Essays in Development Economics: Case Studies from Benin and the DRC
Nik Stoop
Dissertation submitted in fulfilment of the requirements for the joint degree of Doctor in Economics at the University of Leuven, Doctor of Development Studies at the University of Antwerp
Number 573 2017
Doctoral Committee
Supervisors Prof. Dr. Jo Swinnen KU Leuven Prof. Dr. Marijke Verpoorten University of Antwerp & KU Leuven
Members Prof. Dr. Tom De Herdt University of Antwerp Prof. Dr. Joachim De Weerdt University of Antwerp & KU Leuven Prof. Dr. Dominic Parker University of Wisconsin-Madison Prof. Dr. Peter van der Windt New York University Abu Dhabi
Chairman Prof. Dr. Willem Moesen KU Leuven
i
Daar de proefschriften in de reeks van de Faculteit Economie en Bedrijfswetenschappen het persoonlijk werk zijn van hun auteurs, zijn alleen deze laatsten daarvoor verantwoordelijk.
The views expressed in dissertations published by the Faculty of Economics and Business are those of their authors, and not necessarily represent those of the University of Leuven.
ii Acknowledgements
I would like to start these acknowledgements with a fragment from Ariel Rubinstein’s ‘Experienced Advice for “Lost” Graduate Students in Economics’:
“Remember that you are one of the most privileged people on earth. Society has given you a wonderful opportunity. You are supposed to do whatever you want, to think about new ideas, to express your views freely, to do things in the way that you choose, and on top of that you will be rewarded nicely for doing so. These privileges should not be taken for granted. We are extremely lucky, and we owe something in return.”
Indeed, I feel privileged for having had the opportunity to pursue my research interests over the past years. I owe thanks to many people for giving me this opportunity, and for contributing to this thesis in one way or another.
First and foremost, I owe thanks to my supervisors. Marijke, being your PhD student has been a fantastic learning experience. I consider you a great mentor. Your passion for research is inspiring, and I was very happy to pursue our similar research interests – ranging from Voodoo in Benin to mining in Congo. I believe the number of transportation modes that we shared, in different parts of the world, partly reveals the extent of our close collaboration. I count at least nine. A fishing boat, zemidjan and mountain bike in Benin; a taxi-moto, jeep, and kayak in Congo; a bus and taxi in Rwanda; and the train from Leuven to Antwerp. While moving from one place to another, I always enjoyed our conversations about research, and life in general. Jo, I am very grateful for your continued support throughout the years, both as my supervisor and as the director of LICOS. I was happy to be part of the amazing LICOS team, where my PhD journey started. Apart from being an incredibly enriching institute in terms of research capacities, it is also a fantastic working environment, in no small terms thanks to your efforts. As such, working at LICOS not only helped me to grow as a researcher, but also to put my yeast allergy behind me and start enjoying beers.
I would also like to thank the other members of my doctoral committee. Peter, I found our collaboration very enriching. Your many comments and suggestions strongly increased the quality of the research. Joachim, thanks for your helpful feedback and advice at many occasions over the past years. Also, thanks for putting me in touch with the people of SALDRU in Cape Town. It looks like I’ll be spending some more time there. Tom, thanks for your role as chair of my doctoral committee at IOB – the other fantastic institute which I was lucky enough to collaborate with, and
iii which facilitated the Congo-based research. Dominic, many thanks for your helpful comments, for your willingness to share data, and for flying all the way here for my defense. I would like to thank all of you for the time and effort you put in reviewing the chapters of this thesis. Finally, I would like to thank Willem Moesen for his willingness to act as chair during the defense.
This thesis further benefitted from much appreciated feedback and comments from many other people, in particular, Jean-Paul Azam, Jean-Marie Balland, Elena Briones Alonso, Pascaline Dupas, Koen Deconinck, Marcel Fafchamps, Sara Geenen, Catherine Guirkinger, Romain Houssa, Enzo Nusio, Gabriel Picone, Jean-Philippe Platteau, Olivia Rutazibwa, Marco Sanfilippo, Petros Sekeris, Anja Shortland, Nina von Uexküll, Leonard Wantchekon, Joshua Wilde, and participants at various conferences and workshops. I thank Research Foundation Flanders (FWO) and the University Research Fund of the University of Antwerp (BOF) for funding my PhD scholarship. The financial support of FWO, VLIR-UOS, CEGEMI and the IOB Research Fund was crucial to conduct the fieldwork.
I am very grateful that my PhD gave me the opportunity to travel to two fascinating countries: Benin and the Democratic Republic of Congo. Doing fieldwork allowed me to get a much deeper understanding of the local context. Moreover, I got to experience things which you don’t usually get to experience as a tourist, such as transporting IT equipment by fishing boat in Benin and descending in artisanal gold mines in Eastern Congo. The fieldwork in Benin would not have been possible without Romain Houssa. I would like to thank you for your seemingly endless energy, for facilitating the fieldwork in so many ways, and for introducing me to your country of birth. Thanks Elena, for all the (good) time(s) we spent in designing and executing the pilot experiment. Perhaps it will still be scaled up in the future. I also thank the entire team of enumerators, especially Jonas Zoki and Médard Kounou who guided us through the stilt villages on lake Nokoué. Finally, I am very grateful to the fishermen who offered their time to participate in our research activities. In the second half of my PhD, I spent a considerable amount of time in the DRC. From the first moments, the entire Muhanzi family (Mundi, Maguy, Axel, Gabriel, Deborah, Philippe, Michael and Diana) made me feel right at home in avenue Nyo Fu. A special thanks to Mundi, who provided tremendous assistance in navigating the Congolese bureaucracy when our tablets got stuck in Kinshasa, and who facilitated our fieldwork in many ways. I further owe thanks to my colleague Janvier Kilosho, who provided invaluable assistance in preparing and executing the different rounds of fieldwork. The survey team worked hard, in sometimes challenging
iv circumstances. I thank the enumerators: Alex Nyakabasa, Célestin Mukotanyi Munyali, Fortunat, Bamporiki Bisanga, Gabriel Mugisho Dunia, Isidore Barhanywerha Baderhakuguma, John Kadjunga, Jules Nyunda Nkuru, Olivier Rubambura Kabuye, Pascal Barhanywanywa, Serge Nyembo Charles and Teiggy Birhula Mongane. Their work was greatly facilitated by the local guides in Kamituga: Belgique Babingwa, Jean Bisimwa, Leonard Kabungulu, Oswald Bilinganene, Paul Aishi Wabutongo and Waluna Itongwa. Most importantly, this research would not have been possible without the artisanal miners of Kamituga who offered their time to be interviewed. I offer them my sincere gratitude.
Throughout the years, I was lucky enough to interact and collaborate with many wonderful people. I would like to thank all my former and current colleagues at LICOS and IOB for the inspiring and fun times, for the many research-related discussions and suggestions, and for the wonderful lunches and great parties. You are too numerous to mention everyone explicitly, but know that it was great collaborating with you all, and that many of you have become good friends. At LICOS, special thanks go to Elena, my office mate, with whom I had a great time discovering Benin and who made sure our office had an endless stock of healthy snacks; Joachim, for all the nice trips we made to Napoli, Rome, Germany and Stockholm – too bad you got us kicked out of that one nightclub; Andrea, for all the fantastic evenings, for your bad jokes in the morning, and for giving us the opportunity to travel to Stockholm and Germany; Hannah, for being my partner in crime in the Party Committee; Maria, for your energy and your hilarious stories; Jeroen, for introducing me to Kalfort culture; and Emma and Senne, for the competitive ping-pong games. I also thank Elfriede and Nathalie, for their great assistance in many administrative and non-administrative issues. At IOB, I would like to extend special thanks to Sara, for the many helpful suggestions and whose network literally opened many doors for us in Bukavu and Kamituga; Klara, for the good talks we had, both in Antwerp and Bukavu; and An Vermeesch who, together with Ann Hasendonckx at KUL, offered tremendous assistance in managing the DRC fieldwork budgets. Finally, I thank everyone at IPIS for their collaboration and for sharing their invaluable data; and I thank the people of SALDRU – especially those in the ‘fish bowl’ – for hosting me during the last months of my PhD.
I also want to thank the people who helped me take my mind off the PhD. I thank the so called ‘ladies’ (Bertolli, Brambo, Danny, Davidoff, Fabré, Linda, Luitenant Verboven, Matt, Timberly, Tunagu) and the other climbers for the numerous great holidays, weekends, training sessions and
v drinks. I thank the ‘Rijmenamkliek’ (Anneleen, Ans, Bregt, Katrien, Sarah, Sil, as well as their partners and ever-growing offspring) and of course the ‘Geile Schwanz’ (Boelli, den Duerinckx, Karain, PJ, Shari, Michael) for the fun weekends, parties, dinners, running sessions and board game evenings.
I owe much to my family, whose unconditional support has been very important. I thank my parents, for giving me so many opportunities and for making my studies possible. Their adventurous spirit and curiosity to discover new places and cultures – which they passed on to both their children – most certainly influenced many of my life choices, including starting a PhD in development economics. I also thank my sister, who always has my back. Finally, I thank Kim. For the person she is, for the person she makes me, and for co-authoring the first chapter of our life as parents: Anna.
vi Contents
ACKNOWLEDGEMENTS ...... III CONTENTS ...... VII 1. INTRODUCTION ...... 1 REFERENCES ...... 6 2. VOODOO, VACCINES AND BED NETS ...... 9 2.1. INTRODUCTION ...... 10 2.2. BACKGROUND ...... 14 2.2.1. African traditional religion, African cosmology and healthcare ...... 14 2.2.2. The rise, fall and renaissance of Voodoo in Benin ...... 16 2.2.3. Voodoo and healthcare in Benin ...... 17 2.3. DATA DESCRIPTION ...... 18 2.3.1. Religion ...... 18 2.3.2. Health ...... 19 2.3.3. Correlation between ATR and healthcare ...... 20 2.4. ACCOUNTING FOR OBSERVABLES ...... 21 2.5. ACCOUNTING FOR UNOBSERVABLES ...... 24 2.5.1. Subsample analysis: bed net owners and father’s characteristics ...... 25 2.5.2. Using selection on observables to assess the bias from unobservables ...... 26 2.5.3. Instrumental variable approach ...... 26 2.6. TESTING FOR CHANNELS OF CAUSALITY: WORLD VIEW VERSUS TRADITIONAL HEALERS ... 28 2.6.1. The ATR worldview ...... 28 2.6.2. Traditional healers ...... 29 2.7. DISCUSSION ...... 31 REFERENCES ...... 34 FIGURES ...... 40 TABLES ...... 41 APPENDIX ...... 54 3. TO FISH OR NOT TO FISH? RESOURCE DEGRADATION AND INCOME DIVERSIFICATION IN BENIN ...... 71 3.1. INTRODUCTION ...... 72 3.2. BENIN’S INLAND FISHERIES ...... 74 3.3. DATA AND DESCRIPTIVE STATISTICS ...... 75 3.3.1. Survey design ...... 75 3.3.2. Natural resource degradation ...... 76 3.3.3. Income diversification ...... 76 3.3.4. Degradation and Income Diversification ...... 77 3.4. ECONOMETRIC FRAMEWORK ...... 78 3.4.1. Individual level controls ...... 79 3.4.2. Household and community level controls ...... 80 3.4.3. Endogeneity of degradation ...... 80 3.5. MAIN RESULTS ...... 82 3.6. ALTERNATIVE SPECIFICATIONS AND ROBUSTNESS CHECKS ...... 83 3.6.1. Alternative measures of degradation ...... 83 3.6.2. Representativeness of the sample ...... 84 3.6.3. Robustness checks ...... 84 3.7. DISCUSSION ...... 85
vii REFERENCES ...... 87 FIGURES ...... 90 TABLES ...... 93 APPENDIX A: FIGURES & TABLES ...... 97 APPENDIX B: ALTERNATIVE SPECIFICATIONS ...... 108 APPENDIX C: ROBUSTNESS CHECKS ...... 114 4. ARTISANAL OR INDUSTRIAL CONFLICT MINERALS? EVIDENCE FROM EASTERN CONGO ...... 123 4.1. INTRODUCTION ...... 124 4.2. CONTEXT: CONFLICT AND MINING IN EASTERN CONGO ...... 127 4.2.1. Artisanal Mining ...... 127 4.2.2. Large-Scale Mining ...... 128 4.3. CONCEPTUAL FRAMEWORK ...... 129 4.3.1. Previous Literature: Rapacity, Protection and Opportunity Costs ...... 129 4.3.2. Area, Actors and Actions ...... 130 4.3.3. Shocks ...... 130 4.3.4. Hypotheses ...... 131 4.4. DATA SOURCES AND DESCRIPTION ...... 133 4.4.1. Data sources ...... 133 4.4.2. Descriptive Statistics ...... 135 4.5. EMPIRICAL STRATEGY AND IDENTIFICATION ...... 137 4.6. RESULTS ...... 139 4.7. ROBUSTNESS ...... 141 4.8. CONCLUSION ...... 143 REFERENCES ...... 145 FIGURES ...... 150 TABLES ...... 154 APPENDIX A: WORLD PRICE OF COPPER ...... 157 APPENDIX B: EXCLUDING THE DODD-FRANK PERIOD ...... 158 APPENDIX C: EXCLUDING TANTALUM ...... 161 APPENDIX D: ADDRESSING REVERSE CAUSALITY IN THE EXPANSION OF LSM ...... 162 APPENDIX E: ROBUSTNESS CHECKS ...... 166 5. WOULD YOU REBEL? AN INQUIRY AMONG HIGH-RISK YOUTH IN EASTERN CONGO ...... 173 5.1. INTRODUCTION ...... 174 5.2. BACKGROUND ...... 177 5.3. WHY REBEL ? ...... 180 5.3.1. Grievances...... 180 5.3.2. Selective incentives ...... 181 5.3.3. Exposure to conflict ...... 182 5.4. DATA COLLECTION ...... 183 5.5. DATA DESCRIPTION ...... 185 5.5.1. Self-reported motivations to join an armed group ...... 185 5.5.2. The intention to rebel ...... 185 5.5.3. Determinants to rebel ...... 187 5.5.4. Control variables ...... 190 5.6. EMPIRICAL STRATEGY AND DATA ANALYSIS ...... 191 5.6.1. Results ...... 192 5.6.2. Correlation versus causation ...... 193 5.6.3. Nested models and interactions ...... 193
viii 5.6.4. Robustness checks ...... 194 5.7. DISCUSSION ...... 194 REFERENCES ...... 197 FIGURES ...... 202 TABLES ...... 204 APPENDIX A: ACTS OF REBELLION BY ARTISANAL MINERS ...... 210 APPENDIX B: SELECTION ON UNOBSERVABLES ...... 213 APPENDIX C: ROBUSTNESS CHECKS ...... 216 6. CONCLUSION ...... 219
ix
x Chapter 1
1. Introduction
This thesis comprises four essays that fall under the broad umbrella of ‘development economics’. In writing it, I am completing a joint PhD in economics and ‘development studies’. It is worthwhile briefly discussing these notions before introducing the actual essays. The definition of what constitutes ‘development’ is controversial, and changes over time. Yet, common among most definitions is “that ‘development’ encompasses ‘change’ in a variety of aspects of the human condition” (Sumner, 2008: p.10). The dimensions of human development are manifold, including amongst others, economic, social, political and legal structures, the environment, peace and religion. This multi-dimensionality is also reflected in the Human Development Report, the annual flagship publication of the United Nations Development Programme. The latest report states that “Human development is about enlarging freedoms so that all human beings can pursue choices that they value” (UNDP 2016: p.1). The report provides a Human Development Index which ranks countries worldwide on three main dimensions: health, education and income. Additional indices aim to provide measures of inequality, women’s empowerment and non-income dimensions of poverty such as access to public services, nutrition and housing. Given the many dimensions of development, it is not surprising that ‘development studies’ is characterized as a multidisciplinary and interdisciplinary field of study, drawing on knowledge of various academic disciplines, including economics, political science, sociology, anthropology and demography (EADI 2005; Sumner 2006; Potter 2014). Other distinctive features are its focus on developing countries and its normative point of departure – trying to improve people’s lives – translating to a commitment to conduct policy-relevant research (EADI 2005; Sumner 2006). Indeed, development studies “aims at contributing to possible solutions to societal problems that development or its absence may produce” (EADI 2005: p.4). Being closely related, ‘development economics’ has been defined as “a subject that studies the economics of the developing world, and has made excellent use of economic theory, econometric methods, sociology, anthropology, political science, biology and demography” (Debraj 2008, p.1). Compared to development studies more general, it can be said to emphasize methods of economic analysis. The four essays that constitute this thesis have in common many of the above-mentioned characteristics. First, all essays empirically analyze various aspects of human development. Chapter 2 investigates the uptake of preventive healthcare measures; Chapter 3 studies the link between employment and environmental degradation; while Chapters 4 and 5 are concerned with armed
1 Chapter 1
conflict. Second, the focus is on two developing countries in Sub-Saharan Africa. Chapters 2 and 3 present case studies from Benin, while Chapters 4 and 5 study the Democratic Republic of Congo (DRC). Benin is a small country in West-Africa with approximately 11 million inhabitants. The DRC is vast country in Central-Africa with population estimates exceeding 82 million. The latest Human Development Report categorizes Benin and the DRC as having low human development, ranking them at the very bottom of the Human Development Index, occupying positions 167 and 176 out of 188 listed countries (UNDP 2016: p.22–25). Third, the essays exploit insights from various academic disciplines. For instance, Chapter 2 relies on anthropological and ethnographic research in formulating the hypothesis and devising the econometric strategy; in Chapter 3, anthropologic accounts provide background information and guide the empirical analysis; while the conceptual framework of Chapters 4 and 5 strongly relies on political science theories. Fourth, although applied econometrics – “Economists’ use of data to answer cause-and-effect questions” (Angrist and Pischke 2015: p.xi) – takes a central role, the analysis in each essay draws on a mix of methodologies. For each country, the first chapter makes us of existing databases that are representative of a large area (Chapters 2 and 4), while the second chapter zooms in on a local area and relies on original data, collected during several rounds of fieldwork (Chapters 3 and 5). Moreover, the fieldwork that underlies Chapters 3 and 5 was characterized by a mix of qualitative and quantitative methods, where the design of structured surveys strongly relied on field observations, focus group discussions and open interviews. Finally, underlying each essay is a policy-relevant question, and the analyses aim to provide policy recommendations. I now briefly introduce each essay.
Chapter 2 starts from the observation that Sub-Saharan Africa is the region with the highest under-five mortality rates, at 92 deaths per 1,000 live births (UN IGME 2014). Vaccines and bed nets rank among the most cost-effective measures to reduce child mortality, and their increased supply has greatly contributed to a decline of mortality rates (Bloom et al. 2005; Kenny 2009; Laxminarayan et al. 2006; Miller and Sentz 2006; Webster et al. 2005; WHO 2008). Yet, although often available at low costs, households in developing countries tend to underinvest in preventive healthcare measures (Banerjee and Duflo 2009; Dupas 2011). The lower-than-optimal uptake is explained among others by liquidity constraints, time-inconsistent preferences and a lack of verifiable information on the cost-effectiveness of the measures (Dupas 2011). Lacking such information, caretakers turn to heuristic decision-making, looking for rules of thumb, opinions of others, behavior of neighbors or their own understanding of sickness and health. Based on a large ethnographic literature, we conjecture that African Traditional Religion (ATR) may be an
2 Chapter 1
important input for such heuristic decision-making, through the authority of its religious leaders, or because of its distinct worldview. Our case study, Benin, is typical for a Sub-Saharan African country as ATR-related beliefs and practices are widespread. On the other hand, Benin is atypical as people freely report ATR adherence. Its main ATR – Voodoo – is awarded the same status as monotheistic religions, and about 20% of the population reports adherence. The data come from four nationally representative rounds of Demographic and Health Surveys, collected between 1996 and 2012. The revealed ATR belief and its substantial within-village and within-household variation allow us to estimate its impact on the uptake of several healthcare measures and outcomes, while controlling for a large set of confounding factors.
While remaining in Benin, Chapter 3 moves the focus from health to employment in the fisheries sector. In Benin, the fisheries sector is of great importance to both national and rural economic development. Most fish are caught in the three main coastal lakes: Lake Nokoué, lake Ahémé and Porto-Novo lagoon. Resource degradation has however strongly reduced the fish stock available to the communities located at these lakes, threatening their livelihood (Gnohossou 2006; Niyonkuru and Lalèyè 2010). This chapter relies on original survey data, collected from fishing communities at the three main coastal lakes. The communities we study are remote and characterized by underdeveloped labor and credit markets. We examine if and to what extent resource degradation induces artisanal fishermen to reallocate their labor away from fishing activities. Understanding this impact is important. Not only because diversification directly affects income and thus poverty (Barrett et al. 2001), but also because it may alleviate environmental degradation (Reardon and Vosti 1995; Forsyth et al. 1998; Swinton et al. 2003; Ellis and Allison 2004).
In Chapter 4 we move from Benin to the Democratic Republic of Congo. The DRC is a textbook case when it comes to the resource curse. Its untapped deposits of raw minerals are estimated to be worth US$24 trillion (UNEP 2011), but the majority of its population is dismal poor, mainly because both war and political mismanagement have ravaged the country. The eastern part of the country is home to many artisanal mining sites and a growing number of industrial mining concessions. At the same time, it is the part of Congo that is most affected by persistent episodes of conflict, even after the formal end of the two Congo wars (1996-1998 and 1998-2003). Several advocacy groups have argued that ‘conflict minerals’ contribute to perpetuating the violence (Autesserre 2012; Prendergast 2009; Seay 2012). This ‘conflict minerals’ narrative has
3 Chapter 1
mainly targeted artisanal mining. However, to date, we know very little about how different modes of mineral extraction relate to conflict. There are two main ways to extract mineral resources: artisanal mining (ASM) and large- scale mining (LSM). ASM refers to a largely manual mode of extraction, practiced by individuals, groups or communities. LSM refers to a mechanized mode of production, practiced by large, often international, companies. To understand how ASM and LSM may impact local conflict dynamics, we exploit detailed, geo-referenced information about artisanal mining sites, large-scale mining concessions and conflict events in Eastern Congo. We take advantage of two types of shocks that affected ASM and LSM activities in Congo in the period 2004 to 2015. First, a surge in world mineral prices translated to large increases in the value of Congolese minerals. Second, the introduction of a new Mining Code and Mining Regulations in 2002 and 2003 led to a strong increase in the number of granted LSM research and production permits.
Chapter 5 complements Chapter 4, as it provides a case study of a mining site in Eastern Congo where an LSM company is about to move to the production phase, while thousands of artisanal miners are operating within its concession. Within this context, we study what motivates individuals to take up arms and fight. Several decades of research on armed conflict have yielded relatively few quantitative empirical analyses on the individual propensity to rebel (compared to the large number of ethnographic case studies and cross-country studies). The gap in the literature is understandable however, given the difficulty of approaching (past or potential) recruits in conflict and post-conflict areas. To address this gap, we collected original data at the mining site of Kamituga in South-Kivu. We inquired about the ‘intention to rebel’ among a representative sample of artisanal miners who live and work in Kamituga. The miners are all (young) men; the vast majority was exposed to the violence of the first and second Congo wars, and some participated in the violence. Despite the formal end of the war in 2003, pockets of chronic violence remained in the surroundings of Kamituga at the time of our survey in May 2015 (Stearns and Vogel 2015). Moreover, as the mining concession on which Kamituga is located was granted to a large-scale mining company, the miners faced an uncertain economic future – risking eviction once the company would move from the research to the production phase. It is to these so-called ‘high-risk youth’, who experienced violent conflict and were at risk of losing their employment, that we asked the question “would you rebel?”. We framed this question as a manifestation against the large- scale mining company, and enquire about miners’ intention to engage in four concrete rebellious actions: destroying property, attacking employees, using fire arms and joining an armed group.
4 Chapter 1
Chapter 6 concludes by briefly summarizing the main findings of each essay and its policy recommendations.
5 Chapter 1
References
Angrist, Joshua David, and Jörn-Steffen Pischke. 2015. Mastering ’Metrics: The Path from Cause to Effect. Princeton; Oxford: Princeton University Press. Autesserre, Séverine. 2012. “Dangerous Tales: Dominant Narratives on the Congo and Their Unintended Consequences.” African Affairs 111 (443): 202–22. Banerjee, Abhijit V., and Esther Duflo. 2009. “The Experimental Approach to Development Economics.” Annual Review of Economics 1 (1): 151–78. Barrett, C.B, T Reardon, and P Webb. 2001. “Nonfarm Income Diversification and Household Livelihood Strategies in Rural Africa: Concepts, Dynamics, and Policy Implications.” Food Policy 26 (4): 315–31. Bloom, David E., David Canning, and Mark Weston. 2005. “The Value of Vaccination.” World Economics 6 (3): 15. Debraj, Ray. 2008. “Development Economics.” In The New Palgrave Dictionary of Economics, edited by Steven N. Durlauf and Lawrence E. Blume, 2nd ed. Palgrave Macmillan. Dupas, Pascaline. 2011. “Health Behavior in Developing Countries.” Annual Review of Economics 3 (1): 425–449. EADI. 2005. “Development Studies, Accreditation and EADI. a Vision Paper Presented to the EADI Executive Committee.” European Association of Development Research and Training Institutes. Ellis, Frank, and Edward Allison. 2004. “Livelihood Diversification and Natural Resource Access.” Overseas Development Group, University of East Anglia. Forsyth, Tim, Melissa Leach, and Tim Scoones. 1998. “Poverty and Environment: Priorities for Research and Study-an Overview Study, Prepared for the United Nations Development Programme and European Commission.” Gnohossou, P.M. 2006. “La Faune Benthique D’une Lagune Ouest Africaine (Le Lac Nokoué Au Benin), Diversité, Abondance, Variations Temporelles et Spatiales, Place Dans La Chaine Tropique.” Dissertation. Institut Nationale Polytechnique de Toulouse, France. Kenny, Charles. 2009. “There’s More to Life than Money: Exploring the Levels/Growth Paradox in Income and Health.” Journal of International Development 21 (1): 24–41. Laxminarayan, Ramanan, Jeffrey Chow, and Sonbol A. Shahid-Salles. 2006. “Intervention Cost- Effectiveness: Overview of Main Messages.” In Disease Control Priorities in Developing Countries. Washington, DC: World Bank and Oxford University Press. Miller, Mark A., and John T. Sentz. 2006. “Vaccine-Preventable Diseases.” In Disease and Mortality in Sub-Saharan Africa, edited by Dean T Jamison, Richard G Feachem, Malegapuru W Makgoba, Eduard R Bos, Florence K Baingana, Karen J Hofman, and Khama O Rogo, 2nd ed. Washington (DC): World Bank. Niyonkuru, C., and P.A. Lalèyè. 2010. “Impact of Acadja Fisheries on Fish Assemblages in Lake Nokoué, Benin, West Africa.” Knowledge and Management of Aquatic Ecosystems, no. 399 (November): 5. Potter, Robert B. 2014. “The Nature of Development Studies.” In The Companion to Development Studies, edited by Vandana Desai and Robert B. Potter, Third edition. London; New York: Routledge. Prendergast, John. 2009. “Can You Hear Congo Now? Cell Phones, Conflict Minerals, and the Worst Sexual Violence in the World.” The Enough Project. Reardon, Thomas, and Stephen A. Vosti. 1995. “Links between Rural Poverty and the Environment in Developing Countries: Asset Categories and Investment Poverty.” World Development 23 (9): 1495–1506.
6 Chapter 1
Seay, L. 2012. “What’s Wrong with Dodd-Frank 1502? Conflict Minerals, Civilian Livelihoods, and the Unintended Consequences of Western Advocacy.” 284. Working Paper. Center for Global Development. Stearns, Jason, and Christoph Vogel. 2015. “The Landscape of Armed Groups in the Eastern Congo.” Congo Research Group, Center on International Cooperation. Stoop, Nik, Janvier Kilosho, and Marijke Verpoorten. 2016. “Relocation, Reorientation, or Confrontation? Insights from a Representative Survey Among Artisanal Miners in Kamituga, South-Kivu.” IOB Working Paper 2016.09. Institute of Development Policy and Management, University of Antwerp. Sumner, Andrew. 2006. “What Is Development Studies?” Development in Practice 16 (6): 644–50. ———. 2008. “What Is ‘Development’?” In International Development Studies: Theories and Methods in Research and Practice, edited by Andrew Sumner and Michael A. Tribe. SAGE Publications, Inc. Swinton, Scott M, Germán Escobar, and Thomas Reardon. 2003. “Poverty and Environment in Latin America: Concepts, Evidence and Policy Implications.” World Development 31 (11): 1865–72. UN IGME. 2014. “Levels & Trends in Child Mortality - Report 2014.” UN Inter-Agency Group for Child Mortality Estimation. UNDP. 2016. “Overview Human Development Report 2016. Human Development for Everyone.” United Nations Development Programme. UNEP. 2011. “Post-Conflict Environmental Assessment of the Democratic Republic of Congo: Synthesis Report for Policy Makers.” Nairobi, Kenya: United Nations Environment Programme. Webster, Jayne, Jo Lines, Jane Bruce, Joanna Armstrong Schellenberg, and Kara Hanson. 2005. “Which Delivery Systems Reach the Poor? A Review of Equity of Coverage of Ever- Treated Nets, Never-Treated Nets, and Immunisation to Reduce Child Mortality in Africa.” The Lancet Infectious Diseases 5 (11): 709–17. WHO. 2008. “Vaccination Greatly Reduces Disease, Disability, Death and Inequity Worldwide.” Bulletin of the World Health Organization.
7 Chapter 1
8 Chapter 2
2. Voodoo, vaccines and bed nets *
Summary
This chapter provides the first quantitative analysis to scrutinize the ample ethnographic evidence that magico-religious beliefs affect the demand for conventional healthcare in Sub-Saharan Africa. We rely on the unique case of Benin, where Voodoo-adherence is freely reported, and varies greatly within villages and even within households, yet can be traced to historic events that are arguably exogenous to present-day healthcare behavior. These features allow us to account for confounding village- and household-factors, and address self-selection into Voodoo. We find that Voodoo adherence of the mother is associated with lower uptake of preventive healthcare measures and worse child health outcomes, a relationship that weakens but remains when controlling for village dummies and a large set of observables. We employ three different strategies to test for the potential influence of unobservables. The results suggest that the estimated Voodoo-effects are partly causal. A tentative exploration of the causal mechanisms suggests a mediating role of traditional healers.
* This chapter is based on a paper written with Marijke Verpoorten (University of Antwerp, IOB) and Koen Deconinck (University of Leuven, Licos). The original paper is forthcoming in Economic Development and Cultural Change. We received much appreciated comments from Jean-Marie Balland, Elena Briones Alonso, Pascaline Dupas, Tom de Herdt, Joachim de Weerdt, Marcel Fafchamps, Catherine Guirkinger, Romain Houssa, Gabriel Picone, Jean-Philippe Platteau, Olivia Rutazibwa, Marco Sanfilippo, Petros Sekeris, Leonard Wantchekon, Joshua Wilde and two anonymous referees. We also benefited from useful comments from participants at seminars, conferences and workshops in Leuven (LICOS-KULeuven), Oxford (CSAE conference), San Francisco (ASREC conference), Antwerp (IOB-UA), Benin (joint CRED-ASE workshop), Brussels (St. Louis) and Helsinki (UNU-WIDER).
9 Chapter 2
2.1. Introduction
Sub-Saharan Africa (SSA) is the region with the highest under-five mortality rates, at 92 deaths per 1,000 live births (UN IGME 2014). Vaccines and bed nets rank among the most cost-effective measures to reduce child mortality in SSA, and their increased supply has greatly contributed to a decline of mortality rates (Bloom et al. 2005; Kenny 2009; Laxminarayan et al. 2006; Miller and Sentz 2006; Webster et al. 2005; WHO 2008). Yet, although often available at low costs, their uptake is far from perfect (Dupas 2011). In order to bridge the last mile, we need to understand what is holding back uptake. In recent years, scholars have made progress in this direction. Several studies have shown that the demand for preventive healthcare measures is extremely price-sensitive, falls victim to procrastination, and suffers from a lack of information on their cost-effectiveness (Banerjee et al. 2010; Cohen and Dupas 2010; Dupas 2009; Hoffmann et al. 2009; Jalan and Somanathan 2008; Kremer and Miguel 2007; Madajewicz et al. 2007). The findings call for more subsidies, more incentives for parents to act now rather than later, and more information. Providing information is crucial because parents cannot empirically observe the efficacy of some healthcare measures (e.g. vaccines), and because learning may be slow and costly for many other measures (e.g. bed nets). Simply receiving information is however rarely sufficient to change behavior (Das and Das 2003; Dupas 2011). The message (provider) needs to be perceived as credible. Whether or not this is the case ultimately depends on heuristics – defined as “a simple procedure that helps find adequate, though often imperfect, answers to difficult questions” (Kahneman 2011). Heuristics may be based on the observation of comparable outcomes (Were other – more easily observed – programs of the healthcare provider successful?), perceptions of the larger system (Do I trust the public health system?), the behavior of others (What are my neighbors doing?), the opinion of leaders (What does my (religious) leader say about the healthcare provider?) or one’s own understanding of disease and healing (Do the actions of the provider make sense to me?). Religion may thus affect the demand for healthcare, through the authority of religious leaders or by acting as a frame of reference for evaluating healthcare measures (McCullough and Willoughby 2009). That religion affects disease and healing in ways not in tune with conventional medicine is well documented for the main monotheistic religions and their spin-offs.1 In this chapter, we study if and
1 For instance, some groups of Orthodox Protestants in the Netherlands oppose vaccination because of a perceived obligation to trust in Divine Providence (Ruijs et al. 2011). (In)Famous is the Roman Catholic Church’s discouragement of condom use (Joshua 2010; Bokenkotter 1985). In Islam, there is resistance against vaccines that contain haram
10 Chapter 2
how African Traditional Religion (ATR)2 affects health behavior and outcomes. ATR’s influence may occur through the authority of its religious leaders (who often act as traditional healers) or through its sense-making role, in particular its understanding of disease (as stemming from a conflict with the spiritual world) and healing (as the result of reconciliation with the spirits or ancestors). While a vast quantitative literature exists on religiosity and health behavior in the West and in the Muslim world3, there is a lack of quantitative literature on the ATR-health linkage.4 The gap in the literature is surprising, given the continued high child mortality in SSA and the rising appreciation of behavioral economics as a useful lens for studying health demand in developing countries (World Bank 2015). Moreover, it stands in contrast with the numerous ethnographic studies on the ATR-health linkage (cf. Section 2.2). One reason for the lacuna may be related to the dearth of data on ATR beliefs in SSA, and other empirical challenges. There is widespread under-reporting of ATR adherence which can be traced back to colonial, post-colonial and missionary efforts in SSA to promote monotheistic religion as the only socially acceptable choice (Neill 1991). Self-reported ATR adherence therefore tends to be a poor measure of actual ATR beliefs and practices. Another empirical difficulty (shared with other religions) is that ATR beliefs are often clustered in space and correlate with several community-, household- and individual- level characteristics, reflecting the location- or ethnicity-specific (historic) spread of religions. This reduces the ceteris paribus variation in ATR and therefore hampers a meaningful quantitative analysis of its relationship with healthcare. Finally, although people often grow up with religion and are thus influenced by parents and their neighborhood (Iannaccone 1998), religious adherence is to some extent an individual choice because conversion remains possible. Therefore, any analysis of the impact of religious adherence needs to deal with significant endogeneity issues.
substances, such as the anti-meningitis vaccine that includes pork derivatives (Padela 2013). And, suspicion of the Muslim world against the West hampered vaccination campaigns in Nigeria and Pakistan (Heymann and Aylward 2004; McGirk 2015). 2 The term ‘African Traditional Religion’ was launched by Parrinder (1954) to denote African beliefs and practices that are religious but neither Christian nor Islamic. While ‘traditional’ suggests that ATR is a thing of the past, in reality, it is lived and practiced by Africans today (cf. Section 2.2.1). The term ‘traditional’ needs therefore to be understood as “handed down from generation to generation by the forebears of the present generation of Africans” (Awolalu 1976). 3 Regarding the uptake of preventive healthcare measures, the focus lies mainly on opposition against vaccination (see for instance Grabenstein 2013; Ruijs et al. 2011; Streefland 2001). Other themes vary widely, including the link between religion and risky health behaviors (e.g. Mellor and Freeborn 2011), religion and subjective well-being (e.g. Dolan et al. 2008), the health consequences of the Ramadan (e.g. van Ewijk 2011), and the Muslim advantage in child survival in India (Bhalotra et al. 2010). 4 There exist a handful of quantitative studies in the medical literature, that include ATR as a regressor when studying health outcomes (Antai 2009; Antai et al. 2009; Cau et al. 2013; Gyimah 2007; Gyimah et al. 2006), but none of these studies explicitly focuses on ATR, nor addresses omitted variable- and endogeneity bias.
11 Chapter 2
In the case of Benin, the first two of these three caveats are less severe, and history provides us with plausibly exogenous variation in self-reported ATR to address the third caveat. First, self- reported ATR in Benin is a uniquely credible indicator for actual ATR-beliefs because Benin’s main ATR – Voodoo – is awarded the same status as the monotheistic religions. It is mentioned explicitly in the constitution as an official religion, there is a yearly national Voodoo holiday, and the country is patched with Voodoo convents where Voodoo priests receive training. Because Voodoo is not marginalized socially or politically, people freely report adherence. About 20% of respondents did so in the past Demographic and Health Survey (DHS) rounds. Second, Benin is among the countries with the highest religious diversity and the lowest government restrictions on religion (Pew Research Center 2014a, 2014b). This freedom translates into considerable within-village and within-household variation in religious adherence. For instance, the average DHS survey cluster counts 3 to 4 different religious affiliations for an average sample size of only 24 mothers; and 27% of couples in Benin’s DHS do not share the same religious affiliation. Third, the history of Voodoo in Benin is well-documented, among others by missionaries who faced fierce resistance to evangelization by the kingdom of Dahomey and its initial founders, the Adja (see section 2.2.2). Relying on this recorded history, one can predict the spatial and inter-ethnic group variation in Voodoo that is inherited rather than a result of individual choice. Armed with these unique empirical advantages and four waves of nationally representative DHS surveys, we quantify the relation between a mother’s ATR adherence and two preventive healthcare measures that are known to have a strong impact on child morbidity and mortality: child immunization and the use of bed nets. We also look at two health outcomes: child mortality and malaria incidence. To identify the ATR-health relation, we first control for a large set of potentially confounding observables. In particular, we control for observables at the level of the household (e.g. asset wealth), mother (e.g. education) and child (e.g. age). Moreover, we plug in the entire set of survey cluster dummies to account for confounding supply side factors at the local level.5 As such, we are always comparing children of ATR adherents to other children in the same community. We find that a mother’s ATR adherence is associated with lower uptake of preventive healthcare measures and worse child health outcomes, a finding that is consistent with ethnographic accounts on the impact of ATR on health behavior.
5 DHS survey clusters correspond to villages in rural areas and city blocks in urban areas.
12 Chapter 2
This need not point to a causal relationship. One concern relates to the poor measurement of household-level income by the DHS proxy for asset wealth. If income is poorly controlled for, it may confound the relation between ATR and health behavior. To deal with this concern, we look at two subsample analyses. First, in the subsample of DHS households where also the husband was interviewed, we control for a husband’s religion, thereby isolating the relation between our outcome variables and the ATR adherence of the mother, who is the child’s primary caretaker. Second, when studying the use of bed nets, we further abstract from household-level access issues by restricting our sample to bed net owning households. We find that the association between a mother’s ATR adherence and our health variables weakens, but remains negative, indicating that we are not merely picking up the influence of household-level confounding factors. A remaining concern relates to certain characteristics of mother and child that may be poorly measured (e.g. education) or remain entirely unobserved. An example of the latter is a mother’s cognitive type: an intuitive rather than an analytical type could imply a “taste for superstition”, causing a mother to self-select in certain religions and health systems (Pennycook et al. 2012; Svedholm et al. 2010). Alternatively, a child may suffer from a chronic and innate poor health condition, that affects its health outcome but could also drive its mother - being disillusioned with the biomedical health system - towards ATR and its traditional healers. To assess the potential threat of these unobservables, our second strategy turns to the procedures developed by Altonji et al. (2005) and refined by Oster (2015) to investigate how much greater the influence of unobservable factors would need to be, relative to observable factors, to completely explain away the negative relationship between ATR and our health measures. Depending on the specification used, we find that the influence of unobservable factors would either have to be between 2 and 5 times greater than observable factors, or almost 10 times greater. The value of these findings depends however entirely on the unverifiable assumption that the unobservables influence selection in a similar way as the observables. Because this assumption may not be plausible, we turn to a third strategy to deal with mother- and child-level unobservables. Specifically, to further counter the concern that our results are driven by mothers’ self-selection into Voodoo, we instrument a mother’s ATR-adherence with a dummy that takes the value one for Adja mothers who currently live within the boundaries of the ancient Dahomey kingdom. The instrument exploits the fact that present-day ATR-adherence is not merely an individual choice, but is shaped by history and tradition, most notably by the fierce resistance of the ancient Dahomey kingdom and its Adja founders against evangelization (cf. section 2.2). The IV-results
13 Chapter 2
provide further evidence to support the claim that the ATR-health relationship is not entirely confounded or driven by self-selection. We conclude that, while much of the relation between ATR and health is spurious, a non-negligible part of it is extremely robust to empirical scrutiny, suggesting a causal pathway between ATR and health. As mentioned above, this causal pathway may be mediated by a certain belief system, or by ATR leaders. The worldview embedded in African cosmology does not seem to be the driving factor: the negative ATR-health relation is ATR specific, and not akin to African cosmology in general as we do not find a relation between the uptake of conventional healthcare measures and the adherence to African Independent Churches, or the belief in witchcraft. A tentative exploration of the role of traditional healers, often taken up by Voodoo priests, indicates that ATR-mothers rely more on traditional healers and – when they do – make less use of conventional medicine. Before turning to the empirical analysis, we provide background information on ATR, Voodoo in particular, and its relation with traditional and conventional healthcare.
2.2. Background
2.2.1. African traditional religion, African cosmology and healthcare
ATR beliefs and practices are firmly rooted in African cosmology6 , which is characterized by a continuum between the visible and invisible world (Geschiere 2013). Whether manifested in ATR or witchcraft beliefs, an essential characteristic of African cosmology is the day-to-day intimacy with the spiritual world: ancestors-turned-spirits can directly affect your life, living family members can turn to occult forces to bewitch you, and your local pharmacist may be your religious leader (Geschiere 2013). The intimacy between the physical and spiritual world shapes the cultural understanding of illness and healing in SSA. A disease is often not seen as a consequence of a virus, parasite or malfunctioning of the body, but as the result of witchcraft, attacks by evil spirits or a conflict between humans and their ancestors (Aikins et al. 1994; Comoro et al. 2003; Kale 1995; LeMay-Boucher et al. 2013; Maslove et al. 2009; Omonzejele 2003, 2008). To be effective, disease prevention and treatment should include contact with the spiritual world, through divination (the consultation of spirits) and sacred rituals. Only
6 Although Africa’s cosmology is diverse, the literature has distinguished a core of shared beliefs which may be said to constitute ‘African’ cosmology (Akoto and Akoto 2005; Awolalu 1976; Kanu 2013; Nyang 1982).
14 Chapter 2
then can harmony between the spiritual and physical world be maintained or restored, and the disease prevented or cured. Conventional medicine and its products, being solely focused on the physical world, may not easily take root in such a belief system. Numerous qualitative studies have indeed argued that African traditional beliefs affect the demand for healthcare, and that conventional health programs should take these beliefs into account to improve their effectiveness (Aikins et al. 1994; Comoro et al. 2003; de Sousa et al. 2011; Kale 1995; LeMay-Boucher et al. 2013; Maslove et al. 2009; Muela et al. 1998; Omonzejele 2008). These arguments were further underlined in research on HIV/AIDS prevention and treatment (Awusabo-Asare and Anarfi 1997; Kalichman and Simbayi 2004; Thomas 2007; Van Dyk 2001), and also the recent Ebola outbreak in West-Africa gave way to a similar discussion.7 In contrast to conventional medicine, traditional healers provide health services that are consistent with the local magico-religious understanding of illness and health.8 There is wide consensus that such healers have considerable authority and influence in SSA, playing an important role by providing traditional healthcare based on sacred rituals and medicinal plants, and – in some cases – by rejecting conventional medicine (e.g. Aujoulat et al. 2003; Maslove et al. 2009; Soumonni 2012; WHO 2012). Apart from ethnographic studies, a systematic data collection project on spiritual life in SSA documents the wide prevalence of magico-religious beliefs and practices (Pew Research Center 2010). Among the 25,091 respondents of 19 SSA countries in the PEW dataset, a large share mentioned having consulted a traditional healer (42.4%), believing in witchcraft (45.6%) and evil spirits (50.6%), and believing that “sacrifices to spirits or ancestors can protect you from bad things happening” (33.0%). 9 Most of the respondents are however self-reported Christians or Muslims; only 2% mentioned ATR as their religious affiliation. In most SSA countries, self-reported ATR adherence thus tends to be a poor proxy for the beliefs and rituals that are characteristic of traditional African religions. It is likely to be a much better proxy in Benin. 10 In Benin’s four DHS rounds, about 20% of
7 See for instance: BBC (2014); IFRC (2014); and Telegraph (2014). 8 Some authors also argue that traditional healers appeal to the public in another way, i.e. by offering outcome-contingent contracts (Leonard and Zivin 2005). The enforcement of this contract hinges on the patient’s fear of the power of traditional healers (who may curse them when they do not respect their payment obligations). 9 The 19 countries are: Botswana, Cameroon, Chad, DR Congo, Djibouti, Ethiopia, Ghana, Guinea Bissau, Kenya, Liberia, Mali, Mozambique, Nigeria, Rwanda, Senegal, South Africa, Tanzania, Uganda and Zambia. 10 Self-reported ATR adherence is likely to be a better proxy in countries with relatively high self-reported ATR adherence. In the full PEW sample, ATR adherence is significantly correlated with the practice of visiting traditional healers, witchcraft beliefs, belief in evil spirits, and sacrifices to ancestors – but the correlation coefficients are rather low, at 0.13, 0.10, 0.07 and 0.20 respectively. In Liberia, the PEW sample country with the highest self-reported ATR adherence (12%), the correlation coefficients are much higher, at 0.35, 0.50, 0.40 and 0.53 (own calculations from PEW dataset).
15 Chapter 2
respondents report ATR as their religious affiliation. For the vast majority of these respondents, ATR adherence equals adherence to Voodoo, which is Benin’s main traditional religion.
2.2.2. The rise, fall and renaissance of Voodoo in Benin
Voodoo became the dominant religion in Benin in the 17th century as a result of the supremacy of the Dahomey kingdom (Janssen, 2010; Law, 2004; Tall, 1995). The Dahomey kingdom was founded by the Adja, who migrated from Togo to southern Benin around the 13th century (Bourgoignie 1972; Glele 1974; Herskovits 1938; Le Hérissé 1911). Dahomey became one of the most powerful kingdoms along the West-African coast and its warfare and slave trade activities gave way to a fusion of religious beliefs and practices from various ethnic groups. This fusion led to the development of Voodoo as a ‘new’ supra-clan religion (Geschiere 2013; Manning 1982; Soumonni 2012). Voodoo priests took up powerful positions in the kingdom, and advised the king to resist evangelization. As a result, conversion to Christianism was made punishable by the death penalty and missionaries faced restricted access to the kingdom (Dupuis 1998). In his “Histoire de l’église du Bénin” father Paul-Henry Dupuis concludes that “Dahomey remained the least welcoming, the most ‘closed’ of all the kingdoms in the Gulf of Guinea”11 (Dupuis 1998, 220). In the 19th and 20th century, under colonial rule and post-independence Marxist-Leninist dictatorship, Voodoo and other ATR in Benin were marginalized and socially stigmatized, much like in other SSA counties. For instance, in 1976, an ‘anti-witchcraft’ law was put in place, which was especially harmful to Voodoo priests who risked being persecuted. The intimidation led to a sharp decline of self-reported Voodoo adherents and a promotion of monotheistic religions, although ATR influences survived, leading to considerable religious syncretism (Barbier and Dorier-Apprill, 2002; Tall, 1995a). The advent of democracy, in the early 1990s, allowed ATR to resurface in Benin. The new democratic leadership promoted Voodoo as part of a (new) national identity, but also because of political opportunism, to win votes in the democratic elections through the support of traditional religious leaders (Mayrargue 1995; Tall 1995a). In 1992, Voodoo became enlisted in the constitution as an official religion and received its own public holiday, on January 10. Presently, Voodoo is widely cherished as a national heritage, preached openly by voodoo priests, and further cultivated in Voodoo convents.
11 Own translation from “le Dahomey restait le moins accueillant, le plus ‘fermé’ de tous les royaumes de ce Golfe de Guinée”.
16 Chapter 2
The exceptional renaissance of ATR in Benin allows us to safely assume that self-reported ATR adherence in Benin is a reasonably good proxy for ATR beliefs. It is however by no means a perfect proxy. The religious landscape in Benin is characterized by great syncretism, and Voodoo faces competition by African Independent Churches that blend Christian traditions with ATR-like rituals, miracles and charismatic healing (Barbier and Dorier-Apprill 2002; Olupona 2014; Tall 1995b). We will return to this syncretism in the empirical analysis.
2.2.3. Voodoo and healthcare in Benin
Similar to other ATR adherents, Voodoo adherents believe that the dead turn into spirits who interact with humanity, playing a significant role in human destiny and wellbeing (Bourgoignie 1972).12 Several studies indicate that Voodoo’s powerful traditions and mystic beliefs constitute obstacles to the demand for healthcare in Benin. Aujoulat et al. (2003), for instance, argue that cultural beliefs in southern Benin affect the demand for treatment of Buruli ulcer, a skin infection which can lead to disfigurement and disability. Magico-religious beliefs in Benin are also said to affect malaria prevention and treatment (Rashed et al. 1999), and have been blamed for a low demand for vaccines, because some traditional healers have advised against their use (ONE 2011a, 2011b).13 Magico-religious beliefs do not always go against conventional medical prescriptions. For instance, Jenkins and Curtis (2005) argue that such beliefs are a motivation for latrine adoption in Benin, among others because of “fear of enemies stealing your feces for sorcery against you”. According to a WHO (2002) report on traditional medicine, up to 80% of the Beninese population relies on traditional healers for their primary healthcare.14 This reliance is stronger among Voodoo adherents. For instance, LeMay-Boucher et al. (2013), studying magico-religious expenditures in Benin, find that self-declared Voodoo adherents report significantly higher such expenditures
12 Some of the dead can gain the high status of deified ancestors and become protectors of the clan. Higher up in the cosmologic hierarchy are several deities which are ambivalent beings (good/evil) that relate to different natural elements (e.g. gods of the earth, sky, water, iron, forests). What distinguishes Voodoo from mere ancestor worship or animism is the recognition of a Supreme Being (Mawu or God), who leads the cosmologic hierarchy but does not concern itself directly with man. The communication with Mawu passes through the worship of deities. Especially feared is Sakpata, the Voodoo of the earth, who inflicts disease on humans (Henry 2010; Soumonni 2012). 13 The campaigning and advocacy organization ONE collaborated with VBS television network on a documentary about the relationship between Voodoo and conventional medicine in Benin. The documentary points to the importance of traditional healers, who are among the most respected members of society and have considerable authority. Dr. Roch Hougnihin, director of the national program of traditional medicine in Benin, argues that “if village traditional healers are opposed to vaccination, parents will not allow their children to be vaccinated”(Gerson 2011; ONE 2011a, 2011b). 14 This may, in part, be due to supply side issues: it is estimated that Benin has only one doctor for each 10,000 individuals, but one traditional healer for each 800 individuals (Ministère de la santé, Benin 2008).
17 Chapter 2
compared to individuals from other religious groups. If these expenditures act as a substitute for conventional health care, we expect to find a negative relation between ATR adherence and the uptake of bed nets and vaccines.
2.3. Data description
We pool data from four DHS surveys, conducted in 1996, 2001, 2006 and 2012. All surveys are nationally representative, covering the six provinces of Benin.15 Geographic stratification was based on survey clusters, corresponding to villages in rural areas and city blocks in urban areas. Information collection in DHS mostly concerns children younger than 6 years and their mothers. Our sample is composed of 35,121 children, from 23,801 households and 1,777 survey clusters, for whom information is available on the mother’s socio-demographic characteristics. Information on father’s characteristics is only available for a subsample of 6,533 households. Mothers are on average 29 years old, while fathers are on average 37 years old. The level of parents’ education is generally low, averaging about 1.4 years of schooling for mothers and 3 years for fathers. Parents in our sample are mainly Fon (41%) or Adja (16%); others belong to one of six smaller ethnic groups. These and other summary statistics are presented in Table A.1 in the supplementary appendix. In what follows, we provide a brief description of religion and health in the DHS data, and how they relate to each other.
2.3.1. Religion
Table 1 reveals that 19% of mothers in our sample report to be ATR adherents (85% of which are Voodoo-adherents)16; other important religious affiliations among mothers include Catholicism (26%), Islam (24%), Protestantism (6%) and other Christian churches (18%). The other Christian churches include Anglo-American variants as well as African Independent Churches, most prominently the Celestial Church (Barbier and Dorier-Apprill 2002). Despite regional differences, all religions are practiced throughout the country. Within DHS survey clusters there is large religious heterogeneity: the average cluster counts just 24 mothers, but includes more than 3 different religious affiliations; and about half of the 1,777 survey clusters included in our analysis comprise both ATR and other mothers. Even within households we find substantial
15 The six provinces were transformed into 12 departments by an administrative reorganization in 1999. 16 The share of 85% is calculated from the 2006 and 2012 DHS survey rounds, which are the only rounds that distinguish between Voodoo and ‘other traditional religions’.
18 Chapter 2
religious heterogeneity: in 27% of the 6,533 households for which we have information on both parents, the parents do not have the same religious affiliation (see Table 1); and in almost half of the 1,825 couples in which at least one partner reports to be an ATR adherent, the other partner belongs to a different religion. Several descriptive statistics indicate that individual ATR-adherence is not just determined by individual choice, but also by history and tradition. First, the share of ATR-adherents is much higher among the descendants of the founders of the Dahomey kingdom: about half of Adja mothers (47%) and fathers (57%) are ATR-adherents versus 13% and 18% among mothers and fathers from other ethnicities (see Panel A of Table 2). Furthermore, parents currently living within the boundaries of the ancient Dahomey kingdom are significantly more likely to be ATR-adherents – whether they belong to the Adja ethnicity or not (see Panel B of Table 2).17
2.3.2. Health
A child is fully immunized if it received all eight vaccines which are required by the WHO Expanded Program on Immunization, protecting the child among others from polio, tetanus, and measles.18 It is recommended that children are fully immunized by the age of one. In our sample, we have information on 26,359 children aged 1-5 for which we have data on vaccination rates and the socio-demographic characteristics of their mother. Table 3 shows that the first-time vaccination rates are fairly high for the individual vaccines – between 60% and 90%. The full immunization rate is driven down by a failure of subsequent vaccinations. For instance, vaccination rates for the second and third vaccine of DPT and polio are about 10 and 20 percentage points lower compared to the first. Because of the incomplete uptake of the various vaccines, full immunization rates in Benin are dangerously low, at only 37% in the latest DHS round (2012).19 In order to turn the tide and sustain progress in immunization coverage rates, Benin relies on large-scale immunization campaigns. Distribution posts go from village to village and health workers go from door to door in an attempt to reach all children, especially those in remote areas. In 2009 about 2.7 million children under the age of five were to be vaccinated during two rounds
17 Historical maps from (Dupuis 1998) indicate that the boundaries of the ancient Dahomey kingdom roughly corresponded to the present-day departments Atlantique, Kouffo and Zou – located in the South-West of Benin (see Figure 1). 18 Routine immunization schedules in Benin stipulate that infants should be vaccinated with: (1) a dose of Bacillus Calmette- Guerin (BCG) vaccine at birth; (2) three doses of diphtheria, pertussis and tetanus (DPT) vaccine at 6, 10 and 14 weeks after birth; (3) at least three doses of oral polio vaccine (OPV) at birth and at 6, 10 and 14 weeks after birth; and (4) one dose of measles vaccine 9 months after birth. 19 When considering the 34 Sub Saharan African countries where the most recent DHS survey was organized between 2006-2015, full immunization rates range from 24% to 93% with an average of 60% (DHS 2017).
