources available for public spending. This would, however, be short-sighted. Not only is it unlikely that the local labor market will be able to absorb a large 21 Unfortunately, in the expenditure category of water and sanitation, this indicator contains a large number of missing values. Models employing the SGP control for water and sanitation consequently have much lower number of observations. Results for the full model in this expenditure category should therefore be interpreted with great care. Mandated expenditure levels per SGP are also available for the school food program (Programa de Alimentaci on Escolar) which is intended to combat malnutrition. However, since data on this program are not available for all regions of Colombia, this variable was excluded from the analysis. 84 amount of refugees in a a short amount of time; many refugees might in fact not even be eligible to become part of the local work force, due to traumatization, injury, or age. I therefore control for the natural log of both the in ow and out ow of IDPs. The set of controls also includes the natural log of the total population in a municipality. This controls for the eect of economies of scale in the provision of public services. Lastly, I control for the ratio of the rural population to the total population. Rural environments with a widely dispersed population and a lower level of development typically make public goods provision harder, and hence more expensive. This indicator can take on values between zero and one, with higher values indicating a higher level of rurality. 3.6.3.2 Alternative explanations The existing literature highlights a number of other mechanisms that could lead governments to adjust their welfare investments in times of insecurity | mechanisms that are not directly attributable to the government seeking to improve its ghting capacity. I employ three additional sets of variables that engage these literatures and probe their validity. First, the aforementioned guns versus butter debate claims that governments will reduce, not increase, their welfare investments in light of insecurity (Sexton et al., 2019). As mentioned above, because the Colombian army and national police remain centralized, this theory is not directly applicable to the context of municipal spending. It might be that security spending crowds out welfare investments at the national level, with no measurable eect at the local level if municipalities can compensate reduced central government transfers with own resources. Given the current research, I cannot directly test whether this is the case. However, as a preliminary test for the `guns versus butter' mechanism, I estimate a set of models that controls for per capita municipal investments in the justice sector (justicia), obtained through the CEDE panel. 22 If I observe a positive eect of violence on welfare investments upon controlling for these expenditures, 22 According to the CEDE codebook, this investment category among others covers expenditures such as the salaries of police inspectors or the contracting of special services in agreement with the national police. 85 this provides tentative evidence that security spending is not crowding out welfare expenditure in Colombian municipalities. Second, it has been argued that both guerrillas and paramilitaries in some cases exerted a strong in uence on municipal politics, mainly through violent pressure of voters and narco money (Cardenas et al., 2016, 4). Hence, municipalities might increase investments not in an attempt to ght the armed challenger, but because they have become the `puppet on a string' of precisely those actors. The likelihood of this type of municipal capture should be increasing in the strength of non-state armed actors. I employ an interaction term between the presence of either guerrillas or paramilitaries and the proportion of the municipal area in which coca is produced as a proxy for the strength of non-state armed actors. Both guerrillas and paramilitaries rely heavily on drug trade for their nancing (Dube and Naidu, 2015, 251{252). Hence, higher levels of coca production in municipalities where either group is present indicate stronger non-state actors and a higher probability of municipal capture by these actors. Data on the presence of guerrillas, paramilitaries, and the area covered by coca is obtained through the CEDE panel, and supplemented with data from Dube and Naidu (2015). Third, existing research has established both a strong eect of political budget cycles on municipal spending (Drazen and Eslava, 2010), and the increase of political assassinations by paramilitaries in election years (though there is a substitution away from other assassinations during those times, Dube and Naidu 2015). Hence, if political violence and elections are cor- related, and under the presence of political budget cycles, we cannot separate out the eect of insecurity on welfare investments. The year xed eects that are included in all model specica- tions should control for a simple form of political budget cycles, where spending is increased in election years to buy voter support, as the timing of council and mayoral elections is the same across all municipalities. 23 However, it has been shown that highly competitive municipalities are 23 Local elections that fall in the period of observation in this study were conducted in 1994, 1997, 2000, 2003, and 2007, based on information from the National civil status register http://www.registraduria.gov. co/-Elecciones-Regionales-.html, accessed 17 February 2017. 86 more susceptible to election-related violence (Dube and Naidu, 2015, 264). To control for com- petitiveness in local elections, I control for the winning vote margin in the last mayoral elections, measured as the percentage of total votes that made the elected candidate win, from Machado and Vesga (2016). 24 3.6.4 Estimation To estimate the impact that insecurity has on government investment in social welfare, I estimate the following xed-eect model via Ordinary Least Squares (OLS): Exp i;t =Ins i;t1 + 0 Budget i;t + 0 Budget i;t1 + 0 Controls i;t1 + i + t + i;t ; where Exp i;t denotes the observed expenditure in the given category (education, health care, or water/sanitation) by municipality i and year t, Ins i;t1 denotes the one-year lagged indica- tor of insecurity, Budget i;t and Budget i;t1 indicate the contemporaneous and one-year lagged budgetary controls, Controls i;t1 corresponds to the matrix of other municipality time-varying one-year lagged controls, and i and t are xed eects by municipality and year, respectively. For each dependent variable investment category, I estimate ve versions of this model: a) Using only the insecurity indicator, b) using a full set of base controls, c) base controls plus investments in the justice sector, d) variables representing the municipal capture alternative explanation in addition to the base controls, and nally e) a model with base controls and the vote margin in the last municipal elections. The CEDE panel contains information on the executed, rather than planned budget of munic- ipalities, that is spending amounts that are recoded at the end of the year. Therefore, I consider both contemporaneous and one-year lagged versions of the budgetary variables in the base model. While the budgeted investment amounts will be determined in the year prior to the execution, 24 I thank Fabiana Machado for graciously sharing her data and patiently answering numerous questions. 87 current year spending may depend on both nancial planning in the previous year and the evo- lution of municipal nances in the current year, such as tax yield. 25 In contrast, and to omit endogeneity concerns for the main eect as much as possible, I lag all other predictor variables, including violence indicators, by one year, to ensure that I am appropriately capturing the in u- ence of insecurity on the budgeting process. Year dummies are included in all model specications to control for factors that in uence the spending of all Colombian municipalities simultaneously, such as changes in legislation and macroeconomic impacts. 3.7 Empirical Analysis The theoretical considerations in this project rely on the assumption that insecurity has an in u- ence on government investment in social welfare and that this eect can be positive. Governments may choose to channel funds toward areas that see higher levels of insecurity in an attempt to increase their technical ability to project control, or to increase popular support. Figure 3.3 plots the eect of guerrilla presence on a logged indicator of municipal expenditure on education, health care, and water/sanitation over time. The grey trend line indicates the number of municipalities in which guerrillas as present. The eect of guerrilla presence on mu- nicipal investments on education and health care increases over time | an eect that especially for health-related investments is statistically signicant at the minimum 5% level of signicance. The eect of guerrilla presence on water investments is negative for all years, as compared to the excluded year 1994, but in most instances not statistically signicant at the minimum 5% level. Guerrilla presence in Colombian municipalities peaked around 2003 and then quickly falls back to average levels. Despite this national trend of a return to lower levels of guerrilla presence, a proxy for an increased intensity and spatial coverage of the con ict, the eect of insecurity on invest- ments in education and health in individual municipalities continues to increase, in comparison 25 While large changes between the planned and executed budget are unlikely, the results are very similar to using one-year lags for all predictor variables. 88 to 1994. This might indicate an increased sensibility of municipalities to the necessity to invest in social welfare over time. Figure 3.3: Eect of guerrilla presence on municipal expenditure in education, health care, and water/sanitation. The points plot the coecient for a regression of municipal welfare investments on the interaction between a dummy for one-year lagged guerrilla presence and the year (including lower order terms) with 95% condence intervals. The model is estimated via OLS and includes municipality xed eects. Standard errors are clustered by municipality. 1994 is the excluded category for the year dummies. Positive coecients denote eects of guerrilla presence on welfare investments that are larger than in 1994; negative coecients smaller eects than in 1995. Shown in the background by a thick grey line is the standardized number of municipalities with guerrilla presence in a given year. Positive values denote above average numbers of municipalities in which guerrillas are present. Figure 3.3 illustrates the applicability of my prior assumptions to the Colombian context: Guerrilla presence does appear to have an eect on municipal welfare spending. The graphs demonstrate that insecurity can have a positive and statistically signicant eect on expenditure (indicated by points that lie above the horizontal line of no discernible dierence between munic- ipalities in which guerrillas are present and those in which FARC and/or ELN do not operate). The graph shows that especially toward the end of the period of observation, municipalities in which guerrilla groups operate, on average have higher expenditures and health than municipali- ties with no presence of the FARC and/or ELN. However, the graph also demonstrates variation 89 concerning the eect of insecurity on welfare spending remains to be explained. I will therefore now turn to the discussion of the multivariate regression models that oer more nuanced insights into the underlying relationships. 3.7.1 Main regression results Figure 3.4 plots the coecients for all insecurity indicators, based on models with a complete baseline set of controls for the three dependent variables municipal expenditure on education, health care, and water/sanitation. 26 For ease of comparability, the insecurity indicators have been standardized to have a mean of zero and variance of one (z-scores). This makes the size of the impact of each insecurity indicator on municipal investment in social welfare directly com- parable. Since the dependent variables are log transformed and the coecients of the insecurity indicators are rather small, the size of the regression coecient times 100 can be interpreted as the percent change in expenditure that is associated with a one standard deviation change in inse- curity. In Figure 3.4, I separate the coecient plot into panels that correspond to the theoretical expectations for ease of interpretation. Based on the rst vertical panel in Figure 3.4, I do not nd empirical support for the notion that insecurity alone incentivizes governments to invest in social welfare (H1). Neither the coecient for the total municipal homicide rate nor the coecient for the non con ict-related fatality rate are statistically signicant for either dependent variable at the minimum 5% level of signicance. This is a rst piece of evidence for my claim that governments, who are faced with violence, use welfare spending strategically only in a few select circumstances. The second and third vertical panels in Figure 3.4 allow me to assess whether violence against state actors will trigger municipal investments in citizen welfare. Armed forces fatalities that are attributable to non-state armed actors are associated with higher investments in education and health and the eect on education is robust when considering only fatal incidents perpetrated by 26 Tables with a full set of regression specications, including covariates, and robustness checks are presented in the appendix. 90 Figure 3.4: Coecient plot for three dependent variables: Annual per capita municipal investment in education, health care, and water/sanitation. Dependent variables are measured in thousands of per capita constant 2010 pesos and log-transformed. Budgetary controls are included as con- temporaneous and one-year lagged regressors. All other explanatory variables are lagged by one year. All insecurity indicators have been standardized to have a mean of zero and a standard deviation of one. All models include the full set of baseline controls. 0.0001 is added to all vari- ables before logging. Horizontal lines indicate the 90% and 95% condence intervals. Solid shapes correspond to coecients that are signicant at the minimum 5% level of signicance. 91 guerrillas. A one standard deviation increase in the number of armed forces fatalities perpetrated by non-state armed actors ( 0:6) is associated with 2.2% higher municipal investments in education and 2.6% higher investments in health care per capita. In addition, investments in health care appear to be in uenced by guerrilla attacks on government forces, but not overall non-state actor oensives. The comparative eect of oensives by guerrilla groups on health investment is lower than the eect of armed forces fatalities, despite the former constituting the more common event. The estimated eect of a one standard deviation increase in oensives through guerrillas ( 3:3) corresponds an average of 1.6% higher per capita health investments. The comparison of the eect size, frequency of events, and severity of the assault lends itself to the interpretation that is not so much the quantity, but the quality, of insecurity that triggers governments to strategically invest in social welfare. Taken together with the insight that political assassinations eect neither education nor health care investments (third vertical panel), the results of the empirical analysis yield empirical support for the theoretical expectation (H2): When under attack, governments use welfare spending to improve their strategic situation, not their overall ability to govern. I nd a weak positive eect of guerrilla attacks against civilians on municipal investments in education (fourth panel in Figure 3.4). A one standard deviation higher number of attacks against civilians ( 1:25) is associated with on average 1.5% higher per capita educational investments. The eect is positive, but does not remain statistically signicant at the minimum 5% level of signicance when considering assaults against the civilian population by all other armed non-state actors. When contrasted with the insight that the non-con ict related fatality rate | a large proportion of which will civilian fatalities through common crime | does not yield a signicant scal response, this results suggest weak support for the third theoretical expectation (H3). Only when civilians are caught in the front lines between the state and its armed challenger, will governments resort to social welfare investments. This allows for the cautious interpretation 92 that governments' motivation to increase service provision in the education sector relies on the expectation of strategic, but not political, public support. What about investments in the water and sanitation sector? The negative eect of guerrilla oensives on water investments is the sole instance in 40 models that include a full set of baseline controls in which the coecient an insecurity indicator is statistically signicant at the minimum 95% level, when using water investments as the dependent variable. 27 This result is not robust to alternative explanations. One explanation for the non-eect of insecurity on the provision of water and sewage services might be the comparatively high level of privatization in that sector. A law passed in 1996 opened the way for \private participation" in the water sector (Andres et al., 2010, 30). In 2010, 38% of Colombian municipalities featured some level of private participation in the provision of water and sewage services (Andres et al., 2010, 52). Hence, municipalities might be less visible as the actor who brings improvements to the living situation of the people | rendering investments in the water sector a subprime tool when seeking to garner strategic popular support. With respect to the technical ghting ability of governments, the private participation in the water sector could also make it harder for governments to utilize investments toward their strategic needs. For this reason, I exclude investments in the water sector from further discussion. 28 3.7.2 Alternative explanations Figure 3.5 illustrates the eects of adding controls for the three alternative explanations. The rst row re-plots the eects of insecurity on education and health care investments in the baseline model. The second row of panels adds a control for the guns versus butter eect, the third considers indicators that proxy municipal capture by armed non-state actors, the and the fourth row controls for the in uence of political budget cycles. The non-signicant eects of overall levels of violence and political assassinations are robust to alternative explanations. When it comes to 27 See the full regression tables provided in the appendix. 28 Future research should study the eect of insecurity on service provision through public-private partnerships in more depth. 93 violence directed against armed state agents and events where the security of citizens is threatened by armed non-state actors, the interpretations are more nuanced. The second vertical panel in Figure 3.5 suggests that the eects of both armed forces fatalities and oensives are, by and large, robust to controlling for municipal investments in the justice sector, as well as the probability of municipal capture through non-state armed actors. In many models, I estimate a positive and statistically signicant eect of contemporaneous investments in the justice sector | suggesting that in the case of Colombian municipalities, the guns versus butter trade-o does not exist. For all models employing education expenditures as the depen- dent variable, I estimate a large negative and statistically signicant interaction between the level of coca production and the presence of guerrillas in a municipality. I do not nd a statistically signicant eect of municipal capture indicators on health investments. Contrary to the alter- native view according to which non-state actors seizing control over municipal politics may lead to increases in welfare spending, the opposite appears to be the case. Municipalities with strong guerrillas seem to invest less, not more, in education. In addition, and despite the strength of guerrilla groups and the accompanying trend to divest, municipal education investments continue to be increasing in the level of armed forces fatalities that are attributable to either the FARC or ELN. While guerrilla groups may exert political in uence on education investments, governments appear to retain their motivation for educational investments that are of strategic importance on the battleeld. Upon controlling for political budget cycle eects, I do not nd statistically signicantly discernible eects of violence against armed state actors on neither education nor health investments; despite the coecient of the winning vote margin the in last municipal elec- tions remaining insignicant across all models. The results on the eect of political budget cycles are inconclusive. 29 29 Future work on this project should explore alternative measurement strategies for the competitiveness of the municipal political environment, for example via the Golosov index measuring of the eective number of parties in mayoral or council elections. This measurement approach has previously been employed for example in the study by Dube and Naidu (2015). 94 The eects of violence perpetrated by non-state armed actors against civilians are by and large robust to alternative explanations considered in this study. In fact, upon controlling for municipal investments in justice, guerrilla attacks against civilians are found to have an even larger eect on investments in social welfare. A one standard deviation increase in the number of assaults by guerrillas is now associated with on average 2.6% (compared to 1.5% without the guns versus butter control) higher educational and 2% higher health investments, all else equal. This suggests that while the eect of citizen insecurity is not numerically larger than the eect of state insecurity on welfare investments, it may be more robust to alternative mechanisms. 95 Figure 3.5: Coecient plot illustrating the eects of control variables capturing the three alternative explanations for increased municipal social welfare investments in light of insecurity. Point estimates represent the coecients of the insecurity indicators for education and welfare per capita investments. Investments in the water sector are excluded from the graph, but can be found in the full regression tables in the appendix. Dependent variables are measured in thousands of per capita constant 2010 pesos and log-transformed. Budgetary controls are included as contemporaneous and one-year lagged regressors. All other explanatory variables are lagged by one year. All insecurity indicators have been standardized to have a mean of zero and a standard deviation of one. All models include the full set of baseline controls. 0.0001 is added to all variables before logging. Horizontal lines indicate the 90% and 95% condence intervals. Solid shapes correspond to coecients that are signicant at the minimum 5% level of signicance. 96 3.8 Conclusion The results of the empirical analysis shed light on the patterns of government usage of welfare spending as a strategic con ict management tool. Not all types of intrastate insecurity motivate governments to `win the hearts and minds' of the people. The nature and source of insecurity matters for whether governments choose to react with investments in social welfare. Governments are highly strategic spenders. The empirical analysis in this paper suggests that governments increase welfare spending under two circumstances. First, investments in education and health care are found to increase when strategic, but not political, government forces are under attack. Second, we see higher municipal investments in education and health care when civilians are victimized as a result of the ght between the government and an armed challenger. Together with the nding that overall civilian fatalities itself do not induce increased investments in social welfare, the results paint a grim picture of the state seeking to improve the living situation of a con ict-ridden population only to the extent that its own survival is suciently threatened | either forcefully through an armed challenger, or indirectly through the revocation of citizen support. The analysis in this paper illustrates why we do not see governments universally increasing their welfare spending in light of insecurity, even though the existing literature suggests it as a highly eective preventive measure, or even remedy for violence (Taydas and Peksen, 2012; Dasgupta et al., 2015). If citizens are harmed through violence that is not immediately attributable to an organized armed challenger of the state, governments might simply not see a need to use welfare spending as a way to ensure citizen support. Research on insurgencies identies civilian information sharing (or the lack thereof) as the factor that makes or breaks the success of the government in ghting the rebels (Berman and Matanock, 2015; Berman et al., 2011). Insurgencies in this literature have been modeled as a three-sided game between the state, rebels, and civilians (Berman et al., 2011). However, in the case of criminal violence, no well-dened link or direct 97 strategic interaction exists between the government and an organized armed actor that citizens may pinpoint as the source of their suering. This turns the three-sided game of insurgency into two binary struggles that are fought in a high degree of isolation. One contest is fought between the state and the non-state armed actor(s), and another struggle exists between non-state armed actors and the civilian population. In these scenarios, garnering popular support may neither be a strategic necessity, nor a primary concern for the government. 98 3.9 Acknowledgements I am grateful to the audiences of 2015 and 2017 International Studies Association (ISA) Annual Conventions and ISA West 2016 for helpful comments on earlier versions of this study, in particular Joseph Young, David Carter, Charles Mahoney, and William Thompson. Special thanks goes to Patrick James, Brian Rathbun, Jerey Nugent, Carol Wise, and Joey Huddleston for providing comments throughout various stages of this project. 99 Appendix 3.A Summary statistics Table 3.3: Summary statistics. Statistic Min Median Mean Max St. Dev. N Education investment p.c. 0.001 35.841 130.614 261,120.400 2,759.017 14,232 Health investment p.c. 0.005 121.011 462.753 520,642.300 9,104.046 13,986 Water/sanitation investment p.c. 0.001 42.231 198.830 322,054.500 4,190.964 14,024 Homicides p. 100,000 0.000 32.616 52.949 1,327.945 70.606 17,684 Non-con ict related homicides p. 100,000 149.589 29.331 48.396 1,327.945 66.839 17,676 Armed Forces (AF) deaths (NSAA) 0 0 0.087 30 0.566 18,828 AF deaths (GUER) 0 0 0.030 7 0.225 18,828 Oensives against AF (NSAA) 0 0 1.684 146 5.978 18,820 Oensives against AF (GUER) 0 0 0.951 74 3.304 18,828 Political homicides (NSAA) 0 0 0.067 17 0.368 18,828 Political homicides (GUER) 0 0 0.014 4 0.132 18,828 Attacks against civilians (NSAA) 0 0 2.045 326 8.097 18,828 Attacks against civilians (GUER) 0 0 0.405 32 1.250 18,828 Total income p.c. 0.000 431.731 517.377 8,551.180 467.197 18,459 Decit dummy 0 1 0.505 1 0.500 18,666 Transfer dependence 0.000 0.075 0.091 0.729 0.063 17,257 Proportion of income from taxes 0.000 11.872 20.854 349.152 25.902 17,257 SGP Education investment p.c. 0.108 28.725 78.851 70,176.700 1,048.961 15,922 SGP Health investment p.c. 0.026 78.572 260.442 201,947.700 4,783.339 15,905 SGP Water/sanitation investment p.c. 0.005 30.071 40.238 1,139.328 41.807 8,893 Rural population of total population 0.001 0.648 0.601 1.000 0.239 18,801 Total population 0 12,355 36,791.520 7,363,782 222,506.500 19,074 Internally displaced persons (IDP) expulsion 0 40 309.200 53,322 1,072.778 19,074 IDP reception 0 28 307.500 52,260 1,572.623 19,074 Justice investment p.c. 0.0002 3.936 18.613 16,100.160 323.667 9,930 Presence of guerrillas dummy 0 0 0.446 1 0.497 19,074 Presence of paramilitaries dummy 0 0 0.134 1 0.341 19,074 Proportion of area with coca cultivation 0.000 0.000 0.001 0.241 0.006 15,922 Vote margin in last mayoral election 0.000 12.107 16.761 100.000 16.294 12,857 Notes: NSAA refers to all non-state actors (guerrillas, paramilitaries, and unidentied actors). GUER refers to the two guerrilla groups FARC and ELN. 3.B Regression Tables 100 Table 3.4: Regression results for the con ict variable: Total fatalities per 100,000 population. Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.007 0.008 0.008 0.0004 0.004 0.017 0.013 0.005 0.011 0.004 0.008 0.027 0.008 0.010 0.017 (0.009) (0.008) (0.010) (0.008) (0.007) (0.010) (0.009) (0.007) (0.009) (0.007) (0.010) (0.034) (0.056) (0.039) (0.031) Income p.c. i;t 0.491 0.422 0.468 0.569 0.252 0.126 0.265 0.389 0.798 0.486 0.789 0.676 (0.098) (0.143) (0.107) (0.060) (0.067) (0.058) (0.074) (0.046) (0.108) (0.207) (0.126) (0.158) Deficit dummy i;t 0.112 0.093 0.105 0.116 0.129 0.082 0.126 0.133 0.329 0.227 0.311 0.252 (0.017) (0.017) (0.018) (0.016) (0.015) (0.012) (0.016) (0.014) (0.045) (0.069) (0.049) (0.050) Transfer dependence i;t 0.415 0.147 0.295 0.037 0.959 0.506 0.883 0.699 0.187 0.119 0.073 0.205 (0.242) (0.343) (0.249) (0.220) (0.270) (0.213) (0.283) (0.264) (0.637) (1.223) (0.668) (0.674) Tax revenue i;t 0.002 0.003 0.001 0.003 0.002 0.002 0.001 0.002 0.002 0.006 0.001 0.002 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.125 0.170 0.172 0.113 0.025 0.095 0.020 0.029 0.0001 0.346 0.112 0.150 (0.035) (0.036) (0.032) (0.036) (0.041) (0.033) (0.042) (0.034) (0.090) (0.143) (0.097) (0.109) Deficit dummy i;t1 0.013 0.018 0.029 0.022 0.028 0.042 0.033 0.044 0.052 0.014 0.028 0.029 (0.014) (0.013) (0.015) (0.015) (0.014) (0.011) (0.015) (0.013) (0.042) (0.066) (0.049) (0.048) Transfer dependence i;t1 0.223 0.208 0.109 0.097 0.045 0.032 0.020 0.108 0.443 1.712 0.215 1.302 (0.199) (0.234) (0.208) (0.205) (0.215) (0.188) (0.231) (0.189) (0.617) (1.021) (0.660) (0.615) Tax revenue i;t1 0.0001 0.001 0.0004 0.001 0.001 0.001 0.00003 0.0001 0.001 0.004 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.572 0.812 0.113 0.286 0.111 0.452 0.551 0.125 0.174 0.643 0.220 1.208 (0.490) (0.719) (0.505) (0.539) (0.706) (0.817) (0.714) (0.677) (0.708) (2.113) (0.774) (0.908) Total population i;t1 0.158 0.195 0.133 0.068 0.551 0.035 0.453 0.388 0.819 0.453 0.950 0.380 (0.157) (0.228) (0.172) (0.209) (0.178) (0.219) (0.193) (0.173) (0.224) (0.589) (0.265) (0.338) IDP expulsion i;t1 0.001 0.001 0.0002 0.001 0.002 0.001 0.007 0.004 0.001 0.003 0.002 0.008 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.006 0.002 0.002 0.003 0.007 0.002 0.008 0.001 0.006 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.584 0.670 0.596 0.630 (0.034) (0.036) (0.036) (0.035) SGP education p.c. i;t1 0.020 0.037 0.030 0.044 (0.019) (0.014) (0.016) (0.014) Justice investment p.c. i;t 0.077 0.064 0.129 (0.013) (0.012) (0.106) Justice investment p.c. i;t1 0.001 0.010 0.031 (0.007) (0.008) (0.047) Coca production i;t1 2.221 3.600 10.012 (0.885) (2.441) (7.791) Guerrilla dummy i;t1 0.002 0.028 0.007 (0.019) (0.020) (0.054) Paramilitary dummy i;t1 0.010 0.013 0.071 (0.018) (0.019) (0.093) Coca x guerrillas i;t1 11.209 2.779 10.549 (3.532) (3.611) (3.086) Coca x paramilitaries i;t1 7.026 3.290 3.693 (3.355) (2.134) (11.238) Winning vote margin i;t1 0.0003 0.0001 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.556 0.683 0.571 0.625 (0.042) (0.046) (0.044) (0.042) SGP health p.c. i;t1 0.010 0.015 0.008 0.027 (0.015) (0.012) (0.014) (0.012) SGP water/sanitation p.c. i;t 0.086 0.294 0.115 0.067 (0.049) (0.127) (0.058) (0.062) SGP water/sanitation p.c. i;t1 0.071 0.026 0.041 0.074 (0.040) (0.074) (0.043) (0.046) Observations 13,562 11,851 7,430 10,201 8,714 13,320 11,675 7,448 10,098 8,718 13,357 4,436 2,381 3,626 3,204 Adjusted R 2 0.546 0.691 0.790 0.701 0.731 0.731 0.804 0.832 0.800 0.800 0.607 0.624 0.715 0.633 0.688 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 101 Table 3.5: Regression results for the con ict variable: Total non-con ict related fatalities per 100,000 population. Con ict variable is standardized (mean zero, sd one). Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.006 0.008 0.006 0.001 0.003 0.015 0.012 0.005 0.011 0.002 0.011 0.034 0.014 0.015 0.025 (0.010) (0.008) (0.010) (0.009) (0.007) (0.010) (0.009) (0.007) (0.009) (0.007) (0.010) (0.035) (0.055) (0.039) (0.031) Income p.c. i;t 0.489 0.422 0.468 0.569 0.254 0.126 0.265 0.390 0.799 0.483 0.788 0.675 (0.098) (0.143) (0.107) (0.060) (0.067) (0.058) (0.074) (0.046) (0.108) (0.207) (0.126) (0.158) Deficit dummy i;t 0.112 0.093 0.105 0.116 0.128 0.082 0.126 0.132 0.329 0.227 0.311 0.252 (0.017) (0.017) (0.018) (0.016) (0.015) (0.012) (0.016) (0.014) (0.045) (0.069) (0.049) (0.050) Transfer dependence i;t 0.416 0.145 0.295 0.039 0.956 0.506 0.883 0.697 0.191 0.133 0.073 0.201 (0.241) (0.343) (0.249) (0.220) (0.270) (0.213) (0.283) (0.264) (0.637) (1.227) (0.668) (0.674) Tax revenue i;t 0.002 0.003 0.001 0.003 0.002 0.002 0.001 0.002 0.002 0.006 0.001 0.002 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.123 0.170 0.172 0.113 0.024 0.095 0.020 0.029 0.001 0.344 0.112 0.150 (0.035) (0.036) (0.032) (0.037) (0.040) (0.033) (0.042) (0.034) (0.091) (0.143) (0.097) (0.109) Deficit dummy i;t1 0.013 0.018 0.029 0.022 0.029 0.042 0.033 0.044 0.050 0.015 0.029 0.029 (0.014) (0.013) (0.015) (0.014) (0.014) (0.011) (0.015) (0.013) (0.042) (0.066) (0.049) (0.048) Transfer dependence i;t1 0.221 0.210 0.110 0.096 0.043 0.032 0.020 0.106 0.442 1.688 0.212 1.291 (0.199) (0.234) (0.208) (0.205) (0.214) (0.188) (0.231) (0.189) (0.617) (1.018) (0.659) (0.613) Tax revenue i;t1 0.0001 0.001 0.0004 0.001 0.001 0.001 0.00003 0.00004 0.001 0.004 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.581 0.811 0.113 0.287 0.151 0.452 0.551 0.124 0.116 0.653 0.218 1.207 (0.491) (0.719) (0.505) (0.540) (0.705) (0.817) (0.714) (0.678) (0.707) (2.115) (0.774) (0.908) Total population i;t1 0.157 0.198 0.133 0.067 0.554 0.035 0.453 0.387 0.824 0.446 0.950 0.381 (0.157) (0.228) (0.172) (0.209) (0.178) (0.219) (0.193) (0.174) (0.224) (0.590) (0.265) (0.338) IDP expulsion i;t1 0.001 0.001 0.0001 0.001 0.002 0.001 0.007 0.004 0.002 0.004 0.002 0.008 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.006 0.002 0.002 0.003 0.007 0.002 0.007 0.001 0.006 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.584 0.670 0.596 0.630 (0.034) (0.036) (0.036) (0.035) SGP education p.c. i;t1 0.020 0.037 0.030 0.044 (0.019) (0.014) (0.016) (0.014) Justice investment p.c. i;t 0.077 0.064 0.129 (0.013) (0.012) (0.106) Justice investment p.c. i;t1 0.001 0.010 0.031 (0.007) (0.008) (0.047) Coca production i;t1 2.225 3.600 9.977 (0.884) (2.441) (7.782) Guerrilla dummy i;t1 0.002 0.028 0.007 (0.019) (0.020) (0.054) Paramilitary dummy i;t1 0.010 0.013 0.071 (0.018) (0.019) (0.093) Coca x guerrillas i;t1 11.211 2.779 10.577 (3.532) (3.611) (3.079) Coca x paramilitaries i;t1 7.024 3.290 4.084 (3.355) (2.134) (11.107) Winning vote margin i;t1 0.0003 0.0001 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.556 0.683 0.571 0.625 (0.042) (0.046) (0.044) (0.042) SGP health p.c. i;t1 0.009 0.015 0.008 0.027 (0.015) (0.012) (0.014) (0.012) SGP water/sanitation p.c. i;t 0.086 0.295 0.115 0.067 (0.049) (0.128) (0.058) (0.062) SGP water/sanitation p.c. i;t1 0.071 0.027 0.041 0.074 (0.040) (0.074) (0.043) (0.046) Observations 13,559 11,849 7,430 10,201 8,714 13,318 11,673 7,448 10,098 8,718 13,355 4,434 2,381 3,626 3,204 Adjusted R 2 0.547 0.692 0.790 0.701 0.731 0.731 0.804 0.832 0.800 0.800 0.607 0.624 0.715 0.633 0.688 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 102 Table 3.6: Regression results for the con ict variable: Armed forces fatalities by non-state armed actors. Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.015 0.022 0.020 0.018 0.012 0.020 0.026 0.031 0.025 0.008 0.008 0.008 0.019 0.014 0.052 (0.008) (0.008) (0.010) (0.011) (0.012) (0.010) (0.013) (0.011) (0.012) (0.014) (0.011) (0.017) (0.034) (0.026) (0.050) Income p.c. i;t 0.486 0.435 0.472 0.574 0.264 0.142 0.275 0.401 0.824 0.483 0.799 0.684 (0.094) (0.144) (0.107) (0.058) (0.066) (0.060) (0.075) (0.045) (0.109) (0.202) (0.127) (0.156) Deficit dummy i;t 0.113 0.097 0.107 0.116 0.125 0.085 0.124 0.133 0.331 0.230 0.309 0.251 (0.016) (0.017) (0.018) (0.015) (0.014) (0.012) (0.016) (0.014) (0.044) (0.068) (0.049) (0.049) Transfer dependence i;t 0.442 0.160 0.355 0.035 0.913 0.510 0.894 0.607 0.268 0.143 0.100 0.304 (0.234) (0.328) (0.243) (0.217) (0.253) (0.206) (0.277) (0.244) (0.631) (1.174) (0.672) (0.665) Tax revenue i;t 0.002 0.002 0.001 0.003 0.002 0.002 0.001 0.002 0.003 0.005 0.001 0.003 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.117 0.160 0.162 0.115 0.021 0.082 0.020 0.020 0.008 0.332 0.104 0.151 (0.034) (0.035) (0.032) (0.037) (0.039) (0.030) (0.042) (0.033) (0.090) (0.140) (0.096) (0.108) Deficit dummy i;t1 0.012 0.017 0.031 0.023 0.030 0.041 0.033 0.046 0.050 0.023 0.032 0.025 (0.014) (0.013) (0.015) (0.014) (0.014) (0.010) (0.015) (0.013) (0.042) (0.065) (0.049) (0.047) Transfer dependence i;t1 0.056 0.139 0.077 0.014 0.019 0.019 0.075 0.074 0.355 1.332 0.215 1.200 (0.219) (0.224) (0.205) (0.228) (0.204) (0.180) (0.228) (0.175) (0.598) (0.920) (0.650) (0.584) Tax revenue i;t1 0.0002 0.001 0.0001 0.001 0.001 0.0002 0.0001 0.001 0.001 0.003 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.469 0.957 0.109 0.289 0.032 0.431 0.483 0.121 0.101 0.471 0.169 1.110 (0.480) (0.705) (0.501) (0.530) (0.679) (0.787) (0.704) (0.659) (0.692) (2.047) (0.770) (0.912) Total population i;t1 0.135 0.235 0.134 0.056 0.479 0.013 0.462 0.309 0.823 0.442 0.957 0.366 (0.154) (0.216) (0.170) (0.205) (0.176) (0.209) (0.191) (0.170) (0.221) (0.581) (0.265) (0.337) IDP expulsion i;t1 0.001 0.0001 0.001 0.001 0.003 0.001 0.008 0.005 0.001 0.004 0.002 0.007 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.007 0.002 0.003 0.003 0.007 0.002 0.008 0.001 0.006 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.593 0.668 0.596 0.627 (0.032) (0.036) (0.036) (0.035) SGP education p.c. i;t1 0.022 0.035 0.030 0.044 (0.018) (0.014) (0.016) (0.013) Justice investment p.c. i;t 0.073 0.065 0.126 (0.013) (0.012) (0.105) Justice investment p.c. i;t1 0.003 0.012 0.032 (0.007) (0.007) (0.046) Coca production i;t1 1.832 3.812 10.942 (1.149) (2.518) (6.521) Guerrilla dummy i;t1 0.003 0.032 0.006 (0.018) (0.020) (0.054) Paramilitary dummy i;t1 0.007 0.009 0.067 (0.018) (0.018) (0.094) Coca x guerrillas i;t1 10.336 2.330 10.673 (3.981) (3.658) (3.041) Coca x paramilitaries i;t1 6.313 3.108 2.093 (3.287) (1.966) (10.155) Winning vote margin i;t1 0.0002 0.0001 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.566 0.678 0.570 0.619 (0.039) (0.044) (0.043) (0.041) SGP health p.c. i;t1 0.009 0.013 0.006 0.024 (0.014) (0.012) (0.014) (0.011) SGP water/sanitation p.c. i;t 0.069 0.291 0.114 0.060 (0.051) (0.128) (0.058) (0.060) SGP water/sanitation p.c. i;t1 0.081 0.022 0.045 0.073 (0.041) (0.074) (0.043) (0.046) Observations 14,225 12,372 7,784 10,392 9,086 13,980 12,200 7,805 10,293 9,090 14,018 4,581 2,486 3,661 3,323 Adjusted R 2 0.543 0.690 0.788 0.699 0.727 0.731 0.805 0.831 0.801 0.802 0.602 0.625 0.715 0.632 0.689 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 103 Table 3.7: Regression results for the con ict variable: Armed forces fatalities by guerrillas (FARC and ELN). Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.012 0.019 0.018 0.019 0.008 0.012 0.015 0.014 0.014 0.008 0.009 0.027 0.056 0.016 0.045 (0.008) (0.008) (0.010) (0.008) (0.008) (0.011) (0.009) (0.008) (0.009) (0.006) (0.011) (0.025) (0.030) (0.022) (0.024) Income p.c. i;t 0.486 0.437 0.473 0.574 0.264 0.142 0.275 0.400 0.824 0.492 0.799 0.686 (0.092) (0.141) (0.104) (0.058) (0.065) (0.059) (0.074) (0.045) (0.109) (0.203) (0.127) (0.156) Deficit dummy i;t 0.113 0.097 0.108 0.116 0.125 0.085 0.124 0.133 0.330 0.232 0.309 0.250 (0.016) (0.017) (0.018) (0.015) (0.014) (0.011) (0.016) (0.014) (0.044) (0.067) (0.049) (0.049) Transfer dependence i;t 0.441 0.162 0.351 0.036 0.911 0.516 0.893 0.606 0.260 0.164 0.086 0.261 (0.233) (0.325) (0.242) (0.217) (0.252) (0.205) (0.277) (0.243) (0.631) (1.180) (0.670) (0.657) Tax revenue i;t 0.002 0.002 0.001 0.003 0.002 0.002 0.001 0.002 0.003 0.005 0.001 0.003 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.118 0.161 0.162 0.115 0.020 0.084 0.020 0.020 0.008 0.313 0.105 0.153 (0.034) (0.035) (0.032) (0.037) (0.039) (0.031) (0.042) (0.033) (0.090) (0.138) (0.096) (0.108) Deficit dummy i;t1 0.013 0.018 0.032 0.024 0.031 0.041 0.033 0.046 0.050 0.024 0.032 0.027 (0.014) (0.013) (0.015) (0.014) (0.014) (0.010) (0.015) (0.013) (0.041) (0.065) (0.049) (0.048) Transfer dependence i;t1 0.060 0.128 0.083 0.011 0.018 0.013 0.075 0.077 0.383 1.408 0.224 1.246 (0.219) (0.225) (0.205) (0.228) (0.204) (0.180) (0.228) (0.175) (0.596) (0.889) (0.650) (0.574) Tax revenue i;t1 0.0001 0.001 0.0001 0.001 0.001 0.001 0.0001 0.001 0.001 0.003 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.465 0.959 0.107 0.289 0.041 0.393 0.492 0.122 0.078 0.263 0.170 1.103 (0.480) (0.702) (0.501) (0.530) (0.679) (0.783) (0.704) (0.658) (0.693) (2.026) (0.771) (0.913) Total population i;t1 0.129 0.237 0.131 0.058 0.477 0.008 0.462 0.311 0.821 0.490 0.957 0.338 (0.154) (0.214) (0.170) (0.205) (0.176) (0.208) (0.191) (0.170) (0.221) (0.577) (0.264) (0.339) IDP expulsion i;t1 0.001 0.0001 0.001 0.001 0.003 0.001 0.008 0.005 0.001 0.004 0.002 0.007 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.007 0.002 0.003 0.003 0.007 0.002 0.008 0.001 0.007 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.592 0.667 0.595 0.627 (0.032) (0.036) (0.035) (0.035) SGP education p.c. i;t1 0.022 0.036 0.031 0.045 (0.018) (0.014) (0.016) (0.013) Justice investment p.c. i;t 0.073 0.065 0.124 (0.013) (0.012) (0.106) Justice investment p.c. i;t1 0.003 0.012 0.032 (0.007) (0.007) (0.046) Coca production i;t1 1.807 3.836 11.128 (1.148) (2.512) (6.580) Guerrilla dummy i;t1 0.004 0.031 0.008 (0.018) (0.020) (0.054) Paramilitary dummy i;t1 0.007 0.010 0.069 (0.018) (0.018) (0.092) Coca x guerrillas i;t1 10.443 2.404 10.668 (3.921) (3.607) (3.075) Coca x paramilitaries i;t1 6.408 3.168 0.251 (3.205) (1.929) (10.397) Winning vote margin i;t1 0.0002 0.0001 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.566 0.678 0.571 0.619 (0.039) (0.044) (0.043) (0.041) SGP health p.c. i;t1 0.009 0.013 0.006 0.024 (0.014) (0.012) (0.014) (0.011) SGP water/sanitation p.c. i;t 0.071 0.302 0.115 0.069 (0.051) (0.130) (0.058) (0.062) SGP water/sanitation p.c. i;t1 0.081 0.013 0.044 0.072 (0.040) (0.071) (0.043) (0.046) Observations 14,225 12,372 7,784 10,392 9,086 13,980 12,200 7,805 10,293 9,090 14,018 4,581 2,486 3,661 3,323 Adjusted R 2 0.543 0.690 0.788 0.700 0.727 0.731 0.805 0.831 0.800 0.802 0.602 0.625 0.717 0.632 0.690 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 104 Table 3.8: Regression results for the con ict variable: Oensives against armed forces by non-state armed actors. Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.007 0.007 0.014 0.003 0.013 0.027 0.019 0.017 0.018 0.012 0.026 0.010 0.019 0.007 0.031 (0.015) (0.008) (0.006) (0.007) (0.008) (0.012) (0.011) (0.011) (0.011) (0.008) (0.012) (0.032) (0.040) (0.029) (0.051) Income p.c. i;t 0.484 0.433 0.471 0.574 0.263 0.139 0.273 0.401 0.823 0.482 0.799 0.684 (0.095) (0.145) (0.107) (0.058) (0.067) (0.061) (0.076) (0.045) (0.109) (0.204) (0.127) (0.156) Deficit dummy i;t 0.113 0.097 0.107 0.116 0.125 0.085 0.124 0.133 0.331 0.230 0.310 0.251 (0.016) (0.017) (0.018) (0.015) (0.015) (0.012) (0.016) (0.014) (0.044) (0.068) (0.049) (0.049) Transfer dependence i;t 0.448 0.174 0.357 0.031 0.915 0.528 0.897 0.609 0.276 0.138 0.091 0.304 (0.235) (0.329) (0.244) (0.217) (0.253) (0.208) (0.278) (0.244) (0.634) (1.190) (0.671) (0.667) Tax revenue i;t 0.002 0.002 0.001 0.003 0.002 0.002 0.001 0.002 0.003 0.006 0.001 0.003 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.120 0.162 0.163 0.116 0.019 0.085 0.018 0.021 0.008 0.336 0.106 0.152 (0.034) (0.034) (0.032) (0.037) (0.040) (0.030) (0.042) (0.033) (0.090) (0.143) (0.096) (0.109) Deficit dummy i;t1 0.013 0.018 0.031 0.023 0.031 0.041 0.032 0.046 0.051 0.023 0.032 0.026 (0.014) (0.013) (0.015) (0.014) (0.014) (0.010) (0.015) (0.013) (0.041) (0.066) (0.049) (0.047) Transfer dependence i;t1 0.051 0.141 0.072 0.016 0.026 0.024 0.066 0.072 0.359 1.333 0.206 1.185 (0.219) (0.223) (0.205) (0.228) (0.204) (0.180) (0.228) (0.175) (0.605) (0.963) (0.653) (0.599) Tax revenue i;t1 0.00004 0.001 0.0002 0.001 0.001 0.0004 0.00004 0.001 0.001 0.003 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.466 0.970 0.104 0.288 0.038 0.387 0.487 0.123 0.086 0.650 0.183 1.130 (0.481) (0.707) (0.501) (0.530) (0.679) (0.788) (0.704) (0.658) (0.692) (2.045) (0.770) (0.907) Total population i;t1 0.132 0.247 0.134 0.056 0.477 0.003 0.462 0.309 0.822 0.438 0.954 0.360 (0.154) (0.218) (0.171) (0.205) (0.176) (0.211) (0.191) (0.170) (0.221) (0.578) (0.264) (0.338) IDP expulsion i;t1 0.001 0.0001 0.0005 0.001 0.004 0.001 0.008 0.005 0.001 0.005 0.002 0.007 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.007 0.002 0.002 0.003 0.007 0.002 0.007 0.001 0.006 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.592 0.667 0.595 0.627 (0.032) (0.036) (0.036) (0.035) SGP education p.c. i;t1 0.023 0.036 0.031 0.045 (0.018) (0.014) (0.016) (0.013) Justice investment p.c. i;t 0.073 0.065 0.124 (0.013) (0.012) (0.106) Justice investment p.c. i;t1 0.003 0.012 0.032 (0.007) (0.007) (0.047) Coca production i;t1 1.821 3.770 10.961 (1.147) (2.477) (6.502) Guerrilla dummy i;t1 0.002 0.030 0.005 (0.018) (0.020) (0.054) Paramilitary dummy i;t1 0.007 0.014 0.065 (0.018) (0.018) (0.093) Coca x guerrillas i;t1 10.276 2.120 10.707 (3.996) (3.705) (3.041) Coca x paramilitaries i;t1 6.247 2.795 3.480 (3.320) (2.013) (9.978) Winning vote margin i;t1 0.0002 0.0001 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.566 0.678 0.570 0.619 (0.039) (0.044) (0.043) (0.041) SGP health p.c. i;t1 0.009 0.013 0.006 0.024 (0.014) (0.011) (0.014) (0.011) SGP water/sanitation p.c. i;t 0.069 0.297 0.115 0.060 (0.051) (0.128) (0.058) (0.059) SGP water/sanitation p.c. i;t1 0.081 0.022 0.045 0.071 (0.041) (0.076) (0.043) (0.046) Observations 14,221 12,369 7,781 10,392 9,083 13,976 12,197 7,802 10,293 9,087 14,014 4,577 2,482 3,661 3,319 Adjusted R 2 0.543 0.690 0.788 0.699 0.727 0.731 0.805 0.831 0.800 0.802 0.602 0.625 0.714 0.632 0.689 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 105 Table 3.9: Regression results for the con ict variable: Oensives against armed forces by guerrillas (FARC and ELN). Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.001 0.002 0.009 0.0003 0.006 0.025 0.017 0.016 0.016 0.010 0.015 0.055 0.026 0.039 0.055 (0.011) (0.008) (0.009) (0.007) (0.007) (0.009) (0.008) (0.007) (0.008) (0.006) (0.009) (0.028) (0.028) (0.025) (0.029) Income p.c. i;t 0.484 0.433 0.471 0.574 0.263 0.139 0.273 0.400 0.825 0.492 0.802 0.687 (0.095) (0.145) (0.107) (0.058) (0.067) (0.061) (0.076) (0.045) (0.109) (0.202) (0.127) (0.157) Deficit dummy i;t 0.113 0.097 0.107 0.116 0.125 0.085 0.123 0.132 0.333 0.232 0.312 0.254 (0.016) (0.017) (0.018) (0.015) (0.014) (0.012) (0.016) (0.014) (0.044) (0.067) (0.049) (0.049) Transfer dependence i;t 0.447 0.172 0.357 0.033 0.915 0.530 0.897 0.609 0.292 0.207 0.113 0.320 (0.235) (0.329) (0.244) (0.217) (0.253) (0.207) (0.278) (0.244) (0.634) (1.182) (0.673) (0.667) Tax revenue i;t 0.002 0.003 0.001 0.003 0.002 0.002 0.001 0.002 0.003 0.005 0.001 0.003 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.119 0.162 0.163 0.116 0.018 0.086 0.017 0.021 0.012 0.331 0.101 0.147 (0.035) (0.034) (0.032) (0.037) (0.040) (0.030) (0.043) (0.032) (0.090) (0.141) (0.097) (0.109) Deficit dummy i;t1 0.013 0.018 0.031 0.023 0.031 0.042 0.032 0.046 0.050 0.022 0.032 0.026 (0.014) (0.013) (0.015) (0.014) (0.014) (0.010) (0.015) (0.013) (0.041) (0.065) (0.049) (0.047) Transfer dependence i;t1 0.050 0.144 0.072 0.017 0.027 0.029 0.066 0.070 0.383 1.355 0.228 1.189 (0.219) (0.222) (0.205) (0.228) (0.204) (0.179) (0.228) (0.175) (0.600) (0.942) (0.653) (0.590) Tax revenue i;t1 0.00003 0.001 0.0002 0.001 0.001 0.0005 0.0001 0.001 0.001 0.003 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.463 0.983 0.103 0.294 0.043 0.374 0.492 0.118 0.078 0.546 0.172 1.149 (0.481) (0.706) (0.501) (0.530) (0.679) (0.784) (0.704) (0.658) (0.692) (2.041) (0.770) (0.905) Total population i;t1 0.131 0.245 0.133 0.056 0.477 0.002 0.462 0.309 0.822 0.424 0.955 0.357 (0.154) (0.218) (0.170) (0.205) (0.175) (0.210) (0.191) (0.170) (0.221) (0.579) (0.264) (0.337) IDP expulsion i;t1 0.001 0.0001 0.0005 0.001 0.004 0.001 0.008 0.005 0.001 0.004 0.002 0.007 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.007 0.002 0.002 0.003 0.007 0.002 0.008 0.001 0.007 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.592 0.668 0.595 0.627 (0.032) (0.036) (0.036) (0.035) SGP education p.c. i;t1 0.022 0.036 0.031 0.045 (0.018) (0.014) (0.016) (0.013) Justice investment p.c. i;t 0.073 0.065 0.125 (0.013) (0.012) (0.105) Justice investment p.c. i;t1 0.003 0.012 0.033 (0.007) (0.007) (0.046) Coca production i;t1 1.814 3.770 11.176 (1.149) (2.455) (6.474) Guerrilla dummy i;t1 0.002 0.029 0.014 (0.018) (0.020) (0.054) Paramilitary dummy i;t1 0.007 0.012 0.056 (0.018) (0.018) (0.093) Coca x guerrillas i;t1 10.295 2.030 10.594 (3.998) (3.716) (3.120) Coca x paramilitaries i;t1 6.280 2.719 3.154 (3.320) (2.014) (9.937) Winning vote margin i;t1 0.0002 0.0001 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.566 0.678 0.570 0.619 (0.039) (0.044) (0.043) (0.041) SGP health p.c. i;t1 0.009 0.013 0.006 0.024 (0.014) (0.011) (0.014) (0.011) SGP water/sanitation p.c. i;t 0.069 0.292 0.114 0.062 (0.051) (0.128) (0.057) (0.061) SGP water/sanitation p.c. i;t1 0.081 0.025 0.044 0.072 (0.040) (0.074) (0.043) (0.046) Observations 14,225 12,372 7,784 10,392 9,086 13,980 12,200 7,805 10,293 9,090 14,018 4,581 2,486 3,661 3,323 Adjusted R 2 0.543 0.690 0.788 0.699 0.727 0.731 0.805 0.831 0.800 0.802 0.602 0.625 0.714 0.632 0.690 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 106 Table 3.10: Regression results for the con ict variable: Political homicides by non-state armed actors. Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.033 0.011 0.012 0.007 0.012 0.006 0.011 0.016 0.010 0.013 0.009 0.011 0.014 0.006 0.025 (0.010) (0.009) (0.007) (0.009) (0.011) (0.018) (0.016) (0.015) (0.016) (0.013) (0.010) (0.018) (0.017) (0.019) (0.033) Income p.c. i;t 0.485 0.433 0.472 0.575 0.264 0.140 0.274 0.402 0.822 0.483 0.800 0.681 (0.095) (0.145) (0.107) (0.058) (0.067) (0.061) (0.076) (0.045) (0.109) (0.204) (0.127) (0.156) Deficit dummy i;t 0.113 0.097 0.107 0.116 0.125 0.085 0.124 0.133 0.330 0.228 0.310 0.251 (0.016) (0.017) (0.018) (0.015) (0.015) (0.012) (0.016) (0.014) (0.044) (0.068) (0.049) (0.049) Transfer dependence i;t 0.449 0.167 0.358 0.033 0.920 0.526 0.900 0.612 0.277 0.161 0.091 0.295 (0.235) (0.329) (0.244) (0.217) (0.253) (0.208) (0.279) (0.244) (0.632) (1.178) (0.670) (0.672) Tax revenue i;t 0.002 0.003 0.001 0.003 0.002 0.002 0.001 0.002 0.003 0.006 0.001 0.002 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.118 0.160 0.162 0.115 0.020 0.083 0.019 0.020 0.008 0.333 0.106 0.152 (0.034) (0.035) (0.032) (0.037) (0.039) (0.030) (0.042) (0.033) (0.090) (0.142) (0.097) (0.108) Deficit dummy i;t1 0.013 0.018 0.031 0.023 0.031 0.041 0.033 0.046 0.050 0.021 0.032 0.026 (0.014) (0.013) (0.015) (0.014) (0.014) (0.010) (0.015) (0.013) (0.041) (0.065) (0.049) (0.047) Transfer dependence i;t1 0.055 0.142 0.075 0.013 0.021 0.022 0.072 0.077 0.372 1.346 0.204 1.204 (0.219) (0.223) (0.205) (0.228) (0.203) (0.180) (0.227) (0.175) (0.597) (0.935) (0.651) (0.593) Tax revenue i;t1 0.0001 0.001 0.0001 0.001 0.001 0.0003 0.00001 0.001 0.001 0.004 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.472 0.960 0.110 0.288 0.038 0.409 0.487 0.126 0.083 0.541 0.184 1.143 (0.482) (0.708) (0.502) (0.530) (0.681) (0.788) (0.706) (0.657) (0.692) (2.031) (0.770) (0.908) Total population i;t1 0.128 0.240 0.131 0.056 0.471 0.002 0.457 0.307 0.824 0.431 0.952 0.361 (0.154) (0.217) (0.170) (0.205) (0.176) (0.211) (0.191) (0.170) (0.221) (0.582) (0.263) (0.338) IDP expulsion i;t1 0.001 0.0001 0.001 0.001 0.003 0.001 0.008 0.005 0.001 0.004 0.002 0.007 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.007 0.002 0.003 0.003 0.007 0.002 0.008 0.001 0.007 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.593 0.669 0.596 0.628 (0.032) (0.036) (0.036) (0.035) SGP education p.c. i;t1 0.021 0.035 0.030 0.044 (0.018) (0.014) (0.016) (0.013) Justice investment p.c. i;t 0.073 0.065 0.125 (0.013) (0.012) (0.105) Justice investment p.c. i;t1 0.003 0.012 0.033 (0.007) (0.007) (0.046) Coca production i;t1 1.840 3.794 11.005 (1.152) (2.517) (6.506) Guerrilla dummy i;t1 0.003 0.031 0.005 (0.018) (0.020) (0.054) Paramilitary dummy i;t1 0.007 0.009 0.063 (0.018) (0.018) (0.094) Coca x guerrillas i;t1 10.302 2.280 10.675 (3.993) (3.681) (3.061) Coca x paramilitaries i;t1 6.256 3.019 3.266 (3.312) (1.988) (9.892) Winning vote margin i;t1 0.0002 0.00005 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.566 0.678 0.570 0.619 (0.039) (0.044) (0.043) (0.041) SGP health p.c. i;t1 0.009 0.013 0.007 0.024 (0.014) (0.012) (0.014) (0.011) SGP water/sanitation p.c. i;t 0.069 0.288 0.115 0.058 (0.051) (0.128) (0.057) (0.061) SGP water/sanitation p.c. i;t1 0.082 0.022 0.044 0.073 (0.041) (0.074) (0.043) (0.046) Observations 14,225 12,372 7,784 10,392 9,086 13,980 12,200 7,805 10,293 9,090 14,018 4,581 2,486 3,661 3,323 Adjusted R 2 0.544 0.690 0.788 0.699 0.727 0.731 0.805 0.831 0.800 0.802 0.602 0.625 0.714 0.632 0.689 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 107 Table 3.11: Regression results for the con ict variable: Political homicides by guerrillas (FARC and ELN). Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.010 0.005 0.011 0.008 0.006 0.001 0.003 0.003 0.001 0.002 0.003 0.005 0.060 0.014 0.031 (0.008) (0.008) (0.007) (0.010) (0.012) (0.010) (0.009) (0.009) (0.010) (0.008) (0.009) (0.023) (0.043) (0.026) (0.032) Income p.c. i;t 0.484 0.433 0.471 0.575 0.263 0.140 0.273 0.401 0.823 0.483 0.801 0.679 (0.095) (0.145) (0.107) (0.058) (0.067) (0.061) (0.076) (0.045) (0.109) (0.202) (0.127) (0.155) Deficit dummy i;t 0.113 0.097 0.107 0.116 0.125 0.085 0.124 0.133 0.331 0.235 0.309 0.250 (0.016) (0.017) (0.018) (0.015) (0.014) (0.012) (0.016) (0.014) (0.044) (0.069) (0.049) (0.049) Transfer dependence i;t 0.446 0.166 0.356 0.036 0.914 0.518 0.896 0.606 0.268 0.067 0.093 0.269 (0.235) (0.329) (0.244) (0.217) (0.253) (0.208) (0.278) (0.243) (0.631) (1.181) (0.672) (0.664) Tax revenue i;t 0.002 0.003 0.001 0.003 0.002 0.002 0.001 0.002 0.003 0.005 0.001 0.002 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.119 0.160 0.163 0.115 0.020 0.083 0.019 0.020 0.008 0.336 0.104 0.155 (0.034) (0.034) (0.032) (0.037) (0.039) (0.030) (0.042) (0.033) (0.090) (0.140) (0.097) (0.108) Deficit dummy i;t1 0.013 0.018 0.031 0.023 0.031 0.041 0.032 0.046 0.050 0.031 0.032 0.026 (0.014) (0.013) (0.015) (0.014) (0.014) (0.010) (0.015) (0.013) (0.041) (0.067) (0.049) (0.047) Transfer dependence i;t1 0.048 0.143 0.070 0.017 0.027 0.026 0.066 0.070 0.363 1.356 0.203 1.176 (0.219) (0.223) (0.205) (0.228) (0.204) (0.180) (0.228) (0.175) (0.597) (0.931) (0.651) (0.588) Tax revenue i;t1 0.00002 0.001 0.0002 0.001 0.001 0.001 0.0001 0.001 0.001 0.003 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.459 0.986 0.093 0.297 0.046 0.365 0.497 0.117 0.086 0.709 0.207 1.152 (0.481) (0.706) (0.502) (0.529) (0.681) (0.785) (0.704) (0.659) (0.695) (2.035) (0.774) (0.901) Total population i;t1 0.132 0.243 0.135 0.056 0.475 0.002 0.462 0.310 0.823 0.451 0.949 0.358 (0.154) (0.217) (0.170) (0.205) (0.176) (0.211) (0.191) (0.170) (0.222) (0.574) (0.263) (0.339) IDP expulsion i;t1 0.001 0.00005 0.0004 0.001 0.004 0.001 0.008 0.005 0.001 0.005 0.002 0.007 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.007 0.002 0.003 0.003 0.007 0.002 0.007 0.001 0.007 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.592 0.668 0.595 0.627 (0.032) (0.036) (0.036) (0.035) SGP education p.c. i;t1 0.023 0.036 0.031 0.045 (0.018) (0.014) (0.016) (0.013) Justice investment p.c. i;t 0.073 0.065 0.121 (0.013) (0.012) (0.106) Justice investment p.c. i;t1 0.004 0.012 0.028 (0.007) (0.007) (0.047) Coca production i;t1 1.807 3.824 11.039 (1.154) (2.518) (6.515) Guerrilla dummy i;t1 0.001 0.032 0.003 (0.018) (0.020) (0.054) Paramilitary dummy i;t1 0.007 0.009 0.060 (0.018) (0.018) (0.094) Coca x guerrillas i;t1 10.326 2.295 10.607 (3.982) (3.673) (3.094) Coca x paramilitaries i;t1 6.319 3.071 3.771 (3.304) (1.999) (10.004) Winning vote margin i;t1 0.0002 0.0001 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.566 0.678 0.570 0.619 (0.039) (0.044) (0.043) (0.041) SGP health p.c. i;t1 0.009 0.013 0.007 0.024 (0.014) (0.011) (0.014) (0.011) SGP water/sanitation p.c. i;t 0.070 0.298 0.114 0.064 (0.051) (0.124) (0.057) (0.062) SGP water/sanitation p.c. i;t1 0.081 0.019 0.045 0.074 (0.041) (0.076) (0.043) (0.046) Observations 14,225 12,372 7,784 10,392 9,086 13,980 12,200 7,805 10,293 9,090 14,018 4,581 2,486 3,661 3,323 Adjusted R 2 0.543 0.690 0.788 0.699 0.727 0.731 0.805 0.830 0.800 0.802 0.602 0.625 0.715 0.632 0.689 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 108 Table 3.12: Regression results for the con ict variable: Attacks against civilians by non-state armed actors. Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.051 0.019 0.020 0.010 0.011 0.011 0.014 0.007 0.016 0.005 0.011 0.023 0.022 0.033 0.013 (0.017) (0.011) (0.010) (0.008) (0.006) (0.015) (0.010) (0.010) (0.010) (0.006) (0.021) (0.029) (0.035) (0.035) (0.040) Income p.c. i;t 0.484 0.433 0.471 0.574 0.262 0.139 0.272 0.400 0.824 0.487 0.800 0.682 (0.095) (0.145) (0.107) (0.058) (0.067) (0.061) (0.076) (0.045) (0.109) (0.205) (0.127) (0.157) Deficit dummy i;t 0.113 0.097 0.107 0.116 0.125 0.085 0.124 0.133 0.331 0.230 0.310 0.249 (0.016) (0.017) (0.018) (0.015) (0.015) (0.012) (0.016) (0.014) (0.044) (0.068) (0.049) (0.049) Transfer dependence i;t 0.455 0.184 0.361 0.027 0.918 0.524 0.901 0.609 0.260 0.126 0.081 0.275 (0.235) (0.329) (0.244) (0.217) (0.253) (0.209) (0.278) (0.244) (0.632) (1.178) (0.671) (0.664) Tax revenue i;t 0.002 0.003 0.001 0.003 0.002 0.002 0.001 0.002 0.003 0.006 0.001 0.002 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.119 0.161 0.162 0.115 0.020 0.083 0.020 0.020 0.009 0.333 0.104 0.150 (0.034) (0.034) (0.032) (0.037) (0.039) (0.030) (0.042) (0.033) (0.090) (0.142) (0.097) (0.108) Deficit dummy i;t1 0.013 0.018 0.031 0.023 0.031 0.041 0.032 0.046 0.050 0.023 0.031 0.026 (0.014) (0.013) (0.015) (0.014) (0.014) (0.010) (0.015) (0.013) (0.041) (0.065) (0.049) (0.047) Transfer dependence i;t1 0.056 0.129 0.075 0.012 0.024 0.021 0.070 0.073 0.354 1.338 0.204 1.168 (0.219) (0.223) (0.205) (0.228) (0.204) (0.180) (0.228) (0.175) (0.598) (0.940) (0.652) (0.589) Tax revenue i;t1 0.00005 0.001 0.0002 0.001 0.001 0.001 0.0001 0.001 0.001 0.003 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.463 0.969 0.103 0.293 0.046 0.369 0.494 0.117 0.106 0.642 0.194 1.145 (0.480) (0.706) (0.501) (0.530) (0.680) (0.785) (0.704) (0.658) (0.692) (2.026) (0.770) (0.909) Total population i;t1 0.137 0.249 0.136 0.057 0.480 0.0003 0.467 0.310 0.824 0.438 0.959 0.354 (0.154) (0.217) (0.170) (0.205) (0.175) (0.211) (0.191) (0.170) (0.221) (0.577) (0.264) (0.339) IDP expulsion i;t1 0.001 0.0001 0.001 0.001 0.004 0.001 0.008 0.005 0.001 0.004 0.002 0.007 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.007 0.002 0.003 0.003 0.007 0.002 0.007 0.001 0.006 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.591 0.666 0.594 0.626 (0.032) (0.036) (0.036) (0.035) SGP education p.c. i;t1 0.023 0.036 0.031 0.045 (0.017) (0.014) (0.016) (0.013) Justice investment p.c. i;t 0.074 0.065 0.124 (0.013) (0.012) (0.105) Justice investment p.c. i;t1 0.003 0.012 0.032 (0.007) (0.007) (0.047) Coca production i;t1 1.778 3.874 10.916 (1.155) (2.522) (6.494) Guerrilla dummy i;t1 0.002 0.032 0.004 (0.018) (0.020) (0.054) Paramilitary dummy i;t1 0.004 0.014 0.082 (0.018) (0.018) (0.087) Coca x guerrillas i;t1 10.316 2.323 10.742 (3.982) (3.672) (3.030) Coca x paramilitaries i;t1 6.343 3.158 4.269 (3.304) (2.001) (10.162) Winning vote margin i;t1 0.0002 0.0001 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.566 0.678 0.570 0.619 (0.039) (0.044) (0.043) (0.041) SGP health p.c. i;t1 0.009 0.013 0.007 0.024 (0.014) (0.011) (0.014) (0.011) SGP water/sanitation p.c. i;t 0.070 0.296 0.115 0.066 (0.051) (0.127) (0.058) (0.062) SGP water/sanitation p.c. i;t1 0.081 0.023 0.044 0.073 (0.040) (0.074) (0.043) (0.047) Observations 14,225 12,372 7,784 10,392 9,086 13,980 12,200 7,805 10,293 9,090 14,018 4,581 2,486 3,661 3,323 Adjusted R 2 0.544 0.690 0.788 0.699 0.727 0.731 0.805 0.830 0.800 0.802 0.602 0.625 0.714 0.632 0.689 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 109 Table 3.13: Regression results for the con ict variable: Attacks against civilians by guerrillas (FARC and ELN). Dependent variable: Education investment p.c. i;t Health care investment p.c. i;t Water/sanitation investment p.c. i;t Binary Base GB Capture PBC Binary Base GB Capture PBC Binary Base GB Capture PBC (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Conflict variable i;t1 0.001 0.015 0.026 0.013 0.012 0.015 0.017 0.020 0.021 0.015 0.008 0.035 0.025 0.012 0.039 (0.009) (0.007) (0.007) (0.007) (0.006) (0.009) (0.009) (0.008) (0.008) (0.006) (0.010) (0.030) (0.031) (0.028) (0.029) Income p.c. i;t 0.485 0.435 0.472 0.573 0.264 0.141 0.275 0.400 0.822 0.487 0.799 0.684 (0.094) (0.142) (0.106) (0.058) (0.066) (0.060) (0.075) (0.045) (0.109) (0.203) (0.127) (0.156) Deficit dummy i;t 0.113 0.097 0.107 0.116 0.126 0.085 0.124 0.133 0.330 0.230 0.310 0.249 (0.016) (0.017) (0.018) (0.015) (0.014) (0.011) (0.016) (0.014) (0.044) (0.068) (0.049) (0.049) Transfer dependence i;t 0.451 0.184 0.359 0.028 0.920 0.534 0.899 0.615 0.288 0.205 0.100 0.343 (0.235) (0.326) (0.243) (0.217) (0.253) (0.206) (0.277) (0.244) (0.635) (1.191) (0.673) (0.668) Tax revenue i;t 0.002 0.003 0.001 0.003 0.002 0.002 0.001 0.002 0.003 0.006 0.001 0.002 (0.001) (0.002) (0.001) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.005) (0.003) (0.004) Income p.c. i;t1 0.119 0.164 0.163 0.116 0.019 0.086 0.018 0.021 0.007 0.331 0.105 0.152 (0.035) (0.034) (0.032) (0.036) (0.040) (0.030) (0.043) (0.032) (0.090) (0.141) (0.097) (0.108) Deficit dummy i;t1 0.013 0.019 0.032 0.024 0.031 0.042 0.034 0.047 0.051 0.021 0.032 0.026 (0.014) (0.013) (0.015) (0.014) (0.014) (0.010) (0.015) (0.013) (0.041) (0.066) (0.049) (0.047) Transfer dependence i;t1 0.065 0.111 0.084 0.004 0.009 0.0003 0.088 0.088 0.394 1.362 0.222 1.209 (0.220) (0.224) (0.206) (0.229) (0.204) (0.180) (0.227) (0.176) (0.598) (0.937) (0.652) (0.590) Tax revenue i;t1 0.0001 0.001 0.0001 0.001 0.001 0.001 0.0001 0.001 0.001 0.003 0.001 0.002 (0.001) (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (0.002) (0.002) (0.004) (0.002) (0.003) Rural population i;t1 0.457 0.941 0.095 0.297 0.048 0.402 0.503 0.115 0.102 0.483 0.184 1.178 (0.481) (0.701) (0.502) (0.530) (0.680) (0.779) (0.705) (0.658) (0.694) (2.057) (0.770) (0.908) Total population i;t1 0.131 0.247 0.132 0.057 0.475 0.003 0.458 0.310 0.833 0.421 0.959 0.362 (0.154) (0.215) (0.170) (0.205) (0.175) (0.209) (0.191) (0.170) (0.222) (0.581) (0.265) (0.339) IDP expulsion i;t1 0.001 0.0004 0.001 0.001 0.003 0.001 0.008 0.005 0.001 0.004 0.002 0.007 (0.003) (0.003) (0.004) (0.004) (0.003) (0.002) (0.004) (0.003) (0.006) (0.008) (0.009) (0.007) IDP reception i;t1 0.003 0.003 0.002 0.007 0.002 0.002 0.003 0.007 0.002 0.008 0.001 0.007 (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003) (0.007) (0.011) (0.008) (0.008) SGP education p.c. i;t 0.592 0.667 0.595 0.627 (0.032) (0.036) (0.035) (0.035) SGP education p.c. i;t1 0.022 0.034 0.030 0.044 (0.018) (0.014) (0.016) (0.013) Justice investment p.c. i;t 0.074 0.066 0.126 (0.013) (0.012) (0.105) Justice investment p.c. i;t1 0.003 0.012 0.032 (0.007) (0.007) (0.047) Coca production i;t1 1.815 3.824 11.001 (1.142) (2.492) (6.509) Guerrilla dummy i;t1 0.006 0.027 0.009 (0.018) (0.020) (0.054) Paramilitary dummy i;t1 0.004 0.014 0.063 (0.018) (0.019) (0.093) Coca x guerrillas i;t1 10.224 2.165 10.709 (3.999) (3.685) (3.056) Coca x paramilitaries i;t1 6.218 2.956 2.742 (3.322) (1.999) (9.928) Winning vote margin i;t1 0.0002 0.0001 0.001 (0.0004) (0.0004) (0.001) SGP health p.c. i;t 0.566 0.677 0.570 0.619 (0.039) (0.044) (0.043) (0.041) SGP health p.c. i;t1 0.009 0.012 0.007 0.023 (0.014) (0.012) (0.014) (0.011) SGP water/sanitation p.c. i;t 0.070 0.292 0.114 0.064 (0.051) (0.126) (0.057) (0.061) SGP water/sanitation p.c. i;t1 0.081 0.026 0.044 0.071 (0.041) (0.074) (0.043) (0.047) Observations 14,225 12,372 7,784 10,392 9,086 13,980 12,200 7,805 10,293 9,090 14,018 4,581 2,486 3,661 3,323 Adjusted R 2 0.543 0.690 0.788 0.699 0.727 0.731 0.805 0.831 0.801 0.802 0.602 0.625 0.714 0.632 0.689 Note: p<0.05; p<0.01; p<0.001 Linear regression results with year and municipality fixed effects. Standard errors clustered by municipality. Conflict variable is standardized (mean zero, sd one). GB stands for models testing the guns vs. butter effect, Capture refers to municipal capture by non-state armed actors, and PBC stands for political budget cycles. 110 3.C Auxiliary results The discussion in the main manuscript focuses on the eects that violence has on municipal investments in public good provision. Below, I engage with the auxiliary results of the statistical analysis, focusing on the estimated eects for the battery of control variables. Since the results on the covariates are extremely similar across models incorporating alternative measures of violence, I limit the narrative to the discussion of Table 3.13. As one would expect, more auent communities are found to make higher municipal in- vestments in education health care, and water/sanitation. The eects are largest for the wa- ter/sanitation sector. Comparing the base models across sectors, a one unit increase in per capita income (i.e. revenue) is on average associated with a 0.82 unit increase in municipal investment in water/sanitation, a 0.48 unit increase in educational investments, and a 0.26 unit increase in health-related investments. Interestingly, one-year lagged municipal per capita income does only have an eect on the education sector, suggesting a potentially longer time horizon for investment decisions in this realm compared to health care and sanitation. At least part of these investments appear to be nanced via decit spending, as the the coecients of the contemporary decit dummies are positive and statistically signicant at the minimum 99% level in all models. Recall that all base models account for the amount of investment that is mandated via the SGP scheme. Controlling for these mandated investments, municipalities that rely more on transfers and royalties from the central or departmental levels invest less in health care, but not education or sanitation. Future research should seek to model these investment patterns in more detail and incorporate interactions between scal variables. Most of the non-scal control variables are not found to be statistically signicantly related to public good investments, with the exception of population. More populous municipalities invest less per capita in the health care and the water sectors, however, the population size does not have an eect on educational investments. This suggests scale eects for the provision of health care and 111 sanitation, however the amount of investment in teachers and schools appears to be proportional to the size of the population. Neither the rural footprint, not the reception of expulsion of IDPs, are statistically signicantly related to municipal public good investments. A fruitful avenue for future work will be to more explicitly model whether the proportion of physical infrastructure versus non-infrastructure investments diers across sectors. Systematic dierences in the degree to which investments in education versus health versus water are charac- terized by expenditures for the building and maintenance of physical infrastructure could explain some of the dierences in the scal patterns. Moreover, it could shed light on the causal pathway between violence and welfare investments. If the eect of violence diers between infrastruc- ture versus non-infrastructure investments, this could indicate that the government is rebuilding destroyed assets, as opposed to investing in capacity building or modernization. 112 Chapter 4 Bread before guns or butter: Introducing Surplus Domestic Product (SDP) The physical product of hundreds of millions of peasants may dwarf that of ve million factory workers, but since most of it is immediately consumed, it is far less likely to lead to surplus wealth or decisive military striking power. { Paul Kennedy, The Rise and Fall of Great Powers. 4.1 Introduction International relations scholarship systematically mis-measures both power-resources and military burdens because the operationalization of these variables depends on Gross Domestic Product (GDP) as an indicator of the income states can devote to arming and projecting power. The core problem is that GDP confounds two conceptually distinct forms of income into one aggregate in- dicator. Subsistence resources are the income necessary to provide the minimal amount of `bread' that the population needs to survive. Surplus resources are the remaining economic income that can be invested in `guns' or `butter' (Garnkel and Syropoulus, 2019). As a consequence of con- ating these two types of income into a single indicator, existing inferences about the distribution of power-resources, states' capacity to build power projection capabilities, and the costs of arming 113 are biased. In particular, the misuse of GDP as a proxy measure of power-resources systematically overestimates power-resources of low-income states with large populations and underestimates the rate at which these resources increase when low-income countries begin to experience economic growth. We address this bias by developing a new measure called Surplus Domestic Product (SDP). SDP is created by decomposing GDP into surplus income and subsistence income. This paper yields three contributions for scholars of international relations, and political science more broadly. First, we introduce surplus income (SDP) and subsistence income, which are both part of total income (measured in GDP). We demonstrate that once we account for both types of income, SDP is a more conceptually appropriate measure of the power-resources available for states to arm and project military force abroad. For illustration, scholars mis-measure military burdens by using military expenditures as share of GDP. When SDP is used in the denominator instead, we demonstrate that, for most of history, nearly all countries endured much higher military burdens than previously realized. Failing to take into account that only surplus resources can sustainably be invested in the military concealed just how rapid and steep the fall in military burdens has been after the Cold War. This drop is especially pronounced for developing states around the world, particularly in Asia. Our results suggest that scholars underestimated the size of the peace dividend associated with the end of the Cold War and the gains from entering into hierarchical security relationships (e.g., Lake 2009). Our recommendation of using SDP while also accounting for subsistence income as a separate variable|instead of total GDP|is potentially critical to scholars who currently use GDP as a proxy for the capacity of states to devote resources to non-military purposes, for instance, in work evaluating states' decisions to invest in education, healthcare, or other quasi-public goods (e.g., Baum 2001; Tanzi and Schuknecht 2000). Second, in the process of building comprehensive data on surplus and subsistence income we provide new data for some of the most widely used variables in political science and economics, 114 with coverage from 1800-2018. 1 In particular, we improve existing cross-sectional and tempo- ral data coverage for both GDP and population (the components of our new SDP indicator). Currently, cross-national data on military expenditure as a percentage of GDP go back to 1950 (Nordhaus et al., 2012). Our new data allow us to extend the data coverage for military expen- diture as a percentage of economic resources (SDP or GDP) back to 1816 for most of the world's states. Third, we use SDP to improve measurement of relative power-resources between states, allow- ing us to provide a more accurate identication of the most powerful and potentially threatening states in the world each year. Based on existing scholarship, which relies on GDP as a proxy for power-resources, China is ranked as the world's most powerful country in the early and mid-19th century. However, until the 1990s, nearly all of China's income was used to sustain its large, impoverished population and little surplus remained to invest in guns or butter. Using SDP to measure relative power-resources leads to a markedly dierent set of countries topping the rankings of powerful states. In the early 19th century, our estimates place the United Kingdom rather than China in the top spot, which comports much better with arming and power-projection behaviors of countries during this time period. Decomposing GDP into surplus and subsistence income thus provides new insights for scholars working on a broad range of topics related to the distribution of power such as arming, alliances, power transition, peace, and great power politics. While nancial resources are not the only dimension of a state's power-resources, they rep- resent a particularly important, if not the most critical, dimension (Beckley, 2018; Norrlof and Wohlforth, 2019; Zielinski, 2016). The importance of nancial resources led prior scholarship to use GDP as the primary measure of states' power-resources. Our claim is not that SDP measures every dimension of states' power-resources, or that other dimensions of power should not be con- sidered, but only that SDP outperforms GDP as an approximation for the income a state can devote to pursuing various objectives, which may include not only arming, but also investment 1 See Coyle (2014) for a discussion of the history of the statistic known as GDP. 115 in industrialization. As we demonstrate in this paper, SDP is more closely related than GDP to alternative measures of power-resources, such as industrial capacity-related components of the Composite Index of National Capabilities (CINC) data. We also develop and examine solely population-based measures of power-resources. We show that GDP measures are more closely related to population measures than SDP. In fact, in the early to mid-19th century, variation in GDP is almost entirely explained by variation in population|a re ection of the Malthusian constraint faced by most countries at the time. Thus, SDP captures a concept of power-resources distinct from GDP or population. Our new measure of SDP outperforms GDP in three validation exercises. First, SDP fares better than GDP when compared to alternative indicators that approximate states' relative power- resources. 2 Second, SDP more accurately ranks states with the greatest power-resources. We modify a recently developed operational strategy to incorporate SDP into measuring the level of potential threat in states' geopolitical environments (Markowitz and Fariss, 2018). Specically, we use SDP to measure relative power-resources between pairs of states and weight these ratios by distance and preference compatibility. Third, we show that weighted power-resource ratios between states based on SDP produce more valid rankings of countries each state nds most threatening, as compared to the same approach using GDP. For our primary empirical application, we generate an aggregate country-year measure of the potential threat each state faces in its strategic environment. We illuminate how the potential threat states face in uences the degree to which they invest scarce resources into arming and power projection capabilities. We demonstrate that when SDP is incorporated into this model, it outperforms the same model using GDP in predicting military investments. Strikingly, our model reveals that key variables suggested by existing theories of arming, such as relative power, geographic proximity, and preference compatibility cannot explain the degree to which states arm when power-resources are measured using GDP. However, these same variables can explain the 2 Following Beckley (2018), we dene a state's power-resources as the pool of resources a state can potentially invest in generating in uence. 116 degree to which states arm, when power-resources are measured using SDP and military burdens are measured using military expenditures as share of SDP. It is only the misuse of GDP as a measure of available power-resources that makes existing theories of arming appear empirically unsupported. Our results are robust to using naval tonnage relative to SDP as an alternative measure of arms levels. Decomposing GDP into surplus income and subsistence income allows us to examine patterns of SDP historically. We show that in the 19th and early 20th centuries, military burdens|the percentage of income devoted to arming|were, on average, higher than suggested by existing research. This dierence exists because, until recently, most states were able to generate only small amounts of economic surplus. States that generated surplus spent a large proportion of it on arming. As a result, many states in this earlier period had military burdens as high as 25%{50% of SDP. To put this in perspective, even during the Cold War, the United States spent only about 10% of its SDP on the military. Finally, SDP reveals that military burdens have fallen faster than previously realized. Newly industrialized states like China are seeing large gains in economic surplus, which they could invest in arming. Yet, when measured as a percentage of SDP, the military burdens of these states came tumbling down over the last several decades. Today, most governments are prioritizing butter over guns and, as a result, military burdens as a percentage of SDP are far lower than in the past. However, for the least developed states, military burdens are still much higher than previously realized. 