19 Chapter 2
of National Immunization Days (UNICEF 2009). The negative side-effect is that mothers do not spontaneously bring their children to health centers to get them vaccinated. They rather wait for health workers to pass by their house (INSAE 2013, 2007). This perverse effect may have contributed to the decline in immunization rates in the past decade (see Panel A of Table 3). To combat malaria, Benin organized large-scale campaigns in 2007 and 2011. Almost 6 million long lasting insecticide nets were distributed (INSAE 2013). This explains the rising trend in bed net ownership and use, shown in Panel B of Table 3. Household ownership of bed nets doubled from 44% in 2001 to 90% in 2012, when 3 in 4 bed nets were obtained from a distribution campaign. Over the same period, the share of households in which all children slept under a bed net the night before the interview increased from 35% to 74%; and between 2006 and 2012, the share of children who slept under a treated net increased from 30% to 89%. Malaria nevertheless remains a major issue in Benin. In 2012 it was reported to be the main cause of under-5 mortality and responsible for nearly half (42%) of hospitalizations of children under five (INSAE 2013). During the 2012 DHS round, a malaria blood test was administered on children of 6 to 59 months old. Among the 3,134 children in our database that were tested, 26% tested positive. Overall, substantial progress in health outcomes has been made over the past decades. The neonatal-, infant-, and under-five mortality rates have all decreased between 1996 and 2012. For instance, under-five (and infant) mortality rates more than halved, from 166 (99) deaths per 1,000 live births in 1996 to 67 (43) deaths in 2012 (see Panel C of Table 3).
2.3.3. Correlation between ATR and healthcare
In Table 4 we present the simple two-way relationship between mothers’ ATR adherence and the uptake of preventive healthcare measures. The results indicate that children whose mother is an ATR adherent are less likely to be fully immunized (37% compared to 44%) and more likely to have received none of the eight vaccinations required by the WHO (15% compared to 10%). Children of ATR mothers are also less likely to live in a household that owns a bed net (56% compared to 74%), and even when the household owns a bed net they are less likely to sleep under it (73% compared to 80%). The differences are all statistically significant at the 1%-level. Turning to health outcomes, the results in Table 4 indicate that children whose mother is an ATR adherent are more likely to test positive for malaria (40% compared to 23%) and have a higher risk of dying, as indicated by significantly higher neonatal-, infant- and under-five mortality rates. As these bivariate correlations may be contaminated by confounding factors, we now turn to a multivariate regression analysis.
20 Chapter 2
2.4. Accounting for observables
Informed by ethnographic studies, we hypothesize that ATR adherence reduces household uptake of conventional healthcare measures and worsens health outcomes. We test this hypothesis for Benin because ATR adherence in Benin is freely reported and varies considerably within villages and within households. However, we still face two major empirical challenges. First, even in Benin, self-reported ATR adherence is by no means a perfect proxy for traditional religious beliefs. There still is a fair amount of underreporting, especially in the northern part of the country where Islam dominates, and there is lots of religious syncretism, which goes unnoticed in the DHS as it only asks about the primary religious affiliation. Besides, ATR does not have a monopoly on beliefs and practices that go against conventional medicine. In particular, Charismatic Churches and African Independent Churches also put a lot of emphasis on spiritual healing. Combined, these features imply that, when estimating the relation between ATR adherence and conventional healthcare, our control group (no self-reported ATR adherence) is contaminated by both ATR- and non-ATR related beliefs that also affect the demand for healthcare. Our estimated relation will therefore be a lower bound. Second, ATR correlates significantly with several community-, household- and individual-level characteristics. As such, households in which the mother is an ATR adherent are more likely to live in rural areas (81% vs 62%) and tend to be less wealthy (see Table 5). ATR mothers are also more likely to live in a polygamous household (51% compared to 38%) and ATR adherents are less educated; a finding that holds both for mothers (0.4 years of schooling compared to 1.6 years) and fathers (1.6 years compared to 3 years). To account for these confounding factors, we include the entire set of survey cluster dummies as well as a large set of socio-economic characteristics. Formally, the empirical model can be written as follows:
0 0 0 0 0 !"#$% = '( + '* ,-.#$% + /"#$% Ω + /#$% Ε + 3$%∆ + 536% Γ + 8% Λ + :"#$% (1a)
0 0 0 0 !#$% = '(′ + '*′ ,-.#$% + /#$% Ω′ + 3$%∆′ + 536% Γ′ + 8% Λ′ + :#$% (1b)
, where < indexes children, = mothers, ℎ households and ? DHS survey clusters. In Equation (1a) children are the units of observation, and the outcome variables, denoted by !"#$%, are 'not having received any vaccines', 'full immunization', 'the use of bed nets', and 'malaria incidence'. In Equation
21 Chapter 2
(1b), mothers are the unit of observation and the outcome variables are 'bed net ownership' and 'child mortality', denoted by !#$% . The sample size in the regressions varies depending on the unit of observation (children or mothers) and the availability of data (e.g. information on bed nets was not collected in the 1996 DHS survey, and the malaria test was only administered in the 2012 DHS survey).
The main explanatory variable is mothers’ adherence to ATR, captured by ,-.#$% which is a dummy variable that takes the value one if the mother of the child reports to be an ATR adherent.20
/"#$%, /#$% and 3$% are vectors containing child-, mother- and household-level covariates which are likely to influence the uptake of conventional healthcare measures. At the level of the child, we include: gender, age (in months), and a birth-order and -interval variable.21 We control for the mother’s age, age at first birth, years of schooling, and her ethnicity. At the household level we include wealth quintiles22, a dummy indicating if the household is polygamous, and the number of children under five.
The model also controls for the year in which the DHS survey took place (536%) and the survey cluster in which the household lives (8%); '( is a constant, εCD% and εED% are error terms. Following Angrist and Pischke (2009), we estimate the equations using a Linear Probability Model (LPM). As a robustness check, we compare the LPM estimates to marginal effects estimated by a Logit model. Standard errors are clustered at the household level to account for within-household correlation of the residuals.23 In Table 6 we look at the determinants of not having received any vaccination among children aged 1-5.24 We present five models, going from parsimonious to more inclusive specifications. In the first column, we only control for the survey year and the geographical department. We find a large and
20 The DHS data only record reported affiliations, not the intensity of religious participation (e.g. frequency of ceremony attendance, prayer, etc.). 21 Studies have shown that birth order and the time between births matter for the vaccination of children. We merged the DHS variables “birth order” and “preceding birth interval” into one variable. In doing so, we follow the categorization of Antai (2010): (1) first births; (2) birth order 2-4 with short birth interval (<24 months); (3) birth order 2-4 with medium birth interval (24-47 months); (4) birth order 2-4 with long birth interval (48+ months); (5) birth order 5+ with short birth interval (<24 months); (6) birth order 5+ with medium birth interval (24-47 months); and (7) birth order 5+ with long birth interval (48+ months) 22 We calculated a wealth index as the first principal component of a large number of household assets including: source of water, type of toilet facility, type of floor/wall/roof-material, and the ownership of radio, television, telephone, refrigerator, car. From the index we calculated wealth quintiles that range from 1 to 5 with a mean value of 2.8 and a standard deviation of 1.36. We control for the wealth quintiles as it allows for an easier interpretation, but all results are robust to controlling for the wealth index in levels. 23 Following Cameron and Miller (2015), we progressively clustered the standard errors at broader levels. In our baseline estimates, reported in the paper, we chose to cluster the standard errors at the household-level because equations (1a) and (1b) include several household-level controls. All results are however highly robust to clustering the standard errors at the level of the survey clusters (see infra and Table A.7 in the supplementary appendix). 24 As explained in section 3.2, it is recommended that children are fully immunized by age one. When analyzing the determinants of vaccination, we therefore consider children aged 1-5 rather than children aged 0-5.
22 Chapter 2
significantly positive coefficient estimate on ATR adherence indicating that children whose mother is an ATR adherent are 8 percentage points more likely not to have received a single recommended vaccination. When adding cluster fixed-effects in the second column, the ATR effect is reduced considerably suggesting that local- or supply-side factors are important confounding factors. Adding controls for individual- and household-level characteristics further reduces the estimated ATR effect (see columns 3-5). We find a similar change in estimated coefficients across parsimonious and more inclusive specifications when looking at the determinants of full immunization, and the ownership and use of bed nets (Columns 2-4 in Table 7 report the inclusive model specifications while the full set of model specifications can be consulted in Tables A.2-A.4 in the supplementary appendix). Even when controlling for the full set of socio-economic characteristics and when comparing children within the same survey cluster, we find that children with an ATR mother are 3 percentage points more likely not to have received any vaccine (13% vs 10%), 3 percentage points less likely to be fully immunized (40.5% vs 43.3%), 6 percentage points less likely to live in a household that owns a bed net (65% vs 71%) and 6 percentage points less likely to sleep under a bed net (61% vs 67%) (see Columns 1-4 in Table 7). Overall, the estimated ATR effects are larger than the effect of an additional six years of schooling for the mother or a change from the first to the second household wealth quintile. The signs on the control variables are generally in line with our expectations.25 In columns 5-6 of Table 7 we turn to the determinants of health outcomes (we present the inclusive specifications, and report the more parsimonious ones in the supplementary appendix, in Tables A.5 and A.6). Column 5 presents the estimated determinants of malaria incidence, obtained from blood tests administered in the 2012 DHS survey, indicating that children whose mother is an ATR adherent are 6 percentage points more likely to test positive for malaria (32.5% vs 26.9%) – even when controlling for bed net ownership. The results in Column 6 indicate that the under-five mortality rate for ATR mothers is higher with approximately 9 deaths per 1,000 live births compared to non- ATR mothers (90.8 vs 81.7). In sum, when adding socio-economic characteristics and especially cluster fixed-effects to the model, the ATR effect gradually diminishes. However, even in the inclusive specifications, ATR
25 For instance, living in a richer household and having a more educated mother significantly improves children’s uptake of preventive healthcare measures.
23 Chapter 2
adherence of the mother remains significantly associated with a lower uptake of preventive healthcare measures and worse child health outcomes.26
2.5. Accounting for unobservables
The above estimated ATR-health relationship may be spurious. First, while the survey cluster dummies control for the supply of health care at the level of the village or city block, they do not control for effective household access to various health facilities, e.g. related to a household’s financial wealth. In our estimating equations wealth is proxied by an asset index, which is far from perfect. Second, mothers could self-select into ATR for various reasons. One reason relates to cognitive style. For instance, an intuitive cognitive style, as opposed to a more analytical one, predicts religious beliefs, paranormal beliefs, anthropomorphism and attitudes towards alternative medicine (Pennycook et al. 2012; Svedholm et al. 2010). In other words, some mothers may have a ‘taste for the supernatural’, thus self-selecting both in ATR and other-than-biomedical medical forms of healing. Another driver of self-selection may relate to the health history of the mother or of one of her children. Chronic or mental illness, or an unexpected death, generally trigger more suspicion of supernatural causes than other medical conditions or misfortunes. A typical example is epilepsy, which is strongly associated with spirit possession (Carrazana et al. 1999; Khoury 2012). Children suffering from epileptic seizures, or other chronic health conditions, may be confined to Voodoo convents. The innate health condition of the child may thus be a possible confounder, affecting both our outcome variable and explanatory variable of interest.
26 These results are robust to clustering the standard errors at the level of the survey cluster (see Table A.7 in the supplementary appendix) and to the inclusion of DHS sample weights (results not reported but available upon request). The same holds for all other results presented in the paper. In Table A.8 we further compare the results estimated with LPM and Logit models. In order to make the samples comparable, we restrict the LPM estimates to survey clusters that have variation in the dependent variable (Beck 2015). To calculate marginal effects after the Logit estimations we use the procedure suggested by Beck (2015): We first estimate the regression coefficients with a conditional logit model (clogit); then we run a fixed effects logit model, constraining the coefficients to those estimated by clogit, from which we calculate the marginal effects. The thus calculated marginal effects are highly comparable to the constrained LPM estimates. In addition, in Table A.9 in the supplementary appendix, we estimate equations 1a and 1b for each DHS survey year separately. Although the results do not allow us to say much about the evolution of the ATR-effect over time (due to a lack of data availability and relatively small sample sizes in some survey years), the ATR-health relationship seems to be persistent over time and is relevant in the latest DHS survey round of 2012. Finally, we explore heterogeneity in the ATR-effect by interacting a mother’s ATR-adherence with (1) her level of education; (2) a dummy indicating if a mother is literate; and (3) the household wealth quintiles. None of these specifications yield a significant interaction term however (results not reported but available upon request).
24 Chapter 2
We adopt three strategies to determine whether the ATR-health relation is causal, rather than driven by various unobservables.
2.5.1. Subsample analysis: bed net owners and father’s characteristics
To better rule out the influence of financial wealth, we look at two subsamples. First, when estimating the determinants of bed net use, we confine the sample to bed net owning households. Second, we consider the subsample of 6,533 households for which we have information on both parents. This allows us to control for a father’s ATR-adherence. If mother-specific factors rather than household- level factors drive our ATR estimate, we should find that ATR adherence of the mother has a larger effect on our outcome variables than ATR adherence of the father, because women in Benin have the primary responsibility when it comes to children’s healthcare decisions (de Sousa et al. 2011; ONE 2011b; Rashed et al. 1999). The results in column 1 of Table 8 indicate that even when comparing households with similar socio-economic characteristics, who live in the same survey cluster and who own a bed net, children of ATR mothers are 3 percentage points less likely to actually sleep under the bed net (82.5% vs 85.5%). In columns 2-7 of Table 8, we control for the father’s characteristics (his ATR adherence, age, years of schooling and ethnicity) and still find a large and significant negative relation between mother’s ATR adherence and the uptake of all preventive healthcare measures. 27 In contrast, for father’s ATR adherence we only find a slightly significant negative relation with the ownership and use of bed nets. With respect to health outcomes, we find that only mother’s ATR adherence is associated with higher child mortality and a higher incidence of malaria among children, although this relation is no longer significant.28 These results indicate that the ATR-health correlation is specific to the mother of the child, reducing concerns that household-level characteristics associated with ATR-adherence are driving the results.
27 The estimated ATR-effects are even slightly higher compared to those estimated in Table 7: children whose mother is an ATR-adherent are 12 percentage points more likely not to be vaccinated, 8 percentage points less likely to be fully immunized, 11 percentage points less likely to own a bed net and 11 percentage points less likely to sleep under a bed net. 28 When interpreting the results on health outcomes, one should take into account that the sample of children tested for malaria is relatively small – especially within this subsample of observations with information on both parents – and that the death of a child is a relatively rare event.
25 Chapter 2
2.5.2. Using selection on observables to assess the bias from unobservables
To further assess the influence of unobservables, we turn to the approach proposed by Altonji et al. (2005) and fine-tuned by Oster (2015). The approach uses the selection on observable variables as a guide to assess the potential bias from unobserved variables. Put very simply: if adding a battery of relevant observables does not affect our coefficient of interest much, then it is unlikely that there exist many unobservables that would completely cancel out our result. The selection on observable variables can be evaluated by looking at coefficient movements in the ATR-estimate while gradually adding additional control variables; their relevance is assessed by the associated movements in the R-squared. Based on these insights, Oster (2015) develops a measure that indicates how large selection on unobservable variables has to be, relative to selection on observables, to fully explain away the estimated effect. In Appendix A.10, we detail how we implement the procedure in our setting. We find that selection on unobservables would need to be about twice as large as selection on observables to fully explain away the estimated ATR-effects reported in Table 7. Moreover, when additionally including father’s characteristics, the results suggest that selection on unobservables would need to be 12.0, 9.4 and 9.0 times as large as selection on observables to fully explain away the estimated ATR-effects (reported in Table 8) on ‘not having received a single vaccination’, ‘full immunization’ and ‘under 5 mortality’. Overall, these findings suggest it is unlikely that the ATR effect is entirely driven by omitted variable bias. On the other hand, the value of the test depends entirely on the unverifiable assumption that the unobservables influence selection in roughly the same way as the observables. Because this assumption may not be plausible, we turn to a third strategy to deal with unobservables.
2.5.3. Instrumental variable approach
This strategy exploits the fact that present-day ATR-adherence is shaped by history and tradition. We instrument a mother's ATR-adherence with a dummy that takes the value one for Adja mothers who currently live within the boundaries of the ancient Dahomey kingdom. The instrument is a relevant predictor of a mother's ATR-adherence: the Adja in our sample are the descendants of the initial founders of the Dahomey kingdom, where Voodoo originated and evangelization was strongly resisted (see Section 2.2.2). These historical relationships still influence present-day ATR-adherence: Adja
26 Chapter 2
mothers are significantly more likely to be ATR-adherents, and this is even more apparent for Adja mothers currently living within the boundaries of the ancient Dahomey kingdom (see Section 2.3.1).29 For the exclusion restriction to be satisfied, the instrument should only impact the outcome variables through a mother's ATR-adherence. The exclusion restriction is violated if there exist certain (cultural) traits specific to the Adja ethnicity, that are associated with the outcome variables but are not captured by a mother's ATR-adherence. As Voodoo originated from and further shaped the Adja culture, we argue that – conditional on the included control variables – the exclusion restriction is defendable in this respect. A second concern is that, due to e.g. geographical reasons, the outcome variables might be directly affected by living within the boundaries of the Dahomey kingdom. We implement two strategies to address this concern. First, we additionally control for a set of geography- related covariates (i.e. a dummy indicator for rural areas, a dummy indicator for the geographical South of Benin, the survey cluster latitude and longitude, and the survey cluster distance to the closest city).30 Second, we run two sets of IV regressions: one within the full sample, and a second in which we restrict the sample to the Southern departments of Benin – which are geographically closely related to the Dahomey kingdom.31 The first-stage results, presented in column 1 of Table 9, indicate a strong and significant relationship between our instrument and a mother's ATR-adherence: both in the full sample and when restricting the sample to the Southern departments we find that Adja mothers who currently live within the boundaries of the Dahomey kingdom are about 35 percentage points more likely to be ATR- adherents. Moreover, the first-stage F-tests are large and significant, giving further confidence that the instrument is sufficiently relevant. The second-stage results are presented in columns 2-8 of Table 9. It is important to keep in mind that they capture the local average treatment effect, i.e. they are driven by the variation in a mother's ATR-adherence caused by the instrument; and additionally deal with attenuation bias (not unimportant given the high religious syncretism). With respect to preventive
29 53% of the 3,888 Adja mothers in our sample currently live within the boundaries of the ancient Dahomey kingdom. 30 In the IV regressions we no longer add survey cluster fixed effects, as this would absorb much of the variation in our instrument. To assess the strength of the geography-related covariates we run two sets of LPM regressions: one in which we control for the baseline covariates (all those reported in Table 7) and the geography covariates, but without adding cluster fixed effects – and another in which we additionally add cluster fixed effects. We do not find a statistically significant difference between both sets of ATR-effects (except for the malaria test results), indicating that the geography covariates provide a reasonable proxy for local unobserved covariates (results not reported but available upon request). 31 The South of Benin counts seven departments: Atlantique, Kouffo, Littoral, Mono, Ouèmè, Plateau and Zou (see Figure 1). Although they cover a relatively small surface area, about 60% of the sample lives in these Southern departments. The South of Benin differs from the North, both in terms of geography (e.g. the South is characterized by its large lakes and lagoons) and culture (e.g. in the Northern departments about 50% of our sample reports to adhere to Islam, while this is only the case for 6% in the South).
27 Chapter 2
healthcare measures, the estimated ATR-effects are in line with those reported in Tables 7 and 8: we find that children whose mother is an ATR-adherent are 14 percentage points more likely not to have received any of the eight recommended vaccines; 15 percentage points less likely to live in a household which owns a bed net; and 11 percentage points less likely to sleep under a bed net (even if the household owns one). With respect to health outcomes: the ATR-effect is no longer significant for the under-5 mortality rate, while it is large and highly significant for the malaria test (indicating that children whose mother is an ATR-adherent are almost 50 percentage points more likely to test positive for malaria). The first- and second-stage IV results remain qualitatively unchanged when running the estimation on the full sample or when restricting the sample to the South of Benin, suggesting that local geography is not a strong mediating channel between the instrument and our outcome variables.
2.6. Testing for channels of causality: world view versus traditional healers
Overall, the results of the three strategies discussed above indicate that the ATR-health relationship is not entirely spurious. In this section, we explore two main channels through which ATR may influence heuristic decision making about healthcare measures: the ATR worldview and the reliance on traditional religious leaders.
2.6.1. The ATR worldview
ATR adherence may promote a worldview that is not conducive to the uptake of conventional healthcare treatments. As explained above, this worldview is akin to African cosmology. If the worldview is the mediating channel, we would therefore expect that other beliefs that strongly relate to African cosmology are also associated with lower uptake of conventional healthcare measures. To test for this possibility, we look more closely at African Independent Churches and witchcraft beliefs. The DHS records adherence to African Independent Churches in the response category ‘other Christian religions’. Regarding witchcraft beliefs, the DHS data includes one single proxy, i.e. whether a mother believes that HIV/AIDS can be transmitted by witchcraft. This question was only asked in the 2006 and 2012 DHS surveys. Table 10 shows that on average 48% of the mothers in our sample believe in HIV/AIDS-transmission through witchcraft. Among ATR mothers the belief is slightly higher, at 52% but it is highest (56%) among mothers from ‘other Christian religions’. Especially
28 Chapter 2
Pentecostal and Celestial churches are known for their explicit recognition of and fight against witchcraft (Casanova 2001; Geschiere 2013). In Table 11 we replace mother’s ATR adherence with a dummy indicating mother’s adherence to ‘other Christian religions’ as the variable of interest. The results indicate that, in terms of preventive healthcare and health outcomes, children of these mothers are not significantly different from other children (although they are slightly more likely to live in a household that owns a bed net). In Table 12, we add the variable FGH<ℎ 2.6.2. Traditional healers Since the DHS only collects limited information on the use of traditional healthcare, we can only tentatively explore the influence of traditional healers on the demand for healthcare, relying on information on the type of person who assisted with the birth delivery, consultations and treatments for children who had a fever or diarrhea in the two weeks prior to the survey, and visits to biomedical health facilities in the past 12 months. The results in Table 13 indicate that, all else equal, ATR mothers are significantly more likely to have a traditional birth attendant assist during the delivery of their child (2 percentage points), and are significantly more likely to consult a traditional healer to treat the diarrhea or fever of their child (2 32 This finding may relate to self-selection: when a child dies young, a witch may be blamed for its death. The belief in witchcraft may especially be high for mothers who have lost more than one child. On the other hand, it could relate to a belief in ‘witch babies’ whose presence may be detected at birth. Sargent (1988, 1981), doing research among the Bariba in Benin, notes that unusual features of either the infant or the birth process (e.g. birth occurring at eight months or babies born with teeth) are signs indicating the possible presence of a witch baby. Babies displaying such signs at birth or during teething were customarily killed or abandoned (Sargent 1988). 29 Chapter 2 percentage points). At the same time, ATR mothers are 5 percentage points less likely to have visited a biomedical health facility in the 12 months prior to the survey.33 Finally, in columns 4-5 of Table 13 we look at the determinants of using ORS to treat diarrhea and using conventional medication (which includes aspirin, ibuprofen, paracetamol and several anti-malarial medications) to treat fever, for the subsample of children who suffered from these conditions in the two weeks prior to the survey. 8GLGH_ℎNJONI#$% is a dummy indicating if a mother visited a traditional healer in order to treat the diarrhea or fever of her child(ren). We find that mothers who visited a traditional healer are 8 percentage points less likely to have used ORS to treat the diarrhea of their children, and 18 percentage points less likely to have used conventional medication to treat their children’s fever. In sum, ATR mothers are more likely to visit traditional healers and less likely to visit health centers. Mothers who visit a traditional healer are less likely to use ORS and other conventional medication to treat their children’s illness. Given that we carefully accounted for access issues (cf. Table A.14 in the supplementary Appendix), these results suggest that the demand for conventional healthcare is lower for ATR mothers. Why their demand is lower, is a question that we cannot answer with the data at hand. One possibility is that traditional healers offer services and products that function as substitutes for conventional medicine (and that ATR-mothers are more exposed to these healers because traditional religious leaders often also function as healers). Alternatively, it could be that the ATR-affiliation makes mothers more susceptible to the influence and authority of traditional healers, which in turn may affect their heuristics about the usefulness of conventional health care. We can neither exclude that there is something about the ATR worldview, distinct from African cosmology in general (as discussed in the previous section), that leads both to an increased use of traditional healers and a reduced reliance on conventional medicine. 33 Besides the inclusion of our battery of controls, we use additional information from the DHS survey to rule out that this lower use reflects restricted access rather than lower demand of Voodoo mothers for health services offered by biomedical health facilities. The DHS records information on the potential hurdles to visiting a biomedical health facility. These include: not knowing where to go; getting permission; finding money to pay for the treatment; the distance to the health center; having to take transport; not wanting to go alone and the fear that there may not be a female health worker. For each of these potential hurdles, mothers indicate whether they present ‘no problem’, ‘a small problem’ or ‘a big problem’. In Table A.14 (in the supplementary appendix) we present the marginal effects of mothers’ ATR adherence calculated after Ordered Probit regressions on the determinants of each hurdle. Controlling for the full set of covariates and including cluster-level fixed effects, we find that ATR mothers consistently report having less problems with each of these potential obstacles. 30 Chapter 2 2.7. Discussion One of the stylized facts in the recent literature on health behavior in developing countries is that households tend to underinvest in preventive healthcare measures (Banerjee and Duflo 2009; Dupas 2011). The lower-than-optimal uptake is explained among others by liquidity constraints, time- inconsistent preferences and a lack of verifiable information on the cost-effectiveness of the measures (Dupas 2011). Lacking such information, caretakers turn to heuristic decision-making, looking for rules of thumb, opinions of others, behavior of neighbors or their own understanding of sickness and health. Based on a large ethnographic literature, we conjectured that ATR may be an important input for such heuristic decision-making, either because of the authority of religious leaders, or because of its distinct worldview. Our case study, Benin, is typical for a SSA country as ATR-related beliefs and practices are widespread. On the other hand, Benin is atypical as people freely report ATR adherence. The revealed ATR belief and its substantial within-village and within-household variation allowed us to estimate its impact on the uptake of several DHS healthcare measures and outcomes, while controlling for a large set of confounding factors. We find that, ceteris paribus, children whose mother is an ATR adherent are 3 percentage points more likely not to have received any vaccination, 3 percentage points less likely to be fully immunized, 6 percentage points less likely to live in a household which owns a bed net and 6 percentage points less likely to sleep under a bed net. Overall, these estimated ATR effects are larger than the effect of an additional six years of schooling for the mother or a change from the first to the second household wealth quintile. Even when comparing households with similar socio-economic characteristics, who live in the same survey cluster and who own a bed net, we find that children of an ATR mother are 3 percentage points less likely to sleep under the bed net. Mothers’ ATR adherence is further associated with a 6 percentage point increase in their children’s likelihood of testing positive for malaria and an under-five mortality rate which is higher with 9 deaths per 1,000 live births. We also find that the ATR effect is driven by the mother, who is the primary child care taker in Benin. The fact that father’s ATR adherence is not important for the uptake of preventive healthcare measures or child health outcomes further increases confidence that our results are not driven by unobserved household factors that are related to ATR and influence the demand for preventive healthcare or child health outcomes. Nevertheless, the ATR-health relationship may be spurious, as mothers could self-select into ATR. To formally investigate to what extent the results are driven by selection on unobservable 31 Chapter 2 variables, we turn to the procedures developed by Altonji et al. (2005) and fine-tuned by Oster (2015). In addition, we turn to an instrumental variables approach in which we take advantage of the fact that present-day ATR-adherence is not merely an individual choice, but is shaped by history and tradition. The combined results suggest that self-selection does not entirely drive our results. Even if the ATR- health relationship is partly spurious, our results are important from a policy perspective as they establish a highly robust correlation. This suggests that the uptake of preventive healthcare, and ultimately child health outcomes, may be improved by targeting ATR mothers. We test for two main channels through which ATR may influence heuristic decision-making about healthcare measures: the ATR worldview and the reliance on traditional religious leaders. The analysis suggests that the mediating channel underlying our results relates to ATR in specific, and not to magico-religious beliefs in general. We find that ATR-mothers rely more on the services of traditional healers, but – due to issues of self-selection – this result does not allow us to draw strong conclusions on the role of traditional healers e.g. in providing substitutes for conventional medicine, or actively exercising influence in another way. Millennium Development Goal 4 called for a two-thirds reduction in child mortality between 1990 and 2015. This goal has not been attained, but progress has been made: under-five mortality rate has dropped from 90 to 43 deaths per 1000 live births, globally, and from 179 to 86 deaths in SSA (UN 2015). To achieve further declines, finding ways to improve the uptake of (preventive) healthcare is of critical importance. Our results suggest that properly acknowledging the role of ATR beliefs can help in bridging the last mile. Although the data at hand did not allow us to fully uncover the role of traditional healers, it is possible that they provide ATR-mothers with off-the-shelve answers on what (not) to do in terms of healthcare. This does not mean that the caretakers’ minds are immune to new information. It does mean that it will take more than just information to persuade them. Acknowledging this means directing efforts at building trust in conventional healthcare providers and the health system, and working closely with traditional healers to persuade people. Building partnerships between public health providers and traditional healers is easier said than done. A pilot-program doing exactly that was initiated in the South-West of Benin in 2009. Run by the Université Libre de Bruxelles and financed by the European Union, ‘Interface entre prestataires de soins modernes et traditionnels’, created a platform where modern and traditional health providers could interact and exchange information; traditional healers also received medical training allowing them to quickly recognize severe cases of illness that needed referral to health centers. The project’s evaluation report 32 Chapter 2 mentions that, as a result, referrals from traditional healers to health centers increased (Aissan et al. 2013). But, it also mentions that the referral system reduced the perceived contributions made by the traditional healer, thereby demotivating traditional healers to continue their collaboration with health centers. This account resonates the one made by French colonial administrators in 1906, who explained the opposition of Voodoo priests against smallpox vaccinations as resulting from a conflict of interest: “their benefits are reduced when they have few patients to treat, smallpox being their assured commission money” (as cited in Soumonni 2012). In combination with our results, these accounts suggest that any collaboration with traditional healers should be cleverly designed, duly considering the incentives on the part of traditional healers. 33 Chapter 2 References Aikins, M.I.C., H. Pickering, and B. M. Greenwood. 1994. “Attitudes to Malaria, Traditional Practices and Bednets (Mosquito Nets) as Vector Control Measures: A Comparative.” Journal of Tropical Medicine and Hygiene 97: 8l–86. Aissan, J., Y. Bokossa, A. Dresse, and F. Zinsou. 2013. “Les Acquis Du Projet de Recherche-action ‘Interface Entre Prestataires de Sois Officiels et Traditionnels’ dans La Zone Sanitaire de Klouékanmè-Toviklin-Lalo.” Cotonou, Benin: European Union and Université Libre de Bruxelles. Akoto, Kwame A., and Akua N. Akoto. 2005. “African Cosmology.” In Encyclopedia of Black Studies, edited by Molefi Asante and Ama Mazama. Thousand Oaks California: SAGE Publications, Inc. Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber. 2005. “Selection on Observed and Unobserved Variables: Assessing the Effectiveness of Catholic Schools.” Journal of Political Economy 113 (1): 151–84. Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An Empiricist’s Companion. 1 edition. Princeton: Princeton University Press. Antai, Diddy. 2009. “Faith and Child Survival: The Role of Religion in Childhood Immunization in Nigeria.” Journal of Biosocial Science 41 (1): 57–76. ———. 2010. “Migration and Child Immunization in Nigeria: Individual-and Community-Level Contexts.” BMC Public Health 10 (1): 116. Antai, Diddy, Gebrenegus Ghilagaber, Sara Wedrén, Gloria Macassa, and Tahereh Moradi. 2009. “Inequities in Under-Five Mortality in Nigeria: Differentials by Religious Affiliation of the Mother.” Journal of Religion and Health 48 (3): 290–304. Aujoulat, Isabelle, Christian Johnson, Claude Zinsou, Augustin Guédénon, and Françoise Portaels. 2003. “Psychosocial Aspects of Health Seeking Behaviours of Patients with Buruli Ulcer in Southern Benin.” Tropical Medicine & International Health: 8 (8): 750–59. Awolalu, J.O. 1976. “What Is African Traditional Religion?” Studies in Comparative Religion 10 (2). Awusabo-Asare, K., and J. K. Anarfi. 1997. “Health-Seeking Behaviour of Persons with HIV/AIDS in Ghana.” Health Transition Review: The Cultural, Social, and Behavioural Determinants of Health 7 Suppl: 243–56. Banerjee, A. V., E. Duflo, R. Glennerster, and D. Kothari. 2010. “Improving Immunisation Coverage in Rural India: Clustered Randomised Controlled Evaluation of Immunisation Campaigns with and without Incentives.” BMJ 340 (may17 1). Banerjee, Abhijit V., and Esther Duflo. 2009. “The Experimental Approach to Development Economics.” Annual Review of Economics 1 (1): 151–78. Barbier, Jean-Claude, and Élisabeth Dorier-Apprill. 2002. “Cohabitations et Concurrences Religieuses Dans Le Golfe de Guinée, Le Sud-Bénin, Entre Vodun, Islam et Christianismes.” Bulletin de L’association Des Géographes Français, 223–236. BBC. 2014. “Ebola outbreak: ‘Witchcraft’ hampering Treatment, Says Doctor.” BBC News: Health. August. http://www.bbc.com/news/health-28625305. Beck, Nathaniel. 2015. “Estimating Grouped Data Models with a Binary Dependent Variable and Fixed Effects: What Are the Issues?” University of Rochester. Bhalotra, Sonia, Christine Valente, and Arthur van Soest. 2010. “The Puzzle of Muslim Advantage in Child Survival in India.” Journal of Health Economics 29 (2): 191–204. Bloom, David E., David Canning, and Mark Weston. 2005. “The Value of Vaccination.” World Economics 6 (3): 15. 34 Chapter 2 Bokenkotter, Thomas S. 1985. Essential Catholicism. 1st edition. Garden City, N.Y: Doubleday. Bourgoignie, Georges E. 1972. Les Hommes de L’eau: Ethno-Écologie Du Dahomey Lacustre. Paris, France: Editions universitaires. Cameron, Colin A., and Douglas L. Miller. 2015. “A Practitioner’s Guide to Cluster-Robust Inference.” The Journal of Human Resources 50 (2): 317–72. Carrazana, E., J. DeToledo, W. Tatum, R. Rivas-Vasquez, G. Rey, and S. Wheeler. 1999. “Epilepsy and Religious Experiences: Voodoo Possession.” Epilepsia 40 (2): 239–41. Casanova, José. 2001. “Religion, the New Millennium, and Globalization.” Sociology of Religion 62 (4): 415–41. Cau, Boaventura M., Arusyak Sevoyan, and Victor Agadjanian. 2013. “Religious Affiliation and under-Five Mortality in Mozambique.” Journal of Biosocial Science 45 (3): 415–29. Cohen, Jessica, and Pascaline Dupas. 2010. “Free Distribution or Cost Sharing? Evidence from a Randomized Malaria Prevention Experiment.” The Quarterly Journal of Economics CXXV (1). Comoro, C., S. E. D. Nsimba, M. Warsame, and G. Tomson. 2003. “Local Understanding, Perceptions and Reported Practices of Mothers/Guardians and Health Workers on Childhood Malaria in a Tanzanian District--Implications for Malaria Control.” Acta Tropica 87 (3): 305–13. Das, Jishnu, and Saumya Das. 2003. “Trust, Learning, and Vaccination: A Case Study of a North Indian Village.” Social Science & Medicine 57 (1): 97–112. DHS. 2017. “STATcompiler. The Demographic and Health Survey Program.” Dolan, Paul, Tessa Peasgood, and Mathew White. 2008. “Do We Really Know What Makes Us Happy? A Review of the Economic Literature on the Factors Associated with Subjective Well-Being.” Journal of Economic Psychology 29 (1): 94–122. Dupas, Pascaline. 2009. “What Matters (and What Does Not) in Households’ Decision to Invest in Malaria Prevention?” American Economic Review 99 (2): 224–30. ———. 2011. “Health Behavior in Developing Countries.” Annual Review of Economics 3 (1): 425–449. Dupuis, Père Paul-Henry. 1998. “Histoire de L’église Du Bénin”, Tome 1, Le Temps Des Semeurs (1494- 1901). Cotonou: Notre Dame. Ewijk, Reyn van. 2011. “Long-Term Health Effects on the next Generation of Ramadan Fasting during Pregnancy.” Journal of Health Economics 30 (6): 1246–60. Gerson, M. 2011. “Going from Voodoo to Vaccines.” CNN International. June. http://www.cnn.com/2011/OPINION/06/21/gerson.vaccines.africa/index.html. Geschiere, Peter. 2013. Witchcraft, Intimacy, and Trust: Africa in Comparison. Chicago: University Of Chicago Press. Glele, Maurice Ahanhanzo. 1974. Le Danxome. Du Pouvoir Aja À La Nation Fon. Paris: Nubia. González, Felipe, and Edward Miguel. 2015. “War and Local Collective Action in Sierra Leone: A Comment on the Use of Coefficient Stability Approaches.” Journal of Public Economics 128 (August): 30–33. Grabenstein, John D. 2013. “What the World’s Religions Teach, Applied to Vaccines and Immune Globulins.” Vaccine 31 (16): 2011–23. Gyimah, Stephen O. 2007. “What Has Faith Got to Do with It? Religion and Child Survival in Ghana.” Journal of Biosocial Science 39 (6): 923–37. Gyimah, Stephen O., Baffour K. Takyi, and Isaac Addai. 2006. “Challenges to the Reproductive- Health Needs of African Women: On Religion and Maternal Health Utilization in Ghana.” Social Science & Medicine 62 (12): 2930–44. Henry, Christine. 2010. “La terre de Sakpata.” Journal des africanistes, no. 80–1/2 (June): 253–65. Herskovits, Melville J. 1938. Dahomey: An Ancient African Kingdom. 2 vols. New York: J. J. Augustin. 35 Chapter 2 Heymann, David L., and R. Bruce Aylward. 2004. “Eradicating Polio.” New England Journal of Medicine 351 (13): 1275–77. Hoffmann, Vivian, Christopher B. Barrett, and David R. Just. 2009. “Do Free Goods Stick to Poor Households? Experimental Evidence on Insecticide Treated Bednets.” World Development 37 (3): 607–17. Iannaccone, Laurence. 1998. “Introduction to the Economics of Religion.” Journal of Economic Literature 36 (3): 1465–95. IFRC. 2014. “Ebola, Snakes and Witchcraft: Stopping the Deadly Disease in Its Tracks in West Africa.” International Federation of Red Cross and Red Crescent Societies. June. INSAE. 2007. “Enquête Démographique et de Santé, Benin 2006.” Cotonou, Benin: Institut National de la Statistique et de l’Analyse Economique. ———. 2013. “Enquête Démographiqe et de Santé 2011-2012.” Cotonou, Benin: Institut National de la Statistique et de l’Analyse Economique. Jalan, Jyotsna, and E. Somanathan. 2008. “The Importance of Being Informed: Experimental Evidence on Demand for Environmental Quality.” Journal of Development Economics 87 (1): 14– 28. Janssen, H.T. 2010. “Stewardship in West African Vodun: A Case Study of Ouidah, Benin.” Missoula, MT: University of Montana. Jenkins, Marion W., and Val Curtis. 2005. “Achieving the ‘good Life’: Why Some People Want Latrines in Rural Benin.” Social Science & Medicine 61 (11): 2446–59. Joshua, Stephen Muoki. 2010. “A Critical Historical Analysis of the South African Catholic Church’s HIV/AIDS Response between 2000 and 2005.” African Journal of AIDS Research 9 (4): 437– 47. Kahneman, Daniel. 2011. Thinking, Fast and Slow. 1St Edition edition. Farrar, Straus and Giroux. Kale, Rajendra. 1995. “South Africa’s Health: Traditional Healers in South Africa: A Parallel Health Care System.” BMJ 310 (6988): 1182–85. Kalichman, S.C., and L. Simbayi. 2004. “Traditional Beliefs about the Cause of AIDS and AIDS- Related Stigma in South Africa.” AIDS Care 16 (5): 572–80. Kanu, Ikechukwu Anthony. 2013. “The Dimensions of African Cosmology.” Kenny, Charles. 2009. “There’s More to Life than Money: Exploring the Levels/Growth Paradox in Income and Health.” Journal of International Development 21 (1): 24–41. Khoury, Nayla M. 2012. “Explanatory Models and Mental Health Treatment: Is Vodou an Obstacle to Psychiatric Treatment in Rural Haiti?” Culture, Medicine and Psychiatry 36 (3): 514–34. Kremer, Michael, and Edward Miguel. 2007. “The Illusion of Sustainability.” The Quarterly Journal of Economics 122 (3): 1007–65. Law, Robin. 2004. Ouidah: The Social History of a West African Slaving “Port”, 1727-1892. Ohio University Press. Laxminarayan, Ramanan, Jeffrey Chow, and Sonbol A. Shahid-Salles. 2006. “Intervention Cost- Effectiveness: Overview of Main Messages.” In Disease Control Priorities in Developing Countries. Washington, DC: World Bank and Oxford University Press. Le Hérissé, Auguste. 1911. L’ancien Royaume Du Dahomey, Moeurs, Religion, Histoire. Paris: E. Larose. LeMay-Boucher, Philippe, Joël Noret, and Vincent Somville. 2013. “Facing Misfortune: Expenditures on Magico-Religious Powers for Cure and Protection in Benin.” Journal of African Economies 22 (2): 300–322. Leonard, Kenneth L., and Joshua Graff Zivin. 2005. “Outcome versus Service Based Payments in Health Care: Lessons from African Traditional Healers.” Health Economics 14 (6): 575–93. 36 Chapter 2 Madajewicz, Malgosia, Alexander Pfaff, Alexander van Geen, Joseph Graziano, Iftikhar Hussein, Hasina Momotaj, Roksana Sylvi, and Habibul Ahsan. 2007. “Can Information Alone Change Behavior? Response to Arsenic Contamination of Groundwater in Bangladesh.” Journal of Development Economics 84 (2): 731–54. Manning, Patrick. 1982. Slavery, Colonialism and Economic Growth in Dahomey, 1640-1960. Cambridge University Press. Maslove, David M, Anisa Mnyusiwalla, Edward J Mills, Jessie McGowan, Amir Attaran, and Kumanan Wilson. 2009. “Barriers to the Effective Treatment and Prevention of Malaria in Africa: A Systematic Review of Qualitative Studies.” BMC International Health and Human Rights 9 (1): 26. Mayrargue, C. 1995. “Le Religieux et Les Législatives de Mars 1995 Au Bénin.” Politique Africaine, no. 58 (June): 157–62. McCullough, Michael E., and Brian L. B. Willoughby. 2009. “Religion, Self-Regulation, and Self- Control: Associations, Explanations, and Implications.” Psychological Bulletin 135 (1): 69–93. McGirk, Tim. 2015. “Taliban Assassins Target Pakistan’s Polio Vaccinators.” National Geographic News. http://news.nationalgeographic.com/2015/03/150303-polio-pakistan-islamic-state- refugees-vaccination-health/. McKenzie, David. 2012. “Beyond Baseline and Follow-up: The Case for More T in Experiments.” Journal of Development Economics 99 (2): 210–21. Mellor, Jennifer M., and Beth A. Freeborn. 2011. “Religious Participation and Risky Health Behaviors among Adolescents.” Health Economics 20 (10): 1226–40. Miller, Mark A., and John T. Sentz. 2006. “Vaccine-Preventable Diseases.” In Disease and Mortality in Sub-Saharan Africa, edited by Dean T Jamison, Richard G Feachem, Malegapuru W Makgoba, Eduard R Bos, Florence K Baingana, Karen J Hofman, and Khama O Rogo, 2nd ed. Washington (DC): World Bank. Ministère de la santé, Benin. 2008. “Politique Nationale de La Pharmacopée et de La Médicine Traditionnnelles Du Bénin.” Cotonou, Benin: Ministère de la santé - Pharmacopée Béninoise. Muela, Susanna Hausmann, Joan Muela Ribera, and Marcel Tanner. 1998. “Fake Malaria and Hidden Parasites—the Ambiguity of Malaria.” Anthropology & Medicine 5 (1): 43–61. Neill, Stephen. 1991. A History of Christian Missions. Edited by Owen Chadwick. 2 edition. London; New York: Penguin Books. Nyang, Sulayman S. 1982. “An African Cosmology.” The Unesco Courrier 35 (2): 27–33. Olupona, Jacob K. 2014. African Religions: A Very Short Introduction. OUP USA. Omonzejele, Peter F. 2003. “Current Ethical and Other Problems in the Practice of African Traditional Medicine.” Medicine and Law 22 (1): 29–38. ———. 2008. “African Concepts of Health, Disease, and Treatment: An Ethical Inquiry.” EXPLORE: The Journal of Science and Healing 4 (2): 120–26. ONE. 2011a. “ONE’s ‘Voodoo and Vaccines’ Film Makes CNN Headlines.” ONE. June. http://www.one.org/us/2011/06/21/ones-voodoo-and-vaccines-film-makes-cnn- headlines/. ———. 2011b. “Watch Our New Documentary ‘Voodoo and Vaccines.’” ONE. June. http://www.one.org/international/blog/watch-the-premiere-of-our-new-documentary- voodoo-and-vaccines-7pm-on-facebook/. Oster, Emily. 2015. “Unobservable Selection and Coefficient Stability: Theory and Evidence.” Working Paper. Brown University and National Bureau of Economic Research. 37 Chapter 2 Padela, Aasim I. 2013. “Islamic Verdicts in Health Policy Discourse: Porcine-Based Vaccines as a Case Study.” Journal of Religion and Science 48 (3): 655–70. Parrinder, E. G. 1954. African Traditional Religion. London: SPCK. Pennycook, Gordon, James Allan Cheyne, Paul Seli, Derek J. Koehler, and Jonathan A. Fugelsang. 2012. “Analytic Cognitive Style Predicts Religious and Paranormal Belief.” Cognition 123 (3): 335–46. Pew Research Center. 2010. “Tolerance and Tension: Islam and Christianity in Sub-Saharan Africa.” Washington, D.C. ———. 2014a. “Global Religious Diversity: Half of the Most Religiously Diverse Countries Are in Asia-Pacific Region.” ———. 2014b. “Religious Hostilities Reach Six-Year High.” Rashed, S, H Johnson, P Dongier, R Moreau, C Lee, R Crépeau, J Lambert, V Jefremovas, and C Schaffer. 1999. “Determinants of the Permethrin Impregnated Bednets (PIB) in the Republic of Benin: The Role of Women in the Acquisition and Utilization of PIBs.” Social Science & Medicine 49 (8): 993–1005. Ruijs, Wilhelmina LM, Jeannine LA Hautvast, Koos van der Velden, Sjoerd de Vos, Hans Knippenberg, and Marlies EJL Hulscher. 2011. “Religious Subgroups Influencing Vaccination Coverage in the Dutch Bible Belt: An Ecological Study.” BMC Public Health 11 (1): 102. Sargent, Carolyn F. 1981. “Choosing Among Parallel Health Care Systems in a Bariba Community.” Canadian Journal of African Studies / Revue Canadienne Des Études Africaines 15 (2): 303–9. ———. 1988. “Born to Die: Witchcraft and Infanticide in Bariba Culture.” Ethnology 27 (1): 79–95. Soumonni, Elisée. 2012. “Disease, Religion and Medicine: Smallpox in Nineteenth-Century Benin.” História, Ciências, Saúde-Manguinhos 19 (December): 35–45. Sousa, Alexandra de, Leon P. Rabarijaona, Jean L. Ndiaye, Doudou Sow, Mouhamed Ndyiae, Jacques Hassan, Nilda Lambo, et al. 2011. “Acceptability of Coupling Intermittent Preventive Treatment in Infants with the Expanded Programme on Immunization in Three Francophone Countries in Africa: Acceptability of Coupling Intermittent Preventive Treatment in Infants.” Tropical Medicine & International Health, November. Streefland, Pieter H. 2001. “Public Doubts about Vaccination Safety and Resistance against Vaccination.” Health Policy 55 (3): 159–72. Svedholm, Annika M., Marjaana Lindeman, and Jari Lipsanen. 2010. “Believing in the Purpose of Events-Why Does It Occur, and Is It Supernatural?” Applied Cognitive Psychology 24 (2): 252– 65. Tall, Emmanuel Kadya. 1995a. “De La Démocratie et Des Cultes Voduns Au Bénin.” Cahiers D’études Africaines 35 (137): 195–208. ———. 1995b. “Dynamique Des Cultes Voduns et Du Christianisme Céleste Au Sud-Bénin.” Cah. Sci. Hum 31 (4): 797–823. Telegraph. 2014. “Ebola Outbreak: Fight against Disease Hampered by Belief in Witchcraft, Warns British Doctor.” The Telegraph. July. http://www.telegraph.co.uk/news/worldnews/africaandindianocean/sierraleone/11001610/ Ebola-outbreak-fight-against-disease-hampered-by-belief-in-witchcraft-warns-British- doctor.html. Thomas, Felicity. 2007. “‘Our Families Are Killing Us’: HIV/AIDS, Witchcraft and Social Tensions in the Caprivi Region, Namibia.” Anthropology & Medicine 14 (3): 279–91. UN. 2015. “The Millenium Development Goals Report 2015.” New York: United Nations. 38 Chapter 2 UN IGME. 2014. “Levels & Trends in Child Mortality - Report 2014.” UN Inter-Agency Group for Child Mortality Estimation. UNICEF. 2009. “Benin Completes First Round of National Immunization Days against Polio.” UNICEF. Van Dyk, Alta. 2001. “Traditional African Beliefs and Customs: Implications for AIDS Education and Prevention in Africa.” South African Journal of Psychology 31 (2): 60–66. Webster, Jayne, Jo Lines, Jane Bruce, Joanna Rm Armstrong Schellenberg, and Kara Hanson. 2005. “Which Delivery Systems Reach the Poor? A Review of Equity of Coverage of Ever-Treated Nets, Never-Treated Nets, and Immunisation to Reduce Child Mortality in Africa.” The Lancet Infectious Diseases 5 (11): 709–17. WHO. 2002. “World Health Organization Traditional Medicine Strategy: 2002-2005.” Geneva: World Health Organization. ———. 2008. “Vaccination Greatly Reduces Disease, Disability, Death and Inequity Worldwide.” Bulletin of the World Health Organization. ———. 2012. “Global Routine Vaccination Coverage, 2011.” 44. Weekly Epidemiological Record. World Health Organization. World Bank. 2015. World Development Report 2015: Mind, Society, and Behavior. Washington, D.C.: The World Bank. 39 Chapter 2 Figures Figure 1: Southern Benin, and the approximate boundaries of the ancient Dahomey kingdom Notes: The map indicates the seven departments of Southern Benin (Atlantique, Kouffo, Littoral, Mono, Ouémé, Plateau and Zou). The ancient Dahomey kingdom was located in the South-West of Benin. Historical maps from Dupuis (1998) indicate that the boundaries of the ancient Dahomey kingdom roughly corresponded to the present-day departments Atlantique, Kouffo and Zou (indicated in orange). 40 Chapter 2 Tables Table 1: Religious heterogeneity in our sample mothers’ religious share of couples that affiliation are discordant Catholic 25.7% 38.6% Islam 23.9% 18.2% Traditional religion 18.7% 44.7% Protestant 5.5% 62.8% Other Christian religions 17.9% 45.3% Other religions 1.5% 87.8% No religion 6.9% 86.0% Observations 23,801 6,533 Notes: The shares in the second column are based on the 6,533 households for which we have information on both parents. They are calculated using PQ.ST UU VP WDVCD SPXY SPZ [\QZP] D\^ QZXV_VSP ` the following formula: with PQ.ST UU VP WDVCD \] XZ\^] SPZ [\QZP] D\^ QZXV_VSP ` X taking on the different religions as reported in the DHS survey. 41 Chapter 2 Table 2: ATR-adherence by historical relationships Panel A: Adja ethnicity Adja obs. No Yes ATR-adherence among mothers 13% 47% *** 23,801 ATR-adherence among fathers 18% 57% *** 6,533 Panel B: ancient Dahomey kingdom Dahomey obs. No Yes ATR-adherence among Adja mothers 35% 58% *** 3,888 ATR-adherence among other mothers 10% 22% *** 19,792 ATR-adherence among Adja fathers 47% 65% *** 1,047 ATR-adherence among other fathers 13% 31% *** 5,421 Notes: *** p<0.01, ** p<0.05, * p<0.1; In Panel A we compare the ATR- adherence of Adja mothers and fathers with parents who belong to another ethnic group. In Panel B, we compare the ATR-adherence of mothers and fathers living within the boundaries of the ancient Dahomey kingdom with parents living outside those boundaries. 42 Chapter 2 Table 3: Healthcare in Benin Panel A: Immunization rates, by DHS survey year none full bcg dpt1 dpt2 dpt3 polio1 polio2 polio3 measles 1996 15.81% 45.55% 83.00% 79.55% 73.47% 64.13% 78.85% 71.94% 60.67% 54.59% 2001 8.18% 51.98% 88.68% 85.50% 78.32% 69.10% 88.41% 80.94% 65.79% 64.35% 2006 8.55% 43.54% 86.40% 82.46% 75.94% 66.67% 86.74% 79.34% 61.79% 61.43% 2012 12.65% 36.60% 85.95% 74.56% 71.05% 63.71% 82.54% 77.13% 52.32% 67.21% Total 10.57% 42.24% 86.29% 79.70% 74.25% 65.71% 84.82% 78.19% 58.70% 63.50% Notes: Column 1 represents the share of children who are at least one year old and have received none of the eight vaccinations required by WHO. The full immunization rates in column 2 represent the share of children who are at least one year old and have received all eight required vaccines: bcg, dpt1-3, polio 1-3 and measles. Panel B: Bed nets B.1 Ownership and use of bed nets, by DHS survey year own bed net use bed net (some) use bed net (all) 1996 n.a. n.a. n.a. 2001 44.0% 38.3% 35.3% 2006 63.4% 53.6% 47.0% 2012 89.8% 81.8% 74.2% B.2 Characteristics of the bed nets used the night before the interview share of children who share of nets obtained from nr. of people who slept under a treated net a distribution campaign slept under the net 2006 30.15% n.a. 2.98 2012 89.20% 74% 2.85 Notes: In the 1996 DHS round, information on the use of bed nets was not collected. The first column of Panel B.1 indicates the share of households which owns a bed net. The second (third) column indicates the share of households in which at least some (all) kids slept under a bed net the night before the interview. Panel C: Mortality rates neonatal infant under five 1996 39 99 166 2001 39 91 151 2006 33 69 120 2012 24 43 67 Notes: The mortality rates are calculated following the procedures detailed in the 2012 Guide to DHS statistics. They are calculated for the five years preceding the survey-year and represent the number of deaths per 1,000 live births. The rates are calculated using the subsample of children for which complete information on the mother’s demographic information is available. The neonatal-, infant- and under five- mortality rates represent, respectively, the probability of dying within the first month of life, before the first birthday, between birth and the fifth birthday. 