4.2 Subsistence and Surplus Scholars long recognized the close relationship between states' income and the resources available for military use. For example, Sandler and Hartley argue that \as GDP rises, a nation has both more resources to protect and greater means to provide protection." (Sandler and Hartley, 1995, 117 60) However, scholars acknowledge that states' ability to generate military power depends not only on the size and sustainability of its resource base, but also on the degree to which the state is constrained from extracting and mobilizing those resources (Sandler and Hartley, 1995; Lamborn, 1983; Milward, 1977). Our argument and measure of SDP build on these insights. While most previous research highlights how domestic political factors constrain the amount of resources a state can extract, we focus on how biological factors limit the amount of resources that are potentially available for extraction. Political constraints are important, but we cannot accurately estimate their impact without rst creating a measure of the resources that could, given the political will, be extracted by the state. We argue that when estimating the amount of resources that are potentially available for extraction and arming, scholars should account for a state's surplus and subsistence income sepa- rately, rather than adding them together. Surplus income (SDP) is calculated by removing from GDP the resources the population must consume to survive. Biology determines the number of calories required for survival and this lower caloric bound is largely stable across time and space. If citizens do not survive, they cannot use their labor to produce income. It is possible for states to extract subsistence income from their population, and states some- times do, particularly in times of crisis. However, this level of predation results in the population growing sicker and weaker|decreasing their ability to produce income and consequently reducing the resources available for the state to extract. Thus, the subsistence needs of the population place an upper bound to the amount of income states can sustainably extract. Resources remaining after subsistence needs of the population are met represent surplus income that can potentially be extracted by the government. Minimum subsistence is a key concept in the dual sector model of development. In developing economies that are characterized by surplus labor in the agricultural sector, the minimum subsis- tence level, or average product of farmers, determines the wage at which surplus labor is available 118 for employment in the manufacturing sector (Lewis, 1954, 189). 3 As capital accumulation occurs, surplus labor transfers to the manufacturing sector where higher wages can be earned. Only once surplus labor is exhausted will wages rise above the existing rates. Due to surplus labor, aggregate income in the agricultural sector remains constant throughout the process of economic development (Lewis, 1954, 157). Increased capitalist prots generate a higher aggregate national income over time and the proportion of national income from manu- facturing prots exhibits an upward trajectory. Our proposition that states may extract income from their population only up to the minimum subsistence level holds in economies characterized by surplus labour, because the computation of SDP is agnostic as to to how national income is distributed among the sectors of the economy. How do we estimate subsistence needs? The World Bank monitors health and wellbeing associated with extreme poverty at several thresholds. We argue that a $3-per-day threshold, which the World Bank calls \close to extreme poverty," is a conservative estimate of the subsistence resources required per capita. 4 It is conservative because even at $3 per day, people tend to be chronically malnourished. As a result, they are more likely to succumb to disease and generate less surplus income. The World Bank estimates that in low-income countries (dened as countries having a per-capita gross national income of $1095 or just under $3 per day), 27% of the population were undernourished in 2015 (World Bank, 2017). Moving from a $3 to a $1.90-per-day threshold is associated with chronic malnourishment|causing approximately 10% to 40% of children under the age of ve to be underweight (Ezzati, 2004, pp. 1949 and 1985). We use the $3-per-day threshold to calculate the SDP available for the state to extract and spend on public or private goods|in particular, arming and power projection capabilities|and the remaining subsistence income. While it is possible for states to extract subsistence income via 3 Lewis (1954) does not provide a numerical value for the minimum subsistence level. 4 The thresholds are calculated in constant 2011 purchasing power parity dollars. The World Bank's poverty threshold underwent adjustments over time|the threshold of $1.90 today is equivalent to $1.08 in 1993, and $1.00 in 1985 (Ezzati, 2004). The statistic on malnourishment referenced here was originally calculated based on the $1.08 threshold and adjusted to re ect the World Bank's new poverty threshold of $1.90 (Ferreira et al., 2015). 119 taxation, we argue that the costs of doing so are very high and that it cannot be done sustainably. Taxing the population to below the subsistence threshold undermines economic productivity in even the very near future. Biological constraints are not the only limits states face in taxing their population (e.g., Lamborn 1983; Milward 1977). Historically, most states lacked the political will or capacity to extract the entire surplus. SDP estimates the upper bound on the resources a state can sustainably extract if it has the capacity and political will to do so. To compute SDP for each state i in year t, we rst calculate the minimum subsistence value v it , which is the level of income necessary to sustain the state's population, and then use this value to determine how much surplus income remains. We let v it = [(365)Population it ]; where is the per-day, per-person subsistence threshold. Based on our discussion above, we use a subsistence threshold of $3 per-day per-person. We also assess the sensitivity of our results to per-day subsistence thresholds at $3, $2, $1, and $0 (standard GDP). If GDP it >v it , then SDP it = GDP it v it and subsistence it = v it . If GDP it v it , then SDP it = 0 and subsistence it =GDP it . For a state to have surplus income, it must generate enough subsistence income to meet the needs of the population. If the state's income does not exceed this minimum surplus value, it has a SDP of zero and only has subsistence income to work with. We thus decompose GDP into surplus income (SDP) and subsistence income (see Supplementary Appendix Section 4.B for further details about this formalization). We account for both income values in the measurements and regression models we develop below. 120 4.2.1 Converging and diverging trends in the international system: GDP vs. SDP Having decomposed GDP into surplus income (SDP) and subsistence income, we ask: Is SDP a more valid measure of power-resources than GDP? As a rst concurrent validity test, we compare temporal trends in states' shares of global economic power for GDP versus SDP. 5 We demonstrate that SDP produces a better representation of historical trends and a more valid list of the global top ten powers than GDP for the past 200 years. Second, we demonstrate the benet of using surplus as a measure of power-resources by showing that SDP correlates more strongly with alternative indicators of power-resources than GDP. Figure 4.1 displays the evolution of states' shares of global power-resources for six countries, for SDP and GDP. Using SDP, wealthier states like the United States and the United Kingdom are estimated to have a much higher share of global power-resources than when using GDP. 6 The opposite is true for most poorer states. For countries like China and South Korea, measuring power-resources via GDP overestimates relative power-resources prior to 1980, and underestimates the rate of their rise since. The eect of economic development is illustrated by the case of South Korea. Prior to its burst of economic growth, the county's share of global GDP was substantially higher than its share of global SDP. Beginning in 1985, this trend reversed and South Korea's share of global SDP started to become larger than its share of global GDP. Figure 4.2 displays the top ten powers based on their average share of global SDP or GDP for four time periods. The dierence is most striking for China. In the 19th century, the economic income produced by hundreds of millions of peasant-farmers dwarfed the production of other states. Based on its share of global GDP, China would be considered the most powerful country in the period from 1816 to 1869. However, because most of this income was immediately consumed 5 Supplementary Appendices 4.C and 4.C present formal validity test descriptions. 6 Figure 4.15 in the Supplementary Appendix provides scatterplots of SDP versus GDP for select years. Figure 4.16 in the Supplementary Appendix illustrates the relationship between SDP and per capita GDP and demon- strates that while economic development increases the correlation between SDP and per capita GDP, individual countries vary considerably regarding the strength of this relationship. 121 Figure 4.1: Evolution of economic power-resources based on a state's share of global SDP versus GDP. SDP is based on a $3 per diem subsistence level. A state either has positive surplus or no surplus at all. If a state has no SDP, then it would need to extract from the available subsistence resources, which reduces the ability of individuals to produce economic surplus. 122 Figure 4.2: Top 10 powers ranked by their average share of global SDP or GDP. With the exception of China and a few other powers, the membership in the top 10 club is similar between the two measures of economic income. What does change is the rank-ordering of the countries. SDP produces a more historically valid ranking of the great powers than GDP. for subsistence needs, China had little surplus wealth to invest in arming or power projection. SDP takes this into account and does not place China in the top ten powers in the early 19th century. These cases provide evidence that SDP produces a more valid ranking of great powers than GDP over several historic periods. For a second set of validations, we assess convergent validity by comparing a country's share of global SDP to several CINC component variables. While CINC has well-known drawbacks (Beckley, 2018; Kadera and Sorokin, 2004), 7 its components are useful for comparison as an established, widely-used source of variables related to states' economic power. Because SDP and GDP are indicators of economic resources, we assess them in relation to four CINC variables that measure resources that could potentially be invested in military capabilities. The remaining 7 CINC's restrictive approach toward including countries as members of the international system leads to dis- agreements in the power estimates for some years (Supplementary Appendix 4.D). 123 CINC components, namely military expenditure and personnel, are related to actual military capabilities, which we use as a dependent variable to measure military investment below. (a) CINC industrial capacity variables (b) CINC population variables Figure 4.3: Yearly correlation coecients with 95% condence intervals. In each panel, we assess the degree to which a state's share of global SDP and GDP correlate with each of four com- ponent variables of CINC. Note that we utilize an alternative to CINC's population measure (Supplementary Appendix 4.D). 124 In the top panel of Figure 4.3, we display correlations between a country's share of global SDP or GDP and two components of CINC, which are yearly shares of global (1) iron and steel production and (2) primary energy consumption. 8 The year-by-year correlation coecient with SDP ranges between 0.69 and 1 for iron and steel production, and between 0.62 and 1 for primary energy consumption. Until World War II, a state's share of global SDP correlates more strongly with related measures of power-resources than a state's share of global GDP. Between 1860 and 1960, the 95% condence intervals do not overlap|suggesting a statistically signicant dierence between GDP and SDP. In the bottom panel of Figure 4.3, we present analogous graphs for the correlation between the yearly share of SDP or GDP with a country's share of global (3) urban and (4) total population. 9 The contrast with the industrial capacity variables is striking: GDP is more closely related to a country's share of global population than SDP. The discrepancy between the two series changes over time. In the early 19th century, population is perfectly predictive of GDP, while SDP is not at all correlated with a country's population. In pre-industrial years, and extending into the World War II period for many states, national wealth is primarily a function of how many citizens a government can exploit. The surplus resources most countries can extract from their population pre-1800 are nearly zero. Thus, the development of a force structure that goes beyond feeding soldiers and obtaining basic equipment is cost-prohibitive for most states. The Industrial Revolution changed this pattern because states began to produce economic wealth beyond the basic subsistence needs. These resources were, in turn, invested into the equipment and technology necessary to project force over distance. The historical patterns depicted in Figure 4.3 demonstrate that accounting for subsistence income is crucial for the study of global power relationships, particularly in historical international 8 The break in the correlation between industrial capacity and GDP in 1860 is caused by China entering the CINC data set (with an approximate global population proportion of 47% according to CINC and 31% according to our computation). 9 To alleviate concerns regarding the coverage of CINC's total population gures, we use our extended series of population estimates (Supplementary Appendix 4.D). 125 relations analyses. By shifting the conceptual focus from gross to surplus domestic product, we can more accurately identify which states possess the greatest power-resources and potential to generate the military capabilities to threaten other countries. States with low surplus economic resources are largely incapable of projecting power, even if their total GDP is large. Next, we show how SDP, compared to GDP, represents a more valid measure of relative power- resources between pairs of states. This dyadic-level measure, along with additional information about the state-pair in question, allows us to assess the level of potential threat one state might expect from its interactions with other states. 4.3 Measuring Relative Power-Resources and Potential Threat To date, international relations scholarship had diculty explaining patterns of arming. Even though theoretical expectations suggest that states arm in response to threats in the interna- tional system, empirical results are inconclusive (Zielinski et al., 2017; Nordhaus et al., 2012; Sandler and Hartley, 1995). Our argument suggests that inconclusive results emerged because prior scholarship mis-measured power-resources by relying on GDP. Our SDP-based approach addresses this systematic bias and reveals new patterns in line with theoretical expectations. States' eorts to arm, i.e. their military expenditures, should be scaled by the potential surplus resources they have available to arm. While previous scholars scaled military expenditures by GDP, we are the rst to argue they should instead be scaled by SDP. Our ndings demonstrate that a strong relationship between states' threat environment and eorts to arm exists, but only when military expenditures are scaled by SDP. Relative power-resources between two states is the key construct we utilize to understand patterns of arming and power projection. If military capabilities represent the latent power to hurt, then power-resources represent the latent power to arm, or states' military potential (Kennedy, 126 1987). To account for the fact that not all dyadic economic power relationships are created equal, we additionally consider how other dyad-level features, specically shared preferences and the loss of strength gradient, mitigate the level of threat associated with dierences in relative power-resources between states. 4.3.1 Measuring the dierence in relative power-resources between states using SDP For each country-year unit, i = 1;:::;N indexes states and t = 1;:::;T indexes years. For every country-year unit, we assess information for each dyadic relationship between statei and all other states in the international system that year, indexed byj = 1;:::;J. For every statei, we measure i's annual surplus economic resources, SDP it , as well as the surplus resources of each opponent, SDP jt . For eachij pair, we compute power-resource ratiosr ijt as the proportion of the opponent state's SDP relative to the total SDP in each dyad, such that r ijt = SDPjt (SDPjt+SDPit) . The relative power-resource variable r ijt is bounded between 0 and 1. When SDP j t is large, the relative power-resources of j compared to i will be greater than 0:5 and represent a state that is potentially threatening to i, because j has more resources to invest in arming and power projection. The least powerful state's power-resource ratio will be close to 1. The most powerful state's power-resource ratio will be close to 0. If two states have equal power-resources, they nd each other equally threatening and the power-resource ratio is 0:5. These ratios allow us to compare the relative power-resources of one dyad to the relative power-resources of another dyad, thus relativizing these comparisons and facilitating further measurement aggregation. 127 4.3.2 Relative power-resource relationships between one state and all other states The relative distribution of power-resources between two states is useful for dyadic-level analyses. However, we also want to understand how individual states respond to the total level of potential threat they face from all potential opponents in the international system. We dene a country's level of potential threat as its expectations regarding other states' potential ability to harm it. All else equal, the more power-resources a state has, the more potentially threatening it is. States shift from potentially threatening to actually threatening when they engage in behavior that is perceived as harmful, such as arming, coercing, or attacking. Our goal in this paper is to measure the degree to which a given state is potentially threatening based on its relative power-resources, rather than actually threatening based on its actions. To measure the level of potential threat, we follow an existing operational procedure to cre- ate an aggregated country-year measure using each of the dyadic power-resource ratios. 10 This measure is simply the sum of all power-resources ratios, weighted by the loss of strength gradient and preference compatibility, between each state i and all opponent states j. This generates a country-year measure that summarizes all dyadic relationships of each country in each year and represents the unique level of potential threat each state faces. We call this new variable potential threat it , which is the sum of weighted relative power-resource ratios that relate state i to every opponent state j in year t. threat it = J X j=1 (r ijt p ijt w ijt ) r ijt denotes relative power-resources between two states. Following Markowitz and Fariss (2018), we include weights for preference compatibility between two states p ijt , and the loss of strength gradient between two states w ijt . The rst weight p ijt is based on the preference 10 We modify the formalization by Markowitz and Fariss (2018) to better capture the intuition that the global environment might be transitioning into an increasingly competitive space (Supplementary Appendix 4.F). 128 compatibility between states. p ijt = 0 if state i and state j have compatible preferences with each other in year t; otherwise, p ijt = 1 when states do not share compatible preferences. We operationalize joint democracy as our indicator of compatible preferences using the Polity IV data (Marshall et al., 2016). If both states have a Polity2 value greater or equal to 6, they have compatible preferences and are not threatening to one another|and thus coded 0. If a state has incompatible preferences with other countries, they are potentially threatening, especially those with the greatest power-resources. We consider alternative measures of preference compatibility in the supplementary appendix (Figure 4.35), which include other measures of joint democracy (Boix et al., 2013; Pemstein et al., 2010), defense pacts and alliances (Gibler, 2009; Leeds et al., 2002), United Nations General Assembly (UNGA) voting similarity (Bailey et al., 2017), rivalry (Klein et al., 2006), energy consumption (Markowitz et al., 2019; Greig and Enterline, 2017), bilateral trade (Barbieri and Keshk, 2012; Barbieri et al., 2009), diplomatic exchange (Bayer, 2006), and shared Intergovernmental Organization (IGO) membership (Pevehouse et al., 2004). The inverse logged distance between state capitals creates a second weight, w ijt = 1 ln(dijt) , which operationalizes the loss of strength of state j if it were to attempt to project power into state i. Short distances d ijt between the capital cities of two states yield values closer to 1 and further distances yield values closer to 0. Substantively, the operationalization captures the total power-resources of state i relative to all of the opponent states j in the international system, weighed by distance and preference compatibility. In Supplementary Appendix 4.F, we provide additional detail, a visual guide, and examples for each of the operational steps that generate the country-year values for the potential threat it variable. 129 4.4 Potential Threat using SDP vs. GDP In this section, we compare the performance of SDP-based and GDP-based measures of potential threat. First, we compare the ability of distance- and preference-weighted power-resource ratios to correctly categorize country-year units as potentially threatening (concurrent validity). Next, we use SDP-based and GDP-based potential threat measures to explain variation in dependent variables that measure arming and power projection. We show that SDP-based models of states? military investments conform better with existing theoretical expectations than models based on GDP. 4.4.1 Assessing the most potentially threatening states using SDP vs. GDP Recall that the potential threat measure is the summation the distance- and preference-weighted power-resource ratios for all pairwise relationships for each country-year unit. If SDP better measures relative power-resources than GDP, then the potential threat variable using SDP should more accurately rank countries of greatest concern to other countries. As a control variable, we also generate a potential threat variable that uses population to measure relative power-resources within each pair of states. We show that the results using the population-based measure are very similar to the GDP version, which both contrast with SDP. This provides further evidence that poor, populous countries appear potentially threatening only because their large population gives them a large GDP. Figure 4.4 displays the top ten potentially threatening states within the strategic environment of the United States by decade. 11 The upper panel illustrates the ranking based on distance- weighted relative power-resources using SDP; the middle panel plots the ranking for an analogous measure using GDP; the lower panel shows this ranking for the same measure using population. 11 For graphs for Japan and the United Kingdom, see Supplementary Appendix Figures 4.11 and 4.12. 130 Figure 4.4: Top 10 potentially threatening states for the United States by decade for $3 per diem subsistence level relative SDP (upper panel), relative GDP (middle panel), and relative population (lower panel). Power-ratios are weighted by the inverse of the logged distance between Washington D.C. and the other state's capital. Countries that are potentially threatening to the United States are not democratic, which is denoted through darker shading (not jointly democratic). 131 States with the largest weighted power-resource ratios rank highest on the list as the adversaries that potentially threaten the United States. The upper panel's order of states diers substan- tially from the middle and lower panels. The distance-weighted relative power-resource measures incorporating GDP or population produce similar rankings and place countries that were unlikely to threaten the United States at the top. In particular, China is ranked as the most potentially threatening country for the United States during the entire 19th and early 20th centuries|a pe- riod China could not develop a military force structure capable of projecting force to the shores of the United States. The measure using SDP provides a more historically valid ranking of states posing a potential threat to the United States than the measures using GDP or population. Using a lighter shading of tiles for states with compatible preferences, Figure 4.4 highlights the role of preference compatibility for the United States' assessment of potential threats. Recall that the p ijt component of the potential threat variable down-weights power-resource ratios of jointly democratic dyads. The ranking corresponds to the top ten potentially threatening states for the United States based on economic might and geographic proximity, but this potential threat is mitigated if the country is democratic. Though they might have the economic capability to project power abroad, democratic states are not considered threatening by the United States because of the compatibility of their preferences. As former geopolitical rivals democratize, they stop contributing to the total potential threat faced by the United States. As a result, the strategic environment of the United States has become less threatening over time. Figure 4.5 illustrates this downward trend. The height of each bar denotes the total level of potential threat faced by the United States each year. The potential threat faced by the United States fell sharply over the course of the 19th century and remains much lower today than in the past. Colored values indicate how much China and Russia contribute to the total level of potential threat for the United States. The left panel plots potential threat incorporating GDP as an indicator for economic resources; the right panel illustrates the same measure using SDP. 12 12 Preference compatibility is measured using Polity2 scores, and supplemented with data from Boix et al. (2013) to reduce the number of missing values. 132 Figure 4.5: China's and Russia's contribution to the total potential threat faced by the United States for SDP versus GDP. Preference compatibility is measured using Polity scores, and sup- plemented with data from Boix et al. (2013) to reduce the number of missing values. The dierence between measuring power-resources via SDP versus GDP for the United States' threat assessment is striking. Historically, when the United States considered potential threats, it paid careful attention to states with the greatest power-resources. For over two hundred years, China was one of the largest economies, but only recently became one of the states with the greatest power-resources in the world. SDP yields a more historically valid representation of countries' contributions to the total level of potential threat in the United States' geopolitical environment. Today, China is the largest contributor to the total potential threat faced by the United States. In fact, contemporary China makes up a larger proportion of the total potential threat faced by the United States than Russia did at the height of the Cold War. The rise of China as a major economic power in the late 20th century dramatically increased the total potential threat that the United States experiences in its geopolitical environment. 133 4.4.2 Modeling arming and power projection As previously discussed, while international relations theory suggests that the level of potential threat states face should explain their eorts to arm, existing research nds only mixed empirical support for this proposition (Zielinski et al., 2017; Nordhaus et al., 2012; Sandler and Hartley, 1995). We suggest a potential resolution for this puzzle by demonstrating that once we scale military expenditures by SDP, which corrects for systematic measurement error inherent in using GDP, we nd a strong relationship between potential threat and military burdens. SDP not only does a better job of measuring the distribution of relative power-resources|a core component in the level of potential threat states face|it also does a better job of measuring the power-resources a given state could invest in arming. The example of China in 1990 illustrates this point. Military expenditures represented approximately 2.5 percent of China's GDP|a modest military burden. However, even if China's 1.135 billion citizens in 1990 could survive on just $2 per day and the state could seize the entire remainder of economic income, SDP would be half the value of its GDP. Hence, when military burden is measured using military expenditures as a percentage of SDP instead of GDP, China's military burden in 1990 was approximately twice as large as previously estimated. We apply our potential threat measure to investigate the relationship between threat and mil- itary burden; comparing SDP-based and GDP-based approaches. We estimate regression models in which we vary the measurement of SDP at $3, $2, and $1 per-day subsistence thresholds and compare those to GDP (equivalent to a $0 per-day threshold). We also control for states' subsis- tence income at these thresholds. We assess the relationship between the level of potential threat a state faces in its strategic environment (explanatory variable) and two measures of arming (dependent variables). The rst dependent variable is military burden, operationalized as military expenditure relative to income. Fearon (2018) argues that this is a reasonable proxy for states' resources that could 134 be dedicated to arming, and captures the magnitude of the social welfare costs of arming. In contrast to Fearon and others (Rasler and Thompson, 1985; Khanna et al., 1998), our preferred measure of military burden is a state's military expenditure as a proportion of the state's SDP, rather than GDP or Gross National Product (GNP). This operational choice better approximates surplus resources available for arming. We employ a revised and extended series of military expenditure as a proportion of income. This new indicator is created by rst converting military expenditure values from CINC into constant monetary units (Singer, 1987). We then use new GDP and population estimates to measure the proportion of a state's income (SDP or GDP) devoted to the military. 13 This allows us to extend data coverage to cover most countries in the world from 1816 to 2012. As a robustness check, we assess the relationship between potential threat and a second dependent variable that captures states' investments in arming: power projection capabilities. We operationalize power projection capabilities via states' naval tonnage relative to income. Data on naval tonnage come from Crisher and Souva (2014). 14 States with higher military spending as a proportion of income have higher military burdens, as do states that have more naval tonnage relative to their income. Figure 4.6 displays coecients and 95% condence intervals of standardized potential threat variables with and without the subsistence income control variable. 15 For each dependent variable, we estimate a series of country-year xed-eect regression models. Right-hand side variables are lagged by one year. All models include controls for the natural log of income (SDP or GDP), Polity2 score, and a measure of potential threat based on population. When applicable, we control for the natural log of subsistence (or population for GDP models). 13 Supplementary Appendix 4.I describes in detail new estimates of GDP and population. 14 Supplementary Appendix Figure 4.27 compares temporal trends of dependent variables. 15 Table 4.1 contains regression results based on SDP for the $3 per-day subsistence threshold. Table 4.2 contains analogous results using GDP. For most models, we do not observe statistically signicant or substantively mean- ingful interaction eects between potential threat based on SDP (or GDP) and population. We therefore limit the results presented in Figure 4.6 to additive model specications. In the Supplementary Appendix, we demonstrate that the results are robust to excluding the control for population-based threat (Figure 4.31), omitting controls (Figure 4.32), and to limiting observations to the post-World War II period (Figure 4.33). 135 Figure 4.6: Coecients and 95% condence intervals of standardized potential threat variables for regression models of two dependent variables|the military expenditure index and naval tonnage index|on potential threat and control variables. See Tables 4.1 and 4.2 for model specication information. Our preferred SDP-based approach to measuring power-resources using a $3 dollar-per-day threshold produces statistically signicant relationships between the level of potential threat states face and both measures of arming, while a GDP-based approach does not. As we ad- just our measure of SDP to use lower subsistence levels, the coecients become smaller and cease to be statistically signicant at conventionally accepted levels. For many models, decreas- ing the subsistence level to $1, or switching to GDP, renders the eect of potential threat on military investments negative and insignicant for arming and power projection. Controlling for population-based potential threat increases the size of the estimated eects. Overall, higher levels of potential threat are associated with larger investments in military and naval capabilities when measuring economic power-resources using SDP. Crucially, all results depend on accurately measuring economic resources that states have at their disposal to invest in guns or butter. The conventionally used GDP measure does not yield a statistically signicant 136 association between the level of potential threat and arming|a result that runs counter to ex- isting theoretical expectations, but is consistent with the mixed empirical ndings in the current literature. Only when measuring power-resources via SDP can we explain arming decisions based on the level of potential threat in states' geopolitical environment. 4.5 Evaluating Military Burdens Figure 4.7 illustrates the evolution of states; military burdens. Scaling military expenditures by SDP rather than GDP reveals that, historically, military burdens were higher than existing research suggests (Fearon, 2018). In particular, many Asian states that spent a relatively small percentage of their GDP on the military were, in fact, laboring under extraordinarily high military burdens|spending 25% to 50% of surplus. Figure 4.7: Change in military burden over time for regions: the Americas (including the US and Canada), Europe (including Russia), Asia, the Middle East and North Africa, and Sub-Saharan Africa. Lines represent the smoothed average over all countries in the region for two indicators of military burden: military expenditure as a proportion of SDP versus as a proportion of GDP. Figure 4.8 shows these trends for select countries. Scaling by SDP reveals that military burdens of poor states are much higher than the conventionally used measure of military expenditure as 137 a percentage of GDP suggests. This divergence is particularly large for poor, populous countries like China. Over time, military burdens fall for most states|especially for major powers|but these costs remain high for poor states where most GDP is needed to cover basic subsistence and SDP is low. Figure 4.8: Change in military burden over time for select countries. The good news for states in the Western Hemisphere and Europe is that military burdens are much lower than in the past. Additionally, despite alarmist warnings of impending arms races and con ict, most states in Asia today face dramatically lower military burdens than they did in the past two hundred years. However, this decline becomes apparent only upon using SDP instead of GDP. For illustration, as a share of GDP, South Korea's military spending decreased only slightly 138 from an average of 3.7% during the Cold War (1954{1991) to 3.0% after (1992{2012). However, as a share of SDP, military spending plunges from 9.9% during the Cold War to 3.2% after. Scaling by SDP reveals a sharp decline in South Korea's willingness to prioritize guns over butter and implies that leaders in Seoul believe that the level of threat they face fell enough to justify lower military burdens (Lind, 2011). While South Korea may choose to increase its military burden in the future, they will do so from a historically low baseline. This point informs the debate over the degree to which states are balancing China's rise; in recent years, the willingness of states in the region to bear high military burdens is generally lower than commonly recognized. 4.6 Conclusion GDP is a widely adopted measure of the nancial resources that states can potentially invest in guns or butter (Coyle, 2014). We introduce the concept of SDP, which separates the subsistence income, or `bread,' needed for the population to survive from the surplus income, or `butter,' that can potentially be extracted and invested. Using GDP as a measure of power-resources instead of SDP systematically overestimates the economic resources available to governments of poor, populous countries and underestimates the speed with which these resources increase during the early stages of industrialization (i.e. during the stage when countries rst begin to produce signicant surplus and governments could extract income without starving citizens). Similarly, using military expenditures as a share of GDP to measure military burdens leads scholars to underestimate the size of the military burdens born by poor states. These conceptual errors are particularly problematic for historical-comparative work because, for most of human history, virtually all states had incomes at or near subsistence levels|at least on average | creating a large divergence between estimates of SDP and GDP. This is a major issue. For illustration, as recently as 2007, the gap between SDP and GDP was still large for over half the world's states, which were classied as either low income (Gross National 139 Income (GNI) per-capita below $995) or lower-middle income (GNI per-capita between $996 and $3895, World Bank 2018). At the time, more than 70% of the world's population lived in such states. Thus, using new data on SDP, we reveal that poor, populous states are far less powerful than generally assumed, and that low-income countries historically and today face more severe guns-butter tradeos and higher military burdens than GDP-based measures suggest. Previous scholarship dramatically underestimated the benets of factors that allow poor states to lower defense burdens, such as hierarchy and the liberal peace. Additionally, our results oer a potential solution to the puzzle of why previous scholarship found only mixed support for one of the core propositions of international relations theory: States arm against potential threats. Once we apply SDP to correct for the systematic measurement error associated with GDP, we nd strong empirical support for this proposition. In addition to our theoretical contributions, we provide new data that extend cross-national coverage of GDP, SDP, and population back to 1816 for nearly every country. These new data allow scholars to apply our measure of SDP to reexamine a broad range of research questions in which GDP is frequently used as a proxy for potential or actual state capacity. An area directly parallel to states' capacity to arm is states' capacity to repay debts. For illustration, Malawi in 2015 had a plausibly manageable debt burden of 39.5% of GDP (World Bank, 2017; Reinhart and Rogo, 2010). However, Malawi's debt burden amounts to a crushing 226.5% of SDP. The annual payments on a debt of that magnitude consume a signicant propor- tion of the country's surplus income, even if they constitute a seemingly manageable share of total economic income. This matters for understanding the ability of states to manage debt and engage with international institutions like the World Bank and the International Monetary Fund. More broadly, SDP represents the resources a government may potentially draw upon to build physical infrastructure, establish the rule of law, or provide public services such as education and health 140 care. SDP is not a direct measure of government capacity, but it measures the upper bound of the income states may sustainably extract from citizens to develop that capacity. When measuring the upper bound of states' extractable income, SDP compared to GDP, showcases just how constrained state capacity is in low-income countries, compared to middle- income and higher-income countries. However, SDP also reveals that, in the early stages of economic development, government capacity expands more rapidly than currently realized | increasing the compound returns to growth for countries near subsistence levels. SDP-based assessments are unlikely to lead to exclusively pessimistic or exclusively optimistic new conclusions regarding the prospects for peace, prosperity, and democracy in the developing world. Thus, the analysis of SDP has the potential to radically reshape our understanding of the extent to which dierent scal strategies are plausible or desirable in lower-income countries. 