43 Chapter 2 Table 4: Preventive healthcare and child health outcomes by mothers' ATR adherence Mother is an ATR adherent? Yes No Full immunization rate 36.7% 43.5% *** Share of children that did not receive any of the eight required vaccinations 14.7% 9.6% *** Share of households that own a bed net 56.0% 74.4% *** Share of bed net-owning households in which all children slept under a bed 73.4% 79.8% *** net the night preceding the interview Share of children that tested positive for malaria 39.5% 23.3% *** Neonatal mortality rate 32.2 30.8 *** Infant mortality rate 69.3 64.6 *** Under-five mortality rate 116.3 108.0 *** Notes: *** p<0.01, ** p<0.05, * p<0.1; The mortality rates are calculated for the five years preceding the survey-year and represent the number of deaths per 1,000 live births. 44 Chapter 2 Table 5: Relevant characteristics, by parents’ ATR adherence Mother is an ATR adherent? Yes No Household wealth (wealth quintile) 2.33 3.01 *** Share of households living in a rural area 81% 62% *** Share of mothers living in a polygamous household 51% 38% *** Number of children under five 1.79 1.71 *** Mother's age at first birth 19.11 19.58 *** Mother / father is an ATR adherent? Yes No Schooling Years of schooling mother 0.43 1.57 *** Years of schooling father 1.55 3.00 *** Age Age mother 29.90 29.20 *** Age father 39.37 36.63 *** Notes: *** p<0.01, ** p<0.05, * p<0.1; A wealth index was calculated at the household-level as the first principal component of a large number of household assets including: source of water; type of toilet facility; material used to construct floor, wall, and roof; and the ownership of radio, television, telephone, refrigerator, car. From the index we calculated wealth quintiles which range from 1 to 5 with a mean value of 2.8 and a standard deviation of 1.36. 45 Chapter 2 Table 6: Determinants of not being vaccinated, for children aged 1-5 (1) (2) (3) (4) (5) mother is an ATR adherent 0.078*** 0.036*** 0.032*** 0.033*** 0.032*** (0.007) (0.007) (0.007) (0.007) (0.007) Wealth quintile: 2 -0.024*** -0.019** -0.022*** (0.008) (0.008) (0.008) 3 -0.045*** -0.036*** -0.041*** (0.008) (0.008) (0.008) 4 -0.056*** -0.049*** -0.050*** (0.009) (0.009) (0.009) 5 -0.077*** -0.069*** -0.067*** (0.010) (0.010) (0.010) age of mother 0.000 0.000 (0.000) (0.001) mother's age at first birth 0.001** 0.001* (0.001) (0.001) years of schooling mother -0.003*** -0.003*** (0.001) (0.001) polygamous household 0.002 0.001 (0.005) (0.005) gender of child (girl=1) -0.004 (0.004) age of child (in months) 0.001*** (0.000) nr. of children < 5 in HH 0.004 (0.003) ethnicity of mother No No No Yes Yes birth order No No No No Yes DHS survey year Yes Yes Yes Yes Yes geographical department Yes No No No No cluster fixed effects No Yes Yes Yes Yes Observations 26,359 26,359 26,359 26,359 26,359 R2 0.03 0.26 0.26 0.27 0.27 Adjusted R2 0.03 0.21 0.21 0.21 0.22 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; All specifications are estimated with a Linear Probability Model and include all children aged 1-5; The dependent variable is a dummy which takes the value one if a child did not receive any of the eight vaccinations recommended by the WHO. 46 Chapter 2 Table 7: Determinants of preventive healthcare measures and health outcomes Preventive healthcare measures Health outcomes full own use malaria under 5 no vaccines immunization bed net bed net positive mortality (1) (2) (3) (4) (5) (6) mother is an ATR adherent 0.032*** -0.028*** -0.062*** -0.064*** 0.056* 9.122** (0.007) (0.010) (0.010) (0.012) (0.031) (4.182) Wealth quintile: 2 -0.022*** 0.016 0.054*** 0.047*** 0.043 5.755 (0.008) (0.010) (0.010) (0.011) (0.034) (4.181) 3 -0.041*** 0.052*** 0.114*** 0.098*** 0.037 3.453 (0.008) (0.011) (0.011) (0.013) (0.040) (4.528) 4 -0.050*** 0.072*** 0.180*** 0.143*** -0.057 -13.673*** (0.009) (0.013) (0.013) (0.015) (0.042) (4.966) 5 -0.067*** 0.110*** 0.277*** 0.233*** -0.057 -28.541*** (0.010) (0.019) (0.017) (0.019) (0.050) (6.694) age of mother 0.000 0.001 -0.002*** -0.002** -0.000 0.651*** (0.001) (0.001) (0.000) (0.001) (0.003) (0.223) mother's age at first birth 0.001* 0.000 0.002*** 0.001 -0.005 -2.275*** (0.001) (0.001) (0.001) (0.001) (0.004) (0.358) years of schooling mother -0.003*** 0.009*** 0.008*** 0.008*** -0.001 -0.978** (0.001) (0.001) (0.001) (0.001) (0.003) (0.483) polygamous household 0.001 -0.000 -0.008 -0.030*** 0.009 28.736*** (0.005) (0.007) (0.007) (0.008) (0.022) (2.828) gender of child (girl=1) -0.004 -0.001 0.006 -0.014 (0.004) (0.006) (0.005) (0.018) age of child (in months) 0.001*** 0.003*** -0.002*** 0.003*** (0.000) (0.000) (0.000) (0.001) nr. of children < 5 in HH 0.004 -0.010*** 0.021*** -0.010** 0.013 -38.148*** (0.003) (0.003) (0.004) (0.004) (0.014) (1.721) household owns a bed net -0.051 (0.046) child slept under bed net -0.015 night before interview (0.030) ethnicity of mother Yes Yes Yes Yes Yes Yes birth order Yes Yes No Yes Yes No DHS survey year Yes Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Yes Observations 26,359 26,359 20,327 25,038 2,757 22,821 R2 0.27 0.24 0.41 0.31 0.43 0.16 Adjusted R2 0.22 0.18 0.36 0.27 0.24 0.08 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; No Vaccines, full immunization, own bed net, use bed net and malaria positive are dummy variables which take the value one if, respectively: a child did not receive a single vaccination, a child is fully immunized, a household owns a bed net, a child slept under a bed net the night preceding the interview, a child tested positive for malaria; The ownership of bed nets and the under-five mortality rate are estimated at the level of the mother. The determinants of immunization rates are estimated for all children aged 1-5, while the determinants of bed net use and a positive malaria test are estimated for all children aged 0-5 for which data was available; Information on bed nets was not collected in DHS 1996, the malaria test was only administered in DHS 2012; All specifications are estimated using a Linear Probability Model; This Table summarizes the most inclusive model specifications for each dependent variable - a full set of model specifications is available in the online appendix (Tables A.2 – A.6). 47 Chapter 2 Table 8: Subsample analysis: bed net owners and father's characteristics Subsample: Bed net Households with information on father's characteristics owners Preventive healthcare measures Health outcomes use bed no full ownership use bed malaria under 5 net if own vaccination immunization bed net net positive mortality (1) (2) (3) (4) (5) (6) (7) mother is an ATR adherent -0.029*** 0.116*** -0.080** -0.113*** -0.110** 0.020 55.198*** (0.011) (0.032) (0.039) (0.040) (0.051) (0.085) (17.883) father is an ATR adherent / 0.018 0.001 -0.047* -0.069* -0.034 16.923 / (0.018) (0.031) (0.027) (0.035) (0.053) (11.555) both parents are ATR / 0.048*** -0.033 -0.069*** -0.066** 0.042 18.895* adherents / (0.017) (0.027) (0.027) (0.030) (0.048) (10.966) age of mother -0.001 -0.000 0.002 -0.005*** -0.003 -0.001 1.178* (0.001) (0.002) (0.002) (0.001) (0.002) (0.004) (0.675) age of father / 0.001 -0.001 -0.000 -0.000 0.001 -0.967* / (0.001) (0.001) (0.001) (0.001) (0.002) (0.494) mother's age at first birth 0.001 0.003 -0.000 0.003 0.001 -0.006 -2.485*** (0.001) (0.002) (0.003) (0.002) (0.003) (0.004) (0.866) years of schooling mother 0.002** -0.001 0.006* 0.005* 0.002 -0.003 0.559 (0.001) (0.002) (0.004) (0.003) (0.003) (0.005) (1.409) years of schooling father / -0.002 0.006** 0.010*** 0.008*** -0.001 -1.877* / (0.002) (0.003) (0.002) (0.002) (0.004) (0.961) polygamous household -0.037*** -0.008 0.019 0.015 -0.002 0.023 43.981*** (0.007) (0.012) (0.017) (0.017) (0.019) (0.030) (7.020) gender of child (girl=1) 0.009* 0.003 -0.016 -0.003 -0.038* (0.005) (0.008) (0.013) (0.011) (0.023) age of child (in months) -0.002*** 0.001** 0.004*** -0.001*** 0.003*** (0.000) (0.000) (0.000) (0.000) (0.001) nr. of children < 5 in HH -0.027*** 0.006 -0.004 0.018** -0.035*** 0.012 -35.939*** (0.004) (0.006) (0.008) (0.009) (0.009) (0.018) (3.888) ethnicity of mother Yes Yes Yes Yes Yes Yes Yes ethnicity of father No Yes Yes Yes Yes Yes Yes birth order Yes Yes Yes No Yes Yes No wealth quintiles Yes Yes Yes Yes Yes Yes Yes DHS survey year Yes Yes Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Yes Yes Observations 19,436 6,826 6,826 5,138 6,266 2,032 5,809 R2 0.21 0.49 0.43 0.60 0.52 0.49 0.33 Adjusted R2 0.14 0.32 0.25 0.42 0.37 0.26 0.05 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; The model specifications are based on those presented in Table 7; In the first column we restrict the sample to bed net-owning households. In columns 2-7, we restrict the sample to households with information on both parents and additionally control for father’s characteristics; All specifications are estimated using a Linear Probability Model; Parent’s ATR adherence is a categorical variable indicating if only the mother, only the father or both parents are ATR adherents - ‘no ATR parents’ is the reference category. 48 Chapter 2 Table 9: IV – instrumenting mother's ATR-adherence using 2SLS 1st stage 2nd stage (1) (2) (3) (4) (5) (6) (7) (8) Mother's full no own use bed use bed malaria under 5 ATR- immuni- vaccines bed net net net if own positive mortality adherence zation Panel A: Full sample mother is an ATR-adherent 0.143** -0.153** -0.103 -0.111* -0.183*** 0.472*** 17.296 (0.056) (0.071) (0.083) (0.067) (0.056) (0.100) (20.312) Adja Dahomey 0.358*** (0.032) Observations 15,741 15,741 15,741 11,394 12,859 11,737 3,079 13,770 First stage F-test 152*** Panel B: sample restricted to the South of Benin mother is an ATR-adherent 0.147*** -0.139* -0.105 -0.117* -0.198*** 0.504*** 16.817 (0.057) (0.074) (0.090) (0.068) (0.056) (0.104) (20.591) Adja Dahomey 0.351*** (0.032) Observations 9,528 9,528 9,528 6,979 8,036 7,342 1,992 8,252 First stage F-test 123*** all baseline controls Yes Yes Yes Yes Yes Yes Yes Yes dummy indicating rural area Yes Yes Yes Yes Yes Yes Yes Yes dummy indicating south Yes Yes Yes Yes Yes Yes Yes Yes cluster longitude Yes Yes Yes Yes Yes Yes Yes Yes cluster latitude Yes Yes Yes Yes Yes Yes Yes Yes cluster distance to closest Yes Yes Yes Yes Yes Yes Yes Yes city Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the cluster-level and reported in parentheses; We instrument a mother's ATR-adherence with a dummy indicating if the mother belongs to the Adja ethnicity and currently lives within the boundaries of the ancient Dahomey kingdom; Columns 2-8 present second-stage results; The first-stage output presented in Column 1 is associated with the second-stage results presented in Column 2 – the first-stage results for the other outcome variables are comparable and omitted from the Table; All specifications are estimated using 2SLS. 49 Chapter 2 Table 10: Belief that AIDS can be caused by witchcraft Panel A: Share of mothers who believe HIV can be caused by witchcraft Not DHS survey Overall Voodoo Voodoo 2006 51.0 55.5 49.9 *** 2012 43.7 47.7 42.9 *** Overall 47.5 52.2 46.5 *** Panel B: Share of belief in AIDS through witchcraft, by religious affiliation Traditional religion 52.2 Catholic 46.2 Protestant 44.7 Other Christian religions 56.3 Islam 38.4 Other religions 50.4 No religion 44.1 Notes: *** p<0.01, ** p<0.05, * p<0.1; This question was only asked in the 2006 and 2012 DHS surveys 50 Chapter 2 Table 11: Including mother's adherence to “other Christian religions” as the variable of interest no full ownership malaria under 5 use bed net vaccination vaccination bed net positive mortality (1) (2) (3) (4) (5) (6) mother adheres to “another -0.005 -0.008 0.018** 0.008 0.013 4.355 Christian religion” (0.005) (0.009) (0.008) (0.010) (0.028) (3.737) all baseline controls Yes Yes Yes Yes Yes Yes DHS survey year Yes Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Yes Observations 26,359 26,359 20,327 25,038 2,757 22,821 R2 0.27 0.24 0.41 0.31 0.43 0.16 Adjusted R2 0.21 0.18 0.36 0.26 0.24 0.08 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; The explanatory variable of interest is a dummy variable which takes the value one if a mother reports to adhere to “other Christian religions” (which includes Evangelicalism and Pentecostalism as well as African Independent Churches such as the Celestial Church); All specifications are estimated using a Linear Probability Model; Table A.11 in the supplementary appendix shows the coefficients for the baseline controls. 51 Chapter 2 Table 12: Including mother’s belief that AIDS can be caused by witchcraft no vaccination full vaccination ownership bed net use bed net malaria positive under 5 mortality (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) mother is an ATR adherent 0.023** 0.024** -0.019 -0.019 -0.063*** -0.063*** -0.067*** -0.067*** 0.054 0.049 7.407 7.013 (0.009) (0.010) (0.014) (0.014) (0.013) (0.013) (0.015) (0.015) (0.038) (0.038) (5.739) (5.738) mother beliefs AIDS can -0.007 0.002 -0.010 -0.002 0.045 11.859*** be caused by witchcraft (0.006) (0.010) (0.008) (0.009) (0.031) (3.806) all baseline controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes DHS survey year Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 14,472 14,472 14,472 14,472 11,958 11,958 16,312 16,312 1,976 1,976 11,958 11,958 R2 0.27 0.26 0.24 0.24 0.37 0.37 0.32 0.32 0.49 0.49 0.18 0.18 Adjusted R2 0.19 0.18 0.16 0.16 0.28 0.28 0.26 0.26 0.27 0.27 0.07 0.07 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; Information on the relationship between witchcraft and AIDS was only asked in the 2006 and 2012 DHS survey rounds; We compare estimates that (do not) control for mother’s reported belief that AIDS can be caused by witchcraft; To allow for a meaningful comparison, we restrict the sample to the observations for which information on this belief is available; All specifications are estimated using a Linear Probability Model; Table A.12 in the supplementary appendix shows the coefficients for the baseline controls. 52 Chapter 2 Table 13: Healthcare services chosen by ATR mothers traditional use of visit health traditional birth healer to treat use of ORS to medication to facility attendant diarrhea or treat diarrhea treat fever fever (1) (2) (3) (4) (5) mother is an ATR adherent -0.047*** 0.020*** 0.017*** 0.000 -0.081*** (0.010) (0.005) (0.006) (0.031) (0.021) visited a traditional healer -0.082** -0.176*** (0.040) (0.046) all baseline controls Yes Yes Yes Yes Yes DHS survey year Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Observations 35,106 34,984 10,318 3,262 6,430 R2 0.29 0.28 0.26 0.53 0.42 Adjusted R2 0.24 0.23 0.10 0.24 0.26 Notes: *** p<0.01, ** p<0.05, * p<0.1; The dependent variables are dummy variables which take the value one if, respectively: (1) a mother visited a health facility in the 12 months prior to the interview; (2) a traditional birth attendant assisted with the delivery of the child; (3) a mother visited a traditional healer to treat the diarrhea or fever of her child in the two weeks prior to the interview; (4) a mother used ORS to treat her child’s diarrhea; (5) a mother used conventional medication (which includes aspirin, ibuprofen, paracetamol and several anti-malarial medications) to treat her child’s fever. In columns 3-5 the sample is restricted to mothers whose child had diarrhea or fever in the two weeks prior the interview; Robust standard errors are clustered at the household-level and reported in parentheses; Table A.13 in the supplementary appendix shows the coefficients for the baseline controls. 53 Chapter 2 Appendix Table A.1: Summary statistics Observations by DHS survey year obs. % 1996 2,833 8.07% 2001 5,067 14.43% 2006 14,645 41.70% 2012 12,576 35.81% Total 35,121 100% Household demographics mean st.dev. min. max. obs. age mother 29.3 6.7 15 49 23,801 age father 37.3 9.1 17 64 6,533 years of schooling mother 1.36 2.9 0 21 23,801 years of schooling father 3.0 4.4 0 21 6,533 mother’s age at first birth 19.5 3.8 8 43 23,801 age of child (in months) 28.1 17.1 0 59 36,797 number of children < 5 in the household 1.7 1 0 9 23,801 child is a girl 49.3% 0.5 0 1 36,797 household lives in rural area 65.7% 0.47 0 1 23,801 polygamous household 40.4% 0.49 0 1 23,801 Ethnicity of parents mother father Adja 16.4% 16.2% Bariba 9.7% 10.7% Betamaribe 7.6% 7.4% Dendi 3.4% 3.6% Fon 41.0% 41.2% Peulh 5.5% 5.9% Yoa & Lokpa 4.6% 4.9% Yoruba 10.1% 9.2% Other 1.7% 1.1% Notes: This Table includes observations for whom we have complete information on mother’s demographic characteristics. The full sample comprises 35,121 children aged 0-5, living in 23,801 households. Information on fathers’ characteristics is available for a sub- sample of 6,533 households. 54 Chapter 2 Table A.2: Determinants of full immunization among children aged 1-5 (1) (2) (3) (4) (5) mother is an ATR adherent -0.095*** -0.035*** -0.029*** -0.028*** -0.028*** (0.009) (0.010) (0.010) (0.010) (0.010) Wealth quintile: 2 0.020** 0.016 0.016 (0.010) (0.010) (0.010) 3 0.062*** 0.054*** 0.052*** (0.011) (0.011) (0.011) 4 0.086*** 0.074*** 0.072*** (0.013) (0.013) (0.013) 5 0.143*** 0.112*** 0.110*** (0.018) (0.019) (0.019) age of mother 0.002*** 0.001 (0.001) (0.001) mother's age at first birth -0.000 0.000 (0.001) (0.001) years of schooling mother 0.009*** 0.009*** (0.001) (0.001) polygamous household -0.006 -0.000 (0.007) (0.007) gender of child (girl=1) -0.001 (0.006) age of child (in months) 0.003*** (0.000) nr. of children < 5 in HH -0.010*** (0.003) ethnicity of mother No No No Yes Yes birth order No No No No Yes DHS survey year Yes Yes Yes Yes Yes geographical department Yes No No No No cluster fixed effects No Yes Yes Yes Yes Observations 26,359 26,359 26,359 26,359 26,359 R2 0.02 0.22 0.22 0.22 0.24 Adjusted R2 0.02 0.16 0.16 0.16 0.18 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; All specifications are estimated with a Linear Probability Model and include all children aged 1-5; The dependent variable is a dummy which takes the value one if a child is fully immunized (i.e. received all eight vaccinations recommended by the WHO). 55 Chapter 2 Table A.3: The ownership of bed nets within a household (1) (2) (3) (4) (5) mother is an ATR adherent -0.147*** -0.081*** -0.068*** -0.062*** -0.062*** (0.009) (0.010) (0.010) (0.010) (0.010) Wealth quintile: 2 0.059*** 0.055*** 0.054*** (0.010) (0.010) (0.010) 3 0.122*** 0.114*** 0.114*** (0.011) (0.011) (0.011) 4 0.195*** 0.182*** 0.180*** (0.013) (0.013) (0.013) 5 0.311*** 0.278*** 0.277*** (0.017) (0.017) (0.017) age of mother -0.002*** -0.002*** (0.000) (0.000) mother's age at first birth 0.002*** 0.002*** (0.001) (0.001) years of schooling mother 0.008*** 0.008*** (0.001) (0.001) polygamous household 0.002 -0.008 (0.006) (0.007) nr. of children < 5 in HH 0.021*** (0.004) ethnicity of mother No No No Yes Yes DHS survey year Yes Yes Yes Yes Yes geographical department Yes No No No No cluster fixed effects No Yes Yes Yes Yes Observations 20,327 20,327 20,327 20,327 20,327 R2 0.16 0.39 0.40 0.41 0.41 Adjusted R2 0.16 0.33 0.35 0.36 0.36 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; All specifications are estimated with a Linear Probability Model and include all the mothers in our sample; Information on bed nets was not collected in DHS 1996; The dependent variable is a dummy which takes the value one if the household owns a bed net. 56 Chapter 2 Table A.4: Determinants of sleeping under a bed net the night before the interview (1) (2) (3) (4) (5) mother is an ATR adherent -0.159*** -0.086*** -0.076*** -0.065*** -0.064*** (0.011) (0.012) (0.012) (0.012) (0.012) Wealth quintile: 2 0.047*** 0.047*** 0.047*** (0.012) (0.011) (0.011) 3 0.100*** 0.097*** 0.098*** (0.013) (0.013) (0.013) 4 0.148*** 0.141*** 0.143*** (0.015) (0.015) (0.015) 5 0.260*** 0.232*** 0.233*** (0.019) (0.019) (0.019) age of mother -0.004*** -0.002** (0.001) (0.001) mother's age at first birth 0.002** 0.001 (0.001) (0.001) years of schooling mother 0.008*** 0.008*** (0.001) (0.001) polygamous household -0.035*** -0.030*** (0.008) (0.008) gender of child (girl=1) 0.006 (0.005) age of child (in months) -0.002*** (0.000) nr. of children < 5 in HH -0.010** (0.004) ethnicity of mother No No No Yes Yes birth order No No No No Yes DHS survey year Yes Yes Yes Yes Yes geographical department Yes No No No No cluster fixed effects No Yes Yes Yes Yes Observations 25,038 25,038 25,038 25,038 25,038 R2 0.09 0.30 0.30 0.31 0.31 Adjusted R2 0.08 0.25 0.26 0.26 0.27 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; Information on bed nets was not collected in DHS 1996; All specifications are estimated with a Linear Probability Model and include all children aged 0-5; The dependent variable is a dummy which takes the value one if the child slept under a bed net the night preceding the interview. 57 Chapter 2 Table A.5: Determinants of a positive malaria test (1) (2) (3) (4) (5) (6) mother is an ATR adherent 0.188*** 0.076** 0.071** 0.065* 0.061* 0.056* (0.030) (0.031) (0.031) (0.035) (0.031) (0.031) Wealth quintile: 2 0.033 0.034 0.040 0.043 (0.034) (0.039) (0.034) (0.034) 3 0.033 0.033 0.037 0.037 (0.040) (0.045) (0.040) (0.040) 4 -0.056 -0.057 -0.058 -0.057 (0.042) (0.049) (0.042) (0.042) 5 -0.068 -0.065 -0.063 -0.057 (0.049) (0.058) (0.050) (0.050) age of mother 0.001 0.000 -0.000 (0.002) (0.003) (0.003) mother's age at first birth -0.005* -0.005 -0.005 (0.003) (0.004) (0.004) years of schooling mother -0.002 -0.001 -0.001 (0.004) (0.003) (0.003) polygamous household 0.015 0.010 0.009 (0.024) (0.022) (0.022) gender of child (girl=1) -0.015 -0.014 (0.018) (0.018) age of child (in months) 0.003*** 0.003*** (0.001) (0.001) nr. of children < 5 in HH 0.014 0.013 (0.014) (0.014) HH owns a bed net -0.051 (0.046) child slept under bed net -0.015 night before interview (0.030) ethnicity of mother No No No Yes Yes Yes birth order No No No No Yes Yes geographical department Yes No No No No No cluster fixed effects No Yes Yes Yes Yes Yes Observations 2,757 2,757 2,757 2,757 2,757 2,757 R2 0.05 0.41 0.42 0.42 0.43 0.43 Adjusted R2 0.04 0.23 0.24 0.23 0.24 0.24 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; Information on this test was only collected in DHS 2012; All specifications are estimated using a Linear Probability Model and include all children aged 0-5; The dependent variable is a dummy which takes the value one if the child tested positive for malaria. 58 Chapter 2 Table A.6: Determinants of the under-five mortality rate (1) (2) (3) (4) mother is an ATR adherent 20.639*** 12.680*** 9.224** 9.122** (3.617) (4.252) (4.265) (4.182) Wealth quintile: 2 4.105 5.755 (4.266) (4.181) 3 3.522 3.453 (4.626) (4.528) 4 -16.805*** -13.673*** (5.063) (4.966) 5 -30.336*** -28.541*** (6.813) (6.694) age of mother 1.112*** 0.651*** (0.222) (0.223) mother's age at first birth -2.325*** -2.275*** (0.365) (0.358) years of schooling mother -0.625 -0.978** (0.491) (0.483) polygamous household 28.736*** (2.828) nr. of children < 5 in HH -38.148*** (1.721) ethnicity of mother No No Yes Yes DHS survey year Yes Yes Yes Yes geographical department Yes No No No cluster fixed effects No Yes Yes Yes Observations 22,821 22,821 22,821 22,821 R2 0.02 0.12 0.13 0.16 Adjusted R2 0.02 0.04 0.05 0.08 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; All specifications are estimated with a Linear Probability Model and include all the mothers in our sample; The dependent variable is the under-five mortality rate calculated at the level of the mother. 59 Chapter 2 Table A.7: Determinants of preventive healthcare measures and health outcomes, clustering the standard errors at the level of the survey cluster Preventive healthcare measures Health outcomes full own use malaria under 5 no vaccines immunization bed net bed net positive mortality (1) (2) (3) (4) (5) (6) mother is an ATR adherent 0.032*** -0.028** -0.062*** -0.064*** 0.056 9.122** (0.008) (0.011) (0.010) (0.013) (0.035) (4.402) Wealth quintile: 2 -0.022** 0.016 0.054*** 0.047*** 0.043 5.755 (0.009) (0.011) (0.011) (0.013) (0.038) (4.442) 3 -0.041*** 0.052*** 0.114*** 0.098*** 0.037 3.453 (0.009) (0.012) (0.012) (0.015) (0.045) (4.754) 4 -0.050*** 0.072*** 0.180*** 0.143*** -0.057 -13.673** (0.010) (0.014) (0.015) (0.017) (0.049) (5.317) 5 -0.067*** 0.110*** 0.277*** 0.233*** -0.057 -28.541*** (0.011) (0.020) (0.021) (0.023) (0.057) (7.282) age of mother 0.000 0.001 -0.002*** -0.002** -0.000 0.651*** (0.001) (0.001) (0.000) (0.001) (0.003) (0.236) mother's age at first birth 0.001* 0.000 0.002*** 0.001 -0.005 -2.275*** (0.001) (0.001) (0.001) (0.001) (0.004) (0.385) years of schooling mother -0.003*** 0.009*** 0.008*** 0.008*** -0.001 -0.978* (0.001) (0.002) (0.001) (0.001) (0.004) (0.522) polygamous household 0.001 -0.000 -0.008 -0.030*** 0.009 28.736*** (0.005) (0.007) (0.007) (0.008) (0.025) (3.017) gender of child (girl=1) -0.004 -0.001 0.006 -0.014 (0.004) (0.006) (0.005) (0.019) age of child (in months) 0.001*** 0.003*** -0.002*** 0.003*** (0.000) (0.000) (0.000) (0.001) nr. of children < 5 in HH 0.004 -0.010** 0.021*** -0.010** 0.013 -38.148*** (0.003) (0.004) (0.004) (0.005) (0.017) (1.965) Household owns a bed net -0.051 (0.055) child slept under bed net -0.015 night before interview (0.034) ethnicity of mother Yes Yes Yes Yes Yes Yes birth order Yes Yes No Yes Yes No DHS survey year Yes Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Yes Observations 26,359 26,359 20,327 25,038 2,757 22,821 R2 0.27 0.24 0.41 0.31 0.43 0.16 Adjusted R2 0.22 0.18 0.36 0.27 0.24 0.08 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the level of the survey cluster and reported in parentheses; No Vaccines, full immunization, own bed net, use bed net and malaria positive are dummy variables which take the value one if, respectively: a child did not receive a single vaccination, a child is fully immunized, a household owns a bed net, a child slept under a bed net the night preceding the interview, a child tested positive for malaria; The ownership of bed nets and the under-five mortality rate are estimated at the level of the mother. The determinants of immunization rates are estimated for all children aged 1-5, while the determinants of bed net use and a positive malaria test are estimated for all children aged 0-5 for which data was available; Information on bed nets was not collected in DHS 1996, the malaria test was only administered in DHS 2012; All specifications are estimated using a Linear Probability Model. 60 Chapter 2 Table A.8: Comparing the estimates of LPM and Logit models no vaccines full immunization own bed net use bed net malaria positive LPM Logit LPM Logit LPM Logit LPM Logit LPM Logit (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) mother is an ATR adherent 0.055*** 0.087*** -0.028*** -0.027*** -0.078*** -0.078*** -0.065*** -0.067*** 0.095** 0.098*** (0.011) (0.001) (0.011) (0.000) (0.013) (0.000) (0.013) (0.001) (0.043) (0.003) age of mother -0.000 0.000*** 0.001 0.001*** -0.003*** -0.003*** -0.002** -0.002*** -0.001 -0.001*** (0.001) (0.000) (0.001) (0.000) (0.001) (0.000) (0.001) (0.000) (0.005) (0.000) mother's age at first birth 0.003* 0.003*** 0.001 0.001*** 0.003*** 0.004*** 0.002 0.002 -0.006 -0.007* (0.001) (0.000) (0.001) (0.000) (0.001) (0.000) (0.001) (0.001) (0.006) (0.003) years of schooling mother -0.008*** -0.016*** 0.009*** 0.008*** 0.012*** 0.017*** 0.009*** 0.012*** -0.003 -0.002*** (0.002) (0.000) (0.002) (0.000) (0.002) (0.000) (0.002) (0.000) (0.007) (0.000) polygamous household 0.001 0.006 -0.003 -0.001 -0.014* -0.013 -0.034*** -0.038*** 0.003 0.015 (0.008) (0.010) (0.008) (0.006) (0.009) (0.008) (0.009) (0.008) (0.035) (0.032) gender of child (girl=1) -0.008 -0.014*** 0.001 0.001*** 0.006 0.008*** -0.020 -0.019*** (0.006) (0.000) (0.006) (0.000) (0.006) (0.000) (0.030) (0.001) age of child (in months) 0.001*** 0.002*** 0.004*** 0.004*** -0.002*** -0.002*** 0.004*** 0.004*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.001) (0.000) nr. of children < 5 in HH 0.004 0.005*** -0.008** -0.008*** 0.028*** 0.031*** -0.007 -0.012*** 0.020 0.021 (0.004) (0.000) (0.004) (0.000) (0.005) (0.000) (0.005) (0.004) (0.021) (0.018) household owns a bed net -0.055 -0.088 (0.073) (0.068) child slept under bed net night -0.018 -0.013 before the interview (0.043) (0.040) wealth quintiles Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes ethnicity of mother Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes birth order Yes Yes Yes Yes No No Yes Yes Yes Yes DHS survey year Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 14,293 14,293 25,235 25,235 15,205 15,205 22,145 22,145 1,621 1,621 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; The LPM specifications are those reported in Table 7, now restricting the sample to survey clusters which have variation in the dependent variable; The under-five mortality rate is not an indicator variable and is hence not included in this robustness check; The coefficients reported in columns with the header ‘Logit’ represent marginal effects. They were calculated following the procedure suggested by Beck (2015): We first estimate the regression coefficients with a conditional logit model (clogit). Then we run a fixed effects logit model, constraining the coefficients to those estimated by clogit, from which we calculate the marginal effects. 61 Chapter 2 Table A.9: Estimated ATR-effects, by DHS survey year full own bed use malaria under 5 no vaccines immunization net bed net positive mortality (1) (2) (3) (4) (5) (6) 2012 ATR effect 0.017 -0.032* -0.046*** -0.047*** 0.069** 23.854*** (0.014) (0.018) (0.013) (0.018) (0.030) (6.923) obs. 9,108 9,108 7,443 10,030 2,757 7,443 2006 ATR effect 0.042*** -0.013 -0.071*** -0.063*** 1.724 (0.010) (0.016) (0.017) (0.017) n.a. (6.668) obs. 11,412 11,412 9,562 13,242 9,562 2001 ATR effect 0.021 -0.041 -0.065*** -0.082* 5.407 (0.015) (0.026) (0.023) (0.046) n.a. (12.378) obs. 3,815 3,815 3,322 1,766 3,327 1996 ATR effect 0.046** -0.043 9.346 (0.023) (0.033) n.a. n.a. n.a. (12.925) obs. 2,024 2,024 2,489 Notes: *** p<0.01, ** p<0.05, * p<0.1; The Table presents the ATR-effects estimated in four sets of regressions, one for each available DHS survey year; The regressions control for the same set of covariates as reported in Table 7; Robust standard errors are clustered at the household-level and reported in parentheses; Information on bed nets was not collected in DHS 1996, the malaria test was only administered in DHS 2012; All specifications are estimated using a Linear Probability Model. 62 Chapter 2 Table A.10: Using selection on observables to assess the bias from unobservable Baseline controls Baseline controls + father's characteristics d !"#$ d !"#$ Received not a single vaccination 5.25 0.35 12.01 0.64 Full immunization 2.08 0.31 9.44 0.56 Ownership of bed net 1.88 0.53 1.74 0.77 Use of bed net 2.06 0.42 1.03 0.68 Malaria positive 1.73 0.56 / / Under 5 mortality 2.40 0.21 8.99 0.52 Notes: d is a measure that indicates how large selection on unobservables needs to be, relative to selection on observables, to fully explain away the estimated ATR-effects reported in Table 7 (for d in column 2) and Table 8 (for d in column 4). R&'( is the R-squared from a hypothetical regression that controls for all observables and . unobservables. As suggested by Oster (2015), we set R&'( = 1.3 R . For the outcome "malaria positive" no values were reported in columns 4-5 because the estimated ATR-effect is not significant in the subsample with information on the characteristics of both parents. 63 Chapter 2 Explanatory notes on the ‘Oster method' To assess the influence of unobservables, we turn to the approach proposed by Altonji et al. (2005) and fine-tuned by Oster (2015). The approach uses the selection on observable variables as a guide to assess the potential bias from unobserved variables. Put very simply: if adding a battery of relevant observables does not affect our coefficient of interest much, then it is unlikely that there exist many unobservables that would completely cancel out our result. The selection on observable variables can be evaluated by looking at coefficient movements in the ATR-estimate while gradually adding additional control variables; their relevance is assessed by the associated movements in the R-squared. Based on these insights, Oster (2015) develops a measure that indicates how large selection on unobservable variables has to be, relative to selection on observables, to fully explain away the estimated effect.35 The larger the measure, denoted by d, the less likely the threat of omitted variable bias. To calculate d, we first run two regressions for each outcome variable: an uncontrolled and a controlled regression. In the uncontrolled regression, we only regress the outcome variable on a mother’s ATR adherence. In the controlled regression we control for the observed covariates discussed above. Denote the estimated coefficient on a mother's ATR adherence /0 in the uncontrolled regression and /1 in the controlled regression; !0 and !1 are the R-squared values associated with these regressions. Next, the procedure requires to make an assumption about !"#$, which is defined as the R-squared from a hypothetical regression that controls for all 1 36 observed and unobserved covariates. We follow Oster (2015) in setting !"#$ = 1.3 ! . d is then 4 4 7 / (! −! ) 37 calculated as follows: 3 = 7 4 4 . Oster (2015) argues that a value of d > 1 (i.e. / −/ (!9:;−! ) that selection on observables is at least as important as selection on unobservables) indicates a result that is robust to omitted variable bias. 35 The calculations can be performed with the Stata Code ‘psacalc’, provided by Oster (2015) and freely available through ssc. 36 Oster (2015) derives this value by analyzing coefficient movements for 65 randomized studies, published in five top economic journals between 2008-2013, that provide estimates with and without controls. With a value of !"#$ = 1.3 !1, 90% of the evaluated randomized results survived. 37 Consider the intuition behind this expression. We find /1 in the numerator, indicating that the larger /1, the larger the effect that needs to be explained away by selection on unobservables. In the denominator we find (/0 − /1): the smaller the difference between /0 and /1, the less the ATR-estimate is affected by selection on observables, and the larger selection on unobservables needs to be, relative to selection on observables, to fully explain away the estimated 1 0 1 ATR-effect. The strength of the observed covariates increases in (! − ! ) and decreases in (!"#$ − ! ): the larger the difference between !1 and !0, the more variation in the outcome variable is accounted for by observed covariates; 1 on the other hand, the smaller the difference between !"#$ and ! , the more of the ‘explainable’ variation is accounted for by the observed covariates. 64 Chapter 2 We consider two sets of controlled regressions; the first includes all covariates discussed in section 2.4, the second also includes fathers' characteristics (discussed in section 2.5.1). The results in the first and third column of Table A.10 report the values of d for both sets of controlled regressions. For every outcome variable we find that d > 1. When the controlled regressions include only the baseline controls (column 1), we find that for most outcome variables, selection on unobserved covariates has to be about twice as important as selection on the included covariates to fully explain away the estimated ATR-effects (reported in Table 7). When looking at the outcome ‘received not a single vaccination’, selection on unobserved covariates would need to be five times more important. The value of d strongly increases for several outcome variables when the controlled regressions also include father's characteristics (column 3). Specifically, we find that that selection on unobservables would need to be 12.0, 9.4 and 9.0 times as large as selection on observables to fully explain away the estimated ATR-effects (reported in Table 8) on ‘not having received a single vaccination’, ‘full immunization’ and ‘under 5 mortality’.38 38 When following the procedures outlined in Oster (2015), the assumed values for R&'( are rather low for the first set of controlled regressions (varying between 0.21 and 0.56). Both Oster (2015) and González and Miguel (2015) argue that R&'( is bounded below one when there is measurement error in the dependent variable. Work by McKenzie (2012: p.214) further suggests that measurement errors may be substantial in the context of low income country household datasets; he demonstrates that "for many economic outcomes, the autocorrelations are typically lower than 0.5, with many around 0.3" and that "autocorrelations are often in the 0.2-0.3 range for household income and consumption". In our case, the outcome variables with the lowest R&'( – ‘under-five mortality’ (0.21), ‘not being vaccinated’ (0.35) and ‘being fully immunized’ (0.31) – are indeed most likely to suffer from substantial measurement error. The under-five mortality rate is a composite measure which depends on the correct measurement of several variables (e.g. number of children born, number of children still alive, exact date of birth of all children, age in months at death). Moreover, measurement error may be aggravated due to the sensitive nature of the questions. Outcomes pertaining to vaccination are also composite measures, relying on the correct measurement of (not) having received the eight recommended vaccines. On the other hand, the outcome variables that are relatively easier to measure – ‘the ownership of bed nets’, and ‘testing positive for malaria’ – yield higher values for R&'( (0.53 and 0.56 respectively). Finally, all values of R&'( increase substantially (varying between 0.52 and 0.77) when the controlled regressions additionally include father's characteristics. 65 Chapter 2 Table A.11: Including mother's adherence to “other Christian religions” as the variable of interest no full ownership malaria under 5 use bed net vaccination vaccination bed net positive mortality (1) (2) (3) (4) (5) (6) mother adheres to “another -0.005 -0.008 0.018** 0.008 0.013 4.355 Christian religion” (0.005) (0.009) (0.008) (0.010) (0.028) (3.737) Wealth quintile: 2 -0.022*** 0.017 0.055*** 0.048*** 0.044 5.640 (0.008) (0.010) (0.010) (0.011) (0.034) (4.181) 3 -0.042*** 0.053*** 0.116*** 0.101*** 0.036 3.036 (0.008) (0.011) (0.011) (0.013) (0.040) (4.525) 4 -0.052*** 0.073*** 0.183*** 0.146*** -0.058 -14.256*** (0.009) (0.013) (0.013) (0.015) (0.043) (4.959) 5 -0.070*** 0.112*** 0.282*** 0.238*** -0.066 -29.343*** (0.010) (0.019) (0.017) (0.019) (0.050) (6.691) age of mother 0.000 0.001 -0.002*** -0.002** 0.000 0.680*** (0.001) (0.001) (0.000) (0.001) (0.003) (0.222) mother's age at first birth 0.001* 0.000 0.002*** 0.001 -0.005 -2.285*** (0.001) (0.001) (0.001) (0.001) (0.004) (0.358) years of schooling mother -0.003*** 0.009*** 0.009*** 0.008*** -0.001 -1.057** (0.001) (0.001) (0.001) (0.001) (0.003) (0.481) polygamous household 0.002 -0.001 -0.010 -0.032*** 0.013 28.959*** (0.005) (0.007) (0.007) (0.008) (0.022) (2.825) gender of child (girl=1) -0.004 -0.001 0.007 -0.014 (0.004) (0.006) (0.005) (0.018) age of child (in months) 0.001*** 0.003*** -0.002*** 0.003*** (0.000) (0.000) (0.000) (0.001) nr. of children < 5 in HH 0.004 -0.010*** 0.021*** -0.010** 0.014 -38.134*** (0.003) (0.003) (0.004) (0.004) (0.014) (1.720) ethnicity of mother Yes Yes Yes Yes Yes Yes birth order Yes Yes No Yes Yes No DHS survey year Yes Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Yes Observations 26,359 26,359 20,327 25,038 2,757 22,821 R2 0.27 0.24 0.41 0.31 0.43 0.16 Adjusted R2 0.21 0.18 0.36 0.26 0.24 0.08 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; The explanatory variable of interest is a dummy variable which takes the value one if a mother reports to adhere to “other Christian religions” (which includes Evangelicalism and Pentecostalism as well as African Independent Churches such as the Celestial Church); All specifications are estimated using a Linear Probability Model. 66 Chapter 2 Table A.12: Including mother’s belief that AIDS can be caused by witchcraft no vaccination full vaccination ownership bed net use bed net malaria positive under 5 mortality (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) mother is an ATR adherent 0.023** 0.024** -0.019 -0.019 -0.063*** -0.063*** -0.067*** -0.067*** 0.054 0.049 7.407 7.013 (0.009) (0.010) (0.014) (0.014) (0.013) (0.013) (0.015) (0.015) (0.038) (0.038) (5.739) (5.738) mother beliefs AIDS can be -0.007 0.002 -0.010 -0.002 0.045 11.859*** caused by witchcraft (0.006) (0.010) (0.008) (0.009) (0.031) (3.806) Wealth quintile: 2 -0.020* -0.016 0.013 0.013 0.055*** 0.056*** 0.037*** 0.038*** 0.060 0.062 5.610 5.236 (0.011) (0.011) (0.015) (0.015) (0.013) (0.013) (0.015) (0.015) (0.044) (0.044) (5.852) (5.848) 3 -0.035*** -0.030** 0.029* 0.029* 0.114*** 0.114*** 0.086*** 0.086*** 0.054 0.056 4.836 4.843 (0.011) (0.012) (0.016) (0.016) (0.014) (0.014) (0.016) (0.016) (0.049) (0.048) (6.361) (6.361) 4 -0.034*** -0.034*** 0.045** 0.045** 0.190*** 0.190*** 0.138*** 0.138*** -0.054 -0.054 -13.412** -13.353** (0.012) (0.012) (0.018) (0.018) (0.016) (0.016) (0.018) (0.018) (0.051) (0.051) (6.780) (6.782) 5 -0.047*** -0.049*** 0.076*** 0.076*** 0.269*** 0.268*** 0.234*** 0.234*** -0.026 -0.024 -12.748 -12.368 (0.013) (0.013) (0.025) (0.025) (0.021) (0.021) (0.023) (0.023) (0.061) (0.061) (8.966) (8.973) age of mother -0.000 -0.000 0.001 0.001 -0.003*** -0.003*** -0.003*** -0.003*** 0.000 0.000 0.032 0.026 (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.004) (0.004) (0.302) (0.302) mother's age at first birth 0.003*** 0.003*** 0.001 0.001 0.002* 0.002* 0.003* 0.003* -0.007 -0.007 -1.982*** -1.914*** (0.001) (0.001) (0.002) (0.002) (0.001) (0.001) (0.001) (0.001) (0.004) (0.004) (0.457) (0.456) years of schooling mother -0.003*** -0.002*** 0.007*** 0.007*** 0.007*** 0.006*** 0.008*** 0.008*** -0.001 -0.001 -0.975* -0.775 (0.001) (0.001) (0.002) (0.002) (0.001) (0.001) (0.002) (0.002) (0.004) (0.004) (0.583) (0.583) polygamous household -0.000 -0.001 -0.005 -0.005 -0.002 -0.002 -0.026*** -0.026*** 0.009 0.005 23.621*** 23.583*** (0.006) (0.006) (0.010) (0.010) (0.008) (0.008) (0.010) (0.010) (0.027) (0.027) (3.846) (3.847) gender of child (girl=1) 0.002 0.002 0.009 0.009 0.011 0.011* -0.020 -0.020 (0.005) (0.005) (0.008) (0.008) (0.007) (0.007) (0.023) (0.023) age of child (in months) 0.001*** 0.001*** 0.003*** 0.003*** -0.002*** -0.002*** 0.003*** 0.003*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.001) (0.001) nr. of children < 5 in HH 0.000 -0.002 -0.003 -0.003 0.023*** 0.023*** -0.006 -0.006 0.009 0.009 -36.945*** -36.995*** (0.003) (0.004) (0.005) (0.005) (0.004) (0.004) (0.005) (0.005) (0.018) (0.018) (2.389) (2.389) ethnicity of mother Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes birth order Yes Yes Yes Yes No No Yes Yes Yes Yes No No DHS survey year Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 14,472 14,472 14,472 14,472 11,958 11,958 16,312 16,312 1,976 1,976 11,958 11,958 R2 0.27 0.26 0.24 0.24 0.37 0.37 0.32 0.32 0.49 0.49 0.18 0.18 Adjusted R2 0.19 0.18 0.16 0.16 0.28 0.28 0.26 0.26 0.27 0.27 0.07 0.07 Notes: *** p<0.01, ** p<0.05, * p<0.1; Robust standard errors are clustered at the household-level and reported in parentheses; Information on the relationship between witchcraft and AIDS was only asked in the 2006 and 2012 DHS survey rounds; We compare estimates that (do not) control for mother’s reported belief that AIDS can be caused by witchcraft; To allow for a meaningful comparison, we restrict the sample to the observations for which information on this belief is available; All specifications are estimated using a Linear Probability Model. 67 Chapter 2 Table A.13: Healthcare services chosen by ATR mothers traditional use of visit health traditional birth healer to treat use of ORS to medication to facility attendant diarrhea or treat diarrhea treat fever fever (1) (2) (3) (4) (5) mother is an ATR adherent -0.047*** 0.020*** 0.017*** 0.000 -0.081*** (0.010) (0.005) (0.006) (0.031) (0.021) visited a traditional healer -0.082** -0.176*** (0.040) (0.046) Wealth quintile: 2 0.035*** -0.008 -0.006 -0.017 0.007 (0.011) (0.005) (0.007) (0.033) (0.022) 3 0.066*** -0.009* 0.001 0.072** 0.008 (0.011) (0.005) (0.007) (0.037) (0.023) 4 0.079*** -0.021*** -0.014* 0.067 0.086*** (0.013) (0.005) (0.008) (0.044) (0.025) 5 0.117*** -0.019*** -0.014 0.171** 0.123*** (0.018) (0.005) (0.010) (0.070) (0.038) age of mother -0.004*** 0.001* 0.001** -0.001 0.002 (0.001) (0.000) (0.001) (0.003) (0.002) mother's age at first birth 0.000 0.000 -0.002** 0.009** -0.006** (0.001) (0.001) (0.001) (0.004) (0.003) years of schooling mother 0.009*** -0.001*** -0.000 0.006 0.007** (0.001) (0.000) (0.001) (0.006) (0.003) polygamous household -0.030*** 0.005* 0.002 -0.025 -0.023 (0.007) (0.003) (0.004) (0.023) (0.015) nr. of children < 5 in HH 0.023*** 0.002 -0.003 0.030*** 0.007 (0.004) (0.002) (0.003) (0.010) (0.008) age of child (in months) -0.002*** 0.000** -0.000 -0.001 0.000 (0.000) (0.000) (0.000) (0.001) (0.000) gender of child (girl=1) -0.002 0.001 -0.002 -0.024 -0.005 (0.005) (0.002) (0.004) (0.020) (0.012) ethnicity of mother Yes Yes Yes Yes Yes birth order Yes Yes Yes Yes Yes DHS survey year Yes Yes Yes Yes Yes cluster fixed effects Yes Yes Yes Yes Yes Observations 35,106 34,984 10,318 3,262 6,430 R2 0.29 0.28 0.26 0.53 0.42 Adjusted R2 0.24 0.23 0.10 0.24 0.26 Notes: *** p<0.01, ** p<0.05, * p<0.1; The dependent variables are dummy variables which take the value one if, respectively: (1) a mother visited a health facility in the 12 months prior to the interview; (2) a traditional birth attendant assisted with the delivery of the child; (3) a mother visited a traditional healer to treat the diarrhea or fever of her child in the two weeks prior to the interview; (4) a mother used ORS to treat her child’s diarrhea; (5) a mother used conventional medication (which includes aspirin, ibuprofen, paracetamol and several anti-malarial medications) to treat her child’s fever. In columns 3-5 the sample is restricted to mothers whose child had diarrhea or fever in the two weeks prior the interview; Robust standard errors are clustered at the household-level and reported in parentheses. 68 Chapter 2 Table A.14: Potential hurdles to visiting a health center No problem Small problem Big problem Obs. (1) (2) (3) (1) Knowing where to go 0.021 -0.007 -0.014 3,305 (0.020) (0.007) (0.013) (2) Maybe no female health worker 0.030 -0.008 -0.022 3,305 (0.019) (0.005) (0.014) (3) Having to take transport 0.044** -0.015** -0.030** 3,301 (0.021) (0.007) (0.014) (4) Getting permission 0.054*** 0.006** -0.060*** 11,594 (0.016) (0.003) (0.018) (5) Money for treatment 0.028*** 0.030*** -0.058*** 11,591 (0.009) (0.009) (0.018) (6) Distance 0.038*** 0.022*** -0.060*** 11,586 (0.012) (0.007) (0.018) (7) Not wanting to go alone 0.082*** -0.005 -0.078*** 11,593 (0.016) (0.004) (0.014) Notes: *** p<0.01, ** p<0.05, * p<0.1; We estimate seven specifications, looking at the determinants of each potential hurdle to visiting a health center; For each hurdle, mothers indicate whether it presents ‘no problem’, ‘a small problem’ or ‘a big problem’; The reported coefficients represent marginal effects of mothers’ ATR adherence for each answer category, calculated after an Ordered Probit regression; Robust standard errors are clustered at the level of the survey cluster and reported in parentheses; In every specification we control for the set of baseline covariates (all those reported in Table 7) and we include cluster-level fixed effects; Information on the first three hurdles was only available in the 2001 survey, information on the other hurdles was available in the 2001 and 2012 surveys. 69 Chapter 2 70 Chapter 3 3. To fish or not to fish? Resource degradation and income diversification in Benin* Summary This chapter studies the impact of natural resource degradation on income diversification in Beninese fishing communities. Using survey data and econometric analysis, we show that fishermen are more likely to diversify their income when the degradation of the fish stock is more severe. However, the level of income diversification that we find is surprisingly low and far from sufficient to relieve the stress on the lakes. The latter relates to low levels of formal education among fishermen and the unregulated use of highly productive, but damaging, fishing gear. These two factors result in a high return to fishing relative to non-fishing activities, even amidst degradation. * This chapter is based on a paper written with Romain Houssa (University of Namur, CRED) and Marijke Verpoorten (University of Antwerp, IOB). The original paper has been published in the journal Environment and Development Economics (volume 21, issue 6, pages 669-689). We thank participants at seminars and conferences in Leuven (LICOS seminar and EAAE workshop), Oxford (CSAE conference), Antwerp (IOB research day) and Berlin (DIW) for helpful comments and suggestions. This research was funded by the Research Foundation Flanders (FWO) – Grant Nr. S_2/2_2012 & F_6/12_2010 and by the Centre for Institutions and Economic Performance (LICOS). Romain Houssa acknowledges financial support from ACROPOLIS (Academic Research Group for Policy Support). We are grateful to Pierre Midogbo Gnohossou for making his bio-index available and to LICOS PhD-students for assistance in data-cleaning. The usual disclaimer applies. 71 Chapter 3 3.1. Introduction The World Commission on Environment and Development (WCED) coined the idea that poor people are stuck in a poverty-environment trap. In its report ‘Our Common Future’, the commission stated that poor people in developing countries are left with little choice but to overexploit the available natural resources in order to survive (World Commission on Environment and Development 1987). The subsequent natural resource degradation further impoverishes them, making them even more reliant on the available natural resources. In reality the relationship between poverty and the environment is more complex, and poor people are not doomed to depend on ever decreasing natural resources. Dasgupta (1993: pp.477–511), for instance, develops a model in which, instead of living on the exploitation of common-property resources, rural households have alternatives in terms of labour market participation. The attractiveness of these alternatives is directly linked to the productivity of common-property resources. Specifically, a degradation of common property resources reduces the revenues obtained from them, thus lowering the reservation wage to engage in outside employment and stimulating resource-users to reallocate their labour away from resource- dependent activities. In the extreme case, this reallocation is complete, but in many cases (including our case study), some resource dependence remains, and reallocation essentially implies income diversification. Hence, we refer to Dasgupta’s hypothesis as the income diversification hypothesis.1 Barbier (2010) develops a theoretical model in which he shows that this diversification hypothesis holds, provided that outside options are accessible and markets operate well. In the absence of well-functioning markets for labour, credit or products, a poverty-environment trap may loom at the horizon, in which degradation continues, further impoverishing communities and exacerbating market failures. Understanding the impact of resource degradation on income diversification is important. Not only because diversification directly affects income and thus poverty (Barrett et al. 2001), but also because diversification may alleviate environmental degradation (Reardon and Vosti 1995; Forsyth et al. 1998; Swinton et al. 2003; Ellis and Allison 2004). The latter may occur directly through a diminished use of natural resources, or indirectly because lower resource dependence has been shown to play a role in the emergence of successful common pool resource management systems (Cinner et al. 2013; Ernst et al. 2013; MacNeil and Cinner 2013). 1 Panel A of Figure 1 graphically presents the hypothesized relationship between degradation and income diversification that follows from this proposition. The reservation wage must as a minimum be offered if an individual is to accept a job in the labour market (Dasgupta 1993: p.479). 72 Chapter 3 In this chapter, we examine if and to what extent natural resource degradation induces artisanal fishermen and fishmongers in Benin to reallocate their labour away from resource- dependent activities. The (male) fishermen and (female) fishmongers in our study belong to communities located at the three main coastal lakes of Benin. The communities are remote and characterized by underdeveloped labour and credit markets. In the absence of effective common pool resource management, highly productive but damaging fishing gear has proliferated. By now, resource degradation has strongly reduced the fish stock available to these communities, threatening their livelihood (Gnohossou 2006; Niyonkuru and Lalèyè 2010). In our empirical framework, we disentangle three channels which could potentially yield a positive relationship between resource degradation and income diversification in our sample area (Panel B of Figure 1 offers a graphical representation). The first channel follows from Dasgupta’s (1993) model: the degradation of the fish stock decreases fishing revenues, which lowers the reservation wage for non-fishing activities and stimulates labour reallocation away from the fisheries sector.2 The other two channels are instead driven by the ownership of highly productive but damaging fishing gear. Channel 2 assumes that fishermen who own such gear have higher incomes, allowing them to more easily invest in outside options. A third channel assumes that non-fishing activities generate the income needed to invest in more productive but damaging fishing gear, resulting in higher degradation. Our study relates to a growing literature that analyses the state of common pool resources as a determinant of employment decisions among households in developing countries. For instance, Barbier (2007) finds that mangrove deforestation in coastal communities of Thailand encourages female labour participation to outside employment. Pascual and Barbier (2006, 2007) show that declining soil fertility in Mexico induces income diversification for wealthier households, but encourages poorer households to clear more forest for agricultural use. We contribute to this empirical literature, firstly by providing a case-study on African small-scale fisheries, thereby studying a context in which the quality of institutions for common pool resource management is critical as fishermen can use a wide variety of fishing gear, including gear that greatly aggravates the natural resource degradation. Secondly, unlike previous studies, we empirically address the endogeneity of resource degradation. In particular, we use an instrumental variables approach to isolate the exogenous variation in resource degradation. This allows us to deal with the two-way causality between degradation and income diversification, and study the causal impact of degradation on income diversification away from the fisheries sector. Furthermore, although we 2 In the context of our study, we define ‘reservation wage’ as the minimum wage that a fisher is willing to accept to substitute labour from fishing activities to outside employment. 73 Chapter 3 apply a quantitative method to study the impact of degradation on income diversification, our analysis builds on several months of fieldwork in the study area that generated less tangible though equally valuable qualitative observations, which helped immensely in the design of our econometric framework and interpretation of estimation results. In what follows, we first provide a background on Benin’s inland fisheries sector. Next, we present our data and the econometric framework. The final sections discuss the results and potential policy implications. 3.2. Benin’s inland fisheries Benin is a small country in West Africa (Figure 2). About half of its 9 million inhabitants live in the most southern part of the country, close to the coastal lakes and lagoons. The fisheries sector is of great importance to both national and rural economic development. Fish is the main source of animal protein consumed in the country and, in 2008, approximately 600,000 people were employed in the fisheries sector (USAID 2007; FAO 2008). Inland fisheries dominate, accounting for 75% of Benin’s fisheries production (FAO 2008). The majority of fish is caught in the three main coastal lakes: Lake Nokoué (150 km2), lake Ahémé (85 km2) and Porto-Novo lagoon (35 km2). Lake Nokoué and Porto Novo lagoon are connected by the Totché channel, making it the largest water body of the country. Two other channels connect the lakes to the Atlantic ocean. These connections are vital to the ecosystem as the inflow of marine water creates seasonal variations in the salinity, temperature and oxygen level of the water, which promotes diversity and the reproduction of aquatic fauna and flora (Amoussou 2004). Villages are located along the border of the lakes and on the water surface. They are inhabited by different ethnic groups which strongly depend on the fisheries sector for their livelihoods. In the past few decades the lakes have experienced dramatic environmental degradation leading to a loss of biodiversity and a decline of the fish stock (Amoussou 2004; Allan et al. 2005; Gnohossou 2006; USAID 2007; FAO 2008). The socio-economic changes associated with the colonization of Benin (1894-1960) gradually raised pressure on the lakes. The emergence of markets and transport systems enhanced the economic value of fishery resources beyond subsistence level, with a consequent increase in the number of fishers (Dangbégnon 2000). In addition, both the number of fishers and the demand for fish increased due to high population growth as well as migration flows to the coastal region. The colonization, as well as the increased monetization and commercialization of the economy, also marked the breakdown of the traditional natural resource management system 74 Chapter 3 which was embedded in the Voodoo religion. This system prevented overfishing and the use of damaging fishing instruments through a set of rules and sanctions. The rules were enforced by local spiritual leaders who gained legitimacy from their close association with the ancestors (Dangbégnon 2000; USAID 2007; Briones Alonso et al. 2016). The colonization and subsequent promotion of monotheistic religions in Benin reduced the influence of Voodoo and undermined the authority of local spiritual leaders. In recent years, new management institutions were introduced but they have not been able to effectively regulate fishing activities due to problems with monitoring and enforcement (Maarleveld and Dangbégnon 1999; Dangbégnon 2000; USAID 2007; Briones Alonso et al. 2016).3 The lack of effective management gave way to the introduction of highly yielding but damaging fishing instruments from the 1960s onward (Dangbégnon 2000). Especially the increased use of acadja and konou has contributed to the problem of overfishing and resource degradation (USAID 2007; FAO 2008; Niyonkuru and Lalèyè 2010). Acadja resembles a fishing pond. It is constructed by placing wooden branches in the lake and fencing them with fishing nets (USAID 2007). Konou is also a fixed fishing installation, with fine-mesh nets which are set in such a way that fish get trapped. Both instruments are more expensive to acquire than other, less productive, instruments. The Acadja can be purchased or inherited, but the market for Acadja is informal (there are no legal titles) and very thin, as it is constrained by the size of the lakes which are already saturated with Acadja. Figure A.1 in Appendix A provides a visual representation of these instruments as well as detailed information on their contribution to the resource degradation in our sample area. 3.3. Data and descriptive statistics 3.3.1. Survey design The data are obtained from a survey among 418 households, implemented in the period March- July 2009. The sample was randomly selected from Benin’s 2006 fisheries census and stratified geographically. The sample area comprises 18 villages in three communes (Kpomasse, So-Ava and Aguesgues), which are located at the three main lakes of Benin (Figure 2). We collected recall information on economic activities in the 12 months prior to the interview as well as on the evolution of economic activities for the period 2002-2009. The full survey sample includes 1,873 individuals aged 15 years or older, who individually responded to the questions. 3 The new management institutions are situated both at the national and local level. The way national regulations are implemented de facto depends on the local context. As such, at lake Ahémé, where degradation is most severe, the lake-level fishing committees are much less lenient towards the use of highly productive but damaging fishing gear. A detailed discussion of the fishing committees and their operation can be found in Briones Alonso et al. (2016). 75 Chapter 3 3.3.2. Natural resource degradation We asked fishermen and fishmongers about their opinion on the evolution of the fish stock: “In your opinion how did the fish stock evolve in the past 10 years?”