141 Table 4.1: Regression models relating dierent specications of the potential threat variable to investments in arming and power projection. Power-resources are measured using SDP at a $3 per diem subsistence level. The loss of strength gradient is conceptualized as curvilinear using the formula 1 log(distance) . Interest compatibility based joint democracy using Polity scores. Dependent variables Military expenditure/SDP Naval tonnage/SDP (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Potential threat (SDP) i;t1 0.62 0.26 0.72 0.72 1.87 1.60 2.53 2.87 (0.09) (0.10) (0.16) (0.15) (0.49) (0.55) (1.00) (1.15) Potential threat (Population) i;t1 0.05 0.74 0.76 0.82 1.64 0.19 (0.13) (0.21) (0.25) (0.50) (1.04) (1.08) ln SDPi;t1 0.08 0.09 0.06 0.06 0.04 0.09 0.01 0.01 (0.01) (0.01) (0.01) (0.01) (0.06) (0.06) (0.06) (0.06) ln Subsistencei;t1 0.21 0.19 0.38 0.38 1.95 2.45 1.54 2.03 (0.18) (0.17) (0.18) (0.17) (1.00) (0.91) (1.02) (0.99) Polity2 i;t1 0.03 0.04 0.04 0.04 0.01 0.02 0.002 0.04 (0.01) (0.01) (0.01) (0.01) (0.04) (0.05) (0.05) (0.04) Interaction Potential threati;t1 0.01 1.28 (0.08) (0.43) Fixed-eects CY CY CY CY CY CY CY CY CY CY Observations 11,616 11,616 11,616 11,616 11,616 12,033 12,033 12,033 12,033 12,033 Adjusted R 2 0.02 0.09 0.09 0.11 0.11 0.01 0.02 0.01 0.03 0.05 Note: p<0.05; p<0.01; p<0.001 Clustered standard errors by country (Satterthwaite correction) in parentheses. Potential threat variables are standardized. CY denotes two-way xed eects. Period of observation: 1816-2012. Table 4.2: Regression models relating dierent specications of the potential threat variable to investments in arming and power projection. Power-resources are measured using GDP. The loss of strength gradient is conceptualized as curvilinear using the formula 1 log(distance) . Interest compatibility based joint democracy using Polity scores. Dependent variables Military expenditure/GDP Naval tonnage/GDP (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Potential threat (GDP) i;t1 0.23 0.05 0.36 0.59 0.23 0.04 0.37 0.23 (0.18) (0.15) (0.28) (0.39) (0.14) (0.14) (0.51) (0.49) Potential threat (Population) i;t1 0.01 0.36 0.18 0.03 0.39 0.59 (0.19) (0.39) (0.32) (0.16) (0.56) (0.61) ln GDPi;t1 0.05 0.03 0.17 0.22 0.01 0.01 0.14 0.13 (0.25) (0.25) (0.31) (0.32) (0.23) (0.21) (0.33) (0.32) ln Population i;t1 0.17 0.15 0.32 0.25 1.59 1.58 1.76 1.83 (0.18) (0.20) (0.26) (0.23) (0.32) (0.31) (0.38) (0.38) Polity2 i;t1 0.05 0.04 0.05 0.05 0.02 0.02 0.02 0.02 (0.01) (0.01) (0.01) (0.02) (0.01) (0.01) (0.01) (0.01) Interaction Potential threati;t1 0.23 0.18 (0.19) (0.10) Fixed-eects CY CY CY CY CY CY CY CY CY CY Observations 11,616 11,616 11,616 11,616 11,616 12,033 12,033 12,033 12,033 12,033 Adjusted R 2 0.03 0.01 0.01 0.01 0.002 0.02 0.08 0.08 0.08 0.08 Note: p<0.05; p<0.01; p<0.001 Clustered standard errors by country (Satterthwaite correction) in parentheses. Potential threat variables are standardized. CY denotes two-way xed eects. Period of observation: 1816-2012. 142 4.7 Acknowledgements This chapter is co-authored with Jonathan N. Markowitz and Christopher J. Fariss and has been accepted for publication at International Studies Quarterly, a peer-reviewer political science journal. Earlier versions of this chapter were presented at the Annual Meeting of the American Political Science Association in Boston, August 30{September 2, 2018, the Annual Convention of the International Studies Association in San Francisco, April 4{7, 2018. Special thanks go to Benjamin Graham, Andrew Coe, and James Lo for helpful comments and suggestions. All data and code for this project will be made publicly available at the time of publication. All authors contributed to the design of the study, data collection and analysis, as well as the preparation of this manuscript. 143 Appendix 4.A Introduction The supplementary material provides additional graphs and details about both the construction of the Surplus Domestic Product (SDP) and subsistence measures, the potential threat measure, as well as the the latent variable model developed in Chapter 4. The main manuscript makes reference to the materials contained here. The estimates presented in this appendix along with the code necessary to implement the models in R will be made publicly available at the time of publication. 4.B GDP = surplus + subsistence Gross Domestic Product (GDP) is technically an accounting identity, made up of four component parts, such thatGDP =consumption +investments + (importsexports). Our decomposition of GDP into surplus and subsistence is also an accounting identity. Subsistence is technically part of the consumption component of the GDP accounting identity. Surplus is also part of the consumption component in addition to the other component parts of the identity. For country i2f1;:::;Ng in yeart2f1800;:::; 2018g, the equation for Gross Domestic Product as an additive identity of two income components is: 144 GDP it =surplus it +subsistence it Before we dene surplus it income and subsistence it income, we rst have to dene the mini- mum surplus value: v it for country i in year t, which is calculated as: v it = 365population it where 2f$0; $1; $2; $3g is the daily surplus threshold. Conceptually, represents the min- imal amount of income necessary for an individual to meet her caloric needs. As we describe in detail in the main manuscript, in the contemporary period, it is at least $2 in constant United States (US) dollars and the World Bank recommends $3 in constant US dollars. 16 The variable subsistence it takes on positive dollar values that are less than or equal to the surplus value v it such that: subsistence it = 8 > > > < > > > : v it if GDP it >v it GDP it if GDP it v it The variable surplus it takes on positive dollar values only if GDP it is greater than the value of the surplus value v it such that: surplus it = 8 > > > < > > > : GDP it v it if GDP it >v it 0 if GDP it v it In the paper we refer to surplus it as surplus domestic product SDP it . 17 16 The value of has likely changed over time. In a future project, we plan to try to estimate this value based on historic information about the subsistence behaviors of individuals living and working in dierent periods of time and dierent countries. Such a measurement project is outside the scope of the current paper. Thus, we opted to set to one of four dierent constant values that we use in our statistical models. 17 It is likely the case that many of the states without surplus income are still importing and exporting some goods and making some investments. However, to do this, the state must extract from the basic subsistence income of the citizens. 145 Next we dene the level of investment in military expenditures that a state makes each year. This is an important quantity that international relations theorists demonstrate is related to milex ratio it . This is our main dependent variable. We calculate this in two ways. Both are ratios of the dollars spent out of all the available income to be spent by the state; given the political ability and willingness to extract it. In both cases, we place the total amount of military dollars spent as a ratio of either surplus domestic product milex ratio it = milexit SDPit , or gross domestic product milex ratio it = milexit GDPit . With these two alternative versions of the dependent variable, we then specify a regression model to analyze the correlation between this quantity and several important covariates. We specify the following primary estimating equation: milex ratio it = 1 surplus it + 2 subsistence it + X it +a i +u t + it ; where X it is a matrix of additional covariates. We specify two-way xed eects; using a i , the country xed eect, and u t , the time-period xed eect. 146 Surplus Income GDP = Surplus + Subsistence v 0 2 4 6 8 10 0 2 4 6 8 10 Subsistence Income GDP = Surplus + Subsistence v 0 2 4 6 8 10 0 2 4 6 8 10 Figure 4.9: The dollar values displayed on the x-axes and y-axes in the panels above are in billions of $US. Suppose a country with a population of 2,739,726 people. Such a country needs to generate 3 billion $US dollars (365 days $3 per-day 2,739,726 people) per year to healthfully sustain each member of the population over the long term, which isv, the minimum surplus value. Such a country is consuming all of its income for subsistence up until it generates income surpassing this minimum surplus value v. Once such a country generates income greater than v, the country is generating positive surplus income which it can invest in items other than \bread" (e.g., \butter" or \guns"). Poor and under-developed countries do exist today and in earlier periods of history with income levels at and below this threshold. Indeed, some state governments have worked diligently to develop extractive institutions to take even the subsistence income of the population. However, these states do not maintain the levels of healthy adults necessary for other state-making tasks (e.g., conscription) to sustain such a strategy over the long run. 147 Figure 4.10: The top row of panels shows the yearly correlation between GDP and surplus income (SDP). The middle row of the panels shows the yearly correlation between GDP and subsistence income. The bottom row of panels shows the yearly proportion of countries that generate enough income to pass above the subsistence threshold at $1, $2, or $3 per person per day. The columns indicate these subsistence thresholds for each set of panels. 148 4.C Rankorder Graphs The gures below display the top ten potentially threatening states within the strategic envi- ronment of Japan and the United Kingdom | analogous to the graph representing the strategic environment of the United States presented in the main manuscript. The highest panel illustrates the rank order of the top ten potentially threatening states when using the distance-weighted rela- tive power ratio that incorporates SDP; the middle panel plots an analogous ranking for the same measure using GDP; the lowest panel shows weighted relative power ratios based on a distance- weighted relative population measure. Opponent states with a large SDP that are geographically proximate to each of these states should have higher weighted relative power ratios than states with either low levels of SDP, or that are geographically distant, or both. The Loss of Strength Gradient (LSG) is conceptualized as the inverse of the logged distance between capital cities. We use concurrent validity to make these assessments. Concurrent validity is an assessment of the ability of an empirical measure to distinguish between cases that are distinct based on some prior theoretical knowledge about the status of those cases (Trochim and Donelly, 2008, 60). To have concurrent validity, the potential threat measure should be able to accurately categorize the opponent states that are the most threatening to any individual state in any historic period. The denition for concurrent validity is analogous to face validity, except that face validity assesses the link between the theory and the operational protocol, while concurrent validity assesses the link between the operational protocol and data. It should also be able to categorize states that face highly threatening strategic environments and those that do not. 149 Figure 4.11: Top 10 potentially threatening states for Japan by decade for $3 per diem subsistence level relative SDP on the upper, standard relative GDP in the middle, and relative population on the lower panel. Dyads that are not jointly democratic are potentially threatening and denoted through opaque shading. Dyads that are jointly democratic are not potentially threatening and denoted through brighter shading. 150 Figure 4.12: Top 10 potentially threatening states for the United Kingdom by decade for $3 per diem subsistence level relative SDP on the upper, standard relative GDP in the middle, and relative population on the lower panel. Dyads that are not jointly democratic are potentially threatening and denoted through opaque shading. Dyads that are jointly democratic are not potentially threatening and denoted through brighter shading. 151 4.D Coverage of Composite Index of National Capabilities (CINC) variables In the main manuscript, we assess convergent validity by comparing a country's share of global SDP to several component variables from CINC. Convergent validity is dened as\the degree to which the operationalization is similar to (converges on) other operationalizations that it theoret- ically should be similar to." (Trochim and Donelly, 2008) CINC's restrictive approach toward including countries as members of the international system leads to distortions in the estimates of power. For example, based on the Correlates of War (COW) classication, China does not become a member of the international system of states until 1860, while Gleditsch and Ward code it as a system member since 1816. As a result, CINC population totals are likely undercounting global population and therefore in ating other countries' relative population gures prior to 1860. Figure 4.13 illustrates the eect that China has on the total CINC score. Plotted in Figure 4.13 is the annual correlation between the original CINC score and a re-computed CINC score that drops China from the global sums of the component variables. Before China enters the COW system of states (and CINC), the two correlate perfectly, because China is included in neither of the series between 1816 and 1859. When China enters the National Material Capabilities (NMC) data (Greig and Enterline, 2017) in 1960, the correlation drops to approximately 0.9976, mostly because China has such a large total population relative to other countries (see below). Our new measurement approach below makes population estimates available for a larger set of countries in the pre-industrial period and correct part of the bias resulting from the exclusion of units in CINC. For example, while CINC codes China as having 47% of the global population in 1860, our data code the population share to be 31%. The exclusion of China in the CINC scores from 1816 to 1859 also aects the population shares of other countries. The United States drops from having 8.3% of global population in 1859 to 3.9% in 1960 in CINC; our estimates are 2.6% 152 Figure 4.13: Annual correlation with 95% condence intervals between the original CINC score and a re-computation of CINC that drops China from the global sums of iron and steel produc- tion, primary energy consumption, total population, urban population, military expenditure, and military personnel. The annual observation for China is dropped from both series. and 2.6%, respectively. We this use our revised series of population data to compute a country's share of global population (Figure 4.3 in the main manuscript). These population estimates are available for a larger set of country-year units. They contain data on China and Japan (that are missing in CINC) prior to 1860. Figure 4.14 below re-plots the upper panel of Figure 4.3 from the main manuscript. In panel (a), we exclude China observations from the global sums of iron and steel production, primary energy consumption, SDP, and GDP, respectively. In panel (b), we exclude China observations from the global sums of iron and steel production and primary energy consumption, but keep China observations in the computation of global SDP and GDP. The graphs are virtually identical. The plots demonstrate that the drastic drop in the correlation between the share of global GDP and the CINC component variables is caused by the exclusion of China from CINC prior to 1860, not by an in issue our GDP or SDP estimates. 153 Note that in Figure 4.3 in the main manuscript, the drastic drop in the annual correlation between CINC component variables and a country's share of global GDP is not replicated in the correlation with a country's share of global GDP because China does not contribute much to global surplus until the post-WWII period. Based on our estimates, China starts to consistently have GDP income that exceeds subsistence needs in 1964. 154 (a) Excluding China from all series (b) Excluding China from CINC, but not SDP or GDP series Figure 4.14: The plots display yearly correlation coecients with 95% condence intervals. In each of the panels, we assess the degree to which SDP (orange) and GDP (grey) correlate with the iron and steel production and primary energy consumption variables of CINC. 155 4.E Comparing SDP with GDP and GDP per capita Figure 4.15: The graph plots the natural logarithm of GDP against the natural logarithm of SDP for select years. As time progresses and countries develop, SDP and GDP correlate highly. An exception are least developed countries, mostly in Sub-Saharan counties, who do not have a positive surplus in 2010. The SDP measure is based on a $3 per day subsistence threshold and is truncated to 1 for countries with no surplus resources in order to allow for a transformation via the natural logarithm. 156 Figure 4.16: The graph plots GDP per capita against SDP across all country-years in the sam- ple. The linear patterns of dots show individual countries' trajectories over time. Labeled are observations for select countries in 1990. 157 4.F Measuring potential threat in the strategic environment In the main manuscript, we dened three component variables of potential threat that capture information about dyads, which we review here. The variables exists for each unit, i = 1;:::;N across each time period t = 1;:::;T . For each country-year variable, we make use of information about each of the dyadic relationships between state i and the other j states in the international system each year, which j = 1;:::;J indexes the other states in relationship with state i. We consider three types of relationships between statei and statej all of which are bounded between 0 and 1: 1. Relative power ratio in terms of the dierence in power-resources between state i and state j in year t (i.e., is the opponent state j is relatively larger or smaller than state i). 2. LSG over geographic distance between state i and state j in year t. 3. Preference compatibility between state i and state j in year t. Relative power ratio for state i with opponent state j is dened based on the ratio of the power-resources as measured by the SDP of the opponent state j as a proportion of the sum of the SDP values for both state i and the opponent state j. SDP is measured as a function of each state's GDP it , Population it , and the subsistence level threshold which we set to either $3, $2, or $1 dollars per day as dened in equation 4.B above. 18 The relative power ratio between two states is measured using the estimate of surplus domestic product SDP it for state i and the SDP jt for the opponent state j as r ijt = SDP jt (SDP jt +SDP it ) 18 We are currently working on collecting additional data that will help us model this threshold parameter as a latent variable. 158 This quantity falls on the unit interval [0; 1] such that r ijt = 8 > > > > > > > > < > > > > > > > > : (0:5; 1] if SDP it <SDP jt 0:5 if SDP it =SDP jt [0; 0:5) if SDP it >SDP jt These relative power ratios capture the intuition that powerful states will nd less powerful countries less threatening because they are the weaker state in the ij pairing. The most powerful state in the system will fear all other countries less than those countries fear it. The most powerful state's relative power ratio will be close to 0. The least powerful state's relative power ratios will be close to 1. If two states have equal power, they will each nd the other equally threatening and the relative power ratio for two equal states is 0.5. We weight these relative power ratios using two additional relational features between pairs of states: the preference relationship between states (preference compatibility) and the relative position of a state within the geographic arena (loss of strength gradient). Preference Compatibility: Only certain powerful states are potentially threatening to oth- ers and observable indicators of shared preferences can help to identify these relationships. When preferences between pairs of states are compatible, the probability of con ict between the two is reduced and, as such, should minimize the importance of power-resource dierences between the two states. Though we consider many alternative indicators of preference compatibility in a related project, 19 in this paper, we focus on insights from the democratic peace literature to assess degree to which two state have compatible preferences. We assume that all states are potentially threatening to one another, unless they are both democracies. States with democratic institutions 19 Further below, Figures 4.34 and 4.35 in this supplementary appendix illustrate that our results are largely robust to indicators of preference compatibility that do not rely on joint democracy, such as rivalry, alliances, bilateral trade relationships, or United Nations General Assembly voting. 159 have more compatible preferences and are therefore not as threatening to one another. We make no claims regarding whether it is democratic institutions themselves or some other variable that co-varies with democracy that causes states to have more compatible preferences. Thus, we make a descriptive, rather than causal claim, when arguing that democratic states, and only democratic states, do not nd each other democracies threatening. Thus, we assume that democracies nd autocracies threatening, and autocracies nd all states threatening regardless of their regime type. We use utilize this assumption regarding which states will nd each other threatening, to dene a preference compatibility measure that we use to down-weight each power-resource ratio r ijt . Preference compatibility is dened as p ijt , which is a measure of the shared preferences of state i and state j in year t. This quantity falls on the unit interval [0; 1]. For some of the preference indicators we consider, this variable takes only integer valuesf0; 1g. Specically, the value is 0 if statei and statej both have compatible interests in yeart based on the Polity2 or Boix et al. (2013) democracy variables. p ijt is otherwise coded as 1 when this is not the case. A coding of 1 captures incompatible relationships, which could potentially be threatening depending on the value ofr ijt . Using one of two binary democracy variables, we dene the preference compatibility between two states as p ijt = 8 > > > < > > > : 0 if i and j jointly democratic 1 otherwise For the continuous measure of preference compatibility, we use the Unied Democracy Scale (UDS) for each state to dene preferences as p ijt = (UDS it ) (UDS jt ): Thus, if a pair of states does not have compatible preferences, then the relative power-resource measure is not changed. If a pair of states has compatible preferences, then the power-resource ratio is reduced to 0 | eectively making states both non-threatening to one another. 160 Loss of Strength Gradient over geographic distance: The costs associated with con ict and arming are increasing in the distance over which power must projected to state i by an opponent state j. We therefore assume that the LSG, which increases over distance, reduces the level of threat between two states. Contiguous or geographically proximate states should be more in uential or potentially threatening than states that are far away because the LSG results in power-resources dissipating over distance (Markowitz and Fariss, 2013; Gleditsch and Ward, 2001; Boulding, 1962). LSG over geographic distance is dened as d ijt , which is the distance between the capital city of country i and the capital city of neighbor j in yeart. d ijt is dened for each country-year pair in each year using the longitude and latitude coordinates for each state's capital city d ijt =acos(sin(lat it )sin(lat jt ) +cos(lat it )cos(lat jt )cos(lon it lon jt ))radius Where d ijt is the distance between state i's capital city and state j's capital city. lat i , lat j , lon i , lon j , are the latitude and longitude locations for state i and state j. These values vary little over time but we calculate d ijt for each year t. We transform the distance values into a proportion w ijt , so that it falls on the unit interval [0; 1]. This captures the intuition that states that are geographically proximate (short distance between i and j) should have more in uential relationships than states that are geographically distant from one another. The LSG increases the costs associated with projecting power. In many existing empirical applications, the transformation of distance to the unit interval is accomplished using either inverse distance or the inverse natural logarithm of this quantity. For the inverse natural logarithm, this is dened as: w ijt = 1 ln(d ijt ) : 161 In words, w ijt is the the inverse of the natural log of distance d ijt in km between state i and state j in year t. The measure captures the intuition that neighbors, which are geographically proximate (close neighbors), are more in uential on the behavior of countryi than neighbors that are far away. Figure 4.17 provides visual examples of the distribution of this component measure. Potential threat is dened as the total of each of these weighted relative power ratios for country i in year t, based on state i's relationship with all other j states in the international system in each year. It is formally dened as Potential threat it = X j2J [r ijt w ijt p ijt ]: Table 4.3 provides a summarization of each component part of Potential threat it for the eco- nomic resource-based version. Table 4.4 provides an analogous specication for the population- based potential threat measure. Figure 4.18 provides a step by step illustration of the construction of this measure. Concept Measurement Relative power ratio r ijt = SDPjt SDPjt+SDPit Loss of strength gradient w ijt = 1 ln(dij ) Preference compatibility (binary) p ijt = ( 0 if i and j jointly democratic 1 otherwise Preference compatibility (continuous) p ijt = (UDS it ) (UDS jt ) Total potential threat (economic) Potential threat it = P j2J [r ijt w ijt p ijt ] Table 4.3: Concepts and operational denitions of each of the component parts of the country-year potential threat measure based on economic resources. We brie y describe the measurement process that generates the total relative power variable for a hypothetical three-state system. Suppose that in the year 1900 there are only three countries in the world: the United Kingdom, Germany, and the United States. The table below shows 162 Concept Measurement Relative power ratio r ijt = Population jt Population jt +Population it Loss of strength gradient w ijt = 1 ln(dij ) Preference compatibility (binary) p ijt = ( 0 if i and j jointly democratic 1 otherwise Preference compatibility (continuous) p ijt = (UDS it ) (UDS jt ) Total potential threat (population) Potential threat it = P j2J [r ijt w ijt p ijt ] Table 4.4: Concepts and operational denitions of each of the component parts of the country-year potential threat measure based on economic resources. the computation of the level of potential threat that the United Kingdom faces if its strategic environment consists of only Germany and the United States. In 1900, the United Kingdom is coded as a democracy based on the categorical value of its democracy score based on Polity2. Its SDP was approximately 265 billion in constant 2011 international Purchasing Power Parity (PPP) dollars. In this year, the United Kingdom does not have compatible preferences with then- autocratic Germany, but is jointly democratic with the United States. 20 Based on the binary specication of regime type, only the relative power-resources of Germany, weighted by distance, contribute to the total level of potential threat, which is the sum of the relative power ratios that the United Kingdom faces in this three-state international environment. 21 Table 4.5 below illustrates the computation of the United Kingdom's potential threat in this three-state example. the United Kingdom's potential threat score in this hypothetical three-state international system is 0.07. 20 Please note that this is only true for the binary joint democracy measures using the Polity and Boix et al. data. When using the continuous joint democracy measure based on the UDS scale, the distance-weighted power of the United States, relative to the United Kingdom, would contribute to the total potential threat faced by the United Kingdom. However, the United States' contribution would be very small because the preference-weight will be close to zero 21 The maximum distance between capital cities in our data is approximately 19949km. The distance London{ Berlin is approximately 916km; London{Washington D.C. approximately 5954km. 163 Relative Preference Loss of Strength Weighted relative power-resources Compatibility Gradient power-resources Germany 234 234+265 = 0:47 1 1 ln(916) = 0:15 0:47 1 0:15 = 0:07 United States 457 457+265 = 0:63 0 1 ln(5954) = 0:12 0:63 0 0:11 = 0 Table 4.5: Hypothetical example of a three-state system. The example demonstrates how each component part of the potential threat variable is combined into the nal value for this country- year variable. 164 4.F.1 Geographic proximity and the loss of strength gradient Figure 4.17: Comparing the binning of countries by a) distance, b) a linear transformation of distance computed as max(Distance)Distance max(Distance) , and c) the inverse of the logged distance from the perspective of the United States, France, and China. 165 4.F.2 Construction of the potential threat measure Figure 4.18: The graph illustrates the construction of the potential threat measure for the USin 1935. The SDP is based on a $3 per diem subsistence value. 166 4.F.3 Correlation between alternative potential threat measures Figure 4.19: Correlation plot for alternative potential threat measures, using SDP ($3 subsistence level) to measure economic resources. Colored cells denote values that are signicant at the minimum 5% level of signicance. 167 4.F.4 Global trends of potential threat Figure 4.20: The graphs show the evolution of potential threat over time for alternative indicators of preference compatibility (see the caption of Figure 4.19 for data sources). Plotted is each country-year observation with the line denoting the loess smoothed trend over time across all countries. 168 Figure 4.21: Evolution of potential threat over time for alternative indicators of preference com- patibility. Plotted is each country-year observation with the lines denoting the loess smoothed trend over time for three groups of states: a) states that enter the international system before 1900, b) states that enter between 1900 and 1945, and c) states that enter after 1945 based on Gleditsch and Ward (1999). 169 Figure 4.22: Evolution of potential threat over time for each indicator of preference compatibility (see the caption of Figure 4.19 for data sources). Plotted is each country-year observation with the lines denoting the line of best t over time for the rst, second, third, and fourth quartile of states based on the distribution of the per capita SDP (using a $3 per diem subsistence level). 170 4.F.5 Spatiotemporal variation of potential threat Figure 4.23: Maps plotting the spatiotemporal distribution of the natural log of the potential threat variable for the years 1965 and 2000. Potential threat is measured through joint democracy based on the Polity, Boix et al., and UDS scores, respectively. Power-resources are measured using the SDP indicator with a $3 per diem subsistence level. Grey shaded areas denote missing values. The maps are based on the borders for 1 January 1965 and 2000, respectively, using data from the cshapes library in R (Weidmann et al., 2010). The operationalization of each indicator is based on substantive choices of each coding team. Therefore, the coverage does not always perfectly map, either spatially or temporally. For example, some geographical spaces such as Greenland or former colonies in Africa are missing. 171 4.F.6 Top 20 states facing the most threatening strategic environment Figure 4.24: Top 20 states facing the most potentially threatening strategic environment in 1816, 1910, 1935, 1965, 1990, and 2010. The red dots show our estimate of the total level of potential threat each country faces when using the Polity2 score to measure preference compatibility; green triangles the Boix et al. estimates, and blue squares the UDS potential threat scores. Countries are ranked based on the potential threat variable that measures preference compatibility via the Polity2 score. Error bars indicate the 95% condence intervals for the average of all alterna- tive potential threat measures.All potential threat variables are standardized; hence, the x-axis measures are expressed in standard deviations. 172 4.F.7 Economic-based potential threat versus population-based potential threat Figure 4.25: Relationship between country-year values of potential threat based on population on the y-axis and potential threat based on economic resources for various subsistence thresholds on the x-axis. 173 Figure 4.26: Annual correlation between three alternative Potential Threat (PT) measures over time. We vary how power resources are measured across the three indicators: using GDP, using SDP ($3 subsistence threshold), and using population. For all indicators, preference compatibility is measured via joint democracy (Polity) and power resources are weighted by the inverse of the logged distance between capital cities. 174 4.G Dependent variables: Military investments 4.G.1 Evolution of the military investments over time Figure 4.27: The two upper plots illustrate the temporal evolution of the military expenditure index (military expenditure/SDP in 2011 constant US PPP dollars) and naval tonnage index (naval tonnage/SDP in constant 2011 international PPP dollars), respectively. 175 4.G.2 Correlations Figure 4.28: The plot illustrates the correlations between the two alternative dependent variables over time. Points denote the correlation coecients for each year between 1965 and 2007. Lines represent the Loess smooth over those points. 176 4.G.3 Comparing individual countries over time Figure 4.29: The plot demonstrates the ability of our measurement strategy to obtain scores for the level of potential threat that individual countries face at any given point in time (granted data availability). Plotted in the rst row are economic resource-based potential threat scores using alternative regime type indicators to measure preference compatibility for the United States, the United Kingdom, France, Japan, China, and Brazil in the 20th century | the line representing a smoothed trend across all variables. In the rows below, we graph the time trends for the two dependent variables military expenditure as a proportion of SDP and naval tonnage as a proportion of SDP. All variables are shown on a logarithmic scale with base 10. 177 Figure 4.30: The graph shows the evolution of the potential threat across world regions for alternative measures of preference compatibility. 178 4.H Regression models 4.H.1 Dropping the population-based potential threat variable Figure 4.31: Coecients and 95% condence intervals of standardized potential threat variables for regression models of two dependent variables | the military expenditure index and naval tonnage index | on potential threat and control variables. All models include controls for the natural log of income (SDP or GDP) and a country's Polity2 score. Standard errors are clustered by country; right-hand side variables are lagged by one year. 179 4.H.2 Bivariate regressions Figure 4.32: Regression models upon dropping all control variables. 4.H.3 Post-WWII sample Figure 4.33: Regression models for a post-WWII sample. For model specication details, see the caption of Figure 4.31. 180 4.H.4 Alternative interest compatibility measures Figure 4.34: Regression models for alternative preference compatibility measures. For model specication details, see the caption of Figure 4.31. 181 Figure 4.35: Regression models for alternative preference compatibility measures. For model specication details, see the caption of Figure 4.31. 