. Answer categories included (1) increased, (2) decreased, (3) unchanged, (4) don’t know. The large majority of our respondents (66%) report a decreasing fish stock. Our main measure for local natural resource degradation is the village-level share of respondents who report a decreasing fish stock (Table 1). We are confident that this is a reasonable proxy. First, the variation of responses across lakes makes sense as the effects of degradation are more apparent in smaller lakes (Ahémé and Porto-Novo), which start off with less abundant natural resources and are more prone to silting. Second, self-reported degradation is higher in areas where the lake is more shallow - making it more prone to silting, low oxygen levels, poor water circulation and higher levels of salinity (Gnohossou 2006).4 Lake Ahémé stands out in this respect, as it has a maximal depth of only 2.35 meters. In addition, with a length of 24km inwards, it touches on many agricultural land parcels, of which erosion exacerbates the problem of silting (Amoussou 2004). Finally, the data reported in Table 1 are in line with existing studies on aquatic resource degradation in the sample area (Atti-Mama 1998; Cledjo 2006; Roche International 2000). 3.3.3. Income diversification We define five different income sectors: the fisheries sector, agriculture and livestock-keeping, petty trade, other self-employment in the non-farm sector (e.g. barbers, tailors, etc.) and wage employment (e.g. government officials or people employed by an NGO or private company). The fisheries sector is by far the most important sector, both in terms of employment and contribution to annual income. The large majority of the economically active sample (86%) remains at least partly employed in the fisheries sector, and at every lake the sector accounts for about 80% of annual income. Fishing activities are marginally more important at lake Ahémé and Nokoué (84.7% and 82.3% of annual income) compared to Porto Novo (77.9%). Petty trade and self-employment prevail at Porto Novo, while agriculture and livestock keeping are mostly practiced at lake Ahémé, where people have more access to farm land. Based on the 12-month recall module we construct two standard measures of income diversification. First, we derive the number of income sources. Secondly, we construct the Herfindahl index of diversification which takes into account both the number of income sources 4 In addition, where the lake is shallow, one can find the largest human settlements (on pile villages). Gnohossou (2006) finds that pollution peaks around these villages, probably because of human waste. He also finds that the lower the water quality, the lower the presence of micro-species that constitute the feed of fish species. 76 Chapter 3 and the income share derived from each source (Barrett and Reardon 2000). Among the existing specifications of the index, we use the complementary proportion (1-proportion) such that a higher index indicates stronger income diversification.5 We are specifically interested in the effect of resource degradation on the reallocation of labour away from resource-dependent activities. As the large majority of our sample is engaged in the fisheries sector, both measures of income diversification essentially capture the diversification away from the fisheries sector. To make this very explicit, our first measure will directly consider the number of activities outside the fisheries sector. Table 2 confirms that income diversification is rather limited in the sample area. People on average derive their income from 1.26 sources, of which only 0.4 sources are outside the fisheries sector. The average Herfindahl index measures 0.06 – with the index ranging from 0 to 0.5. 3.3.4. Degradation and Income Diversification A bivariate analysis shows that four factors correlate significantly with the level of income diversification (Table 2). First, income diversification is more wide-spread at the smallest water body (Ahémé). Second, income diversification is significantly higher for individuals who reported a degradation of the fish stock compared to those who did not. Third, the data show that literate individuals have higher income diversification compared to illiterate individuals. Finally, individuals who do not own the most productive fishing instruments (acadja or konou) are more likely to diversify their economic activities. Despite the degradation of the fish stock, the majority of individuals remain extremely dependent on fishing activities. This could be the case because fishing is, on average, still more profitable than other income activities, and especially so for the owners of acadja or konou. Figure 3 illustrates the relative profitability of fishing activities. The figure clearly shows that fishing is by far the most profitable activity for those who own acadja or konou. For the average fisherman who does not own these productive instruments, fishing is only marginally more profitable than self-employment and petty trade. The above statistics provide a snapshot of cross-correlations at one point in time. We are also interested in the evolution of income diversification over time. On average, 7.5% of our sample reported having abandoned an economic activity in the period 2002-2009. At every lake, fishing was the main abandoned activity. At lake Ahémé 92% of individuals who abandoned an activity 5 & + The index is defined as ! = 1 − '() *' , where *' is the income share of activity i. It ranges from zero, ) indicating that all income is generated by a single income source, to 1 − , indicating that all . income sources & equally contribute to total income. 77 Chapter 3 were fishermen, compared to 73% and 52% at lake Nokoué and Porto Novo lagoon. Over the same period, 14% of our sample reported having started a new activity. There was a strong proportional growth in non-fisheries sectors at all three lakes. Overall, the number of individuals involved in agriculture, self-employment and petty-trade grew by 20%, 34% and 75%. Regarding the motives to diversify income towards non-fishing activities, the large majority of respondents (64%) reported a decline in the relative profitability of fishing activities: 26% indicated that non-fishing activities have a higher return, while 38% mention other reasons which directly affect the reservation wage of fishermen (the degradation of the fish stock (11%), the ban on shrimp exports to the EU (15%)6 and a lack of access to productive fishing instruments (12%)). We also find that in villages with higher levels of self-reported degradation, daily fishing revenues were significantly lower in 2009 and a larger share of the population had abandoned fishing activities in the period 2002-2009 in favour for the non-fisheries sector (see Table A.1 in Appendix A). In sum, the descriptive evidence is in line with Dasgupta’s (1993) hypothesis that the degradation of the fish stock has led to a decrease in fishing revenues and thus lowered the reservation wage to engage in non-fishing activities. This is especially the case for individuals who do not possess the most productive fishing instruments. We now move to an econometric analysis to verify whether these relations also hold when controlling for potentially confounding factors. 3.4. Econometric framework Our empirical model reads: /0 = 3 + 3 6789:6:;<=> + ?@ Ω + B@ ∆ + 3 D=D + 3 6 , where < indexes individuals, ℎ households and M villages. ID'12 denotes income diversification at the individual level; 34 is a constant; 6789:6:;<=>2 denotes village-level natural resource degradation; ?'12 and B12 are vectors of individual- and household-level covariates; D=D2 denotes total village population; 6 .=H=Ié2 and L=9;= .=M=2 represent lake dummies (lake Ahémé is the base category), and εST2 6 In 2003 a ban on shrimp exports to the EU was imposed due to a lack of compliance with food standards (Houssa and Verpoorten 2015). 78 Chapter 3 is the standardized error term.7 The determinants of the number of income sources are estimated with an ordered probit model, while we estimate the determinants of the Herfindahl index with a tobit model. We estimate these models on the economically active sample, which comprises 1,220 individuals. 3.4.1. Individual level controls We study income diversification at the individual level because the financial spheres of husband and wife in Benin are largely disconnected, i.e. expenditure and investment decisions are based on individual, rather than common, budgets and preferences (LeMay-Boucher and Dagnelie 2012).8 Individual level control variables include gender, age, education, ethnicity and religion. Gender is an important determinant in the choice of economic activities, mainly due to social norms or because of differences in physical strength (Ellis 1999; Barrett et al. 2001; Lanjouw et al. 2001; Smith et al. 2001). In Benin’s coastal fisheries, the actual fishing is reserved for men while women operate as small or intermediate traders of fish, or process the catch (by smoking or drying it). In their role as fishmongers and processors, women may also see their revenues being threatened by the degradation of the fish stock, and thus be pushed to diversify their activity portfolio. Age may also play a role in the choice of economic activities. Using the konou for instance, requires physical strength which reduces with age. Young men also tend to be more mobile, which may enhance their opportunities for income diversification. Older men, on the other hand, may have been able to acquire acadja on a first come first serve basis, leaving younger men empty-handed until they inherit one. Educational attainment is another important determinant of nonfarm earnings and self-employment in rural Africa (Barrett et al. 2001). In our setting, education may increase productivity both in fishing and non-fishing activities, and can improve access to non-fishing activities. We measure educational attainment by an indicator variable which equals one if an individual is able to read a small note and write numbers.9 We control for ethnicity as some groups are traditionally more involved in fishing activities than others due to different times and patterns of settlement around and on the lakes (Pliya 1980). We include dummy variables 7 We allow for within-household dependence in the standard errors by estimating cluster-robust standard errors (see Moulton (1990) on the use of corrected standard errors). 8 LeMay-Boucher and Dagnelie (2012) analyze individual-level income and expenditure data from Beninese couples and find that not only do husbands and wives not pool their incomes, but that more than three out of four individuals are unable to estimate their spouse’s income. These findings are in line with other economic intra-household studies of West-Africa (e.g. Udry 1996 and Doss 2001), and confirm numerous anthropological accounts from West Africa that conclude that husband and wife are secretive and individually allocate their personal income to private and public goods (e.g. Clark 1994; Falen 2003 and Mandel 2006). 9 A substantial fraction of individuals who attended school for several years are still illiterate (see Table A.2 in the Appendix A). As such, the literacy indicator is a better proxy for human capital than the years of schooling an individual attended. 79 Chapter 3 for the three most prevalent ethnicities (i.e. Goun, Houedah and Tofin) which jointly account for 91% of the sample. The remaining eleven smaller ethnicities form the base category. Finally, we control for religion by including an indicator variable which equals one if an individual is a Voodoo adherent. As previously explained, the traditional natural resource management system was embedded in the Voodoo religion. The Voodoo governance system prevented overfishing and the use of damaging fishing instruments through a set of rules and sanctions (Briones Alonso et al. 2016). Voodoo adherents may therefore be less likely to use acadja or konou. 3.4.2. Household and community level controls Although budgets are separately managed by husband and wife, the size and structure of a family may still influence the activity portfolio of its members. We therefore include the size and the dependency ratio of the household as possible determinants of income diversification. The dependency ratio is calculated as the ratio of the number of inactive individuals to the number of active individuals in the household. We further control for the ownership of high-yielding fishing gear (acadja or konou). Finally, community level covariates may be important. The population size of a village may influence the opportunities of engaging in outside employment and affect the potential gains from both fishing and non-fishing activities through the availability of markets. The village distance to the nearest town may influence outside employment opportunities and occupational choices. Lake dummies are included to control for the different characteristics and location of the lakes.10 3.4.3. Endogeneity of degradation Our measure for natural resource degradation may be endogenous. First, there may be reversed causality: income diversification may affect degradation. After all, if people choose to diversify their activity portfolio, this may protect the fish stock from degradation. This endogeneity problem is exacerbated because self-reported degradation is subjective and likely depends on individual characteristics which could also be related to income diversification. Second, in the absence of well-functioning credit markets, a positive relationship between degradation and income diversification could be driven by the ownership of high-yielding fishing gear: channels 2 and 3 presented in Figure 1 both imply that individuals who own productive fishing gear are more likely to diversify their income. If either channel holds, not controlling for the ownership of high-yielding fishing gear would give rise to an omitted variable bias, making our degradation measure endogenous. 10 Table A.3 in Appendix A provides an overview of descriptive statistics for the control variables. 80 Chapter 3 We deal with these endogeneity issues in two ways. First, we use an instrumental variables (IV) approach to account for reversed causality. Second, we assess the likelihood of channels 2 and 3. Instrumental variables approach The three lakes in our sample are so-called estuaries, meaning that they are both subject to marine and riverine influences. During the rainy season, the rivers grow and fill the lakes with fresh water; during the dry season, rivers dry up and seawater enters the lakes. We instrument our degradation measure with the village-distance to the closest influx of water. For the villages at lake Ahémé this equals the distance to the channel connecting the lake with the ocean. For the villages at lake Nokoué and Porto-Novo lagoon it is the distance to the river entering the lakes from the North (see Figure 2). The instrument is relevant as it is highly correlated with self-reported degradation (0.43***). Several factors underlie this positive correlation. First, the further from an influx of water, the more shallow the lake is – which makes it more prone to silting, low oxygen levels, poor water circulation and higher levels of salinity (Gnohossou 2006).11 Second, acadja can only be placed in areas where the lake is shallow. Finally, large fishing nets such as konou are usually placed close to an influx of water, thereby leaving fewer resources for fishermen further downstream. In a first step we estimate village-level degradation by regressing it on the set of included instruments and the distance to the closest influx of water. In a second step we estimate individual- level income diversification by regressing /0'12 on 6789:6:;U=>2, estimated in the first step, and the other explanatory variables. The IV procedure is given by the following equations: 6789:6:;<=> = V + V <>WXIY + ? @Ω@ + B@ ∆@ + V D=D + 2 4 ) 2 '12 12 + 2 (2) VE 6 /0 = 3′ + 3′ 6789:6:;U=> + ? ′Ω′′ + B ′∆′′ + 3′ D=D + '12 4 ) 2 '12 12 + 2 (3) +3′E 6 The results of the first-stage estimation (Eq. 2) confirm the relevance of our instrument: distance to the closest influx of water is positively and significantly related to degradation and the F-test of the first-stage estimation equals 23.78*** (see column 3 in Table A.4). When comparing the estimates of Eq. (1) and Eq. (3), we find qualitatively similar results, though a significant difference in quantitative terms between the estimated coefficients on degradation. This indicates there is 11 At lake Nokoué, we find a negative correlation of – 0.59*** between depth of the lake and distance to the closest influx of water (we don’t have detailed information about depth for the other two lakes). 81 Chapter 3 indeed endogeneity, and an IV approach is more appropriate. Consequently, we present IV results in the remainder of this article. The Eq.(1) estimates are available in Table A.4 (in Appendix A). Likelihood of channels (2) and (3) Channels 2 and 3 necessitate a positive correlation between the ownership of acadja or konou and income diversification. If channel 1, i.e. ‘the reservation wage channel’ holds we expect a negative correlation: the use of productive fishing gear renders reservation wages more robust to a degradation of the fish stock, thus spurring less income diversification among its owners. The estimated coefficients on the ownership of these fishing instruments will therefore give a first indication of these channels. In addition, we can assess the likelihood of channels (2) and (3) by estimating two model specifications; one in which we control for the ownership of acadja or konou and one in which we do not. If either channel holds, we would find a significant change in the coefficient estimates on degradation between these two specifications (because of omitted variable bias). 3.5. Main Results Our results are in line with Dasgupta's (1993) hypothesis: the degradation of the fish stock correlates positively and significantly with income diversification. This result is robust across both measures of income diversification. The data in Table 3 indicate that a 10% increase in village- level self-reported degradation increases a fisherman’s likelihood of having an income source outside the fisheries sector by 12.19 percentage points while it increases the Herfindahl index by 0.16 units. The likelihood of having one, two or three income sources outside the fisheries sector increases respectively with 4, 3 and 5 percentage points (although the latter change is not statistically significant). The results further indicate that literate fishermen are 8 percentage points more likely to have an income source outside the fisheries sector. Tables A.5 and A.6 in Appendix A show the estimated marginal effects for all control variables. Turning to the alternative channels: The ownership of acadja or konou is negatively related to income diversification (although the coefficient is only statistically significant with respect to having three income sources outside the fisheries sector, see Panel A of Table 3). This provides first evidence against channels 2 and 3 of Figure 1, which both imply a positive and significant coefficient. When controlling separately for the ownership of these fishing instruments, we find that fishermen who own both instruments or only a konou are respectively 14 and 23 percentage points less likely to have an income source outside the fisheries sector. The ownership of only 82 Chapter 3 acadja is also associated with lower income diversification, but not significantly so (see Table A.7). Finally, comparing model estimations with and without controlling for the ownership of acadja or konou, we find that the magnitude and significance level of the degradation coefficient remain largely unchanged (see Table A.8). Taken together, these results provide evidence that our estimated effect of degradation on income diversification is not driven by channels (2) or (3), but by channel (1), i.e. a change in the reservation wage for non-fishing activities. In the following section, we test how sensitive these findings are to alternative specifications and a series of robustness checks. 3.6. Alternative specifications and robustness checks 3.6.1. Alternative measures of degradation First, we exploit the fact that we have individual-level information on degradation. Specifically, we estimate equation (1) using individual-level self-reported degradation as a measure for natural resource degradation. Doing so allows us to add village fixed-effects and therefore to check the sensitivity of our results to omitted variables at the village-level. Table B.1 in Appendix B reports the results of two regressions, one with and the other without village fixed-effects. We find that individuals who indicated a decreasing fishing stock are 11 percentage points more likely to have an income source outside the fisheries sector. When adding village fixed-effects the estimate slightly increases to 13 percentage points, indicating that the degradation effect is largely insensitive to omitted variables at the village-level. Literate fishermen are found to be 11 percentage points more likely to have an income source outside the fisheries sector, while the owners of acadja or konou are 11 percentage points less likely to diversify their income away from the fisheries sector. Second, we construct an alternative measure for natural resource degradation using a bio- index. The index was constructed by Gnohossou (2006) who collected information on physicochemical parameters (e.g. temperature, depth, salinity, transparency, oxygen levels) and the presence of aquatic vertebrates from different measuring stations at lake Nokoué. Based on this information he conducted a factor analysis and derived a pollution-sensitivity score for all types of aquatic vertebrates based on their presence at the different measuring stations throughout the year. Finally, a bio-index was calculated at the level of the measuring station by taking a weighted sum of the pollution-sensitivity scores of the present aquatic vertebrates. Thirty-four measuring stations are located within our sample area. Using their GPS locations, we calculate a village-level bio-index 83 Chapter 3 for the villages located at lake Nokoué.12 The results indicate that a 0.1 unit increase in the bio- index increases the likelihood of having an income source outside the fishing sector by 13 percentage points. Literate fishermen are found to be 19 percentage points more likely to have an income source outside the fishing sector, while the ownership of acadja or konou decreases the likelihood of diversification with 23 percentage points (see Table B.2). 3.6.2. Representativeness of the sample In order to address the concern that our sample may not be representative, we use information from Benin’s 2006 fisheries census. The census includes rather crude information on income diversification but has the advantage of a large coverage. This allows us to expand our analysis to 109 villages and 10,850 individuals located in the five communes which border lake Nokoué – the lake for which the bio-index is available. As a dependent variable, we construct an indicator variable which takes the value zero if an individual is a full-time fisher, while it equals one if this is not the case. We use Gnohossou's (2006) bio-index to create a commune-level index in ArcGIS. For each commune, we identify the 10 closest measuring stations and use their average score as a commune-level bio-index. We estimate an IV probit model, using distance to the closest influx of water as an instrument for degradation. The results in Table 4 confirm our baseline findings. A 0.1 unit increase in the bio- index increases the likelihood of having a diversified income with 13 percentage points. Furthermore, literate individuals are 29 percentage points more likely to diversify their income away from the fishing sector, while individuals who own acadja or konou are 22 percentage points less likely to have a diversified income.13 3.6.3. Robustness checks Finally, we perform a number of robustness checks. The analyses and results are reported in Appendix C. First, we repeat our analysis, using the household rather than the individual as a unit of observation. The results in Table C.1 indicate that a 10% increase in village-level self-reported degradation is associated with an increase of 1.3 income sources at the HH-level. Second, we run the regression separately for the fishermen and fishmongers in our sample. The results in Table C.2 indicate that the degradation-effect holds to a similar extent for men and for women. Third, 12 See section B.2 in Appendix B for detailed information on Gnohossou's (2006) bio-index. The village-level bio- index ranges from 2.1 to 2.4 with a standard deviation of 0.09. 13 Figure A.2 in Appendix A shows the location of the five communes which border lake Nokoué. The commune- level bio-index ranges from 1.2 to 2.4 with a standard deviation of 0.25 (see section B.2 in the Appendix B for detailed information). 84 Chapter 3 the results in Tables C.3 and C.4 show that the estimated marginal effects of degradation on income diversification are hardly affected when additionally controlling for access to credit or the ownership of various assets. Fourth, as our analysis only covers economically active individuals, we use a Heckman (1976) selection model to investigate if our findings are influenced by selection bias. The results in Table C.6 indicate this is not the case. Finally, the results in Table C.7 show that our estimate for the degradation effect is robust to using different estimation methods (IV Poisson and 2SLS instead of IV ordered probit and IV tobit). 3.7. Discussion Severe overfishing has led to a marked decline of the fish stock in southern Benin. In order to curb this declining trend, fishing communities need to alleviate the pressure on the lakes’ resources. One way to do so is to diversify income and develop activities outside the fisheries sector. As hypothesized by Dasgupta (1993), such diversification becomes more attractive as degradation worsens, lowering fishermen’s reservation wage for outside employment. Testing this hypothesis for fishing communities in southern Benin, we find that fishermen are more likely to reallocate labour towards activities outside the fishing sector in areas where natural resource degradation is more severe, a result that remains throughout a number of robustness checks and when controlling for a large number of individual-, household- and village-level covariates. The result also holds (i) when instrumenting degradation using distance to the closest influx of water, (ii) when using a bio-index to measure degradation rather than self-reported degradation, and (iii) when using a larger sample of households from the 2006 census. Although we find support for Dasgupta’s hypothesis, the level of diversification away from the fisheries sector remains low and far from sufficient to relieve the stress on the lakes. Fishing communities in our sample remain extremely dependent on the local fish stock, with fishing activities contributing up to 80% of annual income. Such dependency continues to put pressure on the lakes. Without access to attractive outside options, there may be a real danger for these fishing communities to fall into a poverty-environmental trap. We find indications that the limited levels of diversification stem from the lack of effective institutions to manage the commons, and the limited access to attractive outside options for a large part of our sample. Indeed, the level of income diversification is especially weak for fishermen that use productive, but highly damaging, fishing gear and for illiterate fishermen. Our analysis thus confirms the importance of education for diversification, showing that higher educational attainment leads to higher levels of income diversification. Schooling levels in 85 Chapter 3 our sample are, however, very low, with 70% of our respondents being illiterate. In future research we aim to look more in depth at the relationship between degradation and investment in schooling in the sample area. So far, the data show that younger generations are increasingly attending school. This is especially the case in villages with higher levels of degradation. From interviews with fishermen, we know that parents used to keep their children from school in order to train them as fishers from a young age. As the continuing degradation of the fish stock and the associated decline in fishing revenues decrease the attractiveness of fishing activities, parents seem to become more eager to send their children to school. Since educational attainment is positively related to income diversification, this may indicate a potential ‘generational effect’ of resource degradation on income diversification, mediated by schooling. Benin’s 2011 poverty reduction strategy paper explicitly aims to link the sustainable use of natural resources, including inland fisheries, to poverty reduction (OECD 2012). In that respect, our findings suggest that policy makers should promote economic activities outside the fisheries sector e.g. by stimulating entrepreneurship through micro-credit programs, improving transportation networks, and promoting education. Especially functional schooling and specific vocational training may help fishermen to access microfinance and digital technologies (such as satellite navigation, mobile phones and the internet) (FAO 2001). At the same time, the use of highly productive but damaging fishing instruments needs to be discouraged by effective regulation and monitoring. Such a two-track policy could enhance the fishermen’s access to attractive outside options, and safeguard fishing communities from a poverty-environment trap. 86 Chapter 3 References Allan, J. David, Robin Abell, Zeb Hogan, Carmen Revenga, Brad W. Taylor, Robin L. Welcomme, and Kirk Winemiller. 2005. “Overfishing of Inland Waters.” BioScience 55 (12): 1041. Amoussou, E. 2004. “Systèmes Traditionnels de Gestion Durable Du Lac Ahémé Au Bénin.” Développement Durable: Leçons et Perspectives, Acte de Colloque AUF Ouagadougou,Benin, , 263–70. Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An Empiricist’s Companion. 1 edition. Princeton: Princeton University Press. Atti-Mama, C. 1998. “Co-Management in Continental Fishing in Benin: The Case of Lake Nokoué.” In The International Workshop on Fisheries Co-Management. Maputo, Mozambique. Barbier, Edward B. 2007. “Natural Capital and Labor Allocation Mangrove-Dependent Households in Thailand.” The Journal of Environment & Development 16 (4): 398–431. ———. 2010. “Poverty, Development, and Environment.” Environment and Development Economics 15 (6): 635–60. Barrett, C.B, T Reardon, and P Webb. 2001. “Nonfarm Income Diversification and Household Livelihood Strategies in Rural Africa: Concepts, Dynamics, and Policy Implications.” Food Policy 26 (4): 315–31. Barrett, Christopher B., and Thomas Reardon. 2000. “Asset, Activity, And Income Diversification Among African Agriculturalists: Some Practical Issues.” Working Paper 14734. Cornell University, Department of Applied Economics and Management. Briones Alonso, Elena, Romain Houssa, and Marijke Verpoorten. 2016. “Voodoo versus Fishing Committees: The Role of Traditional and Contemporary Institutions in Fisheries Management.” Ecological Economics 122 (February): 61–70. Cinner, Joshua E., M. Aaron MacNeil, Xavier Basurto, and Stefan Gelcich. 2013. “Looking beyond the Fisheries Crisis: Cumulative Learning from Small-Scale Fisheries through Diagnostic Approaches.” Global Environmental Change 23 (6): 1359–65. Clark, Gracia. 1994. Onions Are My Husband: Survival and Accumulation by West African Market Women. Chicago: University of Chicago Press. Cledjo, P. 2006. “Genre de Vie et Problèmes Environnementaux Du Lac Nokoué. Thèse de Doctorat En Géographie et Gestion de l’Environnement et Aménagement.” L’Université d’Abomey-Calavi. Dangbégnon, C. 2000. “Governing Local Commons: What Can Be Learned from the Failures of Lake Ahémé’s Institutions in Benin?” Bloomington, Indiana, USA. Dasgupta, Partha. 1993. An Inquiry Into Well-Being and Destitution. Oxford University Press. Doss, Cheryl R. 2001. “Is Risk Fully Pooled within the Household? Evidence from Ghana.” Economic Development and Cultural Change 50 (1): 101–30. Ellis, F. 1999. “Rural Livelihood Diversity in Developing Countries: Evidence and Policy Implications.” ODI Natural Resources Perspective, no. 40. Ellis, Frank, and Edward Allison. 2004. “Livelihood Diversification and Natural Resource Access.” Overseas Development Group, University of East Anglia. Ernst, Billy, Julio Chamorro, Pablo Manríquez, J. M. Lobo Orensanz, Ana M. Parma, Javier Porobic, and Catalina Román. 2013. “Sustainability of the Juan Fernández Lobster Fishery (Chile) and the Perils of Generic Science-Based Prescriptions.” Global Environmental Change 23 (6): 1381–92. Falen, D. 2003. “Paths of Power: Control, Negotiation and Gender among the Fon of Benin.” Ph.D. Thesis. University of Pennsylvania. FAO. 2001. “Sustainable Fisheries Livelihoods Programme - New Directions in Fisheries.” New Directions in Fisheries. Rome, Italy: FAO. 87 Chapter 3 ———. 2008. “Vue Générale Du Secteur Des Pêches National - La République Du Bénin.” Forsyth, Tim, Melissa Leach, and Tim Scoones. 1998. “Poverty and Environment: Priorities for Research and Study-an Overview Study, Prepared for the United Nations Development Programme and European Commission.” Gnohossou, P.M. 2006. “La Faune Benthique D’une Lagune Ouest Africaine (Le Lac Nokoué Au Benin), Diversité, Abondance, Variations Temporelles et Spatiales, Place Dans La Chaine Tropique.” Dissertation. Institut Nationale Polytechnique de Toulouse, France. Heckman, James. 1976. “The Common Structure of Statistical Models of Truncation, Sample Selection and Limited Dependent Variables and a Simple Estimator for Such Models.” Annals of Economic and Social Measurement 5 (4): 475–92. Houssa, Romain, and Marijke Verpoorten. 2015. “The Unintended Consequence of an Export Ban: Evidence from Benin’s Shrimp Sector.” World Development 67 (March): 138–50. Lanjouw, Peter, Jaime Quizon, and Robert Sparrow. 2001. “Non-Agricultural Earnings in Peri- Urban Areas of Tanzania: Evidence from Household Survey Data.” Food Policy 26 (4): 385–403. LeMay-Boucher, Philippe, and Olivier Dagnelie. 2012. “The Divorced Financial Spheres of Beninese Spouses.” Journal of International Development. Maarleveld, M, and C. Dangbégnon. 1999. “Managing Natural Resources: A Social Learning Perspective,” September. MacNeil, M., and Joshua E. Cinner. 2013. “Hierarchical Livelihood Outcomes among Co- Managed Fisheries.” Global Environmental Change 23 (6): 1393–1401. Mandel, Jennifer L. 2006. “Creating Profitable Livelihoods: Mobility as a ‘Practical’ and ‘Strategic’ Gender Need in Porto Novo, Benin.” Tijdschrift Voor Economische En Sociale Geografie 97 (4): 343–63. Moulton, Brent R. 1990. “An Illustration of a Pitfall in Estimating the Effects of Aggregate Variables on Micro Unit.” The Review of Economics and Statistics 72 (2): 334–38. Niyonkuru, C., and P.A. Lalèyè. 2010. “Impact of Acadja Fisheries on Fish Assemblages in Lake Nokoué, Benin, West Africa.” Knowledge and Management of Aquatic Ecosystems, no. 399 (November): 5. OECD. 2012. Strategic Environmental Assessment in Development Practice a Review of Recent Experience. Paris: Organisation for Economic Co-operation and Development. Pascual, Unai, and Edward B. Barbier. 2006. “Deprived Land-Use Intensification in Shifting Cultivation: The Population Pressure Hypothesis Revisited.” Agricultural Economics 34 (2): 155–165. ———. 2007. “On Price Liberalization, Poverty, and Shifting Cultivation: An Example from Mexico.” Land Economics 83 (2): 192–216. Pliya, J. 1980. “La Pêche Dans Le Sud-Oest Du Bénin. Etude de Géographie Appliquée Sur La Partie Continentale et Maritime.” Paris, France: AGECOOP. Reardon, Thomas, and Stephen A. Vosti. 1995. “Links between Rural Poverty and the Environment in Developing Countries: Asset Categories and Investment Poverty.” World Development 23 (9): 1495–1506. Roche International. 2000. “Etude Du Projet d’Amenagement Des Plan d’Eau Du Sud Benin.” Republique du Benin. Roodman, David. 2009. “Estimating Fully Observed Recursive Mixed-Process Models with Cmp.” Centre for Global Development, Working paper 168. Smith, Davd Rider, Ann Gordon, Kate Meadows, and Karen Zwick. 2001. “Livelihood Diversification in Uganda: Patterns and Determinants of Change across Two Rural Districts.” Food Policy 26 (4): 421–35. Swinton, Scott M, Germán Escobar, and Thomas Reardon. 2003. “Poverty and Environment in Latin America: Concepts, Evidence and Policy Implications.” World Development 31 (11): 1865–72. 88 Chapter 3 Udry, Christopher. 1996. “Gender, Agricultural Production, and the Theory of the Household.” Journal of Political Economy, 1010–1046. USAID. 2007. “Biodiversity and Tropical Forest Assessment for Benin.” USAID. Verbeek, Marno. 2008. A Guide to Modern Econometrics. John Wiley & Sons. World Commission on Environment and Development. 1987. “Our Common Future.” New York: Oxford University Press. 89 Chapter 3 Figures Figure 1: Relationship between degradation and income diversification (ID) Notes: Panel A represents the hypothesized relationship between the degradation of common property resources and income diversification that follows from Dasgupta (1993: pp.477–511) model of labour market participation. Panel B offers an overview of the channels that could yield a positive relationship between natural resource degradation and income diversification in our sample area. Channel 1 follows from Dasgupta's (1993) model. Channels 2 and 3 crucially depend on non-convexities in production, such as fixed costs and missing credit markets – if not, earning higher incomes would not be a necessity to invest in outside options or productive fishing gear. In addition, both channel 2 and 3 embody a failure of common-pool resource management as they entail the intensive use of productive but highly damaging fishing instruments. 90 Chapter 3 Figure 2: Location of the sampling area and the three main lakes of southern Benin Notes: Our sample area comprises three communes (Kpomassè, Sô-Ava and Aguégués), located at the three main lakes of Benin (lake Ahémé, lake Nokoué and Porto Novo lagoon). 91 Chapter 3 Figure 3: Average daily income (US$), by sector Notes: The graphs represent the average daily income (in US$) for the following income categories (in order of appearance): fishing, self-employment, petty trade and agriculture. Wage employment is not considered due to the low number of observations (n=4). A t-test indicates a statistically significant difference in average daily income for fishing (p-value=0.00) between Panels A and B. As the use of Acadja and Konou is prohibited at lake Ahémé, these graphs only include observations from lake Nokoué and Porto-Novo lagoon. 92 Chapter 3 Tables Table 1: Village-level self-reported degradation (%) Ahémé Nokoué Porto Novo Adja Tokpa I 82 Agoundankomey 68 Aholoukome 37 Segbohoue I 100 Sokomey 42 Akpoloukome 73 Segbohoue II 97 Tohokomey 19 Dogodo 98 Gbetozo 91 Gbetigao 44 Djigbekome 32 Lokogbo I 100 Todo 60 Houndekome 95 Tokpa-Dome II 95 Vekky Daho 29 Kindji 91 Total 93 Total 47 Total 70 Notes: The reported percentages indicate the village-level share of respondents who indicated that the fish stock decreased over the 10 years prior to the survey. 93 Chapter 3 Table 2: Mean income diversification Nr. of income Nr. of sources outside Herfindahl income the fisheries index sources sector Lake Ahémé 1.35 0.42 0.10 Nokoué 1.27 *** 0.40 *** 0.04 *** Porto Novo 1.16 *** 0.35 *** 0.05 *** Self-reported degradation No 1.15 0.32 0.03 Yes 1.30 *** 0.43 *** 0.08 *** Literate No 1.23 0.36 0.06 Yes 1.35 *** 0.52 *** 0.09 *** Owns acadja / konou No 1.33 0.45 0.08 Yes 1.21 *** 0.36 *** 0.05 *** Total 1.26 0.40 0.06 Notes: *** p<0.01, ** p<0.05, * p<0.1; the significance levels of the differences in means were obtained from an ANOVA-test. 94 Chapter 3 Table 3: Determinants of income diversification, marginal effects Panel A: Number of income sources outside the fisheries sector 0 1 2 3 degradation -1.219*** 0.376*** 0.335*** 0.508 (0.148) (0.145) (0.063) (0.324) literacy -0.082** 0.025 0.022 0.034** (0.040) (0.022) (0.015) (0.015) ownership acadja / konou 0.102 -0.031 -0.028 -0.042* (0.067) (0.032) (0.023) (0.025) observations 1,220 1,220 1,220 1,220 Panel B: Herfindahl index degradation 1.582** (0.621) literacy 0.028 (0.032) ownership acadja / konou -0.032 (0.067) observations 1,220 1,220 1,220 1,220 Notes: *** p<0.01, ** p<0.05, * p<0.1. The coefficients represent marginal effects calculated after an IV ordered probit regression on the determinants of the number of income sources (Panel A) and an IV Tobit regression on the determinants of the Herfindahl index (Panel B). Robust standard errors are adjusted for clustering by household and are reported in parentheses. In Panel A, columns represent the number of income sources. 95 Chapter 3 Table 4: Expansion of the analysis using Benin’s 2006 fisheries census ID degradation 1.269 *** (0.167) literacy 0.292*** (0.037) ownership acadja / konou -0.220*** (0.034) observations 10,850 Notes: *** p<0.01, ** p<0.05, * p<0.1. This Table expands the analysis to 109 villages in the five communes bordering lake Nokoué (see section 3 in Appendix B for detailed information and the full regression output). The coefficients represent marginal effects calculated after an IV probit model on the determinants of income diversification. ID is a dummy variable which takes the value 0 if an individual is a full-time fisher and 1 otherwise. 96 Chapter 3 Appendix A: Figures & Tables Figure A.1: Productive, but damaging fishing instruments Panel A: Acadja Notes: The acadja resembles a fishing pond. It is constructed by placing wooden branches in the lake and fencing them with fishing nets (USAID 2007). The nets protect fish against predators and food is provided abundantly by algae and other micro-organisms which grow on the immersed parts of the branches. The acadja became a popular fishing instrument because of its high yield. From 1981-1996 the number of acadja in Benin’s coastal lakes increased by more than 1500% (USAID 2007). In 2010, they covered about 35% of lake Nokoué’s water surface (Niyonkuru and Lalèyè 2010). As mangrove wood is used to construct the acadja, the popularity of the fishing instrument has led to a decrease of mangrove cover with serious implications for the productivity and diversity of coastal ecosystems (USAID 2007). With the disappearance of mangrove cover, several fish species lost their natural habitat and the shores of the lakes have become vulnerable to erosion and silting (Amoussou 2004; Gnohossou 2006; Pliya 1980). The high concentration of acadja further causes an accumulation of mud and silt in the lakes, reducing oxygen levels and hindering water circulation (Amoussou 2004; Gnohossou 2006). 97 Chapter 3 Panel B: Konou Notes: The konou is a fixed fishing installation with fine-mesh nets. Contrary to the acadja, it is not closed by nets, but nets of several meters are set in such a way that fish get trapped. Its fine- mesh nets also trap young fish and fish eggs, thereby reducing the reproductive potential of the fisheries stock (USAID 2007). In addition, the konou is a direct source of conflict: to make use of the water current to trap fish, the nets are usually set out close to narrow channels thereby catching basically all fish entering the lake and leaving few resources for fishermen further downstream. 98 Chapter 3 Figure A.2: location of the sampling area when expanding the analysis to the villages included in Benin’s 2006 fisheries census Notes: Our sample area comprises three communes (Kpomassè, Sô-Ava and Aguégués), located at the three main lakes of Benin (lake Ahémé, lake Nokoué and Porto Novo lagoon). When expanding our sample to include all communes located at lake Nokoué, we include fishermen living in Sô-Ava, Aguégués, Abomey-Calavi (1.), Littoral (2.) and Sèmè-Podji (3.). 99 Chapter 3 Table A.1: Resource degradation, fishing revenues and the non-fisheries sector Panel A: Correlation between village-level self-reported degradation and daily fishing revenues in 2009 Ahémé -0.26 *** Nokoué -0.11 ** Porto Novo -0.13 *** Panel B: Correlation between village-level self-reported degradation and the village-share of individuals who entered the non-fisheries sector between 2002 – 2009 Petty trade 0.21 *** Agriculture & livestock keeping 0.39 *** Other self-employment outside fisheries sector 0.28 *** Wage-employment 0.39 *** Notes: *** p<0.01, ** p<0.05, * p<0.1; The data in this Table comes from a survey module on the evolution of economic activities over the period 2002-2009; The level of self-reported degradation for each lake and village is reported in Table 1 of the main manuscript. 100 Chapter 3 Table A.2: Literacy by years of schooling able to read a able to write small note numbers no schooling 2.66% 8.94% 1-2 years 18.92% 48.65% 2-5 years 68.83% 94.81% > 5 years 96.88% 97.66% Notes: The shares are calculated for 1,297 individuals of the economically active population 101 Chapter 3 Table A.3: Descriptive statistics, by lake Ahémé Nokoué Porto-Novo Age 35.6 35.8 35.2 Share of women 44.6% 49.8% 46.5% Literacy 22.6% 11.2% 15.2% Ethnicity Goun 3.3% 0.0% 73.5% Houedah 79.7% 1.0% 0.0% Tofin 0.8% 98.5% 14.6% Other 16.2% 0.5% 11.9% Share of Voodoo-adherents 74.9% 28.6% 3.4% Household size 3.9 4.9 5.7 Dependency ratio 27.6% 30.7% 65.1% Village population size 1,276 2,303 1,600 Distance to nearest town (km.) 15.7 5.6 7.7 Asset ownership acadja / konou 0.0% 76.6% 89.6% non-motorized canoe 43.9% 95.8% 93.1% outboard motor 1.0% 29.0% 13.9% generator 7.3% 21.5% 9.1% television 12.7% 16.4% 13.5% radio 73.5% 69.1% 64.9% mobile phone 49.4% 64.1% 71.0% Access to credit bank account 7.8% 6.8% 3.7% tontines (ROSCA) 83.6% 84.4% 80.4% Nr. of observations 390 412 495 Notes: The descriptive statistics in this Table represent averages by lake for the economically active population. The nearest towns to lake Ahémé, lake Nokoué and Porto-Novo lagoon are Ouidah, Abomey-Calavi and Porto-Novo, respectively. Further information on asset ownership and access to credit can be found in section 3 of Appendix C. 102 Chapter 3 Table A.4: Determinants of income diversification No IV (Eq. 1) IV (Eq. 2 and 3) 1st stage 2nd stage 2nd stage nr. of income Herfindahl nr. of income Herfindahl Dependent variable: Degradation sources index sources index (1) (2) (3) (4) (5) log distance to influx of water 0.064*** (0.013) Nokoué 0.787 0.032 -0.570*** 3.694*** 3.577*** (0.507) (0.298) (0.108) (0.620) (1.332) Porto Novo -0.263 -0.159 -0.205*** 1.653*** 1.482** (0.365) (0.214) (0.080) (0.388) (0.669) log age 0.672*** 0.300*** 0.001 0.288*** 0.294*** (0.107) (0.061) (0.014) (0.104) (0.094) female -0.096 -0.033 0.012 0.160** -0.097 (0.080) (0.047) (0.008) (0.079) (0.068) literate 0.403*** 0.143** 0.005 0.273** 0.093 (0.121) (0.070) (0.017) (0.127) (0.108) Goun -0.198 -0.138 0.065 -0.452** -0.438* (0.186) (0.107) (0.045) (0.203) (0.262) Houedah -0.647*** -0.210* 0.027 -0.605*** -0.342** (0.188) (0.109) (0.028) (0.173) (0.156) Tofin -0.529** -0.289* 0.249*** -1.372*** -1.506*** (0.248) (0.150) (0.059) (0.316) (0.549) Voodoo -0.078 -0.142 0.018 -0.129 -0.229 (0.167) (0.091) (0.024) (0.134) (0.143) household size -0.035 -0.023 -0.015*** 0.068*** 0.045 (0.026) (0.016) (0.005) (0.026) (0.035) dependency ratio 0.235*** 0.132*** 0.052*** -0.077 -0.093 (0.064) (0.035) (0.015) (0.089) (0.109) log village population size -0.581*** -0.441*** 0.131*** -0.870*** -0.971*** (0.136) (0.086) (0.023) (0.123) (0.226) log distance to nearest city 0.178 -0.203 0.023 0.481 0.028 (0.362) (0.203) (0.068) (0.332) (0.385) ownership acadja / konou -0.236 -0.137 -0.008 -0.341 -0.107 (0.188) (0.114) (0.040) (0.217) (0.223) degradation 0.750*** 0.561*** 4.080*** 5.263*** (0.259) (0.151) (0.672) (1.656) Observations 1,220 1,220 1,220 1,220 1,220 F-test of excluded instruments 23.78*** Notes: *** p<0.01, ** p<0.05, * p<0.1. The determinants of the number of income sources outside the fisheries sector were estimated using an (IV) Ordered Probit model, while the determinants of the Herfindahl index were estimated using a (IV) Tobit model. The robust standard errors are adjusted for clustering by household and are reported in parentheses. Columns 1-2 present the estimates without IV (Eq. 1) while columns 3-5 present the IV- estimates (Eq. 2 and 3). 103 Chapter 3 Table A.5: Determinants of the number of income sources outside the fisheries sector, marginal effects. Number of income sources 0 1 2 3 outside the fisheries sector: degradation -1.219*** 0.376*** 0.335*** 0.508 (0.148) (0.145) (0.063) (0.324) log age -0.086*** 0.027 0.024* 0.036** (0.033) (0.021) (0.013) (0.015) female -0.048* 0.015 0.013 0.020*** (0.025) (0.014) (0.009) (0.007) literate -0.082** 0.025 0.022 0.034** (0.040) (0.022) (0.015) (0.015) Goun 0.135** -0.042 -0.037* -0.056 (0.061) (0.030) (0.020) (0.036) Houedah 0.181*** -0.056 -0.050** -0.075*** (0.056) (0.042) (0.025) (0.029) Tofin 0.410*** -0.126* -0.113*** -0.171* (0.087) (0.065) (0.035) (0.101) Voodoo 0.039 -0.012 -0.011 -0.016 (0.040) (0.014) (0.012) (0.018) household size -0.020*** 0.006 0.006** 0.008 (0.008) (0.004) (0.003) (0.006) dependency ratio 0.023 -0.007 -0.006 -0.010 (0.026) (0.007) (0.007) (0.014) Nokoué -1.104*** 0.340** 0.303*** 0.460* (0.155) (0.157) (0.075) (0.273) Porto Novo -0.494*** 0.152** 0.136*** 0.206 (0.108) (0.075) (0.041) (0.127) log village population size 0.260*** -0.080* -0.071*** -0.108** (0.038) (0.046) (0.025) (0.052) log distance to nearest city -0.144 0.044 0.039 0.060 (0.100) (0.041) (0.030) (0.048) ownership acadja / konou 0.102 -0.031 -0.028 -0.042* (0.067) (0.032) (0.023) (0.025) Observations 1,220 1,220 1,220 1,220 Notes: *** p<0.01 ** p<0.05 * p<0.1 The coefficients represent marginal effects calculated after an IV ordered probit regression on the determinants of the number of income sources. The robust standard errors are adjusted for clustering by household and are reported in parentheses. 104 Chapter 3 Table A.6: Determinants of the Herfindahl index, marginal effects Herfindahl index degradation 1.582** (0.621) log age 0.088*** (0.029) female -0.029 (0.021) literate 0.028 (0.032) Goun -0.132 (0.083) Houedah -0.103** (0.050) Tofin -0.453** (0.195) Voodoo -0.069 (0.044) household size 0.013 (0.011) dependency ratio -0.028 (0.035) Nokoué 1.075** (0.476) Porto Novo 0.445* (0.229) log village population size -0.292*** (0.086) log distance to nearest city 0.009 (0.116) ownership acadja / konou -0.032 (0.067) Observations 1,220 Notes: *** p<0.01 ** p<0.05 * p<0.1 The coefficients represent marginal effects calculated after an IV tobit regression on the determinants of the Herfindahl index. The robust standard errors are adjusted for clustering by household and are reported in parentheses. 105 Chapter 3 Table A.7: Determinants of income diversification, marginal effects – looking at the ownership of acadja and konou separately Panel A: Number of income sources outside the fisheries sector 0 1 2 3 degradation -1.333*** 0.321* 0.323*** 0.688* (0.129) (0.166) (0.100) (0.376) literacy -0.068* 0.016 0.017 0.035** (0.039) (0.017) (0.014) (0.016) ownership acadja -0.052 0.013 0.013 0.027 (0.077) (0.016) (0.017) (0.047) ownership konou 0.227*** -0.055 -0.055 -0.117** (0.081) (0.044) (0.034) (0.054) ownership acadja and konou 0.137* -0.033 -0.033 -0.071* (0.076) (0.033) (0.027) (0.037) observations 1,220 1,220 1,220 1,220 Panel B: Herfindahl index degradation 1.925** (0.805) literacy 0.024 (0.036) ownership acadja 0.124 (0.118) ownership konou -0.175 (0.116) ownership acadja and konou -0.087 (0.087) observations 1,220 Notes: *** p<0.01, ** p<0.05, * p<0.1. The coefficients represent marginal effects calculated after an IV ordered probit regression on the determinants of the number of income sources (Panel A) and an IV Tobit regression on the determinants of the Herfindahl index (Panel B). Robust standard errors are adjusted for clustering by household and are reported in parentheses. In Panel A, columns represent the number of income sources. The coefficients in Panel A (Panel B) were estimated with the same specification as in Table A.5 (Table A.6), now controlling separately for the ownership of acadja and konou (the reference category being individuals who do not own acadja or konou). 106 Chapter 3 Table A.8: Marginal effect of degradation on income diversification – with and without controlling for the ownership of acadja / konou Panel A: Number of income sources outside the fisheries sector Controlling for ownership acadja / Konou ? 0 1 2 3 Yes -1.219*** 0.376*** 0.335*** 0.508 (0.148) (0.145) (0.063) (0.324) No -1.222*** 0.384*** 0.336*** 0.502 (0.149) (0.142) (0.062) (0.318) observations 1,220 1,220 1,220 1,220 Panel B: Herfindahl index Controlling for ownership acadja / Konou ? Yes 1.582** (0.621) No 1.560** (0.607) observations 1,220 Notes: *** p<0.01, ** p<0.05, * p<0.1. The coefficients represent marginal effects calculated after an IV ordered probit regression on the determinants of the number of income sources (Panel A) and an IV Tobit regression on the determinants of the Herfindahl index (Panel B). Robust standard errors are adjusted for clustering by household and are reported in parentheses. In Panel A, columns represent the number of income sources. The coefficients in Panel A (Panel B) were estimated with the same specification as in Table A.5 (Table A.6). 107 Chapter 3 Appendix B: Alternative specifications 1. Individual-level information on degradation We exploit the fact that we have individual-level information on degradation. Specifically, we estimate equation (1) using individual-level self-reported degradation as a measure for natural resource degradation. Doing so allows us to add village fixed-effects and therefore to check the sensitivity of our results to omitted variables at the village-level. Table B.1 reports the results of two regressions, one with and one without village fixed-effects. In both specifications we find that individuals who indicated a decreasing fishing stock are more likely to have an income source outside the fisheries sector. Moreover, the size of the degradation effect is highly comparable across specifications (11 vs 13 percentage points). From specification A to B, the pseudo R2-value strongly increases from 0.09 to 0.20, indicating that omitted variables at the village-level explain quite some additional variation in the number of income sources. The estimated degradation effect however barely changes, providing evidence that it is largely insensitive to such omitted variables. Although the use of individual-level self-reported degradation allows us to control for omitted variables at the village-level, there are two main reasons why we prefer the use of the village-level aggregate. First, the individual-level measure may suffer more strongly from endogeneity as it is related to individual characteristics which may both influence perceived degradation and income diversification. For instance, full-time fishers may be better informed about the evolution of the fishing stock and may hence report higher levels of degradation. The opposite might hold for fishers who have diversified their income, leading to a negative relationship between income diversification and degradation. Aggregating self-reported degradation at the village-level is therefore likely to give a more accurate representation of the actual level of degradation. Second, as our instrumental variable is specified at the village-level, we cannot include village-fixed effects in the IV-approach. 108 Chapter 3 Table B.1: Using individual-level degradation, with and without including village FE Panel A: without village FE Number of income sources 0 1 2 3 outside the fisheries sector: degradation -0.106*** 0.077*** 0.023*** 0.005** (0.027) (0.019) (0.007) (0.002) literacy -0.153*** 0.110*** 0.035*** 0.008*** (0.028) (0.019) (0.009) (0.003) ownership acadja / konou 0.191*** -0.135*** -0.045*** -0.011*** (0.032) (0.020) (0.010) (0.004) Observations 1,220 1,220 1,220 1,220 Panel B: with village FE Number of income sources 0 1 2 3 outside the fisheries sector: degradation -0.132*** 0.098*** 0.027*** 0.007** (0.032) (0.024) (0.008) (0.003) literacy -0.106*** 0.079*** 0.022*** 0.005** (0.035) (0.026) (0.008) (0.002) ownership acadja / konou 0.110** -0.082** -0.023** -0.006* (0.046) (0.035) (0.010) (0.003) Observations 1,220 1,220 1,220 1,220 Notes: *** p<0.01 ** p<0.05 * p<0.1 The coefficients represent marginal effects calculated after an ordered probit regression on the determinants of the number of income sources. The robust standard errors are adjusted for clustering by household and are reported in parentheses. We use individual-level self-reported degradation as a measure for natural resource degradation. In Panel B, we estimate the same specification as in Panel A, but we include village fixed effects. In every specification we control for the following covariates: age, gender, literacy, ethnicity, religion, household size, dependency ratio, lake, village population size, distance to the nearest city and ownership of acadja / konou. The pseudo R2 – values for the two specifications are respectively 0.12 and 0.27. 2. Bio-index based on physicochemical parameters As a second sensitivity check we construct an alternative measure for natural resource degradation, using a bio-index. The index was constructed by Gnohossou (2006) who studied the impact of water pollution on aquatic fauna at lake Nokoué. In order to study this relationship, Gnohossou (2006) focused on 33 different types of aquatic vertebrates. He justifies this approach mentioning three main points : 1) aquatic vertebrates play an important role in the food chain as they constitute the main source of food for fish in the lake; 2) they facilitate organic matter degradation and thus play a key role in the ecological functioning of aquatic ecosystems; 3) due to the specific development stages of aquatic invertebrates, they allow to study acute types of pollution even after the toxic substances which caused it are no longer measurable in the water. First, he installed 79 different measuring stations across lake Nokoué, for which he collected physicochemical parameters of water quality: temperature, depth, salinity, transparency and oxygen levels. These parameters were used in a factor analysis to calculate a pollution gradient 109 Chapter 3 for each measuring station. A pollution-sensitivity score was then calculated for all types of aquatic vertebrates based on their presence at the different measuring stations throughout the year. To take into account seasonal variations, measurements took place both during the rainy season and the dry season. Finally, a bio-index was calculated at the level of the measuring station by taking a weighted sum of the pollution-sensitivity scores of the present aquatic vertebrates. The index ranges from 1 (very polluted water) to 5 (water of high quality). 34 of the 79 measuring stations in Gnohossou's (2006) study are located within our sample area. Using the GPS locations of these stations, we calculate a village-level bio-index for the six villages in our sample which are located at lake Nokoué. In ArcGIS, we calculate the average score for the measuring stations within a 4 km buffer around each village. We recoded the index such that a higher score relates to higher levels of degradation. Specifically, we recoded any value Y of the index in the following way Y è 5 − (Y − 1). For instance, if Y = 2.9 in the original index, it equals 3.1 in the recoded index. The village-level bio-index ranges from 2.1 to 2.4 with a standard deviation of 0.09. Table B.2 presents the marginal effects calculated after estimating equation (3) using the bio-index as a measure for natural resource degradation. The results indicate that a 0.1 unit increase in the bio-index increases the likelihood of having an income source outside the fishing sector by 12.66 percentage points. Literate fishermen are found to be 19 percentage points more likely to have an income source outside the fishing sector. Table B.2: Using the bio-index Number of income sources 0 1 2 3 outside the fisheries sector degradation -1.266** 0.937** 0.206* 0.123 (0.605) (0.445) (0.114) (0.077) literacy -0.185*** 0.137*** 0.030** 0.018* (0.068) (0.051) (0.013) (0.010) ownership acadja / konou 0.231*** -0.171*** -0.038*** -0.022** (0.056) (0.042) (0.013) (0.010) observations 357 357 357 357 Notes: *** p<0.01 ** p<0.05 * p<0.1 The coefficients represent marginal effects which were calculated after an IV ordered probit regression on the determinants of the number of income sources. As a measure for degradation, the village-level bio-index was used. Columns represent the number of income sources. We additionally control for the following covariates: age, gender, literacy, religion, village population size, household size, dependency ratio, distance to the nearest city and ownership of acadja / konou. 110 Chapter 3 Our two measures of degradation differ across several dimensions. Self-reported degradation relates to the perceived change of the fish stock for respondents on the three lakes, while the bio- index relates to the measured water quality at a certain point in time, for lake Nokoué only. Illustrative of these differences, is the correlation coefficient between the two measures, which only amounts to 0.09 (***). On the other hand, the correlation coefficient is significant, and positive, which is in line with our claim on p. 8 that “self-reported degradation is higher in areas where the lake is more shallow - making it more prone to silting, low oxygen levels, poor water circulation and higher levels of salinity”. It is also in line with the analysis presented by Gnohossou (2006), who studied the relation between his index of water quality and the presence of crustacean and micro-organism that constitute the feed of fish species caught by the fishermen (to establish this feed, he investigated the stomach content of 754 fish from 22 different species). Although this evidence is from one point in time, and therefore does not reveal anything about the evolution of the fish stock over time, it is very plausible to assume that “areas of low water quality have not only a lower fish density but also more fish stock reduction”, because the variation in water quality captured by Ghonossou’s index is very much prone to factors that change over time, in particular the increased population of human settlement. As such, Gnohossou (2006) finds the highest levels of pollution nearby the fastest growing and most dense settlement on the water, i.e. Ganvié. 3. Representativeness of the sample As a third sensitivity check we address the concern that our sample may not be representative. To do so, we use information from Benin’s 2006 fisheries census. The census includes rather crude information on income diversification but has the advantage of a large coverage. This allows us to expand our analysis to 109 villages and 10,850 individuals located in the five communes which border lake Nokoué – the lake for which the bio-index is available. As a dependent variable, we construct an indicator variable which takes the value zero if an individual is a full-time fisher, while it is equal to one if this is not the case. We use Gnohossou's (2006) index to create a commune-level bio-index in ArcGIS. For each commune, we identify the 10 closest measuring stations and use their average score as a commune-level bio-index. As with the village-level index, we recoded the bio-index such that a higher score relates to higher levels of degradation. The commune-level bio-index ranges from 1.2 to 2.4 with a standard deviation of 0.25. We estimate an IV probit model, using the average distance from the 10 measuring stations to the closest influx of water as an instrument for degradation. The results in Table B.3 confirm 111 Chapter 3 our baseline findings. A 0.1 unit increase in the bio-index increases the likelihood of having a diversified income with 12.69 percentage points. Furthermore, literate individuals are 29.2 percentage points more likely to diversify their income away from the fishing sector, while individuals who own acadja or konou are 22 percentage points less likely to have a diversified income. 112 Chapter 3 Table B.3: Expansion of the analysis using Benin’s 2006 fisheries census IV: 1st stage 2nd stage Dependent variable: degradation ID (1) (2) log distance to influx of water 0.247*** (0.007) log age 0.027*** 0.264*** (0.005) (0.064) female -0.007 -0.522 (0.012) (0.349) literate 0.031*** 0.292*** (0.004) (0.037) Goun -0.365*** -0.225** (0.009) (0.111) Tofin -0.037*** -0.798*** (0.012) (0.108) Houédah -0.040 -0.833** (0.025) (0.348) Aïzo 0.113*** 0.464*** (0.012) (0.090) Wémè -0.486*** 1.157*** (0.008) (0.123) Xwla -0.004 -0.506*** (0.007) (0.083) Voodoo 0.049*** -0.264*** (0.003) (0.047) total nr. of children -0.002*** -0.016** (0.001) (0.008) nr. of dependent children -0.001** 0.013 (0.001) (0.009) log village population size 0.045*** -0.122*** (0.003) (0.025) log distance to nearest city 0.197*** -0.233*** (0.004) (0.060) ownership acadja / konou 0.001 -0.220*** (0.003) (0.034) degradation 1.269*** (0.167) Observations 10,850 10,850 Notes: *** p<0.01, ** p<0.05, * p<0.1. Using data from Benin’s 2006 fisheries census, this table expands the analysis to 109 villages in the five communes bordering lake Nokoué. The coefficients represent marginal effects calculated after an IV probit model. ID is a dummy variable which takes the value 0 if an individual is a full-time fisher and 1 otherwise. We include the six largest ethnicities in the regression, the 10 remaining smaller ethnicities form the base category. The first stage estimates the determinants of the commune-level bio-index using the distance to the closest influx of water as an instrument; the second stage looks at the determinants of fishing full time. 113 Chapter 3 Appendix C: Robustness checks 1. Household-level analysis Table C.1: Income diversification at the HH-level IV: 1st stage 2nd stage nr. of HH income Dependent variable: degradation sources (1) (2) log distance to influx of water 0.049*** (0.011) Nokoué -0.592*** 9.997*** (0.104) (3.810) Porto-Novo -0.218*** 3.910** (0.082) (1.892) log age of HH head -0.049** 0.968** (0.023) (0.471) HH head is female 0.048** -0.724 (0.023) (0.461) HH head is literate -0.024 0.811** (0.024) (0.347) HH head is Goun 0.090* -1.355 (0.054) (0.852) HH head is Houédah 0.006 -0.917* (0.028) (0.549) HH head is Tofin 0.248*** -3.849** (0.062) (1.511) HH head is Voodoo adherent 0.022 -0.654 (0.025) (0.466) household size -0.002 0.214*** (0.005) (0.081) dependency ratio 0.034** -0.622*** (0.014) (0.217) log village population size 0.114*** -2.316*** (0.021) (0.608) log distance to nearest city 0.025 0.445 (0.062) (0.952) ownership acadja / konou -0.050 0.093 (0.031) (0.587) degradation 13.350*** (4.595) Observations 418 418 Notes: *** p<0.01, ** p<0.05, * p<0.1. The robust standard errors are adjusted for clustering by household and are reported in parentheses. The coefficients were estimated using a 2SLS regression. 114 Chapter 3 We study income diversification at the individual level because the financial spheres of husband and wife in Benin are largely disconnected. Expenditure decisions are based on individual budgets rather than a common one (LeMay-Boucher and Dagnelie 2012). As a first robustness check, we repeat the analysis at the household-level. The results in Table C.1 indicate that a 10% increase in village-level self-reported degradation is associated with an increase of 1.3 income sources at the household level. 2. Estimating the degradation effect separately for men and women About 47% of our sample observation are women. In Benin’s coastal fisheries, the actual fishing is reserved for men (with the exception of setting traps for crabs and looking for oysters) while women operate as small or intermediate traders of fish, or process the catch (by smoking or drying it). In their role as fishmongers and processors, women may also see their revenues being threatened by the degradation of the fisheries stock, and thus be pushed to diversify their activity portfolio. As a second robustness check, we estimate the degradation effect separately for men and women. The coefficients in Table C.2 represent marginal effects calculated after an IV Ordered Probit regression on the determinants of the number of income sources. The results indicate that the degradation-effect holds both for men and for women. When running the same specification on the full sample (men and women) with an interaction term between degradation and gender, the results indicate that the gender-difference in the degradation effect is statistically insignificant. Table C.2: Estimating the degradation effect separately for men and women Number of income sources 0 1 2 3 outside the fisheries sector -1.202*** 0.433** 0.336*** 0.433 Women (567 obs.) (0.180) (0.177) (0.073) (0.361) -1.239*** 0.344** 0.334*** 0.561 Men (653 obs.) (0.176) (0.149) (0.087) (0.374) Notes: *** p<0.01, ** p<0.05, * p<0.1 The coefficients represent marginal effects calculated after an IV Ordered Probit regression on the determinants of the number of income sources. The robust standard errors are adjusted for clustering by household and are reported in parentheses. We ran two regressions, one on the female sample and one on the male sample. We additionally control for the following covariates: age, literacy, lake, village population size, ethnicity, household size, dependency ratio, distance to the nearest city and ownership of acadja / konou. 115 Chapter 3 3. Controlling for access to credit and asset ownership The level of income diversification may be affected by an individual’s access to credit. As a third robustness check, we include two measures to control for access to credit. The first one is an indicator variable which takes the value one if an individual has a bank account (this is the case for about 6% of the sample). The second is an indicator variable which takes the value of one if an individual is involved in a ‘tontine’, i.e. ROSCA or Rotating Savings and Credit Association (this is the case for about 80% of the sample). The coefficients reported in Table C.3 are marginal effects calculated after an IV Ordered Probit estimation on the determinants of the number of income sources. The estimates do not show a strong correlation between access to credit and income diversification. Having a bank account or being involved in a ‘tontine’ is not significantly related to the number of income sources. The estimated marginal effects of self-reported degradation on income diversification are hardly affected by additionally controlling for access to credit, which suggests that access to credit does not dramatically alter the relationship between degradation and income diversification. Although these results are interesting, we do not include them in the baseline analysis for two main reasons. First, and most importantly, access to credit is endogenous: those who have a more diversified income, or use highly productive fishing gear, may have a higher income, which could facilitate their access to credit – especially when some kind of collateral is necessary. It is therefore not clear in which direction this relationship goes. Second, we do not have information on access to credit for the full sample; adding these variables therefore significantly reduces our sample size (from 1,220 to 862 observations). 116 Chapter 3 Table C.3: Controlling for access to credit Number of income sources 0 1 2 3 outside the fisheries sector A) degradation -1.029*** 0.479*** 0.300*** 0.250 (0.277) (0.075) (0.099) (0.257) B) degradation -1.038*** 0.477*** 0.301*** 0.259 (0.267) (0.079) (0.094) (0.258) bank account -0.060 0.028 0.017 0.015 (0.059) (0.035) (0.018) (0.013) C) degradation -1.058*** 0.472*** 0.306*** 0.280 (0.242) (0.086) (0.084) (0.253) bank account -0.062 0.028 0.018 0.016 (0.059) (0.033) (0.018) (0.014) tontines -0.028 0.012 0.008 0.007 (0.043) (0.021) (0.012) (0.011) Observations 862 862 862 862 Notes: *** p<0.01, ** p<0.05, * p<0.1 The coefficients represent marginal effects calculated after an IV Ordered Probit regression on the determinants of the number of income sources. The robust standard errors are adjusted for clustering by household and are reported in parentheses. We estimate three different specifications (A, B, C), gradually adding two variables which capture access to credit (having a bank account and being involved in a tontine). In every specification we additionally control for the following covariates: age, gender, literacy, ethnicity, religion, household size, dependency ratio, lake, village population size, distance to the nearest city and ownership of acadja / konou. Since wealth plays a role in the acquirement of fishing methods, and may thus both directly and indirectly affect diversification, we further explore this variable by controlling for the ownership of six assets: outboard motor, mobile phone, non-motorized canoe, generator, radio and TV (Table A.3 presents summary statistics on the ownership of these assets at the three lakes). The coefficients in Table C.4 represent marginal effects calculated after an IV Ordered Probit regression on the determinants of the number of income sources. The results indicate that ownership of the above assets is not significantly related to income diversification. Furthermore, the degradation effect barely changes when additionally controlling for asset ownership of different kinds. 117 Chapter 3 Table C.4: Additionally controlling for the ownership of different assets Number of income sources 0 1 2 3 outside the fisheries sector: A) degradation -1.219*** 0.376*** 0.335*** 0.508 (0.148) (0.145) (0.063) (0.324) B) degradation -1.205*** 0.412*** 0.341*** 0.453 (0.166) (0.137) (0.055) (0.310) owns outboard motor -0.007 0.002 0.002 0.003 (0.059) (0.021) (0.017) (0.022) owns mobile phone 0.051 -0.017 -0.014 -0.019 (0.035) (0.013) (0.010) (0.019) owns non-motorized canoe 0.018 -0.006 -0.005 -0.007 (0.038) (0.014) (0.011) (0.014) owns generator -0.072 0.025 0.020 0.027 (0.062) (0.022) (0.017) (0.030) owns radio -0.036 0.012 0.010 0.013 (0.035) (0.014) (0.010) (0.015) owns tv -0.029 0.010 0.008 0.011 (0.060) (0.020) (0.017) (0.025) Observations 1,220 1,220 1,220 1,220 Notes: *** p<0.01, ** p<0.05, * p<0.1 The coefficients represent marginal effects calculated after an IV Ordered Probit regression on the determinants of the number of income sources. The robust standard errors are adjusted for clustering by household and are reported in parentheses. We estimate two different specifications (A, B). In specification B we control for the ownership of 6 different assets. In every specification we control for the following covariates: age, gender, literacy, ethnicity, religion, household size, dependency ratio, lake, village population size, distance to the nearest city and ownership of acadja / konou. 4. Heckman selection model As our analysis only covers economically active individuals, we use a Heckman selection model to investigate if our findings are influenced by selection bias. The selection equation in the model estimates the determinants of being economically active, the outcome equation estimates the determinants of the number of income sources among the economically active sample. To satisfy the exclusion restriction we use an indicator variable which takes the value one if an individual is between 15 and 25 years of age and zero otherwise. Individuals in this age- category are more likely to attend school and hence less likely to be economically active. Table C.5 shows that those in the age cohort 15-25 on average have 3.62 years of schooling, while those who are older on average have less than 1 year of education. Furthermore, only 39% of those within the 15-25 age cohort are economically active compared to 85% of those who are older. 118 Chapter 3 Table C.5: Summary statistics on age cohort 15-25 Aged 15-25 ? years of schooling % economically active Yes 3.62 0.39 No 0.99 *** 0.85 *** Notes: *** p<0.01 ** p<0.05 * p<0.1; The significance levels were obtained from a t-test. Conditional on being economically active, we do not expect this age group to strongly influence the level of income diversification. Table C.6 presents the results of the Heckman selection model. In Columns 1 and 2 we first estimate the determinants of the number of income sources with and without controlling for the 15-25 age cohort. The results imply that belonging to this age cohort does not have a significant impact on the number of income sources an individual has. In Column 3 we estimate the determinants of being economically active; i.e. the Heckman selection equation. The results indicate that individuals in the 15-25 age cohort are 76% less likely to be economically active. Finally, in Column 4 we estimate the Heckman outcome equation. The estimated coefficient on degradation is practically unchanged with respect to those estimated in Columns 1 and 2 (although it loses some significance). Importantly, we find that the correlation between the error terms of the selection and outcome equation is low and statistically insignificant, indicating that our results are not influenced by selection bias. 119 Chapter 3 Table C.6: Heckman selection model Estimation method Ordered Probit Ordered Probit Heckman nr. of income nr. of income nr. of income Dependent variable: economically active sources sources sources (1) (2) (3) (4) degradation 0.750*** 0.746*** 0.464* 0.767* (0.259) (0.260) (0.257) (0.414) log age 0.672*** 0.492*** 0.765*** 0.723 (0.107) (0.171) (0.206) (0.914) female -0.096 -0.110 -0.155** -0.099 (0.080) (0.082) (0.077) (0.090) literate 0.403*** 0.396*** -0.609*** 0.379 (0.121) (0.121) (0.097) (0.503) Goun -0.198 -0.216 -0.109 -0.197 (0.186) (0.187) (0.143) (0.189) Houedah -0.647*** -0.635*** -0.292 -0.656*** (0.188) (0.189) (0.210) (0.225) Tofin -0.529** -0.534** 0.520** -0.518 (0.248) (0.250) (0.212) (0.341) Voodoo -0.078 -0.082 0.094 -0.075 (0.167) (0.168) (0.142) (0.170) household size -0.035 -0.033 -0.064** -0.037 (0.026) (0.026) (0.027) (0.052) dependency ratio 0.235*** 0.233*** -0.652*** 0.203 (0.064) (0.065) (0.063) (0.606) Nokoué 0.787 0.812 -0.995** 0.756 (0.507) (0.506) (0.505) (0.797) Porto Novo -0.263 -0.237 -0.346 -0.275 (0.365) (0.366) (0.369) (0.426) log village population size -0.581*** -0.590*** -0.322*** -0.588*** (0.136) (0.136) (0.108) (0.181) log distance to nearest city 0.178 0.184 -0.193 0.170 (0.362) (0.361) (0.386) (0.388) ownership acadja / konou -0.236 -0.248 0.267 -0.223 (0.188) (0.192) (0.165) (0.313) age cohort 15-25 -0.208 -0.761*** (0.159) (0.186) Observations 1,220 1,220 1,873 1,220 athanrho 0.113 (2,17) rho 0.112 (2,14) Notes: *** p<0.01, ** p<0.05, * p<0.1; The selection equation (column 3) measures the determinants of being economically active, while the outcome equation (column 4) measures the determinants of the number of income sources for the economically active sample. As an exclusion restriction, we use an indicator variable which takes the value one if an individual belongs to the 15-25 age-cohort and zero otherwise. Columns 1 and 2 estimate the determinants of the number of income sources with and without including this indicator variable. The Heckman model was estimated with the stata-command ‘cmp’ which allows us to estimate the outcome equation with an Ordered Probit model (Roodman 2009). The (athan)rho variables indicate that the correlation between the error terms of the selection and outcome equations is low and statistically insignificant; i.e. our results are not influenced by selection bias. 120 Chapter 3 5. Using different estimation methods We assume there is an upper bound to the number of income sources from which a person can derive income. We therefore opt for an Ordered Probit model to estimate the determinants of the number of income sources in the baseline estimation. As a robustness check we use a Poisson model for count data, which has two important differences: First, the values of the outcome have a cardinal rather than just ordinal meaning (i.e. 2 income sources are considered to be twice as much as 1 income source; 4 double as much as 2, etc.). Second, the Poisson model does not assume a natural upper bound to the outcome variable (Verbeek 2008). With respect to the Herfindahl index, we use a Tobit model in the baseline estimation as the index is continuous, but its range is constrained. The Tobit model is particularly useful in our case, where the dependent variable is zero for a substantial part of the population but positive for the rest of the population (Verbeek 2008). Angrist and Pischke (2009: pp.102–107) argue however that there is a case for estimating Linear Probability Models instead of using nonlinear models such as Tobit. As a robustness check, we compare the 2SLS estimate to the marginal effect calculated after estimation with an IV Tobit model. The IV results in Table C.7 shows that using different estimation methods yields qualitatively similar marginal effects of degradation on income diversification. Table C.7: The degradation effect when using different estimation methods Panel A: Number of income sources 0 1 2 3 outside the fisheries sector -1.219*** 0.376*** 0.335*** 0.508 1. IV Ordered Probit (0.148) (0.145) (0.063) (0.324) 1.677*** 2. IV Poisson (0.464) Panel B: Herfindahl index 1.582** 1. IV Tobit (0.621) 0.923*** 2. 2SLS (0.259) Notes: *** p<0.01 ** p<0.05 * p<0.1 The coefficients represent marginal effects of degradation on income diversification calculated with different estimation methods. For specification 1 in Panel A, the columns represent the number of income sources. The robust standard errors are adjusted for clustering by household and are reported in parentheses. In every specification we control for the following covariates: age, gender, literacy, ethnicity, religion, household size, dependency ratio, lake, village population size, distance to the nearest city and ownership of acadja / konou. 121 Chapter 3 122 Chapter 4 4. Artisanal or industrial conflict minerals? Evidence from Eastern Congo* Summary Existing research suggests a strong link between mining and local conflict, but makes no distinction between artisanal and industrial mining. We conceptualize and hypothesize how these extraction modes have very different implications for the level and type of conflict. To test the hypotheses, we rely on a sample of 2,026 artisanal mining sites and 3,700 large-scale mining concessions in Eastern Congo, and exploit the quasi-exogenous variation in mineral values and the granting of industrial mining concessions. We find that a rise in the value of artisanal mining sites increases battles between armed actors, indicating competition between rapacious rebel groups. In contrast, the expansion of industrial mining decreases battles, suggesting that companies can secure their concessions; yet such expansion increases riots. Finally, when expanding into areas where artisanal miners are active, industrial mining increases violence against civilians and looting. * This chapter is based on a paper written with Marijke Verpoorten (University of Antwerp, IOB) and Peter van der Windt (New York University Abu Dhabi). We thank Tom De Herdt, Joachim De Weerdt, Dominic Parker and Jo Swinnen for comments. We also thank IPIS and CAMI for sharing data. 123 Chapter 4 4.1. Introduction For many countries in the world, resources are a curse rather than a blessing (Angrist and Kugler 2008; Bulte et al. 2005; Collier and Hoeffler 2004, 1998; Fearon 2005; Fearon and Laitin 2003; M. Humphreys 2005; Lujala 2010; Ross 2004; van der Ploeg 2011). Also at the subnational level, resources seem to be cursed, with local mineral extraction shown to increase the risk of local conflict (Berman et al. 2017; Buhaug and Rød 2006; Dube and Vargas 2013). However, to date, we know very little about how different modes of mineral extraction relate to conflict. There are two main ways to extract mineral resources: artisanal mining (ASM) and large- scale mining (LSM). ASM refers to a largely manual mode of extraction, practiced by individuals, groups or communities.1 LSM refers to a mechanized mode of production, practiced by large, often international, companies. These forms of mineral extraction are very different. ASM provides working opportunities for millions of families around the world. It is estimated that more than 100 million people worldwide depend directly or indirectly on ASM for their livelihoods, with up to 20 million living in Africa (CASM 2009: p.9). In contrast, LSM is highly capital-intensive. Furthermore, while ASM is intimately connected with the livelihoods of local communities, LSM is often disruptive for local communities and struggles to establish forward and backward linkages with the local economy (Banchirigah 2008; Carstens and Hilson 2009; Geenen 2013). ASM activities are often controlled and taxed by local elites (or warlords). Companies that undertake LSM activities, on the other hand, often maintain close relations with national elites. Because of these differences, ASM and LSM are likely to relate very differently to conflict. To understand how the different modes of mineral extraction relate to local conflict, we study the Democratic Republic of Congo (DRC). The DRC is a textbook case when it comes to the resource curse. Its untapped deposits of raw minerals are estimated to be worth US$24 trillion (UNEP 2011), but the majority of its population is dismally poor, mainly because both war and political mismanagement have ravaged the country. The eastern part of the country is home to a large number of artisanal mining sites and a growing number of industrial mining concessions. The Congo is also well-known for its persistent episodes of conflict, even after the formal end of the two Congo wars (1996-1998 and 1998-2003).2 Several advocacy groups have argued that ‘conflict minerals’ contribute to perpetuating the violence (Autesserre 2012; Prendergast 2009; 1 The term ASM incorporates both artisanal and small-scale mining. In small-scale mining, mineral extraction is partly mechanized and may involve small companies. In this study, the abbreviation ASM refers to artisanal mining only. 2 Both wars are described and discussed at length in among others Autesserre (2010), Reyntjens (2010), and Stearns (2011). 124 Chapter 4 Seay 2012). The subject of this ‘conflict minerals’ narrative has been ASM, while LSM has been completely overlooked. To understand how ASM and LSM may impact local conflict dynamics, we present a conceptual framework, featuring ‘a rapacity effect’, ‘an opportunity cost effect’, and ‘a protection effect’. Next, we test the resulting hypotheses empirically. To do so we build on detailed, geo- referenced information about 2,026 artisanal mining sites, 3,700 LSM concessions and 6,542 conflict events in Eastern Congo. Specifically, we explore four types of local conflict events: battles between armed actors, violence against civilians, riots and looting. The units of observation are 25x25 kilometer grid cells. The analysis exploits within-cell variations over time. To identify the impact of ASM and LSM on local conflict, we take advantage of two types of shocks that affected ASM and LSM activities in Congo in the period 2004 to 2015. First, a surge in world mineral prices translated to large increases in the value of Congolese minerals. Second, the introduction of a new Mining Code and Mining Regulations in 2002 and 2003 led to a strong increase in the number of granted LSM research and production permits in subsequent years. The data show that artisanal and industrial mining have very different implications for conflict. A rise in the value of artisanal mining sites increases battles, violence against civilians and looting. This suggests that the multiple armed actors that are present in ASM sites intensify their fighting efforts over the increased value of the extraction site. In terms of our conceptual framework, this means that the ‘rapacity effect’ dominates the ‘opportunity cost’ and ‘protection’ effects. In contrast, the expansion of LSM production activities leads to fewer incidences of battles between armed actors. This is consistent with the idea that LSM companies have the means and incentives to protect their site. On the other hand, we find that riots increase when LSM moves to the production phase, suggesting that such a move negatively affects the livelihoods of local communities. Finally, when expanding into areas where artisanal miners are active, industrial mining increases violence against civilians and looting. This result suggests that armed actors who previously ‘protected’ and taxed artisanal miners turn to alternative sources of finance, looting and attacking civilians. Another possible channel is the ‘opportunity cost effect’: having lost their livelihood, artisanal miners reallocate their labor to criminal activities. This study makes two contributions. First, we further unravel the complexities of the resource curse. To the best of our knowledge, this is the first study that explicitly analyses the effect of both artisanal and industrial mining – as well as the interaction between both modes of extraction – on local conflict events. In doing so, we contribute to the small but growing literature that uses subnational data to study the relationship between natural resources and conflict (see e.g. Berman et al. 2017; Buhaug and Rød 2006; Dube and Vargas 2013). We specifically add to a recent 125 Chapter 4 literature on ‘conflict minerals’ in the DRC. So far, three papers have used subnational data to analyze the local dynamics between mineral resources and conflict in the DRC (Maystadt et al. 2014; Parker and Vadheim 2017; Sanchez de la Sierra 2016).3 These papers provided important insights which serve as a foundation for our conceptual framework. However, none of the papers makes a distinction between ASM and LSM sites, which is an important lacuna given the current high-profile debate on the ASM-LSM balance in DRC’s mining policy. Furthermore, in contrast to Maystadt et al (2014) and Parker and Vadheim (2017), our analysis does not focus on territories (with an average surface area of almost 18,000 km2), but on much smaller grid cells of 625 km2. We do so because we believe that the dynamics between the different modes of mineral extraction and conflict are best studied at the most local level, near to mineral extraction sites. Further, Parker and Vadheim (2017: p.45) recognize that “a more disaggregated analysis could reveal additional detail about militia activity and movement”. The 25 x 25 kilometer cells that we study are large enough to encompass LSM concessions, while being small enough to allow us to focus on the near vicinity of the extraction sites. Second, we add to the literature on the relation between ASM and LSM in general (Banchirigah 2008; Bush 2009; Carstens and Hilson 2009) and in the DRC in particular (Geenen 2014; Geenen and Claessens 2013; Stoop and Verpoorten mimeo; van Puijenbroeck and Schouten 2013). In the DRC, as well as in many other African countries, the pendulum has been swinging back and forth between the two production modes (Campbell 2009). The pendulum’s movements are influenced by national interests and world mineral prices, but also by the agenda and policy advice of international institutions such as the World Bank and the IMF. Often this advice is based on criteria related to economic efficiency and does not take wider political economy considerations into account. Acemoglu and Robinson (2013: p.173) fittingly state that “politics is largely absent from the scene”. This study brings politics onto the scene by spelling out how the resource curse operates differently for ASM and LSM. 3 Maystadt et al. (2014) exploit monthly variations in world mineral prices between 1997-2007 to show that the granting of industrial mining permits in DRC has no effect on conflict at the territory-level, while it does foster violence at the district level (one administrative unit higher). Sanchez de la Sierra (2016) constructed a yearly panel database for the period 1998-2013, on the formation of monopolies of violence and taxation by armed groups in North- and South- Kivu. The results indicate that a demand shock for coltan (a bulky mineral) leads violent actors to organize monopolies of violence, tax output and provide protection at coltan mines. Instead, a similar shock for gold (which can easily be concealed to avoid taxes) leads violent actors to form such monopolies in the support-villages where gold miners spend their earnings. Parker and Vadheim (2017) investigate the impact of section 1502 of the Dodd-Frank Act, which aims to break the link between minerals and conflict by cutting the funding of warlords. Using a geo-referenced dataset on conflicts and artisanal mining sites for the period 2004-2012, they find that the legislation increased looting of civilians and shifted battles between armed groups from artisanal 3T mining towards relatively unregulated gold mining. 126 Chapter 4 The remainder of the chapter is structured as follows. In the next section, we anchor our study in the Congolese context. Section 4.3 presents the conceptual framework. Sections 4.4 and 4.5 present the data and empirical strategy, respectively. We present the results in Section 4.6. Section 4.7 presents robustness tests, and Section 4.8. concludes. 4.2. Context: Conflict and Mining in Eastern Congo Our study focuses on Eastern Congo, which consists of the eleven provinces highlighted in panel (a) of Figure 1. Eastern Congo was home to the start of the First and Second Congolese Wars (1996-1997 and 1998-2003). The latter, with the direct involvement of eight African nations and 25 armed groups, has been the deadliest war in modern African history (IRC 2007). Despite the formal end to the war in July 2003, much of Eastern Congo continues to experience conflict. In the remainder of this section, we give background information on ASM and LSM activities in this region, and how they relate to conflict. 4.2.1. Artisanal Mining Artisanal mining gradually developed from the 1960s onward and accelerated after 1982, when Mobutu liberalized the exploitation and trade in minerals. During the two Congo wars, when LSM came to a standstill, ASM continued to expand. Presently, ASM is an important livelihood strategy in DRC. The World Bank (2008: p.56) puts employment in the artisanal mining sector in the range of 0.8 to 2 million. Using an average of four to five dependents for each miner, this implies that up to 10 million people, or 16 percent of DRC’s population, are dependent on artisanal mining for their livelihood (World Bank 2008: p.7). The ASM-based livelihood is, however, under pressure. In 2002 and 2003, a new Mining Code and Mining Regulations were developed under the guidance of the World Bank and the IMF. These institutions, as well as the Congolese government, prioritized LSM for tax reasons. 4 Although the Mining Code recognizes ASM as a valid production mode, it specifies that artisanal activities should take place in clearly demarcated Artisanal Exploitation Zones (AEZ). In practice, very few AEZ were created (see Section 4.4.2), and the code provides the opportunity to close them down if “a new deposit which does not lend itself to artisanal mining has been discovered”.5 Consequently, LSM companies have the upper hand, while ASM largely takes place outside the 4 Artisanal miners are largely able to escape taxation. According to the World Bank (2008: p.56): “when compared to official statistics of gold production […], it would appear that more than half of DRC gold production is smuggled out of the country.” 5 DRC mining code 2002, Title 4, Chapter 1, Article 110. 127 Chapter 4 state’s regulatory framework (Geenen 2013; Geenen and Claessens 2013; Kilosho, Stoop, and Verpoorten Forthcoming). Artisanal mining in Eastern Congo does not only involve artisanal miners. In about 56% of the ASM sites, armed actors are present on a permanent or regular basis (Weyns, et al. 2016). The Congolese army (the FARDC) is present in 38% of ASM sites, while in 25% of sites the armed presence consists of various rebel groups and local self-defense militias (“Mai-Mai groups”).6 These armed actors mostly profit from ASM through illegal taxation, but are also known to engage in mineral trade, monopolize the sale of certain commodities (e.g. beer, cigarettes or palm oil), force artisanal miners to work for them, or resort to looting and pillaging (Geenen 2014; Laudati 2013; Seay 2012; Weyns et al., 2016). 4.2.2. Large-Scale Mining In the colonial era, mineral deposits in the DRC were mined industrially. However, LSM came to a standstill due to a combination of Mobutu’s nationalization campaigns in the 1970’s and the two Congo wars (Geenen 2014). The 2002 Mining Code and 2003 Mining Regulations aimed to attract industrial mining companies again by offering an advantageous fiscal regime (Mazalto 2005). In recent years, large-scale mining has increased in importance and has been the major driver of GDP growth in the country (African Economic Outlook 2014). In 2013, mineral rents accounted for an estimated 18.6% of GDP, up from 6.5% in 2006 and 0.2% in 2002 (World Bank 2016). FDI to DRC increased from 1.6% of GDP in 2002 to 11% in 2007 (World Bank 2016). Concurrently, government revenue from the minerals sector increased, reaching a record high of US$1.4 billion in 2011, corresponding to about 10% of total government revenue (EITI 2014). There are two types of LSM permits. Initially, a mining company must obtain a research permit, which gives the holder the right to carry out mineral exploration works. To extract minerals, the company needs to transform its research permit into a production permit. A research permit usually does not entail immediate consequences for artisanal miners and local communities operating and living within the concession’s perimeter. However, in the production phase, local communities may have to be relocated, and artisanal miners may have to cease their activities (e.g. Geenen and Claessens 2013; Kilosho, Stoop, and Verpoorten Forthcoming; van Puijenbroeck and Schouten 2013). Several studies indicate that artisanal miners experience confrontation with mining companies as a war-like situation, and often (violently) defend their ‘customary right’ to 6 For detailed information on armed actors in Eastern Congo see e.g. (Stearns 2013a, 2013b; Stearns and Vogel 2015; Vlassenroot 2013; Vogel and Mvano 2016). 128 Chapter 4 dig for minerals (Geenen and Claessens 2013; Stoop and Verpoorten mimeo; van Puijenbroeck and Schouten 2013). 4.3. Conceptual framework In this section, we first define three key concepts that were introduced in related literature and serve as foundations for our conceptual framework. We then define the different actors in our study area and the possible actions they can undertake. Finally, we discuss two ASM and LSM related shocks observed in the area during our period of study, and hypothesize how they may affect local conflict dynamics. 4.3.1. Previous Literature: Rapacity, Protection and Opportunity Costs The literature on mining and conflict has drawn on three concepts to discern how shocks to the value of resource endowments affect conflict. The rapacity effect states that a rise in the value of minerals leads to increased violence because there are larger spoils to be had from fighting (Grossman 1991; Hirshleifer 1991). In contrast, the opportunity cost effect states that negative shocks to peaceful economic activities (farming, artisanal mining) decrease the opportunity cost of fighting, drawing labor into conflict activities (Chassang and Miquel 2009; Grossman 1999). The reverse holds for positive shocks to peaceful economic activities: they increase the opportunity cost to fight, thus depressing conflict activities. A set of recent papers that have used subnational data to analyze local conflict dynamics makes use of a third concept – the protection effect (Berman et al. 2017; Maystadt et al. 2014; Parker and Vadheim 2017; Sanchez de la Sierra 2016). This concept builds on the (Olson 1993) stationary bandit methapor. The prototypical example is the Mafia, an organization that collects tax in exchange for protection (Skaperdas 2001). In the context of the DRC, various armed groups play this role. Filling the power vacuum left by an absent state, they establish a monopoly on violence and, in return for taxes, offer protection against violence they would have committed themselves and/or against violence committed by others (Parker and Vadheim 2017; Sanchez de la Sierra 2016). These armed groups mainly ‘protect and tax’ artisanal miners or farmers. While they could, in theory, also provide these ‘services’ for LSM companies, in general, this is not the case for our 129 Chapter 4 study area in the period we consider;7 instead, LSM infrastructure and employees are protected by private security forces and by the Congolese army. 4.3.2. Area, Actors and Actions We consider a theoretical area of 25x25 km in eastern DRC. In line with the empirical reality, we allow for two modes of mineral extraction: the area may encompass one or more ASM sites, operated by artisanal miners, and/or one or more LSM concessions, which may either be in the research or the production phase. The area is further inhabited by local communities who engage in other economic activities like farming. Next, we allow for the presence of three types of armed actors: the private security forces of the LSM company, rebel groups and the Congolese army. We assume that the private security forces hired by the LSM company only engage in one activity: protecting the interests of the company. We assume that rebel groups engage in three activities. First, they may act like a stationary bandit and tax artisanal miners in return for protection. Second, they may battle to obtain control of an ASM site to act like a stationary bandit. Third, they may loot and attack artisanal miners and local communities. We assume that the Congolese army also engages in these three activities.8 However, in their capacity as state actors, the FARDC may likewise protect local communities, as well as the interests of the LSM company – since LSM companies are the legal titleholders of a mining concession and receive official backing from the Congolese government.9 Finally, individual artisanal miners as well as other members (farmers) of the local community may riot, e.g. when their livelihood is under threat, and they may also take up arms, e.g. when the opportunity cost to fight declines. 4.3.3. Shocks We focus on two shocks that impacted artisanal and large-scale mining activities in Eastern Congo over the period under study (2004 to 2015).10 The first shock relates to the value of minerals. The main minerals extracted in Eastern Congo (gold and 3T minerals: tin, tungsten and tantalum) strongly increased in value in the period 7 Human Right Watch reports on several symbiotic company-rebel group relationships in earlier periods in the DRC (Human Rights Watch 2005). 8 See also: Stearns et al. (2013); Stearns and Vogel (2015); Weyns et al. (2016). 9 Case studies in South-Kivu suggest that LSM companies can count on armed support of the FARDC (see e.g. Geenen 2013; Geenen and Claessens 2013; Stoop et al. 2016). 10 We focus on the period 2004-2015 to exclude the period of the two Congo wars from our analysis. First, it is unlikely that our data on the location of artisanal and industrial mining sites provides an accurate picture of mining activities during this chaotic period. Second, industrial mining came to a standstill during the two Congo wars and only gradually resumed after a new Mining Code and Mining Regulations were introduced in 2002 and 2003. 130 Chapter 4 that we study. On average, the world prices of gold and 3T minerals tripled (see Figure 2). A troy ounce of gold, for instance, was on average valued at $326 in 2004, while it was worth $1.160 in 2015 in constant 2015 US dollars. This boom in mineral prices is generally explained by an increasing demand in emerging economies, particularly China (Canuto 2014; Humphreys 2010). Mineral traders in Eastern Congo closely monitor world mineral prices, and use them to set local prices (Geenen 2014). The boom thus strongly increased the value of mineral deposits in both the ASM and LSM sites in our study area. Second, in response to the 2002 Mining Code and 2003 Mining Regulations, the number of LSM research and production concessions strongly increased in subsequent years. The total number of granted research permits increased from 221 to 3,368 between 2004 and 2015, while the total number of granted production permits increased from 87 to 332. These two shocks give way to four events: 1) the value of an existing ASM site increases; 2) the value of an existing LSM site increases; 3) LSM obtains a research permit for a mining site; and 4) LSM obtains a production permit for a mining site. 4.3.4. Hypotheses We now hypothesize how these four events may affect local conflict. We consider four types of conflict: battles between armed actors, violence against civilians, riots and looting. First, we consider the impact of an exogenous increase in the value of an ASM site. Let’s call this ‘ASM_value’. Theoretically, ASM_value may trigger opposite effects on local conflict events. On the one hand, it increases incentives for a stationary bandit controlling an ASM site to levy taxes and protect its area. This ‘protection effect’ could reduce violence, or leave it unchanged. Furthermore, ASM_value raises the opportunity costs to fight, reducing the incentives of artisanal miners to shift to criminal activities. This ‘opportunity cost effect’ could thus further contain violence. On the other hand, ASM_value increases the spoils to fighting and may induce contestation of the stationary bandit by other armed actors. This ‘rapacity effect’ predicts that armed actors will intensify or re-allocate their fighting efforts to sites in which the valuable resources are extracted. As argued by Parker and Vadheim (2017), and elsewhere in the literature, the ‘rapacity effect’ is likely to dominate when the stationary bandit and other armed actors are relatively homogenous in size and strength, increasing incentives for other armed actors to contest the stationary bandit (Anderson and MC Chesney 1994; Fearon 1995; Hirshleifer 1991; Ralston 2012; Skaperdas 1992; Umbeck 1981). Such power symmetry applies to Eastern Congo, where artisanal mining sites are controlled by a plethora of relatively small groups of armed actors, with continuously changing allegiances, reorganizations and reincarnations (Stearns and Vogel 2015; 131 Chapter 4 Vogel and Mvano 2016; Weyns et al. 2016). We expect the rapacity effect to increase battles, violence against civilians and looting. With respect to the incidence of riots, we have no a priori reason to expect a significant change. We therefore hypothesize that: H1. An increase in ASM_value significantly increases the incidence of battles, violence against civilians and looting. Second, we consider an exogenous increase in the value of an LSM site’s mineral endowments. Let’s call this ‘LSM_value’. As the potential production value of a site increases, LSM companies may step up their production. However, protection measures are unlikely to fluctuate with mineral prices as they mostly concern employees (to safeguard them from kidnapping) and the company’s infrastructure and equipment (to safeguard them from vandalism and theft). Second, given that LSM companies do not employ a lot of workers, and given that they in any case receive a fixed wage, we do not expect the ‘opportunity cost effect’ to apply either. Finally, given the strong security apparatus of LSM companies, we assume they are not challenged by other armed actors when the value of their mining site increases. Therefore, the ‘rapacity effect’ neither applies. In sum, we do not expect an increase in the value of LSM sites to have a significant impact on local conflict. H2. An increase in LSM_value does not significantly affect local conflict events. Third, we consider how the arrival of a new LSM research concession affects local conflict events. Let’s call this ‘LSM_research’. Case studies have indicated that LSM activities have a limited impact on local communities and artisanal miners during the research phase (Geenen 2014, 2013; Geenen and Claessens 2013; Kilosho, Stoop, and Verpoorten Forthcoming; Stoop, Kilosho, and Verpoorten 2016; van Puijenbroeck and Schouten 2013). Many concessions covered by research permits are inactive, and some are held by companies that never had the intention to explore or extract minerals (so-called ‘dormant titles’) (Geenen and Radley 2014). In addition, during the research phase, while LSM companies are not allowed to extract minerals, artisanal mining is still tolerated and the local population can remain where they are. We therefore conjecture that: H3. LSM_research does not significantly affect conflict events. Fourth, we consider the impact of an LSM company entering the production phase, ‘LSM_production’. In this phase, LSM companies make considerable investments in infrastructure. To safeguard their investments, the company relies on private security forces and the FARDC for protection. Backed by the government, the LSM company and its ‘protectors’ likely outweigh other armed actors in terms of size and strength, thus obtaining a monopoly on violence. As such, we hypothesize a ‘protection effect’: H4. LSM_production significantly decreases the incidence of battles. 132 Chapter 4 The positive externalities of the protection effect by LSM companies may or may not extend to the civilian population. Further, LSM_production may also negatively impact local communities, because its activities may require the relocation of the local population. As a result, some locals may see their opportunity cost to fight decline, for instance, upon the loss of farmland. We thus do not have strong priors on looting and violence against civilians. Finally, with respect to riots, a number of case studies lead us to expect confrontations between the company and local communities (Geenen and Claessens 2013; Kilosho et al. Forthcoming; Stoop and Verpoorten mimeo; van Puijenbroeck and Schouten 2013). LSM concessions generate heightened but oft- unfulfilled expectations in host communities. The discontent can boil over into protests and riots, most likely at the start of the production phase, which is often associated with the dislocation of local communities. We therefore conjecture: H5. LSM_production significantly increases the incidence of riots. LSM research and production activities can also start in an area that encompasses a number of ASM sites (LSMr_in_ASM and LSMp_in_ASM). If, as a result, artisanal mining is halted, there is no longer a stable surplus for the stationary bandit to expropriate. This may lead to the breakdown of an existing ‘protection effect’: armed actors who previously profited from taxing the artisanal activities may look for alternative sources of income, resorting to looting and other forms of violence against civilians. Moreover, the ‘opportunity cost effect’ may add to this, as a lack of alternative economic opportunities may incentivize certain artisanal miners to join an armed group and engage in looting or violence. Based on the above-mentioned case-studies, we hypothesize that these effects only materialize when the company moves from the research to the production phase: H6. LSMr_in_ASM does not significantly affect conflict events. H7. LSMp_in_ASM significantly increases looting and violence against civilians. Table 1 provides an overview of the hypothesized impact of the events on local violence. 4.4. Data sources and description 4.4.1. Data sources To measure local conflict, we build on the Armed Conflict Location and Event Data Project (ACLED). ACLED provides information on the date and location of conflict events in 60 developing countries in Africa and Asia. The dataset has been widely used in recent academic research (Berman et al. 2017; Maystadt et al. 2014; Minoiu and Shemyakina 2014; Parker and Vadheim 2017). The data, however, mainly relies on reports from local and regional news sources, 133 Chapter 4 as well as humanitarian agencies (Raleigh et al. 2010). Worries might thus exist over data quality. For instance, it could be the case that conflict events in very remote or insecure areas get little media coverage, leading to an underestimation of conflict in those areas (Van der Windt and Humphreys 2016; Verpoorten 2012). Besides ACLED, there is one other source of systematically collected conflict data that covers the entire study area for the period 2004-2015: the Uppsala Conflict Data Program Georeferenced Events Dataset (UCDP GED). The UCDP database relies on similar sources of information as ACLED (Sundberg and Melander 2013) and may hence suffer from a similar bias. We prefer to use ACLED, because the UCDP dataset is limited to events that resulted in fatalities, while we are interested in a wider set of conflict events.11 For Eastern Congo, ACLED contains information on the date and location of 6,542 conflict events that occurred between 2004 and 2015. The database allows us to separate out the type of conflict. We focus on four different types. First are battles, which are defined as “a violent interaction between two politically organized armed groups at a particular time and location” (Raleigh and Dowd 2016: p.10), where armed groups include both rebel movements and the FARDC. The database contains 2,748 battle events. Second, we focus on violence against civilians, which occurs when armed groups attack civilians. The database contains 2,487 such events. Third, we consider riots and protests, which are coded as potentially violent public demonstrations by groups. There are 518 riot events. Finally, we use the description of the conflict events to construct a variable that indicates events of looting. We follow Parker and Vadheim (2017) and consider an event as looting if an armed group’s actions are described by the words ‘loot’, ‘pillage’, ‘plunder’, ‘rob’, ‘steal’, ‘ransack’, ‘sack’, or ‘seize’.12 In total, we have information on 718 looting events. Information about artisanal mining sites comes from the International Peace Information Service (IPIS). In a collaboration with the Congolese ministry of mines and other local stakeholders, IPIS research teams equipped with GPS devices and surveys mapped the artisanal mining sites in Eastern Congo.13 The data was collected between 2009 and 2015, and contains the 11 Specifically, the UCDP dataset defines an event as “An incident where armed force was used by an organized actor against another organized actor, or against civilians, resulting in at least 1 direct death at a specific location and a specific date” (Croicu and Sundberg 2016: p.2). Recently, Eck (2012) compared the quality of the ACLED and UCDP data and concluded that "those interested in subnational analyses of conflict should be wary of ACLED's data". However, Eck (2012) mainly criticizes ACLED on the precision of its geocoding which may introduce an urban bias. In our case, this is not an issue since Eck (2012: p.139) acknowledges that “The quality of geocoding depends in large part on the quality of … maps used to identify locales. … in DRC, for example, the International Peace Information Service has provided detailed maps which allow researchers to find small villages.”. 12 Examples of looting events include “Rebels loot 10 houses in Kiwanja and 3 others in Rubare and Kako”; “FDLR rebels established a base for looting gold and cassiterite from the mines at Kasiyiro”; and “Soldiers erected illegal barriers at Mangi and Panga mining sites in Banalia area since the beginning of June, extorting and seizing goods from mine workers and merchants”. The looting events are mostly (73%) coded as violence against civilians by ACLED. 13 Other local stakeholders include: the Congolese Mining Registry (CAMI); the Congolese Public Service for Assistance to Artisanal- and Small-scale Mining (SAESSCAM), the provincial Mining Divisions and local civil society 134 Chapter 4 location and minerals of 2,026 artisanal mining sites. Although the IPIS dataset does not provide information about when the sites were established, almost all sites are thought to have existed before 2004.14 The data and the collection process are described in detail in Spittaels and Hilgert (2013), Spittaels et al. (2014), and Weyns et al. (2016). To learn about large scale mining concessions, we rely on a dataset of the Congolese Mining Registry (CAMI). CAMI is a public entity under the supervision of the Congolese ministry of mines, and oversees the granting and renewal of LSM research and production permits. The CAMI dataset provides information about the geographical boundaries of 3,700 concessions covered by a valid mining permit. The dataset also records the granting date of these permits, and whether the permit is for mineral research (3,368) or production (332). Finally, we obtain time-series data on international mineral prices from metalprices.com. 4.4.2. Descriptive Statistics Our database counts 2,176 grid cells and 144 months. Summary information at the level of the grid cells can be found in Table 2. Figure 3 maps the location of conflict events, ASM sites and LSM concessions. Panel (a) of Figure 3 offers a geographical presentation of our conflict data. Table 2 indicates the share of grid cells that witnessed conflict events over the period of study: battles (13%), violence against civilians (14%), riots (5%) and looting (7%). It further shows the number of months during which these conflict events occurred within a cell: the maxima are 68 for battles, 59 for violence against civilians, 44 for riots and 26 for looting. Since the empirical analysis is based on cell-month observations, Table 2 also indicates the monthly probability of battles (0.5%), violence against civilians (0.5%), riots (0.1%) and looting (0.2%). Panel (b) maps the location of artisanal mining sites. Gold and 3T are by far the most important minerals: 68% of ASM sites contain gold, while 29% contain on one of the 3T minerals.15 About 14% of the grid cells contain ASM sites; 9% have ASM gold mines and 6% have ASM 3T mines. The number of gold mines in a cell varies between zero and fifty, while the number of 3T mines varies between zero and thirty. Only about 2% of cells contain both ASM gold and 3T sites. The IPIS database further provides information on the number of artisanal miners organizations. The data collection and analysis were funded by the World Bank’s PROMINES program, the International Organization for Migration and the Belgian Ministry of Foreign Affairs (Weyns et al. 2016). 14 IPIS believes that the large majority of sites were established before 2004 (personal communication with IPIS, October 2016). Also Sanchez de la Sierra (2016) found that, of the 411 artisanal mining sites in his sample from North and South Kivu, all but one existed before 1995. 15 Only 3% of ASM mines focus on other minerals, mostly diamonds. However, all cells with ASM sites are either dominated by gold or 3T mining. 135 Chapter 4 employed in the mines. It records a total of 307,555 artisanal miners. With four to five dependents for each artisanal miner, this implies that somewhere between 1.2 and 1.6 million people are dependent on ASM in eastern DRC. While this is significantly lower than the (World Bank 2008) estimate (which took into account all mineral-producing regions, not just eastern DRC), it is substantial. Panel (c) of Figure 3 maps the LSM concessions covered by a research or production permit during the period of our study. The majority of LSM permits (80%) cover gold or one of the 3T minerals, while copper (16%) is also important.16 As mentioned in Section 4.3.3, the total number of granted research permits increased from 221 in 2004 to 3,368 in 2015, while the total number of granted production permits increased from 87 to 332. For each cell-month observation, we calculate the share of the cell surface area that is covered by LSM concessions. When considering the entire period of study (2004-2015), 28% of the average cell is covered by research permits, while 2% is covered by production permits (Table 2). The Congolese Mining Code states that ASM can only take place in Artisanal Exploitation Zones (AEZ). However, only 177 AEZ have been created in eastern DRC, covering about 1% of the total mineral concession surface area (see Panel (c) of Figure 3).17 When we combine information from the IPIS and CAMI databases, we find that less than 1% of the artisanal miners registered by IPIS operate in an AEZ. About 29% of the ASM sites registered by IPIS are located on an LSM research concession and about 32% are located on an LSM production concession, while the remaining sites are located in areas that are not covered by a mining permit. It is thus pertinent to not only explore the effect of ASM and LSM in itself, but also of their interaction. 16 On average, the world price of copper also increased over the period of study, but to a much smaller extent than gold and 3T minerals. Appendix A shows the monthly price evolution of copper from 2004 to 2015. The remaining 4% of LSM companies focuses on diamonds. Panel data on diamond prices are not available. 17 Based on information from the CAMI database, we calculate that about 80% of the surface area covered by mineral concessions is covered by research permits, 15% by production permits, 4% by Small Mine Exploitation Permits and 1% by AEZ. 136 Chapter 4 4.5. Empirical strategy and identification We conduct our empirical investigation at the level of 2,176 grid cells of 25x25 kilometers.18 They are illustrated in panel (b) of Figure 1. We work with monthly observations of these cells over the period 2004-2015, adding up to a total of 144 months. We estimate the following equation using a linear probability model: d=>WX , where d=>WX Congo as a whole. We cluster our errors, Nit, at the cell-level to account for serial correlation within cells. Following the conceptual framework, we estimate the impact of two types of shocks: the increase in value of ASM and LSM sites, and the granting of LSM research and production permits. The variable h*i_M:XI7'f captures changes in the value of ASM sites. It equals the monthly world price of the most prevalent artisanally mined mineral (gold, tantalum, tungsten or tin) in cell i. Hence, in a cell where artisanal gold mining dominates, h*i_M:XI7'f equals the monthly price of gold. In a cell where artisanal tantalum mining dominates, h*i_M:XI7'f 19 indicates the monthly price of tantalum. In a cell without ASM sites, h*i_M:XI7'f equals zero. Similarly, to measure changes in the value of LSM sites, we interact the presence of LSM research 18 The 25 x 25 kilometer cells that we study are large enough to encompass LSM concessions, while being small enough to allow us to focus on the near vicinity of the extraction sites. The DRC Mining Code limits the size of an LSM research permit to 400 km2 (DRC Mining Code 2002, Title 3, Chapter 1, Article 53). The average LSM research permit in our sample covers 190 km2, while the average production permit has a surface area of 104 km2. In choosing the size of the cells, we further follow Parker and Vadheim (2017: p.D1) who provide a robustness check of their results using grid cells of approximately 25x25 kilometers. Further spatial disaggregation is not desirable due to potential measurement errors at the very local level in the geo-referenced mining and conflict data. 19 To be precise: h*i_M:XI7',f = q 0=o<>:>; i<>79:X'×Lq,f , where j is the type of mineral. To facilitate interpretation of the results, mineral prices are standardized over the period of study such that the mean equals zero and the standard deviation equals one. Note that the dominant mineral in a cell does not change over time. Since we add cell-fixed effects, the main terms for 0=o<>:>; i<>79:X' drop out of the estimation. Similarly, since we add month-fixed effects, the main terms for Lq,f drop out. All results are highly robust to defining h*i_M:XI7',f as a weighted index, according to the relative importance of the minerals in each cell: h*i_M:XI7',f = q *ℎ:97 i<>79:Xq,'×Lq,f. With a positive correlation of 0.94 (significant at the 99% significance-level), there is little difference between the two definitions of the price index. 