182 4.H.5 Summary statistics Table 4.6: Summary statistics for key variables. Statistic Min Median Mean Max St. Dev. N ln Military expenditure/GDP 20.72 3.98 4.55 0.00 3.69 13,097 ln Naval tonnage/GDP 20.72 20.72 18.74 11.74 2.80 13,321 ln Military Expenditure/SDP 20.72 3.35 3.68 0.00 3.75 13,097 ln Naval tonnage/SDP 20.72 20.72 17.36 12.02 6.70 13,321 Potential threat (SDP) joint democracy (Polity) 0.06 3.61 4.92 18.58 4.16 16,001 Potential threat (SDP) joint democracy (Boix et al.) 0.04 3.63 5.05 20.82 4.44 15,284 Potential threat (SDP) joint democracy (UDS) 0.002 1.74 2.29 9.90 2.08 9,694 Potential threat (SDP) rivalry 0.00 0.00 0.06 1.79 0.12 12,427 Potential threat (SDP) defense alliances 0.07 4.56 6.17 21.16 4.78 17,355 Potential threat (SDP) s-scores 0.04 1.47 1.59 12.88 0.98 9,424 Potential threat (SDP) absolute ideal dierence 0.03 1.70 1.72 10.10 0.94 9,537 Potential threat (SDP) joint IGO membership 0.11 7.25 7.71 19.64 4.96 7,933 Potential threat (SDP) diplomatic exchange 0.00 2.34 4.29 19.24 4.81 2,660 Potential threat (SDP) bilateral trade 0.12 5.80 7.24 21.27 5.20 11,519 Potential threat (SDP) per capita energy consumption 0.01 1.54 2.11 9.31 1.91 14,402 Potential threat (SDP) ATOP Alliances (continuous) 0.04 3.14 4.32 13.82 3.41 15,081 Potential threat (SDP) ATOP Alliances (binary) 0.05 4.70 6.39 21.16 5.03 15,081 Potential threat (SDP) no interest variable 0.15 6.22 7.62 21.66 5.19 26,067 Potential threat (GDP) joint democracy (Polity) 0.07 3.93 5.00 18.33 3.94 16,001 Potential threat (GDP) joint democracy (Boix et al.) 0.08 3.98 5.14 20.10 4.18 15,284 Potential threat (GDP) joint democracy (UDS) 0.003 1.85 2.30 10.24 1.98 9,694 Potential threat (GDP) rivalry 0.00 0.00 0.06 1.70 0.11 12,427 Potential threat (GDP) defense alliances 0.10 4.90 6.29 20.86 4.61 17,355 Potential threat (GDP) s-scores 0.05 1.46 1.59 13.69 0.99 9,424 Potential threat (GDP) absolute ideal dierence 0.04 1.70 1.72 10.76 0.94 9,537 Potential threat (GDP) joint IGO membership 0.16 7.53 7.73 19.67 4.82 7,933 Potential threat (GDP) diplomatic exchange 0.00 2.43 4.31 20.22 4.77 2,660 Potential threat (GDP) bilateral trade 0.17 5.93 7.27 20.81 5.06 11,519 Potential threat (GDP) per capita energy consumption 0.01 1.63 2.13 9.76 1.83 14,402 Potential threat (GDP) ATOP Alliances (continuous) 0.05 3.26 4.34 13.94 3.30 15,081 Potential threat (GDP) ATOP Alliances (binary) 0.06 4.89 6.42 20.86 4.87 15,081 Potential threat (GDP) no interest variable 0.15 7.28 7.98 21.21 5.01 26,067 Potential threat (population) joint democracy (Polity) 0.05 4.14 5.00 17.29 3.52 16,001 Potential threat (population) joint democracy (Boix et al.) 0.05 4.12 5.14 19.64 3.81 15,284 Potential threat (population) joint democracy (UDS) 0.01 1.85 2.30 10.08 1.91 9,694 Potential threat (population) rivalry 0.00 0.00 0.06 1.11 0.10 12,427 Potential threat (population) defense alliances 0.08 5.02 6.29 22.18 4.49 17,355 Potential threat (population) s-scores 0.05 1.39 1.59 13.88 1.08 9,424 Potential threat (population) absolute ideal dierence 0.05 1.59 1.72 10.92 1.01 9,537 Potential threat (population) joint IGO membership 0.11 7.29 7.73 21.17 4.68 7,933 Potential threat (population) diplomatic exchange 0.00 2.58 4.31 21.43 4.62 2,660 Potential threat (population) bilateral trade 0.12 6.31 7.30 21.93 4.94 11,519 Potential threat (population) per capita energy consumption 0.05 1.69 2.13 9.39 1.68 14,402 Potential threat (population) ATOP Alliances (continuous) 0.05 3.59 4.34 13.63 3.16 15,081 Potential threat (population) ATOP Alliances (binary) 0.07 5.21 6.42 22.18 4.71 15,081 Potential threat (population) no interest variable 0.13 6.80 7.98 22.39 4.93 26,067 ln GDP 15.36 22.77 22.55 30.74 2.78 27,321 ln SDP 0.00 21.70 17.78 30.68 9.52 27,321 ln Subsistence 13.80 21.86 21.44 28.00 2.50 27,321 ln Population 6.80 14.99 14.76 21.00 2.21 27,321 Polity2 score 10.00 3.00 0.55 10.00 7.07 16,974 Notes: GDP measure in constant 2011 international PPP dollars. SDP is based on a $3 per diem subsistence level. Loss of strength gradient measured using the following formula 1 log(distance) . Very small values are rounded to 0 in the output above. 183 4.I GDP, Population, and GDPpc Component Datasets Total power-resources are measured using GDP data in constant 2011 international dollars from the World Development Indicators (World Bank, 2016), and supplemented with a number of historic GDP data estimates that are combined using a measurement model to estimate a GDP series that covers the entire period of observation 1816{2012. The latent variable model that is used to compute the GDP and population data for the analysis is estimated based on data for GDP 22 , GDP per capita 23 , and population. 24 Details on the sources, measurement choices, and coverage of the component variables are provided in Table 4.7. For each component dataset, we extract relevant indicators, attach unique country identiers, and reshape the data into a common country-year format. We consulted the codebooks of each dataset to drop observations that are interpolated or extrapolated by the authors of the dataset, or already covered by other datasets 25 . Details on the underlying source materials for each component measure and coding decisions are provided below and are documented in the R code we use to merge the constituent datasets together. When merging the dierent variables together we relied on the available country-year units as prepared by the authors of the original datasets. We use the Gleditsch and Ward (1999) revised list of independent states as the base set of units. For years prior to the start year of this data set (1816 A.D.) we again use the date the year the unit enters the dataset or 1500 A.D. As we discussed in each dataset description, dierent datasets sometimes use dierent spatial denitions for units. We have matched country-year units across datasets using the best match available. In some cases, units exist in the dataset that are not historically accurate such as a unied Germany prior to 1871. Maddison includes this unit in his historic data series, aggregating information 22 For observed data on GDP see World Bank (2016); Feenstra et al. (2015); Broadberry and Klein (2012); Maddison (2010); Gleditsch (2002); Bairoch (1976). 23 For observed data on GDP per capita see World Bank (2016); Broadberry (2015); The Maddison-Project (2013); Broadberry and Klein (2012); Gleditsch (2002); Bairoch (1976). 24 For observed data on population see World Bank (2016); Feenstra et al. (2015); Broadberry and Klein (2012); Maddison (2010); Gleditsch (2002); Singer et al. (1972). 25 For example, the data generated by Gleditsch (2002) includes some interpolated values and values taken from the Maddison Project 184 across the various principalities and other administrative districts that existed until Germany had completely unied in 1871. As another example, Maddison also disaggregates information about North and South Korea backwards in time. Additional details about these unit specic issues are available in the original source material. Documentation about how we merged all of the data sources together are available in our code les, which are publicly accessible. Importantly, because many of these units are subsets of larger ones (for example North and South Korea), analysts can aggregate the estimates of these two units together if necessary for a specic empirical application. 185 Table 4.7: Component Measures for GDP, GDP per capita, and Population Latent Variable Model Variable Descriptions Coverage in Original Coverage in Model Source Material and Citations GDP data are measured in 1990 international dollars. 1AD{2008 1500{2008 Historical GDP data collected by Angus Maddison (Maddison, 2010). GDP data are measured as total real GDP at 2005 prices (PPP). 1950-2011 1950-2011 Expanded GDP data version 6.0 beta, September 2014 (Gleditsch, 2002). GDP data are measured in constant 2010 USD. 1960{2015 1960{2015 World Development Indicators (World Bank, 2016) GDP data are measured in constant 2011 international dollars (PPP). 1990{2015 1960{2015 World Development Indicators (World Bank, 2016) GDP data limited to European countries and the United States, after accounting for changing country boundaries. GDP is measured in millions of 1990 international dollars (national currencies are converted to international dollars using Angus Maddison's purchasing power parities) 1870{2001 1870{2001 Broadberry and Klein (2012). Gross National Product (GNP) data limited to European countries, after accounting for changing country boundaries. GNP is measured at market prices and expressed in constant 1960 USdollars. 1830{1973 1830{1973 Bairoch (1976). GDP (expenditure oriented) in millions of constant 2011 international dollars (PPP). 1950{2014 1950{2014 Penn World Tables (PWT) version 9.0 (Feenstra et al., 2015). GDP (output oriented) in millions of constant 2011 international dollars (PPP). 1950{2014 1950{2014 PWT version 9.0 (Feenstra et al., 2015). GDP per capita data are measured in 1990 international dollars. 1AD-2010 1500{2010 Extension of Angus Maddison's historical GDP and population estimates (The Maddison-Project, 2013). GDP per capita data are measured as total real GDP at 2005 prices (PPP). 1950-2011 1950-2011 Expanded GDP data version 6.0 beta, September 2014 (Gleditsch, 2002). GDP per capita are measured in constant 2010 USD. 1960{2015 1960{2015 World Development Indicators (World Bank, 2016) GDP per capita are measured in constant 2011 international dollars (PPP). 1990{2015 1960{2015 World Development Indicators (World Bank, 2016) GDP per capita data limited to European countries and the United States, after accounting for changing country boundaries. GDP is measured in millions of 1990 international dollars. 1870{2001 1870{2001 Broadberry and Klein (2012). GNP per capita data are limited to European countries, after accounting for changing country boundaries. GNP is measured at market prices and expressed in constant 1960 USdollars. 1830{1973 1830{1973 Bairoch (1976). GDP per capita data limited England/Great Britain, Holland/Netherlands, Italy, Spain, Japan, China, and India. GDP is measured in millions of 1990 international dollars. 725{1850 1500{1850 Broadberry (2015). Total population measured in thousands at mid-year. 1AD{2030 1500{2010 Historical population data collected by Angus Maddison (Maddison, 2010). Total population measured in thousands. 1950-2011 1950-2011 Expanded GDP data version 6.0 beta, September 2014 (Gleditsch, 2002). Population data limited to European countries and the United States. 1870{2001 1870{2001 Broadberry and Klein (2012). 186 Total population. 1960{2015 1960{2015 World Development Indicators (World Bank, 2016) Total population measured in thousands. 1816{2001 1816{2001 The Correlates of War Project's National Material Capabilities data version 4.0 (Singer et al., 1972) Population (in millions). 1950{2014 1950{2014 PWT version 9.0 (Feenstra et al., 2015). 187 The Maddison Project (Maddison, 2010; The Maddison-Project, 2013): Maddison's original GDP, GDP per capita, and population variables are derived from a large number of country-level sources (Maddison, 2003, 2001, 1995). Because the underlying source materials employed by Maddison are expansive and country-specic, we refrain from describing them in detail. The more recent version of these data, The Maddison-Project (2013), is based on a collaboration of researchers dedicated to continuing Angus Maddison's data collection eorts by extending and, if warranted, revising his estimates. Due to the collaborative nature of the eort, dierent research teams use dierent methods and source material to obtain their estimates. With a few exceptions, data from 1990{2010 were revised using gures from the Total Economy Database of the Conference Board (Bolt and van Zanden, 2014). Other estimates are based on historical national statistics from country-specic sources (Bolt and van Zanden, 2014). We subset the data from the Maddison Project to include only country-year observations starting in 1500. The original Maddison (2010) data includes both GDP and population values. The updated version only included GDP per capita estimates. We include both data versions in our model since, as we describe below, it is capable of linking all of these observed indicators together in united model that leverages the information from each type of variable. Unlike some of the other datasets we describe below, these datasets do not contain origin codes that classify the source material used to inform the country-year values. Expanded GDP data version 6.0 beta (Gleditsch, 2002): Gleditsch (2002)'s (beta) ver- sion 6.0 of the Expanded GDP data is based primarily on the PWT 8.0, and supplemented with data from the PWT 5.6, the Maddison Project Database, and the World Bank Global develop- ment indicators. In addition, Gleditsch (2002) constructed his data using imputations for the lead and tail values, as well as interpolation for estimates within the series. We use only the values that stem from the PWT gures in the latent variable model (origin codes 0, -1, and 3) and exclude data from the Maddison Project, as well as interpolated or imputed gures (origin codes -2, 1, and 2). In the Validity section below, we consider the model t for the latent variables estimates that do include these variables compared to the latent variable model estimates that exclude them and demonstrate the model t is improved by estimating these missing values using our model-based approach instead of using interpolation or extrapolation. 188 World Development Indicators (World Bank, 2016): We include GDP, GDP per capita, and population from the World Bank (2016). Where possible, we use the metadata for each indicator provided by the World Bank's DataBank to determine the underlying source material of the GDP, GDP per capita, and population values. As with the Gleditsch (2002) data, we drop values that are interpolated or extrapolated and allow our model to generate new estimates for these units. We describe each of these variables in turn. We include the World Bank (2016) GDP indicator measured in constant 2010 USdollars in our latent variable model. The gures are compiled from the World Bank and Organisation for Economic Co-operation and Development (OECD) national accounts data. The documentation in the metadata le indicates that the series is based on an underlying interpolation of component data upon aggregating it to a \gap-lled total." Unfortunately, we do not have information on the details of this aggregation process. We therefore use the full series of GDP as provided by the World Bank (2016)'s online data portal DataBank. In future versions of our model, we plan to identify these cases when possible and adjust our model accordingly. The per capita GDP series is based on the World Bank (2016)'s GDP in constant 2010 USdol- lars and the total population gures (for the underlying source material see below). According to the metadata, the data is aggregated using weighted averages. We exclude observations from our model that the metadata indicates as being preliminary, extrapolated, or interpolated. Informa- tion on which country-years were excluded based on the metadata is provided in the replication material that accompanies this paper. The population gures from World Bank (2016) are based on national population censuses. The census data that informs this measure stem from a variety of sources, including the United Nations World Population Prospects (for the majority of developing countries), Eurostat (for European countries), and national statistical agencies. The data are interpolated for all years between census years. Since we do not have information on the years that a census was conducted for each country, we retain the interpolated data for the use in the latent variable model. We do, however, exclude population gures that are explicitly indicated as being extrapolated, interpo- lated, or preliminary in the metadata. Information on which country-years-units were excluded is provided in the replication material that accompanies this paper. In future versions of our model, we plan to identify the other interpolated cases when possible and again adjust our model accordingly. 189 Broadberry and Klein (2012): The GDP, GDP per capita, and population variables in Broadberry and Klein (2012) are limited to European nations, including Russia and Turkey, as well as the United States. A detailed list of underlying source material is available in the paper's appendix (Broadberry and Klein, 2012, pp. 105). For GDP, these sources include the data from Maddison (2010), ocial national account statistics, and the work of country-expert historians. Data on population are drawn mainly from Mitchell (2003) and Maddison (2010), and supplemented with country-specic data from ocial national statistics and historians. We exclude those country-year observations that are taken from Maddison (2010) in our model. Bairoch (1976): The underlying source material for the data by Bairoch is detailed in the paper's methodological appendix. For GNP, these sources include the work of historians and ocial national statistics for earlier country-years, as well as OECD gures for years starting in 1950 (Bairoch, 1976, 329 et seq.). For the 19th century and the year 1900, three-year annual averages are available for every decade starting from 1830 and expressed in 1960 U.S. dollars (Bairoch, 1976, 286). For the 20th century, data are available for select years between 1913 and 1973 and expressed in 1960 U.S. dollars as well (Bairoch, 1976, 297). For population, Bairoch relies on United Nations Demographic yearbooks, data from the League of Nations, and national statistical agencies to assemble his data (321). We incorporate all of Bairoch's estimates in our model, including the ones agged as having a larger-than-average margin of error (the gures presented in parentheses). The data from Bairoch (1976) cover the total and per capita GNP, not GDP. Bairoch's denition is based on the United Nations' 1953 System of National accounts (United Nations, 1953). With the exception of the data from Bairoch (1976), the data on economic size are measured as GDP. Bairoch (1976) uses GNP instead. While the GNP excludes value added by foreign rms, this measure is highly correlated with GDP. The correlation between GNP and GDP is quite high, with correlation coecients between 0.865 and 0.995 for country-year units within the period 1830{1973. The strength of the positive relationship varies over time but rarely falls below 0.9. We anticipate that in future years, the correlation between the two measures should drop as globalization increases and the internationalization of production and investment increases the relevance of the conceptual dierence between GNP and GDP. Additional estimates of GNP and GDP from more recent years would help researchers determine how this empirical 190 relationship evolves over time. The evaluation of this distinction is one possible avenue that our new latent variable model opens up for exploration, which we discuss below. Broadberry (2015): The GDP per capita estimates in Broadberry (2015) are based on histor- ical national accounting data that is constructed from documents such as \government accounts, customs accounts, poll tax returns, parish registers, city records, trading company records, hospi- tal and educational establishment records, manorial accounts, probate inventories, farm accounts, tithe les and other records of religious institutions." (Broadberry, 2015, 5). Broadberry lists the data sources for each country in the main text. 26 As with the Maddison data, we exclude cases for years prior to 1500 from our model. COW National Military Capabilities data v4.0 (Singer et al., 1972): The COW Project provides a variety of country-level estimates including population beginning in the year 1816. For country-years starting in 1919, the population estimates by Singer et al. (1972) are based primarily on the estimates of the United Nations Statistical Oce. The population estimates for years prior to 1919 are based on national government censuses. For these earlier years in the series, the authors of the population dataset selected country-specic data that presents the greatest continuity with the data from the United Nations. 27 The authors of the data use a variety of methods to bridge gaps in the data, including interpolation, regression, and extrapolation. Quality codes for the estimates of the total population gure are specied | indicating whether a data point stems from an identied source, is missing, derived through interpolation, regression, or extrapolation. We retain only those data points that stem from an identied source (quality code A). 4.J Latent Variable Model Specication To specify the dynamic latent variable model, leti = 1;:::;N, index cross-sectional units andt = 1;:::T , index time periods. For each country-year unit,j = 1;:::;J indexes the observed variables y itj . Because the observed variables that enter the model represent three dierent concepts | GDP, population, and GDP per capita | we estimate three latent variable parameters, where 26 Pages 6 and 7 contain the underlying source material for Britain, the Netherlands, Italy, and Spain; page 8 contains the data for China, Japan, and India. 27 For details, please refer to the codebook for version 4.0 of the data: Correlates of War Project National Material Capabilities Data Documentation Version 4.0, http://cow.dss.ucdavis.edu/data-sets/ national-material-capabilities/nmc-codebook/at_download/file, accessed 1 December 2016. 191 k = 1; 2; 3, indexes the three categories gdp; pop; gdppc. This allows us to dene the set of y itj that we observe for each of the k dimensions of the latent variable model, where y itj 1fy 2 k g. This notation allows us to denote the set of observed variables used to estimate each of the three underlying latent variables such that gdp =fy it1 ;y it2 ;y it3 ;y it4 ;y it5 g, pop = fy it6 ;y it7 ;y it8 ;y it9 ;y it10 g, gdppc =fy it11 ;y it12 ;y it13 ;y it14 ;y it15 ;y it16 g. 28 With knowledge of how the observed variables relate to each categoryk, we can denote how the three dimensions of the latent variable relate to them as well. The model assumes that the latent variables take the form: itk N (0; 1) for all i when t = 1 (the rst year a country enters the dataset). When t> 1, the standard normal prior is centered around the latent variable estimate from the previous year such that: itk N ( it1;k ; k ). The latent variables themselves are estimated with uncertainty. The rst year each country enters the model, the variances for these parameters are set to 1. For all years after t = 1, gdp and pop are drawn from a uniform distribution U(0; 1). For the latent GDP per capita variable, the latent estimates and associated uncertainty are deterministically determined by the GDP and Population latent variables themselves such that it;gdppc it;gdp it;pop . This modeling innovation allows information form the three types of observed variables to inform more than just one of the latent variables. The latent variables are estimated by linking each of these parameters to the sets of observed GDP, population, or GDP per capita variables. Since all of the GDP, population, and GDP per capita variables are continuous, we specify a Gaussian link function with a unique error term for each of the the three types of variables k :f gdp , pop , gdppc g. These k parameters are estimates of model level uncertainty, which link each of the latent variables to the sets of observed GDP, population, or GDP per capita variables. Shape parameters translate the observed variables from their original unit-of-measurement into the latent variable unit-of-measurement. Because we specify a Gaussian link function, these shape parameters are the intercept and slope from the linear model. For the intercept parameters j , we center the standard normal prior around the the mean value of the observed data with a relatively large variance (low precision): j N ( y j ; 4). We choose the mean value of the observed variables because the mean of latent traits 28 A useful feature of this notation is that the sets of observed variables do not need to be mutually exclusive. Though we do not allow the observed variables to inform the estimation of multiple latent variables in the appli- cation presented here, this is a possibility in other applications. See Gelman and Hill (2007); Imai et al. (2016) for more details. We thank James Lo for this notational suggestion. 192 themselves are centered around 0. 29 The intercept parameter therefore transforms the latent trait into the unit-of-measurement of the original observed variable. For identication of the model we set j = 1 because we assume a one-unit change in the latent trait is equivalent to a one- unit change in the original observed variable. 30 All of the prior distributions are summarized in Table 4.8. Recall that we organize the three types of observed variables in three sets such that y itj 1fy2 k g. Therefore, the likelihood function that links the observed data to the estimated parameters is: L(;;;jy itj 1fy2 k g) = N Y i=1 T Y t=1 J Y j=1 K Y k=1 N ( j + itk j ; k ) Table 4.8: Prior Distribution for Latent Variables and Model Level Parameter Estimates Parameter Prior Country i latent GDP estimate in rst year t it=1;gdp N (0; 1) Country i latent GDP estimate in all other years it;gdp N ( t1;gdp ; gdp ) Latent GDP uncertainty gdp U(0; 1) Country i latent population estimate in rst year t it=1;pop N (0; 1) Country i latent population estimate in all other years it;pop N (t1;pop;pop) Latent population uncertainty pop U(0; 1) Country i latent GDP per capita estimate it;gdppc it;gdp it;pop Model j intercept \diculty parameter" j N ( yitj; 4) Model j slope \discrimination parameter" j 1 Model uncertainty for all GDP items gdp G(0:001; 0:001) Model uncertainty for all population items pop G(0:001; 0:001) Model uncertainty for all GDP per capita items gdppc G(0:001; 0:001) The model is estimated with ve Markov Chain Monte Carlo (MCMC) chains, run for 100,000 iterations each. The rst 50,000 iterations were thrown away as burn-in and the rest were used to generate the posterior prediction intervals for the original observed variables. 31 29 We set this parameter to the empirical mean of the Maddison GDP and population variables as an identication constraint. 30 This assumption can be relaxed to examine the relative strength of the relationship between one measure compared to another. We leave this analysis to future research. Relaxing this assumption would allow for analysts to explore the relative relationship between measures of GDP and GNP as functions of the underlying latent trait. We view this as a useful extension to the model we present here. 31 The Gibbs sampler was implemented in Martyn Plummer's Just Another Gibbs Sampler (JAGS) software (Plummer, 2010). Conventional diagnostics all suggest convergence. 193 Chapter 5 Re ections and future research The three essays that comprise this dissertation illuminate facets of the interplay between states' military and administrative capacity. The analyses emphasize that the coercive and non-coercive aspects of state capacity cannot be studied in isolation. In interstate con icts, a state's ability to develop military capacity in part depends on the degree to which it can extract surplus resources that can be invested in the military, or other goods and services. In intrastate con icts, and particularly in asymmetric civil wars, a state's administrative capacity determines not just its material military capability, but crucially, its ability to defeat insurgents that seek cover among the civilian population. The analyses demonstrate that measurement innovation can help overcome some of the short- comings in the availability of data regarding state capacity in con ict. The work contributes new or improved indicators to measure theoretical concepts that are central to the study of the political geography of con ict: territorial control in asymmetric civil war, subnational con ict exposure, states' economic resources available for investment in coercive and non-coercive capabilities, and potential threat in the international system. Many of the measures advanced in this dissertation leverage spatiotemporal variation of con- ict or state characteristics to create indicators that allow for comparisons across space and over time. In Chapter 2, I utilize the spatiotemporal variation of guerrilla versus terrorist rebel tactics 194 to compute estimates of territorial control. Also in Chapter 2, 1 I develop an improved measure of con ict exposure that uses spatial and temporal weights to aggregate con ict events to subnational areal units, as opposed to discrete assignment commonly used in the existing literature. Contin- uously assigning events to grid cells allows for a more realistic assessment of the spatiotemporal distribution of events that pinpoint weaknesses in states' capacity to deter violent challengers. Similarly, Chapter 4 utilizes spatial weights to account for the attenuating eect large geographic distances have on the perception of threat potential between pairs of states. In this concluding chapter, I re ect on the contributions this work provides for the study of the political geography of con ict processes and outline avenues for future research. 5.1 Territorial control and subnational con ict processes The study in Chapter 2 provides a new measurement framework that facilitates the estimation of territorial control via publicly available event data. The methodology is applicable across countries, but allows for small-scale spatiotemporal variation at the subnational level. The use of existing geo-coded event data in a machine learning framework reduces the amount of resources that are necessary to compute estimates of territorial control. The project is embedded in a broader research agenda that investigates the drivers and consequences of intrastate territorial contestation in a quest to better understand subnational con ict dynamics. A number of ongoing and planned projects will extend this work. First, I seek to increase the accuracy of the existing estimates of territorial control and con- duct in-depth sensitivity analyses. To this end, in future work I plan to dive deeper into the Colombian case to further corroborate estimates of territorial control. In Colombia, as in many other con icts, variation in territorial control is not observed directly and needs to be estimated 1 See the appendix to Chapter 2. 195 retroactively. The Colombian peace process provides a unique opportunity for out-of-sample val- idation of my estimates using a number of proxy measures that correlate highly with territorial control. 2 In addition to the rate of deforestation studied in Chapter 2, these observable prox- ies include the level of coca production and the intensity of common crime|which were heavily regulated and/or controlled by the Fuerzas Armadas Revolucionarias de Colombia{Ej ercito del Pueblo (FARC) guerrillas within their territorial strongholds. Regarding crime in the post-peace period, a recent report from Reuters suggests that \[n]ationwide, murders have declined in recent years. In Tumaco, and other former FARC bastions, homicides are soaring."(Murphy and Acosta, 2018) The signing of the peace agreement between the Colombian government and the FARC in 2016 introduced a sudden change in territorial control, because it forced rebels to disarm and abandon their strongholds. The timing of the eventual signing of the peace agreement in 2016 was plausibly unexpected, given the long history of failed peace negotiations. The unanticipated timing minimizes potential endogeneity bias when relating the dierence between pre- and post- peace levels in the proxy variables to pre-peace levels of FARC territorial control. Similar to the analysis conducted in Chapter 2, I expect post-peace rates of coca production and common crime to be higher in former rebel strongholds. Information on rebel tactics obtained via global coverage data suer from systematic under- reporting of con ict events, particularly in less populated regions of Colombia. Future research should therefore consider the inclusion of country-specic data sources on political violence in an eort to decrease concerns about a potential under-reporting of con ict events in global coverage databases such as the Georeferenced Event Dataset (GED) and the Global Terrorism Database (GTD). For Colombia in particular, the impending publication of geo-coded event data by the Violent Presence of Armed Actors in Colombia project promises further improvements of the es- timates. 3 The availability of new software tools to cross-reference and remove duplicates between 2 Out-of-sample validation refers to the practice of corroborating estimates of territorial control using data that is not part of the estimation process. Territorial control estimates are produced using data on the variation in rebel tactics (terrorism versus guerrilla tactics), obtained from con ict event data. 3 See https://www.colombiaarmedactors.org. 196 multiple sources of geo-coded event data aids this integration of global-coverage and country- specic data sources (Donnay et al., 2019). Second, I seek to increase the spatiotemporal coverage of the territorial control estimates beyond the cases of Nigeria and Colombia in Chapter 2. The scarcity of existing ne-grained data on territorial control motivated the development of the new measurement model developed in this dissertation. However, the dearth of existing information also presented a signicant challenge for validating model estimates. Nigeria and Colombia were chosen as prototype cases because they allow for the validation of the estimates via the construction of comparison data from Armed Con ict Location and Event Data (ACLED) and variation in deforestation rates pre- and post-peace accord, respectively. Extending the spatiotemporal coverage of the estimates involves further validation 4 and the extension of the theoretical model beyond two actors. This extension of the model will make the methodology applicable to civil wars involving more than two main actors, such as Syria or Iraq, that are of central interest to con ict scholars and development practitioners. The validation exercises in Chapter 2 show that despite similarity in general patterns, deviation exists between the Hidden Markov Model (HMM) estimates and existing measures of territorial control. These discrepancies warrant further work on the model and validation data presented here. Avenues for an improvement of the accuracy of the HMM estimates include the modeling of spatial autocorrelation within the framework of Hidden Markov Random Field (HMRF) models, 5 the inclusion of context variables such as terrain or forest cover, and an exploration of a possible variation in the transition probabilities across con ict contexts, for example secessionist versus non-secessionist civil wars or protracted versus highly active con icts. In the current project, the values of model parameters such as transition and emission matrices, as well as the margin parameter to compute overlap between the relative frequencies of terrorism and conventional war 4 The future publication of territorial control estimates by the Resources and Con ict project provides a promis- ing avenue for further in-depth validation of estimates beyond the cases of Colombia and Nigeria (Tao et al., 2016). 5 See Anders et al. (2017) for a discussion of this approach and its limitations. 197 acts are xed and informed by theoretical and empirical considerations. A critical next step is to quantitatively assess the sensitivity of the model to changes in these parameters and/or to estimate them in a quest to improve the accuracy of the data and extend coverage to con icts beyond Nigeria and Colombia. Third, I plan to utilize the new estimates of territorial control to study micro-dynamics of subnational con ict. Territorial control is one of the central variables for understanding patterns of violence and governance in civil wars. Establishing territorial control is a pre-condition for actors to move from purely coercive interactions with the population into a space of non-coercive governance. Recent years have seen increased scholarly interest regarding the determinants of public good provision in con ict zones as an important component of non-coercive governance. 6 However, territorial control is astonishingly absent as a variable in most empirical studies of gov- ernance in con ict zones. This is mainly due to the|to date|lack of ne-grained data that captures subnational variation of territorial control across space and time. The estimates devel- oped in Chapter 2 and further improvements in the accuracy and coverage of the data open a number of promising avenues for future research on the relationship between territorial control and subnational governance regimes in particular, which I elaborate in more detail in section 5.2 below. 5.2 Subnational governance in con ict The analysis in Chapter 3 shows that governments use welfare investments to supplement their military ghting capacity when faced with an internal armed challenger to their statehood. These welfare investments are strategic, because they are a reaction to violence that threatens the survival of the state|either directly through atrocities against armed state agents or indirectly because civilians, whose support is vital to the military success of the government, are harmed as 6 See for example Arjona 2016; Findley et al. 2016; Hollenbach et al. 2016; De Juan and Bank 2015. 198 a corollary of ghting between armed forces and insurgents. The analysis of data from Colombian municipalities in Chapter 3 suggests that violence experienced by the civilian population that is not a direct consequence of the ght between non-state armed actors and the government, such as homicides, or violence against unarmed state agents, does not elicit a government response in terms of higher investments in citizen welfare. Due to constraints in the availability of the underlying municipal-level data, the analysis in Chapter 3 is limited to the years 1994 to 2010. It does not cover the period after the signing of the peace agreement between the FARC rebels and the Colombian government in late 2016. The theoretical argument developed in Chapter 3 can, however, be extended to generate predic- tions about the likely relationship between violence and the welfare-mindedness of the Colombian government in the aftermath of the peace accord. Based on the theory developed in Chapter 3, with the major armed competitor to statehood removed as a result of the peace agreement, I do not expect the Colombian government to utilize welfare spending as a tool to combat insecurity and violence in former FARC strongholds. While the in uence of rebel groups like the Ej ercito de Liberaci on Nacional (ELN), FARC splinter factions, and criminal groups has strengthened in former FARC strongholds (Murphy and Acosta, 2018), due to their smaller size and sphere of in uence, these groups currently do not represent as powerful of a challenger to the government as the FARC did. General insecurity experienced by the civilian population is not expected to elicit a government response in terms of higher welfare spending, based on the theory developed in Chapter 3. Recent journalistic and scholarly accounts suggest that despite pledges by the Colombian gov- ernment to invest in infrastructure and service provision in former FARC strongholds, the progress lacks behind the promises. The 2016 nal peace accord signed in Havana details investment promises for rural development in the areas of road, irrigation, and electricity infrastructure, as well as social development in the areas of health, education, housing, drinking water, and poverty 199 eradication. 7 A recent survey by the Monitoring Attitudes, Perceptions and Support (MAPS) project found that while there appear to be some improvements in the areas of policing, road construction, and education, particularly the provision of health services remains inadequate. 8 These accounts of the insucient provision of public goods and services, despite the governments' pledges to invest in rural development in former rebel strongholds, are in line with the theoretical expectations developed in Chapter 3. The impending publication of municipal welfare spending data for the period after the signing of the peace agreement, and further improvements in the availability of con ict data, 9 will allow future research to systematically study whether the Colom- bian government abandoned the strategic use of welfare spending as a violence management tool in the post-peace period. In addition to extending the analysis to the post-peace period, I plan to combine municipal welfare spending data from four sources in a Bayesian measurement model to obtain improved estimates of Colombian subnational investments in the areas of health care, education, and san- itation. I seek to supplement the Panel Municipal of the Centro de Estudios sobre Desarrollo Econ omico (CEDE) institute at the Universidad de los Andes 2000{2010 used in Chapter 3 with data from Departamento Administrativo Nacional de Estad stica (DANE) 1973-1999, the Con- tralor a General 1984{2014, and Departamento nacional de planeaci on (DNP) 1993{2016. 10 The procedure extends the methodology to create new estimates of indicators such as Gross Domes- tic Product (GDP), population, and GDP per capita that underlies the data creation process in Chapter 4 to the subnational realm. 11 7 See Acuerdo nal para la terminaci on del con icto y la construccti on de una paz estable y duradera (https: //tinyurl.com/yah9u8e7), accessed 24 August 2019. 