137 Chapter 4 and production concessions with the monthly world price of the dominant mineral that companies (plan to) extract. This is captured by the interaction terms k*i_97F7:9eℎ'f ∗ D9 20 k*i_D9=6Ie;<=>'f ∗ D9 The variables k*i_97F7:9eℎ'f and k*i_D9=6Ie;<=>'f indicate, for each cell-month, the share of the cell surface area that is covered by a valid research or production permit. Concessions must pass through the research phase before going to the production phase. An increase in k*i_97F7:9eℎ'f thus indicates the arrival of new LSM research activities in a cell, while an increase in k*i_D9=6Ie;<=>'f indicates a move from the research to the production phase. The variables k*i9_<>_h*i'f and k*iD_<>_h*i'f capture the start of LSM research and production activities in an area within cell i where artisanal miners are active. Specifically, the variables indicate the number of artisanal mining sites encompassed by the LSM research and production concessions. Figure 4 gives an illustration. Identification We are interested in the causal effect of both shocks on local conflict outcomes. However, we did not randomly assign the value of minerals nor the granting of LSM permits. Following a number of recent articles (Berman et al. 2017; Dube and Vargas 2013; Maystadt et al. 2014; Sanchez de la Sierra 2016), we assume that the strong increase in world mineral prices provides us with plausibly exogenous variation in the value of ASM and LSM sites. This assumption rests on three conditions with the third condition relating to the granting of LSM permits. First, local Congolese mineral prices follow international price trends. Fieldwork by Geenen (2014) indicates that local mineral traders in Eastern Congo closely monitor world mineral prices and use them to set local prices.21 However, in July 2010, section 1502 of the Dodd-Frank Act was passed in US legislation. It requires all companies listed on the US stock market to determine the exact origin of minerals sourced from conflict areas and to reveal their supply chains to the US Securities and Exchange Commission. The period after the introduction of this act may provide a notable exception to the transmission from international to local mineral prices. We elaborate in Appendix B, and show that our results remain stable when dropping this period from the analysis. 20 The price interactions for LSM include gold, tin, tantalum, tungsten and copper. Diamonds are excluded since panel data on diamond prices are not available. This does not bias the results, as diamonds are the main LSM mineral in only 0.1% of the grid cells. 21 This is also happens at the very local level: “Even small traders who are based near the mining sites say they regularly check the price online, on their phone, or on TV5 Afrique” (Geenen 2014: p.249). Geenen (2014: p.249) further quotes a local mineral trader stating that “Following the world market price is the least we can do. If you don’t do it, you lose money”. 138 Chapter 4 Second, world mineral prices should not be affected by local conflict events in the DRC. This is reasonable, since the large majority of ASM (68%) and LSM (69%) sites in our sample focus on the production of gold, while the DRC supplies less than 1% of total gold production (The Enough Project 2009). However, one may argue that it does not hold for tantalum since an estimated 15-20% of the total world production of tantalum originates from the DRC (The Enough Project 2009). Cells dominated by tantalum mining represent about 6% of cells with ASM sites and about 2% of cells with LSM concessions. The analysis in Appendix C indicates that our results remain stable when dropping the price interactions for these cells. Third, the coefficients V), V+ and VE are identified through an interaction between the presence of mining sites in a cell and variations in the world price of the extracted minerals. To identify causal effects, the presence of ASM and LSM sites in a cell should not be affected by conflict events. For h*i_M:XI7'f this is not an issue, since the artisanal mining sites in our database existed throughout the period of our study.22 However, the number of LSM research and production permits strongly increased between 2004 and 2015, affecting all LSM-related variables. Since LSM companies may choose to pursue such permits for cells that are relatively conflict-free, we need to be wary of reverse causality. In Appendix D we demonstrate that this concern is unlikely to drive our results. We show descriptive evidence that goes against the idea that industrial mining companies choose to operate in cells that are relatively conflict free. In addition, we show that the results are robust to the inclusion of dynamic conflict lags, effectively controlling for the conflict situation up to one year before the arrival of the LSM production concession. 4.6. Results Table 3 presents our results. First, we explore how changes in the value of mining sites affect local conflict events. The results offer support for hypothesis H1, which posits a ‘rapacity effect’ for ASM sites: h*i_M:XI7'f leads to a significant increase in battles, looting and violence against civilians, as the stationary bandit gets challenged by other armed actors that are relatively homogenous in strength. Specifically, we find that a one standard deviation increase in the world prices of gold and 3T minerals is associated with an increase of 0.4 percentage points in the monthly probability of battles between armed actors. We find similar results for violence against 22 IPIS collected the information between 2009-2015 and believes that the large majority of sites were established before 2004 (see Section 4.1). Exploiting variation in world mineral prices to identify the impact of industrial mining on conflict, Berman et al. (2017) use a similar argument to establish exogeneity in the presence of mining sites; i.e. they focus on a subsample of cells for which the LSM concessions in their sample were active throughout the entire period of their study. 139 Chapter 4 civilians (a 0.4 percentage point increase) and looting (a 0.2 percentage point increase). Compared to the average monthly probability of conflict events in cells with ASM (1.5% for battles, 1.3% for violence against civilians and 0.5% for looting) these estimates imply an increase of 27%, 31% and 40% respectively. These effects are sizable, considering that the increase in world mineral prices of gold and 3T averages about two standard deviations over the period of our study. The monthly probability of riots is not significantly affected by the increase in the value of ASM sites. In line with hypothesis H2, we do not find evidence that changes to the value of LSM research and production concessions are significantly related to local conflict events. Second, we explore how the rise of LSM research and production activities since 2004 affected local conflict events. As hypothesized (H3), we find no evidence of a significant relationship between the expansion of LSM research activities and local conflict. In contrast, column 1 of Table 3 shows that an expansion of LSM production concessions is associated with a significant decrease in the incidence of battles between armed groups. Specifically, a 10 percentage point increase in k*i_D9=6Ie;<=>'f is associated with a 0.26 percentage point decrease in the monthly probability of battles in a cell, which represents a decrease of 33% with respect to the average monthly probability of battles in cells with an LSM production concession (0.8%).23 This finding is in line with the idea that LSM companies, backed by the government, obtain a monopoly on violence which allows them to protect their investment (H4). This does not necessarily mean that the local population is better off. The ‘protection effect’ does not seem to deter armed actors from committing other violent actions in the vicinity of the LSM concession: we find no significant change in the incidence of looting and violence against civilians. In line with hypothesis H5, the results indicate that local populations resist the expansion of LSM production activities. An increase of 10 percentage points in k*i_D9=6Ie;<=>'f raises the monthly probability of riots with 0.12 percentage points. Compared to the average probability of riots in cells with an LSM production concession (0.3%), this represents an increase of 40%. How is local conflict affected when LSM research and production activities expand in areas where artisanal miners are active? In line with hypothesis H6, local conflict events are not significantly affected in the LSM research phase. In contrast, the results in Table 3 show support for H7: when moving to the production phase (k*iD_<>_h*i'f), LSM significantly increases looting and violence against civilians. Specifically, the monthly probability of violence against civilians increases with 0.3 percentage points for every additional ASM site located in a production concession. The result is significant at the 99% significance level and represents an increase of 23 In cells with LSM production activities, the concessions covered by a production permit on average cover 20% of the cell surface area. 140 Chapter 4 25% compared to the average monthly probability of violence against civilians in cells with ASM sites (1.2%). Similarly, we find that the monthly probability of looting increases with 0.2 percentage points for every additional ASM site located in a production concession. This result is significant at the 90% level and represents a 40% increase compared to the average monthly probability of looting in cells with ASM sites (0.5%). These are sizable effects, considering that the number of ASM sites located in a production concession averages 7, with their number ranging between 1 and 48, for cells where ASM and LSM production activities coincide. In sum, we find distinct, even opposite conflict effects by extraction mode. A rise in the value of an ASM increases battles by rapacious rebel groups that challenge each other. Such an effect is entirely absent in the case of LSM. While there are no candidates to challenge the LSM monopoly on violence, its mining monopoly and claim on the land does not go unchallenged, as is testified by the rise in riots when the company moves to the production phase. In addition, in cases where LSM crowds out ASM, our results suggest that the artisanal miners and/or their parasitic stationary bandits may turn to criminal activities, ransacking civilians.24 4.7. Robustness Before we move to the conclusion we show how the results are robust to a large number of checks. The results are presented in Appendix E. The first set of robustness tests relates to the mineral price interactions. We first construct alternative measures of h*i_M:XI7'f , k*i_97F7:9eℎ'f ∗ D9 D9 24 Note that both shocks we study may affect the migration of individuals. Such migration could potentially have a ‘mechanical’ effect on our estimates as there are more / or less individuals around to loot and be looted, and to attack and suffer from attacks. Unfortunately, reliable data on the number of artisanal miners does not exist, so we cannot measure or control for such migration directly. However, we conjecture that the impact is limited. Specifically, both shocks have a potential opposite effect on migration: while higher mineral prices may attract artisanal miners to mining sites, the start of LSM production activities may crowd them out. Yet, for both shocks, our results indicate a significant increase in the incidence of looting and violence against civilians. 141 Chapter 4 presented in Table E.2. In columns 1 to 4 we construct the additional price interactions using the difference in world mineral price with respect to the previous month, which can be positive or negative; in columns 5 to 8, we construct them using the absolute difference in world mineral price with respect to the previous month. All the month-to-month changes yield zero-coefficients. The results hence seem to be driven by price changes over a longer period of time. As a third check, we therefore collapse our dataset to cell-year observations. The dependent variables now count the number of conflict events in cell < and year ;, while the other variables represent the yearly average of their baseline counterparts. The results are presented in Table E.3 and are again in line with the baseline findings. Finally, when including the LSM price interactions in the regression, the variables k*i_97F7:9eℎ'f and k*i_D9=6Ie;<=>'f measure the impact of LSM expansions with mineral prices at their average levels.25 The results in Table E.4 indicate that the results remain stable when the LSM price interactions are excluded. The second set of robustness checks tweaks our conflict measure. Conflict may not only persist through time, but also across space. In response, we re-run our baseline regressions while including lags for the incidence of conflict events in previous months and in adjacent grid cells. For each cell-month observation, we create a variable that indicates the number of adjacent cells that witnessed any of the conflict events in that month. In the regressions, we include contemporaneous conflict incidence in adjacent cells, as well as a one-month lag.26 We further include dynamic conflict lags up to six months. The results are presented in Table E.5, and are in line with the baseline findings. Finally, we control for rainfall as a proxy for exogenous shocks to agricultural income. In doing so, we follow a number of recent papers (Maystadt et al. 2014; Miguel et al. 2004; Parker and Vadheim 2017). On the one hand, an increase in agricultural income may raise the opportunity cost to join armed groups; on the other hand, it may increase armed groups' incentives to loot farmer communities. Heavy rainfall could also hinder mining activities and the movement of armed groups. We use monthly rainfall data from the Climatic Research Unit of the University of East Anglia to calculate contemporaneous and lagged rainfall anomalies (that measure the monthly deviation from the long-term monthly mean) and to construct indicators for cell-specific dry and 25 Recall that mineral prices are standardized over the period of study such that the mean equals zero and the standard deviation equals one. 26 The results are robust to adding up to 12-month spatial lags, but only the first lag was found to be statistically significant. Introducing spatial lags gives rise to a simultaneity or reflection problem, since it is unclear if conflict in a specific cell is driven by conflict in adjacent cells, or the other way around (Anselin 2002; Manski 1993). A positive correlation of conflict across adjacent cells would overstate the estimated coefficients on the spatial conflict lags. The coefficients on spatial lags may thus be estimated with bias, but we only introduce them to check the robustness of our b coefficients. 142 Chapter 4 wet seasons. Details are provided in section 3 of Appendix E. The results are presented in Table E.6, and confirm the baseline findings. 4.8. Conclusion This chapter uses spatially disaggregated data to study the impact of artisanal and industrial mining, as well as their interaction, on local conflict. We find that an exogenous rise in the value of ASM sites leads to increases in battles, attacks against civiilans and looting. These increases can be explained by the dominance of the rapacity effect over the protection and opportunity cost effect. The entry of LSM, on the other hand, poses a negative shock – not only to the value of ASM activities, but also for local communities who may have to relocate. Consistent with the fact that this negative shock does not materialize during the LSM research phase, we do not find any effect of LSM research activities on local conflict. However, when LSM moves to the production phase, we find that riots increase. In cases where LSM production activities crowd out ASM, we further find an increase in attacks against civilians and looting. We argue that individuals who previously profited from ASM, as artisanal miners or armed actors, turn to alternative sources of finance (by looting and attacking civilians). Finally, we find that battles decrease as LSM production activities expand. This is consistent with the idea that LSM has the means and incentives to establish a monopoly on power. Previous studies have also found a link between mining and local conflict, but could not credibly attribute the effect to ASM or LSM (see e.g. Berman et al. 2017). Our study sheds light on this blind spot, demonstrating that ASM and LSM, as well as their interaction, have distinct and sometimes opposite effects on the level and type of violence. While we show this for the case of Eastern Congo, the conceptual framework we developed relies on largely generalizable assumptions, and therefore has considerable external validity. Whether in DRC or elsewhere, the results have to be interpreted through a political economy lens. In the case of the DRC, the government is not a neutral bystander. It has chosen to back up the power monopoly of LSM, and not support security around ASM sites. This choice has far-reaching consequences, not only directly and economically for the millions of Congolese whose livelihoods depend on artisanal mining, but also indirectly and politically, in that it changes the distribution of natural resource rents in a direction benefiting already powerful groups. In particular, the promotion of LSM shifts power to the political elite in Kinshasa, away from local elites and artisanal mining communities in eastern DRC. In the words of Acemoglu and Robinson (2013: p.189) “well-intentioned economic policies might tilt the balance of political power even further in favor of 143 Chapter 4 dominant groups, creating significant adverse consequences for future political equilibria.” Such concentration of power is bad news for the ‘resource curse’, which depends more on the political institutions and governance of minerals, rather than on the intrinsic characteristics of minerals or whether the extraction is organized efficiently from an economic point of view (Acemoglu and Robinson 2013: p.179–81). At first sight, our results – in particular the relation between mineral price increases and local conflict at ASM sites – may add to the arguments of those who seek to replace ASM by LSM. However, when, looking through the political economy lens, this conclusion is recognized as short-sighted. First, LSM could fuel the national resource curse by further enriching and empowering a small national elite. Second, the association of ASM and local conflict stems from a conscious policy decision to only secure LSM sites, not ASM sites. Political will aside, there is nothing inherent about ASM sites that prevents the same type of security. In view of these considerations, we end this chapter with the warning that the results of the local resource curse literature cannot be interpreted without taking into account the broader political economy and resource curse at the national level. 144 Chapter 4 References Acemoglu, Daron, and James Robinson. 2013. “Economics versus Politics: Pitfalls of Policy Advice.” Journal of Economic Perspectives 27 (2): 173–92. African Economic Outlook. 2014. “Congo, Democratic Republic.” African Economic Outlook. Anderson, Terry L., and Fred S. MC Chesney. 1994. “Raid or Trade? An Economic Model of Indian-White Relations.” Journal of Law and Economics XXXVII (April). Angrist, Joshua D., and Adriana D. Kugler. 2008. “Rural Windfall or a New Resource Curse? Coca, Income, and Civil Conflict in Colombia.” Review of Economics and Statistics 90 (2): 191–215. Anselin, Luc. 2002. “Under the Hood Issues in the Specification and Interpretation of Spatial Regression Models.” Agricultural Economics 27 (3): 247–67. Autesserre, Séverine. 2010. The Trouble with the Congo. Local Violence and the Failure of International Peacebuilding. Cambridge: Cambridge University Press. ———. 2012. “Dangerous Tales: Dominant Narratives on the Congo and Their Unintended Consequences.” African Affairs 111 (443): 202–22. Banchirigah, Sadia Mohammed. 2008. “Challenges with Eradicating Illegal Mining in Ghana: A Perspective from the Grassroots.” Resources Policy 33 (1): 29–38. Berman, Nicolas, Mathieu Couttenier, Dominic Rohner, and Mathias Thoenig. 2017. “This Mine Is Mine! How Minerals Fuel Conflicts in Africa.” American Economic Review 107 (6): 1564– 1610. Buhaug, Halvard, and Jan Ketil Rød. 2006. “Local Determinants of African Civil Wars, 1970– 2001.” Political Geography 25 (3): 315–35. Bulte, Erwin H., Richard Damania, and Robert T. Deacon. 2005. “Resource Intensity, Institutions, and Development.” World Development 33 (7): 1029–44. Bush, Ray. 2009. “‘Soon There Will Be No-One Left to Take the Corpses to the Morgue’: Accumulation and Abjection in Ghana’s Mining Communities.” Resources Policy, Small- Scale Mining, Poverty and Development in Sub-Saharan Africa, 34 (1): 57–63. Campbell, Bonnie, ed. 2009. Mining in Africa: Regulation and Development. New York: Pluto Press. Canuto, Otaviano. 2014. “The Commodity Super Cycle: Is This Time Different?” The World Bank. Carisch, Enrico. 2012. “Conflict Gold to Criminal Gold: The New Face of Artisanal Gold Mining in Congo.” Southern Africa Resource Watch. Carstens, Johanna, and Gavin Hilson. 2009. “Mining, Grievance and Conflict in Rural Tanzania.” International Development Planning Review 31 (3): 301–26. CASM. 2009. “Mining Together. Large-Scale Mining Meets Artisanal Mining: A Guide for Action.” Communities and Small-scale Mining. Chassang, Sylvain, and Gerard Padró i Miquel. 2009. “Economic Shocks and Civil War.” Quarterly Journal of Political Science 4 (3): 211–28. Collier, Paul, and Anke Hoeffler. 1998. “On Economic Causes of Civil War.” Oxford Economic Papers 50 (4): 563–73. ———. 2004. “Greed and Grievance in Civil War.” Oxford Economic Papers 56 (4): 563–95. Croicu, Mihai, and Ralph Sundberg. 2016. “UCDP GED Codebook Version 5.0.” Department of Peace and Conflict Research, Uppsala University. Cuvelier, Jeroen, Steven Van Bockstael, Koen Vlassenroot, and Claude Iguma. 2014. “Analyzing the Impact of the Dodd-Frank Act on Congolese Livelihoods.” SSRC Conflict Prevention and Peace Forum. Dube, Oeindrila, and Juan F. Vargas. 2013. “Commodity Price Shocks and Civil Conflict: Evidence from Colombia.” The Review of Economic Studies 80 (4): 1384–1421. 145 Chapter 4 Eck, K. 2012. “In Data We Trust? A Comparison of UCDP GED and ACLED Conflict Events Datasets.” Cooperation and Conflict 47 (1): 124–41. EITI. 2014. “Generating ‘Ripple Effects’ in DR Congo.” Extractive Industries Transparency Initiative. January 24. http://eiti.org/news/generating-ripple-effects-dr-congo. Fearon, James D. 1995. “Rationalist Explanations for War.” International Organization 49 (3): 379– 414. ———. 2005. “Primary Commodity Exports and Civil War.” Journal of Conflict Resolution 49 (4): 483–507. Fearon, James D., and David D. Laitin. 2003. “Ethnicity, Insurgency, and Civil War.” American Political Science Review 97 (1): 75–90. Geenen, Sara. 2012. “A Dangerous Bet: The Challenges of Formalizing Artisanal Mining in DRC.” Resources Policy 37 (3): 322–30. ———. 2013. “Dispossession, Displacement and Resistance: Artisanal Miners in a Gold Concession in South-Kivu, Democratic Republic of Congo.” Resources Policy. ———. 2014. “‘Qui Cherche, Trouve’ the Political Economy of Access to Gold Mining and Trade in South Kivu, DRC.” Geenen, Sara, and Klara Claessens. 2013. “Disputed Access to the Gold Sites in Luhwindja, Eastern DRC.” Journal of Modern African Studies 51: 85–108. Geenen, Sara, and Ben Radley. 2014. “In the Face of Reform: What Future for ASM in the Eastern DRC?” Futures, “The Futures of Small-Scale Mining in Sub-Saharan Africa,” 62, Part A (October): 58–66. Grossman, Herschell I. 1991. “A General Equilibrium Model of Insurrections.” The American Economic Review 81 (4): 912–21. ———. 1999. “Kleptocracy and Revolutions.” Oxford Economic Papers 51 (2): 267–83. Hirshleifer, Jack. 1991. “The Technology of Conflict as an Economic Activity.” The American Economic Review 81 (2): 130–34. Human Rights Watch. 2005. “The Curse of Gold. Democratic Republic of Congo.” New York: Human Rights Watch. Humphreys, David. 2010. “The Great Metals Boom: A Retrospective.” Resources Policy 35 (1): 1– 13. Humphreys, Macartan. 2005. “Natural Resources, Conflict, and Conflict Resolution Uncovering the Mechanisms.” Journal of Conflict Resolution 49 (4): 508–37. ILO. 1999. “Social and Labour Issues in Small-Scale Mines.” Geneva: International Labour Organization. IRC. 2007. “Mortality in the Democratic Republic of Congo: An Ongoing Crisis.” https://www.rescue.org/report/mortality-democratic-republic-congo-ongoing-crisis. Kilosho, Janvier, Nik Stoop, and Marijke Verpoorten. Forthcoming. “Defusing the Social Minefield of Gold Sites in Kamituga, South Kivu. from Legal Pluralism to the Re- Making of Institutions?” Resources Policy. Koning, Ruben de. 2011. Conflict Minerals in the Democratic Republic of the Congo: Aligning Trade and Security Interventions. SIPRI Policy Paper, no. 27. Solna, Sweden: Stockholm International Peace Research Institute. Laudati, Ann. 2013. “Beyond Minerals: Broadening ‘Economies of Violence’ in Eastern Democratic Republic of Congo.” Review of African Political Economy 40 (135): 32–50. Lujala, Päivi. 2010. “The Spoils of Nature: Armed Civil Conflict and Rebel Access to Natural Resources.” Journal of Peace Research 47 (1): 15–28. Manski, Charles F. 1993. “Identification of Endogenous Social Effects: The Reflection Problem.” The Review of Economic Studies 60 (3): 531–42. Maystadt, Jean-François, Giacomo De Luca, Petros G. Sekeris, and John Ulimwengu. 2014. “Mineral Resources and Conflicts in DRC: A Case of Ecological Fallacy?” Oxford Economic Papers 66: 721–49. 146 Chapter 4 Mazalto, M. 2005. “La Réforme Des Législations Minières En Afrique Et Le Rôle Des Institutions Financières Internationales: La République Démocratique Du Congo.” In L’Afrique Des Grands Lacs Annuaire 2004-2005, edited by Stefaan Marysse and Filip Reyntjens, 263–87. Paris: L’Harmattan. Miguel, Edward, Shanker Satyanath, and Ernest Sergenti. 2004. “Economic Shocks and Civil Conflict: An Instrumental Variables Approach.” Journal of Political Economy 112 (4): 725– 53. Minoiu, Camelia, and Olga N. Shemyakina. 2014. “Armed Conflict, Household Victimization, and Child Health in Côte d’Ivoire.” Journal of Development Economics 108 (May): 237–55. Nickell, Stephen. 1981. “Biases in Dynamic Models with Fixed Effects.” Econometrica 49 (6): 1417–26. Olson, Mancur. 1993. “Dictatorship, Democracy, and Development.” American Political Science Review 87 (3): 567–76. Parker, Dominic P., Jeremy D. Foltz, and David Elsea. 2016. “Unintended Consequences of Economic Sanctions for Human Rights. Conflict Minerals and Infant Mortality in the Democratic Republic of the Congo.” WIDER Working Paper 2016/124. UNU-WIDER. Parker, Dominic P., and Bryan Vadheim. 2017. “Resource Cursed or Policy Cursed? US Regulation of Conflict Minerals and Violence in the Congo.” Journal of the Association of Environmental and Resource Economists 4 (1): 1–49. Ploeg, Frederick van der. 2011. “Natural Resources: Curse or Blessing?” Journal of Economic Literature 49 (2): 366–420. doi:10.1257/jel.49.2.366. Prendergast, John. 2009. “Can You Hear Congo Now? Cell Phones, Conflict Minerals, and the Worst Sexual Violence in the World.” The Enough Project. Puijenbroeck, J van, and P. Schouten. 2013. “Le 6ième Chantier? L’économie Politique de L’exploitation Auifière Artisanale et Le Sous-Développement En Ituri.” L’Afrique Des Grands Lacs. Annuaire 2012-2013. Raleigh, Clionadh, and Caitriona Dowd. 2016. “Armed Conflict Location and Event Data Project (ACLED) Codebook 2016.” ACLED. Raleigh, Clionadh, Andrew Linke, Håvard Hegre, and Joakim Karlsen. 2010. “Introducing Acled.” Journal of Peace Research 47 (5): 651–60. Ralston, Laura. 2012. “Less Guns, More Violence: Evidence from Disarmament in Uganda.” Unpublished paper. Boston, MA: Massachusetts Institute of Technology. Reyntjens, Filip. 2010. The Great African War: Congo and Regional Geopolitics, 1996 - 2006. Cambridge: Cambridge Univ. Press. Roodman, David. 2009. “How to Do xtabond2: An Introduction to Difference and System GMM in Stata.” The Stata Journal 9 (1): 86–136. Ross, Michael L. 2004. “What Do We Know About Natural Resources and Civil War?” Journal of Peace Research 41 (3): 337–56. Sanchez de la Sierra, Raul. 2016. “On the Origins of the State: Stationary Bandits and Taxation in Eastern Congo.” Seay, L. 2012. “What’s Wrong with Dodd-Frank 1502? Conflict Minerals, Civilian Livelihoods, and the Unintended Consequences of Western Advocacy.” 284. Working Paper. Center for Global Development. Skaperdas, Stergios. 1992. “Cooperation, Conflict, and Power in the Absence of Property Rights.” The American Economic Review 82 (4): 720–39. ———. 2001. “The Political Economy of Organized Crime: Providing Protection When the State Does Not.” Economics of Governance 2 (3): 173–202. Spittaels, Steven, and Filip Hilgert. 2013. “Analysis of the Interactive Map of Artisanal Mining Areas in Eastern DR Congo.” Antwerp: IPIS. 147 Chapter 4 Spittaels, Steven, Ken Matthysen, Yannick Weyns, Filip Hilgert, and Anna Bulzomi. 2014. “Analysis of the Interactive Map of Artisanal Mining Areas in Eastern DR Congo: May 2014 Update.” Antwerp: IPIS. Stearns, Jason. 2011. Dancing in the Glory of Monsters: The Collapse of the Congo and the Great War. New York: Public Affairs. ———. 2013a. Banyamulenge: Insurgency and Exclusion in the Mountains of South Kivu. Usalama Project. London: Rift Valley Institute. ———. 2013b. “Raia Mutomboki. the Flawed Peace Process in the Drc and the Birth of an Armed Franchise.” Usalama Project. Rift Valley Institute. Stearns, Jason, Judith Verweijen, and Maria Eriksson Baaz. 2013. “The National Army and Armed Groups in the Eastern Congo: Untangling the Gordian Knot of Insecurity.” Rift Valley Institute - Usalama Project. Stearns, Jason, and Christoph Vogel. 2015. “The Landscape of Armed Groups in the Eastern Congo.” Congo Research Group, Center on International Cooperation. Stoop, Nik, Janvier Kilosho, and Marijke Verpoorten. 2016. “Relocation, Reorientation, or Confrontation? Insights from a Representative Survey Among Artisanal Miners in Kamituga, South-Kivu.” IOB Working Paper 2016.09. Institute of Development Policy and Management, University of Antwerp. Stoop, Nik, and Marijke Verpoorten. mimeo. “Would You Rebel? An Inquiry Among High-Risk Youth in Eastern DRC.” Stoop, Nik, Marijke Verpoorten, and Peter Van der Windt. mimeo. “More Legislation, More Violence? The Impact of Dodd-Frank in the DRC.” Sundberg, Ralph, and Erik Melander. 2013. “Introducing the UCDP Georeferenced Event Dataset.” Journal of Peace Research 50 (4): 523–32. The Enough Project. 2009. “A Comprehensive Approach to Congo’s Conflict Minerals.” The Enough Project. Umbeck, John. 1981. “Might Makes Rights: A Theory of the Formation and Initial Distribution of Property Rights.” Economic Inquiry 19 (1): 38–59. UN security council. 2011. “Letter Dated 29 November 2011 from the Chair of the Security Council Committee Established Pursuant to Resolution 1533 (2004) Concerning the Democratic Republic of the Congo Addressed to the President of the Security Council.” United Nations Security Council. ———. 2012. “Letter Dated 12 November 2012 from the Chair of the Security Council Committee Established Pursuant to Resolution 1533 (2004) Concerning the Democratic Republic of the Congo Addressed to the President of the Security Council.” United Nations Security Council. UNEP. 2011. “Post-Conflict Environmental Assessment of the Democratic Republic of Congo: Synthesis Report for Policy Makers.” Nairobi, Kenya: United Nations Environment Programme. Van der Windt, Peter, and Macartan Humphreys. 2016. “Crowdseeding in Eastern Congo: Using Cell Phones to Collect Conflict Events Data in Real Time.” Journal of Conflict Resolution 60 (4): 748–81. Verpoorten, Marijke. 2012. “Detecting Hidden Violence: The Spatial Distribution of Excess Mortality in Rwanda.” Political Geography 31 (1): 44–56. Vlassenroot, Koen. 2013. “South Kivu. Identity, Territory, and Power in the Eastern Congo.” Usalama Project. Rift Valley Institute. Vogel, Christoph, and Chrispin Mvano. 2016. “The Dogged Persistance of the FDLR.” Congo Research Group. March 4. http://congoresearchgroup.org/guest-blog-the-dogged- persistence-of-the-fdlr/. Weyns, Yannick, Lotte Hoex, and Ken Matthysen. 2016. “Analysis of the Interactive Map of Artisanal Mining Areas in Eastern DR Congo. 2015 Update.” Antwerp: IPIS. 148 Chapter 4 Wimmer, Sarah Zingg, and Filip Hilgert. 2011. “Bisie. A One-Year Snapshot of the DRC’s Principal Cassiterite Mine.” International Peace Information Service. World Bank. 2008. “Democratic Republic of Congo - Growth with Governance in the Mining Sector.” 43402–ZR. World Bank. ———. 2016. “World Development Indicators.” World DataBank. http://databank.worldbank.org. 149 Chapter 4 Figures Figure 1: Eastern DRC Panel (a): Provinces Panel (b): Grid cells Notes: Panel (a) shows the 26 provinces of the DRC. The shaded area covers the eleven eastern provinces which encompass the former province of Orientale (now: Bas-Uele, Haut-Uele, Tshopo and Ituri), North- Kivu, South-Kivu, Maniema and the former province of Katanga (now: Tanganyka, Haut-Lomami, Lualaba and Haut-Katanga). Panel (b) shows the 2,176 grid cells of 25 x 25 kilometer that we take as our units of analysis. 150 Chapter 4 Figure 2: World prices of gold and 3T minerals Notes: These graphs show the monthly averages of world prices for gold, tin, tantalum and tungsten for the period 2004-2015 in 2015 US dollars. Gold prices are reported in US dollars per troy ounce. 3T prices are reported in US dollars per pound. The data was obtained from metalprices.com. 151 Chapter 4 Figure 3: Location of conflict events, ASM sites and LSM concessions Panel (a): Conflict events (ACLED) Panel (b): Artisanal mining sites (IPIS) Panel (c): Mining concessions (CAMI) Notes: Panel (a) shows the location of ACLED conflict events that occurred between 2004-2015. Looting events are not shown, as they are a sub-category of the other events, mostly violence against civilians. Panel (b) shows the location of artisanal mining sites in the most recent database of IPIS. Panel (c) shows the location of the large-scale mining concessions in the CAMI database that were covered by a valid permit at the end of 2015. 152 Chapter 4 Figure 4: Example of a grid cell Panel (a) Panel (b) Panel (c) Panel (d) Notes: The squares in Panel (a)-(d) represent an example of a grid cell in our analysis. The red dots denote the location of ASM sites, yellow areas indicate concessions covered by an LSM research permit, while green areas indicate concessions covered by an LSM production permit. In Panel (a), the cell counts three ASM sites, and no LSM concessions. Moving to Panel (b), an LSM research permit is added, which covers about ¼ of the surface area and encompasses two ASM sites. k*i_97F7:9eℎ'f hence increases from 0 to 0.25 and k*i9_<>_h*i'f from 0 to 2. Moving to Panel (c), an additional research permit is added, further increasing k*i_97F7:9eℎ'f . In Panel (d), the first research permit moves to the production phase, increasing k*i_D9=6Ie;<=>'f from 0 to 0.25 and k*iD_<>_h*i'f from 0 to 2. In this example k*i_97F7:9eℎ'f and k*i9_<>_h*i'f decrease when moving from Panel (c) to Panel (d). Note, however, that multiple research and/or production permits may be added in a cell from one month to the next, that may or may not encompass ASM sites. We find a rather weak negative correlation between k*i_97F7:9eℎ'f and k*i_D9=6Ie;<=>'f (-0.08), while k*i9_<>_h*i'f and k*iD_<>_h*i'f are positively correlated (0.06). 153 Chapter 4 Tables Table 1: Overview of hypothesized impacts Looting Shock Event Battles & Riots Description violence Rapacity effect > opportunity cost & Increase in ASM_value + + protection effect world mineral prices LSM_value 0 0 0 No effect LSM_research 0 0 0 No effect LSMr_in_ASM 0 0 0 No effect Expansion Protection effect, discontent local LSM permits LSM_production - + community Breakdown of existing protection LSMp_in_ASM + effect, opportunity cost effect 154 Chapter 4 Table 2: Summary statistics Obs. Mean Std. Dev. Min Max incidence of battles 2,176 0.13 0.34 0 1 incidence of violence 2,176 0.14 0.35 0 1 incidence of riots 2,176 0.05 0.21 0 1 incidence of looting 2,176 0.07 0.26 0 1 # of months with battle 2,176 0.66 3.77 0 68 # of months with violence 2,176 0.67 3.63 0 59 # of months with riots 2,176 0.18 1.65 0 44 # of months with looting 2,176 0.23 1.46 0 26 monthly incidence of battles 313,344 0.005 0.067 0 1 monthly incidence of violence 313,344 0.005 0.068 0 1 monthly incidence of riots 313,344 0.001 0.035 0 1 monthly incidence of looting 313,344 0.002 0.040 0 1 ASM site indicator 2,176 0.14 0.35 0 1 ASM gold indicator 2,176 0.09 0.29 0 1 ASM 3T indicator 2,176 0.06 0.24 0 1 # ASM sites 2,176 0.90 3.65 0 50 # ASM gold sites 2,176 0.64 3.09 0 50 # ASM 3T sites 2,176 0.27 1.60 0 30 share LSM research concessions 313,344 0.28 0.32 0 1 share LSM production concessions 313,344 0.02 0.09 0 1 Notes: This Table shows summary statistics at the level of the grid cell. Information on conflict events was calculated from the ACLED database; information on ASM sites was calculated from the IPIS database; information on LSM concessions was calculated from the CAMI database. 155 Chapter 4 Table 3: Results Battles violence riots looting (1) (2) (3) (4) ASM_value 0.004** 0.004*** 0.001 0.002** (0.002) (0.001) (0.000) (0.001) LSM_research * price 0.002 0.001 -0.000 0.001 (0.001) (0.001) (0.000) (0.000) LSM_production * price 0.003 -0.006 0.002 0.003 (0.004) (0.004) (0.003) (0.003) LSM_research -0.000 -0.000 -0.000 0.000 (0.001) (0.001) (0.000) (0.000) LSM_production -0.026*** 0.003 0.012** 0.002 (0.009) (0.005) (0.006) (0.004) LSMr_in_ASM 0.000 -0.000 -0.000 0.000 (0.000) (0.000) (0.000) (0.000) LSMp_in_ASM -0.000 0.003*** 0.001 0.002* (0.001) (0.001) (0.001) (0.001) Observations 313,344 313,344 313,344 313,344 R2 0.146 0.134 0.102 0.059 R2 (within) 0.002 0.003 0.002 0.002 Cell FE Yes Yes Yes Yes month FE Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using a Linear Probability Model; Standard errors are clustered at the level of the cell. 156 Chapter 4 Appendix A: World price of copper Notes: This graph shows the monthly average of the world price for copper for the period 2004-2015 in 2015 US dollars. The prices are reported in US dollars per pound. The data was obtained from metalprices.com. 157 Chapter 4 Appendix B: Excluding the Dodd-Frank period The Dodd-Frank period may provide a notable exception to the transmission from international to local mineral prices. In July 2010, section 1502 of the Dodd-Frank Act was passed in US legislation. It requires all companies listed on the US stock market to determine the exact origin of minerals sourced from conflict areas and to reveal their supply chains to the US Securities and Exchange Commission. The goal was to end the illegal exploitation of minerals, to the interest of ending the ongoing conflict in DRC. In practice, the act created a de-facto embargo on artisanal mining when two global coalitions of major electronic companies stopped buying minerals from smelters who couldn’t prove that they did not source minerals that fund conflict in the DRC (Cuvelier et al. 2014; Seay 2012; Wimmer and Hilgert 2011).27 Although artisanal mining communities were affected by the embargo28, mineral trade did not stop entirely. First, minerals were smuggled across the DRC’s eastern borders (UN security council 2011). This was especially the case for gold, which is easy to conceal and for which most of the production was already smuggled out of the country before the introduction of Dodd-Frank (Koning 2011; World Bank 2008). Second, Chinese buyers, who were not affected by the Dodd- Frank act, continued to export 3T minerals from the DRC (UN security council 2011: p.105). Research by the Southern Africa Research Watch does indicate that buyers took advantage of the situation to buy minerals at heavily discounted prices from artisanal miners (Carisch 2012: p.15). The impact of Dodd-Frank on local conflict events is analysed in detail by Parker and Vadheim (2017) and Stoop et al. (mimeo). Here, we simply drop the Dodd-Frank period to account for the fact that world mineral prices may not have fully transmited to local prices during this period. Specifically, we drop all observations between July 2010 and December 2012. We end the Dodd-Frank period in December 2012, as several mining sites in North-Kivu, South-Kivu and Maniema had been validated ‘conflict free’ by that time. Moreover, by the end of 2012 local traders identified alternative markets for untagged minerals, allowing them to resume export (UN security council 2012).29 The results, presented in Table B.1, confirm our results. The coefficient estimates 27 The two coalitions are the Electronic Industry Citizenship Coalition (EICC - which includes a.o. Apple, HP, Dell and Microsoft), and the Global e-Sustainable Initiative (GeSI – which includes a.o. Motorola and Nokia). Because of this decision, the Malaysia Smelting Corporation (MSC), which previously purchased up to 80% of eastern Congolese tin, stopped sourcing minerals from the DRC. 28 Qualitative evidence suggests that people in mining communities could no longer afford to visit healthcare facilities or pay for their children’s schooling; moreover, the economic effects where felt throughout the eastern provinces as artisanal miners could no longer afford to pay for goods, services and agricultural products (Cuvelier et al. 2014; Geenen 2012; Seay 2012; Wimmer and Hilgert 2011). Using quantitative data, Parker et al. (2016) further find that the probability of infant deaths increased by at least 143% in villages near artisanal mines targeted by the Dodd-Frank act. 29 Especially see pages 47, 48 and 193 of the report. For instance: “By 16 august 2012, thirteen tantalum smelters and refiners had been awarded ‘conflict free’ status” (UN security council 2012: p.193); “In July 2012, the Minister of Mines authorized all export houses, including Huaying and TTT/CCM, to export minerals that they purchased from Maniema … The provincial Minister 158 Chapter 4 for h*i_M:XI7'f on battles, violence against civilians and riots are slightly larger, implying that the impact of world price-induced increases to the value of ASM sites was somewhat smaller during the Dodd-Frank period. of Mines in North-Kivu extended the provision to also include validated mines in Masisi in a subsequent letter. Consequently, by the end of August 2012, Huaying had exported … for a total of 248 tons of tin ore, up to and including 24 September 2012 … TTT/CCM officially exported 86 tons of tin ore” (UN security council 2012: p.47); “In North Kivu, the export house AMR Mugote has lawfully exported minerals purchased from ‘green’ mine sites in Masisi” (UN security council 2012: p.47). 159 Chapter 4 Table B.1: Excluding the Dodd-Frank period battles violence riots looting (1) (2) (3) (4) ASM_value 0.006*** 0.004** 0.001* 0.002** (0.002) (0.002) (0.001) (0.001) LSM_research * price 0.002 0.001 -0.000 0.000 (0.001) (0.001) (0.000) (0.001) LSM_production * price 0.005 -0.007 0.001 0.002 (0.006) (0.007) (0.003) (0.002) LSM_research -0.000 0.000 -0.000 0.000 (0.001) (0.001) (0.000) (0.000) LSM_production -0.028*** -0.000 0.014** 0.002 (0.009) (0.005) (0.006) (0.004) LSMr_in_ASM 0.000 -0.000 -0.000 -0.000 (0.000) (0.000) (0.000) (0.000) LSMp_in_ASM 0.000 0.004*** 0.001 0.002** (0.001) (0.001) (0.001) (0.001) Observations 254,592 254,592 254,592 254,592 R2 0.135 0.123 0.096 0.050 R2 (within) 0.003 0.004 0.003 0.002 Cell FE Yes Yes Yes Yes Month FE Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using a Linear Probability Model; Standard errors are clustered at the level of the cell; All observations during the Dodd-Frank period (July 2010 - December 2012) are dropped. 160 Chapter 4 Appendix C: Excluding tantalum The assumption that world mineral prices are not affected by local conflict events in Eastern Congo may not hold for tantalum, since an estimated 15-20% of the total world production of tantalum originates from the DRC (The Enough Project 2009). Cells dominated by tantalum mining represent about 6% of cells with ASM sites and about 2% of cells with LSM concessions. The results in Table C.1 indicate that our results remain stable when dropping the price interactions for these cells. Table C.1: Excluding tantalum battles violence riots looting (1) (2) (3) (4) ASM_value 0.004** 0.003*** 0.000 0.002** (0.002) (0.001) (0.000) (0.001) LSM_research * price 0.002 0.001 -0.000 0.001 (0.001) (0.001) (0.000) (0.000) LSM_production * price 0.003 -0.006 0.002 0.003 (0.004) (0.004) (0.003) (0.003) LSM_research -0.000 -0.000 -0.000 0.000 (0.001) (0.001) (0.000) (0.000) LSM_production -0.026*** 0.003 0.012** 0.002 (0.009) (0.005) (0.006) (0.004) LSMr_in_ASM 0.000 0.000 -0.000 0.000 (0.000) (0.000) (0.000) (0.000) LSMp_in_ASM -0.000 0.003*** 0.001 0.002* (0.001) (0.001) (0.001) (0.001) Observations 313,344 313,344 313,344 313,344 R2 0.146 0.134 0.102 0.059 R2 (within) 0.002 0.003 0.002 0.002 Cell FE Yes Yes Yes Yes Month FE Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using a Linear Probability Model; Standard errors are clustered at the level of the cell; Price interactions for cells in which tantalum is the dominant mineral are dropped. 161 Chapter 4 Appendix D: Addressing reverse causality in the expansion of LSM If industrial mining companies choose to invest in cells that are relatively conflict free, we expect a negative relationship between the incidence of conflict events and the expansion of any type of LSM activity. However, we find that the expansion of production concessions is associated with an increase, rather than a decrease, in the incidence of violence against civilians, riots and looting. In the case of reverse causality, these effects would thus represent a lower bound estimate. We do find a negative relationship between the expansion of production concessions and the incidence of battles between armed actors. In our interpretation, this finding is in line with the idea that LSM companies, backed by the government, obtain a monopoly on violence. Although the data do not allow us to entirely rule out reverse causality, we provide descriptive evidence and a sensitivity check that make it unlikely that the battle results are entirely driven by reverse causality. First, LSM companies do not seem to start production activities in areas with significantly lower battles. Indeed, conditional on having any LSM permit in a cell – implying the presence of mineral deposits suitable for LSM production – we find a higher monthly incidence of battles in cells with an LSM production permit. While the average monthly probability of battles equals 0.4% for cells with only a research permit, it equals 0.8% for cells with a production permit. Moreover, LSM production concessions are concentrated in the three Kivu provinces (North-Kivu, South- Kivu and Maniema), which jointly account for 69% of the battle events in our sample and hence have the highest monthly probability of battles. While the average share of the cell-surface area covered by research permits is highest for the provinces in Orientale (42%) and similar for the provinces in Katanga and Kivu (19% and 20%), the average share covered by production permits is higher in the Kivu provinces (4%) compared to the provinces in Orientale and Katanga (1.9% and 1%). Second, we conduct a sensitivity check that re-runs our baseline regressions but includes lags for the incidence of conflict events in previous months. By doing so, we effectively control for the incidence of conflict events prior to the expansion of LSM activities. These specifications further allow us to account for the fact that conflict often persists through time. The lags capture the incidence of all types of conflict events, thus taking into account that past battles may affect e.g. future violence against civilians and looting, or the other way around. We realize that the coefficients on the dynamic lags may be estimated with bias.30 However, we only introduce them to check the robustness of our b coefficients. 30 The introduction of lagged conflict variables gives rise to ‘dynamic panel bias’, i.e. the lags are correlated with the error term (Nickell 1981). Since we perform within-cell estimations, our estimates of the lags would understate the 162 Chapter 4 We introduce dynamic conflict lags in Table D.1. The lags take the form of a dummy variable which indicates if any of the conflict events (battles between armed groups, violence against civilians, riots or looting) took place within a cell in previous months. We control for six- month lags in columns (1) to (4) and for twelve month lags in columns (5) to (8). Adding this battery of lags only slightly changes the estimated coefficient sizes; except for ‘violence against civilians’ where we now find a larger effect of k*i_D9=6Ie;<=>'f, suggesting that our baseline estimate was indeed a lower bound. Importantly, the estimated coefficient size on battles hardly changes: we still find a rather strong negative relationship between the expansion of LSM production concessions and the incidence of battles. actual persistence of conflict. However, the bias is likely to be small since it decreases with the number of time periods, which is large in our case; i.e. 144 months (Roodman 2009). 163 Chapter 4 Table D.1: Including dynamic conflict lags battles violence riots looting battles violence riots looting (1) (2) (3) (4) (5) (6) (7) (8) ASM_value 0.003** 0.002** 0.000 0.001* 0.003** 0.002*** 0.000 0.001* (0.001) (0.001) (0.000) (0.001) (0.001) (0.001) (0.000) (0.001) LSM_research * price 0.001 0.001 -0.000** 0.000 0.001 0.000 -0.000** 0.000 (0.001) (0.001) (0.000) (0.000) (0.001) (0.001) (0.000) (0.000) LSM_production * price 0.004 -0.005 0.003 0.004 0.005 -0.006 0.003 0.004 (0.003) (0.003) (0.003) (0.003) (0.004) (0.004) (0.003) (0.003) LSM_research -0.000 -0.000 -0.000 0.000 -0.000 0.000 -0.000 0.000 (0.001) (0.001) (0.000) (0.000) (0.001) (0.001) (0.000) (0.000) LSM_production -0.024*** 0.006 0.013** 0.002 -0.025*** 0.009* 0.013** 0.003 (0.008) (0.004) (0.005) (0.004) (0.008) (0.006) (0.005) (0.004) LSMr_in_ASM 0.000 -0.000 -0.000 0.000 0.000 -0.000 -0.000* 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) LSMp_in_ASM -0.001 0.002*** 0.000 0.001* -0.001 0.002*** 0.000 0.001* (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Conflict lag 1 0.079*** 0.085*** 0.020*** 0.031*** 0.077*** 0.082*** 0.019*** 0.030*** (0.011) (0.010) (0.005) (0.006) (0.011) (0.010) (0.005) (0.006) Conflict lag 2 0.044*** 0.049*** 0.012** 0.024*** 0.041*** 0.045*** 0.012** 0.024*** (0.009) (0.008) (0.005) (0.005) (0.008) (0.008) (0.005) (0.005) Conflict lag 3 0.023*** 0.036*** 0.012** 0.009* 0.020*** 0.033*** 0.012** 0.009* (0.006) (0.007) (0.005) (0.005) (0.006) (0.006) (0.005) (0.005) Conflict lag 4 0.025*** 0.030*** 0.008* 0.012** 0.022*** 0.026*** 0.006 0.012** (0.008) (0.008) (0.004) (0.005) (0.007) (0.007) (0.004) (0.005) Conflict lag 5 0.014** 0.022** 0.008* 0.003 0.011* 0.015* 0.007* 0.002 (0.007) (0.009) (0.004) (0.004) (0.006) (0.008) (0.004) (0.004) Conflict lag 6 0.030*** 0.028*** 0.017*** 0.010* 0.026*** 0.021*** 0.014*** 0.008 (0.006) (0.008) (0.005) (0.005) (0.006) (0.007) (0.004) (0.005) Table continues on the next page 164 Chapter 4 Table continues from the previous page Conflict lag 7 0.016** 0.010 0.007* 0.004 (0.007) (0.007) (0.004) (0.004) Conflict lag 8 0.017** 0.017** 0.004 0.001 (0.007) (0.007) (0.004) (0.004) Conflict lag 9 0.003 0.006 0.000 -0.003 (0.008) (0.008) (0.004) (0.004) Conflict lag 10 0.006 0.019*** 0.007* 0.006 (0.008) (0.007) (0.004) (0.004) Conflict lag 11 0.016* 0.024*** 0.009** 0.008** (0.009) (0.006) (0.003) (0.004) Conflict lag 12 0.003 0.012 0.006 -0.000 (0.007) (0.008) (0.005) (0.004) Observations 300,288 300,288 300,288 300,288 287,232 287,232 287,232 287,232 R2 0.169 0.161 0.110 0.070 0.173 0.166 0.108 0.072 R2 (within) 0.028 0.034 0.013 0.013 0.029 0.037 0.014 0.013 Cell FE Yes Yes Yes Yes Yes Yes Yes Yes Month FE Yes Yes Yes Yes Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using a Linear Probability Model; Standard errors are clustered at the level of the cell; Columns (1) -(4) include up to 6-month dynamic conflict lags, while Columns (5) -(8) include up to 12-month dynamic conflict lags. 165 Chapter 4 Appendix E: Robustness checks 1. Robustness checks related to mineral price interactions Table E.1: Using lagged world mineral prices battles violence riots looting battles violence riots looting (1) (2) (3) (4) (5) (6) (7) (8) ASM_value (1m lag) 0.004*** 0.004*** 0.001 0.002** (0.002) (0.001) (0.000) (0.001) LSM_research * price (1m lag) 0.002 0.001 -0.000* 0.001 (0.001) (0.001) (0.000) (0.000) LSM_production * price (1m lag) 0.003 -0.006 0.002 0.003 (0.004) (0.004) (0.003) (0.003) ASM_value (6m lag) 0.004*** 0.003*** 0.001 0.001** (0.002) (0.001) (0.000) (0.001) LSM_research * price (6m lag) 0.001 0.000 -0.000** 0.000 (0.001) (0.001) (0.000) (0.000) LSM_production * price (6m lag) 0.003 -0.003 0.001 0.004 (0.004) (0.005) (0.003) (0.003) LSM_research -0.000 -0.000 -0.000 0.000 -0.000 0.000 -0.000 0.000 (0.001) (0.001) (0.000) (0.000) (0.001) (0.001) (0.000) (0.000) LSM_production -0.026*** 0.003 0.012** 0.002 -0.026*** 0.001 0.012** 0.002 (0.009) (0.005) (0.006) (0.004) (0.009) (0.005) (0.005) (0.004) LSMr_in_ASM 0.000 -0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) LSMp_in_ASM -0.000 0.003*** 0.001 0.002* -0.000 0.003*** 0.001 0.002* (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Observations 313,343 313,343 313,343 313,343 313,338 313,338 313,338 313,338 R2 0.146 0.134 0.102 0.059 0.146 0.134 0.102 0.059 R2 (within) 0.002 0.003 0.002 0.002 0.002 0.003 0.002 0.002 Cell FE Yes Yes Yes Yes Yes Yes Yes Yes Month FE Yes Yes Yes Yes Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using a Linear Probability Model; Standard errors are clustered at the level of the cell; In Columns (1) -(4) the ASM index is calculated with 1 month lags of world mineral prices, while we use 6- month lags in Columns (5) -(8). 166 Chapter 4 Table E.2: Controlling for month-to-month price changes battles violence riots looting battles violence riots looting (1) (2) (3) (4) (5) (6) (7) (8) ASM_value 0.004*** 0.004*** 0.001 0.002** 0.004** 0.004*** 0.001 0.002** (0.002) (0.001) (0.000) (0.001) (0.002) (0.001) (0.000) (0.001) LSM_research * price 0.002 0.001 -0.000 0.001 0.002 0.001 -0.000* 0.001 (0.001) (0.001) (0.000) (0.000) (0.001) (0.001) (0.000) (0.001) LSM_production * price 0.003 -0.006 0.002 0.003 0.003 -0.005 0.003 0.003 (0.004) (0.004) (0.003) (0.003) (0.004) (0.004) (0.003) (0.003) ASM_value (D m-1) -0.000* -0.000 -0.000 -0.000 (0.000) (0.000) (0.000) (0.000) LSM_research * price (D m-1) -0.000 0.000 0.000** 0.000 (0.000) (0.000) (0.000) (0.000) LSM_production * price (D m-1) 0.000 -0.000 -0.000 -0.000 (0.000) (0.000) (0.000) (0.000) ASM_value (abs. D m-1) -0.000 -0.000 -0.000 -0.000 (0.000) (0.000) (0.000) (0.000) LSM_research * price (abs. D m-1) 0.000 0.000 0.000** 0.000 (0.000) (0.000) (0.000) (0.000) LSM_production * price (abs. D m-1) -0.000 -0.000 -0.000 0.000 (0.000) (0.000) (0.000) (0.000) LSM_research -0.000 -0.000 -0.000 0.000 -0.001 -0.000 -0.000 0.000 (0.001) (0.001) (0.000) (0.000) (0.001) (0.001) (0.000) (0.000) LSM_production -0.026*** 0.003 0.012** 0.002 -0.025*** 0.005 0.013** -0.001 (0.009) (0.005) (0.006) (0.004) (0.008) (0.005) (0.006) (0.004) LSMr_in_ASM 0.000 -0.000 -0.000 0.000 0.000 -0.000 -0.000 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) LSMp_in_ASM -0.000 0.003*** 0.001 0.002* -0.000 0.003*** 0.001 0.002* (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Observations 313,343 313,343 313,343 313,343 313,343 313,343 313,343 313,343 R2 0.146 0.134 0.102 0.059 0.146 0.134 0.102 0.059 R2 (within) 0.002 0.003 0.002 0.002 0.002 0.003 0.002 0.002 Cell FE Yes Yes Yes Yes Yes Yes Yes Yes Month FE Yes Yes Yes Yes Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using a Linear Probability Model; Standard errors are clustered at the level of the cell; In Columns (1) -(4) we construct the month-to-month price interactions using the difference in world mineral price with respect to the previous month, which can be positive or negative; in Columns (5) -(8), we construct them using the absolute difference in world mineral price with respect to the previous month. 167 Chapter 4 Table E.3: Looking at yearly observations battles violence riots looting (1) (2) (3) (4) ASM_value 0.051*** 0.045*** 0.009 0.021** (0.019) (0.017) (0.006) (0.009) LSM_research * price 0.022 0.014 -0.005 0.009 (0.016) (0.015) (0.003) (0.007) LSM_production * price 0.027 -0.085 0.010 0.043 (0.047) (0.059) (0.028) (0.039) LSM_research -0.003 -0.002 -0.003 -0.000 (0.014) (0.014) (0.005) (0.006) LSM_production -0.298*** 0.029 0.163** 0.014 (0.090) (0.069) (0.076) (0.056) LSMr_in_ASM 0.003 -0.001 -0.002 0.000 (0.004) (0.004) (0.002) (0.003) LSMp_in_ASM -0.003 0.034*** 0.008 0.019* (0.013) (0.012) (0.011) (0.010) Observations 26,112 26,112 26,112 26,112 R2 0.503 0.457 0.382 0.323 R2 (within) 0.012 0.015 0.012 0.015 Cell FE Yes Yes Yes Yes Year FE Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using OLS; the dependent variables count the number conflict events that occurred in cell < and year ;, while the other variables represent the yearly average of their baseline counterparts; Standard errors are clustered at the level of the cell. 168 Chapter 4 Table E.4: Excluding the LSM price interactions battles violence riots looting (1) (2) (3) (4) ASM_value 0.004*** 0.003*** 0.001 0.002** (0.002) (0.001) (0.001) (0.001) LSM_research -0.000 0.000 -0.000 0.000 (0.001) (0.001) (0.000) (0.000) LSM_production -0.025*** -0.000 0.013** 0.003 (0.008) (0.005) (0.006) (0.005) LSMr_in_ASM 0.000 0.000 -0.000 0.000 (0.000) (0.000) (0.000) (0.000) LSMp_in_ASM -0.000 0.003*** 0.001 0.002* (0.001) (0.001) (0.001) (0.001) Observations 313,344 313,344 313,344 313,344 R2 0.146 0.134 0.102 0.059 R2 (within) 0.002 0.003 0.002 0.002 Cell FE Yes Yes Yes Yes month FE Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using a Linear Probability Model; Standard errors are clustered at the level of the cell. 169 Chapter 4 2. Robustness check controlling for conflict persistence through time and across space Table E.5: Including dynamic and spatial conflict lags battles violence riots looting (1) (2) (3) (4) ASM_value 0.002* 0.002** 0.000 0.001 (0.001) (0.001) (0.000) (0.001) LSM_research * price 0.001 0.001 -0.000** 0.000 (0.001) (0.001) (0.000) (0.000) LSM_production * price 0.004 -0.005 0.003 0.004 (0.003) (0.004) (0.003) (0.003) LSM_research -0.000 0.000 -0.000 0.000 (0.001) (0.001) (0.000) (0.000) LSM_production -0.023*** 0.007 0.013** 0.003 (0.007) (0.005) (0.005) (0.004) LSMr_in_ASM 0.000 -0.000 -0.000 -0.000 (0.000) (0.000) (0.000) (0.000) LSMp_in_ASM -0.001 0.002*** 0.000 0.001* (0.001) (0.001) (0.001) (0.001) Adjacent cell conflict 0.017*** 0.009*** 0.002** 0.005*** (0.002) (0.002) (0.001) (0.001) Adjacent cell conflict lag 1 0.004*** 0.003** 0.001* 0.001 (0.001) (0.001) (0.001) (0.001) Observations 300,288 300,288 300,288 300,288 R2 0.174 0.163 0.110 0.071 R2 (within) 0.033 0.036 0.013 0.014 Cell FE Yes Yes Yes Yes Month FE Yes Yes Yes Yes Dynamic conflict lags 1-6 Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using a Linear Probability Model and include dynamic conflict lags up to six months; Standard errors are clustered at the level of the cell. 170 Chapter 4 3. Robustness check controlling for rainfall shocks and seasons We control for rainfall as a proxy for exogenous shocks to agricultural income. In doing so, we follow a number of recent papers (Maystadt et al. 2014; Miguel et al. 2004; Parker and Vadheim 2017). On the one hand, an increase in agricultural income may raise the opportunity cost to join armed groups; on the other hand, it may increase armed groups' incentives to loot farmer communities. Heavy rainfall could also hinder mining activities and the movement of armed groups. We use monthly rainfall data from the Climatic Research Unit (CRU – University of East Anglia). The spatial resolution of the CRU grid cells is larger than the grid cells in our analysis. In ArcGIS, we therefore assign to each grid cell the rainfall data from the nearest CRU centroid. The distance varies between 0 and 52 kilometers, with a mean of 8.3 and a standard deviation of 7.7 kilometers. First, we follow Maystadt et al. (2014) in calculating rainfall anomalies; these measure deviations from normal rainfall conditions for each cell-month observation. Specifically, the anomalies measure the monthly deviation from the long-term monthly mean, divided by the monthly long-term standard deviation.31 The monthly basis of the rainfall anomalies corrects for seasonal patterns in rainfall. To control for the possibility that seasonal patterns may matter, we follow the example of Parker and Vadheim (2017) in constructing variables to indicate wet and dry seasons. Specifically, based on the long-run monthly rainfall averages, we create two dummy variables that indicate the three driest and the three wettest months for each cell. Table E.6 presents the results. In Columns (1) -(4) we control for contemporaneous and lagged rainfall anomalies.32 The baseline findings remain unchanged. The coefficients on the rainfall anomalies are negative for all conflict events, but always small and mostly insignificant. In Columns (5) -(8) we additionally include the indicator variables for cell-specific wet and dry seasons. The baseline findings are again confirmed. The seasonal indicators are not significant in any of the specifications. 31 The long-term average and standard deviation are calculated over the period 1997-2015. 32 We include lags for the previous six months, as rainfall lags further in time are mostly insignificant. The results are however robust to adding rainfall lags for the previous twelve months. We recognize that including rainfall lags also gives rise to the ‘dynamic panel bias’ discussed in Appendix D. 171 Chapter 4 Table E.6: Controlling for rainfall shocks and seasons battles violence riots looting battles violence riots looting (1) (2) (3) (4) (5) (6) (7) (8) ASM_value 0.004** 0.004*** 0.001 0.002** 0.004** 0.004*** 0.001 0.002** (0.002) (0.001) (0.000) (0.001) (0.002) (0.001) (0.000) (0.001) LSM_research * price 0.002 0.001 -0.000 0.001 0.002 0.001 -0.000 0.001 (0.001) (0.001) (0.000) (0.000) (0.001) (0.001) (0.000) (0.000) LSM_production * price 0.003 -0.006 0.002 0.003 0.003 -0.006 0.002 0.003 (0.004) (0.004) (0.003) (0.003) (0.004) (0.004) (0.003) (0.003) LSM_research -0.000 0.000 -0.000 0.000 -0.000 0.000 -0.000 0.000 (0.001) (0.001) (0.000) (0.000) (0.001) (0.001) (0.000) (0.000) LSM_production -0.027*** 0.003 0.012** 0.001 -0.027*** 0.003 0.012** 0.001 (0.009) (0.005) (0.006) (0.004) (0.009) (0.005) (0.006) (0.004) LSMr_in_ASM 0.000 0.000 -0.000 0.000 0.000 0.000 -0.000 0.000 (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) LSMp_in_ASM -0.000 0.003** 0.001 0.002* -0.000 0.003** 0.001 0.002* (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Observations 300,288 300,288 300,288 300,288 300,288 300,288 300,288 300,288 R2 0.147 0.134 0.101 0.060 0.147 0.134 0.101 0.060 R2 (within) 0.002 0.003 0.002 0.002 0.002 0.003 0.002 0.002 Cell FE Yes Yes Yes Yes Yes Yes Yes Yes Month FE Yes Yes Yes Yes Yes Yes Yes Yes Rainfall anomaly lags 0-6 Yes Yes Yes Yes Yes Yes Yes Yes Seasonal indicators No No No No Yes Yes Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.1; All specifications are estimated using a Linear Probability Model; Standard errors are clustered at the level of the cell. 172 Chapter 5 5. Would you rebel? An inquiry among high-risk youth in Eastern Congo* Summary Why would an individual choose to take up arms and fight? Several decades of research on armed conflict have yielded relatively few quantitative empirical analyses on the individual propensity to rebel (compared to the large number of ethnographic case studies and cross-country studies). This chapter looks at the intention to rebel in a high-risk population of artisanal miners in a mining town in eastern Congo. The large majority of our respondents have been exposed to armed conflict in the past and some have participated in the activities of armed groups. We inquire about their intention to rebel at a time when their main income source is under threat because of the arrival of a large-scale mining company. We then identify how their responses vary with four motivations that have been highlighted in the theoretical literature: grievances, material incentives, social incentives and previous exposure to conflict. The results suggest that all four motivations play a significant role. * This chapter is based on a paper written with Marijke Verpoorten (University of Antwerp, IOB). The research was funded by FWO – Research Foundation Flanders (grant numbers 1517614N & 11Q2816N), and by CEGEMI – Centre d'Expertise en Gestion Minière. Permission to conduct the research was granted by the University of Antwerp’s Ethics Committee for the Social Sciences and Humanities (file nr. SHW_15_06), by the Congolese Ministry of Mines, by SAESSCAM and by the local authorities in Kamituga. This research would not have been possible without the artisanal miners of Kamituga who offered their time to be interviewed, without the enumerators who worked in sometimes challenging circumstances, and without our colleague Janvier Kilosho who assisted with the field work. We offer them our sincere gratitude. We further owe thanks to employees of Banro, IPIS and INSO for their assistance and for sharing valuable information and data with us. We are also grateful for comments received on conference and seminar presentations at the 16th Jan Tinbergen Peace Science Conference in Milan, the UNU-MERIT Maastricht seminar, and the conference on transition and local development in Eastern DRC in Bukavu. Finally, we received much appreciated comments and suggestions from Jean-Paul Azam, Tom De Herdt, Enzo Nusio, Anja Shortland, Jo Swinnen, Peter van der Windt and Nina von Uexküll. 173 Chapter 5 5.1. Introduction In the 25-year period between 1989 and 2014, the world witnessed an average of 38 armed conflicts each year. About one in three of these conflicts took place in Africa, and about a quarter surpassed the threshold of 1,000 battle deaths a year, thus qualifying as civil wars. The majority – three quarters – were so-called ‘minor’ conflicts, surpassing the threshold of 25 battle-deaths a year (Pettersson and Wallensteen 2015). Whether major or minor, these conflicts could not take place without foot soldiers: the men (and women) who take up arms and fight. Why do these individuals take up arms and fight? This question remains largely unanswered. Most quantitative studies on armed conflict have analyzed data at the country or regional level – a level of analysis that is not well-suited to pin down individual motivations to fight. For instance, one of the most robust empirical associations is the one between low GDP and the occurrence of civil war (Collier and Hoeffler 1998), but this does not mean that it is the poor who fight (Blattman and Miguel 2010). Low GDP could also point to weak state capacity to fight off rebels (Fearon and Laitin 2003). Another stylized fact in cross-country studies is the recurrence of civil war (Collier and Sambanis 2002). Without individual-level data, it remains guessing what drives this conflict trap: e.g. the mobilization of individuals following conflict- induced poverty or victimization, or rebel networks? To a somewhat lesser extent, the same caveats apply to a more recent wave of papers that rely on within-country regional data (e.g. Dube and Vargas 2013; Verpoorten 2012). The bottom line is that we need individual-level data to test theories about individual behavior. The gap in the literature is understandable however, given the difficulty of approaching (past or potential) recruits in conflict and post-conflict areas. To address this gap, we inquired about the ‘intention to rebel’ among a sample of high- risk youth in eastern DR Congo. Our sample consists of 469 artisanal miners who live and work in Kamituga, a mining town in South-Kivu. The miners are all (young) men; the vast majority was exposed to the violence of the first and second Congo wars, and some participated in the violence. Despite the formal end of the war in 2003, pockets of chronic violence remained in the surroundings of Kamituga at the time of our survey in May 2015 (Stearns and Vogel 2015). Moreover, as the mining concession on which Kamituga is located was granted to a large-scale mining company, the miners faced an uncertain economic future – risking eviction once the company would move from the research to the production phase (Stoop et al. 2016). It is to these so-called ‘high-risk youth’, who experienced violent conflict and were at risk of losing their employment, that we asked the question “would you rebel?”. We framed this question as a manifestation against the large-scale mining company, and enquire about miners’ intention to 174 Chapter 5 engage in four concrete rebellious actions: destroying property, attacking employees, using fire arms and joining an armed group. Clues to the answer to our question can be found in theories on individual participation in collective violence. These theories have highlighted four groups of motivations. A first school of thought focusses on grievances and feelings of discontent and frustration as the drivers of rebellious action (Gurr 1970; Lichbach 1989). Influenced by Olson's (1965) analysis of collective action, a second school of thought argues that rational individuals will only rebel when the individual benefits from participation outweigh the costs. Two types of ‘selective incentives’ may influence the cost-benefit analysis. Material incentives include material and financial benefits from participation, as well as having low opportunity costs (Blattman and Annan 2016; Lichbach 1995; Tullock 1971). Social incentives emphasize the importance of an individual’s social network and are largely non- material and psychological in nature (Muller, Dietz, and Finkel 1991; Muller and Opp 1986). Finally, recent experimental studies have built on evolutionary theories of human prosocial behavior, and indicate that experiences of victimization and exposure to conflict may affect the costs and non- material benefits of participation in collective violence (e.g. Bauer et al. 2014; Gneezy and Fessler 2011; Voors et al. 2012). Our data provides proxies for each of these motivations, which we relate to miners’ intention to rebel. The associations that emerge allow us to sketch the profile of ‘would- be-rebels’. Of course, ‘would-be-rebels’ are not per se ‘rebels-to-be’. In other words, the self-reported intentions that we pick up could differ from actual decision-making in the future. This discrepancy can result from untruthful answering (because of strategic or social desirability considerations) or because participating in violence is costly and risky, and the effect of the actual costs may only set in when decision time has come. We address both issues and argue that our measures pick up meaningful variation in the intention to rebel. Our approach builds on prominent examples in the literature that measure the intention to rebel, and this is the first strand of literature we contribute to. Muller et al. (1991) for instance, test theories of rebellion in Peru using data on the intention of rebellious political action, both in a national sample and a sample of students at a protest-prone university. MacCulloch (2004) tests the relation between income and self-reported revolutionary support in a large sample of individuals taken from the World Values and Eurobarometer surveys. Exploiting the same data, Thyne and Schroeder (2012) investigate how marriage, unemployment and military involvement affect the individual-level taste for revolt. Finally, Argo (2009) and Ginges and Atran (2009) study the professed willingness to rebel in the context of the Israeli- Palestinian conflict. 175 Chapter 5 A second relevant strand in the literature has looked at ‘have-been-rebels’. Humphreys and Weinstein (2008) test for the theorized determinants of participation in Sierra Leone’s civil war, relying on a sample that includes both ex-combatants and non-combatants. Verwimp (2005) studies the characteristics of perpetrators of the Rwandan genocide in a small panel dataset of Rwandan households. Nussio (Forthcoming) compares characteristics across voluntarily and forcibly recruited members of Colombian armed groups. Although observing have-been rebels has its advantages, these studies have their own caveats: most importantly the non-randomness of the sample (due to attrition stemming from e.g. mortality or migration) and recall bias of pre- conflict characteristics (unless pre-conflict data is available). Overall, the quantitative studies looking at ‘have-been-rebels’ conclude that selective incentives perform better at predicting individual behavior than proxies for political grievances, but absence of evidence is not evidence of absence (Blattman and Miguel 2010). In fact, most studies on ‘have-been-rebels’ are ethnographic in nature, and argue that grievances take a prominent place, thus challenging the rational actor theory (e.g. Kalyvas and Kocher 2007; Scott 1976; Wood 2003). The contrasting lack of evidence for grievances in quantitative studies could signify that it is very difficult to capture political deprivation and grievances with quantitative data. Looking at the studies on individual participation in violence in DRC, we find that they are all ethnographic in nature, and award a role to a range of possible motivations: greed, low opportunity costs, grievances, community cohesion and related non- material incentives, protection from harm and insecurity, and forced recruitment (Jourdan 2011; Laudati 2013; Richards 2014; Van Acker and Vlassenroot 2001). A final relevant strand of qualitative literature deals with artisanal miners in DRC and their confrontation with mining companies (e.g. Geenen and Claessens, 2013; van Puijenbroeck and Schouten, 2013). A reading of this literature bears out a noteworthy element: judging from their use of language, miners experience the confrontation with mining companies as a war-like situation. For instance, van Puijenbroeck and Schouten (2013: p.29) quote a miner in Ituri Province, saying: “Here it is war. When the mining company arrives to relocate us, we will be ready: even the mothers are angry and have their axes and spears ready”.1 Geenen and Claessens (2013: p.102) cite a miner in South-Kivu Province: “They say they will give you work and the next day they chase you off. … They threatened us with policemen and dogs. We told them to do what they want, but we will not die because of hunger! [. . .] We would rather die by a bullet than die of hunger (Miners focus group 2011 int.)” Such confrontations between artisanal miners and companies are far from isolated cases. It is estimated that about 61 % of 1 Own translation of “Ici c’est la guerre. Quand la société minière vienne ici nous déplacer, nous serons prêtes ; même les mamans sont en colère et ont leurs arcs et flèches prêtes”. 176 Chapter 5 artisanal miners in eastern DRC are operating illegally in concessions of large-scale mining companies (Stoop et al. mimeo). From this overview of the literature, we conclude that our study makes three contributions. First, by looking at the intention to rebel of high-risk youth in a post-conflict environment, we are at the intersection of studies that look at would-be-rebels and studies that look at have-been rebels. This allows us to make a unique contribution, i.e. we are the first to empirically verify in a sizeable sample if and how past victimization and rebel networks are associated with the individual-level propensity to participate in violent conflict. By doing so, we can start opening the black box of the micro-dynamics underlying the recurrence of armed conflict. Second, the setting of artisanal miners in DRC is one in which all types of incentives to fight are present and about to rise. Interestingly, our setting provides for a ‘most likely case’ to pick up the role of grievances. Indeed, the concrete and imminent threat of the mining company and the associated perceived injustice allow us to sidestep the difficulty of measuring general and abstract grievances, and instead focus on concrete grievances. If, despite the concrete grievances and specific framing, grievances do not turn up significant, we would move closer in the direction of ‘evidence of absence’, rather than merely ‘absence of evidence’. Third, we add a quantitative analysis to the qualitative body of two strands of research on DRC: on the propensity to rebel, and on the confrontation between artisanal and large-scale mining. This allows us to triangulate findings across qualitative and quantitative methods, which in itself is an interesting exercise, and responds to the plea for more mixed methods research in conflict studies (Thaler 2015). We begin our analysis by discussing the background of mining and conflict in eastern DRC. Drawing on theories of individual participation in armed conflict, we then formulate a number of hypotheses about the intention to rebel. Next, we describe our research design and data, and present the results of the analysis. We conclude with a discussion of our main findings and their implications for policymaking. 5.2. Background We briefly sketch the two Congo wars, the post-war environment, and the situation of artisanal miners. First we discuss these issues for (eastern) DRC, then we zoom in on Kamituga. Eastern DRC includes the country’s seven most eastern provinces.2 The area is home to many of DRC’s untapped deposits of raw minerals, estimated to be worth US$24 trillion (UNEP 2 Ituri, Haut-Uele, North-Kivu, South-Kivu, Maniema, Haut-Katanga and Tanganyika. 177 Chapter 5 2011). Despite its vast mineral wealth, the majority of DRC’s population is dismal poor, as indicated by the country’s poverty headcount of 72.5%, and its bottom ranking on the Human Development Index.3 The First Congo War started in 1996, when a rebel force crossed the country from east to west, to overthrow the dictatorship of Mobutu Sésé Seko. Although nominally led by Laurent- Désiré Kabila, the rebellion was initiated and supported by Rwandan and Ugandan forces. The new government was headed by Laurent Désiré Kabila, who soon turned his back on his former allies. This planted the seed for the Second Congo war (1998-2003), the so-called ‘Great War of Africa’ that involved nine African countries (and even more separate armed groups), and became the deadliest conflict worldwide since World War II. Whereas the initial objective was to overthrow Kabila, the various countries and rebel groups also developed their own objectives, which involved the settling of local disputes and the appropriation of mineral wealth. Both wars are described and discussed at length by among others Autesserre (2010), Reyntjens (2010) and Stearns (2011). Despite the formal end of the Second Congo War and a national unity government in 2003, violence continued. In 2015, the year of our interviews, more than seventy armed groups were active in eastern Congo, and approximately 1.6 million people remained displaced (Stearns and Vogel 2015: p.7). Several factors explain the continued violence. For instance, due to the infusion of arms and rebels, many dormant local conflicts turned violent, and these local conflicts were not addressed in the peace accords that focused on national and international issues (Autesserre 2010). Furthermore, Kabila junior’s rule has so far proved inapt to turn the war logic around and transit to a peace economy. In the words of Stearns and Vogel (2015: p.8): “the government and its foreign partners have been unable to create a virtuous cycle of economic development in the rural Kivus that could entice local leaders to invest in stability rather than conflict”. In 2002, a new Mining Code was developed under the guidance of the World Bank and the IMF. The Code was designed to restore DRC’s reputation in terms of business environment after the debacle of the nationalizations of mining companies by Mobutu in the 1970s. The Code succeeded in attracting FDI – by offering advantageous fiscal regimes to private companies – but it has been criticized for remaining extremely vague on the use of mineral revenues, by the government or private companies, and on how these revenues should benefit the Congolese population (Mazalto 2009, 2005). Moreover, in its search to maximize fiscal revenue for the State, the Code has prioritized large-scale mining to the detriment of the sector’s artisanal- and small- scale mining segment. It limits artisanal mining to a handful of relatively small Artisanal 3 UNDP, 2015 (http://hdr.undp.org/en/countries/profiles/COD). The DRC is ranked 176 out of 188 countries. 178 Chapter 5 Exploitation Zones (AEZ). As a result, the vast majority of artisanal miners operates illegally in large-scale mining concessions.4 Kamituga is a mining town of 190,000 inhabitants, located in South-Kivu in the territory of Mwenga, at 180 kilometers of the provincial capital Bukavu (see Figure). Gold deposits were discovered in the 1920s and the Belgian company ‘Minière des Grands Lacs Africains’ (MGL) started commercial gold exploitation in the 1930s (Geenen 2014; Kyanga Wasso 2013; Vlassenroot and Raeymaekers 2004). Due to Mobutu’s disastrous economic policies, the instability of world mineral prices, and eventually the Congo wars, industrial production came to a halt, and artisanal mining got the upper hand. During the two Congo wars, Kamituga was the scene of several atrocities, including public executions, massacres and mutilations of civilians (UN 2010a). The town was also occupied by several armed groups that benefitted from the mineral sector by taxing artisanal miners and traders (Geenen 2014; Vlassenroot and Raeymaekers 2004). After the Congo wars, armed actors continued to benefit from Kamituga’s artisanal mining sector. The Congolese national army, the FARDC (Forces Armées de la République Démocratique du Congo) reportedly took over the existing taxation systems, while the rebel group FDLR (Forces Démocratiques pour la Libération du Rwanda) remained active in Kamituga’s surroundings setting up ‘tax barriers’ and relying on ambush attacks against mineral traders (Geenen 2014; Spittaels et al. 2014; UN 2010b). In 2002, a Canada-based multinational – Banro – acquired the right to exploit minerals in Kamituga. Banro started the research phase in 2011, and at the time of our interviews they were hoping to move to the production phase. Between 13,000 and 15,000 artisanal miners are however operating in its concessions.5 While still tolerating artisanal miners, Banro already restricted their activities. For instance, miners were not allowed to open new pits or make use of dynamite, crushing mills and electricity, all of which enhance the productivity of artisanal mining. To enforce the rules, Banro mainly relies on the Mining Police and at times on the FARDC. The enforcement of these rules often leads to friction and incidents, and such tensions are likely to increase when Banro decides to move to the production phase (Stoop et al. 2016). For instance, in Namoya and Twangiza – two mining sites in eastern DRC where Banro already moved to the production phase – artisanal miners engaged in various rebellious actions against the company, including protest 4 Stoop et al. (mimeo) estimate that AEZ cover only about 1% of the total surface area of mineral concessions in Eastern DRC. Furthermore, we find that less than 1% of artisanal miners operates in AEZ, while about 61% operates in concessions that have been granted to large-scale mining companies. 5 Geenen (2013) estimates the number of artisanal miners in Kamituga between 10,000 and 15,000. During our fieldwork in 2015, the representatives of several local mining committees communicated that a census undertaken in 2013 counted 13,600 artisanal miners. We counted 15,250 artisanal miners on the combined membership lists of two local committees of artisanal miners (COKA and CRC). 179 Chapter 5 marches, destruction of Banro property, kidnapping of Banro employees, and the forcible reoccupation of a mining site. Regarding the latter, a miner explains: “Banro had closed down Kaduma and Lukunguri, but we retook them by force. We marched, we barricaded the road. Banro tried to drive us out with policemen and dogs. But we told them that whatever action they took, we would stand firm” (Geenen 2013: p.6). Appendix A provides details on these and other rebellious actions on the part of miners. 5.3. Why rebel ? A vast literature deals with theories of individual participation in rebellion. In this section, we focus on four main arguments brought forth in the literature: grievances, material incentives, social incentives and previous exposure to conflict. We relate them to the context of Kamituga and derive testable hypotheses. 5.3.1. Grievances A first school of thought focusses on grievances and feelings of discontent as the drivers of rebellious action. In an overview of these deprived actor theories, Lichbach (1989) mentions a number of underlying processes, such as frustration from unfulfilled needs, political alienation, and a sense of social injustice. Central is the idea that individuals form expectations about what they are entitled to. A discrepancy between their expectations and what they actually get, may make individuals frustrated and angry enough to rebel (Gurr 1970). While empirical research on grievances originally focused on peasant revolt (Paige 1975; Scott 1976), recent studies indicate that they also play a role in conflicts between artisanal miners and industrial mining companies. Relying on evidence from case studies across the developing world, Carstens and Hilson (2009) argue that artisanal miners’ grievances – resulting from what they consider as unfair treatment and illegitimate claims to their land by large scale mining companies – lie at the basis of numerous conflicts that have led to casualties and costly damage to infrastructure. And, the United Nations Economic Commission for Africa acknowledges that conflicts between large and small-scale miners arise among others from “legitimate and illegitimate resource claims by the two groups” and “unfulfilled promises by large-scale mining companies” (UNECA 2003: p.8). Grievances are likely to play a role in Kamituga as well. While Banro refers to their legal right to exploit the minerals in their concession, artisanal miners refer to their traditional rights to live and work on the land of their ancestors (Geenen and Claessens 2013). And, while Banro’s presence already depressed revenues of artisanal miners, the expected compensation for such 180 Chapter 5 losses is not forthcoming (Kilosho et al. Forthcoming). To voice and claim their rights through peaceful means, miners count to some extent on local authorities. At the same time, trust in local authorities is not very high, and miners fear unjust dispossession and relocation of their activities and livelihoods. We hypothesize that artisanal miners’ intention to rebel increases the more they are aggrieved with Banro and with the local authorities who could defend their interests in negotiations with the company. Miners’ intention to rebel increases with: (H1) the level of grievances with Banro. (H2) the level of grievances with the local authorities. 5.3.2. Selective incentives A second school of thought, influenced by Olson's (1965) analysis of collective action, puts forward rational actor theories. These theories postulate that individuals will only rebel when the benefits from participation outweigh the costs. The costs are borne individually and can be high (e.g. involving the risk of personal injury); while the benefits of a successful rebellion are largely available to all regardless of participation. Rational actors will therefore only rebel if they expect to receive private benefits for participating, called ‘selective incentives’ (Tullock 1971). The initial focus was on material incentives. Lichbach (1995) offers extensive examples of cases where money, loot and other material rewards motivated participation in various rebellious acts, including peasant resistance, protests, riots, strikes, revolutions and coups. On the cost side of collective action, the opportunity cost motive postulates that participation in the activities of rebel groups is more likely for individuals who are unemployed or have little alternative economic opportunities (Collier and Hoeffler 1998). Blattman and Annan (2016) provided the first rigorous individual-level evidence linking increased employment to a reduction in rebellion. They experimentally evaluated a program of agricultural training, capital inputs and counselling directed at ex-fighters in Liberia who were engaged in illegal mining or occupied rubber plantations. The findings indicate that the increased earnings potential and the expectation of future transfers raised the opportunity cost to engage in mercenary activities in neighboring Côte d’Ivoire. The arguments and findings with respect to material incentives lead us to assume that artisanal miners’ intention to rebel: (H3) increases, the greater the individual material benefits are in rebelling against Banro (H4) decreases, the higher the individual material costs are in rebelling against Banro 181 Chapter 5 Selective incentives may also be non-material. Muller and Opp (1986) argue that the relevant incentives must be psychological in nature, simply because it is unlikely that the average citizen can expect substantial personal material gains from participating in rebellion. They put forward ‘social rewards’, stemming from feelings of solidarity with a group and gratification from conforming to a social norm. Similarly, Muller et al. (1991: p.1264) argue that “If an individual’s friends – or groups the individual belongs to or identifies with – believe that he or she ought to participate in collective action, then a social norm of participation exists. […] The individual who is part of the social network will derive benefit in the form of social approval from conforming to the social norm.” In line with this argument, Humphreys and Weinstein (2008) find that having social ties with members increases the likelihood of participation in the activities of a rebel group in Sierra Leone’s civil war. Rejecting the standard rational actor model, Muller and Opp (1986) propose a model in which the public goods value of rebellion is a relevant incentive for participation. They argue that individuals may adopt a collectivist notion of rationality, recognizing that rebellious collective action will not succeed if everyone freerides, making it collectively rational for all to participate. This mechanism is especially supposed to operate when a group’s influence in attaining the public good is perceived to be high – which is more likely when similar dissident groups are perceived as having been successful in the past (Bandura 1973; Muller and Opp 1986). Analyzing data from two surveys, Muller and Opp (1986) provide empirical evidence that supports these hypotheses. Based on the arguments and findings with respect to these social incentives, we hypothesize that artisanal miners are more likely to display an intention to rebel if: (H5) they are more exposed to a social norm to rebel. (H6) they perceive similar previous rebellious actions as having been successful. 5.3.3. Exposure to conflict Evolutionary theories posit that war and intergroup conflicts likely played an important role in shaping human behavior. In an environment of lethal group competition over scarce resources, evolution may have favored groups that displayed high levels of within-cooperation and out-group hostility (Bowles 2006; Choi and Bowles 2007; Darwin 1998). In line with these arguments, recent studies have shed light on how the experience of conflict may affect the cost of collective action as well as non-material benefits to participation. Experimental games indicate for instance that individuals exposed to war-related violence show an increased willingness to cooperate with their in-group and punish non-cooperation (Bauer et al. 2014; Gneezy and Fessler 2011; Voors et al. 2012). 182 Chapter 5 The findings of these studies thus award a role to experiences of victimization and exposure to violent conflict, which could influence the taste for revenge and justice for the community, but also foster the intra-group cohesion necessary to organize collective violence.6 We hypothesize that artisanal miners in Kamituga are more likely to display an intention to rebel if: (H7) they have been exposed to violent conflict. 5.4. Data collection In our research design, we aimed to reach a representative sample of artisanal miners in Kamituga, present them with the question “Would you rebel?”, and collect individual characteristics to study the determinants of individual intentions to participate in collective violence. In practice, we had to overcome many challenges. First, as there was no reliable list of artisanal miners that would allow us to draw a random sample, we had to establish one. To do so, we took advantage of the hierarchical structure of the mining site, which is divided in different zones (headed by ‘zone managers’), that consist of several mining pits (supervised by ‘pit managers’) who have a number of artisanal miners working with them. We first constituted a list of all active mining zones in Kamituga. From the resulting list of forty mining zones, we selected nine zones, seeking variation in terms of geographical location, the number of artisanal miners and the presence of Banro. In a second step, we asked the zone managers in the selected zones to provide us with a list of all pit managers, who in their turn provided us with a list of all artisanal miners working with them. The complete list for the nine mining zones consisted of 1,254 artisanal miners, working in 72 different pits. We randomly selected half of the pits in each zone, and – in each of the selected pits – randomly selected ten or fifteen miners to be included in the survey.7 The pit managers of selected pits were also included. Our final sample comprises 430 artisanal miners and 39 pit managers. All selected miners were individually interviewed with a structured survey.8 Prior to the implementation of the actual survey in May 2015, we conducted two rounds of exploratory fieldwork in Kamituga, Twangiza and Bukavu. In June and December 2014, we held focus group discussions with artisanal miners, organized in-depth interviews with different 6 While these factors could be categorized under ‘grievances’ or ‘social incentives’, we chose to make a separate category, because of their specific origin in previous violence. 7 The average pit on the list counts 17 miners. For selected pits with more than 30 miners, we randomly selected 15 instead of 10 miners. When a pit with less than 10 miners was selected, we randomly selected an additional pit in the same mining zone. 8 In total, we made 21 replacements (4% of our sample). The large majority of the miners who were replaced (86%) were not found by the enumerators. In most cases, these miners no longer worked in the selected pit. 183 Chapter 5 stakeholders9, and tested our survey instruments. This allowed us to get a good understanding of the research context and the applicability of our questions. In doing so, we could build on the extensive local network of our colleague Sara Geenen, who had been working with artisanal miners in the area for over five years (see e.g. Geenen, 2014, 2013; Geenen and Claessens, 2013). Combined with our repeated visits, this allowed us to create the necessary trust and network to implement a representative structured survey. The final questionnaire was shaped by the usual financial constraints, but also by several practical and security challenges. For instance, the initial plan was to survey miners not only in Kamituga, but also in Twangiza – where Banro already moved to the production phase and miners resisted dislocation. However, the climate in Twangiza proved to be too tense, making it impossible to draw a random sample and conduct independent research. Although the situation in Kamituga was relatively safe, it was not safe enough to walk around with a lot of cash. In the absence of a single bank in Kamituga, we had to drop most of the incentive-compatible games that we designed to measure attitudes (such as trust, fairness and cooperation). We did play a simple lottery game, to measure risk preferences, but also because the foresight to win a small sum of cash turned out to be important to motivate miners to participate in the survey.10 Finally, since we noticed that, overall, miners freely talked about their exposure to and participation in violent conflict events, we decided to drop the time-consuming and cognitively demanding ‘list experiment’ to inquire about the intention to rebel, and simply ask the question directly. The survey team was drawn from a large pool of potential candidates with extensive experience in conducting surveys in eastern DRC. Fifteen potential enumerators were selected and took part in a three-week training program. Based on their performance during the training program and a pilot-test of the survey, we selected twelve enumerators for the actual fieldwork. The survey was conducted in Swahili, using tablets. 9 Interviewed stakeholders include representatives of artisanal mining committees, Banro, civil society organizations, SAESSCAM (the Congolese public Service for Assistance to Artisanal and Small-Scale mining), the Congolese Ministry of Mines, the FARDC and local authorities. 10 We opted for a simple method of eliciting risk preferences, as these have been shown to be substantially easier to understand for (illiterate) participants (Dave et al. 2010). Specifically, we follow the Eckel & Grossman method (see (Dave et al. 2010; Eckel and Grossman 2002) which gives respondents a single choice between six gambles. Each gamble involves a 50% chance of winning either a high or a low payoff. Moving from gamble one (a safe alternative involving a sure payoff with zero variance) to gamble five, the gambles increase in both expected return and risk. From gamble five to six, only the variance increases while the expected payoff remains the same. The average expected return for respondents was equal to about 5 US$. 184 Chapter 5 5.5. Data description 5.5.1. Self-reported motivations to join an armed group To study why individuals may join an armed group, we first used a direct approach by asking miners the following question: “I will now read several motivations that could explain why people may join an armed group. Could you please indicate how important each of these reasons is in your opinion?” We presented the miners with six potential motivations: ‘defend the community against external aggression’, ‘fight against injustice, for the rights of the community’, ‘personal insecurity’, ‘to gain respect’, ‘forced participation’, and ‘to gain money’. The answer categories were: not at all important, very little importance, rather important, important, very important and refuse to respond. Table 1 gives an overview of the distribution of answers. Interestingly, even when talking about people’s motivations in general, respondents award more importance to the somewhat more laudable motivations ‘defend the community against external aggression’ and ‘fight against injustice, for the rights of the community’ than to the more selfish motivation ‘to gain money’. When asked in an open follow-up question if there were other important reasons, 131 responded affirmatively. Strikingly, 85 mentioned revenge as an important reason. Furthermore, 26 mentioned that rebel groups are a safe haven for bandits (i.e. those who risk being convicted may flee and join an armed group to escape a prison sentence), and 20 mentioned ‘the lack of employment, poverty’, which suggests need rather than greed as a motive. Overall, this direct line of questioning awards some importance to all motivations, yet gives more prominence to the motivations that can be linked to grievances, social incentives and victimization. To study the determinants of participation in collective violence in a more indirect way, we measure miners’ individual intention to rebel, and relate it to proxy measures for grievances, material incentives, social incentives and exposure to conflict. 5.5.2. The intention to rebel We framed the intention to rebel as a reaction against Banro. We first asked our respondents to imagine the following realistic scenario: “Imagine a situation where Banro moves to the production phase in Kamituga. Imagine that they organize professional training programs and authorize some artisanal miners to continue operating in selected mining sites at Kamituga. However, their budget is not sufficiently large to accommodate all miners in the training programs, and the selected mining sites are not sufficiently large to accommodate all artisanal miners.” We asked them if such a situation would lead to a conflict between the company and artisanal miners. Answer categories included: refuse to respond (0), certainly not (1), probably not (2), maybe (3), probably (4) and certainly (5). We additionally asked if they thought the conflict 185 Chapter 5 would be violent. The large majority of miners indicates that the above scenario would certainly lead to a conflict (72%) and that it would be violent (64%). Next, we presented each miner with four rebellious actions: (1) destroying Banro property; (2) attacking Banro employees; (3) using fire arms; and (4) joining or revitalizing an armed group or local defense force (“like Mai-Mai Shikito”11). We first asked them about the likelihood that, in the event of the above scenario, artisanal miners would engage in these actions. The large majority of respondents indicates that artisanal miners would probably or certainly destroy Banro property (80%), attack Banro employees (76%), use fire arms (63%), join or revitalize a rebel group (64%). We then asked about their personal intentions to engage in them. Figure 2 displays the miners’ personal intentions to engage in these actions. We use these answers to construct our proxies for the ‘intention to rebel’. Specifically, we construct four dummy variables that indicate a miner’s intention to certainly or probably destroy Banro property (48%), attack Banro employees (36%), use fire arms (29%), join or revitalize a rebel group (19%). Of course, would-be-rebels are not per se rebels-to-be. The self-reported intentions that we pick up could differ from actual decision- making in the future. This discrepancy can result from untruthful answering (because of strategic considerations) or because participating in violence is costly and risky, and the effect of the actual costs may only set in when decision time has come. We now address both issues. Since the question is framed with respect to a concrete threat and enemy, and because negotiations between Banro and the community are ongoing, miners may strategically overstate their intention to rebel. To elicit truthful and accurate answers, we started each interview with a standardized introduction in which we presented ourselves, the research and its purpose, and guaranteed anonymity to the respondent as well as the option to refuse to respond to questions or abort the interview at any point. We directly encouraged respondents to give truthful answers by stating: “It must be said that this research has no direct benefit for you. If you decide to participate and respond frankly, you will help us to better understand the situation of miners in Kamituga.”. While this introduction can by no means guarantee truthful answers, we do think it can increase the quality of the interview and the answers given. To further reduce the probability of strategic responses in our variables of interest, we first asked miners about the intention to rebel in the population of miners. Strategic respondents could hence overstate the general intention to rebel, while there was no need to overstate their personal intentions. Table 2 shows summary statistics for both the general and the 11 Mai-Mai are community based self-defense militia. They originated in the 1960’s, when education minister Piere Mulele organized youth into militia to rebel against Mobutu’s government. In the 1990’s, Mai-Mai groups resurfaced in Eastern DRC to protect local communities from Mobutu’s army and invading foreign armed forces after the Rwandan genocide (Bøås and Dunn 2014; IRIN 2006; Verweijen 2015). See infra for detailed information on Mai- Mai Shikito. 186 Chapter 5 personal intention to rebel. While the general intention to rebel is on average answered with ‘probably’, the personal intention to rebel is much lower, averaging between ‘probably not’ and ‘maybe’. Actual behavior of miners also suggests that the intention to rebel is not greatly overstated. As described in the Background Section and detailed in Appendix A, miners have undertaken very concrete rebellious actions against Banro and other mining companies. Banro also acknowledges that the threat of rebellion by artisanal miners is real; their 2013-2014 Annual Information Form mentions: “Some or all of the Company's properties are inhabited by artisanal miners. These conditions may interfere with work on the Company's properties and present a potential security threat to the Company's employees. There is a risk that operations of the Company may be delayed or interfered with, due to the conditions of political instability, violence and the inhabitation of the properties by artisanal miners.” (Banro 2014: p.16). In sum, we argue that our proxies provide meaningful variation in the intention to rebel, and that the likelihood of rebellious action is real. 5.5.3. Determinants to rebel Grievances Hypotheses (1) and (2) postulate that the intention to rebel increases as miners are more aggrieved with Banro and local authorities. The relevant local authorities that could defend miners’ interests in negotiations with Banro are the Chef de Poste and the Mwami. The Mwami is the ‘traditional’ local chief that played (and can sometimes still play) an important role in the allocation of land. The Chef de Poste is the ‘modern’ local authority. Respondents were asked: “Now I would like to ask some questions about actors that play a role in the management of mineral resources in Kamituga. How do you think these actors contribute to the well-being of artisanal miners?” The miners were shown a set of five smileys, corresponding to the opinions: very negative (1), rather negative (2), no effect (3), rather positive (4), very positive (5). To capture ‘grievances’ we recoded these answers, coding very positive as (1) and very negative as (5). The results, shown in Table 3, indicate that miners are very much aggrieved with Banro, reaching an average value of 4.6. There is however still variation, with 11% of miners indicating that Banro has no, or a positive impact. The Chef de Poste and the Mwami are, on average, thought to have little impact on artisanal miners’ well-being, reaching values of 3.1 and 3.3. Responses vary widely however, with 38% and 28% of miners expressing a positive opinion about the Chef de Poste and the Mwami, compared to 37% and 48% expressing a negative opinion. 187 Chapter 5 Material incentives According to hypothesis (3), the intention to rebel increases the greater are the individual material benefits in rebelling against Banro. While there are no benefits that are strictly individual, there is a group of miners that stands to lose more than others should Banro move to the production phase; they therefore have more to gain from a rebellion. This group of miners operates in the mining zones that are of most interest to Banro, because they are more suitable for industrial mining development. These zones are frequently visited by Banro employees, for purposes of geological exploration. To capture the variation in benefits from rebellion against Banro, we purposefully stratified our sample across zones suited for industrial exploitation and other zones. In addition, we asked each miner how many times Banro employees visited their mining zone in the month prior to the interview. The answers range from 0 to 31, with an average of 3 (see Table 3). Fourth, we hypothesize that the intention to rebel decreases with miners’ personal cost to rebel against Banro. While the risk of injury may be similar across miners that decide to rebel, we do have a proxy for a cost that varies across miners, namely the foregone employment opportunity at Banro (assuming that Banro will not hire miners who engaged in rebellious actions against the company). We asked miners whether they would be interested in working for Banro, in case artisanal mining would no longer be possible in Kamituga. Answer categories included: very interested (1), interested (2), rather indifferent (3), not interested (4). Just over half of the miners (55%) expressed interest, choosing options (1) or (2). We create a dummy variable that equals one if a miner is interested in working for Banro; thus indicating relatively high costs to participation compared to miners who are not interested in working for Banro (see Table 3). Social incentives Hypothesis (5) conjectures that being exposed to norms that value rebellious activity may positively affect a miner’s intention to rebel. Such exposure is more likely if the miner’s network includes (former) rebels. In the survey, we directly asked miners if they had ever participated in the activities of an armed group: 3.2% of the miners indicated they had. Since miners who work together in a mining pit form a close network, we defined a variable that indicates whether a miner works in a pit that includes at least one self-reported (former) rebel. This is the case for 29% of our respondents; we assume they have a stronger rebel network (see Table 3).12 12 2.8% of miners refused to answer the question on participation in the activities of an armed group. If we assume that these respondents did participate, 48% of miners works in a pit with an ex-combatant. All results are robust to using this alternative specification of having a rebel network. 188 Chapter 5 According to Hypothesis (6), the intention to rebel increases if previous rebellious actions are perceived as having been successful. In the case of Kamituga, we evaluate how the actions of Mai-Mai Shikito are perceived. In the focus group discussions that we held during the exploratory rounds of fieldwork, miners referred to Mai-Mai Shikito as an armed group that attracted many young men of Kamituga in the past. Allegedly the group was created by the vice-president of CEPACAM (a committee of artisanal miners in Kamituga) and consisted mainly of artisanal miners from Kamituga (focus group discussion, December 2014). In 2008, the Harvard Humanitarian Initiative interviewed 25 members of Mai-Mai Shikito in Kamituga and surroundings. The militiamen stated that their aim was to protect Congo’s population from foreign invaders: “we came to realize that our country has been invaded by foreign troops, and that we needed, ourselves, to fight for our country … us … we have waited for government support for so long, it did not come, so we decided to fight ourselves for the country” (Harvard Humanitarian Initiative 2009: p.33). Moreover, they stress the importance of natural resources: “The goal of this group is to protect natural resources that are in this part of the country. We know already that natural resources are what motivate the enemy to come here. […] So it is mostly to protect natural resources and protect those who are weak and fearful, those who say they can’t do that. We just sacrifice ourselves.” (Harvard Humanitarian Initiative 2009: p.34). The interviews further indicate that the militiamen supplemented their income from the group with artisanal mining activities. According to our focus group discussions in 2014, most of the members had returned to mining activities, but reportedly still have their arms at home and could pick them up when necessary. One miner said: “If you take about 100 miners today, you may find 3 to 5 who were part of Shikito. But if we are all chased away in the future and Banro does not leave us with any alternative, everyone could join” (open-ended interview, December 2014). To measure the perceived successfulness of previous rebellion, we asked artisanal miners to evaluate the impact of Mai-Mai Shikito’s actions on the well-being of artisanal miners in Kamituga. The answer categories included: very negative (1), rather negative (2), no effect (3), rather positive (4), very positive (5). On average, their impact is evaluated at 2.8, between having a rather negative effect and no effect at all (see Table 3). Yet, about 14% of miners indicates that Mai-Mai Shikito had a positive impact. Exposure to conflict Finally, we hypothesize that exposure to violent conflict may increase miner’s intention to rebel. Nearly all miners in our sample have been exposed to violent conflict events. The large majority of miners (79%) were at a certain point internally displaced due to armed combat. Four out of ten were forced to perform labor under armed threat (41%), or have relatives who were physically hurt 189 Chapter 5 during armed combat (40%). About one third of the miners witnessed killings or rape (29%), were forced to give away revenue under armed threat (29%), or had their house pillaged by an armed group (28%). About 25% have close family members who were physically hurt during armed combat and about 5% were physically hurt themselves. A whopping 93% of the miners in our sample were exposed to at least one of these events, while 65% of miners were exposed to one or more of the most extreme forms of these events (i.e. having close family members or relatives who were physically hurt during armed combat, being physically hurt themselves, or having witnessed killings or rape). Bauer et al. (2014) find that the impact on behavior is largest when traumatic conflict events are experienced within the developmental age window of 7 to 20 years. We therefore create a dummy variable to capture exposure to the most extreme forms of conflict events between the age of 7 and 20; it equals one for 36% of the sample (see Table 3). Table 4 provides an overview of the general hypotheses as formulated in Section 5.3 and their concrete operationalization. 5.5.4. Control variables Apart from these characteristics of interest, we control for a number of variables that may enter a monetary cost-benefit calculation, but can also constitute the source of grievances in the mind-set of a deprived actor (see Table 3). To account for economic deprivation, we calculate an asset index as the principal component of several household assets.13 We further control for the respondent’s level of education, by adding a dummy that indicates whether he finished high school. This is the case for 19% of respondents. Next, we control for two variables that proxy for alternative economic opportunities outside the mining sector: a dummy indicating whether the household owns agricultural plots (29% do), and a dummy indicating whether the household has an income source outside the mining sector (16% do). We further consider family responsibilities and community ties (see Table 3). These characteristics are important according to so-called social control theory, which focuses on factors that are likely to produce conformity with norms and laws and may constrain motivated actors from participating in rebellion; such as social attachments and community involvement (see e.g. Thyne and Schroeder 2012). We proxy family responsibilities by age, marital status, and number of children. Nearly half of the miners is married and about one in three lives together with his partner without being married. Most miners (84%) have children and about half of them has three 13 A household was defined as the group of individuals who usually sleep in the same house and share their meals. The calculation of the index included: the ownership of a house and the number of rooms; the quality of the walls and the floor; and the ownership of a mattress, radio and television. 190 Chapter 5 or more children. The miners in our sample are on average 33 years old, with ages ranging from 16 to 65 years. Community ties are captured by ethnicity and place of birth. Just over half of the miners (52%) was born in Kamituga, and the large majority of miners (84%) belong to the ‘Lega’ ethnic group. We also include a dummy that takes the value one for pit managers; this is the case for eight percent of the miners in our sample. Pit managers are often wealthier and well-connected, which may increase the opportunity cost to rebel. On the other hand, they stand to lose most, because many among them have invested several thousands of dollars in the exploration and preparation of a mining pit. As long as the pit has not been in the production phase, they are therefore indebted to money lenders (often gold traders), and have most at stake should artisanal mining be prohibited in Kamituga. Finally, we collected a measure of interview quality. As some questions are somewhat sensitive and others demand cognitive effort, a trusting relation as well as the physical shape of the miner (i.e. not intoxicated by alcohol) can boost the reliability of responses. We therefore asked enumerators to score the interview quality as follows: (1) very bad, (2) bad, (3) reasonable, (4) good, (5) very good. The average interview was given a score of 3.93. 5.6. Empirical strategy and data analysis We test our hypotheses by estimating the following equation: @ @ rstsk' = 34 + u9<7M:>e7F' Α + i:;79<:X <>e7>; , where < indexes the 469 miners and ~ the 9 mining zones. The outcome variables, denoted by @ rstsk' are the four measures of a miner’s intention to rebel presented in Figure. u9<7M:>e7F' , @ @ @ i:;79<:X <>e7>; 191 Chapter 5 exogenous variations, we turn to the procedures developed by Altonji et al. (2005) and refined by Oster (2015) to formally assess omitted variable bias. Finally, different strands of the literature stress different motivations for individual participation in collective violence. Using hypothesis testing techniques for nested models, we empirically investigate if each of the four groups of incentives significantly contributes to the model fit. By introducing interaction terms, we further analyze whether these incentives operate independently, or rather complement each other. 5.6.1. Results The results are presented in Table 5. We find empirical support for each of the four motivations to participate in collective violence. First, grievances matter. The more miners are aggrieved with Banro, the higher their intention to rebel. Specifically, a one-unit increase in grievances with Banro is associated with an increase in the intention to destroy Banro property, attack Banro employees and use fire arms with respectively 9, 8 and 12 percentage points. Grievances with Banro are not significantly related to joining a rebel group however. We find more mixed results for grievances with the local authorities. Grievances with the Chef de Poste are found to significantly increase all four rebellious actions, but grievances with the Mwami are only slightly significantly related to attacking employees, and negatively so. Material incentives are also found to be significant motivators. Miners’ intention to engage in any of the rebellious actions increases with about 1 percentage point for each additional visit of Banro to their mining zone in the month before the interview. Moreover, miners who are interested in working for Banro are 11 to 15 percentage points less likely to display any intention to rebel. The results confirm that non-material, social incentives also have a role to play. Miners operating in a pit with a self-reported ex-rebel are between 11 and 14 percentage points more likely to display an intention to attack employees, use fire arms or join a rebel group. Miners’ intention to rebel further increases if they believe that the actions of Mai-Mai Shikito contributed to the well-being of artisanal mines. Specifically, a one-unit increase in the appreciation of Mai-Mai Shikito increases the intention to engage in any of the rebellious actions with 5 to 8 percentage points. Previous exposure to conflict further significantly increases the intention to rebel. Miners who have been exposed to extreme conflict events between the age of 7 and 20 display a higher 192 Chapter 5 intention to destroy property (10 percentage points), use fire arms (15 percentage points) and join a rebel group (8 percentage points). 5.6.2. Correlation versus causation Given the nature of the data, in particular the lack of exogenous variation in the explanatory variables, the results of the above analyses merely provide evidence of correlations, not of causal relations. The main difficulty to infer causation from these correlations lies with the miners’ unobserved characteristics that can influence both the explanatory and dependent variables, thus causing spurious correlation. This issue is particularly relevant in the case of previous exposure to extreme conflict events and having a rebel network. Victimization only partly relates to random bad luck, and partly to the unobserved war-time decisions and behavior of the men in our sample. Moreover, miners with a larger rebel network may be more likely to have participated in the activities of an armed group themselves, even if they report that they did not.14 Past participation, except in the case of forced recruitment, clearly is a decision variable, and thus highly prone to endogeneity. To formally assess omitted variable bias, we turn to the approach proposed by Altonji et al. (2005) and fine-tuned by Oster (2015). It uses the selection on observable variables as a guide to assess the potential bias from unobserved variables. Put very simply: if adding a battery of relevant observables does not affect our coefficient of interest much, then it is unlikely that there exist many unobservables that would completely cancel out our results. Appendix B provides detailed information on the methodology and reports the results. The findings suggest it is unlikely that the effects on victimization and rebel networks are entirely driven by omitted variable bias.15 5.6.3. Nested models and interactions We now test more formally whether the four families of incentives significantly improve the model fit. To do so, we estimate a range of alternative models for every outcome variable, that include a combination of either one, two or three of the incentive families. The results from likelihood ratio and Wald tests indicate that the comprehensive model, incorporating all four incentive families, significantly outperforms any of the restricted models, except for two cases: exposure to conflict 14 Our measure for rebel networks, i.e. working in a pit with a self-reported ex-rebel, is by construction positively correlated with past participation in the activities of armed groups (we find a correlation coefficient of 0.28, significant at the 99% significance level). 15 It should be noted however, that the value of these findings depends entirely on the unverifiable assumption that unobservables influence selection in a similar way as the observables. 193 Chapter 5 does not significantly matter in explaining the intention to attack Banro employees, while grievances do not significantly matter in explaining the intention to join a rebel group. The above results suggest that all four motivations stressed by different strands of the literature significantly relate to the intention to rebel among artisanal miners in Kamituga. What remains to be addressed is whether these incentives operate independently, or instead work as complements. We verify this by including, one by one, the interaction terms between the eight proxies for grievances, material incentives, social incentives and exposure to conflict. None of the interaction terms were found to be significant, suggesting that – at least in our case study – the four families of incentives operate largely independent of each other. We do find a positive interaction between the two measures of social incentives: having a rebel network and an appreciation of Mai-Mai Shikito mutually reinforce miners’ intention to rebel (see Table 6). 5.6.4. Robustness checks We run a number of checks to gauge the validity and robustness of our findings. First, we estimate equation 1 using different models. Tables C.1 and C.2 in Appendix C report the results when using a probit model and ordinary least squares; they are highly comparable to the results reported in Table 5. Second, we additionally control for individual risk- and time preferences. Risk preferences were elicited using a simple incentive-compatible Eckel and Grossman lottery game (see section 5.4). Time preferences were elicited using a hypothetical game.16 The elicited risk- and time preferences do not turn up significant in any of the model specifications, and including them does not affect the baseline findings (results not reported, but available upon request). 5.7. Discussion Our study of the individual intention to rebel has three unique features. First, the data were collected in a post-war context, which allows us to study not only the usual material and non- material incentives, but also the role of past victimization and rebel networks. Second, given that the presence of Banro is perceived as illegitimate in the eyes of the miners and poses a concrete and imminent threat to their livelihoods, our case study is arguably a worst-case scenario on both 16 In the hypothetical game, we presented miners with two choices between two options. First, they had to choose between receiving 10$ today or receiving 30$ in two weeks. Subsequently, those who chose 10$ today were presented with the choice of receiving 10$ today or 60$ in two weeks; while those who preferred to wait for 30$ in two weeks were presented with the choice of receiving 10$ today and 20$ in two weeks. As such, we derive four categories going from impatient to patient. Category A: 10$>30$, 10$>60$; category B: 10$>30$, 10$<60$; category C: 10$<30$, 10$>20$; and category D: 10$<30$, 10$<20$. 194 Chapter 5 the front of the rational- and deprived actor theories. Finally, in contrast with the existing body of literature on motivations to rebel in eastern DRC that is largely qualitative, we conducted a quantitative analysis based on a representative survey of high-risk youth. While it remains a challenge to distinguish correlations from causations with non-experimental data, we believe that these three features allow us to provide important insights. First, the variables that could uniquely be studied in our set-up, i.e. past victimization and rebel networks, turn out to be fairly strong predictors of the intention to rebel. Furthermore, the self-reported motivations to fight suggest that revenge, a feeling closely related to victimization, plays an important role. These results start to open the black box of the micro-dynamics behind the stylized fact of armed conflict recurrence. The ‘conflict trap’ is unlikely to be driven only by post-war poverty; instead psychological and social processes likely play a role as well. Such processes align well with recent studies showing a positive effect of war exposure on intra-group cohesion and pro-social attitudes. These attitude changes could constitute the mechanism linking up victimization and rebel networks with the intention to fight. Second, in our worst-case scenario, we find that all four groups of incentives significantly contribute to explaining miners’ intention to rebel. Thus, in contrast to most quantitative analyses, we find that miners’ grievances also significantly relate to their intention to engage in rebellious actions. While we find that these different mechanisms operate at the same time, they seem to do so without much interdependencies – as we found no significant interaction terms. Finally, compared to our direct questioning about the motivations to join a rebel group, our quantitative analysis reveals a larger relative importance of material incentives. Indeed, when asked about the motivations to rebel, respondents award most importance to rather laudable motivations such as protecting the community from harm and injustice. This difference in result across both approaches is in line with the general gap in findings between largely quantitative and largely qualitative studies. We speculate that one explanation for the gap lies in social desirability bias. In terms of future research, these insights lead to a plea for more work on how exactly rebel networks and victimization affect the intention to rebel; for more effort to measure (concrete) grievances in surveys; and for experimenting with a mixed methods approach that combines the best of both methodological worlds. Turning back to the context of eastern DRC, it should be highlighted that Kamituga’s mining site is far from unique in its tension between artisanal and industrial mining. Artisanal mining is an important livelihood strategy in DRC. The World Bank (2008: p.7) estimates that up to 10 million people, or 16 percent of DRC’s population, are dependent on artisanal mining. Yet, 195 Chapter 5 their livelihood is under pressure, as the Congolese state prioritizes the development of industrial mining (Stoop et al. 2016). Stoop et al. (mimeo) estimate that approximately 61% of artisanal miners in eastern DRC operate in concessions that have been granted to large-scale mining companies – creating a palpable tension between both modes of production. In terms of policy recommendations to keep eastern DRC safe, our findings support a focus both on the creation of employment and on the fight against impunity. The first of these recommendations is hardly new. The US general in Iraq said in 2006 that finding jobs for ‘‘angry young men’’ was ‘‘absolutely critical to lowering the level of violence’’ (Department of Defence 2006). This recommendation also finds support in a recent study by Blattman and Annan (2016), who experimentally evaluate a program of agricultural training, capital inputs, and counselling directed at ex-fighters in Liberia. The men included in the program showed reduced interest to engage in mercenary activities in neighboring Côte d’Ivoire. In the context of Kamituga and eastern DRC in general, policymakers should safeguard employment in the ASM sector and provide young men losing their livelihood with alternative opportunities. In Stoop et al. (2016) we detail how this can be done according to our respondents. Regarding transitory justice, the DRC has achieved very little. The various peace building interventions have given priority to the disarmament, demobilization, and reintegration (DDR) of fighters, often at the expense of justice. On this, Autesserre (2010: p.141) writes that “high-profile perpetrators of past war crimes not only enjoyed impunity, but were also rewarded with positions of authority”. Apart from impunity overshadowing the DDR programs, there is also a general failure of the national justice system. It is understaffed and underfunded, has little independence from the executive power, operates fees that are unaffordable for most ordinary Congolese, and is prone to bribing. Because of these hurdles, Congolese do not count on justice as administered by the state, and communities feel they need to take their protection in their own hands, which constitutes a disincentive for local militia to disarm (Autesserre 2010). Despite the challenges involved, rebuilding the local and national justice system is absolutely essential, in order to allow communities to overcome resentment over past injustices and human rights violations, end impunity and deter violence. 196 Chapter 5 References Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber. 2005. “Selection on Observed and Unobserved Variables: Assessing the Effectiveness of Catholic Schools.” Journal of Political Economy 113 (1): 151–84. Argo, Nichole. 2009. “Why Fight? Examining Self-Interested Versus Communally-Oriented Motivations in Palestinian Resistance and Rebellion.” Security Studies 18 (4): 651–80. Autesserre, Séverine. 2010. The Trouble with the Congo. Local Violence and the Failure of International Peacebuilding. Cambridge: Cambridge University Press. Bandura, Albert. 1973. Aggression: A Social Learning Analysis. Prentice Hall PTR. Banro. 2014. “2013-2014 Annual Information Form.” Banro Corporation. Bauer, Michal, Alessandra Cassar, Julie Chytilová, and Joseph Henrich. 2014. “War’s Enduring Effects on the Development of Egalitarian Motivations and In-Group Biases.” Psychological Science 25 (1): 47–57. Blattman, Christopher, and Jeannie Annan. 2016. “Can Employment Reduce Lawlessness and Rebellion? A Field Experiment with High-Risk Men in a Fragile State.” American Political Science Review 110 (1): 1–17. Blattman, Christopher, and Edward Miguel. 2010. “Civil War.” Journal of Economic Literature 48 (1): 3–57. Bøås, Morten, and Kevin Dunn. 2014. “Peeling the Onion: Autochthony in North Kivu, DRC.” Peacebuilding 2 (2): 141–56. Bowles, Samuel. 2006. “Group Competition, Reproductive Leveling, and the Evolution of Human Altruism.” Science 314 (5805): 1569–72. Carstens, Johanna, and Gavin Hilson. 2009. “Mining, Grievance and Conflict in Rural Tanzania.” International Development Planning Review 31 (3): 301–26. Choi, Jung-Kyoo, and Samuel Bowles. 2007. “The Coevolution of Parochial Altruism and War.” Science 318 (5850): 636–40. Collier, Paul, and Anke Hoeffler. 1998. “On Economic Causes of Civil War.” Oxford Economic Papers 50 (4): 563–73. Collier, Paul, and Nicholas Sambanis. 2002. “Understanding Civil War: A New Agenda.” The Journal of Conflict Resolution 46 (1): 3–12. Darwin, Charles. 1998. The Descent of Man. Amherst, NY: Prometheus Books. Dave, Chetan, Catherine C. Eckel, Cathleen A. Johnson, and Christian Rojas. 2010. “Eliciting Risk Preferences: When Is Simple Better?” Journal of Risk and Uncertainty 41 (3): 219–43. Department of Defence. 2006. DoD Press Briefing with Lt. Gen. Peter Chiarelli from the Pentagon, U.S. Army Commander, Multinational Corps Iraq. http://www.globalsecurity.org/military/library/news/2006/12/mil-061208-dod01.htm. Dube, Oeindrila, and Juan F. Vargas. 2013. “Commodity Price Shocks and Civil Conflict: Evidence from Colombia.” The Review of Economic Studies 80 (4): 1384–1421. Eckel, Catherine C., and Philip J. Grossman. 2002. “Sex Differences and Statistical Stereotyping in Attitudes Toward Financial Risk.” Evolution and Human Behavior 23 (4): 281–95. Fearon, James D., and David D. Laitin. 2003. “Ethnicity, Insurgency, and Civil War.” American Political Science Review 97 (1): 75–90. Geenen, Sara. 2012. “A Dangerous Bet: The Challenges of Formalizing Artisanal Mining in DRC.” Resources Policy 37 (3): 322–30. ———. 2013. “Dispossession, Displacement and Resistance: Artisanal Miners in a Gold Concession in South-Kivu, Democratic Republic of Congo.” Resources Policy. ———. 2014. “‘Qui Cherche, Trouve’ the Political Economy of Access to Gold Mining and Trade in South Kivu, DRC.” 197 Chapter 5 Geenen, Sara, and Klara Claessens. 2013a. “Disputed Access to the Gold Sites in Luhwindja, Eastern DRC.” The Journal of Modern African Studies 51 (1): 85–108. ———. 2013b. “Disputed Access to the Gold Sites in Luhwindja, Eastern DRC.” Journal of Modern African Studies 51: 85–108. Ginges, Jeremy, and Scott Atran. 2009. “What Motivates Participation in Violent Political Action: Selective Incentives or Parochial Altruism?” Annals of the New York Academy of Sciences 1167 (1): 115–23. Gneezy, Ayelet, and Daniel M. T. Fessler. 2011. “Conflict, Sticks and Carrots: War Increases Prosocial Punishments and Rewards.” Proceedings of the Royal Society B: Biological Sciences. González, Felipe, and Edward Miguel. 2015. “War and Local Collective Action in Sierra Leone: A Comment on the Use of Coefficient Stability Approaches.” Journal of Public Economics 128 (August): 30–33. Gurr, R.T. 1970. Why Men Rebel. Princeton University Press. Harvard Humanitarian Initiative. 2009. “Characterizing Sexual Violence in the Democratic Republic of the Congo. Profiles of Violence, Community Responses, and Implications for the Protection of Women.” Harvard Humanitarian Initiative. Humphreys, Macartan, and Jeremy M. Weinstein. 2008. “Who Fights? The Determinants of Participation in Civil War.” American Journal of Political Science 52 (2): 436–455. IRIN. 2006. “From Protection to Insurgency - History of the Mayi-Mayi.” IRIN. March 16. http://www.irinnews.org/report/58443/drc-protection-insurgency-history-mayi-mayi. Jourdan, L. 2011. “Mayi-Mayi: Young Rebels in Kivu, Drc.” Africa Development 36 (3–4). Kalyvas, Stathis N., and Matthew Adam Kocher. 2007. “How ‘Free’ Is Free Riding in Civil Wars?: Violence, Insurgency, and the Collective Action Problem.” World Politics 59 (2): 177–216. Kilosho, Janvier, Nik Stoop, and Marijke Verpoorten. Forthcoming. “Defusing the Social Minefield of Gold Sites in Kamituga, South Kivu. from Legal Pluralism to the Re- Making of Institutions?” Resources Policy. Kyanga Wasso, Athanase. 2013. “Sominki En Liquidation: Aide-Memoire Sur L’evolution De La Societe À Kamituga.” Bukavu: Sominki. Laudati, Ann. 2013. “Beyond Minerals: Broadening ‘Economies of Violence’ in Eastern Democratic Republic of Congo.” Review of African Political Economy 40 (135): 32–50. Lichbach, Mark Irving. 1989. “An Evaluation of ‘Does Economic Inequality Breed Political Conflict?’ Studies.” World Politics 41 (4): 431–70. ———. 1995. The Rebel’s Dilemma. University of Michigan Press. MacCulloch, Robert. 2004. “The Impact of Income on the Taste for Revolt.” American Journal of Political Science 48 (4): 830–48. Mazalto, M. 2005. “La Réforme Des Législations Minières En Afrique Et Le Rôle Des Institutions Financières Internationales: La République Démocratique Du Congo.” In L’Afrique Des Grands Lacs Annuaire 2004-2005, edited by Stefaan Marysse and Filip Reyntjens, 263–87. Paris: L’Harmattan. ———. 2009. “La Réforme Du Secteur Minier En République Démocratique Du Congo: Enjeux De Gouvernance Et Perspectives De Reconstruction.” Afrique contemporaine 227 (3): 53–80. McKenzie, David. 2012. “Beyond Baseline and Follow-up: The Case for More T in Experiments.” Journal of Development Economics 99 (2): 210–21. Muller, Edward N., Henry A. Dietz, and Steven E. Finkel. 1991. “Discontent and the Expected Utility of Rebellion: The Case of Peru.” American Political Science Review 85 (4): 1261–82. Muller, Edward N., and Karl-Dieter Opp. 1986. “Rational Choice and Rebellious Collective Action.” The American Political Science Review 80 (2): 471–88. Nussio, Enzo. Forthcoming. “The Role of Sensation Seeking in Violent Armed Group Participation.” Terrorism & Political Violence. 198 Chapter 5 Olson, Mancur. 1965. The Logic of Collective Action. Harvard University Press. Oster, Emily. 2015. “Unobservable Selection and Coefficient Stability: Theory and Evidence.” Working Paper. Brown University and National Bureau of Economic Research. Paige, Jeffrey. 1975. Agrarian Revolution: Social Movements and Export Agriculture in the Underdeveloped World. New York: Free Press. Pettersson, Therése, and Peter Wallensteen. 2015. “Armed Conflicts, 1946–2014.” Journal of Peace Research 52 (4): 536–50. Puijenbroeck, J van, and P. Schouten. 2013. “Le 6ième Chantier? L’économie Politique de L’exploitation Auifière Artisanale et Le Sous-Développement En Ituri.” L’Afrique Des Grands Lacs. Annuaire 2012-2013. Radio Okapi. 2007. “Kolwezi: Affrontement Sanglant Entre Creuseurs et Policiers.” Radio Okapi. September 27. http://www.radiookapi.net/sans-categorie/2007/09/27/kolwezi- affrontement-sanglant-entre-creuseurs-et-policiers. ———. 2012. “Maniema : Un Accrochage Entre Exploitants Artisanaux D’or et Policiers Paralyse Les Activités À Salamabila.” Radio Okapi. September 15. http://www.radiookapi.net/actualite/2012/09/15/salamabila-vives-tensions-causees- par-mouvement-de-colere-des-creuseurs-dor. ———. 2014. “Maniema : Les Habitants de Salamabila Demandent À Banro de Respecter Ses Engagements.” Radio Okapi. January 15. http://www.radiookapi.net/regions/maniema/2014/01/15/maniema-les-habitants-de- salamabila-demandent-banro-de-respecter-ses-engagements. ———. 2015. “Bunia: 3 000 Orpailleurs Protestent Contre Leur Délogement Forcé À Mungwalu.” Radio Okapi. July 7. http://www.radiookapi.net/actualite/2015/07/07/bunia-3-000-orpailleurs-protestent- contre-leur-delogement-force-mungwalu. ———. 2016a. “Maniema: Un Activiste Des Droits de L’homme Abattu À Salamabila.” Radio Okapi. January 26. http://www.radiookapi.net/2016/01/26/actualite/societe/maniema- un-activiste-des-droits-de-lhomme-abattu-salamabila. ———. 2016b. “Maniema : Un Député Appelle Banro À Dédommager Les Victimes Délogées de Sa Concession.” Radio Okapi. January 28. http://www.radiookapi.net/2016/01/28/actualite/societe/maniema-un-depute-appelle- banro-dedommager-les-victimes-delogees-de-sa. ———. 2016c. “Kolwezi: Un Mort Dans Des Échauffourées Entre Creuseurs Clandestins et Policiers.” Radio Okapi. May 18. http://www.radiookapi.net/2016/05/19/actualite/societe/kolwezi-un-mort-dans-des- echauffourees-entre-creuseurs-clandestins-et. ———. 2016d. “Lualaba: Près de 10 000 Creuseurs Clandestins Envahissent Le Site de Tenke Fungurume.” Radio Okapi. November 21. http://www.radiookapi.net/2016/11/21/actualite/societe/lualaba-pres-de-10-000- creuseurs-clandestins-envahissent-le-site-de. ———. 2017a. “Kidnappées Depuis Deux Jours, 8 Personnes Relâchées Au Maniema.” Radio Okapi. January 4. http://www.radiookapi.net/2017/01/04/actualite/securite/kidnappees-depuis-deux- jours-8-personnes-relachees-au-maniema. ———. 2017b. “Ituri: Quatre Blessés Lors Des Manifestations Des Creuseurs Artisanaux Contre MGM.” Radio Okapi. February 27. http://www.radiookapi.net/2017/02/27/actualite/societe/ituri-quatre-blesses-lors-des- manifestations-des-creuseurs-artisanaux. ———. 2017c. “Maniema : Sept Employés D’une Entreprise Minière Pris En Otages À Salamabila.” Radio Okapi. March 9. 199 Chapter 5 http://www.radiookapi.net/2017/03/09/actualite/securite/maniema-sept-employes- dune-entreprise-miniere-pris-en-otages. ———. 2017d. “Kindu : Libération D’un Employé de BANRO Sur Les Cinq Pris En Otage.” Radio Okapi. April 20. http://www.radiookapi.net/2017/04/21/actualite/securite/kindu-liberation-dun- employe-de-banro-sur-les-cinq-pris-en-otage. Reyntjens, Filip. 2010. The Great African War: Congo and Regional Geopolitics, 1996 - 2006. Cambridge: Cambridge Univ. Press. Richards, Joanne. 2014. “Forced, Coerced and Voluntary Recruitment into Rebel and Militia Groups in the Democratic Republic of Congo.” The Journal of Modern African Studies 52 (2): 301–26. Scott, James. 1976. The Moral Economy of the Peasant: Rebellion and Subsistence in Southeast Asia. New Haven: Yale University Press. Spittaels, Steven, Ken Matthysen, Yannick Weyns, Filip Hilgert, and Anna Bulzomi. 2014. “Analysis of the Interactive Map of Artisanal Mining Areas in Eastern DR Congo: May 2014 Update.” Antwerp: IPIS. Stearns, Jason. 2011. Dancing in the Glory of Monsters: The Collapse of the Congo and the Great War. New York: Public Affairs. Stearns, Jason, and Christoph Vogel. 2015. “The Landscape of Armed Groups in the Eastern Congo.” Congo Research Group, Center on International Cooperation. Stoop, Nik, Janvier Kilosho, and Marijke Verpoorten. 2016. “Relocation, Reorientation, or Confrontation? Insights from a Representative Survey Among Artisanal Miners in Kamituga, South-Kivu.” IOB Working Paper 2016.09. Institute of Development Policy and Management, University of Antwerp. Stoop, Nik, Marijke Verpoorten, and Peter Van der Windt. mimeo. “Artisanal or Industrial Conflict Minerals? Evidence from Eastern DR Congo.” Thaler, Kai M. 2015. “Mixed Methods Research in the Study of Political and Social Violence and Conflict.” Journal of Mixed Methods Research. Thyne, Clayton L., and Ryan D. Schroeder. 2012. “Social Constraints and Civil War: Bridging the Gap with Criminological Theory.” The Journal of Politics 74 (4): 1066–78. Tullock, Gordon. 1971. “The Paradox of Revolution.” Public Choice 11: 89–99. UN. 2010a. “Democratic Republic of the Congo, 1993-2003. Report of the Mapping Exercise Documenting the Most Serious Violations of Human Rights and International Humanitarian Law Committed within the Territory of the Democratic Republic of the Congo between March 1993 and June 2003.” United Nations. ———. 2010b. “Letter Dated 15 November 2010 from the Chair of the Security Council Committee Established pursuant to Resolution 1533 (2004) Concerning the Democratic Republic of the Congo Addressed to the President of the Security Council.” S/2010/596. United Nations. UNECA. 2003. “Reports on Selected Themes in Natural Resources Development in Africa: Artisanal and Small-Scale Mining and Technology Challenges in Africa.” Addis Ababa, Ethiopia: United Nations Economic Commission for Africa. UNEP. 2011. “Post-Conflict Environmental Assessment of the Democratic Republic of Congo: Synthesis Report for Policy Makers.” Nairobi, Kenya: United Nations Environment Programme. Van Acker, F., and K. Vlassenroot. 2001. “War as Exit from Exclusion? The Formation of Mayi- Mayi Militias in Eastern Congo.” Afrika Focus 17 (1–2): 51–77. Verpoorten, Marijke. 2012. “Leave None to Claim the Land a Malthusian Catastrophe in Rwanda?” Journal of Peace Research 49 (4): 547–63. Verweijen, Judith. 2015. “From Autochthony to Violence? Discursive and Coercive Social Practices of the Mai-Mai in Fizi, Eastern DR Congo.” African Studies Review 58 (2). 200 Chapter 5 Verwimp, Philip. 2005. “An Economic Profile of Peasant Perpetrators of Genocide: Micro-Level Evidence from Rwanda.” Journal of Development Economics 77 (2): 297–323. Vlassenroot, Koen, and Timothy Raeymaekers. 2004. Conflict and Social Transformation in Eastern DR Congo. Gent [Belgium]: Academia Press Scientific Publishers. Voors, Maarten J., Eleonora E.M. Nillesen, Philip Verwimp, Erwin H. Bulte, Robert Lensink, and Daan P. Van Soest. 2012. “Violent Conflict and Behavior: A Field Experiment in Burundi.” The American Economic Review 102 (2): 941–64. Wood, Elisabeth Jean. 2003. Insurgent Collective Action and Civil War in El Salvador. 1st edition. New York: Cambridge University Press. World Bank. 2008. “Democratic Republic of Congo - Growth with Governance in the Mining Sector.” 43402–ZR. World Bank. 201 Chapter 5 Figures Figure 1: Mining in eastern DRC Notes: This figure shows the location of artisanal mining sites registered by the International Peace Information Service (IPIS), as well as Artisanal Exploitation Zones and large-scale mining concessions registered by the Congolese Mining Cadaster (CAMI) in eastern DRC. It further shows the location of Kamituga and Bukavu, as well as the boundaries of the eastern provinces: Ituri, Haut-Uele, North-Kivu, South-Kivu, Maniema, Haut-Katanga and Tanganyika. Source: own compilation based on data from IPIS and CAMI. 202 Chapter 5 Figure 2: Intention to rebel Notes: This Figure illustrates our four measures of ‘intention to rebel’: destroying Banro property, attacking Banro employees, using fire arms, and joining a rebel group. In the empirical analysis, we use dummy variables that equal one if a miner ‘probably’ or ‘certainly’ intends to participate in a rebellious action. 203 Chapter 5 Tables Table 1: Importance of different motivations to participate in the activities of an armed group little important importance refuse to not at all & & total respond important very rather important important N=469 to defend the community against 0.9 13.7 19.6 65.9 100% external aggression to fight injustice, for the rights of the 0.2 12.2 21.3 66.3 100% community to gain respect 0.2 21.5 29.2 49.0 100% personal insecurity 0.9 21.8 28.6 48.8 100% forced participation 0.4 25.0 38.0 36.7 100% to gain money 0.4 36.0 30.3 33.3 100% Other reasons revenge (N=85) 1.2 0.0 18.8 80.0 100% safe haven for bandits (N=26) 0.0 0.0 3.9 96.2 100% lack of employment, poverty (N=20) 5.0 0.0 15.0 80.0 100% Notes: We asked all our respondents the following question: “I will now read several motivations that could explain why people may join an armed group. Could you please indicate how important each of these reasons is in your opinion?”. After this question, respondents were asked whether they could mention other reasons that were important, but not listed. This Table represents the share of responses to the different answer categories. We grouped some answer categories to facilitate interpretation. 204 Chapter 5 Table 2: General and personal intention to rebel Std. Obs. Mean Min Max Dev. destroy Banro property general 469 4.2 1.1 1 5 personal 469 3.1 1.6 0 5 attack Banro employees general 469 4.0 1.2 1 5 personal 469 2.7 1.7 0 5 use fire arms general 469 3.7 1.4 1 5 personal 469 2.4 1.6 0 5 join rebel group general 469 3.6 1.3 1 5 personal 469 2.0 1.5 0 5 Notes: This Table displays summary statistics for the general and personal intention to engage in four rebellious actions against Banro. Answer categories included: refuse to respond (0), certainly not (1), probably not (2), maybe (3), probably (4) and certainly (5). 205 Chapter 5 Table 3: Summary statistics obs. mean st.dev. min. max. Intention to rebel Destroy Banro property 469 0.48 0.50 0 1 Attack Banro employees 469 0.36 0.48 0 1 Use fire arms 469 0.29 0.45 0 1 Join rebels 469 0.19 0.39 0 1 Grievances Grievances with Banro 469 4.6 0.8 1 5 Grievances with Chef de Poste 469 3.1 1.2 1 5 Grievances with Mwami 469 3.3 1.2 1 5 Material incentives Nr. Banro visits to zone last month 469 2.9 4.5 0 31 Interested in working for Banro 469 0.55 0.50 0 1 Social incentives Ex-rebel in pit 469 0.29 0.46 0 1 Favorable to Mai-Mai Shikito 469 2.8 1.0 1 5 Exposure to conflict Extreme exposure to conflict 7-20 469 0.36 0.48 0 1 Control variables Asset index 469 0.0 1.7 -3.3 6.6 Household owns plots 469 0.29 0.45 0 1 Finished high school 469 0.19 0.39 0 1 Household has income source 469 0.16 0.36 0 1 outside artisanal mining Age 469 33 10 16 65 Lives with partner 469 0.79 0.41 0 1 Children younger than 10 in 469 0.70 0.46 0 1 household Was born in Kamituga 469 0.52 0.50 0 1 Belongs to Lega ethnicity 469 0.84 0.37 0 1 Is a pit manager 469 0.08 0.28 0 1 Interview quality 469 3.9 0.6 1 5 206 Chapter 5 Table 4: Overview of hypotheses and their operationalization Why rebel? Hypotheses: The intention to rebel… Operationalization Grievances (H1) increases with the level of grievances with Banro Grievances with Banro (H2) increases with the level of grievances with local Grievances with Chef de Poste authorities Grievances with Mwami Material incentives (H3) increases, the greater the individual material Number of Banro visits to zone benefits are in rebelling against Banro last month (H4) decreases, the higher the individual material costs Interested in working for Banro are in rebelling against Banro Social incentives (H5) increases with exposure to a social norm to rebel Ex-rebel in pit (H6) increases with the perception that similar rebellious Favorable to Mai-Mai Shikito actions have been successful. Conflict exposure (H7) increases with exposure to violent conflict. Exposure to extreme conflict, when aged 7-20 207 Chapter 5 Table 5: Results, marginal effects (1) (2) (3) (4) Destroy Attack Use Join property employees arms rebels Grievances H1 grievances with Banro 0.087** 0.076*** 0.123*** -0.018 (0.036) (0.026) (0.031) (0.027) H2 grievances with Chef de Poste 0.043** 0.043*** 0.047** 0.030*** (0.021) (0.013) (0.018) (0.011) H2 grievances with Mwami -0.037 -0.046** -0.009 -0.003 Material Incentives H3 nr. Banro visits to zone last month 0.014** 0.014*** 0.012*** 0.010*** (0.006) (0.004) (0.002) (0.003) H4 interested in working for Banro -0.150*** -0.152** -0.112** -0.119*** (0.031) (0.062) (0.050) (0.045) Social incentives H5 ex-rebel in pit 0.084 0.114** 0.134*** 0.140*** (0.075) (0.058) (0.047) (0.041) H6 favorable to Mai-Mai Shikito 0.076*** 0.071*** 0.069** 0.048** (0.026) (0.021) (0.028) (0.021) Conflict exposure H7 extreme exposure to conflict 7-20 0.098* 0.056 0.152*** 0.081** (0.052) (0.045) (0.038) (0.036) Control variables asset index 0.010 0.003 -0.007 -0.022*** (0.013) (0.005) (0.010) (0.007) household owns plots -0.102** -0.039 -0.042 -0.040 (0.050) (0.049) (0.032) (0.040) finished high school 0.010 0.006 0.018 -0.017 (0.068) (0.076) (0.078) (0.053) HH has income outside ASM -0.167*** -0.106 -0.044 0.018 (0.051) (0.066) (0.030) (0.034) age -0.003 -0.002 -0.001 -0.001 (0.003) (0.003) (0.003) (0.002) lives with his partner -0.006 -0.020 0.031 0.008 (0.065) (0.074) (0.084) (0.093) children younger than 10 in HH -0.043 -0.081* -0.082 -0.050 (0.044) (0.048) (0.063) (0.065) was born in Kamituga -0.018 0.034 0.077 0.034 (0.047) (0.044) (0.060) (0.043) belongs to Lega ethnicity 0.011 -0.018 -0.085* -0.051 (0.067) (0.057) (0.051) (0.062) is a pit manager 0.014 0.038 0.130 0.046 (0.045) (0.060) (0.098) (0.063) interview quality -0.041 0.036 0.011 0.023 (0.026) (0.033) (0.026) (0.031) mining zone fixed effects Yes Yes Yes Yes clustered s.e. Yes Yes Yes Yes Observations 469 469 469 469 Pseudo R2 0.13 0.14 0.16 0.17 Notes: *** p<0.01, ** p<0.05, * p<0.1; the coefficients represent marginal effects calculated after estimating a probit model; standard errors are reported between brackets and clustered at the level of the mining zones. 208 Chapter 5 Table 6: Interaction between social incentives Destroy Attack Use Join property employees fire arms rebels ex-rebel in pit 0.087 0.116 0.146** 0.151** (0.081) (0.066) (0.054) (0.050) favorable to Mai-Mai Shikito 0.050 0.047* 0.044 0.024 (0.028) (0.024) (0.030) (0.024) interaction 0.084** 0.075** 0.080* 0.084** (0.026) (0.030) (0.035) (0.025) Notes: *** p<0.01, ** p<0.05, * p<0.1; standard errors are clustered at the level of the mining zone and reported between brackets; all specifications include mining zone fixed effects; the appreciation of Mai-Mai Shikito is centered around its mean; the specifications further include all other covariates reported in Table 5. 209 Chapter 5 Appendix A: Acts of rebellion by artisanal miners In this section, we detail acts of rebellion by artisanal miners against large-scale mining companies that occurred in Eastern DRC in the past decade. We start by highlighting incidents that occurred in two of Banro’s concessions (Namoya and Twangiza), and then provide some examples for other mining sites. Banro’s Namoya concession The development of Banro’s Namoya concession, located in the province of Maniema on the border with South-Kivu, has caused numerous acts of rebellion, which have taken a particularly violent turn over the past year. These incidents are reported on by the UN-affiliated Radio Okapi. In September 2012, artisanal miners in Salamabila, located in the territory of Kabambara, organized a protest march against Banro. The demonstrators ransacked two offices and two houses at the Banro site. Local Banro employees were evacuated by helicopter, and brought to Bukavu. The demonstrators would not leave the site, and demanded that Banro compensate them for their incurred losses. Finally, the police intervened, wounding at least three demonstrators with gunfire (Radio Okapi 2012). In January 2014, the population of Salamabila organized another protest march against Banro. The local civil society and customary chiefs argued that Banro was not upholding the promises it had made towards the local community. Specific points mentioned were the rehabilitation of roads and offering employment opportunities to competent local people (Radio Okapi 2014). In January 2016, the leader of a civil society organization was killed by a police officer during another protest march against Banro. The deputy of the local constituency held Banro indirectly responsible for his death: “This dramatic event is caused because Banro did not respect the agreements they made with the local population. This serious situation, which happened today in Salamabila, is caused by the violation of the DRC laws by this company”17 (Radio Okapi 2016a, 2016b). In December 2016, two Banro vehicles where pillaged and set on fire by a group of men armed with guns and machetes. The eight passengers, employees of a Banro subcontracting company, where kidnapped and forced to carry the stolen goods through the forest before being released after two days (Radio Okapi 2017a). The same article mentions that, in 2016, seven 17 Translated from French: “Ce drame est la conséquence du non-Respect par Banro des accords qu’elle a conclus avec la population autochtone. Cet état grave, enregistré aujourd’hui à Salamabila est causé par la violation des lois de la RDC par cette société”. 210 Chapter 5 vehicles belonging to Banro were pillaged and set on fire on the road between Salamabila and Bukavu. Finally, in the beginning of March 2017, seven Banro employees were kidnapped by a group of armed men and kept hostage in the forest (Radio Okapi 2017c). By the end of April 2017, one of the hostages was released, while negotiations are still ongoing to release the others (Radio Okapi 2017d). The hostage takers claim that their land has been unrightfully expropriated by Banro, and demand the restitution of their land to continue their artisanal mining activities. Banro’s Twangiza concession Geenen (2014, 2012) and Geenen and Claessens (2013) have analyzed how Banro’s development of the Twangiza mine has led to the displacement of local communities, and left artisanal miners with few alternative livelihoods. Although Banro set up a community forum – to discuss issues of resettlement and compensation – and organized training and employment programs, Geenen and Claessens (2013) argue that these measures have only benefitted a relatively small part of the affected population and are unlikely to bring relief in the long run. Lacking alternative livelihood options, artisanal miners have resisted the dispossession, both in words and actions. For instance, 500-900 artisanal miners reoccupied sites within the Twangiza concession in April 2011 (Geenen 2013). Regarding this incident, Geenen (2013) cites artisanal miners making the following statements: “Banro had closed down Kaduma and Lukunguri, but we retook them by force. We marched, we barricaded the road. Banro tried to drive us out with policemen and dogs. But we told them that whatever action they took, we would stand firm”. and “They deceive us. They promise us work and the next day they drive us out. If they were planning to stay, they would offer us contracts. We have no prospect of work. So our only option is to reoccupy this concession. They threatened us with policemen and dogs. We told them that whatever they did, but would never die of hunger! […] We would rather be killed by bullets than starve to death”. Examples from other mining sites In September 2007, artisanal miners protested against dams erected by the mining company Gécamines in Kolwezi, Lualaba province. The dams aimed to impede artisanal alluvial mining activities. In a violent confrontation with the police, two miners were hit by gunfire and eleven police officers were injured (Radio Okapi 2007). In May 2016, one person got killed in another 211 Chapter 5 confrontation between the police and artisanal miners who refused to leave the same mining site. According to the artisanal miners, the provincial governor had promised them an alternative artisanal mining site, and they refused to vacate the current site until this promise was met. The miners further ransacked two office buildings of the police, and burnt down two jeeps and a truck belonging to the Kamoto Copper Company (Radio Okapi 2016c). Also in Lualaba province, in November 2016, nearly 10,000 artisanal miners invaded a concession belonging to the mining company Tenke Fungurume Mining for a week. They were reported to have robbed and ransacked property of the company (Radio Okapi 2016d). In July 2015, more than 3,000 artisanal miners protested against their forced removal from an artisanal mining site just north of Bunia in the province of Orientale. The miners blocked roads and refused to leave the site unless the mining company Mungwalu Gold Mining (MGM) compensated them for their losses (Radio Okapi 2015). In February 2017, four artisanal miners were hit by gunfire in a confrontation with the police during a protest march against MGM (Radio Okapi 2017b). 212 Chapter 5 Appendix B: Selection on unobservables To formally assess omitted variable bias, we turn to the approach proposed by Altonji et al. (2005) and fine-tuned by Oster (2015). It uses the selection on observable variables as a guide to assess the potential bias from unobserved variables. Put very simply: if adding a battery of relevant observables does not affect our coefficient of interest much, then it is unlikely that there exist many unobservables that would completely cancel out our results. The selection on observable variables can be evaluated by looking at coefficient movements in the estimates of past victimization and rebel networks while gradually adding additional control variables; their relevance is assessed by the associated movements in the R-squared. Based on these insights, Oster (2015) developed a measure that indicates how large selection on unobservable variables has to be, relative to selection on observables, to fully explain away the estimated effect.18 The larger the measure, denoted by d, the less likely the threat of omitted variable bias. To calculate d, we first run two regressions for each outcome variable: an uncontrolled and a controlled regression. In the uncontrolled regression, we only regress the outcome variable on past victimization or rebel networks. In the controlled regression we control for the remaining observed covariates. Denote the estimated coefficient on past victimization or rebel networks V in the uncontrolled regression and VÄ in the controlled regression; r and rÄ are the R-squared values associated with these regressions. Next, the procedure requires making an assumption about rÅÇÉ, which is defined as the R-squared from a hypothetical regression that controls for all Ä 19 observed and unobserved covariates. We follow Oster (2015) in setting rÅÇÉ = 1.3 r . d is then Ü Ü â Ö (á àá ) 20 calculated as follows: g = â Ü Ü . Oster (2015) argues that a value of d > 1 (i.e. that Ö àÖ (áäãåàá ) selection on observables is at least as important as selection on unobservables) indicates a result that is robust to omitted variable bias. 18 The calculations can be performed with the Stata Code ‘psacalc’, provided by (Oster 2015) and freely available through Stata’s ssc. 19 (Oster 2015) derives this value by analysing coefficient movements for 65 randomized studies, published in five top economic journals between 2008-2013, that provide estimates with and without controls. With a value of rÅÇÉ = 1.3 rÄ, 90% of the evaluated randomized results survived. 20 Consider the intuition behind this expression. We find VÄ in the numerator, indicating that the larger VÄ, the larger the effect that needs to be explained away by selection on unobservables. In the denominator we find (V − VÄ): the smaller the difference between V and VÄ, the less the estimate is affected by selection on observables, and the larger selection on unobservables needs to be, relative to selection on observables, to fully explain away the estimated effect. Ä Ä The strength of the observed covariates increases in (r − r ) and decreases in (rÅÇÉ − r ): the larger the difference between rÄ and r, the more variation in the outcome variable is accounted for by observed covariates; Ä on the other hand, the smaller the difference between rÅÇÉ and r , the more of the "explainable" variation is accounted for by the observed covariates. 213 Chapter 5 The results in Panel A of Table B.1 report the values of d; in all cases we find that d > 1. For victimization we find that selection on unobserved covariates has to be between 13 and 34 times as important as selection on the included covariates to fully explain away the estimates in Table 5. Similarly, for rebel networks, we find that selection on unobserved covariates has to be between 5 and 14 times as important as selection on the included covariates. These findings suggest it is unlikely that the effects on victimization and rebel networks are entirely driven by omitted variable bias. However, when following the procedures outlined in Oster (2015), the assumed values for Réèê are rather low (varying between 0.21 and 0.22). Both Oster (2015) and González and Miguel (2015) argue that Réèê is bounded below one when there is measurement error in the dependent variable. Work by McKenzie (2012) further suggests that measurement errors may be substantial in the context of low income country household datasets; he demonstrates that "for many economic outcomes, the autocorrelations are typically lower than 0.5, with many around 0.3" and that "autocorrelations are often in the 0.2-0.3 range for household income and consumption". In Panel B, we report the highest value of rÅÇÉ for which d ³ 1. Both for victimization and rebel networks, the values of d are still well above the threshold when rÅÇÉ is set to its maximum value of 1. There is only one case where rÅÇÉ does not reach its maximum value when d is set to one: rÅÇÉ equals 0.67 for the effect of rebel networks on attacking employees. This value of rÅÇÉ is however still high in comparison to those reported by McKenzie (2012). 214 Chapter 5 Table B.1: Using selection on observables to assess the bias from unobservables Ä Panel A: Value of d for rÅÇÉ = 1.3 r extreme exposure ex-rebel in pit aged 7-20 d rÅÇÉ d rÅÇÉ Destroy property 13.8 0.22 14.0 0.22 Attack employees / / 5.3 0.22 Use fire arms 13.2 0.22 10.7 0.22 Join rebels 34.0 0.21 13.8 0.21 Panel B: Highest value of rÅÇÉ with d ³ 1 extreme exposure ex-rebel in pit aged 7-20 d rÅÇÉ d rÅÇÉ Destroy property 1.5 1 1.9 1 Attack employees / / 1 0.67 Use fire arms 2.2 1 1.7 1 Join rebels 3.7 1 2.8 1 Notes: d is a measure that indicates how large selection on unobservables needs to be, relative to selection on observables, to fully explain away the estimated effects for past victimization and rebel networks in Table 5. Réèê is the R-squared from a hypothetical regression that controls for all observables and unobservables. As suggested by (Oster ë 2015), we set Réèê = 1.3 R in panel A. In Panel B we report the highest value for Réèê under the condition that d ³ 1. Past victimization is not significantly related to ‘attack employees’, so no values are reported. As ‘ex-rebel in pit’ no longer significantly affects ‘destroy property’ after adding mining zone fixed effects, we report the values for the specification without fixed effects. 215 Chapter 5 Appendix C: Robustness checks Table C.1: The intention to rebel, using a logit model (1) (2) (3) (4) Destroy Attack Use Join property employees arms rebels Grievances H1 grievances with Banro 0.086** 0.076*** 0.125*** -0.021 (0.037) (0.027) (0.035) (0.025) H2 grievances with Chef de Poste 0.044** 0.043*** 0.046** 0.028** (0.022) (0.014) (0.019) (0.011) H2 grievances with Mwami -0.037 -0.046** -0.009 -0.001 (0.025) (0.022) (0.014) (0.016) Material Incentives H3 nr. Banro visits to zone last month 0.014** 0.014*** 0.011*** 0.010*** (0.007) (0.004) (0.002) (0.003) H4 interested in working for Banro -0.146*** -0.146** -0.105** -0.119** (0.031) (0.063) (0.050) (0.047) Social incentives H5 ex-rebel in pit 0.086 0.114* 0.136*** 0.148*** (0.076) (0.060) (0.048) (0.043) H6 favorable to Mai-Mai Shikito 0.076*** 0.070*** 0.068** 0.048** (0.027) (0.021) (0.029) (0.020) Conflict exposure H7 extreme exposure to conflict 7-20 0.096* 0.053 0.144*** 0.082** (0.050) (0.045) (0.036) (0.035) Control variables asset index 0.010 0.004 -0.007 -0.018** (0.013) (0.005) (0.011) (0.007) household owns plots -0.104** -0.039 -0.041 -0.033 (0.049) (0.049) (0.033) (0.039) finished high school 0.008 0.005 0.012 -0.023 (0.069) (0.077) (0.080) (0.054) HH has income outside ASM -0.170*** -0.112 -0.042 0.025 (0.051) (0.069) (0.030) (0.033) age -0.003 -0.002 -0.002 -0.001 (0.002) (0.003) (0.003) (0.002) lives with his partner -0.007 -0.014 0.033 0.007 (0.067) (0.077) (0.089) (0.102) children younger than 10 in HH -0.041 -0.083 -0.082 -0.050 (0.043) (0.051) (0.066) (0.070) was born in Kamituga -0.019 0.033 0.072 0.039 (0.046) (0.044) (0.064) (0.046) belongs to Lega ethnicity 0.017 -0.008 -0.079 -0.044 (0.070) (0.059) (0.056) (0.065) is a pit manager 0.010 0.035 0.136 0.042 (0.045) (0.061) (0.102) (0.064) interview quality -0.040 0.035 0.009 0.019 (0.028) (0.034) (0.028) (0.035) mining zone fixed effects Yes Yes Yes Yes clustered s.e. Yes Yes Yes Yes Observations 469 469 469 469 Pseudo R2 0.13 0.14 0.16 0.17 Notes: *** p<0.01, ** p<0.05, * p<0.1; the coefficients represent marginal effects calculated after estimating a logit model; standard errors are reported between brackets and clustered at the level of the mining zones. 216 Chapter 5 Table C.2: The intention to rebel, using OLS (1) (2) (3) (4) Destroy Attack Use Join property employees arms rebels Grievances H1 grievances with Banro 0.087** 0.072** 0.099*** -0.027 (0.034) (0.022) (0.020) (0.033) H2 grievances with Chef de Poste 0.041* 0.039** 0.040** 0.031** (0.021) (0.014) (0.017) (0.011) H2 grievances with Mwami -0.035 -0.045* -0.007 -0.002 (0.026) (0.023) (0.016) (0.019) Material Incentives H3 nr. Banro visits to zone last month 0.013** 0.014** 0.012*** 0.013*** (0.005) (0.004) (0.002) (0.003) H4 interested in working for Banro -0.153*** -0.150* -0.109* -0.116** (0.032) (0.066) (0.055) (0.046) Social incentives H5 ex-rebel in pit 0.087 0.116 0.146** 0.151** (0.081) (0.065) (0.054) (0.051) H6 favorable to Mai-Mai Shikito 0.077** 0.071** 0.070* 0.052** (0.027) (0.022) (0.032) (0.021) Conflict exposure H7 extreme exposure to conflict 7-20 0.100* 0.056 0.153*** 0.083 (0.053) (0.046) (0.042) (0.045) Control variables asset index 0.011 0.004 -0.005 -0.020** (0.014) (0.006) (0.012) (0.007) household owns plots -0.103* -0.040 -0.059* -0.030 (0.049) (0.051) (0.031) (0.037) finished high school 0.012 0.009 0.009 -0.032 (0.072) (0.081) (0.086) (0.055) HH has income outside ASM -0.172*** -0.114 -0.046 0.023 (0.050) (0.065) (0.026) (0.035) age -0.003 -0.002 -0.002 -0.001 (0.003) (0.003) (0.003) (0.002) lives with his partner -0.010 -0.022 0.033 0.013 (0.066) (0.075) (0.082) (0.103) children younger than 10 in HH -0.040 -0.080 -0.085 -0.063 (0.044) (0.051) (0.067) (0.071) was born in Kamituga -0.016 0.037 0.074 0.036 (0.050) (0.045) (0.062) (0.050) belongs to Lega ethnicity 0.018 -0.010 -0.077 -0.047 (0.072) (0.060) (0.060) (0.076) is a pit manager 0.021 0.052 0.157 0.060 (0.044) (0.062) (0.111) (0.061) interview quality -0.040 0.035 0.010 0.023 (0.028) (0.034) (0.029) (0.036) mining zone fixed effects Yes Yes Yes Yes clustered s.e. Yes Yes Yes Yes Observations 469 469 469 469 Pseudo R2 0.17 0.17 0.17 0.16 Notes: *** p<0.01, ** p<0.05, * p<0.1; all specifications are estimated using OLS; standard errors are reported between brackets and clustered at the level of the mining zones. 217 Chapter 5 218 Chapter 6 6. Conclusion Does Voodoo affect the uptake of preventive healthcare? Do fishermen diversify their income when the fish stock degrades? How do mineral resources relate to armed conflict? Why do individuals take up arms and fight? This thesis consists of four essays that empirically study various dimensions of human development in two developing countries: Benin and the Democratic Republic of Congo. Underlying each essay is a policy-relevant question. This concluding chapter provides a concise overview of the main findings and their policy implications. Does Voodoo affect the uptake of preventive healthcare? Chapter 2 provides the first quantitative analysis to scrutinize the ample ethnographic evidence that magico-religious beliefs affect the demand for conventional healthcare in Sub- Saharan Africa. We rely on the unique case of Benin, where ATR-adherence is freely reported. Its main ATR – Voodoo – is awarded the same status as monotheistic religions, and about 20% of the population reports adherence. The revealed ATR belief and its substantial within-village and within-household variation allowed us to estimate its impact on the uptake of several healthcare measures and outcomes, while controlling for a large set of confounding factors. The data come from four nationally representative rounds of Demographic and Health Surveys, collected between 1996 and 2012. We find that children whose mother is an ATR adherent are significantly more likely not to have received any vaccination, less likely to be fully immunized, less likely to live in a household which owns a bed net and less likely to sleep under a bed net. Overall, the estimated effects are larger than the effect of an additional six years of schooling for the mother or a change from the first to the second household wealth quintile. Mothers’ ATR adherence is further associated with an increased likelihood of testing positive for malaria and a higher under-five mortality rate. Nevertheless, the ATR-health relationship may be spurious, as mothers could self-select into ATR. We employ three different strategies to test for the potential influence of unobservables, including an instrumental variable approach which exploits the fact that present-day ATR- adherence is not merely an individual choice, but is shaped by history and tradition. The combined results suggest that self-selection does not entirely drive our results. Even if the ATR-health relationship is partly spurious, our results are important from a policy perspective as they establish a highly robust correlation. This suggests that the uptake of 219 Chapter 6 preventive healthcare, and ultimately child health outcomes, may be improved by targeting ATR mothers. A tentative exploration of the causal mechanisms suggests a mediating role of traditional healers, who may provide ATR-mothers with off-the-shelve answers on what (not) to do in terms of healthcare. Acknowledging this means directing efforts at building trust in conventional healthcare providers and the health system, and working closely with traditional healers to persuade people. Do fishermen diversify their income when the fish stock degrades? Chapter 3 studies the impact of natural resource degradation on income diversification in Beninese fishing communities. The analysis relies on original survey data, collected from fishing communities in Southern Benin. We find that fishermen are more likely to reallocate labor towards activities outside the fishing sector in areas where natural resource degradation is more severe. The result remains throughout several robustness checks and when controlling for many individual-, household- and village-level covariates. The result also holds when using an instrumental variables approach to deal with the two-way causality between degradation and income diversification. The level of income diversification that we find is however surprisingly low. The fishing communities in our sample remain extremely dependent on the local fish stock, with fishing activities contributing up to 80% of annual income. Such dependency continues to put pressure on the lakes. Without access to attractive outside options, there may be a real danger for these communities to fall into a poverty-environment trap. We find indications that the limited levels of diversification stem from the lack of effective institutions to manage the commons, and the limited access to attractive outside options for a large part of our sample. Indeed, the level of income diversification is especially weak for fishermen that use productive, but highly damaging, fishing gear and for illiterate fishermen. Our findings suggest that policy makers should promote economic activities outside the fisheries sector. At the same time, the use of highly productive but damaging fishing instruments needs to be discouraged by effective regulation and monitoring. Such a two-track policy could enhance the fishermen’s access to attractive outside options, and safeguard fishing communities from a poverty-environment trap. How do mineral resources relate to armed conflict? Existing research suggests a strong link between mining and local conflict, but makes no distinction between artisanal and industrial mining. Chapter 4 conceptualizes and hypothesizes how these extraction modes have very different implications for the level and type of conflict. To test the hypotheses, we rely on a sample of 2,026 artisanal mining sites and 3,700 large-scale mining 220 Chapter 6 concessions in Eastern Congo, and exploit the quasi-exogenous variation in mineral values and the granting of industrial mining concessions. The data show that artisanal and industrial mining have very different implications for conflict. A rise in the value of artisanal mining sites increases battles, violence against civilians and looting. This suggests that the multiple armed actors that are present in ASM sites intensify their fighting efforts over the increased value of the extraction site. The entry of LSM, on the other hand, poses a negative shock – not only to the value of ASM activities, but also for local communities who may have to relocate. Consistent with the fact that this negative shock does not materialize during the LSM research phase, we do not find any effect of LSM research activities on local conflict. However, when LSM moves to the production phase, we find that riots increase. In cases where LSM production activities crowd out ASM, we further find an increase in attacks against civilians and looting. We argue that individuals who previously profited from ASM, as artisanal miners or armed actors, turn to alternative sources of finance (by looting and attacking civilians). Finally, we find that battles decrease as LSM production activities expand. This is consistent with the idea that LSM has the means and incentives to establish a monopoly on power. The results have to be interpreted through a political economy lens. In the case of the DRC, the government is not a neutral bystander. It has chosen to back up the power monopoly of LSM, and not support security around ASM sites. This choice has far-reaching consequences, not only directly and economically for the millions of Congolese whose livelihoods depend on artisanal mining, but also indirectly and politically, in that it changes the distribution of natural resource rents in a direction benefiting already powerful groups. In particular, the promotion of LSM shifts power to the political elite in Kinshasa, away from local elites and artisanal mining communities in eastern DRC. At first sight, our results – in particular the relation between mineral price increases and local conflict at ASM sites – may add to the arguments of those who seek to replace ASM by LSM. However, when, looking through the political economy lens, this conclusion is recognized as short-sighted. First, LSM could fuel the national resource curse by further enriching and empowering a small national elite. Second, the association of ASM and local conflict stems from a conscious policy decision to only secure LSM sites, not ASM sites. Political will aside, there is nothing inherent about ASM sites that prevents the same type of security. In view of these considerations, we end this chapter with the warning that the results of the local resource curse literature cannot be interpreted without taking into account the broader political economy and resource curse at the national level. 221 Chapter 6 Why do individuals take up arms and fight? Several decades of research on armed conflict have yielded relatively few quantitative empirical analyses on the individual propensity to rebel (compared to the large number of ethnographic case studies and cross-country studies). To address this gap, Chapter 5 looks at the intention to rebel in a high-risk population of artisanal miners. The data was collected at the mining site of Kamituga in South-Kivu, Eastern Congo. The large majority of our respondents have been exposed to armed conflict in the past and some have participated in the activities of armed groups. We inquire about their intention to rebel at a time when their main income source is under threat because of the arrival of a large-scale mining company. Clues to the answer to our question can be found in theories on individual participation in collective violence. These theories have highlighted four groups of motivations: grievances, material incentives, social incentives and previous exposure to conflict. Our data provides proxies for each of these motivations, which we relate to miners’ intention to rebel. The associations that emerge allow us to sketch the profile of ‘would-be-rebels’. Of course, ‘would-be-rebels’ are not per se ‘rebels-to-be’. In other words, the self-reported intentions that we pick up could differ from actual decision-making in the future. This discrepancy can result from untruthful answering (because of strategic or social desirability considerations) or because participating in violence is costly and risky, and the effect of the actual costs may only set in when decision time has come. We address both issues and argue that our measures pick up meaningful variation in the intention to rebel. The setting of artisanal miners in Kamituga is one in which all types of incentives to fight are present and about to rise. Indeed, we find that all four groups of incentives significantly contribute to explaining miners’ intention to rebel. In line with Chapter 4, the results suggest that when LSM production activities crowd out ASM, the incidence of looting and violence may increase. Turning back to the context of Eastern Congo, it should be highlighted that Kamituga is far from unique in its tension between artisanal and industrial mining. Artisanal mining is an important livelihood strategy in DRC. The World Bank estimates that up to 10 million people, or 16 percent of DRC’s population, are dependent on artisanal mining. Yet, their livelihood is under pressure, as the Congolese state prioritizes the development of industrial mining. In Chapter 4 we estimate that approximately 61% of artisanal miners in eastern DRC operate in concessions that have been granted to large-scale mining companies – creating a palpable tension between both modes of production. Policymakers should safeguard employment in the ASM sector and provide young men losing their livelihood with alternative opportunities. 222