8 The Washington Post Monkey Cage: \Colombia's historic peace agreement with the FARC is fraying. We talked to 1,700 Colombians to understand why." (https://www.washingtonpost.com/politics/2019/08/06/ colombias-historic-peace-agreement-with-farc-is-fraying-we-talked-colombians-understand-why/, last accessed 24 August 2019.) See also New York Times : \Colombia's Peace Deal Promised a New Era. So Why Are These Rebels Rearming? (https://www.nytimes.com/2019/05/17/world/americas/colombia-farc-peace-deal. html, last accessed 24 August 2019). 9 See https://www.colombiaarmedactors.org. Future research should also explore the use of alternative data on violent incidents collected by the Centro the Memoria Hist orica (http://www.centrodememoriahistorica.gov. co/micrositios/informeGeneral/basesDatos.html). 10 Special thanks goes to Fabio Sanchez for sharing these data. 11 See also Fariss et al. (2017). 200 Furthermore, I plan to explore the link between the validated estimates of territorial control and subnational governance in civil war. The measure of territorial control developed in Chapter 2 is an important new source of data for one of the most scarcely measured, but most theoretically important, variables in the study of civil war. In particular the question regarding the strategic importance of welfare spending (Chapter 3) will benet from a future integration of the territorial control measure. A recent study by Sexton (2016) highlights the importance of considering the status of territorial control when examining the relationship between violence and good provision. Attesting to the strategic component of local public good provision, Sexton (2016) shows that in Afghanistan, aid delivery in districts already controlled by pro-government forces reduced violence, while aid targeted to contested districts signicantly increased insurgent attacks in an eort by the non-state actors to curb increased opponent territorial control. Aided by the additional work to validate the territorial control measure for the Colombian case above, in future research, I seek to supplement the analysis in Chapter 3 with an indicator of the underlying status of control in Colombian municipalities. Governments may be hesitant to invest resources in areas they might not control in the fore- seeable future, due to the risk of their investment falling into the hands of the rebels. On the other hand, welfare spending can be used as a tool to increase popular support and prevent territorial gains by the adversary. Con ict actors are risk-averse and likely to utilize an area's history of control as a heuristic to determine whether or not to invest in citizen welfare. In areas that were previously controlled by insurgents, the government faces a high level of uncertainty regarding the allegiance of the civilian population. Areas that have long been under the control of the rebels should therefore not see signicant investments in citizen welfare upon a re-capture by the govern- ment. 12 In contrast, areas that were only brie y under insurgent control, or in which territorial control is contested, are expected to enjoy higher welfare investments. Areas that are uncontested and rmly under government control, when matched on important dimensions, serve as a baseline 12 The risk of insurgent re-capture sets the determinants of public good provision during con ict apart from governance dynamics in post-con ict contexts. 201 for comparison. To show that variation is driven by strategic considerations, as opposed to purely humanitarian need, I plan to contrast subnational patterns of government welfare spending with aid provision from external actors. 5.3 Threat and arming in the international system The study in Chapter 4 is part of a broader research agenda that seeks to explain patterns of threat and arming in the international system. To this end, Chapter 4 introduces Surplus Domestic Product (SDP) as a better measure of states' economic power in the international system than the conventionally used indicator of GDP. My co-authors and I show that SDP produces more concurrently valid orderings of the most powerful states in the international system in the 19th, 20th, and early 21st centuries. When used as a component of potential threat and military burden indicators, SDP fares better than GDP in terms of its predictive validity. Specically, we show that when we use SDP to measure relative dyadic power and a state's resources available for investment, higher levels of potential threat are associated with higher levels of military investment|which is not the case when GDP is used. High levels of preference compatibility attenuate threats stemming from states that are geo- graphically proximate and powerful in terms of their relative dyadic economic might. The con- ceptualization of the potential threat measure in the main analysis in Chapter 4 is limited to the measurement of preference compatibility between states via joint democracy. States are deemed to have compatible interests, and are thus not threatening to each other, if they are both democratic. However, democracy is only one facet of what renders interests between states compatible. In a accompanying project, we explore the determinants of preference compatibility between states and the degree to which preference compatibility drives the transformation of con ictual into coopera- tive interactions between states. We study the relative predictive power of alternative measures of interest compatibility as constituent parts of our potential threat measure for patterns of arming, 202 power projection, and con ict initiation. These alternative measures include, but are not limited to, some of the indicators mentioned in Section 4.3.2, 13 that is alliances and defense pacts, shared Intergovernmental Organization (IGO) membership, United Nations General Assembly (UNGA) voting similarity, energy consumption, bilateral trade, and diplomatic exchange. We further extend and improve the accuracy of measurement of the GDP, population, and GDP per capita indicators used to create the SDP indicator in Chapter 4, in a connected project. We include a wider range of historical and contemporary sources of GDP, population, and GDP per capita and alter the Bayesian measurement model to account for dierences in the concep- tualization of alternative GDP indicators. GDP indicators dier regarding the methodology that is used to make the gures comparable across space and/or over time. We distinguish between series that make GDP data comparable between countries via exchange rates versus Purchasing Power Parity (PPP). In addition, we adjust our measurement model to incorporate parameters for the distinction between GDP indicators that are comparable between countries at a single point in time, those that are suitable for the computation of growth rates within countries but should not be used for cross-national comparison, and measures that allow for the contrasting of GDP levels across countries and over time via the chaining of multiple rounds of international price comparisons. 14 These improvements further reduce uncertainty in the current estimates and render predictions regarding the relationship between potential threat and arming or power projection more precise. Finally, future research should engage in more depth with the measurement of a populations' subsistence needs. In the main manuscript in Chapter 4, we stipulate a level of $3 per day per person in 2011 US PPP dollars as the minimum resources needed ensure the survival of the pop- ulation. This minimum subsistence threshold is informed by relative poverty and malnutrition 13 See also Figures 4.19, 4.34 and 4.35 in Appendix 4.7. 14 See Bolt et al. (2018) and Feenstra et al. (2015). 203 rates among people who have $3 or less per day to survive. 15 The choice of the $3 PPP sub- sistence threshold is based on the assumption that the minimum caloric needs of the population are relatively stable across space and time. Future work should seek to estimate this minimum subsistence threshold. In particular the use of ratios of urban to rural population as an indica- tor for production beyond agricultural subsistence presents a fruitful avenue toward the goal of estimating, rather than assuming, the minimum resources needed to sustain the population. As Angus Maddison states in his 2004 Ruggles lecture: When countries are able to expand their urban ratios, it indicates that there was a growing surplus beyond subsistence in agriculture, and that the non-agricultural component of economic activity was increasing. (Maddison, 2004, 29) This dissertation contributes new methods and theoretical insights to enhance our understand- ing of the interplay between states' administrative and military capacity. Crucially, they form the foundation for substantive and methodological future research endeavors at the interstate and intrastate levels. The innovations in the measurement of territorial control in civil war, economic power, and potential threat in states' geopolitical environment provide new sources of indicators for scholars studying a wide range of topics pertaining to the political geography of con ict pro- cesses. These indicators are of particular value for a better understanding of the nexus between security and economic development, as developing countries tend to be disproportionately strained by intrastate political violence as well as higher levels of potential threat and military burdens due to, on average, lower surplus resources. 15 See section 4.2 for more details. 204 Bibliography Albertus, M. and O. Kaplan (2013). Land reform as a counterinsurgency policy: Evidence from Colombia. Journal of Con ict Resolution 57 (2), 198{231. Anders, T., H. Xu, C. Cheng, and T. S. Kumar (2017). Measuring territorial control in civil wars using Hidden Markov Models: A data informatics-based approach. Proceedings of the NIPS 2017 Workshop on Machine Learning for the Developing World. arXiv:1711.06786. Andres, L. A., D. Sislen, and P. Marin (2010). Charting a new course: Structural reforms in Colombia's water supply and sanitation sector. Technical report, The International Bank for Reconstruction and Development / The World Bank, Washington, D.C. Angrist, J. D. and A. D. Kugler (2008). Rural windfall or a new resource curse? Coca, income, and civil con ict in Colombia. The Review of Economics and Statistics 90 (2), 191{215. Arjona, A. (2016). Rebelocracy. Social Order in the Colombian Civil War. New York: Cambridge University Press. Aronson, J., D. Ciland, P. K. Huth, and J. I. Walsh (2017, September). An enhanced dataset of territorial control by con ict actors. Paper presented at the 2017 Meeting of the American Political Science Association. Asal, V., L. De La Calle, M. Findley, and J. Young (2012). Killing civilians or holding territory? How to think about terrorism. International Studies Review 14 (3), 475{497. Bailey, M. A., A. Strezhnev, and E. Voeten (2017). Estimating dynamic state preferences from United Nations voting data. Journal of Con ict Resolution 61 (2), 430{456. Bairoch, P. (1976). Europe's gross national product, 1800-1975. Journal of European Economic History 5 (2), 273{340. Bakker, R., J. Daniel W Hill, and W. H. Moore (2016). How much terror? Dissidents, governments, institutions, and the cross-national study of terror attacks. Journal of Peace Research 53 (5), 711{726. Balcells, L. (2017). Rivalry and Revenge: The Politics of Violence during Civil War. Cambridge and New York: Cambridge University Press. Barbieri, K. and O. Keshk (2012). Correlates of War project trade data set codebook. Online: http://correlatesofwar.org. Barbieri, K., O. M. Keshk, and B. M. Pollins (2009). Trading data: Evaluating our assumptions and coding rules. Con ict Management and Peace Science 26 (5), 471{491. Barnett, M. and R. Duvall (2005). Power in international politics. International Organiza- tion 59(1), 39{75. 205 Barron, P., K. Kaiser, and M. Pradhan (2009). Understanding variations in local con ict: Evi- dence and implications from Indonesia. World Development 37 (3), 698{713. Baum, D. A. L. M. A. (2001). The invisible hand of democracy: Political control and the provision of public services. Comparative Political Studies 34 (6), 587{621. Bayer, R. (2006). Diplomatic exchange data set, v2006.1. Online: http://correlatesofwar.org, retreived 10 October 2016. Bazzi, S. and M. Gudgeon (2015). Local government proliferation, diversity, and con ict. Working Paper 205, Households in Con ict Network (HiCN), Brighton, UK. Beath, A., F. Christia, and R. Enikolopov (2012). Winning hearts and minds through develop- ment? Evidence from a eld experiment in Afghanistan. Policy Research Working Paper 6129, World Bank, Washington DC. Beckley, M. (2018). The power of nations: Measuring what matters. International Security 43 (2), 7{44. Berman, E., J. Felter, J. N. Shapiro, and E. Troland (2013, February). Modest, secure and informed: Successful development in con ict zones. NBER Working Paper No. 18674. www. nber.org/papers/w18674.pdf, accessed 30 October 2013. Berman, E. and A. M. Matanock (2015). The empiricists' insurgency. Annual Review of Political Science 18(1), 443{464. Berman, E., J. N. Shapiro, and J. H. Felter (2011). Can hearts and minds be bought? The economics of counterinsurgency in Iraq. Journal of Political Economy 119 (4), 766{819. Bhavnani, R., D. Miodownik, and H. J. Choi (2011). Three two tango: Territorial control and selective violence in Israel, the West Bank, and Gaza. The Journal of Con ict Resolution 55 (1), 133{158. Boix, C., M. K. Miller, and S. Rosato (2013). A complete data set of political regimes, 1800{2007. Comparative Political Studies 46 (12), 1523{1554. Bolt, J., R. Inklar, H. de Jong, and J. L. van Zanden (2018). Rebasing `Maddison': New income comparisons and the shape of long-run economic development. GGDC Research Memorandum 174, Groningen Growth and Development Centre, Groningen. Bolt, J. and J. L. van Zanden (2014). The Maddison Project: Collaborative research on historical national accounts. The Economic History Review 67 (3), 627{651. B orzel, T. A. (2012). How much statehood does it take - and what for? SFB-Governance Working Paper Series No. 29. Berlin: DFG Sonderforschungsbereich 700. Boulding, K. E. (1962). Con ict and defense. New York: Harper and Brothers. Bozzoli, C., T. Br uck, and N. Wald (2013). Self-employment and con ict in Colombia. Journal of Con ict Resolution 57 (1), 117{142. Brecher, M. and J. Wilkenfeld (2000). A Study of Crisis. Ann Arbor: University of Michigan Press. Brecher, M., J. Wilkenfeld, K. Beardsley, P. James, and D. Quinn (2017). International crisis behavior data codebook, version 12. http://sites.duke.edu/icbdata/data-collections/. 206 Broadberry, S. (2015, September). Accounting for the great divergence. Online. https:// www.nuffield.ox.ac.uk/users/Broadberry/AccountingGreatDivergence6.pdf, accessed 23 November 2016. Broadberry, S. and A. Klein (2012). Aggregate and per capita GDP in Europe, 1870{2000: Continental, regional and national data with changing boundaries. Scandinavian Economic History Review 60 (1), 79{107. Buhaug, H. (2010). Dude, where's my con ict? LSG, relative strength, and the location of civil war. Con ict Management and Peace Science 27 (2), 107{128. Burgoon, B. (2006). On welfare and terror: Social welfare policies and political-economic roots of terrorism. Journal of Con ict Resolution 50 (2), 176{203. Cardenas, M., M. Eslava, and S. Ramirez (2016). Why internal con ict deteriorates state capacity? Evidence from Colombian municipalities. Defence and Peace Economics 27 (3), 353{377. Carter, D. B. (2015). When terrorism is evidence of state success: Securing the state against territorial groups. Oxford Economic Papers 67 (1), 116. Cederman, L.-E., K. S. Gleditsch, and H. Buhaug (2013). Inequality, Grievances, and Civil War. New York: Cambridge University Press. Chaparro, J. C., M. Smart, and J. G. Zapata (2004). Intergovernmental transfers and municipal nance in Colombia. ITP Paper 0403, University of Toronto, Joseph L. Rotman School of Management, Toronto, Ontario. Chenoweth, E. (2013). Terrorism and democracy. Annual Review of Political Science 16 (1), 355{378. Collier, P. and A. Hoeer (2002). Military expenditure. Threats, aid, and arms races. Policy Research Working Paper 2927, The World Bank Development Research Group, Washington DC. Collier, P. and A. Hoeer (2004). Greed and grievance in civil war. Oxford Economic Pa- pers 56(4), 563{595. Cort es, D. and D. Montolio (2014). Provision of public goods and violent con ict: Evidence from Colombia. Peace Economics, Peace Science and Public Policy 20 (1), 143{167. Coyle, D. (2014). GDP: A Brief But Aectionate History. Princeton, NJ: Princeton University Press. Crisher, B. B. and M. Souva (2014). Power at sea: A naval power dataset, 1865{2011. International Interactions 40 (4), 602{629. Croicu, M. and R. Sundberg (2017). UCDP GED codebook version 17.1. Uppsala University. Crost, B., J. H. Felter, and P. B. Johnston (2016). Conditional cash transfers, civil con ict and insurgent in uence: Experimental evidence from the Philippines. Journal of Development Economics 118, 171 { 182. Cunningham, D. E., K. S. Gleditsch, and I. Salehyan (2013). Non-state actors in civil wars: A new dataset. Con ict Management and Peace Science 30 (5), 516{531. Dasgupta, A., K. Gawande, and D. Kapur (2015). (When) do anti-poverty programs reduce violence? India's rural employment guarantee and Maoist con ict. Available at SSRN: http: //ssrn.com/abstract=2495803. 207 De Juan, A. and A. Bank (2015). The Ba`athist blackout? Selective goods provision and political violence in the Syrian civil war. Journal of Peace Research 52 (1), 91{104. de la Calle, L. and I. S anchez-Cuenca (2012). Rebels without a territory: An analysis of nonter- ritorial con icts in the world, 1970{1997. Journal of Con ict Resolution 56 (4), 580{603. de la Calle, L. and I. S anchez-Cuenca (2015). How armed groups ght: Territorial control and violent tactics. Studies in Con ict & Terrorism 38 (10), 795{813. Donnay, K., E. T. Dunford, E. C. McGrath, D. Backer, and D. E. Cunningham (2019). Integrating con ict event data. Journal of Con ict Resolution 63 (5), 1337{1364. Drazen, A. and M. Eslava (2010). Electoral manipulation via voter-friendly spending: Theory and evidence. Journal of Development Economics 92 (1), 39 { 52. Dube, O. and S. Naidu (2015). Bases, bullets and ballots: The eect of U.S. military aid on political con ict in Colombia. Journal of Politics 77 (1), 249{267. Dube, O. and J. Vargas (2013). Commodity price shocks and civil con ict: Evidence from Colom- bia. The Review of Economic Studies 80, 1384{1421. Eaton, K. (2006). The downside of decentralization: Armed clientelism in Colombia. Security Studies 15 (4), 533{562. Eck, K. (2012). In data we trust? A comparison of UCDP GED and ACLED con ict events datasets. Cooperation and Con ict 47 (1), 124{141. Ezzati, M. (2004). Comparative quantication of health risks: Global and regional burden of disease attributable to selected major risk factors. World Health Organization. http://apps. who.int/iris/bitstream/10665/42792/1/9241580348_eng_Volume1.pdf. Faguet, J.-P. and F. S anchez (2014). Decentralization and access to social services in Colombia. Public Choice 160 (1-2), 227{249. Fariss, C. J., C. D. Crabtree, T. Anders, Z. M. Jones, F. J. Linder, and J. N. Markowitz (2017). Latent estimation of GDP, GDP per capita, and population from historic and contemporary sources. Available at arXiv https://arxiv.org/abs/1706.01099. Fearon, J. D. (2018). Cooperation, con ict, and the costs of anarchy. International Organiza- tion 72(3), 523{559. Fearon, J. D. and D. D. Laitin (2003). Ethnicity, insurgency, and civil war. American Political Science Review 97(1), 75{90. Feenstra, R. C., R. Inklaar, and M. P. Timmer (2015, October). The next generation of the Penn World Table. American Economic Review 105 (10), 3150{82. Ferreira, F. H. G., S. Chen, A. L. Dabalen, Y. M. Dikhanov, N. Hamadeh, D. M. Jollie, A. Narayan, E. B. Prydz, A. L. Revenga, P. Sangraula, U. Serajuddin, and N. Yoshida (2015). A global count of the extreme poor in 2012: Data issues, methodology and initial results. Policy Research Working Paper 7432, World Bank, Washington D.C. Findley, M. G., R. Sexton, and R. L. Wellhausen (2016). Dissident violence, government budget- ing, and social welfare: Evidence from Peru. Paper Presented at the 2016 Annual Convention of the International Studies Association, Atlanta, GA, March 16-19. 208 Findley, M. G. and J. K. Young (2012). Terrorism and civil war: A spatial and temporal approach to a conceptual problem. Perspectives on Politics 10, 285{305. Fortna, V. P. (2015). Do terrorists win? Rebels' use of terrorism and civil war outcomes. Inter- national Organization 69, 519{556. Fox, S. and K. Hoelscher (2012). Political order, development and social violence. Journal of Peace Research 49 (3), 431{444. Garc a-S anchez, M. (2016). Territorial control and vote choice in Colombia. A multilevel approach. Pol tica y gobierno 23 (1), 53{96. Garnkel, M. R. and C. Syropoulus (2019). Problems of commitment in arming and war: How insecurity and destruction matter. Public Choice 178 (3{4), 349{369. Gelman, A. and J. Hill (2007). Data Analysis Using Regression and Multilevel/Hierarchical Mod- els. Cambridge: Cambridge University Press. Ghahramani, Z. (2001). An introduction to Hidden Markov Models and Bayesian networks. Journal of Pattern Recognition and Articial Intelligence 15 (1), 9{42. Gibler, D. M. (2009). International Military Alliances, 1648{2008. Washington, D.C.: CQ Press. Gleditsch, K. S. (2002). Expanded trade and GDP data. Journal of Con ict Resolution 46 (5), 712{724. Gleditsch, K. S. and M. D. Ward (1999). Interstate system membership: A revised list of the independent states since 1816. International Interactions 25, 393{413. Gleditsch, K. S. and M. D. Ward (2001). Measuring space: A minimum-distance database and applications to international studies. Journal of Peace Research 38 (6), 739{758. Gleditsch, N. P., P. Wallensteen, M. Eriksson, M. Sollenberg, and H. Strand (2002). Armed con ict 1946{2001: A new dataset. Journal of Peace Research 39 (5), 615{637. Greig, J. M. and A. J. Enterline (2017). National material capabilities (NMC) data documentation version 5.0. Technical report, Department of Political Science, University of North Texas. Harbom, L., H. Strand, and H. M. Nyg ard (2009). UCDP/PRIO Armed Con ict Dataset Code- book. http://www.prio.no/Global/upload/CSCW/Data/UCDP/2009/Codebook_UCDP_PRIO% 20Armed%20Conflict%20Dataset%20v4_2009.pdf, accessed October 28, 2013. Hart, J. (1976). Three approaches to the measurement of power in international relations. Inter- national Organization 30 (2), 289{305. Hendrix, C. S. (2010). Measuring state capacity: Theoretical and empirical implications for the study of civil con ict. Journal of Peace Research 47 (3), 273{285. Hendrix, C. S. and J. K. Young (2014). State capacity and terrorism: A two-dimensional approach. Security Studies 23 (2), 329{363. Himmelmann, L. (2010). HMM - Hidden Markov Models. R Package version 1.0, https://CRAN. R-project.org/package=HMM. Hollenbach, F. M., E. Wibbels, and M. D. Ward (2016, May). State building and the geogra- phy of governance: Evidence from satellites. Working paper. http://www.fhollenbach.org/ wp-content/uploads/2016/05/HollenbachWibbelsWard_2016.pdf. 209 Holtermann, H. (2012). Explaining the development{civil war relationship. Con ict Management and Peace Science 29(1), 56{78. Ib a~ nez, A. M. and C. E. V elez (2008). Civil con ict and forced migration: The micro determinants and welfare losses of displacement in Colombia. World Development 36 (4), 659{676. Imai, K., J. Lo, and J. Olmsted (2016). Fast estimation of ideal points with massive data. American Political Science Review 110 (4), 631{656. Iqbal, Z. (2006). Health and human security: The public health impact of violent con ict. Inter- national Studies Quarterly 50 (3), 631{649. Jurafsky, D. and J. H. Martin (2017, August). Hidden markov models. Chapter 9. In Speech and Language Processing. Online https://www.cs.jhu.edu/ ~ jason/papers/jurafsky+martin. slp3draft.ch9.pdf. Justino, P. (2015). Civil unrest and government transfers in India. Evidence report no. 108, Institute of Development Studies. Kadera, K. M. and G. L. Sorokin (2004). Measuring national power. International Interac- tions 30(3), 211{230. Kalyvas, S. N. (2006). The Logic of Violence in Civil War. Cambridge, UK: Cambridge University Press. Kalyvas, S. N. (2015). How civil wars help explain organized crime|and how they do not. Journal of Con ict Resolution 59 (8), 1517{1540. Kalyvas, S. N. and L. Balcells (2010). International system and technologies of rebellion: How the end of the cold war shaped internal con ict. American Political Science Review 104, 415{429. Kalyvas, S. N. and M. A. Kocher (2009). The dynamics of violence in Vietnam: An analysis of the Hamlet evaluation system (hes). Journal of Peace Research 46 (3), 335{355. Kennedy, P. (1987). The Rise and Fall of the Great Powers: Economic Change and Military Power from 1500 to 2000. New York: Random House. Khanna, G. and L. Zimmermann (2017). Guns and butter? Fighting violence with the promise of development. Journal of Development Economics 124, 120 { 141. Khanna, J., T. Sandler, and H. Shimizu (1998). Sharing the nancial burden for U.N. and NATO peacekeeping, 1976{1996. Journal of Con ict Resolution 42 (2), 176{95. Klein, J. P., G. Goertz, and P. F. Diehl (2006). The new rivalry dataset: Procedures and patterns. Journal of Peace Research 43 (3), 331{348. Krasner, S. D. and T. Risse (2014). External actors, state-building, and service provision in areas of limited statehood: Introduction. Governance 27(4), 545{567. Krcmaric, D. (2018). Varieties of civil war and mass killing: Reassessing the relationship between guerrilla warfare and civilian victimization. Journal of Peace Research 55 (1), 18{31. Lake, D. A. (2009). Open economy politics: A critical review. Review of International Organiza- tions 4, 219{244. Lamborn, A. C. (1983). Power and the politics of extraction. International Studies Quar- terly 27(2), 125{146. 210 Lee, M. M., G. Walter-Drop, and J. Wiesel (2014). Taking the state (back) out? statehood and the delivery of collective goods. Governance 27(4), 635{654. Leeds, B., J. Ritter, S. Mitchell, and A. Long (2002). Alliance treaty obligations and provisions, 1815-1944. International Interactions 28 (3), 237{260. Lewis, A. W. (1954). Economic development with unlimited supplies of labour. The Manchester school of economic and social studies 22 (2), 139{191. Lidow, N. H. (2016). Violent Order. Understanding Rebel Governance through Liveria's Civil War. New York: Cambridge University Press. Lind, J. (2011). Democratization and stability in East Asia. International Studies Quarterly 55 (2), 409{436. Lloyd, R. B., M. Haussman, and P. James (2019). Religion and health care in East Africa: Lessons from Uganda, Mozambique and Ethiopia. Bristol: Bristol University Press. Lyall, J. and I. Wilson (2009). Rage against the machines: Explaining outcomes in counterinsur- gency wars. International Organization 63 (1), 67{106. Machado, F. and G. Vesga (2016). Water and sanitation sector: A Colombian overview. Technical Note IDB-TN-713, Inter-American Development Bank, Washington, D.C. Maddison, A. (1995). Monitoring the World Economy, 1820-1992. Development Centre Studies. Paris: OECD. Maddison, A. (2001). The World Economy. A Millenial Perspective. Development Centre Studies. Paris: OECD. Maddison, A. (2003). The World Economy. Historial Statistics. Development Centre Studies. Paris: OECD. Maddison, A. (2004). Contours of the world economy and the art of macro-measurement 1500- 2001. Ruggles Lecture, IARIW 28th General Conference, Cork, Ireland. Maddison, A. (2010). Statistics on world population, GDP and per capita GDP, 1-2008 AD. Online. http://www.ggdc.net/maddison/oriindex.htm, accessed 20 August 2016. Mampilly, Z. C. (2011). Rebel Rulers: Insurgent Governance and Civilian Life During War. Ithaca and London: Cornell University Press. Markowitz, J., C. Fariss, and R. B. McMahon (2019). Producing goods and projecting power: How what you make in uences what you take. Journal of Con ict Resolution 63 (6), 1368{1402. Markowitz, J. N. and C. J. Fariss (2013). Going the distance: The price of projecting power. International Interactions 39 (2), 119{143. Markowitz, J. N. and C. J. Fariss (2018). Power, proximity, and democracy: Geopolitical compe- tition in the international system. Journal of Peace Research 55 (1), 78{93. Marshall, M. G., T. R. Gurr, and K. Jaggers (2016). Polity IV Project. Political regime charater- istics and transitions, 1800-2015. Dataset users' manual, Center for Systemic Peace, Vienna, VA. Matanock, A. M. and M. Garcia-Sanchez (2018). Does counterinsurgent success match social support? Evidence from a survey experiment in Colombia. Journal of Politics 80 (3), 800{814. 211 Milward, A. S. (1977). War, Economy and Society, 1939-1945. Berkeley: University of California Press. Mitchell, B. R. (2003). International Historical Statistics: Europe, 1750 2000 (5 ed.). New York: Palgrave Macmillan. Murphy, H. and L. J. Acosta (2018, April). A fractured peace. Reuters Special Report. On- line https://www.reuters.com/investigates/special-report/colombia-peace/, accessed 12 September 2018. National Consortium for the Study of Terrorism and Responses to Terrorism (START) (2016). Global terrorism database [data le]. Nordhaus, W., J. R. Oneal, and B. Russett (2012). The eects of the international security environment on national military expenditures: A multicountry study. International Organiza- tion 66(3), 491{513. Norrlof, C. and W. C. Wohlforth (2019). Raison de l'h eg emonie (the hegemon's interest): Theory of the costs and benets of hegemony. Security Studies 28 (3), 422{450. Obinger, H. and K. Petersen (2016). Mass warfare and the welfare state { causal mechanisms and eects. Forthcoming in British Journal of Political Science. Openshaw, S. and P. J. Taylor (1979). A millon or so correlation coecients: three experiments on the modiable areal unit problem. In N. Wrigley (Ed.), Statistical applications in the spatial sciences, Chapter 5, pp. 127{144. London: Pion. Oyelere, R. U. and K. Wharton (2013). The impact of con ict on education attainment and en- rollment in Colombia: lessons from recent IDPs. HiCN Working Paper 141. Online http: //www.hicn.org/wordpress/wp-content/uploads/2012/06/HiCN-WP-141.pdf, accessed 26 November 2014. Pemstein, D., S. A. Meserve, and J. Melton (2010). Democratic compromise: A latent variable analysis of ten measures of regime type. Political Analysis 18 (4), 426{449. Perry, G., E. Garc a, and P. Jim enez (2015, May). State capabilities in Colombian municipalities: Measurement and determinants. Documentos cede, Universidad de los Andes. Facultad de Econom a. Centro de Estudios sobre Desarrollo Econ omico, Bogot a, Colombia. Pettersson, T. and P. Wallensteen (2015). Armed con icts, 1946{2014. Journal of Peace Re- search 52(4), 536{550. Pevehouse, J. C., T. Nordstrom, and K. Warnke (2004). The COW-2 International Organizations dataset version 2.0. Con ict Management and Peace Science 21 (2), 101{119. Pierskalla, J. H. (2012). Urban-Rural Bias and the Political Geography of Distributive Con icts. PhD thesis, Duke University, Durham, NC. Plummer, M. (2010). JAGS (Just Another Gibbs Sampler) 1.0.3 universal. Online. http:// www-fis.iarc.fr/ ~ martyn/software/jags/. Poast, P. (2019). Beyond the \Sinew of war": The political economy of security as a subeld. Annual Review of Political Science 22 (1), 223{239. Polo, S. M. and K. S. Gleditsch (2016). Twisting arms and sending messages: Terrorist tactics in civil war. Journal of Peace Research 53 (6), 815{829. 212 Powell, R. (1993). Guns, butter, and anarchy. American Political Science Review 87 (1), 115{132. Quinn, J. M. (2015). Territorial contestation and repressive violence in civil war. Defence and Peace Economics 26 (5), 536{554. Raleigh, C., A. Linke, H. Hegre, and J. Karlsen (2010). Introducing ACLED: An armed con ict location and event dataset: Special data feature. Journal of Peace Research 47 (5), 651{660. Rasler, K. A. and W. R. Thompson (1985). War making and state making: Governmental expenditures, tax revenues, and global wars. American Political Science Review 79 (2), 491{ 507. Reeder, B. W. (2018). The political geography of rebellion: Using event data to identify insurgent territory, preferences, and relocation patterns. International Studies Quarterly 62 (3), 696{707. Reinhart, C. M. and K. S. Rogo (2010). Growth in a time of debt. American Economic Re- view 100(2), 573{578. Reuter, P. (2008). Can production and tracking of illicit drugs be reduced or merely shifted? Policy Research Working Paper 4564, The World Bank Development Research Group, Wash- ington DC. Risse, T. (2011). Governance in areas of limited statehood. introduction and overview. In T. Risse (Ed.), Governance Without a State? Policies and Politics in Areas of Limited Statehood. New York: Columbia University Press. Rodr guez, C. and F. S anchez (2012). Armed con ict exposure, human capital investments, and child labor: Evidence from Colombia. Defence and Peace Economics 23 (2), 161{184. Rozo, S. V. (2015). Is murder bad for business and real income? the eects of violent crime on economic activity. Unpublished manuscript. http://www.sandravrozo.com/uploads/2/9/3/ 0/29306259/rozo_2014_-_jmp.pdf, accessed 03 September 2015. Rubin, M. A. (2020). Rebel territorial control and civilian collective action in civil war: Evidence from the Communist insurgency in the Philippines. Journal of Con ict Resolution 64 (2{3), 459{489. Sambanis, N. (2004). What is civil war? Conceptual and empirical complexities of an operational denition. The Journal of Con ict Resolution 48 (6), 814{858. S anchez, F. and J. N u~ nez (2007). Determinantes del crimen violento en un pa s altamente violento: El case de Colombia. In F. S anchez (Ed.), Las cuentas de la violencia: Ensayos econ omicos sobre el conlicto y el crimen en Colombia, Chapter 2, pp. 25{61. Bogot a: Grupo Editorial Norma. Sandler, T. and K. Hartley (1995). The Economics of Defense. New York: Cambridge University Press. Schelling, T. (1960). Arms and In uence. Hartford: Yale University Press. Schr oder, U. C. (2010). Measuring Security Sector Governance - A Guide to Relevant Indicators. Occasional Paper No. 20. Geneva: Geneva Centre for Democratic Control of Armed Forces (DCAF). Schutte, S. (2017). Geographic determinants of indiscriminate violence in civil wars. Con ict Management and Peace Science 34 (4), 380{405. 213 Schutte, S. and K. Donnay (2014). Matched wake analysis: Finding causal relationships in spa- tiotemporal event data. Political Geography 41, 1 { 10. Sexton, R. (2016). Aid as a tool against insurgency: Evidence from contested and controlled territory in Afghanistan. American Political Science Review 110 (4), 731{749. Sexton, R., R. L. Wellhausen, and M. G. Findley (2019). How government reactions to violence worsen social welfare: Evidence from Peru. American Journal of Political Science 63, 353{367. Singer, D. J. (1987). Reconstructing the Correlates of War dataset on material capabilities of states, 1816{1985. International Interactions 14 (2), 115{132. Singer, D. J., S. Bremer, and J. Stuckey (1972). Capability distribution, uncertainty, and major power war, 1820-1965. In B. Russett (Ed.), Peace, War, and Numbers, pp. 19{48. Beverly Hills: Sage. Singer, J. D. and M. Small (1972). The Wages of War, 1816{1965: A Statistical Handbook. New York: Wiley. Staniland, P. (2012). States, insurgents, and wartime political orders. Perspectives on Poli- tics 10(2), 243{264. Stewart, F. (2008). Horizontal inequalities and con ict: An introduction and some hypothesis. In F. Stewart (Ed.), Horizontal Inequalities and Con ict. Understanding Group Violence in Multiethnic Societies, pp. 3{24. Houndsmills, UK: Palgrave Macmillan. Stewart, M. A. (2018). Civil war as state-making: Strategic governance in civil war. International Organization 72 (1), 205{226. Stewart, M. A. and Y.-M. Liou (2017). Do good borders make good rebels? Territorial control and civilian casualties. The Journal of Politics 79 (1), 284{301. Sundberg, R. and E. Melander (2013). Introducing the UCDP Georeferenced Event Dataset. Journal of Peace Research 50 (4), 523{532. Tajima, Y. (2013). The institutional basis of intercommunal order: Evidence from Indonesia's democratic transition. American Journal of Political Science 57 (1), pp. 104{119. Tal, E. (2017). Measurement in science. In E. N. Zalta (Ed.), The Stanford Encyclopedia of Philosophy (Fall 2017 ed.). Metaphysics Research Lab, Stanford University. Tanzi, V. and L. Schuknecht (2000). Public Spending in the 20th Century. A Global Perspective. Cambridge: Cambridge University Press. Tao, R., D. Strandow, M. Findley, J.-C. Thill, and J. Walsh (2016). A hybrid approach to modeling territorial control in violent armed con icts. Transactions in GIS 20 (3), 413{425. Taydas, Z. and D. Peksen (2012). Can states buy peace? Social welfare spending and civil con icts. Journal of Peace Research 49 (2), 273{287. The Maddison-Project (2013). Online. http://www.ggdc.net/maddison/maddison-project/ home.htm, retrieved August 2016. Thyne, C. L. (2006). ABC's, 123's, and the golden rule: The pacifying eect of education on civil war, 1980{1999. International Studies Quarterly 50 (4), 733{754. Tilly, C. (2004). Terror, terrorism, terrorists. Sociological Theory 22 (1), 5{13. 214 Torres, F. S. and M. Pach on (2013). Decentralization, scal eort and social progress in Colombia at the municipal level, 1994-2009: Why does national politics matter? IDB Working Paper Series IBD-WP-396, Inter-American Development Bank, Washington DC. Trochim, W. M. K. and J. P. Donelly (2008). The Research Methods Knowledge Base (3rd ed.). Mason, OH: Atomic Dog. United Nations (1953). A system of national accounts and supporting tables. Studies in Methods 2, Department of Economic Aairs Statistical Oce, New York. U.S. Army/Marine Corps (2014). Insurgencies and countering insurgencies. Field Manual 3-24, Headquarters, Department of the Amry, Washington, D.C. Valentino, B. A. (2014). Why we kill: The political science of political violence against civilians. Annual Review of Political Science 17 (1), 89{103. Wald, N. and C. Bozzoli (2011). Bullet proof? program evaluation in con ict areas: Evidence from rural Colombia. Proceedings of the German Development Economics Conference, Berlin 2011, No. 80 , http://econstor.eu/bitstream/10419/48329/1/80_wald.pdf, accessed 22 October 2014. Weidmann, N. B. (2015). On the accuracy of media-based con ict event data. Journal of Con ict Resolution 59 (6), 1129{1149. Weidmann, N. B., D. Kuse, and K. S. Gleditsch (2010). The geography of the international system: The cshapes dataset. International Interactions 36 (1), 86{106. World Bank (2016). World development indicators. World Bank (2017). World development indicators. http://data.worldbank.org/ data-catalog/world-development-indicators. World Bank (2018). Classifying countries by income. http:// datatopics.worldbank.org/world-development-indicators/stories/ the-classification-of-countries-by-income.html, accessed 15 July 2019. Yesilyurt, M. E. and J. P. Elhorst (2017). Impacts of neighboring countries on military expendi- tures: A dynamic spatial panel approach. Journal of Peace Research 54 (6), 777{790. Zielinski, R. C. (2016). How States Pay for Wars. Ithaca, NY: Cornell University Press. Zielinski, R. C., B. O. Fordham, and K. E. Schilde (2017). What goes up, must come down? The asymmetric eects of economic growth and international threat on military spending. Journal of Peace Research 54 (6), 791{805. 215

Abstract (if available)

Abstract This research illuminates facets of the interplay between states’ military and administrative capacity—emphasizing that coercive and non-coercive aspects of state capacity cannot be studied in isolation. A guiding question for this dissertation is how to measure the capacity of states engulfed in conflictual interactions with internal and external adversaries? I center on the link between states’ administrative/bureaucratic and military capacity. In interstate conflicts, a state’s ability to develop military capacity in part depends on the degree to which it can extract surplus resources that can be invested in the military, or other goods and services. In intrastate conflicts, and particularly in asymmetric civil wars, a state’s administrative capacity determines not just its material military capability, but also its ability to defeat insurgents that seek cover among the civilian population. ❧ This collection of essays demonstrates that measurement innovation can help overcome some of the shortcomings in the availability of data regarding state capacity in conflict. The work contributes new insights and methods to the study of conflict processes at the intrastate and interstate levels by offering improved strategies toward the measurement of key concepts of interest, such as territorial control in civil war and economic power in the international system. ❧ In Chapter 2, I develop and validate a measurement model for the estimation of territorial control in asymmetric civil war. For most contemporary and historic conflicts, we lack detailed information on who controls a subnational area at a given point in time. I advance a theoretical model of actor behavior in asymmetric civil war and leverage geo-coded conflict event data to compute territorial control estimates for insurgencies in Nigeria and Colombia. In addition, I advance an improved method to estimate the exposure of subnational geographic units to conflict events. ❧ In Chapter 3, I engage with one of the dominant paradigms in the study of asymmetric civil war: the notion that counterinsurgents seek to capture the ‘hearts and minds’ of the population in order to gain the upper hand in the conflict and establish territorial control. Using data from Colombian municipalities, I highlight the strategic importance of welfare spending by showing that governments do react to violence with higher investment in citizen welfare, but only if violence presents a direct threat to the survival of the state. ❧ Chapter 4 champions a measurement correction for one of the most widely used theoretical constructs in international relations research: economic power in the international system. Together with my co-authors, I show that the mis-measurement of economic power via gross domestic product (GDP) can be mended by accounting for subsistence, that is the resources a state’s population needs for survival. We introduce surplus domestic product (SDP) as a better measure for the resources that states can invest in the military or citizen welfare. In addition, the chapter provides an improved method to estimate potential threat in states’ geopolitical environment and an extended data series of military expenditure as a proportion of economic resources (GDP or SDP) from 1816–2012. ❧ Conflict zones in particular are characterized by a dearth of information—hindering academic research and data-driven policy interventions. It is my hope and ambition for this dissertation to provide the field of conflict studies with higher quality data to stimulate future research and innovation.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Public opinion and international affairs: a multi-method approach to foreign policy attitudes

PDF

Essays in political economy and mechanism design

PDF

Essays on the politics of the American welfare state

PDF

Essays on health insurance programs and policies

PDF

International politics and domestic institutional change: the rise of executive war-making autonomy in the United States

PDF

Like father, like son? A succession-based explanation for conflict initiation by authoritarian regimes

PDF

Economic policy-making in a changing environment: the politics of success in Peru vis-à-vis China

PDF

Selling virtue: how human rights NGOs and their donors work together to create a better world ... for themselves

PDF

Essays on the empirics of risk and time preferences in Indonesia

PDF

Three essays on social policy: institutional development, and subjective well-being as a cause and consequence of labor market outcomes

PDF

Essays on sovereign debt

PDF

Essays on political economy and corruption

PDF

Habits and policy: the socio-cognitive foundations of foreign policy stability

PDF

The origins and evolution of the U.S. alliance network: how military allies transform and transact

PDF

When passions run high: emotions and the communication of intentions in face-to-face diplomacy

PDF

Organized intraethnic conflict and cooperation in divided stateless nations: the competitive dynamics of rival ethnonationalist factions

PDF

Essays on firm investment, innovation and productivity

PDF

Essays on the estimation and inference of heterogeneous treatment effects

PDF

Essays on macroeconomics and income distribution

PDF

Challenges from below: the origins of status competitions in world politics

Asset Metadata

Creator Anders, Therese (author)

Core Title Perspectives on state capacity and the political geography of conflict

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Political Science and International Relations

Publication Date 05/15/2020

Defense Date 05/14/2020

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag Civil War,economic resources,OAI-PMH Harvest,public goods,territorial control

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor James, Patrick (committee chair), Fariss, Christopher (committee member), Nugent, Jeffrey (committee member), Rathbun, Brian (committee member)

Creator Email tanders.usc@gmail.com,tanders@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-312764

Unique identifier UC11663631

Identifier etd-AndersTher-8538.pdf (filename),usctheses-c89-312764 (legacy record id)

Legacy Identifier etd-AndersTher-8538.pdf

Dmrecord 312764

Document Type Dissertation

Rights Anders, Therese

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA